Recursive Models of Dynamic Linear Economies
Recursive Models of Dynamic Linear Economies
Lars Hansen University of C...

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!

Recursive Models of Dynamic Linear Economies

Recursive Models of Dynamic Linear Economies

Lars Hansen University of Chicago Thomas J. Sargent New York University and Hoover Institution

c

Lars Peter Hansen and Thomas J. Sargent

6 September 2005

Contents

Acknowledgements

xii

Preface

xiii

Part I: Components of an economy

1. Introduction

3

1.1. Introduction. 1.2. Computer Programs. 1.3. Organization.

2. Stochastic Linear Difference Equations

9

2.1. Introduction. 2.2. Notation and Basic Assumptions. 2.3. Prediction Theory. 2.4. Transforming Variables to Uncouple Dynamics. 2.5. Examples. 2.5.1. Deterministic seasonals. 2.5.2. Indeterministic seasonals. 2.5.3. Univariate autoregressive processes. 2.5.4. Vector autoregressions. 2.5.5. Polynomial time trends. 2.5.6. Martingales with drift. 2.5.7. Covariance stationary processes. 2.5.8. Multivariate ARMA processes. 2.5.9. Prediction of a univariate first order ARMA. 2.5.10. Growth. 2.5.11. A rational expectations model. 2.6. The Spectral Density Matrix. 2.7. Computer Examples. 2.7.1. Deterministic seasonal. 2.7.2. Indeterministic seasonal, unit root. 2.7.3. Indeterministic seasonal, no unit root. 2.7.4. First order autoregression. 2.7.5. Second order autoregression. 2.7.6. Growth with homoskedastic noise. 2.7.7. Growth with heteroskedastic noise. 2.7.8. Second order vector autoregression. 2.7.9. A rational expectations model. 2.8. Conclusion.

3. The Economic Environment 3.1. Information. 3.2. Taste and Technology Shocks. 3.3. Technologies. 3.4. Examples of Technologies. 3.4.1. Other technologies. 3.5. Preferences and Household Technologies. 3.6. Examples of Household Technology Preference Structures. 3.7. Constraints to Keep the Solutions “Square Summable”. 3.8. Summary.

–v–

39

vi

Contents

4. Optimal Resource Allocation

57

4.1. Planning problem. 4.2. Lagrange Multipliers. 4.3. Dynamic programming. 4.4. Lagrange multipliers as gradients of value function. 4.5. Planning problem as linear regulator. 4.6. Solutions for five economies. 4.6.1. Preferences. 4.6.2. Technology. 4.6.3. Information. 4.6.4. BrockMirman model. 4.6.5. A growth economy fueled by habit persistence. 4.6.6. Lucas’s pure exchange economy. 4.6.7. An economy with a durable consumption good. 4.7. Hall’s model. 4.8. Higher Adjustment Costs. 4.9. Altered ‘growth condition’. 4.10. A Jones-Manuelli economy. 4.11. Durable consumption goods. 4.12. Summary. A. Synthesizing the linear regulator. B. A Brock-Mirman model. 4.B.1. Uncertainty. 4.B.2. Optimal Stationary States.

5. The Commodity Space

105

5.1. Valuation. 5.2. Price systems as linear functionals. 5.3. A one period model under certainty. 5.4. One period under uncertainty. 5.5. An infinite number of periods and uncertainty. 5.5.1. Conditioning information. 5.6. Lagrange multipliers. 5.7. Summary. A. Appendix.

6. A Competitive Economy

113

6.1. Introduction. 6.2. The Problems of Households and Firms. 6.2.1. Households. 6.2.2. Firms of type I. 6.2.3. Firms of type II. 6.3. Competitive Equilibrium. 6.4. Lagrangians. 6.4.1. Households. 6.4.2. Firms of type I. 6.4.3. Firms of type II. 6.5. Equilibrium Price System. 6.6. Asset Pricing. 6.7. Term Structure of Interest Rates. 6.8. Re-opening Markets. 6.8.1. Recursive price system. 6.8.2. Non-Gaussian asset prices. 6.9. Summary of Pricing Formulas. 6.10. Asset Pricing Example. 6.10.1. Preferences. 6.10.2. Technology. 6.10.3. Information. 6.11. Exercises.

7. Applications 7.1. Introduction. 7.2. Partial Equilibrium Interpretation. 7.2.1. Partial equilibrium investment under uncertainty. 7.3. Introduction. 7.4. A Housing Model. 7.4.1. Demand. 7.4.2. House producers. 7.5. Cattle Cycles. 7.5.1. Mapping cattle farms into our framework. 7.5.2. Preferences. 7.5.3. Technology. 7.6. Models of Occupational Choice and Pay. 7.6.1. A one-occupation model. 7.6.2. Skilled and unskilled workers. 7.7. A Cash-in-Advance Model. 7.7.1. Reinterpreting the household technology. 7.8. Taxation in a Vintage Capital Model. A. Decentralizing the Household.

139

Contents

8. Efficient Computations 8.1. Introduction. 8.2. The Optimal Linear Regulator Problem. 8.3. Transformations to eliminate discounting and cross-products. 8.4. Stability Conditions. 8.5. Invariant Subspace Methods. 8.5.1. P x as Lagrange multiplier. 8.5.2. Invariant subspace methods. 8.5.3. Distorted Economies. 8.5.4. Transition Dynamics. 8.6. The Doubling Algorithm. 8.7. Partitioning the State Vector. 8.8. The Periodic Optimal Linear Regulator. 8.9. A Periodic Doubling Algorithm. 8.9.1. Partitioning the state vector. 8.10. Linear Exponential Quadratic Gaussian Control. 8.10.1. Doubling algorithm. A. Concepts of Linear Control Theory. B. Symplectic Matrices. C. Alternative forms of Riccati equation.

vii

157

viii

Contents

Part II: Representations and Properties

9. Representation and Estimation

187

9.1. The Kalman Filter. 9.2. Innovations Representation. 9.3. Convergence results. 9.3.1. Time-Invariant Innovations Representation. 9.4. Serially Correlated Measurement Errors. 9.5. Combined System. 9.6. Recursive Formulation of Likelihood Function. 9.6.1. Initialization. 9.6.2. Non-existence of a stationary distribution. 9.6.3. Serially correlated measurement errors. 9.7. Wold Representation. 9.8. Vector Autoregression for {yt }. 9.8.1. The factorization identity. 9.8.2. Location of zeros of characteristic polynomial. 9.8.3. Wold and autoregressive representations (white measurement errors). 9.8.4. Serially correlated measurement errors. 9.9. Innovations in yt+1 as Functions of Innovations wt+1 and ηt+1 . 9.10. Innovations in the yt ’s and the wt ’s in a Permanent Income Model. 9.10.1. Preferences. 9.10.2. Technology. 9.10.3. Information. 9.11. Frequency Domain Estimation. 9.12. Approximation Theory. 9.13. Aggregation Over Time. 9.14. Simulation Estimators. A. Initialization of the Kalman Filter.

10. Semiparametric Estimation with Limited Information

227

10.1. Introduction. 10.2. Underlying Economic Model. 10.3. Econometrician’s information and the implied orthogonality conditions. 10.4. An Adjustment Cost Example. 10.5. A Slightly Simpler Estimation Problem. 10.5.1. Scalar Parameterizations of B . 10.6. Multidimensional Parameterizations of B . 10.7. Nonparametric Estimation of B . 10.8. Back to the Adjustment Cost Model.

11. Representation of Demand 11.1. Introduction. 11.2. Canonical Representations of Services. 11.3. Dynamic Demand Functions for Consumption Goods. 11.3.1. The multiplier µw 0 . 11.3.2. Dynamic Demand System. 11.3.3. Foreshadow of Gorman aggregation. 11.4. Computing Canonical Representations. 11.4.1. Heuristics. 11.4.2. An auxiliary problem that induces a canonical representation. 11.5. Operator Identities. 11.6. Becker-Murphy Model of Rational Addiction. A. Fourier transforms. 11.A.1. Primer on transforms. 11.A.2. Time reversal and Parseval’s formula. 11.A.3. One sided

239

Contents

ix

sequences. 11.A.4. Useful properties. 11.A.5. One sided transforms. 11.A.6. Discounting. 11.A.7. Fourier transforms. 11.A.8. Verifying Equivalent Valuations. 11.A.9. Equivalent representations of preferences. 11.A.10. First term: factorization identity. 11.A.11. Second term. 11.A.12. Third term.

12. Gorman Heterogeneous Households

265

12.1. Introduction. 12.2. A Digression on Gorman Aggregation. 12.3. An Economy with Heterogeneous Consumers. 12.4. Allocations. 12.4.1. Consumption sharing rules. 12.5. Risk Sharing Implications. 12.6. Implementing the Allocation Rule with Limited Markets. 12.7. A Computer Example. 12.8. Exercises. 12.8.1. Part one. 12.8.2. Part two. 12.9. Economic integration. 12.9.1. Preferences:. 12.9.2. Technology. 12.9.3. Information.

13. Permanent Income Models

287

13.1. Technology. 13.2. Two Implications. 13.3. Solution. 13.4. Deterministic Steady States. 13.5. Cointegration. 13.6. Constant Marginal Utility of Income. 13.7. Consumption Externalities. 13.8. Tax Smoothing Models.

14. Non-Gorman Heterogeneity Among Households 14.1. Introduction. 14.2. Households’ Preferences. 14.2.1. Technology. 14.3. A Pareto Problem. 14.4. Competitive Equilibrium. 14.4.1. Households. 14.4.2. Firms of type I and II. 14.4.3. Definition of competitive equilibrium. 14.5. Computation of Equilibrium. 14.5.1. Candidate equilibrium prices. 14.5.2. A Negishi algorithm. 14.6. Mongrel Aggregation. 14.6.1. Static demand. 14.6.2. Frequency domain representation of preferences. 14.7. A Programming Problem for Mongrel Aggregation. 14.7.1. Factoring S ′ S . 14.8. Summary of Findings. 14.9. The Mongrel Preference Shock Process. 14.9.1. Interpretation of sˆt component. 14.10. Choice of Initial Conditions.

307

x

Contents

Part III: Extensions

15. Equilibria with Distortions

333

15.1. Introduction. 15.2. A Representative Agent Economy with Distortions. 15.2.1. a. Consumption externalities. 15.2.2. b. Production externalities. 15.2.3. c. Taxes. 15.3. Households. 15.4. Firms. 15.5. Information. 15.6. Equilibrium. 15.7. Heterogeneous Households with Distortions. 15.7.1. Households. 15.7.2. Firms of type I. 15.7.3. Firms of type II. 15.7.4. Government. 15.7.5. Definition of equilibrium. 15.7.6. Equilibrium computation. 15.8. Government Deficits and Debt. 15.9. Examples. 15.9.1. A production externality. 15.9.2. Consumption tax only. 15.9.3. Machinery investment subsidy. 15.9.4. ‘Personal’ habit persistence. 15.9.5. ‘Social’ habit persistence. 15.10. Conclusions. A. Invariant subspace equations for first specification. 15.A.1. Household’s Lagrangian. 15.A.2. Firm’s first order conditions. 15.A.3. Representativeness conditions. B. Invariant subspace equations for heterogeneous agent model.

16. Recursive Risk Sensitive Control

369

16.1. Introduction. 16.2. A Control Problem. 16.3. Pessimistic Interpretation. 16.4. Recursive Preferences. 16.4.1. Endowment economy. 16.5. Asset Pricing. 16.6. Characterizing the Pricing Expectations Operator. 16.7. Production Economies. 16.8. Risk-Sensitive Investment under Uncertainty. 16.9. Equilibrium Prices in the Adjustment Cost Economies.

17. Periodic Models of Seasonality 17.1. Introduction. 17.2. A Periodic Economy. 17.3. Asset Pricing. 17.4. Prediction Theory. 17.5. The Term Structure of Interest Rates. 17.6. Conditional Covariograms. 17.7. The Stacked and Skip-Sampled System. 17.8. Covariances of the Stacked, Skip Sampled Process. 17.9. The Tiao-Grupe Formula. 17.9.1. A state space realization of the TiaoGrupe formulation. 17.10. Some Calculations with a Periodic Hall Model. 17.11. Periodic Innovations Representations for the Periodic Model. A. A Model of Disguised Periodicity. 17.13. A1. Two Illustrations of Disguised Periodicity. 17.14. A2. Mathematical Formulation of Disguised Periodicity.

387

Contents

xi

Part IV: Economies as Objects

18. Introduction to Objects

429

18.1. Matlab Objects. 18.1.1. Definitions. 18.1.2. Matlab Specifics. 18.1.3. How to Define a Matlab Class. 18.2. Summary.

19. Economies as Matlab Objects

435

19.1. Introduction. 19.2. Parent Classes: Information. 19.2.1. Structure. 19.2.2. Functions. 19.3. Parent Classes: Technology. 19.3.1. Structure. 19.3.2. Functions. 19.4. Parent Classes: Preferences. 19.4.1. Structure. 19.4.2. Functions. 19.5. Child Class: Economy. 19.5.1. Structure. 19.5.2. Fields containing the history of the economy. 19.5.3. Functions. 19.5.4. Constructing the object and changing parameters. 19.5.5. Analyzing the economy. 19.6. Working with economies. 19.6.1. The built-in economies. 19.6.2. Mixing and matching built-in parent objects. 19.6.3. Building your own economy. 19.7. Tutorial.

20. MATLAB Programs

445

20.1. Matlab programs.

21. References

499

22. Index

513

23. Author Index

517

24. Matlab Index

519

Acknowledgements

– xii –

Preface

– xiii –

Part I Components of an economy

Chapter 1 Introduction

1.1. Introduction This book views many apparently disparate dynamic economic models as examples of a single class of models that can be adapted and specialized to study diverse economic phenomena. The class of models was created by using recent advances in (i) the theory of recursive dynamic competitive economies; 1 (ii) methods for estimating and interpreting vector autoregression; 2 (iii) linear optimal control theory; 3 and (iv) computer languages for rapidly manipulating linear optimal control systems. 4 We combine these elements to build a class of models for which the competitive equilibria are vector autoregressions that can be swiftly computed, represented, and simulated using the methods of linear optimal control theory. We use the computer language MATLAB to implement the computations. This language has a powerful vocabulary and a convenient structure that liberate time and energy from programming, and thereby spur creative application of linear control theory. Our goal has been to create a class of models that merge recursive economic theory and with dynamic econometrics. Systems of autoregressions and of mixed autogregressive, moving average processes are a dominant setting for dynamic econometrics. We constructed our economic models by adopting a version of recursive competitive theory in which an outcome of theorizing is a vector autoregression. We formulated this class of models because practical difficulties of computing and estimating recursive equilibrium models still limit their use as a tool for thinking about applied problems in economic dynamics. Recursive competitive equilibria were themselves developed as a special case of the Arrow-Debreu competitive equilibrium, both to restrict the range of outcomes possible in the 1 2 3 4

This work is summarized by Harris (1987) and Stokey, Lucas, and Prescott (1989). See Sims (1980), Hansen and Sargent (1980, 1981, 1990). For example, see Kwakernaak and Sivan (1972), and Anderson and Moore (1979). See the MATLAB manual.

–3–

4

Introduction

Arrow-Debreu setting and to create a framework for studying applied problems in dynamic economies of long duration. Relative to the general Arrow-Debreu setting, the great advantage of the recursive competitive equilibrium formulation is that equilibria can be computed by solving a discounted dynamic programming problem. Further, under particular additional conditions, an equilibrium can be represented as a Markov process in the state variables. When that Markov process has an invariant distribution to which the process converges, there exists a vector autoregressive representation. Thus, the theory of recursive competitive equilibria holds out the promise of making closer contact with econometric theory than did previous formulations of equilibrium theory. Two computational difficulties have left much of this promise unrealized. The first is Bellman’s “curse of dimensionality” which usually makes dynamic programming a costly procedure for systems with even small numbers of state variables. The second problem is that after a dynamic program has been solved and the equilibrium Markov process computed, the vector autoregression implied by the theory has to be computed by applying classic projection formulas to a large number of second moments of the stationary distribution associated with that Markov process. Typically, each of these computational problems can be solved only approximately. Good research along a number of lines is now being directed at evaluating alternative ways of making these approximations. 5 The need to make these approximations originates in the fact that for general functional forms for objective functions and constraints, even one iteration on the functional equation of Richard Bellman cannot be performed analytically. It so happens that the functional forms economists would most like to use have been of this general class for which Bellman’s equation cannot be iterated upon analytically. Linear control theory studies the most important special class of problems for which iterations on Bellman’s equation can be performed analytically: problems with a quadratic objective function and a linear transition function. Application of dynamic programming leads to a system of well understood and rapidly solvable equations known as the matrix Riccati equation. The philosophy of this book is to swallow hard and to accept up front as primitive descriptions of tastes, technology, and information specifications that satisfy the assumptions of linear optimal control theory. This approach 5 See Marcet (1989) and Judd (1990). Also see Coleman (1990) and Tauchen (1990).

Organization

5

purchases the ability rapidly to compute equilibria together with a form of equilibrium that is automatically in the form of a vector autoregression. A cost of the approach is that it does not accommodate many specifications that we would like to be able to analyze. The purpose of this book is to display the versatility and tractability of our class of models. Versions of a wide range of models from modern capital theory and asset pricing theory can be represented within our framework. The equilibria of these models can be computed so easily that we hope that the reader will soon be thinking of improvements to our specifications. We provide formulas and software for the reader to experiment.

1.2. Computer Programs In writing this book, we put ourselves under a restriction that we should supply the reader with a computer program that implements every equilibrium concept and mathematical representation that we describe. The programs are written in MATLAB, and are described throughout the book. When a MATLAB program is referred to in the text, we place it in typewriter font. Similarly, all computer code is placed in typewriter font. 6 You will get much more out of this book if you use and modify our programs as you read.

1.3. Organization This book is organized as follows. Chapter 10 describes the first order linear vector stochastic difference equation, and shows how special cases of it are formed by a variety of models of time series processes that have been studied by economists. This difference equation will be used to represent the information flowing to economic agents within our models. It will also be used to represent the equilibrium of the model. Chapter 3 defines an economic environment in terms of the preferences of a representative agent, the technology for producing goods, stochastic processes 6 To run our programs, you will need MATLAB’s Control Toolkit in addition to the basic MATLAB software.

6

Introduction

disturbing preferences and the technology, and the information structure of the economy. The stochastic processes fit into the model introduced in chapter 10, while the preferences, technology, and information structure are specified with an eye toward making the competitive equilibrium one that can be computed by the application of linear control theory. Chapter 4 describes a social planning problem associated with the equilibrium of the model. The problem is formulated in two ways, first as a variational problem using stochastic Lagrange multipliers, and then as a dynamic programming problem. We describe how to compute the solution of the dynamic programming problem using formulas from linear control theory. The solution of the social planning problem is a first order vector stochastic difference equation of the form studied in chapter 10. We also show how to use the value function for the social planning problem to compute the Lagrange multipliers associated with the planning problem. These multipliers are later used in chapter 6 to compute the equilibrium price system. Chapter 5 describes the price system and the commodity space that support a competitive equilibrium. We use a formulation that lets the values that appear in agents’ budget constraints and objective functions be represented as conditional expectations of geometric sums of streams of future “prices” times quantities. Chapter 5 relates these prices to Arrow-Debreu state contingent prices. Chapter 6 describes a decentralized version of our economy, and defines and computes a competitive equilibrium. Competitive equilibrium quantities solve a social planning problem. The price system can be deduced from the stochastic Lagrange multipliers associated with the social planning problem. Chapter 7 describes versions of several dynamic models from the literature that fit easily within our class of models. Chapter 9 describes the links between our theoretical equilibrium and autoregressive representations of time series of observables. We show how to obtain an autoregressive representation for a list of observable variables that are linear functions of the state variables of the model. The autoregressive representation is naturally affiliated with a recursive representation of the likelihood function for the observable variables. In describing how to deduce the autoregressive representation from the parameters determining the equilibrium of the model, and possibly also from parameters of measurement error processes, we are completing a key step needed to permit econometric estimation of the model’s free

Organization

7

parameters. Chapter 9 also treats two other topics intimately related to econometric implementation of the models; aggregation over time, and the theory of approximation of one model by another. Chapter 8 describes fast methods to compute equilibria. We describe how doubling algorithms can speed the computation of expectations of geometric sums of quadratic forms, and help to solve dynamic programming problems. Chapter 11 describes alternative ways to represent demand. It identifies an equivalence class of preference specifications that imply the same demand functions, and characterizes a special subset of them as canonical household preferences. Canonical representations of preferences are useful for describing economies with heterogeneity among household’s preferences. Chapter 12 describes a version of our economy with the type of heterogeneity among households allowed when preferences aggregate in a sense introduced by Terrance Gorman . In this setting, affine Engle curves of common slope prevail and give rise to a representative consumer. This representative consumer is ‘easy to find,’ and from the point of view of equilibrium computation of prices and aggregate quantities, adequately stands in for the household of chapters 3–6. The allocations to individual consumers require additional computations, which this chapter describes. Chapter 13 uses our model of preferences to represent multiple goods versions of permanent income models along the lines of Robert Hall’s (1978). We retain Hall’s specification of the ‘storage’ technology for accumulating physical assets, and also the restriction on the discount factor, depreciation rate, and gross return on capital that delivered to Hall a martingale for the marginal utility of consumption. Adopting Hall’s specification of the storage technology imparts a martingale characterization to the model, but it is hidden away in an ‘index’ whose increments drive the behavior of consumption demands for various goods, which themselves are not martingales. This model forms a convenient laboratory for thinking about the sources in economic theory of ‘unit roots’ and ‘co-integrating vectors.’ Chapter 14 describes a setting in which there is more heterogeneity among households’ preferences, causing the conditions for Gorman aggregation to fail. Households’ Engle curves are still affine, but dispersion of their slopes arrests Gorman aggregation. There is another sense, originating with Negishi, in which there is a representative household whose preferences represent a complicated kind of average over the preferences of different types of households. We show

8

Introduction

how to compute and interpret this preference ordering over economy-wide aggregates. This average preference ordering cannot be computed before one knows the distribution of wealth evaluated at equilibrium prices. Chapter 15 describes economies with production and consumption externalities and also distortions due to a government’s imposing distorting flat rate taxes. Equilibria of these economies has to be computed by a direct attack on Euler equations and budget constraints, rather than via dynamic programming for an artificial social planning problem. Chapter 16 describes a recursive version of Jacobson’s and Whittle’s ‘risk sensitive’ preferences. This preference specification has the features that, although it violates certainty equivalence – so that the conditional covariance of forecast error distributions impinge on equilibrium decision rules – it does so in a way that preserves linear equilibrium laws of motion, and retains calculation of equilibria and asset prices via simple modifications of our standard formulas. These preferences are a version of those studied by Epstein and Zin ( ) and Weil ( ). Chapter 17 describes how to adapt our setup to include features of the periodic models of seasonality that have been studied by Osborne (1988), Todd (1990), and Ghysels (1993). Chapter 20 is a manual of the MATLAB programs that we have prepared to implement the calculations described in this book. The design is consistent with other MATLAB manuals. The notion of duality and the ‘factorization identity’ from recursive linear optimal control theory are used repeatedly in Chapter 9 (on representing equilibria econometrically), and chapters 11, 12, and 14 (on representing and aggregating preferences). ‘Duality’ is the observation that recursive filtering problems (Kalman filtering) have the same mathematical structure as recursive formulations of linear optimal control problems (leading to Riccati equations via dynamic programming). That duality applies so often in our settings in effect ‘halves’ the mathematical apparatus that we require.

Chapter 2 Stochastic Linear Difference Equations

2.1. Introduction This chapter introduces the first-order vector linear stochastic difference equation, which we use in two important ways. We use it first to represent the information flowing to economic agents, then again to represent equilibria of our models. The first-order linear stochastic difference equation is associated with a tidy theory of prediction and a host of procedures for econometric application. Their ease of analysis has prompted us to adopt economic specifications that cause our equilibria to have representations in terms of a first-order linear stochastic difference equation. The first order vector stochastic difference equation is recursive because it expresses next period’s vector of state variables as a linear function of this period’s state vector and a vector of new disturbances to the system. These disturbances form a “martingale difference sequence,” and are the basic building block out of which the time series are created. Martingale difference sequences are easy to forecast, a fact that delivers convenient recursive formulas for optimal predictions.

2.2. Notation and Basic Assumptions Let {xt : t = 1, 2, . . .} be a sequence of n -dimensional random vectors, i.e. an n -dimensional stochastic process. The vector xt contains variables observed by economic agents at time t . Let {wt : t = 1, 2, . . .} be a sequence of N dimensional random vectors. The vectors {wt } will be treated as building blocks for {xt : t = 1, 2, . . .}, in the sense that we shall be able to express xt as the sum of two terms. The first is a moving average of past wt ’s. The second describes the effects of an initial condition. The {wt } process is used to generate a sequence of information sets {Jt : t = 0, 1, . . .}. Let J0 be generated by x0 and Jt be generated by x0 , w1 , . . . , wt , which means that Jt consists of the set

–9–

10

Stochastic Linear Difference Equations

of all measurable functions of {x0 , w1 , . . . , wt }. 1 The building block process is assumed to be a martingale difference sequence adapted to this sequence of information sets. We explain what this means by advancing the following Definition 1: The sequence {wt : t = 1, 2, . . .} is said to be a martingale difference sequence adapted to {Jt : t = 0, 1, . . .} if E(wt+1 |Jt ) = 0 for t = 0, 1, . . . . In addition, we assume that the building block process is conditionally homoskedastic, a phrase whose meaning is conveyed by Definition 2: The sequence {wt : t = 1, 2, . . .} is said to be conditionally ′ homoskedastic if E(wt+1 wt+1 | Jt ) = I for t = 0, 1, . . . . It is an implication of the law of iterated expectations that {wt : t = 1, 2, . . .} is a sequence of (unconditional) mean zero, serially uncorrelated random vectors. 2 In addition, the entries of wt are assumed to be mutually uncorrelated. The process {xt : t = 1, 2, . . .} is constructed recursively using an initial random vector x0 and a time invariant law of motion: xt+1 = Axt + Cwt+1 ,

for t = 0, 1, . . . ,

(2.2.1)

where A is an n by n matrix and C is an n by N matrix. Representation (2.2.1) will be a workhorse in this book. First, we will use (2.2.1) to model the information upon which economic agents base their decisions. Information will consist of variables that drive shocks to preferences and to technologies. Second, we shall specify the economic problems faced by the agents in our models and the economic process through which agents’ decisions 1 The phrase “ J is generated by x ” means that J can be expressed as a measurable 0 0 0 function of x0 . 2 Where φ and φ are information sets with φ ⊂ φ , and x is a random variable, the 1 2 1 2 law of iterated expectations states that E (x | φ1 ) = E (E (x | φ2 ) | φ1 ) . Letting φ1 be the information set corresponding to no observations on any random variables, letting φ2 = Jt , and applying this law to the process {wt } , we obtain

E wt+1 = E E wt+1 | Jt

= E (0) = 0.

Prediction Theory

11

are coordinated (competitive equilibrium) so that the state of the economy has a representation of the form (2.2.1).

2.3. Prediction Theory A tractable theory of prediction is associated with (2.2.1). This theory is used extensively both in computing the equilibrium of the model and in representing that equilibrium in the form of (2.2.1). The optimal forecast of xt+1 given current information is E (xt+1 | Jt ) = Axt ,

(2.3.1)

and the one-step-ahead forecast error is xt+1 − E (xt+1 | Jt ) = Cwt+1 .

(2.3.2)

The covariance matrix of xt+1 conditioned on Jt is just CC ′ : ′

E (xt+1 − E (xt+1 | Jt )) (xt+1 − E (xt+1 | Jt )) = CC ′ .

(2.3.3)

Sometimes we use a nonrecursive expression for xt as a function of x0 , w1 , w2 , . . . , wt . Using (2.2.1) repeatedly, we obtain xt = Axt−1 + Cwt = A2 xt−2 + ACwt−1 + Cwt t−1 i hX Aτ Cwt−τ + At x0 . =

(2.3.4)

τ =0

Representation (2.3.4) is one type of moving-average representation. It expresses {xt : t = 1, 2, . . .} as a linear function of current and past values of the building block process {wt : t = 1, 2, . . .} and an initial condition x0 . 3 3 Slutsky (1937) argued that business cycle fluctuations could be well modelled by moving average processes. Sims (1980) showed that a fruitful way to summarize correlations between time series is to calculate an impulse response function. In chapter 8, we study the relationship between the impulse response functions calculated by Sims (1980) and the impulse response function associated with ( 2.3.4 ).

12

Stochastic Linear Difference Equations

The moving average piece of representation (2.3.4) is often called an impulse response function. An impulse response function depicts the response of current and future values of {xt } to an imposition of a random shock wt . In representation (2.3.4), the impulse response function is given by entries of the vector sequence {Aτ C : τ = 0, 1, . . .}. 4 Shift (2.3.4) forward in time: xt+j =

j−1 X

As Cwt+j−s + Aj xt .

(2.3.5)

s=0

Projecting both sides of (2.3.5) on the information set {x0 , wt , wt−1 , . . . , w1 } gives 5 Et xt+j = Aj xt . (2.3.6) where Et (·) ≡ E[(·) | x0 , wt , wt−1 , . . . , w1 ] = E(·) | Jt , and xt is in Jt . Equation (2.3.6) gives the optimal j step ahead prediction. It is useful to obtain the covariance matrix of the j -step ahead prediction error j−1 X xt+j − Et xt+j = As Cwt−s+j (2.3.7) s=0

We have

E(xt+j − Et xt+j ) (xt+j − Et xt+j )′ =

j−1 X

k=0

′

Ak CC ′ Ak ≡ vj

(2.3.8a)

Note that vj defined in (2.3.8a) can be calculated recursively via v1 = CC ′ vj = CC ′ + Avj−1 A′ ,

j ≥ 2.

(2.3.8b)

The matrix vj is the covariance matrix of the errors in forecasting xt+j on the basis of time t information xt . To decompose these covariances into parts attributable to the individual components of wt , we let iτ be an N -dimensional 4 Given matrices A and C , the impulse response function can be calculated using the MATLAB program dimpulse.m. 5 For an elementary discussion of linear least squares projections, see Sargent (1987b, chapter IX).

Transforming Variables to Uncouple Dynamics

13

column vector of zeroes except in position τ , where there is a one. Define a matrix υj,τ by j−1 X ′ (2.3.8c) Ak Ciτ i′τ C ′ A k . υj,τ = k=0

Note that

PN

′ τ =1 iτ iτ

= I , so that from (2.3.8a) and (2.3.8c) we have N X

υj,τ = υj .

τ =1

Evidently, the matrices {υj,τ , τ = 1, . . . , N } give an orthogonal decomposition of the covariance matrix of j -step ahead prediction errors into the parts attributable to each of the components τ = 1, . . . , N . 6 The “innovation accounting” methods of Sims (1980) are based on (2.3.8). Sims recommends computing the matrices vj,τ in (2.3.8) for a sequence j = 0, 1, 2, . . . . This sequence represents the effects of components of the shock process wt on the covariance of j -step ahead prediction errors for each series in xt .

2.4. Transforming Variables to Uncouple Dynamics A convenient analytical device for the analysis of linear system (2.2.1) is to uncouple the dynamics using the distinct eigenvalues of the matrix A . We use the Jordan decomposition of the matrix A : A = T DT −1 ,

(2.4.1)

where T is a nonsingular matrix and D is a matrix constructed as follows. Recall that the eigenvalues of A are the zeroes of the polynomial det (ζI − A). This polynomial has n zeroes because A is n by n . Not all of these zeroes are necessarily distinct, however. 7 Suppose that there are m ≤ n distinct zeroes 6 For given matrices A and C , the matrices v j,τ and vj are calculated by the MATLAB program evardec.m. 7 In the case in which the eigenvalues of A are distinct, D is taken to be the diagonal matrix whose entries are the eigenvalues and T is the matrix of eigenvectors corresponding to those eigenvalues.

14

Stochastic Linear Difference Equations

of this polynomial, denoted δ1 , δ2 , . . . , δm . For each δj , we construct a matrix Dj that has the same dimension as the number of zeroes of det (ζI − A) that are equal to δj . The diagonal entries of Dj are δj and the entries in the single diagonal row above the main diagonal are all either zero or one. The remaining entries of Dj are zero. Then the matrix D is block diagonal with Dj in the j th diagonal block. Transform the state vector xt as follows: x∗t = T −1 xt .

(2.4.2)

Substituting into (2.2.1), we have that x∗t+1 = Dx∗t + T −1 Cwt+1 .

(2.4.3)

Since D is block diagonal, we can partition x∗t according to the diagonal blocks of D or, equivalently, according to the distinct eigenvalues of A . In the law of motion (2.4.3), partition j of x∗t+1 is linked only to partition j of x∗t . In this sense, the dynamics of system (2.4.3) are uncoupled. To calculate multi-period forecasts and dynamic multipliers, we must raise the matrix A to integer powers (see (2.3.6)). It is straightforward to verify that Aτ = T (Dτ )T −1 .

(2.4.4)

Since D is block diagonal, Dτ is also block diagonal, where block j is just (Dj )τ . The matrix (Dj )τ is upper triangular with δjτ on the diagonal, with all entries of the k th upper right diagonal given by (δj )τ −k τ !/[k!(τ − k)!] for 0 ≤ k ≤ τ,

(2.4.5)

and zeroes elsewhere. Consequently, raising D to an integer power involves raising the eigenvalues to integer powers. Some of the eigenvalues of A may be complex. In this case, it is convenient to use the polar decomposition of the eigenvalues. Write eigenvalue δj in polar form as δj = ρj exp(iθj ) = ρj [cos(θj ) + i sin(θj )]

(2.4.6)

where ρj =| δj |. Then δjτ = (ρj )τ exp(iτ θj ) = (ρj )τ [cos(τ θj ) + i sin(τ θj )].

(2.4.7)

Examples

15

We shall often assume that ρj is less than or equal to one, which rules out instability in the dynamics. Whenever ρj is strictly less than one, the term (ρj )τ decays to zero as τ → ∞ . When θj is different from zero, eigenvalue j induces an oscillatory component with period (2π/ | θj |).

2.5. Examples Next we consider some examples of processes that can be accommodated by (2.2.1).

2.5.1. Deterministic seasonals We use (2.2.1) to represent the model yt = yt−4 . Let n = 4, C = 0, xt = (yt , yt−1 , yt−2 , yt−3 )′ , x0 = (0 0 0 1)′ , 0 1 A= 0 0

0 0 1 0

0 0 0 1

1 0 , 0 0

0 0 C= 0.

(2.5.1)

0

In this case the A matrix has four distinct eigenvalues and the absolute values of each of these eigenvalues is one. Two eigenvalues are real (1, −1) and two eigenvalues are imaginary (i, −i), and so have period four. The resulting sequence {xt : t = 1, 2, . . .} oscillates deterministically with period four. It can be used to model deterministic seasonals in quarterly time series.

16

Stochastic Linear Difference Equations

2.5.2. Indeterministic seasonals We want to use (2.2.1) to represent the model yt = α4 yt−4 + wt ,

(2.5.2)

where wt is a martingale difference sequence and | α4 |≤ 1. We define xt = [yt , yt−1 , yt−2 , yt−3 ]′ , n = 4, 0 1 A= 0 0

0 0 1 0

0 0 0 1

α4 1 0 0 . , C= 0 0 0 0

With these definitions, (2.2.1) represents (2.5.2). This model displays an “indeterministic” seasonal. Realizations of (2.5.2) display recurrent, but aperiodic, seasonal fluctuations.

2.5.3. Univariate autoregressive processes We can use (2.2.1) to represent the model yt = α1 yt−1 + α2 yt−2 + α3 yt−3 + α4 yt−4 + wt ,

(2.5.3)

where wt is a martingale difference sequence. We set n = 4, xt = [yt yt−1 yt−2 yt−3 ]′ , α1 1 A= 0 0

α2 0 1 0

α3 0 0 1

α4 1 0 0 , C = . 0 0 0 0

The matrix A has the form of the companion matrix to the vector [α1 α2 α3 α4 ].

Examples

17

2.5.4. Vector autoregressions Reinterpret (2.5.3) as a vector process in which yt is a (k × 1) vector, αj a (k × k) matrix, and wt a k × 1 martingale difference sequence. Then (2.5.3) is termed a vector autoregression. To map this into (2.2.1), we set n = k · 4, I α1 α2 α3 α4 0 I 0 0 0 , C = A= 0 0 I 0 0 0

0

I

0

0

where I is the (k × k) identity matrix.

2.5.5. Polynomial time trends Let n = 2, x0 = [0 1]′ , and A=

1 0

1 0 , C= . 1 0

(2.5.4)

Notice that D = A in the Jordan decomposition of A . It follows from (2.4.5) that 1 t t . (2.5.5) A = 0 1

Hence xt = (t, 1)′ , so that the first component of xt is a linear time trend and the second component is a constant. It is also possible to use (2.2.1) to represent polynomial trends of any order. For instance, let n = 3, C = 0, x0 = (0, 0, 1)′ , and 1 1 0 A = 0 1 1. (2.5.6) 0 0 1

Again, A = D in the Jordan decomposition of A . It follows from (2.4.5) that 1 t t(t − 1)/2 . At = 0 1 t (2.5.7) 0 0 1

Then x′t = [t(t−1)/2, t, 1], so that xt contains linear and quadratic time trends.

18

Stochastic Linear Difference Equations

2.5.6. Martingales with drift We modify the linear time trend example by Suppose that making C nonzero. 1 t 1 1 t ′ , it follows that and A = N is one and C = [1 0]. Since A = 0 1 0 1 1 τ A C= . (2.5.8) 0 Substituting into the moving-average representation (2.3.4), we obtain (2.25) x1t =

t−1 X

wt−τ + [1 t]x0

τ =0

where x1t is the first entry of xt . The first term on the right-hand side of the preceding equation is a cumulated sum of martingale differences, and is called a martingale, while the second term is a translated linear function of time.

2.5.7. Covariance stationary processes Next we consider specifications of x0 and A which imply that the first two moments of {xt : t = 1, 2, . . .} are replicated over time. Let A satisfy A11 A12 A= , (2.5.9) 0 1 where A11 is an (n − 1) × (n − 1) matrix with eigenvalues that have moduli strictly less than one and A12 is an (n − 1) × 1 column vector. In addition, let C ′ = [C1′ 0]. We partition x′t = [x′1t x′2t ] where x1t has n − 1 entries. It follows from (2.2.1) that x1t+1 = A11 x1t + A12 x2t + C1 wt+1 (2.5.10) x2t+1 = x2t .

(2.5.11)

By construction, the second component, x2t , simply replicates itself over time. For convenience, take x20 = 1 so that x2t = 1 for t = 1, 2, . . . . We can use (2.5.10) to compute the first two moments of x1t . Let µt = Ex1t . Taking unconditional expectations on both sides of (2.5.10) gives µt+1 = A11 µt + A12 .

(2.5.12)

Examples

19

We can solve the nonstochastic difference equation (2.5.12) for the stationary value of µt . Define µ as the stationary value of µt , and substitute µ for µt and µt+1 in (2.5.12). Solving for µ gives µ = (I − A11 )−1 A12 . Therefore, if Ex10 = (I − A11 )−1 A12 ,

(2.5.13)

then Ex1t will be constant over time and equal to the value on the right side of (2.5.13). Further, if the eigenvalues of A11 are less than unity in modulus, then starting from any initial value of µ0 , µt will converge to the stationary value (I − A11 )−1 A12 . Next we use (2.5.10) to compute the unconditional covariances of xt . Subtracting (2.5.12) from (2.5.10) gives (x1t+1 − µt+1 ) = A11 (x1t − µt ) + C1 wt+1

(2.5.14)

From (2.5.14) it follows that (x1t+1 − µt+1 )(x1t+1 − µt+1 )′ = A11 (x1t − µt )(x1t − µt )′ A′11

′ ′ + C1 wt+1 wt+1 C1′ + C1 wt+1 (x1t − µt )′ A′11 + A11 (x1t − µt )wt+1 C1′ .

The law of iterated expectations implies that wt+1 is orthogonal to (x1t − µt ). Therefore, taking expectations on both sides of the above equation gives Vt+1 = A11 Vt A′11 + C1 C1′ , where Vt ≡ E(x1t − µt )(x1t − µt )′ . Evidently, the stationary value V of the covariance matrix Vt must satisfy V = A11 V A′11 + C1 C1′ .

(2.5.15)

It is straightforward to verify that V is a solution of (2.5.15) if and only if V =

∞ X

Aj11 C1 C1′ Aj′ 11 .

(2.5.16)

j=0

The infinite sum (2.5.16) converges under the condition that the eigenvalues of A11 are less in modulus than unity. 8 If the covariance matrix of x10 is V and 8 Equation ( 2.5.15 ) is known as the discrete Lyapunov equation. Given the matrices A 11 and C1 , this equation is solved by the MATLAB program dlyap.m.

20

Stochastic Linear Difference Equations

the mean of x10 is (I − A11 )−1 A12 , then the covariance and mean of x1t remain constant over time. In this case, the process is said to be covariance stationary. If the eigenvalues of A11 are all less than unity in modulus, then Vt → V as t → ∞ , starting from any initial value V0 . From (2.3.8) and (2.5.16), notice that if all of the eigenvalues of A11 are less than unity in modulus, then limj→∞ vj = V . That is, the covariance matrix of j -step ahead forecast errors converges to the unconditional covariance matrix of x as the horizon j goes to infinity. 9 The matrix V can be decomposed according to the contributions of each entry of the process {wt }. Let ιτ be an N -dimensional column vector of zeroes except in position τ , where there is a one. Then N X

ιτ ι′τ .

(2.5.17)

(A11 )j C1 ιτ ι′τ C1′ (A11 )j′

(2.5.18)

I=

τ =1

Define a matrix V˜τ V˜τ ≡

∞ X j=o

We have, by analogy to (2.5.15) and (2.5.16), that V˜τ satisfies V˜τ = A11 V˜τ A′11 + C1 iτ i′τ C1′ . In light of (2.5.17), (2.5.18), and (2.5.16) we have that V =

N X

V˜τ .

(2.5.19)

τ =1

The matrix V˜τ has the interpretation of being the contribution to V of the τ th component of the process {wt : t = 1, 2, . . .}. Hence, (2.5.19) gives a decomposition of the covariance matrix V into the portions attributable to each of the underlying economic shocks. Next, consider the autocovariances of {xt : t = 1, 2, . . .}. From the law of iterated expectations, it follows that E[(x1t+τ − µ)(x1t − µ)′ ] = E{E[(x1t+τ − µ) | Jt ](x1t − µ)′ } = E[Aτ11 (x1t − µ)(x1t − µ)′ ]

=

(2.5.20)

Aτ11 V.

9 The doubling algorithm described in chapter 9 can be used to compute the solution of ( 2.5.15 ) via iterations that approximate ( 2.5.16 ). The algorithm is implemented in the MATLAB programs doublej.m and doublej2.m .

Examples

21

Notice that this expected cross-product or autocovariance does not depend on calendar time but only on the gap τ between the time indices. 10 Independence of means, covariances, and autocovariances from calendar time defines covariance stationary processes. For the particular class of processes we are considering, if the covariance matrix does not depend on calendar time, then none of the autocovariance matrices does.

2.5.8. Multivariate ARMA processes Specification (2.2.1) assumes that xt contains all the information that is available at time t to forecast xt+1 . In many applications, vector time series are modelled as multivariate autoregressive moving-average (ARMA) processes. Let yt be a vector stochastic process. An ARMA process {yt : t = 1, 2, . . .} has a representation of the form: yt = α1 yt−1 + α2 yt−2 + · · · + αk yt−k

+ γ0 wt + γ1 wt−1 + · · · + γk wt−k .

(2.5.21)

where E[wt | yt−1 , yt−2 , · · · yt−k+1 , wt−1 , wt−2 , · · · wt−k+1 ] = 0. The requirement that the same number of lags of y enter (2.5.21) as the number of lags of w is not restrictive because some coefficients can be set to zero. Hence we can think of k as being the greater of the two lag lengths. A representation such as (2.5.21) can be shown to satisfy (2.2.1). To see this, we define

xt =

yt α2 yt−1 + α3 yt−2 · · · + αk yt−k+1 + γ1 wt + γ2 wt−1 · · · + γk−1 wt−k+2 + γk wt−k+1 α3 yt−1 · · · + αk yt−k+2 + γ2 wt · · · + γk−1 wt−k+3 + γk wt−k+2 . .. αk yt−1 + γk−1 wt + γk wt−1 γ k wt (2.5.22)

10 Equation ( 2.5.20 ) shows that the matrix autocovariogram of x (i.e., Γτ ≡ E[(x 1t 1t+τ − µ)(x1t − µ)′ ] taken as a function of τ ) satisfies the nonrandom difference equation Γt+1 = A11 Γt with initial condition Γ0 = V .

22

Stochastic Linear Difference Equations

γ0 γ1 C = .. .

(2.5.23)

γk

and

α1 α2 . A = .. αk 0

I 0 .. .

··· ··· .. .

0 0

··· ···

0 0 .. . I

(2.5.24)

0

It is straightforward to verify that the resulting process {xt : t = 1, 2, . . .} satisfies (2.2.1).

2.5.9. Prediction of a univariate first order ARMA Consider the special case of (2.5.21) yt = α1 yt−1 + γ0 wt + γ1 wt−1

(2.5.25)

where yt is a scalar stochastic process and wt is a scalar white noise. Assume that | α1 |< 1 and that | γ1 /γ0 |< 1. Applying (2.5.22), we define the state xt as xt =

yt . γ1 wt

Applying (2.5.23) and (2.5.24), we have γ0 α1 C= , A= γ1 0

1 . 0

We can apply (2.3.6) to obtain a formula for the optimal j -step ahead prediction of yt . Using (2.3.6) in the present example gives j yt+j α1 α1j−1 yt Et = γ1 wt+j 0 0 γ1 wt which implies that Et yt+j = α1j yt + α1j−1 γ1 wt .

(2.5.26)

Examples

23

We can use (2.5.26) to derive a famous formula of John F. Muth (1960). Assume that the system (2.5.25) has been operating forever, so that the initial time is infinitely far in the past. Then using the lag operator L, express (2.5.25) as (1 − α1 L)yt = (γ0 + γ1 L)wt . Solving for wt gives wt = γ0−1

1 − α L 1 yt , 1 + γγ10 L

which expresses wt as a geometric distributed lag of current and past yt ’s. Substituting this expression for wt into (2.5.26) and rearranging gives Et yt+j = α1j−1

h α1 + 1+

γ1 γ0

γ1 γ0 L

i

yt .

In the limiting case as α1 → 1 from below, this formula becomes Et yt+j =

h 1+ 1+

γ1 i γ0 yt , γ1 γ0 L

(2.5.27)

which is independent of the forecast horizon j . In the limiting case of α1 = 1, it is optimal to forecast yt for any horizon as a geometric distributed lag of past y ’s. This is Muth’s finding that a univariate process whose first difference is a first order moving average is optimally forecast via an “adaptive expectations” scheme (i.e., a geometric distributed lag with the weights adding up to unity).

2.5.10. Growth In much of our analysis, we assume that the eigenvalues of A have absolute values less than or equal to one. We have seen that such a restriction still allows for polynomial growth. Geometric growth can also be accommodated by suitably scaling the state vector. For instance, suppose that {x+ t : t = 1, 2, . . .} satisfies: + + + x+ (2.5.28) t+1 = A xt + Cwt+1 + + + where E(wt+1 | Jt ) = 0 and E[wt+1 (wt+1 )′ | Jt ] = (ε)t I . The positive number ε can be bigger than one. The eigenvalues of A+ are assumed to have absolute 1 values that are less than or equal to ε 2 , an assumption that we make to assure

24

Stochastic Linear Difference Equations

that the matrix A to be defined below has eigenvalues with modulus bounded above by unity. We transform variables as follows: t

xt = (ε)− 2 x+ t

(2.5.29)

t

wt = (ε)− 2 wt+ .

(2.5.30)

The transformed process {wt : t = 1, 2, . . .} is now conditionally homoskedastic as required because E[wt+1 (wt+1 )′ | Jt ] = I . Furthermore, the transformed 1 process {xt : t = 1, 2, . . .} satisfies (2.2.1) with A = ε− 2 A+ . The matrix A now satisfies the restriction that its eigenvalues are bounded in modulus by unity. The original process {x+ t : t = 1, 2, . . .} is allowed to grow over time at a rate of up to .5 log (ε).

2.5.11. A rational expectations model Consider a model in which a variable pt is related to a variable mt via pt = λEt pt+1 + γmt ,

0 0

gt2

kt = δk kt−1 + it

(3.4.1)

, 0 < δk < 1

where d1t is a random endowment of the consumption good at time t , and d2t is a random disturbance to adjustment costs at time t . Given d2t , investment can be increased or decreased only by adjusting the amount of the intermediate good employed. The larger is the parameter φ1 , the higher are adjustment costs. Employment of the intermediate good requires labor input on a one-forone-basis. Physical capital depreciates over time. To capture this technology, we specify Φc =

1 0 0 , Φg = , Φi = , 0 −1 φ1

γ Γ= , ∆k = δk , Θk = 1. 0 We set A22 , C2 and Ud to make (d1t , d2t )′ = dt follow one of the stochastic processes described in chapter 2. This technology embodies a linear quadratic, general equilibrium version of the adjustment-cost technology used in Lucas and Prescott’s [1971] model of investment under uncertainty. Technology 3: Multi-Period Adjustment Costs and “Time to Build” A single consumption good is produced by a single capital good. The capital good can be produced in two ways: a fast and relatively resource-intensive way, and a slow and less resource intensive way. Different amounts of intermediate goods are absorbed in producing investment goods in the fast and the slow ways.

44

The Economic Environment

We model this by positing that there are two capital stocks, two investment goods, and four intermediate goods, and that adjustment costs are larger for the faster investment technology. This technology is represented as ct = γk1t−1 + d1t ,

γ>0

k1t = δk k1t−1 + k2t−1 + i1t

(3.4.2a) ,

0 < δk < 1

k2t = i2t

(3.4.2b) (3.4.2c)

g1t = φ1 (i1t + i2t )

, φ1 > 0

g2t = φ2 (i1t + k2t−1 )

, φ2 > 0

(3.4.2d) (3.4.2e)

g3t = φ3 i1t

, φ3 > 0

(3.4.2f )

g4t = φ4 i2t

, φ4 > 0

(3.4.2g)

ℓ2t = gt · gt

(3.4.2h)

Equation (3.4.2a) describes how physical capital, k1t , and an endowment shock, d1t , are transformed into the consumption good. Equations (3.4.2b ) and (3.4.2c) tell how capital, k1t , can be augmented by “quick investment”, i1t , and by “slow investment”, i2t . Notice that (3.4.2b ) and (3.4.2c) imply that physical capital, k1t , is determined by k1t = δk k1t−1 + i1t + i2t−1 , an equation that exhibits the status of i1t and i2t as ‘fast’ and ‘slow’ investment processes, respectively. Equations (3.4.2d ) and (3.4.2e) describe how the intermediate goods, g1t and g2t , are required to produce investment goods. According to (3.4.2d ) and (3.4.2e), it is as though two stages of production are required to produce capital, the first stage using intermediate good g1t , and the second stage using intermediate good g2t . According to (3.4.2d ) and (3.4.2e), fast investment i1t undergoes both stages of production in the same period t , while slow investment i2t undergoes the first stage described by (3.4.2d ) in period t and the second stage described by (3.4.2e) in period (t + 1). Equations (3.4.2f ) and (3.4.2g ) describe some additional inputs of intermediate goods that are specific to the two types of investment processes. We can set φ3 > φ4 to capture the notion that it is more resource intensive to invest quickly. In equation (3.4.2h ), ‘·’ denotes an inner product.

Examples of Technologies

45

To map this technology into our setup, we set ∆k =

δk 0

1 0

, Θk = I

0 0 0 0 0 φ1 0 0 , Φi = φ2 φ3 −1 0 0 0 −1 γ 0 0 0 Γ = 0 −φ2 0 0 0 0

0 0 1 −1 0 0 Φc = 0 , Φg = 0 −1 0 0 0 0 0 0

0 φ1 0 0

(3.4.3)

φ4

Recall that the matrices Φc , Φg , Φi multiply the vectors ct , [g1t g2t g3t g4t ]′ , and [i1t i2t ]′ , respectively, while Γ multiplies the vector [k1t−1 , k2t−1 ]′ . Again, we set Ud , A22 , C2 to make d1t obey one of the processes described in Chapter 2. This technology captures aspects of those used by Park (1984) and Kydland and Prescott (1982). Technology 4: Growth There are a single consumption good, a single investment good, a single capital good, and no intermediate good. Output obeys ct + it = γkt−1 + dt where dt is a random endowment of output at time t . The motion of capital obeys kt = δk kt−1 + it . To represent this technology, we could set Φc = 1, Φi = 1, Φg = 0, Γ = γ, ∆k = δk , Θk = 1. The reader can verify that this specification of the technology violates assumption 3 ( [Φc Φg ] is singular). To analyze such an economy, we could modify some of our calculations to dispense with assumption 3. An alternative way is

46

The Economic Environment

to approximate the technology with another one that satisfies assumption 3. In particular, assume that ct + it = γkt−1 + d1t gt = φ1 it kt = δk kt−1 + it where φ1 is a very small positive number and d2t ≡ 0. To implement this technology, set γ 1 0 1 , Γ= , Φi = , , Φg = Φc = −φ1 1 0 0

∆k = δk , Θk = 1. 2 This technology can be used to create a model of consumption along the lines of Hall (1978) and Flavin (1981), and a linear quadratic version of a model of capital accumulation along the lines of Cass (1965), Koopmans (1965), and Brock and Mirman (1972) . We shall also use later it to represent aspects of a model of economic growth authored by Jones and Manuelli (1988). Technology 5: Depletable Resource There is a single consumption good, a single investment good, two intermediate goods and one capital stock. The capital stock is the cumulative stock of the resource that has been extracted. We let investment it be the extraction rate, so that kt = kt−1 + it . (3.4.4a) All of the amount extracted is consumed, so that ct = it .

(3.4.4b)

There are two sources of extraction costs. The first, which is coincident with using the first intermediate good g1t , depends on the amount extracted in the current time period g1t = φ1 it . (3.4.4c) 2 In effect, the modification induces investment to be associated with the use of a small (because φ1 ≈ 0 ) amount of intermediate goods, which require labor input. The matrix [Φc Φg ] is now nonsingular, so that assumption 3 is satisfied. When φ1 > 0 , technical conditions are satisfied that are required for the solution of the social planning problem automatically to lie in the space L20 (see Chapters 4 and 5). When φ1 is close to zero, the solution of the social planning problem will closely approximate the solution of the social planning problem for φ1 = 0 , augmented with the restriction that the solution lie in L20 .

Examples of Technologies

47

The second source of extraction costs, captured by the intermediate good g2t , depends on the cumulative amount extracted at period t , which we approximate as (it /2 + kt−1 ): 3 g2t = φ2 (it /2 + kt−1 ).

(3.4.4d)

To represent this technology, we set 0 −1 0 0 1 Φc = 0 , Φg = −1 0 , Φi = φ1 Γ = 0 , −φ2 φ2 /2 0 −1 0

∆k = 1, Θk = 1.

In this technology, we have included no endowment shock process dt , so that we can take Ud = 0, A22 = 0, C2 = 0. It would be possible to modify the technology in various ways to provide a role for an endowment or technology shock. Such a technology was used by Hansen, Epple and Roberds [1985] to study alternative arrangements for an exhaustible resource market. Technology 6: Learning by Doing There is a single consumption good, a single investment good, a single intermediate good, and a single capital stock. The capital stock is interpreted as the cumulative stock of knowledge, the accumulation of which requires expenditure of current output and the intermediate good. Thus, we set ct + it = γ1 kt−1 + dt kt = δk kt−1 + (1 − δk )it

(3.4.5)

Setting Θk = (1 − δk ) makes kt a weighted average of current and past rates of investment. Possession of knowledge (capital) lowers the amount of intermediate goods required to accumulate more knowledge: gt = φit − γ2 kt−1 , where φ ≥ γ2 > 0. 3 We add half the current extraction rate i to k t t−1 to approximate the average amount over the period that has been extracted cumulatively.

48

The Economic Environment

To represent this economy, we set 1 0 1 Φc = , Φg = , Φi = 0 −1 φ γ1 Γ= , ∆k = δk , Θk = (1 − δk ). γ2

(3.4.6)

Technology 7: Fixed Proportions There is a single consumption good, a single capital good, and a single “intermediate good” to be interpreted as labor. Labor and capital are required in fixed proportions, apart from the effects of a random “labor-requirements” shock d2t . The technology requires ct + it = γ1 kt−1 + d1t gt = γ2 kt−1 + d2t gt2 = ℓ2t kt = δk kt−1 + it . Here gt represents employment of labor input. The parameter γ2 determines the nonstochastic part of the capital-labor ratio. To map this technology into our setup, we set 1 1 0 Φc = , Φi = , Φg = , 0 0 1 γ1 Γ= , ∆k = δk , Θk = 1. γ2 Technology 8: Interrelated Factor Demand with Costs of Adjustment To produce output requires physical capital, k1t , and labor, k2t . It is costly to adjust the stock of either factor of production. To adjust capital, the intermediate good g1t must be employed, while to adjust labor, the intermediate good g2t must be employed. To implement this technology, we require k2t = g3t ,

Examples of Technologies

49

which identifies k2t with the direct input of labor. The technology satisfies

k1t−1 c1t + it = [γ1 γ2 ] + d1t k2t−1 k1t = δk k1t−1 + i1t

k2t = k2t−1 + i2t g1t = φ2 i1t g2t = φ3 i2t g3t = k2t . When φ3 < φ2 , it is more costly to adjust capital than labor. To capture this technology, we set δk 0 1 0 ∆k = , Θk = 0 1 0 1 1 1 0 0 0 0 γ1 γ2 0 0 1 0 0 1 0 0 Φc = 0 , Φi = φ2 0 , Φg = 0 −1 0 , Γ = 0 0 0 0 φ3 0 0 −1 0 0 This technology is a version of one used by Mortensen [1973] and Hansen and Sargent [1981].

3.4.1. Other technologies Alternative technologies can be constructed that blend features of two or more of those described here. For instance, multiple-period adjustment costs can be incorporated into the growth technology, while learning by doing can be introduced into one of the adjustment cost technologies. Also versions of these single consumption good technologies can be combined to yield technologies for the production of multiple consumption goods.

50

The Economic Environment

3.5. Preferences and Household Technologies We assume a representative household. We postpone until Chapter 12 discussing ways that heterogeneity among consumers can be accommodated within this assumption. We describe preferences in terms of two elements. First we describe a household technology for accumulating a vector of household capital and for using it to produce a vector of consumption services. Then we specify intertemporal preferences for consumption services in different dates and states of the world. We assume that there is an nh –dimensional vector of household capital stocks ht−1 brought into time t . The vector h−1 is taken as an initial condition. The vectors of consumption goods ct and household capital stocks ht−1 are inputs into the household technology at time t . The outputs of this technology are an ns –dimensional vector of household services st and a new vector of stocks of household capital ht . The relation between inputs and outputs is described by ht = ∆h ht−1 + Θh ct (3.5.1) and st = Λht−1 + Πct .

(3.5.2)

We maintain the following technical assumption: 4 Assumption 5: The absolute values of the eigenvalues of ∆h are less than or equal to one. Preferences are defined over stochastic processes for household services and household inputs into production. These preferences are separable across components of services, across states of the world, and over time. In particular, preferences are described by the quadratic utility functional: ∞ 1 X t −( )E β (st − bt ) · (st − bt ) + (ℓt )2 | J0 2 t=0

, 0 < β < 1.

(3.5.3)

where β is a subjective discount factor. The household services in this economy play the role of characteristics or attributes in the analyses of Gorman (1980) and Lancaster (1966). We can think 4 The purpose of this assumption is to assure that under the equilibrium (optimal) decision rule, the state vector for the economy has a transition matrix that is ‘stable’.

Examples of Household Technology Preference Structures

51

of consumption ct at date t as generating a bundle of consumption services in current and future time periods. Thus, the consumption vector ct generates a vector Πct of consumption services at time t and a vector Λ(∆h )j−1 Θh ct of consumption services at time t+j , for j ≥ 1. In effect, the household technology puts time and component nonseparabilities into the indirect preference ordering for consumption goods induced by (3.5.3). We do not impose nonnegativity constraints on consumption goods.

3.6. Examples of Household Technology Preference Structures We describe five examples of household technology-preference structures. Household Technology 1: Time Separability There is a single consumption good which is identical with the single service. There is no household capital. Preferences are described by ∞ 1 X t − E β (ct − bt )2 + ℓ2t | J0 2 t=0

, 0 0.

P∞ Here the bliss point is in effect bt + λ(1 − δh ) j=0 δhj ct−j−1 , so that the bliss point shifts in response to a moving average of past consumption. Preferences P∞ j in this form require an initial condition for the geometric sum j=0 δh ct−j−1 , which we specify as an initial condition for the ‘stock of household durables,’ h−1 . To implement these preferences, let the household capital stock be ht = δh ht−1 + (1 − δh )ct

, 0 < δh < 1.

This implies that ht = (1 − δh )

t X j=0

δhj ct−j + δht+1 h−1

Let consumption services be st = −λht−1 + ct

, λ > 0.

We can represent the desired preferences by setting Λ = −λ, Π = 1, ∆h = δh , Θh = 1 − δh .

Examples of Household Technology Preference Structures

53

The parameter λ governs the strength of habit persistence. When λ = 0, we recover a version of household technology 1. Household technology-preferences 3 is a version of the model of habit persistence of Ryder and Heal [1973]. Later we shall use this specification to represent aspects of some ideas of Jones and Manuelli [1988]. Household Technology 4: Adjustment Costs There is a single consumption good, a single household capital stock equal to consumption, and two consumption services. We want to represent preferences of the form ∞

1 X t − E β [(ct −b1t )2 + λ2 (ct − ct−1 )2 + ℓ2t ] | J0 2 t=0

(3.6.4)

0 0

4.6.3. Information zt+1

1 0 = 0 .8 0 0

0 0 0 0 zt + 1 0 wt+1 .5 0 1

Ub = [ 30 0 0 ] 5 1 0 Ud = 0 0 0 x0 = [ 5

150

1 0

′

0]

Notice that the information process and the initial condition are specified so that the constant is the third state variable. Notice that we have set the bt process equal to a constant value of 30. There is no random component of the preference shock process. Notice that there is a single nontrivial endowment shock, the second component of dt having been set to zero via the specification of the matrix Ud . The first component of dt has been specified to follow a first order autoregressive process with positive mean. The autoregressive parameter for the endowment process has been set at .8. Notice that the third component of

Solutions for five economies

77

the zt vector is a first order autoregressive process with coefficient .5. However, this component of the zt vector impinges neither on bt nor on dt , given the way that we have specified Ub and Ud . We include the third component of the zt process in case the reader would like to edit one our files, say, to add a random component to the preference shock bt . These specifications of preferences and technology are rich enough to encompass versions of several models that have been popular in the recent macroeconomic literature. The preference specification can accommodate preferences that are quadratic in consumption, as used by Hall [1978]; preferences incorporating habit persistence, as used recently by Becker and Murphy [1988]; and preferences for a durable consumption good, as used by Mankiw [1982]. The technology specification is a version of the one-good ‘growth’ technology of chapter 2, modified to include costs of adjusting capital. 9 We shall initially set the parameters of the technology to satisfy the necessary condition for consumption to be a random walk in Hall’s model, namely, the condition β(γ1 + δk ) = 1. This is also the condition for the ‘growth condition’ of Jones and Manuelli just to be satisfied. For all of the specifications, we set Ub so that bt = 30 for all t . By setting the parameter values of this general model to particular values, we can capture the following models.

4.6.4. Brock-Mirman model Set the preference parameters as λ = 0, π = 1, δh and θh arbitrarily. This makes preferences take the form −.5E

∞ X t=0

β t [(ct − bt )2 + ℓ2t ]|J0 .

Set the technology parameters so that γ1 > 0, φ1 > 0 but φ1 ≈ 0, (γ1 +δk )β = 1.

9 The parameters for our first version of Hall’s economy are in clex11.m; those for our second version of Hall’s economy are in clex12.m; those for our third version of Hall’s economy are in clex13.m; those for the Jones-Manuelli model are in clex10.m; those for the model with durable consumption goods are in clex15.m; and those for Lucas’s economy are in clex14.m.

78

Optimal Resource Allocation

4.6.5. A growth economy fueled by habit persistence Set the technology parameters as in Hall’s model, but set the preference parameters to capture preference specification 3 of chapter 2. In particular, set 1 > δh > 0, θh = (1 − δh ), π = 1, λ = −1. This makes preferences assume the form −.5E

∞ X t=0

β t [(ct − bt − λ(1 − δh )

∞ X

δhj ct−j−1 )2 + ℓ2t ]|J0 .

j=0

4.6.6. Lucas’s pure exchange economy Set preference parameters as in Hall’s model, but alter the technology to render capital unproductive, i.e., set γ1 = 0.

4.6.7. An economy with a durable consumption good Set the technology as in Hall’s model, but alter preferences to capture the idea that the consumption good is durable. Set π = 0, λ > 0, 0 < δh < 1, θh = 1. We now illustrate how the solutions of the social planning problem associated with several of these models can be computed and analyzed. Generally, we proceed as follows. First we read in the parameters that represent our economy by way of the matrices listed in Table 1. We have prepared a set of ‘.m’ files that read in these matrices for the several economies listed above. Thus, clex11.m, clex12.m, and clex13.m are files that read in matrices corresponding to Hall’s model for various different parameter settings. Next, we use solvea.m to compute all of the matrices listed in Table 2, which characterize the solution of the planning problem. To compute the vector ARMA representation of any subset of quantities or Lagrange multipliers, we use aarma.m. To compute the impulse response functions of any set of quantities and/or Lagrange multipliers to components of w(t), we use the program aimpulse.m. Finally, we can use simul.m or asimul.m to simulate the solution of the model.

Hall’s model

79

4.7. Hall’s model We begin with the version of Hall’s model which we solved by hand earlier in this chapter. We begin by setting the parameters in a way that is designed to make consumption follow a random walk. In particular, we set φ1 = .00001, γ1 = .1, δk = .95, β = 1/1.05. Notice that β(γ1 + δk ) = 1. We set the remaining parameters to the values described above. After reading in the matrices by typing clex11, we compute the solution of the planning problem by typing solvea. Issuing this command causes the computer to respond as follows: Calculating, please wait The matrix ao has been calculated for the law of motion of the entire state vector. This matrix satisfies x(t+1) = ao*x(t) + c*w(t+1). The endogenous eigenvalues are in the vector endo, and the exogenous eigenvalues are in the vector exog. The solution to the model is given by c(t) = sc*x(t), g(t) = sg*x(t), h(t) = sh*x(t), i(t) = si*x(t) k(t) = sk*x(t), and s(t) = ss*x(t). The matrices sc, sg, sh, si, sk, and ss have now been computed and can be used in other matlab programs. The matrices sb and sd are constructed so that b(t) = sb*x(t) and d(t) = sd*x(t) and can be used in other matlab programs. The shadow price vectors satisfy Mc(t) = mc*x(t), Mg(t) = mg*x(t), Mh(t) = mh*x(t), Mi(t) = mi*x(t), Mk(t) = mk*x(t), Ms(t) = ms*x(t), and Md(t) = md*x(t). The matrices of these linear combinations can be used in other matlab programs. Your equilibrium has been calculated. You are now ready to experiment with the economy. This is the end of the output that appears on the screen. The solution of the planning problem is stored in the matrices listed in table 2. To inspect these matrices, we just ask MATLAB to show them to us. Thus, issuing the

80

Optimal Resource Allocation

MATLAB command ao results in the output

0.9000 0.0000 ao = 0.0000 0.0000 0.0000

0.0050 1.0000 0.0000 0.0000 0.0000

0.5000 0.0000 1.0000 0.0000 0.0000

0.0000 0.0000 0.0000 0.0000 0.5000

0.0200 0.8000 0.0000 0.8000 0.0000

To see the matrix c, we type c and elicit the response

0.0000 0.0000 c = 0.0000 1.0000 0.0000

0.0000 0.0000 0.0000 0.0000 1.0000

Recall that various quantities in the model are determined by premultiplying the state xt by matrices Sj which are stored by MATLAB in sj. For various purposes, it is useful to create a matrix by stacking various sj’s on top of one another. For example, we can stack the s matrices for consumption, household durables, services, physical investment, and physical capital by issuing the MATLAB command G=[sc;sh;ss;si;sk], which evokes the response

0.0000 0.9000 G = 0.0000 0.0000 0.0000

0.0500 0.0050 0.0500 0.0500 1.0000

5.0000 0.5000 5.0000 0.0000 0.0000

0.2000 0.0200 0.2000 0.8000 0.8000

0.0000 0.0000 0.0000 0.0000 0.0000

The first row of G is Sc , and so on. Similarly, various Lagrange multipliers in the model are determined by premultiplying xt by the matrices Mj , which are stored by MATLAB in mj. We can create a matrix by stacking various mj’s by issuing the command H=[mc;ms;mh;mi;mk], which evokes

0.0000 0.0000 H = 0.0000 0.0000 0.0000

−0.0500 −0.0500 0.0000 −0.0500 −0.0500

25.0000 25.0000 0.0000 25.0000 25.0000

−0.2000 −0.2000 0.0000 −0.2000 −0.2000

0.0000 0.0000 0.0000 0.0000 0.0000

Hall’s model

81

The endogenous and exogenous eigenvalues of Ao or ao are stored in endo and exo, respectively. For the present model, they are given by .90 endo = 1.0 1.00 exo = .80 .50

The exogenous eigenvalue of unity corresponds to the constant (unity) in the state vector, while the other two exogenous eigenvalues are also directly inherited from our specification of the A22 matrix. The endogenous eigenvalue of .9 is inherited from the depreciation factor of .9 which we set for consumer durables, which is irrelevant in Hall’s model because we set λ = 0. This eigenvalue will become relevant below in specifications in which λ 6= 0. The eigenvalue of unity reflects the random walk character of consumption in Hall’s model. Actually, the second endogenous eigenvalue is not really unity, it is only close to unity. To see this, we switch to a long format in MATLAB by typing format long and then we type endo to receive the response 0.90000000000000 endo = 0.99999999999048 The eigenvalue is not exactly unity because of the very small costs of adjusting capital that we have imposed. The fact that the endogenous eigenvalues of this model are below unity means that it possesses a nonstochastic steady state. To compute the steady state, we set nnc=3, which tells the computer that the constant term is the third component of the state vector. Then we type steadst, which causes the steady state to be computed and stored in zs. To compute the steady state value of consumption, just type sc*zs, and so on. For the present model, we obtain 5.0003 0.0061 zs = 1.0000 0.0000 0.0000

The steady state value of consumption is given by sc*zs, which is sc ∗ zs = [ 5.0003 ]

82

Optimal Resource Allocation

The steady state value of investment is given by si*zs, which is si ∗ zs = [ 0.0003 ] For the present model, these stationary steady state values are of little practical value because of the near unit endogenous eigenvalue. It will take very many periods for the effect of the initial conditions to die out in this model, despite the fact that a steady state for the nonstochastic version of the model does exist. We can compute an ARMA representation for the impulse response of any quantities or Lagrange multipliers to a given component of the white noise process wt . We can learn how aarma.m works by typing help aarma, which delivers the response function[num,den]=aarma(ao,c,sy,ii) Creates ARMA Representation for linear recursive equilibrium models. The equilibrium is x(t+1) = ao*x(t) + c*w(t+1) and is created by running SOLVEA. A vector of observables is given by y(t) = sy*x(t) where sy picks off the desired variables. For example, if we want y=[c’,i’], we set sy=[sc;si]. AARMA creates the representation den(L)y(t) = num(L)wi(t) This is an arma representation for the response of y(t) to the i-th component of w(t). For example, to compute the ARMA representation for the impulse response of ct , it to the first component of wt , we type sy=[sc;si] and [num,den]=aarma(ao, c,sy,1) which gives the response num =

0.0000 0.0000

den = [ 1.0000

0.2000 0.8000

−0.6400 −2.6800

0.7540 3.3040

−0.3860 −1.7660

0.0720 0.3420

−4.2000

6.9700

−5.7000

2.2900

−0.3600 ]

Hall’s model

83

This output is to be interpreted as follows. For i = 0, . . . , 5. Define αi as the element in the (i + 1) column of den. For i = 0, . . . 5 define ξi as the 3 × 1 matrix that is the i + 2st column of num. Define two polynomials in the lag operator L by P5 α(L) = i=0 αi Li P5 ξ(L) = i=0 ξi Li Let w1t be the first innovation in the system, which drives the endowment process. Then we have the representation ct α(L) it = ξ(L)w1t mct For example, the first row of this representation is

(1 − 4.2L + 6.97L2 − 5.7L3 + 2.29L4 − .36L5 )ct

= (.2 − .64L + .754L2 − .386L3 + .0072L4 )w1t

We can also create the impulse response function for a list of variables in response to a particular innovation. We shall compute the impulse response function for the two variables, c, i. To accomplish this, we set sy by typing sy = [sc;si]. We set ii at 1 (we want the response to the first innovation), and specify the number of lags we want to perform the calculation for. We want the impulse response out to forty lags, so we specify ni=40. To compute the impulse response, we issue the MATLAB function aimpulse, which has the syntax [z]=aimpulse(ao,c,sy,ii,ni), where sy,ii,ni have the settings just described. 10 The impulse response function is returned in z. In Fig. 4.7.1.a we plot the impulse response functions for this model in response to the first innovation, which is the innovation in the endowment shock. These impulse response functions have shapes that are characteristic of a random walk for consumption and a unit root in capital. For consumption, the impulse response is an open “box” which attains its maximum height immediately. This impulse response is characteristic of a random walk consumption process. For investment, the impulse response has an asymptote. 11 10 The MATLAB program aimpulse.m takes the inputs we have created from the solution of the social planning problem and feeds them into the MATLAB program dimpulse.m, which computes impulse response functions. 11 In actuality, there is really no asymptote for the impulse response function for either consumption or investment, because the largest eigenvalue is just a little bit less than unity.

84

Optimal Resource Allocation

16

1 0.9

14 0.8 * consumption 12

0.7

*

0.6

*

0.5

10 *

investment *

0.4

8 * * *

0.3 0.2 *

*

6

* * * * * * * * * * * * * * * * * * * * * * * * ** ** ** ** * * * * * * * * * * * * * * * * * *

consumption

investment

4

0.1 0

0

5

10

15

20

25

30

35

40

Fig. 4.7.1.a. Impulse response of consumption and investment to an endowment innovation in a version of Hall’s model.

2

0

20

40

60

80

100

120

140

160

Fig. 4.7.1.b. Simulation of consumption and investment for Hall’s model.

We now generate a random simulation of the model for 150 periods. We use the non-interactive program asimul.m to generate this simulation. To use this program we must specify an observer matrix sy that links the called-for variables to the state. Since we want to simulate the four series c, i, k , and the shadow price of consumption, we set sy=[sc;si;sk;mc]. We also have to specify the length of the simulation t1, whether we want a random (k=1) or nonrandom (k=2) simulation, and the initial state vector x0. We want a random simulation of length 150 with the initial condition specified above. After setting these parameters, we execute the simulation by commanding asimul. We obtain the response: Your simulated vector is in the vector ‘‘y’’. We display aspects of this simulation in Fig. 4.7.1.b.The sample paths of c, k , and the shadow price drift in the fashion that random walks do. For paths that are long enough, a random simulation of this model will eventually In fact, the impulse response functions for both consumption and investment are ‘square summable’, but it would take a very long realization of them for this behavior to become apparent.

Higher Adjustment Costs

85

encounter negative values for capital and consumption. The key to rigging samples so that capital and consumption for a long time remain positive with high probability is to select the initial condition for capital large enough and the elements of c2 small enough. Figure 4.7.1.b indicates that investment is relatively more variable than consumption, a pattern that is found in aggregate data for a variety of countries. The fact that this version of Hall’s model, like the stochastic growth model of Brock and Mirman [1972], so easily delivers this pattern is an important feature that has attracted adherents to this and other versions of ‘real business cycle’ theories.

4.8. Higher Adjustment Costs We now turn to a second model which is created by making one modification to the economy we have just studied. The one change we make is to raise the costs associated with adjusting capital. We raise the absolute value of the cost parameter to φ1 = .2. All other parameters remain as in the previous economy. We computed the solution of the social planning problem using solvea.m. The endogenous eigenvalues were computed to be:

endo =

0.9000 0.9966

Notice that relative to the previous economy, one endogenous eigenvalue is left unaltered at .9, while the other endogenous eigenvalue has fallen below unity. The endogenous eigenvalue of .9 is inherited from the law of accumulation that we posit for household capital (which in this model is again irrelevant). The drop below unity of the second endogenous eigenvalue is the result of our having increased the costs of adjusting capital. The analysis that we performed on pages BLANK indicates that this is exactly what should occur when adjustment costs increase. Figure 4.8.1.a reports impulse response functions for the response of ct and it and to an innovation in the endowment process. Notice how these no longer have the tell tale signs of the presence of an endogenous unit eigenvalue. The

86

Optimal Resource Allocation

1

14

0.9

12 consumption

0.8 *

10

0.7 *

8

0.6 *

0.5

6

*

investment *

0.4

4

* *

0.3 0.2

*

investment

*

2

* * * * * * * * * * * * * * * ** * * * * * * * * * * * * * * * * * * ** ** ** ** ** ** ** ** * * * * * * * * * * * * * * * * consumption

0

0.1 0

0

5

10

15

20

25

30

35

40

Fig. 4.8.1.a. Impulse response of consumption and investment to an endowment innovation in a version of Hall’s model with higher costs of adjusting capital and no random walk in consumption.

-2 0

20

40

60

80

100

120

140

160

Fig. 4.8.1.b. Simulation of a version of Hall’s model with higher costs of adjusting capital and no random walk in consumption.

impulse response for consumption and investment now both appear to be convergent and ‘square summable’. Figure 4.8.1.b shows a random simulation beginning from the same value for x0 used with the earlier version of Hall’s model. Notice how consumption, while still smoother than income, has increased high frequency volatility relative to that depicted in figure 4.7.1.a, while the high frequency volatility of investment has decreased. This pattern is a response to the higher costs for adjusting capital. Notice also that there seems to be a downward ‘trend’ in both consumption and investment. This is a consequence of the decrease in the largest endogenous eigenvalue from being very nearly one in the earlier economy. The present economy has a nonstochastic steady state value for capital of .0000, for consumption of 5.00 (which is the mean of the endowment process), and for investment of .0000, each of which we computed using steadst.m. These nonstochastic steady state values correspond to the unconditional means from the asymptotic stationary distribution of our variables. Because the largest endogenous eigenvalue for this economy is .9966 rather than

Altered ‘growth condition’

87

.9999, the economy is headed toward these mean values much more rapidly than for our previous economy.

4.9. Altered ‘growth condition’ We generate our next economy by making two alterations in the preceding economy. First, we raise the adjustment cost parameter from .2 to 1. This will have the effect of further lowering the endogenous eigenvalue that is not .9, and of causing the impulse response functions to dampen faster than they did in the previous economy. Second, we raise the production function parameter from .1 to .15. This will have the effect of raising the optimal stationary value of capital to a positive value for the nonstochastic version of the model. Recall that the optimal stationary value of capital was zero in the previous economy. The nonstochastic steady state values of consumption, investment, and capital are 17.5, 6.25, and 125, respectively, for this economy. The endogenous eigenvalues are

0.9000 endo = 0.9524

We also created the impulse response function for c and i, which is reported in figure 4.9.1.a. Notice the much faster rate of damping relative to the impulse responses displayed for the previous economies. Figure 4.9.1.b displays a random simulation of this economy. Notice that the “transient” behavior displayed by our simulation of the previous economy is not present here. This is a consequence of our having altered the production function parameter value to induce a positive optimal stationary value for the capital stock of 125, and from our having started the simulation at an initial condition of 125 for the capital stock.

88

Optimal Resource Allocation

1

22

0.9

20

0.8 0.7

*

16 *

0.6

14

*

0.5

*

0.4

12 *

consumption

*

0.3

10

* *

*

* *

0.2

* * *

0.1 0

consumption

18

0

*

* * * * * * * * * * * * * * * * * * * investment * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *

5

10

15

20

25

30

35

8

investment

6 4

40

Fig. 4.9.1.a. Impulse response of consumption and investment to an endowment innovation in a version of Hall’s model with higher adjustment costs and the ‘growth condition’ altered.

0

20

40

60

80

100

120

140

160

Fig. 4.9.1.a. Simulation of consumption and investment in a version of Hall’s model with higher adjustment costs and the ‘growth condition’ altered.

4.10. A Jones-Manuelli economy A notable feature of the models for the previous simulations is that consumption, investment, and capital generally failed to grow. We now define the matrices and set parameters with a view toward attaining a version of Jones and Manuelli’s model of economic growth. We set the parameters of the technology so that Jones and Manuelli’s “growth condition” is just satisfied. 12 Our version of Jones and Manuelli’s model has the feature that their growth condition is a necessary but not a sufficient condition for growth to occur. Their growth condition makes sustained growth feasible in our model. In order for growth to occur, it is also necessary that it be desirable, a condition that is determined by the preference parameters λ , δh , and θh . We set these parameters in order to generate growth. 12 The Jones-Manuelli growth condition on the technology in our notation is β(γ + δ ) ≡ 1 . k This is also a condition that makes the marginal utility of consumption follow a martingale in Hall’s model.

A Jones-Manuelli economy

89

In particular, setting λ equal to minus one turns out to generate a preference for growth. 13 As usual, we compute the equilibrium by using asolve.m. For this model, the endogenous eigenvalues are

1.0000 + 0.0000i endo = 1.0000 − 0.0000i

The exogenous eigenvalue of unity is inherited from the law of motion of the unit vector, which is the third state variable. Notice that there are two unit endogenous eigenvalues. With some experimentation, the reader can determine how these two unit endogenous eigenvalues result from specifying the parameters of technology to obey the growth condition, and the parameters of preferences (especially λ ) to capture a longing for consumption growth. 14 Figure 4.10.1.a displays impulse responses of consumption and investment to an innovation in the endowment process. For both consumption and investment, the effect of an innovation actually grows indefinitely over time. This is a product of the second unit endogenous eigenvalue that is inherited from the preference parameter λ . Figure 4.10.1.b displays a simulation of consumption and investment for this economy. The economy grows. Notice that consumption is much smoother than investment. Notice also that investment typically exceeds consumption. In order to support the ‘habit’ that fuels growth, the economy has to accumulate physical capital. 15 We invite the reader to experiment with this economy by altering the settings of some parameter values one at a time relative to the parameter settings that we have made. In particular, we recommend that the following experiments be tried: 13 The parameter values for this economy are stored in clex10.m 14 One unit endogenous eigenvalue stems from setting β, Γ , and ∆ at the boundary of the k Jones-Manuelli growth condition. The other unit endogenous eigenvalue results from setting λ = −1 . The presence of very small positive adjustment costs for capital is what prevents these two endogenous eigenvalues from being exactly unity. The reader can check that they are not exactly unity by using the format long command in MATLAB. 15 It is a feature of models of addiction based on the type of preference specification used here, e.g., Becker and Murphy [1988], that ‘addicts’ grow wealthier and wealthier over time as they follow a consumption plan that allows for enough accumulation to support their growing addiction.

90

Optimal Resource Allocation

70

1 0.9

*

60 *

0.8

*

0.7

50 *

0.6

*

0.5

investment * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *

40 investment 30

0.4 0.3 0.2 0.1 0

* * * * * * *

0

5

consumption

* * * * * * * * * * * * * * * * * * consumption * * * * * * * * * * * * * *

10

15

20

25

30

35

20

10

40

Fig. 4.10.1.a. Impulse response of consumption and investment to an endowment innovation in a Jones-Manuelli economy.

0

0

20

40

60

80

100

120

140

160

Fig. 4.10.1.b. Simulation of consumption and investment in a Jones-Manuelli economy.

1. Change the value of λ to −.7, leaving the other parameters unaltered. Obtain the solution of the planning problem, and inspect the endogenous eigenvalues. Also compute the impulse response function and simulate the model in response to the same initial condition that we used above. Does the economy still grow? Explain. 2. Change the value of β to .94. Recompute the solution of the planning problem. Does the economy grow? Link your explanation to the JonesManuelli growth condition. 3. Change the value of Γ(1) to .09. Does the economy still grow?

Durable consumption goods

91

4.11. Durable consumption goods For our next example economy, we restore the productivity of capital to a value of .1 and raise the level of the parameter measuring adjustment costs for capital to a value of 1. We change the specification of preferences to make the consumption good durable. In particular, we adopt a version of preference specification 2. We implement this by setting λ equal to .1, π equal to zero, and θh equal to one. We leave δh at the value .9. 16

1.2

20 18

1*

consumption 16

0.8

*

14 *

0.6

12

*

consumption

0.4

*

10

* *

0.2

*

8

*

0

* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * ** ** ** ** ** ** ** ** ** ** ** ** ** * * investment

-0.2 0

5

10

15

20

25

30

35

investment

6

40

Fig. 4.11.1.a. Impulse response of consumption and investment to an endowment innovation in an economy with a durable consumption good.

4

0

20

40

60

80

100

120

140

160

Fig. 4.11.1.b. Simulation of consumption and investment in an economy with a durable consumption good.

Figure 4.11.1.a displays the impulse response functions to an innovation in the endowment process. The impulse response function for consumption and for investment are very different than for our first model. In particular, from the impulse response function, we can see that in choosing consumption, the social planner ‘smooths’ the endowment shock much less than he does in Hall’s original model, in which the planner in effect makes consumption an equal-weight moving average of current and lagged innovations to the endowment process. In the 16 These parameters settings are created by the file clex15.m.

Optimal Resource Allocation

92

present model, the planner makes consumption a much shorter, more peaked moving average of the endowment process. This shows up in the simulation of consumption and investment, which is reported in figure 4.11.1.b. Notice that now, in contrast to Hall’s model, it is investment that is much smoother than consumption. This example thus illustrates how making consumption goods durable tends to undo the strong consumption smoothing result which Hall obtained.

4.12. Summary In this chapter, we have formulated a planning problem, and described how to compute its solution. We have also described computer programs that solve the planning problem, and that characterize the solution in a variety of ways. Associated with the solution of the planning problem are a set of Lagrange multipliers, which we have shown how to compute in terms of the derivatives of the value function for the planners dynamic programming problem. In the next two chapters, we shall show how those Lagrange multipliers are related to the price system for a competitive equilibrium. We begin by describing how to represent values.

A. Synthesizing the linear regulator The social planning problem is to maximize

−.5E

∞ X t=0

β t (st − bt ) · (st − bt ) + gt · gt

(4.A.1)

subject to Φc ct + Φg gt + Φi it = Γkt−1 + dt

(4.A.2)

kt = ∆k kt−1 + Θk it

(4.A.3)

ht = ∆h ht−1 + Θh ct

(4.A.4)

st = Λht−1 + Πct

(4.A.5)

zt+1 = A22 zt + C2 wt+1

(4.A.6)

Synthesizing the linear regulator

93

bt = U b z t

(4.A.7)

dt = Ud zt

"

#

ht−1 We define the state of the system as xt = kt−1 and the control as ut = it . In zt defining the control to be it , we exploit the assumption that [Φc Φg ] is nonsingular. Solve ( 4.A.2 ) for (ct , gt ) :

h

ct gt

i

= [Φc Φg ]−1 {Γkt−1 + Ud zt − Φi it }.

Let Uc and Ug be selector matrices that pick off the first nc and the last ng rows, respectively, of the right side of the above expression, so that the expression can be written ct = Uc [Φc Φg ]−1 {Γkt−1 + Ud zt − Φi it }

(4.A.8)

gt = Ug [Φc Φg ]−1 {Γkt−1 + Ud zt − Φi it }. Substituting ( 4.A.8 ) into ( 4.A.4 ) and ( 4.A.5 ) gives ht = ∆h ht−1 + Θh Uc [Φc Φg ]−1 {Γkt−1 + Ud zt − Φi it }

(4.A.9)

st = Λht−1 + ΠUc [Φc Φg ]−1 {Γkt−1 + Ud zt − Φi it }

(4.A.10)

Combining ( 4.A.3 ), ( 4.A.9 ), and ( 4.A.6 ) gives the law of motion for the linear regulator ht kt zt+1

!

Θh Uc [Φc Φg ]−1 Γ ∆k 0

=

∆h 0 0

+

−Θh Uc [Φc Φg ]−1 Φi Θk 0

!

Θh Uc [Φc Φg ]−1 Ud 0 A22 it +

0 0 C2

!

!

ht−1 kt−1 zt

! (4.A.11)

wt+1

or xt+1 = Axt + But + Cwt+1

(4.A.12)

where the matrices A, B , and C in ( 4.A.12 ) equal the corresponding matrices in ( 4.A.11 ). Now use ( 4.A.10 ) to compute (st −bt ) = Λht−1 +ΠUc [Φc Φg ]−1 Γkt−1 +(ΠUc [Φc Φg ]−1 Ud − Ub )zt − ΠUc [Φc Φg ]−1 Φi it . Express this in matrix notation as

ht−1 . . . kt−1 . . . −1 −1 −1 (st − bt ) = [Λ . ΠUc [Φc Φg ] Γ . ΠUc [Φc Φg ] Ud − Ub . − ΠUc [Φc Φg ] Φi ] zt it (4.A.13) or h i .. xt (4.A.14) (st − bt ) = [Hs . Hc ] it

94

Optimal Resource Allocation

.. where the matrix [Hs . Hc ] in ( 4.A.14 ) equals the corresponding matrix in ( 4.A.13 ). Next, use ( 4.A.8 ) to express gt as

or

ht−1 . . . k . . . −1 −1 −1 ′ gt = [0 . Ug [Φc Φg ] Γ . Ug [Φc Φg ] Ud . − Ug [Φc Φg ] Φi ] t−1 zt it

h i . xt . gt = [Gs . Gc ] it

(4.A.15)

(4.A.16)

. where the matrix [Gs .. Gc ] in ( 4.A.16 ) equals the corresponding matrix in ( 4.A.15 ). Define the matrices R = .5(Hs′ Hs + G′s Gs ), Q = .5(Hc′ Hc + G′c Gc ), W = .5(Hc′ Hs + G′c Gs ).

(4.A.17)

Then the current period return function for the social planning problem is −(x′t Rxt + u′t Qut + 2u′t W xt ).

(4.A.18)

In view of ( 4.A.14 ), ( 4.A.16 ), ( 4.A.17 ) and ( 4.A.18 ), we can represent the objective function in the social planning problem as

−E

∞ X

β t (x′t Rxt + u′t Qut + 2u′t W xt ),

(4.A.19)

t=0

which is to be maximized over {ut }∞ t=0 subject to xt+1 = Axt + But + Cwt+1 ,

t ≥ 0,

(4.A.20)

x0 given. Thus, we have mapped the social planning problem into a discounted optimal linear regulator problem.

A Brock-Mirman model

95

B. A Brock-Mirman model We shall usually use the recursive numerical methods described above to compute a solution of a social planning problem. These computational methods are quick and easy to use. However, to deepen our understanding of the structure of the social planning problem and the role played by various technical assumptions, and also to heighten our appreciation of the ease and power of those recursive numerical methods, it is useful to solve one problem by hand. We solve a social planning problem for a model with one consumption good and one capital good. We include costs of adjusting the capital stock, but permit them to be zero as a special case. When these costs of adjustment are zero (i.e., when the parameter φ in the model is set to zero), the model becomes a linear - quadratic, equilibrium version of Hall’s consumption model. To recover Hall’s solution of the model when φ = 0 , it is necessary to impose a side condition in the form of a version of our restriction (2.24) that forces the capital stock sequence {kt } to belong to L20 . The example is a useful laboratory for illustrating the relationships among the presence of costs to control (φ > 0) , the transversality condition, and the side condition that the solution lie in L20 . After we work out the answer by hand, we can solve the problem by using the MATLAB program solvea.m . The social planning problem comes from combining versions of our preference specification number 1 and our technology specification number 4: choose a contingency plan for {ct , kt }∞ t=0 to maximize: −E0

∞ X t=0

β t [(ct − bt )2 + ℓ2t ] , 0 < β < 1

(4.B.1)

subject to ct + it = γkt−1 + d1t , γ > 0

(4.B.2)

φit = gt , φ ≥ 0

(4.B.3)

kt = δkt−1 + it , 0 < δ < 1

(4.B.4)

gt2 = ℓ2t

(4.B.5)

k−1 given

(4.B.6)

The stochastic processes bt and d1t are given by bt = Ub zt and d1t = Ud1 zt , where zt obeys a version of (1.1). We assume that {d1t } and {bt } each belong to L20 , and do not impose that {kt } belongs to L20 . We begin by forming the Lagrangian

J = −E0

∞ X t=0

1 β t { [(ct − bt )2 + ℓ2t ] − λ1t [γkt−1 + d1t − ct − it ] 2

− λ2t [gt − φit ]

(4.B.7)

1 − λ3t [δkt−1 + it − kt ] − λ4t [ (ℓ2t − gt2 )]} 2

Here {λ1t , λ2t , λ3t , λ4t }∞ t=0 is a 4-tuple of stochastic Lagrange multipliers. We obtain the first order necessary conditions for a saddle point with respect to {ct , it , kt , ℓt , gt , λ1t , λ2t ,

96

Optimal Resource Allocation

λ3t , λ4t }∞ t=0 , and display the transversality condition for capital. First order conditions with respect to ct , it , kt , lt , and gt are:

ct : −(ct − bt ) − λ1t = 0, t ≥ 0

(4.B.8)

it : −λ1t − φλ2t + λ3t = 0, t ≥ 0

(4.B.9)

kt : γβEt λ1t+1 + βδEt λ3t+1 − λ3t = 0, t ≥

(4.B.10)

ℓt : −ℓt + λ4t ℓt = 0, t ≥ 0

(4.B.11)

gt : λ2t − λ4t gt = 0, t ≥ 0

(4.B.12)

In addition, we have the transversality condition lim E0 β t kt λ3t = 0.

t→∞

(4.B.13)

Equation ( 4.B.10 ) can be solved forward to yield

λ3t = γβ

∞ X

(δβ)j−1 Et λ1t+j .

(4.B.14)

j=1

Our strategy is to substitute the above expressions for the multipliers into the first-order condition with respect to kt to obtain an ‘Euler equation, and to study under what conditions, if any, this equation implies that the marginal utility of consumption is a martingale. Solving the first order conditions for the multipliers, we obtain λ1t = bt − ct

(4.B.15)

λ2t = gt

(4.B.16)

λ3t = φgt + (bt − ct )

(4.B.17)

λ4t = 1

(4.B.18)

Substituting ( 4.B.17 ) into ( 4.B.10 ) gives the “Euler equation” γβEt (bt+1 − ct+1 ) + βδEt (φgt+1 + bt+1 − ct+1 ) = φgt + (bt − ct )

(4.B.19)

or βδEt φgt+1 + β(γ + δ)Et (bt+1 − ct+1 ) = φgt + (bt − ct ).

(4.B.20)

Under the special condition that φ = 0 , this equation becomes Et (bt+1 − ct+1 ) = [β(γ + δ)]−1 (bt − ct ),

(4.B.21)

A Brock-Mirman model

97

which states that the shadow price of consumption (λ1t = bt − ct ) follows a first-order autoregressive process. Under the further special condition that β(γ + δ) = 1 , the shadow price of consumption follows a martingale. 17 Finally, under the even further special condition that bt is a martingale, ( 4.B.21 ) asserts that consumption is a martingale. The Euler equation ( 4.B.21 ) is satisfied by the consumption plan ct = bt for t ≥ 0.

(4.B.22)

Solving ( 4.B.2 ) and ( 4.B.4 ) for it under this plan gives kt = (γ + δ)kt−1 + d1t − bt .

(4.B.23)

Note that in the special case that λ1t (and maybe also ct ) is a martingale, (γ + δ) = 1/β , so that {kt } given by ( 4.B.23 ) is a “process of exponential order 1/β ”. This implies that kt does not belong to L20 . Nevertheless, the transversality condition ( 4.B.13 ) is satisfied because λ3t = φgt + (bt − ct ) = 0 along this solution, so that lim β t λ3t kt = 0

t→∞

along this solution. Thus, when φ = 0 , it is optimal to consume bliss consumption always and to adjust the capital stock to support this consumption plan. The difference equation ( 4.B.23 ) implies that kt = ξ t k0 +

t−1 X j=0

ξ j (d1t−j − bt−j )

where ξ ≡ γ + δ . If bt − d1t > α > 0 for some α for all t , then kt will eventually become negative and, indeed, will eventually fall below any finite negative number. Such a consumption path is eventually being supported by “borrowing” or by ‘negative capital.’ In the interests of attaining an ‘Euler equation’ for capital, we substitute the following two implications of the constraints into the Euler equation: ct = (γ + δ)kt−1 + d1t − kt gt = φkt − φδkt−1 After rearrangement, this gives the following Euler equation for capital: ηEt {kt+1 − ψkt + β −1 kt−1 } = Et zt where

(4.B.24)

η = β[δφ2 + (γ + δ)] ψ=

βδ 2 φ2 + β(γ + δ)2 + φ2 + 1 β(δφ2 + (γ + δ))

(4.B.25)

zt = bt − β(γ + δ)bt+1 − d1t + β(γ + δ)d1t+1 17 The condition that β(γ + δ) ≡ 1 plays the role of a “growth condition” in the model of Jones and Manuelli [1988].

Optimal Resource Allocation

98

We will solve the Euler equation ( 4.B.24 ) using the “certainty equivalence” methods described in Sargent [1987, ch. XIV] and Hansen and Sargent [1980, 1981]. This involves first solving the deterministic version of ( 4.B.25 ), and then replacing “feedforward” terms with their expectations conditioned on time t information. We begin by solving the deterministic version of the Euler equation ( 4.B.24 ): η{kt+1 − ψkt + β −1 kt−1 } = zt

(4.B.26)

ηL−1 {1 − ψL + β −1 L2 }kt = zt .

(4.B.27)

Write this as

We seek a factorization of the polynomial in L : (1 − ψL + β −1 L2 ) = (1 − λ1 L)(1 − λ2 L)

(4.B.28)

Evidently ψ = λ 1 + λ2 λ1 λ2 = β −1 . Thus we have 1 λ1 β

(4.B.29)

1 = ψ. λ1 β

(4.B.30)

λ2 = and λ1 +

Equations ( 4.B.29 ) and ( 4.B.30 ) imply that λ1 and λ2 = λ1β are the intersections of the 1 1 in figure 4.B.1. Since the function line of zero slope and height ψ with the curve λ + λβ

p

1 achieves a minimum of 2/ β at the value λ = 1/ f (λ) = λ + λβ solution of ( 4.B.30 ) exists, it satisfies, without loss of generality,

0 < λ1 < λ2 >

Substituting ( 4.B.28 ) into ( 4.B.27 ) gives

η[(1 − λ1 L)(1 −

p

β , it follows that if a

1

p

β

1

p

.

β

1 L)]kt+1 = zt . λ1 β

(4.B.31)

A Brock-Mirman model

99

b λ + λ -1

bλ

ψ

1+b

λ1

-.5

λ2

b

λ

Figure 4.B.1: The function bλ + 1/λ and its intersections with ψ , which determine the roots λ of the characteristic polynomial ( 4.B.28 ). We start analyzing the solution of ( 4.B.31 ) by returning to the special case in which φ = 0 . In this case, ( 4.B.25 ) implies that ψ=ξ+

1 , βξ

ξ = γ + δ,

(4.B.32)

η = βξ. It then follows immediately from ( 4.B.30 ) that we can take λ1 =

1 βξ

(4.B.33)

λ2 = ξ. In the special case that the shadow price of consumption is a martingale, βξ = 1 , so that 1 . The Euler equation thus becomes, in the special case that φ = 0 , λ1 = 1 and λ2 = β βξ{(1 −

1 L)(1 − ξL)}kt+1 = zt . βξ

But from the constraints to our problem, (1 − ξL)kt+1 = d1t+1 − ct+1

Optimal Resource Allocation

100

Substituting this and the last line of ( 4.B.25 ) into the Euler equation gives βξ(1 −

1 L)(d1t+1 − ct+1 ) = (βξ − L)(d1t+1 − bt+1 ) βξ

or (βξ − L){(d1t+1 − ct+1 } = (βξ − L)(d1t+1 − bt+1 ), an equation that is satisfied by setting ct = bt for all t . Thus, our analysis of the Euler equation for capital in the case that φ = 0 reconfirms our earlier derivation that the optimal plan involves setting ct = bt and choosing whatever capital path is required to support this. We begin to study the case when φ > 0 by considering the special case in which φ is positive but arbitrarily close to zero. In particular, φ can be chosen sufficiently close to zero that in the Euler equation for capital, η{(1 − λ1 L)(1 − λ2 L)}kt+1 = zt , 1 , and λ is arbitrarily close to ξ . This η is arbitrarily close to βξ, λ1 is arbitrarily close βξ 2 can be verified by using a version of figure 4.B.1. It is tempting to suppose that since the Euler equation is arbitrarily close to that for the φ = 0 case, the optimal solution for kt will be close to the solution for kt found in the φ = 0 case, namely,

kt = ξ t k0 +

t−1 X j=0

ξ j (d1t−j − bt−j ).

(4.B.34)

We now show that this supposition is wrong. Note that when kt obeys ( 4.B.34 ), it = kt − δkt−1 , obeys it = ξ t−1 (ξ − δ)k0 + d1t − bt + (ξ − δ)

t−2 X j=0

ξ j (d1t+j−1 − bt−j−1 ).

(4.B.35)

Also, ct = bt ∀t in this case. When it follows ( 4.B.35 ), it is a process of exponential order ξ . It follows that φit is also a process of exponential order ξ when φ > 0 . Now since ℓt = φit along the optimal path, we have that ∞ X i=0

β t ℓ2t = φ2

∞ X

β t i2t .

(4.B.36)

t=0

The process i2t is of exponential order ξ 2t along the solution ( 4.B.35 ). The infinite series ( 4.B.36 ) will converge if and only if β · ξ 2 < 1, or ξ

√1 , so β

that this condition is violated. In this case, ( 4.B.36 ) diverges to +∞ . This means that when investment follows the path ( 4.B.35 ), the objective function for the social planning problem diverges to −∞ when φ = 0 . Since it is possible to find investment paths that leave the value of the objective function finite, a plan in which the objective function diverges to −∞ cannot be optimal. Notice the role that the assumption that φ > 0 plays in the above argument. An alternative argument can be used to show that the path ( 4.B.35 ), or one close to it, cannot be optimal when φ > 0 and ξ > √1 . This argument involves checking the β

transversality condition, which is lim β t kt λ3t = 0.

t→∞

Computing, we have lim β t kt λ3t

t→∞

= lim β t kt (φgt + λ1t ) t→∞

= lim β t kt [φ2 (kt − δkt−1 ) + (bt − ct )]. t→∞

= lim β t [φ2 (kt2 − δkt kt−1 ) + (bt − ct )kt ] t→∞

For a solution that involves setting bt = ct , this becomes lim β t [φ2 (kt2 − δkt kt−1 )] = 0

(4.B.37)

t→∞

A necessary and sufficient condition for ( 4.B.37 ) to be satisfied is that {kt } be of exponential order less than √1 . Along a solution like ( 4.B.34 ), this requires that ξ < √1 , which is β

β

ruled out in the special case that the shadow price of consumption is a martingale. Arguments along these lines can be used to establish generally that when φ > 0 , the solutions for it and for kt are required to be of exponential order less than √1 . β

To solve for the optimal plan when φ > 0 , we return to the factored Euler equation ( 4.B.31 ): η[(1 − λ1 L)(1 −

1 L)]kt+1 = zt λ1 β

(4.81)

p

where 0 < λ1 < 1/ β . Formally, express (1 − λ1β L) = − λ1β L(1 − λ1 βL−1 ) . Substitute 1 1 this into ( 4.B.31 ) to get −η [(1 − λ1 βL−1 )(1 − λ1 L)]kt = zt λ1 β

(4.B.38)

Operating on both sides of ( 4.B.38 ) with the stable (forward) inverse of (1 − λ1 βL−1 ) gives λ β 1 zt (1 − λ1 L)kt = − 1 η (1 − λ1 βL−1 )

(4.B.39)

Optimal Resource Allocation

102

or kt = λ1 kt−1 −

p

Since λ1 < 1/

β, λ1 β

0 , equation ( 4.B.40 ) gives the unique p solution of the Euler equation that satisfies the transversality condition. Because λ1 < 1/ β , kt belongs to L20 .

4.B.1. Uncertainty In the case that zt is a random sequence, the solution when φ1 > 0 is given by kt = λ1 kt−1 −

λ1 β η

∞ X

(λ1 β)j Et zt+j

(4.B.41)

j=0

That this is the solution can be verified by applying the methods of Sargent [1987, chapter XIV]. Consider applying ( 4.B.41 ) in the special case that makes consumption a martingale: βξ = 1, η = βξ = 1, λ1 = 1, bt = ¯b for all t . In this case ( 4.B.41 ) becomes, kt − kt−1 = −β

∞ X j=0

β j Et (d1t+j+1 − d1t+j )

(4.B.42)

We can use a summation by parts argument to show that Et

∞ X j=0

β j (d1t+j+1 − d1t+j ) = (β −1 − 1)Et

∞ X j=0

(4.B.43) β j d1t+j − β −1 d1t

In particular, note that ∞ X j=0

β j (d1t+j+1 − d1t+j )

=

∞ X j=1

β j−1 d1t+j −

= (β −1 − 1)

∞ X j=0

∞ X

β j d1t+j

j=0

β j d1t+j − β −1 d1t .

A Brock-Mirman model

103

Note that from the constraints ct = (γ + δ)kt−1 − kt + d1t or ct =

1 kt−1 − kt + d1t β

(4.B.44)

in the special case that (γ + δ)β = 1 , which we are studying. Substituting ( 4.B.42 ) and ( 4.B.43 ) into ( 4.B.44 ) and rearranging gives

ct =

∞

X 1 β j Et d1t+j . − 1 kt−1 + (1 − β) β

(4.B.45)

j=0

With kt−1 interpreted as “assets” and {d1t } interpreted as “labor income”, representation ( 4.B.45 ) matches the representation of the permanent income theory of consumption that is associated with a linear quadratic version of Hall’s model. In this model, φ = 0 , so that ( 4.B.42 ) and ( 4.B.45 ), which emerge from imposing that {kt } reside in L20 , are not optimal for the original problem as stated. The solution ( 4.B.45 )results from imposing as a side condition on the problem a version of ( 4.A.10 ). This side condition is intended to capture the idea that it is not really feasible to drive capital to negative infinity as quickly as the (unrestricted) φ = 0 solution would require. The solution ( 4.B.45 ) is well approximated by the solution of the original problem with φ > 0 but φ very close to zero. Instead of imposing the requirement that {kt }ǫL20 as a sort of “feasibility” condition, setting φ > 0 rigs preferences so that the social planner always prefers to make {kt } ∈ L20 .

4.B.2. Optimal Stationary States Temporarily assume that bt = ¯b and d1t = d¯ for all t . To solve for the optimal stationary values of ct and kt (if they exist), we can use equation ( 4.B.20 ) and the following constraints: φit = gt it = kt − δkt−1 ct + it = γkt−1 + d1t

(4.52) (4.53) (4.51)

¯ for all t gives Evaluating these at steady state levels ct = c¯ and kt = k ¯ + d. ¯ c¯ = (γ + δ − 1)k Substituting the constraints into the Euler equation ( 4.B.20 ) and evaluating at ct = c¯ and ¯ gives kt = k ¯ = [1 − β(γ + δ)](¯b − c¯) φ2 (βδ − 1)(1 − δ)k

104

Optimal Resource Allocation

¯ gives Solving the two preceding equations for c¯ and k ¯ = [φ2 (βδ − 1)(1 − δ) + (1 − β(γ + δ))(γ + δ − 1)]−1 k ¯ (1 − β(γ + δ)) · (¯b − d) c¯ =

(γ + δ − 1)(1 − β(γ + δ)) [φ2 (βδ − 1)(1 − δ) + (1 − β(γ + δ))(γ + δ − 1)]

(4.B.46)

(4.B.47)

¯ + d. ¯ (¯b − d) In the special case that φ = 0 , these solutions imply that c¯ = ¯b , so that consumption is at bliss consumption and the steady state value of the multiplier λ1t is zero. When φ = 0 , the steady state value of k can be taken to be ¯= k

1 [d¯ − ¯b], 1 − (γ + δ)

(4.B.48)

a solution that makes sense only when (γ + δ) < 1 . Note that the constraints imply that capital evolves according to kt = (γ + δ)kt−1 − ct + d1t . Setting ct = c¯ and d1t = d¯ implies ¯ kt = (γ + δ)kt−1 − c¯ + d. The solution of this equation is

kt = (γ + δ)t k0 + (d¯ − c¯)

t−1 X

(γ + δ)j .

j=0

¯ when c¯ = ¯b and (γ + δ) < 1 . This solution converges to the solution ( 4.B.48 ) for k

Chapter 5 The Commodity Space

5.1. Valuation This chapter describes a concept of value that we shall later use to formulate a decentralized version of our model in which the decisions of agents are reconciled in a competitive equilibrium. We describe a commodity space in which both the quantities and prices will reside. The stochastic Lagrange multipliers of chapter 4 are very closely related to the equilibrium prices that we shall compute, and live in the same mathematical space with prices. The social planning problem studied in chapter 4 produces an outcome in which the process for consumption {ct } is an n -dimensional stochastic process that belongs to L20 . To calculate the value π(c) of a particular consumption plan c = {ct } from the vantage of time zero, we shall use the representation π(c) = E

∞ X t=0

β t p0t · ct | J0 ,

where p0t belongs to L20 . The text of this chapter presents a heuristic justification for so representing the value of {ct }. We proceed by reviewing several examples of commodity spaces and valuation functions. The appendix contains a more formal treatment.

– 105 –

106

The Commodity Space

5.2. Price systems as linear functionals We follow Debreu (1954) and express values by using a linear functional π that maps elements of a linear space L into the real line. The space L is taken as the commodity space, elements of which are the vectors of commodities to be evaluated. The functional π assigns values to points in L. It is convenient when the functional π has an inner-product representation, which is a representation in which the value π(c) of a commodity point c equals the inner product of c ˜ When such a representation exists, with a point p in another linear space L. we can write π(c) =< c | p > for all c in L (5.2.1) ˜ and < · | · > denotes an inner product. In all of the cases that we where pε L ˜ = L, so that c and p reside in the same linear consider, it turns out that L space. Next we consider several examples of a commodity space L, a valuation functional π , and an inner product representation for π .

5.3. A one period model under certainty Suppose that there is one period and no uncertainty. Let there be n consumption goods. Let c be an n × 1 vector of consumption goods. Let the commodity space L be Rn , the n -dimensional Euclidean space. In this case, the value of a vector c is given by n X π(c) =< c | p >≡ ci pi i=1

where < · | · > denotes the inner product, and p is an n -dimensional price vector that belongs to L = Rn . Note that both c and p belong to the same linear space L.

One period under uncertainty

107

5.4. One period under uncertainty Suppose there is again one period, but now there is uncertainty about economic outcomes. Prior to the resolution of uncertainty, the quantity of the ith consumption good is a random variable ci (ω), where ω is the state of the world to be realized. Let c = c(ω) be an n -dimensional random vector whose ith component is ci (ω). Let prob(ω ) be the probability density function of ω . We want to evaluate a bundle of consumption goods prior to the resolution of uncertainty. Introducing uncertainty serves to increase the dimension of the commodity space, there being a vector c(ω) for each state of the world ω ∈ Ω, where Ω is the set of possible states of the world. When there is an infinite number of states of the world Ω, the commodity space L becomes infinite dimensional. To evaluate a state-contingent bundle of consumption goods prior to the resolution of uncertainty requires a well defined notion of “adding up” or integrating across states of the world. When the number of states of the world is finite (or countable), it is natural to follow Arrow and Debreu and to define an n -dimensional vector of statecontingent prices q(ω), where Ω = [ω1 , ω1 , . . . , ωN ] is the set of possible states of the world. The value of the random vector c can then be represented as π(c) =

N X j=1

c(ωj ) · q(ωj ) ≡< c | q > .

(5.4.1)

Here both c and q are elements of L, the space of n -dimensional random vectors indexed by the state of the world. The ith component of q(ω), qi (ω), is to be interpreted as price of one unit of the ith consumption good contingent on the state of the world being ω . It is convenient to represent π(c) in the alternative form π(c) =

N X j=1

c(ωj ) · p(ωj ) prob (ωj ),

(5.4.2)

where q(ωj ) = p(ωj ) prob (ωj ). Here c and p are each vectors in L, the space of n -dimensional random vectors. Notice that (5.4.2) implies π(c) = Ec · p ≡< c | p > . Representation (5.4.1) is often used in contexts in which there is a finite or countable number of states of the world. We find it easier to use representations

108

The Commodity Space

that build upon (5.4.2) because we shall be dealing with environments with an uncountable number of states of the world.

5.5. An infinite number of periods and uncertainty We now come to the main case studied in this book. The n -dimensional vector of consumption goods ct is indexed both by states of the world and by time. We define an information set Jt as in chapters 2 and 3. Let L be the space of all n -dimensional stochastic processes {ct : t = 0, 1, . . .} for which ct is in Jt for all t and for which ∞ X β t E(ct · ct ) < ∞. (5.5.1) t=0

The constraint that ct be in Jt is imposed because we want to represent the values only of contingent claims that depend on information available when the contingency is realized. The inequality restriction in (5.5.1) identifies which claims might have finite value. In addition to integrating over states of the world, we also must sum over points in time. We find it convenient to use the discount factor β in performing this summation. Hence we use the following inner product: < c | p >=

∞ X t=0

β t E(ct · pt ).

(5.5.2)

In this case, the price system used to represent the valuation functional is an n -dimensional stochastic process {pt : t = 0, 1, . . .} in L.

An infinite number of periods and uncertainty

109

5.5.1. Conditioning information So far we have considered valuation functions that map into the real numbers IR . This approach suffices for representing competitive equilibrium prices for markets that meet and clear prior to the realization of any information. However, we also want to reopen markets and to study valuations at later points in time, conditioned on information available then. Consider valuation from the vantage point of time τ . Let valuation be conditioned on the time τ information set Jτ . Let πτ be a time τ valuation function. We take the domain of πτ to be the space Lτ consisting of all n dimensional processes {ct+τ : t = 0, 1 . . .} where ct+τ is in Jτ and ∞ X t=0

β t E(ct+τ · ct+τ ) | Jτ < ∞

(5.5.3)

with probability one. The range of πτ is Jτ because valuations reflect the available conditioning information. There is no longer an inner-product representation for πτ because the range of πτ is not the real line. Rather, the range is the space of random variables depending on Jτ . However, we can follow Harrison and Kreps (1979) and Hansen and Richard (1987) by using a conditional inner-product representation: πτ (c) =< c | p >τ =

∞ X t=0

β t E(ct+τ · pt+τ | Jτ )

(5.5.4)

where {pt+τ : t = 0, 1, . . .} is a price process in L. The value assigned by πτ is a random variable in L2τ .

110

The Commodity Space

5.6. Lagrange multipliers While we have focused on representing valuation in a competitive equilibrium, much of our discussion applies to using the method of Lagrange multipliers for solving constrained optimization problems. The vector of Lagrange multipliers for a vector of constraints indexed by states of the world and calendar time can be regarded as a stochastic processes {Mt : t = 0, 1, . . .} in a space L. The contribution to the Lagrangian is given by a corresponding linear functional µ with an inner product representation µ(ε) =< ε | M >=

∞ X t=0

β t E(εt · Mt )

(5.6.1)

where εt is the deviation of the constraint at time t .

5.7. Summary Our purpose in this chapter has been to lay groundwork necessary to decentralize the economy described in chapter 3 into one with a collection of price-taking agents whose decisions are coordinated through markets. The Appendix to this chapter describes the valuation functions that we use in more mathematical detail.

A. Appendix As was indicated above, we model π as a linear functional on a space L . The space L is assumed to be a linear space, by which we mean that for any two members x1 and x2 in L and any two real numbers c1 and c2 in R , c1 x1 + c2 x2 are in L . In addition, we suppose that there is an inner product < · | · > defined on L . This inner product can be used to define a norm || x ||=< x | x >1/2 and hence a metric. We take L to be complete. This means that all Cauchy sequences in L converge to an element of L . The commodity spaces in all of the examples described in the text are complete linear spaces. The restriction that π be linear requires that π(c1 x1 + c2 x2 ) = c1 π(x1 ) + c2 π(x2 ) . According to the Riesz Representation Theorem, π has an inner product representation whenever π is continuous at zero. When conditioning information is introduced, it is convenient to work with a space LJ that is linear conditioned on J . For the moment, consider LJ to be a collection of

Appendix

111

random variables. Products and sums of random variables are also random variables. For LJ to be linear conditioned on J , for any two elements x1 and x2 of LJ and any w1 and w2 in J, we require that w1 x1 + w2 x2 is in LJ . Similarly, πJ is conditionally linear if πJ (w1 x1 +w2 x2 ) = w1 πJ (x1 )+w2 πJ (x2 ) . The rationale for focusing on conditional linearity is that information in J can be used to construct consumption plans or trading strategies. Hansen and Richard (1987) obtained a conditional counterpart to the Riesz Representation Theorem that establishes the existence of a representation πJ (x) = E(x · p | J) for some p in LJ . The restriction that LJ be a space of random variables is too limited for our purposes. Instead, we are interested in spaces of n -dimensional stochastic processes. Given an initial probability space (Ω, F, Pr) and a sequence {Ft : t = 0, 1, . . .} of subsigma algebras of F , we construct a new probability space (Ω+ , F + , Pr+ ) where Ω+ is the Cartesian product of Ω+ , the nonnegative integers, and the set {1, 2, . . . , n} , and where Pr+ is the product measure of Pr , a measure that assigns β t (1 − β) to integer t , and 1/n to integer j . The sigma algebra F + is generated by sets of the form: {(w, t, j) : w ∈ ft,j }

(5.A.1)

where {ft,j : t = 0, 1, . . . ; , j = 1, 2, . . . , n} is a collection of sets in F such that ft,j is in Ft for all t and j . An n -dimensional stochastic process defined on the original space can be viewed as a random variable on the product space. Thus we can apply the preceding analysis to obtain a conditional inner product representation for πτ described in the text.

Chapter 6 A Competitive Economy

6.1. Introduction This chapter describes a decentralized version of our economy. We assign ownership and decision making to three distinct economic entities, a household and two kinds of firms. We define a competitive equilibrium. Versions of the two fundamental theorems of welfare economics are true. We establish these theorems by exhibiting the connection between a competitive equilibrium and a social planning problem. A price system supports the competitive equilibrium, and implies interest rates and prices for derivative assets. The representative household can be interpreted as a single individual drawn from a population that is homogeneous in all respects. Alternatively, the representative household can be interpreted along lines to be described in chapter 12, as an artificial or “average” household that emerges from aggregating over the preferences and endowments of a collection of households. The representative household owns the technology shock process dt , and each period sells to firms the current period’s realization of the shock process. The household owns the initial stocks h−1 of household capital and k−1 of productive capital, the latter of which it sells to firms. It sells this initial capital for a value v0 · k−1 . The household sells its input ℓt to firms. The household uses its resources to purchase consumption goods, which add to its stocks of consumer durables and thereby generate consumption services and utility. Of the two types of firms, the first type rents capital from firms of type II, rents labor from the household, and buys the current period’s realization of the technology shock process dt from the household. A firm of type I produces new consumption and investment goods, sells the consumption goods to the household, and sells the investment goods to the firms of type II. A firm of type II purchases the initial capital stock k−1 and all of the investment goods produced each period, then rents capital to firms of type I. We use a formulation of a price system which is mathematically convenient, as well as economically interpretable. We let the price system be [v0 , {p0t , wt0 , αt0 , qt0 , – 113 –

114

A Competitive Economy

0 rt0 }∞ t=0 ], where v0 is a vector that prices the initial capital stock k−1 ; pt is an 0 nc × 1 stochastic process that prices the consumption process ct ; wt is a scalar stochastic process that prices ℓt ; αt0 is a vector stochastic process that prices the process {dt }; qt0 is an nk ×1 vector stochastic process that prices new investment goods; and rt0 is an nk ×1 vector stochastic process of capital rental rates. Each 2 component of [{p0t , wt0 , αt0 , qt0h, rt0 }∞ t=0 ] resides in the mathematical space L0 defined earlier, namely, L20 = {yt }∞ t=0 : yt is a random variable in Jt for t ≥ i P∞ t 2 0, and E t=0 β yt | J0 < +∞ . That ‘yt is in Jt ’ means that yt can be

expressed as a measurable function of Jt = [wt , x0 ], where J0 = [x0 ]. The P∞ square summability requirement, E t=0 β t yt2 | J0 < ∞ , imposes a stochastic version of a requirement that yt not grow too fast in absolute value. Stochastic processes for both prices and quantities in our economy must reside in L20 . By virtue of a Cauchy-Schwartz inequality, this makes the conditional inner products to be used in the budget constraints and objective functions below well defined and finite in equilibrium. This chapter formulates and computes a competitive equilibrium. We proceed by first describing the problem for each of our three classes of agents in terms of a Lagrangian. Next we obtain the first order conditions from these Lagrangians. By “matching up” these first-order conditions to the first order conditions found in chapter 3 for the social planning problem, we accomplish two goals. First, we can verify the two fundamental theorems of welfare economics for our economy. Second, we can describe an efficient algorithm for computing the equilibrium price system in terms of the matrices Mk , Mh , Ms , Md , Mc , and Mi of chapter 3 associated with the multipliers for the social planning problem. We first describe the problems faced by each of our three types of agents.

The Problems of Households and Firms

115

6.2. The Problems of Households and Firms

6.2.1. Households The household chooses stochastic processes for {ct , st , ht , ℓt }∞ t=0 , each element of which is in L20 , to maximize −

∞ h i X 1 β t (st − bt ) · (st − bt ) + ℓ2t E0 2 t=0

(6.2.1)

subject to E

∞ X t=0

β t p0t · ct | J0 = E

∞ X t=0

β t (wt0 ℓt + αt0 · dt ) | J0 + v0 · k−1

st = Λht−1 + Πct ht = ∆h ht−1 + Θh ct ,

h−1 , k−1 given.

(6.2.2) (6.2.3) (6.2.4)

The household and each firm acts as a price taker. The optimal contingency plan for (ct , st , ht , ℓt ) must be “realizable” in the sense that time t decisions must be contingent only on information available at time t , i.e., it must reside in L20 .

6.2.2. Firms of type I A firm of type I rents capital and labor, and buys the realization of the endowment process dt . It uses these inputs to produce consumption goods and investment goods, which it sells. The firm of type I chooses stochastic processes for {ct , it , kt , ℓt , gt , dt }, each element of which is in L20 , to maximize E0

∞ X t=0

β t (p0t · ct + qt0 · it − rt0 · kt−1 − wt0 ℓt − αt0 · dt )

(6.2.5)

subject to Φc ct + Φg gt + Φi it = Γkt−1 + dt

(6.2.6)

116

A Competitive Economy

− ℓ2t + gt · gt = 0

(6.2.7)

6.2.3. Firms of type II A firm of type II is in the business of purchasing investment goods and renting capital to firms of type I. A firm of type II faces as a price taker the vector v0 and the stochastic processes {rt0 , qt0 }. The firm chooses k−1 and stochastic processes for {kt , it }∞ t=0 to maximize E

∞ X t=0

subject to

β t (rt0 · kt−1 − qt0 · it ) | J0 − v0 · k−1 kt = ∆k kt−1 + Θk it .

(6.2.8)

(6.2.9)

6.3. Competitive Equilibrium We define a competitive equilibrium for this economy. Definition: A competitive equilibrium is a price system [v0 , {p0t , wt0 , αt0 , qt0 , rt0 }∞ t=0 ] and an allocation {ct , it , kt , ht , gt , dt }∞ that satisfy the following conditions: t=0 a. Each component of the price system and the allocation resides in the space L20 .

b. Given the price system and given h−1 , k−1 , the stochastic process {ct , st , ℓt , ht }∞ t=0 solves the consumer’s problem.

c. Given the price system, the stochastic process {ct , it , kt , ℓt , dt , gt } solves the problem of the firm of type I.

d. Given the price system, the vector k−1 and the stochastic process {kt , it }∞ t=0 solve the problem of the firm of type II.

Lagrangians

117

6.4. Lagrangians We now formulate each agent’s problem as a Lagrangian, and obtain the associated first order conditions.

6.4.1. Households The household’s Lagrangian is ∞ X βt (st − bt ) · (st − bt ) + ℓ2t /2 L = − E0 t=0

0 0 0 + µw 0 [pt · ct − wt ℓt − αt · dt ]

+ µs′ t [st − Λht−1 − Πct ] + µh′ [h − ∆ h − Θ c ] + µw t h t−1 h t t 0 v0 · k−1 . h s Here µw 0 is a scalar and {µt , µt } are sequences of vectors of stochastic Lagrange multipliers. The first order conditions with respect to st , ℓt , ct , and ht , respectively, are:

st : ℓt : ct : ht :

(st − bt ) + µst = 0,

ℓt −

wt0

·

µw 0

t≥0

t≥0

= 0,

′ h ′ s 0 µw 0 pt − Π µt − Θh µt = 0,

− β Et Λ

′

µst+1

−β

Et ∆′h

µht+1

+

µht

= 0,

t≥0

t≥0

Solving these equations, we obtain

µht = Et

µst = bt − st ,

t≥0

(6.4.1)

wt0 = ℓt /µw 0,

t≥0

(6.4.2)

∞ X

β τ (∆′h )τ −1 Λ′ µst+τ

τ =1 0 ′ s ′ h µw 0 pt = Π µt + Θh µt ,

, t≥0

t≥0

(6.4.3) (6.4.4)

118

A Competitive Economy

6.4.2. Firms of type I The Lagrangian of a type I firm is ∞ X t LI = E0 β [p0t · ct + qt0 · it − rt0 · kt−1 − wt0 ℓt − αt0 · dt ] t=0

+ Ld′ t [Γkt−1 + dt − Φc ct − Φg gt − Φi it ] ℓ′ 2 + Lt [(ℓt − gt · gt )/2] .

Here {Ldt , Lℓt } is a vector stochastic process of Lagrange multipliers. The first order conditions associated with interior solutions for ct , it , kt , ℓt , dt , and gt , respectively, are ct :

p0t − Φ′c Ldt

= 0,

t≥0

(6.4.5)

it :

qt0 − Φ′i Ldt

= 0,

t≥0

(6.4.6)

kt :

0 rt+1 − Γ′ Ldt+1

= 0,

t ≥ −1

(6.4.7)

ℓt :

− wt0 + Lℓt ℓt

= 0,

t≥0

(6.4.8)

dt :

− αt0 + Ldt

= 0,

t≥0

(6.4.9)

gt :

−

Φ′g Ldt

−

gt Lℓt

= 0,

t≥0

Solving (6.4.5) and (6.4.10) for Ldt gives ′ −1 Φc p0t d . Lt = Φ′g −gt Lℓt

(6.4.10)

(6.4.11)

From (6.4.8), the solution for Lℓt satisfies Lℓt = wt0 /ℓt .

(6.4.12)

Equations (6.4.6), (6.4.7) and (6.4.9) imply qt0 = Φ′i Ldt

(6.4.13)

rt0 = Γ′ Ldt

(6.4.14)

αt0 = Ldt

(6.4.15)

Lagrangians

119

6.4.3. Firms of type II The Lagrangian of a type II firm is LII = E0

∞ X t=0

β (rt0 · kt−1 − qt0 · it ) t

+ ηt′ (∆k kt−1 + Θk it − kt ) − v0 · k−1 where {ηt } is a sequence of stochastic Lagrange multipliers. The first order conditions for interior solutions with respect to kt , it , and k−1 , respectively, are 0 kt : βEt rt+1 − ηt + βEt ∆′k ηt+1 = 0,

it :

− qt0 + Θ′k ηt

= 0,

k−1 : r00 + ∆′k η0 − v0

t≥0

(6.4.16)

t≥0

(6.4.17)

=0

(6.4.18)

, t≥0

(6.4.19)

Solving (6.4.16) for ηt gives ηt = Et

∞ X

′(j−1) 0 rt+j

β j ∆k

j=1

Equation (6.4.17) implies qt0 = Θ′k ηt ,

t≥0

(6.4.20)

Equation (6.4.18) implies v0 = r00 + ∆′k η0

(6.4.21)

120

A Competitive Economy

6.5. Equilibrium Price System Our task now is to find stochastic processes of prices, quantities, and Lagrange multipliers that satisfy the first-order conditions for each of our three classes of agents for all time and contingencies. We proceed constructively to link equilibrium prices to the Lagrange multipliers for the planning problem. Recall the following equations obeyed by the Lagrange multipliers associated with the social planning problem: Mst = bt − st

(4.8)

(4.9)

(4.11)

(4.19)

X ∞

Mht

=E

Mdt

Φ′c = Φ′g

Mkt

β

τ

τ =1

=E

−1

X ∞

τ

(∆′h )τ −1 Λ′ Mst+τ

| Jt

Φ′h Mht + Π′ Mst −gt

′ τ −1

β (∆ )

Γ

τ =1

′

Mdt+τ

| Jt

In chapter 3, we gave formulas for these multipliers along the optimum of the social planning problem, namely, (4.21)

Mkt = Mk xt and Mht = Mh xt

(4.22)

Mst = Ms xt

(4.23)

Mdt = Md xt .

We also defined shadow prices for consumption and investment: (4.24)

Mct = Mc xt , Mc = Θ′h Mh + Π′ Ms

(4.25)

Mit = Mi xt , Mi = Θ′i Mk .

Equilibrium Price System

121

We gave formulas for the matrices Ms , Mk , Mh and Md in terms of the optimal value function of the social planning problem. The formulas (4.21), (4.22), (4.23), (4.24), (4.25) for the multipliers are evaluated along the solution xt+1 = Ao xt + Cwt+1 of the social planning problem. We can compute the equilibrium price system in terms of the multipliers from the social planning problem. For the time being let µw 0 be a free parameter. Later we shall indicate how choosing the scalar marginal utility of wealth at time zero, µw 0 , amounts to specifying a numeraire for our price system. We propose to set c w p0t = Π′ Mst + Θ′h Mht /µw (6.5.1) 0 = Mt /µ0 wt0 =| Sg xt | /µw 0

(6.5.2)

rt0 = Γ′ Mdt /µw 0

(6.5.3)

i w qt0 = Θ′k Mkt /µw 0 = Mt /µ0

(6.5.4)

αt0 = Mdt /µw 0

(6.5.5)

′ k w v0 = Γ′ Md0 /µw 0 + ∆k M0 /µ0 .

(6.5.6)

We shall verify that with this price system, values can be assigned to the Lagrange multipliers for each of our three classes of agents that cause all of their first-order necessary conditions to be satisfied at these prices and at the quantities associated with the optimum of the social planning problem. For the household, we set (6.5.7) µst = Mst µht = Mht

(6.5.8)

With these choices of multipliers, equations (6.4.1), (6.4.2), (6.4.3) and (6.4.4) are evidently satisfied at the proposed equilibrium prices (6.5.1) – (6.5.6) and at the quantities associated with the optimum of the social planning problem. For the firm of type I, we set Ldt = Mdt /µw 0

(6.5.9)

Lℓt = 1/µw 0.

(6.5.10)

With the settings (6.5.9) for Ldt , (6.5.10) for Lℓt , and the price process (6.5.1), equation (6.4.11) becomes equivalent with (4.11) from the social planning problem. Equation (6.5.3) for rt0 implies that the firm’s marginal condition (6.4.14)

122

A Competitive Economy

is satisfied along the solution of the social planning problem. Similarly, (6.4.20) implies that (6.4.15) is satisfied. Formula (6.5.4) for qt0 together with the fourth equation of (4.8) (−Φ′i Mdt +Θ′k Mkt = 0) implies that (6.4.13) is satisfied along the solution of the social planning problem. Finally, (6.5.9)–(6.5.10) imply that (6.4.10) is equivalent with the second equation of (4.8) (−gt − Φ′g Mdt = 0). Thus, with settings (6.5.8), (6.5.9), price system (6.5.1)–(6.5.6) implies that firm I’s first order necessary conditions are satisfied along the quantity path implied by the social optimum. For the firm of type II, we set ηt = Mkt /µw 0.

(6.5.11)

With this setting, (6.5.6) and (3.19) imply that (6.4.19) (and thus (6.4.16)) is satisfied. Also, (6.4.20) is evidently satisfied as is (6.4.21). Thus, the first order conditions for firms of type II are all satisfied at price system (6.5.1)–(6.5.6) along the solution of the social planning problem. We are finished. In summary, the price system (6.5.1)–(6.5.6) supports the allocation associated with the optimum of the social planning problem as a competitive equilibrium. The direction of argument can be reversed to establish that a competitive equilibrium solves the social planning problem. This argument uses a competitive equilibrium allocation and price system to define multiplier processes that satisfy the first order conditions for the social planning problem. 1 The scalar µw 0 that appears as a free parameter in (6.5.1)–(6.5.6) is evidently the marginal utility of wealth at time zero. In setting this parameter, we select a numeraire for our price system. For example, the j th consumption good at time zero can be selected as the numeraire by setting ¯j Mct = e¯j Mc x0 µw 0 =e where e¯j is a (1 × nc ) vector consisting of zeros in each location except the j th, where there is a one. For the j th consumption good at time zero to be a valid numeraire, we require that e¯j Mc x0 not equal zero. This is imposed in: Assumption 5.1: The random variable e¯j Mc x0 selected as numeraire differs from zero with probability one.

1 Since the solution of the social planning problem is unique, so is the competitive equilibrium.

Asset Pricing

123

6.6. Asset Pricing We can use the main idea behind “arbitrage pricing theory” to motivate asset pricing formulas. Arbitrage pricing theory extracts solely from the weak implication of equilibrium that assets must be priced so that budget sets offer no opportunities for earning sure returns with a zero commitment of resources. To illustrate this approach, imagine altering the representative household’s problem (6.2.1) – (6.2.4) by supplying it with one additional opportunity. The household can go into the securities business on the side by issuing securities that promise to pay off a stream of the (nc × 1) vector of consumption goods {yt }. We assume that {yt } ∈ L20 . Suppose there is a market in such securities and that the price at time 0 of one unit of such security is a0 . If the household sells S of these securities, its revenue at time 0 is Sa0 . To cover itself in all contingencies, the household must purchase state contingent claims to consumption in the amount {yt } for each unit of the security issued. The cost of purchasing these claims to support the sale of S securities is S·E

∞ X t=0

β t p0t · yt | J0 .

With this opportunity opened up to the household, the following term must be added to the right side of household’s budget constraint (6.2.2): S(a0 − E P∞

t

p0t

∞ X t=0

β t p0t · yt | J0 ).

· yt | J0 , the household can make the present value of conIf a0 > E t=0 β sumption as large as it wants by setting S equal to a suitable positive number, i.e., by selling the security whose price is a0 . However, for our economy, it is not feasible for the consumer to achieve any such desired present value of conP∞ sumption. Therefore, in equilibrium we cannot have a0 > E t=0 β t p0t · yt | J0 . P∞ Similarly, we cannot have a0 < E t=0 β t p0t · yt | J0 , because that would confront the household with the opportunity to make the present value of consumption as large as it wants by buying the security at prices a0 , then selling the returns yt in the market for state contingent claims. Therefore, we must have ∞ X a0 = E β t p0t · yt | J0 . (6.6.1) t=0

124

A Competitive Economy

We can use (6.6.1) to derive formulas for various special {yt } processes, and thereby recover versions of Lucas’s asset pricing model [1978], and theories of the term structure of interest rates. We derive more explicit formulas for assets with payoffs of the form yt = Ua xt

(6.6.2)

where Ua is an nc × n matrix. Substituting (6.6.2) and the pricing formula p0t = Mc xt /µw 0 into (6.6.1) gives a0 = E

∞ X t=0

β t x′t Za xt | J0

(6.6.3)

where Za = Ua′ Mc /µw 0.

(6.6.4)

We shall now show that a0 can be represented as a0 = x′0 µa x0 + σa where µa =

∞ X

β τ (Ao′ )τ Za Aoτ

(6.6.5)

(6.6.6)

τ =0

σa =

∞ X β trace Za β τ (Ao )τ CC ′ (Ao′ )τ . 1−β τ =0

(6.6.7)

According to (6.6.5), the asset price a0 turns out to be the sum of a constant σa , which reflects a “risk premium,” and a quadratic form in the state vector xt . To understand why σa reflects a risk premium, notice how the parameters in C influence σa but do not influence µa . To derive (6.6.5), first express (6.6.3) as 2 ∞ X

β t trace [Za xt x′t ] | J0 .

(6.6.8)

(Ao )τ CC ′ (Ao′ )τ + (Ao )t x0 x′0 (Ao′ )t .

(6.6.9)

a0 = E

t=0

For t ≥ 1, (1.5) implies that Ext x′t | J0 =

t−1 X

τ =0

2 An alternative way to derive these formulas is described in chapter [ ] (seasonality).

Asset Pricing

125

Substituting (6.6.9) into (6.6.8) and rearranging gives

a0 =

∞ X

β t trace

t=1

t−1 h X i Za (Ao )τ CC ′ (Ao′ )τ τ =0

+ trace Za

∞ X

t

o t

β (A )

x0 x′0

(6.6.10)

o′ t

(A ) .

t=0

Exchanging orders of summation in the first term on the right of (6.6.10) gives ∞ X t=1

t−1 h i X β t trace Za (Ao )τ CC ′ (Ao′ )τ τ =0

= trace Za =

∞ ∞ X X

τ =0 t=τ +1 ∞ X

β trace Za 1−β

≡ σa

β t (Ao )τ CC ′ (Ao′ )τ β τ (Ao )τ CC ′ (Ao′ )τ

τ =0

which establishes (6.6.7). The second term on the right side of (6.6.10) can be transformed (by repeatedly using the rule trace AB = trace BA) to x′0

∞ X t=0

β t (Ao′ )t Za (Ao )t x0 ≡ x′0 µa x0 ,

which defines the matrix µ given in (6.6.6). This completes our verification of the asset pricing formulas (6.6.5) – (6.6.7). To implement (6.6.5) requires the application of numerical methods to calculate the matrices µa and σa that satisfy (6.6.6) and (6.6.7). An efficient ‘doubling algorithm’ for calculating these matrices is described in chapter 8. As an application of (6.6.3) – (6.6.5), let us compute the value of a title to one unit of the stream of the j th endowment shock, {djt }∞ t=0 . Let djt = ej xt , where ej is a selection vector that picks off the appropriate linear combination of xt . From (6.5.5) we have that the time zero value of the time t shock djt is ′ ′ d w djt M d xt /µw 0 = xt ej M xt /µ0 .

126

A Competitive Economy

The value of the entire stream is then given by E

∞ X t=0

β t x′t Za xt | J0

where Za = e′j M d /µw 0 . This matches (6.6.3), so that formulas (6.6.5)–(6.6.7) are applicable.

6.7. Term Structure of Interest Rates The value at time zero of a sure claim to one unit of the first consumption good at time zero is evidently given by R10 = βE[¯ e1 · p01 ] | J0 or R10 = β e¯1 · Mc Ao x0 /µw 0.

(6.7.1)

Here R10 is the reciprocal of the gross one-period sure interest rate at time zero. For longer horizons, we have Rj0 = β j E[¯ e1 · p0j ] | J0 , j ≥ 1 or Rj0 = β j e¯1 · Mc Aoj x0 /µw 0.

(6.7.2)

Here Rj0 is the reciprocal of the gross interest factor for a sure claim on the first consumption good j periods into the future at time zero.

Re-opening Markets

127

6.8. Re-opening Markets The competitive equilibrium prices state– and date–contingent commodities that are traded at time zero. After time zero, markets are “closed,” with traders simply executing agreements entered into at time zero. As usual in ArrowDebreu models, markets can be opened in subsequent time periods, but are redundant in the sense that zero trades occur. However, for the purpose of extracting the time series implications of our model, it is useful to compute the prices in such re-opened markets. Suppose that markets re-open at some time t ≥ 1, and that the household and firms re-evaluate their contingency plans at new prices. The household now values consumption services from time t forward. Only goods from time t forward enter the valuations appearing in the budget sets and objective functions of each of our agents. We use L2t as the commodity space, defined as L2t = [{ys }∞ s=t : ys is a random variable in Js for s ≥ t ∞ X and E β s−t ys2 | Jt < +∞] s=t

Expectations conditioned on Jt replace those conditioned on J0 in the intertemporal budget constraint of the household and the cash flow evaluations of the firms. For convenience, we use the j th consumption good at time t as the numeraire. For this choice to deliver a valid numeraire, we invoke Assumption 5.2: The random variable e¯j M c xt differs from zero with probability one. We set the household’s marginal utility of time t wealth, µw t , equal to c th e¯j M xt in order to select the time t , j consumption good as numeraire. With these specifications, we can simply replicate the time zero analysis to obtain equilibrium prices from the vantage point of time t . This yields the following price system: pts = Mc xs /[¯ ej Mc xt ],

s≥t

(6.8.1)

wst =| Sg xs |/[¯ ej Mc xt ], s ≥ t

(6.8.2)

rst = Γ′ Md xs /[¯ ej Mc xt ], s ≥ t

(6.8.3)

qst = Mi xs /[¯ ej Mc xt ],

(6.8.4)

s≥t

128

A Competitive Economy

αst = Md xs /[¯ ej Mc xt ], s ≥ t

(6.8.5)

vt = [Γ′ Md + ∆′k Mk ]xt / [¯ ej Mc xt ]

(6.8.6)

Of particular interest are the spot market prices implied by (6.8.1) – (6.8.6), namely, ptt , wtt , rtt , qtt , αtt .

6.8.1. Recursive price system Prescott and Mehra [1980] and Lucas [1982, JME] extensively utilized recursive formulas expressed in terms of one period forward state contingent claims prices. Counterparts to their recursive pricing formulas are easy to express for our framework. In particular, one-period forward claims on consumption are priced by the function ptt+1 = Mc xt+1 /¯ ej Mc xt . At time t , claims on consumption j -period forward are priced by ej Mc xt . ptt+j = Mc xt+j /¯ Evidently, ptt+j can be built up recursively using the equality ¯j ptt+1 ptt+j = pt+1 t+j e =

Mc xt+j e¯j Mc xt+1

e¯j

Mc xt+1 . e¯j Mc xt

This is a version of a recursive pricing formula often used in formulations of recursive competitive equilibria.

Summary of Pricing Formulas

129

6.8.2. Non-Gaussian asset prices The time t value of a permanent claim to a stream ys = Ua xs , s ≥ t is given by at = (x′t µa xt + σa )/(¯ ej Mc xt ) (6.8.7) where µa and σa satisfy (6.6.6) and (6.6.7) with Za = Ua′ Mc . Notice how (6.8.7) makes the asset price at a nonlinear function of the state vector xt . Suppose, for example, that the wt process is Gaussian. This implies that the equilibrium xt process given by is a multivariate normal process. Even so, the asset prices determined by (6.8.7) are not normally distributed, being determined as the ratio of a quadratic form in the Gaussian state vector xt to a linear form in xt . Thus, the asset prices generated by this “most Gaussian of economies” are highly nonlinear stochastic processes. The term structure of interest rates on perfectly safe claims on the first consumption good j periods ahead is characterized by the gross interest factors Rjt = β j e¯1 · Mc Aoj xt /[¯ ej Mc xt ],

j ≥ 1, t ≥ 0

(6.8.8)

which generalizes (6.7.2).

6.9. Summary of Pricing Formulas For convenience, we now summarize our formulas for the competitive equilibrium price system. They are: (6.58)

ej Mc xt ], s ≥ t pts = Mc xs /[¯

(6.59)

wst =| Sg xs | /[¯ ej Mc xt ], s ≥ t

(6.60)

rst = Γ′ Md xs /[¯ ej Mc xt ], s ≥ t

(6.61)

qst = Mi xs /[¯ ej Mc xt ], s ≥ t

(6.62)

αst = Md xs /[¯ ej Mc xt ], s ≥ t

130

A Competitive Economy

υt = [Γ′ Md + ∆′k Mk ] xt /¯ ej Mc xt

(6.63)

The asset that entitles the owner to the stream of returns yt = Ua xt is priced according to at = (x′t µa xt + σa )/[¯ ej Mc xt ]

(6.64) where (6.51)

µa =

∞ X

β τ (Ao′ )τ Za Aoτ

τ =0

(6.52)

σa =

∞ X β trace Za β τ (Ao )τ CC ′ (Ao′ )τ 1−β τ =0

Za = Ua′ Mc The term structure of interest rates is determined by (6.65)

Rjt = β j e¯1 Mc Aoj xt /[¯ ej Mc xt ],

which gives the price at t of a sure claim on the first consumption good j periods ahead.

6.10. Asset Pricing Example We 3 use the simple pure exchange one good model that is contained in clex14.m to illustrate our asset pricing formulas. The economy in clex14.m is a linearquadratic version of an economy that Lucas (1978) used to develop an equilibrium theory of asset prices. The economy is a member of the special class of structures described in chapter 3. The economy is described as follows:

3 Note to Sargent. This section was 0 in /mnt2/linquad on the SUN. This file is hsch5in.tex. The figures are in hsch*.ps. The MATLAB program used to generate this is hschap5.m. A diary of these runs is in hsch5 in the /linquad directory.

Asset Pricing Example

131

6.10.1. Preferences −.5E

∞ X t=0

β t [(ct − bt )2 + ℓ2t ]|J0

st = ct bt = Ub zt

6.10.2. Technology ct = d1t kt = δk kt−1 + it

gt = φ1 it , d1t = Ud zt 0

φ1 > 0

6.10.3. Information 0 0 1 0 0 zt+1 = 0 .8 0 zt + 1 0 wt+1 0 1 0 0 .5 Ub = [ 30 0 0 ] 5 1 0 Ud = 0 0 0

x0 = [ 5

150

1 0

′

0]

To compute the asset prices in this economy we issue the following MATLAB commands: clex14 solvea t1 = 100; nt = t1; sy=sc; asimul pay=sd(1,:) asseta

132

A Competitive Economy

The program asseta constructs a simulation of length nt of the price and rate of return of an asset that yields a stream of returns equal to pay ∗ xt , where the user specifies the matrix pay . Here we specified that pay = sd(1, :), so that we are pricing a perpetual claim on the endowment process d1t , which is the asset that Lucas priced in his 1978 paper. If the user desires to price a vector of assets, he should simply feed in the matrix pay such that pay ∗ xt is the payout vector of those assets. Let nn be the number of rows of pay , i.e., nn is the number of assets being priced. The program asseta creates a vector y of length nt that equals the vector [mrs, payoff, asset prices, returns], where mrs is the one period intertemporal marginal rate of substitution; payoff is the payoff on the asset(s), which equals pay ∗ xt ; asset prices is the series of asset prices; and ret is the one period gross realized rate of return on the asset(s). For j = 1, 2, 5, the program also creates the reciprocals of the j-period ahead gross rates of return on safe assets, and stores them in the vectors R1, R2, R5. We have computed asset prices for two versions of this economy. The first has the parameter settings listed above, while the second alters the autoregressive parameter in the endowment process to be .4 rather than .8. Figures 6.10.1 through 6.10.3 record the results of one hundred period simulations for each of these two economies. Figure 6.10.1 displays the simulated value of the asset price for the first economy. Figures 6.10.2 displays the gross rates of return on the ‘Lucas tree’ and on a sure one-period bond. We computed the correlation coefficient between these two returns to be -.49. For this economy, the ‘risk premium’ term in the price of the Lucas tree, namely σa in formula (6.8.7), is calculated to be -12.80. To give an idea of how the term structure of interest rates moves in this economy, Figure 6.10.2.b displays the net rates of return on one period and five period sure bonds. (We computed the net rate of return on j -periods bonds by taking the log of the gross rate of return and dividing by j .) Notice the tendency of the term structure to slope upward when rates are low, and to slope downward when rates are high. Figures 6.10.3.a and 6.10.3.b record rates of return for the ‘Lucas tree’ and for sure bonds in the economy with the autoregressive parameter for the endowment process equaling .4. Figure 6.10.3.a shows the gross rates of return on the ‘Lucas tree’ and on a sure one-period bond. The correlation between these two was computed to be -.62. From Figure 6.10.3.b, we see that the tendency for the yield curve to slope upward when rates are low and to slope downward when rates are high has been accentuated relative to our first economy. For the

Asset Pricing Example

133

125 120 115 110 105 100 95 90 85 0

10

20

30

40

50

60

70

80

90

100

Figure 6.10.1: Price of a ‘stock’ entitling the owner to a perpetual claim on the dividends of a ‘Lucas tree’ when the autoregressive parameter for the endowment process is .8.

1.2

0.065

0.06

1.15

0.055 1.1 0.05 1.05 0.045 1 0.04 0.95

0.9 0

0.035

10

20

30

40

50

60

70

80

90

100

Fig. 6.10.2.a. Realized one period gross rates of return on a Lucas tree (solid line) and on a sure one period bond (dotted line) when the autoregressive parameter for the endowment process is .8.

0.03 0

10

20

30

40

50

60

70

80

90

100

Fig. 6.10.2.b. Net rates of return on a one-period (solid line) and a five period (dotted line) when the autoregressive parameter for the endowment process is .8.

134

A Competitive Economy

second economy, the ‘risk premium’ term σa in the price of the Lucas tree is calculated to be -5.90.

1.12

0.09

1.1

0.08 0.07

1.08

0.06 1.06 0.05 1.04 0.04 1.02

0.03

1

0.98 0

0.02

10

20

30

40

50

60

70

80

90

100

Fig. 6.10.3.a. Realized one period gross rates of return on a Lucas tree (solid line) and on a sure one period bond (dotted line) when the autoregressive parameter for the endowment process is .4.

0.01 0

10

20

30

40

50

60

70

80

90

100

Fig. 6.10.2.b. Net rates of return on a one-period (solid line) and a five period (dotted line) when the autoregressive parameter for the endowment process is .4.

The pure exchange economy in clex14.m is one of the simplest to which our asset pricing formulas can be applied. Indeed, for this simple economy, the pricing formulas can be worked out by hand, as the exercises at the end of this chapter indicate. In chapter 4, we shall apply these formulas and our computer programs in much richer contexts in which one can’t get very far ‘by hand’

Exercises

135

6.11. Exercises 1. Consider an economy that consists of technology specification 1 and preference specification 1. The social planning problem is simply to maximize E0 −

∞ X t=0

βt {

1 (ct − bt )2 } 2

subject to ct = dt Assume that bt = ¯b > 0 for all t and that dt = ξ0 + ξ1 dt−1 + εdt , where εdt is √ a white noise with mean zero and variance σε2 , ξ0 > 0, and | ξ1 |< 1/ β . The endowment process {dt } is produced by “trees”, there being one tree for each (representative) household. The household owns the “tree” at the beginning of time (time t = 0). a. Carefully define a competitive equilibrium for this economy. In your definition, describe a particular decentralization scheme, being careful to tell who owns what and who trades with whom. b. Calculate the time-zero equilibrium price system. c. Let υ0 be the time zero value in terms of the time zero consumption good of a title to the entire stream of dividends {dt }∞ t=0 . Prove that υ0 satisfies ∞ ∞ i h X X β t d2t /(b − d0 ) β t ¯b dt − E0 υ 0 = E0 t=0

t=0

d. Compute the gross one period sure rate of interest. e. Compute the gross two period sure rate of interest. 2. Consider an economy with preference specification 1. The technology is specified as ct = dt − Gt where Gt is government purchases. We assume that Gt = UG zt , and dt = Ud zt . Assume that bt = ¯b > 0 for all t . Assume that the government levies state-contingent lump sum taxes τt on the household at time t , where τt = τt (wt , x0 )

136

A Competitive Economy

where wt = (w1 , w2 , . . . , wt ). Lump sum taxes τt are denominated in units of the time t consumption good. a. Formulate the government’s time zero budget constraint. b. Define a competitive equilibrium for this economy. c. Compute a time zero equilibrium price system. d. Define a formula like the one derived under (c) in problem 1 for the time zero value of a title to the dividends from the tree. e. Suppose that the lump sum taxes are on trees, not on the household. Derive a formula for the value of a tree at time zero. f. Compute the gross interest rate on sure one period loans. 3. Consider an economy defined by the social planning problem: maximize the utility of the representative household (1)

−

∞ X 1 E β t (ct − b)2 + ℓ2t | J0 , 0 < β < 1, b > 0 2 t=0

subject to the technology (2)

ct = dt + φgt ,

(3)

gt = ℓt .

φ>0

Here dt is an exogenous process describing the flow of dividends from a single tree (per representative household). The dividend obeys the stochastic process d0 given ,

b > d0 > 0

dt = d0 + wt ,

t≥1

where wt is an independently and identically distributed random process with Ewt = 0 2 Ewt2 = σw .

In (1), b is a constant, ct is consumption at t , and ℓt is labor supplied at t ; E is the mathematical expectation operator, and J0 is information available at

Exercises

137

t = 0, namely, d0 . Equation (2) describes how consumption is related to the exogenous level of dividends at t and the amount φgt = φℓt produced through the application of labor. a. Solve the social planning problem, finding the optimal strategy for consumption and the labor supply. Now decentralize the economy as follows. Let households own the stock of one tree initially. Households sell the tree to a representative firm at time zero (before d0 has been realized). Households sell their labor to the firm each period. The firm buys the tree at the beginning of time zero, hires labor, and sells output to the household. b. State the maximum problems of the representative household and the representative firm for the decentralized economy. c. Find a representation for the time zero Arrow-Debreu price system that supports the solution of the social planning problem as a competitive equilibrium. Give formulas for the price of consumption goods and for the wage of labor. d. Derive a formula for the time zero price of trees in terms of the parameters of preferences, technology, and stochastic process for dividends. (Get as far you can in deriving a closed form). e. Give a formula for the gross interest rate on sure one period loans.

Chapter 7 Applications

7.1. Introduction

7.2. Partial Equilibrium Interpretation The models studied in this book can be reinterpreted as partial equilibrium models which employ the notion of a representative firm, and which generalize the preference and technology specifications of Lucas and Prescott (1971). The idea is that there is a large number of identical firms that produce the same goods and sell them in competitive markets. Because they are all identical, we carry along only one of these firms, and let it produce the entire output in the industry (which is harmless under constant returns to scale). But we have to be careful in our analysis because this representative firm’s decisions play two very different roles: as a stand–in for the ‘average’ competitive producer, and as producer of the entire industry’s output. We make the firm act as a competitor in solving its optimum problem. Demand is governed by the system (11.3.14), with p0t simply being replaced by pt , namely, ct = −Π−1 Λht−1 + Π−1 bt − Π−1 Et {Π′ −1 − Π′ −1 Θ′h

[I − (∆′h − Λ′ Π′ −1 Θ′h )βL−1 ]−1 Λ′ Π′−1 βL−1 }pt

(7.2.1)

ht = ∆h ht−1 + Θh ct

Here ct is a vector of consumption goods. Through this demand system, the representative firm’s output decisions influence the evolution of the market price. However, we want the representative firm to ignore this influence in making its output decisions. A representative firm takes as given and beyond its control the stochastic process {pt }∞ t=0 . The firm sells its output ct in a competitive market each – 139 –

140

Applications

period. Only spot markets convene at each date t ≥ 0. The firm also faces an exogenous process of cost disturbances dt . The firm chooses stochastic processes {ct , gt , it , kt }∞ t=0 to maximize E0

∞ X t=0

β t {pt · ct − gt · gt /2}

subject to Φc ct +Φi it + Φg gt = Γkt−1 + dt kt = ∆k kt−1 + Θk it

(7.2.2)

given k−1 . This problem is not well posed until we describe perceived laws of motion for the processes {pt , dt }∞ t=0 that the firm does not control, but which influence its returns. Specifying the law for the exogenous process {dt } is easy, because the representative firm’s decisions are assumed not to influence it. The situation is different with the price process, because the price is influenced by the output decisions of the representative firm. Despite this influence, we want the firm to behave competitively, that is, to regard the price process as beyond its control. We want to specify the firm’s beliefs about the evolution of the price so that: (a) the firm has ‘rational expectations’, i.e., its beliefs about the evolution of prices allow it to forecast future prices optimally, given the information that it has at each moment; and (b) the firm acts competitively and treats the price process as given and beyond its control. We assume that the firm takes as given a law of motion for spot prices and for the information variables that help to predict spot prices. We model this forecasting problem as follows. The firm observes the state of the market Xt at t , and believes that the law of motion for the spot price is p t = mp X t Xt+1 = ap Xt + Cwt+1

(7.2.3)

′ where Xt = [h′t−1 , Kt−1 , zt′ ]′ , where Kt is the market-wide capital stock, which the firm takes as given and beyond its control. The firm believes that the cost shock process evolves according to dt = Sd Xt . The state for the firm at date t is ′ x ˜t = [Xt′ , kt−1 ]′ .

The firm’s problem is a discounted linear regulator problem. Under our assumptions about the technology, the firm’s control can be taken to be it . The

Partial Equilibrium Interpretation

141

solution of the firm’s problem is a decision rule for investment of the form it = −fi x ˜t .

(7.2.4)

This decision rule and equations (7.2.2) then determine [ct , gt , kt ] as linear functions of x ˜t . The matrix fi in the above equation is a function of all of the matrices describing the firm’s constraints, including ap and mp . The firm’s rule for ct , implied by (7.2.2) and (7.2.4) can be represented as ct = fc x ˜t .

(7.2.5)

Equation (7.2.2) implies that the firm’s capital evolves according to kt = ∆k kt−1 − Θk fi x ˜t .

(7.2.6)

At this point, but not earlier, impose that the ‘representative firm is representative’ by setting kt ≡ Kt in (7.2.6), use it to deduce the actual law of motion for Kt , and then use this to fill in the rows corresponding to Kt of the actual law of motion for Xt : Xt+1 = aa Xt + Cwt+1 . (7.2.7) To get the rows corresonding to ht , use (7.2.5) together with the law of motion ht = ∆h ht−1 + Θh ct . To get a formula for the actual law of motion of the price, use (11.1.1) and the actual law of motion (7.2.7) for xt = Xt to solve for a consumption process. Put the consumption process and preference shock into (11.3.1) and solve for µst . Then solve (11.3.3) forward for µht ; substitute into (11.3.2) to solve for p0t . Set pt = p0t , then express the motion of prices as p t = ma X t .

(7.2.8)

The system (7.2.7), (7.2.8) describes the actual law of motion for spot prices that is induced by the firm’s optimizing behavior and market clearing when the firm’s perceived law of motion for the spot prices is (7.2.3). The firm’s optimization problem and market clearing thus induce a mapping from a perceived law of motion (ap , mp ) for spot prices to an actual law (aa , ma ). Definition: A rational expectations equilibrium (or a partial equilibrium) is a fixed point of the mapping from the perceived law of motion for spot prices to the actual law of motion for spot prices.

142

Applications

An equivalent definition is: Definition: A partial equilibrium is a stochastic process {pt , ct , it , gt , kt , Kt , ht }∞ t=0 , 2 each element of which belongs to L0 , such that: i. Given {pt }∞ t=0 , in particular given the law of motion (7.2.3), {ct , it , ∞ gt , kt }t=0 solve the firm’s problem. ii. {pt , ct , ht }∞ t=0 satisfy the demand system (7.2.1). ∞ iii. {kt }∞ t=0 = {Kt }t=0 .

This is a version of Lucas and Prescott’s (1971) rational expectations competitive equilibrium, which they used to study investment under uncertainty with adjustment costs. The following proposition states the relationship between a partial equilibrium and our earlier notion of competitive equilibrium: Proposition: Let {ct , st , it , gt , kt , p0t , wt0 , αt0 , rt0 , qt0 }∞ t=0 , v0 be a competitive equi0 ∞ librium. Then {pt , ct , it , gt , kt , kt , ht }t=0 is a partial equilibrium. This proposition can be proved directly by verifying that the proposed partial equilibrium satisfies the first order necessary and sufficient conditions for the firm’s problem in the partial equilibrium, and that the proposed {pt , ct , ht }∞ t=0 process satisfies the demand system (11.3.14).

7.2.1. Partial equilibrium investment under uncertainty Our partial equilibrium structure includes many examples of linear rational expectations models (e.g., Sargent (1987, chapter XVI), Eichenbaum (1983), and Hansen and Sargent (1991, Two Difficulties). Here is how we can apply these ideas to a version of Lucas and Prescott’s (1971) model of investment under uncertainty. There is one good produced with one factor of production (capital) via a linear technology. A representative firm maximizes E

∞ X t=0

β t {pt ct − gt2 /2},

subject to the technology ct = γkt−1 kt = δk kt−1 + it gt = f1 it + f2 dt ,

Partial Equilibrium Interpretation

143

where dt is a cost shifter, γ > 0, and f1 > 0 is a cost parameter and f2 = 1. Demand is governed by pt = α0 − α1 ct + ut , where ut is a demand shifter with mean zero and α0 , α1 are positive parameters. Assume that ut , dt are uncorrelated first-order autoregressive processes. Lucas and Prescott computed rational expectations equilibrium quantities by forming a social planning problem with criterion E

∞ X t=0

β

t

Z

0

ct

(α0 − α1 ν + ut )d ν −

.5gt2

,

where the integral under the demand curve is ‘consumer surplus.’ Consumer surplus equals α1 2 (α0 + ut )ct − c . 2 t To map this model into our framework, set Λ = 0, ∆h = 0, Θh = 0, Π2 = α1 , bt = αΠ0 + Π1 ut . Notice that with this specification, (st − bt )2 /2 = (α0 + ut )ct −

α1 2 c + b2t /2. 2 t

The term in b2t can be ignored because it influences no decisions. With this specification, our social planning problem is equivalent with Lucas and Prescott’s. After we have computed the equilibrium quantities by solving the social planning problem, we can compute the ‘marginal utility price’ pt = Π(bt − st )

= α0 + ut − α1 ct ,

where we are using α1 = Π2 .

144

Applications

7.3. Introduction The remainder of chapter provides more examples of models that conform to our framework. Most of these examples were originally stated as partial equilibrium models. The appendix of the chapter describes a scheme for pricing objects that until now were unpriced because they were sheltered from the market by the warmth of the household. We use this decentralization when we want to price some of the household capital stocks.

7.4. A Housing Model Rosen and Topel (1988) formulated a partial equilibrium model of a housing market consisting of a linear demand curve relating a stock of housing inversely to a rental rate; an equilibrium condition relating the price of houses to the discounted present value of rentals, adjusted for depreciation; and a quadratic cost curve for producing houses.

7.4.1. Demand We can capture Rosen and Topel’s specification by sweeping house rentals into the household sector. See the appendix of this chapter for an account of a decentralization that supports this interpretation. Rosen and Topel expressed the demand side of their model in terms of the two equations Rt = bt + αht ∞ X p t = Et (βδh )τ Rt+τ τ =0

where ht is the stock of housing at time t , Rt is the rental rate for housing, pt is the price of new houses, and bt is a demand shifter; α < 0 is a demand parameter, and δh is the depreciation factor for houses. We cast this demand specification within our class of models by letting the stock of houses ht evolve according to ht = δh ht−1 + ct , δh ∈ (0, 1), where ct is the rate of production of new houses. Houses produce services st ¯ t or st = λht−1 + πct , where λ = λδ ¯ h , π = λ. ¯ We can take according to st = λh

A Housing Model

145

¯ 0 = Rt as the rental rate on housing at time t , measured in units of time t λρ t consumption (housing). Demand for housing services is st = bt − µ0 ρ0t , where the price of new houses pt is related to ρ0t by ρ0t = π −1 [pt − βδh Et pt+1 ]. This equation, which is a special case of equation (11.3.7) from chapter 11, imposes the feature of the present specification that δh − λπ −1 θh = 0, is a version of Rosen and Topel’s equation (12). It can be solved to yield pt = ¯ t P∞ (βδh )τ ρ0 , a version of Rosen and Topel’s equation (14). The parameλE t τ =0 ¯ governs the slope of the demand curve for housing, in terms of the rental ter λ rate for housing.

7.4.2. House producers Rosen and Topel’s representative firm maximizes E0

∞ X t=0

β t [pt ct − Ω(ct , ct − ct−1 , et )],

where Ω(ct , ct − ct−1 , et ) is the cost of producing new houses, and {et } is a cost shifter. The function Ω incorporates costs of adjusting the rate of production of new houses. The firm takes the stochastic process for pt as given. Costs are given by Ω(ct , ct − ct−1 , et ) = gt · gt where

g1t = f1 ct + f2 et g2t = f3 (ct − ct−1 ),

where et is our cost-shifter. To map this into our specification, we use the technology f1 ct − g1t = 0kt−1 − f2 et ct − it = 0

f3 ct − g2t = f3 kt−1

kt = 0kt−1 + it .

146

Applications

7.5. Cattle Cycles Rosen, Murphy, and Scheinkman (1994) used a partial equilibrium model to interpret recurrent cycles in U.S. cattle prices. Their model has a static linear demand curve interacting with a ‘time-to-grow’ structure for raising cattle. Let pt be the price of freshly slaughtered beef, mt the feeding cost of preparing an ˜ t the one-period holding cost for a mature animal, γ1 h ˜t animal for slaughter, h ˜ the one-period holding cost for a yearling, and γ0 ht the one period holding cost ˜ t , mt }∞ are exogenous stochastic processes, while the for a calf. The costs {h t=0 stochastic process {pt }∞ is determined by a rational expectations equilibrium. t=0 Let x ˜t be the breeding stock, and y˜t be the total stock of animals. The law of motion for stocks is x ˜t = (1 − δ)˜ xt−1 + g˜ xt−3 − ct , (7.5.1) where ct is a rate of slaughtering. The total head count of cattle is y˜t = x ˜t + g˜ xt−1 + g˜ xt−2 ,

(7.5.2)

which is the sum of adults, calves, and yearlings, respectively. A representative farmer maximizes E0

∞ X t=0

˜ tx ˜ t )(g˜ ˜ t )(g˜ β t {pt ct −h ˜t − (γ0 h xt−1 ) − (γ1 h xt−2 ) − mt ct

(7.5.3)

− Ψ(˜ xt , x ˜t−1 , x ˜t−2 , ct )},

where

ψ3 2 ψ4 2 ψ1 2 ψ2 2 x ˜ + x ˜ + x ˜ + c . (7.5.4) 2 t 2 t−1 2 t−2 2 t The maximization is subject to the law of motion (7.5.1), taking as given the stochastic laws of motion for the exogenous random processes and the equilibrium price process, and the initial state [˜ x−1 , x ˜−2 , x ˜−3 ]. Here (ψj , j = 1, 2, 3) are small positive parameters that represent quadratic costs of carrying stocks, and ψ4 is a small positive parameter. The costs in (7.5.4) are implicitly taken into account by Rosen, Murphy, and Scheinkman, and motivate their decision to “solve stable roots backwards and unstable roots forwards.” To capture Rosen, Murphy, and Scheinkman’s solution, we shall set each of the φj ’s to a positive but very small number. Demand is governed by Ψ=

(5)

ct = α0 − α1 pt + d˜t ,

Cattle Cycles

147

where α0 > 0, α1 > 0, and {d˜t }∞ t=0 is a stochastic process with mean zero representing a demand shifter.

7.5.1. Mapping cattle farms into our framework We show how to map the model of Rosen, Murphy, and Scheinkman into our general setup.

7.5.2. Preferences Set Λ = 0, ∆h = 0, Θh = 0, Π = α1−1 , bt = Πd˜t + Πα0 . With these settings, first-order condition (6.13) for the household’s problem becomes ct = Π−1 bt − Π−2 pt , or ct = α0 − α1 pt + d˜t .

7.5.3. Technology The law of motion for capital is (1 − δ) 0 x ˜t x 0 ˜t−1 = 1 0 1 x ˜t−2 or

1 g x ˜t−1 0x ˜t−2 + 0 it , 0 0 x ˜t−3

kt = ∆k kt−1 + Θh it . Here it = −ct . We use adjustment costs to capture the holding and slaughtering costs. We set ˜ t, g1t = f1 x ˜t + f2 h or ˜ t. g1t = f1 [(1 − δ)˜ xt−2 + g˜ xt−3 − ct ] + f2 h

148

Applications

We set

Notice that

˜t g2t = f3 x ˜t−1 + fr h ˜ t. g3t = f5 x ˜t−1 + f6 h 2 ˜ 2 + 2f1 f2 x ˜t g1t = f12 x ˜2t + f2 h ˜t h t 2 2 2 2 ˜ + 2f3 f4 x ˜t g =f x ˜ + f2 h ˜t−1 h 2t 2 g3t

Thus, we set

=

3 t−1 f52 x ˜2t−2

t

˜ 2 + 2f5 f6 x ˜ t. + f6 h ˜t−2 h t

ψ1 ψ2 ψ3 f22 = f32 = 2 2 2 2f1 f2 = 1 2f3 f4 = γ0 g 2f5 f6 = γ1 g f12 =

To capture the feeding costs we set g4t = f7 ct +f8 mt , and set f72 = ψ24 1. Thus, we set 0 0 0 0 1 1 1 0 0 0 g1t 0 f 1 g 2t 0 ct + 0 it + 0 1 0 0 g 0 0 1 0 3t 0 0 g4t 0 0 0 1 0 −f7 0 0 0 0 ˜ f (1 − δ) 0 gf x ˜t−1 1 1 f2 ht ˜ = f3 0 0 x ˜t−2 + f4 h t . ˜ f6 ht 0 f5 0 x ˜t−3 0

We set dt = Ud zt where

0

0

2f7 f8 =

f8 mt

0 f2 Uh Ud = f4 Uh , f6 Uh f8 Um

˜ t and mt from the exogenous where [Uh , Um ] are selector vectors that pick off h state vector zt . We specify the information matrices [A22 , C2 ] to incorporate ˜ t , mt , d˜t ] consists of three Rosen, Murphy, and Scheinkman’s specification that [h uncorrelated first order autoregressive processes. 1 1 This model is estimated by Anderson, Hansen, McGrattan, and Sargent (1996).

Models of Occupational Choice and Pay

149

7.6. Models of Occupational Choice and Pay Aloyisius Siow (1984) and Sherwin Rosen (1995) and have used pure ‘timeto-build’ structures to represent entry cycles into occupations, and also interoccupational wage movements. It is easiest to incorporate these models into our framework by putting production into the household technology, using the decentralization described in the appendix to generate prices.

7.6.1. A one-occupation model Rosen [1995] studied a partial equilibrium model determining a stock of ‘engineers’ Nt ; the number of new entrants into engineering school, nt ; and the wage level wt of engineers. It takes k periods of schooling to become an engineer. The model consists of the following equations: first, a demand curve for engineers wt = −αd Nt + ǫ1t , αd > 0;

(7.6.1)

second, a time-to-build structure of the education process Nt+k = δN Nt+k−1 + nt , 0 < δN < 1;

(7.6.2)

third, a definition of the discounted present value of each new engineering student ∞ X v t = β k Et (βδN )j wt+k+j ; (7.6.3) j=0

and fourth, a supply curve of new students driven by vt nt = αs vt + ǫ2t , αs > 0.

(7.6.4)

Here {ǫ1t , ǫ2t } are stochastic processes of labor demand and supply shocks. A partial equilibrium is a stochastic process {wt , Nt , vt , nt }∞ t=0 satisfying these four equations, and initial conditions N−1 , n−s , s = 1, . . . , −k .

We can represent this model by sweeping the time-to-build structure and the demand for engineers into the household technology, and putting the supply of new engineers into the technology for producing goods. Here is how. We take

150

Applications

the household technology to be

h1t h2t .. .

st = [λ1 0 . . . 0]

δN 0 . . = . hk,t 0 0 hk+1,t

bt = ǫ1t

h1t−1 h2t−1 .. .

hk+1,t−1

1 0 .. .

0 1 .. .

··· ··· .. .

··· 0

··· 0

0 ···

+ 0 · ct

0 0 .. . 1 0

h 1t−1 h2t−1 .. . hk,t−1

0 0 . + .. ct 0 1 hk+1,t−1

This specification sets Rosen’s Nt = h1t−1 , nt = ct , hs+1,t−1 = nt−s , s = 1, . . . , k , and uses the home-produced good to capture the demand for labor. Here λ1 embodies Rosen’s demand parameter αd . 2 To capture Rosen’s supply curve, we use the physical technology ct = it + d1t ϕ1 it = gt where d1t is proportional to Rosen’s supply shock ǫ2t , and where the adjustment cost parameter ϕ1 varies directly with Rosen’s supply curve parameter αs . Rosen showed that the equilibrium decision role for new entrants (our ct ) must satisfy the condition nt = f1 Et Nt+k + f2 ǫ1t + f3 ǫ2t where f1 < 0.

2 In the definition of Λ in the household technology, we would replace the zeros with ε > 0 as a trick to acquire detectability; see chapter 9 and its appendix for the definition and role of detectability.

Models of Occupational Choice and Pay

151

7.6.2. Skilled and unskilled workers We can generalize the preceding model to two occupations, called skilled and unskilled, to obtain alternative versions of a model estimated by A. Siow (1984). The model consists of the following elements: first, a demand curve for labor

wut wst

= αd

Nut + ǫ1t ; Nst

where αd is a (2 × 2) matrix of demand parameters and ǫ1t is a vector of demand shifters; second, time-to-train specifications for skilled and unskilled labor, respectively: Nst+k = δN Nst+k−1 + nst Nut = δN Nut−1 + nut ; where Nst , Nut are stocks of the two types of labor, and nst , nut are entry rates into the two occupations; third, definitions of discounted present values of new entrants to the skilled and unskilled occupations, respectively: vst = Et β

k

∞ X

(βδN )j wst+k+j

j=0

vut = Et

∞ X

(βδN )j wut+j ,

j=0

where wut , wst are wage rates for the two occupations; and fourth, supply curves for new entrants: nst vut = αs + ǫ2t . (7.6.5) nut vst As an alternative to (7.6.5), Siow simply used the ‘equalizing differences’ condition vut = vst . (7.6.6) We capture this model by pushing most of the ‘action’ into the household sector. Households decide what kind of durable good to accumulate, namely, unskilled labor or skilled labor. Unskilled labor and skilled labor can be combined to produce services, which we specify to generate the demands labor. We let c1t , c2t be rates of entry nut , nst into unskilled and skilled labor, and constrain

152

Applications

them to satisfy c1t + c2t = it + d1t , the rate of total new entrants. To generate the upward sloping supply curves (7.6.5), we specify that φ1 it + φ2 c2t = gt . The technology is thus 1 0 0

1 −1 c1t −φ2 + −φ1 c2t 1 0

d1t 0 0 0 i1t 0 + 1 gt = 0 kt−1 + 0 , i2t 0 0 −1 0

where d1t is a supply shifter. To get Siow’s model, we set φ1 = φ2 = 0, in which case d1t becomes an exogenous supply of new entrants into the labor force. We specify the law of motion for household capital

h1t h2t h3t .. .

hk+1,t hk+2,t

δN 0 0 = . .. 0

0

0 δN 0

0 1 0

0 ··· 0 ··· 1 ···

0 0

0 0

0 ··· 0 ···

h 0 1 1t−1 h2t−1 0 0 0 0 h3t−1 + .. .. . . 1 hk+1,t−1 0 0 0 hk+2,t−1

0 0 0 c1t . .. c2t . 0 1

where h1t−1 = Nut−1 , h2t−1 = Nst , hj+2,t−1 = ns,t−j , j = 1, . . . , k . We generate the demand for labor by specifying services as

s1t s2t

¯ h1t = Λe∆ ¯ h ht−1 + ΛeΘ ¯ h ct =Λ h2t

shock process bt = [ b1t labor.

b2t

0

h1t = eht We set the preference h2t ′ 0 ] to capture the shifters in the demands for

where e is a selector vector that verifies

A Cash-in-Advance Model

153

7.7. A Cash-in-Advance Model We want to use our framework to mimic a situation in which households are in a cash-in-advance environment in which they face a sequence of budget constraints mt mt−1 c˜t + + Bt ≤ yt + Rgt−1 Bt−1 + pt pt mt−1 c˜t ≤ pt Here Rgt−1 is the gross rate of return on indexed bonds Bt−1 held from t − 1 to t ; pt is the price level at t ; c˜t is time t consumption; and mt is currency held from t to t + 1. The household’s preferences P∞ are ordered by E0 t=0 β t u(˜ ct ). (We use c˜t to denote consumption in order to separate this notation from the ct of our framework, which is soon to be defined.) Using the cash-in-advance constraint at equality in the budget constraint gives Bt +

c˜t+1 = yt + Bt−1 Rgt−1 , Rt

pt where Rt = pt+1 is the gross rate of return on currency between t and t+1. The force of the cash-in-advance restriction is that decisions about time t moneyholding influence time t + 1 consumption c˜t+1 , but time t consumption is predetermined.

7.7.1. Reinterpreting the household technology We can specify the household technology to capture key elements of the cash-inpt advance specification. We can use a ‘back-solving’ approach, and let R = pt+1 be a constant rate of return on currency. We shall set c˜t = st and ct = mt /pt , and sweep the cash-in-advance specification into a one-period time delay between a decision to consume (i.e., hold real balances) and when consumption goods are actually enjoyed. Thus, we take the household technology to be st = Rct−1 , which we accomplish by taking Λ = R, Π = 0, δh = 0, θh = 1. When there is inflation, R < 1. When R > 1, there is deflation. Preferences are of the usual kind ∞ X β t [(st − bt ) · (st − bt ) + ℓ2t ]. −.5E0 t=0

154

Applications

With these matchups, the time t ‘seignorage’ component of government revenue is mt − mt−1 r˜t = pt = ct − st . This means that given R , the present value of seignorage revenues can be computed using the methods in chapter 8. With these specifications, an equilibrium with present value government budget balance can be computed, possibly including the inflation rate parameter R along with some of the τ ’s over which we search for an equilibrium. Once an equilibrium is computed, the time series for real balances can be found from mt = ct , pt and the price level and nominal level of currency can be computed using the assumed R .

7.8. Taxation in a Vintage Capital Model Owens (1994) has studied the effects of taxation on prices of new and old commercial buildings. His analysis requires keeping track of the age distribution of capital, which we can accomplish by specifying, for example, k1t 0 0 0 0 k1t−1 1 k2t δh 0 0 0 k2t−1 0 k3t = 0 δh 0 0 k3t−1 + 0 it . k4t

0

0

δh

1

k4t−1

0

Here k1t is new capital, k2t is one year old capital, and so on. We could also include a time-to-build aspect, but have not here. To differentiate among the services produced by capital of different ages, we specify ct = Γc kt−1 , where we make ct a vector that is comparable in dimension with kt−1 and Γc is diagonal. ‘Office services’ are then produced according to st = [ π1

π2

...

πn ] ct .

Decentralizing the Household

155

We can set the π vector to make new office space more desirable than old office space. We let the government tax capital of different ages differently, which puts P P terms like E0 β t τk rt0 · kt or E0 β t τkk p0kt · kt−1 in the budget constraints of households and the government, as in the framework of chapter 14. The tax matrices τk , τkk can be chosen to model different kinds of policies for depreciating capital for tax purposes. This framework can be used to model the effects on the prices of capital of alternative policies for capital taxation.

A. Decentralizing the Household It can be useful to decentralize the household sector in order to price household services and stocks. Suppose that the household buys a vector of services from firms of type III at the price of services ρ0t . The household sells its initial stocks of both physical and household capital and also its labor and endowment process to firms. The price of the initial stock of household capital is v˜0 . The household maximizes ∞ X E0 β t [(st − bt ) · (st − bt ) + ℓ2t } t=0

subject to the budget constraint E0

∞ X t=0

β t ρ0t · st = E0

∞ X t=0

β t (wt0 ℓt + αt0 · dt ) + (v0 · k−1 + v˜0 · h−1 ).

(7.A.1)

Firms of type III Firms of type III purchase the consumption vector ct , rent household capital, and produce and sell household services and additions to the stocks of household capital. Type III firms sell st to households at price ρ0t and new household capital Θh ct to firms of type IV at price p0ht . Firms of type III rent 0 household capital from firms of type IV at a rental price rht , and maximize E0

∞ X t=0

0 β t {ρ0t · st + p0ht Θh ct − rht · ht−1 − p0t · ct }

156

Applications

subject to st = Λht−1 + Πct . Firms of type IV Firms of type IV purchase new household capital from firms of type III, and 0 rent existing household capital to firms of type III at rental price rht . Firms of type IV maximize E0

∞ X t=0

0 β t {rht · ht−1 − p0ht Θh ct } − v˜0 · h−1

subject to ht = ∆h ht−1 + Θh ct . Computing Prices If we formulate the optimum for a firm of type III, obtain the associated first order necessary conditions and rearrange, we get the following restrictions on prices: p0t = Θ′h pht + Π′ ρ0t (7.A.2) 0 = Λ′ ρ0t . rht From the first order conditions for a firm of type IV, obtain 0 p0ht = Et β[∆′h p0ht+1 + rht+1 ].

(7.A.3)

We can use (7.A.2) with (7.A.3) to obtain ∞ X 0 p0t = Θ′h Et [ β (j) ∆′(j−1) rht+j ] + Π′ ρ0t . j=1

This is a generalization of Siow’s equilibrium condition (7.6.6). For us p0t = Mc xt is the vector of shadow prices of new entrants into the two types of profession. We have already shown how to compute the price ρ0t . Indeed, this decentralization is a way to set up an explicit market in the ‘implicit’ services priced by ρ0t . The prices of stocks of household capital can be computed from the multipliers on ht−1 and ct in the social planning problem.

Chapter 8 Efficient Computations

8.1. Introduction This chapter describes fast algorithms for computing the value function and the optimal decision rule of our social planning problem. 1 The decision rule determines the allocation. The value function determines competitive equilibrium prices. The optimal value function and the optimal decision rules can be computed by iterating to convergence on the T operator associated with Bellman’s functional equation. These iterations can be accelerated by using ideas from linear optimal control theory. We avail ourselves of these faster methods because we want to analyze high dimensional systems. This chapter is organized as follows. First, we display a transformation that removes both discounting and cross-products between states and controls. This transformation simplifies the algebra without altering the substance. Next we describe invariant subspace methods for solving an optimal linear regulator problem, which are typically faster than iterating on the Bellman equation. We then describe a closely related method called the doubling algorithm, which ‘skips steps’ in iterating on the Bellman equation. The calculations can be further accelerated by partitioning the state vector to take advantage of the pattern of zeros in A and B . Next we discuss fast methods for computing equilibria for periodic economies. We describe the periodic optimal linear regulator problem, and show how to solve it rapidly. We conclude the chapter by describing how our calculations can be adapted to handle Hansen and Sargent’s (1995) recursive formulation of Jacobsen’s and Whittle’s risk-sensitive preferences, which will be used in Chapter @[email protected] This chapter focuses mostly on nonstochastic optimal linear regulator problems. As indicated in chapter 3, the optimal decision rule for a stochastic optimal 1 Parts of this chapter use results described in Anderson, Hansen, McGrattan, and Sargent (1995). Also see Kwakernaak and Sivan [1972] for what is mostly a treatment of continuous time systems.

– 157 –

158

Efficient Computations

linear regulator problem equals the optimal decision rule for the associated nonstochastic optimal linear regulator problem. Furthermore, from chapter 3, the matrices determining the Lagrange multipliers depend only on the piece of the optimal value function associated with the nonstochastic part of our problem. Throughout this chapter, we study solutions of our control problem that satisfy the additional condition E

∞ X t=0

β t (|xt |2 + |ut |2 ) < ∞,

(8.1.1)

where xt is the state and ut is the control. In an appendix, we describe conditions on the matrices determining returns and the transition law that are sufficient by themselves to imply condition (8.1.1). 2

8.2. The Optimal Linear Regulator Problem Consider the following version of the optimal linear regulator problem: choose a contingency plan for {ut }∞ t=0 to maximize −E subject to

∞ X

β t [x′t Rxt + u′t Qut + 2u′t W xt ], 0 < β < 1

(8.2.1)

t=0

xt+1 = Axt + But + Cwt+1 , t ≥ 0,

(8.2.2)

where x0 is given. In (8.2.1) – (8.2.2), xt is an n × 1 vector of state variables, and ut is a k × 1 vector of control variables. In (8.2.2), we assume that wt+1 is a martingale difference sequence with Ewt wt′ = I , and that C is a matrix conformable as required to x and w . We also impose condition (8.1.1). We temporarily assume that R and Q are positive definite matrices, although in practice we use weaker assumptions about both matrices. A standard way to solve this problem is the method of dynamic programming. Let V (x) be the optimal value associated with the program starting from initial state vector x0 = x . Bellman’s functional equation is n o V (xt ) = max −(x′t Rxt + u′t Qut + 2u′t W xt ) + βEt V (xt+1 ) (8.2.3) ut

2 For conditions sufficient to imply this condition, see Kwakernaak and Sivan [1972], Anderson and Moore [1979], and Anderson, Hansen, McGrattan, and Sargent (1995).

The Optimal Linear Regulator Problem

159

where the maximization is subject to (8.2.2). One way to solve this functional equation is to iterate on a version of (8.2.3), thereby constructing a sequence Vj (xt ) of successively better approximations to V (xt ). In particular, let n o Vj+1 (xt ) = max −(x′t Rxt + u′t Qut + 2u′t W xt ) + βEt Vj (xt+1 ) , ut

(8.2.4)

where again the maximization is subject to (8.2.2). Suppose that we initiate the iterations from V0 (x) = 0 (which is the appropriate terminal value function for a one-period problem). Then direct calculations show that successive iterates on (8.2.4) take the quadratic form Vj (xt ) = −x′t Pj xt − ρj ,

(8.2.5)

where Pj and ρj satisfy the equations Pj+1 = R + βA′ Pj A − (βA′ Pj B + W )

× (Q + βB ′ Pj B)−1 (βB ′ Pj A + W ′ )

ρj+1 = βρj + β trace Pj CC ′ .

(8.2.6)

(8.2.7)

Equation (8.2.6) is the matrix Riccati difference equation. Notice that it involves only {Pj } and is independent of {ρj }. Notice also that C , which multiplies the noises impinging on the system and so determines the variances of innovations to information in the system, affects the {ρj } sequence but not the {Pj } sequence. We can say that {Pj } is independent of the system’s noise statistics. 3 Let P and ρ be the limits of (8.2.6) and (8.2.7), respectively. Then the value function V (xt ) that satisfies the Bellman equation (8.2.3) is given by V (xt ) = −x′t P xt − ρ, where P and ρ are the limit points of iterations on (8.2.6) and (8.2.7) starting from P0 = 0, ρ = 0. The decision rule that attains the right side of (8.2.4) is given by ut = −Fj xt 3 This fact is what permits us to focus on nonstochastic problems in devising our algorithms.

160

Efficient Computations

where Fj = (Q + βB ′ Pj B)−1 (βB ′ Pj A + W ′ ).

(8.2.8)

The optimal decision rule for the original problem is given by ut = −F xt , where F = limj→∞ Fj , or F = (Q + βB ′ P B)−1 (βB ′ P A + W ′ ).

(8.2.9)

According to (8.2.9), the optimum decision rule for ut is independent of the parameters C , and so of the noise statistics. The limit point P of iterations on (8.2.6) evidently satisfies P = R + βA′ P A − (βA′ P B + W )

× (Q + βB ′ P B)−1 (βB ′ P A + W ′ )

(8.2.10)

This equation in P is called the algebraic matrix Riccati equation. One way to solve an optimal linear regulator problem is to iterate directly on (8.2.6) and (8.2.7). Faster algorithms are available. First, we describe a useful transformation that simplifies some of the formulas.

8.3. Transformations to eliminate discounting and crossproducts The following transformation eliminates both discounting and cross-products between states and controls. Define the transformed control vt and transformed state x ˆt by vt = β t/2 (ut + Q−1 W ′ xt ), x ˆt = β t/2 xt . (8.3.1) Notice that vt′ Qvt

=β

t

[ x′t

u′t

W Q−1 W ′ ] W′

It follows that β t [ x′t u′t ]

R W

W′ Q

xt ut

W Q

xt . ut

=x ˆ′t R∗ x ˆt + vt′ Qvt

where R∗ = R − W Q−1 W ′ . The transition law (8.2.2) can be represented as x ˆt+1 = A∗ x ˆ t + B ∗ vt + β

t+1 2

Cwt+1

Stability Conditions

161

where A∗ = β 1/2 (A − BQ−1 W ′ ), B ∗ = β 1/2 B . Therefore, regulator problem (8.2.1) – (8.2.2) is equivalent to the following regulator problem without crossproducts between states and controls and without discounting: choose {vt } to maximize ∞ X −E [ˆ x′t R∗ x ˆt + vt′ Qvt ] (8.3.2) t=0

subject to

x ˆt+1 = A∗ x ˆ t + B ∗ vt + β

t+1 2

Cwt+1 ,

(8.3.3)

where ′

P = R∗ + A∗′ P A∗ − A∗′ P B ∗ (Q + B ∗′ P B ∗ )−1 B ∗ P A∗ F ∗ = (Q + B ∗′ P B ∗ )−1 B ∗′ P A∗ ,

(8.3.4) (8.3.5)

it being understood that P is the positive semi-definite solution of (8.3.4). The optimal closed loop system in terms of transformed variables is x ˆt+1 = (A∗ − B ∗ F ∗ )ˆ xt + β Multiplying both sides of this equation by β −( 1

t+1 2

t+1 2 )

Cwt+1

(8.3.6)

gives

xt+1 = β − 2 (A∗ − B ∗ F ∗ )xt + Cwt+1 .

(8.3.7)

8.4. Stability Conditions We shall typically restrict the undiscounted linear regulator (8.3.2), (8.3.3) defined by the matrices (A∗ , B ∗ , R∗ , Q) to satisfy some conditions from control theory designed to render the problem well behaved. In particular, let DD′ = R∗ , so that D is said to be a factor of R∗ . Our conditions are cast in terms of the concepts of stabilizability and detectability defined in Appendix A. We make Assumption A1: The pair (A∗ , B ∗ ) is stabilizable. The pair (A∗ , D) is detectable. Then there obtains:

162

Efficient Computations

Stability Theorem: Under assumption A1: (i.) starting from any negative semi-definite matrix Po , iterations on the matrix Riccati difference equation converge; and (ii.) The eigenvalues of (A∗ − B ∗ F ∗ ) are stable. In the next section, we describe a class of algorithms that exploit the stabilizing property of the optimal (A∗ − B ∗ F ). 4

8.5. Invariant Subspace Methods Following Vaughan [1970], a literature has developed fast algorithms for computing the limit point of the matrix Riccati equation (8.2.6), based on an eigenstructure of a matrix associated with the Riccati equation. These methods work with a Lagrangian formulation of the problem and with the linear restrictions that stability condition (8.1.1) imposes on the multipliers and the state vector. These conditions restrict the matrix P that solves the algebraic matrix Riccati equation. Without loss of generality, we work with the undiscounted deterministic optimal linear regulator problem: choose {ut }∞ t=0 to maximize −

∞ X t=0

{x′t Rxt + u′t Qut }

(8.5.1)

subject to xt+1 = Axt + But .

(8.5.2)

4 Because the eigenvalues of (A∗ − B ∗ F ∗ ) are less than unity in modulus, it follows that 1 the eigenvalues of Ao = β − 2 (A∗ − B ∗ F ∗ ) are less than √1 in modulus. β

Invariant Subspace Methods

163

8.5.1. P x as Lagrange multiplier It is convenient to write a Lagrangian for the Bellman equation: V (x) = max{−(x′ Rx + u′ Qu + 2µ′ [Ax + Bu − x ˜]) + V (˜ x)}, where x ˜ is next period’s value of the state, µ is a vector of multipliers, and V (x) = −x′ P x where the matrix P solves the matrix Riccati equation. The first-order condition for the above Lagrangian with respect to x ˜ implies that µ = P x . Thus, as usual, the multipliers are linked to the gradient of the value function.

8.5.2. Invariant subspace methods Invariant subspace methods compute P indirectly by vector of the multipliers µ to stabilize the solution for (8.1.1). For now, we assume that A is invertible. We sequences, and let {µt }∞ t=0 be a sequence of vectors of Form the Lagrangian J =−

∞ X t=0

restricting the initial xt , ut , as required by move to the space of Lagrange multipliers.

{x′t Rxt + u′t Qut + 2µ′t+1 [Axt + But − xt+1 ]} − 2µ′0 (¯ x0 − x0 ). (8.5.3)

Here x ¯0 is the given initial level of x0 . First order necessary conditions for the ∞ maximization of J with respect to {ut }∞ t=0 and {xt }t=0 are ut :

Qut + B ′ µt+1 = 0 ,

t≥0

(8.5.4)

xt :

µt = Rxt + A′ µt+1 ,

t ≥ 0.

(8.5.5)

Solve (8.5.4) for ut and substitute into (8.5.2) to obtain xt+1 = Axt − BQ−1 B ′ µt+1 .

(8.5.6)

Represent (8.5.5) and (8.5.6) as L

xt+1 µt+1

=N

xt , µt

(8.5.7)

164

Efficient Computations

where

I L= 0

BQ−1 B ′ , A′

A N= −R

0 . I

Represent (8.5.7) as

or

xt+1 µt+1

xt µt

xt µt

(8.5.8)

xt+1 µt+1

(8.5.9)

= Mf

= Mb

where −1

Mf = L

A + BQ−1 B ′ A′ N= −1 −A′ R

and Mb = N

−1

−1

A−1 L= RA−1

R

−BQ−1 B ′ A′ −1 A′

−1

A−1 BQ−1 B ′ . RA−1 BQ−1 B ′ + A′

,

(8.5.10)

(8.5.11)

Evidently Mb = Mf−1 . The matrices Mf and Mb each have the property that their eigenvalues occur in reciprocal pairs: if λo is an eigenvalue, then so is λ−1 o . We postpone a proof of the ‘reciprocal pairs’ property of the eigenvalues to the subsequent section on the doubling algorithm, where it will follow simply by verifying that Mb and Mf are examples of symplectic matrices. Because its eigenvalues occur in reciprocal pairs, we can represent the matrix Mf in (8.5.8) via a Schur decomposition Mf = V W V −1 ,

(8.5.12)

where V is a nonsingular matrix,

W11 W = 0

W12 , W22

where W11 is a stable matrix, and W22 is an unstable matrix. In terms of ∗ −1 −1 xt , the system can be written transformed variables yt = V yt ≡ V µt ∗ yt+1 = W yt∗ .

(8.5.13)

Invariant Subspace Methods

165

V 11 V 12 , where the partitions conform in size to those of W . Let V = V 21 V 22 The solution of (8.5.13) is −1

yt∗

t W11 = 0

φt t W22

V 11 x0 + V 12 µ0 , V 21 x0 + V 22 µ0

(8.5.14)

j where φ0 = W12 , φj+1 = W11 W12 + φj W22 for j ≥ 0. Because W22 is an unstable matrix, to guarantee that limt→∞ yt∗ = 0, we require that

V 21 x0 + V 22 µ0 = 0,

(8.5.15)

which sets an initial condition that replicates itself over time in the sense that recursions on (8.5.14) imply V 21 xt + V 22 µt = 0,

(8.5.16)

for all t ≥ 0. Equation (8.5.15) implies µ0 = −(V 22 )−1 V 21 x0 . Substituting (8.5.16) into (8.5.13) and using

xt µt

= V yt∗ gives

xt+1 = V11 W11 (V 11 − V 12 (V 22 )−1 V 21 )xt

µt+1 = V21 W11 (V 11 − V 12 (V 22 )−1 V 21 )xt .

(8.5.17)

However, as noted above, µt = P xt , where P solves the algebraic Riccati equation (8.3.4). Therefore, (8.5.17) implies that P V11 = V21 or −1 P = V21 V11 = −(V 22 )−1 V 21 .

Equation (8.5.18) is our formula for P .

(8.5.18)

166

Efficient Computations

8.5.3. Distorted Economies The invariant subspace method can also be applied to compute solutions of distorted economies whose equilibrium conditions can be arranged into the form (8.5.7). Examples of such economies are described in Chapter 15, where equilibrium conditions of the form (8.5.7) cannot be interpreted as the first order conditions of any linear quadratic control problem. For these economies, the matrix Mf in general fails to have eigenvalues in reciprocal pairs. It may or may not be possible to sort the eigenvalues into equal numbers of stable and unstable ones, which are to become the eigenvalues of W11 and W22 , respectively. Whether it is possible becomes a check for the existence and uniqueness of a stable solution of the model. The condition that there exist a unique solution of (8.5.8) with |xt |2 < ∞ is that there exists a Schur decomposition (8.5.12) of Mf in which half the eigenvalues of Mf are stable, and the other half are unstable. An excess of stable eigenvalues indicates nonuniqueness; an excess of unstable eigenvalues indicates nonexistence of a stable solution. Where a unique solution exists, it can be computed using formula (8.5.18). 5

8.5.4. Transition Dynamics Invariant subspace algorithms can be adapted to solve models in which elements of the matrices determining preferences, technologies, information, and government policies vary deterministically over time, before some date T1 , after which they are constant. The procedure is to use the algorithm (8.5.18) to solve the model for t ≥ T1 , then to work backwards for earlier dates. We want to compute an equilibrium in which the L, N matrices are timevarying in a simple deterministic way, say, due to once and for all changes in tax rates at some date t = T1 > 0. Suppose that we want to solve xt xt+1 N =L , t ≥ T1 (8.5.19) µt µt+1 and

xt xt+1 ˜ ˜ N =L , µt µt+1

0 ≤ t < T1 ,

(8.5.20)

˜, L ˜ are the ‘temporary’ versions of the matrices whose ‘permanent’ where N values are L, N . 5 See Blanchard and Kahn (1981) and Whiteman (1983).

Invariant Subspace Methods

167

For t ≥ T1 , we use the solution of the ‘permanent’ system, with V 21 xt + V 22 µt = 0.

(8.5.21)

In particular, (8.5.21) implies that G1T1 xT1 + G2T1 µT1 = 0, where G1T1 = V 21 , G2T1 = V 22 . We know that xT1 V11 V12 W11 = µT1 V21 V22 0

W12 W22

V 11 xT1 −1 + V 12 µT1 −1 . V 21 xT1 −1 + V 22 µT1 −1

(8.5.22)

(8.5.23)

We want to impose restriction (8.5.22) on (8.5.23) and use it to solve for xT1 −1 as a linear function of µT1 −1 . A couple of lines of algebra leads to the restriction [G1T1 V11 + G2T1 V21 ][W11 V 11 + W12 V 21 ] + [G1T1 V12 + G2T1 V22 ]W22 V 21 xT1 −1 o n + [G1T1 V11 + G2T1 V21 ][W11 V 12 + W12 V 22 ] + [G1T1 V12 + G2T1 V22 ]W22 V 22 µT1 −1 ≡ G1,T1 −1 xT1 −1 + G2,T1 −1 µT1 −1 = 0.

(8.5.24)

This equation can be written as G1T1 −1 xT1 −1 + G2T1 −1 µT1 −1 = 0,

(8.5.25)

and it can be solved for µT1 −1 as a linear function of xT1 −1 . Equations (8.5.24) and (8.5.25) define (G1t , G2t ) as a function of (G1t+1 , G2t+1 ). We use (8.5.25) to ‘backdate’ the Git , i = 1, 2, matrices, and iterate back to t = 0. These calculations will produce time-varying versions of all of our equilibrium matrices Ao , C, Sc , Mc , . . . for t = 0, 1, . . . , T1 described in chapters 4 and 6.

168

Efficient Computations

8.6. The Doubling Algorithm The algebraic matrix Riccati equation can be solved with a doubling algorithm. 6 The algorithm shares with invariant subspace methods the prominent role it assigns to the matrix Mb of equation (8.5.9). We start with a finite horizon version of our problem for horizon t = 0, . . . , τ − 1, which leads to a two point boundary problem. We continue to assume that A is nonsingular, iterate on (8.5.8), and impose the boundary condition µτ = 0 to get xτ x0 ˆ M = , (8.6.1) 0 µ0 where ˆ = M −τ = M τ . M (8.6.2) b f We want to solve (8.6.2) for µ0 as a function of x0 , and from this solution ˆ conformably with deduce a finite-horizon approximation to P . Partitioning M ˆ 11 xτ = x0 , M ˆ 21 xτ = µ0 . Therefore, the state-co-state partition, we deduce M −1 ˆ ˆ we choose µ0 = M21 (M11 ) x0 , and set the matrix ˆ 21 (M ˆ 11 )−1 . P =M

(8.6.3)

ˆ for large horizon τ , then use (8.6.3) to The plan is efficiently to compute M compute P . We can accelerate the computations by choosing τ to be a power of two and using k k k+1 (8.6.4) = (Mf−2 )Mf−2 . Mf−2 ˆ = M −τ can be computed in j iterations Thus, for τ = 2j , the matrix M f instead of 2j iterations, inspiring the name doubling algorithm. Because Mf−1 has unstable eigenvalues, direct iterations on (8.6.4) can be unreliable. Therefore, the doubling algorithm transforms iterations on (8.6.4) into other iterations whose important objects converge. These iterations exploit the fact that the matrix Mf is symplectic (see Appendix B). The eigenvalues of symplectic matrices come in reciprocal pairs. The product of symplectic matrices is symplectic; for any symplectic matrix S , the matrices S21 (S11 )−1 and (S11 )−1 S12 are both symmetric; and ′ −1 S22 = (S11 ) + S21 (S11 )−1 S12 ′ −1 = (S11 ) + S21 (S11 )−1 S11 (S11 )−1 S12 .

6 This section is based on Anderson, Hansen, McGrattan, and Sargent (1995). For another discussion of the doubling algorithm, see Anderson and Moore [1979, pp. 158–160].

Partitioning the State Vector

169

Therefore, a (2n × 2n) symplectic matrix can be represented in terms of three (n × n) matrices α = (S11 )−1 , β = (S11 )−1 S12 , γ = S21 (S11 )−1 , the latter two matrices being symmetric. These properties of symplectic matrices inspire the following parameterizak tion of Mf−2 αk−1 αk−1 βk = , γk αk−1 αk′ + γk αk−1 βk where the n × n matrices αk , βk , γk satisfy the recursions k Mf−2

(8.6.5)

αk+1 = αk (I + βk γk )−1 αk

βk+1 = βk + αk (I + βk γk )−1 βk αk′ γk+1 = γk +

αk′ γk (I

−1

+ βk γk )

(8.6.6)

αk .

To initialize, we use representation (8.5.11) for Mb = Mf−1 to induce the settings: α0 = A, γ0 = R, β0 = BQ−1 B ′ . Anderson, Hansen, McGrattan, and Sargent (1996) describe a version of the doubling algorithm modified to build in a positive definite terminal value matrix Po . Their scheme initializes iterations on (8.6.6) as follows: α0 = (I + BQ−1 B ′ Po )−1 A β0 = (I + BQ−1 B ′ Po )−1 BQ−1 B ′ ′

−1

γ0 = R − Po + A Po (I + BQ

′

(8.6.7) −1

B Po )

A.

The modified algorithm then works as follows: 1. Initialize α0 , β0 , γ0 according to (8.6.7). 2. Iterate on (8.6.6). 3. Form P as the limit of γk + Po .

We have assumed that A is nonsingular, but Anderson (1985) argues that the doubling algorithm is applicable also in circumstances where A is singular. 7 Anderson, Hansen, McGrattan, and Sargent (1996) report the results of computations in which the doubling algorithm is among the fastest and most reliable available algorithms for solving several example economies.

7 See Anderson, Hansen, McGrattan, and Sargent (1996) for conditions under which the matrix sequences {αk }, {βk }, {γk } converge.

170

Efficient Computations

8.7. Partitioning the State Vector Undiscounted versions of the control problem solved by our social planner assume a form for which it is natural to partition the state vector to take advantage of the pattern of zeros in A and B . This leads to a control problem of the form: choose {ut }∞ t=0 to maximize −

′ ∞ X x1t R11 { x2t R21 t=0

x1t+1 x2t+1

R12 R22

x1t + u′t Qut } x2t

(8.7.1)

x1t B1 + ut , x2t 0

(8.7.2)

subject to

=

A11 0

A12 A22

with [x′10 , x′20 ]′ given. 8 For this problem, the operator associated with Bellman’s equation is T (P ) = R + A′ P A − A′ P B(Q + B ′ P B)−1 B ′ P A. (8.7.3) x1t Partitioning P and T (P ) conformably with the partition makes the x2t (1, 1) and (1, 2) components of T (P ) satisfy T11 (P11 ) = R11 + A′11 P11 A11 − A′11 P11 B1 (Q + B1′ P11 B1 )−1 B1′ P11 A11 (8.7.4) T12 (P11 ,P12 ) = R12 + A′11 P11 A12 − A′11 P11 B1 (Q + B1′ P11 B1 )−1 B1′ P11 A12

+

[A′11

−

A′11 P11 B1 (Q

+

(8.7.5)

B1′ P11 B1 )−1 B1′ ]P12 A22

Equation (8.7.4) shows that T11 depends on P11 , but not on other elements of the partition of P . From (8.7.5), T12 depends on P11 and P12 , but not on P22 . Because T maps symmetric matrices into symmetric matrices, the (2, 1) 8 System ( 8.7.1 ) – ( 8.7.2 ) is called a controllability canonical form (see Kwakernaak and Sivan [1972]). Two things distinguish a controllability canonical form: (1) the pattern of zeros in the pair (A, B, ) and (2) a requirement that (A11 , B1 ) be a controllable pair (see Appendix A of this chapter). A controllability canonical form adopts a description of the state vector that separates it into a part x2t that cannot be affected by the controls, and a part x1t that can be controlled in the sense that there exists a sequence of controls {ut } that sends x1 to any arbitrarily specified point within the space in which x1 lives.

Partitioning the State Vector

171

block of T is just the transpose of the (1, 2) block. Finally, the (2, 2) block of T depends on P11 , P12 , and P22 . Partition the optimal feedback matrix F = [F1 F2 ], where the partition is conformable with that of xt . Then the optimal control is ut = [F1 F2 ]

x1t . x2t

f f f Let P11 be the fixed point of (8.7.4) and let P12 be the fixed point of T12 (P11 , P12 ). Then F1 and F2 are given by f f F1 = (Q + B1′ P11 B1 )−1 B1′ P11 A11

(8.7.6)

f f f F2 = (Q + B1′ P11 B1 )−1 (B1′ P11 P12 + B1′ P12 A22 )

(8.7.7)

f f Equation (8.7.6) shows that F1 depends only on P11 , while F2 depends on P11 f f and P12 , but not on P22 , the fixed point of T22 . We aim to compute [F1 , F2 ] and the multipliers described in chapter 3, f f which turn out only to depend on P11 and P12 . We can compute these objects rapidly by using the structure exposed by (8.7.4) and (8.7.5). First, note that the T11 operator identified by (8.7.4) is formally equivalent with the T operator of (8.7.3), except that (1, 1) subscripts appear on A and R , and a (1) subscript appears on B . Thus, the T11 operator is simply the operator whose iterations define the matrix Riccati difference equation for the small optimal regulator f problem determined by the matrixes (A11 , B1 , Q, R11 ). We can compute P11 by using any of the algorithms described above for this smaller problem. We have chosen to use the doubling algorithm (8.6.6). f Second, given a fixed point P11 of T11 , we apply another sort of doubling f algorithm to compute the fixed point of T12 (P11 , ·). This mapping has the form f T12 (P11 , P12 ) = D + G′ P12 H

(8.7.8)

f f f f where D = R12 +A′11 P11 A12 −A′11 P11 B1 (Q+B1′ P11 B1 )−1 B1′ P11 A12 , G = [A11 − ′ f −1 ′ f B1 (Q + B1 P11 B1 ) B1 P11 A11 ], H = A22 . Notice that G = A11 − B1 F1 , where F1 is computed from (8.7.6). When x2t is set to zero for all t , the law of motion for x1t under the optimal control is thus given by

x1t+1 = Gx1t .

172

Efficient Computations

For problems for which condition (8.1.1) is either automatically satisfied or else imposed, the eigenvalues of G and H each have absolute values strictly less than unity. That the eigenvalues of G and H are both less than unity assures the existence of a limit point to iterations on (8.7.8). The limit point satisfies the Sylvester equation P12 = D + G′ P12 H, (8.7.9) which is to be solved for P12 . The limit point of iterations on T12 initiated from P12 (0) = 0 can be represented f P12

=

∞ X

G′j DH j ,

(8.7.10)

j=0

f , ·) can be verified directly. However, whose status as a fixed point of T12 (P11 iterations on (8.7.9) would not be an efficient way to compute P12 . Instead, we recommend using this doubling algorithm. Compute the following objects recursively: Gj = Gj−1 Gj−1

Hj = Hj−1 Hj−1 P12,j = P12,j−1 +

(8.7.11)

G′j−1 P12,j−1 Hj−1

where we set P12,0 = D, G0 = G, H0 = H . By repeated substitution it can be shown that j 2X −1 P12,j = Gi′ DH i . (8.7.12) i=0

Each iteration doubles the number of terms in the sum. 9 , 10

9 This algorithm is implemented in the MATLAB program double2j.m. 10 The (1, 2) partition of P is simply P f ′ . We could derive an algorithm similar to ( 8.7.11 ) 12 f

f

to compute P22 , but we don’t need to compute P22 , which is used to compute neither [F1 F2 ] nor the Lagrange multipliers that determine the price system associated with our equilibrium.

A Periodic Doubling Algorithm

173

8.8. The Periodic Optimal Linear Regulator In chapter 17, we study a class of models of seasonality whose social planning problems form a periodic optimal linear regulator problem: choose {ut }∞ t=0 to maximize ∞ X {x′s(t) Rs(t) xt + u′t Qs(t) ut } (8.8.1) t=0

subject to

xt+1 = As(t) xt + Bs(t) ut .

(8.8.2)

Here s(t) is a periodic function that maps the integers into a subset of the integers: s : (· · · − 1, 0, 1, · · ·) → [1, 2, · · · , p] s(t + p) = s(t) for all t.

In problem (8.8.1) - (8.8.2), the matrices As , Bs , Qs , and Rs that define the linear regulator problem are each periodic with common period p . Associated with problem (8.8.1) – (8.8.2) is the following version of the matrix Riccati difference equation: Pt = Rs(t) + A′s(t) Pt+1 As(t) (8.8.3) ′ ′ − A′s(t) Pt+1 Bs(t) (Qs(t) + Bs(t) Pt+1 Bs(t) )−1 Bs(t) Pt+1 As(t) .

Under conditions that generalize assumption A1, which were discussed by Richard Todd [1983], iterations on (8.8.3) yield p convergent subsequences, whose limit points we denote P1 , P2 , . . . , Pp . The optimal decision rule in period t is ut = −Fs(t) xt ,

(8.8.4)

where ′ ′ Fs(t) = −(Qs(t) + Bs(t) Ps(t+1) Bs(t) )−1 Bs(t) Ps(t+1) , As(t) .

(8.8.5)

Thus, the optimal decision rules themselves have period p . One way to compute the optimal decision rules is to iterate on (8.8.3) to convergence of the p subsequences, and then to use (8.8.5). Faster algorithms can be obtained by adapting calculations described earlier in this chapter. In the next section, we show how doubling algorithms apply to the periodic linear regulator problem, and also how the ‘controllability canonical form’ can be exploited.

174

Efficient Computations

8.9. A Periodic Doubling Algorithm First-order conditions for the periodic linear regulator can be represented as

xt+1 µt+1

= Mf,s(t)

xt , µt

(8.9.1)

where Mf,s(t) is the periodic counterpart to the matrix Mf defined in (8.5.10). Iterating this equation p times and using the periodic structure of s(t) gives xt+p xt = Γp , (8.9.2) µt+p µt where Γp ≡ Mf,p−1 Mf,p−2 · · · Mf,1 Mf,p .

(8.9.3)

The matrix Γp is the product of p symplectic matrices, and therefore is symplectic. Equation (8.9.2) at t = p can be represented x2p xp Γ−1 = , (8.9.4) p µ2p µp where −1 −1 −1 Γ−1 p = Mf,p Mf,1 · · · Mf,p−1 .

(8.9.5)

Iterating (8.9.4) τ − 1 ≥ 1 times and imposing the same boundary condition used in (8.6.1) gives ˆ xpτ = xp , (8.9.6) M µp 0 ˆ = Γ−τ . An argument used earlier implies that the doubling algorithm where M p ˆ to compute can be applied to our redefined M ˆ 21 (M ˆ 11 )−1 . Pp = M

(8.9.7)

It is straightforward to compute the remaining p−1 value functions. Notice that (8.9.4) implies ˆ xpτ = Mf,p−1 xp−1 , M 0 µp−1 or −1 ˆ xpτ = xp−1 . Mf,p−1 M 0 µp−1

A Periodic Doubling Algorithm

175

The same argument used above now implies that ˜ 21 (M ˜ 22 )−1 x1 ≡ P1 x1 , µ1 = M ˜ =M ˆ p−1 ≡ M −1 M ˆ is symplectic because it is the product of two where M f,p−1 symplectic matrices. The product of two symplectic matrices Z1 , Z2 has representation −1 α ˜ α ˜ −1 β˜ Z1 Z2 = Z¯ = γ˜ α ˜ −1 α ˜ 1 + γ˜ α ˜ −1 β˜ where

α ˜ = α2 (I + β1 γ2 )−1 α1 γ˜ = γ1 + α1′ γ2 (I + β1 γ2 )−1 α1 β˜ = β2 + α2 (I + β1 γ2 )−1 β1 α′ .

(8.9.8)

2

We can use this feature to compute Pp−1 from the γ term produced by this representation of multiplication. Iterating this argument leads us to compute Pp−2 , . . . , P1 as the correˆ p−2 = sponding γ matrices in the successive multiplications used to form M −1 −1 ˆ p−1 , . . . , M ˆ1 = M M ˆ2 . Mf,p−2 M f,1 Thus, the algorithm works as follows. 1. Initialize α0 , β0 , γ0 according to (8.6.7). 2. Use the algorithm (8.9.8) for multiplying symplectic matrices to form Γ−1 p defined as in (8.9.5). 3. Iterate on (8.6.6). 4. Form Pp as the limit of γk + Po . ˆ p−1 , M ˆ p−2 , . . . , M ˆ 1 using (8.9.8), and set the corre5. Successively form M sponding γ terms to Pp−1 , Pp−2 , . . . , P1 . Having computed P1 , . . . , Pp , we can use (8.8.5) to compute the optimal decision rules. The optimal feedback laws are periodic, so that ut = −Fs(t) xt . The matrices F1 , . . . Fp are computed from Fj = (Qj + Bj ′ Pj+1 Bj )−1 Bj ′ Pj+1 Aj , where it is understood that Pp+1 = P1 .

176

Efficient Computations

8.9.1. Partitioning the state vector We can also apply the partitioning technique to the periodic optimal linear regulator i problem h x1t exactly as in order to accelerate the computations. We partition the state vector into x2t above. With the appropriate specification of Rs , Qs , As , and Bs , we obtain a periodic version of the T11 (P11 ) mapping described in equation ( 8.7.4 ). Use our procedures to compute f P1 , P2 , . . . , Pp as described above, then set P11,j = Pj for j = 1, . . . , p . The T12 mapping for the periodic model becomes f

T12,k (P11,k+1 , P12,k+1 ) = Dk + Gk ′ P12,k+1 Hk

(8.9.9)

where f

f

Dk = R12,k + A′11,k P11,k+1 A12,k − A′11,k P11,k+1 B1k f

′ P ′ P × (Qk + B1k B )−1 B1k 11,k+1 A12,k 11,k+1 1k f

′ P ′ P Gk = [A11,k − B1k (Qk + B1k B )−1 B1k 11,k+1 A11,k 11,k+1 1k

(8.9.10)

Hk = A22,k f

In ( 8.9.9 ) – ( 8.9.10 ), P11k+1 is the fixed point for P11,k+1 corresponding to period k + 1 . Iterations on ( 8.9.9 ) will give rise to a sequence consisting of p convergent subsequences, f f whose limit points we call P12,1 , . . . , P12,p . We desire to compute these limiting matrices. f We begin by creating an operator T¯12,1 whose fixed point is P12,1 . We define

¯1 ¯1 + G ¯ ′ P12,1 H T¯12,1 (P12,1 ) = D 1

(8.9.11)

′ ′ ′ ¯ ¯′ ¯ 1 = D1 +G′ D2 H1 +· · ·+G′ G′ · · · G′ where D 1 1 2 p−1 Dp Hp−1 Hp−2 · · · H1 G1 = G1 G2 · · · Gp H1 = Hp Hp−1 · · · H1 . We can compute the fixed point of ( 8.9.11 ) by using the standard doubling algorithm that is described in section blank and that is implemented in the MATLAB program double2j.m. f f Once we have computed P12,1 , we can compute P12,j for j = p, p − 1, . . . , 2 by using f

f

f

f

P12,p = Dp + G′p P12,1 H1 P12,j = Dj + G′j P12,j+1 Hj

(8.9.12) , j = p − 1, p − 2, · · · , 2

The optimal feedback laws ut ≡ −Fs(t) , xt can be computed as follows. Let Fs(t) = [F1s(t) F2s(t) ] , where the partition of Fs(t) matches that of the state vector into x1 (t), x2 (t) . Then we have f

f

′ P −1 B ′ P F1j = (Qj + B1j 1j 11,j+1 A11,j 11,j+1 B1j ) f

f

f

′ −1 (B ′ P ′ P F2j = (Qj + B1j 1j 11,j+1 A12,j + B1j P12,j+1 A22,j ) 11,j+1 B1j )

for j = 1, . . . , p.

(8.9.13)

Linear Exponential Quadratic Gaussian Control

177

The optimal closed loop system is then xt+1 = (As(t) − Bs(t) Fs(t) )xt .

(8.9.14)

8.10. Linear Exponential Quadratic Gaussian Control In chapter 16, we shall reinterpret some of our economies in terms of risksensitive control theory. In this section, we describe how to adapt the preceding computational strategies to handle versions of the ‘risk-sensitivity corrections’ of Jacobsen (1973, 1977) and Whittle (1990). We use Hansen and Sargent’s (1995) method of implementing discounting. The resulting specification preserves the computational ease of the original linear quadratic specification, while relaxing ‘certainty equivalence.’ Let Vt1 (xt1 ) = −x′t1 Pt1 xt1 − ηt1 . Let β ∈ (0, 1) and 1 consider the sequence {Vt (xt )}tt=t of value functions generated by the following 0 constrained optimization problems: o n σ 2 Vt (xt ) = max −(x′t Rxt + u′t Qut ) + β log Et exp Vt+1 (xt+1 ) ut σ 2

(8.10.1)

subject to xt+1 = Axt + But + Cwt+1 ,

(8.10.2)

where wt+1 is an (N × 1) martingale difference sequence with Gaussian density f (wt+1 ) =

1 1 ′ exp{− wt+1 Σ−1 wt+1 }. 2 (2π)N/2 |Σ|1/2

(8.10.3)

Usually, we shall set the covariance matrix Σ = Ewt wt′ = I . We momentarily retain the more general notation in order to state a useful lemma in greater generality. In solving this discounted linear exponential quadratic Gaussian (LEQG) control problem, we use the following lemma due to Jacobson (1973). Lemma (Jacobson): Let wt+1 ∼ N (0, Σ) and xt+1 = Axt + But + Cwt+1 . Suppose that the matrix (Σ−1 − σC ′ Pt+1 C) is positive definite. Then Et exp{

σ ′ x Pt+1 xt+1 } = 2 t+1

178

Efficient Computations

Z

∞

−∞

where

1 1 ′ σ exp{− wt+1 Σ−1 wt+1 } exp{ x′t+1 Pt+1 xt+1 } 2 2 (2π)N/2 |Σ|1/2

(8.10.4)

nσ o = k exp (Axt + But )′ P˜t+1 (Axt + But ) 2 P˜t+1 = Pt+1 + σPt+1 C(Σ−1 − σC ′ Pt+1 C)−1 C ′ Pt+1

k=

s

| (Σ−1 − σC ′ Pt+1 C)−1 | . |Σ|

(8.10.5)

(8.10.6)

This concludes the statement of the lemma. Let Vt+1 (xt+1 ) = −x′t+1 Pt+1 xt+1 − ηt+1 , and apply the lemma to evaluate the term inside the braces on the right side of (8.10.1): nσ o 2 log Et exp [x′t+1 Pt+1 xt + ηt+1 ] σ 2 ′ ′ = x Rxt + u Qut + β(Axt + But )′ P˜t+1 (Axt + But )

x′t Rxt +u′t Qut + β t

(8.10.7)

t

+ constant where P˜t+1 is given by equation (8.10.5). Maximizing the right hand side of (8.10.7) with respect to ut gives the linear decision rule ut = −Ft xt , where Ft is determined by the recursions: P˜t+1 = Pt+1 + σPt+1 C(Σ−1 − σC ′ Pt+1 C)−1 C ′ Pt+1 Ft = {Q + βB ′ P˜t+1 B}−1 βB ′ P˜t+1 A

Pt = R + βA′ P˜t+1 A − β 2 A′ P˜t+1 B(Q + βB ′ P˜t+1 B)−1 B ′ P˜t+1 A.

(8.10.8) (8.10.9) (8.10.10)

Notice that in the special case that σ = 0, these equations are versions of the Riccati difference equation and the associated decision rule. Notice also that when σ 6= 0, equations (8.10.8), (8.10.9), and (8.10.10) imply that the decision rules Ft depend on the innovation variances of the exogenous processes (note the appearance of C in (8.10.8)).

Linear Exponential Quadratic Gaussian Control

179

We can obtain a more compact version of these recursions as follows. Apply the matrix identity (a − b d−1 c)−1 = a−1 + a−1 b(d − ca−1 b)−1 ca−1 to (8.10.10) using the settings a−1 = β P˜t+1 , b = −B, d = Q, c = B ′ to obtain β P˜t+1 − β P˜t+1 B(B ′ (β P˜t+1 ) B + Q)−1 B ′ (β P˜t+1 ) 1 = ( P˜t+1 + BQ−1 B ′ )−1 . β Substituting into the right side of (8.10.10) gives 1 Pt = R + A′ ( P˜t+1 + BQ−1 B ′ )−1 A . β

(8.10.11)

Now apply the same matrix identity to the right side of (8.10.8) to obtain −1 P˜t+1 = (Pt+1 − σCΣC ′ )−1 .

(8.10.12)

Substituting (8.10.12) into (8.10.11) gives the version −1 Pt = R + A′ (β −1 Pt+1 + BQ−1 B ′ − σβ −1 CΣC ′ )−1 A.

(8.10.13)

Collecting results, we have that the solution of the problem can be represented via the recursions (8.10.13), (8.116), (8.10.9). We are interested in problems for which recursions on these equations converge as t → −∞ . In situations in which convergence prevails, we can avail ourselves of a doubling algorithm to accelerate the computations.

8.10.1. Doubling algorithm It suffices to consider the undiscounted (β = 1) version of our problem, because we can transform a discounted problem into an undiscounted one. Represent the Riccati equation (8.10.13) in the form (see Appendix C) −1 Pt = R + A′ (Pt+1 + J)−1 A

where J = BQ−1 B ′ − σCΣC ′ . The doubling algorithm applies with −1 A A−1 J Mf−1 = Mb = , RA−1 A′ + RA−1 J

(8.10.14)

180

Efficient Computations

and with the settings α0 = A, γ0 = R, β0 = J . To compute the solution with terminal value matrix Po , use the initializations α0 = (I + JPo )−1 A, β0 = (I + JPo )−1 J, γ0 = −Po + R + A′ Po (I + JPo )−1 A . The algorithm then works as follows. 1. Initialize α0 , β0 , γ0 according to the formulas just given. 2. Iterate on (8.6.6). 3. Form P as the limit of γk + Po .

A. Concepts of Linear Control Theory Assume in the deterministic linear regulator (8.5.1)–(8.5.2) that matrix R is positive semi-definite and that Q is positive definite. Sufficient conditions for existence and stability of a solution of the deterministic linear regulator are typically stated in terms defined in the following four definitions. Definition: The pair (A, B) is stabilizable if y ′ B = 0 and y ′ A = λy ′ for some complex number λ and some complex vector y implies that |λ| < 1 or y = 0. Definition: The pair (A, B) is controllable if y ′ B = 0 and y ′ A = λy ′ for some complex number λ and some complex vector y implies that y = 0. Definition: The pair (A, D) is detectable if D′ y = 0 and Ay = λy for some complex number λ and some complex vector y implies that |λ| < 1 or y = 0. Definition: The pair (A, D) is observable if D′ y = 0 and Ay = λy for some complex number λ and some complex vector y implies y = 0. Stabilizability and controllability evidently form a pair of concepts, with controllability implying stabilizability, but not vice versa (i.e., controllability is a more restrictive assumption. Similarly, detectability and observability form a pair of concepts, with observability implying detectability, but not vice versa. Stabilizability is equivalent with existence of a time-invariant control law that stabilizes the state vector. Controllability implies that there exists a sequence of controls that can attain an arbitrary value for the state vector starting from any initial state vector, within n periods, where n is the dimension of the state. When (A, B) is controllable, the entire state vector is ‘endogenous,’ in the sense of being potentially ‘under control.’

Concepts of Linear Control Theory

181

The concepts of detectability and observability are applied to the pair of matrices (A, D), where DD′ = R (i.e., D is a factor of R ). Assume (a.) that the pair (A, B) is stabilizable, which implies that it is feasible to stabilize the state vector; and that (b.) the pair (A, D) is detectable, which means that it is desirable to stabilize the state vector. Together, assumptions (a.) and (b.) imply that the optimal control stabilizes the state vector. When R is nonsingular, the pair (A, D) is observable, and the value function is strictly concave.

182

Efficient Computations

B. Symplectic Matrices We now define symplectic matrices 11 and state some of their properties. Definition: A (2n × 2n) matrix Z is said to be symplectic if Z ′ JZ = J , where 0 −In . J= In 0 The following properties of symplectic matrices follow directly from the definition of a symplectic matrix: Property 1: If the matrix Z is symplectic, then so is any positive integer power of Z . Property 2: If Z1 and Z2 are both (2n × 2n) symplectic matrices, then their product Z1 Z2 is also symplectic. Property 3: If a symplectic matrix Z is written in partitioned form Z=

Z11 Z21

Z12 , Z22

−1 and if Z11 exists, then −1 ′ −1 ) + Z21 Z11 Z12 Z22 = (Z11

Property 4: The eigenvalues of any symplectic matrix Z occur in reciprocal pairs, i.e., if λi is an eigenvalue of a symplectic matrix Z , then so is λ−1 i . To establish property 4, that from the definition that any symplectic matrix Z satisfies Z −1 = J −1 Z ′ J . Since Z −1 and Z ′ are thus related by a similarity transformation, they have common eigenvalues. This implies that the eigenvalues of Z must occur in reciprocal pairs. −1 Property 3 means that if Z11 exists, then a symplectic matrix Z can be represented in the form α−1 Z= γα−1

α−1 β ′ α + γα−1 β

(8.B.1)

11 See Anderson and Moore [1979, pp. 160–161] and also Anderson, Hansen, McGrattan, and Sargent (1996).

Alternative forms of Riccati equation

183

Let Zj , for j = 1, 2, be two symplectic matrices, each represented in the form (8.B.1): −1 αj αj−1 βj Zj = . (8.B.2) γj αj−1 αj′ + γj αj−1 βj

It can be verified directly that the product Z1 Z2 = Z¯ has the same form, namely, −1 α ˜ α ˜ −1 β˜ Z1 Z2 = Z¯ = . (8.B.3) γ˜ α ˜ −1 α ˜ ′ + γ˜ α ˜ −1 β˜ where

α ˜ = α2 (I + β1 γ2 )−1 α1 γ˜ = γ1 + α1′ γ2 (I + β1 γ2 )−1 α1 β˜ = β2 + α2 (I + β1 γ2 )−1 β1 α′

(8.B.4)

2

This algorithm is implemented in our MATLAB program mult.m.

C. Alternative forms of Riccati equation It is useful to display alternative forms of the Riccati equation P = R + A′ P A − A′ P B(Q + B ′ P B)−1 B ′ P A.

(8.C.1)

We first apply the following matrix identity from Noble and Daniel [1977, p. 29]. Assume that d−1 and a−1 exist. Then (a − bd−1 c)−1 = a−1 + a−1 b[d − ca−1 b]−1 ca−1 . Apply this identity, setting a−1 = Pt+1 , b = −B, d = Q, c = B ′ to obtain −1 (Pt+1 + BQ−1 B ′ )−1 = Pt+1 − Pt+1 B(B ′ Pt+1 B + Q)−1 B ′ Pt+1 .

Substituting the above identity into (8.C.1) establishes −1 Pt = R + A′ (Pt+1 + BQ−1 B ′ )−1 A.

Now write (8.C.2) as −1 −1 Pt = R + A′ Pt+1 Pt+1 (Pt+1 + BQ−1 B ′ )−1 A −1 Pt = R + A′ Pt+1 (Pt+1 Pt+1 + BQ−1 B ′ Pt+1 )−1 A

Pt = R + A′ Pt+1 (I + BQ−1 B ′ Pt+1 )−1 A

(8.C.2)

Assume that A−1 exists, and write the preceding equation as Pt = R + A′ Pt+1 (A−1 + A−1 BQ−1 B ′ Pt+1 )−1 Pt = A′ Pt+1 (A−1 + A−1 BQ−1 B ′ Pt+1 )−1 + R(A−1 + A−1 BQ−1 B ′ Pt+1 )(A−1 + A−1 BQ−1 B ′ Pt+1 )−1 This equation can be represented as Pt = {RA−1 + [A′ + RA−1 BQ−1 B ′ ]Pt+1 } {A−1 + A−1 BQ−1 B ′ Pt+1 }−1 .

(8.C.3)

Equation (8.C.3) takes the form Pt = {C + DPt+1 } × {E + F Pt+1 }−1 where

(8.C.4)

C = RA−1 D = A′ + RA−1 BQ−1 B ′ E = A−1 F = A−1 BQ−1 B ′ ,

which can be represented as

Xt Yt

=

E C

F D

Xt+1 , Yt+1

where Pt = Yt Xt−1 . Notice that

E C

F D

= Mf−1 = Mb ,

and that limt→−∞ Pt = limt→−∞ Yt Xt−1 can be computed as the limit of the j γj term in the representation of the symplectic matrix Mf−2 .

Part II Representations and Properties

Chapter 9 Representation and Estimation This chapter shows how our models restrict moments of observed prices and quantities, and how observations can be used to make inferences about the parameters of our models. Earlier chapters have prepared a state-space representation that expresses states xt and observables yt as linear functions of an initial x0 and histories of martingale difference sequences wt . The wt ’s are shocks to endowments and preferences whose histories are observed by the agents in the economy. The econometrician does not see those shocks, at least directly. Therefore, to prepare a model for estimation we obtain another representation of it that is cast in terms of shocks that can in principle be recovered from an econometrician’s observations. By using the Kalman filter we shall obtain what is known as an ‘innovations representation’. It is a workhorse. It can be transformed to yield a Wold representation or a vector autoregression for observables. 1 An important approach to estimation, approximation, and aggregation over time is to deduce the restrictions that the models of the economy and of data collection impose on the innovations representation. The Kalman filter does this efficiently, and enables a recursive way of calculating a Gaussian likelihood function. We describe how to obtain maximum likelihood and generalized method of moments estimates of a model’s parameters, using both time domain and frequency domain methods. As by-products of time domain estimation, we deduce autoregressive and Wold representations for observables. As a by-product of frequency domain estimation, we recover a theory of specification error. We also study the effects of aggregation over time, and how to estimate a model specified at a finer time interval than pertains to the available data. These methods must be augmented to incorporate data on asset prices, which are non-linear functions of the state of the economy. The last part of the chapter describes how asset prices, returns, and other nonlinear functions of the state can be used in estimation. The Kalman filter is intimately connected to the optimal linear regulator (i.e., the linear-quadratic dynamic programming problem). Remarkably, the key 1 See Sims (1972, 1980), Whittle (1983), and Sargent (1987) for definitions and discussions of the Wold and autoregressive representations.

– 187 –

188

Representation and Estimation

mathematical formula associated with the Kalman filter is the same matrix Riccati equation that solves the linear regulator. Furthermore, the same key ‘factorization identity’ occurring with the Kalman filter plays a role in linear-quadratic optimization theory. In chapters 11 and 14, we shall use a ‘factorization identity’ to provide information about alternative representations of preferences.

9.1. The Kalman Filter We regard a vector of time t data yt as error-ridden measures of linear combinations of the state vector xt . We append a measurement equation to the equilibrium law of motion of the state to attain the following state space system: xt+1 = Ao xt + Cwt+1 (9.1.1) yt = Gxt + vt where vt is a martingale difference sequence of measurement errors that satisfies Evt vt′ = R, Ewt+1 vs′ = 0 for all t + 1 ≥ s . Here G is a matrix whose rows are composed of entries of the Sj and Mj matrices, computed for example in chapters 4, 6, and 7, that select those components of quantities and prices for which data are available. 2 We assume that x0 is a random vector with known mean x ˆ0 and covariance matrix E(x0 − x ˆ0 )(x0 − x ˆ0 )′ = Σ0 . Using (9.1.1), we have Ey0 = Gˆ x0 . t . For any variable We adopt the notation y0t = [yt , yt−1 , . . . , y0 ], y t = y−∞ t−1 ˆ t |y , x ˆ zt , t ≥ 0, we let zˆt = E[z ˆ ], where E(·) is the linear least squares 0 0 ˆ ˆ t |y t , x projection operator. Also, we occasionally use the notation Et zt = E[z 0 ˆ0 ]. We want recursive formulas for yˆt , x ˆt . We begin by constructing an innovation process {at } such that [at0 , x ˆ0 ] forms an orthogonal basis for the information ˆ0 ]. We recursively calculate the projections x ˆt+1 and yˆt by regressing set [y0t , x t ˆ0 ]. on the orthogonal basis [a0 , x ˆ0 ]is constructed using a Gram-Schmidt proThe orthogonal basis for [y0t , x cess. Begin with the regression equation y0 = Ey0 + a0 = Gˆ x0 + a0 or a0 = y0 − Gˆ x0 , 2 Later we shall permit serially correlated measurement errors. It is easy to modify the calculations to permit Ewt+1 vt′ to be nonzero.

The Kalman Filter

189

where the residual a0 satisfies the least squares normal equation Ea′0 = 0. Evidently, [y0 , x ˆ0 ] and [a0 , x ˆ0 ] span the same linear space. Next, form a1 as the residual from a regression of y1 on [y0 , x ˆ0 ], or equivalently, a regression of ˆ y1 on [a0 , x ˆ0 ]: a1 = y1 − E[y1 | y0 , x ˆ0 ] or ˆ 1 | a0 , x a1 = y1 − E[y ˆ0 ]; a1 is by construction orthogonal to a0 and x ˆ0 ; i.e., E(a1 a′0 ) = 0, E(a1 ) = 0. ˆ t | y t−1 , x ˆ t | at−1 , x Continuing in this way, form at = yt − E[y ˆ0 ] = yt − E[y ˆ0 ], 0 0 where E(at a′s ) = 0 for s = 0, . . . , t−1 and E(at ) = 0. We call at the innovation in yt . It is useful to represent at as follows. From the second equation of (9.1.1) and from the fact that vt is orthogonal to yt−s and xt−s for s ≥ 1, it follows that yˆt = Gˆ xt and that yt = Gˆ xt + G(xt − x ˆ t ) + vt . By subtracting the first equation from the second, we find that the innovation at in yt satisfies at ≡ yt − yˆt = G(xt − x ˆ t ) + vt . (9.1.2) Calculate the second moment matrix of at to be Eat a′t = GE(xt − x ˆt )(xt − x ˆt )′ G′ + Evt vt′ = GΣt G′ + R ≡ Ωt

where Σt ≡ E(xt − x ˆt )(xt − x ˆt )′ . We shall soon give a recursive formula for Σt . From the first equation in (9.1.1), it follows that ˆt xt+1 = Ao E ˆt xt = Ao E ˆt−1 xt + Ao (E ˆt xt − E ˆt−1 xt ), E

(9.1.3)

ˆt denotes projection on [y t , x ˆ where again E 0 ˆ0 ]. Express the projection Et xt = Pt ˆt xt +ψt , ψt is a least squares residual vector, and Ext + j=0 Γj aj where xt = E the regression coefficients Γj are determined by the least squares orthogonality conditions Eψt a′s = 0 for s = 0, . . . , t. Because [at0 , x ˆ0 ] is an orthogonal basis t for [y0 , x ˆ0 ], these orthogonality conditions imply (Ext a′t )(Ωt )−1 = Γt ,

(9.1.4)

190

Representation and Estimation

ˆt−1 xt = Ext + where Eat a′t = Ωt . To compute Ext a′t , first notice that E Pt−1 ˆ j=0 Γj aj . Then xt = Et−1 xt + Γt at + ψt can be interpreted in terms of the regression equation ˆt−1 xt ) = Γt at + ψt , (xt − E (9.1.5) ˆ t−E ˆt−1 xt )|at ]. Evidently, E(xt − E ˆt−1 xt )a′t = Ext a′t . Use where Γt at = E[(x ′ ′ ˆ (9.1.2) to compute E(xt − Et−1 xt )at = Σt G . It follows that (9.1.4) becomes Γt = Σt G′ (GΣt G′ + R)−1 ,

(9.1.6)

ˆt xt = E ˆt−1 xt + Γt at . E

(9.1.7)

and from (9.1.5) that Substituting (9.1.7) into (9.1.3) gives x ˆt+1 = Ao x ˆt + Ao Γt (yt − Gˆ xt ) or x ˆt+1 = Ao x ˆt + Kt at ,

(9.1.8)

where at = yt − Gˆ xt , and where from (9.1.6) Kt must satisfy Kt = Ao Σt G′ (GΣt G′ + R)−1 .

(9.1.9)

Equation (9.1.9) expresses the ‘Kalman gain’ Kt in terms of the state covariance matrix Σt = E(xt − x ˆt )(xt − x ˆt )′ . We need an equation for Σt . Subtract x ˆt+1 = Ao x ˆt + Kt (yt − Gˆ xt ) from o the first equation of (9.1.1) to obtain xt+1 − x ˆt+1 = (A − Kt G)(xt − x ˆt ) + Cwt+1 − Kt vt . Multiply each side of this equation by its own transpose and take expectations to obtain Σt+1 = (Ao − Kt G)Σt (Ao − Kt G)′ + CC ′ + Kt RKt′ .

(9.1.10)

Substituting (9.1.9) into (9.1.10) and rearranging gives a matrix Riccati difference equation for Σt : Σt+1 = Ao Σt Ao′ + CC ′ − Ao Σt G′ (GΣt G′ + R)−1 GΣt Ao′ .

(9.1.11)

The recursive (9.1.9) and (9.1.11) for Σt , Kt determine the Kalman filter. They are to be initialized from a given Σ0 . Later we discuss alternative ways to choose Σ0 .

Innovations Representation

191

9.2. Innovations Representation The Kalman filter lets us associate with representation (9.1.1) an ‘innovations representation’: x ˆt+1 = Ao x ˆt + Kt at (9.2.1) yt = Gˆ xt + at , where Eat a′t ≡ Ωt = GΣt G′ + R . This time varying representation is obtained starting from arbitrary initial conditions x ˆ0 , Σ0 . We can rearrange (9.2.1) into the form of a whitening filter at = yt − Gˆ xt

x ˆt+1 = Ao x ˆt + Kt at ,

(9.2.2)

which can be used for recursively constructing a record of innovations {at }Tt=0 from an x ˆ0 and a record of observations {yt }Tt=0 . The filter defined by (9.2.2) is called a “whitening filter” because it accepts as “input” the serially correlated process {yt } and produces as “output” the serially uncorrelated (i.e., “white”) vector stochastic process {at }. The process {at } is said to be a fundamental white noise for the {yt } process, because it equals the one-step ahead prediction error in a linear least squares projection of yt on the history of y . 3 Later, we shall use the whitening filter in several ways. We shall use it to study how the innovations {at } from a population vector autoregression for {yt } are related to the {yt } process and to the underlying martingale process {wt } of information flowing to agents. We shall also use it to construct a recursive representation of a Gaussian likelihood function for a sample drawn from the {yt } process.

3 See Sims (1972), Hansen and Sargent (1991, chapter 2), and Sargent (1987, chapter XI) for the role such an error process plays in the construction of Wold’s representation theorem.

192

Representation and Estimation

9.3. Convergence results For the purpose of obtaining a time-invariant counterpart to (9.2.1), we introduce two assumptions. Assumption A1: The pair (Ao′ , G′ ) is stabilizable. Assumption A2: The pair (Ao′ , C) is detectable. See the appendix to chapter 8 for definitions of stabilizability and detectability. Assumptions A1 and A2 are typically met for our applications. Under A1 and A2, two useful results occur. The first is that iterations on the matrix Riccati difference equation (9.1.11) converge as t → ∞ , starting from any positive semi-definite initial matrix Σ0 . The limiting matrix Σ∞ ≡ limt→∞ Σt is the unique positive semi-definite matrix Σ that satisfies the algebraic matrix Riccati equation 4 Σ = Ao ΣAo′ + CC ′ (9.3.1) − Ao ΣG′ (GΣG′ + R)−1 GΣAo′ . If we initiate the Kalman filter by choosing Σ0 = Σ∞ , then from (9.1.11) and (9.1.9), we obtain a time invariant Kt matrix, call it K . Under this circumstance, representation (9.2.1) becomes time invariant. The stationary covariance matrix of the innovations is given by Ω = Eat a′t = GΣG′ + R , where Σ = Σ∞ = Σ0 . The second useful result is that Assumptions A1 and A2 imply that Ao − KG is a stable matrix, i.e., its eigenvalues are strictly less than unity in modulus. Later we shall see how the stability of the matrix Ao − KG plays a key role in a convenient formula for the autoregressive representation for the {yt } process.

4 The limiting form of ( 9.1.10 ) is evidently a discrete Lyapunov or Sylvester equation. See chapter 8.

Serially Correlated Measurement Errors

193

9.3.1. Time-Invariant Innovations Representation The infinite-horizon time invariant Kalman filter defines a matrix valued function, which we express as h i K, Σ = kfilter (Ao , G, V1 , V2 , V3 ) (9.3.2)

where V1 = CC ′ , V2 = Evt vt′ , V3 = Ewt+1 vt′ , Σ = Et−1 (xt − x ˆt )(xt − x ˆt )′ . For our model, we can use (9.3.2) with the following settings for the matrices V1 , V2 , V3 : V1 = CC ′ , V2 = R, V3 = a matrix of zeros conformable to x and y. 5 By using the function kfilter, we can evidently associate with representation (9.1.1) a time-invariant innovations representation (9.2.1) in which Kt is constant.

9.4. Serially Correlated Measurement Errors It is useful to adapt the preceding calculations to cover the case in which the measurement errors vt in (9.1.1) are serially correlated. 6 Modify (9.1.1) to be xt+1 = Ao xt + Cwt+1 yt = Gxt + vt

(9.4.1)

vt = Dvt−1 + ηt where D is a matrix whose eigenvalues are strictly below unity in modulus and ηt is a martingale difference sequence that satisfies Eηt ηt′ = R Ewt+1 ηs′ = 0 for all t and s. In (9.4.1), vt is a serially correlated measurement error process that is orthogonal to the xt process. Define the quasi-differenced process y t ≡ yt+1 − Dyt .

(9.4.2)

5 The function kfilter defined in ( 9.3.2 ) solves a version of ( 9.1.9 ) and ( 9.1.11 ) for Σ∞ and K∞ , a version that has been generalized to permit arbitrary covariance between wt+1 and vt , which is required for several of our applications. 6 The calculations in this section imitate those of Anderson and Moore [1979].

194

Representation and Estimation

From (9.4.1) and the definition (9.4.2) it follows that y t = (GAo − DG)xt + GCwt+1 + ηt+1 Thus, (xt , y t ) is governed by the state space system xt+1 = Ao xt + Cwt+1 y t = Gxt + GCwt+1 + ηt+1

(9.4.3)

where G = GAo − DG . This state space system has nonzero covariance between the state noise Cwt+1 and the “measurement noise” (GCwt+1 + ηt+1 ). Define the covariance matrices V1 = CC ′ , V2 = GCC ′ G′ + R, V3 = CC ′ G′ . By applying the Kalman filter to (9.4.3), we obtain a gain sequence Kt with which to construct the associated innovations representation x ˆt+1 = Ao x ˆ t + Kt u t y t = Gˆ xt + ut

(9.4.4)

′ ˆ t | y t−1 , x ˆ t | y t−1 , x where x ˆt = E[x ˆ0 ], ut = y t − E[y ˆ0 ], Ω1 ≡ Eut u′t = GΣt G +V2 . 0 0 Using definition (9.4.2), it follows that [y0t+1 , x ˆ0 ] and [y t0 , x ˆ0 ] span the same t t ˆ ˆ space, so that x ˆt = E[xt | y0 , x ˆ0 ], ut = yt+1 − E[yt+1 | y0 , x ˆ0 ]. Thus, ut is the innovation in yt+1 .

9.5. Combined System It is useful to have a formula that gives a state space representation for yt driven by the innovations to yt . We obtain this by combining the innovations system (9.4.4) for y t with the system yt+1 = Dyt + y t .

(9.5.1)

The system (9.5.1) accepts {y t } as an “input” and produces {yt } as an “output”. The two systems (9.4.4) and (9.5.1) can be combined in a series to give the state space system: o x ˆt Kt x ˆt+1 A 0 + ut = G D yt I yt+1 (9.5.2) x ˆt yt = [ 0 I ] + [0]ut yt

Recursive Formulation of Likelihood Function

195

The MATLAB program evardec.m uses the time-invariant version of (9.5.2), obtained using kfilter.m, to obtain a decomposition of the j -step ahead prediction error variance associated with the Wold representation for yt . 7

9.6. Recursive Formulation of Likelihood Function The Kalman filter enables a recursive algorithm for computing a Gaussian likelihood function for a sample of observations {ys }Ts=0 on a (p × 1) vector yt . We assume that these data are governed by the innovations representation (9.2.1) The likelihood function of {ys }Ts=0 is defined as the density f (yT , yT −1 , . . . , y0 ). It is convenient to factor the likelihood function f (yT , yT −1 , . . . , y0 ) = fT (yT |yT −1 , . . . , y0 )fT −1 (yT −1 |yT −2 , . . . , y0 ) · · · f1 (y1 |y0 )f0 (y0 ).

(9.6.1)

The Gaussian likelihood function for an n × 1 random vector y with mean µ and covariance matrix V is 1 1 N (µ, V ) = (2π)−n/2 |V |− 2 exp − (y − µ)′ V −1 (y − µ) . (9.6.2) 2 Evidently, from (9.1.1), the distribution f0 (y0 ) is N (Gx0 , Ω0 ), where Ωt = GΣt G′ +R and Σt is the covariance matrix of xt around x ˆt . Further, f (yt |yt−1 , . . . , y0 ) = N (Gˆ xt , Ωt ). It is easy to verify that the distribution gt (at ) of the innovation at is N (0, Ωt ) Thus, f0 (y0 ) equals g0 (a0 ), the distribution of the initial innovation. More generally, from (9.2.1), the conditional density ft (yt |yt−1 , . . . , y0 ) equals the density gt (at ) of at . Then the likelihood (9.6.1) can be represented as gT (aT )gT −1 (aT −1 ) . . . g1 (a1 )g0 (a0 ). (9.6.3) Expression (9.6.3) implies that the logarithm of the likelihood function for y0T is T X −.5 {p ln(2π) + ln |Ωt | + a′t Ω−1 (9.6.4) t at }. t=0

7 The MATLAB program series.m can be used to obtain the time-invariant system ( 9.5.2 ) from the two systems ( 9.4.4 ) and ( 9.5.1 ).

196

Representation and Estimation

9.6.1. Initialization Two alternative sets of assumptions are commonly used to initiate the Kalman filter, corresponding to different information about y0 . (a.) The distribution of the initial y0 is treated as if it were conditioned on an infinite history of y ’s. This idea is implemented by specifying that x0 has mean x ˆ0 = E[x0 |y−1 , y−2 , . . .], and a covariance matrix Σ0 = Σ∞ coming from the steady state of the Kalman filter. In this case, the time-invariant Kalman filter can be used to construct the Gaussian log likelihood: −.5

T X t=0

{p ln(2π) + ln |Ω| + a′t Ω−1 at },

(9.6.5)

where Ω = GΣ∞ G′ + R , and where the innovations at are computed using the steady state Kalman gain K . This procedure amounts to replacing f0 (y0 ) in −1 (9.6.1) with f (y0 |y−∞ ). (b.) The initial value y0 is drawn from the stationary distribution of y , meaning that it is associated with an x0 governed by the stationary distribution of xt , an assumption implemented by initiating the Kalman filter with Σ0 = Σx , where Σx is the asymptotic stationary covariance matrix of x . Assumptions (a) and (b) pertain to how one selects the matrix Σ0 . Under each of assumptions (a) and (b), it is common to set x ˆ0 equal to the unconditional mean of x , provided that this exists.

9.6.2. Non-existence of a stationary distribution Approach (b) assumes that the law of motion xt+1 = Ao xt + Cwt+1 is such that the {xt } process has an asymptotic stationary distribution, and cannot be used without modification in models that violate this assumption. When an asymptotic stationary distribution doesn’t exist, one procedure is to assume a ‘diffuse’ initial distribution over the piece of x0 that has no stationary distribution. The models described in chapter 13, with their co-integrated equilibrium consumption processes, necessitate such a procedure. In the appendix, we describe a method for coping with this situation, inspired by ideas of Kohn and Ansley (19XXX). It is most useful for us to describe

Recursive Formulation of Likelihood Function

197

their idea in the context of models with serially correlated measurement errors, which we treat in the next section.

9.6.3. Serially correlated measurement errors When we use the state space model with serially correlated measurement errors (9.4.1), some adjustments are called for in the above procedures for forming the log likelihood. These adjustments are occasioned by the timings in the definitions of x ˆt , ut . In particular, the notation now t denotes x ˆt = E[xt |y ] and Σt = E(xt − x ˆt )(xt − x ˆt )′ . These changes mean that the distribution gt−1 (ut−1 ) equals ft (yt |yt−1 , . . . , y0 ). So corresponding to the factorization (9.6.1) we have gT −1 (uT −1 )gT −2 (uT −2 ) . . . g0 (u0 )g−1 (u−1 ).

(9.6.6)

To deduce the appropriate distribution of y0 , or equivalently, of u−1 , notice that the time 0 version of the ‘whitener’ is u−1 = y0 − Dy−1 − Gˆ x−1 x ˆ0 = Ao x ˆ−1 + K0 u−1 ,

where K0 is the time 0 value for the Kalman gain. It is natural to start the system with y−1 = GEx and x ˆ−1 = Ex , where Ex is the stationary mean of xt , 8 and to initiate the Kalman filter from the mean of the stationary distribution of x . So the Gaussian log likelihood function is −.5

T −1 X

t=−1

{p ln(2π) + ln |Ωt | + u′t Ω−1 t ut }.

(9.6.7)

We now indicate how these procedures can be adapted to handle models for which no stationary distribution for xt exists, following procedures of Kohn and Ansley (BLANK). The idea is to factor the likelihood function as f (yT , yT −1 , . . . , y0 ) = fT (yT |yT −1 , . . . , y0 )fT −1 (yT −1 |yT −2 , . . . , y0 ) · · · fm (ym |ym−1 , . . . , y0 )f (ym−1 , . . . , y0 ).

8 Notice that G and not G appears in the equation for the unconditional mean.

(9.6.8)

198

Representation and Estimation

Kohn and Ansley assign a ‘diffuse prior’ to that subset of the state vector that does not possess a stationary distribution, and let the remaining piece of x0 be distributed according to its stationary distribution. This specification embodies an ‘improper prior’ distribution for (ym−1 , . . . , y0 ). Under this specification, we use the first m observations of yt to estimate xm−1 , then form x ˆm−1 , Σm−1 from which to initiate the Kalman filter for the system (9.4.3) with serially correlated measurement errors. The Kalman filter is applied to compute the likelihood for the sample {ys }Ts=m . In addition, we can adjust (9.6.8) to account for the first m observations. Details are given in the appendix.

9.7. Wold Representation For the purpose of describing the relationship of the time-invariant innovations representation to the Wold and autoregressive representations, we shall avail ourselves when needed of: Assumption A3: The eigenvalues of Ao are all less than unity in modulus, except possibly for one associated with a constant. A Wold representation for a stationary stochastic process yt is a moving average of the form ∞ X yt = Ey + Γj ǫt−j , j=0

ˆ t |y t−1 ], and P∞ traceΓj Γ′ < +∞ . (Below, we shall where ǫt = yt − E[y j j=0 for the most part set the unconditional mean vector Ey to zero, to conserve on notation.) We can attain a Wold representation by manipulating the innovations system in a way that amounts to driving the date for the initial x ˆ0 arbitrarily far into the past. Thus, the first equation of (9.4.4) can be solved recursively for t X x ˆt+1 = (Ao )j Kut−j + (Ao )t+1 x ˆ0 . j=0

Now assume that x ˆ0 was itself formed by having observed the history y −1 , so that x ˆ0 = (I − Ao L)−1 Ku−1 + µx ,

Vector Autoregression for {yt }

199

where µx is the unconditional mean of x . Under this specification for x ˆ0 , x ˆt+1 = (I − Ao L)−1 Kut + µx .

(9.7.1)

Below, we shall omit the unconditional mean term, by assuming that µx = 0. To get a Wold representation for yt , substitute (9.4.2) into (9.4.4) to obtain x ˆt+1 = Ao x ˆt + Kut yt+1 − Dyt = Gˆ xt + ut .

(9.7.2)

Then (9.7.2) and (9.7.1) can be used to get a Wold representation for yt : yt+1 = [I − DL]−1 [I + G(I − Ao L)−1 KL]ut ,

(9.7.3)

where again L is the lag operator. Also, from (9.7.2) a “whitening filter” for obtaining {ut } from {yt } is given by ut = yt+1 − Dyt − Gˆ xt

x ˆt+1 = Ao x ˆt + Kut .

(9.7.4)

9.8. Vector Autoregression for {yt } We can use the innovations representation and some results from linear algebra to derive a convenient formula for the one-step-ahead linear least squares forecast of yt based on the history of the {yt } process. We begin by deriving a version of the factorization identity, which asserts equality between two representations of the spectral density matrix of the observables. We will encounter a mathematically equivalent form of this identity in Chapter 11 when we discuss observationally equivalent representations of preferences.

200

Representation and Estimation

9.8.1. The factorization identity For the model with serially uncorrelated measurement errors, we have two alternative representations for an observed process {yt }, the original state space representation (9.1.1) and the innovations representation (9.2.1). Because they describe the same stochastic process {yt }, they give two alternative representations of the spectral density matrix of {yt }, an outcome that expresses the factorization identity. The original state space representation is xt+1 = Ao xt + Cwt+1 yt = Gxt + vt ,

(9.8.1)

where wt+1 is a martingale difference sequence of innovations to agents’ information sets, and vt is another martingale difference sequence of measurement errors. We assumed that wt+1 , vt are mutually orthogonal stochastic processes. The first line of representation (9.8.1) can be written L−1 xt = (I−Ao L)−1 Cwt+1 or xt = (L−1 − Ao )−1 Cwt+1 . It follows that the covariance generating function of {xt } satisfies Sx (z) = (zI − Ao )−1 CC ′ (z −1 I − (Ao )′ )−1 . Using this expression and the second line of (9.1.1), together with the observation that vt is orthogonal to the process xt , shows that the covariance generating function of yt is given by Sy (z) = G(zI − Ao )−1 CC ′ (z −1 I − (Ao )′ )−1 G′ + R.

(9.8.2)

Representation (9.2.1) implies x ˆt = (L−1 − Ao )−1 Kat , and yt = [G(L−1 − Ao )−1 K + I]at .

(9.8.3)

Because at is a white noise with covariance matrix GΣG′ + R , it follows that the covariance generating function of {yt } equals Sy (z) = [G(zI − Ao )−1 K + I][GΣG′ + R][K ′ (z −1 I − Ao′ )−1 G′ + I].

(9.8.4)

Expressions (9.8.2) and (9.8.4) are alternative representations for the covariance generating function Sy (z). Equating them leads to the factorization identity: G(zI − Ao )−1 CC ′ (z −1 I − Ao′ )−1 G′ + R =

[G(zI − Ao )−1 K + I][GΣG′ + R][K ′ (z −1 I − Ao′ )−1 G′ + I].

(9.8.5)

Vector Autoregression for {yt }

201

The importance of the factorization identity hinges on the fact that, under assumptions A1 and A2, the zeros of the polynomial det[G(zI − Ao )−1 K + I] all lie inside the unit circle, which means that in the representation (9.8.3) for yt , the polynomial in L on the right hand side has a one-sided inverse in nonnegative powers of L, so that at lies in the space spanned by y t . We establish this result in the next section, then apply it in subsequent ones.

9.8.2. Location of zeros of characteristic polynomial We utilize two theorems from the algebra of partitioned matrices. Let a, b, c, d be appropriately conformable and invertible matrices. Then (a − bd−1 c)−1 = a−1 + a−1 b(d − ca−1 b)−1 ca−1

(9.8.6)

det(a) det(d + ca−1 b) = det(d) det(a + bd−1 c).

(9.8.7)

and Apply equality ( 9.8.6 ) to [I + G(zI − Ao )−1 K]−1 with the settings a = I, b = −G, d = (zI − Ao ), c = K , to get [I + G(zI − Ao )−1 K]−1 = I − G[zI − (Ao − KG)]−1 K.

(9.8.8)

Apply equality ( 9.8.7 ) with the settings a = I, b = G, d = (zI − Ao ), c = K to get det(zI − (Ao − KG)) = det(zI − Ao ) det(I + G(zI − Ao )−1 K), or det[I + G(zI − Ao )−1 K] =

det(zI − (Ao − KG)) . det(zI − Ao )

(9.8.9)

It follows from ( 9.8.9 ) that the zeros of det[I + G(zI − Ao )−1 K] are the eigenvalues of

Ao − KG , and the poles of det[I + G(zI − Ao )−1 K] are the eigenvalues of Ao . Assumptions A1 and A2 guarantee that the eigenvalues of Ao − KG are less than unity in modulus. We

have already made assumptions that assure that the eigenvalues of Ao are less than unity in modulus. These conditions on the eigenvalues together with equations ( 9.8.8 ) and ( 9.8.9 ) permit us to obtain the Wold and autoregressive representations of {yt } in convenient forms.

202

Representation and Estimation

9.8.3. Wold and autoregressive representations (white measurement errors) From (9.8.8), we have that [G(I − Ao L)−1 KL + I]−1 = I − G[I − (Ao − KG)L]−1 KL.

(9.8.10)

For the model with serially uncorrelated measurement errors, the Wold representation for {yt } is yt = [G(I − Ao L)−1 KL + I]at .

(9.8.11)

Applying the inverse of the operator on the right of (9.8.11) and using (9.8.10) gives yt = G[I − (Ao − KG)L]−1 Kyt−1 + at , (9.8.12) which decomposes yt into an innovation at and a one-step ahead linear least squares predictor E[yt |y t−1 ] = G[I − (Ao − KG)L]−1 Kyt−1 .

(9.8.13)

Equation (9.8.12) is equivalent with yt =

∞ X j=1

G(Ao − KG)j−1 Kyt−j + at .

(9.8.14)

Equation (9.8.14) is a vector autoregressive representation for yt . Thus, the Kalman filter allows us to move from the original state space representation to a vector autoregression.

Vector Autoregression for {yt }

203

9.8.4. Serially correlated measurement errors With few modifications, the preceding analysis can be adapted to calculate the vector autoregressive representation and the one-step ahead prediction for yt for the case in which the measurement errors are vector first-order autoregressive processes. We have seen that the Wold representation in this case takes the form yt+1 = [I − DL]−1 [I + G(I − Ao L)−1 KL]ut . (9.8.15) Operating on both sides of (9.8.15) with the inverse of the operator in L on the right side, and using (9.8.10), we obtain [I − DL] {I − G[I − (Ao − KG)L]−1 KL}yt+1 = ut , or yt+1 = {D + (I − DL) G[I − (Ao − KG)L]−1 K} yt + ut ,

(9.8.16)

ˆ t+1 |y t ]. The above equation can be expressed where recall that ut = yt+1 − E[y in the alternative forms yt+1 = Dyt + G

∞ X

(Ao − KG)j−1 Kyt−j+1

j=1 ∞ X

− DG

j=1

(Ao − KG)j−1 Kyt−j + ut ,

or yt+1 =[D + GK]yt +

∞ X [G(Ao − KG)j K j=1

(9.8.17)

− DG(Ao − KG)j−1 K]yt−j + ut .

These equations express yt+1 as the sum of the one-step ahead linear least squares forecast and the one-step prediction error. 9

9 The MATLAB program varrep.m uses ( 9.8.17 ) to obtain a vector autoregressive representation for an equilibrium set of yt ’s, given [Ao , C, G, D, R] .

204

Representation and Estimation

9.9. Innovations in yt+1 as Functions of Innovations wt+1 and ηt+1 By coupling the original state space system with the associated innovations representation, it is possible to express the innovations in the {yt } process as functions of the disturbances {wt } and the measurement errors {vt }. Having a method for expressing this connection can be useful when we want to interpret the innovations in {yt } as functions of the shocks impinging on agents’ information sets and the measurement errors. The state space system is xt+1 = Ao xt + Cwt+1 y t = Gxt + GCwt+1 + ηt+1 ,

(9.9.1)

which corresponds to an innovations representation, which can be expressed as the “whitener” xt + Ky t x ˆt+1 = (Ao − KG)ˆ (9.9.2) ut = y t − Gˆ xt .

Substituting the second equation of (9.4.3) into the first equation of (9.9.2) gives x ˆt+1 = (Ao − KG)ˆ xt + KGxt + KGCwt+1 + Kηt+1 . (9.9.3)

Using (9.9.3), systems (9.4.3) and (9.9.2) can be combined to give the consolidated system o xt Cwt+1 xt+1 A 0 + (9.9.4) = x ˆt KGCwt+1 + Kηt+1 x ˆt+1 KG Ao − KG xt + [GCwt+1 + ηt+1 ] ut = [G − G] x ˆt

In system (9.9.4), the “inputs” are the innovations to agents’ information sets, namely, wt+1 , and the innovations to the measurement errors, namely, ηt+1 . ˆ t+1 | The “output” of the system is the innovation to yt+1 , namely ut = yt+1 −Ey t y . By computing the impulse response function of system (9.9.4), we can study how the innovations ut depend on current and past values of wt+1 and ηt+1 . Versions of formula (9.9.4) are useful for studying the range of issues considered by Hansen and Sargent [1991, “Two Difficulties”]. 10 In the next section, we illustrate one such issue in the context of a permanent income example. 10 The MATLAB programs white1.m and white2.m use formula ( 9.9.4 ) to compute impulse response functions of ut with respect to wt and ηt , respectively.

Innovations in the yt ’s and the wt ’s in a Permanent Income Model 205

9.10. Innovations in the yt ’s and the wt ’s in a Permanent Income Model This section illustrates some of the preceding ideas in the context of an economic model that implies that the econometrician’s information set spans a smaller space than agents’ information. The context is a class of models which impose a form of expected present value budget balance. As we shall see, expected present value budget balance is characterized by a condition that implies that the moving average representation associated with the model, which records the response of the system to the wt ’s, fails to be invertible. The outcome is that the innovations in the autoregressive representation don’t coincide with the wt ’s. Representation (9.9.4) can be used to compute a distributed lag expressing the innovations as functions of the lagged wt ’s. We consider the following version of Hall’s model in which the endowment process is the sum of two orthogonal autoregressive processes. Preferences, technology, and information are specified as follows:

9.10.1. Preferences ∞

1 X t − E β [(ct − bt )2 + ℓ2t ] | J0 2 t=0

9.10.2. Technology ct + it = γkt−1 + dt φ1 it = gt kt = δk kt−1 + it gt · gt = ℓ2t

206

Representation and Estimation

9.10.3. Information 1 0 0 = 0 0 0

A22

0 .9 0 0 0 0

Ud =

0 0 0 1 0 0

0 0 0 0 1 0

5 1 0 0

Ub = [ 30

0

0 0 0 1 0 0 , C2 = 0 0 0 0 0 0 1 .8 .6 .4 0 0 0 0 0 0 0 0 0 1

0 0

0 0 4 0 0 0

0 0]

We specify that γ = .05, δk = 1, β = 1/1.05, φ1 = .00001. Note that β(δk + γ) = 1, which is the condition for consumption to be a random walk in Hall’s model. The preference shock is constant at 30, while the endowment process is the sum of a constant (5) plus two orthogonal processes. In particular, we have specified that dt = 5 + d1t + d2t , where d1t = .9d1t−1 + w1t d2t = w ˜2t + .8w ˜2t−1 + .6w ˜2t−2 + .4w ˜2t−3 where (w1t , w ˜2t ) = (w1t , 4w2t ). Notice that we have set ′ w1t w1t 1 0 . E = 0 16 w ˜2t w ˜2t Here d1t is a first order autoregressive process, while d2t is a third order pure moving average process. We define the household’s net of interest deficit as ct − dt . Hall’s model imP∞ poses “expected present value budget balance,” in the sense that E j=0 β j (ct+j − dt+j ) | Jt = β −1 kt−1 for all t , 11 which implies that the present value of the moving average coefficients in the response of the deficit to innovations in agents’ information sets must be zero. That is, let the moving average representation of (ct , ct − dt ) in terms of the wt ’s be 12 ct σ1 (L) = wt , (9.10.1) ct − dt σ2 (L) 11 See Sargent [1987] and Hansen, Roberds, and Sargent [1990]. 12 Without loss of generality, the covariance matrix of w can be chosen to be the identity t matrix.

Innovations in the yt ’s and the wt ’s in a Permanent Income Model 207

where σ1 (L) and σ2 (L) are each (1 × 2) matrix polynomials, and σ(L) = P∞ j j=0 σj L . Then Hall’s model imposes the restriction σ2 (β) = [ 0

0].

(9.10.2)

The agents in this version of Hall’s model observe Jt at t , which includes the history of each component of wt up to t . This means that agents see histories of both components of the endowment process d1t and d2t . Let us now put ourselves in the shoes of an econometrician who has data on the history of the pair [ ct , dt ], but not directly on the history of wt . We imagine the econometrician to form a record of consumption and the deficit [ ct , ct − dt ], and to obtain a Wold representation for the process [ ct , ct − dt ]. Let this representation be denoted 13 ∗ σ1 (L) ct = ut , (9.10.3) σ2∗ (L) ct − dt where σ ∗ (L) is one-sided in nonnegative powers of L, and ut is a serially uncorrelated process with mean zero and Eut u′t = I ; ut is the innovation of [ ct , ct − dt ] relative to the history [ct−1 , ct−1 −dt−1 ]. In representation (9.10.3), ut is the object that would appear in the Gaussian log likelihood function, as in formula (9.6.4). It is natural to ask whether the impulse response functions σ ∗ (L) in the Wold representation (or vector autoregression) (9.10.3) estimated by the econometrician “resemble” the impulse response functions σ(L) that depict the response of ( ct , ct − dt ) to the innovations to agents’ information. A way to attack this question is to ask whether the history of the {ut } process of innovations to the econometrician’s information set in (9.10.3) reveals the history of the {wt } process impinging on agents’ information sets. In the present model, the answer to this question is ‘no’ precisely because restriction (9.10.2) holds. In particular, (9.10.2) implies that the history of ut ’s in (9.10.3) spans a smaller linear space than does the history of wt ’s. Here is the reason. The ut ’s in (9.10.3) are constructed to lie in the space spanned by the history of the [ ct , ct − dt ] process. 14 Technically, this implies 13 Without loss of generality, the covariance matrix of u can be chosen to be the identity t matrix. 14 Recall the construction underlying Wold’s representation theorem, e.g., see Sargent [1987, chapter XI].

208

Representation and Estimation

that the operator σ ∗ (L) in (9.10.3) is invertible, so that (9.10.3) can be expressed as ct ut = σ ∗ (L)−1 , ct − dt where σ ∗ (L)−1 is one-sided in nonnegative powers of L, and where the coefficients in its power series expansion are square summable. Given that σ ∗ (z)σ ∗ (z −1 )′ is of full rank, a necessary condition for σ ∗ (L)−1 to exist (i.e., to have a representation as a square-summable polynomial in nonnegative powers of L) is that det σ ∗ (z) have no zeros inside the unit circle. Condition (9.10.2) then rules out the possibility that σ ∗ (L) is related to σ(L) by a relation of the form σ ∗ (L) = U σ(L) where U is a nonsingular 2 × 2 matrix. For (9.10.2) implies that det σ(z) has a zero at β , which is inside the unit circle. In circumstances in which [ ct , ct − dt ] is a full rank process, 15 the history of [ ct , ct − dt ] generates a smaller information set than does the history of the wt process. When ut spans a smaller space than wt , ut will typically be a distributed lag of wt that is not concentrated at zero lag: ut =

∞ X

αj wt−j .

(9.10.4)

j=0

Thus the econometrician’s news ut potentially responds with a lag to the agents’ news wt . The calculations leading to representation (9.9.4) can be used to compute the vector distributed lag αj . To illustrate these ideas in the context of the present version of Hall’s model, figures 9.10.1.a and 9.10.1.b display the impulse response functions of [ ct , ct − dt ] to the two innovations in the endowment process. Consumption displays the characteristic “random walk” response with respect to each innovation. Each endowment innovation leads to a temporary surplus followed by a permanent net-of-interest deficit. The temporary surplus is used to accumulate a stock of capital sufficient to support the permanent net of interest deficit that is to follow it. Restriction (9.10.2) states that the temporary surplus just offsets the permanent deficit in terms of expected present value. For each innovation, we computed the present value of the response of (ct − dt ) to be zero, as predicted by (9.10.2). 15 By a full rank process we mean that σ ∗ (z)σ ∗ (z −1 ) .

Innovations in the yt ’s and the wt ’s in a Permanent Income Model 209

0.4

1 0.5

0.2 0 0

-0.5 -1

-0.2 -1.5 -0.4

-2 -2.5

-0.6 -3 -0.8 0

5

10

15

20

25

30

35

40

45

50

Fig. 9.10.1.a. Impulse response of consumption and deficit to first endowment innovation. The dotted line denotes the deficit, the dark line consumption.

-3.5 0

5

10

15

20

25

30

35

40

45

50

Fig. 9.10.1.b. Impulse response of consumption and deficit to second endowment innovation. The dotted line denotes the deficit, the dark line consumption.

Figures 9.10.2.a and 9.10.2.b report the impulse responses from the Wold representation, which we obtained using the programs varma.m and varma2.m. The innovation covariance matrix for the ut ’s was .3662 −1.9874 ∗ ∗′ Eσ0 σ0 = . −1.9874 12.8509

Notice that consumption responds only to the first innovation in the Wold representation, and that it responds with an impulse response symptomatic of a random walk. That consumption responds only to the first innovation in the vector autoregression is indicative of the Granger-causality imposed on the [ ct , ct − dt ] process by Hall’s model: consumption Granger causes ct − dt , with no reverse causality. Unlike consumption, the response of the deficit (ct − dt ) to the innovations in the vector autoregression depicted in figures 9.10.2.a and 9.10.2.b fail to match up qualitatively with the patterns displayed in figures 9.10.1.a and 9.10.1.b. In particular, the present values (σ2∗ (β)) of the response of ct − dt to ut are (6.0963, 6.6544). By construction, σ2∗ (β) cannot be zero because σ2∗ (L) is invertible.

210

Representation and Estimation

0.7

4

0.6

3.5

0.5

3

0.4

2.5

0.3 2 0.2 1.5 0.1 1

0

0.5

-0.1 -0.2

0

-0.3 0

-0.5 0

5

10

15

20

25

30

35

40

45

50

Fig. 9.10.2.a. Impulse response of consumption and deficit to first innovation in Wold representation. The dotted line denotes the deficit, the dark line consumption.

5

10

15

20

25

30

35

40

45

50

Fig. 9.10.2.b. Impulse response of consumption and deficit to second innovation in Wold representation. The dotted line denotes the deficit, the dark line consumption.

Figures 9.10.3.a and 9.10.3.b display the impulse responses of ut to wt , the kind of representation depicted in equation (9.10.4). While the responses of the innovations to consumption are concentrated at lag zero for both components of wt , the responses of the innovations to (ct − dt ) are spread over time (especially the response to w1t ). Thus, the innovations to (ct − dt ) as revealed by the vector autoregression depend on what to economic agents is “old news”. Hansen, Roberds, and Sargent [1991] describe how such issues impinge on strategies for econometrically testing present value budget balance. Hansen and Sargent [1991] and Marcet [1991] more generally study the link between innovations in a vector autoregression and the innovations in agents’ information sets.

Frequency Domain Estimation

0.4

211

1 0.5

0.2 0 0

-0.5 -1

-0.2 -1.5 -0.4

-2 -2.5

-0.6 -3 -0.8 0

2

4

6

8

10

12

14

16

18

-3.5 0

20

Fig. 9.10.3.a. Impulse response of innovations in Wold representation for consumption and deficit to first endowment innovation. The dotted line denotes the deficit, the dark line consumption.

2

4

6

8

10

12

14

16

18

20

Fig. 9.10.3.b. Impulse response of innovations in Wold representation for consumption and deficit to second endowment innovation. The dotted line denotes the deficit, the dark line consumption.

9.11. Frequency Domain Estimation We now describe how to estimate the free parameters of the model (9.4.1) using the frequency domain approximation to the likelihood function of Hannan [1970].We assume a model for which yt is asymptotically stationary. Let the mean vector for the observable {yt } process be denoted µ. The mean vector µ is a function of the parameters of the model. The spectral density matrix of the {yt } process is defined as Sy (ω) =

∞ X

Cy (τ )e−iωτ

(9.11.1)

τ =−∞

where Cy (τ ) = E[yt − µ][yt−τ − µ]′ . For the model (9.4.1), the spectral density can be represented as Sy (ω) = G(I − Ao e−iω )−1 CC ′ (I − Ao′ e+iω )−1 G′ + (I − De−iω )−1 R(I − D′ e+iω )−1

(9.11.2a)

212

Representation and Estimation

and the unconditional means can be represented via a function Eyt ≡ µ = µ(Ao , G).

(9.11.2b)

The autocovariances can be recovered from Sy (ω) via the inversion formula 16 Cy (τ ) =

1 2π

Z

π

Sy (ω)e+iωτ dω.

(9.11.3)

−π

Let yt be a (p × 1) vector. Suppose that a sample of observations on {yt }Tt=1 is available. Define the Fourier transform of {yt }Tt=1 as y(ωj ) =

T X

yt e−iωj t , ωj =

t=1

2πj , j = 1, . . . , T. T

(9.11.4)

The periodogram of {yt }Tt=1 is defined as Jy (ωj ) =

1 y(ωj )y(ωj )′ , T

(9.11.5)

where the overbar denotes complex conjugation. Following Hannan [1970], the Gaussian log likelihood of {yt }Tt=1 as a function of the free parameters determining Ao , C, D, and R can be approximated as T /2+1 X 1 (T + T p) log 2π − log{det Sy (ωj )} L∗ = − 2 j=1 T /2+1

− −

X j=1

trace Sy (ωj )−1 Jy (ωj )

(9.11.6)

T T n X X ′ o T yt − µ yt − µ T −1 trace Sy (0)−1 T −1 2 t=1 t=1

In (9.11.6), p is the dimension of the yt vector. The free parameters determining Ao , C, D , and R can be estimated by maximizing the right side of (9.11.6) with respect to them. Notice that the data {yt }Tt=1 enter the right side of (9.11.6) only through the sample mean 16 The MATLAB programs spectral.m and spectr1.m can be used to compute a spectral density matrix for one of our models. These programs implement formula ( 9.11.2 ).

Approximation Theory

213

PT T −1 t=1 yt and the periodogram Jy (ωj ), while the theory enters through relation (9.11.2) which determines µ and Sy (ωj ) as functions of the free parameters. Parameter estimation uses any of a variety of hill-climbing algorithms on (9.11.6). 17 An advantage of frequency domain estimation is that it avoids the need, associated with time domain estimation, to deduce a Wold representation for yt . Notice that estimation proceeds without factoring the spectral density matrix (9.11.2).

9.12. Approximation Theory When an economist estimates a misspecified model, how are the probability limits of the parameters that he estimates related to the parameters of a “true” model? This question is not well posed until one states an alternative model relative to which the model at hand is regarded as misspecified. If such an alternative model is on the table, then the question can be answered by adapting the analysis of approximation used by Christopher Sims [1972] and Halbert White [1982]. A modification of (9.11.6) underlies the theory of approximation. To state a complete theory of approximation, these elements are required: (1) a model that in truth generates the data (to speak of approximation, it is necessary to specify what is being approximated); (2) the model being estimated; and (3) the method of parameter estimation. We make the following assumption about these three elements. The true model is a member of the class of models described earlier in this book, with parameters denoted by a vector δ . The true mean vector for the observables is ν(δ), and the true spectral density matrix is S y (ω, δ), where S y (ω, δ) is determined by a version of (9.11.2a) with parameters o A , C, G, R, D , which depend on the parameter vector δ . The estimated model is another version of (9.11.2), where the parameters determining the matrices Ao , C, G, R, D , are denoted α , the spectral density matrix is Sy (ω, α), and the mean vector is µ(α). The method of estimation is maximum likelihood. It can be shown (see Hansen and Sargent (1993)) that the probability limits of the free

17 For example, see Bard (1974 XXXX).

214

Representation and Estimation

parameters α satisfy n 1 Z π plim α ˆ = arg max − log det Sy (ω, α) dω α 2π −π Z π 1 − trace Sy (ω, α)−1 S y (ω, δ) dω 2π −π o − [ν − µ(α)] Sy (0, α)−1 [ν − µ(α)]′ .

(9.12.1)

The right side of (9.12.1) is obtained from (9.11.6) by appropriately taking limits as T → ∞ . Roughly speaking, taking limits replaces the periodogram Jy (ωj ) with the spectral density for the true model S y (ωj ), and replaces the sample mean with the true mean vector.

9.13. Aggregation Over Time In this section, we describe how to use the Kalman filter to calculate the likelihood for data that are “aggregated over time.” We formulate a model that generates observations in state space form and then use the Kalman filter to derive an associated innovations representation from which the Gaussian log likelihood function can be constructed. Let the original equilibrium model have the state space form xt+1 = Ao xt + Cwt+1 yt = Gxt

(9.13.1)

′ where wt+1 is a martingale difference sequence with Ewt+1 wt+1 |Jt = I . We assume that the model is formulated to apply at a finer time interval than that for which data are available. For example, the model (9.13.1) may apply to weekly or monthly data, while only quarterly or annual data may be available to the economist. Furthermore, some of the observed data may be averages over time of the {yt } data generated by (9.13.1), as when “flow” data are generated by averaging over time. (Data on output, consumption, and investment flows are usually generated in this way.) Others of the data may simply be point-intime “skip sampled” versions of the data. That is, “quarterly” data are formed by sampling every thirteenth observation of the “weekly” data. We want to catalogue the restrictions imposed on these time aggregated data by the model

Aggregation Over Time

215

(9.13.1). We accomplish this by deducing the likelihood function of these data as a function of the free parameters for (9.13.1). We perform our analysis of aggregation over time in two steps. First, we expand the state space by including enough lagged states to accommodate whatever averaging over time of data is occurring. Let m be the number of dates over which data are potentially to be averaged. Then we form the augmented system x t+1 xt xt−1 .. .

xt−m+2

Ao I 0 = .. .

0

0 0 I .. .

··· ··· ··· .. .

0 0 0 .. .

0

···

I

0 0 0 .. . 0

xt xt−1 xt−2 .. .

xt−m+1

C 0 0 + wt+1 .. .

0

or xt+1 = Axt + Cwt+1

(9.13.2)

where

xt+1

x t+1 xt = xt−1 .. .

Ao I , A= 0 .. .

xt−m+2

0

0 0 I .. .

··· ··· ··· .. .

0 0 0 .. .

0

···

I

0 C 0 0 0, C = 0 . .. .. . . 0 0

(9.13.3)

Once we have formed A and C , it is easy to form the appropriate model for averaged data. For example, suppose that we are interested in forming the model governing three period averages of consumption. We would set m equal to 3 in (9.13.2) and (9.13.3), and could then model averaged consumption via the observer equation yt = Gxt where G = [Sc Sc Sc ]. The MATLAB program avg.m obtains the matrices A and C of (9.13.3) for a given m , thereby accomplishing the first step in our analysis of aggregation over time. The second step is actually to perform the aggregation over time by skipping observations on a representation of the form (9.13.1) or (9.13.2). Let an equilibrium be represented in the state space form xt+1 = Axt + Cwt+1 , yt = Gxt

t = 0, 1, 2, . . .

(9.13.4)

216

Representation and Estimation

where the first line of (9.13.4) could correspond either to (9.13.2) or to its special case, the first line of (9.13.1). Suppose that data on yt are available only every r > 1 periods, where r is an integer. Then the data are generated by the model r xt+r = Ar xt + wt+r , t = 0, r, 2r, . . . (9.13.5) yt = Gxt where Ar = Ar r wt+r = Ar−1 Cwt+1 + Ar−2 Cwt+2 + · · · + ACwt+r−1 + Cwt+r

(9.13.6)

Represent (9.13.5),(9.13.6) as the state space system r xs+1 = Ar xs + ws+1

, s = 0, 1, 2, . . .

ys = Gxs

(9.13.7)

r where ws+1 is a martingale difference sequence with contemporaneous covariance matrix

Ewsr wsr′ = CC ′ + ACC ′ A′ + · · · + Ar−1 CC ′ Ar−1′ ≡ V.

(9.13.8)

Now suppose that only error-corrupted observations on the time aggregated {ys } data are available, and that the measurement errors are first-order serially correlated. To capture this assumption, augment (9.13.7) – (9.13.8) to become the state space system r xs+1 = Ar xs + ws+1 ys = Gxs + υs

(9.13.9)

υs = Dυs−1 + ηs where Eηs ηs′ = R and Ews+1 ηs′ = 0 for all t and s . System (9.13.9) is a version of the state space system (9.4.1). Proceeding as in our analysis of (9.4.1), define y s ≡ ys+1 − Dys and Gr = (GAr − DG). Then (9.13.9) implies the system r xs+1 = Ar xs + ws+1 r y s = Gr xs + Gws+1 + ηs+1 . ′

(9.13.10)

r r Define the covariance matrices Ewsr wsr = V ≡ V1 , E(Gws+1 + ηs+1 )(Gws+1 + ′ ′ r r ′ ′ ηs+1 ) = GV G + R ≡ V2 , Ews+1 (Gws+1 + ηs+1 ) = V G = V3 . Use the function

Aggregation Over Time

217

kfilter to obtain [K, Σ] = kfilter(Ar , Gr , V1 , V2 , V3 ). Then an innovations representation for system (9.13.10) is x ˆs+1 = Ar x ˆs + Kas y s = Gr x ˆs + as

(9.13.11)

], as = y s − E[y s | y s−1 ], Ω1 ≡ Eas a′s = Gr ΣG′r + where x ˆs = E[xs | y s−1 0 0 V2 . Once again, the innovations representation (9.13.11) can be used to form the residuals at recursively, and thereby to form the Gaussian log likelihood function. 18 We illustrate the programs avg.m and aggreg.m by showing how they can be used to analyze the effects of aggregation over time in the context of our equilibrium version of Hall’s model. We want to deduce the univariate Wold representation for consumption data that are constructed by taking a three period moving average, and then “skip sampling” every third period. The following MATLAB code performs these calculations: reads in parameters of Hall’s economy computes the equilibrium forms state for three period averaging forms observer for three-period moving average of consumption R = .0001; D = 0; sets parameters of measurement error process [Ar,Cr,aa,bb,cc,dd,V1] = aggreg (AA,CC,G,D,R,3) y = dimpulse(aa,bb,cc,dd,1,20); forms moving average representation

clex11; solvea; [AA,CC]= avg(a0,C,3); G = [sc sc sc];

We have set the parameters of Hall’s model at the values that make unaveraged consumption follow a random walk. Notice that we set R and D so that only a very small measurement error is present in consumption. The impulse response function for skip-sampled three period moving average consumption reveals the following representation for the skip-sampled moving average data ct : ct − ct−1 = at + .2208at−1 where at = ct − E(ct | ct−1 , ct−2 , . . .). Thus, the first difference of ct is a firstorder moving average process. These calculations recover a version of Holbrook 18 The MATLAB program aggreg.m constructs the innovations representation ( 9.13.11 ) from inputs in the form of the state space representation ( 9.13.4 ) and the parameters R and D of the measurement error model ( 9.13.2 ).

218

Representation and Estimation

Working’s [1960] findings about the effects of skip sampling a moving average of a random walk.

9.14. Simulation Estimators We have described how to estimate the free parameters of a model using data that are possibly error-ridden linear functions of the state vector xt . In our models, quantities and (scaled Arrow-Debreu) prices are linear functions of the state, but asset prices and rates of return are non-linear functions of the state. In this section, we describe how observations of non-linear functions of the state can be used in estimation. The equilibrium transition law for the state vector xt is given by xt+1 = Ao (θ)xt + C(θ)wt+1 ,

Ewt wt′ = I

(9.14.1)

where the r × 1 vector θ contains the free parameters of preferences, technologies, and information. We partition the data into two parts, (z1t , t = 0, . . . T ) and (z2t , t = 0, . . . T ), where the z1t ’s are linear functions of the state xt , and the z2t ’s are nonlinear functions of the state. Assume that z1t is k × 1 and z2t is m × 1. The data are related to the state xt and measurement errors vt as follows: z1t = G(θ)xt + v1,t z2t = f (xt , v2,t , θ), where E

wt+1 vt

wt+1 vt

′

=

Q(θ) W (θ)′

W (θ) R(θ)

,

and where Q(θ) = C(θ)C(θ)′ . The Gaussian log likelihood function of {z1t }Tt=0 is L(θ) =

T X t=0

T

ℓt = −

i 1 Xh p log(2π) + log |Ωt | + a′t Ω−1 t at 2 t=0

where zt is p × 1 and at = z1,t − E[z1,t |z1,t−1 , . . . , z1,0 ] is the innovation vector from the ‘innovations representation’ and Ωt = Eat a′t .

Simulation Estimators

219

Maximizing the log likelihood function with respect to θ is equivalent with a particular Generalized Method of Moments (GMM) procedure using observations on (z1t , t = 0, . . . T ). Note that the first-order order conditions for maximizing the log likelihood function are ∂L = 0. ∂θ To see how this matches up with GMM, compute the score vector st = which has elements,

∂ℓt ∂θ

∂a ′ ∂ℓt ∂Ωt 1 t −1 ′ = − tr Ω−1 I − Ω a a − Ω−1 t t t t t at . ∂θi 2 ∂θi ∂θi Using the notation of Hansen (1982), the GMM estimator of θ minimizes JT (θ) = gT (θ)′ WT gT (θ) where

(9.14.2)

T

gT (θ) =

1 X st (θ) T + 1 t=0

and WT is any positive definite r×r weighting matrix. Notice that gT (θ) = ∂L ∂θ , so that for any positive definite weighting matrix, (9.14.2) is minimized by the minimizer of L(θ). The irrelevance of the weighting matrix WT reflects the property that from the viewpoint of GMM, this is a ‘just-identified’ system, with as many moment conditions as free parameters. Suppose that we want to use the observations in z1t and in z2t to estimate θ . We can apply a method described by Ingram and Lee. Given the law of motion in (9.14.1) and a realization from a pseudo-random number generator for {wj+1 , v1j , v2j }N j=0 , we can generate a pseudo-random realization of the series N {z1j , z2j }j=0 . Let q(·) be a given function of the data. Use the data and the simulation of the model, respectively, to compute the two moment vectors: T

1 X HT (z) = q(z1t , z2t ) T + 1 t=0 N

HN (θ) =

1 X q(z1j , z2j ; θ). N + 1 j=0

220

Representation and Estimation

Define hT (θ) as follows hT (θ) =

T n i 1 Xh 1 X q(z1t , z2t ) − q(z1j , z2j ; θ) T + 1 t=0 n + 1 j=0

= HT (z) − HN (θ),

where n + 1 = (N + 1)/(T + 1) and N + 1 is some integer multiple of T + 1. Then the estimation strategy for obtaining θ is to minimize JT (θ) =

∂L/∂θ hT (θ)

′

WT

∂L/∂θ hT (θ)

for some weighting matrix WT . To estimate WT , we can use the two-stage procedure in Hansen (1982), which is to start with WT = I and then construct the weighting matrix associated with the resulting estimate of θ .

A. Initialization of the Kalman Filter This appendix describes how Kohn and Ansley’s idea for estimating the initial state can be applied in the context of our class of models. Aside from numerical issues, Kohn and Ansley’s procedure is equivalent to using all of the data {ys }Ts=0 , and initializing the Kalman filter from a partitioned covariance matrix designed to approximate +∞I +∞1 , Σ0 = +∞Q′ Σ0,22 where Σ0,22 is the asymptotic covariance matrix of that piece of the state vector that has an asymptotically stationary distribution, and 1 is a matrix of ones. The +∞I pertains to elements of the state that have no asymptotic stationary distribution. In practice, +∞ is approximated by a large positive scalar. This procedure was used by Harvey and Pierse. 19 This procedure ought to be close to Kohn and Ansley’s, though the literature contains examples of cases in which the numerical properties of the ‘+∞ ≈ a big number ’ approach are poor. For that reason, it is good to have in hand procedures like the one we shall describe. 19 Another approach has been to use an ‘inverse filter’ in which the recursions are cast in terms of the inverse of Σt .

Initialization of the Kalman Filter

221

For convenience, we temporarily work with the state-space system 20 ∗ xt+1 = Axt + Cwt+1

(9.A.1)

∗ yt = Gxt + Qwt+1 ,

∗ where wt+1 is a martingale difference sequence with identity for its conditional covariance matrix. In the interest of eventually imputing a diffuse prior to the initial values of that part of the state vector that has no stationary distribution, we represent the initial state as

x0 = φη + ψ + N ν, where ψ is an n × 1 vector with all zeros except possibly for one value of one, which locates the constant in the state; and ν is normally distributed with mean zero and covariance I , and η is normally distributed with mean zero and covariance kI , where the random vectors ν and η are assumed to be independent. We use φη to represent the piece of the initial state that has no stationary distribution, and N ν to represent the piece with a stationary distribution. We attain a diffuse prior on the stationary distribution by driving k to +∞ . Our plan is to project xm on ym−1 , . . . , y0 , while driving k → +∞ , and then to initialize the Kalman filter from the resulting estimators of the distribution of xm . By iterating on the state equation (9.A.1), we can write: xm = Am φη + Am ψ + Hm wm where wm ′ = ( ν ′

w1 ∗′

...

∗′ wm ) and

Hm = ( Am−1 N ′

(9.A.2)

Now create a vector Y m−1 = ( y0′

y1′

Am−2 C ...

...

C ).

′ ) that obeys: ym−1

Y m−1 = Mm η + α + Gm wm

(9.A.3)

20 It is easy to map ( 9.4.3 ), which describes the state-space system wih serially correlated wt+1 ∗ measurement errors, into this form. Define wt+1 = and represent ( 9.4.3 ) as ηt+1 ∗ xt+1 = Axt + (C O) wt+1

∗ . yt = Gxt + (GC I) wt+1

222

Representation and Estimation

where

Mm =

and

Gm

GN GAN GA2 N .. .

= GAm−2 N GA

m−1

N

Gφ GAφ .. . GAm−1 φ

,

α=

G GA .. . GAm−1

ψ

Q GC GAC .. .

0 Q GC .. .

0 0 Q .. .

··· ··· ··· .. .

0 0 0 .. .

GAm−3 C GAm−2 C

GAm−4 C GAm−3 C

GAm−5 C GAm−4 C

··· ···

Q GC

0 0 0 .. .

0 Q

Transform equation (9.A.2) as follows. Regress Hm wm onto Gm wm , and denote the residual as Rm wm , to obtain the representation ∗ Hm w m = H m Gm wm + Rm wm ,

(9.A.4)

∗ = (EHm wm wm−1′ G′m )(EGm wm wm−1′ G′m )−1 is a matrix of least where Hm ∗ ∗ = Hm G′m (Gm Gm ′ )−1 . Gm . Thus Hm squares regression coefficients and Rm = Hm −Hm Also, since Gm wm = Y m−1 − Mm η − α , (9.A.4) implies the representation ∗ (Y m−1 − Mm η − α) + Rm wm . Hm w m = H m

Rewrite state equation (9.A.2) as: ∗ ∗ ∗ xm = (Am φ − Hm Mm )η + Am ψ − Hm α + Hm Y m−1 + Rm wm .

(9.A.5)

Next we compute some conditional expectations and covariances. Initially, we use (9.A.5) and the facts that (i) by assumption, wm is orthogonal to η , and (ii) by construction, Rm wm is orthogonal to Gm wm , to compute: ∗ ∗ ∗ E(xm |Y m−1 , η) = (Am φ − Hm Mm )η + Am ψ − Hm α + Hm Y m−1 ,

and ′ cov(xm |Y m−1 , η) = Rm Rm .

Initialization of the Kalman Filter

223

To compute the conditional expectation and covariance matrix conditioning only Y m−1 , we first compute the projection of η on Y m−1 − α : η = β ∗ (Y m−1 − α) + ε, where ε is a least squares residual. We compute Eη(Y m−1 − α) and the second moment matrix of Y m−1 − α and use them in the projection formula: ′ ′ β ∗ = (kMm )(kMm Mm + Gm G′m )−1 . ′ ′ ′ (Gm G′m )−1 Mm ]−1 [Mm (Gm G′m )−1 Mm ] to get β ∗ = [Mm (Gm Premultiply by [Mm ′ −1 −1 ′ ′ −1 ′ ′ ′ −1 Gm ) Mm ] Mm (Gm Gm ) [kMm Mm (kMm Mm +Gm Gm ) ]. If we drive k → +∞ , the last term in square brackets approaches the identity matrix, so that we have ′ ′ E(η|Y m−1 − α) = [Mm (Gm Gm ′ )−1 Mm ]−1 Mm (Gm Gm ′ )−1 (Y m−1 − α). (9.A.6)

Notice that ε = β ∗ (Mm η + Gm wm ) − η = (β ∗ Mn − I)η + β ∗ Gm wm , and that (β ∗ Mm − I) = 0. It follows that ′ cov(η|Y m−1 − α) = [Mm (Gm Gm ′ )−1 Mm ]−1 .

(9.A.7)

Using these results and applying the Law of Iterated Expectations to (9.A.5) gives: 21 ∗ ′ E(xm |Y m−1 ) = (Am−1 φ − Hm Mm )[Mm (Gm G′m )−1 Mm ]−1

′ ∗ ∗ Mm (Gm G′m )−1 (Y m−1 − α) + Am−1 ψ − Hm α + Hm Y m−1 ,

(9.A.8)

and ′ ∗ ′ cov(xm |Y m−1 ) = Rm Rm + (Am−1 φ − Hm Mm )[Mm (Gm G′m )−1 Mm ]−1 ∗ (Am−1 φ − Hm Mm )′ .

(9.A.9)

The Kalman filter is to be initialized by using these values of x ˆm , Σm , then T applied to compute (9.6.8), using observations {ys }s=m . 21 Note that equations ( 9.A.6 ) and ( 9.A.7 ) result from applying generalized least squares to the system of equations ( 9.A.3 ), where η is regarded as a matrix of constants and Mm is a matrix of regressors.

224

Representation and Estimation

When we apply this procedure with (9.A.1) corresponding to the system m−1 (9.4.3), we should interpret Y m−1 in the preceding development as Y which corresponds to Y m in the real data. In this case, we should interpret x ˆm , Σm according to definitions of the (ˆ·) variables defined for the system with serially correlated measurement errors. 22 We can also include a contribution to the likelihood function to account for the initial observations used to form x ˆm . Begin with (9.A.3) and let Ω = ′ Gm Gm , which we take to be nonsingular. Suppose that Mm is dimensioned r by s where r > s so that η is ‘overidentified.’ Construct two matrices labeled M ⊥ and M ∗ , dimensioned (r − s) × r and s × r , respectively, to satisfy: M ⊥ Ω−1 Mm = 0 M ∗ Ω−1 M ⊥′ = 0, and construct the nonsingular matrix: D=

M ∗ Ω−1 M ⊥ Ω−1

.

Define: z1 = M ∗ Ω−1 Y m z2 = M ⊥ Ω−1 Y m . Notice that conditioned on η , z1 and z2 are uncorrelated. Moreover, by construction z2 does not depend on η . We deduce the initial likelihood contribution as follows. First note that z1 = DY m . z2 Transforming the z ’s introduces a Jacobian term: log det D, which is the first contribution to the likelihood. 22 Notice that with serially correlated measurement errors, ( 9.A.8 ), ( 9.A.9 ) give the appropriate initial conditions for the Kalman filter, because of the dating conventions that make ut the innovation to yt+1 .

Initialization of the Kalman Filter

225

The second and third contributions are the likelihoods of the z ’s. Conditioned on η , the likelihood can be factored. Only the first term in the factorization depends on η , and is present in the ‘exactly identified’ case. The quadratic form term converges to zero for this contribution. We deduce the log det term by taking the limit as k goes to infinity of: ′ log det(kM ∗ Ω−1 Mm Mm Ω−1 M ∗′ + M ∗ Gm G′m M ∗′ ) =

1 ∗ M Gm G′m M ∗′ ) k where n1 is the dimension of z1 . Taking the limit and neglecting the term n1 log k , which is the same for all settings of the parameter values and so can be ignored, leaves the term: ′ n1 log k + log det(M ∗ Ω−1 Mm Mm Ω−1 M ∗′ +

′ log det(M ∗ Ω−1 Mm Mm Ω−1 M ∗′ ).

The z2 contribution to the likelihood retains both a log det and a quadratic form contribution. Notice that the z2 term is absent in the ‘exactly identified’ case.

Chapter 10 Semiparametric Estimation with Limited Information

10.1. Introduction This chapter describes semiparametric estimation of transmission mechanisms under limited information.

10.2. Underlying Economic Model Consider the following economic model. The information available to economic agents at time t is denoted Jt . There is an endogenous state vector kt−1 which we will think of as a vector of capital stocks. The capital stocks evolve according to the evolution equation: kt = ∆kt−1 + Θit (10.2.1) where it is a vector of flow variables which we refer to as investment goods. The absolute values of eigenvalues of the matrix ∆ are presumed to be strictly less than one. There is also an exogenous state vector zt with dynamics given by: zt+1 = A22 zt + C2 wt+1

(10.2.2)

where {wt } is a martingale difference sequence adapted to {Jt } with a conditional covariance matrix I . It is known that a first-order linear specification of the dynamics is quite flexible because it can represent multivariate ARMA models of arbitrary orders. The composite state vector at time t is denoted kt−1 . xt ≡ zt The recursive solution to the model gives investment it as a function of the state vector xt : it = Sxt = Sk kt−1 + Sz zt .

– 227 –

228

Semiparametric Estimation with Limited Information

Substituting this solution into equations (10.2.1) and (10.2.2), we find that xt+1 = Axt + Cwt+1 where A≡

∆ − ΘSk 0

ΘSz A22

,C ≡

(10.2.3)

0 C2

.

Thus, equation (10.2.3) gives the evolution of the state vector process when the optimal or equilibrium investment rule is imposed. The state vector zt may enter into the decision rule or equilibrium investment relation for one of two reasons. Some components of zt may enter directly into the objective functions of economic agents; and other components may simply be used in forecasting future values of these variables. Let zt1 denote a vector of the former components, which we presume are related to zt via: zt1 = H1 zt where H1 is just a matrix that selects elements from zt . The solution for investment can often be represented as: it = Sk kt−1 + S0 zt1 + Sf

∞ X j=0

1 (Λ)j S1 E(zt+j |Jt )

(10.2.4)

and hence Sz = S0 H1 + Sf (I − Λ)−1 S1 H1 (see Hansen and Sargent 1981 and Sargent 1987). One way to obtain a solution of this form is to stack the first order conditions for the endogenous state vector and its corresponding co-state vector into an expectational first-order difference equation driven by the forcing process {zt1 }, then to solve that difference equation. A similar structure can be obtained even when the endogenous state vector is not the solution to an optimal resource allocation problem. The estimation method we describe below exploits the feedforward structure of the solution for investment whereby investment depends on current and expected future values of the forcing process {zt1 }.

Econometrician’s information

229

10.3. Econometrician’s information and the implied orthogonality conditions

We presume that the econometrician observes a time series of investment it , t = 1, 2, ..., T . With knowledge of the depreciation matrix ∆, he can construct an approximate capital series, ‘approximate’ because the initial capital stock k0 may be unknown. Because the dominant eigenvalue of ∆ is strictly less than unity in modulus, the approximation error vanishes as the sample size T gets large. The econometrician also observes some but not all of the vector zt . Partition yt zt1 = ut where yt is observed by the econometrician by ut is not. To have any hope of identifying the parameters in the underlying model, we assume that the exogenous state vector process can be uncoupled in the following way: y y zt+1 = Ay zty + Cy wt+1

and u u zt+1 = Au ztu + Cu wt+1

where zt = and

zty ztu

, wt =

wty wtu

yt = Hy zty and ut = Hu ztu . To guarantee asymptotic stationarity, we restrict the absolute values of the eigenvalues of Au to be strictly less than one. The {ut } process gives us one interpretation of why investment is not an exact function of variables observed by an econometrician. Therefore, we rewrite the investment relation as: it = Sk kt−1 + S0,y yt + Sf

∞ X j=0

where et ≡ S0,u ut + Sf

∞ X j=0

(Λ)j S1,y E(yt+j |Jt ) + et

(Λ)j S1,u E(ut+j |Jt )

230

Semiparametric Estimation with Limited Information

and Sj = ( Sj,y

Sj,u ) , j = 0, 1.

The process {et } provides an error term that can be interpreted along the lines of Hansen and Sargent (1980). By construction it is uncorrelated with the process {yt } at all leads and lags. This uncorrelatedness can be directly exploited in estimating the parameters of endogenous dynamics of the model, e.g., the parameters governing the transmission mechanism. In fact, this can be accomplished in ways that permit a robust specification of the dynamics associated with the underlying shock process {ut }. In other words, a semiparametric estimation method is possible in this setting. Another source of omitted information is in the forecasting of future values of the process {yt }. For instance, let Kt denote the information set used by the econometrician, constructed in a way so that at least it is no larger than Jt . This gives rise to an additional model “specification” error, say ft , as emphasized by Shiller (1978) and Hansen and Sargent (1980, 1982). Thus, the investment equation used by an econometrician is given by: it = Sk kt−1 + S0,y yt + Sf

∞ X j=0

where ft ≡ Sf

∞ X j=0

(Λ)j S1,y E(yt+j |Kt ) + et + ft

(10.3.1)

(Λ)j S1,y [E(yt+j |Jt ) − E(yt+j |Kt )].

By the Law of Iterated Expectations, the error term ft is uncorrelated with current and past values of {yt }, but can be correlated with future values of this process. As a consequence, the orthogonality conditions that are robust to misspecifying the information set are: E[(ft + et )yt−j ′ ] = 0 for j = 0, 1, ....

(10.3.2)

The presence of the component ft in the disturbance term is what limits the orthogonality conditions to be one-sided. Future values of yt may be correlated with the disturbance term in the investment equation. Since the information set Kt is a misspecified version of Jt , unless one is willing to specify the omitted information precisely, it is most convenient to envision an econometrician modeling the evolution equation for {yt } in a flexible

An Adjustment Cost Example

231

manner. Although information is omitted, so long as {yt } has a state-space representation, we know that it can be represented as a multivariate version of an ARMA model, although the autoregressive and moving-average orders will be unknown to the econometrician. For the sake of simplicity, we assumethat the {yt } process is stationary and has a moving-average representation: yt = B(L)vt

(10.3.3)

where the operator B has a one-sided inverse, and current and past values of vt also generate the information set Kt . In what follows, it is not necessary to limit B to be a ratio of polynomials, as in ARMA models. More general dynamics can be accommodated. As we will see, this in effect introduces an infinite dimensional nuisance parameter into the moment conditions (10.3.2). Finally, note that by omitting information relative to that used by economic agents, we cannot expect to deduce impulse response functions that are interpretable in terms of the economic shocks impinging on the decision maker. In other words, the response of investment or capital stock to an innovation in {yt } (i.e., in economic agents’ information set) will be contaminated. Nevertheless, we will still be in a position to identify parameters of the endogenous dynamics.

10.4. An Adjustment Cost Example A linear-quadratic model of adjustment costs has a solution for investment that is of the form given by (3) with a scalar investment and capital stock and a scalar Λ that we will denote by λ . For simplicity, we presume that the obervable forcing process {yt } is also scalar. Write the econometric relation for investment in feedforward form as: it = ρkt−1 + ψo yt + ψf

∞ X j=0

kt = δkt−1 + it .

(λ)j E(yt+j |Kt ) + et + ft

(10.4.1)

The operator B enters into the model solution because of its role in the solution to the prediction problem: ∞ X ytp = E( λj yt+j |Kt ). j=0

232

Semiparametric Estimation with Limited Information

It is known from Hansen and Sargent (1980) that ytp = B ∗ (L)vt where B ∗ (ζ) =

ζB(ζ) − λB(λ) . ζ −λ

(10.4.2)

Substituting this formula into (10.4.1) and solving for the econometric disturbance term, we obtain: ft + et = it − ρkt−1 − ψo yt − ψf B ∗ (L)vt .

(10.4.3)

Prior to investigating the estimation of endogenous dynamics as captured by the parameters ρ, ψo , ψf , δ, λ we will study the impact of estimating B in both parametric and nonparametric settings.

10.5. A Slightly Simpler Estimation Problem Let Yt denote a random vector, each entry of which is a linear combination of current and past values of yt . Suppose the unconditional moment condition used in estimation is: E[(ft + et )Yt ] = 0. A component of these moment conditions that depends on B is: βo ≡ E[Yt B ∗ (L)vt ],

(10.5.1)

and for the moment let us suppose that βo is the parameter of interest.

A Slightly Simpler Estimation Problem

233

10.5.1. Scalar Parameterizations of B As a preliminary step to studying nonparametric estimators of B , we initially consider very simple scalar parameterizations of B : Bα = B + αF. We suppose that for sufficiently small values of α we can invert the operator Bα . We must explore what happens to the moment condition for small perturbations in α . Define: −1 φt (α) ≡ Bα (L) yt . Differentiating φt with respect to α and evaluating we find that B(L)Dφt (0) + F (L)vt = 0, or Dφt (0) = −[B(L)]−1 F (L)vt .

(10.5.2)

Then differentiating the moment relation: β(α) = E[zty Bα∗ (L)φt (α)] we find that dβ(0)/dα = E{Yt [F ∗ (L)vt ]} + E{Yt [B ∗ (L)Dφt (0)]}.

(10.5.3)

The * notation is used to denote the transformation of an operator given by (10.4.2). Let αT denote the maximum likelihood estimator for α = 0 for sample size T , and let T X YT Bα∗ T (L)φt,T (αT ) βT ≡ (1/T ) t=1

denote the sample estimator of βo where the notation φt,T (αT ) denotes the time t approximation for vt using the estimator αT . Then the sampling error in βT as an estimator of βo can be decomposed into two components: √

T √ X √ T (βT − βo ) ≈ (1/ T ) [Yt B ∗ (L)vt − βo ] + [dβ(0)/dα] T αT . t=1

(10.5.4)

234

Semiparametric Estimation with Limited Information

The first term is the usual central limit approximation for sample moment estimators while the second term accounts for the additional sampling error induced by having to estimate B . It is the second term that we turn our attention to. The limiting distribution of {αT } is determined by the score of the conditional likelihood of yt conditioned on the past. This score is given by st = −vt Dφt (0) + E[v(t)Dφt (0)].

(10.5.5)

The first term of the score comes from differentiating the quadratic form in the one-step ahead forecast error of yt , and the second term from differentiating the log variance term of the time t contribution to the conditional log likelihood. The score variable has mean zero conditioned on Kt−1 and the resulting score process is a martingale difference sequence. Then PT √ st . (10.5.6) T αT ≈ √ t=1 T E(st 2 ) In light of the fundamental role played by the score variable in deteriming the limiting distribution for the estimator sequence {αT }, it will prove to be very useful to represent the derivative dβ(0)/dα as an expected cross product of some random vector with the score st . We now deduce what that random vector is by obtaining an alternative expression for the right-hand side of (10.5.3). Note that LF (L) ∗ E{Yt [F (L)vt ]} = E Yt vt L−λ (10.5.7) LB(L) Dφt (0) = −E Yt L−λ where the first equality follows because future values of vt are orthogonal to Yt and the second equality follows from formula (10.5.2) for Dφt (0). Substituting (10.5.7) into (10.5.3) and using formula (10.4.2) for B ∗ results in: λB(λ) dβ(0)/dα = −E Yt Dφt (0) L−λ 1 (10.5.8) = −λB(λ)E Yt Dφt (0) L−1 − λ 1 Yt−1 Dφt (0) = −λB(λ)E 1 − λL

where the second equality follows from the joint stationarity of the composite process {[Yt , Dφt (0)]}. Formula (10.5.8) is almost what we want, except that

Multidimensional Parameterizations of B

235

we need an expression in terms of st instead of Dφt (0). This can be obtained by noting that 1 1 E Yt−1 Dφt (0) = −E vt Yt−1 st (10.5.9) 1 − λL 1 − λL

which can be verified as follows. Compute the expectation on the right-hand side by conditioning first on Kt−1 and using the two facts that (i) Dφt (0) is the sum of a term in vt and E[Dφt (0)|Kt−1 ], and (ii) the third moment of vt is zero. Then apply the Law of Iterated Expecations again to obtain the lefthand side of (10.5.9). Combining (10.5.8) and (10.5.9), we obtain the desired formula:

1 Yt−1 st . (10.5.10) 1 − λL Armed with this formula, we can think of the time t contribution of the “correction term” for estimating as the outcome from running a least squares i B h 1 regression of λB(λ)vt 1−λL Yt−1 onto the score st . This interpretation can be seen by substituting (10.5.7) and (10.5.10) into (10.5.4) and interpreting dβ(0)/dα as a population regression coefficient. Although we performed this E(st 2 ) computation for an affine scalar parameterization of B , it can be mimicked for any sufficiently smooth one dimensional parameterization. The correction term will continue be interpretable as a regression score. dβ(0)/dα = λB(λ)E vt

10.6. Multidimensional Parameterizations of B As a further step in studying the impact on βo of using a nonparametric estimator of B, we now briefly consider what happens when we use parameterizations that have more than one dimension but are still finite dimensional. This turns out to be an easy extension of our previous analysis. Let st be the score vector associated with any such nondegenerate parameterization. (By “nondegenerate” we simply mean that the second moment of the score vector is nonsingular, a local identification condition.) The entries of the score vector st can be represented as in (14), and the derivative matrix ∂β(0)/∂α′ is given by the expected cross product: 1 Yt−1 st ′ . ∂β(0)/∂α′ = λB(λ)E vt 1 − λL

236

Semiparametric Estimation with Limited Information

Therefore, √

T √ X [Yt B ∗ (L)vt − βo ] T (βT − βo ) ≈(1/ T ) t=1

+ λB(λ)E vt

T √ X 1 st . Yt−1 st ′ [E(st st ′ )]−1 (1/ T ) 1 − λL t=1

Again the correction term for the h first istage estimation of B has a regression 1 interpretation: regress λB(λ)vt 1−λL Yt−1 onto the score vector st .

10.7. Nonparametric Estimation of B Since the derivative matrix has an expected cross product representation for any finite dimensional parameterization, we can use an insight from Stein (1956) and Levit (1975), developed more fully by Van der Vaart (1991) and Newey (1993), to deduce the asymptotic distribution when B is estimated nonparametrically. ask what happens to the population regression of iWe simply h 1 Yt−1 onto the linear space of time t scores for all possible λB(λ)vt 1−λL parameterizations of B . Since the elements of the regressand can be viewed as scores of hypothetical parameterizations, the resulting limiting distribution for βo is T X √ ∗ T (βT − βo ) ≈ Yt B (L)vt − βo + vt t=1

1 Yt−1 . 1 − λL

(10.7.1)

This additive decomposition gives a time series counterpart to the “correction terms” for semiparametric M-estimators derived by Newey (1993), (e.g., see formula (3.10) in Newey). 1

1 One diffference between Newey’s derivation and ours is that Newey works with score vectors for the entire process of observables. Given the additive structure of our model we can work with the simpler scores of maximum likelihood estimators of B using only data on {yt } .

Back to the Adjustment Cost Model

237

10.8. Back to the Adjustment Cost Model Let us now revert to the estimation problem of interest posed using the adjustment cost model. We let βo denote the parameter vector governing the endogenous dynamics and view ρ, ψo , ψf , δ and ρ as functions of the unknown parameter vector βo . When the capital stock is not directly observable, the generated stock sequence will also depend on βo through its dependence on the depreciation factor δ . Suppose that we estimate βo using a GMM estimator that exploits the unconditional moment restriction: E[Yt (ft + et )] = 0. Then the usual GMM inference works with an additional correction term in which the derivatives of the moment conditions are computed by differentiating with respect to β and evaluating these derivatives at the true value of βo and B . Let this derivative be denoted do , and let ao denote the limiting matrix that selects the moment conditions to be used in estimation. Then T √ X √ T (βT −βo ) ≈ −(ao do )−1 ao (1/ T Yt (ft + et ) + λB(λ) t=1

1 Yt−1 vt . 1 − λL

Chapter 11 Representation of Demand

11.1. Introduction This chapter derives demand schedules from our preference specification ht = ∆h ht−1 + Θh ct st = Λht−1 + Πct

(11.1.1)

with preference shock bt = Ub zt . An equivalence class of preferences (∆h , Θh , Π, Λ, Ub ) give rise to identical demand schedules. Among such preferences, particular ones that we call canonical are easiest to work with. We apply the concept of canonical representation of preferences to a version of Becker and Murphy’s model of rational addiction. The chapter also uses demand curves to to construct partial equilibrium interpretations of our models. This chapter sets the stage for the studies of aggregation in chapters BLANK and BLANK.

11.2. Canonical Representations of Services We begin with a definition. Definition: A representation of a household service technology (∆h , Θh , Π, Λ, Ub ) is said to be canonical if it satisfies the following two requirements: i. Π is nonsingular. ii. The absolute values of the eigenvalues of (∆h − Θh Π−1 Λ) are strictly √ less than 1/ β . A canonical household service technology maps any given service process {st } in L20 into a corresponding consumption process {ct } for which the implied household capital stock process {ht } is also in L20 . To verify this, we use the canonical – 239 –

240

Representation of Demand

representation to obtain a recursive representation for the consumption process in terms of the service process: ct = −Π−1 Λht−1 + Π−1 st

ht = (∆h − Θh Π−1 Λ)ht−1 + Θh Π−1 st

(11.2.1)

The restriction on the eigenvalues of the matrix (∆h − Θh Π−1 Λ) keeps the household capital stock {ht } in L20 .

11.3. Dynamic Demand Functions for Consumption Goods We postpone constructing a canonical representation, and proceed immediately to use one to construct a dynamic demand schedule. In Chapter 6 we derived the following first-order conditions for the household’s optimization problem: st = bt − µst

(11.3.1)

0 Π′ µst = −Θ′h µht + µw 0 pt

(11.3.2)

µht = βEt (Λ′ µst+1 + ∆′h µht+1 ).

(11.3.3)

As a prelude to computing demand for consumption, we compute the demand for services. Our strategy is to use (11.3.2) and (11.3.3) to solve for the multiplier µst , and then to substitute this solution into (11.3.1). Shift (11.3.2) forward one time period and solve (11.3.2) for µst+1 . Substitute this expression into (11.3.3): ′ −1′ 0 µht = βEt (−Λ′ Π−1′ Θ′h µht+1 + µw pt+1 + ∆′h µht+1 ). 0Λ Π

(11.3.4)

Solve (11.3.4) forward to obtain: µht

=

µw 0 Et

∞ X

τ =1

β τ (∆′h − Λ′ Π−1′ Θ′h )τ −1 Λ′ Π−1′ p0t+τ .

(11.3.5)

Dynamic Demand Functions for Consumption Goods

241

Because we are using a canonical household service technology, the infinite sum on the right side of (11.3.5) converges (in L20 ). Therefore, the service demand can be expressed as 0 st = bt − µw 0 ρt

(11.3.6)

where ∞ h i X ρ0t ≡ Π−1′ p0t − Θ′h Et β τ (∆′h − Λ′ Π−1′ Θ′h )τ −1 Λ′ Π−1′ p0t+τ .

(11.3.7)

τ =1

Equations (11.3.6) and (11.3.7) represent the service demands in terms of expected future prices of the consumption good. The random vector ρ0t is the implicit rental price for services expressed in terms of current and expected future prices of consumption goods. Equation (11.2.1) transforms {st } in L20 into {ct } in L20 .

11.3.1. The multiplier µw0 The service demands given in (11.3.6) depend on the endogenous scalar mulw tiplier µw 0 . To compute µ0 , we partition the household capital and service sequences into two components. One component is a service sequence obtained from the initial endowment of household capital. The other component is the service sequence obtained from market purchases of consumption goods. The service sequence {si,t } obtained from the initial endowment of household capital evolves according to: si,t = Λhi,t−1 (11.3.8) hi,t = ∆h hi,t−1 where hi,−1 = h−1 . The service sequence {sm,t } obtained from purchases of consumption satisfies: 0 sm,t = bt − si,t − µw 0 ρt .

(11.3.9)

We can compute the time zero cost of the sequence {sm,t } in one of two equivalent ways. One way is to compute the time zero cost of the consumption sequence {ct } needed to support the service demands using the price sequence

242

Representation of Demand

{p0t }. Another way is to use the implicit rental sequence {ρ0t } directly to compute the time zero costs of {sm,t }. In the appendix to this chapter, we verify that the two measures of costs agree: E0

∞ X t=0

β t ρ0t · sm,t = E0

∞ X t=0

β t p0t · ct .

(11.3.10)

It is reasonable that, starting from h−1 = 0, the value of services equals the value of the associated consumption stream. It follows from (11.3.9) that

E0

∞ X t=0

β t ρ0t · sm,t = E0

∞ X t=0

β t ρ0t · (bt − si,t ) − µw 0 E0

∞ X t=0

β t ρ0t · ρ0t .

(11.3.11)

Substitute (11.3.10) and (11.3.11) into the consumer’s budget constraint (6.2), and solve for the time zero marginal utility of wealth µw 0 : µw 0 =

E0

P∞

β t ρ0t · (bt − si,t ) − W0 P∞ , E0 t=0 β t ρ0t · ρ0t

t=0

(11.3.12)

where W0 denotes initial period wealth given by W0 = E 0

∞ X t=0

β t (wt0 ℓt + αt0 · dt ) + v0 · k−1 .

(11.3.13)

Taken together, (11.3.6), (11.3.7), (11.3.12) and (11.3.13) give the demand functions for consumption services. A recursive representation for the dynamic demand function for consumption goods is obtained by substituting for st in (11.2.1).

Dynamic Demand Functions for Consumption Goods

243

11.3.2. Dynamic Demand System Substituting (11.3.6) and (11.3.7) into (11.2.1) gives ′ −1 ct = −Π−1 Λht−1 + Π−1 bt − Π−1 µw − Π′ −1 Θ′h 0 Et {Π

[I − (∆′h − Λ′ Π′ −1 Θ′h )βL−1 ]−1 Λ′ Π′−1 βL−1 }p0t

(11.3.14)

ht = ∆h ht−1 + Θh ct

Equation system (11.3.14) can be regarded as a system of dynamic demand functions for consumption, which express consumption demand at date t as a function of future scaled Arrow-Debreu prices p0t and, as mediated through the state variable ht−1 , past values of consumption.

11.3.3. Foreshadow of Gorman aggregation In the chapter 12, we shall explore how the dynamic demand schedule for consumption goods opens up the possibility of satisfying Gorman’s (1953) conditions for aggregation in a heterogeneous consumer version of the model. The first equation of (11.3.14) amounts to an Engle curve for consumption that is linear 1 in individual wealth with a coefficient on µw o (which depends on wealth) that only depends on prices. In a model of consumers who have the same household technologies (∆h , Θh , Λ, Π ) but possibly different preference shock processes, the coefficient on wealth is the same for all consumers. Gorman showed that when Engel curves satisfy this property, there exists a unique community or aggregate preference ordering over aggregate consumption that is independent of the distribution of wealth. This property will be exploited in chapter 12 when we solve for the equilibrium of a multiple consumer version of our economy. The community dynamic demand schedule for a heterogeneous agent economy will be obtained by summing the individual Engel curves.

1 Through ( 11.3.12 ) the multiplier µw depends on wealth in an affine relationship. 0

244

Representation of Demand

11.4. Computing Canonical Representations In deriving a dynamic demand function, we assumed that the representation of the household service technology is canonical. Now we start with a preference shock process {bt } and a specification of (∆h , Θh , Λ, Π) that is not necessarily canonical and show how to find a canonical representation. In the appendix, we establish that for any (∆h , Θh , Λ, Π), there exists a canonical service technology ˆ Π) ˆ and accompanying preference shock process {ˆbt } that induces (∆h , Θh , Λ, an identical preference ordering over consumption. In the text, we display the mechanics of how to compute the canonical technology and associated preference shock process, assigning the technical details to the appendix. 2 These mechanics are closely related to mathematics of innovations representations.

11.4.1. Heuristics We study two polynomials in the lag operator L: σ(L) = Π + ΛL[I − ∆h L]−1 Θh ˆ + ΛL[I ˆ σ ˆ (L) = Π − ∆h L]−1 Θh . As explained in the appendix, when ct = 0∀t < 0, applying the operator σ(L) to ct gives st , so that st = σ(L)ct . For two household technologies [∆h , Θh , Π, Λ] ˆ h , Θh , Π, ˆ Λ] ˆ to give rise to the same preference ordering over {ct } it is and [∆ necessary that σ(β .5 L−1 )′ σ(β .5 L) = σ ˆ (β .5 L−1 )′ σ ˆ (β .5 L). ˆ Π] ˆ technology is to be canonical, it is necessary that σ If the [Λ, ˆ (β .5 L) be invertible. In the appendix, we verify the following version of the factorization identity: [Π + β 1/2 L−1 Λ(I − β 1/2 L−1 ∆h )−1 Θh ]′ [Π + β 1/2 LΛ(I − β 1/2 L∆h )−1 Θh ] ˆ + β 1/2 L−1 Λ(I ˆ − β 1/2 L−1 ∆h )−1 Θh ]′ [Π ˆ + β 1/2 LΛ(I ˆ − β 1/2 L∆h )−1 Θh ], = [Π

ˆ Λ] ˆ satisfy (10.16), (10.19), and (10.20) below. As part of the factorwhere [Λ, ˆ Π] ˆ representation satisfies both of the ization identity, it is proved that the [Λ, requirements to achieve the status of a canonical representation. Thus, to attain a canonical household technology, we have to implement this factorization. We can do this by solving a control problem. 2 The MATLAB program canonpr.m computes a canonical representation.

Computing Canonical Representations

245

11.4.2. An auxiliary problem that induces a canonical representation An artificial optimization problem and the associated optimal linear regulator facilitate computing a canonical representation. Thus, confront a household with the optimization problem: choose {ct }∞ t=0 to maximize −.5

∞ X t=0

β t (st − bt ) · (st − bt )

subject to ht = ∆h ht−1 + Θh ct st = Λht−1 + Πct . The recursive solution to this optimization problem contains all of the ingredients for a canonical service technology. This optimization problem is a version of one a household confronts in a competitive equilibrium, except that we have eliminated the budget constraint. For a canonical technology, the solution to this optimization problem is trivial: choose {ct } so that the implied service sequence matches the preference shock sequence, st = bt ∀ t . However, when the service technology is not canonical, it might not be feasible to construct a consumption process that attains that goal, in which case the optimization problem is not trivial. We simplify the household optimization problem further by initially setting the preference shock process to zero for all t ≥ 0. In making this simplification, we are exploiting the fact that for the optimal linear regulator problem, the feedback part of the decision rule can be computed independently of the feedforward part, and that the {bt } process influences only the feedforward part. In this optimization problem it is feasible to stabilize the state vector {ht } so that it satisfies the square summability requirement. For instance, one can set the consumption process to zero for all t ≥ 0. So long as it is also optimal to stabilize the household capital stock process, it is known that there will be a unique positive semidefinite matrix P satisfying the algebraic Riccati equation: 3 P = Λ′ Λ + β∆′h P ∆h − (β∆′h P Θh + Λ′ Π)

(Π′ Π + βΘ′h P Θh )−1 (βΘ′h P ∆h + Π′ Λ).

(11.4.1)

3 We require that assumption A1 and the stability theorem of chapter 9 apply to this control problem.

246

Representation of Demand

The optimal choice of consumption can be represented as ct = −(Π′ Π + βΘ′h P Θh )−1 (βΘ′h P ∆h + Π′ Λ)ht−1 .

(11.4.2)

When this optimal rule is implemented, the evolution equation for the household capital stock is ht = [∆h − Θh (Π′ Π + βΘ′h P Θh )−1 (βΘ′h P ∆h + Π′ Λ)]ht−1 ,

(11.4.3)

where the eigenvalues of the matrix multiplying ht−1 are strictly less than √ ˆ and Λ ˆ so that 1/ β . 4 With this in mind, we choose Π ˆ −1 Λ ˆ = (Π′ Π + βΘ′h P Θh )−1 (βΘ′h P ∆h + Π′ Λ). Π

(11.4.4)

For this choice, condition (ii) for a canonical service technology is met. ˆ . In the appendix, it is shown as an implication We still have to construct Π ˆ to be a factor of the symmetric of the factorization identity that we should set Π ′ ′ positive definite matrix (Π Π + βΘh P Θh ): ˆ ′ Π. ˆ (Π′ Π + βΘ′h P Θh ) = Π

(11.4.5)

ˆ is a square matrix. Since (Π′ Π + Any factorization will work so long as Π ′ ˆ βΘh P Θh ) is nonsingular, Π satisfies condition (i) for a canonical representation. ˆ Λ) ˆ that corIn summary, (11.4.1), (11.4.4), and (11.4.5) compute a (Π, responds to a canonical representation. The service process {ˆ st } for this new household technology satisfies: ˆ t−1 + Πc ˆ t. sˆt = Λh

(11.4.6)

We also need to construct a preference shock process to accompany the canonical service technology. One way to do this is simply to reintroduce the preference shock process {bt } into the auxiliary household optimization problem, and recompute the optimal decision rule for consumption. The decision rule can be represented as: ˆ t−1 + (Π) ˆ −1 U ˆ b zt ˆ −1 Λh ct = −(Π)

(11.4.7)

4 This condition on the eigenvalues of the ‘closed loop system’ follows from the assumption that it is optimal to stabilize the system (i.e., that the system is detectable).

Computing Canonical Representations

247

ˆb . As discussed in chapter 4, the feedback portion of this for some matrix U −1 ˆ ˆ decision rule [(Π) Λ] is the same as for the problem in which the preference ˆ −1 U ˆb ] can be computed shock process was set to zero. The feedforward part [(Π) using the method described in chapter 4, which permits the optimal decision rule to be calculated efficiently in two steps. Using those methods, the shock process associated with the canonical service technology is ˆbt = U ˆ b zt .

(11.4.8)

An alternative method for computing {ˆbt } is more useful and revealing. As shown in the appendix to this chapter, two preference representations having the same demand functions give rise to the same preference ordering over consumption paths. Therefore, the marginal utilities are also the same across the two preference representations, and in particular across the two specifications of household technologies and preference shock processes. Equality between the indirect marginal utility of consumption and the current and expected future marginal utilities of consumption services and (11.3.1), (11.3.2), and (11.3.3) implies that the two preference shock processes must satisfy: ′

Π bt +

Θ′h Et

∞ X

β

τ

(∆′h )τ −1 Λ′ bt+τ

τ =1

ˆ ′ˆbt + Θ′ Et =Π h

∞ X

ˆ ′ˆbt+τ . β τ (∆′h )τ −1 Λ

τ =1

(11.4.9) ˆ Π) ˆ technology Let the left side of (11.4.9) be denoted ˜bt for each t . Since the (Λ, is canonical, it follows that we can solve (11.4.9) for ˆbt : ˆbt = Π ˆ −1′˜bt − Π ˆ −1′ Θ′ Et h

∞ X

τ =1

ˆ −1′ Θ′ )τ −1 Λ ˆ ′Π ˆ −1′˜bt+τ . ˆ ′Π β τ (∆′h − Λ h

(11.4.10)

Relation (11.4.10) is derived by applying an operator identity to equation (11.4.9).

248

Representation of Demand

11.5. Operator Identities For canonical household technologies a matrix identity is [Π + Λ(I − ∆h L)−1 Θh L]−1 =

{Π−1 −Π−1 Λ[I − (∆h − Θh Π−1 Λ)L]−1 Θh Π−1 L}.

The identity shows that for canonical representations of preferences (∆h , Θh , Π, Λ, Ub ) , there are two equivalent ways of expressing the mapping between sequences {st } and sequences {ct } . To establish the identity, assume that h−1 = 0 , or equivalently that ct = 0 ∀ t < 0 . Note that the second equation of representation ( 11.2.1 ) implies ht = [I − (∆h − Θh Π−1 Λ)L]−1 Θh Π−1 st .

Lagging this one period and substituting into the first equation of ( 11.2.1 ) gives ct = {Π−1 − Π−1 Λ[I − (∆h − Θh Π−1 Λ)L]−1 Θh Π−1 L}st .

This equation shows how to obtain sequences {ct } ∈ L20 that are associated with arbitrary sequences {st } ∈ L20 . Now recall that the household technology implies st = [Π + Λ(I − ∆h L)−1 Θh L]ct ,

which expresses {st } ∈ L20 as a function of {ct } ∈ L20 . The assumption that (Λ, Π) is canonical implies that the operator [Π + Λ(I − ∆h L)−1 Θh L] mapping sequences from L20 into L20 is invertible, which implies the identity. Here is how to derive the ‘dual’ or transposed version of the identity, which is the one used to get ( 11.4.10 ). Use ( 11.3.3 ) to deduce ′ −1 )−1 βΛ′ L−1 µs . µh t = (I − β∆h L t+1

Then use ( 11.3.2 ) to deduce (†)

′ ′ ′ −1 )−1 βL−1 ]µs . µw t 0 pt = [Π + Θh (I − β∆h L

Alternatively, solve ( 11.3.2 ) for µst ,

w 0 µst = Π′ −1 (−Θ′h µh t + µ0 pt ).

Substitute this into ( 11.3.3 ) to get (‡)

0 µst = {Π′ −1 − Π′ −1 Θ′h [I − (∆′h − Λ′ Π′ −1 Θ′h )βL−1 ]−1 Λ′ Π′−1 βL−1 }µw 0 pt .

When (∆h , Θh , Π, Λ) is canonical, the operator on the right side of (†) has an inverse equal to the operator on the right side of (‡) : [Π′ + Θ′h (I − β∆′h L−1 )−1 βL−1 ]−1 =

{Π′ −1 −Π′ −1 Θ′h [I − (∆′h − Λ′ Π′ −1 Θ′h )βL−1 ]−1 Λ′ Π′−1 βL−1 }.

In the appendix to this chapter, we use Fourier transforms to show that the ˆ Π) ˆ and preference shock process {ˆbt } induce alternative service technology (Λ, the same preference ordering for consumption goods as did the original ones.

Becker-Murphy Model of Rational Addiction

249

11.6. Becker-Murphy Model of Rational Addiction We illustrate our analysis with a discrete-time version of the habit-persistence model advocated by Becker and Murphy (1988). The household technology is a parametric version of induced preferences for consumption of the form suggested by Pollak (1970), Ryder and Heal (1973), and Stigler and Becker (1977). The household technology has a single consumption good, two services, and a single household capital stock. The household capital measures a habit stock constructed to be a geometrically-weighted average of current and past consumptions: ht = δh ht−1 + (1 − δh )ct , (11.6.1) where 0 < δh < 1. The first service is proportional to consumption, and the second one is a linear combination of consumption and the habit stock: π1 0 ct st = . (11.6.2) π2 π3 ht We normalize π1 and π3 to be strictly positive. Imagine for a moment that ct and ht are distinct consumptions goods and that there is no intertemporal connection between them. Then recall from our discussion of preferences for multiple consumption goods in Chapter 3, that the Frisch classification of complements is equivalent to requiring π2 to be negative. In light of the evolution equation (11.6.1) for the household capital stock, this notion of complementarity is limiting because it ignores the fact that ht is a weighted average of current and past consumptions. For this reason, we consider a related notion of complementarity referred to by Ryder and Heal (1973) and Becker and Murphy (1988) as adjacent complementarity. Substituting (11.6.1) into (11.6.2) we obtain the following service technology: st = Λht−1 + Πct , where Λ=

0 π3 δ h

and Π =

(11.6.3)

π1 . π2 + π3 (1 − δh )

Service technology (11.6.3) is clearly not canonical: simply note that two services are constructed from one underlying consumption good, so we cannot construct a consumption sequence to support any hypothetical admissible service sequence.

250

Representation of Demand

To capture the notion of adjacent complementarity, we consider a canonical representation for household services. The canonical household service technology has a single service and can be expressed as: ˆ t−1 + Πc ˆ t, sˆt = Λh

(11.6.4)

ˆ Π) ˆ satisfies: where {ˆ st } is a scalar service process and (Λ, p ˆ Π ˆ |< 1/ β. | δh − (1 − δh )Λ/

(11.6.5)

p ˆ Π ˆ ≤ 1/ β. 0 ≤ δh − (1 − δh )Λ/

(11.6.6)

ˆ to be positive so that increases in time t consumption We normalize the scalar Π increase the time t canonical service sˆt . When specialized to this parametric model, Ryder and Heal’s (1973) notion of adjacent complementarity becomes ˆ must be negative. In this case, (11.6.5) implies that the restriction that Λ

ˆ ≤ 0) As shown by Becker and Murphy (1988), adjacent complementarity (Λ implies that π2 ≤ 0. The converse is not true, however. To see the relation ˆ and π2 , multiply both sides of (9.64) by (1−β 1/2 ζ −1 δh )(1−β 1/2 ζδh ) between Λ to obtain: Π′ Π (1 − β 1/2 ζ −1 δh )(1 − β 1/2 ζδh ) + βΛ′ Λ(1 − δh )2 +

β 1/2 ζ −1 (1 − δh )(1 − β 1/2 ζδh )Λ′ Π + β 1/2 ζ(1 − δh )(1 − β 1/2 ζ −1 δh )Λ′ Π ˆ 2 (1 − β 1/2 ζ −1 δh )(1 − β 1/2 ζδh ) + β Λ ˆ Λ(1 ˆ − δh )2 + =Π

ˆΠ ˆ + β 1/2 ζ(1 − δh )(1 − β 1/2 ζ −1 δh )Λ ˆ Π. ˆ β 1/2 ζ −1 (1 − δh )(1 − β 1/2 ζδh )Λ (11.6.7) This equality holds for all ζ except ζ = 0. Evaluate both sides of (11.6.7) at ζ = β 1/2 δh : βΛ′ Λ(1 − δh )2 + (1 − δh )(1 − βδh2 )Λ′ Π/δh ˆ 2 (1 − δh )2 + (1 − δh )(1 − βδh2 )Λ ˆ Π/δ ˆ h. = βΛ

(11.6.8)

The right side of (11.6.8) can be expressed as ˆ 2 (1 − δh )2 + (1 − δh )(1 − βδ 2 )Λ ˆ Π/δ ˆ h βΛ h ˆ Π{[(1 ˆ ˆ Π ˆ − δh ] + (1/βδh )}. = β(1 − δh )Λ − δh )Λ/

(11.6.9)

Fourier transforms

251

ˆΠ ˆ < 0 and inequality (11.6.6) is satisfied, it follows that Since Λ ˆ Π{[(1 ˆ ˆ Π ˆ − δh ] + (1/βδh )} β(1 − δh )Λ − δh )Λ/ p ˆ Π[−1/ ˆ ≤ β(1 − δh )Λ β + (1/βδh )]

(11.6.10)

≤ 0.

ˆ ≤ 0, then Combining (11.6.10) and (11.6.8), we have that if Λ βΛ′ Λ(1 − δh )2 + (1 − δh )(1 − βδh2 )Λ′ Π/δh ≤ 0.

(11.6.11)

Inequality (11.6.11) is satisfied only when Λ′ Π ≤ 0. This in turn requires that π2 ≤ 0 because Λ′ Π = π3 δh [π2 + π3 (1 − δh )], (11.6.12) 0 < δh < 1 and π3 > 0. ˆ Λ ˆ to exceed one. In this case, growth Inequality (11.6.6) permits δh −(1−δh )Π/ in consumption is required to support most constant service sequences, although this growth will be dominated by {β t/2 : t = 0, 1, . . .}. This is a household technology with an extreme form of addiction to the consumption good. Note that ˆ Π ˆ = δh (1 + Λ/ ˆ Π) ˆ − Λ/ ˆ Π. ˆ δh − (1 − δh )Λ/ (11.6.13) ˆ exceeds Π ˆ in the canonical houseTherefore, instability is implied whenever −Λ hold service technology.

A. Fourier transforms This appendix applies Fourier transforms to establish some key equalities asserted in the text. We begin with some background on 0.

252

Representation of Demand

11.A.1. Primer on transforms For a two-sided scalar sequence {cj }∞ j=−∞ , the z− transform is defined as the complex valued function ∞ X cj z j , c(z) = j=−∞

5

where z is a scalar complex number. The inversion formula asserts Z 1 ck = c(z)z −k−1 dz 2πi Γ where Γ is any closed contour around zero in the complex plane, and the integration is complex integration counterclockwise along the path Γ . If we take Γ to be the unit circle and set z = e−iω , we get the following version of the inversion formula Z π 1 c(e−iω )eiωk d ω. ck = 2π −π We denote transform pairs with the notation {ck } ↔ c(z). The convolution of two sequences {yk }, {xk }, is denoted {y ∗ x} and is defined as ∞ X {y ∗ x}∞ ≡ { ys xk−s }∞ k=−∞ k=−∞ . s=−∞

Direct calculations establish the convolution property {y ∗ x}∞ k=−∞ ↔ x(z)y(z). We have the linearity property that for any scalars (a, b) a{xk } + b{yk } ↔ ax(z) + by(z).

5 For descriptions of Fourier and z-transforms, see Gabel and Roberts (1973), Liu and Liu (19**). For some of their uses in time series economics see Nerlove, Grether and Carvalho (1979) and Sargent (1979).

Fourier transforms

253

11.A.2. Time reversal and Parseval’s formula Let c˜−k = ck for all k . Then {˜ ck }∞ k=−∞ has transform c˜(z) =

∞ X

∞ X

c˜k z k =

ck z −k = c(z −1 ).

k=−∞

k=−∞

Applying the convolution theorem to c(z)c(z −1 ) gives c(z)c(z

−1

∞ X

)↔{

s=−∞

cs cs−k }∞ k=−∞ .

Applying the inversion formula gives ∞ X

cs cs−k

s=−∞

1 = 2π

Z

π

c(e−iω )c(eiω )eiωk d ω.

−π

If we set k = 0, we obtain Parseval’s equality: ∞ X

s=−∞

c2s

1 = 2π

Z

π

−π

|c(e−iω )|2 d ω.

11.A.3. One sided sequences There are two types of one-sided sequences (also called ‘half-infinite’ sequences). A sequence is called a causal sequence if ck = 0 ∀k < 0, and is anti-causal if it has zero elements ∀ k > 1. A one-sided causal sequence can be obtained by setting to zero all elements of a two-sided sequence with negative subscripts. Let {uk }∞ k=−∞ be the step sequence, which is zero for k < 0, and 1 for k ≥ 0. Evidently {uk ck } is always a one-sided sequence.

254

Representation of Demand

11.A.4. Useful properties 1. z0 is said to be a pole of order m ≥ 1 of c(z) if limz→z0 (z − z0 )m c(z) 6= 0. 2. c(z) is the transform of a causal sequence if all of its poles lie outside the unit circle. 3. c(z) is the transform of an anti-causal sequence if all of its poles lie inside the unit circle. 4. If c(z) is either causal or anti-causal, the inversion formula can be implemented by ‘long division.’ 5. Initial value theorem: lim c(z) = c0 .

z→0

6. Final value theorem: lim ck = lim (1 − z)c(z). z→1

k→∞

11.A.5. One sided transforms A one-sided transform is defined as c+ (z) =

∞ X

k=0

ck z k ≡ [c(z)]+ ,

where [ ]+ is the ‘annihilation operator’ that sets to zero all coefficients on negative powers of z . The same inversion formulas hold, with c+ (z) replacing c(z). Notice that c+ (z) = c(z) only if {ck } is causal. We shall adopt the notation F (c)(z) = c+ (z). For one-sided transforms, we have the shift theorem F ({ct−n })(z) = z n F ({ct })(z) +

n X

k=1

z n−k c−k .

Fourier transforms

255

11.A.6. Discounting For the purpose of introducing discounting, we shall work with the alternative transformation defined by t/2 ∞ T ({ct }∞ }t=0 )(z), t=0 )(z) ≡ F ({ct β

so that T (y) is the ordinary transform of {β t/2 yt }. The inversion formula is then Z π 1 β t/2 yt = T (e−iω )eiωt d ω, 2π −π and the shift theorem is .5

n

T ({ct−n })(z) = (β z) F ({ct })(z) +

n X

(zβ .5 )n−k c−k .

k=1

11.A.7. Fourier transforms Below we shall work with vector versions of the transforms T . Consider a vector sequence y = {yt } satisfying ∞ X t=0

β t yt · yt < ∞,

define the transform: T (y)(ζ) ≡

∞ X

(11.A.1)

β t/2 yt ζ t .

(11.A.2)

t=0

This transform is at least well-defined for | ζ |< 1 and can also be defined through an appropriate limiting argument for | ζ |= 1. 6 For vector sequences {yt } and {ˆ yt } satisfying (11.A.1), Parseval’s formula is (1/2π)

Z

π −π

T (y)[exp(iθ)] · T (ˆ y )[exp(−iθ)]dθ =

∞ X t=0

β t (yt · yˆt ).

(11.A.3)

6 The boundary of the unit circle can be parameterized by ζ = exp(iθ) for θ ∈ (−π, π] . Using this parameterization, the infinite series on the right side of ( 11.4.6 ) converges in L2 where the L2 space is constructed using Lebesgue measure on (−π, π] .

256

Representation of Demand

We use Fourier 0 to represent our dynamic household technologies. It follows from (11.2.1) and the definitions of smt and sit that ΠT (c)(ζ) = −β 1/2 ζΛT (hm )(ζ) + T (sm )(ζ)

T (hm )(ζ) =β 1/2 ζ(∆h − Θh Π−1 Λ)T (hm )(ζ) + Θh Π−1 T (sm )(ζ)

(11.A.4)

where hm,−1 = 0. The transforms of the consumption sequence and the market service sequence are related by T (c)(ζ) = C (ζ)T (sm )(ζ)

(11.A.5)

where o n C (ζ) ≡ Π−1 I − β 1/2 ζΛ[I − β 1/2 ζ(∆h − Θh Π−1 Λ)]−1 Θh Π−1 .

(11.A.6)

The matrix function C of a complex variable ζ represents the mapping from desired consumption services into the consumption goods required to support those services.

11.A.8. Verifying Equivalent Valuations Our derivation of the dynamic demand functions for consumption goods relied on two intermediate results: (a) equivalent 0 of market services and consumption goods asserted in (11.3.10); and (b) for a given specification of preferences and household technology, the existence of a canonical service technology that induces the same preference ordering over consumption streams. To establish these intermediate results we use Fourier transforms. We now show establish the valuation equivalence asserted in (11.3.10). Applying Parseval’s formula (11.A.3), we have that ∞ X

β t p0t

t=0

= (1/2π) = (1/2π)

· ct = (1/2π)

Z

π

−π Z π −π

Z

π −π

T (p0 )[exp(iθ)] · T (c)[exp(−iθ)]dθ

T (p0 )[exp(iθ)] · {C [exp(−iθ)]T (sm )[exp(−iθ)}]dθ {C [exp(−iθ)]′ T (p0 )[exp(iθ)]} · T (sm )[exp(−iθ)]dθ.

(11.A.7)

Fourier transforms

257

Formula (11.A.7) gives us the following candidate for the transform of the rental sequence for consumption services: C (ζ −1 )′ T (p0 )(ζ). The rental sequence {˜ ρ0t } associated with this transform is given by: n o ρ˜0t ≡ Π−1′ I − βL−1 Θ′h [I − βL−1 (∆h − Θh Π−1 Λ)′ ]−1 Λ′ Π−1′ p0t ∞ h i X = Π−1′ p0t − Θ′h β τ (∆h − Θh Π−1 Λ)′τ −1 Λ′ Π−1′ p0t+τ .

(11.A.8)

τ =1

Using this rental sequence, it follows from (11.A.5) that ∞ X t=0

β t p0t · ct =

∞ X t=0

β t ρ˜0t · smt .

(11.A.9)

Notice that the candidate rental sequence {˜ ρ0t } violates the information constraints because ρ˜0t will not necessarily be in Jt . From the vantage point of valuation, all that we require is equality of the expectations of the infinite sums in (11.A.9) conditioned on J0 . It follows from the Law of Iterated Expectations that E0 ρ˜0t · st = E0 ρ0t · st (11.A.10) where ρ0t ≡ Et ρ˜0t ,

(11.A.11)

since hypothetical service vectors st are restricted to be in the information set Jt . Taking expectations of both sides of (11.A.11) conditioned on J0 and substituting from (11.A.11) establishes the value equivalence given in (11.3.10).

258

Representation of Demand

11.A.9. Equivalent representations of preferences We now turn to task (b), to show that the candidate canonical representation of the service technology implies the same induced preference ordering for conˆ Π), ˆ sumption. There are two preference representations on the table (Λ, Π), (Λ, where the objects with hats are canonical. Again we partition the household capital stock and the consumption service process into two components. Similar to (11.A.4) we have that T (sm )(ζ) = β 1/2 ζΛT (hm )(ζ) + ΠT (c)(ζ) T (hm )(ζ) = β 1/2 ζ∆h T (hm )(ζ) + Θh T (c)(ζ).

(11.A.12)

Hence T (sm )(ζ) = S (ζ)T (c)(ζ)

(11.A.13)

S (ζ) ≡ [Π + β 1/2 ζΛ(I − β 1/2 ζ∆h )−1 Θh ].

(11.A.14)

where The function S represents the mapping from consumption goods into market supplied consumption services. An analogous argument leads to the formula: T (ˆ sm )(ζ) = Sˆ(ζ)T (c)(ζ)

(11.A.15)

ˆ + β 1/2 ζ Λ(I ˆ − β 1/2 ζ∆h )−1 Θh ], Sˆ(ζ) ≡ [Π

(11.A.16)

where where objects with hats, including sˆm , correspond to the canonical representation. It is straightforward to show that T (si )(ζ) = Λ(I − β 1/2 ζ∆h )−1 h−1

(11.A.17)

ˆ − β 1/2 ζ∆h )−1 h−1 . T (ˆ si )(ζ) = Λ(I

(11.A.18)

and The time t contribution to the consumers’ utility function can be expressed as: h i −(1/2)β t (bt − si,t − sm,t ) · (bt − si,t − sm,t ) = −(1/2)β t sm,t · sm,t + 2sm,t · si,t − 2sm,t · bt + (11.A.19) (bt − si,t ) · (bt − si,t ) .

Fourier transforms

259

Note that the fourth term is not affected by the consumption choice, and thus can be ignored. We now study the Fourier representations of the sums: ∞ X t=0

β t sm,t · sm,t ,

∞ X t=0

β t sm,t · si,t and

∞ X t=0

β t sm,t · bt .

(11.A.20)

11.A.10. First term: factorization identity The first infinite sum in (11.A.20) can be represented as: ∞ X

(1/2π)

β t sm,t · sm,t =

t=0 πn

Z

−π

o′ T (c)[exp(iθ)] S [exp(iθ)]′ S [exp(−iθ)]

(11.A.21)

T (c)[exp(−iθ)]dθ.

ˆ Λ) ˆ and {ˆbt } imply the same induced preferences for consumpTo show that (Π, tion goods, we must first establish the factorization: S (ζ −1 )′ S (ζ) = Sˆ(ζ −1 )′ Sˆ(ζ).

(11.A.22)

To verify this result, note that [Π + β 1/2 ζ −1 Λ(I − β 1/2 ζ −1 ∆h )−1 Θh ]′ [Π + β 1/2 ζΛ(I − β 1/2 ζ∆h )−1 Θh ] = Π′ Π + βΘ′h (I − β 1/2 ζ −1 ∆h )−1′ Λ′ Λ(I − β 1/2 ζ∆h )−1 Θh

+ β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ Λ′ Π

+ β 1/2 ζΠ′ Λ(I − β 1/2 ζ∆h )−1 Θh .

(11.A.23) Since P satisfies the algebraic Riccati equation(11.4.1), it follows that ˆ ′Λ ˆ Λ′ Λ = P − β∆′h P ∆h + Λ

= (I − β 1/2 ζ −1 ∆h )′ P (I − β 1/2 ζ∆h ) + β 1/2 ζ −1 ∆′h P (I − β 1/2 ζ∆h ) (11.A.24) ˆ ′ Λ. ˆ + β 1/2 ζ(I − β 1/2 ζ −1 ∆h )′ P ∆h + Λ

260

Representation of Demand

Therefore, Θ′h (I − β 1/2 ζ −1 ∆h )−1′ Λ′ Λ(I − β 1/2 ζ∆h )−1 Θh

= Θ′h P Θh + β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ ∆′h P Θh

+ β 1/2 ζΘ′h P ∆h (I − β 1/2 ζ∆h )−1 Θh ˆ ′ Λ(I ˆ − β 1/2 ζ∆h )−1 Θh . + Θ′ (I − β 1/2 ζ −1 ∆h )−1′ Λ

(11.A.25)

h

Furthermore, it follows from (11.4.4) and (11.4.5) that ˆ ′Λ ˆ =Π ˆ ′ Π( ˆ Π) ˆ −1 Λ ˆ Π = (βΘ′h P Θh + Λ′ Π).

(11.A.26)

Substituting (11.A.25) and (11.A.26) into (11.A.23) results in [Π + β 1/2 ζ −1 Λ(I − β 1/2 ζ −1 ∆h )−1 Θh ]′ [Π + β 1/2 ζΛ(I − β 1/2 ζ∆h )−1 Θh ] ˆ ′ Λ(I ˆ − β 1/2 ζ∆h )−1 Θh = Π′ Π + βΘ′h P Θh + βΘ′h (I − β 1/2 ζ −1 ∆h )−1′ Λ + β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ (β∆′h P Θh + Λ′ Π) + β 1/2 ζ(ΠΛ′ + βΘ′h P ∆h )(I − β 1/2 ζ∆h )−1 Θh ˆ ′Π ˆ + βΘ′h (I − β 1/2 ζ −1 ∆h )−1′ Λ ˆ ′ Λ(I ˆ − β 1/2 ζ∆h )−1 Θh =Π ˆ ′Π ˆ + β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ Λ ˆ − β 1/2 ζ∆h )−1 Θh ˆ ′ Π(I + β 1/2 ζ Λ ˆ + β 1/2 ζ −1 Λ(I ˆ − β 1/2 ζ −1 ∆h )−1 Θh ]′ [Π ˆ + β 1/2 ζ Λ(I ˆ − β 1/2 ζ∆h )−1 Θh ] = [Π (11.A.27)

which proves factorization (11.A.22).

Fourier transforms

261

11.A.11. Second term The second infinite sum in (11.A.20) can be represented as Z π ∞ X β t sm,t · si,t = (1/2π) {T (c)[exp(iθ)]}′ S [exp(iθ)]′ Λ −π

t=0

[I − β

1/2

−1

exp(−iθ)∆h ]

(11.A.28)

h−1 dθ.

We will verify that S (ζ −1 )′ Λ∆h (I − β 1/2 ζ∆h )−1 = ˆ h (I − β 1/2 ζ∆h )−1 + β 1/2 ζ −1 Θ′ (I − β 1/2 ζ −1 ∆h )−1′ P ∆h . Sˆ(ζ −1 )′ Λ∆ h (11.A.29) It then follows that (1/2π)

Z

π

−π π

= (1/2π)

Z

−π

{T (c)[exp(iθ)]}′ S[exp(iθ)]′ Λ∆h [I − β 1/2 exp(−iθ)∆h ]−1 h−1 dθ ′ˆ ˆ {T (c)[exp(iθ)]}′ S[exp(iθ)] Λ∆h [I − β 1/2 exp(−iθ)∆h ]−1 h−1 dθ

(11.A.30)

because

(1/2π)

Z

π −π

{T (c)[exp(iθ)]}′ β 1/2 exp(iθ) Θ′h [I

−β

1/2

−1′

exp(iθ)∆h ]

(11.A.31)

P ∆h h−1 dθ = 0.

Relation (11.A.31) holds since T (c)(ζ)′ β 1/2 ζ∆′h (I − β 1/2 ζ∆h )−1′ has a power series expansion and is zero when ζ = 0 and P ∆h h−1 can be viewed a constant function with a trivial power series expansion. Relation (11.A.31) then follows from Parseval’s formula (11.A.3) where β t/2 yt is constructed from the tth coefficient of the power series expansion for the first function and β t/2 y˜t from the tth coefficient of the power series expansion for the second function. It remains to establish (11.A.29). Note that the left side of (11.A.29) can be expanded as follows: [Π + β 1/2 ζ −1 Λ(I − β 1/2 ζ −1 ∆h )−1 Θh ]′ Λ(I − β 1/2 ζ∆h )−1 ∆h

= Π′ Λ(I − β 1/2 ζ∆h )−1 ∆h + β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ Λ′ Λ (11.A.32)

(I − β 1/2 ζ∆h )−1 ∆h .

262

Representation of Demand

It follows from the algebraic Riccati equation (11.4.1) that ˆ ′Λ ˆ Λ′ Λ = P (I − β 1/2 ζ∆h ) + β 1/2 ζP ∆h − β∆′h P ∆h + Λ

ˆ ′ Λ, ˆ = P (I − β 1/2 ζ∆h ) + β 1/2 ζ(I − β 1/2 ζ −1 ∆′h )P ∆h + Λ

(11.A.33)

and hence β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ Λ′ Λ(I − β 1/2 ζ∆h )−1 ∆h

= β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ P ∆h + βΘ′h P ∆h (I − β 1/2 ζ∆h )−1 ∆h ˆ ′ Λ(I ˆ − β 1/2 ζ∆h )−1 ∆h + β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ Λ (11.A.34) Substituting (11.A.34) and (11.A.26) into (11.A.32) gives [Π + β 1/2 ζ −1 Λ(I − β 1/2 ζ −1 ∆h )−1 Θh ]′ Λ(I − β 1/2 ζ∆h )−1 ∆h = Π′ Λ(I − β 1/2 ζ∆h )−1 ∆h +

β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ Λ′ Λ(I − β 1/2 ζ∆h )−1 ∆h = (Π′ Λ + βΘ′h P ∆h )(I − β 1/2 ζ∆h )−1 ∆h +

β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ P ∆h + ˆ ′ Λ(I ˆ − β 1/2 ζ∆h )−1 ∆h β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ Λ ˆ ′ Λ/(I ˆ =Π − β 1/2 ζ∆h )−1 ∆h +

ˆ − β 1/2 ζ∆h )−1 ∆h + ˆ ′ Λ(I β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ Λ

β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ P ∆h ˆ − β 1/2 ζ∆h )−1 ∆h + ˆ − β 1/2 ζ −1 ∆h )−1 Θh ]′ Λ(I ˆ + β 1/2 ζ −1 Λ(I = [Π β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ P ∆h

which establishes (11.A.29).

(11.A.35)

Fourier transforms

263

11.A.12. Third term The third sum in (11.A.20) can be represented as ∞ X t=0 π

(1/2π)

Z

β t sm,t · bt =

(11.A.36) ′

−π

′

{T (c)[exp(iθ)]} S [exp(iθ)] T (b)[exp(−iθ)]dθ.

Note that

S (ζ −1 )′ T (b)(ζ) = Sˆ(ζ −1 )′ Sˆ(ζ −1 )′ −1 S (ζ −1 )′ T (b)(ζ).

(11.A.37)

With this in mind, we define ˆbt = E{[Sˆ(L−1 )′ ]−1 S (L−1 )′ bt | Jt }.

(11.A.38)

Then by reasoning similar to that leading to result (a), we have that ∞ X t=0

t

β sm,t · bt =

∞ X t=0

β t sˆm,t · ˆbt .

(11.A.39)

Taken together (11.A.22), (11.A.29) and (11.A.37) show that the induced prefˆ Π) ˆ and {ˆbt } as it is for the erence ordering for consumption is the same for (Λ, original specification (Λ, Π) and {bt }. This establishes result (b).

Chapter 12 Gorman Heterogeneous Households

12.1. Introduction This chapter and the next describe methods for computing equilibria of versions of our economies in which consumers have heterogeneous preferences and endowments. In each chapter, we adopt simplifications that make it possible for us to cope with the complications introduced by heterogeneity. In the present chapter, we adopt a version of Terrance Gorman’s (1953) specification. We describe a class of heterogeneous consumer economies that satisfy Gorman’s conditions for aggregation, which lets us compute equilibrium aggregate allocations and prices before computing allocations to individuals. 1 In the following chapter, we adopt a more general specification of heterogeneity that causes us to depart from the representative consumer framework of Gorman. In particular, we adapt the idea of Negishi (1960), who described a social welfare function that is maximized, subject to resource and technological constraints, by a competitive equilibrium allocation. For Negishi, that social welfare function is a “linear combination of the individual utility functions of consumers, with the weights in the combination in inverse proportion to the marginal utilities of income.” Because Negishi’s weights depend on the allocation through the marginal utilities of income, computing a competitive equilibrium via constrained maximization of a Negishi-style welfare function requires solving a fixed point problem in the weights. In the following chapter, we apply such a fixed point approach. The beauty of Gorman’s aggregation conditions is that, when they apply, time series aggregates and market prices can be computed without resorting to Negishi’s fixed point approach. In the present chapter, consumers are permitted to differ only with respect to their endowments and the process {bt } that disturbs their preferences. We 1 The discussion in this chapter is patterned after the material in section 3 of Hansen (1987).

– 265 –

266

Gorman Heterogeneous Households

assume that all consumers have a common information set that includes observations on past values of the economy wide capital stocks ht−1 , kt−1 , and the common exogenous state variables in zt that drive each of the individual preference shock processes and the technology shock process {dt }. Preferences of the individual consumers can be aggregated simply by summing both preference shocks and initial endowments across consumers, thereby forming a representative consumer. We can compute all aggregate aspects of a competitive equilibrium of the economy with heterogeneous consumers by forming the representative consumer and proceeding as in chapters 3, 4, and 5. We show how to calculate individual allocations by using the demand functions that were described in the previous chapter. In the next section, we briefly describe Gorman aggregation in a standard static section before adapting it to our purposes.

12.2. A Digression on Gorman Aggregation Suppose for the moment that there are n consumption goods, taking into account indexation by dates and states, and that consumption of person j = 1, . . . , J is denoted cj . Let ca denote the aggregate amount of consumption to be allocated among consumers. Associated with ca is an Edgeworth box and a collection of Pareto optimal allocations. From the Pareto optimal allocations, one can construct utility allocation surfaces describing the frontier of alternative feasible utility assignments to individual consumers. Imagine moving from the aggregate vector ca to some other vector c˜a and hence to a new Edgeworth box. If neither the original box nor the new box contain one another, then it is possible that the utility allocation surfaces for the two boxes may cross, in which case there exists no ordering of aggregate consumption that is independent of the utility weights assigned to individual consumers. Before describing a special case in which an aggregate social preference ordering does exist, we illustrate a situation in which there doesn’t exist a social preference ordering that is independent of the aggregate allocation. Figures 12.2.1 and Fgpareto2f describe efficient allocations in a two person, two good, pure exchange economy with a structure of preferences that violate the

A Digression on Gorman Aggregation

267

6

Utility of Agent B

5

4

3

2

1

0 0

1

2

3

4

5

6

Utility of Agent A

Figure 12.2.1: Utility allocation for Agents A and B for endowment vectors E = (8,3) and E = (3,8). 1/3

2/3

Gorman aggregation conditions. Agent A has utility function U A = XA YA , 2/3 1/3 while consumer B has utility function given by U B = XB YB , where the aggregate endowment pair is E = (XA + XB , YA + YB ). Figure 12.2.1 shows two utility possibility frontiers, one associated with E = (8, 3), a second one associated with E = (3, 8). 2 The fact that the utility possibility frontiers in figure 12.2.1 cross indicates that the two aggregate endowment vectors (8, 3), (3, 8) cannot be ranked in a way that ignores how utility is distributed between consumers A and B. In the same economy, Figure 12.2.2 shows the Edgeworth boxes and contract curves with the two allocations E = (8, 3) and E = (3, 8). For a given endowment, the slope of the consumers’ indifference curves at the tangencies between indifference curves that determines the contract curve varies as one moves along the contract curve. This means that for a given aggregate endowment, the competitive equilibrium price depends on the allocation between consumers A and B. It follows that for this economy, one cannot expect to determine equilibrium prices independently of the equilibrium allocation. 1/3 2/3 2 A utility possibility frontier is the locus of pairs (U , U ) that solve U = max A B A XA ,YA XA YA 2/3 1/3 subject to the constraints XB YB ≥ U B , (XA + XB , YA + YB ) = E .

268

Gorman Heterogeneous Households

Gorman (1953) described restrictions on preferences under which it is possible to obtain a community preference ordering. Whenever Gorman’s conditions are satisfied, there occur substantial simplifications in solving multiple-consumer optimal resource allocation problems: in intertemporal contexts, it becomes possible first to determine the optimal allocation of aggregate resources over time. Then the aggregate consumption can be allocated among consumers by allocating utility levels to each person. To understand Gorman’s restrictions, imagine specifying the preferences of consumer j in one of two equivalent ways: in terms either of a family of indifference curves indexed by the utility level, or in terms of a family of compensated demand functions. Following Gorman (1953), let ψj (p) denote the baseline indifference curve for person j parameterized in terms of a price vector (or vector of utility gradients) p . In addition, let ψc (p) denote a common indifference curve for all consumers used to measure deviations from the baseline curves. This lets the compensated demand function for person j be represented as

8 7 6 5 4 3 2 1 0 0

1

2

3

4

5

6

7

8

Figure 12.2.2: Overlapping Edgeworth Boxes for endowment vectors E = (8,3) and E = (3,8).

cj = ψj (p) + uj ψc (p) j

(12.2.1)

where u is a scalar utility index for person j . The baseline functions ψj and the common function ψc are the derivatives of concave functions that are

A Digression on Gorman Aggregation

269

positively homogeneous of degree 1. Hence these functions are homogeneous of degree zero in prices, assuring that the slopes of indifference curves should depend only on the ratio of prices. The baseline indifference curves are either the highest or lowest indifference curves, corresponding respectively to cases in which the utility indices are restricted to be nonpositive or nonnegative. As noted by Gorman, when preferences are of this form, there is a well defined compensated demand function for a fictitious representative consumer obtained by aggregating (12.2.1): ca = ψa (p) + ua ψc (p) (12.2.2) where ua =

X

uj and ψa =

X

ψj .

(12.2.3)

In this case, optimal resource allocation in a heterogeneous consumer economy simplifies as follows. Preferences (12.2.2), define a community preference ordering for aggregate consumption. This preference–ordering can be combined with a specification of the technology for producing consumption goods to determine the optimal allocation of aggregate consumption. Mapping (12.2.2) can be inverted to obtain a gradient vector p that is independent of how utilities are allocated across consumers. Since ψc and ψa are homogeneous of degree zero, gradients are only determined up to a scalar multiple. Armed with p , we can then allocate utility among J consumers while respecting the adding up constraint given in (12.2.3). The allocation of aggregate consumption across goods and the associated gradient are determined independently of how the aggregate utility is divided among the individual consumers. A decentralized version of this analysis proceeds as follows. Let W j denote the wealth of consumer j and W a denote aggregate wealth. Then W j should satisfy: W j = p · cj = p · ψj (p) + uj p · ψc (p).

(12.2.4)

Solving (12.2.4) for uj gives uj = [W j − p · ψj (p)]/p · ψc (p). Hence the Engel curve for consumer j is given by

(12.2.5)

270

Gorman Heterogeneous Households

cj = ψj (p) − p · ψj (p)/p · ψc (p) + W j ψc (p)/p · ψc (p).

(12.2.6)

Notice that the coefficient on W j is the same for all j since ψc (p)/p · ψc (p) is only a function of the price vector p in a decentralized economy. The individual allocations can be determined from the Engel curves by substituting for p the gradient vector obtained from the single consumer optimal allocation problem. Individual consumption cj as given by (12.2.6) depends directly on prices through the functions ψj and ψc and indirectly through the evaluation of wealth. For the specifications of preferences adopted in this book, the baseline indifference curves are degenerate because they do not depend on p . A finitedimensional counterpart to this circumstance occurs when ψj (p) = χj ,

(12.2.7)

where χj is a vector with the same dimension as cj . With this specification, the rules for allocating consumption across individuals become linear in aggregate consumption. To see this, observe that an implication of (12.2.2) is ψc (p) = (ca − χa )/ua .

(12.2.8)

Substituting (12.2.8) into (12.2.1) gives cj − χj = (uj /ua )(ca − χa ),

(12.2.9)

so that there is a common scale factor (uj /ua ) across all goods for person j . Hence the fraction of total utility assigned to consumer j determines his fraction of the vector (ca − χa ). Here is an example. Suppose that the preferences of consumer j are represented using the utility function: U j (cj ) = −[(cj − χj )′ V (cj − χj )]1/2 .

(12.2.10)

The compensated demand schedule is then obtained by solving the first-order conditions: V (cj − χj )/U j (cj ) = µj p

(12.2.11)

An Economy with Heterogeneous Consumers

271

U j (cj ) = uj , where µj is a Lagrange multiplier. Substitute the second equation into the first and solve for cj − χj : cj − χj = uj µj V −1 p.

(12.2.12)

Substitute the right side of (12.2.12) into the utility function and solve for the multiplier µj : µj = 1/(p′ V −1 p)1/2

(12.2.13)

Hence the compensated demand function is given by cj = bj + uj V −1 p/(p′ V −1 p)1/2 .

(12.2.14)

In this example, ψj (p) = χj and ψc (p) = V −1 p/(p′ V −1 p)1/2 .

(12.2.15)

Notice that to obtain a representation of preferences linear in the utility index requires using a particular monotonic transformation of the utility function. In our example, the quadratic form on the right side of (12.2.10) is raised to the one-half power.

12.3. An Economy with Heterogeneous Consumers We now specify a multi-consumer version of our dynamic linear economy designed to satisfy counterparts to Gorman’s conditions for aggregation. There is a collection of consumers, indexed by i = 1, 2, . . . , I . Consumers differ both in their preferences and in their endowments, but not in their information. Consumer i has preferences that are ordered by X ∞ 1 − E β t (sit − bit ) · (sit − bit ) + ℓi2 | J0 t 2 t=0

(12.3.1)

272

Gorman Heterogeneous Households

where {sit } is linked to {hit } and {cit } via sit = Λ hit−1 + Π cit

(12.3.2)

hit = ∆h hit−1 + Θh cit ,

(12.3.3)

and hi−1 is given. In (12.3.1), (12.3.2), (12.3.3), the i superscript pertains to consumer i. The preference disturbance bit is determined by bit = Ubi zt

(12.3.4)

where zt continues to be governed by (3.2). The ith consumer maximizes (12.3.1) subject to (12.3.2), (12.3.3) and the budget constraint E

∞ X t=0

β t p0t · cit | J0 = E

∞ X t=0

i β t (wt0 ℓit + αt0 · dit ) | J0 + v0 · k−1 ,

(12.3.5)

i where k−1 is given. The ith consumer owns an endowment process dit , governed by the stochastic process dit = Udi zt . Each consumer observes the aggregate information Jt at time t , as well as i and hit−1 . The information set Jt continues the idiosyncratic capital stocks kt−1 t to be defined as Jt = [w , x0 ]. This specification confines heterogeneity among consumers to differences in the preference {bit } processes, represented by different selections of Ubi ; differences in the endowment {dit } processes, represented by different selections i of Udi ; differences in hi−1 ; and differences in k−1 . The matrices Λ, Π, ∆h , Θh do not depend on i. This makes everybody’s demand curve have the form of (10.15), with different µw 0 (reflecting different wealth levels) and different bt processes. Prices and the aggregate real variables can be computed by synthesizing a representative consumer and solving a version of the social planning probP lem that was described in chapter 3. Use the settings h−1 = i hi−1 , k−1 = P i P i P i i Ud . This gives us aggregate quantities and i Ub , and Ud = i k−1 Ub , = wa prices. We let µo denote the multiplier on wealth in the budget constraint of the representative (or average) household. To compute individual individual allocations requires more work, to which we now turn.

Allocations

273

12.4. Allocations A direct way to compute individual allocations would be to solve the problem each household faces in the competitive equilibrium at the competitive equilibrium prices. For a fixed Lagrange multiplier on the household’s budget constraint, the household’s problem can be expressed as an optimal linear regulator, with a state vector augmented to reflect the aggregate state variables determining the scaled Arrow-Debreu prices. It is possible to compute the allocation to a particular household by using an iterative scheme to calculate the Lagrange multiplier that assures that the household’s budget constraint is satisfied, but this is not the procedure that we recommend. Instead note that the allocation rule for labor is wa a ℓjt = (µwj (12.4.1) 0 /µ0 )ℓt . If we substitute this expression for ℓjt into versions of (10.13) and (10.14) for the j th consumer, we get the following version of the household’s budget constraint: µjw 0 E0

∞ X t=0

a β t {ρ0t · ρ0t + (wt0 /µaw 0 )ℓt } = E0

∞ X t=0

j β t {ρ0t · (bjt − sjti ) − αt0 · djt } − v0 k−1 .

jw in hand, Solve this equation for µwj 0 , using a doubling algorithm. With µ0 we can use the first-order conditions for services and the canonical service technology to solve for the equilibrium allocation to household j . For a canonical service technology, the first-order conditions for consumption services are: 0 sjt − bjt = µjw 0 ρt .

(12.4.2)

Given ρ0t , which we know from the aggregate allocation and (10.8), we can solve (12.4.2) for sjt , then plug sjt into the ‘inverse canonical representation’ to solve for cjt : cjt = −Π−1 Λhjt−1 + Π−1 sjt (12.4.3) hjt = (∆h − Θh Π−1 Λ)hjt−1 + Π−1 Θh sjt , hj−1 given.

274

Gorman Heterogeneous Households

12.4.1. Consumption sharing rules Our preference specification is an infinite-dimensional generalization of the one described in our digression on Gorman aggregation, a version in which goods are indexed by both dates and states of the world. The counterpart to the matrix V is determined by the probability distribution over states of the world conditioned on J0 and on the parameters of the household technology. The counterpart to χj is determined by the preference shock process {bjt } and the initial endowment of household capital hj−1 . The allocation rule for consumption has the form: cjt − χjt = (uj /ua )(cat − χat ), (12.4.4) where the ratio (uj /ua ) is time invariant and depends only on information available at time zero. We can express (12.4.4) as cjt = (uj /ua )cat + χ ˜jt c˜jt = χ ˜jt , where χ ˜jt ≡ χjt − (uj /ua )χat . Our goal is to show how to compute χ ˜jt and (uj /ua ). We shall show that the utility indexes can be set at the consumers’ marginal utilities of wealth µjw 0 , and that the ‘deviation’ baseline process for consumption {χ ˜jt } can be computed by initializing the inverse canonical repre˜ j and using a ‘deviation’ preference process {˜bj } as the sentation at a vector h t −1 ‘driving’ service process. In terms of ‘deviation’ processes, the allocation rule for consumption services is aw a a sjt − bjt = (µjw (12.4.5) 0 /µ0 )(st − bt ) or s˜jt = ˜bjt , aw a where y˜tj ≡ ytj − (µjw 0 /µ0 )yt . The beauty of this representation is that it does not directly involve prices. The ˜ version of (12.4.3) is

˜ j + Π−1 s˜j c˜jt = −Π−1 Λh t t−1 j j −1 ˜ ˜ ht = (∆h − Θh Π Λ)ht−1 + Π−1 Θh s˜jt ,

(12.4.6)

˜ j given. Associated with s˜j is a synthetic consumption process χ h ˜jt such t −1 that c˜jt = χ ˜tj is the optimal sharing rule. To construct χ ˜jt we simply substitute

Risk Sharing Implications

275

s˜jt = ˜bjt into the inverse canonical representation: j χ ˜jt = −Π−1 Λ˜ ηt−1 + Π−1˜bjt

j η˜tj = (∆h − Θh Π−1 Λ)˜ ηt−1 + Π−1 Θh˜bjt ˜j . η˜j = h −1

(12.4.7)

−1

j ˜ j , it follows from (12.4.6) and (12.4.7) that c˜j = Since s˜jt = ˜bjt and η˜−1 =h t −1 χ ˜jt . Equivalently, allocation rule (12.4.4) holds with {χjt } given by recursion aw (12.4.7), {χat }by its aggregate counterpart, and (uj /ua ) = (µjw 0 /µ0 ). Since the allocation rule for consumption can be expressed as aw a cjt = (µjw ˜jt , 0 /µ0 )ct + χ

(12.4.8)

we can append the recursion in (10.27) for ct and χt from the aggregate, singleconsumer economy to obtain a recursion for generating cjt .

12.5. Risk Sharing Implications Because the coefficient (uj /ua ) is invariant over time and across goods, allocation rule (12.4.4) implies a form of risk pooling in the deviation process {cjt − χjt }. Nonseparabilities (either over time or across goods) in the induced preference ordering for consumption goods appear only in the construction of the baseline process {χjt } and in calculation of the risk-sharing coefficient (uj /ua ) implied by the distribution of wealth. In the special case in which the preference shock processes {bjt } are deterministic (in the sense that they reside in the information set J0 ), individual consumption goods will be perfectly correlated with their aggregate counterparts (conditioned on J0 ). 3

3 Need to add references to literature on risk sharing: Altug-Miller, MaCurdy, Mace, Cochrane, Townsend etc.

276

Gorman Heterogeneous Households

12.6. Implementing the Allocation Rule with Limited Markets We have seen that one way to implement the allocation rule (12.4.4) is to introduce a complete set of markets in state and date contingent consumption. In some environments, a much smaller set of security markets suffices. An example occurs where a single consumption good is produced according to the linear technology: a cat + iat = γkt−1 + dat a kta = δk kt−1 + iat ,

β = 1/(γ + δk ).

(12.6.1)

Each consumer has a common household technology with a heterogeneous preference shock process {bjt } and a heterogeneous initial endowment of household capital hj−1 . The preference shock process is constrained to be in J0 . Each j consumer is endowed with a heterogeneous initial level of capital k−1 and an j endowment process {dt } for consumption. Instead of introducing a full array of contingent claims markets, there is a stock market for J securities that pay dividends {djt }. In addition, oneperiod riskless claims to consumption are traded. To devise a way to implement allocation rule (12.4.4), note that

a cjt − χjt + (uj /ua )χat + (uj /ua )kta = (uj /ua )[(δk + γ)kt−1 + dat ].

(12.6.2)

Let consumer j sell all its shares of stock j and purchase (uj /ua ) shares of all securities traded in the stock market. Once purchased at date zero, let consumer j hold onto this portfolio for all time periods. The total dividends paid in period t will be (uj /ua )dat . Suppose that the consumer purchases fraction (uj /ua ) of the capital stock each period in the one period bond market. The time t payoff a to the t − 1 purchase will be (uj /ua )(δk + γ)kt−1 and the time t purchase will a be (uj /ua )kt . Taken together, these market transactions have a time t receipt a of (uj /ua )[(δk + γ)kt−1 + dat ] and a time t payout of (uj /ua )kta for t = 1, 2, . . . . The difference between payouts and receipts in time t is not equal to cjt , but to cjt − χjt + (uj /ua )χat . This deviation induces trading in the bond market. Note that χjt − (uj /ua )χat is in the time zero information set J0 by assumption.

Implementing the Allocation Rule with Limited Markets

277

Let kˆtj denote additional purchases in the bond market by person j at time t . Construct kˆtj so that j , χjt − (uj /ua )χat + kˆtj = (δk + γ)kˆt−1

t = 1, 2, . . .

(12.6.3)

Solve this equation forward to determine an initial value kˆ0j : kˆ0j =

∞ X t=1

β t [χjt − (uj /ua )χat ].

(12.6.4)

Notice that kˆ0j is in J0 so that it is feasible to construct the sequence {kˆtj }. Modify the previous investment strategy so that the bond market purchases of person j at time t equals (uj /ua )kta + kˆtj for t = 1, 2, . . . . The time t receipts from the previous period purchases in the bond and stock markets equal j a (δk + γ)[(uj /ua )kt−1 + kˆt−1 ]. In light of (12.6.2) and (12.6.3), the difference between time t payouts and receipts is cjt for t = 1, 2, . . . . The coefficient (uj /ua ) in the allocation rules is determined so that initial period consumption cj0 can be purchased from the difference between the time zero receipts and payouts. In this implementation, all consumers hold the same stock portfolio or mutual fund, but make a sequence of person-specific trades in the market for oneperiod bonds. We have allowed for nonseparabilities over time in the induced preference ordering for consumption goods, which have important effects on bond market transactions. This construction displays a multiperiod counterpart to an aggregation result for security markets that was derived by Rubinstein (1974). In a two-period model, Rubinstein provided sufficient conditions on the preferences of consumers and asset market payoffs for the implementation of an Arrow-Debreu contingent claims economy in an environment with incomplete security markets. In Rubinstein’s implementation, all consumers hold the same portfolio of risky assets. In our construction, consumers also hold the same portfolio of risky assets, and portfolio weights do not vary over time. All of the changes over time in portfolio composition take place through transactions in the bond market.

278

Gorman Heterogeneous Households

12.7. A Computer Example The MATLAB program computes the allocation to individual i by executing the computations described above. The program heter.m requires that be run first, and that its output reside in memory. The program heter.m computes individual allocations in the form i X , cit = Sci Xt , hit = Sh t a, and so on. The matrices Sji are returned. The program also computes the matrices Sca , Sh and so on, which determine the aggregate allocations ct , ht , . . . as functions of the augmented state variable Xt : ct = Sca Xt aX , ht = Sh t

and so on. The MATLAB program can then be used to simulate the allocation to individual i and the aggregate allocation. The programs and must both be run for each individual i in a heterogeneous consumer economy. We illustrate the workings of these programs with the following pure exchange economy. There are two households, each with identical preferences −

1 2

E

∞ X t=0

β t (cit − bit )2 + ℓ2t | J0 , i = 1, 2

We specify that bit = 15 for i = 1, 2 . The aggregate preference shock is bt = We specify the following endowment processes. For consumer 1, d1t = 4 + .2 wt1 ,

P

bi = 30 . i t

where wt1 is a Gaussian white noise with variance (.2) 2 . For consumer 2, we specify d2t = 3 + d˜2t d˜2t = 1.2 d˜2t−1 − .22 d˜2t−2 + .25 wt2 where wt2 is a Gaussian white noise with variance (.25) 2 . To capture the pure exchange setup, we specify ∆k = 0, Θk = 0, ∆h = 0, Θh = 0, Λ = 0, Π = 1 . We set β = 1/1.05 . We have used and to simulate a realization of this economy. Figure 12.7.1 reports the individual allocations to consumers 12.2.1 and 12.2.2. Notice how they appear perfectly correlated. Household one is wealthier than the other and so always consumes more (notice that the mean of the first household’s endowment process is 4, while the mean of the second household’s is 3). The perfect correlation between the two consumption services reflects the sharing present in Arrow-Debreu models with time separable preferences. Figure 12.7.2.a graphs d1t − c1t while figure 12.7.2.b graphs d2t − c2t . These figures indicate the “balance of payments” between the two households.

Exercises

279

5

4.5

c1

4

3.5 c2 3

2.5

10

0

20

30

40

50

60

70

80

90

100

Figure 12.7.1: Consumption allocations of consumers one and two in pure endowment economy.

1

0.8

0.8

0.6

0.6

0.4

0.4

0.2

0.2

0

0

-0.2

-0.2

-0.4

-0.4

-0.6

-0.6 -0.8 0

-0.8

10

20

30

40

50

60

70

80

90

100

Fig. 12.7.2.a. Saving of consumer one.

12.8. Exercises

-1 0

10

20

30

40

50

60

70

80

90

100

Fig. 12.7.2.ab5. Saving of consumer two.

280

Gorman Heterogeneous Households

Exercise 1: The first part of this exercise is to be answered using “pencil and paper” as your tools. The second part is to be done on a computer using MATLAB.

12.8.1. Part one An economy consists of identical numbers of two types of infinitely lived consumers. Type i consumers all have preferences that are ordered by

(1)

∞ X − .5 E (cit − bi )2 |J0

, i = 1, 2 ,

0 < β < 1,

t=0

where cit is consumption of a single good by an consumer of type i, E is the mathematical expectation operator, and J0 is information known at time 0. In (1), bi is a parameter that determines the satiation level of consumption for consumers of type i. This is a pure endowment (or pure exchange) economy. The only sources of the single consumption good are two types of trees, the first initially being owned by the first type of consumer, the second type of tree initially being owned by the second type of consumer. Initially, there is one tree owned by each consumer in the economy. The first type of tree yields a constant “dividend” of fruit at the rate (2a)

d1t = d1

∀ t ≥ 0.

The second type of tree yields dividends at time t of (2b)

d2,t+1 = d2 + g wt+1 , d20 = d2

t ≥ 0,

where d2 is a constant and where {wt+1 }∞ t=0 is a martingale difference sequence 2 t with Ewt+1 |Jt = 1, where Jt = w ≡ (wt , wt−1 , . . . , w1 ). Notice that at time 0, d2t = d2 , where d2 is the stationary mean of {d2t }. The feasibility constraint is 2 X i=1

cit =

2 X i=1

dit .

Exercises

281

You are to consider the following decentralized version of this economy. At time 0, there are open a complete array of Arrow-Debreu contingent claims markets. We use the Harrison-Kreps commodity space and pricing system. At time 0, households of type i face the problem of maximizing (1) subject to (3)

E

∞ X t=0

β t p0t cit |J0 =

∞ X t=0

β t p0t dit |J0 .

A. Define a competitive equilibrium for this economy. B. Construct an argument to show that this is a representative consumer economy in the sense that the equilibrium price system can be determined without simultaneously determining individual allocations. C. Describe the preferences and constraints faced by the representative consumer in this economy. D. Compute the equilibrium price system {p0t }∞ t=0 .

E. At the equilibrium prices, compute the right hand side of (3), namely, (4)

a0i = E

∞ X t=0

β t p0t dit |J0 ,

for i = 1, 2. Go as far as you can in getting an analytic, closed form expression for a0i . F. Use your answer to E to establish that the value of trees of type 2 is smaller the larger is the absolute value of g . Interpret this result. G. Give an argument to establish that a0i given by (4) would be the equilibrium price of type i trees if a market in trees (or equivalently, perpetual claims to the dividends from a tree) were opened at time 0. H. Suppose that d1 = d2 < b1 = b2 . Suppose that g = 0, so that the second tree has a perfectly sure yield. Compute the equilibrium consumption allocation. I. Suppose that d1 = d2 = 8, that b1 = b2 = 15, and that g = 1. Compute the equilibrium consumption allocation (for both types of consumer). i. Verify that the equilibrium consumption allocations satisfy a “sharing rule”.

282

Gorman Heterogeneous Households

ii. Which type of consumer consumes more in equilibrium? Why? J. For this economy, compute the price at time t of a perfectly sure claim to one unit of consumption at time t + 1.

12.8.2. Part two Now you are to use the computer. Use MATLAB to do the computations. Hint: We have written some programs that should be a big help in doing this problem. The main programs are and ; does the main calculations while reads in the parameter values for the economy. A program called does the asset pricing calculations. You can edit these files and run them to answer the question. A. For the economy described in Exercise 1 of Part One, compute the equilibrium consumption allocation and simulate it. Write down the consumption allocations and the endowment realizations for t = 0, . . . , 10. B. For the economy described in Exercise 1 of Part One, change the value of g from 1 to 0. Recompute the equilibrium consumption allocation, and simulate it. Write down the consumption allocations and endowment realizations for t = 0, . . . , 10. C. How do the results of A and B conform to your answers in Part One? D. Use the program to compute the value of trees at time 0 under the specification of parameter values given in Exercise 1 of Part One (in particular, set g = 1). Do the results conform with your reasoning in Part One?

Economic integration

283

12.9. Economic integration A. The world consists of two virtually identical but separated economies, j = 1, 2. The economies never trade with one another. Within each economy there exist Arrow-Debreu markets at time 0. Economy j has the following structure:

12.9.1. Preferences: ∞ 1 X t − E0 β [(cjt − bj )2 + ℓ2t ] 2 t=0

12.9.2. Technology cjt + ijt = γkjt−1 + djt ,

γ>0

φijt = gjt , φ > 0 kjt = δk kjt−1 + ijt , 0 < δk < 1 2 gjt = ℓ2jt

kj,−1 given

12.9.3. Information zt+1 = A22 zt + C2 wt+1 djt = Udj zt z0 given. Here cjt is consumption at t , ijt is investment at t , kjt is the capital stock at 2 is the square of labor absorbed in adjusting the capital the end of period t , gjt stock, djt is an endowment shock, and bj is a fixed (across time) preference parameter; all of these objects are scalars. The vector zt is a set of information variables common to the two economies, and {wt+1 } is a martingale difference ′ = I . The parameters γ, φ, β, δk , A22 , and C2 are sequence with Ewt+1 wt+1 common across the two economies. The two economies differ in their values for bj and Udj .

284

Gorman Heterogeneous Households

a. Define a competitive equilibrium for economy j . b. Describe how to compute a competitive equilibrium for economy j using dynamic programming. c. Let p0jt be the (scaled) time zero Arrow-Debreu price for consumption in economy j . Show that there exists a representation for the equilibrium for economy j of the form o kjt A11 Ao12,j kjt−1 0 = + wt+1 zt+1 0 A22 zt C2 (1)

cjt p0jt

=

Scj Mcj

kjt−1 . zt

In particular, argue that Ao11 is the same across the two economies, but that Ao12,j , Scj , and Mcj depend on j . Hint: Two approaches to this problem will bear fruit. First, one can obtain an Euler equation for capital and solve it, as in our treatment of Hall’s model in chapter 3. Second, one can use the method of “adding speed by partitioning the state vector” described in chapter 9. B. Consider a world consisting of two economies with preferences, technology, information, and initial capital stocks identical to the previous one. Now, however, the two economies are integrated, there being world-wide time zero ArrowDebreu markets. Residents of country j own the initial capital stock kj,−1 and the endowment process {dj,t }. a. Define a competitive equilibrium for the integrated economy. b. Argue that the integrated economy is a representative consumer economy, being careful to define what you mean by a representative consumer economy. c. Describe how to compute the equilibrium of the representative consumer economy. d. Let pˆ0t be the (scaled) time zero Arrow-Debreu price for consumption in the integrated economy, and let kˆt = kˆ1t + kˆ2t , and cˆt = cˆ1t +ˆ c2t , where (ˆ·) denotes equilibrium objects for the integrated economy. Show that for the integrated economy, the equilibrium has a representation ˆ o ˆ kt A11 Ao12 kt−1 0 = + wt+1 zt+1 0 A22 zt C2

Economic integration

cˆt pˆ0t

Sc = Mc

ˆ kt−1 , zt

where Ao11 and A22 and the same objects that appear in (1). e. Show that kˆt = k1t + k2t and cˆt = c1t + c2t . f. Find formulas for cˆ1t and cˆ2t . Show that cˆ1t 6= c1t and cˆ2t 6= c2t .

g. Show that p01t + p02t = α pˆ0t for some positive α .

285

Chapter 13 Permanent Income Models This chapter describes a class of permanent income models of consumption, which stress a connection between consumption and income implied by present value budget balance, and which generate interesting predictions about the responses of various components of consumption to identifiable shocks to the information sets of economic agents. The models allow us to characterize how consumption of durables act as a form of savings and how habit persistence alters consumption-savings profiles.

13.1. Technology To focus on dynamics induced by the household technology, it serves our purposes to adopt the following technology specification: φc · ct + it = γkt−1 + et kt = kt−1 + it

(13.1.1)

where φc is a vector of positive real numbers with nc elements, et is a scalar exogenous endowment of consumption and kt−1 is a scalar capital stock. We set δk = 1, thereby ignoring depreciation in capital so that it is net investment. Introducing depreciation in capital would add nothing to our analysis because we shall eliminate any additional input requirement for making new capital productive. With no intermediate inputs required for investment, even if there were depreciation in the capital stock, a version of the first equation of (13.1.1) would apply to net investment by suitably altering the marginal product of capital parameter γ . The empirical counterpart to the scalar endowment process {et } is typically labor income (e.g., see Flavin 1981, and Deaton 1992). Labor is supplied inelastically and produces et units of output independently of the level of capital. The absence of curvature in the technology has some troublesome implications for equilibrium prices that we will discuss later. Nevertheless, this technology provides a good laboratory for studying how the household technology alters consumption-savings profiles. Moreover, this specification has played

– 287 –

288

Permanent Income Models

a prominent role in the empirical literature on the permanent income theory of consumption. To make the model behave well, we impose the restriction that (1+γ)β = 1. Relaxing this restriction to make capital more or less productive has unpleasant implications for the solution to the model. Thus, reducing the marginal product of capital will produce a solution in which the capital stock must eventually become negative. 4 Increasing the marginal product of capital typically produces a solution with asymptotic satiation in a deterministic version of the model; stochastic versions yield a marginal utility vector with mean zero in a stochastic steady state. This razor’s edge linkage between the marginal product of capital and subjective discount factor is the price we pay for eliminating the role of intermediate goods in making new capital productive. To put this technology within the general specification of Chapter 3, we include an additional equation φi it − gt = 0,

(13.1.2)

where φi is a small positive number. Strictly speaking, this introduces a form of adjustment cost by requiring a household input be used to make capital productive. This small penalty makes capital satisfy the square summability constraint. When there are multiple consumption goods, to make the resulting matrix [Φc Φg ] nonsingular, we introduce nc − 1 additional investment goods equal to the last nc − 1 entries of ct Thus, combining these constraints, when nc is one, we form φc 1 0 Φc = Φi = , Φg = , (13.1.3) 0 φi −1 γ Γ= , ∆k = 1, Θk = 1; 0 and when nc is greater than one, φ′c 1 Φc = 0 Φi = φi 0 −I 0

0 0 0 , Φg = −1 , I 0

(13.1.4)

4 A more interesting solution of the model imposes a period-by-period nonnegativity constraint on capital. Instead we limit the terminal behavior of the capital stock by imposing square summability.

Two Implications

γ Γ = 0 , ∆k = 1, Θk = ( 1 0

289

0);

Thus, to embed this model into the general setup of Chapter 3, we have nc investment goods (the original good and nc − 1 additional ones introduced for technical purposes), and one intermediate good (used to enforce square summability in the capital stock).

13.2. Two Implications We extract two sharp implications of this class of permanent income models of consumption. We obtain the first by substituting kt − kt−1 for it in (13.1.1) and solving the resulting difference equation for kt−1 : kt−1 = β

∞ X

β j (φc · ct+j − et+j ).

(13.2.1)

β j E(φc · ct+j − et+j )|Jt .

(13.2.2)

j=0

From this formula, it follows that kt−1 = β

∞ X j=0

Relation (13.2.2) will help us to find recursive solutions to the model. Formula (13.2.1) also has important implications for how consumption and (endowment) income respond to the underlying shocks, as displayed by a set of dynamic multipliers - or impulse responses – {χj } and {ǫj } for {ct } and {et }, respectively, where χj wt gives the response of ct+j to wt and ǫj wt the response of et+j to wt . Since the capital stock kt−1 cannot depend on wt it follows from (13.2.1) that the shock must be present-value neutral. In other words, the impact of wt on current and future values of et must be offset in a present-value sense by its impact on current and future values of ct : ∞ X j=0

β j (φc )′ χj =

∞ X

β j ǫj .

(13.2.3)

j=0

Equality (13.2.3) is precisely the present-value relation studied by Flavin (1981), Hamilton and Flavin (1986), Sargent (1987), Hansen, Roberds and Sargent (1991), and Gali (1991).

290

Permanent Income Models

The second implication pertains to the martingale behavior of the shadow price vector for consumption and capital. To begin, note that the forward evolution equation for the shadow price of capital is Mkt = E(βMkt+1 |Jt ) + E(βγMet+1 |Jt ).

(13.2.4)

The first-order conditions for the first component of investment imply that Met + φi Mgt = Mkt ,

(13.2.5)

where the left side captures the cost of an additional unit of investment and right side the benefit. The second term on the left side reflects the adjustment cost, and is zero in the limiting φi = 0 case. By substituting (13.2.5) into (13.2.4) and using the fact that β(1 + γ) = 1, we obtain the martingale implication for the shadow price of capital: Mkt = E(Mkt+1 |Jt ),

(13.2.6)

and likewise for the multiplier process {Met }. Finally, the shadow price of consumption is given by: Mct = (Φc )′ Mdt = φc Met ,

(13.2.7)

since the last nc − 1 components of Mdt are zero because these are the multipliers on a set of bookkeeping identities. Hence the shadow price process for consumption depends on a single scalar multiplier process {Met }, a martingale that we call the marginal utility process for income. We shall pursue the presentvalue budget balance and martingale implications further, and use them to find and represent the solution of the model.

Solution

291

13.3. Solution To solve the model, we begin by deriving allocation rules for consumption and investment which can be represented in terms of the scalar martingale process {Met }. Then we use present-value relation (13.2.2) to compute Met . Our focus on the marginal utility of income imitates an aspect of the analysis in Bewley (1977). To accomplish the first step, we use the notion of a canonical household technology. Recall that the household technology determines the sequence of consumption services associated with a given sequence of consumption goods and an initial condition for the household capital stock. When the household technology is canonical, we can construct an inverse system which maps a given sequence of consumption services and an initial condition on the household capital stock uniquely into a sequence of consumption goods required to support that service sequence. For a household technology to be canonical, there must be the same number of services as goods, the matrix Π must be nonsingular, and the absolute values of the eigenvalues of the matrix (∆h − Θh Π−1 Λ) must be strictly less than β −1/2 . Under these restrictions, the inverted system can be represented recursively as: ct = Λ∗ ht−1 + Π∗ st

(13.3.1)

ht = ∆∗h ht−1 + Θ∗h st where Λ∗ ≡ −Π−1 Λ, Π∗ ≡ Π−1 ; and ∆∗h ≡ (∆h − Θh Π−1 Λ), Θ∗h ≡ Θh Π−1 . In chapter @[email protected], we showed that there always exists a representation of the induced preferences for consumption goods in terms of a canonical technology. Governing the multipliers is an analogous dual system: Mht = E[β(∆h )′ Mht+1 |Jt ] + E[βΛ′ Mst+1 |Jt ] Mct = (Θh )′ Mht + Π′ Mst ;

(13.3.2)

292

Permanent Income Models

and an associated inverse system: Mht = E[β(∆∗h )′ Mht+1 |Jt ] − E[β(Λ∗ )′ Mct+1 |Jt ]

(13.3.3)

Mst = −(Θ∗h )′ Mht + (Π∗ )′ Mct . Since {Mct } is a martingale sequence, it follows from the inverse dual system that {Mht } and {Mst } are both martingales. In fact they are linear combinations of the scalar martingale sequence Met . For instance, Mst = Ms Met

(13.3.4)

Ms ≡ {(Π∗ )′ + (Θ∗h )′ [I − β(∆∗h )′ ]−1 βΛ∗′ }φc

(13.3.5)

where

Consequently, we can solve for the service sequence in terms of the scalar martingale {Met } from the simple link between the vector of services and the corresponding marginal utility vector: st = −Ms Met + bt .

(13.3.6)

From this relation and from the inverse household technology (13.3.1) it follows that ct = Λ∗ ht−1 − Π∗ Ms Met + Π∗ bt (13.3.7) ht = ∆∗h ht−1 − Θ∗h Ms Met + Θ∗h bt . To characterize the decision rule for investment, we solve (13.1.1) for it and substitute the right-hand side of (13.3.7) for ct : it = γkt−1 + et − φc · ct

(13.3.8)

= γkt−1 + et − (φc )′ Λ∗ ht−1 − (φc )′ Π∗ (bt − Ms Met ). So far, we have derived a recursive representation for consumption, investment and household capital in terms of the scalar multiplier process {Met }. However, we have not shown how to initialize this sequence, and we do not yet have a formula relating the time t increment of this process to the underlying martingale difference sequence {wt }. We now derive formulas for both of these objects.

Solution

293

To find an expression for the marginal utility of income process, we exploit the present-value budget balance restriction (13.2.2). In light of the inverse system (13.3.1) for the household technology, we are led to compute the expected discounted sum of services and household capital: ∞ X j=0

β j E(st+j |Jt ) = −[1/(1 − β)]Ms Met +

∞ X j=0

β j E(bt+j |Jt ),

(13.3.9)

and β

∞ X j=0

β j E(ht+j |Jt ) = β

∞ X j=0

β j E(∆∗h ht+j−1 + Θ∗h st+j |Jt ).

(13.3.10)

Rewriting (13.3.10), we see that (I − β∆∗h ) or

∞ X j=0

∞ X j=0

β j E(ht+j−1 |Jt ) = ht−1 + βΘ∗h

∞ X j=0

β j E(st+j |Jt ),

β j E(ht+j−1 |Jt ) =

(I −

β∆∗h )−1 ht−1

+ β(I −

β∆∗h )−1 Θ∗h

∞ X j=0

(13.3.11) j

β E(st+j |Jt ).

Since φc · ct = (φc )′ (Λ∗ ht−1 + Π∗ st ), it follows from equation (13.3.1) and (13.3.9) that (1 + γ)kt−1 − (φc )′ Λ∗ (I − β∆∗h )−1 ht−1 = (13.3.12) − Ms′ Ms Met /(1 − β) +

∞ X j=0

β j E(Ms′ bt+j − et+j |Jt ).

Solving for {Met } results in Met = 1/(Ms′ Ms )[(1 − β)

∞ X j=0

β j E(Ms′ bt+j − et+j |Jt ) −

(13.3.13)

γkt−1 + (1 − β)(φc )′ Λ∗ (I − β∆∗h )−1 ht−1 ]. To interpret this solution, it is useful to decompose the right-hand side of (13.3.13) into three components. First, we follow the permanent income

294

Permanent Income Models

literature by defining permanent income to be that amount of income to be spent on consumption that can be expected to persist in the future and still satisfy (13.2.2): ytp ≡ γkt−1 + (1 − β)

∞ X j=0

E(β j et+j |Jt ).

(13.3.14)

Formally, this is obtained by letting {ytp } be a hypothetical expenditure process p for consumption, assuming it is a martingale, substituting yt+j for φc · ct+j in p equation (13.2.2), and solving for yt . Note that this measure of permanent income does not adjust for risk in the endowment sequence and hence even when divided by (1 − β) is distinct from equilibrium wealth. Nevertheless, it is an important component of the solution to the model. In fact, ct = ytp is the solution for consumption in the in the special case of single good, no preference shocks ( bt constant), and time separable preferences (Λ zero), which is Hall’s (1978) permanent income model of consumption. In this simple case, the marginal utility process for endowment income is formed by translating the negative of permanent income: Met = b−ytp . More generally, when the preference shock process is not expected to be constant, the term of interest is a ‘permanent’ measure of the preference shock sequence: ∞ X bpt ≡ (1 − β) E[β j (Ms )′ bt+j |Jt ]. (13.3.15) j=0

Finally, when preferences are not separable over time, the household capital stock also enters the solution for the marginal utility of endowment income. To interpret its coefficient, consider the sequence of consumption goods required to support zero consumption services from now into the future. To compute this sequence, simply feed a sequence of zeros into the inverse of the household technology. Discounting the resulting consumption sequence and premultiplying by (φc )′ results in: yth ≡ (1 − β)(φc )′ Λ∗ (I − β∆∗h )−1 ht−1 .

(13.3.16)

Hence yth adjusts the permanent income measure to account for implicit consumption associated with a “baseline” zero service sequence. To summarize, the marginal utility of endowment income can be represented as: Met = (1 − β)(1/Ms′ Ms )(bpt − ytp + yth ). (13.3.17)

Deterministic Steady States

295

Relation (13.3.17) gives a decomposition of marginal utility of income into three components: bpt , ytp and yth . Increases in bpt result in an outward shift in preference for consumption, which increases the marginal utility of income; increases in ept correspond to an outward shift of permanent income, which reduces the marginal utility of income; and alterations in yth reflect movements in the initial household capital stock. The measurement of this third component is the discounted endowment-equivalent consumption sequence associated with a baseline (zero) sequence of services. Increasing it has an opposite impact on the marginal of income from increasing permanent income. The final task of this section is to deduce a formula for the increment of Met of the form µwt for some µ. Note that yth depends on time t − 1 information. Hence only the bpt and ytp terms enter into consideration. Let {ψj } denote the sequence of matrices of dynamic multipliers for the preference shock process {bt }. It follows from (13.3.13) that µ = [(1 − β)/(Ms′ Ms )]

∞ X j=0

β j [(Ms )′ ψj − ǫj ].

(13.3.18)

The dynamic multipliers {χj } for consumption can then be computed recursively from (13.3.7), and by construction satisfy the present-value budget balance restriction (13.2.3).

13.4. Deterministic Steady States It is useful to study consumption in a deterministic steady state, partly to verify that there exist configurations of the model for which consumption of all goods is positive in this steady-state. Otherwise, the stochastic versions of the model would likely have some perverse implications. We consider cases in which {bt } and {et } are constants set at the values b and e, respectively. For a deterministic version of the model, the marginal utility of income is constant over time. Of course, we are only interested in initial conditions such that the initial marginal utility is positive and hence the entire sequence is positive. Thus from equation (13.3.17), we are lead to require that: Me0 = (1/Ms′ Ms )(Ms′ b − e − γk−1 + y0h ) > 0.

(13.4.1)

296

Permanent Income Models

Since 1/(Ms′ Ms ) is positive by construction, we only need to be concerned with requiring that (Ms′ b − e − γk−1 + y0h ) be positive. Any changes in (b, e, h−1 , k−1 ) that alter the right side of (13.4.1) will clearly change the marginal utility of income (in all time periods). Since the preference shifter sequence is fixed at a constant level, the constant marginal utility of income sequence implies a constant service sequence given by: s = b − Ms Me0 . (13.4.2) The corresponding sequences of consumption goods and household capital need not be constant, and we now investigate the limiting behavior of these objects. Armed with the consumption service sequence, the consumption and household capital sequences can be computed from the inverse household technology. The absolute values of the eigenvalues of ∆∗h are less than β −1/2 but can be greater than or equal to one. Without further restricting the eigenvalues of ∆∗h to have absolute values that are strictly less than one, the consumption sequence may not converge to a steady state. 5 With the additional restriction that the absolute values of eigenvalues are strictly less than one, the consumption and household capital sequences will converge with limits: h∞ = (I − ∆∗h )−1 Θ∗h s c∞ = Λ∗ h∞ + Π∗ s

(13.4.3)

where variables with subscript ∞ denote limit points. By combining (13.4.2) and (13.4.3), it can be checked whether the limiting consumption vector is strictly positive. As an illustration, consider a setting with a single consumption good, a single physical capital stock, and the following household technology: ht = δh ht−1 + (1 − δh )ct st = λht−1 + ct ,

(13.4.4)

where δ is a depreciation factor between zero and one. Notice that the household capital stock is constructed to be a weighted average of current and past consumption. The inverse system is given by: ht = [δh − λ(1 − δh )]ht−1 + (1 − δh )st ct = −λht−1 + st .

(13.4.5)

5 As emphasized by Becker and Murphy (1988), configurations with explosive eigenvalues can arise in models of ‘rational addiction’.

Cointegration

297

In this simple case, the eigenvalue of ∆∗h is simply the coefficient on the lagged capital stock in the evolution equation for household capital. This coefficient has an absolute value less than one if: −1 < λ < (1 + δ)/(1 − δ).

(13.4.6)

When these inequalities are satisfied consumption and the household capital stock both converge to s/(1 + λ). Negative values of λ that violate the inequalities in (13.4.6) display a form of ‘rational addiction’ as analyzed by Becker and Murphy (1988). For instance, when λ is −1, the coefficient on lagged capital is unity, and the consumption sequence required to support a constant service sequence must grow linearly over time. Lower values of λ (i.e., negative ones with larger absolute values) imply geometric growth in consumption. Simply requiring the limiting value of consumption to be positive guarantees that consumption will be positive for initial levels of household capital close to the steady state, but it would be nice to obtain a stronger result. One strategy would require entries of the matrices of the inverse household technology all to be positive. Unfortunately, this restriction that would eliminate some important examples in which there is substitutability across goods or over time. For instance, in (13.4.4), when λ is positive as in the case of a durable good, the inverse household technology has a negative coefficient. Nevertheless, starting from an initial level of household capital below the steady state will result in a positive consumption sequence.

13.5. Cointegration A key feature of our solution is that the marginal utility of income process is a martingale, which implies via (13.3.6) that the consumption service process is nonstationary. If in addition the preference shock process {bt } is asymptotically stationary, then the service process and consumption are each cointegrated. Suppose that the preference shock process {bt } is asymptotically stationary, but unobservable to the econometrician. This implies that there are ns − 1 linear combinations of consumption services that are asymptotically stationary. To show this, take any vector ψ that is orthogonal to Ms . It follows from

298

Permanent Income Models

(13.3.6) that ψ ′ st = ψ ′ bt .

(13.5.1)

Evidently there are ns − 1 linearly independent ψ ’s. Each such ψ is referred to as a cointegrating vector, in the terminology of Granger and Engle (19??). An extensive literature treats the efficient estimation of cointegrating vectors. However, what interest us are not the cointegrating vectors but rather the vector Ms that is orthogonal to all of the cointegrating vectors for consumption services. The cointegrating vectors for consumption services differ from the cointegrating vectors for consumption goods. To deduce the cointegrating vectors for the consumption flows, we shall build upon the deterministic steady-state calculations reported in (13.4.3). From (13.4.3), we know that for a deterministic steady state c∞ = [Λ∗ (I − ∆∗ )−1 Θ∗ + Π∗ ]s∞ . (13.5.2) The matrix on the right-hand side of (13.5.2) also gives the transformation mapping a date t shock to consumption services to the limiting response of consumption. Any vector ψ satisfying ψ ′ [Λ∗ (I − ∆∗ )−1 Θ∗ + Π∗ ]Ms = 0

(13.5.3)

is a cointegrating vector for consumption. Let Ψ denote an nc − 1 by nc matrix with rows that are linearly independent cointegrating vectors. Notice that the implied model for Ψct and c1,t − c1,t−1 contains only stationary endogenous variables, so that it can be estimated using methods that require asymptotic stationarity, like the frequency-domain methods of chapter 9. An estimation strategy based on recursive formulations of the Gaussian conditional likelihood function would not require such a transformation to a stationary set of endogenous variables, but would require confronting the nonexistence of an asymptotically stationary distribution of the state vector from which to draw an initial estimator of the state. In chapter 9, we described a method based on ideas of Kohn and Ansley (1985) designed to construct an initial estimator of the state in such a circumstance. If we were to assume that the preference shock process is itself nonstationary and that there does not exist any nontrivial cointegrating vector for this process, then it would follow that there exist no cointegrating vectors for either the service process or the consumption process. In this case, to utilize estimation methods

Constant Marginal Utility of Income

299

requiring stationary, we would base parameter estimation on the implications for the differenced processes for consumption and household capital.

13.6. Constant Marginal Utility of Income In the absence of uncertainty, the marginal utility of income process will be constant. In this section we introduce uncertainty in the endowment and preference shock processes, and ask: when will the marginal utility of income process remain constant through time? Constancy of the marginal utility of income is an extreme version of the prediction of permanent income theory that the ability to transfer consumption over time results in “smoothness of consumption” over time. In the absence of preference shocks, a constant marginal utility of income process implies that consumption services will also be constant through time. We address this question from two angles. Initially, we investigate the limiting behavior as the subjective discount factor β approaches unity, and provide conditions on the stochastic structure sufficient to imply constant marginal utilities of income in the limit. In taking this limit, we will not concern ourselves with interpreting directly the limit economy. Instead we will study the limiting behavior of the solutions to the optimal resource allocation problems along with the associated marginal utility of income processes. The second attack on the question is to characterize the specifications of uncertainty that imply a constant marginal utility of income process for a given β that is strictly less than one. The initial portion of our investigation will imitate and replicate features of Bewley’s (1977) study of the permanent income model of consumption. Our analysis is mechanically simpler than Bewley’s, but different. When we drive β to one, we maintain the link between the subjective discount factor and the marginal product of capital. Hence as β tends to one in our analysis, the marginal product of capital as measured by γ is simultaneously being driven to zero. In contrast, Bewley considered setups in which the counterpart to the marginal product of capital is always zero. Since capital is less productive, nonnegativity constraints are a central feature of his analysis of the economies with β strictly less than one. Suppose that {zt } is a stationary stochastic process and that zt has a finite PN −1 second moment. Then it is known that the time series average (1/N ) j=0 zt+j

300

Permanent Income Models

converges. 6 Moreover, the limit vector is invariant to the starting date t of the average. Consistent with our setup in previous chapters, we assume that both bt and et are linear functions of zt . Recall that the portion of the solution for the marginal utility of income that is not predetermined (the portion that can respond to a current-period shock) is a linear combination, say ν , of a P∞ conditional expectation of the geometric average (1 − β) j=0 β j zt+j where ν ≡ [1/(Ms′ Ms )](Ms Ub − Ue )

(13.6.1)

bt = Ub zt and et = Ue zt [see (13.3.13)]. While the simple time-series average and the geometric average will not typically agree, they can be made arbitrarily close by driving N to infinity and β to one. Under both limits, tail terms in the average become relatively more important as the limit point is approached. Formally, it follows from the theory of Cesaro and Abel summability that lim (1/N )

N →∞

N −1 X j=0

zt+j = lim (1 − β) β→1

∞ X

β j zt+j

(13.6.2)

j=0

(e.g. see Zygmond 1979, Theorem 1.33, page 80). Therefore, as the discount factor tends to one, the right side of (13.6.2) converges to a vector independent of t . Moreover, for the information structures we impose, the limit vector must be in the initial period information set. 7 The constancy of the marginal utility of income as β goes to unity follows immediately. The argument just provided relies on stationarity but does not require linearity in the evolution equation for {zt }. In fact, stationarity often can be replaced by a weaker notion of “asymptotic stationarity” as we now illustrate using the linear specification zt+1 = A22 zt + C2 wt+1

(13.6.3)

imposed elsewhere in this book. This specification can be exploited to obtain an alternative demonstration of the constancy of the marginal utility of income. 6 The convergence is both with probability one and in mean square, where mean-square convergence is defined using the square root of the second moment as a norm. We use meansquare convergence in our subsequent analysis. 7 This follows because the limiting random variable has a finite second moment. As long as the forecast error variance is independent of calendar time and information accumulates over time, the forecast error variance must be zero.

Constant Marginal Utility of Income

301

Recall that (1 − β)

∞ X j=0

β j E(zt+j |Jt ) = (1 − β)(I − βA22 )−1 zt .

(13.6.4)

To investigate the limit as β tends to one, it is convenient to uncouple the dynamics according to the eigenvalues. Let A22 = T DT −1

(13.6.5)

be the Jordan decomposition, and suppose that D can be partitioned as: I 0 D= (13.6.6) 0 D2 where the absolute values of the diagonal entries of D2 are all strictly less than one. Using the Jordan decomposition, it follows that I 0 −1 (1 − β)(I − βA22 ) = T T −1 . (13.6.7) 0 (1 − β)(I − βD2 )−1 Taking limits, we see that −1

lim (1 − β)(I − βA22 )

β→1

Therefore, (1 − β)

P∞

j=0

=T

I 0

0 0

T −1 .

(13.6.8)

β j E(zt+j |Jt ) depends only on zt∗ ≡ ( I

0 ) T −1 zt ,

(13.6.9)

where zt∗ has law of motion ∗ zt+1 = zt∗ + C ∗ wt+1 ,

(13.6.10)

C∗ ≡ ( I

(13.6.11)

where 0 ) T −1 C2 .

Sufficient conditions for the marginal utility of income to be constant are that the Jordan decomposition of A22 satisfies (13.6.6) and C ∗ be zero. When these restrictions are satisfied, the process {zt } will be asymptotically stationary because the process {zt∗ } will be constant over time and because {( 0 I ) T −1 zt } will converge to a stationary process. This latter result follows since the diagonal

302

Permanent Income Models

entries of D2 have absolute values that are strictly less than unity. Stationarity is only implied when the {zt } is initialized appropriately. The arguments just given cannot be extended to {zt } processes with more fundamental forms of nonstationarity. For instance, time trends or unit roots in the endowment process would suffice to overturn constancy of the marginal utility of income in the limit. In the case of time trends, the averages may diverge as the limits are taken. For unit root processes (without drifts) the limits are well defined, but the uncertainty in the marginal utility of income process does not vanish. We now change experiments and hold fixed the subjective discount factor and ask if it is still possible for the marginal utility of income to be constant. The answer to this question turns out to be yes. Assume the Jordan decomposition (13.6.5) and (13.6.6) along with restriction (13.6.11), except that D2 can now have eigenvalues with absolute values that are equal or even greater than one (but less than β −1/2 ). If ν(I − βA22 )−1 = ( ν ∗

0)

(13.6.12)

for some vector ν ∗ , then the marginal utility of income will be constant over time. While this clearly imposes a restriction on the matrix A22 , it is satisfied by some stationary and nonstationary specifications of the endowment and preference shock processes.

13.7. Consumption Externalities One of the prime motivations for intertemporal complementarities put forth by Ryder and Heal (1973) is that individual consumers are concerned in part about their consumption relative to the past community average. In other words, there is an externality in consumption. This motivation is in contrast to that given by Becker and Murphy (1988) in which the complementarities are purely private. The solution described previously is applicable even if this consumption externality is present as a solution to an optimal resource allocation problem. However, the link between optimal resource allocation and competitive equilibrium may vanish when there is a consumption externality. We now investigate this issue.

Consumption Externalities

303

To capture the externality, we endow the consumer with the household technology: Ht = ∆h Ht−1 + Θh Ct (13.7.1) st = ΛHt−1 + Πct where Ht , Ht−1 and Ct are interpreted as community-wide vectors that are beyond the control of the private consumers but are equal to their lower case counterparts in equilibrium. The previous argument justifying the martingale properties of the marginal utilities of income and consumption relied only on the technology specification and still applies when the externality is present. In light of the externality interpretation, the marginal utility of services now satisfies: Mst = (Π∗ )′ Mct

= (Π∗ )′ φc Met .

(13.7.2)

Recall that Π∗ is equal to Π−1 . Although the link between the marginal utility of services and marginal utility of consumption goods is altered, the marginal utility of service process remains a martingale. The previous solution method can now be imitated by substituting the matrix (Π∗ )′ φc for Ms given by (13.3.4). When there is a single consumption good and the household technology is canonical, the two solutions coincide. This alteration can be verified by taking the previous solution for the marginal utility of income and showing that all of the equilibrium conditions and first-order conditions remain satisfied. While the marginal utility of services is altered by a constant scale factor over time, this clearly has no impact on the implied marginal rates of substitution for consumption services and therefore the original quantity allocation remains intact with the externality interpretation. When there are multiple consumption goods, the quantity allocations can be altered. Also, when the original household technology is not ‘canonical,’ the quantity allocations can be altered even when there is a single consumption good. While there generally exists a canonical household technology that implies the same induced preferences for consumption goods, the externality version of the specification can give rise to a fundamentally different canonical technology, breaking the simple link between solution to the resource allocation problem and the decentralized economy. To elaborate on this last point, suppose the original household technology is not canonical. In Chapter ? we showed how to find the corresponding canonical

304

Permanent Income Models

technology to be used in solving the optimal resource allocation problem. In the presence of consumption externalities, we can find the analog to a canonical household technology by first noting that bt − st = Bt − Πct ,

(13.7.3)

where Bt ≡ bt − ΛHt−1 . Consequently, (bt − st ) · (bt − st ) = Bt · Bt − 2(Bt )′ Πct + (ct )′ Π′ Πct .

(13.7.4)

Suppose that Π′ Π is nonsingular, and obtain a factorization: ˆ ′Π ˆ = Π′ Π Π

(13.7.5)

ˆ is nonsingular. Also, define where Π ˆ ≡Π ˆ ′−1 Π′ Λ Λ ˆbt ≡ Π ˆ ′−1 Π′ bt

(13.7.6)

ˆ t−1 + Πc ˆ t. sˆt ≡ ΛH Then (bt − st ) · (bt − st ) and (ˆbt − sˆt ) · (ˆbt − sˆt ) agree except for a term that is not controllable by the individual consumer. Consequently, technology (13.7.6) and the implied preferences for the original household technology are the same. For this solution method to apply, we need this transformed version of ˆ is nonsingular the household technology to be canonical. Since the matrix Π ˆ to have eigenvalues ˆ −1 Λ by construction, it suffices for the matrix ∆h − Θh Π −1/2 with absolute values that are strictly less than β . If this restriction is not satisfied then there fails to exist a competitive equilibrium.

Tax Smoothing Models

305

13.8. Tax Smoothing Models By reinterpreting variables, our model can represent a class of linear models of optimal taxation, versions of which were studied by Barro (1979) and Judd (1990). Let τt be a vector of taxes collected from various sources (e.g., capital, labor, imports, etc.); Gt a scalar stochastic process of government expenditures; Bt−1 the stock of risk-free one-period government debt bearing net one-period interest rate of γ = β1 − 1; and def t the gross-of-interest 0 deficit. Match up variables as follows: ct ∼ τt , et ∼ Gt , kt−1 ∼ Bt−1 , it ∼ def t . Set φc to a vector of ones, so that φ′c τt measures total time t government tax revenues. With these associations, (13.1.1) become def t = γBt−1 + Gt − φ′c τt Bt = Bt−1 + def t .

The preference ordering is interpreted as minus the loss function associated with taxation, and measures the dynamics of tax distortions. Consider three special cases of this model, each of which sets bt to a vector of zeroes. 1. Random walk taxes. To capture Robert Barro’s specification, set φc = 1 (so there is only one kind of tax revenues), Λ = 0, Π = 1, Θh = ∆h = 0. This version makes taxes follow a random walk. It is a relabelling of Hall’s model of consumption. 2. White noise taxes. To capture a one-tax version of Judd’s specification, again set φc = 1, but now set Π = ∆h = 1, Θh = Λ = −1. With these P∞ Pt settings, the government’s objective function is −.5E0 t=0 β t ( j=0 τt−j )2 . This specification is intended to capture the long-lived adverse effects of taxation on capital. The optimal policy makes taxes a white noise process, a feature that characterizes the asymptotic behavior of capital taxation in the model of Chari, Christiano, and Kehoe (1994). To deduce the white noise property for this model, use (13.3.7) and the relations defining Λ∗ , Π∗ , ∆∗h , Θ∗h under (13.3.1). In particular, we obtain τt = −Ms (Met − Met−1 ). 3. Two taxes. Set φc = [ 1 1 ], and specify two taxes whose ‘distortion technology’ is obtained by stacking the two technologies described in examples 1 and 2. This is the kind of setup advocated by Judd (1990), and makes one tax a random walk, the other a white noise.

Chapter 14 Non-Gorman Heterogeneity Among Households

14.1. Introduction The previous chapter studied a setting in which households have heterogeneous endowments and preference shock, but otherwise have identical preferences and household technologies, implying that all Engle curves are linear with the same slopes. The property of identically sloped linear Engle curves delivers a tidy and tractable theory of aggregation assuring the existence of a representative household. This theory applies when different households share the same household technology (Λ, Π, ∆h , Θh ). In this chapter, we maintain the linearity of households’ Engle curves, but permit their slopes to vary across classes of households. In particular, we now allow households’ technology parameters (Λi , Πi , ∆hi , Θhi ) to differ across classes of households indexed by i. This alteration causes the existence of a representative household, in the sense of there being a preference ordering over stochastic processes for aggregate consumption that is independent of the initial wealth distribution, to vanish. Nevertheless, the structure still fits within a class that readily yields to linear quadratic dynamic programming algorithms. Competitive equilibria can be calculated using an algorithm based on Negishi’s idea of finding a fixed point within a class of Pareto problems, where the fixed point is a list of Pareto weights that deliver budget balance at candidate equilibrium prices. In this chapter, we describe how the algorithms can be applied very efficiently within our class of economies. We also describe how a more limited form of aggregation than Gorman’s can be carried out for this economy. In particular, implementation of the Negishi algorithm enables us to uncover a ‘mongrel’ preference ordering over aggregate consumption streams, where the preference ordering depends on the initial distribution of wealth, as do the parameters of any ‘household technology’ for representing those preferences.

– 307 –

308

Non-Gorman Heterogeneity Among Households

14.2. Households’ Preferences There are equal numbers of two types of households, indexed by i = 1, 2. Households of type i have preferences ordered by ∞ 1 X t − E β [(sit − bit ) · (sit − bit ) + ℓ2it ] | J0 . 2 t=0

(14.2.1)

Here sit is a consumption service vector for consumer i, bit is a preference shock process, and ℓit is labor supplied by consumer i. Services sit are produced via the technology sit = Λi hit−1 + Πi cit

(14.2.2)

hit = ∆hi hit−1 + Θhi cit , i = 1, 2

(14.2.3)

Here hit is consumer i’s stock of household durables at the end of period t , and cit is consumer i’s rate of consumption. The preference shock process bit is governed by bit = Ubi zt (14.2.4) where zt continues to be governed by zt+1 = A22 zt + C2 wt+1 Notice that this specification permits each class of households to have its own list of matrices (Λi , Πi , ∆hi , Θhi ) that determine a technology for converting consumption goods into services.

A Pareto Problem

309

14.2.1. Technology Consumption goods (c1t , c2t ) are produced via the technology Φc (c1t + c2t ) + Φg gt + Φi it = Γkt−1 + d1t + d2t

(14.2.5)

kt = ∆k kt−1 + Θk it

(14.2.6)

gt · gt = ℓ2t ,

ℓt = ℓ1t + ℓ2t .

(14.2.7)

As before, gt denotes the quantity of labor-using intermediate production activities; dit is the amount of the endowment vector of household i used in the production process. We assume that dit = Udi zt

, i = 1, 2.

(14.2.8)

14.3. A Pareto Problem The social welfare function is a weighted average of the utilities of the two households, with weight on household 1’s utility being λ, 0 < λ < 1. For fixed λ , we want to find an allocation that maximizes ∞ X 1 − λE0 β t [(s1t − b1t ) · (s1t − b1t ) + ℓ21t ] 2 t=0

∞ X 1 − (1 − λ)E0 β t [(s2t − b2t ) · (s2t − b2t ) + ℓ22t ] 2 t=0

subject to the constraints that describe the household and production technologies. By way of fitting it into an optimal linear regulator, it is convenient to note a property of the solution to the problem that permits us to avoid carrying along ℓ1t and ℓ2t as variables, and to replace them by functions of ℓt . The solution of the social planning problem implies a pair of simple ‘sharing rules’ for labor. We deduce these sharing rules before solving the full problem in order to economize the number of control variables.

310

Non-Gorman Heterogeneity Among Households

Let Mtℓ be the stochastic Lagrange multiplier associated with the constraint ℓ1t + ℓ2t = ℓt . With respect to ℓ1t and ℓ2t , the first order conditions are Mtℓ = λℓ1t and Mtℓ = (1 − λ)ℓ2t . These conditions imply that ℓt = ℓ1t + ℓ2t = Mtℓ /(λ(1 − λ)), or Mtℓ = λ(1 − λ)ℓt . Substituting this last equality for Mtℓ into the marginal conditions for ℓ1t and ℓ2t gives the ‘sharing rules’ ℓ1t = (1 − λ)ℓt , ℓ2t = λℓt . Use these two equations to represent the terms in ℓ1t and ℓ2t in the social planning criterion as λℓ21t + (1 − λ)ℓ22t = λ(1 − λ)ℓ2t . Substituting in the constraint gt · gt = ℓ2t , we can represent the social planning criterion as ∞ X 1 − E0 β t [λ(s1t − b1t ) · (s1t − b1t ) + (1 − λ)(s2t − b2t ) · (s2t − b2t ) 2 (14.3.1) t=0

+ λ(1 − λ)gt · gt ].

The objective function (14.3.1) is to be maximized subject to the following constraints: (12.2)

sit = Λi hit−1 + Πi cit

, i = 1, 2

(12.3)

hit = ∆hi hit−1 + Θhi cit

, i = 1, 2

(12.5)

Φc (c1t + c2t ) + Φg gt + Φi it = Γkt−1 + d1t + d2t

(12.6)

kt = ∆k kt−1 + Θk it

(12.4)

dit = Udi zt ,

bit = Ubi zt ,

zt+1 = A22 zt + C2 wt+1

i = 1, 2

A Pareto Problem

311

This problem can be set up as an optimal linear regulator problem by following steps paralleling those for the single-household economy described in Chapter 4. Define the state and controls of the system as h1t−1 h2t−1 xt = kt−1 zt

,

ut =

it c1t

.

Notice that from (14.2.5), (c2t , gt ) can be expressed as functions of the state and controls at t : n o c2t = [Φc Φg ]−1 Γ kt−1 + (Ud1 + Ud2 ) zt − Φc c1t − Φi it . (14.3.2) gt Substitution from the above equation into (14.2.3) for i = 2 shows that the law of motion for xt+1 can be represented

h1t h2t = k t zt+1

∆h1 0 0 0

0 ∆h2 0 0

0 Θh2 Uc [Φc Φg ]−1 Γ ∆k 0

0 −Θh2 Uc [Φc Φg ]−1 Φi + Θk 0

or

0 0 + 0 C2

Θh1 −Θh2 0 0

0 Θh2 Uc [Φc Φg ]−1 (Ud1 + Ud2 ) 0 A22

it c1t

h1t−1 h2t−1 k t−1 zt

wt+1

xt+1 = Axt + But + C wt+1 .

(14.3.3)

Here Uc is a matrix that selects the first nc rows of the right hand side of (14.3.2), the rows corresponding to c2t , and Ug is a matrix that selects the rows of (14.3.2) corresponding to g1t . Here and below, we use the equalities

312

Non-Gorman Heterogeneity Among Households

Uc [Φc Φg ]−1 Φc = I, Ug [Φc Φg ]−1 Φg = I, Uc [Φc Φg ]−1 Φg = 0, Ug [Φc Φg ]−1 Φc = 0. Now substitute from (14.3.2) into (14.2.2) for i = 2 to get n o s2t = Λ2 h2t−1 + Π2 Uc [Φc Φg ]−1 Γkt−1 + (Ud1 + Ud2 )zt − Φc c1t − Φi it . Use this equation and (14.2.2) for i = 1 to deduce ′ Λ1 0 0 (s1t − b1t ) = −Ub1 0 Π1

h1t−1 h2t−1 k t−1 zt it

c1t

′ 0 Λ2 −1 Π2 Uc [Φc Φg ] Γ (s2t − b2t ) = Π2 Uc [Φc Φg ]−1 (Ud1 + Ud2 ) − Ub2 −Π2 Uc [Φc Φg ]−1 Φi

−Π2

or

(s1t − b1t ) = H1 (s2t − b2t ) = H2

c1t

(14.3.4)

(14.3.5)

xt ut

xt ut

h1t−1 h2t−1 k t−1 zt it

.

Similarly, we have ′ 0 0 Γ −1 gt = Ug [Φc Φg ] (Ud1 + Ud2 ) −Φi 0

or

gt = G

xt ut

.

h1t−1 h2t−1 k t−1 zt it

c1t

(14.3.6)

A Pareto Problem

313

Now notice that the current return function in (14.3.1) can be represented as

λ(s1t − b1t ) · (s1t − b1t )

+ (1 − λ)(s2t − b2t ) · (s2t − b2t ) + λ(1 − λ)gt · gt ′ xt xt = S ut ut

(14.3.7)

where S = λH1′ H1 + (1 − λ)H2′ H2 + λ(1 − λ)G′ G.

(14.3.8)

The analysis on pages 79–80 now applies to the present system. In particular, let x′t Sxt = x′t Rxt + u′t Qut + 2x′t W ut , and write the law of motion in the form (14.3.3). This makes the Pareto problem with weight λ into a discounted optimal linear regulator problem. The solution of the Pareto problem is a law of motion xt+1 = A0 (λ)xt + CWt+1

(14.3.9)

and a list of matrices Sj (λ) such that optimal allocations are given by cit = Sci (λ)xt

, i = 1, 2

it = Si (λ)xt hit = Shi (λ)xt

, i = 1, 2

sit = Ssi (λ)xt

, i = 1, 2

(14.3.10)

The value function for the Pareto problem has the form V (xt ) = x′t V1 (λ)xt + V2 (λ).

(14.3.11)

Associated with the solution of the Pareto problem for a given λ are a set of Lagrange multiplier processes given by

314

Non-Gorman Heterogeneity Among Households

Mth1 (λ) = 2β[I 0 0 0] V1 (λ)A0 (λ)xt Mth2 (λ) = 2β[0 I 0 0] V1 (λ)A0 (λ)xt Mtk (λ) = 2β[0 0 I 0]V1 (λ)A0 (λ)xt Mts1 (λ) = λ(Sb1 − Ss1 (λ))xt

≡ Mk (λ)xt

Mts2 (λ) = (1 − λ)(Sb2 − Ss2 (λ))xt

Mtc1 (λ) = Θ′h1 Mth1 (λ) + Π′1 Mts1 (λ)

(14.3.12)

Mtc2 (λ) = Θ′h2 M h2 (λ) + Π′2 Mts2 (λ) Mti (λ) = Mi (λ)xt , Mi (λ) = Θ′t Mk (λ) ′ −1 ′ Φc Θhi Mthi (λ) + Π′i Mtsi (λ) d Mt (λ) = , Φ′g −λ(1 − λ)gt From the structure of the Pareto problem and the fact that c1t , c2t appear additively in the technology (14.2.5), it follows that Mtc1 (λ) = Mtc2 (λ).

14.4. Competitive Equilibrium We will use the following Definition: A price system is a list of stochastic processes [{p0it , wt0 ; qt0 , rt0 , αt0 }∞ t=0 , v0 ], 2 each element of which belongs to L0 . Definition: An allocation is a list of stochastic processes {cit , sit , hit , ℓit , i = 2 1, 2; kt }∞ t=0 each element of which is in L0 .

Competitive Equilibrium

315

14.4.1. Households Households of type i face the problem of maximizing −

(12.1)

∞ 1 X t E β (sit − bit ) · (sit − bit ) + ℓ2it |J0 2 t=0

subject to the budget constraint E

∞ X t=0

β t p0t · cit | J0 = E

∞ X t=0

i β t [wt0 ℓit + αt0 · dit ] | J0 + v0 k−1 ,

(14.4.1)

the household technology sit = Λi hit−1 + Πi cit

(12.2)

hit = ∆h hit−1 + Θhi cit ,

(12.3)

and the initial conditions hi,−1 , ki,−1 .

14.4.2. Firms of type I and II Firms of type I and II face the problems described in chapter 6, with ct = c1t +c2t and dt = d1t + d2t .

14.4.3. Definition of competitive equilibrium We use the following standard definition: Definition: A competitive equilibrium is an allocation and a price system such that, given the price system, the allocation solves the optimum problem of households of each type and firms of each type.

316

Non-Gorman Heterogeneity Among Households

14.5. Computation of Equilibrium To compute an equilibrium, we use an iterative algorithm based on ideas of Negishi (1960). For each given value of the Pareto weight λ , we know that there exists a competitive equilibrium, though it will typically be associated with some distribution of wealth other than the one associated with the allocation of ownership of capital and endowment processes that we have assigned. An algorithm for computing an equilibrium with a pre-assigned distribution of ownership is to search for a λ ∈ (0, 1) that delivers budget balance for each household.

14.5.1. Candidate equilibrium prices For a given λ , candidate equilibrium prices can be computed from the Lagrange multipliers associated with the solution of the Pareto problem for that value of λ . By pursuing arguments paralleling those of Chapter 6, we find p0t = Mtc1 (λ)/µw 0 rt0 = Γ′ Mtd (λ)/µw 0 qt0 = Θ′k Mtk (λ) αt0 = Mtd (λ)/µw 0

(14.5.1)

w k ′ v0 = Γ′ M0d (λ)/µw 0 + ∆k M0 (λ)/µ0

wt0 = λ(1 − λ) | Sg (λ)xt | /µw 0 These prices and the associated allocations are the ingredients used in the Negishi algorithm.

Computation of Equilibrium

317

14.5.2. A Negishi algorithm The algorithm consists of the following steps. 1. For a given λ ∈ (0, 1), solve the Pareto problem. Compute the Lagrange multipliers from (14.3.12) and use them to compute the candidate competitive equilibrium prices and quantities via (14.5.1). 2. At the candidate equilibrium prices and quantities, compute the left and right side of each household’s budget constraint (14.4.1). In particular, compute Gi = E

∞ X t=0

β t [wt0 ℓit + αt0 · dit ] | J0 + v0 · ki,−1 − E

∞ X t=0

β t p0t · cit | J0

Use the method described in chapter 10 to compute this. For our twohousehold economy, G1 and G2 will either be of opposite signs, or both will equal zero. 3. If G1 > 0, increase λ and return to step 1. If G1 = G2 = 0, terminate the search and accept the allocation and price system associated with the current value of λ as equilibrium objects. 1 In practice, one can improve on this algorithm by using any of a number of root finders to find the zero of the function Gi (λ) defined in step 2. We have found it efficient to use a ‘secant method.’

1 The Negishi algorithm is implemented in the MATLAB program solvehet.m. Some trial inputs are contained in the file clex11h.m, which inputs a two agent version of the economy in clex11.m, the one good stochastic growth model. As a benchmark, clex11h.m has the economy start out with identical endowments for the two households, and has them share identical household technologies. With these inputs, solvehet.m should find Pareto weight (which the program calls ‘alpha’) equal to .5, and should recover the same solution that solvea.m does with inputs clex11.m. Modify the inputs to get a non-Gorman aggregatable example. The program simulhet.m simulates the equilibrium computed by solvehet.m.

318

Non-Gorman Heterogeneity Among Households

14.6. Mongrel Aggregation Except in the special case that Λ1 = Λ2 , Π1 = Π2 , ∆h1 = ∆h2 , Θh1 = Θh2 , the specification of household technologies (14.2.2) – (14.2.3) violates the Gorman conditions for aggregation, so that there does not exist a representative household in the sense described in Chapter 7. However, for each Pareto weight λ , there does exist a representative household in the sense of a mongrel preference ordering over total consumption (c1t + c2t ). This mongrel preference ordering depends on the distribution of wealth, i.e., the value of initial endowments and capital stocks evaluated at equilibrium prices.

14.6.1. Static demand Mongrel aggregation of preferences is easiest to analyze in the special case that the demand curve is ‘static’ in the sense that time t demand is a function only of the current price p0t . Let the household technology be determined by a nonsingular square matrix Π , where each of Λ, ∆h , Θh are matrices of zeros of the appropriate dimensions. For this specification, the (canonical) demand curve is ct = Π−1 bt − µ0 Π−1 Π−1′ pt , (14.6.1) where µ0 is the Lagrange multiplier on the household’s budget constraint. The inverse demand curve is −1 ′ ′ pt = µ−1 0 Π bt − µ0 Π Πct .

(14.6.2)

In equations (14.6.1) and (14.6.2), the price vector pt can be interpreted as the marginal utility vector of the consumption vector pt . Integrating the marginal utility vector shows that preferences can be taken to be (−2µ0 )−1 (Πct − bt ) · (Πct − bt ).

(14.6.3)

From (14.6.2) or (14.6.3) it is evident that the preference ordering is determined only up to multiplication of Π, bt by a common scalar. We are free to normalize preferences by setting µ0 = 1. Now suppose that we have two consumers, i=1,2, with demand curves −1 −1′ cit = Π−1 i bit − µ0i Πi Πi pt .

Mongrel Aggregation

319

Adding these gives the total demand −1 −1 −1′ c1t + c2t = (Π−1 + µ02 Π2 Π−1′ 1 b1t + Π2 b2t ) − (µ01 Π1 Π1 2 )pt .

(14.6.4)

Setting c1t + c2t = ct and solving (14.6.4) for pt gives −1′ −1′ −1 −1 pt = (µ01 Π−1 + µ02 Π−1 (Π−1 1 Π1 2 Π2 ) 1 b1t + Π2 b2t ) −1′ −1′ −1 − (µ01 Π−1 + µ02 Π−1 ct . 1 Π1 2 Π2 )

(14.6.5)

We want to interpret (14.6.5) as an aggregate preference ordering associated with an aggregate demand curve of the form (14.6.2). To do this, we shall evidently have to choose the Π associated with the aggregate ordering to satisfy −1 −1′ −1′ −1 ′ µ−1 + µ02 Π−1 . 0 Π Π = (µ01 Π1 Π2 2 Π2 )

(14.6.6)

To find a matrix Π determining an aggregate preference ordering, we have to form and then factor the matrix on the right side of (14.6.6). This matrix looks like the inverse of a weighted sum of two moment matrices. Even after normalizing Π by setting µ0 = 1, a solution Π will in general depend on the ratio µ01 /µ02 , which functions like a Pareto weight on the two types of consumers. There is a special case for which the aggregate or mongrel preference matrix Π is independent of µ01 /µ02 , namely: Π1 = kΠ2 for scalar k > 0

(14.6.7)

Notice that when Π1 and Π are scalars, condition (14.6.7) is automatically satisfied. So for the one consumption good case with this special specification (i.e., with Λ being zero), Gorman aggregation obtains. In the more general case that demand curves are dynamic (quantities demanded at t depending on future prices), attaining a mongrel preference ordering becomes more difficult. In place of the problem of factoring a moment matrix as required in (14.6.6), we have to factor something that resembles a spectral density matrix, frequency by frequency. For studying mongrel preference orderings in the general case, it is convenient to work with a frequency domain representation of preferences.

320

Non-Gorman Heterogeneity Among Households

14.6.2. Frequency domain representation of preferences From chapter 9, recall the decomposition of services st = smt + sit , where smt are services resulting from market purchases of consumption and sit are services flowing from the initial household capital stock. Let (∆h , Θh , Λ, Π) correspond to a canonical household service technology, and recall that smt = σ(L)ct where σ(L) = [Π + ΛL[I − ∆h L]−1 Θh ]

σ(L)−1 = Π−1 − Π−1 Λ[I − (∆h − Θh Π−1 Λ)L]−1 Θh Π−1 L, and sit = Λ∆th h−1 . We use the transform methods described in the appendix to chapter 9. For P∞ t ′ any matrix sequence {yt } satisfying t=0 β yt yt < +∞ , define T (y) (ζ) = P∞ t/2 t .5 yt ζ . We define S(ζ) = σ(β ζ). Evidently, the transforms obey t=0 β T (sm ) (ζ) = S(ζ)T (c)(ζ)

T (sit )(ζ) = Λ[I − β 1/2 ∆h ζ]−1 h−1 . As in chapter 9, express the one-period return as (st − bt ) · (st − bt ) = smt · smt + 2smt · sit − 2smt · bt + (bt − sit ) · (bt − sit )

(14.6.8)

The term (bt − sit ) · (bt − sit ) is beyond control and therefore influences no decisions. So it can be ignored in describing a preference ordering. In terms of Fourier transforms, we have ∞ X

1 = 2π

Z

π

1 2π

Z

π

1 β smt · bt = 2π t=0

Z

π

t=0

t

β smt · smt

∞ X t=0

∞ X

β t smt · sit = t

T (c)′ S ′ ST (c) dθ

(14.6.9)

−π

T (c)′ S ′ T (si ) dθ

(14.6.10)

T (c)′ S ′ T (b) dθ,

(14.6.11)

π

−π

A Programming Problem for Mongrel Aggregation

321

where it is understood that S = S(ζ), T (c) = T (c)(ζ), T (b) = T (b)(ζ), and ζ = e−iθ . Here ( ′ ) denotes transposition and complex conjugation.

14.7. A Programming Problem for Mongrel Aggregation To find a preference ordering over aggregate consumption, we can pose a nonstochastic optimization problem. 2 We rely on a certainty equivalence result to assert that the preference ordering over random consumption streams is given by the conditional expectation of the optimized value of this nonstochastic problem. Thus, our strategy in deducing the mongrel preference ordering over ct = c1t +c2t is to solve the programming problem: maximize over {c1t , c2t } the criterion ∞ X t=0

β t [λ(s1t − b1t ) · (s1t − b1t ) + (1 − λ)(s2t − b2t ) · (s2t − b2t )]

(14.7.1)

subject to hjt = ∆hj hjt−1 + Θhj cjt sjt = ∆j hjt−1 + Πj cjt ,

j = 1, 2 j = 1, 2

c1t + c2t = ct subject to (h1,−1 , h2,−1 ) given, and {b1t }, {b2t }, {ct } being known and fixed sequences. Substituting the {c1t , c2t } sequences that solve this problem as functions of {b1t , b2t , ct } into the objective (14.7.1) will determine the mongrel preference ordering over {ct }. In solving this problem, it is convenient to proceed by using Fourier transforms. Using versions of (14.6.8), (14.6.9), (14.6.10), and (14.6.11) for households 1 and 2, in terms of transforms we can represent the Pareto-weighted average of discounted utility from consumption as

2 These calculations are done by the MATLAB program .

322

Non-Gorman Heterogeneity Among Households

−

∞ X t=0

β t [λ(s1t − b1t ) · (s1t − b1t ) + (1 − λ)(s2t − b2t ) · (s2t − b2t )] =−

1 2π

Z

π

−π

n [λT (c1 )′ S1′ S1 T (c1 ) + (1 − λ)T (c2 )′ S2′ S2 T (c2 )]

+ 2[λT (c1 )′ S1′ T (s1i ) + (1 − λ)T (c2 )′ S2′ T (s2i )] o − 2[λT (c1 )′ S1′ T (b1 ) + (1 − λ)T (c2 )′ S2′ T (b2 )] dθ

(14.7.2)

+ terms not involving T (c1 ) or T (c2 ). where it is understood that each transform on the right side is to be evaluated at ζ = e−iθ . We want to maximize the right side of (14.7.2) over choice of {c1t , c2t }∞ t=0 or equivalently T (c1 ), T (c2 ), subject to the constraint c1t +c2t = ct , or equivalently the restriction T (c1 ) + T (c2 ) = T (c). To do this optimization, we form the Lagrangian J =− +

1 2π

Z

π

[λT (c1 )′ S1′ S1 T (c1 ) + (1 − λ)T (c2 )′ S2′ S2 T (c2 )]

−π 2[λT (c1 )′ S1′ T (s1i )

+ (1 − λ)T (c2 )′ S2′ T (s2i )]

− 2[λT (c1 )′ S1′ T (b1 ) + (1 − λ)T (c2 )′ S2′ T (b2 )] + µ[T (c) − T (c1 ) − T (c2 )] dθ

(14.7.3)

where it is understood that there is a Lagrange multiplier µ = µ(e−iθ ) for each frequency θ ∈ [−π, π]. We can perform this maximization “frequency by frequency” (i.e., pointwise for each θ ∈ [−π, π]). The first-order necessary conditions with respect to T (c1 ) and T (c2 ), respectively, are λS1′ S1 T (c1 ) + λS1′ T (s1i ) − λS1′ T (b1 ) = µ/2

(1 − λ)S2′ S2 T (c2 ) + (1 − λ)S2′ T (s2i ) − (1 − λ)S2′ T (b2 ) = µ/2

Using these two equations and the constraint T (c1 ) + T (c2 ) = T (c) to solve for µ gives h1 i−1 1 µ=2 (S1′ S1 )−1 + (S2′ S2 )−1 λ 1−λ (14.7.4) n o × T (c) + S1−1 T (s1i ) − T (b1 ) + S2−1 T (s2i ) − T (b2 ) .

A Programming Problem for Mongrel Aggregation

Let S′S =

h1

λ

(S1′ S1 )−1 +

i−1 1 (S2′ S2 )−1 1−λ

323

(14.7.5)

where S is a matrix Fourier transform that satisfies the condition: det S(ζo ) = 0 implies |ζo | > 1. Equation (14.7.5) is the dynamic counterpart of (14.6.6), and collapses to (14.6.6) in the special case in which Sj (L) = Πj . Recall that in the static case, constructing the aggregate preference ordering required factoring −1′ the inverse of a ‘moment’ matrix formed as a weighted sum of Π−1 . j Πj 1 The matrix [ λ1 (S1′ S1 )−1 + 1−λ (S2′ S2 )−1 ]−1 can be regarded as a spectral density matrix. Equation (14.7.5) expresses S ′ S as a spectral factorization 1 of [ λ1 (S1′ S1 )−1 + 1−λ (S2′ S2 )−1 ]−1 . Later, we shall show how to achieve the factorization expressed in (14.7.5). For now, we just assume that we have an S that satisfies (14.7.5) and the condition that the zeros of det S(ζ) all exceed unity in modulus. Using (14.7.5) in (14.7.4) gives µ = 2S ′ S T (c)+S1−1 T (s1i ) − T (b1 ) + S2−1 T (s2i ) − T (b2 ) .

The fact that the Lagrange multiplier is the derivative of the return function with respect to T (c) implies that the mongrel return function has the representation

1 2π

′ T (c) S′S S ′ SS1−1 S ′ SS2−1 T (s1i ) − T (b1 ) S1−1′ S ′ S − − −π −1′ ′ T (s2i ) − T (b2 ) S2 S S − − T (c) T (s1i ) − T (b1 ) dθ T (s2i − T (b2 )

Z

π

where the blank terms do not involve T (c), and do not affect the choice of T (c). Therefore, we can represent the mongrel preference ordering over T (c) by Z π 1 T (c)S ′ ST (c) + (2T (c)′ S ′ )SS1−1 (T (s1i ) − T (b1 )) 2π −π (14.7.6) + (2T (c)S ′ )SS2−1 (T (s2i ) − T (b2 )) dθ.

324

Non-Gorman Heterogeneity Among Households

Compare this with the single agent case of chapter 9, in which the preference ordering was shown to be Z πn o 1 ¯ (¯ T (¯ c)′ S¯′ ST c) + 2T (¯ c)′ S¯′ (T (¯ si ) − T (¯b)) dθ, (14.7.7) 2π −π where we have put bars ( ¯ ) over the objects in (14.7.7) to represent the corresponding single agent-objects. Evidently, for the mongrel preference ordering (14.7.6) to match up with a single agent ordering (14.7.7), we can match objects up as T (c) ∼ T (¯ c) S ∼ S¯ SS1−1 (T (s1i ) − T (b1 ))+

SS2−1 (T (s2i ) − T (b2 )) ∼ T (¯ si ) − T (¯b)

(14.7.8)

where the object on the right of the right of the ∼ corresponds in each case to the single-agent 4. We have two major tasks to complete our work. First, we have to show how to achieve the factorization (14.7.5). Second, we have to show how to interpret and to implement the correspondence given in (14.7.8).

14.7.1. Factoring S ′ S To achieve the spectral factorization (14.7.5), we notice that 1 [ λ1 (S1′ S1 )−1 + 1−λ (S2′ S2 )−1 ] can be regarded as the spectral density matrix of a stochastic process cλt that is generated by the state space system −1 h1t = β .5 (∆h1 − Θh1 Π−1 1 Λ1 )h1t−1 + Θh1 Π1 s1t

−1 h2t = β .5 (∆h2 − Θh2 Π−1 2 Λ2 )h2t−1 + Θh2 Π2 s2t h1t−1 1 1 −1 −1 λ .5 Π2 Λ2 ] ct = β [− √ Π1 Λ1 , − √ h2t−1 1−λ λ s1t √1 + ( √1λ Π−1 Π−1 ) 1 , 1−λ 2 s2t

(14.7.9)

where (∆hi , Θhi , Λi , Πi ) are each associated with a canonical representation, s1t and where st = is a white noise with covariance Est s′t = I . Write this s2t

A Programming Problem for Mongrel Aggregation

system compactly as

325

˜ h ht−1 + Hst ht = ∆ cλt = Gλ ht−1 + Mλ st .

1 The spectral density of cλt can be directly computed to be [ λ1 (S1′ S1 )−1 + 1−λ (S2′ S2 )−1 ]. 3 We can factor the inverse of

1 1 [ (S1′ S1 )−1 + (S ′ S2 )−1 ] λ 1−λ 2 by obtaining an innovations representation for the system (14.7.9), then using it to form a ‘whitener’. The innovations representation is ˆt = ∆ ˆ t−1 + Kat ˜ hh h ˆ t−1 + at cλ = Gλ h

(14.7.10)

t

˜ h , Gλ , HH ′ , Mλ M ′ , where Eat a′t = Ω = Gλ ΣG′λ +Mλ Mλ′ , and [K, Σ] = kfilter (∆ λ ′ HMλ ), where kfilter is the matrix valued function defined in chapter 7 . 1 (S2′ S2 )−1 ], let r′ r = Ω be the To get the inverse of [ λ1 (S1′ S1 )−1 + 1−λ Cholesky decomposition of Ω, and define sˆt by sˆt = r′ −1 at . Then use (14.7.10) to get the ‘whitener’ ˆ t = (∆ ˆ t−1 + Kcλ ˜ h − KGλ )h h t ′−1 ′−1 λ ˆ sˆt = −r Gλ ht−1 + r c t

or

ˆt = ∆ ˆ t−1 + Θ ˆ hh ˆ h cλ h t λ ˆ ˆ ˆ sˆt = Λht−1 + Πc

(14.7.11)

t

where

ˆ h = (∆ ˜ h − KGλ ) ∆ ˆh = K Θ ˆ = −r′−1 Gλ Λ ˆ = r′−1 Π

3 To see this, we make use of the fact that −1 −1 −1 Θ Π−1 ζ. σj (ζ)−1 = Π−1 hj j j − Πj Λj [I − (∆hj − Θhj Πj Λj )ζ]

(14.7.12)

326

Non-Gorman Heterogeneity Among Households

As a consequence of the factorization identity and associated matrix identities described in chapter 7, 4 we have that S ′ S satisfies (14.7.5) where h i ˆ + Λζ[I ˆ ˆ h ζ]−1 Θ ˆh S(ζ) = Π −∆ i h = Π + Λβ .5 ζ[I − ∆h β .5 ζ]−1 Θh . It follows that a (canonical) version of the mongrel household technology is ht = ∆h ht−1 + Θh (c1t + c2t ) st = Λht−1 + Π(c1t + c2t ), where

ˆ h, ∆h = β −.5 ∆ ˆ Λ = β −.5 Λ,

ˆh Θh = Θ ˆ Π = Π.

(14.7.13a)

(14.7.13b)

Collecting results, we have that the canonical household technology is determined by the matrices ˜ h − KGλ ) ∆h = β −.5 (∆ Θh = K

Λ = −r′

Π = r′

−1

−1

Gλ β −.5

.

These equalities imply that ∆h − Θh Π−1 Λ =

∆h1 − Θh1 Π−1 1 Λ1 0

0 . ∆h2 − Θh2 Π−1 2 Λ2

It follows that for the mongrel household technology, the counterpart to (BLANK.1) is 1 1 ′ ct = [− √ Π−1 Π−1 2 Λ2 ]ht−1 + r st 1 Λ1 , − √ 1−λ λ ∆h1 − Θh1 Π−1 0 1 Λ1 ht = ht−1 + Kr′ st . Λ 0 ∆h2 − Θh2 Π−1 2 2 4 In effect, we are using the factorization identity (7.33) and the matrix inversion identity (7.36). We are expressing (S ′ S)−1 from ( 14.7.5 ) first in a form like (7.31a), then via the factorization identity in a form like (7.31b). Then we apply the inversion formula (7.36) to the two factors of (7.31b), replacing Ω with its Cholesky factorization, to construct a factored version of S ′ S .

The Mongrel Preference Shock Process

327

Notice how the weight λ influences this representation: λ appears in the ‘observer’ matrix multiplying ht−1 in the first equation, and it appears indirectly through its influence on the matrices [r′ , K]. However, the state transition matrix ∆h − Θh Π−1 Λ is independent of λ .

14.8. Summary of Findings We have found that the operator σj (L)−1 is implemented by the state space −1 −1 −1 system defined by the four matrices 5 [∆hj − Θhj Π−1 j Λj , Θhj Πj , Πj Λj , Πj ]. The operator σ(L) associated with the mongrel household technology is realized by the state space system [∆h , Θh , Λ, Π] determined by (14.7.12) and (14.7.6). We can use these state space systems to derive a state space system for the mongrel preference shock.

14.9. The Mongrel Preference Shock Process Our next goal is to construct a mongrel preference shock process that achieves the match up given in (14.7.8). Evidently, from (14.7.6), we have to operate on T (s1i ) − T (b1 ) with the “filter” SS1−1 , operate on T (s2i ) − T (b2 ) with the “filter” SS2−1 , then add the results to get a process that we can interpret as the mongrel T (si ) − T (b). The following cascading of state space systems evidently implements the required filtering and adding:

B

(

zt+1 b 1t b2t A hjt sjt

= A22 zt + C2 wt+1 = Ub1 zt = Ub2 zt = ∆hj hjt−1 = Λj hjt−1

j = 1, 2 , j = 1, 2

−1 xjt = (∆hj − Θhj Π−1 j Λj )xjt−1 + Θhj Πj (bjt − sjt ) −1 yjt = −Π−1 j Λj xjt−1 + Πj (bjt − sjt )

, j = 1, 2

5 Defined as usual as the matrices in the state equation followed by the matrices in the observation equation.

328

Non-Gorman Heterogeneity Among Households

C

(

gt = ∆h gt−1 + Θh (y1t + y2t ) (bt − sˆt ) = Λgt−1 + Π(y1t + y2t ).

System A generates the “inputs” (b1t , b2t ), (si1t , si2t ). System B operates on P −1 i (bjt − sijt ) with σj−1 . System C operates on j σj (bjt − sjt ) with σ , as required by (14.7.8). In C , we are free to set sˆt = 0, and to regard the resulting bt as our mongrel preference shock process. A recursive representation of bt − st is attained by linking the three systems in a series. 6 It is evident how to use similar methods to break out the processes bt and sˆt separately.

14.9.1. Interpretation of sˆt component The term SS1−1 T (s1i ) has the following interpretation. T (s1i ) is (the transform of) the contribution of services flowing to the household from the initial household capital stock h−1 . Then S1−1 T (s1i ) is the (transform of the) equivalent amount of consumption that it would have taken to generate those services had they been acquired through new market purchases. The term S1−1 T (s1i ) amounts to a consumption goods equivalent of the transient component of services flowing to the first household.

14.10. Choice of Initial Conditions Our calculations do not tell us how to choose the correct initial condition at time 0 for the mongrel household capital stock vector ht−1 . Here is the reason. The first-order necessary conditions leading to (14.7.4)–(14.7.5) imply the following solution for the transform T (c1 ): 1 1 ′ (S S1 )−1 (S ′ S)T (c) + (S1′ S1 )−1 (S ′ S)S2−1 (T (s2i − T (b2 )) λ 1 λ 1 + ( (S1′ S1 )−1 (S ′ S) − I)S1−1 (T (b1 ) − T (s1i )). λ (14.10.1)

T (c1 ) =

6 The MATLAB command series can be used to link the systems.

Choice of Initial Conditions

329

A similar expression holds for T (c2 ). Our calculations assure that T (c1 ) + T (c2 ) = T (c). We assume that T (c) is the transform of a sequence that is onesided (i.e., ct = 0 ∀t < 0), but this does not guarantee that T (c1 ) and T (c2 ) are each transforms of one-sided sequences, only that their sum is. When Sj′ Sj is not a constant times S ′ S for j = 1, 2, as will generally be the case when the two household technologies are not identical, then T (c1 ) given by (14.10.1) will be the transform of a sequence that is nonzero for t < 0. 7 Thus, our frequency domain programming problem allows the ‘mongrel planner’ to reallocate past consumptions between the two types of households, subject to the restriction c1t + c2t = 0 for t < 0. These choices of cjs for s < 0 translate into choices of initial conditions for h−1 , the vector of mongrel household capital stocks at date t = −1. We will not pursue calculations of the initial conditions here, because they are intricate and only effect the transient responses of services. Our main interest is not in the selection of the initial conditions but in the ‘nontransient’ part of the mapping from total consumption ct = c1t +c2t to the mongrel service vector st , which is given by (14.7.6).

7 The operator (S ′ S )−1 (S ′ S) is two-sided except when it is proportional to the identity 1 1 operator.

Part III Extensions

Chapter 15 Equilibria with Distortions

15.1. Introduction In earlier chapters, we often used the fact that the competitive equilibrium allocations solve a Pareto problem. This chapter describes classes of economies for which the connection between Pareto optimality and equilibrium breaks down, because distortions cause competitive equilibrium allocations not to be Pareto optimal. The distortions occur in the form of externalities in preferences and production technologies, and distorting taxes. These features can be accommodated using the invariant subspace methods described in chapter 8, which provide an approach to studying the existence and uniqueness of an equilibrium, and if one exists, to computing it. This chapter describes ways of adapting our earlier methods to compute, represent, and manipulate equilibria. 1 We describe two classes of economies with distortions. The first class is designed partly as a warmup for the second, but is also interesting in its own right. In this first class of models, an equilibrium can be computed by one pass through the invariant subspace algorithm. This is possible possible because of two sorts of simplifying assumptions: first, that there is a representative agent (which permits Gorman-heterogeneity);and

1 The invariant subspace computational approach follows the tradition of Blanchard and Khan (1980), Whiteman (1983), Anderson and Moore (1985), King, Plosser, and Rebelo (1989), and McGrattan (1991) in expressing the equilibrium conditions as a system of linear difference equations, then adapting the method of Vaughan to solve it. The principal caveat to keep in mind in applying these methods is that in distorted economies, there is no guarantee that the eigenvalues of the matrix ‘characteristic p equation’ of the system of equilibrium conditions will ‘split’ half into those exceeding 1/ β in modulus, and half into those falling

p

short of 1/

β in modulus. Failure of the eigenvalues to divide in that way signals either an

existence or a uniqueness problem.

– 333 –

334

Equilibria with Distortions

second, that the government has the ability to levy lump sum taxes or transfers. 2 Thus, the first setup is a representative agent setting with consumption and production externalities and distorting taxes. 3 . In the second setup, equilibria must be computed with multiple passes through the invariant subspace algorithm. What necessitates this is that there are (non-Gorman) heterogeneous agents with consumption externalities and distorting taxes, but no lump sum taxes or transfers. In this setting, equilibrium computation means finding a set of taxes and households’ marginal utilities of wealth that assure present value budget balance for households and the government. We find the equilibrium taxes and marginal utilities of wealth by repeatedly resorting to the invariant subspace algorithm, constructing candidate prices and quantities at each pass of the invariant subspace algorithm, then checking every agent’s budget constraint at each pass. It is convenient to describe these two classes of models sequentially, partly because the first one is complicated enough, and because it reveals about half of the difficulties involved in calculating models of the second class. We shall use our machinery to compute equilibria of some simple models embodying preference and technology specifications that have occurred in the recent literature.

2 The MATLAB program solvdist.m computes equilibria for the first type of economy, while disthet.m computes equilibria for the second type of economy. The MATLAB program compare.m is useful for comparing the results of disthet.m with related undistorted economies whose equilibria have been computed with solvea.m. 3 The first class of models are linear-quadratic relatives of the ones studied by Braun (1991) and McGrattan (1991b). In particular, both Braun and McGrattan use lump sum transfers to balance the government’s budget.

A Representative Agent Economy with Distortions

335

15.2. A Representative Agent Economy with Distortions We alter the model discussed in chapters 3 and 5 by adding two sorts of distortions: consumption and production externalities, and distorting taxes. We add these distortions in a way that is designed to preserve the applicability of the forms of equilibrium prices and quantities: we want the equilibrium law of motion to be linear in the state, quantities and (scaled Arrow-Debreu) prices to be linear functions of the state, and asset prices to be quadratic functions of the state. We add the following distortions to the Chapter 3–5 model.

15.2.1. a. Consumption externalities The technology for producing household services is now taken to be st = Λh ht−1 + ΛH Ht−1 + Πct + ΠC Ct ht = ∆h ht−1 + ∆H Ht−1 + Θh ct + ΘH Ct , where Ht−1 is the vector of aggregate household capital stocks, and Ct is the vector of aggregate consumption rates. In equilibrium, ct = Ct and ht = Ht , but the representative household is assumed to take {Ht , Ct } as given and beyond control when allocating its resources.

15.2.2. b. Production externalities Firms produce subject to the linear technology Φc (ct + Gt ) + Φi it + Φg gt = Γk kt−1 + ΓK Kt−1 + dt where Kt−1 is the vector of aggregate capital stocks, and Gt is government purchases of consumption goods. In equilibrium, kt−1 = Kt−1 , but the representative firm is assumed to regard {Kt−1 } as beyond its control when choosing its inputs and outputs. The presence of Kt−1 on the right side of the technology constraint is designed to accommodate externalities of the sort analyzed by Paul Romer (1985). Below, we shall assign ownership of the ‘technology’ so that households sell the joint process ΓK Kt−1 + dt to firms. This assignment will confront the firm with a constant returns to scale technology. We assume that the government purchase vector process is exogenous and satisfies Gt = UG zt .

336

Equilibria with Distortions

15.2.3. c. Taxes We add to the model of chapters 3–5 a government that taxes consumption goods, investment goods, capital goods, and the intermediate labor activity. The government’s budget constraint is ∞ X 0′ 0′ 0′ 0′ t β pt τc ct + qt τi it + rt τk kt−1 + wt τℓ gt |J0 E t=0

=E

∞ X t=0

β t p0t · Gt |J0 + T0

where Gt = UG zt . Here τc is a diagonal matrix of tax rates on consumption goods, τi is a diagonal matrix of tax rates on investment goods, τk is a diagonal matrix of tax rates on rentals of capital, and τℓ is a diagonal matrix of tax rates on the labor-using activity vector gt ; T0 is lump sum transfers to the household. The government sets the tax rates τc , τi , τk , τℓ and the expenditure process UG zt = Gt . For this economy, a price system is a collection of stochastic processes {p0t , qt0 , rt0 , 2 wt0 , αt0 }∞ t=0 , each element of which belongs to L0 . A tax system is a list of diagonal matrices [τc , τi , τk , τℓ ]. An allocation is a collection of stochastic processes 2 {st , ct , ht , kt , gt , it }∞ t=0 each element of which belongs to L0 . We now describe the choices faced by households and firms.

15.3. Households Households own an ‘endowment stream’ {dt + ΓK Kt−1 }∞ t=0 and the initial capital stocks, all of which they take as given. Households take the price system, the tax system, and {Ct , Ht−1 , Kt−1 }∞ t=0 as given, and choose contingency plans ∞ for {ct , it , st , ht , kt , gt }t=0 to maximize −

∞

1 X t E β [(st − bt ) · (st − bt ) + gt · gt ] | J0 2 t=0

(15.3.1)

subject to the household technology,

st = Λh ht−1 + ΛH Ht−1 + Πct + ΠC Ct

(15.3.2)

ht = ∆h ht−1 + ∆H Ht−1 + Θh ct + ΘH Ct

(15.3.3)

Firms

337

the law of accumulation for physical capital, kt = ∆k kt−1 + ∆K Kt−1 + Θk it ,

(15.3.4)

and the budget constraint E

∞ X t=0

′

′

β t [p0t (I + τc )ct + qt0 (I + τi )it ] | J0 =E

∞ X t=0

′

β t [wt0 (I − τℓ )gt + αt0 · (dt + ΓK Kt−1 )

(15.3.5)

′

+ rt0 (I − τk )kt−1 ] | J0 + T0

15.4. Firms There is one kind of firm, a production firm that takes the price system 2 and {Kt−1 }∞ t=0 as fixed, and chooses an allocation and a process {ζt } ∈ L0 that maximize E

∞ X t=0

β t {p0t · (ct + Gt ) + qt0 · it − rt0 · kt−1 − αt0 · ζt − wt0 · gt |J0 }

(15.4.1)

subject to the technology Φc (ct + Gt ) + Φg gt + Φi it = Γk kt−1 + ζt · ζt .

(15.4.2)

338

Equilibria with Distortions

15.5. Information We assume that zt+1 = A22 zt + C2 wt+1 , where {wt } is a martingale difference sequence with identity covariance matrix. The stochastic process {zt } drives the exogenous processes bit = Ubi zt dit = Udi zt Gt = UG zt .

15.6. Equilibrium An equilibrium is defined as an allocation, a price system, and a tax system such that (i) Given the price system, the tax system, and {Ht−1 , Ct−1 , Kt−1 }∞ t=0 , the allocation solves the household’s problem; (ii) Given the price system and {Kt−1 }∞ t=0 , the allocation solves the firm’s problem with ζt ≡ ΓK Kt−1 + dt ; (iii) The representative household is representative, Ht−1 = ht−1 , Ct = ct ; and the representative firm is representative, Kt−1 = kt−1 for all t ≥ 0. Given that conditions (i), (ii), (iii), are satisfied, the lump sum tax T0 , which is a present value, can be chosen to make the government budget constraint hold. We have set things up so that the equilibrium law of motion takes the form xt+1 = Ao xt + Cwt+1 , and equilibrium quantities and normalized Arrow-Debreu prices are linear func′ tions of the state x′t = [h′1t−1 h′2t−1 kt−1 zt′ ]. Appendix A describes how to compute the equilibrium of this model by manipulating agents’ first order conditions and the other equilibrium conditions

Heterogeneous Households with Distortions

339

into a system susceptible to the application of Vaughan’s algorithm. The MATLAB program solvdist.m computes an equilibrium. Equilibrium calculation is facilitated by the fact that there is a representative households (so that if there is heterogeneity among households, the features of Gorman aggregation isolate demand functions from effects of redistributions of wealth) and the permission that we give the government to levy lump sum taxes. In the next class of models, we give up both of these simplifying features.

15.7. Heterogeneous Households with Distortions In this section, we describe how to extend the above setup to enable us to compute the equilibrium of a model with externalities, government expenditures and taxes, and two classes of agents with non-Gorman aggregable preferences. To produce the model, we combine elements of the preference specification treated in chapter 8 with the specification and methods described earlier in this chapter. We do not let the government raise lump sum taxes or dispense lump sum transfers, but require the government to balance its budget by levying only flat rate taxes. These changes vis a vis the first class of models have the consequence that equilibrium must now be computed by finding values of tax rates and marginal utilities of wealth for each household that make households’ and the government’s budgets balance. 4

4 Between these two classes of models, we have changed the way we price labor or ‘intermediate goods’ gt . In the first class, the households own and sell to the firm the vector gt , which is priced by the vector wt0 . In the second model, households sell ℓit for wt0 to firms.

340

Equilibria with Distortions

15.7.1. Households There are two classes of households, indexed by i = 1, 2. Preferences of a household of type i are ordered by −.5E

∞ X t=0

β t [(sit − bit ) · (sit − bit ) + ℓ2it ]|J0 ,

(15.7.1)

where the household has access to the household technology sit = Λi1 hit−1 + Λi2 H1t−1 + Λi3 H2t−1 + Πi1 cit−1 + Πi2 C1t + Πi3 C2t hit = ∆hi hit−1 + ∆Hi1 Hit−1 + ∆Hi2 H2t−1 + Θhi cit + ΘH1i C1t + ΘH2i C2t

(15.7.2)

dit = Udi zt bit = Ubi zt . Here Cit , Hit are the aggregate consumption and stock of durables, respectively, of household type i, and cit , hit are the individual consumption and durables, respectively, of a household of type i. There are firms of two types, to be described in detail below. A household of type i is assumed to own a a fraction fi of the ‘technology’, which entitles it to sell fi (Γk + dt ) of an ‘endowment’ to a type I (production) firm, and to rent fi ∆K Kt−1 of capital to a type II (capital renting) firm at time t . Thus, the intertemporal budget constraint of a household of type i is E

∞ X t=0

′

β t {p0t (I + τc )cit − fi αt0 · (ΓK Kt−1 + dt ) −

fi rt0

· ∆K Kt−1 −

wt0 (1

(15.7.3)

+ τℓ )ℓit }|J0 − v0 k−1,i = 0

No lump sum transfers occur in (15.7.3), because we shall require the government to balance its budget using flat rate taxes only. Below, we will carry along a Lagrange multiplier µ0i on (15.7.3) for each type of household.

Heterogeneous Households with Distortions

341

15.7.2. Firms of type I Firms of type I are production firms. They maximize

E

∞ X t=0

β t {p0t · (ct + Gt ) + qt0 · it − rt0 · kt−1 − αt0 · ζt − wt0 ℓt }|J0

(15.7.4)

subject to

Φc (ct + Gt ) + Φi it + Φg gt = Γk kt−1 + ζt , gt · gt = ℓ2t

(15.7.5)

15.7.3. Firms of type II Firms of type II are capital renting firms. They maximize

E0

∞ X t=0

′

′

β t {rt0 (I − τk )kt−1 − qt0 (I + τi )it − νt rt · ∆K Kt−1 }|J0 − v0 k−1 (15.7.6)

subject to kt = ∆k kt−1 + νt ∆K Kt−1 + Θk it . Here νt is the amount of the ‘endowment’ ∆K Kt−1 purchased from households.

15.7.4. Government The government makes a flow of expenditures Gt on consumption goods, governed by Gt = UG zt . (15.7.7) The government’s budget constraint is ∞ X ′ ′ ′ β t {p0t · Gt − p0t τc · (c1t + c2t ) − qt0 τi it − rt0 τk kt−1 E t=0

−

τℓ wt0 (ℓ1t

(15.7.8)

+ ℓ2t )}|J0 = 0

A fiscal policy is a collection of diagonal matrices (τc , τi , τk , τℓ ) determining taxes and a matrix UG determining government expenditures.

342

Equilibria with Distortions

15.7.5. Definition of equilibrium An equilibrium is a price system [{p0t , qt0 , rt0 , αt0 , wt0 }∞ t=0 , v0 ], an individual allocation {c1t , c2t , h1t , h2t , kt , it , gt }, an aggregate allocation {C1t , C2t , H1t , H2t , Kt }, and a fiscal policy (τc , τi , τk , τℓ , UG ) that satisfy the following conditions: i. Given the aggregate allocation, the price system, and the fiscal policy, the individual allocation solves the optimum problems of households and the firms, with ζt ≡ 1 and νt ≡ 1. ii. The allocations satisfy cit = Cit , i = 1, 2 hit = Hit , i = 1, 2 kt = Kt ζ = (ΓK Kt−1 + dt ). iii. The government budget constraint is satisfied. As we have defined it, an equilibrium is typically not unique. In particular, there will usually be a set of taxes that serve to assure equilibrium. Below, we shall select one from among these equilibria partly by fixing enough of these taxes and ‘solving’ for others. 5

15.7.6. Equilibrium computation We compute an equilibrium by finding a fixed point in the parameters that index the tax system τc , τi , τk , τℓ and a pair of marginal utilities of wealth (multipliers) µ01 , µ02 for households. For fixed values of the tax system and multiplier parameters, we can use the modified Vaughan algorithm described in the preceding section to compute a ‘candidate’ 0 allocation and price system. For that allocation and price system, we evaluate all elements of the budget constraints of the government and the two types of households. We use a nonlinear search algorithm We use a secant algorithm. to find a tax system and pair of multipliers µ01 , µ02 that assure that the budget constraints are all satisfied. Thus, we compute an equilibrium by solving a fixed point problem in a space of tax rates and marginal utilities of wealth for the different household types. 5 Even when we fix the ‘right number’ of the taxes there can still be multiple equilibria for reason of ‘Laffer curves’.

Government Deficits and Debt

343

Appendix B describes how the first order conditions for households and firms together with the other equilibrium conditions can be arranged to obtain a system of equations of the form (9.25), which with the invariant subspace methods of chapter 9 permits us to compute our candidate equilibrium. The equilibrium is computed by disthet.m.

15.8. Government Deficits and Debt The government deficit at time t , measured in time t ‘spot’ prices, is Dtt = ptt · (Gt − τc (c1t + c2t )) − qtt · τi it − rtt · τk kt−1 − wtt τℓ (ℓ1t + ℓ2t ). Evidently, Dtt can be represented as Dtt = where

x′t QD xt , e1 Mc xt

(15.8.1)

QD = Mc′ [SG − τc (Sc1 + Sc2 )] − Mi′ τi Si τℓ − Mk′ τk Sk1 − S ′ Sg . (1 − τℓ )(µ01 + µ02 ) g

Let Vt denote the present value of the government deficit, which satisfies the difference equation Vt = −Dtt + βEt {pt1,t+1 Vt+1 }, (15.8.2) where pt1,t+1 = e1 Mc xt+1 /e1 Mc xt is the time t price of a state contingent claim to the first (numeraire) consumption good in time t + 1. Notice that βpt1,t+1 acts as a stochastic discount factor for evaluating government indebtedness next period from the standpoint of this period. This equation is the counterpart of the following version of a one period government budget constraint, which occurs in various nonstochastic macroeconomic models: −Dt + Vt+1 /Rt = Vt , where here Vt is interpreted as one-period debt falling due at time t . Equations (15.8.1) and (15.8.2) imply that Vt =

xt QV xt + σV , e1 Mc xt

where QV , σV are determined as follows. Define v˜ =

β doublej2(βAo , C, Ao , C). 1−β

(15.8.3)

344

Equilibria with Distortions

Then [QV ] = doublej2(βAo ′ , QD , Ao ′ , 1) and σV = trace(QV v˜). The matrix valued function doublej2 was used repeatedly in our asset pricing calculations. The forms of (15.8.2) and (15.8.3) mean that the econometric methods described in Hansen and Sargent (1993) for interpreting observations on asset prices can be used to model government budgets and bond holdings. 6

15.9. Examples

15.9.1. A production externality The following technology is designed to capture features of a specification of DeLong and Summers. There is one consumption good, but two capital goods (‘machines’ and ‘structures’) and two rates of investment. Machines generate a positive production externality, but not structures: ct +Gt + i1t + i2t = γ1 k1t−1 + γ2 k2t−1 + Γ1 K1t−1 + Γ2 K2t−1 + dt φ1 i1t = g1t φ2 i2t = g2t k1t = δk1 k1t−1 + i1t k2t = δk2 k2t−1 + i2t To capture DeLong and Summers’s idea, we set Γ1 = 0, Γ2 > 0, so that we interpret the first capital good as structures and the second as machines. To complete this example, we incorporate a single-agent version of a simple quadratic utility specification. We suppress heterogeneity among consumers and instead make the two types of consumers be identical in their preferences and endowment sequences. Preferences are ordered by −

∞

1X t β [(cit − bit )2 + ℓ2it ]. 2 t=0

6 The restrictions embedded in ( 15.8.2 ) and ( 15.8.3 ) should be compared with those studied in the literature on linear models of ‘present value budget balance.’ See Hansen, Roberds, and Sargent (1991) for a summary of this literature.

Examples

The information process satisfies 1 0 0 0 .95 0 A22 = 0 0 .8 0 0 0

0 0 0 .1

d1t = d2t = [ 3.5 Gt = [ 5

345

0

1√ 20 2 C2 = 0 0

0

1 0

.5

0 ] zt ,

0 0 .5 0

0 0 0 1

1 ] zt .

b1t = b2t = 30. These settings make endowments of the two types of households each the same first order autoregressive processes with mean 3.5 and serial correlation parameter .8, while government expenditures follows the process √ (20 2)−1 .5 Gt − 5 = w1t + w3t , 1 − .95L 1 − .1L where L is the lag operator. This process approximates a mixture of a ‘permanent shock’ (the moving average in w1t ) and a ‘transitory shock’ (the moving average in w3t . The mean of government expenditures is 5. We set γ1 = γ2 = .12, Γ1 = 0, Γ2 = .04, φ1 = φ2 = .5, δk1 = δk2 = .95, β = 1/1.05. These parameter settings make the two types of capital symmetric with respect to adjustment costs and ‘private productivity’ (γi ), but inject a positive production externality for the second capital good.

15.9.2. Consumption tax only With these parameter settings, we computed an equilibrium where the only tax is the scalar consumption tax τc on the single consumption good. The equilibrium tax τc = .2497. Figures 15.9.1.a and 15.9.1.b show a simulation of this equilibrium starting from the initial condition k−1,1 = k−1,2 = 100. Because we start from equal initial stocks of machines and structures, and because their private productivities are the same, the equilibrium must retain equality of structures and machines throughout time. Figures 15.9.1.a and 15.9.1.b embody this property, the rates of investment in machines and structures being identical. Figures 3 display realizations of government indebtedness Vt determined by (15.8.2). Figure 4 shows the government flow deficit.

346

Equilibria with Distortions

110

25

109 20

108

cons

107 15

106 105

10

104 inv

103

5

102 govt

0 0

10

20

30

101 40

50

60

70

80

90

100 0

100

Fig. 15.9.1.a. Total consumption, investment, and government expenditures in Delong-Summers economy with τc = .2497 and no investment subsidy.

10

20

30

40

50

60

70

80

90

100

Fig. 15.9.1.b. Machines and structures in Delong-Summers economy with τc = .2497 and no investment subsidy.

15.9.3. Machinery investment subsidy Figures 15.9.2 and 15.9.3 report the results of recomputing the equilibrium when we impose an investment subsidy τi2 = −.04. The only other tax that we permit the consumption tax, which must be set at τc = .2611 to induce equilibrium. The simulations show how the investment rates for machines and structures now diverge in the direction that we would expect: the economy moves to a path with more machines and fewer structures than the first (no machine subsidy) equilibrium. 7

7 Realizations of the exogenous stochastic processes are held constant across these simulations.

Examples

12

347

15

10 10 8 6

5

4 0

2 0

-5 -2 -4 0

10

20

30

40

50

60

70

80

90

100

Fig. 15.9.2.a. Present value of government surplus in Delong-Summers economy with τc = .2497 and no investment subsidy.

-10 0

10

20

30

40

50

60

70

80

90

100

Fig. 15.9.2.b. Government deficit in Delong-Summers economy with τc = .2497 and no investment subsidy.

15.9.4. ‘Personal’ habit persistence We consider an economy with one consumption good that is allocated to two types of households. The first type of household has preferences ordered by ∞ X β t [(c1t − 15)2 + ℓ21t ], E0 − .5 t=0

while the second has the habit-persistence indicated by the preferences ∞ X β t [(s2t − 15)2 + ℓ22t ] E0 − .5 t=0

where

s2t = −1h2t−1 + 2c2t

h2t = .8h2t−1 + .2c2t We set β = 1/1.05. The production technology has

ct + it = .11kt−1 + .0001Kt−1 + dt kt = .95kt−1 + it g1t = .5it

348

Equilibria with Distortions

125

25

120

cons

20

115 15

machines

110

105

10 machines

100

structures

5 structures

0 0

10

20

30

40

50

60

95

70

80

90

90 0

100

Fig. 15.9.3.a. Total consumption, investment, and government expenditures in Delong-Summers economy with τc = .2611 and investment subsidy, τi2 = −.04 .

10

20

30

40

50

60

70

80

90

100

Fig. 15.9.3.b. Machines and structures in Delong-Summers economy with τc = .2611 and investment subsidy, τi2 = −.04 .

which is a version of our one-good “growth, adjustment cost” technology with a small production externality. We set A22

Ud2

1 = 0 0

0 .8 0

0 0 .5

UG = [ 4 .1 0 ] 3.5 0 1 = Ud1 = 0 0 0 k−1,1 = 125, k−1,2 = 25.

Notice that we endow the first household with more capital and equal claim to the endowment stream, so the first household is richer. We do the following experiments with this economy. First, we set all taxes except τc equal to zero, and solve for the equilibrium values of (µ02 , τc ). They are (2.0765,.2117). A simulation of this economy is in figure 7. Then we reset

Examples

349

τi = .08, and solve again for (µ02 , τc ) = (2.0467, .27). A simulation of this economy is in figure 8. The simulations start from identical initial conditions. In this economy, the habit persistence of the second type of consumer is something of an “engine of growth”. The second type of household accumulates capital to support planned growth in its consumption. Taxing investment causes the household to cut back on these investments.

14

13 12

c1

12 c1

11

10

10

c2 9

8

8

c2 invest

6

7 6

4

invest

govt

5 2 4

govt

3 0

10

20

30

40

50

60

70

80

90

100

Fig. 15.9.4.a. Simulation of two-agent economy with τc = .2117, τi = 0 . Four series are plotted: consumption of type 1 consumer, consumption of type 2 consumer, investment, and government purchases. Equilibrium marginal utilities of wealth are µ01 = 1, µ02 = 2.0765 .

0 0

10

20

30

40

50

60

70

80

90

100

Fig. 15.9.4.b. Simulation of two-agent economy with τc = .1884, τi = .08 . Equilibrium marginal utilities of wealth are (1, 2.0467) .

350

Equilibria with Distortions

15.9.5. ‘Social’ habit persistence Our next example is identical to the previous one, except that we alter the equation generating ‘habits’ of the second type of household to h2t = .8h2t−1 + .2C2t . We computed equilibria of this economy with the same government expenditure process and the same permissible tax instruments as for the previous economy. Figures 15.9.4 and 15.9.5 report simulations of this economy starting from the same initial conditions as for the previous economy. Notice that slightly lower consumption tax rates provide equilibria in this economy, for the reason that due to the ‘social’ rather than ‘personal’ nature of habit persistence, they have slightly lower adverse demand effects on the second type of household.

14

12 c1

11

c1

12 10 9

10

8

c2 8

7 6

c2

6

invest

invest

5 4 4 3 0

govt

govt 10

20

30

40

50

60

70

80

90

100

Fig. @[email protected] Simulation of two-agent economy with τc = .2101, τi = 0 . The second type of agent has ‘keeping up with the Jones’ habit persistence with other agents of his type. Four series are plotted: consumption of type 1 consumer, consumption of type 2 consumer, investment, and government purchases.

2 0

10

20

30

40

50

60

70

80

90

100

Fig. 15.9.5.b. Simulation of two-agent economy with τc = .1883, τi = .08 .

Invariant subspace equations for first specification

351

15.10. Conclusions The models that we have described are rigged to make it practical to extend the data matching exercises performed by Braun (1991) and McGrattan (1991b). Our calculations make possible various extensions of Braun and McGrattan. The ease of computing equilibrium quantities, prices, and present values means that we can use data on quantities, interest rates, and asset prices to do method of moment estimation along the lines described by Hansen and Sargent (1993, chapter 11). In particular, we can use data on government indebtedness and deficits as well as asset prices to help estimate parameters. By accepting our linear-quadratic specifications, we purchase the ability to get our hands on the Arrow-Debreu prices (which is harder with the approximations used by Braun, McGrattan and others in the real business cycle literature), which makes it feasible for us to do without lump sum taxes in our second class of models. It is possible for us to have numbers of capital stocks and enough heterogeneity among households to generate interesting tax incidence effects. Finally, the machinery in this chapter is a useful one within which to revisit issues in the literature on time series implications of ‘present value budget balance’ (e.g., Hansen, Roberds, and Sargent (1991)).

A. Invariant subspace equations for first specification Our strategy is to use Lagrangian methods to obtain first order necessary conditions for households and firms. After obtaining those first order conditions, we substitute the equilibrium conditions (Ht−1 = ht−1 , Kt−1 = kt−1 , Ct = ct , and so on) into them. Then we rearrange the system into the form of (9.25) so that it is susceptible to application of the modified Vaughan method. We can exploit the certainty equivalence principle and solve a nonstochastic version of the model first, and later adjust the solution to accommodate randomness.

352

Equilibria with Distortions

15.A.1. Household’s Lagrangian A Lagrangian for a nonstochastic version of the household’s problem is ∞ X

1 L= β − [(st − Ub zt ) · (st − Ub zt ) + gt · gt ] 2 t=0 t

′

+ µ0 [wt0 (I − τℓ )gt + αt0 (Ud zt + ΓK Kt−1 ) ′

′

′

+ rt0 (I − τk )kt−1 + Tt − p0t (I + τc )ct − qt0 (I + τi )it ]

+ µs′ t [Λh ht−1 + ΛH Ht−1 + Πct + ΠC Ct − st ]

(15.A.1)

+ µh′ t [∆h ht−1 + ∆H Ht−1 + Θh ct + ΘH Ct − ht ]

+ µk′ t [∆k kt−1 + ∆K Kt−1 + Θk it − kt ] z′ + µt [A22 zt − zt+1 ]

We can set the multiplier on the budget constraint µ0 = 1, which will amount to selecting a numeraire. The first-order conditions for the household’s problem are µst = (bt − st )

(15.A.2)

(I + τc )p0t = Π′ µst + Θ′h µht

(15.A.3)

µht = βΛ′h µst+1 + β∆′h µht+1

(15.A.4)

0 µkt = β∆′k µkt+1 + β(I − τk )rt+1

(15.A.5)

gt = (I − τℓ )wt0

(15.A.6)

(I + τi )qt0 = Θ′k µkt

(15.A.7)

0 µzt = β[A′22 µzt+1 + Ub′ (st+1 − Ub zt+1 ) + Ud′ αt+1 ]

(15.A.8)

Invariant subspace equations for first specification

353

The law of motion for the state variables chosen by the household and by nature is ht ∆h 0 0 ht−1 Θh 0 ct kt = 0 ∆k 0 kt−1 + 0 Θk it zt+1 0 0 A22 zt 0 0 (15.A.9) ∆H 0 ΘH Ht−1 + 0 ∆K 0 Kt−1 0 0 0 Ct

15.A.2. Firm’s first order conditions The first-order conditions for the firm’s problem are p0t = Φ′c αt0

(15.A.10)

qt0 = Φ′i αt0

(15.A.11)

wt0 = −Φ′g αt0

(15.A.12)

rt0 = Γ′k αt0

(15.A.13)

The feasibility condition is

Φc (ct + Gt ) + Φi it + Φg gt = Γk kt−1 + ΓK Kt−1 + dt

(15.A.14)

354

Equilibria with Distortions

15.A.3. Representativeness conditions Additional equilibrium conditions are ht−1 = Ht−1 kt−1 = Kt−1

(15.A.15)

ct = Ct We describe in detail how to compute the equilibrium of the first type of model by arranging its equilibrium conditions into the form of equation (9.25). Substituting the equilibrium conditions (15.A.15) into (15.A.9) gives ¯ ht ∆h kt = 0 zt+1 0

0 ¯k ∆

0

¯ 0 ht−1 Θh 0 kt−1 + 0 A22 zt 0

0 ct Θk it 0

(15.A.16)

¯ h = ∆h + ∆H , ∆ ¯ k = ∆k + ∆K , Θ ¯ h = Θh + ΘH . Equations (15.A.10) where ∆ and (15.A.12) imply ′ −1 0 Φc pt . (15.A.17) αt0 = Φ′g −wt0 Substituting from (15.A.3) and (15.A.6) into (15.A.17) gives αt0

Φ′c = Φ′g

−1

(I + τc )−1 [Π′ µst + Θ′h µht ] −(I − τℓ )−1 gt

(15.A.18)

Using kt−1 = Kt−1 in (15.A.14) and solving for [c′t gt′ ]′

ct gt

¯ t−1 + Uf zt − Φi it } = [Φc Φg ]−1 {Γk

(15.A.19)

¯ = Γk + ΓK . where Uf = Ud − Φc UG and Γ Substituting ht−1 = Ht−1 and ct = Ct into equation (15.3.2) st gives ¯ t st = Λht−1 + Πc ¯ = Π + ΠC . where Λ = Λh + ΛH and Π Collecting our results to this point, we want to solve the following system of difference equations µst = (Ub zt − st ) (15.A.20)

Invariant subspace equations for first specification

355

¯ t st = Λht−1 + Πc

(15.A.21)

µht = βΛ′h µst+1 + β∆′h µht+1

(15.A.22)

0 µkt = β∆′k µkt+1 + β(I − τk )Γ′k αt+1

(15.A.23)

0 µzt = β[A′22 µzt+1 + Ub′ (st+1 − Ub zt+1 ) + Ud′ αt+1 ]

¯ ht ∆h kt = 0 zt+1 0

0 ¯k ∆

αt0 =

0

ct gt

Φ′c Φ′g

0 0 A22

¯ ht−1 Θh kt−1 + 0 zt 0

(15.A.24)

0 ct Θk (15.A.25) it 0

¯ t−1 + Uf zt − Φi it } = [Φc Φg ]−1 {Γk

−1

(I + τc )−1 [Π′ µst + Θ′h µht ] −(I − τℓ )−1 gt

(15.A.26)

(15.A.27)

Φ′i αt = (I + τi )−1 Θ′k µkt ,

(15.A.28)

where the last equation comes from combining (15.A.11) (among the firm’s firstorder necessary conditions) with (15.A.7) (among the household’s first-order necessary conditions). Our goal is to manipulate this system into the form (9.25) that is susceptible to the application of the invariant subspace algorithm of chapter 9. Our strategy will be successively to eliminate it , ct , gt , αt0 , st , and µst from the system, so that we are left with a difference equation in the “state” (ht−1 , kt−1 , zt ) and the “co-state” variables µht , µkt , µzt . We begin by substituting (15.A.27) and (15.A.26) into (15.A.28) to get

Φ′i

Φ′c Φ′g

−1

(I + τc )−1 [Π′ µst + Θ′h µht ] ¯ t−1 + Uf zt − Φi it } −(I − τℓ )−1 Ug [Φc Φg ]−1 {Γk

−1

= (I + τi )

Θ′k µkt

(15.A.29)

356

Equilibria with Distortions

where as usual Ug is a selection matrix that picks off the components of the right side of (15.A.26) corresponding to gt . We adopt the partition ˜= Φ

Φ′c Φ′g

−1

˜1 Φ ˜ 2] = [Φ

(15.A.30)

˜ 1 is (nd × nc ) and Φ ˜ 2 is (nd × ng ). Then equation (15.A.29) can be where Φ written as ˜ 1 (I + τc )−1 [Π′ µst + Θ′h µht ] Φ′i Φ ˜ 2 (I − τℓ )−1 Ug [Φc Φg ]−1 {Γk ¯ t−1 + Uf zt } − Φ′i Φ ˜ 2 (I − τℓ )−1 Ug [Φc Φg ]−1 Φi it − (I + τi )−1 Θ′ µk = −Φ′ Φ k t

i

Solving for it gives it = L1 µst + L2 µht + L3 µkt + L4 kt−1 + L5 zt

(15.A.31)

where ′ ˜ −1 ′ L1 = G−1 Π 1 Φi Φ1 (I + τc ) −1 ′ ′ ˜ Θh L2 = G−1 1 Φi Φ1 (I + τc ) −1 ′ L3 = −G−1 Θk 1 (I + τi ) −1 ′ ˜ ¯ L4 = −G1 Φi Φ2 (I − τℓ )−1 Ug [Φc Φg ]−1 Γ −1 ′ ˜ −1 −1 L5 = −G1 Φi Φ2 (I − τℓ ) Ug [Φc Φg ] Uf ˜ 2 (I − τℓ )−1 Ug [Φc Φg ]−1 Φi G1 = −Φ′ Φ

(15.A.32)

i

Substituting (15.A.31) into (15.A.26) and rearranging gives

where

ct = L6 kt−1 + L7 zt + L8 µst + L9 µht + L10 µkt

(15.A.33)

gt = L11 kt−1 + L12 zt + L13 µst + L14 µht + L15 µkt

(15.A.34)

Invariant subspace equations for first specification

357

¯ − Φi L4 ) L6 = Uc [Φc Φg ]−1 (Γ

L7 = Uc [Φc Φg ]−1 (Uf − Φi L5 )

L8 = −Uc [Φc Φg ]−1 Φi L1

L9 = −Uc [Φc Φg ]−1 Φi L2

L10 = −Uc [Φc Φg ]−1 Φi L3 ¯ − Φi L4 ) L11 = Ug [Φc Φg ]−1 (Γ

(15.A.35)

L12 = Ug [Φc Φg ]−1 (Uf − Φi L5 ) L13 = −Ug [Φc Φg ]−1 Φi L1

L14 = −Ug [Φc Φg ]−1 Φi L2

L15 = −Ug [Φc Φg ]−1 Φi L3

˜1 Φ ˜ 2] = Substituting (15.A.34) into (15.A.27) and using our partition [Φ gives αt0 = L16 µst + L17 µht + L18 µkt + L19 kt−1 + L20 zt

Φc Φg

−1

(15.A.36)

where ˜ 1 (I + τc )−1 Π′ + Φ ˜ 2 (I − τℓ )−1 Ug [Φc Φg ]−1 Φi L1 L16 = Φ ˜ 1 (I + τc )−1 Θ′ + Φ ˜ 2 (I − τℓ )−1 Ug [Φc Φg ]−1 Φi L2 L17 = Φ h ˜ 2 (I − τℓ )−1 Ug [Φc Φg ]−1 Φi L3 L18 = Φ

(15.A.37)

˜ 2 (I − τℓ ) Ug [Φc Φg ] {Γ ¯ − Φi L4 } L19 = −Φ ˜ 2 (I − τℓ )−1 Ug [Φc Φg ]−1 (Uf − Φi L5 ) L20 = −Φ −1

−1

Substituting (15.A.33) for ct into (15.A.20) and (15.A.21) gives µst = L21 ht−1 + L22 kt−1 + L23 zt + L24 µht + L25 µkt where

(15.A.38)

L21 = −A −1 Λ ¯ 6 L22 = −A −1 ΠL

¯ 7) L23 = A −1 (Ub − ΠL −1 ¯ L24 = −A ΠL9 ¯ 10 L25 = −A −1 ΠL ¯ 8) A = (I + ΠL

(15.A.39)

358

Equilibria with Distortions

Because st = Ub zt − µst , (15.A.38) implies st = L26 ht−1 + L27 kt−1 + L28 zt + L29 µht + L30 µkt

(15.A.40)

where L26 = −L21

L27 = −L22

L28 = Ub − L23

(15.A.41)

L29 = −L24

L30 = −L25

We can use (15.A.38) to eliminate µst from the right sides of (15.A.31), (15.A.33), (15.A.34), (15.A.36) to obtain it = N1 ht−1 + N2 kt−1 + N3 zt + N4 µht + N5 µkt

(15.A.42)

ct = N6 ht−1 + N7 kt−1 + N8 zt + N9 µht + N10 µkt

(15.A.43)

gt = N11 ht−1 + N12 kt−1 + N13 zt + N14 µht + N15 µkt

(15.A.44)

αt = N16 ht−1 + N17 kt−1 + N18 zt + N19 µht + N20 µkt

(15.A.45)

where N1 N 2 N3 N4 N5

= L1 L21 = L1 L22 + L4 = L1 L23 + L5

(15.A.46)

= L1 L24 + L2 = L1 L25 + L3

N6 = L8 L21 N = L8 L22 + L6 7 N8 = L8 L23 + L7 N9 = L8 L24 + L9 N10 = L8 L25 + L10

(15.A.47)

Invariant subspace equations for first specification

N11 N12 N13 N14 N15 N16 N17 N18 N19 N20

359

= L13 L21 = L13 L22 + L11 = L13 L23 + L12

(15.A.48)

= L13 L24 + L14 = L13 L25 + L15 = L16 L21 = L16 L22 + L19 = L16 L23 + L20

(15.A.49)

= L16 L24 + L17 = L16 L25 + L18

We now substitute (15.A.42), (15.A.43), (15.A.44), (15.A.45) into (15.A.22), (15.A.23), (15.A.24), (15.A.25) to obtain the following system of difference equations: µht = βΛ′h L21 ht + βΛ′h L22 kt + βΛ′h L23 zt+1

(15.A.50)

+ (βΛ′h L24 + β∆′h )µht+1 + βΛ′h L25 µkt+1 µkt = [β∆′k + β(I − τk )Γ′k N20 ]µkt+1

+ β(I − τk )Γ′k N16 ht + β(I − τk )Γ′k N17 kt

+ β(I −

τk )Γ′k N18 zt+1

+ β(I −

(15.A.51)

τk )Γ′k N19 µht+1

µzt = βA′22 µzt+1 + β[Ub′ L26 + Ud′ N16 ]ht + β[Ub′ L27 + Ud′ N17 ]kt + β[Ub′ L28 + Ud′ N18 − Ub′ Ub ] zt+1

(15.A.52)

+ β[Ub′ L29 + Ud′ N19 ]µht+1

+ β[Ub′ L30 + Ud′ N20 ]µkt+1 ¯h + Θ ¯ h N6 )ht−1 + Θ ¯ h N7 kt−1 ht = (∆ ¯ h N10 µk ¯ h N8 z t + Θ ¯ h N9 µh + Θ +Θ t

¯ k + Θk N2 )kt−1 + Θk N1 ht−1 kt = (∆ + Θk N3 zt + Θk N4 µht + Θk N5 µkt zt+1 = A22 zt

(15.A.53)

t

(15.A.54) (15.A.55)

360

Equilibria with Distortions

We can arrange these equations in the form of (9.25) as follows: m2 where m2 =

0 0 0 ¯ ¯ h N6 ) (∆h + Θ Θk N1 0

and

xt µt

= m1

0 0 0 ¯ h N7 Θ ¯ k + Θk N2 ∆ 0

xt+1 , µt+1

0 0 0 ¯ h N8 Θ Θk N3 A22

I 0 0 ¯ h N9 Θ Θk N4 0

0 I 0 ¯ Θh N10 Θk N5 0

0 0 I 0 0 0

m1 =

βΛ′h L21 βΛ′h L22 βΛ′h L23 ′ β(I − τk )Γ′ N16 β(I − τk )Γk N17 β(I − τk )Γ′k N18 k β(U ′ L + U ′ N ) β(U ′ L + U ′ N ) β(U ′ L + U ′ N − U ′ U ) b 26 d 16 b 27 d 17 b 28 d 18 b b I 0 0 0 I 0 0 0 I βΛ′h L24 + β∆′h βΛ′h L25 0 β(I − τk )Γ′k N19 β∆′k + β(I − τk )Γ′k N20 0 ′ ′ ′ ′ β(Ub L29 + Ud N19 ) β(Ub L30 + Ud N20 ) βA′22 0 0 0 0 0 0 0

0

0

′ k′ z′ where we set x′t = [h′t−1 , kt−1 , zt′ ] and µ′t = [µh′ t µt µt ]. Then the equilibrium law of motion for {xt } is given by

xt+1 = Ao xt

(15.A.56)

where −1 V21 ). Ao = W11 ∆1 (V11 − V12 V22

The shadow prices are determined via (12.5b), namely, µt = M xt

(15.A.57)

Invariant subspace equations for heterogeneous agent model

361

−1 where M = W21 W11 . Using (15.A.42), (15.A.43), (15.A.44), and (15.A.45) with (15.A.56) we can write

or

n o it = [N1 N2 N3 ] + [N4 N5 0] M xt n o ct = [N6 N7 N8 ] + [N9 N10 0] M xt n o gt = [N11 N12 N13 ] + [N14 N15 0] M xt n o αt0 = [N16 N17 N18 ] + [N19 N20 0] M xt

(15.A.58) (15.A.59) (15.A.60) (15.A.61)

it = Si xt ct = Sc xt gt = Sg xt

(15.A.62)

αt0 = Sα xt . Substituting (15.A.62) for αt0 into (15.A.10), (15.A.11), (15.A.12), (15.A.13) gives p0t = Sp xt qt0 = Sq xt wt0 = Sw xt

(15.A.63)

rt0 = Sr xt where

We also have

Sp = Φ′c Sα , Sq = Φ′i Sα Sw = −Φ′g Sα , Sr = Γ′k Sα ht = Sh xt , kt = Sk xt

(15.A.64)

(15.A.65)

where Sh = [I 0 0] Ao

,

Sk = [0 I 0] Ao .

(15.A.66)

These formulas express the law of motion for the state xt and all quantities and prices in forms identical to those described in chapters 3–5. With these formulas in hand, all subsequent features of our analysis proceed identically as with that for an undistorted economy.

362

Equilibria with Distortions

B. Invariant subspace equations for heterogeneous agent model

From the first-order conditions of the two types of households and firms the market clearing conditions we can deduce the following set of equations:

s1t = Λ1 h1t−1 + Λ13 h2t−1 + Π1 c1t + Π13 c2t

(15.B.1)

s2t = Λ2 h2t−1 + Λ22 h1t−1 + Π2 c2t + Π22 c1t (15.B.2) ˜ h1 h1t−1 + ∆H12 h2t−1 + Θ ˜ h1 c1t + ΘH12 c2t (15.B.3) h1t = ∆ ˜ h2 h2t−1 + ∆H21 h1t−1 + Θ ˜ h2 c2t + ΘH21 c1t (15.B.4) h2t = ∆ zt+1 = A22 zt ct = Γkt−1 + dt − Φi it − Φc UG zt (Φc Φg ) gt ˜ k kt−1 + Θk it kt = ∆

(15.B.5)

(I +

τc )µ0i p0t Mhi t

= =

Π′i1 (Ubi zt − sit ) + Θ′hi Mhi t ′ hi ′ β∆hi Mt+1 + βΛi1 (Ubi zt+1

(15.B.6) (15.B.7) i = 1, 2

− sit+1 ) i = 1, 2 (15.B.9)

′ zi ′ Mzi t = β[A22 Mt+1 + Ubi (sit+1 − Ubi zt+1 )

αt0 =

Φ′c Φ′g

−1

′ 0 + Udi αt+1 µ0i ] i = 1, 2 0 pt wt0 − lt gt

lit = (1 − τl )µ0i wt0

(I +

Mkt τi )Φ′i αt0

=

=

β∆′k Mkt+1 Θ′k Mkt

i = 1, 2

+ β(I −

(15.B.8)

0 τk )Γ′k αt+1

(15.B.10) (15.B.11) (15.B.12) (15.B.13) (15.B.14)

Invariant subspace equations for heterogeneous agent model

363

where Λ1 = Λ11 + Λ12 Π1 = Π11 + Π12 Λ2 = Λ21 + Λ23 Π2 = Π21 + Π23 ˜ h1 = ∆h1 + ∆H11 ∆ ˜ h2 = ∆h2 + ∆H22 ∆ ˜ h1 = Θh1 + ΘH11 Θ ˜ h2 = Θh2 + ΘH22 Θ ˜ k = ∆k + ∆K ∆ Γ = Γk + ΓK Now, use (15.B.8) twice for i = 1, 2 and (15.B.1), (15.B.2), we get ′ µ−1 01 (Π11 (Ub1 zt − Λ1 h1t−1 − Λ13 h2t−1 − Π1 c1t − Π13 c2t )

+ Θ′h1 Mh1 t )=

′ µ−1 02 (Π21 (Ub2 zt − Λ2 h2t−1 − Λ22 h1t−1 − Π2 c2t − Π22 c1t )

(15.B.15)

+ Θ′h2 Mh2 t )

From (15.B.6) we get c1t + c2t = Uc (Φc Φg )−1 [Γkt−1 − Φi it + Uf zt ] −1

gt = Ug (Φc Φg )

[Γkt−1 − Φi it + Uf zt ]

(15.B.16) (15.B.17)

Also, combining (15.B.14), (15.B.11), (15.B.12) and (15.B.8) gives us ′˜ −1 µ−1 [Π′11 (Ub1 zt − Λ1 h1t−1 01 Φi Φ1 (I + τc )

− Λ13 h2t−1 − Π1 c1t − Π13 c2t ) + Θ′h1 Mh1 t ] ′˜ −1 ′ − Φ Φ2 gt /(1 − τl )(µ01 + µ02 ) = (I + τi ) Θ Mk i

k

(15.B.18)

t

˜1 , Φ ˜ 2 , Uc and Ug are as defined in Chapter 10, Uf = Ud1 + Ud2 − Here Φ Φc UG . Equations (15.B.15) – (15.B.18) can be used to solve 4 variables c1t , k c2t , gt and it in terms of the state variables zt , hit−1 , Mhi t , kt−1 and Mt , i = 1, 2, as:

Equilibria with Distortions

364

zt zt h1t−1 c1t h2t−1 ˜ c2t = N ˜ kt−1 L Mz1 gt t it Mz2 t h1 Mt Mh2

(15.B.19)

t

Mkt

where

˜= L

and

µ01 Π′21 Π22 − µ02 Π′11 Π1 Ic 0 −1 Π′ Π ′˜ µ−1 11 1 01 Φi Φ1 (I + τc )

µ01 Π′21 Π2 − µ02 Π′11 Π13 Ic 0 −1 Π′ Π ′˜ µ−1 11 13 01 Φi Φ1 (I + τc )

0 0 Ig ˜2 (1 − τl )−1 (µ01 + µ02 )−1 Φ′i Φ

0 Uc (Φc Φg )−1 Φi Ug (Φc Φg )−1 Φi 0

Invariant subspace equations for heterogeneous agent model

365

˜ = N [−µ02 Π′11 Ub1 ]′ [µ01 Π′21 Ub2 ]′ [µ02 Π′11 Λ1 − µ01 Π′21 Λ22 ]′ [µ02 Π′11 Λ13 − µ01 Π′21 Λ2 ]′ 0 0 0 [−µ02 Θ′h1 ]′ [µ01 Θ′h2 ]′ 0

[Uc (Φc Φg )−1 Uf ]′ 0 0 0 [Uc (Φc Φg )−1 Γ]′ 0 0 0 0 0

[Ug (Φc Φg )−1 Uf ]′ 0 0 0 [Ug (Φc Φg )−1 Γ]′ 0 0 0 0 0

−1 Π′ U ]′ ′ ′˜ [µ−1 11 b1 01 Φi Φ1 (I + τc ) 0 −1 Π′ Λ ]′ ′˜ [−µ−1 11 1 01 Φi Φ1 (I + τc ) ˜ 1 (I + τc )−1 Π′ Λ13 ]′ [−µ−1 Φ′ Φ 01

11

i

0 0 0 ′Φ ˜ 1 (I + τc )−1 Θ′ ]′ [µ−1 Φ 01 i h1 0 [−(I + τi )−1 Θ′k ]′

From equation (15.B.19) c1t , c2t , gt , it can be expressed as a linear combination of state vector xt and costate vector Mt , as: c1t c2t = N xt gt Mt it

where ˜ ˜ −1 N N =L xt = ( zt′

zt′

h′1t−1

h′2t−1

′ kt−1 )

′

and Mt is the corresponding multipliers of each component of xt .

Divide N into blocks according to the dimensions of cit , gt , it , xt and Mt , such that:

Equilibria with Distortions

366

c1t = N11 xt + N12 Mt = N1 (x′t M′t )′

c2t = N21 xt + N22 Mt = N2 (x′t M′t )′

gt = N31 xt + N32 Mt = N3 (x′t M′t )′

(15.B.20)

it = N41 xt + N42 Mt = N4 (x′t M′t )′

As an intermediate step, we express s1t , s2t and αt0 in terms of xt and Mt . This is done by using (15.B.1), (15.B.2), (15.B.8), (15.B.11) and (15.B.12):

s1t = Ns1 (x′t M′t )′

s2t = Ns2 (x′t M′t )′

(15.B.21)

αt0 = Nα (x′t M′t )′

such that

Ns1 = Π1 N1 + Π13 N2 + (0 0 Λ1 Λ13 0 0 0 0 0 0) Ns2 = Π22 N1 + Π2 N2 + (0 0 Λ22 Λ2 0 0 0 0 0 0) −1 ˜ Nα = µ−1 [−Π′11 Ns1 + (Π′11Ub1 0 0 0 0 0 0 Θ′h1 0 0)] 01 Φ1 (I + τc ) ˜ 2 N3 − (1 + τl )−1 (µ01 + µ02 )−1 Φ

Now, it is possible to combine all equations (15.B.1)–(15.B.14) to write down the Vaughan’s linear equations:

˜1 M

where

xt+1 Mt+1

˜2 =M

xt Mt

(15.B.22)

Invariant subspace equations for heterogeneous agent model

I 0 0 0 0 ˜1 = M ′ U −βUb1 b1 0 βΛ′ Ub1 0 11 0 0

0 I 0 0 0 0 ′ U −βUb2 b2 0 βΛ′21 Ub2 0

0 0 I 0 0 0 0 0 0 0

0 0 0 I 0 0 0 0 0 0

0 0 0 0 I 0 0 0 0 0

0 0 0 0 0 βA′22 0 0 0 0

0 0 0 0 0 0 βA′22 β∆′h1 0 0

0 0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 β∆′h2 0

0 0 0 0 0 0 0 I 0 0

0 0 0 0 0 0 0 0 I 0

0 0 0 0 0 β ′ ′ N α Ub1 Ns1 + µ01 Ud1 U ′ N + µ U ′ Nα 02 d2 b2 s2 ′ N −Λ 11 s1 −Λ′21 Ns2 (I − τk )Γ′k Nα

and A22 ˜2 = M

0 0 0 0 0 0 0 0 0

0 A22 0 0 0 0 0 0 0 0

0 0 0 0 0 0 ˜ ∆ ∆H12 0 h1 ˜ ∆H21 ∆ 0 h2 ˜ 0 0 ∆ k 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 ˜ Θ N2 h1 N1 + ΘH12 ˜ N2 Θ N1 + Θ H21

+

h2

Θk N4 0 0 0 0 0

0 0 0 0 0 I 0 0 0 0

0 0 0 0 0 0 I 0 0 0

0 0 0 0 0 0 0 0 0 I

0 0 0 0 0 0 0 0 β∆′k

367

+

Chapter 16 Recursive Risk Sensitive Control

16.1. Introduction This chapter describes a class of preferences with a single additional parameter that permits us to represent altered attitudes toward risk vis a vis our earlier quadratic specification. We build on work by Jacobson (1973) and Whittle (1990), who proposed an exponential type of risk adjustment that, in linear– quadratic environments with Gaussian disturbances, preserves most of the computational conveniences of the standard undiscounted linear quadratic dynamic programming problem. But Whittle’s (1981, 1990) way of introducing discounting into this problem had the unpleasant feature of introducing time dependence into the optimal decision rules, with the time dependence causing the effects of the risk parameter σ to wear off as the planning horizon is extended. By using a recursive specification of utility along the lines of Koopmans, Lucas and Stokey, Epstein and Zin, and Weil, Hansen and Sargent (1992) altered Whittle’s specification to model discounting in a way that implies time invariance of optimal decision rules. In this chapter we describe this preference specification and how it can be implemented with versions of our earlier formulas for computing equilibria and asset prices. The principal features of our specification are that: (a.) the equations for solving the optimal linear regulator problem are modified by replacing the operator associated with the usual matrix Riccati equation with the composition of that operator with another operator that is easily computed and interpreted; (b.) optimal decision rules remain linear in the state; (c.) ‘certainty equivalence’ no longer holds, i.e., the decision rules, although linear in the state, depend on the conditional covariance matrix of the innovations to the state; (d.) asset pricing formulas have an additional layer of ‘risk adjustments,’ because the covariance conditional covariance matrix of the innovations to the state now enter the quadratic form in the state in our asset pricing formulas.

– 369 –

370

Recursive Risk Sensitive Control

16.2. A Control Problem This section describes a ‘discounted linear quadratic exponential Gaussian control problem that we shall eventually use for the social planning problem of our standard economy. This is an infinite horizon control problem associated with iterations to convergence on the following equation in the value functions Vj (x): Vj+1 (x) = max′ {u′ Qu + x′ Rx + u,x

2β log E(exp(σVj (x′ )/2)|J)}, σ

(16.2.1)

where the maximization is subject to x′ = Ax + Bu + Cw, where w is a Gaussian random vector with Eww′ = I , and where β ∈ (0, 1) is a discount factor. Associated with the solution of this control problem are three operators: D(V ) = V + σV C(I − σC ′ V C)−1 C ′ V

T (W ) = R + A′ (βW − β 2 W B(Q + βB ′ W B)−1 B ′ W )A β d(k, V ) = βk − log det(I − σC ′ V C). σ

(16.2.2)

The operator T is the usual one associated with the matrix Riccati difference equation for the discounted optimal linear regulator problem. We shall give an interpretation of the new operator D shortly. The value function associated with the solution of the infinite horizon control problem is V∞ (x) = x′ U1 x + U0 (16.2.3) where U1 = lim (T ◦ D)j (0) j→∞

U0 = lim dj (0, 0) j→∞

The optimal decision rule is time invariant ut = −F xt where F = β[Q + βB ′ D(U1 )B]−1 B ′ D(U1 )A.

(16.2.4)

These formulas can be derived by solving the maximization problem on the right side of (16.2.1) and using a Lemma stated by Jacobson (1973) on

Pessimistic Interpretation

371

a property of the Gaussian distribution. For details, see Hansen and Sargent (1992). 1 The optimal decision rule is linear in the state. When σ = 0, we get the standard linear–quadratic situation, because in that case D = I , so that the operator T ◦ D = T . When σ = 0, the feedback rule F is independent of the conditional covariance parameters in C , so that ‘certainty equivalence’ holds when σ = 0. When σ 6= 0, the decision rule F depends on the parameters in C.

16.3. Pessimistic Interpretation For some purposes, it is convenient to use an interpretation of the D operator due to Jacobson (1973) and Whittle (1990). Evaluate the following ‘aggregator function’ associated with the functional equation (16.2.1): A(−F x, x, y ′ V y + k|J) = x′ R∗ x +

2β log E{exp(σ(y ′ V y + k)/2)|J} (16.3.1) σ

where A∗ = A − BF, R∗ = F ′ QF + R, and y = A∗ x + Cw. We obtain A(−F x, x, y ′ V y + k|J) = x′ [R∗ + βA∗′ D(V )A∗ ]x + d(k). The piece x′ [R∗ + βA∗′ D(V )A∗ ]x has an interpretation in terms of ‘pessimism,’ to be seen as follows. Consider the following deterministic minimum problem β min{− w′ w + x′ R∗ x + βy ′ V y} w,y σ subject to y = A∗ x + Cw. 1 This specification of discounting differs from the one used by Whittle (1990) in a way designed to assure time-invariant decision rules.

372

Recursive Risk Sensitive Control

In this problem, w is treated as a control vector. As long as (I − σC ′ V C) is positive definite, this problem has a unique solution. The minimized value of the criterion is verified to be x′ [R∗ + βA∗′ D(V )A∗ ]x , which is the quadratic in x part of A(−F x, x, y ′ V y + k|J). The ‘solution’ for w is w = σ(I − σC ′ V C)−1 C ′ V A∗ x. Thus, the optimum problem gives an interpretation of the operator D and a piece σ(I − σC ′ V C)−1 C ′ V A∗ x in terms of a ‘pessimistic’ (when σ < 0) or ‘optimistic’ (when σ > 0) adjustment to the random term wt in the law of motion for the state.

16.4. Recursive Preferences The preceding generalization of linear–quadratic control theory can readily be applied in the context of representative agent versions of the class of economic models that we have been studying. We use the usual household technology st = Λht−1 + Πct ht = ∆h ht−1 + Θh ct . We define the information process zt as usual by zt+1 = A22 zt + C2 zt , and let the law of motion for the complete state xt be denoted xt+1 = Axt + Cwt+1 , where now wt is a Gaussian white noise. We define a utility index recursively by Ut = −(st − bt ) · (st − bt )/2 + βRt (Ut+1 ) where Rt (Ut+1 ) =

2 log E[exp(σUt+1 /2)|Jt ]. σ

(16.4.1)

Asset Pricing

373

Hansen, Sargent, and Tallarini (1993) have shown how to construct the prices induced by this preference specification. We describe these prices first in the context of an endowment economy. 2

16.4.1. Endowment economy Consider a pure endowment economy in which ct = Uc zt is exogenous. Infinite recursions on (16.4.1) lead to the quadratic form Ute = x′t Ωxt + ρ, where Ω, ρ are the limits as j → ∞ of recursions on Ωj+1 = −(Ss − Sb )′ (Ss − Sb )/2 + βA′ D(Ωj )A ρj+1 = d(ρj , Ωj ).

For endowment economies, we use Ω as one of the ingredients in constructing a probability measure appropriate for the inner-product representation for asset pricing.

16.5. Asset Pricing We want to represent asset prices in the usual way as a conditional mathematical expectation of an infinite discounted sum of an inner product of a ‘scaled ArrowDebreu price’ vector and the payout of the asset. Hansen, Sargent, and Tallarini (1993) show that when σ 6= 0, the required conditional expectations operator for representing asset prices corresponds to a probability distribution that is distorted relative to the ‘objective’ one. In particular, they show that to attain the inner product representation of asset prices it is appropriate to construct a conditional expectations operator Ft defined by Ft Ut+1 = E(Vt+1 Ut+1 |Jt )/E(Vt+1 |Jt ), 2 The Matlab program solvex.m computes equilibrium quantities and prices in one of our economies with this preference specification; solvex.m calls doublex.m, which implements a doubling algorithm to solve the control problem.

374

Recursive Risk Sensitive Control

where e Vt+1 = exp(σUt+1 /2).

The operator Ft behaves like a conditional expectation operator in that it is linear, monotone, and maps bounded random variables that are measurable with respect to Jt+1 into bounded random variables that are measurable with respect to Jt . 3 We can use βFt to value a contingent claim to time t + 1 utility, and Mst for the equilibrium valuation of services. Recursively define the sequence of expectations operators St,τ = Ft Ft+1 Ft+2 · · · Ft+τ −1 , where Ft,0 = I . Then the valuation of a stream of consumption services {st } is just ∞ X St,τ Mst+τ · st+τ . t=0

To compute equilibrium asset prices, we have to evaluate consumption rates ct . We compute a multiplier Mct from Mst in the usual way, except that we substitute the conditional expectation operators St,τ for the usual ones in equation (6.*) (refer to asset pricing chapter): Mct = Π′ Mst + Θ′h

∞ X

τ =1

β τ (∆h )τ ′ Λ′ St,τ (Mst+τ ).

3 V t+1 /EVt+1 |Jt is used as a Radon-Nikodym derivative in this construction.

Characterizing the Pricing Expectations Operator

375

16.6. Characterizing the Pricing Expectations Operator Hansen, Sargent, and Tallarini show how to evaluate the conditional expectations operators St,τ by constructing a distorted probability measure for the state xt . 4 In particular, consider the phony law of motion for xt ˆ t + Cw ˆ t+1 xt+1 = Ax where

Aˆ = [I + σC(I − σC ′ ΩC)−1 C ′ Ω]A Cˆ Cˆ ′ = C(I − σC ′ ΩC)−1 C ′ .

Then we can compute Mct = Mc xt where Mc = [Π′ + Θ′h

∞ X

τ =1

β τ ∆τh′ Λ′ Aˆτ ](Sb − Ss ).

We can evaluate a claim on a stream dt = Sd xt by ∞ X

τ =0

ˆt β τ St,τ Mct+τ · dt+τ = E = x′t

∞ X

τ =0 ∞ X

β τ Mct+τ · dt+τ β τ (Aˆ′ )τ Sd′ Mc Aˆτ xt

τ =0

+

β 1−β

trace[Sd′ Mc

∞ X

β τ Aˆτ Cˆ Cˆ ′ (A′ )τ ].

τ =0

4 These calculations are implemented in the program assetxq.m.

376

Recursive Risk Sensitive Control

16.7. Production Economies The preceding formulas also apply to asset pricing in production economies, with the understanding that the matrix Ω now corresponds to the ‘quadratic-form’ matrix in the value function for the social planning problem.

16.8. Risk-Sensitive Investment under Uncertainty For sake of illustration, we now consider a one consumption good, one capital good production economy that is a version of a Brock-Mirman stochastic growth model with adjustment costs. This delivers a linear-quadratic version of Lucas and Prescott’s (1971) theory of investment under uncertainty. Consumption and investment satisfy: ct + it = γkt−1 + dt (16.8.1) where the capital stock kt evolves according to: kt = δk kt−1 + it

(16.8.2)

and {dt } is an exogenous endowment process. Labor input is required to adjust the capital stock, reflected in a quadratic adjustment cost in the preferences of the fictitious social planner. Also, there is an exogenous ideal consumption level process {b1t }. The time t contribution to preferences is: −(ct − b1t )2 − φ2 it 2 = −(γkt−1 + dt − it − b1t )2 − φ2 it 2 which is quadratic in the control it , the endogenous state variable kt−1 and the exogenous states b1t and dt . We begin with a parameter specification that implies no investment in a deterministic steady state, an unrealistic but useful starting point that we adopt to compare two alternative ‘explanations’ for investment. One is that altering the risk parameter σ induces a precautionary savings motive into the preference ordering. The other is that capital is more productive than in the benchmark economy, increasing the physical return to investment. Benchmark Economy with No Steady-State Investment We set β = 1/1.05, γ = .1, δk = .95, φ = .5. With this parameter configuration and constant values of dt and b1t , the model implies steady state values

Risk-Sensitive Investment under Uncertainty

377

of zero for capital and investment. This follows from the fact that γ + δk = β −1 , which equates the physical rate of return to investment net of adjustment costs to the subjective rate of time discount. Since adjustment costs are also present, reflected by a positive value of φ , physical investment in new capital becomes ‘unattractive.’ As a consequence, asymptotically there are no savings in the deterministic version of this model. To see this, the Euler equation for capital is given by E[(1 − L−1 )(1 − β −1 L)kt + φ2 (1 − βδk L−1 )(1 − δk L)kt | Jt ] = E[(1 − L)(b1t+1 − dt+1 ) | Jt ]

(16.8.3)

where L is the lag operator. For future reference we have permitted there to be uncertainty in the forcing processes for {b1t } and {dt }. The solution to this stochastic difference equation has representation: kt = λkt−1 + ψ[(dt − b1t ) − (1 − βλ)

∞ X j=0

(βλ)j E(dt+j − b1t+j | Jt )].

(16.8.4)

where 0 < λ < 1, and 0 < ψ . In contrast to the familiar permanent income model in which λ and ψ are one, the presence of adjustment costs lowers λ . Notice that the term multiplying ψ is the difference between dt − b1t and a geometric average of current and future values of dt − b1t . This simple link to the forcing processes is a result of the γ + δk = β −1 restriction. For a model with b1t and dt constant, the term multiplying ψ will be zero and the capital stock sequence will converge to zero. An Economy with Risk Sensitivity As a precursor to illustrating investment induced by risk sensitivity, we now introduce a specific model of uncertainty in the endowment and preference shock processes. Let an exogenous state vector process {zt } follow a first-order vector autoregression: zt+1 = A22 zt + C2 wt+1 , and suppose that b1t = Sb zt and dt = Sd zt , so that both the endowment and the preference shock are linear functions of exogenous state vector. Construct 0 0 1 0 0 (16.8.5) A22 = 0 .8 0 C2 = 1 0 , 0 .1 0 0 .5

378

Recursive Risk Sensitive Control

and set Sb = [ 30 0 1 ] , Sd = [5 1 0]. We initialize the first component of zt to be one, which replicates itself over time. Consequently, the preference shock process has mean 30 and the endowment shock process has mean 5. For the σ = 0 version of this economy, ordinary certainty equivalence applies. Since {dt } and {b1t } are asymptotically stationary, so are the endogenous processes for consumption, capital, and investment. It again follows from (16.8.4) that investment and capital both have mean zero in the stochastic steady state, and consumption has a mean equal to the mean of the endowment shock dt . Because of certainty equivalence, no precautionary savings occur in this model. When σ is less than zero, ordinary certainty equivalence no longer holds, and mean investment is positive in a stochastic steady state. Table 2.1 reports the means and standard deviations computed with respect to both the objective probability distribution, and the appropriate distorted distribution for σ = −.005. TABLE 2.1: Means and Standard Deviations for the Adjustment Cost Economy with σ = −.005 True Process

c1t it b1t dt

Pessimistic Process

mean

standard deviation

mean

standard deviation

9.83 4.83 30.00 5.00

1.37 1.22 .12 1.67

2.07 0.00 30.01 2.07

1.57 1.32 .12 1.71

The mean of consumption is now notably higher than the mean endowment and the mean of investment is positive. In effect, there is a precautionary savings motive at work here. A version of the pessimistic certainty equivalence principle derived by Jacobson (1973) and Whittle (1981) applies here. Thus, an alternative way to obtain the same solution is to endow the fictitious social planner with a pessimistic view of the world. The stochastic difference equation (16.8.3) and its solution (16.8.4) still apply except that the distorted expectation operator, call ˆ , is used in place of E. As originally suggested by Jacobson (1973), we can it E

Risk-Sensitive Investment under Uncertainty

379

imagine there being a second agent, say distinct from the social planner, that picks future values of the shocks in a pessimistic fashion. The degree of pessimism is governed by the value of σ , and the difference equations associated with these pessimistic forecasts for b1t+j and dt+j used for the time t decisions are given by: (1 − β −1 L)kt+j − dt+j + b1t+j + (100/σ)(1 − .5βL−1 )(1 − .5L)(b1t+j − 30)] = 0 (16.8.6)

−(1 − β −1 L)kt+j + dt+j − b1t+j + (1/σ)(1 − .8βL−1 )(1 − .8L)(dt+j − 5)] = 0. for j = 1, 2, . . . , where kt−1 , b1t and dt are given initial conditions. In effect, we can think of these as the Euler equations for the second agent. Notice that the capital stock enters these difference equations, and hence the pessimistic forecasts of future values of dt and b1t depend on kt−1 . Certainty equivalence works here because the optimal choice for kt can be obtained by (i) shifting (16.8.3) forward j − 1 periods for j = 1, 2, . . . and eliminating the conditional expectation operator; (ii) combining the resulting difference equation with (16.8.6); and (iii) solving the composite system of difference equations for kt , b1t+1 , dt+1 as a function of the state vector kt−1 , b1t and dt , imposing the appropriate terminal conditions. The composite solution gives the evolution equation for the pessimistic forecasts of b1t and dt , and the solution for kt gives the optimal decision rule for kt . This latter equation, when combined with the actual law of motion for b1t and dt , governs the evolution of the optimal capital stock process. While λ is .9852 for this economy, once this forecasting dependence is incorporated, the feedback of kt onto kt−1 drops slightly to .9817. A more dramatic implication is that there is now a positive constant term (1.766) in the decision rule for capital, whereas in the σ = 0 economy the constant term is zero. Table 2.1 also reports the pessimistic means and standard deviations for forcing processes and for consumption and investment. Notice that the perceived means for dt and c1t are considerably less than their true means. The mean of investment for the perceived process is again zero because of the applicability of (16.8.4) with distorted expectations. Hence the perceived long run average of the capital stock is not altered by changing expectations operators. Not surprisingly, the perceived standard deviations are larger than the true ones.

380

Recursive Risk Sensitive Control

An Approximating Economy with σ = 0 Next we show that we can mimic the quantity implications of the preceding model by setting σ = 0 and modifying some of its parameters. The parameters that we alter are the mean of b1t , the mean of dt and the productivity parameter for capital γ . In light of our previous discussion, to make average investment positive it is necessary to relax the restriction that δk + γ = β −1 by making capital more productive. We used the Kullback-Leibler (1951) information criterion as a device for picking the three parameters because of its well known link to maximum likelihood estimation (e.g., see Akaike (1973), Ljung (1978), and White (1982)). 5 The information criterion was constructed in the same manner as in Hansen and Sargent (1993) using the consumption and investment processes implied by the original model and the approximating σ = 0 economy. The parameters that make the approximate (σ = 0) model as close as possible to the original model are .1032 for the productivity parameter γ and 47.0522 and 4.6812 for the means of b1t and dt . In addition to increasing γ to compensate for setting σ = 0, we were led to increase substantially the mean of b1t and decrease slightly the mean of dt . It turns out that these three parameter adjustments are sufficient to make the quantity implications for the two models to be extremely close. This is shown in Figure 16.8.1 which displays the ratios of the spectral densities of consumption and investment, respectively, for the approximating economy to the corresponding spectral densities for the original economy. Departures from 5 Let each model be a member of our class of models, with parameters of the first model being denoted by a vector δ and those of the second model being denoted by α . For the first model, let the mean vector for the observables be ν(δ) , and the spectral density matrix be S¯y (ω, δ) . For the second model let the spectral density matrix be Sy (ω, α) , and the mean vector be µ(α) . The parameters α that make the second model as close as possible to the first are those that minimize the criterion

n

−

1 2π

−

1 2π

Z

π

−π Z π −π

log det Sy (ω, α) dω trace Sy (ω, α)−1 S¯y (ω, δ) dω

− [ν − µ(α)] Sy (0, α)−1 [ν − µ(α)]′

o

.

Risk-Sensitive Investment under Uncertainty

381

unity are small, and confined to very low frequencies. Departures of this small magnitude and frequency-location would be virtually impossible to detect, via say a likelihood ratio test, without an extremely large time series sample. The adjustments in the means of b1t and dt are sufficient to make the means for consumption and investment the same for both economies. Hence, from the standpoint of data on consumption and investment, the precautionary savings version of the model is (almost) observationally equivalent to a specification in which savings are induced by making capital more productive.

1.3

1.25

1.2

1.15

1.1

1.05

1

0.95 0

0.5

1

1.5

2

2.5

3

3.5

Figure 16.8.1: Ratio of spectral density of consumption and investment (dotted line) in the approximating economy to the spectral density of consumption and investment in the true economy.

382

Recursive Risk Sensitive Control

16.9. Equilibrium Prices in the Adjustment Cost Economies We now return to the two adjustment cost economies with their competing explanations for investment. While the quantities for the σ = 0 ‘approximating’ economy are, by design, close to those for the ‘true’ risk-sensitive economy, behaviors of asset prices and rates of return differ markedly. To illustrate this phenomenon, we first consider the implications for the two components of equilibrium wealth: the value of the endowment process from the current period forward (a Lucas tree), and the value of the existing capital stock. Figures 16.9.1.a and 16.9.1.b depict realizations of these two components to wealth for a common realization of the vector of Gaussian noises driving the composite preference and endowment shock process. The value of the endowment process is always lower and the value of the capital stock higher for the original economy than for the approximating σ = 0 economy. Hence a bigger fraction of wealth is held in the form of capital in the original (risk-adjusted) economy. This is true even though the mean of the endowment process is higher for the original economy. The mean gross returns are (1.0444, 1.0786) for capital and the endowment process, respectively, in the true economy; and (1.0500, 1.0504) in the approximating economy. 6 These phenomena trace to the fact that the capital stock is a less risky investment than the endowment and that less risky investments are more highly valued in the original (risk-adjusted) economy than in the approximating economy. Figure 16.9.2 shows probability densities for risk-free gross returns for both economies. Notice that the density for the approximating economy is centered around β −1 , and the modal value is higher than for the original economy with a risk adjustment. The density for the original (risk-adjusted) economy displays more dispersion than its counterpart for the approximating economy, reflecting in part the smaller mean of b1t . Consider next the market prices of risk for the two economies. The probability densities for these prices are reported in Figures 16.9.3.a and 16.9.3.b. The market price of risk is considerably higher in the risk-sensitive (σ = −.005) economy. Finally, Figures 16.9.4.a and 16.9.4.b report spectral densities for the logarithmic one-period returns to holding a claim to the endowment process and 6 These means were computed from simulations of length 100,000.

Equilibrium Prices in the Adjustment Cost Economies

383

to holding capital, respectively. Both returns are more variable in the original (risk-adjusted) economy than in the approximating one. Moreover, the low frequency dip in the spectral density for the return to holding the endowment is substantially more pronounced. Returning to the decomposition of equilibrium wealth described earlier, it is clear from Figures 16.9.4.a and 16.9.4.b that holding capital is much less risky than holding the endowment for both economies. It is also evident from Figure 16.9.2 that a riskless security is more valued (commands a lower equilibrium rate of return) in the original economy. This apparently underlies the fact that the capital stock is more highly valued in the original (σ = −.005) economy as depicted in Figure 16.9.1a. Moreover, as illustrated in Figures 16.9.3.a and 16.9.3.b, the market prices of risk tend to be higher in the original economy, so that the equilibrium risk adjustments are more pronounced. This helps to explain why equilibrium values for claims to the endowment process are lower for the original economy as depicted in Figure 16.9.1.a even though the mean of the endowment process is higher.

125

120

110 120 100 115

90

80

110

70 105 60

50 0

10

20

30

40

50

60

70

80

90

100

Fig. 16.9.1.a. Realizations of price of Lucas tree in true economy ( σ = −.005 ) and approximating ( σ = 0 ) economy (dashed line).

100 0

10

20

30

40

50

60

70

80

90

100

Fig. 16.9.1.b. Realizations of value of capital stock in true economy( σ = −.005 ) and approximating ( σ = 0 ) economy (dashed line).

384

Recursive Risk Sensitive Control

200 180

approximation

160 140 120 100 true

80 60 40 20 0 1.02

1.025

1.03

1.035

1.04

1.045

1.05

1.055

1.06

Figure 16.9.2: Figure 5.3 Densities of risk-free interest rate for true and approximating economies.

1200

14

12

1000

10 800 8 600 6 400 4 200

2

0 0.35

0.4

0.45

0.5

0.55

0.6

Fig. 16.9.3.a. Density for market price of risk for true adjustment cost true economy.

0 0.01

0.01

0.011

0.011

0.012

0.012

0.013

0.013

Fig. 16.9.3.b. Density for market price of risk for approximating adjustment cost economy.

Equilibrium Prices in the Adjustment Cost Economies

385

10 -3

10 -2 true

true 10 -4

approximating approximating

10 -5

10 -3

0

0.5

1

1.5

2

2.5

3

3.5

Fig. 16.9.4.a. Spectral densities of windsorized logarithmic returns to endowment streams for true and approximating adjustment costs economies.

10 -6

0

0.5

1

1.5

2

2.5

3

3.5

Fig. 16.9.4.b. Spectral densities of logarithmic returns on capital in true and approximating adjustment costs economies.

Chapter 17 Periodic Models of Seasonality

17.1. Introduction Until now, each of the matrices defining the preferences, technology and information flows has been specified to be constant over time. In this chapter, we relax this assumption, and let the matrices be strictly periodic functions of time. Our interest is to apply and extend an idea of Denise Osborn (1988) and Richard Todd (1983, 1990) to arrive at a particular model of “seasonality”. Seasonality can be characterized in terms of the spectral density of a variable. A variable is said to “have a seasonal” if its spectral density displays peaks at or in the vicinity of the frequencies commonly associated with the seasons of the year, e.g., every twelve months for monthly data, every four quarters for quarterly data. Within one of our equilibria, it is possible to think of three ways of modelling seasonality. The first two ways can be represented within the time-invariant setup of our previous chapters, while the third way requires following Todd and departing from the assumption that the matrices that define our economies are time invariant. The first model of seasonality is created by specifying the matrices [A22 , C22 , Ub , Ud ] that determine the information structure in the economy. We can exogenously inject a seasonal preference shock into the model by specifying [A22 , Ub ] in such a way that components of the shock process bt have seasonals. Similarly, we can specify [A22 , Ud ] so that components of the endowment shock process dt have seasonals. The seasonality of these exogenous processes will be transmitted to the prices and quantities determined in equilibrium. The way in which this seasonality is transmitted can be subtle, determined as it is by the restrictions across the parameters of the {bt , dt } processes and the price and quantity processes that are determined by the equilibrium. 1 1 Sargent [1976, 1987,chap XI] described some of the ways in which the cross equation restrictions of linear rational expectations models determine the kind of seasonality in endogenous variables that is induced by imposing seasonality in the variables that agents within a model are implicitly forecasting.

– 387 –

388

Periodic Models of Seasonality

The second model of seasonality is created by specifying the matrices [∆h , Θh , Λ, Π] that determine preferences and the matrices [Φc , Φi , Φc , Γ, ∆k , Θk ] that determine the technology so that they make prices and quantities display seasonality even when the preference shocks bt and the endowment shocks dt do not display any seasonality. Seasonality can come either from the technology side or from the preference side. Notice that in the first kind of model the source of seasonality is imposed exogenously, while in this second kind of model the idea is that preferences and technology are such that the equilibrium of the economy creates a “propagation mechanism” that converts nonseasonal impulses into seasonal responses in prices and quantities. This chapter is devoted to studying a third model of seasonality, by following Todd. We now specify an economy in terms of matrices that are periodic functions of time. This specification captures the idea, for example, that the technology is different in Winter than it is in Spring. You will get less corn if you plant in Minnesota in January than if you plant in May. As we shall see, this model of seasonality has properties that contrast in interesting ways to the other two models of seasonality.

17.2. A Periodic Economy The social planner now faces the problem of maximizing −.5

∞ X t=0

subject to

β t (st − bt ) · (st − bt ) + lt2

(17.2.1)

Φc,s(t) ct + Φi,s(t) it + Φg,s(t) gt = Γs(t) kt−1 + dt kt = ∆k,s(t) kt−1 + Θk,s(t) it ht = ∆h,s(t) ht−1 + Θh,s(t) ct st = Λs(t) ht−1 + Πs(t) ct zt+1 = A22,s(t) zt + C22,s(t) wt+1 bt = Ub zt dt = Ud zt

(17.2.2)

A Periodic Economy

389

In (17.2.2), s(t) is a periodic function that assigns integers to integers. In particular, s : (. . . , −1, 0, 1, . . .) → [1, 2, . . . , p] s(t + p) = s(t) ∀ t

(17.2.3)

s(t) = t for t = 1, 2, . . . , p.

A consequence of (17.2.3) is that the constraints in (17.2.2) can be represented in the form Φc,j cp·t+j + Φi,j ip·t+j + Φg,j gp·t+j = Γj kp·t+j+1 + dp·t+j kp·t+j = ∆k,j kp·t+j+1 + Θk,j ip·t+j hp·t+j = ∆h,j hp·t+j+1 + Θh,j cp·t+j sp·t+j = Λj hp·t+j+1 + P e¯j cp·t+j

(17.2.4)

zp·t+j+1 = A22,j zp·t+j + C22,j wp·t+j bp·t+j = Ub zp·t+j dp·t+j = Ud zp·t+j where t = 0, 1, 2, . . . , and j = 1, 2, . . . , p. Notice for t = 0, as j goes from 1 to p , that p · t + j goes from 1 to p ; for t = 1, as j goes from 1 to p , that p · t + j goes from p + 1 to 2p , and so on. Thus, (17.2.4) describes a setting in which the matrices that represent preferences and the technology are periodic with period p . The social planning problem can be expressed in the form of a periodic optimal linear regulator problem. The social planner chooses a sequence of functions expressing ut as functions of xt , for all t ≥ 0, to maximize −E

∞ X t=0

β t {x′t Rs(t) xt + u′t Qs(t) ut + 2u′t Ws(t) xt }

(17.2.5)

subject to the constraints xt+1 = As(t) xt + Bs(t) + Cs(t) wt+1

(17.2.6)

′ , zt ]. In (17.2.5), (17.2.6), the matrices [Rs(t) , Qs(t) , Ws(t) , where x′t = [h′t−1 , kt−1 As(t) , Bs(t) , Cs(t) ] are the same functions of the matrices [Φc,s(t) , Φi,s(t) , Φg,s(t) , Γs(t) , ∆k,s(t) , Θk,s(t) , ∆h,s(t) , Θh,s(t) , Λs(t) , Πs(t) , A22,s(t) , C22,s(t) , Ub , Ud ] that the

390

Periodic Models of Seasonality

matrices [R, Q, W, A, B, C] are of the matrices [Φc , Φi , Φc , Γ, ∆k , Θk , ∆h , Θh , Λ , Π, A22 , C22 , Ub , Ud ] in the constant coefficient case. These functions were described in chapter 3. The Bellman equations for this problem are Vt (xt ) = max{x′t Rs(t) xt + u′t Qs(t) ut + 2u′t Ws(t) xt ut

(17.2.7)

+ βEt Vt+1 (xt )} where the maximization is subject to xt+1 = As(t) xt + Bs(t) + Cs(t) wt+1 .

(13.6)

In (17.2.7), Vt (xt ) is defined as the optimal value of the problem starting from state xt at time t . For the periodic optimal linear regulator problem, the optimal value function is quadratic but time varying: Vt (xt ) = x′t Pt xt + ρt ,

(17.2.8)

where the n × n matrix Pt satisfies the matrix Riccati difference equation ′ Pt = Rs(t) + βA′s(t) Pt+1 As(t) − (βA′s(t) Pt+1 Bs(t) + Ws(t) )

′ ′ Pt+1 As(t) + Ws(t) ), Pt+1 Bs(t) )−1 (βBs(t) × (Qs(t) + βBs(t)

(17.2.9)

while the scalar ρt satisfies ′ ). ρt = βρt+1 + β trace (Pt+1 Cs(t) Cs(t)

(17.2.10)

Now think of solving Bellman’s equation by iterating backwards on (17.2.9), (17.2.10), starting from some terminal values for P and ρ. Because the matrices [Rs(t) , Qs(t) , Ws(t) , As(t) , Bs(t) ] are all functions of time when p ≥ 2, it is too much to hope that {Pt , ρt} will converge in these iterations as t → −∞ to objects that are independent of time. What is reasonable to hope for, and what will indeed obtain under the assumptions made in our setup, is that iterations on (17.2.9) and (17.2.10) will each produce p convergent subsequences. In particular, backwards iterations on (17.2.9) and (17.2.10) will converge to a sequence that oscillates periodically among p value functions associated with

A Periodic Economy

391

the p seasons of the year. Thus, after many iterations, we will eventually have Vt (xt ) = Vs(t) (xt ), where Vs(t) (xt ) = x′t Ps(t) xt + ρs(t)

(17.2.11)

We can also represent these value functions as Vj (xp·t+j ) = x′p·t+j Pj xp·t+j + ρj ,

(17.2.12)

where t = 0, 1, 2, . . . and j = [1, 2, . . . , p]. Equation (17.2.12) summarizes the outcome that there are p value functions, one for each of the p seasons of the year. The optimal decision rules can be represented as ut = −Fs(t) xt

(17.2.13)

where ′ ′ Fs(t) = −(Qs(t) + βBs(t) Ps(t+1) Bs(t) )−1 βBs(t) Ps(t+1) As(t) .

(17.2.14)

The optimal decision rules are thus periodic with period p . Substituting (17.2.13) into the law of motion (17.2.6) gives the following “closed loop” representation of the solution of the social planning problem:

or

xt+1 = As(t) − Bs(t) Fs(t) xt + Cs(t) wt+1

(17.2.15)

xt+1 = Aos(t) xt + Cs(t) wt+1

(17.2.16)

where Aos(t) = As(t) − Bs(t) Fs(t) . We can also represent (17.2.16) in the form xp·t+j+1 = Aoj xp·t+j + Cj wp·t+j+1

(17.2.17)

for t = 0, 1, 2, . . . and j = [1, 2, . . . , p]. Thus the laws of motion are periodic with a periodicity p that is inherited from that of the matrices specifying preferences, technology, and information flows. The matrices [Aoj , Pj ] for j ∈ [1, 2, . . . , p] can be used to construct the quantities and prices associated with the equilibrium of our model. Formulas for the matrices determining our equilibrium, namely the M and S matrices,

392

Periodic Models of Seasonality

are given by the very same formulas described in chapters 3 and 5, with the proviso that in the periodic case s(t) or j subscripts appear on all objects in those formulas. Thus, we have that the quantities determined in our equilibrium are given by ht = Sh,s(t) xt dt = Sd,s(t) xt kt = Sk,s(t) xt

ct = Sc,s(t) xt

kt−1 = Sk1,s(t) xt it = Si,s(t) xt

gt = Sg,s(t) xt

(17.2.18)

st = Ss,s(t) xt

bt = Sb,s(t) xt where

Sh,s(t) Sk,s(t)

Ao11,s(t) = Ao12,s(t)

Sk1,s(t) = [0 I 0]

Si,s(t) = −Fs(t)

Sd,s(t) = [0 0 Ud ] Sb,s(t) = [0 0 Ub ]

(17.2.19)

Sc,s(t) = Uc,s(t) [Φc,s(t) Φg,s(t) ]−1 [−Φi,s(t) Si,s(t) + Γs(t) Sk1,s(t) + Sd,s(t) ] Sg,s(t) = Ug,s(t) [Φc,s(t) Φg,s(t) ]−1 [−Φi,s(t) Si,s(t) + Γs(t) Sk1,s(t) + Sd,s(t) ] Ss,s(t) = Λs(t) [I 0 0] + Πs(t) Sc,s(t) The Lagrange multipliers associated with the social planning problem are determined by the following counterparts of the formulas that we described in chapters 3 and 5: Mk,s(t) = 2β[0 I 0] Ps(t) Aos(t) Mh,s(t) = 2β[I 0 0] Ps(t) Aos(t) Ms,s(t) = Sb,s(t) − Ss,s(t) ′ −1 ′ ′ Φc,s(t) Θh,s(t) Mh,s(t) + Π′s(t) Ms,s(t) Md,s(t) = Φ′g,s(t) −Sg,s(t) ′ ′ Mc,s(t) = Θh,s(t) Mh,s(t) + Πs(t) Ms,s(t) Mi,s(t) = Θk,s(t) Mk,s(t)

(17.2.20)

Asset Pricing

393

Formulas for the equilibrium price system can be stated in terms of the objects defined in (17.2.20):

ptt′ = Mc,s(t′ ) xt′ /[¯ ej Mc,s(t) xt ] wtt′ = |Sg,s(t′ ) xt′ |/[¯ ej Mc,s(t) xt ]

rtt′ = Γ′s(t) Md,s(t′ ) xt′ /[¯ ej Mc,s(t) xt ] qtt′ = Mi,s(t′ ) xt′ /[¯ ej Mc,s(t) xt ]

(17.2.21)

αtt′ = Md,s(t′ ) xt′ /[¯ ej Mc,s(t) xt ] vt = [Γs(t′ ) Md,s(t′ ) + ∆′k,s(t′ ) ]xt′ /[¯ ej Mc,s(t) xt ] These formulas give the time t price system for pricing goods to be delivered at all t′ ≥ t .

17.3. Asset Pricing With the above formulas in hand, we can derive formulas for pricing assets. These formulas generalize those described in chapter 5 to the case in which the economy is strictly periodic. We begin by pricing an asset that entitles its owner to a stream of returns in the form of a vector of consumption goods described by yt = Ua,s(t) xt , where Ua,s(t) is a periodic sequence of matrices. We let at denote the price of this asset at time t . By the same reasoning applied in chapter 5, at satisfies at = Et

∞ X

ej Mc,s(t) xt ], β h x′t+h Za,s(t+h) xt+h /[¯

(17.3.1)

h=0

′ where Zaj = Ua,j Mc,j . We shall show that (17.2.5) can be represented as

at = [x′t µa,s(t) xt + σa,s(t) ]/[¯ ej Mc,s(t) xt ],

where µa,s(t) and σa,s(t) satisfy

(17.3.2)

394

Periodic Models of Seasonality

o 2 o′ o′ o o µa,1 = Za,1 + βAo′ 1 Za,2 A1 + β A1 A2 Za,3 A2 A1 + · · ·

o′ o′ o′ o o o o + β p−1 Ao′ 1 A2 · · · Ap−2 Ap−1 Za,p Ap−1 Ap−2 · · · A2 A1 o′ o′ o′ o o o o + β p Ao′ 1 A2 · · · Ap−1 Ap µa,1 Ap Ap−1 · · · A2 A1

o µa,p = Za,p + βAo′ p µa,1 Ap

(17.3.3)

o′ µa,p−1 = Za,p−1 + βAp−1 µa,p Aop−1 .. . o µa,2 = Za,2 + βAo′ 2 µa,3 Ap−2

and

σa,1 = β trace(µa,2 C1 C1′ ) + βσa,2 σa,2 = β trace(µa,3 C2 C2′ ) + βσa,3 .. .

(17.3.4)

σa,p = β trace(µa,1 Cp Cp′ ) + βσa,1 The matrix µa,1 can be computed from the first equation of (17.3.3) by using a doubling algorithm that is described in chapter 8. Then the remaining equations of (17.3.3) can be used to compute the remaining µa,j ’s. Given the µa,j ’s, (17.3.4) is a system of p equations that can be solved for the p σa,j ’s. To verify (17.3.3), (17.3.4), we can proceed as follows. Let the numerator of (17.3.1), (17.3.2) be denoted a ˜ t = Et =

∞ X

β h x′t+h Za,s(t+h) xt+h

h=0 ′ xt µa,s(t) xt

(17.3.5)

+ σa,s(t)

Recall the equilibrium transition laws (17.2.16): xt+1 = Aos(t) xt + Cs(t) wt+1 . Evidently, (17.3.5) and (17.2.16) imply that a ˜t = x′t Za,s(t) xt + β Et a ˜t+1 or

x′t µa,s(t) xt + σa,s(t) = x′t Za,s(t) xt +βEt (Aos(t) xt + Cs(t) wt+1 )′ µa,s(t+1) (Aos(t) xt + Cs(t) wt+1 ) + βσa,s(t+1)

(13.19)

Prediction Theory

395

The above equation implies that o µa,s(t) = Za,s(t) + βAo′ s(t) µa,s(t+1) As(t)

(17.3.6)

′ σa,s(t) = β σa,s(t+1) + β trace (µa,s(t+1) Cs(t) Cs(t) )

(17.3.7)

Equation (17.3.4) is equivalent with (17.3.7). The first equation of (17.3.3) is the result of recursions on (17.3.6) starting from s(t) = 1, while the remaining equations of (17.3.3) are simply (17.3.6) for s(t) = 2, 3, . . . , p. This completes the verification of (17.3.3), (17.3.4). We shall give a formula for the term structure of interest rates after we have described the prediction theory associated with (17.2.16).

17.4. Prediction Theory For a model with period p ≥ 2, there are two natural alternative ways of specifying the information sets upon which means, covariances, and linear least squares predictions are conditioned. First, we can calculate moments and forecasts by conditioning on the season. This amounts to computing different moments and different forecasting formulas for each of the p seasons. In the appendix to this chapter, we formally describe a sigma algebra, which we denote I p , that contains the information that corresponds to conditioning on the season. Second, we can calculate moments and forecasts by disregarding information about the season, which amounts to averaging data across seasons in a particular way. In the appendix, we formally describe a sigma algebra, denoted I , which corresponds to not conditioning on the season. In this section, we describe parts of the prediction theory for our periodic models that correspond to conditioning on the season. In this section,the notation Et ( · ) denotes a mathematical expectation conditioned on xt , under the assumption that we are also conditioning on the information in I p . We are assuming that the fictitious social planner uses this information to compute all relevant prices. Recursions on (17.2.16) can be used to deduce the linear least squares predictions of the state vector xt . There are p different sets of formulas for the j -step ahead predictions of xt+k conditioned on xt , one for each season of the

396

Periodic Models of Seasonality

year. Recursions on (17.2.16) lead directly to xt+k = Aos(t+k−1) Aos(t+k−2) · · · Aos(t) xt

+ Aos(t+k−1) Aos(t+k−2) · · · Aos(t+1) Cs(t) wt+1 + · · ·

(17.4.1)

+ Aos(t+k−1) Cs(t+k−2) wt+k−1 + Cs(t+k−1) wt+k Equation (17.4.1) implies Et xt+k = Aos(t+k−1) Aos(t+k−2) · · · Aos(t) xt

(17.4.2)

and E(xt+k − Et xt+k )(xt+k − Et xt+k )′ ≡ Σk,s(t)

′ = Aos(t+k−1) Aos(t+k−2) · · · Aos(t+1) Cs(t) Cs(t) Ao′ s(t+1) · · · o′ Ao′ s(t+k−2) As(t+k−1)

+

(17.4.3)

′ · · · + Aos(t+k−1) Cs(t+k−2) Cs(t+k−2) Ao′ s(t+k−1) ′ + Cs(t+k−1) Cs(t+k−1)

Recursive versions of (17.4.2) and (17.4.3) are available. Equation (17.4.2) implies Et xt+k = Aos(t+k−1) Et xt+k−1 . Equation (17.4.3) implies ′ Σk,s(t) = Aos(t+k−1) Σk−1,s(t) Ao′ s(t+k−1) + Cs(t+k−1) Cs(t+k−1) .

The prediction formulas (17.4.2), (17.4.3) are evidently predicated on the assumption that we know the matrices [Aoj , Cj ] for j = [1, . . . , p]. They also assume that xt is in the information set of the forecaster. Later in this chapter, we shall briefly describe how the Kalman filter can be used to compute the linear least squares forecast of yt , conditioned only on the history of observed y ′ s , and also on I p . We shall also describe a different theory of prediction, which assumes that we do not know the values of [Aoj , Cj ], and that we cannot condition on the season, so that all that we possess is a time invariant representation for the {xt , yt } process.

Conditional Covariograms

397

17.5. The Term Structure of Interest Rates In light of formula (17.4.2), the same logic that led to formula (5.65) for the reciprocal of the risk-free interest rate on j -period loans, Rjt , now leads to the following formula: Rjt = β j e¯1 Mc,s(t+j) Aos(t+j−1) Aos(t+j−2) · · · Aos(t) xt /[¯ ej Mc,s(t) xt ]

(17.5.1)

This formula gives the price at time t of a sure claim on the first consumption good j periods ahead.

17.6. Conditional Covariograms In this section, we present formulas for the covariance function of x and y , conditioned on season, i.e., conditioned on I p . The conditional covariogram of {xt } can be expressed in terms of the conditional contemporaneous covariance function cx,t (0) = Ext x′t |I p via the formulas cx,t (−k) ≡ Ext x′t+k |I p

o′ o′ o′ = Ext x′t |I p Ao′ s(t) As(t+1) · · · As(t+k−2) As(t+k−1) , k ≥ 1

or

o′ cx,t (−k) = cx,t (0)Ao′ s(t) As(t+1) · · ·

o′ Ao′ s(t+k−2) As(t+k−1) , k ≥ 1.

(17.6.1)

To solve for the matrices cx,t (0), we can solve the equations ′ Ext+1 x′t+1 |I p = Aos(t) Ext x′t |I p Ao′ s(t) + Cs(t) Cs(t)

or ′ cx,t+1 (0) = Aos(t) cx,t (0)Ao′ s(t) + Cs(t) Cs(t) .

(17.6.2)

By solving the system formed by (17.6.2) for t = 1, 2, . . . , p, we can determine the p contemporaneous covariance matrices cx,1 (0), cx,2 (0), . . . , cx,p (0). Here is a fast way of solving this system. Iterating on (17.6.2) p times yields cx,t+p (0) o′ o′ = Aos(t+p−1) Aos(t+p−2) · · · Aos(t) cx,t (0) Ao′ s(t) · · · As(t+p−2) As(t+p−1)

′ o′ + Aos(t+p+1) Aos(t+p−2) · · · Aos(t+1) Cs(t) Cs(t) Ao′ s(t+1) · · · As(t+p−1)

′ ′ · · · + Aos(t+p−1) Cs(t+p−2) Cs(t+p−2) Ao′ s(t+p−1) + Cs(t+p−1) Cs(t+p−1) .

(17.6.3)

398

Periodic Models of Seasonality

We compute cx,t (0) by setting cx,t+p (0) equal to cx,t (0) in (17.6.3). Equation (17.6.3) is a discrete Lyapunov equation that can be solved by a doubling algorithm that is described in chapter 8. Once (17.6.3) is solved for t = 1 to compute cx,1 (0), (17.6.2) can be used to compute cx,t (0) for t = 2, . . . , p. There is one covariance matrix cx,t (0) for each of the p seasons of the year. ′ Given cx,t (−k) for k ≥ 0, we can compute cy,t+k (−k) = Eyt yt+k |I p by using (17.6.1). We obtain ′ Eyt yt+k | I p = Gs(t) cx,k (−k)G′s(t+k) , k ≥ 0.

(17.6.4)

Although we are starting calendar time at t = 0, cx,t (k) and cy,t (k) are both defined for positive k so long as t ≥ k . For any such t , cx,t (k) = cx,t−k (−k)′ and cy,t (k) = cy,t−k (−k)′ , implying that {cy,t (k)} and {cy,t (k)} are both periodic starting from t = k . For notational convenience, we extend this construction for 0 ≤ t ≤ k by defining cx,t (k) = cx,t+ℓp (k) and cy,t (k) = cy,t+ℓp (k) for any ℓ such that t + ℓp ≥ k . This guarantees that the conditional covariograms are periodic for all values of k .

17.7. The Stacked and Skip-Sampled System The equilibrium has the system of periodic transition laws described in (17.2.16) or (17.2.17). The equilibrium stochastic process for xt is time-varying, albeit in a highly structured way. We have seen that conditional on knowledge of the season, there are p covariograms, and p sets of formulas for linear least squares predictions that apply in the p seasons of the year. Using these formulas requires knowledge of the set of matrices [Aoj , Cj ] for j = [1, . . . , p] that characterize the transition laws (17.2.16). In this section, we describe a time invariant representation that also characterizes the system. We shall use this representation for several purposes. We shall use it to deduce two kinds of impulse response functions or moving average representations that can be defined for periodic models. 2 We shall also use it to compute a population version of a time-invariant vector autoregression for xt . 2 Each of these impulse response functions conditions on knowledge of the season. Later we shall describe yet another moving average representation that does not condition on season.

The Stacked and Skip-Sampled System

399

The p distinct covariograms as described by equations (17.6.1) and (17.6.4) are conditional covariograms, meaning that they are computed by conditioning on the season of the year. Sample counterparts of these conditional covariograms are computed by creating p distinct averages, averaging each over observations p periods apart. Sample covariograms can also be computed ‘unconditionally’, i.e., in a way that ignores the seasonal structure of the transition laws. This amounts to computing sample moments in the standard way, simply by averagPT ′ ing over adjacent observations, namely, as T −1 t=1 yt yt−j . For a periodic model, such averages will converge as T → ∞ , and they will converge to well defined functions of the parameters of the model. In particular, as T → ∞ , PT ′ T −1 t=1 yt yt−k would converge to an average of the p covariograms, namely, −1 p [cy,1 (k) + cy,2 (k) + . . . + cy,p (k)]. The convergence of these sample autocovariances assures the existence of a time invariant vector autoregressive representation for yt . We begin by defining for t = 0, 1, . . . the vector Xt′ = [x′p·t−p+1 , x′p·t−p+2 , . . . , x′p·t ]′ .

(17.7.1)

′ Xt+1 = [x′p·t+1 , x′p·t+2 , . . . , x′p·t+p ].

(17.7.2)

Evidently, we have

To verify this, substitute (t + 1) for t everywhere that t appears on the right side of (17.7.1). We also define ′ ′ ′ ′ Wt+1 = [wp·t+1 , wp·t+2 , . . . , wp·t+p ].

(17.7.3)

It follows from (17.2.17) that DXt+1 = F Xt + GWt+1 , where

I −Ao1 D= 0 . .. 0

0 I −Ao2 .. . 0 F =

0 0 I .. .

··· ··· ··· .. .

0

· · · −Aop−1 Aop 0

0 0

0 0 0 .. .

(17.7.4) 0 0 0 .. . I

(17.7.5)

(17.7.6)

400

Periodic Models of Seasonality

Cp 0 G = .. .

0 C1 .. .

0 ··· 0 ··· .. . . . .

0 0 .. .

0

0 ···

Cp−1

0

Solving (17.7.4) for Xt+1 gives

ˆ t + CW ˆ t+1 Xt+1 = AX

(17.7.7)

(17.7.8)

′ ′ where Aˆ = D−1 F and Cˆ = D−1 G . We also define the vector Yt′ = [yp·t−p+1 , yp·t−p+2 ,..., ′ yp·t ]. Then we have that

Yt = HXt

where H =

G1 0 .. .

0 G2 .. .

··· ··· .. .

0 0 .. .

(17.7.9)

. Thus we have that {Yt } is governed by the

0 0 · · · Gp time invariant state space system

ˆ t + CW ˆ t+1 Xt+1 = AX

(17.7.10)

Yt = HXt

(17.7.11)

Notice that while {xt , yt } is governed by a time varying linear state space system, the stacked and skip sampled process {Xt , Yt } is governed by a time invariant system. 3 From representation (17.7.10) – (17.7.9), we can use standard formulas to deduce the moving average representation of Yt in terms of Wt Yt =

∞ X

C¯j Wt−j .

(17.7.12)

j=0

The moving average representation (17.7.12) implies the following representation for the components of Yt in terms of the components of Wt : ypt−p+k =

p ∞ X X

C¯j (k, h)wp(t−j)−p+h , k = 1, . . . , p,

(17.7.13)

j=0 h=1

3

In terms of the language introduced in the appendix, because S is of period p, S p is of period one.

The Stacked and Skip-Sampled System

401

where C¯j (k, h) denotes the (k, h)th (m × m) block of C¯j , where m is the dimension of yt . According to representation (17.7.13), there are two distinct concepts of a moving average representation, and p embodiments of each of these concepts. The first concept is a representation of ypt−p+k in terms of current and lagged wt ’s. The response of ypt−p+k to lagged w ’s is evidently given by the sequence 4 ¯ ¯ ¯ ¯ ¯ {dk,v }∞ v=0 = {C0 (k, k), C0 (k, k − 1), . . . C0 (k, 1)C1 (k, p), C1 (k, p − 1), . . . , C¯1 (k, 1), C¯2 (k, p), C¯2 (k, p − 1), . . . , C¯1 (k, 1), . . .}. (17.7.14) In particular, we have from (17.7.13) that ypt−p+k =

∞ X

dk,v wpt−p+k−v .

(17.7.15)

v=0

Notice that there is a different moving average of type (17.7.15) for each season k = 1, . . . , p. The second concept of a moving average is the response of the {yt } process to an innovation wpt−p+k in a particular season k . The response of {yt } to wpt−p+k is evidently given by the sequence ¯ ¯ ¯ {gk,v }∞ v=0 = {C0 (k, k), C0 (k + 1, k), . . . , C0 (p, k), C¯1 (1, k), C¯1 (2, k), . . . , C¯1 (p, k), . . .}.

(17.7.16)

In the special case in which the true periodicity is one, it is straightforward to verify that for any p > 1, the impulse functions constructed from the stacked system (17.7.10) – (17.7.9) satisfy the restrictions: C¯j (1, 1) = C¯j (2, 2) = . . . = C¯j (p − 1, p − 1) = C¯j (p, p) C¯j (2, 1) = C¯j (3, 2) = . . . = C¯j (p, p − 1) = C¯j+1 (1, p)

C¯j (3, 1) = C¯j (4, 2) = . . . = C¯j+1 (1, p − 1) = C¯j+1 (2, p) .. . ¯ Cj (p − 1, 1) = C¯j (p, 2) = . . . = C¯j+1 (p − 2, p − 1) = C¯j (p − 2, p) C¯j (p, 1) = C¯j+1 (1, 2) = . . . = C¯j+1 (p − 2, p − 1) = C¯j+1 (p − 1, p) 4 Notice that by construction C ¯0 (k, j) = 0 for k < j .

(4.14)

402

Periodic Models of Seasonality

Under these restrictions, it follows that gk,v = gj,v for all j, k, for all v dk,v = dj,v for all j, k, for all v dk,v = gk,v for all k,

for all v

Thus, in the case in which the hidden periodicity is truly one, all of the impulse response functions defined in (17.7.15) and (17.7.16) are equal, and are equal to each other. However, when the hidden periodicity is truly some p > 1, there are p distinct impulse response functions {dk,v } of ypt−p+k to lagged w ’s, and p distinct impulse responses {gk,v } of {yt } to wpt−pk , for k = 1, . . . p. In general the {dk,v } are different from one another and from the {gk,v }’s for k = 1, . . . , p. These differences provide a useful way of describing how the operating characteristics of a periodic model with p ≥ 2 differ from a period one model. 5 Later in this chapter, we compute the impulse response functions {dk,v } and {gk,v } for investment for a period 4 version of Hall’s model described above. These impulse response functions are depicted in figures 17.10.1.a and 17.10.1.b. The impulse responses are with respect to the one shock in the model, which is a white noise endowment process. Figure 17.10.2 depicts the impulse responses {dk,v } for k = 1, . . . , 4. Notice that they are smooth, but that they vary across quarters. Figure 17.10.3 shows the impulse response {gk,v } for k = 1, . . . , 4. They vary across quarter k , and have shapes that are jagged, in contrast to the smooth {dk,v }’s. Notice how the amplitude of the oscillations in {dk,v } grows as v increases from v = 0 to v about 30.

5 A MATLAB program simpulse performs these calculations.

Covariances of the Stacked, Skip Sampled Process

403

17.8. Covariances of the Stacked, Skip Sampled Process The stacked, skip-sampled process {Yr } is constructed to have periodicity one. We can compute for k ≥ 1, ′ Cr (−k) ≡ E(Yr Yr+k | Ip) =

cy,pr (−pk) cy,pr+1 (−pk + 1) .. . cy,pr+p−1 (−pk + p − 1)

cy,pr (−pk − 1) cy,pr+1 (−pk) .. . cy,pr+p−1 (−pk + p − 2)

··· ··· .. . ···

cy,pr (−pk − p + 1) cy,pr+1 (−pk − p + 2) . .. . cy,pr+p−1(−pk) (17.8.1)

An implication of the period 1 nature of {Yr } is that Cr (−k) is independent of r . This follows immediately from (17.8.1). In particular, we have for k ≥ 1, C(k) ≡ C0 (−k)

cy,0 (−pk) cy,1 (−pk + 1) = .. . cy,p−1 (−pk + p − 1)

cy,0 (−pk − 1) cy,1 (−pk) .. . cy,p−1 (−pk + p − 2)

... ... ..

.

...

cy,0 (−pk − p + 1) cy,1 (−pk − p + 2) (17.8.2) . .. . cy,p−1 (−pk)

The k = 0 term must be treated separately. It is given by Cr (0) ≡ E(Yr Yr′ | I p ) cy,pr (0) cy,pr (−1)′ = .. .

cy,pr (−p + 1)′

cy,pr (−1) cy,pr+1 (0) .. .

··· ··· .. .

cy,pr+1 (−p + 2)′

···

cy,pr (−p + 1) cy,pr+1 (−p + 2) (17.8.3) , .. . cy,pr+p−1 (0)

which can also be shown to be independent of r . The covariance generating function of the {Yr } process is given by S(z) ≡ C(0) +

∞ X

[C(−k)z −k + C(−k)′ z k ].

(17.8.4)

k=1

It is useful to calculate the covariance generating function S(z) of {Yr } by substituting (17.8.2) – (17.8.3) into (17.8.4). We obtain f s11 (z) sf12 (z) sf13 (z) · · · sf1p (z) sf (z) sf (z) sf (z) · · · sf (z) 21 22 23 2p f s31 (z) sf32 (z) sf33 (z) · · · sf3p (z) (17.8.5) S(z) = .. .. .. .. .. . . . . . f f f sp1 (z) sp2 (z) ··· · · · spp (z)

404

Periodic Models of Seasonality

where

sfj,j+ℓ (z) = cj−1 (−ℓ) +

∞ X

k=1

[cy,j−1 (−pk − ℓ)z −k

(17.8.6)

+ cy,j+ℓ−1 (−pk − ℓ)′ z k ], and where the lower triangular terms of S(z) are obtained from the upper by setting S(z) = S(z −1 )′ for z = e−iω . The hypothesis that {yt } is of period one places restrictions on S(z). Period one of {yt } implies that cy,j (k) = cy,1 (k) for all j . By using this equality in (17.8.6) it can be shown that sf11 (z) = sf22 (z) = · · · = sfpp (z)

sf12 (z) = sf23 (z) = · · · sfp−1,p (z) = z −1 sfp,1 (z)

sf13 (z) = sf24 (z) = · · · = sfp−2,p (z) = z −1 sfp−1,1 (z) = z −1 sfp,2 (z)

(17.8.7)

.. .

zsf1,p (z) = sf2,1 (z) = sf3,2 (z) · · · = sfp,p−1 (z) The first line of equalities in (17.8.7) asserts that the block of matrices along the diagonal of S(z) are equal to each other, and to a folded spectrum of the original unsampled {yt } process.

The Tiao-Grupe Formula

405

17.9. The Tiao-Grupe Formula Define ′ ry,t (−k) = E(yt yt+k | I).

It follows that ry,t (−k) = ry,1 (−k) ≡ ry (−k) for all t . Furthermore, by the law of iterated expectations ry (−k) = ry,t (−k) = E cy,t (−k) | I .

(17.9.1)

It follows from (17.9.1) that ry (−k) can be computed either by computing covariances without skip sampling, or by averaging across covariances that have been computed by skip sampling. That is, ry (−k) = p−1

p X

cy,j (−k),

j=1

and by a law of large numbers N 1 X ′ yt yt+k . N →∞ N t=1

ry (−k) = lim

It is useful to derive Tiao and Grupe’s (1980) formula for the covariance generating function of {yt }, not conditioned on season, as a function of the covariance generating function conditioned on season. Tiao and Grupe’s formula expresses the generating function for the covariances not conditioned on season in terms of the (conditional on season) covariance generating function of the stacked and skip sampled process Yt . 6 We define the generating function for the covariances not conditioned on season to be: sy (z) =

∞ X

r(k)z k

k=−∞

or

sy (z) = p−1

∞ X

k=−∞

zk

p X

(17.9.2) cy,j (k).

j=1

6 Gladysev (1960) states a formula restricting the Cramer representations for Y and y t t that has the same content as the Tiao-Grupe formula.

406

Periodic Models of Seasonality

To compute sy (z), define the operator Q(z) = [I zI . . . z p−1 I],

(17.9.3)

where each of the p identity matrices in (17.9.3) is (n×n). Note that for k ≥ 1, Q(z)C(k)Q(z −1 )′ = [cy,0 (−pk) + cy,0 (−pk + 1)z −1 + · · · + cy,0 (−pk − p + 1)z −p+1

+ cy,1 (−pk + 1)z + cy,1 (−pk) + · · · + cy,1 (−pk − p + 2)z −p+2 .. .

+ cy,p−1 (−pk + p − 1)z p−1 + cy,p−1 (−pk + p − 2)z p−2 + . . . + cy,p−1 (−pk)] (17.9.4a) Notice also that for k = 0, we have

Q(z)C(0)Q(z −1 )′ = [I z I . . . z p−1 I]C(0)

I zI −1 .. . z −p+1 I

= cy,0 (0) + cy,1 (0) + . . . + cy,p−1 (0)

+ z[cy,0 (−1)′ + cy,1 (−1)′ + . . . + cy,p−2 (−1)′ ]

(17.9.4b)

+ z −1 [cy,0 (−1) + cy,1 (−1) + . . . + cy,p−2 (−1)]+ .. . + z p−1 cy,0 (−p + 1)′ + z −p+1 cy,0 (−p + 1). Applying (13.67) to (17.9.2) gives sy (z) = p−1

∞ X

z ph Q(z)C(h)Q(z −1 )′

h=−∞

sy (z) = p−1 Q(z)[

∞ X

z ph C(h)]Q(z −1 )′

(17.9.5)

h=−∞

sy (z) = p−1 Q(z)S(z p )Q(z −1 )′ , where S(z) is the generating function for the {Yr } process, which is defined in (17.7.9) and (17.8.4). Equation (17.9.5) is the Tiao–Grupe formula.

The Tiao-Grupe Formula

407

Equation (17.9.5) shows how the generating function of the {yt } process can be obtained by transforming the generating function of the stacked, skip sampled process {Yr }. Equation (17.9.5) is helpful in displaying the types of fluctuations that will occur in a periodic process {yt }. Suppose that we were to take a realization {yt }Tt=1 of the {yt } process, compute the sample covariances as T 1 X rˆ(k) = yt yt−k (17.9.6) T t=k+1

and the sample spectrum as sˆ e−iωh =

T X

w(k)ˆ r(k)e−iωh j , ωh =

k=−T +1

2πh , h = 1, . . . , T T

(17.9.7)

where w(k) is one of the popular windows. Notice that in computing (17.9.6) and (17.9.7) we are ignoring the hidden periodicity. In large samples, rˆ(k) given by (17.9.6) will converge to r(k) defined in (17.9.1), and sˆ e−iωh will converge to sy e−iωh .

17.9.1. A state space realization of the Tiao-Grupe formulation We now return to representation (17.7.10) – (17.7.9). We will use this representation in conjunction with formula (17.9.3) to get a representation for the generating function sy (z) in terms of the parameters of our economic model. Then we shall describe how to use state space methods to factor this covariance generating function, thereby obtaining a Wold representation for yt . If the eigenvalues of Aˆ are bounded in modulus by unity, 7 then {Xt , Yt } will be asymptotically covariance stationary, with covariance generating matrices SX (z) and SY (z) given by SX (z) =

∞ X

RX (k)z k

k=−∞

or

ˆ −1 Cˆ Cˆ ′ (I − Az ˆ −1 )−1′ SX (z) = (I − Az) 7 This is the condition alluded to in section 1.

(17.9.8)

408

Periodic Models of Seasonality

and SY (z) =

∞ X

RY (k)z k

k=−∞

SY (z) = HSX (z)H

(17.9.9)

′

′ ′ where RX (k) = EXt Xt−k , RY (k) = EYt Yt−k . By substituting SX (z) or SY (z) for S(z) in formula (17.9.2), we can compute the covariance generating function for the process {yt } by averaging across covariograms for different periods. For the yt process under study here, we have 8

sy (z) = Q(z)SY (z p )Q(z −1 )′ (17.9.10)

or ˆ p )−1 Cˆ Cˆ ′ (I − Az ˆ −p )−1′ H ′ Q(z −1 )′ . sy (z) = Q(z)H(I − Az

We now show how to use (17.9.10) to deduce a state-space representation for {yt }. The first step involves recognizing that (17.9.10) is realized by the system ˆ t + CV ˆ t+p Zt+p = AZ

(17.9.11)

yt = Q(L)HZt

where L is the lag operator and {Vt } is a vector white noise with identity contemporaneous covariance matrix. It is convenient to stack (17.9.11) into the first order system Z 0 0 t+p Zt+p−1 I 0 Zt+p−2 0 I = .. .. .. . . . 0 0 Zt+1 Z

... ... ... .. . ...

t+p−1

yt+p−1

0 0 0 .. . I

Aˆ 0 0 .. . 0

ˆ C Zt+p−2 0 Zt+p−3 0 + Vt+p .. .. . . 0 Zt Z

t+p−1

(17.9.12)

Zt+p−2 ˜ = H Zt+p−3 . .. Zt

8 A MATLAB program spectrs implements ( 17.9.10 ) for the equilibrium of one of our periodic general equilibrium models.

The Tiao-Grupe Formula

409

where . . . ˜ = p−.5 [G1 0 0 . . . 0 .. 0 G2 0 . . . 0 .. . . . .. 0 0 0 . . . Gp ]. H ˜ takes this form, recall that the operator Q(L) is defined as To see why H p.5 Q(L) = [I IL . . . ILp−1 ] = [I 0 . . . 0] + [0 I . . . 0]L + . . . + [0 0 . . . I]Lp−1 . ˜ take the form This structure for Q(L) and the form of H dictates that H that it does and that the state in (17.9.12) takes the form that it does in order ˜ to map (17.9.11) into a first-order system. Notice that the structure of H implies that yt is formed by averaging over linear combinations of the first n rows of Zt+p−r , the second n rows of Zt+p−2 , . . . , and the pth n rows of Zt . Furthermore, notice that according to (17.9.12), the np×np process Zt consists of p completely uncoupled systems, each of which depends on its own past in exactly the same was as do the others. That is, (17.9.12) has the property that Zt is independent of Zt−1 , Zt−2 , . . . , Zt−p+1 for all t ; and that Zt is correlated with Zt−p in exactly the same way for all t . Thus, the “state equations” of (17.9.12) in effect describe p “parallel realizations” of the process Zt+p defined in (17.9.11). Running p parallel processes is a way of realizing in the time domain the randomization over laws of motion that is involved in adopting a description of {yt } in terms of a stationary probability distribution. As noted above, yt is formed by averaging across these p uncoupled realizations. We can use Kalman filtering methods to derive a Wold representation for {yt }. Modify and represent system (17.9.12) as Z˜t+1 = A˜Z˜t + C˜ V˜t+1 ˜ t + ǫ˜t yt = HZ

(17.9.13)

′ ′ where Z˜t′ = [Z˜t+p−1 , Z˜t+p−2 , . . . , Z˜t′ ] and

0 I A˜ = 0 .. .

0

0 0 I .. . 0

... ... ... .. . .. .

0 0 0 .. . I

Aˆ 0 0 , .. . 0

˜ C 0 C˜ = 0 , .. . 0

410

Periodic Models of Seasonality

where {V˜t } is a white noise, and where ǫt is a (potentially very small) measurement error which is a white noise process that is orthogonal to {V˜t } and satisfies E˜ ǫt ǫ˜′t = R . To obtain the Wold representation for yt which achieves the factorization of the spectral density matrix (17.9.10) for yt , we use the Kalman filter to obtain an innovations representation associated with system (17.9.13). The innovations representation is ˜ at Zˆt+1 = A˜Zˆt + K˜ ˜ Yˆt + a yt = H ˜t ,

(17.9.14)

where a ˜t = yt − E[yt | yt−1 , yt−2 , . . .], Zˆt = E[Zt | yt−1 , yt−2 , . . .], ˜ are the state covariance matrix Σ = E(Zt − Zt )(Zt − Zˆkt )′ and where Σ and K and the Kalman gain computed via the Kalman filter for system (17.9.13). The ˜ H ˜ ′ + R.9 covariance matrix of the innovations is given by E˜ at a ˜′t = HΣ

17.10. Some Calculations with a Periodic Hall Model We use a periodic version of Hall’s model as an example. The model is identical to the version of Hall’s model described in chapters 3 and 5, except that the productivity parameter γ now varies periodically. The social planner chooses contingency plans {ct , kt , it }∞ t=0 to maximize the utility functional ∞

1 X t − ( )E β [(ct − bt )2 + ℓ2t ] | J0 2 t=0 0 0

ℓ2t ,

s(t + p) = s(t),

∀t,

s(t) = t for t = 1, . . . , p

9 These calculations are performed by the MATLAB program .

Some Calculations with a Periodic Hall Model

411

and subject to the (exogenous) laws of motion

bt = 30 dt = .8dt−1 + w1t + 5 ∗ (1 − .8) We set p = 4, and γ1 = .13, γ2 = .1, γ2 = .1, γ4 = .08. We set φ1 = .3, δk = .95, β = 1/1.05. The only source of disturbance in the model is the endowment shock, which is a first order autoregression. The variance of the innovation w1t is unity. The following MATLAB programs can be used to analyze the model. solves.m: simuls.m: steadsts.m: assets.m:

assetss.m: seasla.m:

simpulse.m: spectrs.m: factors.m:

computes the equilibrium of a periodic model, simulates a periodic equilibrium; computes the means of variables from a periodic equilibrium, conditional on the season; computes the objects in the formulas for equilibrium assets prices and the term structure of interest rates for a periodic model; simulates the asset prices in a periodic equilibrium; computes the time invariant state-space representation for the stacked, skip sampled version of a periodic model; computes the two different concepts of perioddependent impulse response functions; computes the spectral density of a periodic model, using the Tiao-Grupe formula; factors a univariate spectral density computed via the Tiao-Grupe formula in order to obtain a univariate Wold representation for a single variable of a periodic equilibrium model.

We computed the equilibrium of the periodic version of Hall’s model using solves.m. Figures 17.10.1.a and 17.10.1.b report the spectral density of consumption and investment, computed by using the Tiao-Grupe formula. Both consumption and investment display seasonality, it being more pronounced in

412

Periodic Models of Seasonality

investment than in consumption. This is a reflection of the consumptionsmoothing property of the model. For the impulse response functions of investment with respect to the innovation in the endowment sequence, we used simpulse.m to compute the {dk,v } and {gk,v } sequences corresponding to the moving average representations defined in (17.7.15) and (17.7.16). Figure 17.10.3 reports {dk,v } . The coefficients for each quarter are smooth functions of the lag, but they vary across quarters. Figure 17.10.3 reports {dk,v }, which are each oscillatory functions of the lag. Recall that our periodic 1 of Hall’s model is a one-shock model, with the only stochastic source of disturbances coming from the white noise endowment process. It follows that if we were to shut down the periodic time variation in the productivity of capital, all of the impulse response functions displayed in figure 17.10.3 and 17.10.4 would be equal to one another. 10 The discrepancies across these impulse response functions is a convenient “window” for examining the hidden periodic structure present in investment in this model. Figure 17.10.4 reports the moving average coefficients associated with the univariate Wold representation for investment, which we have normalized by setting the innovation variance equal to unity (so that it is comparable in units with the impulse response functions in figures 17.10.2 and 17.10.3. The coefficient at zero lag in this moving average is .7528, while the coefficients at zero lag for the moving average kernels in figures 17.10.2 and 17.10.3 are (by quarters) .7075, .7069, .7062, .7002. 11 The squared values of each of these coefficients are the one-step ahead forecast error variances in investment, by quarter, when we condition on knowledge of the quarter. The squared value of the coefficient .7528 from the (time-invariant) Wold representation formed by not conditioning on quarter is larger, as we would expect.

10 For this statement to be true in general requires checking that the first of the “two difficulties” discussed by Hansen and Sargent [1990] is not present. 11 The zero lag coefficients are equal for both the {d } and the {g } sequences. k,v k,v

Periodic Innovations Representations for the Periodic Model

10 3

413

10 3

10 2 10 2

10 1 10 1 10 0

10 0 10 -1

10 -2

0

0.5

1

1.5

2

2.5

3

10 -1

3.5

Fig. 17.10.1.a. Spectral density of consumption for a periodic version of Hall’s model, calculated by applying the Tiao-Grupe formula.

0

0.5

1

1.5

2

2.5

0.7 0.6 0.5 0.4 0.3 0.2 0.1

0

10

20

3.5

Fig. 17.10.1.b2. Spectral density of investment for a periodic version of Hall’s model, calculated by applying the Tiao-Grupe formula.

0.8

0

3

30

40

50

60

Figure 17.10.2: The response of the investment component of yp·t−p+k to an innovation in the endowment shock in a periodic verions of Hall’s model.

414

Periodic Models of Seasonality

0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

0

10

20

30

40

50

60

Figure 17.10.3: The response of investment to wp·t−p+k in a periodic version of Hall’s model.

0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 -0.1 0

10

20

30

40

50

60

Figure 17.10.4: The moving average coefficients for a Wold moving average representation of investment, calculated by factoring the spectral density of investment given by the TiaoGrupe formula.

Periodic Innovations Representations for the Periodic Model

415

17.11. Periodic Innovations Representations for the Periodic Model An equilibrium can be represented as xt+1 = Aos(t) xt + Cs(t) wt+1

(17.11.1)

yt = Gs(t) xt + εyt

(17.11.2)

where yt is a vector of objects that are linear combinations of the state xt , plus a white noise measurement error εyt . The matrix Gs(t) is built up from components of the matrices S·,s(t) and M·,s(t) described above. We assume that the measurement error εyt is orthogonal to the wt+1 process, and that it is serially uncorrelated with contemporaneous covariance matrices ¯ s(t) . E εyt ε′yt = R

(17.11.3)

Associated with system (17.11.1) – (17.11.2) is a periodic innovations representation x ˆt+1 = Aos(t) x ˆt + Ks(t) at (17.11.4) yt = Gs(t) x ˆt + at where x ˆht = E[xt | yt−1 , . . . , y1 , x ˆ0 ], at = yt − E[yt | yt−1 , . . . , y1 , x ˆ0 ], and ′ Eat at = Σs(t) . In (17.11.4), Ks(t) is the periodic Kalman gain. The matrices {Σs(t) , Ks(t) } are the p limits of the p convergent subsequences of the Kalman filtering equations: ′ Σt+1 = Aos(t) Σt Ao′ s(t) + Cs(t) Cs(t)

¯ s(t) )−1 Go Σt A′ − Aos(t) Σt G′s(t) (Gs(t) Σt G′s(t) + R s(t) s(t) ¯ s(t) )−1 . Kt = Ao Σt G′ (Gs(t) Σt G′ + R s(t)

s(t)

(17.11.5)

s(t)

¯ s(t) ] are time-varying, system (17.11.5) Because the matrices [Aos(t) , Cs(t) , Gs(t) , R will not converge. But because the matrices [Aos(t) , Cs(t) , ¯ s(t) ] periodic, there is a prospect that {Σt , Kt }∞ will consist of p conGs(t) , R t=1 vergent subsequences. This prospect is realized under regularity conditions that typically obtain for our problems.

416

Periodic Models of Seasonality

The innovation covariance matrix associated with (17.11.4) is Eat a′t = Ωs(t) ¯ s(t) . = Gs(t) Σs(t) G′s(t) + R

(17.11.6)

Given a sample of observations for {yt }Tt=1 , the likelihood function conditioned in x ˆ0 can be expressed as L∗ = − T ln 2π − .5 − .5

t X

T X t=1

ln | Ωs(t) |

(17.11.7)

a′t Ω−1 s(t) at .

t=1

A. A Model of Disguised Periodicity This appendix characterizes a notion of hidden periodicity in a stationary time series, and describes a strategy for detecting its presence in a given vector time series. 12 The notion of hidden periodicity permits realizations of a stochastic process to be aperiodic, but requires that some particular functions of the tail of the stochastic process be periodic. As we shall see, these particular functions are time series averages of skip-sampled versions of the underlying process. It is averaging and skip sampling that causes the hidden periodicity to drop its 0. Because the apparatus introduced in this appendix is abstract, we begin in section A1 with a heuristic account that is designed to indicate the motivation behind the formal apparatus introduced in section A2.

12 Breiman [1968] is a useful background for the material presented in this section.

A1. Two Illustrations of Disguised Periodicity

417

17.13. A1. Two Illustrations of Disguised Periodicity Let {yt } be an n -dimensional stochastic process that is observed by an econometrician. We can use the Kolmogorov Extension Theorem to construct such a process on a sample space Ω = (Rn )∞ , which is the infinite product space formed by taking copies of n -dimensional Euclidean space. A sample point in Ω can be expressed as an infinite-dimensional vector (r0 , r1 , . . .) where rj is in Rn for each j . Probabilities are then defined over the product sigma algebra generated by taking products of the Borel sets of Rn . Armed with this construction, for any ω = (r0 , r1 , . . .), let yt (ω) = rt . Thus, yt (ω) is simply the tth component of the sample point ω = (r0 , r1 , . . .). An alternative way to represent the process {yt } is in terms of a shift operator S . First, define a random variable y : Ω → Rn as y(ω) = r0 . Define the shift transformation S via: S[(r0 , r1 , r2 , . . .)] = (r1 , r2 , r3 , . . .). Then because yt (ω) = rt , an alternative representation of yt is yt (ω) = y[S t (ω)], where S t is interpreted as applying S t times in succession. In thinking about hidden periodicity, the following example is of pedagogical interest. Example 1: Suppose that n is one and that all of the probability on Ω is concentrated onto two points, say a and b . Let a be a sequence of alternating ones and minus ones, beginning with a one. Let b be a similar sequence except that it begins with a minus one. Note that S(a) = b, and S(b) = a. There are many probability structures that we can impose on Ω in Example 1. We can assign any probability between zero and one to a and the remaining probability to b . This assignment amounts to initializing the process. Unless we assign probability one half to each point, the resulting process will not be stationary.

418

Periodic Models of Seasonality

In one sense, the initial assignment of the probability is quite irrelevant. The {yt } process is deterministic in the sense that given knowledge of the y0 , the entire future of process can be forecast perfectly. Since the initial condition tells the whole story, one might just as well condition on it. However, from the vantage point of interpreting time averages of the process, the initial assignment of probability one half to each point is convenient. Independently of how we initialize the stochastic process, it obeys a Law of Large Numbers. Thus, take any Borel measurable function φ mapping Ω into R and form the sequence {zt } where zt = φ(yt , yt+1 , . . .). Then lim (1/N )

N →∞

N −1 X

zt = (1/2)z(a) + (1/2)z(b)

(17.13.1)

(17.13.2)

t=0

where z0 ≡ z . The equality holds when the left-side of equation (17.13.2) is evaluated at either a or b . When probability one half is assigned to each point, the right side of (17.13.2) can be expressed as Ez , so that we have the usual characterization of the limit points of sample averages as mathematical expectations. With this assignment of probabilities, the process {yt } is both stationary and ergodic. For this particular example, realizations of both the original process {yt } and the constructed process {zt } are periodic sequences. While realizations of {yt } have period two, realizations of the constructed process {zt } can have period one for particular choices of zt . For instance, let zt = yt + yt+1 . Then for either a or b, {zt } is a sequence of zeroes and hence has period one. More generally, for this example the periodicity of {zt } can never exceed two. This follows from the fact that S 2 (a) = a and S 2 (b) = b implying that zt+2 (ω) = zt [S 2 (ω)] = zt (ω). Since the maximum periodicity of any constructed process {zt } is two, we will say that the periodicity of S is two. There is something very special about Example 1. Since realizations of the original {yt } process are periodic, every constructed process {zt } turns out to be periodic. In this paper, we are interested in more general circumstances in which the periodicity is disguised. We do not wish to confine attention to processes {yt }

A1. Two Illustrations of Disguised Periodicity

419

whose realizations have an exact periodicity. The following example embodies what we mean by a hidden periodicity. Example 2: Let {wt } be an nw -dimensional Gaussian white noise with covariance matrix I . Construct an nx -dimensional stochastic process {xt } recursively via xt+1 = At xt + Bt wt+1 where {(At , Bt )} is a periodic sequence with period two, where At is an (nx × nx ) matrix, and {Bt } is an (nx × nw ) matrix. Let {yt } be an n -dimensional process generated as a time-varying function of {xt } yt = ft (xt ) where {ft } is a sequence of Borel measurable functions mapping nx -dimensional Euclidean space into n -dimensional Euclidean space. Let ft be a sequence of period two. Realizations of {yt } will not be periodic, but will inherit a sort of disguised periodicity from {(At , Bt , ft )}. There are two aspects of the process {yt } that we have left unspecified, namely x0 and the periodic sequence {(At , Bt , ft )}. As in Example 1, there is flexibility in the probabilistic specification of {(At , Bt , ft )}. One possibility is, in effect, to condition on {(At , Bt , ft )}, in which case the resulting process {yt } will not, in general, be stationary. Alternatively, we can view {(At , Bt , ft )} as emerging from a random draw from two possible sequences indexed by, say, a and b where [At (b), Bt (b), ft (b)] = [At+1 (a), Bt+1 (a), ft+1 (a)] for all t . As in Example 1, if we assign probability one half to each of these outcomes, under a restriction 13 on a matrix that is a function of At (a) and At (b), we can find an initial specification of x0 under which {yt } is a stationary stochastic process. In this case, we can apply the Law of Large Numbers for stationary processes both to show that time series averages converge and to obtain a characterization of the limit points. Suppose that it is possible to complete the specification in Example 2 so that {yt } is stationary. Consider how the hidden periodicity can be characterized and detected. Let ψ be a Borel measurable function mapping Ω → R and form a 13 The restriction is that the matrix A ˆ in equation (4.7) below have eigenvalues that are bounded in modulus by unity.

420

Periodic Models of Seasonality

scalar stochastic process {zt∗ } via

zt∗ ≡ ψ(yt , yt+1 , . . .) or

zt∗

∗

(17.13.3)

t

= z (S (ω))

We assume that E | z ∗ (ω) |< +∞, where z ∗ (ω) = z0∗ . In contrast to Example 1, when the periodicity is hidden, there is no necessity that {zt∗ } form a periodic sequence. Hence we must have a weaker notion of periodicity, if the S implied by Example 2 is to be classified as periodic with period 2. A workable notion of hidden periodicity can be formulated in terms of a reduced class of constructed processes. Given an integer j ≥ 1 and given ψ , define φ : Ω → R, via φ(yt , yt+1 . . .) = zt where zt ≡ lim (1/N ) N →∞

N −1 X τ =0

∗ zt+τ ·j ,

(17.13.4)

and where the right side of (17.13.4) is defined as an almost sure limit. Note that the process {zt } is constructed by taking time series averages of skip samples of the process {zt∗ } with skip interval j . Notice that zt depends only on the tail of the stochastic process {yt }. It follows by construction that zt is a periodic process with a period not exceeding j . The time series averages of skip samples will reveal the hidden periodicity. The idea is to compute (17.13.4) for j = 2, 3, 4, . . . , and then to determine the period pˆ of this sequence for each j . Thus, in example 2, it will turn out that for j = 1, 3, 5, . . . the number pˆ is one. For j = 2, 4, 6, 8, . . . , the number pˆ will turn out to be 2. We shall define the hidden periodicity p as the maximum of these numbers pˆ over j = 1, 2, 3, . . . , where the maximum is also understood to be taken over a class of “test functions” ψ . Thus, the notion of hidden periodicity in a stochastic process that we shall use is the periodicity to be found in time series averages of skip sampled versions of the data. In the next subsection, we develop these ideas formally, and define hidden periodicity precisely in terms of the properties of the shift operator S and its iterates.

A2. Mathematical Formulation of Disguised Periodicity

421

17.14. A2. Mathematical Formulation of Disguised Periodicity We now use the familiar formalism for stationary stochastic processes. 14 As in the previous subsection, let (Ω, F, P r) denote the underlying probability space, and let S be a measurable, measure-preserving transformation mapping Ω into itself. Definition 1: A transformation S is measure preserving if P r(f ) = P r(S −1 f ) for all f ∈ F . Let I be the collection of invariant sets of the transformation S . Definition 2: f ∈ F is an invariant set of S if S −1 (f ) = f . The collection I turns out to be a sigma algebra of events (see Breiman), so expectations conditioned on I are well defined. The invariant events of the transformation S given in example 1 are the null set and any set containing {a, b}. Definition 3: S is ergodic if all invariant events have probability zero or one. Notice that S in example 1 is ergodic. Let L be the space of random variables with finite absolute first moments, and let M be the subspace of L consisting of the random variables that are I measurable. The expectation operator E(· | I) maps L into M. Throughout this section, we use the common convention that equality between random variables is interpreted formally as equality with probability one. Hence the equivalence class of random variables in L that are equal almost surely are treated as one element. Similarly, for a random variable to be in M, it suffices for it to be in L and to be equal almost surely to a random variable that is measurable with respect to I . When S is ergodic, M contains only random variables that are constant almost surely. A transformation S that is measure-preserving can be used to construct processes that are strictly stationary. Let z be a random variable in L, and construct zt (ω) ≡ z[S t (ω)]. 14 See Breiman [1968, chapter 6].

(17.14.1)

422

Periodic Models of Seasonality

Then {zt } is strictly stationary and hence obeys a Law of Large Numbers. The limit point of the time series averages is given by E(z|I). A z ∈ L has two interpretations. First, it indexes a stochastic process via (17.14.1); and second it denotes the time zero component of that stochastic process. Our purpose is to define a notion of periodicity for the transformation S . Suppose there exists a random variable z such that the realizations of the resulting process {zt } are periodic. That is, for some j the resulting process satisfies: zt+j = zt for all t ≥ 0.

(17.14.2)

The fact that (17.14.2) holds for a particular stochastic process is informative about the periodicity of S but falls short of determining the periodicity of S . Notice that one can always find a random variable z such that (17.14.2) is satisfied for j = 1. In particular, let z be constant over states of the world. Since S is measure-preserving, zt = z for all t . Heuristically, we shall define the periodicity of S by forming a large set of periodic stochastic processes defined as in (17.14.1) and satisfying (17.14.2) for some j , and then calling the periodicity S the maximum j over these processes. Notice that all transformations S have periodicity of at least one. To define formally the periodicity of S , we investigate the collection of invariant events of integer powers of the transformation S . Evidently, if S is measure-preserving, then S j is measure-preserving for any positive integer j . We can think of S j as corresponding to skip-sampling every j time periods. Let I j denote the collection of invariant events of S j , and let Mj denote the corresponding subspace of L of random variables that are I j measurable. Any invariant event of S is also an invariant event of S j . Consequently M ⊂ Mj . The converse is not true, however. Consider example 1. Note that S 2 (a) = a and S 2 (b) = b . Consequently, {a} and {b} are invariant events of S 2 but not of S . In this case I 2 = F . When M consists only of random variables that have the same values on a and b, M2 = L. Processes that are generated (indexed) by elements of M are constant over time and hence have period one. On the other hand, processes generated by elements of M2 can oscillate with period two. It of interest to obtain a characterization of Mj that applies more generally. Lemma 1: For any z ∈ Mj , zt+j = zt for all t ≥ 0. Conversely, for any z ∈ L such that zt = zt+j for all t ≥ 0, zt ∈ Mj for all t ≥ 0.

A2. Mathematical Formulation of Disguised Periodicity

423

Proof: Suppose that z ∈ Mj . Then S −j ({z ∈ b}) = {z ∈ b} for any Borel set b of R. Note that S −j ({z ∈ b}) = {zj ∈ b}. Consequently for any Borel set b , {z ∈ b} = {zj ∈ b}. Equivalently, zj = z . Repeating this same argument, it follows that z = zτ ·j for any positive integer τ . Recall that S is measure-preserving, as is S t . Consequently, for any Borel set b , P r({zt ∈ b} ∩ {zt+τ ·j ∈ b}) = P r({z ∈ b} ∩ {zτ ·j ∈ b}), P r({zt ∈ b}) = P r({z ∈ b}), and P r({zt+τ ·j ∈ b}) = P r({zτ ·j ∈ b}). Since z = zτ ·j , it follows that P r{zt = zt+τ ·j } = 1 for any positive integer τ . Next consider the converse. Suppose that z ∈ L such that zt = zt+j for all t ≥ 0. It remains to show that zt ∈ Mj . The sequence of time series averages {(1/N )

N −1 X τ =0

zt+τ ·j }

converges almost surely to zt as well as to E(zt | I j ). Therefore, P r{zt = E(zt | I)} = 1. In light of Lemma 1, processes generated by elements of Mj are periodic with a period that is no greater than j . We wish to use this insight to construct a formal definition of periodicity. Let Mcℓ be the closed linear space generated by {Mj }∞ j=1 where closure is defined using the standard norm on L, E(| · |). Definition 4: The transformation S is said to have periodicity p if p is the smallest integer such that Mp = Mcℓ . Under this definition, random variables in Mcℓ generate periodic processes with maximum period p . Applying this definition to the transformation S given in example 1, we verify that S has period 2. Next we describe an alternative way to deduce the periodicity of S . Mimicking the previous logic, we can show that for any positive integer j , Mj ⊂ Mτ ·j for τ = 1, 2, . . . .

(17.14.3)

It turns out that if ⊂ in (17.14.3) can be replaced by = , the period of S is no greater than j , and in fact j must be an integer multiple of the actual periodicity p . In other words, once skip-sampling reaches a point where further sampling fails to increase the collection of invariant events, this point is an integer multiple of the periodicity of S .

424

Periodic Models of Seasonality

Lemma 2: Let j be any positive integer such that Mj = Mj·τ for τ = 1, 2, . . . . Then Mcℓ = Mj and S has periodicity p where j = ℓ · p for some positive integer ℓ. Proof: First we show that Mcℓ = Mj . Suppose to the contrary that there is some random variable in Mcℓ that is not in Mj . Since Mj is closed and random variables in Mcℓ are limit points of sequences of random variables in S τ M , there exists a positive integer τ and a random variable z such that z is in Mτ but not in Mj . However, Mτ ·j = Mj by assumption, which is a contradiction. Therefore Mcℓ = Mj and p ≤ j . It remains to show that j = p·ℓ for some integer ℓ. Note that Mp = Mj = Mcℓ . In light of Lemma 1, random variables in Mj generate processes with period p and period j . Let ℓ be the smallest integer such that ℓ · p ≤ j and suppose that ℓ·p < j . Then p > k > 0 where k ≡ j−ℓ·p . For any z ∈ Mp , with probability one z = zp = zℓ·p = zℓ·p+k . Since S is measure-preserving, {zt } is periodic with period k . It follows from Lemma 1 that, z ∈ Mk . Consequently, Mk = Mp which is a contradiction. This in turn implies that the period of S is at least j − ℓ · p , which is a contradiction. Therefore j = ℓ · p . An implication of Lemma 2 is that processes generated by random variables in Mcℓ are periodic with a period equal to j , where j = ℓ · p for some integer ℓ. Note that if S has periodicity p , then S p has periodicity one. Definition 4 of periodicity can be applied to any S transformation that is measure-preserving. Our interest is in the case in which S is the shift transformation described in section 1a. This transformation is measure-preserving by construction as long as the probability measure induced on Ω comes from a process {yt } that is strictly stationary. When the shift transformation is periodic with period p , we say that the process {yt } has hidden periodicity p . Consider again constructions (17.13.3) and (17.13.4). The processes {zt } constructed via (17.13.4) are periodic by construction and hence it follows from Lemma 1 (or from the Law of Large Numbers for Stationary Processes) that the corresponding random variable z is in Mj . The periodicity of {zt } can, in fact, be less than j . By choosing a sufficiently rich collection of test functions, we can span Mj . Let pˆ(j) be the maximum periodicity over such a class of functions. The hidden periodicity p of {yt } is then the supremum of the sequence {ˆ p(j) : j = 1, 2, . . .}. Lemma 2 describes a particular feature of subsequences of {ˆ p(j) : j = 1, 2, . . .}. For instance, for any j = p · ℓ for some ℓ, the subsequence {ˆ p(τ · j) : τ = 1, 2, . . .} is constant. Turning this observation

A2. Mathematical Formulation of Disguised Periodicity

425

around, if one finds a constant subsequence of the form {p(τ · j) : τ = 1, 2, . . .}, then the hidden periodicity of {yt } must satisfy j = p · ℓ.

Part IV Economies as Objects

Chapter 18 Introduction to Objects This help manual is intended to help students use the MATLAB programs referenced in this book. To that end, it is divided into two main chapters, organized by increasing level of difficulty. This first chapter explains a little about object oriented programming (OOP), and why it’s so useful in this context. The second chapter applies these ideas to the construction of an economy object, and offers a more in-depth coverage of all the features of an economy and all the nifty things one can do with it. This second chapter also provides the user with the tools, via examples, to invent new economies to experiment with. Throughout, actual code and MATLAB file names will be in typewriter font, but references to an object will not be. For example, economy.m is in typewriter font, but when an economy is referred to, it is not. Also note that any actual MATLAB file has a short help section at the beginning, which can be accessed by typing help file name at the MATLAB prompt.

18.1. Matlab Objects For those users who have not encountered OOP before, this subsection goes over some definitions and some examples of what one can and cannot do with objects. Of course, all examples are in the context of MATLAB.

18.1.1. Definitions We start with some definitions of basic concepts in OOP. Class The relevant analogy here is that of a type. A class is a new data type that you define. It includes not only the actual structure of the type, but also the functions that operate on it. So, for example, suppose we define a new class called a slde (short for stochastic linear difference equation) which is a collection of two matrices, and we need a function on it that displays it

– 429 –

430

Introduction to Objects

in a nice way. Then both the constructor slde.m and the display function disp.m are in the class. Objects An object is a run-time value which belongs to some class. If a class is a type, an object is a variable. A variable of a given class is called an instance of that class. For example, suppose we define a class that is a matrix. Then the identity matrix would be an object, or an instance of the matrix class. Hierarchy Classes in OOP are arranged in a tree-like hierarchy. A class’ superclass is the class above it in the tree. The classes below it are subclasses. By convention, the root of the tree is called the “Object” class. The semantics of the hierarchy are that any class includes all the properties of its superclasses. In this way the hierarchy is general towards the root and specific towards its leaves. The hierarchy helps add logic to a collection of classes. It also enables similar classes to share properties through inheritance. Superclasses are often referred to as parent classes, and subclasses are often referred to as children. Inheritance A subclass inherits all of the data and functionality of its parent classes. In particular, a class inherits all of the methods. When an object receives a message, it checks for a corresponding method. If one is found, it is executed. Otherwise the search for a matching method travels up the tree to its parent and so on recursively. This means that a class automatically responds to all the messages of its superclasses. Most OOP languages include controls to limit how the data and methods are inherited. A subclass can also extend beyond the inherited functionality by adding data and defining new methods. Overriding and Overloading A class inherits all the methods of its superclasses, but a class can choose to respond to a message in a different way by re-defining a method. When an object receives a message, it checks its own methods before consulting it superclass. If the object’s class and its superclass both contain a method for a message, the object’s method is used. In other words, the first method found in the hierarchy takes precedence. When a subclass responds to a

Matlab Objects

431

message in a different way than its superclass does, the subclass is said to have overriden its superclass’s method—the class overrides and intercepts the message before it gets to the superclass. When a method has more than one definition depending on the context (i.e., it’s defined for both a class and its parent), it is said to be overloaded. In MATLAB, most functions, including addition and subtraction, for example, can be overloaded to be class specific. For example, the function disp.m is overloaded for each class, so that it displays each class properly. Fields A field is a component of a class which is itself a class, or predefined type. For example, if we define a class called foo that consists of an integer and a matrix, then foo has two fields. If we defined a child class called foo child that additionally has a vector, then foo child also has two fields. The first is the foo field that is inherited, and the second is the vector field. There are not three fields, as the first two are subsumed into the foo field that foo child inherits.

18.1.2. Matlab Specifics There are two main ways in which MATLAB departs from the structure laid out above: accessing fields, and overloading certain functions. Additionally, there is a specified way in which MATLAB checks for methods, which we discuss here, as it may sometimes cause confusion. Accessing Fields Generally, the most direct way to access a field of an object is to call it via foo.fieldname. When defining new objects, this method works immediately within the functions of a class. However, to be able to do so at the command line, two extra functions are needed in each class directory: subsref.m and subsasgn.m. The first allows you to reference the value in the field, so if you type foo object.fieldname on the command line, it returns the value contained in the fieldname field of object foo object. The latter function, subsasgn.m, allows you to reassign a field of an object to be a different value. These two functions should be overloaded for every class. They work by taking an object both as an argument and as the

432

Introduction to Objects

output. This will turn out to be especially handy for reassigning values in an object in the economy class, as we will see in the next chapter. Overloading Functions One must be wary about overloading certain functions in MATLAB. For example, we defined our own method sigma.m to extract the sigma field from an economy object. However, sometimes it seems to work, and sometimes it seems to not work. We suspect that this is because there is already a sigma function defined by MATLAB, which it does not like to override. In general, it may be wise to try and name a function a different name, such as sig.m, rather than overloading. One can always check for taken names by typing help function name; if MATLAB says it cannot find the function, it hasn’t been taken. Calling Functions MATLAB has a specified protocol for where to seach when a new name is called. In the following order it looks for a: variable, subfunction, private function, or function on the search path. A subfunction is a function that resides in the same file as the calling function; we almost never use these. A private function is one that is in a private directory, and hence is only accessible to files in the directory immediately above it. An example of this is the solve.m function in the @economy/private directory. The search path is the predefined path along which MATLAB searches, to which we have added paths for examples/econ and clex. For more information on paths in MATLAB, type help path.

18.1.3. How to Define a Matlab Class Suppose one wants to define a MATLAB class called foo. Then in a folder called @foo there must a constructor function called foo.m. If one defines a child class called foo child, then there must be another folder @foo child with a constructor function foo child.m. Note that when constructing a foo object, one must be in a directory such that @foo is a subdirectory. The folders @foo and @foo child may both be in the same directory. The constructor function takes inputs and assigns them to fields. Once all fields are assigned, there is a declaration of the class, and the new object is

Matlab Objects

433

returned. A short example for @foo/foo.m follows, where foo has two fields, an integer and a matrix. Note that it is a good idea to include some comments about what a function does in the function definition file. In Matlab, the first comments section is displayed when help function name is typed. The character for commenting out a line is the percent symbol: %. The help section should include definitions of all variables used, the first line of the function definition, and a short description of what the function does.

function f = foo(integer1, matrix2); % function f = foo(integer1, matrix2); % This function is the constructor called when a new foo object % is defined. It takes two arguments, an integer and a matrix % and assigns them to the two fields of a foo object. foo.integerfield = integer1; foo.matrixfield = matrix2; f = class(f, ‘foo’); To create an object called foo1 using the integer 5 and the matrix eye(4), one simply types at the command line: foo1 = foo(5, eye(4)). Now suppose one wants to define a class called foo child that is a child of foo. In the same directory as @foo create a folder called @foo child that contains the constructor function foo child.m. Suppose that in addition to the two fields in foo, one wants foo child to have a vector field. The file @foo child/foo child.m is given for reference..

function fc = foo child(foo1, vector2); % function fc = foo child(foo1, vector2); % This function is the constructor called when a new foo child % object is defined. It takes two arguments, a foo object % and a vector, and assigns them to the two fields of a foo child % object. foo child.vectorfield = vector2; fc = class(fc, ‘foo child’, foo1);

434

Introduction to Objects

Notice that, since foo1 is an object of the foo class, foo child automatically inherits from foo, and the inherited fields are filled with the values from foo1. For a more thorough introduction to defining MATLAB objects, see Chapter 14 of the MATLAB manual “Using MATLAB”.

18.2. Summary In this chapter we have defined some of the basic concepts in object oriented programming: classes, objects, hierarchy, inheritance, overriding and overloading, and fields. We then went on to point out some pitfalls of MATLAB objects: field access and overloading. Finally, very briefly, we went over how to define a MATLAB class. One can now see why object oriented programming would be a valuable tool in defining and using economies. An economy is defined to be a collection of matrices. These divide naturally into three categories, each of which define the information, technology and preferences structures. Thus one can create three classes representing these structures, and have an economy class inherit from all three. Also, most interesting operations are executed on an economy, and with an economy class, one can simply define functions that operate on an economy object. The exact definitions and functionality of these four classes are discussed in the next chapter.

Chapter 19 Economies as Matlab Objects

19.1. Introduction We describe in the first few sections the structure of the economy class starting with the structure of the parent classes: information, technology and preferences. Notice that the field names correspond as much as possible with the names of matrices and vectors in the main book. We also describe the functions available to manipulate the various objects. To get more information on any of these functions, type help function name in the MATLAB command window. Also, recall that the information, technology and preference divisions are introduced in Chapter 3 of Hansen and Sargent. The last section deals with three different ways of working with objects of this class: using the built-in economies, mixing and matching the built-in parent objects (information, technology and preferences) and building a customized economy.

19.2. Parent Classes: Information

19.2.1. Structure The first parent of the economy class is the information class. An information object contains matrices describing the laws of motion of taste and technology shocks. zt+1 = A22 zt + C2 wt+1 bt = Ub zt dt = Ud zt . An information object has four fields: a22:

The matrix A22

– 435 –

436

c2:

Economies as Matlab Objects

The matrix C2

ub: The selector matrix Ub , which transforms the process zt into taste shocks. ud: The selector matrix Ud , which transforms the process zt into technology shocks.

19.2.2. Functions construction: information, a constructor function that takes as argument the matrices a22, c2, ub, ud

19.3. Parent Classes: Technology

19.3.1. Structure The technology class contains matrices that describe the technology of the economy : Φc ct + Φg gt + Φi it = Γkt−1 + dt kt = ∆k kt−1 + Θk it . Recall that ct is consumption, gt a vector of intermediate goods, it investment, kt capital, and dt the production shock. A technology object has six fields : phic:

The matrix Φc

phig:

The matrix Φg

phii:

The matrix Φi

gamma:

The matrix Γ

deltak:

The matrix ∆k

thetak:

The matrix Θk

Parent Classes: Preferences

437

19.3.2. Functions construction: technology, a constructor function which takes as arguments the matrices phic, phig, phig, gamma, deltak, thetak

19.4. Parent Classes: Preferences

19.4.1. Structure The preferences object contains scalars β, σ and matrices ∆h , Θh , Λ, Π that describe the preferences of the representative agent. The household technology is: ht = ∆h ht−1 + Θh ct st = Λht−1 + Πh ct and preferences are ordered by E0

∞ X t=0

β t (st − bt )2 + ℓ2t .

Recall that ht is a household stock of durables, ct is a consumption, and st is services from the stock of durables. The vector bt is used in the household objective function. A preferences object has six fields : deltah:

The matrix ∆h

thetah:

The matrix Θh

lambda:

The matrix Λ

pihh:

The matrix Πh

beta:

The discount factor β

sigma:

The risk sensitivity σ

438

Economies as Matlab Objects

19.4.2. Functions construction: preferences, a constructor function that takes as argument deltah, thetah, lambda, phih, beta, sigma

19.5. Child Class: Economy

19.5.1. Structure The elements of an economy are contained in the parent fields: information, technology, and preferences. An economy as a child object inherits these three fields from its three parents (technology, preferences, and information). In addition, an economy object has several other fields that are not inherited, namely a set of matrices that characterize a competitive equilibrium. These are calculated automatically when an economy object is defined and become fields in the economy object. Furthermore, the function subsasgn.m for the economy class has been defined so that whenever one changes the value of a field of an object, the equilibrium is automatically recalculated. We shall illustrate this useful feature below. Chapters 4 and 6 showed that an equilibrium has the representation xt+1 =A0 xt + Cwt+1 yt =Gxt . ′ ht−1 0 A11 A12 . , C= where x′t = kt−1 , Ao = C2 0 A22 zt−1 The observables and the shadow prices are in the vector yt . They are all linear combinations of the state variables in xt . The coefficients of these linear combinations are in the matrix G , various rows of which were denoted Mj in Chapter 4 for a price of a quantity j .

The fields containing the solutions of the resource allocation problem are the following :

Child Class: Economy

439

ao : The matrix A0 endo: The eigenvalues of the block A11 of A0 exo: The eigenvalues of the block A22 of A0 nnc: The coordinate of the row of x containing a constant, if any c: The matrix C sj, mj: Quantities and shadow price are stored in sixteen different fields: sb, sc, sd, sg, sh, si, sk, skl, ss, mc, md, mg, mh, mi, mk, ms. These correspond with the matrices of analogous names. For example, sc is the Sc of Chapter 4, which multiplies xt to yield the optimal decision for ct . And mc multiplies xt to yield the Lagrange multiplier Mct (the shadow price of consumption). s space: An s space (state space) object containing the state space representation of the economy (@[email protected] ). To have more information about what a state space object is, type help s space at the MATLAB prompt. This field is useful for computations involving the MATLAB control toolbox.

19.5.2. Fields containing the history of the economy An economy object contains also the history of the economy, that is an initial condition x0 . and a sequence of shocks {wt }Tt=0 . (The sequence of shocks can be initialized at a null matrix, and has been in our sample economies. We include it as a potential field because it can be useful for generating simulations.) This information is stored in the following fields: hinitial: The initial condition for the household capital goods kinitial: The initial condition for the capital stock zinitial: The initial condition for the information process shocks : The sequence of shocks in the information process

440

Economies as Matlab Objects

19.5.3. Functions

19.5.4. Constructing the object and changing parameters This part will be detailed in section 3 where we will explain what sequence of commands is needed to construct an economy object. For the time being, we simply list the functions : construction: economy, a constructor function that takes as arguments an information object, a technology object and a preferences object. display: A function disp which display the structure of the economy

19.5.5. Analyzing the economy steady state: A function steadst to compute the steady state of the model. This uses the resource allocation solution fields. asset pricing: A function asset price to compute and simulate the price of an asset with payoffs that are linear combinations of the state variable of the economy A function riskprem to compute the risk premium on an asset A function sure4j to compute the prices of a j-period sure claim on consumption simulation: A function simulate to compute and graph time path for the observables and shadow prices. reopening markets: A function reopening to compute and graph the time path of prices in markets that reopen every period impulse response: A function impulse to compute and graph the impulse response of the observables and shadow price to the shocks hitting the economy arma representation: A function arma rep to compute the arma representation of the

Working with economies

441

spectral density: A function spect to compute and graph the spectral densities of the observables and shadow prices.

19.6. Working with economies

19.6.1. The built-in economies You will find in the directory examples/econ a series of script files which build standard economies – most of them described in the book. A typical script file is named economy name.m. MATLAB will build the economy when you type economy name.m at the command line. The program will create the following variable : eeconomy name: an economy object

19.6.2. Mixing and matching built-in parent objects By using the built-in economies, your freedom is very restricted: you cannot set any structural parameters, the only thing you can modify is the history of the economy. To give more freedom to your experiments, we have constructed some standard structures corresponding to the examples given in chapter 3. You can mix and match them and set some of their parameters. Technology You will find in the directory examples/tech functions which create the technologies given as examples in chapter 3. A typical function is named techj.m. By typing help techj you will get a description of the parameters you are free to set. Note that all these parameters have default values. You build a technology by typing name of tech = techj(parameters) Preferences You will find in the directory examples/pref functions which create the preferences given as examples in chapter 3. A typical function is named prefj.m. By typing help prefj you will get a description of the parameters you are free to set. Note that all these parameters have default values. You build a technology by typing name of pref = prefj(parameters)

442

Economies as Matlab Objects

Information You will find in the directory examples/info functions which create information processes, most of them following the examples given in chapter 2. A typical function is named infoj.m. You will have to specify parameters of an underlying stochastic linear difference equation and extractor matrices ub and ud. This is a delicate step: you have to make sure that those matrices are comformable in columns with the process zt and in rows with the consumption services vector (for ub) or with the technology shocks vector (for ud). Having a look at the script files in examples/econ can be useful, although the information constructor is designed to warn you when your matrices are not conformable. Economy After you have built the three parents objects info, tech and pref you are ready to create their child, the economy object. To do so you simply type econ name = economy(i,t,p) and a new economy will be born. You will probably immediately want to reset the initial conditions for the h−1 , k−1 , z0 , which the economy constructor sets at vectors of 1’s as their default values. The initial conditions can be assessed from the economy object by typing econ name.hinitial, econ name.kinitial, and econ name.zinitial, respectively. To reset z0 , for example, type econ name.zinitial = [ 5 2 0 ]’. You can also directly reset other objects of any of the three parent objects (preferences, technology, or information), which will then be automatically inherited by the child economy object. We’ll describe how to do this soon.

19.6.3. Building your own economy In the directory examples/econ there is a script file blank.m which may be useful in building your first economies. Just fill in the blanks (the null matrices) with conformable matrices and run the script file; you’ll be ready to experiment with your new economy. % Creates an economy with null matrices everywhere. % Required dimensions are given in comments. %%% Technology %%% deltak=[] ; % n k by n k thetak=[]; % n k by n i phic=[]; % m by n c phig=[] ; % m by n g phii=[]; % m by n i

Working with economies

443

gamma=[] ; % m by n k %%% Preferences %%% deltah=[]; % n h by n h lambda=[]; % n s by n h thetah=[]; % n h by n c pih=[] ; % n s by n c beta=; % scalar sigma = 0; % scalar %%% Information %%% a22=[]; % n z by n z c2=[]; % n z by n w ud=[]; % n d by n z ub=[]; % n b by n z %%% Construction %%% iblank = information (a22, c2, ub, ud); tblank = technology (phic, phig, phii, gamma, deltak, thetak); pblank = preferences (deltah, thetah, lambda, pih, beta, sigma); eblank1 = economy (iblank, tblank, pblank); eblank1.hinitial = []; eblank1.kinitial = []; eblank1.zinitial = []; clear iblank1 tblank1 pblank1; clear phi gam sigma beta; clear a22 c2 u* phi*; clear gamma del* the* lambda pih;

444

Economies as Matlab Objects

19.7. Tutorial Our object oriented programs are contained in a directory called hansar1 that contains various subdirectories. Say that you keep the directory hansar1 in the location c: \projects \hansar1. After you start MATLAB type addpath c: \projects \hansar1 then type startup. To read one of our existing economies, type for example clex11. The object eclex11 is then created. Type eclex11 to display it. Then to conserve notation rename the economy object simply e by typing e=eclex11. To access one of the fields of the economy e, a child object, type e.j where j is one of the economy fields described above, namely, ao, c, endo, exo, sv, mv, hinitial, kinitial, zinitial, where v denotes one of the variables c, i, h, k. To assess one of the parent objects, type either e.information or e.technology or e.preferences. To assess one of the fields of one of the parent objects, type either e.information.j where j=a22, c or e.preferences.j where j= lambda, deltah, pi, thetah,beta, sigma or e.technology.j where j = deltak, thetak, gamma, phig, phii, phic. To reset an element of a parent object, type for example e.information.a22(2,2) = .4, a command that sets A22 (2, 2) = .4, and that then recomputes all of the equilibrium objects in the child economy e.

Chapter 20 MATLAB Programs This chapter consists of a manual of MATLAB programs that implement the calculations described in earlier chapters. Many of the programs use programs in MATLAB’s Control Toolkit. You should load our programs into a subdirectory of MATLAB, and put this subdirectory on the matlabpath statement in your matlab.bat file. There is a demonstration facility for some of our programs, which supplies a small course on how to use many of our programs. To use this program, just get into MATLAB, type hsdemo, and choose one of the options that the menu offers you.

20.1. Matlab programs Our ordinary MATLAB programs are available via ftp at < ftp://zia.stanford.edu/pub/˜sargent/webdocs/matlab/hansar/hansarprograms.zip> . Our object oriented programs are available at < ftp://zia.stanford.edu/pub/˜sargent/webdocs/matlab/hansarobjects.zip>

– 445 –

446

MATLAB Programs

aarma

Purpose: Creates arma representation for a recursive linear equilibrium model. Synopsis: [num,den,p,z]=aarma(ao,c,sy,i)

Description: The equilibrium is computed by first running

solvea.

The equilibrium is

xt+1 = ao xt + c wt+1 A vector of observables is given by yt = sy xt , where sy is formed to pick off the described variables. For example, if we want yt = [c′t , i′t ], we set sy=[sc; si]. aarma creates num and den, which pertain to the representation den (F )yt = num (F )wit where F is the forward shift operator defined by F yt = yt+1 . This is an arma representation for the response of yt to the i-th component of wt . num(F) and den(F) are each stored with the coefficients being arranged in order of descending powers of F . The poles (zeros of den(F)) are returned in the vector p. The zeros of num(F) for each variable are returned in a column vector z, where each column corresponds to a variable. References: [1] Hansen, Lars Peter and T.J. Sargent, Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988.

Matlab programs

447

aggreg

Purpose: Computes state space representation of sampled (time aggregated) data. Synopsis: [ Ar , Cr , aa, bb, cc, dd, V1] = aggreg (A,C,G,D,R). Description: The underlying model is xt+1 = Axt + Cwt+1 yt = Gxt where wt+1 is a martingale difference sequence. Error ridden observations on y are available only every r periods. The state space model for the data is then xs+1 = Ar xs + Cr wrs+1 ys = Gxs + vs vt+1 = Dvs + us+1 ′ where s = t·r, Eut u′t = R, Ar = Ar , Cr = I, Ewrt wrt = Vr Vr = CC ′ +ACC ′ A′ + · · · Ar−1 CC ′ A′ r−1 . The program uses innov to create an innovations representation for the sampled process {yt , t = 0, r, 2r, 3r, . . .} = {ys , s = 0, 1, 2, . . .}. varma2 can be used to compute an arma representation for the sampled data.

References: [1] Hansen, Lars Peter and T.J. Sargent, Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988.

448

MATLAB Programs

aimpulse

Purpose: Computes impulse response function for a recursive linear equilibrium model Synopsis: [z]=aimpulse(ao,c,sy,ii,ni)

Description: The equilibrium is computed by first running

solvea.

The equilibrium is

xt+1 = ao xt + c wt+1 . A vector of observables is given by yt = sy xt where sy is formed to pick off the desired variables. For example, if we want yt = [c′t , i′t ]′ , we set sy=[sc;si]. aimpulse computes the impulse response of yt with respect to component ii of yt for ni periods. References: [1] Hansen, Lars Peter and T.J. Sargent, Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988.

Matlab programs

449

asimul

Purpose: Simulate a recursive linear equilibrium model Synopsis: asimul, a script file. The outputs of solvea must be in memory, as must the matrix sy and the integer t1 . Description: The equilibrium is xt+1 = Ao xt + Cwt+1 A vector of observables yt obeys yt = sy ∗ xt , is to be specified by the user. If we want yt = (c′t i′t )′ , we would set sy = [ sc; si]. asimul computes a simulation of y of length t1 and stores the output in the matrix y .

where

sy

References: [1] Hansen, Lars Peter and T.J. Sargent, Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988.

450

MATLAB Programs

asseta

Purpose: Computes and simulates asset prices for a recursive equilibrium model. Synopsis: asseta is a script file which requires that pay and nt, as well as the output of solvea, reside in memory. Description: Run solvea and

asimul

first. An asset pays out a stream of returns yt = pay ∗ xt

where pay is a vector and where xt is governed by the equilibrium law of motion xt+1 = Ao xt + Cwt+1 The asset is priced by asset price at t = Et

∞ X

β t ptt+j yt+j .

t=0

The program computes the intertemporal marginal rate of substitution, the payoff, the asset price, and the gross rate of return on the asset. A similation of these of length nt is stored in y . The program also calculates the prices of claims on sure j -period forward consumption for j = 1, 2, 5. A simulation of length nt of these for j = 1, 2, 5 are stored in R1, R2, and R5, respectively. References: [1] Hansen, Lars Peter and T.J. Sargent, Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988.

Matlab programs

451

assets

Purpose: Creates matrices and scalars needed to price an asset in a strictly periodic equilibrium model of period p . Synopsis: assets is a script file.

solves

must be run first and its output must be in memory.

Description: An asset with payoff payt = Ua ∗ xt is to be priced, where xt is the state vector for a dynamic linear equilibrium model that is periodic with period p . The asset price at is given by xt = [x′t µa,s(t) xt + σa,s(t) ]/[ij · Mc,s(t) xt ]. This program computes the matrices µa,s(t) and the scalars σa,s(t) for s(t) = 1, 2, . . . , p. These matrices and scalars are stored in memory. To simulate the asset price, use the program assetss. See also: simuls, assetss.

452

MATLAB Programs

assetss

Purpose: Simulates asset price and term structure of interest rates for a strictly periodic equilibrium model with period p . Synopsis: assetss is a script file. The programs solves, first and their outputs must reside in memory.

simuls,

and

assets

must be run

Description: A simulation is constructed for the asset priced in assets. The term structure of interest rates is also computed. The output of the simulation is returned in the vector y , which equals [mrs, pays, as, ret]. Here mrs is the marginal rate of substitution at time, pays is the payoff of the asset, as is the price of the asset and ret is the return on the asset. The prices of risk free claims on comsumption 1, 2, and 5 periods forward are returned in R1, R2, R5, respectively. See also: simuls, solves,

assets

Matlab programs

453

assetx

Purpose: Computes and simulates asset prices for a recursive equilibrium model with Gaussian Exponential Quadratic specification. Synopsis: assetx is a script file which requires that pay and nt , as well as the output of solvex, reside in memory. Description: Run solvex and

asimul

first. An asset pays out a stream of returns yt = pay ∗ xt

where pay is a vector and where xt is governed by the equilibrium law of motion xt+1 = Ao xt + Cwt+1 The asset is priced by asset price at t = Et

∞ X

β t ptt+j yt+j .

t=0

The program computes the intertemporal marginal rate of substitution, the payoff, the asset price, and the gross rate of return on the asset. A similation of these of length nt is stored in y . The program also calculates the prices of claims on sure j -period forward consumption for j = 1, 2, 5. A simulation of length nt of these for j = 1, 2, 5 are stored in R1, R2, and R5, respectively. See also: asseta, solvex

References: [1] Hansen, Lars Peter and T.J. Sargent, Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988.

454

MATLAB Programs

avg

Purpose: Prepares linear system for analysis of aggregation over time with “integrated” or “summed” data Synopsis: [AA, CC] = avg(A, C, m) Description: The state xt evolves according to xt+1 = Axt + Cwt+1 Let zt = [x′t , x′t−1 , . . . , x′t−m+1 ]′ . Then zt evolves according to zt+1 = AA ∗ zt + CC ∗ wt+1 where

A I AA = 0 .. .

0

0 0 I .. .

··· ··· ··· .. .

···

I

The program forms AA and CC . See also: aggreg

0 C 0 0 0 , CC = .. . . .. . 0 0

Matlab programs

455

canonpr

Purpose: Computes canonical representation of preferences. Synopsis: [lamh, pihh] = canonpr (beta, lamba, pih, deltah, thetah)

Description: The program computes a canonical representation of preferences by solving the auxiliary consumer choice problem, maximize ∞ 1 X t β st · st − E0 2 t=0

subject to ht = ∆h ht−1 + Θh ct st = Λht−1 + Πct , h1 given. The solution is a feedback rule ct = −F ht−1 where F = (Π′ Π + βΘ + h′ P Θh )−1 (βΘ′h P ∆h + Π′ Λ), and where P is the nonnegative definite P ˆ Π) ˆ that solves the algebraic Riccati equation for the problem. A canonical (Λ, is chosen for the equations ˆ −1 Λ ˆ =F Π ′ ˆ Π ˆ = (Π′ Π′ + βΘ′ P Θh ). Π h

456

MATLAB Programs

clex 10, 11, 13, 14, 18, 35, 101c, 101f

Purpose: Read in matrices defining an economy. Synopsis: clex*.m is always a script file. Description: Each clex*.m file creates a list of matrices Φc , Φg , Φi , Γ, ∆k , Φk , ∆h , Φh , Γ, Π, A22 , Ud , Ub , and Ud that define an economy. The economies are as follows: clex clex clex clex clex clex clex clex

10 11 13 14 18 35 101c 101f

The Jones-Manuelli examples of chapter 3. Hall’s model of chapter 3. Hall’s model with higher adjustment costs. Lucas’s model of chapter 3. The “seasonal preferences” model of chapter 7. The “heterogeneous agent” example of chapter 6. The hog model of chapter 8. The corn-hog model of chapter 8.

Matlab programs

457

compn

Purpose: compn creates companion matrix. Synopsis: [B] = compn[a] Description: The companion matrix B of the 1 × n row vector a is defined as a1 1 B= 0 .. .

0

a2 0 1 .. .

··· ··· ··· .. .

an−1 0 0 .. .

0

···

1

an 0 0 . .. . 0

458

MATLAB Programs

disthet

Purpose: Compute equilibrium of general equilibrium with two types of households, externalities, distorting taxes, and exogenous government expenditures. Synopsis: disthet is a script file. All matrices must be in memory. Description: disthet computes a competitive equilibrium of a distorted heterogeneous economy. Two types of agents live in an economy with a government. There are externalities. Type i agent’s problem is to maximize: ∞ X β t [(si (t) − bi (t)) · (si (t) − bi (t)) + ℓi (t)2 ]} E0 − 0.5{ t=0

subject to: si (t) = Λi1 hi (t − 1) + Λi2 ∗ H1 (t − 1) + Λi3 H2 (t − 1) + Πi1 ci (t) + Πi2 C1 (t) + Πi3 C2 (t)

hi (t) = ∆hi ∗ hi (t − 1) + ∆Hi1 H1 (t − 1) + ∆Hi2 H2 (t − 1)

+ Θhi ci (t) + ΘHi1 C1 (t) + ΘHi2 C2 (t) X ∞ E{ t=0 β t [(I + τc )p(t) · ci (t) − (1 − τℓ )w(t)li (t) − α(t) · di (t) − fi ∗ (P1 (t) + P2 (t) − Ti (t)]}|I0 − v0 ∗ k0i = 0

where si , hi , ℓi , ci , Ti , Pi are consumption services, household capital stock, labor, consumption, government transfer of type i agent and firms of type i’s profit at t, i=1,2. Capital letters denote aggregate variables. τj is tax on j, j=c,l,k,i. Firms of type 1’s problem is to maximize expected profit: E

∞ X t=0

{β t [p(t)[c(t) + E(t)] + q(t) · i(t) − r(t) · k(t − 1) − α(t) · d(t) − w(t)ℓ(t)]}

subject to: Φc (c(t) + E(t)) + Φi i(t) + Φg g(t) = Γk k(t − 1) + ΓK ∗ K(t − 1) + d(t) g(t) · g(t) = ℓ(t)2

Matlab programs

459

where c(t) = c1 (t)+c2 (t), and similarly for d(t), ℓ(t). E(t), g(t) are government spending and intermediate goods, respectively. Firms of type 2’s problem is to maximize expected profit: E

∞ X t=0

β t [(I − τk )r(t) · k(t − 1) − (I + τi )q(t) · i(t)] − v0 ∗ k0

subject to: k(t) = ∆k k(t − 1) + ∆K K(t − 1) + Θk i(t), where k0 = k01 +k02 . The state vector in this program is defined as [z(t); z(t); h1(t− 1); h2(t − 1); k(t − 1)].

460

MATLAB Programs

dog

Purpose: Compute ‘mongrel’ (i.e., non-Gorman) preference ordering over aggregate consumption for two households. Synopsis: function[Deltah,Thetah,Lambdah,Pih,Am3,Bm3,Cm3]=dog(alpha, beta,lambda1, pih1, deltah1,thetah1,lambda2,pih2,deltah2,thetah2, a22,c2,ub1,ub2) Description: Computes the canonical mongrel service technology for two households with parameter alpha (the Pareto weight on the first consumer). The mongrel household technology is H(t) = ∆h H(t − 1) + Θh c(t) s(t) = ΛH(t − 1) + Πc(t)

The mongrel preference shock is given by the series connection of the three state space systems (A1,B1,C1,D1), (AA,BB,GG,HH), (∆h , Θh , Λ, Π). We calculate a system representation (Am,Bm,Cm,Dm) for the mongrel shock. The mongrel shock is thus described by Z(t + 1) = Am2 Z(t) + Bm3 w(t + 1) bb(t) = Cm3 Z(t). In using this program, it is important to set the initial condition for the state appropriately. The given initial conditions for h01 and h02 are loaded into the SHOCK process, and the initial conditions for the MONGREL h01,h02 are set to zero. Type [Amm,Bmm,Cmm,Dmm]=minreal(Am,Bm,Cm,Dm) to find minimal realization for preference shock.

Matlab programs

461

doubleo

Purpose: Computes time invariant Kalman filter or time invariant linear optimal control. Synopsis: [K, S] = double (A, C, Q, R) Description: The program creates the Kalman filter for the following system: xt+1 = Axt + et+1 yt = Cxt + vt where Eet+1 e′t+1 = Q, Evt vt′ = R , and vt is orthogonal to et for all t and s . Here A is n × n , C is k × n , Q is n × n , and R is k × k . The program creates the observer system x ˆt+1 = Aˆ xt + Kat yt = C x ˆt + at , where K is the Kalman gain, and S = E(xt − x ˆt )(xt − x ˆt )′ where x ˆt = Ext | yt+1 yt−2 , . . . . Also, at = yt − Eyt | yt−1 , yt−2 , . . . . By using duality, the program can be used to solve optimal linear control problems. Let the control problem be to choose a feedback law ut = −F xt to maximize ∞ X {x′t Qxt + u′t Rut } − t=0

subject to

xt+1 = A′ xt + B ′ ut , with x0 given. The optimum control is then given by F = K ′ , where [K, S] = double (A, B, Q, R) and where the optimal value function is x′t Sxt . The doubling algorithm is used to compute the solution. See also: mult and double3. References: [1] Anderson, B.D.O., and J. Moore, Optimal Filtering, 1979, p. 160.

462

MATLAB Programs

doublex

Purpose: Solves recursive undiscounted Gaussian Quadratic Exponential control problem. Synopsis: [K, S, ST ] = doublex (A, C, Q, R, c, sig) Description: This program uses the “doubling algorithm” to solve the Riccati matrix difference equations associated with the undiscounted quadratic-Gauusian linear optimal control problems. The control problem has the form S(t) = max{x(t)′ Qx(t) + u(t)′ Ru(t) + (2/σ) log Et exp(σ/2)S(x(t + 1))}, u(t)

subject to x(t + 1) = A′ x(t) + C ′ u(t) + cw(t + 1), where w(t+1) is a Gaussian martingale difference sequence with unit covariance matrix. The program returns the steady state value function in S . The optimal control law is u(t) = −K ′ ∗ x(t) The program also returns ST , which is the quadratic form in Et exp(sig/2)S(x(t + 1)). See also: mult and double and

solvex.

References: [1] Anderson, B.D.O., and J. Moore, Optimal Filtering, 1979, p. 160. [2] Jacobson, D.H. “Optimal stochastic linear systems with exponential performance criteria and their relation to deterministic differential games.”IEEE Transactions on Automatic Control, AC-18, 124–31.

Matlab programs

463

doublej

Purpose: Computes infinite matrix sums of squares. Synopsis: V = double (a1 , b1 ) Description: The program computes the infinite sum V in V =

∞ X

aj1 b1 aj′ ,

j=0

where a1 and b1 are each n × n matrices. The program iterates to convergence on the following doubling algorithm, starting from V0 = 0: a1j = a1j−1 ∗ a1j−1

Vj = Vj−1 + a1j−1 ∗ Vj−1 ∗ a1j−1 .

The limiting value of Vj is returned in V . References: [1] Hansen, Lars P. and Thomas J. Sargent Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988.

464

MATLAB Programs

doublej2

Purpose: Computes infinite matrix sums of squares. Synopsis: [V ] = doublej2 (a1 , b1 , a2 , b2 ) Description: The program computes the infinite sum V in V =

∞ X

aj1 (b1 b2 )aj2

j=o

where a1 and a2 are each n × n matrices, b1 is n × k and b2 is k × n . The program iterates to convergence on the following doubling algorithm, starting from V0 = 0: a1j = a1j−1 ∗ a1j−1 a2j = a2j−1 ∗ a2j−1

Vj = Vj−1 + a1j−1 Vj−1 a2j−1 .

The limit point is returned in V . References: [1] Hansen, Lars P. and Thomas J. Sargent Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988.

Matlab programs

465

double3

Purpose: Raw doubling algorithm for raising a symplectic matrix to higher and higher powers. Synopsis: [aa, bb, gg] =

double3

(a, b, g)

Description: The algorithm iterates to convergence of gj in the following recursions: aj+1 = aj (I + bj gj )−1 gj gj+1 = gj + a′j gj (I + bj gj )−1 aj , bj+1 = bj + aj (I + bj gj )−1 bj a′j where aj , bj , gj are each n × n matrices. If we let Ej , be the symplectic matrix

a−1 j gj a−1 j

a−1 j bj ′ aj + gj a−1 j bj

j

then Ej = (E0 )2 . References: [1] Anderson, B.D.O., and J. Moore, Optimal Filtering, 1979, p. 160. [2] Hansen, Lars Peter and T.J. Sargent, Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988.

466

MATLAB Programs

heter

Purpose: Computes allocation to an individual who lives within a recursive linear equilibrium model. Synopsis: heter is a script file. The program solvea must be run first, and its inputs and outputs must be in memory. The matrices Udc and Ubi , and the scalars k0i, h0i, and tol > 0 must all be in memory. Description: The consumer maximizes ∞

1 X t i −( )E β [(st − bit ) · (sit − bit ) + ℓ2t ], 0 < β < 1 2 t=o subject to sit = Λhit−1 + Πcit hit = ∆h hit−1 + Θh cit ∞ ∞ X X β t p0t cit | Io = E β t (wt0 ℓit + αt0 dit ) | Io E t=0

bit dit

t=0

i + υ0 k−1 = Ubi zt = Udi zt

i where k−1 = k0i, hi−1 = h0i are parameters to be fed in. The matrices Udi = udi and Ubi = ubi must also be fed in. The parameter tol > 0 must be fed in. The program computes the optimal solution for consumer i in the form cit = Sci xt , hit = Shi xt , sit = Ssi xt , bit = Sbi xt , dit = Sdi xt , where xt is the state variable of the economy augmented by i the state variables kt−1 , hit−1 ideosyncratic to the individual. The program also computes the aggregate allocations ct = Sca xt , ht = Sha xt , and so on. The individual allocations are determined by the matrices sci, shi, . . . , which are placed in memory. The aggregate allocation are placed in the matrices sca, sha, . . . , which are placed in memory.

Algorithm:

Matlab programs

See Hansen and Sargent, Chapter 6 See also: simulh

467

468

MATLAB Programs

innov

Purpose: Compute the innovations representation for a recursive linear model whose observations are corrupted by first-order serially correlated measurement errors. Synopsis: [aa, bb, cc, dd, V1 ] = innov (ao, c, sy, D, R ) Description: The model is assumed to have the state space representation xt+1 = ao xt + cwt+1 yt = Sy xt + et+1 where wt is a white noise with Ewt wt ′ = I and et is a measurement error process governed by et+1 = Det + ηt+1 where ηt is a white noise with contemporaneous covariance matrix R . The matrices R and D must each be m × m where [m, n] = size(Sy ). The program forms the innovations representation for yt , zˆt+1 = aaˆ zt + bbut yt = ccˆ zt + ddut where ut = yt+1 − E[yt+1 | yt , yt−1 , . . .], and Eut u′t = V1 . Algorithm:

ao aa = GG

0 k1 , bb = D I

cc = [0 I], dd = [0], where k1 is the Kalman gain associated with the Kalman filter for the original system. References: [1] Sargent, Thomas, “Two Models of Measurements and the Investment Accelerator,” Journal of Political Economy, April 1989.

Matlab programs

mult

Purpose: Multiplies two symplectic matrices. Synopsis: [a, b, g] = mult (a1 , b1 , g1 , a2 , b2 , g2 ) Description: A symplectic matrix Ei is represented in the form (∗)

Ei =

a−1 i gi a−1 i

a−1 i bi . a′i + gi a−1 i bi

We desire to form E = E2 E1 . We can compute a = a2 (I + b1 g2 )−1 g1 g = g1 + a′1 g2 (I + b1 g2 )−1 a1 b = b2 + a2 (I + b1 g2 )−1 b1 a′2 ,

and represent E as in representation (∗). References: [1] Anderson, B.D.O., and J. Moore, Optimal Filtering, 1979, p. 160.

469

470

MATLAB Programs

seasla

Purpose: Creates a time invariant representation for a strictly periodic, time varying linear equilibrium model Synopsis: seasla is a script file, which requires that the output of

simuls

reside in memory.

Description: Let xt be the state vector for a strictly periodic seasonal process of period p . Let Xt′ = [x′pt−p+1 , x′pt−p+2 , . . . , x′pt ]. The law of motion for Xt is ˆ t + CW ˆ t+1 Xt+1 = AX where Wt+1 is a vector white noise and Aˆ and Cˆ are defined as simuls. The ˆ −1 Cˆ Cˆ ′ (I − spectral density matrix of the Xt process is given by S(z) = (I − AZ) ˆ −1′ . Embedded in the spectral density matrix of the stacked process Xt AZ) are the spectral density matrices s1 (z), s2 (z), . . . , sp (z) for the periodic process {xt }. The process xt whose spectral density is defined to be s(z) = Pp p−1 k=1 sk (z). It can be shown that (∗)

ˆ p )−1 Cˆ Cˆ ′ (I − Az ˆ −p )−1′ , s(z) = p−1 Q(z)(I − Az

where Q(z) = [I zI · · · z p−1 I]. A state space representation for a process xt with spectral density matrix (*) is (†)

ˆ t + CV ˆ t+p Yt+p = AY xt = p−.5 Q(L)Yt

where Vt is a vector white noise with identity covariance matrix. The program seasla creates the spectral density for a univariate process that is a linear function of the state. Let the process be ct = scs(t) xt . We form the time invariant, averaged process, as c˜t which is determined by the system (‡)

ˆ t + CV ˆ t+p Yt+p = AY ct = p−1 Qc (L)Yt

. . . where Qc (z) = [sc1 .. sc2 z .. · · · ..scp z p−1 ].

Matlab programs

471

The program maps into a first-order system, then deduces the impulse response of ct with respect to innovations Vt corresponding to representation (‡). This is stored in z1 . The program also uses the Kalman filter to obtain an innovations representation corresponding to (‡), and returns the impulse response of ct with respect to the innovation in ct in the vector z2 . See also: simuls, assets,

assetss

472

MATLAB Programs

seas1

Purpose: To aid in creating the matrices that define a periodic recursive linear equilibrium model. Synopsis: seas1.m is a script file. Description: seas1 creates matrices Φcs(t) , Φgs(t) , Φis(t) , Γs(t) , ∆ks(t) , ∆hs(t) , A22s(t) , Cs(t) , Φks(t) , Φhs(t) , Λs(t) , and Πs(t) that are needed to define a periodic linear recursive model. It creates time invariant versions of these matrices as follows. It first reads in Φc , Φi , Φy , Γ, ∆k , ∆h , A22 , C, Φk , Φh , Λ1 and Π for a time invariant economy. One of our clex*.m files can be used to read in such matrices. Then seas1 simply sets the matrices Φcs(t) = Φc , and so on. To create a periodic model, the user may find it useful to run seas1 first, and then to modify the resulting time invariant setup, rather than building up all of the matrices from scratch. In a typical periodic model, many of the matrices may in fact be time invariant. See also: solves.m

Matlab programs

473

simpulse

Purpose: Creates different impulse response functions for a periodic linear equilibrium model. Synopsis: simpulse is a script file. memory.

solves

must be run first, and its outputs must be in

Description: A stacked version of a periodic model has state space form (†)

ˆ t + CW ˆ t+1 Xt+1 = AX Yt = HXt ,

′ ′ ′ where Xt′ = [x′tp−p+1 , x′tp−p+2 , . . . , x′tp ], Yt′ = [ytp−p+1 , . . . , ytp ], Wt′ = [wtp−p+1 , ′ ˆ ˆ . . . , wtp ] and where A, C , are as defined in simuls.

This program first uses dimpulse to compute the impulse response function of the stacked system (†). From this impulse response function, it forms two impulse response functions for the periodic process yt . First, it computes {dk,υ } in the representation ypt−p+k =

∞ X

dk,υ wpt−p+k−υ .

υ=0

This is the response of ypt−p+k (i.e., yt in a particular season) to lagged w ’s. Second, the program computes the {hk,υ } that give the response of {yt } to wpt−p+k (i.e. an innovation in a particular season). The value of p must be in memory. The program prompts the user for the index of the innovation whose response functions are to be computed. See also: solves, simuls

474

MATLAB Programs

simulh

Purpose: Simulates allocation of individual i who lives within a recursive linear equilibrium model. Synopsis: simulh is a script file.

heter

must be run first and its output must be in memory.

Description: The user is asked to specify which series he wants to simulate; e.g., to simulate the consumption allocation to agent i and the aggregate consumption allocation, respond [sci; sca]. See also: heter

Matlab programs

475

simulhet

Purpose: Simulates heterogeneous agent economy. Synopsis: simulhet is a script file, not a function. Description: Simulates the prices and quantities for a recursive linear equilibrium model with non-Gorman heterogeneity. solvehet must be run first and its output must be in memory. To simulate the individual consumption allocations, set sy=[sc1;sc2] when asked what series you want to simulate. To simulate the individual consumption service allocations, set sy= [ss1;ss2]. See also: solvea and heter.

476

MATLAB Programs

simuls

Purpose: To simulate a strictly periodic recursive linear model. Synopsis: simuls is a script file, which requires that all of the outputs of solves be in memory. simuls prompts the user for the number of “years” to simulate. Description: simuls creates a simulation of the state vector xt for a strictly periodic model of period p . The stacked state vector Xt is formed, where Xt′ = [x′pt−p+1 , x′pt−p+2 , . . . , x′pt ]. The law of motion for Xt is ˆ t + CW ˆ t+1 Xt+1 = AX where Aˆ = D−1 F, Cˆ = D−1 G , where I −A01 D= 0 . .. 0

0 I −A02 .. .

0 0 I .. .

··· ··· ··· .. .

0 0 0 .. .

0

0

···

−A0p−1

0 F = 0

A0p 0

Cp 0 G = .. .

0

Wt′

1

0 C1 .. .

0 ··· 0 ··· .. . . . .

0 0 .. .

0

0 ···

Cp−1

′ ′ ′ ]′ . [wpt+1 , wpt+2 , . . . , wpt+p

0 0 0 .. .

,

= and where The output of simuls is returned in the matrix X . The matrix X is arranged as follows: x′1 x′2 ··· x′p x′p+1 x′p+2 · · · x′2p X= . .. .. .. .. . . . ′ ′ ′ xT p+1 xT p+2 · · · xT p+p

Matlab programs

477

where T is the number of “years” specified by the user. To simulate ct , it , etc., the user can write a program to put the relevant linear combinations off X . Alternatively, the user can edit the files simulc, simulk, simuli, simulg, simulb, or simuld.

478

MATLAB Programs

solvea

Purpose: Computes solution of recursive linear equilibrium model Synopsis: solvea is a script file. The matrices A22 , C2 , Ud , Ub , Φc , Φg , Φi , Γ, ∆k , Θk , ∆h , Θh , Λ, and Π and the scalar β must be in memory. Description: The social planning problem is to maximize ∞

1 X t −( )E β [(st − bt ) · (st − bt ) + ℓ2t ], 0 < β < 1 2 t=0 subject to Φc ct + Φg gt + Φi it = Γkt−1 + dt gt · gt = ℓ2t

kt = ∆k kt−1 + Θk it

ht = ∆h ht−1 + Θh ct st = Λht−1 + Πct zt+1 = A22 zt + C2 wt+1 bt = Ub zt , dt = Ud zt Here st is consumption services, bt a stochastic bliss process, ℓt is labor services, ct is consumption rates, gt is “intermediate goods”, it is investment goods, dt is an endowment shock process, kt is physical capital, ht is household capital, zt is a vector of exogenous information variables, and wt+1 is a martingale difference sequence. Each of these is a vector, except for ℓt , which is scalar. Let ′ xt = [h′t−1 , kt−1 , zt′ ]′ . The program computes the solution of the social planning problem in the form xt+1 = Ao xt + Cwt+1 kt = Sk xt , gt = Sq xt ht = Sh xt , it = Si xt ct = Sc xt , bt = Sb xt st = Ss xt , dt = Sd xt

Matlab programs

479

The program also computes Lagrange multipliers µjt = Mj xt for variable j = k, c, h, s, i. The program computes and leaves in memory Ao , C, Sj (for j = k, h, c, s, g, i, b, and d ) and Mj (for j = k, c, h, s, and i). Algorithm: The social planning problem is formulated and solved as an optimal linear regulator problem. References: [1] Hansen, Lars Peter and T.J. Sargent, Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988.

480

MATLAB Programs

solvdist

Purpose: Computes equilibrium of representative agent economy with distorting taxes, exogenous govenment expenditures, and externalities. Synopsis: solvdist is a script file. Description: solvdist, a script file (not a function), finds a competitive equilibrium for a representative agent economy with distortions. Households maximize: X E0 − .5 β t [(s(t) − b(t)) · (s(t) − b(t)) + ℓ(t)2 ] subject to g(t) · g(t) = ℓ(t)2

z(t + 1) = a22 ∗ z(t) + c2 ∗ w(t + 1) b(t) = ub ∗ z(t),

d(t) = ud ∗ z(t),

E(t) = ue ∗ z(t)

h(t) = ∆h h(t − 1) + ∆H H(t − 1) + Θh c(t) + θH C(t) s(t) = Λh h(t − 1) + ΛH H(t − 1) + Πc c(t) + ΠC C(t)

∞ X t=0

k(t) = ∆k k(t − 1) + Θk i(t)

β t [(I + τc )p(t) · c(t) + (I + τi )q(t) · i(t) − (1 − τℓ )w(t) · g(t)

−α(t) · (d(t) + γK K(t − 1)) − (I − τk )r(t) · k(t − 1) − T (t)] = 0 Firms maximize profits: ∞ X β t [p(t) · (c(t) + E(t)) + q(t) · i(t) − r(t) · k(t − 1) − α(t) · d(t) − w(t) ∗ g(t)] E0 t=0

subject to g(t) · g(t) = ℓ(t)2

Φc (c(t) + E(t)) + Φg g(t) + Φi i(t) = Γk k(t − 1) + ΓK K(t − 1) + d(t)

Where x(t) = [h(t − 1)′ , k(t − 1)′ , z(t)′ ]′ , the solution of the problem is x(t + 1) = ao ∗ x(t) + c ∗ w(t + 1)

j(t) = sj ∗ x(t), where j = k, h, c, e, s, g, i, b, d, p, q, w, r, α . The program also computes the household’s Lagrange multipliers µj = mj x(t) for j = k, h, s, z . (µ0 is set to 1.)

Matlab programs

481

solvehet

Purpose: Solves Pareto problem for two-agent, non-Gorman preferences. Description: solvehet solves the pareto problem for two agents with heterogeneous household production functions, i.e. to maximize E

∞ X t=0

β t − .5α([(s1 (t) − b1 (t)) · (s1 (t) − b1 (t)) + ℓ1 (t)2 ]) + (1 − α)([(s2 (t) − b2 (t)).(s2 (t) − b2 (t)) + ℓ2 (t)2 ])

subject to Φc c(t) + Φg g(t) + Φi i(t) = Γk(t − 1) + d(t) gi (t) · gi (t) = ℓi (t)2 ,

i = 1, 2

g1 (t) + g2 (t) = g(t)

k(t) = ∆k k(t − 1) + Θk i(t)

hi (t) = ∆hi hi (t − 1) + Θhi ci (t)

si (t) = Λi ∗ hi (t − 1) + Πhi ci (t),

i = 1, 2

c1 (t) + c2 (t) = c(t)

z(t + 1) = a22 ∗ z(t) + c2 ∗ w(t + 1) i(t) = ubi ∗ z(t),

i = 1, 2; d(t) = ud ∗ z(t)

The state vector is x(t) = [h1(t-1)’,h2(t-1)’,k(t-1)’,z(t)]’. The control vector is u(t) = [c1(t)’,i(t)’]. The solution of the problem is given by: x(t + 1) = ao ∗ x(t) + c ∗ w(t + 1) j(t) = sj x(t)

for j = k,c,g,i,d, and ci,bi,gi,hi,si, for i=1,2. The program also computes Lagrange multipliers.

482

MATLAB Programs

solvex

Purpose: Computes solution of recursive linear equilibrium model with Gaussian Quadratic Exponential preference specification. Synopsis: solvex is a script file. The matrices A22 , C2 , Ud , Ub , Φc , Φg , Φi , Γ, ∆k , Θk , ∆h , Θh , Λ, and Π and the scalars σ and β must be in memory. Description: ′ Let xt = [h′t−1 , kt−1 , zt′ ]′ , and let the law of motion for xt be xt+1 = Axt + But + Cwt+1 . The social planning problem is to find a value function V (x(t)) = max{−.5[(s(t) − b(t)).(s(t) − b(t)) + l(t)2 ] +β ∗ (2/σ) ∗ log Et exp(σ/2 ∗ (V (x(t + 1))}

subject to Φc ct + Φg gt + Φi it = Γkt−1 + dt gt · gt = ℓ2t

kt = ∆k kt−1 + Θk it

ht = ∆h ht−1 + Θh ct st = Λht−1 + Πct zt+1 = A22 zt + C2 wt+1 bt = Ub zt , dt = Ud zt Here st is consumption services, bt a stochastic bliss process, ℓt is labor services, ct is consumption rates, gt is “intermediate goods”, it is investment goods, dt is an endowment shock process, kt is physical capital, ht is household capital, zt is a vector of exogenous information variables, and wt+1 is a martingale difference sequence. Each of these is a vector, except for ℓt , which is scalar. The program computes the solution of the social planning problem in the form xt+1 = Ao xt + Cwt+1 kt = Sk xt , gt = Sq xt ht = Sh xt , it = Si xt ct = Sc xt , bt = Sb xt st = Ss xt , dt = Sd xt

Matlab programs

483

The program also computes Lagrange multipliers µjt = Mj xt for variable j = k, c, h, s, i. The program computes and leaves in memory Ao , C, Sj (for j = k, h, c, s, g, i, b, and d) and Mj (for j = k, c, h, s, and i). Algorithm: The social planning problem is formulated and solved using

doublex.

References: [1] Hansen, Lars Peter and T.J. Sargent, Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988.

484

MATLAB Programs

solves

Purpose: Computes the solution of recursive linear equilibrium model with periodic coefficients. Synopsis: solves is a script file. The matrices A22s(t) , C2s(t) , Ud , Ub , Φcs(t) , Φgs(t) , Φis(t) , Γs(t) , ∆ks(t) , Θks(t) , ∆hs(t) , Θhs(t) , Λs(t) , and Πs(t) and the scalar β must be in memory.

Description: The social planning problem is to maximize ∞

1 X t −( )E β [(st − bt ) · (st − bt ) + ℓ2t ], 0 < β < 1 2 t=0 subject to Φcs(t) ct + Φgs(t) gt + Φis(t) it = Γs(t) kt−1 + dt gt · gt = ℓ2t

kt = ∆ks(t) kt−1 + Φks(t) it ht = ∆hs(t) ht−1 + Φhs(t) ct st = Λs(t) ht−1 + Πs(t) ct zt+1 = A22s(t) zt + C2s(t) wt+1 bt = Ub zt , dt = zt . where s(t+p) = s(t), where p is the period of the model. Here st is consumption services, bt is a stochastic bliss process, ℓt is labor services, ct is a vector of consumption rates, gt is “intermediate goods”, it is investment goods, dt is an endowment shock process, kt is physical capital, ht is household capital, zt is a vector of exogenous information variables, and wt+1 is a martingale difference sequence. Each of these is a vector, except for ℓt , which is a scalar. ′ Let xt = [h′t−1 , kt−1 , zt′ ]′ . The program computes the solution of the social

Matlab programs

485

planning problem in the form xt+1 = Aos(t) xt + Cs(t) wt+1 kt = Sks(t) xt , gt = Sgs(t) xt ht = Shs(t) xt , it = Sis(t) xt ct = Scs(t) xt , bt = Sbs(t) xt st = Sss(t) xt , dt = Sds(t) xt The program also computes Lagrange multipliers µjt = Mjs(t) xt for variables j = k, c, h, s, i. The program computes and leaves in memory Aos(t) , Cs(t) , Sjs(t) for j = k, h, c, s, g, i, b and d , and Mj for j = k, c, h, s, and i. The user is advised to use the Matlab program seas1 an an aid in creating the matrices that must be fed into solves. The user must edit solves to set the period p . Also, it will vastly accelerate computations if the user will load either the file seas4.mat (in the case p = 4) or the file seas12.met (in the case p = 4). The lines to edit occur immediately after the information provided by the help command, i.e. the first lines without %. Algorithm: The social planning problem is formulated as a periodic optimal linear regulator problem and solved using doubling algorithms described by Hansen and Sargent. References: [1] Hansen, Lars Peter and T.J. Sargent, Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988.

486

MATLAB Programs

spectr1

Purpose: Computes spectral density of endogenous variables of a dynamic linear equilibrium model. Synopsis: spectr1 is a script file. The matrices ao, c, sy, R, D , and the scalar nnc must be in memory. Description: The equilibrium model is of the form xt+1 = ao xt + cwt+1 yt = sy xt + vt vt+1 = Dvt + ut+1 where Ewt wt′ = I, E, ut u′t = R . The constant corresponds to row number nnc of the state vector xt . The eigenvalues of D and the eigenvalues of A (except for the unit eigenvalue associated with the constant term) must be less than unity in modulus. spectr1 computes the spectral densities variables in yt . Algorithm: spectr1 deletes the nncth row and/or column of ao, c, and sy , which correspond to the constant term. Then spectral is used to compute the spectral density matrix of yt . See also: spectral

Matlab programs

487

spectral

Purpose: Computes spectral density matrix for a linear system. Synopsis: spectral, a script file. The inputs A, C, G, D, R and T must reside in memory. Description: Let the system be xt+1 = Axt + Cet+1 yt = Gxt + vt vt+1 = Dvt + ut Eet e′t

I, Eut u′t

where = = R , and where et and us are orthogonal for all t and s . The vector yt is rg × 1. The spectral density matrix for y is computed for ordinates ωj = 2πj/T, j = o, 1, . . . , T − 1. The spectral density matrix for ordinate j is stored in Syj, j = 0, 1, . . . , T −1. The spectral densities (diagonals of the spectral density matrices) are stored in the matrix S . The matrix S has rg rows and T columns, and S(k, j) = Syj(k, k). The eigenvalues of A and D must all be less than unity in modulus. Algorithm: Let Sy(ωj ) be the spectral density matrix at frequency wj . Then Sy(ωj ) = G(I − Ae−iω−j )−1 CC ′ (I − Aeiωj )−1 G′ + (I − De−iωj )−1 R(I − De+iωj )−1

.

488

MATLAB Programs

spectrs

Purpose: Computes spectral density matrix for set of variables determined by a periodic linear equilibrium model. Synopsis: spectrs is a script file. solves and simuls must be run first, and their outputs must reside in memory. The integer nnc (the index of the constant in the state vector) must be in memory. Description: The spectral density of a process yt with hidden periodicity p is given by the Tiao-Grupe formula ˆ p )−1 Cˆ C(I ˆ − Az ˆ −p )−1′ H ′ Q(z −1 )′ , Sy (z) = Q(z)H(I − Az ˆ Cˆ , and H ˆ are from the stacked state space system where z = e−iωj ; where A, ˆ t + CW ˆ t+1 Xt+1 = AX Yt = HXt , ′ ′ and where Xt′ = [x′pt−p+1 , x′pt−p+2 , . . . , x′pt ], Yt′ = [ypt−p+1 , ypt−p+2 , ′ ′ ′ ′ ′ . . . , ypt ], Wt = [wpt−p+1 , wpt−p+2 , . . . , wpt ] . The program returns the spectral density matrices for frequencies ωj = 2πj/T , for j = 0, 1, . . . , T − 1 in the matrices Sy0, Sy1, . . . , SyT − 1. The spectral densities of the individual series are returned in the matrix S . The user can edit the file to specify T and the particular series whose spectrum is computed.

See also: solves, simuls

Matlab programs

489

steadst

Purpose: steadst computes steady state values of observable variables determined by a recursive linear equilibrium model. Synopsis: steadst is a script file which requires that the scalar nnc and matrices ao, sc, ss, si, sd, sb, sk, sh reside in memory. Description: The equilibrium model is represented as xt+1 = ao ∗ xt + c ∗ wt+1 yt = Gxt

where G = [sc; ss; si; sd; sb; sk; sh]. The integer nnc gives the row in the state vector xt that corresponds to the constant term. steadst assumes that except for the eigenvalue associated with the constant term, all eigenvalues of ao are less than unity in modulus. The program calculates the steady state value of xt , putting its value in zs . Then the program successively calculates the steady state values of c, s, i, d, b, k, and h , which are the components of y . Algorithm: The steady state value of x is obtained as a basis vector for the null space of (I −ao), normalized so that the component corresponding to the constant equals unity. References: [1] Hansen, Lars Peter and T.J. Sargent, Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988. See also: null

490

MATLAB Programs

steadsts

Purpose: Computes seasonal steady states and seasonal means for a periodic recursive linear equilibrium model. Synopsis: steadsts is a script file. solves and simuls must be run first, and their outputs must be in memory. So must nnc, the index of the constant term in the state vector. Description: The equilibrium for the stacked version of a periodic model can be represented as ˆ t + CW ˆ t+1 Xt+1 = AX where Xt′ = [x′tp−p+1,..., x′tp ], Wt′ = [wpt−p+1,..., wpt ]. The program computes the ˆ which gives the steady for Xt = X ¯ . Then seasonal means null space of (I − A), ¯ by matrices formed for individual variables are formed by pre-multiplying X from appropriate seasonal decision rules. The user must edit the file to compute seasonal means of the particular variables he is interested in. See also: steadst, solves,

simuls.

Matlab programs

491

vardec

Purpose: Calculates variance of k -step ahead prediction errors in zt for k = 1, 2, . . . , N for an “innovations system”. Synopsis: [tab] = vardec (A, C, K, V, N ) Description: Consider the innovations system xt+1 = Axt + Kut zt = Cxt + ut Eut u′t = V vardec prepares a table of diagonal elements of the covariance matrices of k -step ahead errors in predicting zt , k = 1, . . . , N . The output is returned in tab, which has N rows and max(size( V )) columns. The (k, h) element of tab gives the variance of the k -step ahead prediction errors for the hth variable in zt .

Algorithm: Let the covariance matrix of k -step ahead prediction error in z be Vk . Then V1 = V V2 = CKV K ′ C ′ + V Vk = Vk−1 + CAk−1 KV K ′ Ak−1 C ′ . References: [1] Sims, Christopher “Macroeconomics and Reality,” Econometrica, 1980. [2] Hansen, Lars Peter and Thomas Sargent, Recursive Linear Models of Dynamic Economies, (manuscript), Dec. 1988.

492

MATLAB Programs

vardeci

Purpose: Compute decomposition of k -step ahead prediction error variances for an “innovations system”. Synopsis: [tab ] = vardeci (A, C, K, V, N, j) Description: Consider an innovations system xt+1 = Axt + Kut zt = Cxt + ut ′

where Eut ut = V. Let r′ r = V be a Cholesky decomposition of V . Form the innovations system with orthogonalized innovations xt+1 = Axt + Bvt zt = Cxt + Dvt where B = K · r′ , D = r′ , and Evt vt′ = I . The program prepares a table of the part of the diagonal elements of the covariance matrix of the k -step ahead prediction errors, k = 1, . . . , N , that is attributable to the j th innovation. The table is returned in tab , which has dimension N × max(size(V )). The (k, h) element of tab gives the variance in the k -step ahead variance in predicting the hth component of z due to the j th orthogonalized innovation in vt . Algorithm: Let Sj be a selector matrix for j , equal to an m × m matrix of zeros except of a one in the (j, j) element. Let Vk be the covariance of the k -step ahead prediction error in z due the jth orthogonalized innovation. The Vk are calculated using the ecursions V1 = DSj Sj′ D′ V2 = CBSj Sj′ B ′ C ′ + V1 Vk = Vk−1 + CAk−1 BSj Sj′ B ′ A′

k−1

C′

References: [1] Sims, Christopher “Macroeconomics and Reality,” Econometrica, 1980. [2] Hansen, Lars Peter and Thomas Sargent, Recursive Linear Models of Dynamic Economies.

Matlab programs

493

varma

Purpose: varma computes an innovations representation for a recursive linear model whose observations are corrupted by first-order serially correlated measurement errors. Synopsis: varma is a script file, which requires that the matrices ao, c, sy, D , and R reside in memory. Description: The model is assumed to have the state space representation xt+1 = ao ∗ xt + c ∗ wt+1 yt = Sy ∗ xt + et+1

where wt is a white noise with Ewt wt ′ = I, and et is a measurement error process governed by et+1 = Det + ηt+1 , where ηt+1 is a vector white noise with contemporaneous covariance matrix R . The matrices R and D must each be m × m , where [m, n] = size(sy). The program uses the Kalman filter to form the innovations representation x ˆt+1 = aoˆ xt + k1 ∗ ut y˜t+1 = GGxt + ut

where GG = [syao − DSy ], y˜t = yt+1 − Dyt , and ut is the innovation in yt+1 , ut = yt+1 − E[yt+1 | yt , yt−1 , . . .]. The program uses evardec to compute a decomposition of variance for the innovations system. References: [1] Hansen, Lars Peter and T.J. Sargent, Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988. [2] Sargent, Thomas, “Two Models of Measurements and the Investment Accelerator,” Journal of Political Economy, April 1989.

494

MATLAB Programs

varma2

Purpose: varma2 creates impulse response functions associated with an innovations representation. Synopsis: varma2 is a script file which requires that the matrices aa, bb, cc, dd, V 1 be in memory. Description: varma2 takes the output of innov and creates impulse response functions of y with respect to components of u . Impulse response functions with respect to −1 the orthogonalized innovations vt = r′ ut are also computed, where r′ r = V 1 is a Cholesky decomposition of V1. Algorithm: dimpulse is applied. References: [1] Sims, Christopher, “Macroeconomics and Reality,” Econometrica, 1980.

Matlab programs

495

varrep

Purpose: Computes a vector autoregressive representation from a state space model with serially correlated measurement errors. Synopsis: function [AA,V1]=varrep(ao,c,sy,D,R,nj,nnc) Description: Computes (an infinite order) vector autoregressive representation for a recursive linear model whose observations are corrupted by first-order serially correlated measurement errors. The model occurs in the state space form x(t + 1) = ao ∗ x(t) + c ∗ w(t + 1) y(t) = sy ∗ x(t) + e(t + 1)

where e(t) is a measurement error process e(t + 1) = D ∗ e(t) + ee(t + 1) and where ee(t+1) is a vector white noise with covariance matrix R. We assume that ee(t+1) and w(t+1) are orthogonal at all leads and lags. The program computes the autoregressive representation y(t) =

∞ X j=1

A(j)y(t − j) + a(t)

where a(t)=y(t) - E[y(t)— y(t-1),y(t-2),...], and the A(j) are square matrices. The program creates the covariance matrix of a, which it stores in V1. The program returns nj of the matrices A(j), stacked into the ((m times nj) by m) matrix AA, where m is the number of rows of y. A(j) occurs in rows ((j-1)*m+1) to row j*m of AA. nnc is the location of the constant term in the state vector.

496

MATLAB Programs

white1

Purpose: Creates a state space system [AA,BB,CC,DD] that accepts the innovation to the (information) state vector wt+1 as in input and puts out the innovation ut to yt as an output. Synopsis: function[AA,BB,CC,DD]=white1(ao,c,sy,D,R) Description: The program couples the systems x(t + 1) = ao ∗ x(t) + c ∗ w(t + 1) y(t) = sy ∗ x(t) + v(t)

v(t) = D ∗ v(t − 1) + η(t)

Eη(t)η(t)′ = R. and

xh(t + 1) = (ao − k1 ∗ GG) ∗ xh(t) + k1 ∗ y(t) u(t) = −GG ∗ xh(t) + u(t)

where w(t+1) is the innovation to agents’ information sets and where u(t) is the fundamental (Wold) representation innovation. A (minimum-realization) state space system [AA,BB,CC,DD] for the coupled system is returned. To compute the impulse response function, use dimpulse.

Matlab programs

497

white2

Purpose: Creates the state space system [AA,BB,CC,DD] that accepts the measurement error v(t + 1) as in input and puts out the innovation u(t) to y(t) as an output. Synopsis: function[AA,BB,CC,DD]=white2(ao,c,sy,D,R) Description: The program couples the systems x(t + 1) = ao ∗ x(t) + c ∗ w(t + 1) y(t) = sy ∗ x(t) + v(t)

v(t) = D ∗ v(t − 1) + η(t)

Eη(t)η(t)′ = R. and

xh(t + 1) = (ao − k1 ∗ GG) ∗ xh(t) + k1 ∗ y(t) u(t) = −GG ∗ xh(t) + u(t)

where w(t + 1) is the innovation to agents’ information sets, η(t) is the innovation to measurement error, and where u(t) is the fundamental (Wold) representation innovation. The (minimum-realization) state space system [AA,BB,CC,DD] for the coupled system is returned. To compute the impulse response function, use dimpulse.

References Akaike, H. (1973). ‘Information Theory and an Extension of the Likelihood Principle’. In B.N. Petrov and F. Csaki (eds.), Proceedings of the Second International Symposium of Information Theory. Budapest: Akademiai Kiado. Anderson, B.D.O. (1978). ‘Second-Order Convergent Algorithms for the Steady-State Riccati Equation’. International Journal Control, 28, 295–306. Anderson, Brian D.O. and John B. Moore (1979). Optimal Filtering. Englewood Cliffs: Prentice Hall. Anderson, Evan W., L.P. Hansen, Ellen R. McGrattan, and T.J. Sargent (1995). ‘Mechanics of Forming and Estimating Dynamic Linear Economies’. In John Rust (ed.), Handbook of Computational Economics, forthcoming. Anderson, Gary S. and George R. Moore (1985). ‘A Linear Algebraic Procedure for Solving Linear, Perfect Foresight Models’. Economic Letters, vol. 17, No. 3, 247–252. Ansley, C.F., and R. Kohn (1983). ‘Exact Likelihood of Vector Autoregressive-Moving Average Process with Missing or Aggregated Data’. Biometrika, 70, 275–278. Ansley, C.F., and R. Kohn (1985). ‘Estimation, Filtering, and Smoothing in State Space Models with Incompletely Specified Initial Conditions’. The Annals of Statistics, Vol. 13, No. 4, 1286–1316. Bell, William (1984). ‘Signal Extraction for Nonstationary Time Series’. The Annals of Statistics, Vol. 12, No. 2, pp. 646–664. Barro, R. J. (1977). ‘Unanticipated Money Growth and Unemployment in the United States’. American Economic Review, 67. Barro, R.J. (1979). ‘On the Determination of the Public Debt’. Journal of Political Economy, 87, 940–71. Becker, Gary and Kevin Murphy (1988). ‘A Model of Rational Addiction’. Journal of Political Economy, 96:4, 675–700. Beltrami, E.J. and M.R. Wohlers (1966). Distributions and the Boundary Values of Analytic Functions. Academic Press, New York. Bernanke, B. (1985). ‘Adjustment Costs, Durables, and Aggregate Consumption’. Journal of Monetary Economics, 15, 41–68. Bernanke, B. (1986). ‘Alternative Explorations of the Money-Income Correlation’. In Karl Brunner and Allan H. Meltzer (eds.), Real Business cycles, Real Exchange Rates, and Actual Policies. Carnegie-Rochester Conference Series on Public Policy. 25, 49–99, Amsterdam: North-Holland Bergstrom, A.R. (1976). Statistical Inference in Continuous Time Economic Models. Amsterdam: North-Holland. Bergstrom, A.R. (1983). ‘Gaussian Estimation of Structural Parameters in Higher Order Continuous Time Dynamic Models’. Econometrica, 51, 117–152. Birchenhall, C.R., R.C. Bladen-Hovell, A.P.L. Chui, D.R. Osborne, and J.P. Smith (1989). ‘A Seasonal Model of Consumption’. Economic Journal, Vol. 99, No. 397, September, pp. 837–843. Blanchard, Olivier J., and C.M. Kahn (1980). ‘The Solution of Linear Difference Models Under Rational Expectations’. Econometrica, 48: pp. 1305–11.

– 499 –

500

References

Blinder, A.S. and A. Deaton (1985). ‘The Time Series Consumption Function Revisited’. Brookings Paper on Economic Activity, 2, 465–511. Braun, R. Anton (1991). ‘The Dynamic Interaction of Distortionary Taxes and Aggregate Variables in Postwar U.S. Data’. Mimeo. University of Virginia. Breeden, D.T. (1979). ‘An Intertemporal Asset Pricing Model with Stochastic Consumption and Investment Opportunities’. Journal of Financial Economics, 7, 265–296. Breiman, Leo (1968). Probability. Addisson-Wesley Publishing Company, Reading, Massachusetts. Brock, William A. (1982). ‘Asset Prices in a Production Economy’. In The Economics of Information and Uncertainty (ed.), John J. McCall Chicago: University of Chicago Press, pp. 1–43. Brock, William A., and Leonard J. Mirman (1972). ‘Optimal Economic Growth and Uncertainty: the Discounted Case’. Journal of Economic Theory, 4: 479–513. Brock, W.A. and M.J.P. MaGill (1979). ‘Dynamics Under Uncertainty’. Econometrica, 47, 843–868. Burdick, Clark, Lars Hansen and T.J. Sargent (1990). ‘Statistics for Detecting Hidden Periodicity as a Source of Seasonality’. Mimeo. forthcoming. Burridge, Peter and Kenneth F. Wallis (1990). ‘Seasonal Adjustment and Kalman Filtering: Extension to Periodic Variances’. Journal of Forecasting, vol. 9, pp. 109-118. Campbell, J.Y. (1986). ‘Bond and Stock Returns in a Simple Exchange Model’. Quarterly Journal of Economics, 101, 785–803. Campbell, J.Y. (1987). ‘Does Saving Anticipate Declining Labor Income? An Alternative Test of the Permanent Income Hypothesis’. Econometrica, 55, 1249–1273. Campbell, J.Y. and R.J. Shiller (1987). ‘Cointegration and Tests of Present Value Models’. Journal of Political Economy, 95, 1062–1088. Campbell, J.Y. and R.J. Shiller (1988). ‘Dividend-Price Ratios and Expectations of Future Dividends and Discount Factors’. Review of Financial Studies, 1, 195–228. Cassing, Shirley and Tryphon Kollintzas (1991). ‘Recursive Factor of Production Interrelations and Endogenous Cycling’. International Economic Review, Vol. 32, No. 2, pp. 417–440. Christiano, L.J. (1980). ‘Notes on Factoring Continuous Time Rational Spectral Densities’. Mimeo. Federal Reserve Bank of Minneapolis. Christiano, L.J. (1984). ‘The Effects of Aggregation over Time on Tests of the Representative Agent Model of Consumption’. Mimeo. University of Chicago Working Papers in Economics and Econometrics, 84–15. December. Christiano, L.J. and M.S. Eichenbaum (1985). ‘A Continuous Time, General Equilibrium, Inventory-Sales Model’. Mimeo. COMPLETE REFERENCE. Christiano, L. J. and M.S. Eichenbaum (1986). ‘Temporal Aggregation and Structural Inference in Macroeconomics’. Carnegie-Rochester Conference on Public Policy, Vol. 26. Christiano, L.J., M.S. Eichenbaum, and D. Marshall (1991). ‘The Permanent Income Hypothesis Revisited’. Econometrica, Vol. 59, No. 2, pp. 397–423. Churchill, R.V., J.W. Brown and R.F. Verhey (1974). Complex Variables and Applications. 3rd edition, New York: McGraw Hill. ´ Cleveland, W.P., and G.C. Tiao (1979). ‘Modeling Seasonal Time Series’. Economie Appliqu´ ee, Tome-XXXII, No. 1, pp. 107–129.

References

501

Cochrane, J.H. (1988). ‘How Big is the Random Walk in GNP?’. Journal of Political Economy, 96, 893–920. Coddington, E. A. and N. Levinson (1955). Theory of Ordinary Differential Equations. New York: McGraw-Hill. Coleman, Wilbur J. II. (1990). ‘Solving the Stochastic Growth Model by Policy Function Iteration’. Journal of Business and Economic Statistics, January, Vol. 8, No. 1, pp. 27–29. Constantinides, G.M. (1990). ‘Habit Formation: A Resolution of the Equity Premium Puzzle’. Journal of Political Economy, 98, 519–543. Cox, John C., Jonathan E. Ingersoll, and Stephen A. Ross (1985). ‘An Intertemporal General Equilibrium Model of Asset Prices’. Econometrica, 53: 363–84. Cox, J.D., J.E. Ingersol, Jr., and S.A. Ross (1985). ‘A Theory of the Term Structure of Interest Rates’. Econometrica, 53, 129–151. Deistler, M., W. Dunsmuir and E. J. Hannan (1976). ‘Vector Linear Time Series Models: Corrections and Extensions’. Advanced in Applied Probability, 8, 360–372. De Jong, Piet (1991). ‘The Diffuse Kalman Filter’. The Annals of Statistics, Vol. 19, No. 2, pp. 1073–1083. De Jong, Piet (1991). ‘Stable Algorithms for the State Space Model’. Journal of Time Series Analysis, Vol. 12, No. 2, pp. 143–157. DeJong, D.N., J.C. Nankervis, N.E. Savin, and C.H. Whiteman (1988). ‘Integration versus Trend-Stationarity in Macroeconomic Time Series’. Mimeo. University of Iowa, Department of Economics. Working Paper #88-27, December. DeJong, D.N. and C.H. Whiteman (1989a). ‘The Temporal Stability of Dividends and Stock Prices: Evidence from the Likelihood Function’. Mimeo. University of Iowa, Department of Economics. Working Paper #89-03, February. DeJong, D.N. and C.H. Whiteman (1989b). ‘Trends and Random Walks in Macroeconomic Time Series: A Reconsideration Based on the Likelihood Principle’. Mimeo. University of Iowa, Department of Economics. Working Paper #89-04, February. De Long, J. Bradford and Lawrence H. Summers (1991). ‘Equipment Investment and Economic Growth’. Quarterly Journal of Economics, pp. 445–501, May. Denman, E.D. and A.N. Beavers (1976). ‘The Matrix Sign Function and Computations in Systems’. Applied Mathematics and Computation, 2, 63–94. Doan, Thomas, Robert Litterman, and Christopher Sims (1984). ‘Forecasting and Conditional Projection using Realistic Prior Distributions’. Econometric Reviews, Vol. 3, pp. 131–44. Dunsmuir, W. and E. J. Hannan (1978). ‘Vector Linear Time Series Models’. Advances in Applied Probability, 8, 339–364. Dunsmuir, W. (1978). ‘A Central Limit Theorem for Parameter Estimation in Stationary Linear Time Series and its Application to Models for Signal Observed with Noise’. Annals of Statistics, 7, 490-506. Eckstein, Z. (1984). ‘Rational Expectations Modeling of Agricultural Supply’. Journal of Political Economy, 93, 1–19. Eichenbaum, M.S. (1983). ‘A Rational Expectations Equilibrium Model of Finished Goods and Employment’. Journal of Monetary Economics, 12, 259–278.

502

References

Eichenbaum, M.S. and L.P. Hansen (1990). ‘Estimating Models with Intertemporal Substitution Using Aggregate Time Series Data’. Journal of Business and Economic Statistics, 8, 53–69. Emre, E. and O. Huseyin (1975). ‘Generalization of Leverrier’s Algorithm to Polynomial Matrices of Arbitrary Degree’. IEEE Transactions on Automatic Control, February. Engle, R.F., and C.W.J. Granger (1987). ‘Cointegration and Error-Correction: Representation, Estimation and Testing’. Econometrica, 55, 251–276. Engle, R. and M. Watson (1981). ‘A One Factor Multivariate Time Series Model of Metropolitan Wage Rates’. Journal of the American Statistical Association, 76, 774–781. Flavin, M.A. (1981). ‘The Adjustment of Consumption to Changing Expectations About Future Income’. Journal of Political Economy, 89, 974–1009. Futia, C.A. (1981). ‘Rational Expectations in Stationary Linear Models’. Econometrica, 49, 171–192. Gantmacher, F.R. (1959). The Theory of Matrices. Vol. 1, New York: Chelsea. Geweke, J.B. (1977a). ‘Wage and Price Dynamics in U.S. Manufacturing’. In C.A. Sims (ed.), New Methods in Business Cycle Research: Proceedings From a Conference 111–56, Minneapolis: Federal Reserve Bank of Minneapolis. Geweke, J.B. (1977b). ‘The Dynamic Factor Analysis of Economic Time Series’. In D.J. Aigner and A.S. Goldberger (eds.), Latent Variables in Socio-Economic Models. Ch. 19, 365–383, Amsterdam: North-Holland. Geweke, J.B. (1978). ‘Temporal Aggregation in the Multivariate Regression Model’. Econometrica, 46, 643–662. Geweke, J.B. and K.J. Singleton (1981a). ‘Maximum Likelihood ‘Confirmations’ Factor Analysis of Economic Time Series’. International Economic Review, 22, 37–54. Geweke, J.B. and K.J. Singleton (1981b). ‘Latent Variable Models for Time Series: A Frequency Domain Approach with an Application to the Permanent Income Hypothesis’. Journal of Economics, 17, 287–304. Gladysev, E.G. (1961). ‘Periodically Correlated Random Sequences’. Sov. Math, 2, 385–88. Gohberg, I., P. Lancaster and L. Rodman (1982). Matrix Polynomials. Academic Press. Gould, J.P. (1968). ‘Adjustment Costs in the Theory of Investment of the Firm’. Review of Economic Studies, 35, 47–56. Granger, C.W.J. (1969). ‘Investigating Causal Relations by Econometric Models and CrossSpectral Methods’. Econometrica, 37, 424–438. Grether, David and M. Nerlove (1970). ‘Some Properties of ‘Optimal’ Seasonal Adjustment’. Econometrica, vol. 38, pp. 682–703. Griliches, Z. (1967). ‘Distributed Lags: A Survey’. Econometrica, 35, 16–40. Grossman, S., A. Melino, and R.J. Shiller (1987). ‘Estimating the Continuous Time ConsumptionBased Asset Pricing Model’. Journal of Business and Economic Statistics, 6, 315–327. Hakkio, C.S. and M. Rush (1986). ‘Co-integration and the Government’s Budget Deficit’. Mimeo. Working Paper, Federal Reserve Bank of Kansas City. Hall, R.E. (1978). ‘Stochastic Implications of the Life Cycle-Permanent Income Hypothesis: Theory and Evidence’. Journal of Political Economy, 86, 971–987. Hall, R.E. (1988). ‘Intertemporal Substitution and Consumption’. Journal of Political Economy, 96, 339–357. Halmos, P.R. (1950). Measure Theory. New York, Van Nostrand.

References

503

Halmos, P.R. (1957). Introduction to Hilbert Space. New York, Chelsea Publishing Company. Hamilton, J.D. and M.A. Flavin (1986). ‘On the Limitations of Government Borrowing: a Framework for Empirical Testing’. American Economic Review, 76, 808–819. Hannan, E. J. (1970). Multiple Time Series. Wiley: New York. Hannan, E. J. (1973). ‘The Asympotic Theory of Linear Time Series Models’. Journal of Applied Probability, 10, 130–145. Hansen, Gary and Thomas J. Sargent (1988). ‘Straight time and Over time in Equilibrium’. Journal of Monetary Economics, 21: 281–308. Hansen, Lars P. (1980). ‘Large Sample Properties of Generalized Method of Moments Estimators’. Econometrica, 50 (July): pp. 1269–1286. Hansen, Lars P. (1987). ‘Calculating Asset Prices in Three Example Economics’. In T.F. Bewley (ed.), Advances in Econometrics Vol. 4, Cambridge University Press, 207–243. Hansen, L.P., D. Epple and W. Roberds (1985). ‘Linear Quadratic Models of Resource Depletion’. In T.J. Sargent (ed.), Energy, Foresight, and Strategy Resources for the Future, Washington, D.C., 101–142. Hansen, Lars Peter, John Heaton, and Thomas J. Sargent (1991). ‘Faster Methods for Solving Continuous Time Recursive Linear Models of Dynamic Economies’. In Lars P. Hansen and Thomas Sargent (eds.), Rational Expectations Econometrics. Westview Press. Hansen, L.P. and R.J. Hodrick (1980). ‘Forward Exchange Rates as Optimal Predictors of Future Spot Rates: An Econometric Analysis’. Journal of Political Economy, 88, 829–853. Hansen, L.P. and R.J. Hodrick (1983). ‘Exchange Rates and International Macroeconomics’. In J. Frenkel (ed.), NBER 113–152. Hansen, Lars Peter, William Roberds, and Thomas J. Sargent (1991). ‘Time Series Implications of Present Value Budget Balance and of Martingale Models of Consumption and Taxes’. In Lars P. Hansen and Thomas Sargent (eds.), Rational Expectations Econometrics. Westview Press. Hansen, L.P. and T.J. Sargent (1980a). ‘Formulating and Estimating Dynamic Linear Rational Expectations Models’. Journal of Economic Dynamics and Control, 2, 7–46. Hansen, L.P. and T.J. Sargent (1980b). ‘Methods for Estimating Continuous Time Rational Expectations Models from Discrete Time Data’. Staff Report 59, Federal Reserve Bank of Minneapolis. Hansen, L.P. and T.J. Sargent (1981a). ‘Linear Rational Expectations Models for Dynamically Interrelated Variables’. In R.E. Lucas, Jr., and T.J. Sargent (eds.), Rational Expectations and Econometric Practice. Minneapolis: University of Minnesota Press. Hansen, L.P. and T.J. Sargent (1981b). ‘A Note of Wiener-Kolmogorov Prediction Formulas for Rational Expectations Models’. Economic Letters, 8, 255–260. Hansen, L.P. and T.J. Sargent (1981c). ‘Identification of Continuous Time Rational Expectations Models from Discrete Time Data’. Mimeo. COMPLETE REFERENCE. Hansen, L.P. and T.J. Sargent (1981d). ‘Formulating and Estimating Continuous Time Rational Expectations Models’. Staff Report 75, Federal Reserve Bank of Minneapolis, October.

504

References

Hansen, L.P. and T.J. Sargent (1981e). ‘Exact Linear Rational Expectations Models: Specification and Estimation’. Staff Report 71, Research Department, Federal Reserve Bank of Minneapolis. Hansen, L.P. and T.J. Sargent (1982). ‘Instrumental Variables Procedures for Estimating Linear Rational Expectations Models’. Journal of Monetary Economics, 9, 263– 296. Hansen, L.P. and T.J. Sargent (1983a). ‘Aggregation over Time and the Inverse Optimal Predictor Problem for Adaptive Expectations in Continuous Time’. International Economic Review, 24, 1–20. Hansen, L.P. and T.J. Sargent (1983b). ‘The Dimensionality of the Aliasing Problem in Models with Rational Spectral Densities’. Econometrica, 51, 377–387. Hansen, L. P. and T.J. Sargent (1984). ‘Two Difficulties in Interpreting Vector Autoregressions’. Mimeo. Working Paper 227, Federal Reserve Bank of Minneapolis. Hansen, L.P. and T.J. Sargent (1990). ‘Recursive Linear Models of Dynamic Economies’. Mimeo. COMPLETE REFERENCE. Hansen, Lars Peter and and Thomas J. Sargent (1991). ‘Lecture Notes on Least Squares Prediction Theory’. In Lars P. Hansen and Thomas Sargent (eds.), Rational Expectations Econometrics. Westview Press. Hansen, Lars Peter, and Thomas J. Sargent (1991). ‘Identification of Continuous Time Rational Expectations Models from Discrete Time Data’. In Lars P. Hansen and Thomas Sargent (eds.), Rational Expectations Econometrics. Westview Press. Hansen, Lars P., and Thomas J. Sargent (1991). ‘Two Difficulties in Interpreting Economic Time Series’. In Lars P. Hansen and T.J. Sargent (eds.), Topics in Rational Expectations Econometrics. Westview Press. Hansen, Lars Peter and Thomas J. Sargent (1990b). ‘Disguised Periodicity as a Source of Seasonality’. Mimeo. June COMPLETE REFERENCE. Hansen, Lars Peter and Thomas J. Sargent (1993). Recursive Linear Models of Dynamic Economies. University of Chicago and Hoover Institution. Hansen, L.P. and K.J. Singleton (1982). ‘Generalized Instrumental Variables Estimators of Nonlinear Rational Expectations Models’. Econometrica, 50, 1269–1286. Hansen, L.P. and K.J. Singleton (1983). ‘Stochastic Consumption, Risk Aversion, and the Temporal Behavior of Asset Returns’. Journal of Political Economy, 91, 249–265. Harris, Milton (1987). Dynamic Economic Analysis. New York: Oxford University Press. Harrison, Michael and D. Kreps (1979). ‘Martingales and Arbitrage in Multiperiod Securities Markets’. Journal of Economic Theory, 20: 381–408. Harvey, A.C. (1981). Time Series Models. New York: Halsted. Harvey, A.C., and G.D.A. Phillips (1979). ‘Maximum Likelihood Estimation of Regression Models with Autoregressive-Moving Average Disturbances’. Biometrika, 66, 1, pp. 49–58. Harvey, A.C., and R.G. Pierse (1984). ‘Estimating Missing Observations in Economic Time Series’. Journal of the American Statistical Association, Vol. 79, No. 385, pp. 125–131. Harvey, A.C., and James H. Stock (1985). ‘The Estimation of Higher Order Continuous Time Autoregressive Models’. Econometric Theory, 1: 97–112.

References

505

Harvey, A.C., and James H. Stock (1987). ‘Estimating Integrated Higher Order Continuous Time Autoregressions with an Application to Money-income Causality’. Mimeo. Working Paper No. E 87-28. Stanford, CA: Hoover Institution Press. Harvey, C. (1988). ‘The Real Term Structure of Consumption Growth’. Journal of Financial Economics, 22, 305–333. Hatanaka, M. (1975). ‘On the Global Identification of the Dynamic Simultaneous Equations Model with Stationary Disturbances’. International Economic Review, 16, 545– 554. Hayashi, F. and C.A. Sims (1983). ‘Nearly Efficient Estimation of Time Series Models with Predetermined, but not Exogenous Instruments’. Econometrica, 51, 3, 783–798. Heaton, John (1988). ‘The Interaction Between Time-Nonseparable Preferences and Time Aggregation’. Mimeo. University of Chicago. Heaton, J. (1989). ‘The Interaction Between Time-Nonseparable Preferences and Time Aggregation’. Ph.D. dissertation, University of Chicago. Heaton, John C. (1990). ‘The Interaction Between Time Nonseparable Preferences and Time Aggregation’. Mimeo. Sloan Working Paper 3181-90-EFA, Massachusetts Institute of Technology. Hildenbrand, W. (1974). Core and Equilibria of a Large Economy. Princeton: Princeton University Press. Houthakker, H. and L. Taylor (1970). Consumer Demand in the United States: Analyses and Projections. Second Edition. Cambridge, MA: Harvard University Press. Jacobson, David H. (1973). ‘Optimal Stochastic Linear Systems with Exponential Performance Criteria and Their Relation to Deterministic Differential Games’. IEEE Transactions on Automatic Control, Vol. AC-18, No. 2, April, 124–131. Jacobson, David H. (1977). Extensions of Linear-Quadratic Control, Optimization and Matrix Theory. Academic Press, New York. Jacobson, D. H. (1973). ‘Optimal Stochastic Linear Systems with Exponential Performance Criteria and their Relation to Deterministic Differential Games’. IEEE Transactions on Automatic Control, AC-18, 124–31. Jacobson, D. H. (1973). Extensions of Linear-quadratic Control, Optimization and Matrix Theory. Academic Press, New York. Jones, Larry and Rodolfo Manuelli (1988). ‘A Model of Optimal Equilibrium Growth’. Mimeo. Stanford University. Jones, R.H. (1980). ‘Maximum Likelihood Fitting of ARMA Models to Time Series with Missing Observations’. Technometrics, 22, 389–395. Jones, R.H. and W.M. Brelsford (1967). ‘Time Series With Periodic Structure’. Biometrika, vol. 54, pp. 403–408. Kalman, R.E. (1960). ‘A New Approach to Linear Filtering and Prediction Problems’. Trans. ASME Ser. D. J. Basic Engrg., 82: 35–45. Kehoe, Timothy and David Levine (1984). ‘Empirical Implications of Complete Contingent Claims Markets’. Mimeo. University of California at Los Angeles. Kennan, J. (1979). ‘The Estimation of Partial Adjustment Models with Rational Expectations’. Econometrica, 47, 1441–1456. Kennan, John (1988). ‘An Econometric Analysis of Fluctuations in Aggregate Labor Supply and Demand’. Econometrica, 56, 2, 317–333.

506

References

King, Robert G., Charles I. Plosser, and Sergio T. Rebelo (1988). ‘Production, Growth, and Business Cycles: I. The Basic Neoclassical Model’. Journal of Money, Credit, and Banking, vol. 21, No. 2/3, March, May, 195–232. King, Robert G., Charles I. Plosser, and Sergio T. Rebelo (1988). ‘Production, Growth, and Business Cycles: II. New Directions’. Journal of Money, Credit, and Banking, vol. 21, No. 2/3, March, May, 309–342. Kohn, Robert and C.F. Ansley (1985). ‘Efficient Estimation and Prediction in Time Series Regression Models’. Biometrika, 72, pp. 694–7. Kohn, Robert and C.F. Ansley (1986). ‘Estimation, Prediction, and Interpolation for ARIMA Models with Missing Data’. Journal of the American Statistical Association, Vol. 81, No. 395, pp. 751–61. Kollintzas, T. (1985). ‘The Symmetric Linear Rational Expectations Model’. Econometrica, 53, 963–985. Kullback, S. and R.A. Leibler (1951). ‘On Information and Sufficiency’. Annals of Mathematical Statistics, 22, 79–86. Kwakernaak, H. and R. Sivan (1972). Linear Optimal Control Systems. New York, Wiley Interscience. Kydland, Finn and Edward C. Prescot (1982). ‘Time to Build and Aggregate Fluctuations’. Econometrica, 50: pp. 1345–70. LeRoy, S.F. (1973). ‘Risk Aversion and the Martingale Property of Stock Prices’. International Economic Review, 14, 436–446. LeRoy, S.F. and R. Porter (1981). ‘The Present Value Relation: Tests Based on Implied Variance Bounds’. Econometrica, 49, 97–113. Litterman, R. and L. Weiss (1985). ‘Money, Real Interest Rates and Output: A Reinterpretation of Post-war Data’. Econometrica, 53, 129–156. Ljung, Lennart (1987). System Identification: Theory for the User. Prentice Hall, Englewood Cliffs. Lo, A. (1988). ‘Maximum Likelihood Estimation of Generalized Ito Processes with Discretely Sampled Data’. Journal of Econometric Theory, 4, 231–247. Lucas, R.E., Jr. (1967). ‘Adjustment Costs and the Theory of Supply’. Journal of Political Economy, 75, 321–334. Lucas, R.E., Jr. (1972). ‘Econometric Testing of the Natural Rate Hypothesis’. In O. Eckstein (ed.), The Econometrics of Price Determination Conference. Washington: Board of Governors of the Federal Reserve System, pp. 50–59. Lucas, R.E., Jr. (1973). ‘Some International Evidence on Output-Inflation Tradeoffs’. American Economic Review, 63 (June): 326–34. Lucas, R.E., Jr. (1976). ‘Econometric Policy Evaluation: A Critique’. In K. Brunner and A.H. Meltzer (eds.), The Phillips Curve and Labor Markets. Carnegie-Rochester Conference on Public Policy 1: 19–46, Amsterdam: North Holland. Lucas, R.E., Jr. (1978). ‘Asset prices in an Exchange Economy’. Econometrica, 46, 1429– 1445. Lucas, R.E., Jr. (1981). ‘Optimal Investments with Rational Expectations’. Rational Expectations and Econometric Practice, Vol. 1, 127–156. Lucas, Robert E., Jr. (1987). ‘Money and Interest in a Cash-in-Advance Economy’. Econometrica, 55: 491–513.

References

507

Lucas, R.E., Jr. and E.C. Prescott (1971). ‘Investment Under Uncertainty’. Econometrica, 39, 659–681. Lucas, Robert E., Jr., and Edward C. Prescott (1974). ‘Equilibrium Search and Unemployment’. Journal of Economic Theory, 7: 188–209. Lucas, R.E., Jr. and T.J. Sargent (1981). Introductory Essay to Rational Expectations and Econometric Practice. Minneapolis: University of Minnesota Press. Luenberger, David G. (1966). ‘Observers for Multivariable System’. IEEE Transactions on Automatic Control, 11: 190–97. Luenberger, David G. (1967). ‘Cononical Forms for Multivariable System’. IEEE Transactions on Automatic Control, 12: 290–93. Luenberger, D.G. (1969). Optimization by Vector Space Methods. John Wiley & Sons. Mace, Barbara (1989). ‘Full Insurance in the Presence of Aggregate Uncertainty’. Mimeo. Rochester Center for Economic Research, Working Paper #197. Mankiw, N.G. (1982). ‘Hall’s Consumption Hypothesis and Durable Goods’. Journal of Monetary Economics, 10, 417–425. Marcet, A. (1987). ‘Approximation of the Continuous Wold Decomposition by Frequent Sampling’. Mimeo. Carnegie Mellon University. Marcet, Albert (1988). ‘Solving Nonlinear Models by Parameterizing Expectations’. Mimeo. Carnegie-Mellon University, G.S.I.A. Marcet, Albert (1991). ‘Temporal Aggregation of Economic Time Series’. In Lars P. Hansen and Thomas Sargent (eds.), Rational Expectations Econometrics. Westview Press. McGrattan, Ellen R. (1991a). ‘A Note on Computing Equilibria in Linear Models’. Mimeo. Duke University, June. McGrattan, Ellen R. (1991b). ‘Fiscal Policy and Technical Change in a Model of Business Cycles’. Mimeo. Duke University, December. Miller, P.J. and W. Roberds (1987). ‘The Quantitative Significance of the Lucas Critique’. Staff Report, Federal Reserve Bank of Minneapolis, #109. Monahan, J. (1984). ‘A Note on Enforcing Stationarity in Autoregressive-Moving Average Models’. Biometrika, 71, 403–404. Monin, A.S. (1963). ‘Stationary and Periodic Time Series in the General Circulation of the Atmosphere’. In M. Rosenblatt (ed.), Proceedings of the Symposium on Time Series Analysis. New York: Wiley. pp. 144–151. Mortenson, D.T. (1973). ‘Generalized Costs of Adjustment and Dynamic Factor Demand Theory’. Econometrica, 41, 657–665. Muth, John F. (1960). ‘Optimal Properties of Exponentially Weighted Forecasts’. Journal of the American Statistical Association, 55, 299–306. Muth, John F. (1961). ‘Rational Expectations and the Theory of Price Movements’. Econometrica, 29: 315–35. Negishi, T. (1960). ‘Welfare Economics and Existence of an Equilibrium for a Competitive Economy’. Metroeconomica, vol. 12: 92–97. Nerlove, M. (1967). ‘Distributed Lags and Unobserved Components in Economic Time Series’. In W. Fellner, et. al. (eds.), Ten Economic Studies in the Tradition of Irving Fisher. New York: Wiley. Nobel, Ben and James W. Daniel (1977). Applied Linear Algebra. 2nd edition, Prentice Hall: Englewood Cliffs, New Jersey.

508

References

Osborn, Denise R. (1991). ‘The Implication of Periodically Varying Coefficients for Time-series Processes’. Journal of Economics, Vol. 43, No. 3, June, pp. 373–384. Osborne, Denise (1990). ‘The Implications of Periodically Varying Coefficients for Seasonal Time Series Processes’. Journal of Econometrics, forthcoming. Osborn, Denise R. (1988). ‘Seasonality and Habit Persistence in a Life Cycle Model of Consumption’. Journal of Applied Econometrics, 3, October-December, pp. 255–266. Osborne, Denise R., and Jeremy P. Smith (1989). ‘The Performance of Periodic Autoregressive Models in Forecasting Seasonal U.K. Consumption’. Journal of Economic and Business Statistics, vol. 7, no. 1, January, pp. 117–128. Osborne, Denise, A.P.L. Chui, J.P. Smith, and C.R. Birchenhall (1988). ‘Seasonality and the Order of Integration for Consumption’. Oxford Bulletin of Economics and Statistics, 50, 4, pp. 361–377. Pagano, M. (1978). ‘On Periodic and Multiple Autoregressions’. The Annals of Statistics, 6, pp. 1310–1317. Park, Jong Ahn (1984). ‘Gestation Lags with Variable Plans: An Empirical Study of Aggregate Investment’. Ph.D. dissertation, Carnegie-Mellon University. Phillips, A.W. (1959). ‘The Estimation of Parameters in Systems of Stochastic Differential Equations’. Biometrika, 59, 67–76. Phillips, P.C.B. (1972). ‘The Structural Estimation of a Stochastic Differential Equation System’. Econometrica, 40, 1021–1041. Phillips, P.C.B. (1973). ‘The Problem of Identification in Finite Parameter Continuous Time Models’. Journal of Econometrics, 1, 351–362. Phillips, P.C.B. (1974). ‘The Estimation of Some Continuous Time Models’. Econometrica, 42, 803–823. Pollak, R.A. (1970). ‘Habit Formation and Dynamic Demand Functions’. Journal of Political Economy, 78, 745–763. Prescott, Edward C. and Robert E. Lucas, Jr. (1972). ‘Price Systems in Infinite Dimensional Space’. International Economic Review, 13: pp. 416–22. Prescott, Edward C., and Rajnish Mehra (1980). ‘Recursive Competitive Equilibrium: the Case of Homogeneous Households’. Econometrica, 48: pp. 1364–79. Roberds, William (1991). ‘Implications of Expected Present Value Budget Balance: Application to Postwar U.S. Data’. In Lars P. Hansen and Thomas Sargent (eds.), Rational Expectations Econometrics. Westview Press. Roberts, J.D. (1971). ‘Linear Model Reduction and Solution of Algebraic Riccati Equations by use of the Sign Function’. CUED/B-Control/TR-13 Report, Cambridge University. Robinson, P.M. (1977). ‘The Construction and Estimation of Continuous Time Models and Discrete Approximations in Econometrics’. Journal of Econometrics, 6, 173–197. Romer, Paul M. (1986). ‘Increasing Returns and Long-Run Growth’. Journal of Political Economy, vol. 94, pp. 1002–37. Rosen, S., K.M. Murphy, and J.A. Scheinkman (1994). ‘Cattle Cycles’. Journal of Political Economy, Vol. 102, No. 3, pp. 468–492. Royden, H.L. (1968). Real Analysis. New York, MacMillan Publishing Co., Inc. Rozanov, Y.A. (1967). Stationary Random Processes. Holden-Day: San Francisco. Rubenstein, Mark (1974). ‘An Aggregation Theorem for Security Markets’. Journal of Financial Economics, Vol. 1, No.3, pp. 225–244.

References

509

Rubinstein, M. (1976). ‘The Valuation of Uncertain Income Streams and the Pricing of Options’. Bell Journal of Economics, 7, 407–425. Rudin, W. (1974). Real and Complex Analysis. McGraw Hill: New York. Ryder, H.E. and G.M. Heal (1973). ‘Optimal Growth with Intertemporally Dependent Preferences’. Review of Economic Studies, 40, 1–31. Sargent, T.J. (1971). ‘A Note on the ‘Accelerationist’ Controversy’. Journal of Money, Credit, and Banking, 3, 721–725. Sargent, T.J. (1978). ‘Estimation of Dynamic Labor Demand Schedules Under Rational Expectations’. Journal of Political Economy, 86, 1009–1044. Sargent, Thomas J. (1978). Comments on. ‘Seasonal Adjustment and Multiple Time Series Analysis’. In A. Zellner (ed.), Seasonal Analysis of Economic Time Series, (by K.F. Wallis). Washington, D.C.: Bureau of the Census, pp. 361–363. Sargent, T.J. (1979). ‘A Note on Maximum Likelihood Estimation of the Rational Expectations of the Term Structure’. Journal of Monetary Economics, 5, 133–143. Sargent, T.J. (1981a). ‘Interpreting Economic Time Series’. Journal of Political Economy, 89, 213–248. Sargent, Thomas J. (1981b). ‘Lecture notes on Filtering, Control, and Rational Expectations’. Mimeo. University of Minnesota. Sargent, T.J. (1983). ‘Preliminary Introduction to Continuous Time Stochastic Processes’. Mimeo. COMPLETE REFERENCE. Sargent, T.J. (1987a). Dynamic Macroeconomic Theory. Cambridge: Harvard University Press. Sargent, Thomas J. (1987b). Macroeconomic Theory. Second Edition, Academic Press, New York, (pp. 336-342). Sargent, Thomas J. (1989). ‘Two Models of Measurements and the Investment Accelerator’. Journal of Political Economy, 97: 251–87. Sargent, T.J. and C.A. Sims (1977). ‘Business Cycle Modeling Without Pretending to Have too Much A Priori Economic Theory’. In C.A. Sims (ed.), New Methods for Business Cycle Research, Federal Reserve Bank of Minneapolis. Sargent, Thomas J. and Neil Wallace (1973). ‘Rational Expectations and the Dynamics of Hyperinflation’. International Economic Review, 63: 328–50. Scheinkman, Jose (1987). ‘Market Incompleteness and the Equilibrium Valuation of Assets’. In S. Bhattacharya and G. Constantinides (eds.), Theory of Valuation: Frontiers of Modern Financial Theory. pp. 45–51. Schwarz, Z.G. (1978). ‘Estimating the Dimension of a Model’. Annals of Statistics, 6, 461–464. Shiller, R.J. (1972). ‘Rational Expectations and the Structure of Interest Rates’. Ph.D. dissertation, (Massachusetts Institute of Technology, Cambridge, MA). Shiller, R.J. (1979). ‘The Volatility of Long-Term Interest Rates and Expectations Models of the Term Structure’. Journal of Political Economy, 87, 1190–1219. Shiller, R.J. (1981). ‘Do Stock Prices Move too much to be Justified by Subsequent Changes in Dividends?’. The American Economic Review, 71, 421–436. Shim, S.D. (1984). ‘Inflation and the Government Budget Constraint: International Evidence’. Ph.D. dissertation, University of Minnesota. Sims, C.A. (1971). ‘Discrete Approximations to Continuous Time Lag Distributions in Econometrics’. Econometrica, 39, 545–564.

510

References

Sims, C.A. (1972a). ‘Approximate Prior Restriction on Distributed Lag Estimation’. Journal of the American Statistical Association, 67, 169–175. Sims, C.A. (1972b). ‘Money, Income and Causality’. American Economic Review, 62, 540–552. Sims, C.A. (1974a). ‘Distributed Lags’. In M.D. Intrilligator and D.A. Kendrick (eds.), Frontiers of Quantitative Economics. Vol. II, Amsterdam: North-Holland, 289– 332. Sims, C.A. (1974b). ‘Seasonality in Regression’. Journal of the American Statistical Association, vol. 69, September pp. 618–626. Sims, Christopher (1976). ‘personal letter to Thomas Sargent’. U.S. mail, October. Sims, Christopher A. (1980). ‘Macroeconomics and Reality’. Econometrica, 48: 1–48. Sims, C.A. (1984). ‘Martingale-Like Behavior of Prices and Interest Rates’. Mimeo. Department of Economics, University of Minnesota, working paper, October. Sims, C.A. (1988). ‘Bayesian Skepticism on Unit Root Econometrics’. Journal of Economic Dynamics and Control, 12, 463–474. Sims, Christopher (1990). ‘Rational Expectations Modeling with Seasonally Adjusted Data’. Mimeo. December. Sims, C.A. and H. Uhlig (1991). ‘Understanding Unit Rooters: A Helicopter Tour’. Econometrica, Vol. 59, No. 6, pp. 1591–1599. Singleton, Kenneth J. (1988). ‘Econometric Issues in the Analysis of Equilibrium Business Cycle Models’. Journal of Monetary Economics, Vol. 21, No. 2/3, March/May; pp. 361–386. Siow, A. (1984). ‘Occupational Choice Under Uncertainty’. Econometrica, Vol. 52, No. 3, pp. 631–645. Stigler, G.J. and G.S. Becker (1977). ‘De Gustibus Non Est Disputandum’. American Economic Review, 67, 76–90. Stokey, Nancy L. and Robert E. Lucas, Jr., with Edward C. Prescott (1989). Recursive Methods in Economic Dynamics. Cambridge, MA: Harvard University Press. Sundaresan, M. (1989). ‘Intertemporally Dependent Preferences and the Volatility of Consumption and Wealth’. The Review of Financial Studies, 2, 73–89. Telser, L.G. and R.J. Graves (1972). Functional Analysis in Mathematical Economics. Chicago: University of Chicago Press. Tiao, G.C., and M.R. Grupe (1980). ‘Hidden Periodic Autoregressive-Moving Average Models in Time Series Data’. Biometrika, 67, 2, pp. 365–73. Todd, Richard (1983). ‘A Dynamic Equilibrium Model of Seasonal and Cyclical Fluctuations in the Corn-Soybean-Hog Sector’. Ph.D. dissertation, University of Minnesota. Todd, Richard (1990). ‘Periodic Linear Quadratic Periodic Models of Seasonality’. Journal of Economic Dynamics and Control, Vol. 14, No 3/4, July/Oct., pp. 763–796. Topel, Robert and S. Rosen (1988). ‘Housing Investment in the United States’. Journal of Political Economy, vol. 96, pp. 718–40. Townsend, R.M. (1983). ‘Forecasting the Forecasts for Others’. Journal of Political Economy, 91, 546–588. Townsend, Robert M. (1987). ‘Arrow-Debreu Programs as Microfoundations for Macroeconomics’. In Truman F. Bewley (ed.), Advances in Econometrica: Fifth World Congress, Cambridge, UK: Cambridge University Press, pp. 379–428. Treadway, A.B. (1969). ‘On Rational Entrepreneur Behavior and the Demand for Investment’. Review of Economic Studies, 367, 227–240.

References

511

Trehan, B. and C.E. Walsh (1988). ‘Common Trends, the Government Budget Constraints, and Revenue Smoothing’. Journal of Economic Dynamics and Control, 12, 425– 446. Vaughan, D.R. (1969). ‘A Negative-Exponential Solution to the Matrix Riccati Equation’. IEEE Transactions on Automatic Control, 14, 72–75. Vaughan, D.R. (1970). ‘ A Nonrecursive Algebraic Solution for the Discrete Riccati Equation’. IEEE Transactions on Automatic Control, Vol. AC-15, No. 5, October, pp. 597– 599. Wallis, Kenneth (1974). ‘Seasonal Adjustment and Relations Between Variables’. Journal of the American Statistical Association, vol. 69, March pp. 18–32. Watson, M. and R. Engle (1983). ‘Alternative Algorithms for the Estimation of Dynamic Factor, MIMIC and Varying Coefficient Regression Models’. Journal of Econometrics, 23, 385–400. West, K.D. (1988). ‘The Insensitivity of Consumption to News About Income’. Journal of Monetary Economics, 21, 17–33. Wheeden, R.L. and A. Zygmund (1977). Measure and Integral; An Introduction to Real Analysis. Marcel Dekker Inc.: New York. White, H. (1982). ‘Maximum Likelihood Estimation of Misspecified Models’. Econometrica, 50, 1-26. Whiteman, Charles (1983). Linear Rational Expectations Models: A User’s Guide. Minneapolis: University of Minnesota Press. Whittle, Peter (1990). Risk Sensitive Optimal Control. Wiley, New York. Wilcox, D.W. (1989). ‘What Do We Know about Consumption?’. Mimeo. Federal Reserve Board of Governors, Washington, D.C. Wilson, Robert, (1968). ‘The Theory of Syndicates’. Econometrica, Vol. 36, No. 1, pp. 119–132. Working, Holbrook (1960). ‘Note on the Correlation of First Differences of Averages in a Random Chain’. Econometrica, 28: 916–18.

Index

periodic model, 351 cross-products between states and controls removal of, 185

adjacent complementarity, 219 aggregation over time, 172 allocation rules, 238 approximation theory, 171 arma autoregressive, moving average processes, 18 Arrow-Debreu competitive equilibrium, 2 asset pricing periodic version, 347 via arbitrage on Arrow-Debreu prices, 114 autocovariance function, 23 autoregression, 27 Bellman equation, 64 Bellman’s “curse of dimensionality”, 2 Bellman’s functional equation, 182 canonical representation of household services computation, 214 canonical representation of service technology, 209 cattle, 134 certainty equivalence, 281 choice household’s, 106 commodity space, 98 competitive equilibrium, 104 definition, 107 computed example a Jones–Manuelli like economy, 72 an economy with a durable consumption good, 72 Brock and Mirman’s model, 71 Brock–Mirman model with adjustment costs, 81 Lucas’s pure exchange economy, 72 periodic Hall model, 364 controllability canonical form, 182, 195 covariance stationarity, 17 covariogram

demand schedules, 209 detectability, 151, 186 directional derivatives, 55 discounted control problem replacement by undiscounted problem, 186 disguised periodicity, 370 two illustrations, 370 doubling algorithm, 69, 116, 182, 192 periodic regulator, 198 dynamic programming linear quadratic, 58 eigenvalues endogenous, 61 exogenous, 61 Engel curve, 235 exogenous state vector, 36 factorization identity, 147 proof, 227 firms leasing, 107 producing, 106 type I, 106 type II, 107 Fourier transforms, 225 frequency domain estimation, 168 Frisch demand functions, 49 fundamental theorems of welfare economics, 104 Gateaux derivative, 55 Gorman aggregation, 232 growth condition Jones–Manuelli, 81 heterogeneous consumers, 237 household technology, 45 adjustment costs, 48 consumer durables, 46 habit persistence, 47 multiple consumption goods, 49 housing market, 133

– 513 –

514

Index

impulse response, 9 inner-product representation, 99 innovation, 148 innovation accounting, 11 innovations functions of model noises, 162 innovations representation, 146 as whitener, 162 periodic version, 368 serially correlated measurement errors, 152 invariant subspace method, 182 invariant subspace methods, 188 Kalman filter, 146, 152 Lagrange multipliers, 52, 53 expressed in terms of gradient of value function, 61 Lagrangian firm of type I, 109 firm of type II, 110 Lagrangian for household’s problem, 108 linear functional, 99 Lyapunov equation, 151 discrete, 17 Martingale difference sequence, 8 measurement errors serially correlated, 152 moving average representation, 9 Negishi weights for social planning problem, 231 occupational choice, 138 optimal linear regulator problem, 183 periodic version, 197 optimal resource allocation problem, 52 Parseval’s formula, 224 partial equilibrium, 128 partitioning state vector, 194 periodic models, 341 periodic optimal linear regulator problem, 182 policy functions dynamic programming, 59

prediction theory linear difference equations, 9 periodic version, 349 preferences, 45 examples, 46 time separable, 46 price system, 111 problem firms of type I, 106 firms of type II, 107 rational expectations hyperinflation – stock price model, 22 reciprocal pairs property, 206 recursive dynamic competitive economies, 1 representative consumer, 231 representative household, 45, 238 Riccati equation, 2, 69, 184, 215, 227 Riccati equations convergence conditions, 151 Riesz Representation Theorem, 103 risk premium, 115 risk sharing, 241 limited markets, 241 risk-sensitive control theory, 201 Schur decomposition, 189 seasonality deterministic, 24 indeterministic, 26 periodic, 341 sharing rules, 238 skip-sampled system, 352 social planning problem, 52 as optimal linear regulator problem, 63 periodic version, 342 variational approach, 53 spectral density matrix, 23 square summability, 50 stability conditions, 151 stabilizability, 151, 186 stacked system, 352 stochastic difference equation linear first order, 7 summation by parts, 95 Sylvester equation, 151, 196

Index

symplectic, 193 symplectic matrices, 189, 206 technology, 36 a depletable resource, 42 examples, 38 interrelated factor demand, 44 learning by doing, 43 multi-period adjustment costs, 39 pure endowment, 38 single-period adjustment costs, 39 time to build, 39 technology: growth, 41 term structure of interest rates, 115, 117 term structure of interest rates, periodic version, 350 Tiao–Grupe formula, 360 Tiao-Grupe formula, 359 state space realization, 361 two fund theorem, 242 uncontrollable stochastic process, 53 unit root, 26 value function dynamic programming, 59 iterations on, 184 Vaughan’s method, 187 vector autoregression, 1, 14 whitener, 162 whitening filter, 150 Wold representation, 157

515

Author Index

Altug, Sumru, 275 Anderson, Brian D. O., 3, 168, 182, 193 Ansley, Craig, 298 Arrow, Kenneth, 3 Barro, Robert, 305 Becker, Gary, 77, 239, 249 Bellman, Richard, 4, 65 Blanchard, Olivier, 166 Breiman, Leo, 416 Brock, William, 46, 95 Browning, Martin, 54

Jones, Larry, 46, 53, 87 Judd, Kenneth, 4, 305 Kahn, Charles, 166 Kehoe, Patrick, 305 Kohn, Robert, 298 Koopmans, T.C., 46 Kreps, David, 109 Kwakernaak, Huibert, 3, 170 Kydland, Finn, 45 Lancaster, Kelvin, 50 Le Roy, Stephen, 24 Liu, blank, 252 Lucas, Robert E., Jr., 3, 64, 78, 124, 142

Eichenbaum, Martin, 142 Engle, Robert, 298

Mace, Barbara, 275 MaCurdy, Thomas, 54 Mankiw,Gregory, 77 Manuelli, Rodolfo, 46, 53, 87 Marcet, Albert, 4 Miller, Robert, 275 Mirman, Leonard, 46, 95 Moore, James, 168 Moore, John, 182, 193 Moore, John B., 3 Mortensen, Dale, 49 Murphey, Kevin, 146 Murphy, Kevin, 77, 239, 249 Muth, John F., 23

Flavin, Marjorie, 46 Frisch, Ragnar, 54

Negishi, T., 265 Nerlove, Marc, 252

Gabel, R. A., 252 Gladysev, E. G., 405 Gorman, Terrance, 7, 50, 243, 265 Granger, C. W. J., 298 Grether, David, 252 Grupe, M. R., 405

Osborn, Denise, 387 Owens, Raymond, 154

Hall, Robert E., 46, 95 Harris, Milton, 3 Harrison, Michael, 109 Heal, Geoffrey, 53, 249 Heckman, James, 54 Houthakker, Hendrik, 53

Richard, Scott, 109 Roberds, William, 206 Roberts, R. A., 252 Rosen, Sherwin, 144, 146, 149 Rubinstein, Mark, 277 Ryder, Harl, 53, 249

Irish,

Scheinkman, Jose, 146

Cagan, Phillip, 24 Carvalho, Jose, 252 Cass, David, 46 Chari, V. V. , 305 Christiano, Lawrence, 305 Cochrane, John , 275 Coleman, Wilbur, 4 Deaton, Angus, 54 Debreu, Gerard, 3, 106

, 54

Park, Jong Ahn, 45 Pollak, Robert, 249 Porter, Richard, 24 Prescott, Edward C., 3, 45, 64, 142

– 517 –

518

Shiller, Robert, 24 Sims, Christopher A., 3, 11, 213 Siow, Aloyisius, 149 Sivan, Raphael, 3, 170 Slutsky, Eugen, 11 Stigler, George, 249 Stokey, Nancy, 3, 64 Tauchen, George, 4 Taylor, Lester, 53 Tiao, George, 405 Todd, Richard, 387 Topel, Robert, 144 Townsend, Robert, 275 Vaughan, D. R., 157 Whiteman, Charles, 166

Author Index

Matlab Index

Matlab Index

dog.m, 321 seasla.m, 410 aarma.m, 74 aggreg.m, 217 aimpulse.m, 74 asimul.m, 74 asseta.m, 75, 132 assets.m, 411 assetss.m, 411 avg.m, 215, 217 bewley4.m, 282 bewley5.m, 282 canonpr.m, 244 clex11.m, 72, 217, 317 clex11h.m, 317 clex14.m, 130, 134 clex35b.m, 282 compare.m, 334 dimpulse.m, 12, 217 disthet.m, 334, 343 dlsim.m, 26 dlyap.m, 19 dog.m, 321 double2.m, 343

519

double2j.m, 75, 172, 176, 343 doublej.m, 20 doublej2.m, 20 doubleo.m, 75 doublex.m, 373 evardec.m, 13 factors.m, 411 heter.m, 278 kfilter.m, 325 seasla.m, 410, 411 series.m, 328 simpulse.m, 411, 412 simulh.m, 278 simulhet.m, 317 simuls.m, 411 solvdist.m, 334, 339 solvea.m, 74, 75, 95, 217, 278, 317, 334 solvehet.m, 317 solves.m, 411 solvex.m, 373 spectral.m, 25 spectrs.m, 411 steadst.m, 74, 411 varrep.m, 203 white1.m, 204 white2.m, 204

Recursive Models of Dynamic Linear Economies

Lars Hansen University of Chicago Thomas J. Sargent New York University and Hoover Institution

c

Lars Peter Hansen and Thomas J. Sargent

6 September 2005

Contents

Acknowledgements

xii

Preface

xiii

Part I: Components of an economy

1. Introduction

3

1.1. Introduction. 1.2. Computer Programs. 1.3. Organization.

2. Stochastic Linear Difference Equations

9

2.1. Introduction. 2.2. Notation and Basic Assumptions. 2.3. Prediction Theory. 2.4. Transforming Variables to Uncouple Dynamics. 2.5. Examples. 2.5.1. Deterministic seasonals. 2.5.2. Indeterministic seasonals. 2.5.3. Univariate autoregressive processes. 2.5.4. Vector autoregressions. 2.5.5. Polynomial time trends. 2.5.6. Martingales with drift. 2.5.7. Covariance stationary processes. 2.5.8. Multivariate ARMA processes. 2.5.9. Prediction of a univariate first order ARMA. 2.5.10. Growth. 2.5.11. A rational expectations model. 2.6. The Spectral Density Matrix. 2.7. Computer Examples. 2.7.1. Deterministic seasonal. 2.7.2. Indeterministic seasonal, unit root. 2.7.3. Indeterministic seasonal, no unit root. 2.7.4. First order autoregression. 2.7.5. Second order autoregression. 2.7.6. Growth with homoskedastic noise. 2.7.7. Growth with heteroskedastic noise. 2.7.8. Second order vector autoregression. 2.7.9. A rational expectations model. 2.8. Conclusion.

3. The Economic Environment 3.1. Information. 3.2. Taste and Technology Shocks. 3.3. Technologies. 3.4. Examples of Technologies. 3.4.1. Other technologies. 3.5. Preferences and Household Technologies. 3.6. Examples of Household Technology Preference Structures. 3.7. Constraints to Keep the Solutions “Square Summable”. 3.8. Summary.

–v–

39

vi

Contents

4. Optimal Resource Allocation

57

4.1. Planning problem. 4.2. Lagrange Multipliers. 4.3. Dynamic programming. 4.4. Lagrange multipliers as gradients of value function. 4.5. Planning problem as linear regulator. 4.6. Solutions for five economies. 4.6.1. Preferences. 4.6.2. Technology. 4.6.3. Information. 4.6.4. BrockMirman model. 4.6.5. A growth economy fueled by habit persistence. 4.6.6. Lucas’s pure exchange economy. 4.6.7. An economy with a durable consumption good. 4.7. Hall’s model. 4.8. Higher Adjustment Costs. 4.9. Altered ‘growth condition’. 4.10. A Jones-Manuelli economy. 4.11. Durable consumption goods. 4.12. Summary. A. Synthesizing the linear regulator. B. A Brock-Mirman model. 4.B.1. Uncertainty. 4.B.2. Optimal Stationary States.

5. The Commodity Space

105

5.1. Valuation. 5.2. Price systems as linear functionals. 5.3. A one period model under certainty. 5.4. One period under uncertainty. 5.5. An infinite number of periods and uncertainty. 5.5.1. Conditioning information. 5.6. Lagrange multipliers. 5.7. Summary. A. Appendix.

6. A Competitive Economy

113

6.1. Introduction. 6.2. The Problems of Households and Firms. 6.2.1. Households. 6.2.2. Firms of type I. 6.2.3. Firms of type II. 6.3. Competitive Equilibrium. 6.4. Lagrangians. 6.4.1. Households. 6.4.2. Firms of type I. 6.4.3. Firms of type II. 6.5. Equilibrium Price System. 6.6. Asset Pricing. 6.7. Term Structure of Interest Rates. 6.8. Re-opening Markets. 6.8.1. Recursive price system. 6.8.2. Non-Gaussian asset prices. 6.9. Summary of Pricing Formulas. 6.10. Asset Pricing Example. 6.10.1. Preferences. 6.10.2. Technology. 6.10.3. Information. 6.11. Exercises.

7. Applications 7.1. Introduction. 7.2. Partial Equilibrium Interpretation. 7.2.1. Partial equilibrium investment under uncertainty. 7.3. Introduction. 7.4. A Housing Model. 7.4.1. Demand. 7.4.2. House producers. 7.5. Cattle Cycles. 7.5.1. Mapping cattle farms into our framework. 7.5.2. Preferences. 7.5.3. Technology. 7.6. Models of Occupational Choice and Pay. 7.6.1. A one-occupation model. 7.6.2. Skilled and unskilled workers. 7.7. A Cash-in-Advance Model. 7.7.1. Reinterpreting the household technology. 7.8. Taxation in a Vintage Capital Model. A. Decentralizing the Household.

139

Contents

8. Efficient Computations 8.1. Introduction. 8.2. The Optimal Linear Regulator Problem. 8.3. Transformations to eliminate discounting and cross-products. 8.4. Stability Conditions. 8.5. Invariant Subspace Methods. 8.5.1. P x as Lagrange multiplier. 8.5.2. Invariant subspace methods. 8.5.3. Distorted Economies. 8.5.4. Transition Dynamics. 8.6. The Doubling Algorithm. 8.7. Partitioning the State Vector. 8.8. The Periodic Optimal Linear Regulator. 8.9. A Periodic Doubling Algorithm. 8.9.1. Partitioning the state vector. 8.10. Linear Exponential Quadratic Gaussian Control. 8.10.1. Doubling algorithm. A. Concepts of Linear Control Theory. B. Symplectic Matrices. C. Alternative forms of Riccati equation.

vii

157

viii

Contents

Part II: Representations and Properties

9. Representation and Estimation

187

9.1. The Kalman Filter. 9.2. Innovations Representation. 9.3. Convergence results. 9.3.1. Time-Invariant Innovations Representation. 9.4. Serially Correlated Measurement Errors. 9.5. Combined System. 9.6. Recursive Formulation of Likelihood Function. 9.6.1. Initialization. 9.6.2. Non-existence of a stationary distribution. 9.6.3. Serially correlated measurement errors. 9.7. Wold Representation. 9.8. Vector Autoregression for {yt }. 9.8.1. The factorization identity. 9.8.2. Location of zeros of characteristic polynomial. 9.8.3. Wold and autoregressive representations (white measurement errors). 9.8.4. Serially correlated measurement errors. 9.9. Innovations in yt+1 as Functions of Innovations wt+1 and ηt+1 . 9.10. Innovations in the yt ’s and the wt ’s in a Permanent Income Model. 9.10.1. Preferences. 9.10.2. Technology. 9.10.3. Information. 9.11. Frequency Domain Estimation. 9.12. Approximation Theory. 9.13. Aggregation Over Time. 9.14. Simulation Estimators. A. Initialization of the Kalman Filter.

10. Semiparametric Estimation with Limited Information

227

10.1. Introduction. 10.2. Underlying Economic Model. 10.3. Econometrician’s information and the implied orthogonality conditions. 10.4. An Adjustment Cost Example. 10.5. A Slightly Simpler Estimation Problem. 10.5.1. Scalar Parameterizations of B . 10.6. Multidimensional Parameterizations of B . 10.7. Nonparametric Estimation of B . 10.8. Back to the Adjustment Cost Model.

11. Representation of Demand 11.1. Introduction. 11.2. Canonical Representations of Services. 11.3. Dynamic Demand Functions for Consumption Goods. 11.3.1. The multiplier µw 0 . 11.3.2. Dynamic Demand System. 11.3.3. Foreshadow of Gorman aggregation. 11.4. Computing Canonical Representations. 11.4.1. Heuristics. 11.4.2. An auxiliary problem that induces a canonical representation. 11.5. Operator Identities. 11.6. Becker-Murphy Model of Rational Addiction. A. Fourier transforms. 11.A.1. Primer on transforms. 11.A.2. Time reversal and Parseval’s formula. 11.A.3. One sided

239

Contents

ix

sequences. 11.A.4. Useful properties. 11.A.5. One sided transforms. 11.A.6. Discounting. 11.A.7. Fourier transforms. 11.A.8. Verifying Equivalent Valuations. 11.A.9. Equivalent representations of preferences. 11.A.10. First term: factorization identity. 11.A.11. Second term. 11.A.12. Third term.

12. Gorman Heterogeneous Households

265

12.1. Introduction. 12.2. A Digression on Gorman Aggregation. 12.3. An Economy with Heterogeneous Consumers. 12.4. Allocations. 12.4.1. Consumption sharing rules. 12.5. Risk Sharing Implications. 12.6. Implementing the Allocation Rule with Limited Markets. 12.7. A Computer Example. 12.8. Exercises. 12.8.1. Part one. 12.8.2. Part two. 12.9. Economic integration. 12.9.1. Preferences:. 12.9.2. Technology. 12.9.3. Information.

13. Permanent Income Models

287

13.1. Technology. 13.2. Two Implications. 13.3. Solution. 13.4. Deterministic Steady States. 13.5. Cointegration. 13.6. Constant Marginal Utility of Income. 13.7. Consumption Externalities. 13.8. Tax Smoothing Models.

14. Non-Gorman Heterogeneity Among Households 14.1. Introduction. 14.2. Households’ Preferences. 14.2.1. Technology. 14.3. A Pareto Problem. 14.4. Competitive Equilibrium. 14.4.1. Households. 14.4.2. Firms of type I and II. 14.4.3. Definition of competitive equilibrium. 14.5. Computation of Equilibrium. 14.5.1. Candidate equilibrium prices. 14.5.2. A Negishi algorithm. 14.6. Mongrel Aggregation. 14.6.1. Static demand. 14.6.2. Frequency domain representation of preferences. 14.7. A Programming Problem for Mongrel Aggregation. 14.7.1. Factoring S ′ S . 14.8. Summary of Findings. 14.9. The Mongrel Preference Shock Process. 14.9.1. Interpretation of sˆt component. 14.10. Choice of Initial Conditions.

307

x

Contents

Part III: Extensions

15. Equilibria with Distortions

333

15.1. Introduction. 15.2. A Representative Agent Economy with Distortions. 15.2.1. a. Consumption externalities. 15.2.2. b. Production externalities. 15.2.3. c. Taxes. 15.3. Households. 15.4. Firms. 15.5. Information. 15.6. Equilibrium. 15.7. Heterogeneous Households with Distortions. 15.7.1. Households. 15.7.2. Firms of type I. 15.7.3. Firms of type II. 15.7.4. Government. 15.7.5. Definition of equilibrium. 15.7.6. Equilibrium computation. 15.8. Government Deficits and Debt. 15.9. Examples. 15.9.1. A production externality. 15.9.2. Consumption tax only. 15.9.3. Machinery investment subsidy. 15.9.4. ‘Personal’ habit persistence. 15.9.5. ‘Social’ habit persistence. 15.10. Conclusions. A. Invariant subspace equations for first specification. 15.A.1. Household’s Lagrangian. 15.A.2. Firm’s first order conditions. 15.A.3. Representativeness conditions. B. Invariant subspace equations for heterogeneous agent model.

16. Recursive Risk Sensitive Control

369

16.1. Introduction. 16.2. A Control Problem. 16.3. Pessimistic Interpretation. 16.4. Recursive Preferences. 16.4.1. Endowment economy. 16.5. Asset Pricing. 16.6. Characterizing the Pricing Expectations Operator. 16.7. Production Economies. 16.8. Risk-Sensitive Investment under Uncertainty. 16.9. Equilibrium Prices in the Adjustment Cost Economies.

17. Periodic Models of Seasonality 17.1. Introduction. 17.2. A Periodic Economy. 17.3. Asset Pricing. 17.4. Prediction Theory. 17.5. The Term Structure of Interest Rates. 17.6. Conditional Covariograms. 17.7. The Stacked and Skip-Sampled System. 17.8. Covariances of the Stacked, Skip Sampled Process. 17.9. The Tiao-Grupe Formula. 17.9.1. A state space realization of the TiaoGrupe formulation. 17.10. Some Calculations with a Periodic Hall Model. 17.11. Periodic Innovations Representations for the Periodic Model. A. A Model of Disguised Periodicity. 17.13. A1. Two Illustrations of Disguised Periodicity. 17.14. A2. Mathematical Formulation of Disguised Periodicity.

387

Contents

xi

Part IV: Economies as Objects

18. Introduction to Objects

429

18.1. Matlab Objects. 18.1.1. Definitions. 18.1.2. Matlab Specifics. 18.1.3. How to Define a Matlab Class. 18.2. Summary.

19. Economies as Matlab Objects

435

19.1. Introduction. 19.2. Parent Classes: Information. 19.2.1. Structure. 19.2.2. Functions. 19.3. Parent Classes: Technology. 19.3.1. Structure. 19.3.2. Functions. 19.4. Parent Classes: Preferences. 19.4.1. Structure. 19.4.2. Functions. 19.5. Child Class: Economy. 19.5.1. Structure. 19.5.2. Fields containing the history of the economy. 19.5.3. Functions. 19.5.4. Constructing the object and changing parameters. 19.5.5. Analyzing the economy. 19.6. Working with economies. 19.6.1. The built-in economies. 19.6.2. Mixing and matching built-in parent objects. 19.6.3. Building your own economy. 19.7. Tutorial.

20. MATLAB Programs

445

20.1. Matlab programs.

21. References

499

22. Index

513

23. Author Index

517

24. Matlab Index

519

Acknowledgements

– xii –

Preface

– xiii –

Part I Components of an economy

Chapter 1 Introduction

1.1. Introduction This book views many apparently disparate dynamic economic models as examples of a single class of models that can be adapted and specialized to study diverse economic phenomena. The class of models was created by using recent advances in (i) the theory of recursive dynamic competitive economies; 1 (ii) methods for estimating and interpreting vector autoregression; 2 (iii) linear optimal control theory; 3 and (iv) computer languages for rapidly manipulating linear optimal control systems. 4 We combine these elements to build a class of models for which the competitive equilibria are vector autoregressions that can be swiftly computed, represented, and simulated using the methods of linear optimal control theory. We use the computer language MATLAB to implement the computations. This language has a powerful vocabulary and a convenient structure that liberate time and energy from programming, and thereby spur creative application of linear control theory. Our goal has been to create a class of models that merge recursive economic theory and with dynamic econometrics. Systems of autoregressions and of mixed autogregressive, moving average processes are a dominant setting for dynamic econometrics. We constructed our economic models by adopting a version of recursive competitive theory in which an outcome of theorizing is a vector autoregression. We formulated this class of models because practical difficulties of computing and estimating recursive equilibrium models still limit their use as a tool for thinking about applied problems in economic dynamics. Recursive competitive equilibria were themselves developed as a special case of the Arrow-Debreu competitive equilibrium, both to restrict the range of outcomes possible in the 1 2 3 4

This work is summarized by Harris (1987) and Stokey, Lucas, and Prescott (1989). See Sims (1980), Hansen and Sargent (1980, 1981, 1990). For example, see Kwakernaak and Sivan (1972), and Anderson and Moore (1979). See the MATLAB manual.

–3–

4

Introduction

Arrow-Debreu setting and to create a framework for studying applied problems in dynamic economies of long duration. Relative to the general Arrow-Debreu setting, the great advantage of the recursive competitive equilibrium formulation is that equilibria can be computed by solving a discounted dynamic programming problem. Further, under particular additional conditions, an equilibrium can be represented as a Markov process in the state variables. When that Markov process has an invariant distribution to which the process converges, there exists a vector autoregressive representation. Thus, the theory of recursive competitive equilibria holds out the promise of making closer contact with econometric theory than did previous formulations of equilibrium theory. Two computational difficulties have left much of this promise unrealized. The first is Bellman’s “curse of dimensionality” which usually makes dynamic programming a costly procedure for systems with even small numbers of state variables. The second problem is that after a dynamic program has been solved and the equilibrium Markov process computed, the vector autoregression implied by the theory has to be computed by applying classic projection formulas to a large number of second moments of the stationary distribution associated with that Markov process. Typically, each of these computational problems can be solved only approximately. Good research along a number of lines is now being directed at evaluating alternative ways of making these approximations. 5 The need to make these approximations originates in the fact that for general functional forms for objective functions and constraints, even one iteration on the functional equation of Richard Bellman cannot be performed analytically. It so happens that the functional forms economists would most like to use have been of this general class for which Bellman’s equation cannot be iterated upon analytically. Linear control theory studies the most important special class of problems for which iterations on Bellman’s equation can be performed analytically: problems with a quadratic objective function and a linear transition function. Application of dynamic programming leads to a system of well understood and rapidly solvable equations known as the matrix Riccati equation. The philosophy of this book is to swallow hard and to accept up front as primitive descriptions of tastes, technology, and information specifications that satisfy the assumptions of linear optimal control theory. This approach 5 See Marcet (1989) and Judd (1990). Also see Coleman (1990) and Tauchen (1990).

Organization

5

purchases the ability rapidly to compute equilibria together with a form of equilibrium that is automatically in the form of a vector autoregression. A cost of the approach is that it does not accommodate many specifications that we would like to be able to analyze. The purpose of this book is to display the versatility and tractability of our class of models. Versions of a wide range of models from modern capital theory and asset pricing theory can be represented within our framework. The equilibria of these models can be computed so easily that we hope that the reader will soon be thinking of improvements to our specifications. We provide formulas and software for the reader to experiment.

1.2. Computer Programs In writing this book, we put ourselves under a restriction that we should supply the reader with a computer program that implements every equilibrium concept and mathematical representation that we describe. The programs are written in MATLAB, and are described throughout the book. When a MATLAB program is referred to in the text, we place it in typewriter font. Similarly, all computer code is placed in typewriter font. 6 You will get much more out of this book if you use and modify our programs as you read.

1.3. Organization This book is organized as follows. Chapter 10 describes the first order linear vector stochastic difference equation, and shows how special cases of it are formed by a variety of models of time series processes that have been studied by economists. This difference equation will be used to represent the information flowing to economic agents within our models. It will also be used to represent the equilibrium of the model. Chapter 3 defines an economic environment in terms of the preferences of a representative agent, the technology for producing goods, stochastic processes 6 To run our programs, you will need MATLAB’s Control Toolkit in addition to the basic MATLAB software.

6

Introduction

disturbing preferences and the technology, and the information structure of the economy. The stochastic processes fit into the model introduced in chapter 10, while the preferences, technology, and information structure are specified with an eye toward making the competitive equilibrium one that can be computed by the application of linear control theory. Chapter 4 describes a social planning problem associated with the equilibrium of the model. The problem is formulated in two ways, first as a variational problem using stochastic Lagrange multipliers, and then as a dynamic programming problem. We describe how to compute the solution of the dynamic programming problem using formulas from linear control theory. The solution of the social planning problem is a first order vector stochastic difference equation of the form studied in chapter 10. We also show how to use the value function for the social planning problem to compute the Lagrange multipliers associated with the planning problem. These multipliers are later used in chapter 6 to compute the equilibrium price system. Chapter 5 describes the price system and the commodity space that support a competitive equilibrium. We use a formulation that lets the values that appear in agents’ budget constraints and objective functions be represented as conditional expectations of geometric sums of streams of future “prices” times quantities. Chapter 5 relates these prices to Arrow-Debreu state contingent prices. Chapter 6 describes a decentralized version of our economy, and defines and computes a competitive equilibrium. Competitive equilibrium quantities solve a social planning problem. The price system can be deduced from the stochastic Lagrange multipliers associated with the social planning problem. Chapter 7 describes versions of several dynamic models from the literature that fit easily within our class of models. Chapter 9 describes the links between our theoretical equilibrium and autoregressive representations of time series of observables. We show how to obtain an autoregressive representation for a list of observable variables that are linear functions of the state variables of the model. The autoregressive representation is naturally affiliated with a recursive representation of the likelihood function for the observable variables. In describing how to deduce the autoregressive representation from the parameters determining the equilibrium of the model, and possibly also from parameters of measurement error processes, we are completing a key step needed to permit econometric estimation of the model’s free

Organization

7

parameters. Chapter 9 also treats two other topics intimately related to econometric implementation of the models; aggregation over time, and the theory of approximation of one model by another. Chapter 8 describes fast methods to compute equilibria. We describe how doubling algorithms can speed the computation of expectations of geometric sums of quadratic forms, and help to solve dynamic programming problems. Chapter 11 describes alternative ways to represent demand. It identifies an equivalence class of preference specifications that imply the same demand functions, and characterizes a special subset of them as canonical household preferences. Canonical representations of preferences are useful for describing economies with heterogeneity among household’s preferences. Chapter 12 describes a version of our economy with the type of heterogeneity among households allowed when preferences aggregate in a sense introduced by Terrance Gorman . In this setting, affine Engle curves of common slope prevail and give rise to a representative consumer. This representative consumer is ‘easy to find,’ and from the point of view of equilibrium computation of prices and aggregate quantities, adequately stands in for the household of chapters 3–6. The allocations to individual consumers require additional computations, which this chapter describes. Chapter 13 uses our model of preferences to represent multiple goods versions of permanent income models along the lines of Robert Hall’s (1978). We retain Hall’s specification of the ‘storage’ technology for accumulating physical assets, and also the restriction on the discount factor, depreciation rate, and gross return on capital that delivered to Hall a martingale for the marginal utility of consumption. Adopting Hall’s specification of the storage technology imparts a martingale characterization to the model, but it is hidden away in an ‘index’ whose increments drive the behavior of consumption demands for various goods, which themselves are not martingales. This model forms a convenient laboratory for thinking about the sources in economic theory of ‘unit roots’ and ‘co-integrating vectors.’ Chapter 14 describes a setting in which there is more heterogeneity among households’ preferences, causing the conditions for Gorman aggregation to fail. Households’ Engle curves are still affine, but dispersion of their slopes arrests Gorman aggregation. There is another sense, originating with Negishi, in which there is a representative household whose preferences represent a complicated kind of average over the preferences of different types of households. We show

8

Introduction

how to compute and interpret this preference ordering over economy-wide aggregates. This average preference ordering cannot be computed before one knows the distribution of wealth evaluated at equilibrium prices. Chapter 15 describes economies with production and consumption externalities and also distortions due to a government’s imposing distorting flat rate taxes. Equilibria of these economies has to be computed by a direct attack on Euler equations and budget constraints, rather than via dynamic programming for an artificial social planning problem. Chapter 16 describes a recursive version of Jacobson’s and Whittle’s ‘risk sensitive’ preferences. This preference specification has the features that, although it violates certainty equivalence – so that the conditional covariance of forecast error distributions impinge on equilibrium decision rules – it does so in a way that preserves linear equilibrium laws of motion, and retains calculation of equilibria and asset prices via simple modifications of our standard formulas. These preferences are a version of those studied by Epstein and Zin ( ) and Weil ( ). Chapter 17 describes how to adapt our setup to include features of the periodic models of seasonality that have been studied by Osborne (1988), Todd (1990), and Ghysels (1993). Chapter 20 is a manual of the MATLAB programs that we have prepared to implement the calculations described in this book. The design is consistent with other MATLAB manuals. The notion of duality and the ‘factorization identity’ from recursive linear optimal control theory are used repeatedly in Chapter 9 (on representing equilibria econometrically), and chapters 11, 12, and 14 (on representing and aggregating preferences). ‘Duality’ is the observation that recursive filtering problems (Kalman filtering) have the same mathematical structure as recursive formulations of linear optimal control problems (leading to Riccati equations via dynamic programming). That duality applies so often in our settings in effect ‘halves’ the mathematical apparatus that we require.

Chapter 2 Stochastic Linear Difference Equations

2.1. Introduction This chapter introduces the first-order vector linear stochastic difference equation, which we use in two important ways. We use it first to represent the information flowing to economic agents, then again to represent equilibria of our models. The first-order linear stochastic difference equation is associated with a tidy theory of prediction and a host of procedures for econometric application. Their ease of analysis has prompted us to adopt economic specifications that cause our equilibria to have representations in terms of a first-order linear stochastic difference equation. The first order vector stochastic difference equation is recursive because it expresses next period’s vector of state variables as a linear function of this period’s state vector and a vector of new disturbances to the system. These disturbances form a “martingale difference sequence,” and are the basic building block out of which the time series are created. Martingale difference sequences are easy to forecast, a fact that delivers convenient recursive formulas for optimal predictions.

2.2. Notation and Basic Assumptions Let {xt : t = 1, 2, . . .} be a sequence of n -dimensional random vectors, i.e. an n -dimensional stochastic process. The vector xt contains variables observed by economic agents at time t . Let {wt : t = 1, 2, . . .} be a sequence of N dimensional random vectors. The vectors {wt } will be treated as building blocks for {xt : t = 1, 2, . . .}, in the sense that we shall be able to express xt as the sum of two terms. The first is a moving average of past wt ’s. The second describes the effects of an initial condition. The {wt } process is used to generate a sequence of information sets {Jt : t = 0, 1, . . .}. Let J0 be generated by x0 and Jt be generated by x0 , w1 , . . . , wt , which means that Jt consists of the set

–9–

10

Stochastic Linear Difference Equations

of all measurable functions of {x0 , w1 , . . . , wt }. 1 The building block process is assumed to be a martingale difference sequence adapted to this sequence of information sets. We explain what this means by advancing the following Definition 1: The sequence {wt : t = 1, 2, . . .} is said to be a martingale difference sequence adapted to {Jt : t = 0, 1, . . .} if E(wt+1 |Jt ) = 0 for t = 0, 1, . . . . In addition, we assume that the building block process is conditionally homoskedastic, a phrase whose meaning is conveyed by Definition 2: The sequence {wt : t = 1, 2, . . .} is said to be conditionally ′ homoskedastic if E(wt+1 wt+1 | Jt ) = I for t = 0, 1, . . . . It is an implication of the law of iterated expectations that {wt : t = 1, 2, . . .} is a sequence of (unconditional) mean zero, serially uncorrelated random vectors. 2 In addition, the entries of wt are assumed to be mutually uncorrelated. The process {xt : t = 1, 2, . . .} is constructed recursively using an initial random vector x0 and a time invariant law of motion: xt+1 = Axt + Cwt+1 ,

for t = 0, 1, . . . ,

(2.2.1)

where A is an n by n matrix and C is an n by N matrix. Representation (2.2.1) will be a workhorse in this book. First, we will use (2.2.1) to model the information upon which economic agents base their decisions. Information will consist of variables that drive shocks to preferences and to technologies. Second, we shall specify the economic problems faced by the agents in our models and the economic process through which agents’ decisions 1 The phrase “ J is generated by x ” means that J can be expressed as a measurable 0 0 0 function of x0 . 2 Where φ and φ are information sets with φ ⊂ φ , and x is a random variable, the 1 2 1 2 law of iterated expectations states that E (x | φ1 ) = E (E (x | φ2 ) | φ1 ) . Letting φ1 be the information set corresponding to no observations on any random variables, letting φ2 = Jt , and applying this law to the process {wt } , we obtain

E wt+1 = E E wt+1 | Jt

= E (0) = 0.

Prediction Theory

11

are coordinated (competitive equilibrium) so that the state of the economy has a representation of the form (2.2.1).

2.3. Prediction Theory A tractable theory of prediction is associated with (2.2.1). This theory is used extensively both in computing the equilibrium of the model and in representing that equilibrium in the form of (2.2.1). The optimal forecast of xt+1 given current information is E (xt+1 | Jt ) = Axt ,

(2.3.1)

and the one-step-ahead forecast error is xt+1 − E (xt+1 | Jt ) = Cwt+1 .

(2.3.2)

The covariance matrix of xt+1 conditioned on Jt is just CC ′ : ′

E (xt+1 − E (xt+1 | Jt )) (xt+1 − E (xt+1 | Jt )) = CC ′ .

(2.3.3)

Sometimes we use a nonrecursive expression for xt as a function of x0 , w1 , w2 , . . . , wt . Using (2.2.1) repeatedly, we obtain xt = Axt−1 + Cwt = A2 xt−2 + ACwt−1 + Cwt t−1 i hX Aτ Cwt−τ + At x0 . =

(2.3.4)

τ =0

Representation (2.3.4) is one type of moving-average representation. It expresses {xt : t = 1, 2, . . .} as a linear function of current and past values of the building block process {wt : t = 1, 2, . . .} and an initial condition x0 . 3 3 Slutsky (1937) argued that business cycle fluctuations could be well modelled by moving average processes. Sims (1980) showed that a fruitful way to summarize correlations between time series is to calculate an impulse response function. In chapter 8, we study the relationship between the impulse response functions calculated by Sims (1980) and the impulse response function associated with ( 2.3.4 ).

12

Stochastic Linear Difference Equations

The moving average piece of representation (2.3.4) is often called an impulse response function. An impulse response function depicts the response of current and future values of {xt } to an imposition of a random shock wt . In representation (2.3.4), the impulse response function is given by entries of the vector sequence {Aτ C : τ = 0, 1, . . .}. 4 Shift (2.3.4) forward in time: xt+j =

j−1 X

As Cwt+j−s + Aj xt .

(2.3.5)

s=0

Projecting both sides of (2.3.5) on the information set {x0 , wt , wt−1 , . . . , w1 } gives 5 Et xt+j = Aj xt . (2.3.6) where Et (·) ≡ E[(·) | x0 , wt , wt−1 , . . . , w1 ] = E(·) | Jt , and xt is in Jt . Equation (2.3.6) gives the optimal j step ahead prediction. It is useful to obtain the covariance matrix of the j -step ahead prediction error j−1 X xt+j − Et xt+j = As Cwt−s+j (2.3.7) s=0

We have

E(xt+j − Et xt+j ) (xt+j − Et xt+j )′ =

j−1 X

k=0

′

Ak CC ′ Ak ≡ vj

(2.3.8a)

Note that vj defined in (2.3.8a) can be calculated recursively via v1 = CC ′ vj = CC ′ + Avj−1 A′ ,

j ≥ 2.

(2.3.8b)

The matrix vj is the covariance matrix of the errors in forecasting xt+j on the basis of time t information xt . To decompose these covariances into parts attributable to the individual components of wt , we let iτ be an N -dimensional 4 Given matrices A and C , the impulse response function can be calculated using the MATLAB program dimpulse.m. 5 For an elementary discussion of linear least squares projections, see Sargent (1987b, chapter IX).

Transforming Variables to Uncouple Dynamics

13

column vector of zeroes except in position τ , where there is a one. Define a matrix υj,τ by j−1 X ′ (2.3.8c) Ak Ciτ i′τ C ′ A k . υj,τ = k=0

Note that

PN

′ τ =1 iτ iτ

= I , so that from (2.3.8a) and (2.3.8c) we have N X

υj,τ = υj .

τ =1

Evidently, the matrices {υj,τ , τ = 1, . . . , N } give an orthogonal decomposition of the covariance matrix of j -step ahead prediction errors into the parts attributable to each of the components τ = 1, . . . , N . 6 The “innovation accounting” methods of Sims (1980) are based on (2.3.8). Sims recommends computing the matrices vj,τ in (2.3.8) for a sequence j = 0, 1, 2, . . . . This sequence represents the effects of components of the shock process wt on the covariance of j -step ahead prediction errors for each series in xt .

2.4. Transforming Variables to Uncouple Dynamics A convenient analytical device for the analysis of linear system (2.2.1) is to uncouple the dynamics using the distinct eigenvalues of the matrix A . We use the Jordan decomposition of the matrix A : A = T DT −1 ,

(2.4.1)

where T is a nonsingular matrix and D is a matrix constructed as follows. Recall that the eigenvalues of A are the zeroes of the polynomial det (ζI − A). This polynomial has n zeroes because A is n by n . Not all of these zeroes are necessarily distinct, however. 7 Suppose that there are m ≤ n distinct zeroes 6 For given matrices A and C , the matrices v j,τ and vj are calculated by the MATLAB program evardec.m. 7 In the case in which the eigenvalues of A are distinct, D is taken to be the diagonal matrix whose entries are the eigenvalues and T is the matrix of eigenvectors corresponding to those eigenvalues.

14

Stochastic Linear Difference Equations

of this polynomial, denoted δ1 , δ2 , . . . , δm . For each δj , we construct a matrix Dj that has the same dimension as the number of zeroes of det (ζI − A) that are equal to δj . The diagonal entries of Dj are δj and the entries in the single diagonal row above the main diagonal are all either zero or one. The remaining entries of Dj are zero. Then the matrix D is block diagonal with Dj in the j th diagonal block. Transform the state vector xt as follows: x∗t = T −1 xt .

(2.4.2)

Substituting into (2.2.1), we have that x∗t+1 = Dx∗t + T −1 Cwt+1 .

(2.4.3)

Since D is block diagonal, we can partition x∗t according to the diagonal blocks of D or, equivalently, according to the distinct eigenvalues of A . In the law of motion (2.4.3), partition j of x∗t+1 is linked only to partition j of x∗t . In this sense, the dynamics of system (2.4.3) are uncoupled. To calculate multi-period forecasts and dynamic multipliers, we must raise the matrix A to integer powers (see (2.3.6)). It is straightforward to verify that Aτ = T (Dτ )T −1 .

(2.4.4)

Since D is block diagonal, Dτ is also block diagonal, where block j is just (Dj )τ . The matrix (Dj )τ is upper triangular with δjτ on the diagonal, with all entries of the k th upper right diagonal given by (δj )τ −k τ !/[k!(τ − k)!] for 0 ≤ k ≤ τ,

(2.4.5)

and zeroes elsewhere. Consequently, raising D to an integer power involves raising the eigenvalues to integer powers. Some of the eigenvalues of A may be complex. In this case, it is convenient to use the polar decomposition of the eigenvalues. Write eigenvalue δj in polar form as δj = ρj exp(iθj ) = ρj [cos(θj ) + i sin(θj )]

(2.4.6)

where ρj =| δj |. Then δjτ = (ρj )τ exp(iτ θj ) = (ρj )τ [cos(τ θj ) + i sin(τ θj )].

(2.4.7)

Examples

15

We shall often assume that ρj is less than or equal to one, which rules out instability in the dynamics. Whenever ρj is strictly less than one, the term (ρj )τ decays to zero as τ → ∞ . When θj is different from zero, eigenvalue j induces an oscillatory component with period (2π/ | θj |).

2.5. Examples Next we consider some examples of processes that can be accommodated by (2.2.1).

2.5.1. Deterministic seasonals We use (2.2.1) to represent the model yt = yt−4 . Let n = 4, C = 0, xt = (yt , yt−1 , yt−2 , yt−3 )′ , x0 = (0 0 0 1)′ , 0 1 A= 0 0

0 0 1 0

0 0 0 1

1 0 , 0 0

0 0 C= 0.

(2.5.1)

0

In this case the A matrix has four distinct eigenvalues and the absolute values of each of these eigenvalues is one. Two eigenvalues are real (1, −1) and two eigenvalues are imaginary (i, −i), and so have period four. The resulting sequence {xt : t = 1, 2, . . .} oscillates deterministically with period four. It can be used to model deterministic seasonals in quarterly time series.

16

Stochastic Linear Difference Equations

2.5.2. Indeterministic seasonals We want to use (2.2.1) to represent the model yt = α4 yt−4 + wt ,

(2.5.2)

where wt is a martingale difference sequence and | α4 |≤ 1. We define xt = [yt , yt−1 , yt−2 , yt−3 ]′ , n = 4, 0 1 A= 0 0

0 0 1 0

0 0 0 1

α4 1 0 0 . , C= 0 0 0 0

With these definitions, (2.2.1) represents (2.5.2). This model displays an “indeterministic” seasonal. Realizations of (2.5.2) display recurrent, but aperiodic, seasonal fluctuations.

2.5.3. Univariate autoregressive processes We can use (2.2.1) to represent the model yt = α1 yt−1 + α2 yt−2 + α3 yt−3 + α4 yt−4 + wt ,

(2.5.3)

where wt is a martingale difference sequence. We set n = 4, xt = [yt yt−1 yt−2 yt−3 ]′ , α1 1 A= 0 0

α2 0 1 0

α3 0 0 1

α4 1 0 0 , C = . 0 0 0 0

The matrix A has the form of the companion matrix to the vector [α1 α2 α3 α4 ].

Examples

17

2.5.4. Vector autoregressions Reinterpret (2.5.3) as a vector process in which yt is a (k × 1) vector, αj a (k × k) matrix, and wt a k × 1 martingale difference sequence. Then (2.5.3) is termed a vector autoregression. To map this into (2.2.1), we set n = k · 4, I α1 α2 α3 α4 0 I 0 0 0 , C = A= 0 0 I 0 0 0

0

I

0

0

where I is the (k × k) identity matrix.

2.5.5. Polynomial time trends Let n = 2, x0 = [0 1]′ , and A=

1 0

1 0 , C= . 1 0

(2.5.4)

Notice that D = A in the Jordan decomposition of A . It follows from (2.4.5) that 1 t t . (2.5.5) A = 0 1

Hence xt = (t, 1)′ , so that the first component of xt is a linear time trend and the second component is a constant. It is also possible to use (2.2.1) to represent polynomial trends of any order. For instance, let n = 3, C = 0, x0 = (0, 0, 1)′ , and 1 1 0 A = 0 1 1. (2.5.6) 0 0 1

Again, A = D in the Jordan decomposition of A . It follows from (2.4.5) that 1 t t(t − 1)/2 . At = 0 1 t (2.5.7) 0 0 1

Then x′t = [t(t−1)/2, t, 1], so that xt contains linear and quadratic time trends.

18

Stochastic Linear Difference Equations

2.5.6. Martingales with drift We modify the linear time trend example by Suppose that making C nonzero. 1 t 1 1 t ′ , it follows that and A = N is one and C = [1 0]. Since A = 0 1 0 1 1 τ A C= . (2.5.8) 0 Substituting into the moving-average representation (2.3.4), we obtain (2.25) x1t =

t−1 X

wt−τ + [1 t]x0

τ =0

where x1t is the first entry of xt . The first term on the right-hand side of the preceding equation is a cumulated sum of martingale differences, and is called a martingale, while the second term is a translated linear function of time.

2.5.7. Covariance stationary processes Next we consider specifications of x0 and A which imply that the first two moments of {xt : t = 1, 2, . . .} are replicated over time. Let A satisfy A11 A12 A= , (2.5.9) 0 1 where A11 is an (n − 1) × (n − 1) matrix with eigenvalues that have moduli strictly less than one and A12 is an (n − 1) × 1 column vector. In addition, let C ′ = [C1′ 0]. We partition x′t = [x′1t x′2t ] where x1t has n − 1 entries. It follows from (2.2.1) that x1t+1 = A11 x1t + A12 x2t + C1 wt+1 (2.5.10) x2t+1 = x2t .

(2.5.11)

By construction, the second component, x2t , simply replicates itself over time. For convenience, take x20 = 1 so that x2t = 1 for t = 1, 2, . . . . We can use (2.5.10) to compute the first two moments of x1t . Let µt = Ex1t . Taking unconditional expectations on both sides of (2.5.10) gives µt+1 = A11 µt + A12 .

(2.5.12)

Examples

19

We can solve the nonstochastic difference equation (2.5.12) for the stationary value of µt . Define µ as the stationary value of µt , and substitute µ for µt and µt+1 in (2.5.12). Solving for µ gives µ = (I − A11 )−1 A12 . Therefore, if Ex10 = (I − A11 )−1 A12 ,

(2.5.13)

then Ex1t will be constant over time and equal to the value on the right side of (2.5.13). Further, if the eigenvalues of A11 are less than unity in modulus, then starting from any initial value of µ0 , µt will converge to the stationary value (I − A11 )−1 A12 . Next we use (2.5.10) to compute the unconditional covariances of xt . Subtracting (2.5.12) from (2.5.10) gives (x1t+1 − µt+1 ) = A11 (x1t − µt ) + C1 wt+1

(2.5.14)

From (2.5.14) it follows that (x1t+1 − µt+1 )(x1t+1 − µt+1 )′ = A11 (x1t − µt )(x1t − µt )′ A′11

′ ′ + C1 wt+1 wt+1 C1′ + C1 wt+1 (x1t − µt )′ A′11 + A11 (x1t − µt )wt+1 C1′ .

The law of iterated expectations implies that wt+1 is orthogonal to (x1t − µt ). Therefore, taking expectations on both sides of the above equation gives Vt+1 = A11 Vt A′11 + C1 C1′ , where Vt ≡ E(x1t − µt )(x1t − µt )′ . Evidently, the stationary value V of the covariance matrix Vt must satisfy V = A11 V A′11 + C1 C1′ .

(2.5.15)

It is straightforward to verify that V is a solution of (2.5.15) if and only if V =

∞ X

Aj11 C1 C1′ Aj′ 11 .

(2.5.16)

j=0

The infinite sum (2.5.16) converges under the condition that the eigenvalues of A11 are less in modulus than unity. 8 If the covariance matrix of x10 is V and 8 Equation ( 2.5.15 ) is known as the discrete Lyapunov equation. Given the matrices A 11 and C1 , this equation is solved by the MATLAB program dlyap.m.

20

Stochastic Linear Difference Equations

the mean of x10 is (I − A11 )−1 A12 , then the covariance and mean of x1t remain constant over time. In this case, the process is said to be covariance stationary. If the eigenvalues of A11 are all less than unity in modulus, then Vt → V as t → ∞ , starting from any initial value V0 . From (2.3.8) and (2.5.16), notice that if all of the eigenvalues of A11 are less than unity in modulus, then limj→∞ vj = V . That is, the covariance matrix of j -step ahead forecast errors converges to the unconditional covariance matrix of x as the horizon j goes to infinity. 9 The matrix V can be decomposed according to the contributions of each entry of the process {wt }. Let ιτ be an N -dimensional column vector of zeroes except in position τ , where there is a one. Then N X

ιτ ι′τ .

(2.5.17)

(A11 )j C1 ιτ ι′τ C1′ (A11 )j′

(2.5.18)

I=

τ =1

Define a matrix V˜τ V˜τ ≡

∞ X j=o

We have, by analogy to (2.5.15) and (2.5.16), that V˜τ satisfies V˜τ = A11 V˜τ A′11 + C1 iτ i′τ C1′ . In light of (2.5.17), (2.5.18), and (2.5.16) we have that V =

N X

V˜τ .

(2.5.19)

τ =1

The matrix V˜τ has the interpretation of being the contribution to V of the τ th component of the process {wt : t = 1, 2, . . .}. Hence, (2.5.19) gives a decomposition of the covariance matrix V into the portions attributable to each of the underlying economic shocks. Next, consider the autocovariances of {xt : t = 1, 2, . . .}. From the law of iterated expectations, it follows that E[(x1t+τ − µ)(x1t − µ)′ ] = E{E[(x1t+τ − µ) | Jt ](x1t − µ)′ } = E[Aτ11 (x1t − µ)(x1t − µ)′ ]

=

(2.5.20)

Aτ11 V.

9 The doubling algorithm described in chapter 9 can be used to compute the solution of ( 2.5.15 ) via iterations that approximate ( 2.5.16 ). The algorithm is implemented in the MATLAB programs doublej.m and doublej2.m .

Examples

21

Notice that this expected cross-product or autocovariance does not depend on calendar time but only on the gap τ between the time indices. 10 Independence of means, covariances, and autocovariances from calendar time defines covariance stationary processes. For the particular class of processes we are considering, if the covariance matrix does not depend on calendar time, then none of the autocovariance matrices does.

2.5.8. Multivariate ARMA processes Specification (2.2.1) assumes that xt contains all the information that is available at time t to forecast xt+1 . In many applications, vector time series are modelled as multivariate autoregressive moving-average (ARMA) processes. Let yt be a vector stochastic process. An ARMA process {yt : t = 1, 2, . . .} has a representation of the form: yt = α1 yt−1 + α2 yt−2 + · · · + αk yt−k

+ γ0 wt + γ1 wt−1 + · · · + γk wt−k .

(2.5.21)

where E[wt | yt−1 , yt−2 , · · · yt−k+1 , wt−1 , wt−2 , · · · wt−k+1 ] = 0. The requirement that the same number of lags of y enter (2.5.21) as the number of lags of w is not restrictive because some coefficients can be set to zero. Hence we can think of k as being the greater of the two lag lengths. A representation such as (2.5.21) can be shown to satisfy (2.2.1). To see this, we define

xt =

yt α2 yt−1 + α3 yt−2 · · · + αk yt−k+1 + γ1 wt + γ2 wt−1 · · · + γk−1 wt−k+2 + γk wt−k+1 α3 yt−1 · · · + αk yt−k+2 + γ2 wt · · · + γk−1 wt−k+3 + γk wt−k+2 . .. αk yt−1 + γk−1 wt + γk wt−1 γ k wt (2.5.22)

10 Equation ( 2.5.20 ) shows that the matrix autocovariogram of x (i.e., Γτ ≡ E[(x 1t 1t+τ − µ)(x1t − µ)′ ] taken as a function of τ ) satisfies the nonrandom difference equation Γt+1 = A11 Γt with initial condition Γ0 = V .

22

Stochastic Linear Difference Equations

γ0 γ1 C = .. .

(2.5.23)

γk

and

α1 α2 . A = .. αk 0

I 0 .. .

··· ··· .. .

0 0

··· ···

0 0 .. . I

(2.5.24)

0

It is straightforward to verify that the resulting process {xt : t = 1, 2, . . .} satisfies (2.2.1).

2.5.9. Prediction of a univariate first order ARMA Consider the special case of (2.5.21) yt = α1 yt−1 + γ0 wt + γ1 wt−1

(2.5.25)

where yt is a scalar stochastic process and wt is a scalar white noise. Assume that | α1 |< 1 and that | γ1 /γ0 |< 1. Applying (2.5.22), we define the state xt as xt =

yt . γ1 wt

Applying (2.5.23) and (2.5.24), we have γ0 α1 C= , A= γ1 0

1 . 0

We can apply (2.3.6) to obtain a formula for the optimal j -step ahead prediction of yt . Using (2.3.6) in the present example gives j yt+j α1 α1j−1 yt Et = γ1 wt+j 0 0 γ1 wt which implies that Et yt+j = α1j yt + α1j−1 γ1 wt .

(2.5.26)

Examples

23

We can use (2.5.26) to derive a famous formula of John F. Muth (1960). Assume that the system (2.5.25) has been operating forever, so that the initial time is infinitely far in the past. Then using the lag operator L, express (2.5.25) as (1 − α1 L)yt = (γ0 + γ1 L)wt . Solving for wt gives wt = γ0−1

1 − α L 1 yt , 1 + γγ10 L

which expresses wt as a geometric distributed lag of current and past yt ’s. Substituting this expression for wt into (2.5.26) and rearranging gives Et yt+j = α1j−1

h α1 + 1+

γ1 γ0

γ1 γ0 L

i

yt .

In the limiting case as α1 → 1 from below, this formula becomes Et yt+j =

h 1+ 1+

γ1 i γ0 yt , γ1 γ0 L

(2.5.27)

which is independent of the forecast horizon j . In the limiting case of α1 = 1, it is optimal to forecast yt for any horizon as a geometric distributed lag of past y ’s. This is Muth’s finding that a univariate process whose first difference is a first order moving average is optimally forecast via an “adaptive expectations” scheme (i.e., a geometric distributed lag with the weights adding up to unity).

2.5.10. Growth In much of our analysis, we assume that the eigenvalues of A have absolute values less than or equal to one. We have seen that such a restriction still allows for polynomial growth. Geometric growth can also be accommodated by suitably scaling the state vector. For instance, suppose that {x+ t : t = 1, 2, . . .} satisfies: + + + x+ (2.5.28) t+1 = A xt + Cwt+1 + + + where E(wt+1 | Jt ) = 0 and E[wt+1 (wt+1 )′ | Jt ] = (ε)t I . The positive number ε can be bigger than one. The eigenvalues of A+ are assumed to have absolute 1 values that are less than or equal to ε 2 , an assumption that we make to assure

24

Stochastic Linear Difference Equations

that the matrix A to be defined below has eigenvalues with modulus bounded above by unity. We transform variables as follows: t

xt = (ε)− 2 x+ t

(2.5.29)

t

wt = (ε)− 2 wt+ .

(2.5.30)

The transformed process {wt : t = 1, 2, . . .} is now conditionally homoskedastic as required because E[wt+1 (wt+1 )′ | Jt ] = I . Furthermore, the transformed 1 process {xt : t = 1, 2, . . .} satisfies (2.2.1) with A = ε− 2 A+ . The matrix A now satisfies the restriction that its eigenvalues are bounded in modulus by unity. The original process {x+ t : t = 1, 2, . . .} is allowed to grow over time at a rate of up to .5 log (ε).

2.5.11. A rational expectations model Consider a model in which a variable pt is related to a variable mt via pt = λEt pt+1 + γmt ,

0 0

gt2

kt = δk kt−1 + it

(3.4.1)

, 0 < δk < 1

where d1t is a random endowment of the consumption good at time t , and d2t is a random disturbance to adjustment costs at time t . Given d2t , investment can be increased or decreased only by adjusting the amount of the intermediate good employed. The larger is the parameter φ1 , the higher are adjustment costs. Employment of the intermediate good requires labor input on a one-forone-basis. Physical capital depreciates over time. To capture this technology, we specify Φc =

1 0 0 , Φg = , Φi = , 0 −1 φ1

γ Γ= , ∆k = δk , Θk = 1. 0 We set A22 , C2 and Ud to make (d1t , d2t )′ = dt follow one of the stochastic processes described in chapter 2. This technology embodies a linear quadratic, general equilibrium version of the adjustment-cost technology used in Lucas and Prescott’s [1971] model of investment under uncertainty. Technology 3: Multi-Period Adjustment Costs and “Time to Build” A single consumption good is produced by a single capital good. The capital good can be produced in two ways: a fast and relatively resource-intensive way, and a slow and less resource intensive way. Different amounts of intermediate goods are absorbed in producing investment goods in the fast and the slow ways.

44

The Economic Environment

We model this by positing that there are two capital stocks, two investment goods, and four intermediate goods, and that adjustment costs are larger for the faster investment technology. This technology is represented as ct = γk1t−1 + d1t ,

γ>0

k1t = δk k1t−1 + k2t−1 + i1t

(3.4.2a) ,

0 < δk < 1

k2t = i2t

(3.4.2b) (3.4.2c)

g1t = φ1 (i1t + i2t )

, φ1 > 0

g2t = φ2 (i1t + k2t−1 )

, φ2 > 0

(3.4.2d) (3.4.2e)

g3t = φ3 i1t

, φ3 > 0

(3.4.2f )

g4t = φ4 i2t

, φ4 > 0

(3.4.2g)

ℓ2t = gt · gt

(3.4.2h)

Equation (3.4.2a) describes how physical capital, k1t , and an endowment shock, d1t , are transformed into the consumption good. Equations (3.4.2b ) and (3.4.2c) tell how capital, k1t , can be augmented by “quick investment”, i1t , and by “slow investment”, i2t . Notice that (3.4.2b ) and (3.4.2c) imply that physical capital, k1t , is determined by k1t = δk k1t−1 + i1t + i2t−1 , an equation that exhibits the status of i1t and i2t as ‘fast’ and ‘slow’ investment processes, respectively. Equations (3.4.2d ) and (3.4.2e) describe how the intermediate goods, g1t and g2t , are required to produce investment goods. According to (3.4.2d ) and (3.4.2e), it is as though two stages of production are required to produce capital, the first stage using intermediate good g1t , and the second stage using intermediate good g2t . According to (3.4.2d ) and (3.4.2e), fast investment i1t undergoes both stages of production in the same period t , while slow investment i2t undergoes the first stage described by (3.4.2d ) in period t and the second stage described by (3.4.2e) in period (t + 1). Equations (3.4.2f ) and (3.4.2g ) describe some additional inputs of intermediate goods that are specific to the two types of investment processes. We can set φ3 > φ4 to capture the notion that it is more resource intensive to invest quickly. In equation (3.4.2h ), ‘·’ denotes an inner product.

Examples of Technologies

45

To map this technology into our setup, we set ∆k =

δk 0

1 0

, Θk = I

0 0 0 0 0 φ1 0 0 , Φi = φ2 φ3 −1 0 0 0 −1 γ 0 0 0 Γ = 0 −φ2 0 0 0 0

0 0 1 −1 0 0 Φc = 0 , Φg = 0 −1 0 0 0 0 0 0

0 φ1 0 0

(3.4.3)

φ4

Recall that the matrices Φc , Φg , Φi multiply the vectors ct , [g1t g2t g3t g4t ]′ , and [i1t i2t ]′ , respectively, while Γ multiplies the vector [k1t−1 , k2t−1 ]′ . Again, we set Ud , A22 , C2 to make d1t obey one of the processes described in Chapter 2. This technology captures aspects of those used by Park (1984) and Kydland and Prescott (1982). Technology 4: Growth There are a single consumption good, a single investment good, a single capital good, and no intermediate good. Output obeys ct + it = γkt−1 + dt where dt is a random endowment of output at time t . The motion of capital obeys kt = δk kt−1 + it . To represent this technology, we could set Φc = 1, Φi = 1, Φg = 0, Γ = γ, ∆k = δk , Θk = 1. The reader can verify that this specification of the technology violates assumption 3 ( [Φc Φg ] is singular). To analyze such an economy, we could modify some of our calculations to dispense with assumption 3. An alternative way is

46

The Economic Environment

to approximate the technology with another one that satisfies assumption 3. In particular, assume that ct + it = γkt−1 + d1t gt = φ1 it kt = δk kt−1 + it where φ1 is a very small positive number and d2t ≡ 0. To implement this technology, set γ 1 0 1 , Γ= , Φi = , , Φg = Φc = −φ1 1 0 0

∆k = δk , Θk = 1. 2 This technology can be used to create a model of consumption along the lines of Hall (1978) and Flavin (1981), and a linear quadratic version of a model of capital accumulation along the lines of Cass (1965), Koopmans (1965), and Brock and Mirman (1972) . We shall also use later it to represent aspects of a model of economic growth authored by Jones and Manuelli (1988). Technology 5: Depletable Resource There is a single consumption good, a single investment good, two intermediate goods and one capital stock. The capital stock is the cumulative stock of the resource that has been extracted. We let investment it be the extraction rate, so that kt = kt−1 + it . (3.4.4a) All of the amount extracted is consumed, so that ct = it .

(3.4.4b)

There are two sources of extraction costs. The first, which is coincident with using the first intermediate good g1t , depends on the amount extracted in the current time period g1t = φ1 it . (3.4.4c) 2 In effect, the modification induces investment to be associated with the use of a small (because φ1 ≈ 0 ) amount of intermediate goods, which require labor input. The matrix [Φc Φg ] is now nonsingular, so that assumption 3 is satisfied. When φ1 > 0 , technical conditions are satisfied that are required for the solution of the social planning problem automatically to lie in the space L20 (see Chapters 4 and 5). When φ1 is close to zero, the solution of the social planning problem will closely approximate the solution of the social planning problem for φ1 = 0 , augmented with the restriction that the solution lie in L20 .

Examples of Technologies

47

The second source of extraction costs, captured by the intermediate good g2t , depends on the cumulative amount extracted at period t , which we approximate as (it /2 + kt−1 ): 3 g2t = φ2 (it /2 + kt−1 ).

(3.4.4d)

To represent this technology, we set 0 −1 0 0 1 Φc = 0 , Φg = −1 0 , Φi = φ1 Γ = 0 , −φ2 φ2 /2 0 −1 0

∆k = 1, Θk = 1.

In this technology, we have included no endowment shock process dt , so that we can take Ud = 0, A22 = 0, C2 = 0. It would be possible to modify the technology in various ways to provide a role for an endowment or technology shock. Such a technology was used by Hansen, Epple and Roberds [1985] to study alternative arrangements for an exhaustible resource market. Technology 6: Learning by Doing There is a single consumption good, a single investment good, a single intermediate good, and a single capital stock. The capital stock is interpreted as the cumulative stock of knowledge, the accumulation of which requires expenditure of current output and the intermediate good. Thus, we set ct + it = γ1 kt−1 + dt kt = δk kt−1 + (1 − δk )it

(3.4.5)

Setting Θk = (1 − δk ) makes kt a weighted average of current and past rates of investment. Possession of knowledge (capital) lowers the amount of intermediate goods required to accumulate more knowledge: gt = φit − γ2 kt−1 , where φ ≥ γ2 > 0. 3 We add half the current extraction rate i to k t t−1 to approximate the average amount over the period that has been extracted cumulatively.

48

The Economic Environment

To represent this economy, we set 1 0 1 Φc = , Φg = , Φi = 0 −1 φ γ1 Γ= , ∆k = δk , Θk = (1 − δk ). γ2

(3.4.6)

Technology 7: Fixed Proportions There is a single consumption good, a single capital good, and a single “intermediate good” to be interpreted as labor. Labor and capital are required in fixed proportions, apart from the effects of a random “labor-requirements” shock d2t . The technology requires ct + it = γ1 kt−1 + d1t gt = γ2 kt−1 + d2t gt2 = ℓ2t kt = δk kt−1 + it . Here gt represents employment of labor input. The parameter γ2 determines the nonstochastic part of the capital-labor ratio. To map this technology into our setup, we set 1 1 0 Φc = , Φi = , Φg = , 0 0 1 γ1 Γ= , ∆k = δk , Θk = 1. γ2 Technology 8: Interrelated Factor Demand with Costs of Adjustment To produce output requires physical capital, k1t , and labor, k2t . It is costly to adjust the stock of either factor of production. To adjust capital, the intermediate good g1t must be employed, while to adjust labor, the intermediate good g2t must be employed. To implement this technology, we require k2t = g3t ,

Examples of Technologies

49

which identifies k2t with the direct input of labor. The technology satisfies

k1t−1 c1t + it = [γ1 γ2 ] + d1t k2t−1 k1t = δk k1t−1 + i1t

k2t = k2t−1 + i2t g1t = φ2 i1t g2t = φ3 i2t g3t = k2t . When φ3 < φ2 , it is more costly to adjust capital than labor. To capture this technology, we set δk 0 1 0 ∆k = , Θk = 0 1 0 1 1 1 0 0 0 0 γ1 γ2 0 0 1 0 0 1 0 0 Φc = 0 , Φi = φ2 0 , Φg = 0 −1 0 , Γ = 0 0 0 0 φ3 0 0 −1 0 0 This technology is a version of one used by Mortensen [1973] and Hansen and Sargent [1981].

3.4.1. Other technologies Alternative technologies can be constructed that blend features of two or more of those described here. For instance, multiple-period adjustment costs can be incorporated into the growth technology, while learning by doing can be introduced into one of the adjustment cost technologies. Also versions of these single consumption good technologies can be combined to yield technologies for the production of multiple consumption goods.

50

The Economic Environment

3.5. Preferences and Household Technologies We assume a representative household. We postpone until Chapter 12 discussing ways that heterogeneity among consumers can be accommodated within this assumption. We describe preferences in terms of two elements. First we describe a household technology for accumulating a vector of household capital and for using it to produce a vector of consumption services. Then we specify intertemporal preferences for consumption services in different dates and states of the world. We assume that there is an nh –dimensional vector of household capital stocks ht−1 brought into time t . The vector h−1 is taken as an initial condition. The vectors of consumption goods ct and household capital stocks ht−1 are inputs into the household technology at time t . The outputs of this technology are an ns –dimensional vector of household services st and a new vector of stocks of household capital ht . The relation between inputs and outputs is described by ht = ∆h ht−1 + Θh ct (3.5.1) and st = Λht−1 + Πct .

(3.5.2)

We maintain the following technical assumption: 4 Assumption 5: The absolute values of the eigenvalues of ∆h are less than or equal to one. Preferences are defined over stochastic processes for household services and household inputs into production. These preferences are separable across components of services, across states of the world, and over time. In particular, preferences are described by the quadratic utility functional: ∞ 1 X t −( )E β (st − bt ) · (st − bt ) + (ℓt )2 | J0 2 t=0

, 0 < β < 1.

(3.5.3)

where β is a subjective discount factor. The household services in this economy play the role of characteristics or attributes in the analyses of Gorman (1980) and Lancaster (1966). We can think 4 The purpose of this assumption is to assure that under the equilibrium (optimal) decision rule, the state vector for the economy has a transition matrix that is ‘stable’.

Examples of Household Technology Preference Structures

51

of consumption ct at date t as generating a bundle of consumption services in current and future time periods. Thus, the consumption vector ct generates a vector Πct of consumption services at time t and a vector Λ(∆h )j−1 Θh ct of consumption services at time t+j , for j ≥ 1. In effect, the household technology puts time and component nonseparabilities into the indirect preference ordering for consumption goods induced by (3.5.3). We do not impose nonnegativity constraints on consumption goods.

3.6. Examples of Household Technology Preference Structures We describe five examples of household technology-preference structures. Household Technology 1: Time Separability There is a single consumption good which is identical with the single service. There is no household capital. Preferences are described by ∞ 1 X t − E β (ct − bt )2 + ℓ2t | J0 2 t=0

, 0 0.

P∞ Here the bliss point is in effect bt + λ(1 − δh ) j=0 δhj ct−j−1 , so that the bliss point shifts in response to a moving average of past consumption. Preferences P∞ j in this form require an initial condition for the geometric sum j=0 δh ct−j−1 , which we specify as an initial condition for the ‘stock of household durables,’ h−1 . To implement these preferences, let the household capital stock be ht = δh ht−1 + (1 − δh )ct

, 0 < δh < 1.

This implies that ht = (1 − δh )

t X j=0

δhj ct−j + δht+1 h−1

Let consumption services be st = −λht−1 + ct

, λ > 0.

We can represent the desired preferences by setting Λ = −λ, Π = 1, ∆h = δh , Θh = 1 − δh .

Examples of Household Technology Preference Structures

53

The parameter λ governs the strength of habit persistence. When λ = 0, we recover a version of household technology 1. Household technology-preferences 3 is a version of the model of habit persistence of Ryder and Heal [1973]. Later we shall use this specification to represent aspects of some ideas of Jones and Manuelli [1988]. Household Technology 4: Adjustment Costs There is a single consumption good, a single household capital stock equal to consumption, and two consumption services. We want to represent preferences of the form ∞

1 X t − E β [(ct −b1t )2 + λ2 (ct − ct−1 )2 + ℓ2t ] | J0 2 t=0

(3.6.4)

0 0

4.6.3. Information zt+1

1 0 = 0 .8 0 0

0 0 0 0 zt + 1 0 wt+1 .5 0 1

Ub = [ 30 0 0 ] 5 1 0 Ud = 0 0 0 x0 = [ 5

150

1 0

′

0]

Notice that the information process and the initial condition are specified so that the constant is the third state variable. Notice that we have set the bt process equal to a constant value of 30. There is no random component of the preference shock process. Notice that there is a single nontrivial endowment shock, the second component of dt having been set to zero via the specification of the matrix Ud . The first component of dt has been specified to follow a first order autoregressive process with positive mean. The autoregressive parameter for the endowment process has been set at .8. Notice that the third component of

Solutions for five economies

77

the zt vector is a first order autoregressive process with coefficient .5. However, this component of the zt vector impinges neither on bt nor on dt , given the way that we have specified Ub and Ud . We include the third component of the zt process in case the reader would like to edit one our files, say, to add a random component to the preference shock bt . These specifications of preferences and technology are rich enough to encompass versions of several models that have been popular in the recent macroeconomic literature. The preference specification can accommodate preferences that are quadratic in consumption, as used by Hall [1978]; preferences incorporating habit persistence, as used recently by Becker and Murphy [1988]; and preferences for a durable consumption good, as used by Mankiw [1982]. The technology specification is a version of the one-good ‘growth’ technology of chapter 2, modified to include costs of adjusting capital. 9 We shall initially set the parameters of the technology to satisfy the necessary condition for consumption to be a random walk in Hall’s model, namely, the condition β(γ1 + δk ) = 1. This is also the condition for the ‘growth condition’ of Jones and Manuelli just to be satisfied. For all of the specifications, we set Ub so that bt = 30 for all t . By setting the parameter values of this general model to particular values, we can capture the following models.

4.6.4. Brock-Mirman model Set the preference parameters as λ = 0, π = 1, δh and θh arbitrarily. This makes preferences take the form −.5E

∞ X t=0

β t [(ct − bt )2 + ℓ2t ]|J0 .

Set the technology parameters so that γ1 > 0, φ1 > 0 but φ1 ≈ 0, (γ1 +δk )β = 1.

9 The parameters for our first version of Hall’s economy are in clex11.m; those for our second version of Hall’s economy are in clex12.m; those for our third version of Hall’s economy are in clex13.m; those for the Jones-Manuelli model are in clex10.m; those for the model with durable consumption goods are in clex15.m; and those for Lucas’s economy are in clex14.m.

78

Optimal Resource Allocation

4.6.5. A growth economy fueled by habit persistence Set the technology parameters as in Hall’s model, but set the preference parameters to capture preference specification 3 of chapter 2. In particular, set 1 > δh > 0, θh = (1 − δh ), π = 1, λ = −1. This makes preferences assume the form −.5E

∞ X t=0

β t [(ct − bt − λ(1 − δh )

∞ X

δhj ct−j−1 )2 + ℓ2t ]|J0 .

j=0

4.6.6. Lucas’s pure exchange economy Set preference parameters as in Hall’s model, but alter the technology to render capital unproductive, i.e., set γ1 = 0.

4.6.7. An economy with a durable consumption good Set the technology as in Hall’s model, but alter preferences to capture the idea that the consumption good is durable. Set π = 0, λ > 0, 0 < δh < 1, θh = 1. We now illustrate how the solutions of the social planning problem associated with several of these models can be computed and analyzed. Generally, we proceed as follows. First we read in the parameters that represent our economy by way of the matrices listed in Table 1. We have prepared a set of ‘.m’ files that read in these matrices for the several economies listed above. Thus, clex11.m, clex12.m, and clex13.m are files that read in matrices corresponding to Hall’s model for various different parameter settings. Next, we use solvea.m to compute all of the matrices listed in Table 2, which characterize the solution of the planning problem. To compute the vector ARMA representation of any subset of quantities or Lagrange multipliers, we use aarma.m. To compute the impulse response functions of any set of quantities and/or Lagrange multipliers to components of w(t), we use the program aimpulse.m. Finally, we can use simul.m or asimul.m to simulate the solution of the model.

Hall’s model

79

4.7. Hall’s model We begin with the version of Hall’s model which we solved by hand earlier in this chapter. We begin by setting the parameters in a way that is designed to make consumption follow a random walk. In particular, we set φ1 = .00001, γ1 = .1, δk = .95, β = 1/1.05. Notice that β(γ1 + δk ) = 1. We set the remaining parameters to the values described above. After reading in the matrices by typing clex11, we compute the solution of the planning problem by typing solvea. Issuing this command causes the computer to respond as follows: Calculating, please wait The matrix ao has been calculated for the law of motion of the entire state vector. This matrix satisfies x(t+1) = ao*x(t) + c*w(t+1). The endogenous eigenvalues are in the vector endo, and the exogenous eigenvalues are in the vector exog. The solution to the model is given by c(t) = sc*x(t), g(t) = sg*x(t), h(t) = sh*x(t), i(t) = si*x(t) k(t) = sk*x(t), and s(t) = ss*x(t). The matrices sc, sg, sh, si, sk, and ss have now been computed and can be used in other matlab programs. The matrices sb and sd are constructed so that b(t) = sb*x(t) and d(t) = sd*x(t) and can be used in other matlab programs. The shadow price vectors satisfy Mc(t) = mc*x(t), Mg(t) = mg*x(t), Mh(t) = mh*x(t), Mi(t) = mi*x(t), Mk(t) = mk*x(t), Ms(t) = ms*x(t), and Md(t) = md*x(t). The matrices of these linear combinations can be used in other matlab programs. Your equilibrium has been calculated. You are now ready to experiment with the economy. This is the end of the output that appears on the screen. The solution of the planning problem is stored in the matrices listed in table 2. To inspect these matrices, we just ask MATLAB to show them to us. Thus, issuing the

80

Optimal Resource Allocation

MATLAB command ao results in the output

0.9000 0.0000 ao = 0.0000 0.0000 0.0000

0.0050 1.0000 0.0000 0.0000 0.0000

0.5000 0.0000 1.0000 0.0000 0.0000

0.0000 0.0000 0.0000 0.0000 0.5000

0.0200 0.8000 0.0000 0.8000 0.0000

To see the matrix c, we type c and elicit the response

0.0000 0.0000 c = 0.0000 1.0000 0.0000

0.0000 0.0000 0.0000 0.0000 1.0000

Recall that various quantities in the model are determined by premultiplying the state xt by matrices Sj which are stored by MATLAB in sj. For various purposes, it is useful to create a matrix by stacking various sj’s on top of one another. For example, we can stack the s matrices for consumption, household durables, services, physical investment, and physical capital by issuing the MATLAB command G=[sc;sh;ss;si;sk], which evokes the response

0.0000 0.9000 G = 0.0000 0.0000 0.0000

0.0500 0.0050 0.0500 0.0500 1.0000

5.0000 0.5000 5.0000 0.0000 0.0000

0.2000 0.0200 0.2000 0.8000 0.8000

0.0000 0.0000 0.0000 0.0000 0.0000

The first row of G is Sc , and so on. Similarly, various Lagrange multipliers in the model are determined by premultiplying xt by the matrices Mj , which are stored by MATLAB in mj. We can create a matrix by stacking various mj’s by issuing the command H=[mc;ms;mh;mi;mk], which evokes

0.0000 0.0000 H = 0.0000 0.0000 0.0000

−0.0500 −0.0500 0.0000 −0.0500 −0.0500

25.0000 25.0000 0.0000 25.0000 25.0000

−0.2000 −0.2000 0.0000 −0.2000 −0.2000

0.0000 0.0000 0.0000 0.0000 0.0000

Hall’s model

81

The endogenous and exogenous eigenvalues of Ao or ao are stored in endo and exo, respectively. For the present model, they are given by .90 endo = 1.0 1.00 exo = .80 .50

The exogenous eigenvalue of unity corresponds to the constant (unity) in the state vector, while the other two exogenous eigenvalues are also directly inherited from our specification of the A22 matrix. The endogenous eigenvalue of .9 is inherited from the depreciation factor of .9 which we set for consumer durables, which is irrelevant in Hall’s model because we set λ = 0. This eigenvalue will become relevant below in specifications in which λ 6= 0. The eigenvalue of unity reflects the random walk character of consumption in Hall’s model. Actually, the second endogenous eigenvalue is not really unity, it is only close to unity. To see this, we switch to a long format in MATLAB by typing format long and then we type endo to receive the response 0.90000000000000 endo = 0.99999999999048 The eigenvalue is not exactly unity because of the very small costs of adjusting capital that we have imposed. The fact that the endogenous eigenvalues of this model are below unity means that it possesses a nonstochastic steady state. To compute the steady state, we set nnc=3, which tells the computer that the constant term is the third component of the state vector. Then we type steadst, which causes the steady state to be computed and stored in zs. To compute the steady state value of consumption, just type sc*zs, and so on. For the present model, we obtain 5.0003 0.0061 zs = 1.0000 0.0000 0.0000

The steady state value of consumption is given by sc*zs, which is sc ∗ zs = [ 5.0003 ]

82

Optimal Resource Allocation

The steady state value of investment is given by si*zs, which is si ∗ zs = [ 0.0003 ] For the present model, these stationary steady state values are of little practical value because of the near unit endogenous eigenvalue. It will take very many periods for the effect of the initial conditions to die out in this model, despite the fact that a steady state for the nonstochastic version of the model does exist. We can compute an ARMA representation for the impulse response of any quantities or Lagrange multipliers to a given component of the white noise process wt . We can learn how aarma.m works by typing help aarma, which delivers the response function[num,den]=aarma(ao,c,sy,ii) Creates ARMA Representation for linear recursive equilibrium models. The equilibrium is x(t+1) = ao*x(t) + c*w(t+1) and is created by running SOLVEA. A vector of observables is given by y(t) = sy*x(t) where sy picks off the desired variables. For example, if we want y=[c’,i’], we set sy=[sc;si]. AARMA creates the representation den(L)y(t) = num(L)wi(t) This is an arma representation for the response of y(t) to the i-th component of w(t). For example, to compute the ARMA representation for the impulse response of ct , it to the first component of wt , we type sy=[sc;si] and [num,den]=aarma(ao, c,sy,1) which gives the response num =

0.0000 0.0000

den = [ 1.0000

0.2000 0.8000

−0.6400 −2.6800

0.7540 3.3040

−0.3860 −1.7660

0.0720 0.3420

−4.2000

6.9700

−5.7000

2.2900

−0.3600 ]

Hall’s model

83

This output is to be interpreted as follows. For i = 0, . . . , 5. Define αi as the element in the (i + 1) column of den. For i = 0, . . . 5 define ξi as the 3 × 1 matrix that is the i + 2st column of num. Define two polynomials in the lag operator L by P5 α(L) = i=0 αi Li P5 ξ(L) = i=0 ξi Li Let w1t be the first innovation in the system, which drives the endowment process. Then we have the representation ct α(L) it = ξ(L)w1t mct For example, the first row of this representation is

(1 − 4.2L + 6.97L2 − 5.7L3 + 2.29L4 − .36L5 )ct

= (.2 − .64L + .754L2 − .386L3 + .0072L4 )w1t

We can also create the impulse response function for a list of variables in response to a particular innovation. We shall compute the impulse response function for the two variables, c, i. To accomplish this, we set sy by typing sy = [sc;si]. We set ii at 1 (we want the response to the first innovation), and specify the number of lags we want to perform the calculation for. We want the impulse response out to forty lags, so we specify ni=40. To compute the impulse response, we issue the MATLAB function aimpulse, which has the syntax [z]=aimpulse(ao,c,sy,ii,ni), where sy,ii,ni have the settings just described. 10 The impulse response function is returned in z. In Fig. 4.7.1.a we plot the impulse response functions for this model in response to the first innovation, which is the innovation in the endowment shock. These impulse response functions have shapes that are characteristic of a random walk for consumption and a unit root in capital. For consumption, the impulse response is an open “box” which attains its maximum height immediately. This impulse response is characteristic of a random walk consumption process. For investment, the impulse response has an asymptote. 11 10 The MATLAB program aimpulse.m takes the inputs we have created from the solution of the social planning problem and feeds them into the MATLAB program dimpulse.m, which computes impulse response functions. 11 In actuality, there is really no asymptote for the impulse response function for either consumption or investment, because the largest eigenvalue is just a little bit less than unity.

84

Optimal Resource Allocation

16

1 0.9

14 0.8 * consumption 12

0.7

*

0.6

*

0.5

10 *

investment *

0.4

8 * * *

0.3 0.2 *

*

6

* * * * * * * * * * * * * * * * * * * * * * * * ** ** ** ** * * * * * * * * * * * * * * * * * *

consumption

investment

4

0.1 0

0

5

10

15

20

25

30

35

40

Fig. 4.7.1.a. Impulse response of consumption and investment to an endowment innovation in a version of Hall’s model.

2

0

20

40

60

80

100

120

140

160

Fig. 4.7.1.b. Simulation of consumption and investment for Hall’s model.

We now generate a random simulation of the model for 150 periods. We use the non-interactive program asimul.m to generate this simulation. To use this program we must specify an observer matrix sy that links the called-for variables to the state. Since we want to simulate the four series c, i, k , and the shadow price of consumption, we set sy=[sc;si;sk;mc]. We also have to specify the length of the simulation t1, whether we want a random (k=1) or nonrandom (k=2) simulation, and the initial state vector x0. We want a random simulation of length 150 with the initial condition specified above. After setting these parameters, we execute the simulation by commanding asimul. We obtain the response: Your simulated vector is in the vector ‘‘y’’. We display aspects of this simulation in Fig. 4.7.1.b.The sample paths of c, k , and the shadow price drift in the fashion that random walks do. For paths that are long enough, a random simulation of this model will eventually In fact, the impulse response functions for both consumption and investment are ‘square summable’, but it would take a very long realization of them for this behavior to become apparent.

Higher Adjustment Costs

85

encounter negative values for capital and consumption. The key to rigging samples so that capital and consumption for a long time remain positive with high probability is to select the initial condition for capital large enough and the elements of c2 small enough. Figure 4.7.1.b indicates that investment is relatively more variable than consumption, a pattern that is found in aggregate data for a variety of countries. The fact that this version of Hall’s model, like the stochastic growth model of Brock and Mirman [1972], so easily delivers this pattern is an important feature that has attracted adherents to this and other versions of ‘real business cycle’ theories.

4.8. Higher Adjustment Costs We now turn to a second model which is created by making one modification to the economy we have just studied. The one change we make is to raise the costs associated with adjusting capital. We raise the absolute value of the cost parameter to φ1 = .2. All other parameters remain as in the previous economy. We computed the solution of the social planning problem using solvea.m. The endogenous eigenvalues were computed to be:

endo =

0.9000 0.9966

Notice that relative to the previous economy, one endogenous eigenvalue is left unaltered at .9, while the other endogenous eigenvalue has fallen below unity. The endogenous eigenvalue of .9 is inherited from the law of accumulation that we posit for household capital (which in this model is again irrelevant). The drop below unity of the second endogenous eigenvalue is the result of our having increased the costs of adjusting capital. The analysis that we performed on pages BLANK indicates that this is exactly what should occur when adjustment costs increase. Figure 4.8.1.a reports impulse response functions for the response of ct and it and to an innovation in the endowment process. Notice how these no longer have the tell tale signs of the presence of an endogenous unit eigenvalue. The

86

Optimal Resource Allocation

1

14

0.9

12 consumption

0.8 *

10

0.7 *

8

0.6 *

0.5

6

*

investment *

0.4

4

* *

0.3 0.2

*

investment

*

2

* * * * * * * * * * * * * * * ** * * * * * * * * * * * * * * * * * * ** ** ** ** ** ** ** ** * * * * * * * * * * * * * * * * consumption

0

0.1 0

0

5

10

15

20

25

30

35

40

Fig. 4.8.1.a. Impulse response of consumption and investment to an endowment innovation in a version of Hall’s model with higher costs of adjusting capital and no random walk in consumption.

-2 0

20

40

60

80

100

120

140

160

Fig. 4.8.1.b. Simulation of a version of Hall’s model with higher costs of adjusting capital and no random walk in consumption.

impulse response for consumption and investment now both appear to be convergent and ‘square summable’. Figure 4.8.1.b shows a random simulation beginning from the same value for x0 used with the earlier version of Hall’s model. Notice how consumption, while still smoother than income, has increased high frequency volatility relative to that depicted in figure 4.7.1.a, while the high frequency volatility of investment has decreased. This pattern is a response to the higher costs for adjusting capital. Notice also that there seems to be a downward ‘trend’ in both consumption and investment. This is a consequence of the decrease in the largest endogenous eigenvalue from being very nearly one in the earlier economy. The present economy has a nonstochastic steady state value for capital of .0000, for consumption of 5.00 (which is the mean of the endowment process), and for investment of .0000, each of which we computed using steadst.m. These nonstochastic steady state values correspond to the unconditional means from the asymptotic stationary distribution of our variables. Because the largest endogenous eigenvalue for this economy is .9966 rather than

Altered ‘growth condition’

87

.9999, the economy is headed toward these mean values much more rapidly than for our previous economy.

4.9. Altered ‘growth condition’ We generate our next economy by making two alterations in the preceding economy. First, we raise the adjustment cost parameter from .2 to 1. This will have the effect of further lowering the endogenous eigenvalue that is not .9, and of causing the impulse response functions to dampen faster than they did in the previous economy. Second, we raise the production function parameter from .1 to .15. This will have the effect of raising the optimal stationary value of capital to a positive value for the nonstochastic version of the model. Recall that the optimal stationary value of capital was zero in the previous economy. The nonstochastic steady state values of consumption, investment, and capital are 17.5, 6.25, and 125, respectively, for this economy. The endogenous eigenvalues are

0.9000 endo = 0.9524

We also created the impulse response function for c and i, which is reported in figure 4.9.1.a. Notice the much faster rate of damping relative to the impulse responses displayed for the previous economies. Figure 4.9.1.b displays a random simulation of this economy. Notice that the “transient” behavior displayed by our simulation of the previous economy is not present here. This is a consequence of our having altered the production function parameter value to induce a positive optimal stationary value for the capital stock of 125, and from our having started the simulation at an initial condition of 125 for the capital stock.

88

Optimal Resource Allocation

1

22

0.9

20

0.8 0.7

*

16 *

0.6

14

*

0.5

*

0.4

12 *

consumption

*

0.3

10

* *

*

* *

0.2

* * *

0.1 0

consumption

18

0

*

* * * * * * * * * * * * * * * * * * * investment * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *

5

10

15

20

25

30

35

8

investment

6 4

40

Fig. 4.9.1.a. Impulse response of consumption and investment to an endowment innovation in a version of Hall’s model with higher adjustment costs and the ‘growth condition’ altered.

0

20

40

60

80

100

120

140

160

Fig. 4.9.1.a. Simulation of consumption and investment in a version of Hall’s model with higher adjustment costs and the ‘growth condition’ altered.

4.10. A Jones-Manuelli economy A notable feature of the models for the previous simulations is that consumption, investment, and capital generally failed to grow. We now define the matrices and set parameters with a view toward attaining a version of Jones and Manuelli’s model of economic growth. We set the parameters of the technology so that Jones and Manuelli’s “growth condition” is just satisfied. 12 Our version of Jones and Manuelli’s model has the feature that their growth condition is a necessary but not a sufficient condition for growth to occur. Their growth condition makes sustained growth feasible in our model. In order for growth to occur, it is also necessary that it be desirable, a condition that is determined by the preference parameters λ , δh , and θh . We set these parameters in order to generate growth. 12 The Jones-Manuelli growth condition on the technology in our notation is β(γ + δ ) ≡ 1 . k This is also a condition that makes the marginal utility of consumption follow a martingale in Hall’s model.

A Jones-Manuelli economy

89

In particular, setting λ equal to minus one turns out to generate a preference for growth. 13 As usual, we compute the equilibrium by using asolve.m. For this model, the endogenous eigenvalues are

1.0000 + 0.0000i endo = 1.0000 − 0.0000i

The exogenous eigenvalue of unity is inherited from the law of motion of the unit vector, which is the third state variable. Notice that there are two unit endogenous eigenvalues. With some experimentation, the reader can determine how these two unit endogenous eigenvalues result from specifying the parameters of technology to obey the growth condition, and the parameters of preferences (especially λ ) to capture a longing for consumption growth. 14 Figure 4.10.1.a displays impulse responses of consumption and investment to an innovation in the endowment process. For both consumption and investment, the effect of an innovation actually grows indefinitely over time. This is a product of the second unit endogenous eigenvalue that is inherited from the preference parameter λ . Figure 4.10.1.b displays a simulation of consumption and investment for this economy. The economy grows. Notice that consumption is much smoother than investment. Notice also that investment typically exceeds consumption. In order to support the ‘habit’ that fuels growth, the economy has to accumulate physical capital. 15 We invite the reader to experiment with this economy by altering the settings of some parameter values one at a time relative to the parameter settings that we have made. In particular, we recommend that the following experiments be tried: 13 The parameter values for this economy are stored in clex10.m 14 One unit endogenous eigenvalue stems from setting β, Γ , and ∆ at the boundary of the k Jones-Manuelli growth condition. The other unit endogenous eigenvalue results from setting λ = −1 . The presence of very small positive adjustment costs for capital is what prevents these two endogenous eigenvalues from being exactly unity. The reader can check that they are not exactly unity by using the format long command in MATLAB. 15 It is a feature of models of addiction based on the type of preference specification used here, e.g., Becker and Murphy [1988], that ‘addicts’ grow wealthier and wealthier over time as they follow a consumption plan that allows for enough accumulation to support their growing addiction.

90

Optimal Resource Allocation

70

1 0.9

*

60 *

0.8

*

0.7

50 *

0.6

*

0.5

investment * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *

40 investment 30

0.4 0.3 0.2 0.1 0

* * * * * * *

0

5

consumption

* * * * * * * * * * * * * * * * * * consumption * * * * * * * * * * * * * *

10

15

20

25

30

35

20

10

40

Fig. 4.10.1.a. Impulse response of consumption and investment to an endowment innovation in a Jones-Manuelli economy.

0

0

20

40

60

80

100

120

140

160

Fig. 4.10.1.b. Simulation of consumption and investment in a Jones-Manuelli economy.

1. Change the value of λ to −.7, leaving the other parameters unaltered. Obtain the solution of the planning problem, and inspect the endogenous eigenvalues. Also compute the impulse response function and simulate the model in response to the same initial condition that we used above. Does the economy still grow? Explain. 2. Change the value of β to .94. Recompute the solution of the planning problem. Does the economy grow? Link your explanation to the JonesManuelli growth condition. 3. Change the value of Γ(1) to .09. Does the economy still grow?

Durable consumption goods

91

4.11. Durable consumption goods For our next example economy, we restore the productivity of capital to a value of .1 and raise the level of the parameter measuring adjustment costs for capital to a value of 1. We change the specification of preferences to make the consumption good durable. In particular, we adopt a version of preference specification 2. We implement this by setting λ equal to .1, π equal to zero, and θh equal to one. We leave δh at the value .9. 16

1.2

20 18

1*

consumption 16

0.8

*

14 *

0.6

12

*

consumption

0.4

*

10

* *

0.2

*

8

*

0

* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * ** ** ** ** ** ** ** ** ** ** ** ** ** * * investment

-0.2 0

5

10

15

20

25

30

35

investment

6

40

Fig. 4.11.1.a. Impulse response of consumption and investment to an endowment innovation in an economy with a durable consumption good.

4

0

20

40

60

80

100

120

140

160

Fig. 4.11.1.b. Simulation of consumption and investment in an economy with a durable consumption good.

Figure 4.11.1.a displays the impulse response functions to an innovation in the endowment process. The impulse response function for consumption and for investment are very different than for our first model. In particular, from the impulse response function, we can see that in choosing consumption, the social planner ‘smooths’ the endowment shock much less than he does in Hall’s original model, in which the planner in effect makes consumption an equal-weight moving average of current and lagged innovations to the endowment process. In the 16 These parameters settings are created by the file clex15.m.

Optimal Resource Allocation

92

present model, the planner makes consumption a much shorter, more peaked moving average of the endowment process. This shows up in the simulation of consumption and investment, which is reported in figure 4.11.1.b. Notice that now, in contrast to Hall’s model, it is investment that is much smoother than consumption. This example thus illustrates how making consumption goods durable tends to undo the strong consumption smoothing result which Hall obtained.

4.12. Summary In this chapter, we have formulated a planning problem, and described how to compute its solution. We have also described computer programs that solve the planning problem, and that characterize the solution in a variety of ways. Associated with the solution of the planning problem are a set of Lagrange multipliers, which we have shown how to compute in terms of the derivatives of the value function for the planners dynamic programming problem. In the next two chapters, we shall show how those Lagrange multipliers are related to the price system for a competitive equilibrium. We begin by describing how to represent values.

A. Synthesizing the linear regulator The social planning problem is to maximize

−.5E

∞ X t=0

β t (st − bt ) · (st − bt ) + gt · gt

(4.A.1)

subject to Φc ct + Φg gt + Φi it = Γkt−1 + dt

(4.A.2)

kt = ∆k kt−1 + Θk it

(4.A.3)

ht = ∆h ht−1 + Θh ct

(4.A.4)

st = Λht−1 + Πct

(4.A.5)

zt+1 = A22 zt + C2 wt+1

(4.A.6)

Synthesizing the linear regulator

93

bt = U b z t

(4.A.7)

dt = Ud zt

"

#

ht−1 We define the state of the system as xt = kt−1 and the control as ut = it . In zt defining the control to be it , we exploit the assumption that [Φc Φg ] is nonsingular. Solve ( 4.A.2 ) for (ct , gt ) :

h

ct gt

i

= [Φc Φg ]−1 {Γkt−1 + Ud zt − Φi it }.

Let Uc and Ug be selector matrices that pick off the first nc and the last ng rows, respectively, of the right side of the above expression, so that the expression can be written ct = Uc [Φc Φg ]−1 {Γkt−1 + Ud zt − Φi it }

(4.A.8)

gt = Ug [Φc Φg ]−1 {Γkt−1 + Ud zt − Φi it }. Substituting ( 4.A.8 ) into ( 4.A.4 ) and ( 4.A.5 ) gives ht = ∆h ht−1 + Θh Uc [Φc Φg ]−1 {Γkt−1 + Ud zt − Φi it }

(4.A.9)

st = Λht−1 + ΠUc [Φc Φg ]−1 {Γkt−1 + Ud zt − Φi it }

(4.A.10)

Combining ( 4.A.3 ), ( 4.A.9 ), and ( 4.A.6 ) gives the law of motion for the linear regulator ht kt zt+1

!

Θh Uc [Φc Φg ]−1 Γ ∆k 0

=

∆h 0 0

+

−Θh Uc [Φc Φg ]−1 Φi Θk 0

!

Θh Uc [Φc Φg ]−1 Ud 0 A22 it +

0 0 C2

!

!

ht−1 kt−1 zt

! (4.A.11)

wt+1

or xt+1 = Axt + But + Cwt+1

(4.A.12)

where the matrices A, B , and C in ( 4.A.12 ) equal the corresponding matrices in ( 4.A.11 ). Now use ( 4.A.10 ) to compute (st −bt ) = Λht−1 +ΠUc [Φc Φg ]−1 Γkt−1 +(ΠUc [Φc Φg ]−1 Ud − Ub )zt − ΠUc [Φc Φg ]−1 Φi it . Express this in matrix notation as

ht−1 . . . kt−1 . . . −1 −1 −1 (st − bt ) = [Λ . ΠUc [Φc Φg ] Γ . ΠUc [Φc Φg ] Ud − Ub . − ΠUc [Φc Φg ] Φi ] zt it (4.A.13) or h i .. xt (4.A.14) (st − bt ) = [Hs . Hc ] it

94

Optimal Resource Allocation

.. where the matrix [Hs . Hc ] in ( 4.A.14 ) equals the corresponding matrix in ( 4.A.13 ). Next, use ( 4.A.8 ) to express gt as

or

ht−1 . . . k . . . −1 −1 −1 ′ gt = [0 . Ug [Φc Φg ] Γ . Ug [Φc Φg ] Ud . − Ug [Φc Φg ] Φi ] t−1 zt it

h i . xt . gt = [Gs . Gc ] it

(4.A.15)

(4.A.16)

. where the matrix [Gs .. Gc ] in ( 4.A.16 ) equals the corresponding matrix in ( 4.A.15 ). Define the matrices R = .5(Hs′ Hs + G′s Gs ), Q = .5(Hc′ Hc + G′c Gc ), W = .5(Hc′ Hs + G′c Gs ).

(4.A.17)

Then the current period return function for the social planning problem is −(x′t Rxt + u′t Qut + 2u′t W xt ).

(4.A.18)

In view of ( 4.A.14 ), ( 4.A.16 ), ( 4.A.17 ) and ( 4.A.18 ), we can represent the objective function in the social planning problem as

−E

∞ X

β t (x′t Rxt + u′t Qut + 2u′t W xt ),

(4.A.19)

t=0

which is to be maximized over {ut }∞ t=0 subject to xt+1 = Axt + But + Cwt+1 ,

t ≥ 0,

(4.A.20)

x0 given. Thus, we have mapped the social planning problem into a discounted optimal linear regulator problem.

A Brock-Mirman model

95

B. A Brock-Mirman model We shall usually use the recursive numerical methods described above to compute a solution of a social planning problem. These computational methods are quick and easy to use. However, to deepen our understanding of the structure of the social planning problem and the role played by various technical assumptions, and also to heighten our appreciation of the ease and power of those recursive numerical methods, it is useful to solve one problem by hand. We solve a social planning problem for a model with one consumption good and one capital good. We include costs of adjusting the capital stock, but permit them to be zero as a special case. When these costs of adjustment are zero (i.e., when the parameter φ in the model is set to zero), the model becomes a linear - quadratic, equilibrium version of Hall’s consumption model. To recover Hall’s solution of the model when φ = 0 , it is necessary to impose a side condition in the form of a version of our restriction (2.24) that forces the capital stock sequence {kt } to belong to L20 . The example is a useful laboratory for illustrating the relationships among the presence of costs to control (φ > 0) , the transversality condition, and the side condition that the solution lie in L20 . After we work out the answer by hand, we can solve the problem by using the MATLAB program solvea.m . The social planning problem comes from combining versions of our preference specification number 1 and our technology specification number 4: choose a contingency plan for {ct , kt }∞ t=0 to maximize: −E0

∞ X t=0

β t [(ct − bt )2 + ℓ2t ] , 0 < β < 1

(4.B.1)

subject to ct + it = γkt−1 + d1t , γ > 0

(4.B.2)

φit = gt , φ ≥ 0

(4.B.3)

kt = δkt−1 + it , 0 < δ < 1

(4.B.4)

gt2 = ℓ2t

(4.B.5)

k−1 given

(4.B.6)

The stochastic processes bt and d1t are given by bt = Ub zt and d1t = Ud1 zt , where zt obeys a version of (1.1). We assume that {d1t } and {bt } each belong to L20 , and do not impose that {kt } belongs to L20 . We begin by forming the Lagrangian

J = −E0

∞ X t=0

1 β t { [(ct − bt )2 + ℓ2t ] − λ1t [γkt−1 + d1t − ct − it ] 2

− λ2t [gt − φit ]

(4.B.7)

1 − λ3t [δkt−1 + it − kt ] − λ4t [ (ℓ2t − gt2 )]} 2

Here {λ1t , λ2t , λ3t , λ4t }∞ t=0 is a 4-tuple of stochastic Lagrange multipliers. We obtain the first order necessary conditions for a saddle point with respect to {ct , it , kt , ℓt , gt , λ1t , λ2t ,

96

Optimal Resource Allocation

λ3t , λ4t }∞ t=0 , and display the transversality condition for capital. First order conditions with respect to ct , it , kt , lt , and gt are:

ct : −(ct − bt ) − λ1t = 0, t ≥ 0

(4.B.8)

it : −λ1t − φλ2t + λ3t = 0, t ≥ 0

(4.B.9)

kt : γβEt λ1t+1 + βδEt λ3t+1 − λ3t = 0, t ≥

(4.B.10)

ℓt : −ℓt + λ4t ℓt = 0, t ≥ 0

(4.B.11)

gt : λ2t − λ4t gt = 0, t ≥ 0

(4.B.12)

In addition, we have the transversality condition lim E0 β t kt λ3t = 0.

t→∞

(4.B.13)

Equation ( 4.B.10 ) can be solved forward to yield

λ3t = γβ

∞ X

(δβ)j−1 Et λ1t+j .

(4.B.14)

j=1

Our strategy is to substitute the above expressions for the multipliers into the first-order condition with respect to kt to obtain an ‘Euler equation, and to study under what conditions, if any, this equation implies that the marginal utility of consumption is a martingale. Solving the first order conditions for the multipliers, we obtain λ1t = bt − ct

(4.B.15)

λ2t = gt

(4.B.16)

λ3t = φgt + (bt − ct )

(4.B.17)

λ4t = 1

(4.B.18)

Substituting ( 4.B.17 ) into ( 4.B.10 ) gives the “Euler equation” γβEt (bt+1 − ct+1 ) + βδEt (φgt+1 + bt+1 − ct+1 ) = φgt + (bt − ct )

(4.B.19)

or βδEt φgt+1 + β(γ + δ)Et (bt+1 − ct+1 ) = φgt + (bt − ct ).

(4.B.20)

Under the special condition that φ = 0 , this equation becomes Et (bt+1 − ct+1 ) = [β(γ + δ)]−1 (bt − ct ),

(4.B.21)

A Brock-Mirman model

97

which states that the shadow price of consumption (λ1t = bt − ct ) follows a first-order autoregressive process. Under the further special condition that β(γ + δ) = 1 , the shadow price of consumption follows a martingale. 17 Finally, under the even further special condition that bt is a martingale, ( 4.B.21 ) asserts that consumption is a martingale. The Euler equation ( 4.B.21 ) is satisfied by the consumption plan ct = bt for t ≥ 0.

(4.B.22)

Solving ( 4.B.2 ) and ( 4.B.4 ) for it under this plan gives kt = (γ + δ)kt−1 + d1t − bt .

(4.B.23)

Note that in the special case that λ1t (and maybe also ct ) is a martingale, (γ + δ) = 1/β , so that {kt } given by ( 4.B.23 ) is a “process of exponential order 1/β ”. This implies that kt does not belong to L20 . Nevertheless, the transversality condition ( 4.B.13 ) is satisfied because λ3t = φgt + (bt − ct ) = 0 along this solution, so that lim β t λ3t kt = 0

t→∞

along this solution. Thus, when φ = 0 , it is optimal to consume bliss consumption always and to adjust the capital stock to support this consumption plan. The difference equation ( 4.B.23 ) implies that kt = ξ t k0 +

t−1 X j=0

ξ j (d1t−j − bt−j )

where ξ ≡ γ + δ . If bt − d1t > α > 0 for some α for all t , then kt will eventually become negative and, indeed, will eventually fall below any finite negative number. Such a consumption path is eventually being supported by “borrowing” or by ‘negative capital.’ In the interests of attaining an ‘Euler equation’ for capital, we substitute the following two implications of the constraints into the Euler equation: ct = (γ + δ)kt−1 + d1t − kt gt = φkt − φδkt−1 After rearrangement, this gives the following Euler equation for capital: ηEt {kt+1 − ψkt + β −1 kt−1 } = Et zt where

(4.B.24)

η = β[δφ2 + (γ + δ)] ψ=

βδ 2 φ2 + β(γ + δ)2 + φ2 + 1 β(δφ2 + (γ + δ))

(4.B.25)

zt = bt − β(γ + δ)bt+1 − d1t + β(γ + δ)d1t+1 17 The condition that β(γ + δ) ≡ 1 plays the role of a “growth condition” in the model of Jones and Manuelli [1988].

Optimal Resource Allocation

98

We will solve the Euler equation ( 4.B.24 ) using the “certainty equivalence” methods described in Sargent [1987, ch. XIV] and Hansen and Sargent [1980, 1981]. This involves first solving the deterministic version of ( 4.B.25 ), and then replacing “feedforward” terms with their expectations conditioned on time t information. We begin by solving the deterministic version of the Euler equation ( 4.B.24 ): η{kt+1 − ψkt + β −1 kt−1 } = zt

(4.B.26)

ηL−1 {1 − ψL + β −1 L2 }kt = zt .

(4.B.27)

Write this as

We seek a factorization of the polynomial in L : (1 − ψL + β −1 L2 ) = (1 − λ1 L)(1 − λ2 L)

(4.B.28)

Evidently ψ = λ 1 + λ2 λ1 λ2 = β −1 . Thus we have 1 λ1 β

(4.B.29)

1 = ψ. λ1 β

(4.B.30)

λ2 = and λ1 +

Equations ( 4.B.29 ) and ( 4.B.30 ) imply that λ1 and λ2 = λ1β are the intersections of the 1 1 in figure 4.B.1. Since the function line of zero slope and height ψ with the curve λ + λβ

p

1 achieves a minimum of 2/ β at the value λ = 1/ f (λ) = λ + λβ solution of ( 4.B.30 ) exists, it satisfies, without loss of generality,

0 < λ1 < λ2 >

Substituting ( 4.B.28 ) into ( 4.B.27 ) gives

η[(1 − λ1 L)(1 −

p

β , it follows that if a

1

p

β

1

p

.

β

1 L)]kt+1 = zt . λ1 β

(4.B.31)

A Brock-Mirman model

99

b λ + λ -1

bλ

ψ

1+b

λ1

-.5

λ2

b

λ

Figure 4.B.1: The function bλ + 1/λ and its intersections with ψ , which determine the roots λ of the characteristic polynomial ( 4.B.28 ). We start analyzing the solution of ( 4.B.31 ) by returning to the special case in which φ = 0 . In this case, ( 4.B.25 ) implies that ψ=ξ+

1 , βξ

ξ = γ + δ,

(4.B.32)

η = βξ. It then follows immediately from ( 4.B.30 ) that we can take λ1 =

1 βξ

(4.B.33)

λ2 = ξ. In the special case that the shadow price of consumption is a martingale, βξ = 1 , so that 1 . The Euler equation thus becomes, in the special case that φ = 0 , λ1 = 1 and λ2 = β βξ{(1 −

1 L)(1 − ξL)}kt+1 = zt . βξ

But from the constraints to our problem, (1 − ξL)kt+1 = d1t+1 − ct+1

Optimal Resource Allocation

100

Substituting this and the last line of ( 4.B.25 ) into the Euler equation gives βξ(1 −

1 L)(d1t+1 − ct+1 ) = (βξ − L)(d1t+1 − bt+1 ) βξ

or (βξ − L){(d1t+1 − ct+1 } = (βξ − L)(d1t+1 − bt+1 ), an equation that is satisfied by setting ct = bt for all t . Thus, our analysis of the Euler equation for capital in the case that φ = 0 reconfirms our earlier derivation that the optimal plan involves setting ct = bt and choosing whatever capital path is required to support this. We begin to study the case when φ > 0 by considering the special case in which φ is positive but arbitrarily close to zero. In particular, φ can be chosen sufficiently close to zero that in the Euler equation for capital, η{(1 − λ1 L)(1 − λ2 L)}kt+1 = zt , 1 , and λ is arbitrarily close to ξ . This η is arbitrarily close to βξ, λ1 is arbitrarily close βξ 2 can be verified by using a version of figure 4.B.1. It is tempting to suppose that since the Euler equation is arbitrarily close to that for the φ = 0 case, the optimal solution for kt will be close to the solution for kt found in the φ = 0 case, namely,

kt = ξ t k0 +

t−1 X j=0

ξ j (d1t−j − bt−j ).

(4.B.34)

We now show that this supposition is wrong. Note that when kt obeys ( 4.B.34 ), it = kt − δkt−1 , obeys it = ξ t−1 (ξ − δ)k0 + d1t − bt + (ξ − δ)

t−2 X j=0

ξ j (d1t+j−1 − bt−j−1 ).

(4.B.35)

Also, ct = bt ∀t in this case. When it follows ( 4.B.35 ), it is a process of exponential order ξ . It follows that φit is also a process of exponential order ξ when φ > 0 . Now since ℓt = φit along the optimal path, we have that ∞ X i=0

β t ℓ2t = φ2

∞ X

β t i2t .

(4.B.36)

t=0

The process i2t is of exponential order ξ 2t along the solution ( 4.B.35 ). The infinite series ( 4.B.36 ) will converge if and only if β · ξ 2 < 1, or ξ

√1 , so β

that this condition is violated. In this case, ( 4.B.36 ) diverges to +∞ . This means that when investment follows the path ( 4.B.35 ), the objective function for the social planning problem diverges to −∞ when φ = 0 . Since it is possible to find investment paths that leave the value of the objective function finite, a plan in which the objective function diverges to −∞ cannot be optimal. Notice the role that the assumption that φ > 0 plays in the above argument. An alternative argument can be used to show that the path ( 4.B.35 ), or one close to it, cannot be optimal when φ > 0 and ξ > √1 . This argument involves checking the β

transversality condition, which is lim β t kt λ3t = 0.

t→∞

Computing, we have lim β t kt λ3t

t→∞

= lim β t kt (φgt + λ1t ) t→∞

= lim β t kt [φ2 (kt − δkt−1 ) + (bt − ct )]. t→∞

= lim β t [φ2 (kt2 − δkt kt−1 ) + (bt − ct )kt ] t→∞

For a solution that involves setting bt = ct , this becomes lim β t [φ2 (kt2 − δkt kt−1 )] = 0

(4.B.37)

t→∞

A necessary and sufficient condition for ( 4.B.37 ) to be satisfied is that {kt } be of exponential order less than √1 . Along a solution like ( 4.B.34 ), this requires that ξ < √1 , which is β

β

ruled out in the special case that the shadow price of consumption is a martingale. Arguments along these lines can be used to establish generally that when φ > 0 , the solutions for it and for kt are required to be of exponential order less than √1 . β

To solve for the optimal plan when φ > 0 , we return to the factored Euler equation ( 4.B.31 ): η[(1 − λ1 L)(1 −

1 L)]kt+1 = zt λ1 β

(4.81)

p

where 0 < λ1 < 1/ β . Formally, express (1 − λ1β L) = − λ1β L(1 − λ1 βL−1 ) . Substitute 1 1 this into ( 4.B.31 ) to get −η [(1 − λ1 βL−1 )(1 − λ1 L)]kt = zt λ1 β

(4.B.38)

Operating on both sides of ( 4.B.38 ) with the stable (forward) inverse of (1 − λ1 βL−1 ) gives λ β 1 zt (1 − λ1 L)kt = − 1 η (1 − λ1 βL−1 )

(4.B.39)

Optimal Resource Allocation

102

or kt = λ1 kt−1 −

p

Since λ1 < 1/

β, λ1 β

0 , equation ( 4.B.40 ) gives the unique p solution of the Euler equation that satisfies the transversality condition. Because λ1 < 1/ β , kt belongs to L20 .

4.B.1. Uncertainty In the case that zt is a random sequence, the solution when φ1 > 0 is given by kt = λ1 kt−1 −

λ1 β η

∞ X

(λ1 β)j Et zt+j

(4.B.41)

j=0

That this is the solution can be verified by applying the methods of Sargent [1987, chapter XIV]. Consider applying ( 4.B.41 ) in the special case that makes consumption a martingale: βξ = 1, η = βξ = 1, λ1 = 1, bt = ¯b for all t . In this case ( 4.B.41 ) becomes, kt − kt−1 = −β

∞ X j=0

β j Et (d1t+j+1 − d1t+j )

(4.B.42)

We can use a summation by parts argument to show that Et

∞ X j=0

β j (d1t+j+1 − d1t+j ) = (β −1 − 1)Et

∞ X j=0

(4.B.43) β j d1t+j − β −1 d1t

In particular, note that ∞ X j=0

β j (d1t+j+1 − d1t+j )

=

∞ X j=1

β j−1 d1t+j −

= (β −1 − 1)

∞ X j=0

∞ X

β j d1t+j

j=0

β j d1t+j − β −1 d1t .

A Brock-Mirman model

103

Note that from the constraints ct = (γ + δ)kt−1 − kt + d1t or ct =

1 kt−1 − kt + d1t β

(4.B.44)

in the special case that (γ + δ)β = 1 , which we are studying. Substituting ( 4.B.42 ) and ( 4.B.43 ) into ( 4.B.44 ) and rearranging gives

ct =

∞

X 1 β j Et d1t+j . − 1 kt−1 + (1 − β) β

(4.B.45)

j=0

With kt−1 interpreted as “assets” and {d1t } interpreted as “labor income”, representation ( 4.B.45 ) matches the representation of the permanent income theory of consumption that is associated with a linear quadratic version of Hall’s model. In this model, φ = 0 , so that ( 4.B.42 ) and ( 4.B.45 ), which emerge from imposing that {kt } reside in L20 , are not optimal for the original problem as stated. The solution ( 4.B.45 )results from imposing as a side condition on the problem a version of ( 4.A.10 ). This side condition is intended to capture the idea that it is not really feasible to drive capital to negative infinity as quickly as the (unrestricted) φ = 0 solution would require. The solution ( 4.B.45 ) is well approximated by the solution of the original problem with φ > 0 but φ very close to zero. Instead of imposing the requirement that {kt }ǫL20 as a sort of “feasibility” condition, setting φ > 0 rigs preferences so that the social planner always prefers to make {kt } ∈ L20 .

4.B.2. Optimal Stationary States Temporarily assume that bt = ¯b and d1t = d¯ for all t . To solve for the optimal stationary values of ct and kt (if they exist), we can use equation ( 4.B.20 ) and the following constraints: φit = gt it = kt − δkt−1 ct + it = γkt−1 + d1t

(4.52) (4.53) (4.51)

¯ for all t gives Evaluating these at steady state levels ct = c¯ and kt = k ¯ + d. ¯ c¯ = (γ + δ − 1)k Substituting the constraints into the Euler equation ( 4.B.20 ) and evaluating at ct = c¯ and ¯ gives kt = k ¯ = [1 − β(γ + δ)](¯b − c¯) φ2 (βδ − 1)(1 − δ)k

104

Optimal Resource Allocation

¯ gives Solving the two preceding equations for c¯ and k ¯ = [φ2 (βδ − 1)(1 − δ) + (1 − β(γ + δ))(γ + δ − 1)]−1 k ¯ (1 − β(γ + δ)) · (¯b − d) c¯ =

(γ + δ − 1)(1 − β(γ + δ)) [φ2 (βδ − 1)(1 − δ) + (1 − β(γ + δ))(γ + δ − 1)]

(4.B.46)

(4.B.47)

¯ + d. ¯ (¯b − d) In the special case that φ = 0 , these solutions imply that c¯ = ¯b , so that consumption is at bliss consumption and the steady state value of the multiplier λ1t is zero. When φ = 0 , the steady state value of k can be taken to be ¯= k

1 [d¯ − ¯b], 1 − (γ + δ)

(4.B.48)

a solution that makes sense only when (γ + δ) < 1 . Note that the constraints imply that capital evolves according to kt = (γ + δ)kt−1 − ct + d1t . Setting ct = c¯ and d1t = d¯ implies ¯ kt = (γ + δ)kt−1 − c¯ + d. The solution of this equation is

kt = (γ + δ)t k0 + (d¯ − c¯)

t−1 X

(γ + δ)j .

j=0

¯ when c¯ = ¯b and (γ + δ) < 1 . This solution converges to the solution ( 4.B.48 ) for k

Chapter 5 The Commodity Space

5.1. Valuation This chapter describes a concept of value that we shall later use to formulate a decentralized version of our model in which the decisions of agents are reconciled in a competitive equilibrium. We describe a commodity space in which both the quantities and prices will reside. The stochastic Lagrange multipliers of chapter 4 are very closely related to the equilibrium prices that we shall compute, and live in the same mathematical space with prices. The social planning problem studied in chapter 4 produces an outcome in which the process for consumption {ct } is an n -dimensional stochastic process that belongs to L20 . To calculate the value π(c) of a particular consumption plan c = {ct } from the vantage of time zero, we shall use the representation π(c) = E

∞ X t=0

β t p0t · ct | J0 ,

where p0t belongs to L20 . The text of this chapter presents a heuristic justification for so representing the value of {ct }. We proceed by reviewing several examples of commodity spaces and valuation functions. The appendix contains a more formal treatment.

– 105 –

106

The Commodity Space

5.2. Price systems as linear functionals We follow Debreu (1954) and express values by using a linear functional π that maps elements of a linear space L into the real line. The space L is taken as the commodity space, elements of which are the vectors of commodities to be evaluated. The functional π assigns values to points in L. It is convenient when the functional π has an inner-product representation, which is a representation in which the value π(c) of a commodity point c equals the inner product of c ˜ When such a representation exists, with a point p in another linear space L. we can write π(c) =< c | p > for all c in L (5.2.1) ˜ and < · | · > denotes an inner product. In all of the cases that we where pε L ˜ = L, so that c and p reside in the same linear consider, it turns out that L space. Next we consider several examples of a commodity space L, a valuation functional π , and an inner product representation for π .

5.3. A one period model under certainty Suppose that there is one period and no uncertainty. Let there be n consumption goods. Let c be an n × 1 vector of consumption goods. Let the commodity space L be Rn , the n -dimensional Euclidean space. In this case, the value of a vector c is given by n X π(c) =< c | p >≡ ci pi i=1

where < · | · > denotes the inner product, and p is an n -dimensional price vector that belongs to L = Rn . Note that both c and p belong to the same linear space L.

One period under uncertainty

107

5.4. One period under uncertainty Suppose there is again one period, but now there is uncertainty about economic outcomes. Prior to the resolution of uncertainty, the quantity of the ith consumption good is a random variable ci (ω), where ω is the state of the world to be realized. Let c = c(ω) be an n -dimensional random vector whose ith component is ci (ω). Let prob(ω ) be the probability density function of ω . We want to evaluate a bundle of consumption goods prior to the resolution of uncertainty. Introducing uncertainty serves to increase the dimension of the commodity space, there being a vector c(ω) for each state of the world ω ∈ Ω, where Ω is the set of possible states of the world. When there is an infinite number of states of the world Ω, the commodity space L becomes infinite dimensional. To evaluate a state-contingent bundle of consumption goods prior to the resolution of uncertainty requires a well defined notion of “adding up” or integrating across states of the world. When the number of states of the world is finite (or countable), it is natural to follow Arrow and Debreu and to define an n -dimensional vector of statecontingent prices q(ω), where Ω = [ω1 , ω1 , . . . , ωN ] is the set of possible states of the world. The value of the random vector c can then be represented as π(c) =

N X j=1

c(ωj ) · q(ωj ) ≡< c | q > .

(5.4.1)

Here both c and q are elements of L, the space of n -dimensional random vectors indexed by the state of the world. The ith component of q(ω), qi (ω), is to be interpreted as price of one unit of the ith consumption good contingent on the state of the world being ω . It is convenient to represent π(c) in the alternative form π(c) =

N X j=1

c(ωj ) · p(ωj ) prob (ωj ),

(5.4.2)

where q(ωj ) = p(ωj ) prob (ωj ). Here c and p are each vectors in L, the space of n -dimensional random vectors. Notice that (5.4.2) implies π(c) = Ec · p ≡< c | p > . Representation (5.4.1) is often used in contexts in which there is a finite or countable number of states of the world. We find it easier to use representations

108

The Commodity Space

that build upon (5.4.2) because we shall be dealing with environments with an uncountable number of states of the world.

5.5. An infinite number of periods and uncertainty We now come to the main case studied in this book. The n -dimensional vector of consumption goods ct is indexed both by states of the world and by time. We define an information set Jt as in chapters 2 and 3. Let L be the space of all n -dimensional stochastic processes {ct : t = 0, 1, . . .} for which ct is in Jt for all t and for which ∞ X β t E(ct · ct ) < ∞. (5.5.1) t=0

The constraint that ct be in Jt is imposed because we want to represent the values only of contingent claims that depend on information available when the contingency is realized. The inequality restriction in (5.5.1) identifies which claims might have finite value. In addition to integrating over states of the world, we also must sum over points in time. We find it convenient to use the discount factor β in performing this summation. Hence we use the following inner product: < c | p >=

∞ X t=0

β t E(ct · pt ).

(5.5.2)

In this case, the price system used to represent the valuation functional is an n -dimensional stochastic process {pt : t = 0, 1, . . .} in L.

An infinite number of periods and uncertainty

109

5.5.1. Conditioning information So far we have considered valuation functions that map into the real numbers IR . This approach suffices for representing competitive equilibrium prices for markets that meet and clear prior to the realization of any information. However, we also want to reopen markets and to study valuations at later points in time, conditioned on information available then. Consider valuation from the vantage point of time τ . Let valuation be conditioned on the time τ information set Jτ . Let πτ be a time τ valuation function. We take the domain of πτ to be the space Lτ consisting of all n dimensional processes {ct+τ : t = 0, 1 . . .} where ct+τ is in Jτ and ∞ X t=0

β t E(ct+τ · ct+τ ) | Jτ < ∞

(5.5.3)

with probability one. The range of πτ is Jτ because valuations reflect the available conditioning information. There is no longer an inner-product representation for πτ because the range of πτ is not the real line. Rather, the range is the space of random variables depending on Jτ . However, we can follow Harrison and Kreps (1979) and Hansen and Richard (1987) by using a conditional inner-product representation: πτ (c) =< c | p >τ =

∞ X t=0

β t E(ct+τ · pt+τ | Jτ )

(5.5.4)

where {pt+τ : t = 0, 1, . . .} is a price process in L. The value assigned by πτ is a random variable in L2τ .

110

The Commodity Space

5.6. Lagrange multipliers While we have focused on representing valuation in a competitive equilibrium, much of our discussion applies to using the method of Lagrange multipliers for solving constrained optimization problems. The vector of Lagrange multipliers for a vector of constraints indexed by states of the world and calendar time can be regarded as a stochastic processes {Mt : t = 0, 1, . . .} in a space L. The contribution to the Lagrangian is given by a corresponding linear functional µ with an inner product representation µ(ε) =< ε | M >=

∞ X t=0

β t E(εt · Mt )

(5.6.1)

where εt is the deviation of the constraint at time t .

5.7. Summary Our purpose in this chapter has been to lay groundwork necessary to decentralize the economy described in chapter 3 into one with a collection of price-taking agents whose decisions are coordinated through markets. The Appendix to this chapter describes the valuation functions that we use in more mathematical detail.

A. Appendix As was indicated above, we model π as a linear functional on a space L . The space L is assumed to be a linear space, by which we mean that for any two members x1 and x2 in L and any two real numbers c1 and c2 in R , c1 x1 + c2 x2 are in L . In addition, we suppose that there is an inner product < · | · > defined on L . This inner product can be used to define a norm || x ||=< x | x >1/2 and hence a metric. We take L to be complete. This means that all Cauchy sequences in L converge to an element of L . The commodity spaces in all of the examples described in the text are complete linear spaces. The restriction that π be linear requires that π(c1 x1 + c2 x2 ) = c1 π(x1 ) + c2 π(x2 ) . According to the Riesz Representation Theorem, π has an inner product representation whenever π is continuous at zero. When conditioning information is introduced, it is convenient to work with a space LJ that is linear conditioned on J . For the moment, consider LJ to be a collection of

Appendix

111

random variables. Products and sums of random variables are also random variables. For LJ to be linear conditioned on J , for any two elements x1 and x2 of LJ and any w1 and w2 in J, we require that w1 x1 + w2 x2 is in LJ . Similarly, πJ is conditionally linear if πJ (w1 x1 +w2 x2 ) = w1 πJ (x1 )+w2 πJ (x2 ) . The rationale for focusing on conditional linearity is that information in J can be used to construct consumption plans or trading strategies. Hansen and Richard (1987) obtained a conditional counterpart to the Riesz Representation Theorem that establishes the existence of a representation πJ (x) = E(x · p | J) for some p in LJ . The restriction that LJ be a space of random variables is too limited for our purposes. Instead, we are interested in spaces of n -dimensional stochastic processes. Given an initial probability space (Ω, F, Pr) and a sequence {Ft : t = 0, 1, . . .} of subsigma algebras of F , we construct a new probability space (Ω+ , F + , Pr+ ) where Ω+ is the Cartesian product of Ω+ , the nonnegative integers, and the set {1, 2, . . . , n} , and where Pr+ is the product measure of Pr , a measure that assigns β t (1 − β) to integer t , and 1/n to integer j . The sigma algebra F + is generated by sets of the form: {(w, t, j) : w ∈ ft,j }

(5.A.1)

where {ft,j : t = 0, 1, . . . ; , j = 1, 2, . . . , n} is a collection of sets in F such that ft,j is in Ft for all t and j . An n -dimensional stochastic process defined on the original space can be viewed as a random variable on the product space. Thus we can apply the preceding analysis to obtain a conditional inner product representation for πτ described in the text.

Chapter 6 A Competitive Economy

6.1. Introduction This chapter describes a decentralized version of our economy. We assign ownership and decision making to three distinct economic entities, a household and two kinds of firms. We define a competitive equilibrium. Versions of the two fundamental theorems of welfare economics are true. We establish these theorems by exhibiting the connection between a competitive equilibrium and a social planning problem. A price system supports the competitive equilibrium, and implies interest rates and prices for derivative assets. The representative household can be interpreted as a single individual drawn from a population that is homogeneous in all respects. Alternatively, the representative household can be interpreted along lines to be described in chapter 12, as an artificial or “average” household that emerges from aggregating over the preferences and endowments of a collection of households. The representative household owns the technology shock process dt , and each period sells to firms the current period’s realization of the shock process. The household owns the initial stocks h−1 of household capital and k−1 of productive capital, the latter of which it sells to firms. It sells this initial capital for a value v0 · k−1 . The household sells its input ℓt to firms. The household uses its resources to purchase consumption goods, which add to its stocks of consumer durables and thereby generate consumption services and utility. Of the two types of firms, the first type rents capital from firms of type II, rents labor from the household, and buys the current period’s realization of the technology shock process dt from the household. A firm of type I produces new consumption and investment goods, sells the consumption goods to the household, and sells the investment goods to the firms of type II. A firm of type II purchases the initial capital stock k−1 and all of the investment goods produced each period, then rents capital to firms of type I. We use a formulation of a price system which is mathematically convenient, as well as economically interpretable. We let the price system be [v0 , {p0t , wt0 , αt0 , qt0 , – 113 –

114

A Competitive Economy

0 rt0 }∞ t=0 ], where v0 is a vector that prices the initial capital stock k−1 ; pt is an 0 nc × 1 stochastic process that prices the consumption process ct ; wt is a scalar stochastic process that prices ℓt ; αt0 is a vector stochastic process that prices the process {dt }; qt0 is an nk ×1 vector stochastic process that prices new investment goods; and rt0 is an nk ×1 vector stochastic process of capital rental rates. Each 2 component of [{p0t , wt0 , αt0 , qt0h, rt0 }∞ t=0 ] resides in the mathematical space L0 defined earlier, namely, L20 = {yt }∞ t=0 : yt is a random variable in Jt for t ≥ i P∞ t 2 0, and E t=0 β yt | J0 < +∞ . That ‘yt is in Jt ’ means that yt can be

expressed as a measurable function of Jt = [wt , x0 ], where J0 = [x0 ]. The P∞ square summability requirement, E t=0 β t yt2 | J0 < ∞ , imposes a stochastic version of a requirement that yt not grow too fast in absolute value. Stochastic processes for both prices and quantities in our economy must reside in L20 . By virtue of a Cauchy-Schwartz inequality, this makes the conditional inner products to be used in the budget constraints and objective functions below well defined and finite in equilibrium. This chapter formulates and computes a competitive equilibrium. We proceed by first describing the problem for each of our three classes of agents in terms of a Lagrangian. Next we obtain the first order conditions from these Lagrangians. By “matching up” these first-order conditions to the first order conditions found in chapter 3 for the social planning problem, we accomplish two goals. First, we can verify the two fundamental theorems of welfare economics for our economy. Second, we can describe an efficient algorithm for computing the equilibrium price system in terms of the matrices Mk , Mh , Ms , Md , Mc , and Mi of chapter 3 associated with the multipliers for the social planning problem. We first describe the problems faced by each of our three types of agents.

The Problems of Households and Firms

115

6.2. The Problems of Households and Firms

6.2.1. Households The household chooses stochastic processes for {ct , st , ht , ℓt }∞ t=0 , each element of which is in L20 , to maximize −

∞ h i X 1 β t (st − bt ) · (st − bt ) + ℓ2t E0 2 t=0

(6.2.1)

subject to E

∞ X t=0

β t p0t · ct | J0 = E

∞ X t=0

β t (wt0 ℓt + αt0 · dt ) | J0 + v0 · k−1

st = Λht−1 + Πct ht = ∆h ht−1 + Θh ct ,

h−1 , k−1 given.

(6.2.2) (6.2.3) (6.2.4)

The household and each firm acts as a price taker. The optimal contingency plan for (ct , st , ht , ℓt ) must be “realizable” in the sense that time t decisions must be contingent only on information available at time t , i.e., it must reside in L20 .

6.2.2. Firms of type I A firm of type I rents capital and labor, and buys the realization of the endowment process dt . It uses these inputs to produce consumption goods and investment goods, which it sells. The firm of type I chooses stochastic processes for {ct , it , kt , ℓt , gt , dt }, each element of which is in L20 , to maximize E0

∞ X t=0

β t (p0t · ct + qt0 · it − rt0 · kt−1 − wt0 ℓt − αt0 · dt )

(6.2.5)

subject to Φc ct + Φg gt + Φi it = Γkt−1 + dt

(6.2.6)

116

A Competitive Economy

− ℓ2t + gt · gt = 0

(6.2.7)

6.2.3. Firms of type II A firm of type II is in the business of purchasing investment goods and renting capital to firms of type I. A firm of type II faces as a price taker the vector v0 and the stochastic processes {rt0 , qt0 }. The firm chooses k−1 and stochastic processes for {kt , it }∞ t=0 to maximize E

∞ X t=0

subject to

β t (rt0 · kt−1 − qt0 · it ) | J0 − v0 · k−1 kt = ∆k kt−1 + Θk it .

(6.2.8)

(6.2.9)

6.3. Competitive Equilibrium We define a competitive equilibrium for this economy. Definition: A competitive equilibrium is a price system [v0 , {p0t , wt0 , αt0 , qt0 , rt0 }∞ t=0 ] and an allocation {ct , it , kt , ht , gt , dt }∞ that satisfy the following conditions: t=0 a. Each component of the price system and the allocation resides in the space L20 .

b. Given the price system and given h−1 , k−1 , the stochastic process {ct , st , ℓt , ht }∞ t=0 solves the consumer’s problem.

c. Given the price system, the stochastic process {ct , it , kt , ℓt , dt , gt } solves the problem of the firm of type I.

d. Given the price system, the vector k−1 and the stochastic process {kt , it }∞ t=0 solve the problem of the firm of type II.

Lagrangians

117

6.4. Lagrangians We now formulate each agent’s problem as a Lagrangian, and obtain the associated first order conditions.

6.4.1. Households The household’s Lagrangian is ∞ X βt (st − bt ) · (st − bt ) + ℓ2t /2 L = − E0 t=0

0 0 0 + µw 0 [pt · ct − wt ℓt − αt · dt ]

+ µs′ t [st − Λht−1 − Πct ] + µh′ [h − ∆ h − Θ c ] + µw t h t−1 h t t 0 v0 · k−1 . h s Here µw 0 is a scalar and {µt , µt } are sequences of vectors of stochastic Lagrange multipliers. The first order conditions with respect to st , ℓt , ct , and ht , respectively, are:

st : ℓt : ct : ht :

(st − bt ) + µst = 0,

ℓt −

wt0

·

µw 0

t≥0

t≥0

= 0,

′ h ′ s 0 µw 0 pt − Π µt − Θh µt = 0,

− β Et Λ

′

µst+1

−β

Et ∆′h

µht+1

+

µht

= 0,

t≥0

t≥0

Solving these equations, we obtain

µht = Et

µst = bt − st ,

t≥0

(6.4.1)

wt0 = ℓt /µw 0,

t≥0

(6.4.2)

∞ X

β τ (∆′h )τ −1 Λ′ µst+τ

τ =1 0 ′ s ′ h µw 0 pt = Π µt + Θh µt ,

, t≥0

t≥0

(6.4.3) (6.4.4)

118

A Competitive Economy

6.4.2. Firms of type I The Lagrangian of a type I firm is ∞ X t LI = E0 β [p0t · ct + qt0 · it − rt0 · kt−1 − wt0 ℓt − αt0 · dt ] t=0

+ Ld′ t [Γkt−1 + dt − Φc ct − Φg gt − Φi it ] ℓ′ 2 + Lt [(ℓt − gt · gt )/2] .

Here {Ldt , Lℓt } is a vector stochastic process of Lagrange multipliers. The first order conditions associated with interior solutions for ct , it , kt , ℓt , dt , and gt , respectively, are ct :

p0t − Φ′c Ldt

= 0,

t≥0

(6.4.5)

it :

qt0 − Φ′i Ldt

= 0,

t≥0

(6.4.6)

kt :

0 rt+1 − Γ′ Ldt+1

= 0,

t ≥ −1

(6.4.7)

ℓt :

− wt0 + Lℓt ℓt

= 0,

t≥0

(6.4.8)

dt :

− αt0 + Ldt

= 0,

t≥0

(6.4.9)

gt :

−

Φ′g Ldt

−

gt Lℓt

= 0,

t≥0

Solving (6.4.5) and (6.4.10) for Ldt gives ′ −1 Φc p0t d . Lt = Φ′g −gt Lℓt

(6.4.10)

(6.4.11)

From (6.4.8), the solution for Lℓt satisfies Lℓt = wt0 /ℓt .

(6.4.12)

Equations (6.4.6), (6.4.7) and (6.4.9) imply qt0 = Φ′i Ldt

(6.4.13)

rt0 = Γ′ Ldt

(6.4.14)

αt0 = Ldt

(6.4.15)

Lagrangians

119

6.4.3. Firms of type II The Lagrangian of a type II firm is LII = E0

∞ X t=0

β (rt0 · kt−1 − qt0 · it ) t

+ ηt′ (∆k kt−1 + Θk it − kt ) − v0 · k−1 where {ηt } is a sequence of stochastic Lagrange multipliers. The first order conditions for interior solutions with respect to kt , it , and k−1 , respectively, are 0 kt : βEt rt+1 − ηt + βEt ∆′k ηt+1 = 0,

it :

− qt0 + Θ′k ηt

= 0,

k−1 : r00 + ∆′k η0 − v0

t≥0

(6.4.16)

t≥0

(6.4.17)

=0

(6.4.18)

, t≥0

(6.4.19)

Solving (6.4.16) for ηt gives ηt = Et

∞ X

′(j−1) 0 rt+j

β j ∆k

j=1

Equation (6.4.17) implies qt0 = Θ′k ηt ,

t≥0

(6.4.20)

Equation (6.4.18) implies v0 = r00 + ∆′k η0

(6.4.21)

120

A Competitive Economy

6.5. Equilibrium Price System Our task now is to find stochastic processes of prices, quantities, and Lagrange multipliers that satisfy the first-order conditions for each of our three classes of agents for all time and contingencies. We proceed constructively to link equilibrium prices to the Lagrange multipliers for the planning problem. Recall the following equations obeyed by the Lagrange multipliers associated with the social planning problem: Mst = bt − st

(4.8)

(4.9)

(4.11)

(4.19)

X ∞

Mht

=E

Mdt

Φ′c = Φ′g

Mkt

β

τ

τ =1

=E

−1

X ∞

τ

(∆′h )τ −1 Λ′ Mst+τ

| Jt

Φ′h Mht + Π′ Mst −gt

′ τ −1

β (∆ )

Γ

τ =1

′

Mdt+τ

| Jt

In chapter 3, we gave formulas for these multipliers along the optimum of the social planning problem, namely, (4.21)

Mkt = Mk xt and Mht = Mh xt

(4.22)

Mst = Ms xt

(4.23)

Mdt = Md xt .

We also defined shadow prices for consumption and investment: (4.24)

Mct = Mc xt , Mc = Θ′h Mh + Π′ Ms

(4.25)

Mit = Mi xt , Mi = Θ′i Mk .

Equilibrium Price System

121

We gave formulas for the matrices Ms , Mk , Mh and Md in terms of the optimal value function of the social planning problem. The formulas (4.21), (4.22), (4.23), (4.24), (4.25) for the multipliers are evaluated along the solution xt+1 = Ao xt + Cwt+1 of the social planning problem. We can compute the equilibrium price system in terms of the multipliers from the social planning problem. For the time being let µw 0 be a free parameter. Later we shall indicate how choosing the scalar marginal utility of wealth at time zero, µw 0 , amounts to specifying a numeraire for our price system. We propose to set c w p0t = Π′ Mst + Θ′h Mht /µw (6.5.1) 0 = Mt /µ0 wt0 =| Sg xt | /µw 0

(6.5.2)

rt0 = Γ′ Mdt /µw 0

(6.5.3)

i w qt0 = Θ′k Mkt /µw 0 = Mt /µ0

(6.5.4)

αt0 = Mdt /µw 0

(6.5.5)

′ k w v0 = Γ′ Md0 /µw 0 + ∆k M0 /µ0 .

(6.5.6)

We shall verify that with this price system, values can be assigned to the Lagrange multipliers for each of our three classes of agents that cause all of their first-order necessary conditions to be satisfied at these prices and at the quantities associated with the optimum of the social planning problem. For the household, we set (6.5.7) µst = Mst µht = Mht

(6.5.8)

With these choices of multipliers, equations (6.4.1), (6.4.2), (6.4.3) and (6.4.4) are evidently satisfied at the proposed equilibrium prices (6.5.1) – (6.5.6) and at the quantities associated with the optimum of the social planning problem. For the firm of type I, we set Ldt = Mdt /µw 0

(6.5.9)

Lℓt = 1/µw 0.

(6.5.10)

With the settings (6.5.9) for Ldt , (6.5.10) for Lℓt , and the price process (6.5.1), equation (6.4.11) becomes equivalent with (4.11) from the social planning problem. Equation (6.5.3) for rt0 implies that the firm’s marginal condition (6.4.14)

122

A Competitive Economy

is satisfied along the solution of the social planning problem. Similarly, (6.4.20) implies that (6.4.15) is satisfied. Formula (6.5.4) for qt0 together with the fourth equation of (4.8) (−Φ′i Mdt +Θ′k Mkt = 0) implies that (6.4.13) is satisfied along the solution of the social planning problem. Finally, (6.5.9)–(6.5.10) imply that (6.4.10) is equivalent with the second equation of (4.8) (−gt − Φ′g Mdt = 0). Thus, with settings (6.5.8), (6.5.9), price system (6.5.1)–(6.5.6) implies that firm I’s first order necessary conditions are satisfied along the quantity path implied by the social optimum. For the firm of type II, we set ηt = Mkt /µw 0.

(6.5.11)

With this setting, (6.5.6) and (3.19) imply that (6.4.19) (and thus (6.4.16)) is satisfied. Also, (6.4.20) is evidently satisfied as is (6.4.21). Thus, the first order conditions for firms of type II are all satisfied at price system (6.5.1)–(6.5.6) along the solution of the social planning problem. We are finished. In summary, the price system (6.5.1)–(6.5.6) supports the allocation associated with the optimum of the social planning problem as a competitive equilibrium. The direction of argument can be reversed to establish that a competitive equilibrium solves the social planning problem. This argument uses a competitive equilibrium allocation and price system to define multiplier processes that satisfy the first order conditions for the social planning problem. 1 The scalar µw 0 that appears as a free parameter in (6.5.1)–(6.5.6) is evidently the marginal utility of wealth at time zero. In setting this parameter, we select a numeraire for our price system. For example, the j th consumption good at time zero can be selected as the numeraire by setting ¯j Mct = e¯j Mc x0 µw 0 =e where e¯j is a (1 × nc ) vector consisting of zeros in each location except the j th, where there is a one. For the j th consumption good at time zero to be a valid numeraire, we require that e¯j Mc x0 not equal zero. This is imposed in: Assumption 5.1: The random variable e¯j Mc x0 selected as numeraire differs from zero with probability one.

1 Since the solution of the social planning problem is unique, so is the competitive equilibrium.

Asset Pricing

123

6.6. Asset Pricing We can use the main idea behind “arbitrage pricing theory” to motivate asset pricing formulas. Arbitrage pricing theory extracts solely from the weak implication of equilibrium that assets must be priced so that budget sets offer no opportunities for earning sure returns with a zero commitment of resources. To illustrate this approach, imagine altering the representative household’s problem (6.2.1) – (6.2.4) by supplying it with one additional opportunity. The household can go into the securities business on the side by issuing securities that promise to pay off a stream of the (nc × 1) vector of consumption goods {yt }. We assume that {yt } ∈ L20 . Suppose there is a market in such securities and that the price at time 0 of one unit of such security is a0 . If the household sells S of these securities, its revenue at time 0 is Sa0 . To cover itself in all contingencies, the household must purchase state contingent claims to consumption in the amount {yt } for each unit of the security issued. The cost of purchasing these claims to support the sale of S securities is S·E

∞ X t=0

β t p0t · yt | J0 .

With this opportunity opened up to the household, the following term must be added to the right side of household’s budget constraint (6.2.2): S(a0 − E P∞

t

p0t

∞ X t=0

β t p0t · yt | J0 ).

· yt | J0 , the household can make the present value of conIf a0 > E t=0 β sumption as large as it wants by setting S equal to a suitable positive number, i.e., by selling the security whose price is a0 . However, for our economy, it is not feasible for the consumer to achieve any such desired present value of conP∞ sumption. Therefore, in equilibrium we cannot have a0 > E t=0 β t p0t · yt | J0 . P∞ Similarly, we cannot have a0 < E t=0 β t p0t · yt | J0 , because that would confront the household with the opportunity to make the present value of consumption as large as it wants by buying the security at prices a0 , then selling the returns yt in the market for state contingent claims. Therefore, we must have ∞ X a0 = E β t p0t · yt | J0 . (6.6.1) t=0

124

A Competitive Economy

We can use (6.6.1) to derive formulas for various special {yt } processes, and thereby recover versions of Lucas’s asset pricing model [1978], and theories of the term structure of interest rates. We derive more explicit formulas for assets with payoffs of the form yt = Ua xt

(6.6.2)

where Ua is an nc × n matrix. Substituting (6.6.2) and the pricing formula p0t = Mc xt /µw 0 into (6.6.1) gives a0 = E

∞ X t=0

β t x′t Za xt | J0

(6.6.3)

where Za = Ua′ Mc /µw 0.

(6.6.4)

We shall now show that a0 can be represented as a0 = x′0 µa x0 + σa where µa =

∞ X

β τ (Ao′ )τ Za Aoτ

(6.6.5)

(6.6.6)

τ =0

σa =

∞ X β trace Za β τ (Ao )τ CC ′ (Ao′ )τ . 1−β τ =0

(6.6.7)

According to (6.6.5), the asset price a0 turns out to be the sum of a constant σa , which reflects a “risk premium,” and a quadratic form in the state vector xt . To understand why σa reflects a risk premium, notice how the parameters in C influence σa but do not influence µa . To derive (6.6.5), first express (6.6.3) as 2 ∞ X

β t trace [Za xt x′t ] | J0 .

(6.6.8)

(Ao )τ CC ′ (Ao′ )τ + (Ao )t x0 x′0 (Ao′ )t .

(6.6.9)

a0 = E

t=0

For t ≥ 1, (1.5) implies that Ext x′t | J0 =

t−1 X

τ =0

2 An alternative way to derive these formulas is described in chapter [ ] (seasonality).

Asset Pricing

125

Substituting (6.6.9) into (6.6.8) and rearranging gives

a0 =

∞ X

β t trace

t=1

t−1 h X i Za (Ao )τ CC ′ (Ao′ )τ τ =0

+ trace Za

∞ X

t

o t

β (A )

x0 x′0

(6.6.10)

o′ t

(A ) .

t=0

Exchanging orders of summation in the first term on the right of (6.6.10) gives ∞ X t=1

t−1 h i X β t trace Za (Ao )τ CC ′ (Ao′ )τ τ =0

= trace Za =

∞ ∞ X X

τ =0 t=τ +1 ∞ X

β trace Za 1−β

≡ σa

β t (Ao )τ CC ′ (Ao′ )τ β τ (Ao )τ CC ′ (Ao′ )τ

τ =0

which establishes (6.6.7). The second term on the right side of (6.6.10) can be transformed (by repeatedly using the rule trace AB = trace BA) to x′0

∞ X t=0

β t (Ao′ )t Za (Ao )t x0 ≡ x′0 µa x0 ,

which defines the matrix µ given in (6.6.6). This completes our verification of the asset pricing formulas (6.6.5) – (6.6.7). To implement (6.6.5) requires the application of numerical methods to calculate the matrices µa and σa that satisfy (6.6.6) and (6.6.7). An efficient ‘doubling algorithm’ for calculating these matrices is described in chapter 8. As an application of (6.6.3) – (6.6.5), let us compute the value of a title to one unit of the stream of the j th endowment shock, {djt }∞ t=0 . Let djt = ej xt , where ej is a selection vector that picks off the appropriate linear combination of xt . From (6.5.5) we have that the time zero value of the time t shock djt is ′ ′ d w djt M d xt /µw 0 = xt ej M xt /µ0 .

126

A Competitive Economy

The value of the entire stream is then given by E

∞ X t=0

β t x′t Za xt | J0

where Za = e′j M d /µw 0 . This matches (6.6.3), so that formulas (6.6.5)–(6.6.7) are applicable.

6.7. Term Structure of Interest Rates The value at time zero of a sure claim to one unit of the first consumption good at time zero is evidently given by R10 = βE[¯ e1 · p01 ] | J0 or R10 = β e¯1 · Mc Ao x0 /µw 0.

(6.7.1)

Here R10 is the reciprocal of the gross one-period sure interest rate at time zero. For longer horizons, we have Rj0 = β j E[¯ e1 · p0j ] | J0 , j ≥ 1 or Rj0 = β j e¯1 · Mc Aoj x0 /µw 0.

(6.7.2)

Here Rj0 is the reciprocal of the gross interest factor for a sure claim on the first consumption good j periods into the future at time zero.

Re-opening Markets

127

6.8. Re-opening Markets The competitive equilibrium prices state– and date–contingent commodities that are traded at time zero. After time zero, markets are “closed,” with traders simply executing agreements entered into at time zero. As usual in ArrowDebreu models, markets can be opened in subsequent time periods, but are redundant in the sense that zero trades occur. However, for the purpose of extracting the time series implications of our model, it is useful to compute the prices in such re-opened markets. Suppose that markets re-open at some time t ≥ 1, and that the household and firms re-evaluate their contingency plans at new prices. The household now values consumption services from time t forward. Only goods from time t forward enter the valuations appearing in the budget sets and objective functions of each of our agents. We use L2t as the commodity space, defined as L2t = [{ys }∞ s=t : ys is a random variable in Js for s ≥ t ∞ X and E β s−t ys2 | Jt < +∞] s=t

Expectations conditioned on Jt replace those conditioned on J0 in the intertemporal budget constraint of the household and the cash flow evaluations of the firms. For convenience, we use the j th consumption good at time t as the numeraire. For this choice to deliver a valid numeraire, we invoke Assumption 5.2: The random variable e¯j M c xt differs from zero with probability one. We set the household’s marginal utility of time t wealth, µw t , equal to c th e¯j M xt in order to select the time t , j consumption good as numeraire. With these specifications, we can simply replicate the time zero analysis to obtain equilibrium prices from the vantage point of time t . This yields the following price system: pts = Mc xs /[¯ ej Mc xt ],

s≥t

(6.8.1)

wst =| Sg xs |/[¯ ej Mc xt ], s ≥ t

(6.8.2)

rst = Γ′ Md xs /[¯ ej Mc xt ], s ≥ t

(6.8.3)

qst = Mi xs /[¯ ej Mc xt ],

(6.8.4)

s≥t

128

A Competitive Economy

αst = Md xs /[¯ ej Mc xt ], s ≥ t

(6.8.5)

vt = [Γ′ Md + ∆′k Mk ]xt / [¯ ej Mc xt ]

(6.8.6)

Of particular interest are the spot market prices implied by (6.8.1) – (6.8.6), namely, ptt , wtt , rtt , qtt , αtt .

6.8.1. Recursive price system Prescott and Mehra [1980] and Lucas [1982, JME] extensively utilized recursive formulas expressed in terms of one period forward state contingent claims prices. Counterparts to their recursive pricing formulas are easy to express for our framework. In particular, one-period forward claims on consumption are priced by the function ptt+1 = Mc xt+1 /¯ ej Mc xt . At time t , claims on consumption j -period forward are priced by ej Mc xt . ptt+j = Mc xt+j /¯ Evidently, ptt+j can be built up recursively using the equality ¯j ptt+1 ptt+j = pt+1 t+j e =

Mc xt+j e¯j Mc xt+1

e¯j

Mc xt+1 . e¯j Mc xt

This is a version of a recursive pricing formula often used in formulations of recursive competitive equilibria.

Summary of Pricing Formulas

129

6.8.2. Non-Gaussian asset prices The time t value of a permanent claim to a stream ys = Ua xs , s ≥ t is given by at = (x′t µa xt + σa )/(¯ ej Mc xt ) (6.8.7) where µa and σa satisfy (6.6.6) and (6.6.7) with Za = Ua′ Mc . Notice how (6.8.7) makes the asset price at a nonlinear function of the state vector xt . Suppose, for example, that the wt process is Gaussian. This implies that the equilibrium xt process given by is a multivariate normal process. Even so, the asset prices determined by (6.8.7) are not normally distributed, being determined as the ratio of a quadratic form in the Gaussian state vector xt to a linear form in xt . Thus, the asset prices generated by this “most Gaussian of economies” are highly nonlinear stochastic processes. The term structure of interest rates on perfectly safe claims on the first consumption good j periods ahead is characterized by the gross interest factors Rjt = β j e¯1 · Mc Aoj xt /[¯ ej Mc xt ],

j ≥ 1, t ≥ 0

(6.8.8)

which generalizes (6.7.2).

6.9. Summary of Pricing Formulas For convenience, we now summarize our formulas for the competitive equilibrium price system. They are: (6.58)

ej Mc xt ], s ≥ t pts = Mc xs /[¯

(6.59)

wst =| Sg xs | /[¯ ej Mc xt ], s ≥ t

(6.60)

rst = Γ′ Md xs /[¯ ej Mc xt ], s ≥ t

(6.61)

qst = Mi xs /[¯ ej Mc xt ], s ≥ t

(6.62)

αst = Md xs /[¯ ej Mc xt ], s ≥ t

130

A Competitive Economy

υt = [Γ′ Md + ∆′k Mk ] xt /¯ ej Mc xt

(6.63)

The asset that entitles the owner to the stream of returns yt = Ua xt is priced according to at = (x′t µa xt + σa )/[¯ ej Mc xt ]

(6.64) where (6.51)

µa =

∞ X

β τ (Ao′ )τ Za Aoτ

τ =0

(6.52)

σa =

∞ X β trace Za β τ (Ao )τ CC ′ (Ao′ )τ 1−β τ =0

Za = Ua′ Mc The term structure of interest rates is determined by (6.65)

Rjt = β j e¯1 Mc Aoj xt /[¯ ej Mc xt ],

which gives the price at t of a sure claim on the first consumption good j periods ahead.

6.10. Asset Pricing Example We 3 use the simple pure exchange one good model that is contained in clex14.m to illustrate our asset pricing formulas. The economy in clex14.m is a linearquadratic version of an economy that Lucas (1978) used to develop an equilibrium theory of asset prices. The economy is a member of the special class of structures described in chapter 3. The economy is described as follows:

3 Note to Sargent. This section was 0 in /mnt2/linquad on the SUN. This file is hsch5in.tex. The figures are in hsch*.ps. The MATLAB program used to generate this is hschap5.m. A diary of these runs is in hsch5 in the /linquad directory.

Asset Pricing Example

131

6.10.1. Preferences −.5E

∞ X t=0

β t [(ct − bt )2 + ℓ2t ]|J0

st = ct bt = Ub zt

6.10.2. Technology ct = d1t kt = δk kt−1 + it

gt = φ1 it , d1t = Ud zt 0

φ1 > 0

6.10.3. Information 0 0 1 0 0 zt+1 = 0 .8 0 zt + 1 0 wt+1 0 1 0 0 .5 Ub = [ 30 0 0 ] 5 1 0 Ud = 0 0 0

x0 = [ 5

150

1 0

′

0]

To compute the asset prices in this economy we issue the following MATLAB commands: clex14 solvea t1 = 100; nt = t1; sy=sc; asimul pay=sd(1,:) asseta

132

A Competitive Economy

The program asseta constructs a simulation of length nt of the price and rate of return of an asset that yields a stream of returns equal to pay ∗ xt , where the user specifies the matrix pay . Here we specified that pay = sd(1, :), so that we are pricing a perpetual claim on the endowment process d1t , which is the asset that Lucas priced in his 1978 paper. If the user desires to price a vector of assets, he should simply feed in the matrix pay such that pay ∗ xt is the payout vector of those assets. Let nn be the number of rows of pay , i.e., nn is the number of assets being priced. The program asseta creates a vector y of length nt that equals the vector [mrs, payoff, asset prices, returns], where mrs is the one period intertemporal marginal rate of substitution; payoff is the payoff on the asset(s), which equals pay ∗ xt ; asset prices is the series of asset prices; and ret is the one period gross realized rate of return on the asset(s). For j = 1, 2, 5, the program also creates the reciprocals of the j-period ahead gross rates of return on safe assets, and stores them in the vectors R1, R2, R5. We have computed asset prices for two versions of this economy. The first has the parameter settings listed above, while the second alters the autoregressive parameter in the endowment process to be .4 rather than .8. Figures 6.10.1 through 6.10.3 record the results of one hundred period simulations for each of these two economies. Figure 6.10.1 displays the simulated value of the asset price for the first economy. Figures 6.10.2 displays the gross rates of return on the ‘Lucas tree’ and on a sure one-period bond. We computed the correlation coefficient between these two returns to be -.49. For this economy, the ‘risk premium’ term in the price of the Lucas tree, namely σa in formula (6.8.7), is calculated to be -12.80. To give an idea of how the term structure of interest rates moves in this economy, Figure 6.10.2.b displays the net rates of return on one period and five period sure bonds. (We computed the net rate of return on j -periods bonds by taking the log of the gross rate of return and dividing by j .) Notice the tendency of the term structure to slope upward when rates are low, and to slope downward when rates are high. Figures 6.10.3.a and 6.10.3.b record rates of return for the ‘Lucas tree’ and for sure bonds in the economy with the autoregressive parameter for the endowment process equaling .4. Figure 6.10.3.a shows the gross rates of return on the ‘Lucas tree’ and on a sure one-period bond. The correlation between these two was computed to be -.62. From Figure 6.10.3.b, we see that the tendency for the yield curve to slope upward when rates are low and to slope downward when rates are high has been accentuated relative to our first economy. For the

Asset Pricing Example

133

125 120 115 110 105 100 95 90 85 0

10

20

30

40

50

60

70

80

90

100

Figure 6.10.1: Price of a ‘stock’ entitling the owner to a perpetual claim on the dividends of a ‘Lucas tree’ when the autoregressive parameter for the endowment process is .8.

1.2

0.065

0.06

1.15

0.055 1.1 0.05 1.05 0.045 1 0.04 0.95

0.9 0

0.035

10

20

30

40

50

60

70

80

90

100

Fig. 6.10.2.a. Realized one period gross rates of return on a Lucas tree (solid line) and on a sure one period bond (dotted line) when the autoregressive parameter for the endowment process is .8.

0.03 0

10

20

30

40

50

60

70

80

90

100

Fig. 6.10.2.b. Net rates of return on a one-period (solid line) and a five period (dotted line) when the autoregressive parameter for the endowment process is .8.

134

A Competitive Economy

second economy, the ‘risk premium’ term σa in the price of the Lucas tree is calculated to be -5.90.

1.12

0.09

1.1

0.08 0.07

1.08

0.06 1.06 0.05 1.04 0.04 1.02

0.03

1

0.98 0

0.02

10

20

30

40

50

60

70

80

90

100

Fig. 6.10.3.a. Realized one period gross rates of return on a Lucas tree (solid line) and on a sure one period bond (dotted line) when the autoregressive parameter for the endowment process is .4.

0.01 0

10

20

30

40

50

60

70

80

90

100

Fig. 6.10.2.b. Net rates of return on a one-period (solid line) and a five period (dotted line) when the autoregressive parameter for the endowment process is .4.

The pure exchange economy in clex14.m is one of the simplest to which our asset pricing formulas can be applied. Indeed, for this simple economy, the pricing formulas can be worked out by hand, as the exercises at the end of this chapter indicate. In chapter 4, we shall apply these formulas and our computer programs in much richer contexts in which one can’t get very far ‘by hand’

Exercises

135

6.11. Exercises 1. Consider an economy that consists of technology specification 1 and preference specification 1. The social planning problem is simply to maximize E0 −

∞ X t=0

βt {

1 (ct − bt )2 } 2

subject to ct = dt Assume that bt = ¯b > 0 for all t and that dt = ξ0 + ξ1 dt−1 + εdt , where εdt is √ a white noise with mean zero and variance σε2 , ξ0 > 0, and | ξ1 |< 1/ β . The endowment process {dt } is produced by “trees”, there being one tree for each (representative) household. The household owns the “tree” at the beginning of time (time t = 0). a. Carefully define a competitive equilibrium for this economy. In your definition, describe a particular decentralization scheme, being careful to tell who owns what and who trades with whom. b. Calculate the time-zero equilibrium price system. c. Let υ0 be the time zero value in terms of the time zero consumption good of a title to the entire stream of dividends {dt }∞ t=0 . Prove that υ0 satisfies ∞ ∞ i h X X β t d2t /(b − d0 ) β t ¯b dt − E0 υ 0 = E0 t=0

t=0

d. Compute the gross one period sure rate of interest. e. Compute the gross two period sure rate of interest. 2. Consider an economy with preference specification 1. The technology is specified as ct = dt − Gt where Gt is government purchases. We assume that Gt = UG zt , and dt = Ud zt . Assume that bt = ¯b > 0 for all t . Assume that the government levies state-contingent lump sum taxes τt on the household at time t , where τt = τt (wt , x0 )

136

A Competitive Economy

where wt = (w1 , w2 , . . . , wt ). Lump sum taxes τt are denominated in units of the time t consumption good. a. Formulate the government’s time zero budget constraint. b. Define a competitive equilibrium for this economy. c. Compute a time zero equilibrium price system. d. Define a formula like the one derived under (c) in problem 1 for the time zero value of a title to the dividends from the tree. e. Suppose that the lump sum taxes are on trees, not on the household. Derive a formula for the value of a tree at time zero. f. Compute the gross interest rate on sure one period loans. 3. Consider an economy defined by the social planning problem: maximize the utility of the representative household (1)

−

∞ X 1 E β t (ct − b)2 + ℓ2t | J0 , 0 < β < 1, b > 0 2 t=0

subject to the technology (2)

ct = dt + φgt ,

(3)

gt = ℓt .

φ>0

Here dt is an exogenous process describing the flow of dividends from a single tree (per representative household). The dividend obeys the stochastic process d0 given ,

b > d0 > 0

dt = d0 + wt ,

t≥1

where wt is an independently and identically distributed random process with Ewt = 0 2 Ewt2 = σw .

In (1), b is a constant, ct is consumption at t , and ℓt is labor supplied at t ; E is the mathematical expectation operator, and J0 is information available at

Exercises

137

t = 0, namely, d0 . Equation (2) describes how consumption is related to the exogenous level of dividends at t and the amount φgt = φℓt produced through the application of labor. a. Solve the social planning problem, finding the optimal strategy for consumption and the labor supply. Now decentralize the economy as follows. Let households own the stock of one tree initially. Households sell the tree to a representative firm at time zero (before d0 has been realized). Households sell their labor to the firm each period. The firm buys the tree at the beginning of time zero, hires labor, and sells output to the household. b. State the maximum problems of the representative household and the representative firm for the decentralized economy. c. Find a representation for the time zero Arrow-Debreu price system that supports the solution of the social planning problem as a competitive equilibrium. Give formulas for the price of consumption goods and for the wage of labor. d. Derive a formula for the time zero price of trees in terms of the parameters of preferences, technology, and stochastic process for dividends. (Get as far you can in deriving a closed form). e. Give a formula for the gross interest rate on sure one period loans.

Chapter 7 Applications

7.1. Introduction

7.2. Partial Equilibrium Interpretation The models studied in this book can be reinterpreted as partial equilibrium models which employ the notion of a representative firm, and which generalize the preference and technology specifications of Lucas and Prescott (1971). The idea is that there is a large number of identical firms that produce the same goods and sell them in competitive markets. Because they are all identical, we carry along only one of these firms, and let it produce the entire output in the industry (which is harmless under constant returns to scale). But we have to be careful in our analysis because this representative firm’s decisions play two very different roles: as a stand–in for the ‘average’ competitive producer, and as producer of the entire industry’s output. We make the firm act as a competitor in solving its optimum problem. Demand is governed by the system (11.3.14), with p0t simply being replaced by pt , namely, ct = −Π−1 Λht−1 + Π−1 bt − Π−1 Et {Π′ −1 − Π′ −1 Θ′h

[I − (∆′h − Λ′ Π′ −1 Θ′h )βL−1 ]−1 Λ′ Π′−1 βL−1 }pt

(7.2.1)

ht = ∆h ht−1 + Θh ct

Here ct is a vector of consumption goods. Through this demand system, the representative firm’s output decisions influence the evolution of the market price. However, we want the representative firm to ignore this influence in making its output decisions. A representative firm takes as given and beyond its control the stochastic process {pt }∞ t=0 . The firm sells its output ct in a competitive market each – 139 –

140

Applications

period. Only spot markets convene at each date t ≥ 0. The firm also faces an exogenous process of cost disturbances dt . The firm chooses stochastic processes {ct , gt , it , kt }∞ t=0 to maximize E0

∞ X t=0

β t {pt · ct − gt · gt /2}

subject to Φc ct +Φi it + Φg gt = Γkt−1 + dt kt = ∆k kt−1 + Θk it

(7.2.2)

given k−1 . This problem is not well posed until we describe perceived laws of motion for the processes {pt , dt }∞ t=0 that the firm does not control, but which influence its returns. Specifying the law for the exogenous process {dt } is easy, because the representative firm’s decisions are assumed not to influence it. The situation is different with the price process, because the price is influenced by the output decisions of the representative firm. Despite this influence, we want the firm to behave competitively, that is, to regard the price process as beyond its control. We want to specify the firm’s beliefs about the evolution of the price so that: (a) the firm has ‘rational expectations’, i.e., its beliefs about the evolution of prices allow it to forecast future prices optimally, given the information that it has at each moment; and (b) the firm acts competitively and treats the price process as given and beyond its control. We assume that the firm takes as given a law of motion for spot prices and for the information variables that help to predict spot prices. We model this forecasting problem as follows. The firm observes the state of the market Xt at t , and believes that the law of motion for the spot price is p t = mp X t Xt+1 = ap Xt + Cwt+1

(7.2.3)

′ where Xt = [h′t−1 , Kt−1 , zt′ ]′ , where Kt is the market-wide capital stock, which the firm takes as given and beyond its control. The firm believes that the cost shock process evolves according to dt = Sd Xt . The state for the firm at date t is ′ x ˜t = [Xt′ , kt−1 ]′ .

The firm’s problem is a discounted linear regulator problem. Under our assumptions about the technology, the firm’s control can be taken to be it . The

Partial Equilibrium Interpretation

141

solution of the firm’s problem is a decision rule for investment of the form it = −fi x ˜t .

(7.2.4)

This decision rule and equations (7.2.2) then determine [ct , gt , kt ] as linear functions of x ˜t . The matrix fi in the above equation is a function of all of the matrices describing the firm’s constraints, including ap and mp . The firm’s rule for ct , implied by (7.2.2) and (7.2.4) can be represented as ct = fc x ˜t .

(7.2.5)

Equation (7.2.2) implies that the firm’s capital evolves according to kt = ∆k kt−1 − Θk fi x ˜t .

(7.2.6)

At this point, but not earlier, impose that the ‘representative firm is representative’ by setting kt ≡ Kt in (7.2.6), use it to deduce the actual law of motion for Kt , and then use this to fill in the rows corresponding to Kt of the actual law of motion for Xt : Xt+1 = aa Xt + Cwt+1 . (7.2.7) To get the rows corresonding to ht , use (7.2.5) together with the law of motion ht = ∆h ht−1 + Θh ct . To get a formula for the actual law of motion of the price, use (11.1.1) and the actual law of motion (7.2.7) for xt = Xt to solve for a consumption process. Put the consumption process and preference shock into (11.3.1) and solve for µst . Then solve (11.3.3) forward for µht ; substitute into (11.3.2) to solve for p0t . Set pt = p0t , then express the motion of prices as p t = ma X t .

(7.2.8)

The system (7.2.7), (7.2.8) describes the actual law of motion for spot prices that is induced by the firm’s optimizing behavior and market clearing when the firm’s perceived law of motion for the spot prices is (7.2.3). The firm’s optimization problem and market clearing thus induce a mapping from a perceived law of motion (ap , mp ) for spot prices to an actual law (aa , ma ). Definition: A rational expectations equilibrium (or a partial equilibrium) is a fixed point of the mapping from the perceived law of motion for spot prices to the actual law of motion for spot prices.

142

Applications

An equivalent definition is: Definition: A partial equilibrium is a stochastic process {pt , ct , it , gt , kt , Kt , ht }∞ t=0 , 2 each element of which belongs to L0 , such that: i. Given {pt }∞ t=0 , in particular given the law of motion (7.2.3), {ct , it , ∞ gt , kt }t=0 solve the firm’s problem. ii. {pt , ct , ht }∞ t=0 satisfy the demand system (7.2.1). ∞ iii. {kt }∞ t=0 = {Kt }t=0 .

This is a version of Lucas and Prescott’s (1971) rational expectations competitive equilibrium, which they used to study investment under uncertainty with adjustment costs. The following proposition states the relationship between a partial equilibrium and our earlier notion of competitive equilibrium: Proposition: Let {ct , st , it , gt , kt , p0t , wt0 , αt0 , rt0 , qt0 }∞ t=0 , v0 be a competitive equi0 ∞ librium. Then {pt , ct , it , gt , kt , kt , ht }t=0 is a partial equilibrium. This proposition can be proved directly by verifying that the proposed partial equilibrium satisfies the first order necessary and sufficient conditions for the firm’s problem in the partial equilibrium, and that the proposed {pt , ct , ht }∞ t=0 process satisfies the demand system (11.3.14).

7.2.1. Partial equilibrium investment under uncertainty Our partial equilibrium structure includes many examples of linear rational expectations models (e.g., Sargent (1987, chapter XVI), Eichenbaum (1983), and Hansen and Sargent (1991, Two Difficulties). Here is how we can apply these ideas to a version of Lucas and Prescott’s (1971) model of investment under uncertainty. There is one good produced with one factor of production (capital) via a linear technology. A representative firm maximizes E

∞ X t=0

β t {pt ct − gt2 /2},

subject to the technology ct = γkt−1 kt = δk kt−1 + it gt = f1 it + f2 dt ,

Partial Equilibrium Interpretation

143

where dt is a cost shifter, γ > 0, and f1 > 0 is a cost parameter and f2 = 1. Demand is governed by pt = α0 − α1 ct + ut , where ut is a demand shifter with mean zero and α0 , α1 are positive parameters. Assume that ut , dt are uncorrelated first-order autoregressive processes. Lucas and Prescott computed rational expectations equilibrium quantities by forming a social planning problem with criterion E

∞ X t=0

β

t

Z

0

ct

(α0 − α1 ν + ut )d ν −

.5gt2

,

where the integral under the demand curve is ‘consumer surplus.’ Consumer surplus equals α1 2 (α0 + ut )ct − c . 2 t To map this model into our framework, set Λ = 0, ∆h = 0, Θh = 0, Π2 = α1 , bt = αΠ0 + Π1 ut . Notice that with this specification, (st − bt )2 /2 = (α0 + ut )ct −

α1 2 c + b2t /2. 2 t

The term in b2t can be ignored because it influences no decisions. With this specification, our social planning problem is equivalent with Lucas and Prescott’s. After we have computed the equilibrium quantities by solving the social planning problem, we can compute the ‘marginal utility price’ pt = Π(bt − st )

= α0 + ut − α1 ct ,

where we are using α1 = Π2 .

144

Applications

7.3. Introduction The remainder of chapter provides more examples of models that conform to our framework. Most of these examples were originally stated as partial equilibrium models. The appendix of the chapter describes a scheme for pricing objects that until now were unpriced because they were sheltered from the market by the warmth of the household. We use this decentralization when we want to price some of the household capital stocks.

7.4. A Housing Model Rosen and Topel (1988) formulated a partial equilibrium model of a housing market consisting of a linear demand curve relating a stock of housing inversely to a rental rate; an equilibrium condition relating the price of houses to the discounted present value of rentals, adjusted for depreciation; and a quadratic cost curve for producing houses.

7.4.1. Demand We can capture Rosen and Topel’s specification by sweeping house rentals into the household sector. See the appendix of this chapter for an account of a decentralization that supports this interpretation. Rosen and Topel expressed the demand side of their model in terms of the two equations Rt = bt + αht ∞ X p t = Et (βδh )τ Rt+τ τ =0

where ht is the stock of housing at time t , Rt is the rental rate for housing, pt is the price of new houses, and bt is a demand shifter; α < 0 is a demand parameter, and δh is the depreciation factor for houses. We cast this demand specification within our class of models by letting the stock of houses ht evolve according to ht = δh ht−1 + ct , δh ∈ (0, 1), where ct is the rate of production of new houses. Houses produce services st ¯ t or st = λht−1 + πct , where λ = λδ ¯ h , π = λ. ¯ We can take according to st = λh

A Housing Model

145

¯ 0 = Rt as the rental rate on housing at time t , measured in units of time t λρ t consumption (housing). Demand for housing services is st = bt − µ0 ρ0t , where the price of new houses pt is related to ρ0t by ρ0t = π −1 [pt − βδh Et pt+1 ]. This equation, which is a special case of equation (11.3.7) from chapter 11, imposes the feature of the present specification that δh − λπ −1 θh = 0, is a version of Rosen and Topel’s equation (12). It can be solved to yield pt = ¯ t P∞ (βδh )τ ρ0 , a version of Rosen and Topel’s equation (14). The parameλE t τ =0 ¯ governs the slope of the demand curve for housing, in terms of the rental ter λ rate for housing.

7.4.2. House producers Rosen and Topel’s representative firm maximizes E0

∞ X t=0

β t [pt ct − Ω(ct , ct − ct−1 , et )],

where Ω(ct , ct − ct−1 , et ) is the cost of producing new houses, and {et } is a cost shifter. The function Ω incorporates costs of adjusting the rate of production of new houses. The firm takes the stochastic process for pt as given. Costs are given by Ω(ct , ct − ct−1 , et ) = gt · gt where

g1t = f1 ct + f2 et g2t = f3 (ct − ct−1 ),

where et is our cost-shifter. To map this into our specification, we use the technology f1 ct − g1t = 0kt−1 − f2 et ct − it = 0

f3 ct − g2t = f3 kt−1

kt = 0kt−1 + it .

146

Applications

7.5. Cattle Cycles Rosen, Murphy, and Scheinkman (1994) used a partial equilibrium model to interpret recurrent cycles in U.S. cattle prices. Their model has a static linear demand curve interacting with a ‘time-to-grow’ structure for raising cattle. Let pt be the price of freshly slaughtered beef, mt the feeding cost of preparing an ˜ t the one-period holding cost for a mature animal, γ1 h ˜t animal for slaughter, h ˜ the one-period holding cost for a yearling, and γ0 ht the one period holding cost ˜ t , mt }∞ are exogenous stochastic processes, while the for a calf. The costs {h t=0 stochastic process {pt }∞ is determined by a rational expectations equilibrium. t=0 Let x ˜t be the breeding stock, and y˜t be the total stock of animals. The law of motion for stocks is x ˜t = (1 − δ)˜ xt−1 + g˜ xt−3 − ct , (7.5.1) where ct is a rate of slaughtering. The total head count of cattle is y˜t = x ˜t + g˜ xt−1 + g˜ xt−2 ,

(7.5.2)

which is the sum of adults, calves, and yearlings, respectively. A representative farmer maximizes E0

∞ X t=0

˜ tx ˜ t )(g˜ ˜ t )(g˜ β t {pt ct −h ˜t − (γ0 h xt−1 ) − (γ1 h xt−2 ) − mt ct

(7.5.3)

− Ψ(˜ xt , x ˜t−1 , x ˜t−2 , ct )},

where

ψ3 2 ψ4 2 ψ1 2 ψ2 2 x ˜ + x ˜ + x ˜ + c . (7.5.4) 2 t 2 t−1 2 t−2 2 t The maximization is subject to the law of motion (7.5.1), taking as given the stochastic laws of motion for the exogenous random processes and the equilibrium price process, and the initial state [˜ x−1 , x ˜−2 , x ˜−3 ]. Here (ψj , j = 1, 2, 3) are small positive parameters that represent quadratic costs of carrying stocks, and ψ4 is a small positive parameter. The costs in (7.5.4) are implicitly taken into account by Rosen, Murphy, and Scheinkman, and motivate their decision to “solve stable roots backwards and unstable roots forwards.” To capture Rosen, Murphy, and Scheinkman’s solution, we shall set each of the φj ’s to a positive but very small number. Demand is governed by Ψ=

(5)

ct = α0 − α1 pt + d˜t ,

Cattle Cycles

147

where α0 > 0, α1 > 0, and {d˜t }∞ t=0 is a stochastic process with mean zero representing a demand shifter.

7.5.1. Mapping cattle farms into our framework We show how to map the model of Rosen, Murphy, and Scheinkman into our general setup.

7.5.2. Preferences Set Λ = 0, ∆h = 0, Θh = 0, Π = α1−1 , bt = Πd˜t + Πα0 . With these settings, first-order condition (6.13) for the household’s problem becomes ct = Π−1 bt − Π−2 pt , or ct = α0 − α1 pt + d˜t .

7.5.3. Technology The law of motion for capital is (1 − δ) 0 x ˜t x 0 ˜t−1 = 1 0 1 x ˜t−2 or

1 g x ˜t−1 0x ˜t−2 + 0 it , 0 0 x ˜t−3

kt = ∆k kt−1 + Θh it . Here it = −ct . We use adjustment costs to capture the holding and slaughtering costs. We set ˜ t, g1t = f1 x ˜t + f2 h or ˜ t. g1t = f1 [(1 − δ)˜ xt−2 + g˜ xt−3 − ct ] + f2 h

148

Applications

We set

Notice that

˜t g2t = f3 x ˜t−1 + fr h ˜ t. g3t = f5 x ˜t−1 + f6 h 2 ˜ 2 + 2f1 f2 x ˜t g1t = f12 x ˜2t + f2 h ˜t h t 2 2 2 2 ˜ + 2f3 f4 x ˜t g =f x ˜ + f2 h ˜t−1 h 2t 2 g3t

Thus, we set

=

3 t−1 f52 x ˜2t−2

t

˜ 2 + 2f5 f6 x ˜ t. + f6 h ˜t−2 h t

ψ1 ψ2 ψ3 f22 = f32 = 2 2 2 2f1 f2 = 1 2f3 f4 = γ0 g 2f5 f6 = γ1 g f12 =

To capture the feeding costs we set g4t = f7 ct +f8 mt , and set f72 = ψ24 1. Thus, we set 0 0 0 0 1 1 1 0 0 0 g1t 0 f 1 g 2t 0 ct + 0 it + 0 1 0 0 g 0 0 1 0 3t 0 0 g4t 0 0 0 1 0 −f7 0 0 0 0 ˜ f (1 − δ) 0 gf x ˜t−1 1 1 f2 ht ˜ = f3 0 0 x ˜t−2 + f4 h t . ˜ f6 ht 0 f5 0 x ˜t−3 0

We set dt = Ud zt where

0

0

2f7 f8 =

f8 mt

0 f2 Uh Ud = f4 Uh , f6 Uh f8 Um

˜ t and mt from the exogenous where [Uh , Um ] are selector vectors that pick off h state vector zt . We specify the information matrices [A22 , C2 ] to incorporate ˜ t , mt , d˜t ] consists of three Rosen, Murphy, and Scheinkman’s specification that [h uncorrelated first order autoregressive processes. 1 1 This model is estimated by Anderson, Hansen, McGrattan, and Sargent (1996).

Models of Occupational Choice and Pay

149

7.6. Models of Occupational Choice and Pay Aloyisius Siow (1984) and Sherwin Rosen (1995) and have used pure ‘timeto-build’ structures to represent entry cycles into occupations, and also interoccupational wage movements. It is easiest to incorporate these models into our framework by putting production into the household technology, using the decentralization described in the appendix to generate prices.

7.6.1. A one-occupation model Rosen [1995] studied a partial equilibrium model determining a stock of ‘engineers’ Nt ; the number of new entrants into engineering school, nt ; and the wage level wt of engineers. It takes k periods of schooling to become an engineer. The model consists of the following equations: first, a demand curve for engineers wt = −αd Nt + ǫ1t , αd > 0;

(7.6.1)

second, a time-to-build structure of the education process Nt+k = δN Nt+k−1 + nt , 0 < δN < 1;

(7.6.2)

third, a definition of the discounted present value of each new engineering student ∞ X v t = β k Et (βδN )j wt+k+j ; (7.6.3) j=0

and fourth, a supply curve of new students driven by vt nt = αs vt + ǫ2t , αs > 0.

(7.6.4)

Here {ǫ1t , ǫ2t } are stochastic processes of labor demand and supply shocks. A partial equilibrium is a stochastic process {wt , Nt , vt , nt }∞ t=0 satisfying these four equations, and initial conditions N−1 , n−s , s = 1, . . . , −k .

We can represent this model by sweeping the time-to-build structure and the demand for engineers into the household technology, and putting the supply of new engineers into the technology for producing goods. Here is how. We take

150

Applications

the household technology to be

h1t h2t .. .

st = [λ1 0 . . . 0]

δN 0 . . = . hk,t 0 0 hk+1,t

bt = ǫ1t

h1t−1 h2t−1 .. .

hk+1,t−1

1 0 .. .

0 1 .. .

··· ··· .. .

··· 0

··· 0

0 ···

+ 0 · ct

0 0 .. . 1 0

h 1t−1 h2t−1 .. . hk,t−1

0 0 . + .. ct 0 1 hk+1,t−1

This specification sets Rosen’s Nt = h1t−1 , nt = ct , hs+1,t−1 = nt−s , s = 1, . . . , k , and uses the home-produced good to capture the demand for labor. Here λ1 embodies Rosen’s demand parameter αd . 2 To capture Rosen’s supply curve, we use the physical technology ct = it + d1t ϕ1 it = gt where d1t is proportional to Rosen’s supply shock ǫ2t , and where the adjustment cost parameter ϕ1 varies directly with Rosen’s supply curve parameter αs . Rosen showed that the equilibrium decision role for new entrants (our ct ) must satisfy the condition nt = f1 Et Nt+k + f2 ǫ1t + f3 ǫ2t where f1 < 0.

2 In the definition of Λ in the household technology, we would replace the zeros with ε > 0 as a trick to acquire detectability; see chapter 9 and its appendix for the definition and role of detectability.

Models of Occupational Choice and Pay

151

7.6.2. Skilled and unskilled workers We can generalize the preceding model to two occupations, called skilled and unskilled, to obtain alternative versions of a model estimated by A. Siow (1984). The model consists of the following elements: first, a demand curve for labor

wut wst

= αd

Nut + ǫ1t ; Nst

where αd is a (2 × 2) matrix of demand parameters and ǫ1t is a vector of demand shifters; second, time-to-train specifications for skilled and unskilled labor, respectively: Nst+k = δN Nst+k−1 + nst Nut = δN Nut−1 + nut ; where Nst , Nut are stocks of the two types of labor, and nst , nut are entry rates into the two occupations; third, definitions of discounted present values of new entrants to the skilled and unskilled occupations, respectively: vst = Et β

k

∞ X

(βδN )j wst+k+j

j=0

vut = Et

∞ X

(βδN )j wut+j ,

j=0

where wut , wst are wage rates for the two occupations; and fourth, supply curves for new entrants: nst vut = αs + ǫ2t . (7.6.5) nut vst As an alternative to (7.6.5), Siow simply used the ‘equalizing differences’ condition vut = vst . (7.6.6) We capture this model by pushing most of the ‘action’ into the household sector. Households decide what kind of durable good to accumulate, namely, unskilled labor or skilled labor. Unskilled labor and skilled labor can be combined to produce services, which we specify to generate the demands labor. We let c1t , c2t be rates of entry nut , nst into unskilled and skilled labor, and constrain

152

Applications

them to satisfy c1t + c2t = it + d1t , the rate of total new entrants. To generate the upward sloping supply curves (7.6.5), we specify that φ1 it + φ2 c2t = gt . The technology is thus 1 0 0

1 −1 c1t −φ2 + −φ1 c2t 1 0

d1t 0 0 0 i1t 0 + 1 gt = 0 kt−1 + 0 , i2t 0 0 −1 0

where d1t is a supply shifter. To get Siow’s model, we set φ1 = φ2 = 0, in which case d1t becomes an exogenous supply of new entrants into the labor force. We specify the law of motion for household capital

h1t h2t h3t .. .

hk+1,t hk+2,t

δN 0 0 = . .. 0

0

0 δN 0

0 1 0

0 ··· 0 ··· 1 ···

0 0

0 0

0 ··· 0 ···

h 0 1 1t−1 h2t−1 0 0 0 0 h3t−1 + .. .. . . 1 hk+1,t−1 0 0 0 hk+2,t−1

0 0 0 c1t . .. c2t . 0 1

where h1t−1 = Nut−1 , h2t−1 = Nst , hj+2,t−1 = ns,t−j , j = 1, . . . , k . We generate the demand for labor by specifying services as

s1t s2t

¯ h1t = Λe∆ ¯ h ht−1 + ΛeΘ ¯ h ct =Λ h2t

shock process bt = [ b1t labor.

b2t

0

h1t = eht We set the preference h2t ′ 0 ] to capture the shifters in the demands for

where e is a selector vector that verifies

A Cash-in-Advance Model

153

7.7. A Cash-in-Advance Model We want to use our framework to mimic a situation in which households are in a cash-in-advance environment in which they face a sequence of budget constraints mt mt−1 c˜t + + Bt ≤ yt + Rgt−1 Bt−1 + pt pt mt−1 c˜t ≤ pt Here Rgt−1 is the gross rate of return on indexed bonds Bt−1 held from t − 1 to t ; pt is the price level at t ; c˜t is time t consumption; and mt is currency held from t to t + 1. The household’s preferences P∞ are ordered by E0 t=0 β t u(˜ ct ). (We use c˜t to denote consumption in order to separate this notation from the ct of our framework, which is soon to be defined.) Using the cash-in-advance constraint at equality in the budget constraint gives Bt +

c˜t+1 = yt + Bt−1 Rgt−1 , Rt

pt where Rt = pt+1 is the gross rate of return on currency between t and t+1. The force of the cash-in-advance restriction is that decisions about time t moneyholding influence time t + 1 consumption c˜t+1 , but time t consumption is predetermined.

7.7.1. Reinterpreting the household technology We can specify the household technology to capture key elements of the cash-inpt advance specification. We can use a ‘back-solving’ approach, and let R = pt+1 be a constant rate of return on currency. We shall set c˜t = st and ct = mt /pt , and sweep the cash-in-advance specification into a one-period time delay between a decision to consume (i.e., hold real balances) and when consumption goods are actually enjoyed. Thus, we take the household technology to be st = Rct−1 , which we accomplish by taking Λ = R, Π = 0, δh = 0, θh = 1. When there is inflation, R < 1. When R > 1, there is deflation. Preferences are of the usual kind ∞ X β t [(st − bt ) · (st − bt ) + ℓ2t ]. −.5E0 t=0

154

Applications

With these matchups, the time t ‘seignorage’ component of government revenue is mt − mt−1 r˜t = pt = ct − st . This means that given R , the present value of seignorage revenues can be computed using the methods in chapter 8. With these specifications, an equilibrium with present value government budget balance can be computed, possibly including the inflation rate parameter R along with some of the τ ’s over which we search for an equilibrium. Once an equilibrium is computed, the time series for real balances can be found from mt = ct , pt and the price level and nominal level of currency can be computed using the assumed R .

7.8. Taxation in a Vintage Capital Model Owens (1994) has studied the effects of taxation on prices of new and old commercial buildings. His analysis requires keeping track of the age distribution of capital, which we can accomplish by specifying, for example, k1t 0 0 0 0 k1t−1 1 k2t δh 0 0 0 k2t−1 0 k3t = 0 δh 0 0 k3t−1 + 0 it . k4t

0

0

δh

1

k4t−1

0

Here k1t is new capital, k2t is one year old capital, and so on. We could also include a time-to-build aspect, but have not here. To differentiate among the services produced by capital of different ages, we specify ct = Γc kt−1 , where we make ct a vector that is comparable in dimension with kt−1 and Γc is diagonal. ‘Office services’ are then produced according to st = [ π1

π2

...

πn ] ct .

Decentralizing the Household

155

We can set the π vector to make new office space more desirable than old office space. We let the government tax capital of different ages differently, which puts P P terms like E0 β t τk rt0 · kt or E0 β t τkk p0kt · kt−1 in the budget constraints of households and the government, as in the framework of chapter 14. The tax matrices τk , τkk can be chosen to model different kinds of policies for depreciating capital for tax purposes. This framework can be used to model the effects on the prices of capital of alternative policies for capital taxation.

A. Decentralizing the Household It can be useful to decentralize the household sector in order to price household services and stocks. Suppose that the household buys a vector of services from firms of type III at the price of services ρ0t . The household sells its initial stocks of both physical and household capital and also its labor and endowment process to firms. The price of the initial stock of household capital is v˜0 . The household maximizes ∞ X E0 β t [(st − bt ) · (st − bt ) + ℓ2t } t=0

subject to the budget constraint E0

∞ X t=0

β t ρ0t · st = E0

∞ X t=0

β t (wt0 ℓt + αt0 · dt ) + (v0 · k−1 + v˜0 · h−1 ).

(7.A.1)

Firms of type III Firms of type III purchase the consumption vector ct , rent household capital, and produce and sell household services and additions to the stocks of household capital. Type III firms sell st to households at price ρ0t and new household capital Θh ct to firms of type IV at price p0ht . Firms of type III rent 0 household capital from firms of type IV at a rental price rht , and maximize E0

∞ X t=0

0 β t {ρ0t · st + p0ht Θh ct − rht · ht−1 − p0t · ct }

156

Applications

subject to st = Λht−1 + Πct . Firms of type IV Firms of type IV purchase new household capital from firms of type III, and 0 rent existing household capital to firms of type III at rental price rht . Firms of type IV maximize E0

∞ X t=0

0 β t {rht · ht−1 − p0ht Θh ct } − v˜0 · h−1

subject to ht = ∆h ht−1 + Θh ct . Computing Prices If we formulate the optimum for a firm of type III, obtain the associated first order necessary conditions and rearrange, we get the following restrictions on prices: p0t = Θ′h pht + Π′ ρ0t (7.A.2) 0 = Λ′ ρ0t . rht From the first order conditions for a firm of type IV, obtain 0 p0ht = Et β[∆′h p0ht+1 + rht+1 ].

(7.A.3)

We can use (7.A.2) with (7.A.3) to obtain ∞ X 0 p0t = Θ′h Et [ β (j) ∆′(j−1) rht+j ] + Π′ ρ0t . j=1

This is a generalization of Siow’s equilibrium condition (7.6.6). For us p0t = Mc xt is the vector of shadow prices of new entrants into the two types of profession. We have already shown how to compute the price ρ0t . Indeed, this decentralization is a way to set up an explicit market in the ‘implicit’ services priced by ρ0t . The prices of stocks of household capital can be computed from the multipliers on ht−1 and ct in the social planning problem.

Chapter 8 Efficient Computations

8.1. Introduction This chapter describes fast algorithms for computing the value function and the optimal decision rule of our social planning problem. 1 The decision rule determines the allocation. The value function determines competitive equilibrium prices. The optimal value function and the optimal decision rules can be computed by iterating to convergence on the T operator associated with Bellman’s functional equation. These iterations can be accelerated by using ideas from linear optimal control theory. We avail ourselves of these faster methods because we want to analyze high dimensional systems. This chapter is organized as follows. First, we display a transformation that removes both discounting and cross-products between states and controls. This transformation simplifies the algebra without altering the substance. Next we describe invariant subspace methods for solving an optimal linear regulator problem, which are typically faster than iterating on the Bellman equation. We then describe a closely related method called the doubling algorithm, which ‘skips steps’ in iterating on the Bellman equation. The calculations can be further accelerated by partitioning the state vector to take advantage of the pattern of zeros in A and B . Next we discuss fast methods for computing equilibria for periodic economies. We describe the periodic optimal linear regulator problem, and show how to solve it rapidly. We conclude the chapter by describing how our calculations can be adapted to handle Hansen and Sargent’s (1995) recursive formulation of Jacobsen’s and Whittle’s risk-sensitive preferences, which will be used in Chapter @[email protected] This chapter focuses mostly on nonstochastic optimal linear regulator problems. As indicated in chapter 3, the optimal decision rule for a stochastic optimal 1 Parts of this chapter use results described in Anderson, Hansen, McGrattan, and Sargent (1995). Also see Kwakernaak and Sivan [1972] for what is mostly a treatment of continuous time systems.

– 157 –

158

Efficient Computations

linear regulator problem equals the optimal decision rule for the associated nonstochastic optimal linear regulator problem. Furthermore, from chapter 3, the matrices determining the Lagrange multipliers depend only on the piece of the optimal value function associated with the nonstochastic part of our problem. Throughout this chapter, we study solutions of our control problem that satisfy the additional condition E

∞ X t=0

β t (|xt |2 + |ut |2 ) < ∞,

(8.1.1)

where xt is the state and ut is the control. In an appendix, we describe conditions on the matrices determining returns and the transition law that are sufficient by themselves to imply condition (8.1.1). 2

8.2. The Optimal Linear Regulator Problem Consider the following version of the optimal linear regulator problem: choose a contingency plan for {ut }∞ t=0 to maximize −E subject to

∞ X

β t [x′t Rxt + u′t Qut + 2u′t W xt ], 0 < β < 1

(8.2.1)

t=0

xt+1 = Axt + But + Cwt+1 , t ≥ 0,

(8.2.2)

where x0 is given. In (8.2.1) – (8.2.2), xt is an n × 1 vector of state variables, and ut is a k × 1 vector of control variables. In (8.2.2), we assume that wt+1 is a martingale difference sequence with Ewt wt′ = I , and that C is a matrix conformable as required to x and w . We also impose condition (8.1.1). We temporarily assume that R and Q are positive definite matrices, although in practice we use weaker assumptions about both matrices. A standard way to solve this problem is the method of dynamic programming. Let V (x) be the optimal value associated with the program starting from initial state vector x0 = x . Bellman’s functional equation is n o V (xt ) = max −(x′t Rxt + u′t Qut + 2u′t W xt ) + βEt V (xt+1 ) (8.2.3) ut

2 For conditions sufficient to imply this condition, see Kwakernaak and Sivan [1972], Anderson and Moore [1979], and Anderson, Hansen, McGrattan, and Sargent (1995).

The Optimal Linear Regulator Problem

159

where the maximization is subject to (8.2.2). One way to solve this functional equation is to iterate on a version of (8.2.3), thereby constructing a sequence Vj (xt ) of successively better approximations to V (xt ). In particular, let n o Vj+1 (xt ) = max −(x′t Rxt + u′t Qut + 2u′t W xt ) + βEt Vj (xt+1 ) , ut

(8.2.4)

where again the maximization is subject to (8.2.2). Suppose that we initiate the iterations from V0 (x) = 0 (which is the appropriate terminal value function for a one-period problem). Then direct calculations show that successive iterates on (8.2.4) take the quadratic form Vj (xt ) = −x′t Pj xt − ρj ,

(8.2.5)

where Pj and ρj satisfy the equations Pj+1 = R + βA′ Pj A − (βA′ Pj B + W )

× (Q + βB ′ Pj B)−1 (βB ′ Pj A + W ′ )

ρj+1 = βρj + β trace Pj CC ′ .

(8.2.6)

(8.2.7)

Equation (8.2.6) is the matrix Riccati difference equation. Notice that it involves only {Pj } and is independent of {ρj }. Notice also that C , which multiplies the noises impinging on the system and so determines the variances of innovations to information in the system, affects the {ρj } sequence but not the {Pj } sequence. We can say that {Pj } is independent of the system’s noise statistics. 3 Let P and ρ be the limits of (8.2.6) and (8.2.7), respectively. Then the value function V (xt ) that satisfies the Bellman equation (8.2.3) is given by V (xt ) = −x′t P xt − ρ, where P and ρ are the limit points of iterations on (8.2.6) and (8.2.7) starting from P0 = 0, ρ = 0. The decision rule that attains the right side of (8.2.4) is given by ut = −Fj xt 3 This fact is what permits us to focus on nonstochastic problems in devising our algorithms.

160

Efficient Computations

where Fj = (Q + βB ′ Pj B)−1 (βB ′ Pj A + W ′ ).

(8.2.8)

The optimal decision rule for the original problem is given by ut = −F xt , where F = limj→∞ Fj , or F = (Q + βB ′ P B)−1 (βB ′ P A + W ′ ).

(8.2.9)

According to (8.2.9), the optimum decision rule for ut is independent of the parameters C , and so of the noise statistics. The limit point P of iterations on (8.2.6) evidently satisfies P = R + βA′ P A − (βA′ P B + W )

× (Q + βB ′ P B)−1 (βB ′ P A + W ′ )

(8.2.10)

This equation in P is called the algebraic matrix Riccati equation. One way to solve an optimal linear regulator problem is to iterate directly on (8.2.6) and (8.2.7). Faster algorithms are available. First, we describe a useful transformation that simplifies some of the formulas.

8.3. Transformations to eliminate discounting and crossproducts The following transformation eliminates both discounting and cross-products between states and controls. Define the transformed control vt and transformed state x ˆt by vt = β t/2 (ut + Q−1 W ′ xt ), x ˆt = β t/2 xt . (8.3.1) Notice that vt′ Qvt

=β

t

[ x′t

u′t

W Q−1 W ′ ] W′

It follows that β t [ x′t u′t ]

R W

W′ Q

xt ut

W Q

xt . ut

=x ˆ′t R∗ x ˆt + vt′ Qvt

where R∗ = R − W Q−1 W ′ . The transition law (8.2.2) can be represented as x ˆt+1 = A∗ x ˆ t + B ∗ vt + β

t+1 2

Cwt+1

Stability Conditions

161

where A∗ = β 1/2 (A − BQ−1 W ′ ), B ∗ = β 1/2 B . Therefore, regulator problem (8.2.1) – (8.2.2) is equivalent to the following regulator problem without crossproducts between states and controls and without discounting: choose {vt } to maximize ∞ X −E [ˆ x′t R∗ x ˆt + vt′ Qvt ] (8.3.2) t=0

subject to

x ˆt+1 = A∗ x ˆ t + B ∗ vt + β

t+1 2

Cwt+1 ,

(8.3.3)

where ′

P = R∗ + A∗′ P A∗ − A∗′ P B ∗ (Q + B ∗′ P B ∗ )−1 B ∗ P A∗ F ∗ = (Q + B ∗′ P B ∗ )−1 B ∗′ P A∗ ,

(8.3.4) (8.3.5)

it being understood that P is the positive semi-definite solution of (8.3.4). The optimal closed loop system in terms of transformed variables is x ˆt+1 = (A∗ − B ∗ F ∗ )ˆ xt + β Multiplying both sides of this equation by β −( 1

t+1 2

t+1 2 )

Cwt+1

(8.3.6)

gives

xt+1 = β − 2 (A∗ − B ∗ F ∗ )xt + Cwt+1 .

(8.3.7)

8.4. Stability Conditions We shall typically restrict the undiscounted linear regulator (8.3.2), (8.3.3) defined by the matrices (A∗ , B ∗ , R∗ , Q) to satisfy some conditions from control theory designed to render the problem well behaved. In particular, let DD′ = R∗ , so that D is said to be a factor of R∗ . Our conditions are cast in terms of the concepts of stabilizability and detectability defined in Appendix A. We make Assumption A1: The pair (A∗ , B ∗ ) is stabilizable. The pair (A∗ , D) is detectable. Then there obtains:

162

Efficient Computations

Stability Theorem: Under assumption A1: (i.) starting from any negative semi-definite matrix Po , iterations on the matrix Riccati difference equation converge; and (ii.) The eigenvalues of (A∗ − B ∗ F ∗ ) are stable. In the next section, we describe a class of algorithms that exploit the stabilizing property of the optimal (A∗ − B ∗ F ). 4

8.5. Invariant Subspace Methods Following Vaughan [1970], a literature has developed fast algorithms for computing the limit point of the matrix Riccati equation (8.2.6), based on an eigenstructure of a matrix associated with the Riccati equation. These methods work with a Lagrangian formulation of the problem and with the linear restrictions that stability condition (8.1.1) imposes on the multipliers and the state vector. These conditions restrict the matrix P that solves the algebraic matrix Riccati equation. Without loss of generality, we work with the undiscounted deterministic optimal linear regulator problem: choose {ut }∞ t=0 to maximize −

∞ X t=0

{x′t Rxt + u′t Qut }

(8.5.1)

subject to xt+1 = Axt + But .

(8.5.2)

4 Because the eigenvalues of (A∗ − B ∗ F ∗ ) are less than unity in modulus, it follows that 1 the eigenvalues of Ao = β − 2 (A∗ − B ∗ F ∗ ) are less than √1 in modulus. β

Invariant Subspace Methods

163

8.5.1. P x as Lagrange multiplier It is convenient to write a Lagrangian for the Bellman equation: V (x) = max{−(x′ Rx + u′ Qu + 2µ′ [Ax + Bu − x ˜]) + V (˜ x)}, where x ˜ is next period’s value of the state, µ is a vector of multipliers, and V (x) = −x′ P x where the matrix P solves the matrix Riccati equation. The first-order condition for the above Lagrangian with respect to x ˜ implies that µ = P x . Thus, as usual, the multipliers are linked to the gradient of the value function.

8.5.2. Invariant subspace methods Invariant subspace methods compute P indirectly by vector of the multipliers µ to stabilize the solution for (8.1.1). For now, we assume that A is invertible. We sequences, and let {µt }∞ t=0 be a sequence of vectors of Form the Lagrangian J =−

∞ X t=0

restricting the initial xt , ut , as required by move to the space of Lagrange multipliers.

{x′t Rxt + u′t Qut + 2µ′t+1 [Axt + But − xt+1 ]} − 2µ′0 (¯ x0 − x0 ). (8.5.3)

Here x ¯0 is the given initial level of x0 . First order necessary conditions for the ∞ maximization of J with respect to {ut }∞ t=0 and {xt }t=0 are ut :

Qut + B ′ µt+1 = 0 ,

t≥0

(8.5.4)

xt :

µt = Rxt + A′ µt+1 ,

t ≥ 0.

(8.5.5)

Solve (8.5.4) for ut and substitute into (8.5.2) to obtain xt+1 = Axt − BQ−1 B ′ µt+1 .

(8.5.6)

Represent (8.5.5) and (8.5.6) as L

xt+1 µt+1

=N

xt , µt

(8.5.7)

164

Efficient Computations

where

I L= 0

BQ−1 B ′ , A′

A N= −R

0 . I

Represent (8.5.7) as

or

xt+1 µt+1

xt µt

xt µt

(8.5.8)

xt+1 µt+1

(8.5.9)

= Mf

= Mb

where −1

Mf = L

A + BQ−1 B ′ A′ N= −1 −A′ R

and Mb = N

−1

−1

A−1 L= RA−1

R

−BQ−1 B ′ A′ −1 A′

−1

A−1 BQ−1 B ′ . RA−1 BQ−1 B ′ + A′

,

(8.5.10)

(8.5.11)

Evidently Mb = Mf−1 . The matrices Mf and Mb each have the property that their eigenvalues occur in reciprocal pairs: if λo is an eigenvalue, then so is λ−1 o . We postpone a proof of the ‘reciprocal pairs’ property of the eigenvalues to the subsequent section on the doubling algorithm, where it will follow simply by verifying that Mb and Mf are examples of symplectic matrices. Because its eigenvalues occur in reciprocal pairs, we can represent the matrix Mf in (8.5.8) via a Schur decomposition Mf = V W V −1 ,

(8.5.12)

where V is a nonsingular matrix,

W11 W = 0

W12 , W22

where W11 is a stable matrix, and W22 is an unstable matrix. In terms of ∗ −1 −1 xt , the system can be written transformed variables yt = V yt ≡ V µt ∗ yt+1 = W yt∗ .

(8.5.13)

Invariant Subspace Methods

165

V 11 V 12 , where the partitions conform in size to those of W . Let V = V 21 V 22 The solution of (8.5.13) is −1

yt∗

t W11 = 0

φt t W22

V 11 x0 + V 12 µ0 , V 21 x0 + V 22 µ0

(8.5.14)

j where φ0 = W12 , φj+1 = W11 W12 + φj W22 for j ≥ 0. Because W22 is an unstable matrix, to guarantee that limt→∞ yt∗ = 0, we require that

V 21 x0 + V 22 µ0 = 0,

(8.5.15)

which sets an initial condition that replicates itself over time in the sense that recursions on (8.5.14) imply V 21 xt + V 22 µt = 0,

(8.5.16)

for all t ≥ 0. Equation (8.5.15) implies µ0 = −(V 22 )−1 V 21 x0 . Substituting (8.5.16) into (8.5.13) and using

xt µt

= V yt∗ gives

xt+1 = V11 W11 (V 11 − V 12 (V 22 )−1 V 21 )xt

µt+1 = V21 W11 (V 11 − V 12 (V 22 )−1 V 21 )xt .

(8.5.17)

However, as noted above, µt = P xt , where P solves the algebraic Riccati equation (8.3.4). Therefore, (8.5.17) implies that P V11 = V21 or −1 P = V21 V11 = −(V 22 )−1 V 21 .

Equation (8.5.18) is our formula for P .

(8.5.18)

166

Efficient Computations

8.5.3. Distorted Economies The invariant subspace method can also be applied to compute solutions of distorted economies whose equilibrium conditions can be arranged into the form (8.5.7). Examples of such economies are described in Chapter 15, where equilibrium conditions of the form (8.5.7) cannot be interpreted as the first order conditions of any linear quadratic control problem. For these economies, the matrix Mf in general fails to have eigenvalues in reciprocal pairs. It may or may not be possible to sort the eigenvalues into equal numbers of stable and unstable ones, which are to become the eigenvalues of W11 and W22 , respectively. Whether it is possible becomes a check for the existence and uniqueness of a stable solution of the model. The condition that there exist a unique solution of (8.5.8) with |xt |2 < ∞ is that there exists a Schur decomposition (8.5.12) of Mf in which half the eigenvalues of Mf are stable, and the other half are unstable. An excess of stable eigenvalues indicates nonuniqueness; an excess of unstable eigenvalues indicates nonexistence of a stable solution. Where a unique solution exists, it can be computed using formula (8.5.18). 5

8.5.4. Transition Dynamics Invariant subspace algorithms can be adapted to solve models in which elements of the matrices determining preferences, technologies, information, and government policies vary deterministically over time, before some date T1 , after which they are constant. The procedure is to use the algorithm (8.5.18) to solve the model for t ≥ T1 , then to work backwards for earlier dates. We want to compute an equilibrium in which the L, N matrices are timevarying in a simple deterministic way, say, due to once and for all changes in tax rates at some date t = T1 > 0. Suppose that we want to solve xt xt+1 N =L , t ≥ T1 (8.5.19) µt µt+1 and

xt xt+1 ˜ ˜ N =L , µt µt+1

0 ≤ t < T1 ,

(8.5.20)

˜, L ˜ are the ‘temporary’ versions of the matrices whose ‘permanent’ where N values are L, N . 5 See Blanchard and Kahn (1981) and Whiteman (1983).

Invariant Subspace Methods

167

For t ≥ T1 , we use the solution of the ‘permanent’ system, with V 21 xt + V 22 µt = 0.

(8.5.21)

In particular, (8.5.21) implies that G1T1 xT1 + G2T1 µT1 = 0, where G1T1 = V 21 , G2T1 = V 22 . We know that xT1 V11 V12 W11 = µT1 V21 V22 0

W12 W22

V 11 xT1 −1 + V 12 µT1 −1 . V 21 xT1 −1 + V 22 µT1 −1

(8.5.22)

(8.5.23)

We want to impose restriction (8.5.22) on (8.5.23) and use it to solve for xT1 −1 as a linear function of µT1 −1 . A couple of lines of algebra leads to the restriction [G1T1 V11 + G2T1 V21 ][W11 V 11 + W12 V 21 ] + [G1T1 V12 + G2T1 V22 ]W22 V 21 xT1 −1 o n + [G1T1 V11 + G2T1 V21 ][W11 V 12 + W12 V 22 ] + [G1T1 V12 + G2T1 V22 ]W22 V 22 µT1 −1 ≡ G1,T1 −1 xT1 −1 + G2,T1 −1 µT1 −1 = 0.

(8.5.24)

This equation can be written as G1T1 −1 xT1 −1 + G2T1 −1 µT1 −1 = 0,

(8.5.25)

and it can be solved for µT1 −1 as a linear function of xT1 −1 . Equations (8.5.24) and (8.5.25) define (G1t , G2t ) as a function of (G1t+1 , G2t+1 ). We use (8.5.25) to ‘backdate’ the Git , i = 1, 2, matrices, and iterate back to t = 0. These calculations will produce time-varying versions of all of our equilibrium matrices Ao , C, Sc , Mc , . . . for t = 0, 1, . . . , T1 described in chapters 4 and 6.

168

Efficient Computations

8.6. The Doubling Algorithm The algebraic matrix Riccati equation can be solved with a doubling algorithm. 6 The algorithm shares with invariant subspace methods the prominent role it assigns to the matrix Mb of equation (8.5.9). We start with a finite horizon version of our problem for horizon t = 0, . . . , τ − 1, which leads to a two point boundary problem. We continue to assume that A is nonsingular, iterate on (8.5.8), and impose the boundary condition µτ = 0 to get xτ x0 ˆ M = , (8.6.1) 0 µ0 where ˆ = M −τ = M τ . M (8.6.2) b f We want to solve (8.6.2) for µ0 as a function of x0 , and from this solution ˆ conformably with deduce a finite-horizon approximation to P . Partitioning M ˆ 11 xτ = x0 , M ˆ 21 xτ = µ0 . Therefore, the state-co-state partition, we deduce M −1 ˆ ˆ we choose µ0 = M21 (M11 ) x0 , and set the matrix ˆ 21 (M ˆ 11 )−1 . P =M

(8.6.3)

ˆ for large horizon τ , then use (8.6.3) to The plan is efficiently to compute M compute P . We can accelerate the computations by choosing τ to be a power of two and using k k k+1 (8.6.4) = (Mf−2 )Mf−2 . Mf−2 ˆ = M −τ can be computed in j iterations Thus, for τ = 2j , the matrix M f instead of 2j iterations, inspiring the name doubling algorithm. Because Mf−1 has unstable eigenvalues, direct iterations on (8.6.4) can be unreliable. Therefore, the doubling algorithm transforms iterations on (8.6.4) into other iterations whose important objects converge. These iterations exploit the fact that the matrix Mf is symplectic (see Appendix B). The eigenvalues of symplectic matrices come in reciprocal pairs. The product of symplectic matrices is symplectic; for any symplectic matrix S , the matrices S21 (S11 )−1 and (S11 )−1 S12 are both symmetric; and ′ −1 S22 = (S11 ) + S21 (S11 )−1 S12 ′ −1 = (S11 ) + S21 (S11 )−1 S11 (S11 )−1 S12 .

6 This section is based on Anderson, Hansen, McGrattan, and Sargent (1995). For another discussion of the doubling algorithm, see Anderson and Moore [1979, pp. 158–160].

Partitioning the State Vector

169

Therefore, a (2n × 2n) symplectic matrix can be represented in terms of three (n × n) matrices α = (S11 )−1 , β = (S11 )−1 S12 , γ = S21 (S11 )−1 , the latter two matrices being symmetric. These properties of symplectic matrices inspire the following parameterizak tion of Mf−2 αk−1 αk−1 βk = , γk αk−1 αk′ + γk αk−1 βk where the n × n matrices αk , βk , γk satisfy the recursions k Mf−2

(8.6.5)

αk+1 = αk (I + βk γk )−1 αk

βk+1 = βk + αk (I + βk γk )−1 βk αk′ γk+1 = γk +

αk′ γk (I

−1

+ βk γk )

(8.6.6)

αk .

To initialize, we use representation (8.5.11) for Mb = Mf−1 to induce the settings: α0 = A, γ0 = R, β0 = BQ−1 B ′ . Anderson, Hansen, McGrattan, and Sargent (1996) describe a version of the doubling algorithm modified to build in a positive definite terminal value matrix Po . Their scheme initializes iterations on (8.6.6) as follows: α0 = (I + BQ−1 B ′ Po )−1 A β0 = (I + BQ−1 B ′ Po )−1 BQ−1 B ′ ′

−1

γ0 = R − Po + A Po (I + BQ

′

(8.6.7) −1

B Po )

A.

The modified algorithm then works as follows: 1. Initialize α0 , β0 , γ0 according to (8.6.7). 2. Iterate on (8.6.6). 3. Form P as the limit of γk + Po .

We have assumed that A is nonsingular, but Anderson (1985) argues that the doubling algorithm is applicable also in circumstances where A is singular. 7 Anderson, Hansen, McGrattan, and Sargent (1996) report the results of computations in which the doubling algorithm is among the fastest and most reliable available algorithms for solving several example economies.

7 See Anderson, Hansen, McGrattan, and Sargent (1996) for conditions under which the matrix sequences {αk }, {βk }, {γk } converge.

170

Efficient Computations

8.7. Partitioning the State Vector Undiscounted versions of the control problem solved by our social planner assume a form for which it is natural to partition the state vector to take advantage of the pattern of zeros in A and B . This leads to a control problem of the form: choose {ut }∞ t=0 to maximize −

′ ∞ X x1t R11 { x2t R21 t=0

x1t+1 x2t+1

R12 R22

x1t + u′t Qut } x2t

(8.7.1)

x1t B1 + ut , x2t 0

(8.7.2)

subject to

=

A11 0

A12 A22

with [x′10 , x′20 ]′ given. 8 For this problem, the operator associated with Bellman’s equation is T (P ) = R + A′ P A − A′ P B(Q + B ′ P B)−1 B ′ P A. (8.7.3) x1t Partitioning P and T (P ) conformably with the partition makes the x2t (1, 1) and (1, 2) components of T (P ) satisfy T11 (P11 ) = R11 + A′11 P11 A11 − A′11 P11 B1 (Q + B1′ P11 B1 )−1 B1′ P11 A11 (8.7.4) T12 (P11 ,P12 ) = R12 + A′11 P11 A12 − A′11 P11 B1 (Q + B1′ P11 B1 )−1 B1′ P11 A12

+

[A′11

−

A′11 P11 B1 (Q

+

(8.7.5)

B1′ P11 B1 )−1 B1′ ]P12 A22

Equation (8.7.4) shows that T11 depends on P11 , but not on other elements of the partition of P . From (8.7.5), T12 depends on P11 and P12 , but not on P22 . Because T maps symmetric matrices into symmetric matrices, the (2, 1) 8 System ( 8.7.1 ) – ( 8.7.2 ) is called a controllability canonical form (see Kwakernaak and Sivan [1972]). Two things distinguish a controllability canonical form: (1) the pattern of zeros in the pair (A, B, ) and (2) a requirement that (A11 , B1 ) be a controllable pair (see Appendix A of this chapter). A controllability canonical form adopts a description of the state vector that separates it into a part x2t that cannot be affected by the controls, and a part x1t that can be controlled in the sense that there exists a sequence of controls {ut } that sends x1 to any arbitrarily specified point within the space in which x1 lives.

Partitioning the State Vector

171

block of T is just the transpose of the (1, 2) block. Finally, the (2, 2) block of T depends on P11 , P12 , and P22 . Partition the optimal feedback matrix F = [F1 F2 ], where the partition is conformable with that of xt . Then the optimal control is ut = [F1 F2 ]

x1t . x2t

f f f Let P11 be the fixed point of (8.7.4) and let P12 be the fixed point of T12 (P11 , P12 ). Then F1 and F2 are given by f f F1 = (Q + B1′ P11 B1 )−1 B1′ P11 A11

(8.7.6)

f f f F2 = (Q + B1′ P11 B1 )−1 (B1′ P11 P12 + B1′ P12 A22 )

(8.7.7)

f f Equation (8.7.6) shows that F1 depends only on P11 , while F2 depends on P11 f f and P12 , but not on P22 , the fixed point of T22 . We aim to compute [F1 , F2 ] and the multipliers described in chapter 3, f f which turn out only to depend on P11 and P12 . We can compute these objects rapidly by using the structure exposed by (8.7.4) and (8.7.5). First, note that the T11 operator identified by (8.7.4) is formally equivalent with the T operator of (8.7.3), except that (1, 1) subscripts appear on A and R , and a (1) subscript appears on B . Thus, the T11 operator is simply the operator whose iterations define the matrix Riccati difference equation for the small optimal regulator f problem determined by the matrixes (A11 , B1 , Q, R11 ). We can compute P11 by using any of the algorithms described above for this smaller problem. We have chosen to use the doubling algorithm (8.6.6). f Second, given a fixed point P11 of T11 , we apply another sort of doubling f algorithm to compute the fixed point of T12 (P11 , ·). This mapping has the form f T12 (P11 , P12 ) = D + G′ P12 H

(8.7.8)

f f f f where D = R12 +A′11 P11 A12 −A′11 P11 B1 (Q+B1′ P11 B1 )−1 B1′ P11 A12 , G = [A11 − ′ f −1 ′ f B1 (Q + B1 P11 B1 ) B1 P11 A11 ], H = A22 . Notice that G = A11 − B1 F1 , where F1 is computed from (8.7.6). When x2t is set to zero for all t , the law of motion for x1t under the optimal control is thus given by

x1t+1 = Gx1t .

172

Efficient Computations

For problems for which condition (8.1.1) is either automatically satisfied or else imposed, the eigenvalues of G and H each have absolute values strictly less than unity. That the eigenvalues of G and H are both less than unity assures the existence of a limit point to iterations on (8.7.8). The limit point satisfies the Sylvester equation P12 = D + G′ P12 H, (8.7.9) which is to be solved for P12 . The limit point of iterations on T12 initiated from P12 (0) = 0 can be represented f P12

=

∞ X

G′j DH j ,

(8.7.10)

j=0

f , ·) can be verified directly. However, whose status as a fixed point of T12 (P11 iterations on (8.7.9) would not be an efficient way to compute P12 . Instead, we recommend using this doubling algorithm. Compute the following objects recursively: Gj = Gj−1 Gj−1

Hj = Hj−1 Hj−1 P12,j = P12,j−1 +

(8.7.11)

G′j−1 P12,j−1 Hj−1

where we set P12,0 = D, G0 = G, H0 = H . By repeated substitution it can be shown that j 2X −1 P12,j = Gi′ DH i . (8.7.12) i=0

Each iteration doubles the number of terms in the sum. 9 , 10

9 This algorithm is implemented in the MATLAB program double2j.m. 10 The (1, 2) partition of P is simply P f ′ . We could derive an algorithm similar to ( 8.7.11 ) 12 f

f

to compute P22 , but we don’t need to compute P22 , which is used to compute neither [F1 F2 ] nor the Lagrange multipliers that determine the price system associated with our equilibrium.

A Periodic Doubling Algorithm

173

8.8. The Periodic Optimal Linear Regulator In chapter 17, we study a class of models of seasonality whose social planning problems form a periodic optimal linear regulator problem: choose {ut }∞ t=0 to maximize ∞ X {x′s(t) Rs(t) xt + u′t Qs(t) ut } (8.8.1) t=0

subject to

xt+1 = As(t) xt + Bs(t) ut .

(8.8.2)

Here s(t) is a periodic function that maps the integers into a subset of the integers: s : (· · · − 1, 0, 1, · · ·) → [1, 2, · · · , p] s(t + p) = s(t) for all t.

In problem (8.8.1) - (8.8.2), the matrices As , Bs , Qs , and Rs that define the linear regulator problem are each periodic with common period p . Associated with problem (8.8.1) – (8.8.2) is the following version of the matrix Riccati difference equation: Pt = Rs(t) + A′s(t) Pt+1 As(t) (8.8.3) ′ ′ − A′s(t) Pt+1 Bs(t) (Qs(t) + Bs(t) Pt+1 Bs(t) )−1 Bs(t) Pt+1 As(t) .

Under conditions that generalize assumption A1, which were discussed by Richard Todd [1983], iterations on (8.8.3) yield p convergent subsequences, whose limit points we denote P1 , P2 , . . . , Pp . The optimal decision rule in period t is ut = −Fs(t) xt ,

(8.8.4)

where ′ ′ Fs(t) = −(Qs(t) + Bs(t) Ps(t+1) Bs(t) )−1 Bs(t) Ps(t+1) , As(t) .

(8.8.5)

Thus, the optimal decision rules themselves have period p . One way to compute the optimal decision rules is to iterate on (8.8.3) to convergence of the p subsequences, and then to use (8.8.5). Faster algorithms can be obtained by adapting calculations described earlier in this chapter. In the next section, we show how doubling algorithms apply to the periodic linear regulator problem, and also how the ‘controllability canonical form’ can be exploited.

174

Efficient Computations

8.9. A Periodic Doubling Algorithm First-order conditions for the periodic linear regulator can be represented as

xt+1 µt+1

= Mf,s(t)

xt , µt

(8.9.1)

where Mf,s(t) is the periodic counterpart to the matrix Mf defined in (8.5.10). Iterating this equation p times and using the periodic structure of s(t) gives xt+p xt = Γp , (8.9.2) µt+p µt where Γp ≡ Mf,p−1 Mf,p−2 · · · Mf,1 Mf,p .

(8.9.3)

The matrix Γp is the product of p symplectic matrices, and therefore is symplectic. Equation (8.9.2) at t = p can be represented x2p xp Γ−1 = , (8.9.4) p µ2p µp where −1 −1 −1 Γ−1 p = Mf,p Mf,1 · · · Mf,p−1 .

(8.9.5)

Iterating (8.9.4) τ − 1 ≥ 1 times and imposing the same boundary condition used in (8.6.1) gives ˆ xpτ = xp , (8.9.6) M µp 0 ˆ = Γ−τ . An argument used earlier implies that the doubling algorithm where M p ˆ to compute can be applied to our redefined M ˆ 21 (M ˆ 11 )−1 . Pp = M

(8.9.7)

It is straightforward to compute the remaining p−1 value functions. Notice that (8.9.4) implies ˆ xpτ = Mf,p−1 xp−1 , M 0 µp−1 or −1 ˆ xpτ = xp−1 . Mf,p−1 M 0 µp−1

A Periodic Doubling Algorithm

175

The same argument used above now implies that ˜ 21 (M ˜ 22 )−1 x1 ≡ P1 x1 , µ1 = M ˜ =M ˆ p−1 ≡ M −1 M ˆ is symplectic because it is the product of two where M f,p−1 symplectic matrices. The product of two symplectic matrices Z1 , Z2 has representation −1 α ˜ α ˜ −1 β˜ Z1 Z2 = Z¯ = γ˜ α ˜ −1 α ˜ 1 + γ˜ α ˜ −1 β˜ where

α ˜ = α2 (I + β1 γ2 )−1 α1 γ˜ = γ1 + α1′ γ2 (I + β1 γ2 )−1 α1 β˜ = β2 + α2 (I + β1 γ2 )−1 β1 α′ .

(8.9.8)

2

We can use this feature to compute Pp−1 from the γ term produced by this representation of multiplication. Iterating this argument leads us to compute Pp−2 , . . . , P1 as the correˆ p−2 = sponding γ matrices in the successive multiplications used to form M −1 −1 ˆ p−1 , . . . , M ˆ1 = M M ˆ2 . Mf,p−2 M f,1 Thus, the algorithm works as follows. 1. Initialize α0 , β0 , γ0 according to (8.6.7). 2. Use the algorithm (8.9.8) for multiplying symplectic matrices to form Γ−1 p defined as in (8.9.5). 3. Iterate on (8.6.6). 4. Form Pp as the limit of γk + Po . ˆ p−1 , M ˆ p−2 , . . . , M ˆ 1 using (8.9.8), and set the corre5. Successively form M sponding γ terms to Pp−1 , Pp−2 , . . . , P1 . Having computed P1 , . . . , Pp , we can use (8.8.5) to compute the optimal decision rules. The optimal feedback laws are periodic, so that ut = −Fs(t) xt . The matrices F1 , . . . Fp are computed from Fj = (Qj + Bj ′ Pj+1 Bj )−1 Bj ′ Pj+1 Aj , where it is understood that Pp+1 = P1 .

176

Efficient Computations

8.9.1. Partitioning the state vector We can also apply the partitioning technique to the periodic optimal linear regulator i problem h x1t exactly as in order to accelerate the computations. We partition the state vector into x2t above. With the appropriate specification of Rs , Qs , As , and Bs , we obtain a periodic version of the T11 (P11 ) mapping described in equation ( 8.7.4 ). Use our procedures to compute f P1 , P2 , . . . , Pp as described above, then set P11,j = Pj for j = 1, . . . , p . The T12 mapping for the periodic model becomes f

T12,k (P11,k+1 , P12,k+1 ) = Dk + Gk ′ P12,k+1 Hk

(8.9.9)

where f

f

Dk = R12,k + A′11,k P11,k+1 A12,k − A′11,k P11,k+1 B1k f

′ P ′ P × (Qk + B1k B )−1 B1k 11,k+1 A12,k 11,k+1 1k f

′ P ′ P Gk = [A11,k − B1k (Qk + B1k B )−1 B1k 11,k+1 A11,k 11,k+1 1k

(8.9.10)

Hk = A22,k f

In ( 8.9.9 ) – ( 8.9.10 ), P11k+1 is the fixed point for P11,k+1 corresponding to period k + 1 . Iterations on ( 8.9.9 ) will give rise to a sequence consisting of p convergent subsequences, f f whose limit points we call P12,1 , . . . , P12,p . We desire to compute these limiting matrices. f We begin by creating an operator T¯12,1 whose fixed point is P12,1 . We define

¯1 ¯1 + G ¯ ′ P12,1 H T¯12,1 (P12,1 ) = D 1

(8.9.11)

′ ′ ′ ¯ ¯′ ¯ 1 = D1 +G′ D2 H1 +· · ·+G′ G′ · · · G′ where D 1 1 2 p−1 Dp Hp−1 Hp−2 · · · H1 G1 = G1 G2 · · · Gp H1 = Hp Hp−1 · · · H1 . We can compute the fixed point of ( 8.9.11 ) by using the standard doubling algorithm that is described in section blank and that is implemented in the MATLAB program double2j.m. f f Once we have computed P12,1 , we can compute P12,j for j = p, p − 1, . . . , 2 by using f

f

f

f

P12,p = Dp + G′p P12,1 H1 P12,j = Dj + G′j P12,j+1 Hj

(8.9.12) , j = p − 1, p − 2, · · · , 2

The optimal feedback laws ut ≡ −Fs(t) , xt can be computed as follows. Let Fs(t) = [F1s(t) F2s(t) ] , where the partition of Fs(t) matches that of the state vector into x1 (t), x2 (t) . Then we have f

f

′ P −1 B ′ P F1j = (Qj + B1j 1j 11,j+1 A11,j 11,j+1 B1j ) f

f

f

′ −1 (B ′ P ′ P F2j = (Qj + B1j 1j 11,j+1 A12,j + B1j P12,j+1 A22,j ) 11,j+1 B1j )

for j = 1, . . . , p.

(8.9.13)

Linear Exponential Quadratic Gaussian Control

177

The optimal closed loop system is then xt+1 = (As(t) − Bs(t) Fs(t) )xt .

(8.9.14)

8.10. Linear Exponential Quadratic Gaussian Control In chapter 16, we shall reinterpret some of our economies in terms of risksensitive control theory. In this section, we describe how to adapt the preceding computational strategies to handle versions of the ‘risk-sensitivity corrections’ of Jacobsen (1973, 1977) and Whittle (1990). We use Hansen and Sargent’s (1995) method of implementing discounting. The resulting specification preserves the computational ease of the original linear quadratic specification, while relaxing ‘certainty equivalence.’ Let Vt1 (xt1 ) = −x′t1 Pt1 xt1 − ηt1 . Let β ∈ (0, 1) and 1 consider the sequence {Vt (xt )}tt=t of value functions generated by the following 0 constrained optimization problems: o n σ 2 Vt (xt ) = max −(x′t Rxt + u′t Qut ) + β log Et exp Vt+1 (xt+1 ) ut σ 2

(8.10.1)

subject to xt+1 = Axt + But + Cwt+1 ,

(8.10.2)

where wt+1 is an (N × 1) martingale difference sequence with Gaussian density f (wt+1 ) =

1 1 ′ exp{− wt+1 Σ−1 wt+1 }. 2 (2π)N/2 |Σ|1/2

(8.10.3)

Usually, we shall set the covariance matrix Σ = Ewt wt′ = I . We momentarily retain the more general notation in order to state a useful lemma in greater generality. In solving this discounted linear exponential quadratic Gaussian (LEQG) control problem, we use the following lemma due to Jacobson (1973). Lemma (Jacobson): Let wt+1 ∼ N (0, Σ) and xt+1 = Axt + But + Cwt+1 . Suppose that the matrix (Σ−1 − σC ′ Pt+1 C) is positive definite. Then Et exp{

σ ′ x Pt+1 xt+1 } = 2 t+1

178

Efficient Computations

Z

∞

−∞

where

1 1 ′ σ exp{− wt+1 Σ−1 wt+1 } exp{ x′t+1 Pt+1 xt+1 } 2 2 (2π)N/2 |Σ|1/2

(8.10.4)

nσ o = k exp (Axt + But )′ P˜t+1 (Axt + But ) 2 P˜t+1 = Pt+1 + σPt+1 C(Σ−1 − σC ′ Pt+1 C)−1 C ′ Pt+1

k=

s

| (Σ−1 − σC ′ Pt+1 C)−1 | . |Σ|

(8.10.5)

(8.10.6)

This concludes the statement of the lemma. Let Vt+1 (xt+1 ) = −x′t+1 Pt+1 xt+1 − ηt+1 , and apply the lemma to evaluate the term inside the braces on the right side of (8.10.1): nσ o 2 log Et exp [x′t+1 Pt+1 xt + ηt+1 ] σ 2 ′ ′ = x Rxt + u Qut + β(Axt + But )′ P˜t+1 (Axt + But )

x′t Rxt +u′t Qut + β t

(8.10.7)

t

+ constant where P˜t+1 is given by equation (8.10.5). Maximizing the right hand side of (8.10.7) with respect to ut gives the linear decision rule ut = −Ft xt , where Ft is determined by the recursions: P˜t+1 = Pt+1 + σPt+1 C(Σ−1 − σC ′ Pt+1 C)−1 C ′ Pt+1 Ft = {Q + βB ′ P˜t+1 B}−1 βB ′ P˜t+1 A

Pt = R + βA′ P˜t+1 A − β 2 A′ P˜t+1 B(Q + βB ′ P˜t+1 B)−1 B ′ P˜t+1 A.

(8.10.8) (8.10.9) (8.10.10)

Notice that in the special case that σ = 0, these equations are versions of the Riccati difference equation and the associated decision rule. Notice also that when σ 6= 0, equations (8.10.8), (8.10.9), and (8.10.10) imply that the decision rules Ft depend on the innovation variances of the exogenous processes (note the appearance of C in (8.10.8)).

Linear Exponential Quadratic Gaussian Control

179

We can obtain a more compact version of these recursions as follows. Apply the matrix identity (a − b d−1 c)−1 = a−1 + a−1 b(d − ca−1 b)−1 ca−1 to (8.10.10) using the settings a−1 = β P˜t+1 , b = −B, d = Q, c = B ′ to obtain β P˜t+1 − β P˜t+1 B(B ′ (β P˜t+1 ) B + Q)−1 B ′ (β P˜t+1 ) 1 = ( P˜t+1 + BQ−1 B ′ )−1 . β Substituting into the right side of (8.10.10) gives 1 Pt = R + A′ ( P˜t+1 + BQ−1 B ′ )−1 A . β

(8.10.11)

Now apply the same matrix identity to the right side of (8.10.8) to obtain −1 P˜t+1 = (Pt+1 − σCΣC ′ )−1 .

(8.10.12)

Substituting (8.10.12) into (8.10.11) gives the version −1 Pt = R + A′ (β −1 Pt+1 + BQ−1 B ′ − σβ −1 CΣC ′ )−1 A.

(8.10.13)

Collecting results, we have that the solution of the problem can be represented via the recursions (8.10.13), (8.116), (8.10.9). We are interested in problems for which recursions on these equations converge as t → −∞ . In situations in which convergence prevails, we can avail ourselves of a doubling algorithm to accelerate the computations.

8.10.1. Doubling algorithm It suffices to consider the undiscounted (β = 1) version of our problem, because we can transform a discounted problem into an undiscounted one. Represent the Riccati equation (8.10.13) in the form (see Appendix C) −1 Pt = R + A′ (Pt+1 + J)−1 A

where J = BQ−1 B ′ − σCΣC ′ . The doubling algorithm applies with −1 A A−1 J Mf−1 = Mb = , RA−1 A′ + RA−1 J

(8.10.14)

180

Efficient Computations

and with the settings α0 = A, γ0 = R, β0 = J . To compute the solution with terminal value matrix Po , use the initializations α0 = (I + JPo )−1 A, β0 = (I + JPo )−1 J, γ0 = −Po + R + A′ Po (I + JPo )−1 A . The algorithm then works as follows. 1. Initialize α0 , β0 , γ0 according to the formulas just given. 2. Iterate on (8.6.6). 3. Form P as the limit of γk + Po .

A. Concepts of Linear Control Theory Assume in the deterministic linear regulator (8.5.1)–(8.5.2) that matrix R is positive semi-definite and that Q is positive definite. Sufficient conditions for existence and stability of a solution of the deterministic linear regulator are typically stated in terms defined in the following four definitions. Definition: The pair (A, B) is stabilizable if y ′ B = 0 and y ′ A = λy ′ for some complex number λ and some complex vector y implies that |λ| < 1 or y = 0. Definition: The pair (A, B) is controllable if y ′ B = 0 and y ′ A = λy ′ for some complex number λ and some complex vector y implies that y = 0. Definition: The pair (A, D) is detectable if D′ y = 0 and Ay = λy for some complex number λ and some complex vector y implies that |λ| < 1 or y = 0. Definition: The pair (A, D) is observable if D′ y = 0 and Ay = λy for some complex number λ and some complex vector y implies y = 0. Stabilizability and controllability evidently form a pair of concepts, with controllability implying stabilizability, but not vice versa (i.e., controllability is a more restrictive assumption. Similarly, detectability and observability form a pair of concepts, with observability implying detectability, but not vice versa. Stabilizability is equivalent with existence of a time-invariant control law that stabilizes the state vector. Controllability implies that there exists a sequence of controls that can attain an arbitrary value for the state vector starting from any initial state vector, within n periods, where n is the dimension of the state. When (A, B) is controllable, the entire state vector is ‘endogenous,’ in the sense of being potentially ‘under control.’

Concepts of Linear Control Theory

181

The concepts of detectability and observability are applied to the pair of matrices (A, D), where DD′ = R (i.e., D is a factor of R ). Assume (a.) that the pair (A, B) is stabilizable, which implies that it is feasible to stabilize the state vector; and that (b.) the pair (A, D) is detectable, which means that it is desirable to stabilize the state vector. Together, assumptions (a.) and (b.) imply that the optimal control stabilizes the state vector. When R is nonsingular, the pair (A, D) is observable, and the value function is strictly concave.

182

Efficient Computations

B. Symplectic Matrices We now define symplectic matrices 11 and state some of their properties. Definition: A (2n × 2n) matrix Z is said to be symplectic if Z ′ JZ = J , where 0 −In . J= In 0 The following properties of symplectic matrices follow directly from the definition of a symplectic matrix: Property 1: If the matrix Z is symplectic, then so is any positive integer power of Z . Property 2: If Z1 and Z2 are both (2n × 2n) symplectic matrices, then their product Z1 Z2 is also symplectic. Property 3: If a symplectic matrix Z is written in partitioned form Z=

Z11 Z21

Z12 , Z22

−1 and if Z11 exists, then −1 ′ −1 ) + Z21 Z11 Z12 Z22 = (Z11

Property 4: The eigenvalues of any symplectic matrix Z occur in reciprocal pairs, i.e., if λi is an eigenvalue of a symplectic matrix Z , then so is λ−1 i . To establish property 4, that from the definition that any symplectic matrix Z satisfies Z −1 = J −1 Z ′ J . Since Z −1 and Z ′ are thus related by a similarity transformation, they have common eigenvalues. This implies that the eigenvalues of Z must occur in reciprocal pairs. −1 Property 3 means that if Z11 exists, then a symplectic matrix Z can be represented in the form α−1 Z= γα−1

α−1 β ′ α + γα−1 β

(8.B.1)

11 See Anderson and Moore [1979, pp. 160–161] and also Anderson, Hansen, McGrattan, and Sargent (1996).

Alternative forms of Riccati equation

183

Let Zj , for j = 1, 2, be two symplectic matrices, each represented in the form (8.B.1): −1 αj αj−1 βj Zj = . (8.B.2) γj αj−1 αj′ + γj αj−1 βj

It can be verified directly that the product Z1 Z2 = Z¯ has the same form, namely, −1 α ˜ α ˜ −1 β˜ Z1 Z2 = Z¯ = . (8.B.3) γ˜ α ˜ −1 α ˜ ′ + γ˜ α ˜ −1 β˜ where

α ˜ = α2 (I + β1 γ2 )−1 α1 γ˜ = γ1 + α1′ γ2 (I + β1 γ2 )−1 α1 β˜ = β2 + α2 (I + β1 γ2 )−1 β1 α′

(8.B.4)

2

This algorithm is implemented in our MATLAB program mult.m.

C. Alternative forms of Riccati equation It is useful to display alternative forms of the Riccati equation P = R + A′ P A − A′ P B(Q + B ′ P B)−1 B ′ P A.

(8.C.1)

We first apply the following matrix identity from Noble and Daniel [1977, p. 29]. Assume that d−1 and a−1 exist. Then (a − bd−1 c)−1 = a−1 + a−1 b[d − ca−1 b]−1 ca−1 . Apply this identity, setting a−1 = Pt+1 , b = −B, d = Q, c = B ′ to obtain −1 (Pt+1 + BQ−1 B ′ )−1 = Pt+1 − Pt+1 B(B ′ Pt+1 B + Q)−1 B ′ Pt+1 .

Substituting the above identity into (8.C.1) establishes −1 Pt = R + A′ (Pt+1 + BQ−1 B ′ )−1 A.

Now write (8.C.2) as −1 −1 Pt = R + A′ Pt+1 Pt+1 (Pt+1 + BQ−1 B ′ )−1 A −1 Pt = R + A′ Pt+1 (Pt+1 Pt+1 + BQ−1 B ′ Pt+1 )−1 A

Pt = R + A′ Pt+1 (I + BQ−1 B ′ Pt+1 )−1 A

(8.C.2)

Assume that A−1 exists, and write the preceding equation as Pt = R + A′ Pt+1 (A−1 + A−1 BQ−1 B ′ Pt+1 )−1 Pt = A′ Pt+1 (A−1 + A−1 BQ−1 B ′ Pt+1 )−1 + R(A−1 + A−1 BQ−1 B ′ Pt+1 )(A−1 + A−1 BQ−1 B ′ Pt+1 )−1 This equation can be represented as Pt = {RA−1 + [A′ + RA−1 BQ−1 B ′ ]Pt+1 } {A−1 + A−1 BQ−1 B ′ Pt+1 }−1 .

(8.C.3)

Equation (8.C.3) takes the form Pt = {C + DPt+1 } × {E + F Pt+1 }−1 where

(8.C.4)

C = RA−1 D = A′ + RA−1 BQ−1 B ′ E = A−1 F = A−1 BQ−1 B ′ ,

which can be represented as

Xt Yt

=

E C

F D

Xt+1 , Yt+1

where Pt = Yt Xt−1 . Notice that

E C

F D

= Mf−1 = Mb ,

and that limt→−∞ Pt = limt→−∞ Yt Xt−1 can be computed as the limit of the j γj term in the representation of the symplectic matrix Mf−2 .

Part II Representations and Properties

Chapter 9 Representation and Estimation This chapter shows how our models restrict moments of observed prices and quantities, and how observations can be used to make inferences about the parameters of our models. Earlier chapters have prepared a state-space representation that expresses states xt and observables yt as linear functions of an initial x0 and histories of martingale difference sequences wt . The wt ’s are shocks to endowments and preferences whose histories are observed by the agents in the economy. The econometrician does not see those shocks, at least directly. Therefore, to prepare a model for estimation we obtain another representation of it that is cast in terms of shocks that can in principle be recovered from an econometrician’s observations. By using the Kalman filter we shall obtain what is known as an ‘innovations representation’. It is a workhorse. It can be transformed to yield a Wold representation or a vector autoregression for observables. 1 An important approach to estimation, approximation, and aggregation over time is to deduce the restrictions that the models of the economy and of data collection impose on the innovations representation. The Kalman filter does this efficiently, and enables a recursive way of calculating a Gaussian likelihood function. We describe how to obtain maximum likelihood and generalized method of moments estimates of a model’s parameters, using both time domain and frequency domain methods. As by-products of time domain estimation, we deduce autoregressive and Wold representations for observables. As a by-product of frequency domain estimation, we recover a theory of specification error. We also study the effects of aggregation over time, and how to estimate a model specified at a finer time interval than pertains to the available data. These methods must be augmented to incorporate data on asset prices, which are non-linear functions of the state of the economy. The last part of the chapter describes how asset prices, returns, and other nonlinear functions of the state can be used in estimation. The Kalman filter is intimately connected to the optimal linear regulator (i.e., the linear-quadratic dynamic programming problem). Remarkably, the key 1 See Sims (1972, 1980), Whittle (1983), and Sargent (1987) for definitions and discussions of the Wold and autoregressive representations.

– 187 –

188

Representation and Estimation

mathematical formula associated with the Kalman filter is the same matrix Riccati equation that solves the linear regulator. Furthermore, the same key ‘factorization identity’ occurring with the Kalman filter plays a role in linear-quadratic optimization theory. In chapters 11 and 14, we shall use a ‘factorization identity’ to provide information about alternative representations of preferences.

9.1. The Kalman Filter We regard a vector of time t data yt as error-ridden measures of linear combinations of the state vector xt . We append a measurement equation to the equilibrium law of motion of the state to attain the following state space system: xt+1 = Ao xt + Cwt+1 (9.1.1) yt = Gxt + vt where vt is a martingale difference sequence of measurement errors that satisfies Evt vt′ = R, Ewt+1 vs′ = 0 for all t + 1 ≥ s . Here G is a matrix whose rows are composed of entries of the Sj and Mj matrices, computed for example in chapters 4, 6, and 7, that select those components of quantities and prices for which data are available. 2 We assume that x0 is a random vector with known mean x ˆ0 and covariance matrix E(x0 − x ˆ0 )(x0 − x ˆ0 )′ = Σ0 . Using (9.1.1), we have Ey0 = Gˆ x0 . t . For any variable We adopt the notation y0t = [yt , yt−1 , . . . , y0 ], y t = y−∞ t−1 ˆ t |y , x ˆ zt , t ≥ 0, we let zˆt = E[z ˆ ], where E(·) is the linear least squares 0 0 ˆ ˆ t |y t , x projection operator. Also, we occasionally use the notation Et zt = E[z 0 ˆ0 ]. We want recursive formulas for yˆt , x ˆt . We begin by constructing an innovation process {at } such that [at0 , x ˆ0 ] forms an orthogonal basis for the information ˆ0 ]. We recursively calculate the projections x ˆt+1 and yˆt by regressing set [y0t , x t ˆ0 ]. on the orthogonal basis [a0 , x ˆ0 ]is constructed using a Gram-Schmidt proThe orthogonal basis for [y0t , x cess. Begin with the regression equation y0 = Ey0 + a0 = Gˆ x0 + a0 or a0 = y0 − Gˆ x0 , 2 Later we shall permit serially correlated measurement errors. It is easy to modify the calculations to permit Ewt+1 vt′ to be nonzero.

The Kalman Filter

189

where the residual a0 satisfies the least squares normal equation Ea′0 = 0. Evidently, [y0 , x ˆ0 ] and [a0 , x ˆ0 ] span the same linear space. Next, form a1 as the residual from a regression of y1 on [y0 , x ˆ0 ], or equivalently, a regression of ˆ y1 on [a0 , x ˆ0 ]: a1 = y1 − E[y1 | y0 , x ˆ0 ] or ˆ 1 | a0 , x a1 = y1 − E[y ˆ0 ]; a1 is by construction orthogonal to a0 and x ˆ0 ; i.e., E(a1 a′0 ) = 0, E(a1 ) = 0. ˆ t | y t−1 , x ˆ t | at−1 , x Continuing in this way, form at = yt − E[y ˆ0 ] = yt − E[y ˆ0 ], 0 0 where E(at a′s ) = 0 for s = 0, . . . , t−1 and E(at ) = 0. We call at the innovation in yt . It is useful to represent at as follows. From the second equation of (9.1.1) and from the fact that vt is orthogonal to yt−s and xt−s for s ≥ 1, it follows that yˆt = Gˆ xt and that yt = Gˆ xt + G(xt − x ˆ t ) + vt . By subtracting the first equation from the second, we find that the innovation at in yt satisfies at ≡ yt − yˆt = G(xt − x ˆ t ) + vt . (9.1.2) Calculate the second moment matrix of at to be Eat a′t = GE(xt − x ˆt )(xt − x ˆt )′ G′ + Evt vt′ = GΣt G′ + R ≡ Ωt

where Σt ≡ E(xt − x ˆt )(xt − x ˆt )′ . We shall soon give a recursive formula for Σt . From the first equation in (9.1.1), it follows that ˆt xt+1 = Ao E ˆt xt = Ao E ˆt−1 xt + Ao (E ˆt xt − E ˆt−1 xt ), E

(9.1.3)

ˆt denotes projection on [y t , x ˆ where again E 0 ˆ0 ]. Express the projection Et xt = Pt ˆt xt +ψt , ψt is a least squares residual vector, and Ext + j=0 Γj aj where xt = E the regression coefficients Γj are determined by the least squares orthogonality conditions Eψt a′s = 0 for s = 0, . . . , t. Because [at0 , x ˆ0 ] is an orthogonal basis t for [y0 , x ˆ0 ], these orthogonality conditions imply (Ext a′t )(Ωt )−1 = Γt ,

(9.1.4)

190

Representation and Estimation

ˆt−1 xt = Ext + where Eat a′t = Ωt . To compute Ext a′t , first notice that E Pt−1 ˆ j=0 Γj aj . Then xt = Et−1 xt + Γt at + ψt can be interpreted in terms of the regression equation ˆt−1 xt ) = Γt at + ψt , (xt − E (9.1.5) ˆ t−E ˆt−1 xt )|at ]. Evidently, E(xt − E ˆt−1 xt )a′t = Ext a′t . Use where Γt at = E[(x ′ ′ ˆ (9.1.2) to compute E(xt − Et−1 xt )at = Σt G . It follows that (9.1.4) becomes Γt = Σt G′ (GΣt G′ + R)−1 ,

(9.1.6)

ˆt xt = E ˆt−1 xt + Γt at . E

(9.1.7)

and from (9.1.5) that Substituting (9.1.7) into (9.1.3) gives x ˆt+1 = Ao x ˆt + Ao Γt (yt − Gˆ xt ) or x ˆt+1 = Ao x ˆt + Kt at ,

(9.1.8)

where at = yt − Gˆ xt , and where from (9.1.6) Kt must satisfy Kt = Ao Σt G′ (GΣt G′ + R)−1 .

(9.1.9)

Equation (9.1.9) expresses the ‘Kalman gain’ Kt in terms of the state covariance matrix Σt = E(xt − x ˆt )(xt − x ˆt )′ . We need an equation for Σt . Subtract x ˆt+1 = Ao x ˆt + Kt (yt − Gˆ xt ) from o the first equation of (9.1.1) to obtain xt+1 − x ˆt+1 = (A − Kt G)(xt − x ˆt ) + Cwt+1 − Kt vt . Multiply each side of this equation by its own transpose and take expectations to obtain Σt+1 = (Ao − Kt G)Σt (Ao − Kt G)′ + CC ′ + Kt RKt′ .

(9.1.10)

Substituting (9.1.9) into (9.1.10) and rearranging gives a matrix Riccati difference equation for Σt : Σt+1 = Ao Σt Ao′ + CC ′ − Ao Σt G′ (GΣt G′ + R)−1 GΣt Ao′ .

(9.1.11)

The recursive (9.1.9) and (9.1.11) for Σt , Kt determine the Kalman filter. They are to be initialized from a given Σ0 . Later we discuss alternative ways to choose Σ0 .

Innovations Representation

191

9.2. Innovations Representation The Kalman filter lets us associate with representation (9.1.1) an ‘innovations representation’: x ˆt+1 = Ao x ˆt + Kt at (9.2.1) yt = Gˆ xt + at , where Eat a′t ≡ Ωt = GΣt G′ + R . This time varying representation is obtained starting from arbitrary initial conditions x ˆ0 , Σ0 . We can rearrange (9.2.1) into the form of a whitening filter at = yt − Gˆ xt

x ˆt+1 = Ao x ˆt + Kt at ,

(9.2.2)

which can be used for recursively constructing a record of innovations {at }Tt=0 from an x ˆ0 and a record of observations {yt }Tt=0 . The filter defined by (9.2.2) is called a “whitening filter” because it accepts as “input” the serially correlated process {yt } and produces as “output” the serially uncorrelated (i.e., “white”) vector stochastic process {at }. The process {at } is said to be a fundamental white noise for the {yt } process, because it equals the one-step ahead prediction error in a linear least squares projection of yt on the history of y . 3 Later, we shall use the whitening filter in several ways. We shall use it to study how the innovations {at } from a population vector autoregression for {yt } are related to the {yt } process and to the underlying martingale process {wt } of information flowing to agents. We shall also use it to construct a recursive representation of a Gaussian likelihood function for a sample drawn from the {yt } process.

3 See Sims (1972), Hansen and Sargent (1991, chapter 2), and Sargent (1987, chapter XI) for the role such an error process plays in the construction of Wold’s representation theorem.

192

Representation and Estimation

9.3. Convergence results For the purpose of obtaining a time-invariant counterpart to (9.2.1), we introduce two assumptions. Assumption A1: The pair (Ao′ , G′ ) is stabilizable. Assumption A2: The pair (Ao′ , C) is detectable. See the appendix to chapter 8 for definitions of stabilizability and detectability. Assumptions A1 and A2 are typically met for our applications. Under A1 and A2, two useful results occur. The first is that iterations on the matrix Riccati difference equation (9.1.11) converge as t → ∞ , starting from any positive semi-definite initial matrix Σ0 . The limiting matrix Σ∞ ≡ limt→∞ Σt is the unique positive semi-definite matrix Σ that satisfies the algebraic matrix Riccati equation 4 Σ = Ao ΣAo′ + CC ′ (9.3.1) − Ao ΣG′ (GΣG′ + R)−1 GΣAo′ . If we initiate the Kalman filter by choosing Σ0 = Σ∞ , then from (9.1.11) and (9.1.9), we obtain a time invariant Kt matrix, call it K . Under this circumstance, representation (9.2.1) becomes time invariant. The stationary covariance matrix of the innovations is given by Ω = Eat a′t = GΣG′ + R , where Σ = Σ∞ = Σ0 . The second useful result is that Assumptions A1 and A2 imply that Ao − KG is a stable matrix, i.e., its eigenvalues are strictly less than unity in modulus. Later we shall see how the stability of the matrix Ao − KG plays a key role in a convenient formula for the autoregressive representation for the {yt } process.

4 The limiting form of ( 9.1.10 ) is evidently a discrete Lyapunov or Sylvester equation. See chapter 8.

Serially Correlated Measurement Errors

193

9.3.1. Time-Invariant Innovations Representation The infinite-horizon time invariant Kalman filter defines a matrix valued function, which we express as h i K, Σ = kfilter (Ao , G, V1 , V2 , V3 ) (9.3.2)

where V1 = CC ′ , V2 = Evt vt′ , V3 = Ewt+1 vt′ , Σ = Et−1 (xt − x ˆt )(xt − x ˆt )′ . For our model, we can use (9.3.2) with the following settings for the matrices V1 , V2 , V3 : V1 = CC ′ , V2 = R, V3 = a matrix of zeros conformable to x and y. 5 By using the function kfilter, we can evidently associate with representation (9.1.1) a time-invariant innovations representation (9.2.1) in which Kt is constant.

9.4. Serially Correlated Measurement Errors It is useful to adapt the preceding calculations to cover the case in which the measurement errors vt in (9.1.1) are serially correlated. 6 Modify (9.1.1) to be xt+1 = Ao xt + Cwt+1 yt = Gxt + vt

(9.4.1)

vt = Dvt−1 + ηt where D is a matrix whose eigenvalues are strictly below unity in modulus and ηt is a martingale difference sequence that satisfies Eηt ηt′ = R Ewt+1 ηs′ = 0 for all t and s. In (9.4.1), vt is a serially correlated measurement error process that is orthogonal to the xt process. Define the quasi-differenced process y t ≡ yt+1 − Dyt .

(9.4.2)

5 The function kfilter defined in ( 9.3.2 ) solves a version of ( 9.1.9 ) and ( 9.1.11 ) for Σ∞ and K∞ , a version that has been generalized to permit arbitrary covariance between wt+1 and vt , which is required for several of our applications. 6 The calculations in this section imitate those of Anderson and Moore [1979].

194

Representation and Estimation

From (9.4.1) and the definition (9.4.2) it follows that y t = (GAo − DG)xt + GCwt+1 + ηt+1 Thus, (xt , y t ) is governed by the state space system xt+1 = Ao xt + Cwt+1 y t = Gxt + GCwt+1 + ηt+1

(9.4.3)

where G = GAo − DG . This state space system has nonzero covariance between the state noise Cwt+1 and the “measurement noise” (GCwt+1 + ηt+1 ). Define the covariance matrices V1 = CC ′ , V2 = GCC ′ G′ + R, V3 = CC ′ G′ . By applying the Kalman filter to (9.4.3), we obtain a gain sequence Kt with which to construct the associated innovations representation x ˆt+1 = Ao x ˆ t + Kt u t y t = Gˆ xt + ut

(9.4.4)

′ ˆ t | y t−1 , x ˆ t | y t−1 , x where x ˆt = E[x ˆ0 ], ut = y t − E[y ˆ0 ], Ω1 ≡ Eut u′t = GΣt G +V2 . 0 0 Using definition (9.4.2), it follows that [y0t+1 , x ˆ0 ] and [y t0 , x ˆ0 ] span the same t t ˆ ˆ space, so that x ˆt = E[xt | y0 , x ˆ0 ], ut = yt+1 − E[yt+1 | y0 , x ˆ0 ]. Thus, ut is the innovation in yt+1 .

9.5. Combined System It is useful to have a formula that gives a state space representation for yt driven by the innovations to yt . We obtain this by combining the innovations system (9.4.4) for y t with the system yt+1 = Dyt + y t .

(9.5.1)

The system (9.5.1) accepts {y t } as an “input” and produces {yt } as an “output”. The two systems (9.4.4) and (9.5.1) can be combined in a series to give the state space system: o x ˆt Kt x ˆt+1 A 0 + ut = G D yt I yt+1 (9.5.2) x ˆt yt = [ 0 I ] + [0]ut yt

Recursive Formulation of Likelihood Function

195

The MATLAB program evardec.m uses the time-invariant version of (9.5.2), obtained using kfilter.m, to obtain a decomposition of the j -step ahead prediction error variance associated with the Wold representation for yt . 7

9.6. Recursive Formulation of Likelihood Function The Kalman filter enables a recursive algorithm for computing a Gaussian likelihood function for a sample of observations {ys }Ts=0 on a (p × 1) vector yt . We assume that these data are governed by the innovations representation (9.2.1) The likelihood function of {ys }Ts=0 is defined as the density f (yT , yT −1 , . . . , y0 ). It is convenient to factor the likelihood function f (yT , yT −1 , . . . , y0 ) = fT (yT |yT −1 , . . . , y0 )fT −1 (yT −1 |yT −2 , . . . , y0 ) · · · f1 (y1 |y0 )f0 (y0 ).

(9.6.1)

The Gaussian likelihood function for an n × 1 random vector y with mean µ and covariance matrix V is 1 1 N (µ, V ) = (2π)−n/2 |V |− 2 exp − (y − µ)′ V −1 (y − µ) . (9.6.2) 2 Evidently, from (9.1.1), the distribution f0 (y0 ) is N (Gx0 , Ω0 ), where Ωt = GΣt G′ +R and Σt is the covariance matrix of xt around x ˆt . Further, f (yt |yt−1 , . . . , y0 ) = N (Gˆ xt , Ωt ). It is easy to verify that the distribution gt (at ) of the innovation at is N (0, Ωt ) Thus, f0 (y0 ) equals g0 (a0 ), the distribution of the initial innovation. More generally, from (9.2.1), the conditional density ft (yt |yt−1 , . . . , y0 ) equals the density gt (at ) of at . Then the likelihood (9.6.1) can be represented as gT (aT )gT −1 (aT −1 ) . . . g1 (a1 )g0 (a0 ). (9.6.3) Expression (9.6.3) implies that the logarithm of the likelihood function for y0T is T X −.5 {p ln(2π) + ln |Ωt | + a′t Ω−1 (9.6.4) t at }. t=0

7 The MATLAB program series.m can be used to obtain the time-invariant system ( 9.5.2 ) from the two systems ( 9.4.4 ) and ( 9.5.1 ).

196

Representation and Estimation

9.6.1. Initialization Two alternative sets of assumptions are commonly used to initiate the Kalman filter, corresponding to different information about y0 . (a.) The distribution of the initial y0 is treated as if it were conditioned on an infinite history of y ’s. This idea is implemented by specifying that x0 has mean x ˆ0 = E[x0 |y−1 , y−2 , . . .], and a covariance matrix Σ0 = Σ∞ coming from the steady state of the Kalman filter. In this case, the time-invariant Kalman filter can be used to construct the Gaussian log likelihood: −.5

T X t=0

{p ln(2π) + ln |Ω| + a′t Ω−1 at },

(9.6.5)

where Ω = GΣ∞ G′ + R , and where the innovations at are computed using the steady state Kalman gain K . This procedure amounts to replacing f0 (y0 ) in −1 (9.6.1) with f (y0 |y−∞ ). (b.) The initial value y0 is drawn from the stationary distribution of y , meaning that it is associated with an x0 governed by the stationary distribution of xt , an assumption implemented by initiating the Kalman filter with Σ0 = Σx , where Σx is the asymptotic stationary covariance matrix of x . Assumptions (a) and (b) pertain to how one selects the matrix Σ0 . Under each of assumptions (a) and (b), it is common to set x ˆ0 equal to the unconditional mean of x , provided that this exists.

9.6.2. Non-existence of a stationary distribution Approach (b) assumes that the law of motion xt+1 = Ao xt + Cwt+1 is such that the {xt } process has an asymptotic stationary distribution, and cannot be used without modification in models that violate this assumption. When an asymptotic stationary distribution doesn’t exist, one procedure is to assume a ‘diffuse’ initial distribution over the piece of x0 that has no stationary distribution. The models described in chapter 13, with their co-integrated equilibrium consumption processes, necessitate such a procedure. In the appendix, we describe a method for coping with this situation, inspired by ideas of Kohn and Ansley (19XXX). It is most useful for us to describe

Recursive Formulation of Likelihood Function

197

their idea in the context of models with serially correlated measurement errors, which we treat in the next section.

9.6.3. Serially correlated measurement errors When we use the state space model with serially correlated measurement errors (9.4.1), some adjustments are called for in the above procedures for forming the log likelihood. These adjustments are occasioned by the timings in the definitions of x ˆt , ut . In particular, the notation now t denotes x ˆt = E[xt |y ] and Σt = E(xt − x ˆt )(xt − x ˆt )′ . These changes mean that the distribution gt−1 (ut−1 ) equals ft (yt |yt−1 , . . . , y0 ). So corresponding to the factorization (9.6.1) we have gT −1 (uT −1 )gT −2 (uT −2 ) . . . g0 (u0 )g−1 (u−1 ).

(9.6.6)

To deduce the appropriate distribution of y0 , or equivalently, of u−1 , notice that the time 0 version of the ‘whitener’ is u−1 = y0 − Dy−1 − Gˆ x−1 x ˆ0 = Ao x ˆ−1 + K0 u−1 ,

where K0 is the time 0 value for the Kalman gain. It is natural to start the system with y−1 = GEx and x ˆ−1 = Ex , where Ex is the stationary mean of xt , 8 and to initiate the Kalman filter from the mean of the stationary distribution of x . So the Gaussian log likelihood function is −.5

T −1 X

t=−1

{p ln(2π) + ln |Ωt | + u′t Ω−1 t ut }.

(9.6.7)

We now indicate how these procedures can be adapted to handle models for which no stationary distribution for xt exists, following procedures of Kohn and Ansley (BLANK). The idea is to factor the likelihood function as f (yT , yT −1 , . . . , y0 ) = fT (yT |yT −1 , . . . , y0 )fT −1 (yT −1 |yT −2 , . . . , y0 ) · · · fm (ym |ym−1 , . . . , y0 )f (ym−1 , . . . , y0 ).

8 Notice that G and not G appears in the equation for the unconditional mean.

(9.6.8)

198

Representation and Estimation

Kohn and Ansley assign a ‘diffuse prior’ to that subset of the state vector that does not possess a stationary distribution, and let the remaining piece of x0 be distributed according to its stationary distribution. This specification embodies an ‘improper prior’ distribution for (ym−1 , . . . , y0 ). Under this specification, we use the first m observations of yt to estimate xm−1 , then form x ˆm−1 , Σm−1 from which to initiate the Kalman filter for the system (9.4.3) with serially correlated measurement errors. The Kalman filter is applied to compute the likelihood for the sample {ys }Ts=m . In addition, we can adjust (9.6.8) to account for the first m observations. Details are given in the appendix.

9.7. Wold Representation For the purpose of describing the relationship of the time-invariant innovations representation to the Wold and autoregressive representations, we shall avail ourselves when needed of: Assumption A3: The eigenvalues of Ao are all less than unity in modulus, except possibly for one associated with a constant. A Wold representation for a stationary stochastic process yt is a moving average of the form ∞ X yt = Ey + Γj ǫt−j , j=0

ˆ t |y t−1 ], and P∞ traceΓj Γ′ < +∞ . (Below, we shall where ǫt = yt − E[y j j=0 for the most part set the unconditional mean vector Ey to zero, to conserve on notation.) We can attain a Wold representation by manipulating the innovations system in a way that amounts to driving the date for the initial x ˆ0 arbitrarily far into the past. Thus, the first equation of (9.4.4) can be solved recursively for t X x ˆt+1 = (Ao )j Kut−j + (Ao )t+1 x ˆ0 . j=0

Now assume that x ˆ0 was itself formed by having observed the history y −1 , so that x ˆ0 = (I − Ao L)−1 Ku−1 + µx ,

Vector Autoregression for {yt }

199

where µx is the unconditional mean of x . Under this specification for x ˆ0 , x ˆt+1 = (I − Ao L)−1 Kut + µx .

(9.7.1)

Below, we shall omit the unconditional mean term, by assuming that µx = 0. To get a Wold representation for yt , substitute (9.4.2) into (9.4.4) to obtain x ˆt+1 = Ao x ˆt + Kut yt+1 − Dyt = Gˆ xt + ut .

(9.7.2)

Then (9.7.2) and (9.7.1) can be used to get a Wold representation for yt : yt+1 = [I − DL]−1 [I + G(I − Ao L)−1 KL]ut ,

(9.7.3)

where again L is the lag operator. Also, from (9.7.2) a “whitening filter” for obtaining {ut } from {yt } is given by ut = yt+1 − Dyt − Gˆ xt

x ˆt+1 = Ao x ˆt + Kut .

(9.7.4)

9.8. Vector Autoregression for {yt } We can use the innovations representation and some results from linear algebra to derive a convenient formula for the one-step-ahead linear least squares forecast of yt based on the history of the {yt } process. We begin by deriving a version of the factorization identity, which asserts equality between two representations of the spectral density matrix of the observables. We will encounter a mathematically equivalent form of this identity in Chapter 11 when we discuss observationally equivalent representations of preferences.

200

Representation and Estimation

9.8.1. The factorization identity For the model with serially uncorrelated measurement errors, we have two alternative representations for an observed process {yt }, the original state space representation (9.1.1) and the innovations representation (9.2.1). Because they describe the same stochastic process {yt }, they give two alternative representations of the spectral density matrix of {yt }, an outcome that expresses the factorization identity. The original state space representation is xt+1 = Ao xt + Cwt+1 yt = Gxt + vt ,

(9.8.1)

where wt+1 is a martingale difference sequence of innovations to agents’ information sets, and vt is another martingale difference sequence of measurement errors. We assumed that wt+1 , vt are mutually orthogonal stochastic processes. The first line of representation (9.8.1) can be written L−1 xt = (I−Ao L)−1 Cwt+1 or xt = (L−1 − Ao )−1 Cwt+1 . It follows that the covariance generating function of {xt } satisfies Sx (z) = (zI − Ao )−1 CC ′ (z −1 I − (Ao )′ )−1 . Using this expression and the second line of (9.1.1), together with the observation that vt is orthogonal to the process xt , shows that the covariance generating function of yt is given by Sy (z) = G(zI − Ao )−1 CC ′ (z −1 I − (Ao )′ )−1 G′ + R.

(9.8.2)

Representation (9.2.1) implies x ˆt = (L−1 − Ao )−1 Kat , and yt = [G(L−1 − Ao )−1 K + I]at .

(9.8.3)

Because at is a white noise with covariance matrix GΣG′ + R , it follows that the covariance generating function of {yt } equals Sy (z) = [G(zI − Ao )−1 K + I][GΣG′ + R][K ′ (z −1 I − Ao′ )−1 G′ + I].

(9.8.4)

Expressions (9.8.2) and (9.8.4) are alternative representations for the covariance generating function Sy (z). Equating them leads to the factorization identity: G(zI − Ao )−1 CC ′ (z −1 I − Ao′ )−1 G′ + R =

[G(zI − Ao )−1 K + I][GΣG′ + R][K ′ (z −1 I − Ao′ )−1 G′ + I].

(9.8.5)

Vector Autoregression for {yt }

201

The importance of the factorization identity hinges on the fact that, under assumptions A1 and A2, the zeros of the polynomial det[G(zI − Ao )−1 K + I] all lie inside the unit circle, which means that in the representation (9.8.3) for yt , the polynomial in L on the right hand side has a one-sided inverse in nonnegative powers of L, so that at lies in the space spanned by y t . We establish this result in the next section, then apply it in subsequent ones.

9.8.2. Location of zeros of characteristic polynomial We utilize two theorems from the algebra of partitioned matrices. Let a, b, c, d be appropriately conformable and invertible matrices. Then (a − bd−1 c)−1 = a−1 + a−1 b(d − ca−1 b)−1 ca−1

(9.8.6)

det(a) det(d + ca−1 b) = det(d) det(a + bd−1 c).

(9.8.7)

and Apply equality ( 9.8.6 ) to [I + G(zI − Ao )−1 K]−1 with the settings a = I, b = −G, d = (zI − Ao ), c = K , to get [I + G(zI − Ao )−1 K]−1 = I − G[zI − (Ao − KG)]−1 K.

(9.8.8)

Apply equality ( 9.8.7 ) with the settings a = I, b = G, d = (zI − Ao ), c = K to get det(zI − (Ao − KG)) = det(zI − Ao ) det(I + G(zI − Ao )−1 K), or det[I + G(zI − Ao )−1 K] =

det(zI − (Ao − KG)) . det(zI − Ao )

(9.8.9)

It follows from ( 9.8.9 ) that the zeros of det[I + G(zI − Ao )−1 K] are the eigenvalues of

Ao − KG , and the poles of det[I + G(zI − Ao )−1 K] are the eigenvalues of Ao . Assumptions A1 and A2 guarantee that the eigenvalues of Ao − KG are less than unity in modulus. We

have already made assumptions that assure that the eigenvalues of Ao are less than unity in modulus. These conditions on the eigenvalues together with equations ( 9.8.8 ) and ( 9.8.9 ) permit us to obtain the Wold and autoregressive representations of {yt } in convenient forms.

202

Representation and Estimation

9.8.3. Wold and autoregressive representations (white measurement errors) From (9.8.8), we have that [G(I − Ao L)−1 KL + I]−1 = I − G[I − (Ao − KG)L]−1 KL.

(9.8.10)

For the model with serially uncorrelated measurement errors, the Wold representation for {yt } is yt = [G(I − Ao L)−1 KL + I]at .

(9.8.11)

Applying the inverse of the operator on the right of (9.8.11) and using (9.8.10) gives yt = G[I − (Ao − KG)L]−1 Kyt−1 + at , (9.8.12) which decomposes yt into an innovation at and a one-step ahead linear least squares predictor E[yt |y t−1 ] = G[I − (Ao − KG)L]−1 Kyt−1 .

(9.8.13)

Equation (9.8.12) is equivalent with yt =

∞ X j=1

G(Ao − KG)j−1 Kyt−j + at .

(9.8.14)

Equation (9.8.14) is a vector autoregressive representation for yt . Thus, the Kalman filter allows us to move from the original state space representation to a vector autoregression.

Vector Autoregression for {yt }

203

9.8.4. Serially correlated measurement errors With few modifications, the preceding analysis can be adapted to calculate the vector autoregressive representation and the one-step ahead prediction for yt for the case in which the measurement errors are vector first-order autoregressive processes. We have seen that the Wold representation in this case takes the form yt+1 = [I − DL]−1 [I + G(I − Ao L)−1 KL]ut . (9.8.15) Operating on both sides of (9.8.15) with the inverse of the operator in L on the right side, and using (9.8.10), we obtain [I − DL] {I − G[I − (Ao − KG)L]−1 KL}yt+1 = ut , or yt+1 = {D + (I − DL) G[I − (Ao − KG)L]−1 K} yt + ut ,

(9.8.16)

ˆ t+1 |y t ]. The above equation can be expressed where recall that ut = yt+1 − E[y in the alternative forms yt+1 = Dyt + G

∞ X

(Ao − KG)j−1 Kyt−j+1

j=1 ∞ X

− DG

j=1

(Ao − KG)j−1 Kyt−j + ut ,

or yt+1 =[D + GK]yt +

∞ X [G(Ao − KG)j K j=1

(9.8.17)

− DG(Ao − KG)j−1 K]yt−j + ut .

These equations express yt+1 as the sum of the one-step ahead linear least squares forecast and the one-step prediction error. 9

9 The MATLAB program varrep.m uses ( 9.8.17 ) to obtain a vector autoregressive representation for an equilibrium set of yt ’s, given [Ao , C, G, D, R] .

204

Representation and Estimation

9.9. Innovations in yt+1 as Functions of Innovations wt+1 and ηt+1 By coupling the original state space system with the associated innovations representation, it is possible to express the innovations in the {yt } process as functions of the disturbances {wt } and the measurement errors {vt }. Having a method for expressing this connection can be useful when we want to interpret the innovations in {yt } as functions of the shocks impinging on agents’ information sets and the measurement errors. The state space system is xt+1 = Ao xt + Cwt+1 y t = Gxt + GCwt+1 + ηt+1 ,

(9.9.1)

which corresponds to an innovations representation, which can be expressed as the “whitener” xt + Ky t x ˆt+1 = (Ao − KG)ˆ (9.9.2) ut = y t − Gˆ xt .

Substituting the second equation of (9.4.3) into the first equation of (9.9.2) gives x ˆt+1 = (Ao − KG)ˆ xt + KGxt + KGCwt+1 + Kηt+1 . (9.9.3)

Using (9.9.3), systems (9.4.3) and (9.9.2) can be combined to give the consolidated system o xt Cwt+1 xt+1 A 0 + (9.9.4) = x ˆt KGCwt+1 + Kηt+1 x ˆt+1 KG Ao − KG xt + [GCwt+1 + ηt+1 ] ut = [G − G] x ˆt

In system (9.9.4), the “inputs” are the innovations to agents’ information sets, namely, wt+1 , and the innovations to the measurement errors, namely, ηt+1 . ˆ t+1 | The “output” of the system is the innovation to yt+1 , namely ut = yt+1 −Ey t y . By computing the impulse response function of system (9.9.4), we can study how the innovations ut depend on current and past values of wt+1 and ηt+1 . Versions of formula (9.9.4) are useful for studying the range of issues considered by Hansen and Sargent [1991, “Two Difficulties”]. 10 In the next section, we illustrate one such issue in the context of a permanent income example. 10 The MATLAB programs white1.m and white2.m use formula ( 9.9.4 ) to compute impulse response functions of ut with respect to wt and ηt , respectively.

Innovations in the yt ’s and the wt ’s in a Permanent Income Model 205

9.10. Innovations in the yt ’s and the wt ’s in a Permanent Income Model This section illustrates some of the preceding ideas in the context of an economic model that implies that the econometrician’s information set spans a smaller space than agents’ information. The context is a class of models which impose a form of expected present value budget balance. As we shall see, expected present value budget balance is characterized by a condition that implies that the moving average representation associated with the model, which records the response of the system to the wt ’s, fails to be invertible. The outcome is that the innovations in the autoregressive representation don’t coincide with the wt ’s. Representation (9.9.4) can be used to compute a distributed lag expressing the innovations as functions of the lagged wt ’s. We consider the following version of Hall’s model in which the endowment process is the sum of two orthogonal autoregressive processes. Preferences, technology, and information are specified as follows:

9.10.1. Preferences ∞

1 X t − E β [(ct − bt )2 + ℓ2t ] | J0 2 t=0

9.10.2. Technology ct + it = γkt−1 + dt φ1 it = gt kt = δk kt−1 + it gt · gt = ℓ2t

206

Representation and Estimation

9.10.3. Information 1 0 0 = 0 0 0

A22

0 .9 0 0 0 0

Ud =

0 0 0 1 0 0

0 0 0 0 1 0

5 1 0 0

Ub = [ 30

0

0 0 0 1 0 0 , C2 = 0 0 0 0 0 0 1 .8 .6 .4 0 0 0 0 0 0 0 0 0 1

0 0

0 0 4 0 0 0

0 0]

We specify that γ = .05, δk = 1, β = 1/1.05, φ1 = .00001. Note that β(δk + γ) = 1, which is the condition for consumption to be a random walk in Hall’s model. The preference shock is constant at 30, while the endowment process is the sum of a constant (5) plus two orthogonal processes. In particular, we have specified that dt = 5 + d1t + d2t , where d1t = .9d1t−1 + w1t d2t = w ˜2t + .8w ˜2t−1 + .6w ˜2t−2 + .4w ˜2t−3 where (w1t , w ˜2t ) = (w1t , 4w2t ). Notice that we have set ′ w1t w1t 1 0 . E = 0 16 w ˜2t w ˜2t Here d1t is a first order autoregressive process, while d2t is a third order pure moving average process. We define the household’s net of interest deficit as ct − dt . Hall’s model imP∞ poses “expected present value budget balance,” in the sense that E j=0 β j (ct+j − dt+j ) | Jt = β −1 kt−1 for all t , 11 which implies that the present value of the moving average coefficients in the response of the deficit to innovations in agents’ information sets must be zero. That is, let the moving average representation of (ct , ct − dt ) in terms of the wt ’s be 12 ct σ1 (L) = wt , (9.10.1) ct − dt σ2 (L) 11 See Sargent [1987] and Hansen, Roberds, and Sargent [1990]. 12 Without loss of generality, the covariance matrix of w can be chosen to be the identity t matrix.

Innovations in the yt ’s and the wt ’s in a Permanent Income Model 207

where σ1 (L) and σ2 (L) are each (1 × 2) matrix polynomials, and σ(L) = P∞ j j=0 σj L . Then Hall’s model imposes the restriction σ2 (β) = [ 0

0].

(9.10.2)

The agents in this version of Hall’s model observe Jt at t , which includes the history of each component of wt up to t . This means that agents see histories of both components of the endowment process d1t and d2t . Let us now put ourselves in the shoes of an econometrician who has data on the history of the pair [ ct , dt ], but not directly on the history of wt . We imagine the econometrician to form a record of consumption and the deficit [ ct , ct − dt ], and to obtain a Wold representation for the process [ ct , ct − dt ]. Let this representation be denoted 13 ∗ σ1 (L) ct = ut , (9.10.3) σ2∗ (L) ct − dt where σ ∗ (L) is one-sided in nonnegative powers of L, and ut is a serially uncorrelated process with mean zero and Eut u′t = I ; ut is the innovation of [ ct , ct − dt ] relative to the history [ct−1 , ct−1 −dt−1 ]. In representation (9.10.3), ut is the object that would appear in the Gaussian log likelihood function, as in formula (9.6.4). It is natural to ask whether the impulse response functions σ ∗ (L) in the Wold representation (or vector autoregression) (9.10.3) estimated by the econometrician “resemble” the impulse response functions σ(L) that depict the response of ( ct , ct − dt ) to the innovations to agents’ information. A way to attack this question is to ask whether the history of the {ut } process of innovations to the econometrician’s information set in (9.10.3) reveals the history of the {wt } process impinging on agents’ information sets. In the present model, the answer to this question is ‘no’ precisely because restriction (9.10.2) holds. In particular, (9.10.2) implies that the history of ut ’s in (9.10.3) spans a smaller linear space than does the history of wt ’s. Here is the reason. The ut ’s in (9.10.3) are constructed to lie in the space spanned by the history of the [ ct , ct − dt ] process. 14 Technically, this implies 13 Without loss of generality, the covariance matrix of u can be chosen to be the identity t matrix. 14 Recall the construction underlying Wold’s representation theorem, e.g., see Sargent [1987, chapter XI].

208

Representation and Estimation

that the operator σ ∗ (L) in (9.10.3) is invertible, so that (9.10.3) can be expressed as ct ut = σ ∗ (L)−1 , ct − dt where σ ∗ (L)−1 is one-sided in nonnegative powers of L, and where the coefficients in its power series expansion are square summable. Given that σ ∗ (z)σ ∗ (z −1 )′ is of full rank, a necessary condition for σ ∗ (L)−1 to exist (i.e., to have a representation as a square-summable polynomial in nonnegative powers of L) is that det σ ∗ (z) have no zeros inside the unit circle. Condition (9.10.2) then rules out the possibility that σ ∗ (L) is related to σ(L) by a relation of the form σ ∗ (L) = U σ(L) where U is a nonsingular 2 × 2 matrix. For (9.10.2) implies that det σ(z) has a zero at β , which is inside the unit circle. In circumstances in which [ ct , ct − dt ] is a full rank process, 15 the history of [ ct , ct − dt ] generates a smaller information set than does the history of the wt process. When ut spans a smaller space than wt , ut will typically be a distributed lag of wt that is not concentrated at zero lag: ut =

∞ X

αj wt−j .

(9.10.4)

j=0

Thus the econometrician’s news ut potentially responds with a lag to the agents’ news wt . The calculations leading to representation (9.9.4) can be used to compute the vector distributed lag αj . To illustrate these ideas in the context of the present version of Hall’s model, figures 9.10.1.a and 9.10.1.b display the impulse response functions of [ ct , ct − dt ] to the two innovations in the endowment process. Consumption displays the characteristic “random walk” response with respect to each innovation. Each endowment innovation leads to a temporary surplus followed by a permanent net-of-interest deficit. The temporary surplus is used to accumulate a stock of capital sufficient to support the permanent net of interest deficit that is to follow it. Restriction (9.10.2) states that the temporary surplus just offsets the permanent deficit in terms of expected present value. For each innovation, we computed the present value of the response of (ct − dt ) to be zero, as predicted by (9.10.2). 15 By a full rank process we mean that σ ∗ (z)σ ∗ (z −1 ) .

Innovations in the yt ’s and the wt ’s in a Permanent Income Model 209

0.4

1 0.5

0.2 0 0

-0.5 -1

-0.2 -1.5 -0.4

-2 -2.5

-0.6 -3 -0.8 0

5

10

15

20

25

30

35

40

45

50

Fig. 9.10.1.a. Impulse response of consumption and deficit to first endowment innovation. The dotted line denotes the deficit, the dark line consumption.

-3.5 0

5

10

15

20

25

30

35

40

45

50

Fig. 9.10.1.b. Impulse response of consumption and deficit to second endowment innovation. The dotted line denotes the deficit, the dark line consumption.

Figures 9.10.2.a and 9.10.2.b report the impulse responses from the Wold representation, which we obtained using the programs varma.m and varma2.m. The innovation covariance matrix for the ut ’s was .3662 −1.9874 ∗ ∗′ Eσ0 σ0 = . −1.9874 12.8509

Notice that consumption responds only to the first innovation in the Wold representation, and that it responds with an impulse response symptomatic of a random walk. That consumption responds only to the first innovation in the vector autoregression is indicative of the Granger-causality imposed on the [ ct , ct − dt ] process by Hall’s model: consumption Granger causes ct − dt , with no reverse causality. Unlike consumption, the response of the deficit (ct − dt ) to the innovations in the vector autoregression depicted in figures 9.10.2.a and 9.10.2.b fail to match up qualitatively with the patterns displayed in figures 9.10.1.a and 9.10.1.b. In particular, the present values (σ2∗ (β)) of the response of ct − dt to ut are (6.0963, 6.6544). By construction, σ2∗ (β) cannot be zero because σ2∗ (L) is invertible.

210

Representation and Estimation

0.7

4

0.6

3.5

0.5

3

0.4

2.5

0.3 2 0.2 1.5 0.1 1

0

0.5

-0.1 -0.2

0

-0.3 0

-0.5 0

5

10

15

20

25

30

35

40

45

50

Fig. 9.10.2.a. Impulse response of consumption and deficit to first innovation in Wold representation. The dotted line denotes the deficit, the dark line consumption.

5

10

15

20

25

30

35

40

45

50

Fig. 9.10.2.b. Impulse response of consumption and deficit to second innovation in Wold representation. The dotted line denotes the deficit, the dark line consumption.

Figures 9.10.3.a and 9.10.3.b display the impulse responses of ut to wt , the kind of representation depicted in equation (9.10.4). While the responses of the innovations to consumption are concentrated at lag zero for both components of wt , the responses of the innovations to (ct − dt ) are spread over time (especially the response to w1t ). Thus, the innovations to (ct − dt ) as revealed by the vector autoregression depend on what to economic agents is “old news”. Hansen, Roberds, and Sargent [1991] describe how such issues impinge on strategies for econometrically testing present value budget balance. Hansen and Sargent [1991] and Marcet [1991] more generally study the link between innovations in a vector autoregression and the innovations in agents’ information sets.

Frequency Domain Estimation

0.4

211

1 0.5

0.2 0 0

-0.5 -1

-0.2 -1.5 -0.4

-2 -2.5

-0.6 -3 -0.8 0

2

4

6

8

10

12

14

16

18

-3.5 0

20

Fig. 9.10.3.a. Impulse response of innovations in Wold representation for consumption and deficit to first endowment innovation. The dotted line denotes the deficit, the dark line consumption.

2

4

6

8

10

12

14

16

18

20

Fig. 9.10.3.b. Impulse response of innovations in Wold representation for consumption and deficit to second endowment innovation. The dotted line denotes the deficit, the dark line consumption.

9.11. Frequency Domain Estimation We now describe how to estimate the free parameters of the model (9.4.1) using the frequency domain approximation to the likelihood function of Hannan [1970].We assume a model for which yt is asymptotically stationary. Let the mean vector for the observable {yt } process be denoted µ. The mean vector µ is a function of the parameters of the model. The spectral density matrix of the {yt } process is defined as Sy (ω) =

∞ X

Cy (τ )e−iωτ

(9.11.1)

τ =−∞

where Cy (τ ) = E[yt − µ][yt−τ − µ]′ . For the model (9.4.1), the spectral density can be represented as Sy (ω) = G(I − Ao e−iω )−1 CC ′ (I − Ao′ e+iω )−1 G′ + (I − De−iω )−1 R(I − D′ e+iω )−1

(9.11.2a)

212

Representation and Estimation

and the unconditional means can be represented via a function Eyt ≡ µ = µ(Ao , G).

(9.11.2b)

The autocovariances can be recovered from Sy (ω) via the inversion formula 16 Cy (τ ) =

1 2π

Z

π

Sy (ω)e+iωτ dω.

(9.11.3)

−π

Let yt be a (p × 1) vector. Suppose that a sample of observations on {yt }Tt=1 is available. Define the Fourier transform of {yt }Tt=1 as y(ωj ) =

T X

yt e−iωj t , ωj =

t=1

2πj , j = 1, . . . , T. T

(9.11.4)

The periodogram of {yt }Tt=1 is defined as Jy (ωj ) =

1 y(ωj )y(ωj )′ , T

(9.11.5)

where the overbar denotes complex conjugation. Following Hannan [1970], the Gaussian log likelihood of {yt }Tt=1 as a function of the free parameters determining Ao , C, D, and R can be approximated as T /2+1 X 1 (T + T p) log 2π − log{det Sy (ωj )} L∗ = − 2 j=1 T /2+1

− −

X j=1

trace Sy (ωj )−1 Jy (ωj )

(9.11.6)

T T n X X ′ o T yt − µ yt − µ T −1 trace Sy (0)−1 T −1 2 t=1 t=1

In (9.11.6), p is the dimension of the yt vector. The free parameters determining Ao , C, D , and R can be estimated by maximizing the right side of (9.11.6) with respect to them. Notice that the data {yt }Tt=1 enter the right side of (9.11.6) only through the sample mean 16 The MATLAB programs spectral.m and spectr1.m can be used to compute a spectral density matrix for one of our models. These programs implement formula ( 9.11.2 ).

Approximation Theory

213

PT T −1 t=1 yt and the periodogram Jy (ωj ), while the theory enters through relation (9.11.2) which determines µ and Sy (ωj ) as functions of the free parameters. Parameter estimation uses any of a variety of hill-climbing algorithms on (9.11.6). 17 An advantage of frequency domain estimation is that it avoids the need, associated with time domain estimation, to deduce a Wold representation for yt . Notice that estimation proceeds without factoring the spectral density matrix (9.11.2).

9.12. Approximation Theory When an economist estimates a misspecified model, how are the probability limits of the parameters that he estimates related to the parameters of a “true” model? This question is not well posed until one states an alternative model relative to which the model at hand is regarded as misspecified. If such an alternative model is on the table, then the question can be answered by adapting the analysis of approximation used by Christopher Sims [1972] and Halbert White [1982]. A modification of (9.11.6) underlies the theory of approximation. To state a complete theory of approximation, these elements are required: (1) a model that in truth generates the data (to speak of approximation, it is necessary to specify what is being approximated); (2) the model being estimated; and (3) the method of parameter estimation. We make the following assumption about these three elements. The true model is a member of the class of models described earlier in this book, with parameters denoted by a vector δ . The true mean vector for the observables is ν(δ), and the true spectral density matrix is S y (ω, δ), where S y (ω, δ) is determined by a version of (9.11.2a) with parameters o A , C, G, R, D , which depend on the parameter vector δ . The estimated model is another version of (9.11.2), where the parameters determining the matrices Ao , C, G, R, D , are denoted α , the spectral density matrix is Sy (ω, α), and the mean vector is µ(α). The method of estimation is maximum likelihood. It can be shown (see Hansen and Sargent (1993)) that the probability limits of the free

17 For example, see Bard (1974 XXXX).

214

Representation and Estimation

parameters α satisfy n 1 Z π plim α ˆ = arg max − log det Sy (ω, α) dω α 2π −π Z π 1 − trace Sy (ω, α)−1 S y (ω, δ) dω 2π −π o − [ν − µ(α)] Sy (0, α)−1 [ν − µ(α)]′ .

(9.12.1)

The right side of (9.12.1) is obtained from (9.11.6) by appropriately taking limits as T → ∞ . Roughly speaking, taking limits replaces the periodogram Jy (ωj ) with the spectral density for the true model S y (ωj ), and replaces the sample mean with the true mean vector.

9.13. Aggregation Over Time In this section, we describe how to use the Kalman filter to calculate the likelihood for data that are “aggregated over time.” We formulate a model that generates observations in state space form and then use the Kalman filter to derive an associated innovations representation from which the Gaussian log likelihood function can be constructed. Let the original equilibrium model have the state space form xt+1 = Ao xt + Cwt+1 yt = Gxt

(9.13.1)

′ where wt+1 is a martingale difference sequence with Ewt+1 wt+1 |Jt = I . We assume that the model is formulated to apply at a finer time interval than that for which data are available. For example, the model (9.13.1) may apply to weekly or monthly data, while only quarterly or annual data may be available to the economist. Furthermore, some of the observed data may be averages over time of the {yt } data generated by (9.13.1), as when “flow” data are generated by averaging over time. (Data on output, consumption, and investment flows are usually generated in this way.) Others of the data may simply be point-intime “skip sampled” versions of the data. That is, “quarterly” data are formed by sampling every thirteenth observation of the “weekly” data. We want to catalogue the restrictions imposed on these time aggregated data by the model

Aggregation Over Time

215

(9.13.1). We accomplish this by deducing the likelihood function of these data as a function of the free parameters for (9.13.1). We perform our analysis of aggregation over time in two steps. First, we expand the state space by including enough lagged states to accommodate whatever averaging over time of data is occurring. Let m be the number of dates over which data are potentially to be averaged. Then we form the augmented system x t+1 xt xt−1 .. .

xt−m+2

Ao I 0 = .. .

0

0 0 I .. .

··· ··· ··· .. .

0 0 0 .. .

0

···

I

0 0 0 .. . 0

xt xt−1 xt−2 .. .

xt−m+1

C 0 0 + wt+1 .. .

0

or xt+1 = Axt + Cwt+1

(9.13.2)

where

xt+1

x t+1 xt = xt−1 .. .

Ao I , A= 0 .. .

xt−m+2

0

0 0 I .. .

··· ··· ··· .. .

0 0 0 .. .

0

···

I

0 C 0 0 0, C = 0 . .. .. . . 0 0

(9.13.3)

Once we have formed A and C , it is easy to form the appropriate model for averaged data. For example, suppose that we are interested in forming the model governing three period averages of consumption. We would set m equal to 3 in (9.13.2) and (9.13.3), and could then model averaged consumption via the observer equation yt = Gxt where G = [Sc Sc Sc ]. The MATLAB program avg.m obtains the matrices A and C of (9.13.3) for a given m , thereby accomplishing the first step in our analysis of aggregation over time. The second step is actually to perform the aggregation over time by skipping observations on a representation of the form (9.13.1) or (9.13.2). Let an equilibrium be represented in the state space form xt+1 = Axt + Cwt+1 , yt = Gxt

t = 0, 1, 2, . . .

(9.13.4)

216

Representation and Estimation

where the first line of (9.13.4) could correspond either to (9.13.2) or to its special case, the first line of (9.13.1). Suppose that data on yt are available only every r > 1 periods, where r is an integer. Then the data are generated by the model r xt+r = Ar xt + wt+r , t = 0, r, 2r, . . . (9.13.5) yt = Gxt where Ar = Ar r wt+r = Ar−1 Cwt+1 + Ar−2 Cwt+2 + · · · + ACwt+r−1 + Cwt+r

(9.13.6)

Represent (9.13.5),(9.13.6) as the state space system r xs+1 = Ar xs + ws+1

, s = 0, 1, 2, . . .

ys = Gxs

(9.13.7)

r where ws+1 is a martingale difference sequence with contemporaneous covariance matrix

Ewsr wsr′ = CC ′ + ACC ′ A′ + · · · + Ar−1 CC ′ Ar−1′ ≡ V.

(9.13.8)

Now suppose that only error-corrupted observations on the time aggregated {ys } data are available, and that the measurement errors are first-order serially correlated. To capture this assumption, augment (9.13.7) – (9.13.8) to become the state space system r xs+1 = Ar xs + ws+1 ys = Gxs + υs

(9.13.9)

υs = Dυs−1 + ηs where Eηs ηs′ = R and Ews+1 ηs′ = 0 for all t and s . System (9.13.9) is a version of the state space system (9.4.1). Proceeding as in our analysis of (9.4.1), define y s ≡ ys+1 − Dys and Gr = (GAr − DG). Then (9.13.9) implies the system r xs+1 = Ar xs + ws+1 r y s = Gr xs + Gws+1 + ηs+1 . ′

(9.13.10)

r r Define the covariance matrices Ewsr wsr = V ≡ V1 , E(Gws+1 + ηs+1 )(Gws+1 + ′ ′ r r ′ ′ ηs+1 ) = GV G + R ≡ V2 , Ews+1 (Gws+1 + ηs+1 ) = V G = V3 . Use the function

Aggregation Over Time

217

kfilter to obtain [K, Σ] = kfilter(Ar , Gr , V1 , V2 , V3 ). Then an innovations representation for system (9.13.10) is x ˆs+1 = Ar x ˆs + Kas y s = Gr x ˆs + as

(9.13.11)

], as = y s − E[y s | y s−1 ], Ω1 ≡ Eas a′s = Gr ΣG′r + where x ˆs = E[xs | y s−1 0 0 V2 . Once again, the innovations representation (9.13.11) can be used to form the residuals at recursively, and thereby to form the Gaussian log likelihood function. 18 We illustrate the programs avg.m and aggreg.m by showing how they can be used to analyze the effects of aggregation over time in the context of our equilibrium version of Hall’s model. We want to deduce the univariate Wold representation for consumption data that are constructed by taking a three period moving average, and then “skip sampling” every third period. The following MATLAB code performs these calculations: reads in parameters of Hall’s economy computes the equilibrium forms state for three period averaging forms observer for three-period moving average of consumption R = .0001; D = 0; sets parameters of measurement error process [Ar,Cr,aa,bb,cc,dd,V1] = aggreg (AA,CC,G,D,R,3) y = dimpulse(aa,bb,cc,dd,1,20); forms moving average representation

clex11; solvea; [AA,CC]= avg(a0,C,3); G = [sc sc sc];

We have set the parameters of Hall’s model at the values that make unaveraged consumption follow a random walk. Notice that we set R and D so that only a very small measurement error is present in consumption. The impulse response function for skip-sampled three period moving average consumption reveals the following representation for the skip-sampled moving average data ct : ct − ct−1 = at + .2208at−1 where at = ct − E(ct | ct−1 , ct−2 , . . .). Thus, the first difference of ct is a firstorder moving average process. These calculations recover a version of Holbrook 18 The MATLAB program aggreg.m constructs the innovations representation ( 9.13.11 ) from inputs in the form of the state space representation ( 9.13.4 ) and the parameters R and D of the measurement error model ( 9.13.2 ).

218

Representation and Estimation

Working’s [1960] findings about the effects of skip sampling a moving average of a random walk.

9.14. Simulation Estimators We have described how to estimate the free parameters of a model using data that are possibly error-ridden linear functions of the state vector xt . In our models, quantities and (scaled Arrow-Debreu) prices are linear functions of the state, but asset prices and rates of return are non-linear functions of the state. In this section, we describe how observations of non-linear functions of the state can be used in estimation. The equilibrium transition law for the state vector xt is given by xt+1 = Ao (θ)xt + C(θ)wt+1 ,

Ewt wt′ = I

(9.14.1)

where the r × 1 vector θ contains the free parameters of preferences, technologies, and information. We partition the data into two parts, (z1t , t = 0, . . . T ) and (z2t , t = 0, . . . T ), where the z1t ’s are linear functions of the state xt , and the z2t ’s are nonlinear functions of the state. Assume that z1t is k × 1 and z2t is m × 1. The data are related to the state xt and measurement errors vt as follows: z1t = G(θ)xt + v1,t z2t = f (xt , v2,t , θ), where E

wt+1 vt

wt+1 vt

′

=

Q(θ) W (θ)′

W (θ) R(θ)

,

and where Q(θ) = C(θ)C(θ)′ . The Gaussian log likelihood function of {z1t }Tt=0 is L(θ) =

T X t=0

T

ℓt = −

i 1 Xh p log(2π) + log |Ωt | + a′t Ω−1 t at 2 t=0

where zt is p × 1 and at = z1,t − E[z1,t |z1,t−1 , . . . , z1,0 ] is the innovation vector from the ‘innovations representation’ and Ωt = Eat a′t .

Simulation Estimators

219

Maximizing the log likelihood function with respect to θ is equivalent with a particular Generalized Method of Moments (GMM) procedure using observations on (z1t , t = 0, . . . T ). Note that the first-order order conditions for maximizing the log likelihood function are ∂L = 0. ∂θ To see how this matches up with GMM, compute the score vector st = which has elements,

∂ℓt ∂θ

∂a ′ ∂ℓt ∂Ωt 1 t −1 ′ = − tr Ω−1 I − Ω a a − Ω−1 t t t t t at . ∂θi 2 ∂θi ∂θi Using the notation of Hansen (1982), the GMM estimator of θ minimizes JT (θ) = gT (θ)′ WT gT (θ) where

(9.14.2)

T

gT (θ) =

1 X st (θ) T + 1 t=0

and WT is any positive definite r×r weighting matrix. Notice that gT (θ) = ∂L ∂θ , so that for any positive definite weighting matrix, (9.14.2) is minimized by the minimizer of L(θ). The irrelevance of the weighting matrix WT reflects the property that from the viewpoint of GMM, this is a ‘just-identified’ system, with as many moment conditions as free parameters. Suppose that we want to use the observations in z1t and in z2t to estimate θ . We can apply a method described by Ingram and Lee. Given the law of motion in (9.14.1) and a realization from a pseudo-random number generator for {wj+1 , v1j , v2j }N j=0 , we can generate a pseudo-random realization of the series N {z1j , z2j }j=0 . Let q(·) be a given function of the data. Use the data and the simulation of the model, respectively, to compute the two moment vectors: T

1 X HT (z) = q(z1t , z2t ) T + 1 t=0 N

HN (θ) =

1 X q(z1j , z2j ; θ). N + 1 j=0

220

Representation and Estimation

Define hT (θ) as follows hT (θ) =

T n i 1 Xh 1 X q(z1t , z2t ) − q(z1j , z2j ; θ) T + 1 t=0 n + 1 j=0

= HT (z) − HN (θ),

where n + 1 = (N + 1)/(T + 1) and N + 1 is some integer multiple of T + 1. Then the estimation strategy for obtaining θ is to minimize JT (θ) =

∂L/∂θ hT (θ)

′

WT

∂L/∂θ hT (θ)

for some weighting matrix WT . To estimate WT , we can use the two-stage procedure in Hansen (1982), which is to start with WT = I and then construct the weighting matrix associated with the resulting estimate of θ .

A. Initialization of the Kalman Filter This appendix describes how Kohn and Ansley’s idea for estimating the initial state can be applied in the context of our class of models. Aside from numerical issues, Kohn and Ansley’s procedure is equivalent to using all of the data {ys }Ts=0 , and initializing the Kalman filter from a partitioned covariance matrix designed to approximate +∞I +∞1 , Σ0 = +∞Q′ Σ0,22 where Σ0,22 is the asymptotic covariance matrix of that piece of the state vector that has an asymptotically stationary distribution, and 1 is a matrix of ones. The +∞I pertains to elements of the state that have no asymptotic stationary distribution. In practice, +∞ is approximated by a large positive scalar. This procedure was used by Harvey and Pierse. 19 This procedure ought to be close to Kohn and Ansley’s, though the literature contains examples of cases in which the numerical properties of the ‘+∞ ≈ a big number ’ approach are poor. For that reason, it is good to have in hand procedures like the one we shall describe. 19 Another approach has been to use an ‘inverse filter’ in which the recursions are cast in terms of the inverse of Σt .

Initialization of the Kalman Filter

221

For convenience, we temporarily work with the state-space system 20 ∗ xt+1 = Axt + Cwt+1

(9.A.1)

∗ yt = Gxt + Qwt+1 ,

∗ where wt+1 is a martingale difference sequence with identity for its conditional covariance matrix. In the interest of eventually imputing a diffuse prior to the initial values of that part of the state vector that has no stationary distribution, we represent the initial state as

x0 = φη + ψ + N ν, where ψ is an n × 1 vector with all zeros except possibly for one value of one, which locates the constant in the state; and ν is normally distributed with mean zero and covariance I , and η is normally distributed with mean zero and covariance kI , where the random vectors ν and η are assumed to be independent. We use φη to represent the piece of the initial state that has no stationary distribution, and N ν to represent the piece with a stationary distribution. We attain a diffuse prior on the stationary distribution by driving k to +∞ . Our plan is to project xm on ym−1 , . . . , y0 , while driving k → +∞ , and then to initialize the Kalman filter from the resulting estimators of the distribution of xm . By iterating on the state equation (9.A.1), we can write: xm = Am φη + Am ψ + Hm wm where wm ′ = ( ν ′

w1 ∗′

...

∗′ wm ) and

Hm = ( Am−1 N ′

(9.A.2)

Now create a vector Y m−1 = ( y0′

y1′

Am−2 C ...

...

C ).

′ ) that obeys: ym−1

Y m−1 = Mm η + α + Gm wm

(9.A.3)

20 It is easy to map ( 9.4.3 ), which describes the state-space system wih serially correlated wt+1 ∗ measurement errors, into this form. Define wt+1 = and represent ( 9.4.3 ) as ηt+1 ∗ xt+1 = Axt + (C O) wt+1

∗ . yt = Gxt + (GC I) wt+1

222

Representation and Estimation

where

Mm =

and

Gm

GN GAN GA2 N .. .

= GAm−2 N GA

m−1

N

Gφ GAφ .. . GAm−1 φ

,

α=

G GA .. . GAm−1

ψ

Q GC GAC .. .

0 Q GC .. .

0 0 Q .. .

··· ··· ··· .. .

0 0 0 .. .

GAm−3 C GAm−2 C

GAm−4 C GAm−3 C

GAm−5 C GAm−4 C

··· ···

Q GC

0 0 0 .. .

0 Q

Transform equation (9.A.2) as follows. Regress Hm wm onto Gm wm , and denote the residual as Rm wm , to obtain the representation ∗ Hm w m = H m Gm wm + Rm wm ,

(9.A.4)

∗ = (EHm wm wm−1′ G′m )(EGm wm wm−1′ G′m )−1 is a matrix of least where Hm ∗ ∗ = Hm G′m (Gm Gm ′ )−1 . Gm . Thus Hm squares regression coefficients and Rm = Hm −Hm Also, since Gm wm = Y m−1 − Mm η − α , (9.A.4) implies the representation ∗ (Y m−1 − Mm η − α) + Rm wm . Hm w m = H m

Rewrite state equation (9.A.2) as: ∗ ∗ ∗ xm = (Am φ − Hm Mm )η + Am ψ − Hm α + Hm Y m−1 + Rm wm .

(9.A.5)

Next we compute some conditional expectations and covariances. Initially, we use (9.A.5) and the facts that (i) by assumption, wm is orthogonal to η , and (ii) by construction, Rm wm is orthogonal to Gm wm , to compute: ∗ ∗ ∗ E(xm |Y m−1 , η) = (Am φ − Hm Mm )η + Am ψ − Hm α + Hm Y m−1 ,

and ′ cov(xm |Y m−1 , η) = Rm Rm .

Initialization of the Kalman Filter

223

To compute the conditional expectation and covariance matrix conditioning only Y m−1 , we first compute the projection of η on Y m−1 − α : η = β ∗ (Y m−1 − α) + ε, where ε is a least squares residual. We compute Eη(Y m−1 − α) and the second moment matrix of Y m−1 − α and use them in the projection formula: ′ ′ β ∗ = (kMm )(kMm Mm + Gm G′m )−1 . ′ ′ ′ (Gm G′m )−1 Mm ]−1 [Mm (Gm G′m )−1 Mm ] to get β ∗ = [Mm (Gm Premultiply by [Mm ′ −1 −1 ′ ′ −1 ′ ′ ′ −1 Gm ) Mm ] Mm (Gm Gm ) [kMm Mm (kMm Mm +Gm Gm ) ]. If we drive k → +∞ , the last term in square brackets approaches the identity matrix, so that we have ′ ′ E(η|Y m−1 − α) = [Mm (Gm Gm ′ )−1 Mm ]−1 Mm (Gm Gm ′ )−1 (Y m−1 − α). (9.A.6)

Notice that ε = β ∗ (Mm η + Gm wm ) − η = (β ∗ Mn − I)η + β ∗ Gm wm , and that (β ∗ Mm − I) = 0. It follows that ′ cov(η|Y m−1 − α) = [Mm (Gm Gm ′ )−1 Mm ]−1 .

(9.A.7)

Using these results and applying the Law of Iterated Expectations to (9.A.5) gives: 21 ∗ ′ E(xm |Y m−1 ) = (Am−1 φ − Hm Mm )[Mm (Gm G′m )−1 Mm ]−1

′ ∗ ∗ Mm (Gm G′m )−1 (Y m−1 − α) + Am−1 ψ − Hm α + Hm Y m−1 ,

(9.A.8)

and ′ ∗ ′ cov(xm |Y m−1 ) = Rm Rm + (Am−1 φ − Hm Mm )[Mm (Gm G′m )−1 Mm ]−1 ∗ (Am−1 φ − Hm Mm )′ .

(9.A.9)

The Kalman filter is to be initialized by using these values of x ˆm , Σm , then T applied to compute (9.6.8), using observations {ys }s=m . 21 Note that equations ( 9.A.6 ) and ( 9.A.7 ) result from applying generalized least squares to the system of equations ( 9.A.3 ), where η is regarded as a matrix of constants and Mm is a matrix of regressors.

224

Representation and Estimation

When we apply this procedure with (9.A.1) corresponding to the system m−1 (9.4.3), we should interpret Y m−1 in the preceding development as Y which corresponds to Y m in the real data. In this case, we should interpret x ˆm , Σm according to definitions of the (ˆ·) variables defined for the system with serially correlated measurement errors. 22 We can also include a contribution to the likelihood function to account for the initial observations used to form x ˆm . Begin with (9.A.3) and let Ω = ′ Gm Gm , which we take to be nonsingular. Suppose that Mm is dimensioned r by s where r > s so that η is ‘overidentified.’ Construct two matrices labeled M ⊥ and M ∗ , dimensioned (r − s) × r and s × r , respectively, to satisfy: M ⊥ Ω−1 Mm = 0 M ∗ Ω−1 M ⊥′ = 0, and construct the nonsingular matrix: D=

M ∗ Ω−1 M ⊥ Ω−1

.

Define: z1 = M ∗ Ω−1 Y m z2 = M ⊥ Ω−1 Y m . Notice that conditioned on η , z1 and z2 are uncorrelated. Moreover, by construction z2 does not depend on η . We deduce the initial likelihood contribution as follows. First note that z1 = DY m . z2 Transforming the z ’s introduces a Jacobian term: log det D, which is the first contribution to the likelihood. 22 Notice that with serially correlated measurement errors, ( 9.A.8 ), ( 9.A.9 ) give the appropriate initial conditions for the Kalman filter, because of the dating conventions that make ut the innovation to yt+1 .

Initialization of the Kalman Filter

225

The second and third contributions are the likelihoods of the z ’s. Conditioned on η , the likelihood can be factored. Only the first term in the factorization depends on η , and is present in the ‘exactly identified’ case. The quadratic form term converges to zero for this contribution. We deduce the log det term by taking the limit as k goes to infinity of: ′ log det(kM ∗ Ω−1 Mm Mm Ω−1 M ∗′ + M ∗ Gm G′m M ∗′ ) =

1 ∗ M Gm G′m M ∗′ ) k where n1 is the dimension of z1 . Taking the limit and neglecting the term n1 log k , which is the same for all settings of the parameter values and so can be ignored, leaves the term: ′ n1 log k + log det(M ∗ Ω−1 Mm Mm Ω−1 M ∗′ +

′ log det(M ∗ Ω−1 Mm Mm Ω−1 M ∗′ ).

The z2 contribution to the likelihood retains both a log det and a quadratic form contribution. Notice that the z2 term is absent in the ‘exactly identified’ case.

Chapter 10 Semiparametric Estimation with Limited Information

10.1. Introduction This chapter describes semiparametric estimation of transmission mechanisms under limited information.

10.2. Underlying Economic Model Consider the following economic model. The information available to economic agents at time t is denoted Jt . There is an endogenous state vector kt−1 which we will think of as a vector of capital stocks. The capital stocks evolve according to the evolution equation: kt = ∆kt−1 + Θit (10.2.1) where it is a vector of flow variables which we refer to as investment goods. The absolute values of eigenvalues of the matrix ∆ are presumed to be strictly less than one. There is also an exogenous state vector zt with dynamics given by: zt+1 = A22 zt + C2 wt+1

(10.2.2)

where {wt } is a martingale difference sequence adapted to {Jt } with a conditional covariance matrix I . It is known that a first-order linear specification of the dynamics is quite flexible because it can represent multivariate ARMA models of arbitrary orders. The composite state vector at time t is denoted kt−1 . xt ≡ zt The recursive solution to the model gives investment it as a function of the state vector xt : it = Sxt = Sk kt−1 + Sz zt .

– 227 –

228

Semiparametric Estimation with Limited Information

Substituting this solution into equations (10.2.1) and (10.2.2), we find that xt+1 = Axt + Cwt+1 where A≡

∆ − ΘSk 0

ΘSz A22

,C ≡

(10.2.3)

0 C2

.

Thus, equation (10.2.3) gives the evolution of the state vector process when the optimal or equilibrium investment rule is imposed. The state vector zt may enter into the decision rule or equilibrium investment relation for one of two reasons. Some components of zt may enter directly into the objective functions of economic agents; and other components may simply be used in forecasting future values of these variables. Let zt1 denote a vector of the former components, which we presume are related to zt via: zt1 = H1 zt where H1 is just a matrix that selects elements from zt . The solution for investment can often be represented as: it = Sk kt−1 + S0 zt1 + Sf

∞ X j=0

1 (Λ)j S1 E(zt+j |Jt )

(10.2.4)

and hence Sz = S0 H1 + Sf (I − Λ)−1 S1 H1 (see Hansen and Sargent 1981 and Sargent 1987). One way to obtain a solution of this form is to stack the first order conditions for the endogenous state vector and its corresponding co-state vector into an expectational first-order difference equation driven by the forcing process {zt1 }, then to solve that difference equation. A similar structure can be obtained even when the endogenous state vector is not the solution to an optimal resource allocation problem. The estimation method we describe below exploits the feedforward structure of the solution for investment whereby investment depends on current and expected future values of the forcing process {zt1 }.

Econometrician’s information

229

10.3. Econometrician’s information and the implied orthogonality conditions

We presume that the econometrician observes a time series of investment it , t = 1, 2, ..., T . With knowledge of the depreciation matrix ∆, he can construct an approximate capital series, ‘approximate’ because the initial capital stock k0 may be unknown. Because the dominant eigenvalue of ∆ is strictly less than unity in modulus, the approximation error vanishes as the sample size T gets large. The econometrician also observes some but not all of the vector zt . Partition yt zt1 = ut where yt is observed by the econometrician by ut is not. To have any hope of identifying the parameters in the underlying model, we assume that the exogenous state vector process can be uncoupled in the following way: y y zt+1 = Ay zty + Cy wt+1

and u u zt+1 = Au ztu + Cu wt+1

where zt = and

zty ztu

, wt =

wty wtu

yt = Hy zty and ut = Hu ztu . To guarantee asymptotic stationarity, we restrict the absolute values of the eigenvalues of Au to be strictly less than one. The {ut } process gives us one interpretation of why investment is not an exact function of variables observed by an econometrician. Therefore, we rewrite the investment relation as: it = Sk kt−1 + S0,y yt + Sf

∞ X j=0

where et ≡ S0,u ut + Sf

∞ X j=0

(Λ)j S1,y E(yt+j |Jt ) + et

(Λ)j S1,u E(ut+j |Jt )

230

Semiparametric Estimation with Limited Information

and Sj = ( Sj,y

Sj,u ) , j = 0, 1.

The process {et } provides an error term that can be interpreted along the lines of Hansen and Sargent (1980). By construction it is uncorrelated with the process {yt } at all leads and lags. This uncorrelatedness can be directly exploited in estimating the parameters of endogenous dynamics of the model, e.g., the parameters governing the transmission mechanism. In fact, this can be accomplished in ways that permit a robust specification of the dynamics associated with the underlying shock process {ut }. In other words, a semiparametric estimation method is possible in this setting. Another source of omitted information is in the forecasting of future values of the process {yt }. For instance, let Kt denote the information set used by the econometrician, constructed in a way so that at least it is no larger than Jt . This gives rise to an additional model “specification” error, say ft , as emphasized by Shiller (1978) and Hansen and Sargent (1980, 1982). Thus, the investment equation used by an econometrician is given by: it = Sk kt−1 + S0,y yt + Sf

∞ X j=0

where ft ≡ Sf

∞ X j=0

(Λ)j S1,y E(yt+j |Kt ) + et + ft

(10.3.1)

(Λ)j S1,y [E(yt+j |Jt ) − E(yt+j |Kt )].

By the Law of Iterated Expectations, the error term ft is uncorrelated with current and past values of {yt }, but can be correlated with future values of this process. As a consequence, the orthogonality conditions that are robust to misspecifying the information set are: E[(ft + et )yt−j ′ ] = 0 for j = 0, 1, ....

(10.3.2)

The presence of the component ft in the disturbance term is what limits the orthogonality conditions to be one-sided. Future values of yt may be correlated with the disturbance term in the investment equation. Since the information set Kt is a misspecified version of Jt , unless one is willing to specify the omitted information precisely, it is most convenient to envision an econometrician modeling the evolution equation for {yt } in a flexible

An Adjustment Cost Example

231

manner. Although information is omitted, so long as {yt } has a state-space representation, we know that it can be represented as a multivariate version of an ARMA model, although the autoregressive and moving-average orders will be unknown to the econometrician. For the sake of simplicity, we assumethat the {yt } process is stationary and has a moving-average representation: yt = B(L)vt

(10.3.3)

where the operator B has a one-sided inverse, and current and past values of vt also generate the information set Kt . In what follows, it is not necessary to limit B to be a ratio of polynomials, as in ARMA models. More general dynamics can be accommodated. As we will see, this in effect introduces an infinite dimensional nuisance parameter into the moment conditions (10.3.2). Finally, note that by omitting information relative to that used by economic agents, we cannot expect to deduce impulse response functions that are interpretable in terms of the economic shocks impinging on the decision maker. In other words, the response of investment or capital stock to an innovation in {yt } (i.e., in economic agents’ information set) will be contaminated. Nevertheless, we will still be in a position to identify parameters of the endogenous dynamics.

10.4. An Adjustment Cost Example A linear-quadratic model of adjustment costs has a solution for investment that is of the form given by (3) with a scalar investment and capital stock and a scalar Λ that we will denote by λ . For simplicity, we presume that the obervable forcing process {yt } is also scalar. Write the econometric relation for investment in feedforward form as: it = ρkt−1 + ψo yt + ψf

∞ X j=0

kt = δkt−1 + it .

(λ)j E(yt+j |Kt ) + et + ft

(10.4.1)

The operator B enters into the model solution because of its role in the solution to the prediction problem: ∞ X ytp = E( λj yt+j |Kt ). j=0

232

Semiparametric Estimation with Limited Information

It is known from Hansen and Sargent (1980) that ytp = B ∗ (L)vt where B ∗ (ζ) =

ζB(ζ) − λB(λ) . ζ −λ

(10.4.2)

Substituting this formula into (10.4.1) and solving for the econometric disturbance term, we obtain: ft + et = it − ρkt−1 − ψo yt − ψf B ∗ (L)vt .

(10.4.3)

Prior to investigating the estimation of endogenous dynamics as captured by the parameters ρ, ψo , ψf , δ, λ we will study the impact of estimating B in both parametric and nonparametric settings.

10.5. A Slightly Simpler Estimation Problem Let Yt denote a random vector, each entry of which is a linear combination of current and past values of yt . Suppose the unconditional moment condition used in estimation is: E[(ft + et )Yt ] = 0. A component of these moment conditions that depends on B is: βo ≡ E[Yt B ∗ (L)vt ],

(10.5.1)

and for the moment let us suppose that βo is the parameter of interest.

A Slightly Simpler Estimation Problem

233

10.5.1. Scalar Parameterizations of B As a preliminary step to studying nonparametric estimators of B , we initially consider very simple scalar parameterizations of B : Bα = B + αF. We suppose that for sufficiently small values of α we can invert the operator Bα . We must explore what happens to the moment condition for small perturbations in α . Define: −1 φt (α) ≡ Bα (L) yt . Differentiating φt with respect to α and evaluating we find that B(L)Dφt (0) + F (L)vt = 0, or Dφt (0) = −[B(L)]−1 F (L)vt .

(10.5.2)

Then differentiating the moment relation: β(α) = E[zty Bα∗ (L)φt (α)] we find that dβ(0)/dα = E{Yt [F ∗ (L)vt ]} + E{Yt [B ∗ (L)Dφt (0)]}.

(10.5.3)

The * notation is used to denote the transformation of an operator given by (10.4.2). Let αT denote the maximum likelihood estimator for α = 0 for sample size T , and let T X YT Bα∗ T (L)φt,T (αT ) βT ≡ (1/T ) t=1

denote the sample estimator of βo where the notation φt,T (αT ) denotes the time t approximation for vt using the estimator αT . Then the sampling error in βT as an estimator of βo can be decomposed into two components: √

T √ X √ T (βT − βo ) ≈ (1/ T ) [Yt B ∗ (L)vt − βo ] + [dβ(0)/dα] T αT . t=1

(10.5.4)

234

Semiparametric Estimation with Limited Information

The first term is the usual central limit approximation for sample moment estimators while the second term accounts for the additional sampling error induced by having to estimate B . It is the second term that we turn our attention to. The limiting distribution of {αT } is determined by the score of the conditional likelihood of yt conditioned on the past. This score is given by st = −vt Dφt (0) + E[v(t)Dφt (0)].

(10.5.5)

The first term of the score comes from differentiating the quadratic form in the one-step ahead forecast error of yt , and the second term from differentiating the log variance term of the time t contribution to the conditional log likelihood. The score variable has mean zero conditioned on Kt−1 and the resulting score process is a martingale difference sequence. Then PT √ st . (10.5.6) T αT ≈ √ t=1 T E(st 2 ) In light of the fundamental role played by the score variable in deteriming the limiting distribution for the estimator sequence {αT }, it will prove to be very useful to represent the derivative dβ(0)/dα as an expected cross product of some random vector with the score st . We now deduce what that random vector is by obtaining an alternative expression for the right-hand side of (10.5.3). Note that LF (L) ∗ E{Yt [F (L)vt ]} = E Yt vt L−λ (10.5.7) LB(L) Dφt (0) = −E Yt L−λ where the first equality follows because future values of vt are orthogonal to Yt and the second equality follows from formula (10.5.2) for Dφt (0). Substituting (10.5.7) into (10.5.3) and using formula (10.4.2) for B ∗ results in: λB(λ) dβ(0)/dα = −E Yt Dφt (0) L−λ 1 (10.5.8) = −λB(λ)E Yt Dφt (0) L−1 − λ 1 Yt−1 Dφt (0) = −λB(λ)E 1 − λL

where the second equality follows from the joint stationarity of the composite process {[Yt , Dφt (0)]}. Formula (10.5.8) is almost what we want, except that

Multidimensional Parameterizations of B

235

we need an expression in terms of st instead of Dφt (0). This can be obtained by noting that 1 1 E Yt−1 Dφt (0) = −E vt Yt−1 st (10.5.9) 1 − λL 1 − λL

which can be verified as follows. Compute the expectation on the right-hand side by conditioning first on Kt−1 and using the two facts that (i) Dφt (0) is the sum of a term in vt and E[Dφt (0)|Kt−1 ], and (ii) the third moment of vt is zero. Then apply the Law of Iterated Expecations again to obtain the lefthand side of (10.5.9). Combining (10.5.8) and (10.5.9), we obtain the desired formula:

1 Yt−1 st . (10.5.10) 1 − λL Armed with this formula, we can think of the time t contribution of the “correction term” for estimating as the outcome from running a least squares i B h 1 regression of λB(λ)vt 1−λL Yt−1 onto the score st . This interpretation can be seen by substituting (10.5.7) and (10.5.10) into (10.5.4) and interpreting dβ(0)/dα as a population regression coefficient. Although we performed this E(st 2 ) computation for an affine scalar parameterization of B , it can be mimicked for any sufficiently smooth one dimensional parameterization. The correction term will continue be interpretable as a regression score. dβ(0)/dα = λB(λ)E vt

10.6. Multidimensional Parameterizations of B As a further step in studying the impact on βo of using a nonparametric estimator of B, we now briefly consider what happens when we use parameterizations that have more than one dimension but are still finite dimensional. This turns out to be an easy extension of our previous analysis. Let st be the score vector associated with any such nondegenerate parameterization. (By “nondegenerate” we simply mean that the second moment of the score vector is nonsingular, a local identification condition.) The entries of the score vector st can be represented as in (14), and the derivative matrix ∂β(0)/∂α′ is given by the expected cross product: 1 Yt−1 st ′ . ∂β(0)/∂α′ = λB(λ)E vt 1 − λL

236

Semiparametric Estimation with Limited Information

Therefore, √

T √ X [Yt B ∗ (L)vt − βo ] T (βT − βo ) ≈(1/ T ) t=1

+ λB(λ)E vt

T √ X 1 st . Yt−1 st ′ [E(st st ′ )]−1 (1/ T ) 1 − λL t=1

Again the correction term for the h first istage estimation of B has a regression 1 interpretation: regress λB(λ)vt 1−λL Yt−1 onto the score vector st .

10.7. Nonparametric Estimation of B Since the derivative matrix has an expected cross product representation for any finite dimensional parameterization, we can use an insight from Stein (1956) and Levit (1975), developed more fully by Van der Vaart (1991) and Newey (1993), to deduce the asymptotic distribution when B is estimated nonparametrically. ask what happens to the population regression of iWe simply h 1 Yt−1 onto the linear space of time t scores for all possible λB(λ)vt 1−λL parameterizations of B . Since the elements of the regressand can be viewed as scores of hypothetical parameterizations, the resulting limiting distribution for βo is T X √ ∗ T (βT − βo ) ≈ Yt B (L)vt − βo + vt t=1

1 Yt−1 . 1 − λL

(10.7.1)

This additive decomposition gives a time series counterpart to the “correction terms” for semiparametric M-estimators derived by Newey (1993), (e.g., see formula (3.10) in Newey). 1

1 One diffference between Newey’s derivation and ours is that Newey works with score vectors for the entire process of observables. Given the additive structure of our model we can work with the simpler scores of maximum likelihood estimators of B using only data on {yt } .

Back to the Adjustment Cost Model

237

10.8. Back to the Adjustment Cost Model Let us now revert to the estimation problem of interest posed using the adjustment cost model. We let βo denote the parameter vector governing the endogenous dynamics and view ρ, ψo , ψf , δ and ρ as functions of the unknown parameter vector βo . When the capital stock is not directly observable, the generated stock sequence will also depend on βo through its dependence on the depreciation factor δ . Suppose that we estimate βo using a GMM estimator that exploits the unconditional moment restriction: E[Yt (ft + et )] = 0. Then the usual GMM inference works with an additional correction term in which the derivatives of the moment conditions are computed by differentiating with respect to β and evaluating these derivatives at the true value of βo and B . Let this derivative be denoted do , and let ao denote the limiting matrix that selects the moment conditions to be used in estimation. Then T √ X √ T (βT −βo ) ≈ −(ao do )−1 ao (1/ T Yt (ft + et ) + λB(λ) t=1

1 Yt−1 vt . 1 − λL

Chapter 11 Representation of Demand

11.1. Introduction This chapter derives demand schedules from our preference specification ht = ∆h ht−1 + Θh ct st = Λht−1 + Πct

(11.1.1)

with preference shock bt = Ub zt . An equivalence class of preferences (∆h , Θh , Π, Λ, Ub ) give rise to identical demand schedules. Among such preferences, particular ones that we call canonical are easiest to work with. We apply the concept of canonical representation of preferences to a version of Becker and Murphy’s model of rational addiction. The chapter also uses demand curves to to construct partial equilibrium interpretations of our models. This chapter sets the stage for the studies of aggregation in chapters BLANK and BLANK.

11.2. Canonical Representations of Services We begin with a definition. Definition: A representation of a household service technology (∆h , Θh , Π, Λ, Ub ) is said to be canonical if it satisfies the following two requirements: i. Π is nonsingular. ii. The absolute values of the eigenvalues of (∆h − Θh Π−1 Λ) are strictly √ less than 1/ β . A canonical household service technology maps any given service process {st } in L20 into a corresponding consumption process {ct } for which the implied household capital stock process {ht } is also in L20 . To verify this, we use the canonical – 239 –

240

Representation of Demand

representation to obtain a recursive representation for the consumption process in terms of the service process: ct = −Π−1 Λht−1 + Π−1 st

ht = (∆h − Θh Π−1 Λ)ht−1 + Θh Π−1 st

(11.2.1)

The restriction on the eigenvalues of the matrix (∆h − Θh Π−1 Λ) keeps the household capital stock {ht } in L20 .

11.3. Dynamic Demand Functions for Consumption Goods We postpone constructing a canonical representation, and proceed immediately to use one to construct a dynamic demand schedule. In Chapter 6 we derived the following first-order conditions for the household’s optimization problem: st = bt − µst

(11.3.1)

0 Π′ µst = −Θ′h µht + µw 0 pt

(11.3.2)

µht = βEt (Λ′ µst+1 + ∆′h µht+1 ).

(11.3.3)

As a prelude to computing demand for consumption, we compute the demand for services. Our strategy is to use (11.3.2) and (11.3.3) to solve for the multiplier µst , and then to substitute this solution into (11.3.1). Shift (11.3.2) forward one time period and solve (11.3.2) for µst+1 . Substitute this expression into (11.3.3): ′ −1′ 0 µht = βEt (−Λ′ Π−1′ Θ′h µht+1 + µw pt+1 + ∆′h µht+1 ). 0Λ Π

(11.3.4)

Solve (11.3.4) forward to obtain: µht

=

µw 0 Et

∞ X

τ =1

β τ (∆′h − Λ′ Π−1′ Θ′h )τ −1 Λ′ Π−1′ p0t+τ .

(11.3.5)

Dynamic Demand Functions for Consumption Goods

241

Because we are using a canonical household service technology, the infinite sum on the right side of (11.3.5) converges (in L20 ). Therefore, the service demand can be expressed as 0 st = bt − µw 0 ρt

(11.3.6)

where ∞ h i X ρ0t ≡ Π−1′ p0t − Θ′h Et β τ (∆′h − Λ′ Π−1′ Θ′h )τ −1 Λ′ Π−1′ p0t+τ .

(11.3.7)

τ =1

Equations (11.3.6) and (11.3.7) represent the service demands in terms of expected future prices of the consumption good. The random vector ρ0t is the implicit rental price for services expressed in terms of current and expected future prices of consumption goods. Equation (11.2.1) transforms {st } in L20 into {ct } in L20 .

11.3.1. The multiplier µw0 The service demands given in (11.3.6) depend on the endogenous scalar mulw tiplier µw 0 . To compute µ0 , we partition the household capital and service sequences into two components. One component is a service sequence obtained from the initial endowment of household capital. The other component is the service sequence obtained from market purchases of consumption goods. The service sequence {si,t } obtained from the initial endowment of household capital evolves according to: si,t = Λhi,t−1 (11.3.8) hi,t = ∆h hi,t−1 where hi,−1 = h−1 . The service sequence {sm,t } obtained from purchases of consumption satisfies: 0 sm,t = bt − si,t − µw 0 ρt .

(11.3.9)

We can compute the time zero cost of the sequence {sm,t } in one of two equivalent ways. One way is to compute the time zero cost of the consumption sequence {ct } needed to support the service demands using the price sequence

242

Representation of Demand

{p0t }. Another way is to use the implicit rental sequence {ρ0t } directly to compute the time zero costs of {sm,t }. In the appendix to this chapter, we verify that the two measures of costs agree: E0

∞ X t=0

β t ρ0t · sm,t = E0

∞ X t=0

β t p0t · ct .

(11.3.10)

It is reasonable that, starting from h−1 = 0, the value of services equals the value of the associated consumption stream. It follows from (11.3.9) that

E0

∞ X t=0

β t ρ0t · sm,t = E0

∞ X t=0

β t ρ0t · (bt − si,t ) − µw 0 E0

∞ X t=0

β t ρ0t · ρ0t .

(11.3.11)

Substitute (11.3.10) and (11.3.11) into the consumer’s budget constraint (6.2), and solve for the time zero marginal utility of wealth µw 0 : µw 0 =

E0

P∞

β t ρ0t · (bt − si,t ) − W0 P∞ , E0 t=0 β t ρ0t · ρ0t

t=0

(11.3.12)

where W0 denotes initial period wealth given by W0 = E 0

∞ X t=0

β t (wt0 ℓt + αt0 · dt ) + v0 · k−1 .

(11.3.13)

Taken together, (11.3.6), (11.3.7), (11.3.12) and (11.3.13) give the demand functions for consumption services. A recursive representation for the dynamic demand function for consumption goods is obtained by substituting for st in (11.2.1).

Dynamic Demand Functions for Consumption Goods

243

11.3.2. Dynamic Demand System Substituting (11.3.6) and (11.3.7) into (11.2.1) gives ′ −1 ct = −Π−1 Λht−1 + Π−1 bt − Π−1 µw − Π′ −1 Θ′h 0 Et {Π

[I − (∆′h − Λ′ Π′ −1 Θ′h )βL−1 ]−1 Λ′ Π′−1 βL−1 }p0t

(11.3.14)

ht = ∆h ht−1 + Θh ct

Equation system (11.3.14) can be regarded as a system of dynamic demand functions for consumption, which express consumption demand at date t as a function of future scaled Arrow-Debreu prices p0t and, as mediated through the state variable ht−1 , past values of consumption.

11.3.3. Foreshadow of Gorman aggregation In the chapter 12, we shall explore how the dynamic demand schedule for consumption goods opens up the possibility of satisfying Gorman’s (1953) conditions for aggregation in a heterogeneous consumer version of the model. The first equation of (11.3.14) amounts to an Engle curve for consumption that is linear 1 in individual wealth with a coefficient on µw o (which depends on wealth) that only depends on prices. In a model of consumers who have the same household technologies (∆h , Θh , Λ, Π ) but possibly different preference shock processes, the coefficient on wealth is the same for all consumers. Gorman showed that when Engel curves satisfy this property, there exists a unique community or aggregate preference ordering over aggregate consumption that is independent of the distribution of wealth. This property will be exploited in chapter 12 when we solve for the equilibrium of a multiple consumer version of our economy. The community dynamic demand schedule for a heterogeneous agent economy will be obtained by summing the individual Engel curves.

1 Through ( 11.3.12 ) the multiplier µw depends on wealth in an affine relationship. 0

244

Representation of Demand

11.4. Computing Canonical Representations In deriving a dynamic demand function, we assumed that the representation of the household service technology is canonical. Now we start with a preference shock process {bt } and a specification of (∆h , Θh , Λ, Π) that is not necessarily canonical and show how to find a canonical representation. In the appendix, we establish that for any (∆h , Θh , Λ, Π), there exists a canonical service technology ˆ Π) ˆ and accompanying preference shock process {ˆbt } that induces (∆h , Θh , Λ, an identical preference ordering over consumption. In the text, we display the mechanics of how to compute the canonical technology and associated preference shock process, assigning the technical details to the appendix. 2 These mechanics are closely related to mathematics of innovations representations.

11.4.1. Heuristics We study two polynomials in the lag operator L: σ(L) = Π + ΛL[I − ∆h L]−1 Θh ˆ + ΛL[I ˆ σ ˆ (L) = Π − ∆h L]−1 Θh . As explained in the appendix, when ct = 0∀t < 0, applying the operator σ(L) to ct gives st , so that st = σ(L)ct . For two household technologies [∆h , Θh , Π, Λ] ˆ h , Θh , Π, ˆ Λ] ˆ to give rise to the same preference ordering over {ct } it is and [∆ necessary that σ(β .5 L−1 )′ σ(β .5 L) = σ ˆ (β .5 L−1 )′ σ ˆ (β .5 L). ˆ Π] ˆ technology is to be canonical, it is necessary that σ If the [Λ, ˆ (β .5 L) be invertible. In the appendix, we verify the following version of the factorization identity: [Π + β 1/2 L−1 Λ(I − β 1/2 L−1 ∆h )−1 Θh ]′ [Π + β 1/2 LΛ(I − β 1/2 L∆h )−1 Θh ] ˆ + β 1/2 L−1 Λ(I ˆ − β 1/2 L−1 ∆h )−1 Θh ]′ [Π ˆ + β 1/2 LΛ(I ˆ − β 1/2 L∆h )−1 Θh ], = [Π

ˆ Λ] ˆ satisfy (10.16), (10.19), and (10.20) below. As part of the factorwhere [Λ, ˆ Π] ˆ representation satisfies both of the ization identity, it is proved that the [Λ, requirements to achieve the status of a canonical representation. Thus, to attain a canonical household technology, we have to implement this factorization. We can do this by solving a control problem. 2 The MATLAB program canonpr.m computes a canonical representation.

Computing Canonical Representations

245

11.4.2. An auxiliary problem that induces a canonical representation An artificial optimization problem and the associated optimal linear regulator facilitate computing a canonical representation. Thus, confront a household with the optimization problem: choose {ct }∞ t=0 to maximize −.5

∞ X t=0

β t (st − bt ) · (st − bt )

subject to ht = ∆h ht−1 + Θh ct st = Λht−1 + Πct . The recursive solution to this optimization problem contains all of the ingredients for a canonical service technology. This optimization problem is a version of one a household confronts in a competitive equilibrium, except that we have eliminated the budget constraint. For a canonical technology, the solution to this optimization problem is trivial: choose {ct } so that the implied service sequence matches the preference shock sequence, st = bt ∀ t . However, when the service technology is not canonical, it might not be feasible to construct a consumption process that attains that goal, in which case the optimization problem is not trivial. We simplify the household optimization problem further by initially setting the preference shock process to zero for all t ≥ 0. In making this simplification, we are exploiting the fact that for the optimal linear regulator problem, the feedback part of the decision rule can be computed independently of the feedforward part, and that the {bt } process influences only the feedforward part. In this optimization problem it is feasible to stabilize the state vector {ht } so that it satisfies the square summability requirement. For instance, one can set the consumption process to zero for all t ≥ 0. So long as it is also optimal to stabilize the household capital stock process, it is known that there will be a unique positive semidefinite matrix P satisfying the algebraic Riccati equation: 3 P = Λ′ Λ + β∆′h P ∆h − (β∆′h P Θh + Λ′ Π)

(Π′ Π + βΘ′h P Θh )−1 (βΘ′h P ∆h + Π′ Λ).

(11.4.1)

3 We require that assumption A1 and the stability theorem of chapter 9 apply to this control problem.

246

Representation of Demand

The optimal choice of consumption can be represented as ct = −(Π′ Π + βΘ′h P Θh )−1 (βΘ′h P ∆h + Π′ Λ)ht−1 .

(11.4.2)

When this optimal rule is implemented, the evolution equation for the household capital stock is ht = [∆h − Θh (Π′ Π + βΘ′h P Θh )−1 (βΘ′h P ∆h + Π′ Λ)]ht−1 ,

(11.4.3)

where the eigenvalues of the matrix multiplying ht−1 are strictly less than √ ˆ and Λ ˆ so that 1/ β . 4 With this in mind, we choose Π ˆ −1 Λ ˆ = (Π′ Π + βΘ′h P Θh )−1 (βΘ′h P ∆h + Π′ Λ). Π

(11.4.4)

For this choice, condition (ii) for a canonical service technology is met. ˆ . In the appendix, it is shown as an implication We still have to construct Π ˆ to be a factor of the symmetric of the factorization identity that we should set Π ′ ′ positive definite matrix (Π Π + βΘh P Θh ): ˆ ′ Π. ˆ (Π′ Π + βΘ′h P Θh ) = Π

(11.4.5)

ˆ is a square matrix. Since (Π′ Π + Any factorization will work so long as Π ′ ˆ βΘh P Θh ) is nonsingular, Π satisfies condition (i) for a canonical representation. ˆ Λ) ˆ that corIn summary, (11.4.1), (11.4.4), and (11.4.5) compute a (Π, responds to a canonical representation. The service process {ˆ st } for this new household technology satisfies: ˆ t−1 + Πc ˆ t. sˆt = Λh

(11.4.6)

We also need to construct a preference shock process to accompany the canonical service technology. One way to do this is simply to reintroduce the preference shock process {bt } into the auxiliary household optimization problem, and recompute the optimal decision rule for consumption. The decision rule can be represented as: ˆ t−1 + (Π) ˆ −1 U ˆ b zt ˆ −1 Λh ct = −(Π)

(11.4.7)

4 This condition on the eigenvalues of the ‘closed loop system’ follows from the assumption that it is optimal to stabilize the system (i.e., that the system is detectable).

Computing Canonical Representations

247

ˆb . As discussed in chapter 4, the feedback portion of this for some matrix U −1 ˆ ˆ decision rule [(Π) Λ] is the same as for the problem in which the preference ˆ −1 U ˆb ] can be computed shock process was set to zero. The feedforward part [(Π) using the method described in chapter 4, which permits the optimal decision rule to be calculated efficiently in two steps. Using those methods, the shock process associated with the canonical service technology is ˆbt = U ˆ b zt .

(11.4.8)

An alternative method for computing {ˆbt } is more useful and revealing. As shown in the appendix to this chapter, two preference representations having the same demand functions give rise to the same preference ordering over consumption paths. Therefore, the marginal utilities are also the same across the two preference representations, and in particular across the two specifications of household technologies and preference shock processes. Equality between the indirect marginal utility of consumption and the current and expected future marginal utilities of consumption services and (11.3.1), (11.3.2), and (11.3.3) implies that the two preference shock processes must satisfy: ′

Π bt +

Θ′h Et

∞ X

β

τ

(∆′h )τ −1 Λ′ bt+τ

τ =1

ˆ ′ˆbt + Θ′ Et =Π h

∞ X

ˆ ′ˆbt+τ . β τ (∆′h )τ −1 Λ

τ =1

(11.4.9) ˆ Π) ˆ technology Let the left side of (11.4.9) be denoted ˜bt for each t . Since the (Λ, is canonical, it follows that we can solve (11.4.9) for ˆbt : ˆbt = Π ˆ −1′˜bt − Π ˆ −1′ Θ′ Et h

∞ X

τ =1

ˆ −1′ Θ′ )τ −1 Λ ˆ ′Π ˆ −1′˜bt+τ . ˆ ′Π β τ (∆′h − Λ h

(11.4.10)

Relation (11.4.10) is derived by applying an operator identity to equation (11.4.9).

248

Representation of Demand

11.5. Operator Identities For canonical household technologies a matrix identity is [Π + Λ(I − ∆h L)−1 Θh L]−1 =

{Π−1 −Π−1 Λ[I − (∆h − Θh Π−1 Λ)L]−1 Θh Π−1 L}.

The identity shows that for canonical representations of preferences (∆h , Θh , Π, Λ, Ub ) , there are two equivalent ways of expressing the mapping between sequences {st } and sequences {ct } . To establish the identity, assume that h−1 = 0 , or equivalently that ct = 0 ∀ t < 0 . Note that the second equation of representation ( 11.2.1 ) implies ht = [I − (∆h − Θh Π−1 Λ)L]−1 Θh Π−1 st .

Lagging this one period and substituting into the first equation of ( 11.2.1 ) gives ct = {Π−1 − Π−1 Λ[I − (∆h − Θh Π−1 Λ)L]−1 Θh Π−1 L}st .

This equation shows how to obtain sequences {ct } ∈ L20 that are associated with arbitrary sequences {st } ∈ L20 . Now recall that the household technology implies st = [Π + Λ(I − ∆h L)−1 Θh L]ct ,

which expresses {st } ∈ L20 as a function of {ct } ∈ L20 . The assumption that (Λ, Π) is canonical implies that the operator [Π + Λ(I − ∆h L)−1 Θh L] mapping sequences from L20 into L20 is invertible, which implies the identity. Here is how to derive the ‘dual’ or transposed version of the identity, which is the one used to get ( 11.4.10 ). Use ( 11.3.3 ) to deduce ′ −1 )−1 βΛ′ L−1 µs . µh t = (I − β∆h L t+1

Then use ( 11.3.2 ) to deduce (†)

′ ′ ′ −1 )−1 βL−1 ]µs . µw t 0 pt = [Π + Θh (I − β∆h L

Alternatively, solve ( 11.3.2 ) for µst ,

w 0 µst = Π′ −1 (−Θ′h µh t + µ0 pt ).

Substitute this into ( 11.3.3 ) to get (‡)

0 µst = {Π′ −1 − Π′ −1 Θ′h [I − (∆′h − Λ′ Π′ −1 Θ′h )βL−1 ]−1 Λ′ Π′−1 βL−1 }µw 0 pt .

When (∆h , Θh , Π, Λ) is canonical, the operator on the right side of (†) has an inverse equal to the operator on the right side of (‡) : [Π′ + Θ′h (I − β∆′h L−1 )−1 βL−1 ]−1 =

{Π′ −1 −Π′ −1 Θ′h [I − (∆′h − Λ′ Π′ −1 Θ′h )βL−1 ]−1 Λ′ Π′−1 βL−1 }.

In the appendix to this chapter, we use Fourier transforms to show that the ˆ Π) ˆ and preference shock process {ˆbt } induce alternative service technology (Λ, the same preference ordering for consumption goods as did the original ones.

Becker-Murphy Model of Rational Addiction

249

11.6. Becker-Murphy Model of Rational Addiction We illustrate our analysis with a discrete-time version of the habit-persistence model advocated by Becker and Murphy (1988). The household technology is a parametric version of induced preferences for consumption of the form suggested by Pollak (1970), Ryder and Heal (1973), and Stigler and Becker (1977). The household technology has a single consumption good, two services, and a single household capital stock. The household capital measures a habit stock constructed to be a geometrically-weighted average of current and past consumptions: ht = δh ht−1 + (1 − δh )ct , (11.6.1) where 0 < δh < 1. The first service is proportional to consumption, and the second one is a linear combination of consumption and the habit stock: π1 0 ct st = . (11.6.2) π2 π3 ht We normalize π1 and π3 to be strictly positive. Imagine for a moment that ct and ht are distinct consumptions goods and that there is no intertemporal connection between them. Then recall from our discussion of preferences for multiple consumption goods in Chapter 3, that the Frisch classification of complements is equivalent to requiring π2 to be negative. In light of the evolution equation (11.6.1) for the household capital stock, this notion of complementarity is limiting because it ignores the fact that ht is a weighted average of current and past consumptions. For this reason, we consider a related notion of complementarity referred to by Ryder and Heal (1973) and Becker and Murphy (1988) as adjacent complementarity. Substituting (11.6.1) into (11.6.2) we obtain the following service technology: st = Λht−1 + Πct , where Λ=

0 π3 δ h

and Π =

(11.6.3)

π1 . π2 + π3 (1 − δh )

Service technology (11.6.3) is clearly not canonical: simply note that two services are constructed from one underlying consumption good, so we cannot construct a consumption sequence to support any hypothetical admissible service sequence.

250

Representation of Demand

To capture the notion of adjacent complementarity, we consider a canonical representation for household services. The canonical household service technology has a single service and can be expressed as: ˆ t−1 + Πc ˆ t, sˆt = Λh

(11.6.4)

ˆ Π) ˆ satisfies: where {ˆ st } is a scalar service process and (Λ, p ˆ Π ˆ |< 1/ β. | δh − (1 − δh )Λ/

(11.6.5)

p ˆ Π ˆ ≤ 1/ β. 0 ≤ δh − (1 − δh )Λ/

(11.6.6)

ˆ to be positive so that increases in time t consumption We normalize the scalar Π increase the time t canonical service sˆt . When specialized to this parametric model, Ryder and Heal’s (1973) notion of adjacent complementarity becomes ˆ must be negative. In this case, (11.6.5) implies that the restriction that Λ

ˆ ≤ 0) As shown by Becker and Murphy (1988), adjacent complementarity (Λ implies that π2 ≤ 0. The converse is not true, however. To see the relation ˆ and π2 , multiply both sides of (9.64) by (1−β 1/2 ζ −1 δh )(1−β 1/2 ζδh ) between Λ to obtain: Π′ Π (1 − β 1/2 ζ −1 δh )(1 − β 1/2 ζδh ) + βΛ′ Λ(1 − δh )2 +

β 1/2 ζ −1 (1 − δh )(1 − β 1/2 ζδh )Λ′ Π + β 1/2 ζ(1 − δh )(1 − β 1/2 ζ −1 δh )Λ′ Π ˆ 2 (1 − β 1/2 ζ −1 δh )(1 − β 1/2 ζδh ) + β Λ ˆ Λ(1 ˆ − δh )2 + =Π

ˆΠ ˆ + β 1/2 ζ(1 − δh )(1 − β 1/2 ζ −1 δh )Λ ˆ Π. ˆ β 1/2 ζ −1 (1 − δh )(1 − β 1/2 ζδh )Λ (11.6.7) This equality holds for all ζ except ζ = 0. Evaluate both sides of (11.6.7) at ζ = β 1/2 δh : βΛ′ Λ(1 − δh )2 + (1 − δh )(1 − βδh2 )Λ′ Π/δh ˆ 2 (1 − δh )2 + (1 − δh )(1 − βδh2 )Λ ˆ Π/δ ˆ h. = βΛ

(11.6.8)

The right side of (11.6.8) can be expressed as ˆ 2 (1 − δh )2 + (1 − δh )(1 − βδ 2 )Λ ˆ Π/δ ˆ h βΛ h ˆ Π{[(1 ˆ ˆ Π ˆ − δh ] + (1/βδh )}. = β(1 − δh )Λ − δh )Λ/

(11.6.9)

Fourier transforms

251

ˆΠ ˆ < 0 and inequality (11.6.6) is satisfied, it follows that Since Λ ˆ Π{[(1 ˆ ˆ Π ˆ − δh ] + (1/βδh )} β(1 − δh )Λ − δh )Λ/ p ˆ Π[−1/ ˆ ≤ β(1 − δh )Λ β + (1/βδh )]

(11.6.10)

≤ 0.

ˆ ≤ 0, then Combining (11.6.10) and (11.6.8), we have that if Λ βΛ′ Λ(1 − δh )2 + (1 − δh )(1 − βδh2 )Λ′ Π/δh ≤ 0.

(11.6.11)

Inequality (11.6.11) is satisfied only when Λ′ Π ≤ 0. This in turn requires that π2 ≤ 0 because Λ′ Π = π3 δh [π2 + π3 (1 − δh )], (11.6.12) 0 < δh < 1 and π3 > 0. ˆ Λ ˆ to exceed one. In this case, growth Inequality (11.6.6) permits δh −(1−δh )Π/ in consumption is required to support most constant service sequences, although this growth will be dominated by {β t/2 : t = 0, 1, . . .}. This is a household technology with an extreme form of addiction to the consumption good. Note that ˆ Π ˆ = δh (1 + Λ/ ˆ Π) ˆ − Λ/ ˆ Π. ˆ δh − (1 − δh )Λ/ (11.6.13) ˆ exceeds Π ˆ in the canonical houseTherefore, instability is implied whenever −Λ hold service technology.

A. Fourier transforms This appendix applies Fourier transforms to establish some key equalities asserted in the text. We begin with some background on 0.

252

Representation of Demand

11.A.1. Primer on transforms For a two-sided scalar sequence {cj }∞ j=−∞ , the z− transform is defined as the complex valued function ∞ X cj z j , c(z) = j=−∞

5

where z is a scalar complex number. The inversion formula asserts Z 1 ck = c(z)z −k−1 dz 2πi Γ where Γ is any closed contour around zero in the complex plane, and the integration is complex integration counterclockwise along the path Γ . If we take Γ to be the unit circle and set z = e−iω , we get the following version of the inversion formula Z π 1 c(e−iω )eiωk d ω. ck = 2π −π We denote transform pairs with the notation {ck } ↔ c(z). The convolution of two sequences {yk }, {xk }, is denoted {y ∗ x} and is defined as ∞ X {y ∗ x}∞ ≡ { ys xk−s }∞ k=−∞ k=−∞ . s=−∞

Direct calculations establish the convolution property {y ∗ x}∞ k=−∞ ↔ x(z)y(z). We have the linearity property that for any scalars (a, b) a{xk } + b{yk } ↔ ax(z) + by(z).

5 For descriptions of Fourier and z-transforms, see Gabel and Roberts (1973), Liu and Liu (19**). For some of their uses in time series economics see Nerlove, Grether and Carvalho (1979) and Sargent (1979).

Fourier transforms

253

11.A.2. Time reversal and Parseval’s formula Let c˜−k = ck for all k . Then {˜ ck }∞ k=−∞ has transform c˜(z) =

∞ X

∞ X

c˜k z k =

ck z −k = c(z −1 ).

k=−∞

k=−∞

Applying the convolution theorem to c(z)c(z −1 ) gives c(z)c(z

−1

∞ X

)↔{

s=−∞

cs cs−k }∞ k=−∞ .

Applying the inversion formula gives ∞ X

cs cs−k

s=−∞

1 = 2π

Z

π

c(e−iω )c(eiω )eiωk d ω.

−π

If we set k = 0, we obtain Parseval’s equality: ∞ X

s=−∞

c2s

1 = 2π

Z

π

−π

|c(e−iω )|2 d ω.

11.A.3. One sided sequences There are two types of one-sided sequences (also called ‘half-infinite’ sequences). A sequence is called a causal sequence if ck = 0 ∀k < 0, and is anti-causal if it has zero elements ∀ k > 1. A one-sided causal sequence can be obtained by setting to zero all elements of a two-sided sequence with negative subscripts. Let {uk }∞ k=−∞ be the step sequence, which is zero for k < 0, and 1 for k ≥ 0. Evidently {uk ck } is always a one-sided sequence.

254

Representation of Demand

11.A.4. Useful properties 1. z0 is said to be a pole of order m ≥ 1 of c(z) if limz→z0 (z − z0 )m c(z) 6= 0. 2. c(z) is the transform of a causal sequence if all of its poles lie outside the unit circle. 3. c(z) is the transform of an anti-causal sequence if all of its poles lie inside the unit circle. 4. If c(z) is either causal or anti-causal, the inversion formula can be implemented by ‘long division.’ 5. Initial value theorem: lim c(z) = c0 .

z→0

6. Final value theorem: lim ck = lim (1 − z)c(z). z→1

k→∞

11.A.5. One sided transforms A one-sided transform is defined as c+ (z) =

∞ X

k=0

ck z k ≡ [c(z)]+ ,

where [ ]+ is the ‘annihilation operator’ that sets to zero all coefficients on negative powers of z . The same inversion formulas hold, with c+ (z) replacing c(z). Notice that c+ (z) = c(z) only if {ck } is causal. We shall adopt the notation F (c)(z) = c+ (z). For one-sided transforms, we have the shift theorem F ({ct−n })(z) = z n F ({ct })(z) +

n X

k=1

z n−k c−k .

Fourier transforms

255

11.A.6. Discounting For the purpose of introducing discounting, we shall work with the alternative transformation defined by t/2 ∞ T ({ct }∞ }t=0 )(z), t=0 )(z) ≡ F ({ct β

so that T (y) is the ordinary transform of {β t/2 yt }. The inversion formula is then Z π 1 β t/2 yt = T (e−iω )eiωt d ω, 2π −π and the shift theorem is .5

n

T ({ct−n })(z) = (β z) F ({ct })(z) +

n X

(zβ .5 )n−k c−k .

k=1

11.A.7. Fourier transforms Below we shall work with vector versions of the transforms T . Consider a vector sequence y = {yt } satisfying ∞ X t=0

β t yt · yt < ∞,

define the transform: T (y)(ζ) ≡

∞ X

(11.A.1)

β t/2 yt ζ t .

(11.A.2)

t=0

This transform is at least well-defined for | ζ |< 1 and can also be defined through an appropriate limiting argument for | ζ |= 1. 6 For vector sequences {yt } and {ˆ yt } satisfying (11.A.1), Parseval’s formula is (1/2π)

Z

π −π

T (y)[exp(iθ)] · T (ˆ y )[exp(−iθ)]dθ =

∞ X t=0

β t (yt · yˆt ).

(11.A.3)

6 The boundary of the unit circle can be parameterized by ζ = exp(iθ) for θ ∈ (−π, π] . Using this parameterization, the infinite series on the right side of ( 11.4.6 ) converges in L2 where the L2 space is constructed using Lebesgue measure on (−π, π] .

256

Representation of Demand

We use Fourier 0 to represent our dynamic household technologies. It follows from (11.2.1) and the definitions of smt and sit that ΠT (c)(ζ) = −β 1/2 ζΛT (hm )(ζ) + T (sm )(ζ)

T (hm )(ζ) =β 1/2 ζ(∆h − Θh Π−1 Λ)T (hm )(ζ) + Θh Π−1 T (sm )(ζ)

(11.A.4)

where hm,−1 = 0. The transforms of the consumption sequence and the market service sequence are related by T (c)(ζ) = C (ζ)T (sm )(ζ)

(11.A.5)

where o n C (ζ) ≡ Π−1 I − β 1/2 ζΛ[I − β 1/2 ζ(∆h − Θh Π−1 Λ)]−1 Θh Π−1 .

(11.A.6)

The matrix function C of a complex variable ζ represents the mapping from desired consumption services into the consumption goods required to support those services.

11.A.8. Verifying Equivalent Valuations Our derivation of the dynamic demand functions for consumption goods relied on two intermediate results: (a) equivalent 0 of market services and consumption goods asserted in (11.3.10); and (b) for a given specification of preferences and household technology, the existence of a canonical service technology that induces the same preference ordering over consumption streams. To establish these intermediate results we use Fourier transforms. We now show establish the valuation equivalence asserted in (11.3.10). Applying Parseval’s formula (11.A.3), we have that ∞ X

β t p0t

t=0

= (1/2π) = (1/2π)

· ct = (1/2π)

Z

π

−π Z π −π

Z

π −π

T (p0 )[exp(iθ)] · T (c)[exp(−iθ)]dθ

T (p0 )[exp(iθ)] · {C [exp(−iθ)]T (sm )[exp(−iθ)}]dθ {C [exp(−iθ)]′ T (p0 )[exp(iθ)]} · T (sm )[exp(−iθ)]dθ.

(11.A.7)

Fourier transforms

257

Formula (11.A.7) gives us the following candidate for the transform of the rental sequence for consumption services: C (ζ −1 )′ T (p0 )(ζ). The rental sequence {˜ ρ0t } associated with this transform is given by: n o ρ˜0t ≡ Π−1′ I − βL−1 Θ′h [I − βL−1 (∆h − Θh Π−1 Λ)′ ]−1 Λ′ Π−1′ p0t ∞ h i X = Π−1′ p0t − Θ′h β τ (∆h − Θh Π−1 Λ)′τ −1 Λ′ Π−1′ p0t+τ .

(11.A.8)

τ =1

Using this rental sequence, it follows from (11.A.5) that ∞ X t=0

β t p0t · ct =

∞ X t=0

β t ρ˜0t · smt .

(11.A.9)

Notice that the candidate rental sequence {˜ ρ0t } violates the information constraints because ρ˜0t will not necessarily be in Jt . From the vantage point of valuation, all that we require is equality of the expectations of the infinite sums in (11.A.9) conditioned on J0 . It follows from the Law of Iterated Expectations that E0 ρ˜0t · st = E0 ρ0t · st (11.A.10) where ρ0t ≡ Et ρ˜0t ,

(11.A.11)

since hypothetical service vectors st are restricted to be in the information set Jt . Taking expectations of both sides of (11.A.11) conditioned on J0 and substituting from (11.A.11) establishes the value equivalence given in (11.3.10).

258

Representation of Demand

11.A.9. Equivalent representations of preferences We now turn to task (b), to show that the candidate canonical representation of the service technology implies the same induced preference ordering for conˆ Π), ˆ sumption. There are two preference representations on the table (Λ, Π), (Λ, where the objects with hats are canonical. Again we partition the household capital stock and the consumption service process into two components. Similar to (11.A.4) we have that T (sm )(ζ) = β 1/2 ζΛT (hm )(ζ) + ΠT (c)(ζ) T (hm )(ζ) = β 1/2 ζ∆h T (hm )(ζ) + Θh T (c)(ζ).

(11.A.12)

Hence T (sm )(ζ) = S (ζ)T (c)(ζ)

(11.A.13)

S (ζ) ≡ [Π + β 1/2 ζΛ(I − β 1/2 ζ∆h )−1 Θh ].

(11.A.14)

where The function S represents the mapping from consumption goods into market supplied consumption services. An analogous argument leads to the formula: T (ˆ sm )(ζ) = Sˆ(ζ)T (c)(ζ)

(11.A.15)

ˆ + β 1/2 ζ Λ(I ˆ − β 1/2 ζ∆h )−1 Θh ], Sˆ(ζ) ≡ [Π

(11.A.16)

where where objects with hats, including sˆm , correspond to the canonical representation. It is straightforward to show that T (si )(ζ) = Λ(I − β 1/2 ζ∆h )−1 h−1

(11.A.17)

ˆ − β 1/2 ζ∆h )−1 h−1 . T (ˆ si )(ζ) = Λ(I

(11.A.18)

and The time t contribution to the consumers’ utility function can be expressed as: h i −(1/2)β t (bt − si,t − sm,t ) · (bt − si,t − sm,t ) = −(1/2)β t sm,t · sm,t + 2sm,t · si,t − 2sm,t · bt + (11.A.19) (bt − si,t ) · (bt − si,t ) .

Fourier transforms

259

Note that the fourth term is not affected by the consumption choice, and thus can be ignored. We now study the Fourier representations of the sums: ∞ X t=0

β t sm,t · sm,t ,

∞ X t=0

β t sm,t · si,t and

∞ X t=0

β t sm,t · bt .

(11.A.20)

11.A.10. First term: factorization identity The first infinite sum in (11.A.20) can be represented as: ∞ X

(1/2π)

β t sm,t · sm,t =

t=0 πn

Z

−π

o′ T (c)[exp(iθ)] S [exp(iθ)]′ S [exp(−iθ)]

(11.A.21)

T (c)[exp(−iθ)]dθ.

ˆ Λ) ˆ and {ˆbt } imply the same induced preferences for consumpTo show that (Π, tion goods, we must first establish the factorization: S (ζ −1 )′ S (ζ) = Sˆ(ζ −1 )′ Sˆ(ζ).

(11.A.22)

To verify this result, note that [Π + β 1/2 ζ −1 Λ(I − β 1/2 ζ −1 ∆h )−1 Θh ]′ [Π + β 1/2 ζΛ(I − β 1/2 ζ∆h )−1 Θh ] = Π′ Π + βΘ′h (I − β 1/2 ζ −1 ∆h )−1′ Λ′ Λ(I − β 1/2 ζ∆h )−1 Θh

+ β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ Λ′ Π

+ β 1/2 ζΠ′ Λ(I − β 1/2 ζ∆h )−1 Θh .

(11.A.23) Since P satisfies the algebraic Riccati equation(11.4.1), it follows that ˆ ′Λ ˆ Λ′ Λ = P − β∆′h P ∆h + Λ

= (I − β 1/2 ζ −1 ∆h )′ P (I − β 1/2 ζ∆h ) + β 1/2 ζ −1 ∆′h P (I − β 1/2 ζ∆h ) (11.A.24) ˆ ′ Λ. ˆ + β 1/2 ζ(I − β 1/2 ζ −1 ∆h )′ P ∆h + Λ

260

Representation of Demand

Therefore, Θ′h (I − β 1/2 ζ −1 ∆h )−1′ Λ′ Λ(I − β 1/2 ζ∆h )−1 Θh

= Θ′h P Θh + β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ ∆′h P Θh

+ β 1/2 ζΘ′h P ∆h (I − β 1/2 ζ∆h )−1 Θh ˆ ′ Λ(I ˆ − β 1/2 ζ∆h )−1 Θh . + Θ′ (I − β 1/2 ζ −1 ∆h )−1′ Λ

(11.A.25)

h

Furthermore, it follows from (11.4.4) and (11.4.5) that ˆ ′Λ ˆ =Π ˆ ′ Π( ˆ Π) ˆ −1 Λ ˆ Π = (βΘ′h P Θh + Λ′ Π).

(11.A.26)

Substituting (11.A.25) and (11.A.26) into (11.A.23) results in [Π + β 1/2 ζ −1 Λ(I − β 1/2 ζ −1 ∆h )−1 Θh ]′ [Π + β 1/2 ζΛ(I − β 1/2 ζ∆h )−1 Θh ] ˆ ′ Λ(I ˆ − β 1/2 ζ∆h )−1 Θh = Π′ Π + βΘ′h P Θh + βΘ′h (I − β 1/2 ζ −1 ∆h )−1′ Λ + β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ (β∆′h P Θh + Λ′ Π) + β 1/2 ζ(ΠΛ′ + βΘ′h P ∆h )(I − β 1/2 ζ∆h )−1 Θh ˆ ′Π ˆ + βΘ′h (I − β 1/2 ζ −1 ∆h )−1′ Λ ˆ ′ Λ(I ˆ − β 1/2 ζ∆h )−1 Θh =Π ˆ ′Π ˆ + β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ Λ ˆ − β 1/2 ζ∆h )−1 Θh ˆ ′ Π(I + β 1/2 ζ Λ ˆ + β 1/2 ζ −1 Λ(I ˆ − β 1/2 ζ −1 ∆h )−1 Θh ]′ [Π ˆ + β 1/2 ζ Λ(I ˆ − β 1/2 ζ∆h )−1 Θh ] = [Π (11.A.27)

which proves factorization (11.A.22).

Fourier transforms

261

11.A.11. Second term The second infinite sum in (11.A.20) can be represented as Z π ∞ X β t sm,t · si,t = (1/2π) {T (c)[exp(iθ)]}′ S [exp(iθ)]′ Λ −π

t=0

[I − β

1/2

−1

exp(−iθ)∆h ]

(11.A.28)

h−1 dθ.

We will verify that S (ζ −1 )′ Λ∆h (I − β 1/2 ζ∆h )−1 = ˆ h (I − β 1/2 ζ∆h )−1 + β 1/2 ζ −1 Θ′ (I − β 1/2 ζ −1 ∆h )−1′ P ∆h . Sˆ(ζ −1 )′ Λ∆ h (11.A.29) It then follows that (1/2π)

Z

π

−π π

= (1/2π)

Z

−π

{T (c)[exp(iθ)]}′ S[exp(iθ)]′ Λ∆h [I − β 1/2 exp(−iθ)∆h ]−1 h−1 dθ ′ˆ ˆ {T (c)[exp(iθ)]}′ S[exp(iθ)] Λ∆h [I − β 1/2 exp(−iθ)∆h ]−1 h−1 dθ

(11.A.30)

because

(1/2π)

Z

π −π

{T (c)[exp(iθ)]}′ β 1/2 exp(iθ) Θ′h [I

−β

1/2

−1′

exp(iθ)∆h ]

(11.A.31)

P ∆h h−1 dθ = 0.

Relation (11.A.31) holds since T (c)(ζ)′ β 1/2 ζ∆′h (I − β 1/2 ζ∆h )−1′ has a power series expansion and is zero when ζ = 0 and P ∆h h−1 can be viewed a constant function with a trivial power series expansion. Relation (11.A.31) then follows from Parseval’s formula (11.A.3) where β t/2 yt is constructed from the tth coefficient of the power series expansion for the first function and β t/2 y˜t from the tth coefficient of the power series expansion for the second function. It remains to establish (11.A.29). Note that the left side of (11.A.29) can be expanded as follows: [Π + β 1/2 ζ −1 Λ(I − β 1/2 ζ −1 ∆h )−1 Θh ]′ Λ(I − β 1/2 ζ∆h )−1 ∆h

= Π′ Λ(I − β 1/2 ζ∆h )−1 ∆h + β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ Λ′ Λ (11.A.32)

(I − β 1/2 ζ∆h )−1 ∆h .

262

Representation of Demand

It follows from the algebraic Riccati equation (11.4.1) that ˆ ′Λ ˆ Λ′ Λ = P (I − β 1/2 ζ∆h ) + β 1/2 ζP ∆h − β∆′h P ∆h + Λ

ˆ ′ Λ, ˆ = P (I − β 1/2 ζ∆h ) + β 1/2 ζ(I − β 1/2 ζ −1 ∆′h )P ∆h + Λ

(11.A.33)

and hence β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ Λ′ Λ(I − β 1/2 ζ∆h )−1 ∆h

= β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ P ∆h + βΘ′h P ∆h (I − β 1/2 ζ∆h )−1 ∆h ˆ ′ Λ(I ˆ − β 1/2 ζ∆h )−1 ∆h + β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ Λ (11.A.34) Substituting (11.A.34) and (11.A.26) into (11.A.32) gives [Π + β 1/2 ζ −1 Λ(I − β 1/2 ζ −1 ∆h )−1 Θh ]′ Λ(I − β 1/2 ζ∆h )−1 ∆h = Π′ Λ(I − β 1/2 ζ∆h )−1 ∆h +

β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ Λ′ Λ(I − β 1/2 ζ∆h )−1 ∆h = (Π′ Λ + βΘ′h P ∆h )(I − β 1/2 ζ∆h )−1 ∆h +

β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ P ∆h + ˆ ′ Λ(I ˆ − β 1/2 ζ∆h )−1 ∆h β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ Λ ˆ ′ Λ/(I ˆ =Π − β 1/2 ζ∆h )−1 ∆h +

ˆ − β 1/2 ζ∆h )−1 ∆h + ˆ ′ Λ(I β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ Λ

β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ P ∆h ˆ − β 1/2 ζ∆h )−1 ∆h + ˆ − β 1/2 ζ −1 ∆h )−1 Θh ]′ Λ(I ˆ + β 1/2 ζ −1 Λ(I = [Π β 1/2 ζ −1 Θ′h (I − β 1/2 ζ −1 ∆h )−1′ P ∆h

which establishes (11.A.29).

(11.A.35)

Fourier transforms

263

11.A.12. Third term The third sum in (11.A.20) can be represented as ∞ X t=0 π

(1/2π)

Z

β t sm,t · bt =

(11.A.36) ′

−π

′

{T (c)[exp(iθ)]} S [exp(iθ)] T (b)[exp(−iθ)]dθ.

Note that

S (ζ −1 )′ T (b)(ζ) = Sˆ(ζ −1 )′ Sˆ(ζ −1 )′ −1 S (ζ −1 )′ T (b)(ζ).

(11.A.37)

With this in mind, we define ˆbt = E{[Sˆ(L−1 )′ ]−1 S (L−1 )′ bt | Jt }.

(11.A.38)

Then by reasoning similar to that leading to result (a), we have that ∞ X t=0

t

β sm,t · bt =

∞ X t=0

β t sˆm,t · ˆbt .

(11.A.39)

Taken together (11.A.22), (11.A.29) and (11.A.37) show that the induced prefˆ Π) ˆ and {ˆbt } as it is for the erence ordering for consumption is the same for (Λ, original specification (Λ, Π) and {bt }. This establishes result (b).

Chapter 12 Gorman Heterogeneous Households

12.1. Introduction This chapter and the next describe methods for computing equilibria of versions of our economies in which consumers have heterogeneous preferences and endowments. In each chapter, we adopt simplifications that make it possible for us to cope with the complications introduced by heterogeneity. In the present chapter, we adopt a version of Terrance Gorman’s (1953) specification. We describe a class of heterogeneous consumer economies that satisfy Gorman’s conditions for aggregation, which lets us compute equilibrium aggregate allocations and prices before computing allocations to individuals. 1 In the following chapter, we adopt a more general specification of heterogeneity that causes us to depart from the representative consumer framework of Gorman. In particular, we adapt the idea of Negishi (1960), who described a social welfare function that is maximized, subject to resource and technological constraints, by a competitive equilibrium allocation. For Negishi, that social welfare function is a “linear combination of the individual utility functions of consumers, with the weights in the combination in inverse proportion to the marginal utilities of income.” Because Negishi’s weights depend on the allocation through the marginal utilities of income, computing a competitive equilibrium via constrained maximization of a Negishi-style welfare function requires solving a fixed point problem in the weights. In the following chapter, we apply such a fixed point approach. The beauty of Gorman’s aggregation conditions is that, when they apply, time series aggregates and market prices can be computed without resorting to Negishi’s fixed point approach. In the present chapter, consumers are permitted to differ only with respect to their endowments and the process {bt } that disturbs their preferences. We 1 The discussion in this chapter is patterned after the material in section 3 of Hansen (1987).

– 265 –

266

Gorman Heterogeneous Households

assume that all consumers have a common information set that includes observations on past values of the economy wide capital stocks ht−1 , kt−1 , and the common exogenous state variables in zt that drive each of the individual preference shock processes and the technology shock process {dt }. Preferences of the individual consumers can be aggregated simply by summing both preference shocks and initial endowments across consumers, thereby forming a representative consumer. We can compute all aggregate aspects of a competitive equilibrium of the economy with heterogeneous consumers by forming the representative consumer and proceeding as in chapters 3, 4, and 5. We show how to calculate individual allocations by using the demand functions that were described in the previous chapter. In the next section, we briefly describe Gorman aggregation in a standard static section before adapting it to our purposes.

12.2. A Digression on Gorman Aggregation Suppose for the moment that there are n consumption goods, taking into account indexation by dates and states, and that consumption of person j = 1, . . . , J is denoted cj . Let ca denote the aggregate amount of consumption to be allocated among consumers. Associated with ca is an Edgeworth box and a collection of Pareto optimal allocations. From the Pareto optimal allocations, one can construct utility allocation surfaces describing the frontier of alternative feasible utility assignments to individual consumers. Imagine moving from the aggregate vector ca to some other vector c˜a and hence to a new Edgeworth box. If neither the original box nor the new box contain one another, then it is possible that the utility allocation surfaces for the two boxes may cross, in which case there exists no ordering of aggregate consumption that is independent of the utility weights assigned to individual consumers. Before describing a special case in which an aggregate social preference ordering does exist, we illustrate a situation in which there doesn’t exist a social preference ordering that is independent of the aggregate allocation. Figures 12.2.1 and Fgpareto2f describe efficient allocations in a two person, two good, pure exchange economy with a structure of preferences that violate the

A Digression on Gorman Aggregation

267

6

Utility of Agent B

5

4

3

2

1

0 0

1

2

3

4

5

6

Utility of Agent A

Figure 12.2.1: Utility allocation for Agents A and B for endowment vectors E = (8,3) and E = (3,8). 1/3

2/3

Gorman aggregation conditions. Agent A has utility function U A = XA YA , 2/3 1/3 while consumer B has utility function given by U B = XB YB , where the aggregate endowment pair is E = (XA + XB , YA + YB ). Figure 12.2.1 shows two utility possibility frontiers, one associated with E = (8, 3), a second one associated with E = (3, 8). 2 The fact that the utility possibility frontiers in figure 12.2.1 cross indicates that the two aggregate endowment vectors (8, 3), (3, 8) cannot be ranked in a way that ignores how utility is distributed between consumers A and B. In the same economy, Figure 12.2.2 shows the Edgeworth boxes and contract curves with the two allocations E = (8, 3) and E = (3, 8). For a given endowment, the slope of the consumers’ indifference curves at the tangencies between indifference curves that determines the contract curve varies as one moves along the contract curve. This means that for a given aggregate endowment, the competitive equilibrium price depends on the allocation between consumers A and B. It follows that for this economy, one cannot expect to determine equilibrium prices independently of the equilibrium allocation. 1/3 2/3 2 A utility possibility frontier is the locus of pairs (U , U ) that solve U = max A B A XA ,YA XA YA 2/3 1/3 subject to the constraints XB YB ≥ U B , (XA + XB , YA + YB ) = E .

268

Gorman Heterogeneous Households

Gorman (1953) described restrictions on preferences under which it is possible to obtain a community preference ordering. Whenever Gorman’s conditions are satisfied, there occur substantial simplifications in solving multiple-consumer optimal resource allocation problems: in intertemporal contexts, it becomes possible first to determine the optimal allocation of aggregate resources over time. Then the aggregate consumption can be allocated among consumers by allocating utility levels to each person. To understand Gorman’s restrictions, imagine specifying the preferences of consumer j in one of two equivalent ways: in terms either of a family of indifference curves indexed by the utility level, or in terms of a family of compensated demand functions. Following Gorman (1953), let ψj (p) denote the baseline indifference curve for person j parameterized in terms of a price vector (or vector of utility gradients) p . In addition, let ψc (p) denote a common indifference curve for all consumers used to measure deviations from the baseline curves. This lets the compensated demand function for person j be represented as

8 7 6 5 4 3 2 1 0 0

1

2

3

4

5

6

7

8

Figure 12.2.2: Overlapping Edgeworth Boxes for endowment vectors E = (8,3) and E = (3,8).

cj = ψj (p) + uj ψc (p) j

(12.2.1)

where u is a scalar utility index for person j . The baseline functions ψj and the common function ψc are the derivatives of concave functions that are

A Digression on Gorman Aggregation

269

positively homogeneous of degree 1. Hence these functions are homogeneous of degree zero in prices, assuring that the slopes of indifference curves should depend only on the ratio of prices. The baseline indifference curves are either the highest or lowest indifference curves, corresponding respectively to cases in which the utility indices are restricted to be nonpositive or nonnegative. As noted by Gorman, when preferences are of this form, there is a well defined compensated demand function for a fictitious representative consumer obtained by aggregating (12.2.1): ca = ψa (p) + ua ψc (p) (12.2.2) where ua =

X

uj and ψa =

X

ψj .

(12.2.3)

In this case, optimal resource allocation in a heterogeneous consumer economy simplifies as follows. Preferences (12.2.2), define a community preference ordering for aggregate consumption. This preference–ordering can be combined with a specification of the technology for producing consumption goods to determine the optimal allocation of aggregate consumption. Mapping (12.2.2) can be inverted to obtain a gradient vector p that is independent of how utilities are allocated across consumers. Since ψc and ψa are homogeneous of degree zero, gradients are only determined up to a scalar multiple. Armed with p , we can then allocate utility among J consumers while respecting the adding up constraint given in (12.2.3). The allocation of aggregate consumption across goods and the associated gradient are determined independently of how the aggregate utility is divided among the individual consumers. A decentralized version of this analysis proceeds as follows. Let W j denote the wealth of consumer j and W a denote aggregate wealth. Then W j should satisfy: W j = p · cj = p · ψj (p) + uj p · ψc (p).

(12.2.4)

Solving (12.2.4) for uj gives uj = [W j − p · ψj (p)]/p · ψc (p). Hence the Engel curve for consumer j is given by

(12.2.5)

270

Gorman Heterogeneous Households

cj = ψj (p) − p · ψj (p)/p · ψc (p) + W j ψc (p)/p · ψc (p).

(12.2.6)

Notice that the coefficient on W j is the same for all j since ψc (p)/p · ψc (p) is only a function of the price vector p in a decentralized economy. The individual allocations can be determined from the Engel curves by substituting for p the gradient vector obtained from the single consumer optimal allocation problem. Individual consumption cj as given by (12.2.6) depends directly on prices through the functions ψj and ψc and indirectly through the evaluation of wealth. For the specifications of preferences adopted in this book, the baseline indifference curves are degenerate because they do not depend on p . A finitedimensional counterpart to this circumstance occurs when ψj (p) = χj ,

(12.2.7)

where χj is a vector with the same dimension as cj . With this specification, the rules for allocating consumption across individuals become linear in aggregate consumption. To see this, observe that an implication of (12.2.2) is ψc (p) = (ca − χa )/ua .

(12.2.8)

Substituting (12.2.8) into (12.2.1) gives cj − χj = (uj /ua )(ca − χa ),

(12.2.9)

so that there is a common scale factor (uj /ua ) across all goods for person j . Hence the fraction of total utility assigned to consumer j determines his fraction of the vector (ca − χa ). Here is an example. Suppose that the preferences of consumer j are represented using the utility function: U j (cj ) = −[(cj − χj )′ V (cj − χj )]1/2 .

(12.2.10)

The compensated demand schedule is then obtained by solving the first-order conditions: V (cj − χj )/U j (cj ) = µj p

(12.2.11)

An Economy with Heterogeneous Consumers

271

U j (cj ) = uj , where µj is a Lagrange multiplier. Substitute the second equation into the first and solve for cj − χj : cj − χj = uj µj V −1 p.

(12.2.12)

Substitute the right side of (12.2.12) into the utility function and solve for the multiplier µj : µj = 1/(p′ V −1 p)1/2

(12.2.13)

Hence the compensated demand function is given by cj = bj + uj V −1 p/(p′ V −1 p)1/2 .

(12.2.14)

In this example, ψj (p) = χj and ψc (p) = V −1 p/(p′ V −1 p)1/2 .

(12.2.15)

Notice that to obtain a representation of preferences linear in the utility index requires using a particular monotonic transformation of the utility function. In our example, the quadratic form on the right side of (12.2.10) is raised to the one-half power.

12.3. An Economy with Heterogeneous Consumers We now specify a multi-consumer version of our dynamic linear economy designed to satisfy counterparts to Gorman’s conditions for aggregation. There is a collection of consumers, indexed by i = 1, 2, . . . , I . Consumers differ both in their preferences and in their endowments, but not in their information. Consumer i has preferences that are ordered by X ∞ 1 − E β t (sit − bit ) · (sit − bit ) + ℓi2 | J0 t 2 t=0

(12.3.1)

272

Gorman Heterogeneous Households

where {sit } is linked to {hit } and {cit } via sit = Λ hit−1 + Π cit

(12.3.2)

hit = ∆h hit−1 + Θh cit ,

(12.3.3)

and hi−1 is given. In (12.3.1), (12.3.2), (12.3.3), the i superscript pertains to consumer i. The preference disturbance bit is determined by bit = Ubi zt

(12.3.4)

where zt continues to be governed by (3.2). The ith consumer maximizes (12.3.1) subject to (12.3.2), (12.3.3) and the budget constraint E

∞ X t=0

β t p0t · cit | J0 = E

∞ X t=0

i β t (wt0 ℓit + αt0 · dit ) | J0 + v0 · k−1 ,

(12.3.5)

i where k−1 is given. The ith consumer owns an endowment process dit , governed by the stochastic process dit = Udi zt . Each consumer observes the aggregate information Jt at time t , as well as i and hit−1 . The information set Jt continues the idiosyncratic capital stocks kt−1 t to be defined as Jt = [w , x0 ]. This specification confines heterogeneity among consumers to differences in the preference {bit } processes, represented by different selections of Ubi ; differences in the endowment {dit } processes, represented by different selections i of Udi ; differences in hi−1 ; and differences in k−1 . The matrices Λ, Π, ∆h , Θh do not depend on i. This makes everybody’s demand curve have the form of (10.15), with different µw 0 (reflecting different wealth levels) and different bt processes. Prices and the aggregate real variables can be computed by synthesizing a representative consumer and solving a version of the social planning probP lem that was described in chapter 3. Use the settings h−1 = i hi−1 , k−1 = P i P i P i i Ud . This gives us aggregate quantities and i Ub , and Ud = i k−1 Ub , = wa prices. We let µo denote the multiplier on wealth in the budget constraint of the representative (or average) household. To compute individual individual allocations requires more work, to which we now turn.

Allocations

273

12.4. Allocations A direct way to compute individual allocations would be to solve the problem each household faces in the competitive equilibrium at the competitive equilibrium prices. For a fixed Lagrange multiplier on the household’s budget constraint, the household’s problem can be expressed as an optimal linear regulator, with a state vector augmented to reflect the aggregate state variables determining the scaled Arrow-Debreu prices. It is possible to compute the allocation to a particular household by using an iterative scheme to calculate the Lagrange multiplier that assures that the household’s budget constraint is satisfied, but this is not the procedure that we recommend. Instead note that the allocation rule for labor is wa a ℓjt = (µwj (12.4.1) 0 /µ0 )ℓt . If we substitute this expression for ℓjt into versions of (10.13) and (10.14) for the j th consumer, we get the following version of the household’s budget constraint: µjw 0 E0

∞ X t=0

a β t {ρ0t · ρ0t + (wt0 /µaw 0 )ℓt } = E0

∞ X t=0

j β t {ρ0t · (bjt − sjti ) − αt0 · djt } − v0 k−1 .

jw in hand, Solve this equation for µwj 0 , using a doubling algorithm. With µ0 we can use the first-order conditions for services and the canonical service technology to solve for the equilibrium allocation to household j . For a canonical service technology, the first-order conditions for consumption services are: 0 sjt − bjt = µjw 0 ρt .

(12.4.2)

Given ρ0t , which we know from the aggregate allocation and (10.8), we can solve (12.4.2) for sjt , then plug sjt into the ‘inverse canonical representation’ to solve for cjt : cjt = −Π−1 Λhjt−1 + Π−1 sjt (12.4.3) hjt = (∆h − Θh Π−1 Λ)hjt−1 + Π−1 Θh sjt , hj−1 given.

274

Gorman Heterogeneous Households

12.4.1. Consumption sharing rules Our preference specification is an infinite-dimensional generalization of the one described in our digression on Gorman aggregation, a version in which goods are indexed by both dates and states of the world. The counterpart to the matrix V is determined by the probability distribution over states of the world conditioned on J0 and on the parameters of the household technology. The counterpart to χj is determined by the preference shock process {bjt } and the initial endowment of household capital hj−1 . The allocation rule for consumption has the form: cjt − χjt = (uj /ua )(cat − χat ), (12.4.4) where the ratio (uj /ua ) is time invariant and depends only on information available at time zero. We can express (12.4.4) as cjt = (uj /ua )cat + χ ˜jt c˜jt = χ ˜jt , where χ ˜jt ≡ χjt − (uj /ua )χat . Our goal is to show how to compute χ ˜jt and (uj /ua ). We shall show that the utility indexes can be set at the consumers’ marginal utilities of wealth µjw 0 , and that the ‘deviation’ baseline process for consumption {χ ˜jt } can be computed by initializing the inverse canonical repre˜ j and using a ‘deviation’ preference process {˜bj } as the sentation at a vector h t −1 ‘driving’ service process. In terms of ‘deviation’ processes, the allocation rule for consumption services is aw a a sjt − bjt = (µjw (12.4.5) 0 /µ0 )(st − bt ) or s˜jt = ˜bjt , aw a where y˜tj ≡ ytj − (µjw 0 /µ0 )yt . The beauty of this representation is that it does not directly involve prices. The ˜ version of (12.4.3) is

˜ j + Π−1 s˜j c˜jt = −Π−1 Λh t t−1 j j −1 ˜ ˜ ht = (∆h − Θh Π Λ)ht−1 + Π−1 Θh s˜jt ,

(12.4.6)

˜ j given. Associated with s˜j is a synthetic consumption process χ h ˜jt such t −1 that c˜jt = χ ˜tj is the optimal sharing rule. To construct χ ˜jt we simply substitute

Risk Sharing Implications

275

s˜jt = ˜bjt into the inverse canonical representation: j χ ˜jt = −Π−1 Λ˜ ηt−1 + Π−1˜bjt

j η˜tj = (∆h − Θh Π−1 Λ)˜ ηt−1 + Π−1 Θh˜bjt ˜j . η˜j = h −1

(12.4.7)

−1

j ˜ j , it follows from (12.4.6) and (12.4.7) that c˜j = Since s˜jt = ˜bjt and η˜−1 =h t −1 χ ˜jt . Equivalently, allocation rule (12.4.4) holds with {χjt } given by recursion aw (12.4.7), {χat }by its aggregate counterpart, and (uj /ua ) = (µjw 0 /µ0 ). Since the allocation rule for consumption can be expressed as aw a cjt = (µjw ˜jt , 0 /µ0 )ct + χ

(12.4.8)

we can append the recursion in (10.27) for ct and χt from the aggregate, singleconsumer economy to obtain a recursion for generating cjt .

12.5. Risk Sharing Implications Because the coefficient (uj /ua ) is invariant over time and across goods, allocation rule (12.4.4) implies a form of risk pooling in the deviation process {cjt − χjt }. Nonseparabilities (either over time or across goods) in the induced preference ordering for consumption goods appear only in the construction of the baseline process {χjt } and in calculation of the risk-sharing coefficient (uj /ua ) implied by the distribution of wealth. In the special case in which the preference shock processes {bjt } are deterministic (in the sense that they reside in the information set J0 ), individual consumption goods will be perfectly correlated with their aggregate counterparts (conditioned on J0 ). 3

3 Need to add references to literature on risk sharing: Altug-Miller, MaCurdy, Mace, Cochrane, Townsend etc.

276

Gorman Heterogeneous Households

12.6. Implementing the Allocation Rule with Limited Markets We have seen that one way to implement the allocation rule (12.4.4) is to introduce a complete set of markets in state and date contingent consumption. In some environments, a much smaller set of security markets suffices. An example occurs where a single consumption good is produced according to the linear technology: a cat + iat = γkt−1 + dat a kta = δk kt−1 + iat ,

β = 1/(γ + δk ).

(12.6.1)

Each consumer has a common household technology with a heterogeneous preference shock process {bjt } and a heterogeneous initial endowment of household capital hj−1 . The preference shock process is constrained to be in J0 . Each j consumer is endowed with a heterogeneous initial level of capital k−1 and an j endowment process {dt } for consumption. Instead of introducing a full array of contingent claims markets, there is a stock market for J securities that pay dividends {djt }. In addition, oneperiod riskless claims to consumption are traded. To devise a way to implement allocation rule (12.4.4), note that

a cjt − χjt + (uj /ua )χat + (uj /ua )kta = (uj /ua )[(δk + γ)kt−1 + dat ].

(12.6.2)

Let consumer j sell all its shares of stock j and purchase (uj /ua ) shares of all securities traded in the stock market. Once purchased at date zero, let consumer j hold onto this portfolio for all time periods. The total dividends paid in period t will be (uj /ua )dat . Suppose that the consumer purchases fraction (uj /ua ) of the capital stock each period in the one period bond market. The time t payoff a to the t − 1 purchase will be (uj /ua )(δk + γ)kt−1 and the time t purchase will a be (uj /ua )kt . Taken together, these market transactions have a time t receipt a of (uj /ua )[(δk + γ)kt−1 + dat ] and a time t payout of (uj /ua )kta for t = 1, 2, . . . . The difference between payouts and receipts in time t is not equal to cjt , but to cjt − χjt + (uj /ua )χat . This deviation induces trading in the bond market. Note that χjt − (uj /ua )χat is in the time zero information set J0 by assumption.

Implementing the Allocation Rule with Limited Markets

277

Let kˆtj denote additional purchases in the bond market by person j at time t . Construct kˆtj so that j , χjt − (uj /ua )χat + kˆtj = (δk + γ)kˆt−1

t = 1, 2, . . .

(12.6.3)

Solve this equation forward to determine an initial value kˆ0j : kˆ0j =

∞ X t=1

β t [χjt − (uj /ua )χat ].

(12.6.4)

Notice that kˆ0j is in J0 so that it is feasible to construct the sequence {kˆtj }. Modify the previous investment strategy so that the bond market purchases of person j at time t equals (uj /ua )kta + kˆtj for t = 1, 2, . . . . The time t receipts from the previous period purchases in the bond and stock markets equal j a (δk + γ)[(uj /ua )kt−1 + kˆt−1 ]. In light of (12.6.2) and (12.6.3), the difference between time t payouts and receipts is cjt for t = 1, 2, . . . . The coefficient (uj /ua ) in the allocation rules is determined so that initial period consumption cj0 can be purchased from the difference between the time zero receipts and payouts. In this implementation, all consumers hold the same stock portfolio or mutual fund, but make a sequence of person-specific trades in the market for oneperiod bonds. We have allowed for nonseparabilities over time in the induced preference ordering for consumption goods, which have important effects on bond market transactions. This construction displays a multiperiod counterpart to an aggregation result for security markets that was derived by Rubinstein (1974). In a two-period model, Rubinstein provided sufficient conditions on the preferences of consumers and asset market payoffs for the implementation of an Arrow-Debreu contingent claims economy in an environment with incomplete security markets. In Rubinstein’s implementation, all consumers hold the same portfolio of risky assets. In our construction, consumers also hold the same portfolio of risky assets, and portfolio weights do not vary over time. All of the changes over time in portfolio composition take place through transactions in the bond market.

278

Gorman Heterogeneous Households

12.7. A Computer Example The MATLAB program computes the allocation to individual i by executing the computations described above. The program heter.m requires that be run first, and that its output reside in memory. The program heter.m computes individual allocations in the form i X , cit = Sci Xt , hit = Sh t a, and so on. The matrices Sji are returned. The program also computes the matrices Sca , Sh and so on, which determine the aggregate allocations ct , ht , . . . as functions of the augmented state variable Xt : ct = Sca Xt aX , ht = Sh t

and so on. The MATLAB program can then be used to simulate the allocation to individual i and the aggregate allocation. The programs and must both be run for each individual i in a heterogeneous consumer economy. We illustrate the workings of these programs with the following pure exchange economy. There are two households, each with identical preferences −

1 2

E

∞ X t=0

β t (cit − bit )2 + ℓ2t | J0 , i = 1, 2

We specify that bit = 15 for i = 1, 2 . The aggregate preference shock is bt = We specify the following endowment processes. For consumer 1, d1t = 4 + .2 wt1 ,

P

bi = 30 . i t

where wt1 is a Gaussian white noise with variance (.2) 2 . For consumer 2, we specify d2t = 3 + d˜2t d˜2t = 1.2 d˜2t−1 − .22 d˜2t−2 + .25 wt2 where wt2 is a Gaussian white noise with variance (.25) 2 . To capture the pure exchange setup, we specify ∆k = 0, Θk = 0, ∆h = 0, Θh = 0, Λ = 0, Π = 1 . We set β = 1/1.05 . We have used and to simulate a realization of this economy. Figure 12.7.1 reports the individual allocations to consumers 12.2.1 and 12.2.2. Notice how they appear perfectly correlated. Household one is wealthier than the other and so always consumes more (notice that the mean of the first household’s endowment process is 4, while the mean of the second household’s is 3). The perfect correlation between the two consumption services reflects the sharing present in Arrow-Debreu models with time separable preferences. Figure 12.7.2.a graphs d1t − c1t while figure 12.7.2.b graphs d2t − c2t . These figures indicate the “balance of payments” between the two households.

Exercises

279

5

4.5

c1

4

3.5 c2 3

2.5

10

0

20

30

40

50

60

70

80

90

100

Figure 12.7.1: Consumption allocations of consumers one and two in pure endowment economy.

1

0.8

0.8

0.6

0.6

0.4

0.4

0.2

0.2

0

0

-0.2

-0.2

-0.4

-0.4

-0.6

-0.6 -0.8 0

-0.8

10

20

30

40

50

60

70

80

90

100

Fig. 12.7.2.a. Saving of consumer one.

12.8. Exercises

-1 0

10

20

30

40

50

60

70

80

90

100

Fig. 12.7.2.ab5. Saving of consumer two.

280

Gorman Heterogeneous Households

Exercise 1: The first part of this exercise is to be answered using “pencil and paper” as your tools. The second part is to be done on a computer using MATLAB.

12.8.1. Part one An economy consists of identical numbers of two types of infinitely lived consumers. Type i consumers all have preferences that are ordered by

(1)

∞ X − .5 E (cit − bi )2 |J0

, i = 1, 2 ,

0 < β < 1,

t=0

where cit is consumption of a single good by an consumer of type i, E is the mathematical expectation operator, and J0 is information known at time 0. In (1), bi is a parameter that determines the satiation level of consumption for consumers of type i. This is a pure endowment (or pure exchange) economy. The only sources of the single consumption good are two types of trees, the first initially being owned by the first type of consumer, the second type of tree initially being owned by the second type of consumer. Initially, there is one tree owned by each consumer in the economy. The first type of tree yields a constant “dividend” of fruit at the rate (2a)

d1t = d1

∀ t ≥ 0.

The second type of tree yields dividends at time t of (2b)

d2,t+1 = d2 + g wt+1 , d20 = d2

t ≥ 0,

where d2 is a constant and where {wt+1 }∞ t=0 is a martingale difference sequence 2 t with Ewt+1 |Jt = 1, where Jt = w ≡ (wt , wt−1 , . . . , w1 ). Notice that at time 0, d2t = d2 , where d2 is the stationary mean of {d2t }. The feasibility constraint is 2 X i=1

cit =

2 X i=1

dit .

Exercises

281

You are to consider the following decentralized version of this economy. At time 0, there are open a complete array of Arrow-Debreu contingent claims markets. We use the Harrison-Kreps commodity space and pricing system. At time 0, households of type i face the problem of maximizing (1) subject to (3)

E

∞ X t=0

β t p0t cit |J0 =

∞ X t=0

β t p0t dit |J0 .

A. Define a competitive equilibrium for this economy. B. Construct an argument to show that this is a representative consumer economy in the sense that the equilibrium price system can be determined without simultaneously determining individual allocations. C. Describe the preferences and constraints faced by the representative consumer in this economy. D. Compute the equilibrium price system {p0t }∞ t=0 .

E. At the equilibrium prices, compute the right hand side of (3), namely, (4)

a0i = E

∞ X t=0

β t p0t dit |J0 ,

for i = 1, 2. Go as far as you can in getting an analytic, closed form expression for a0i . F. Use your answer to E to establish that the value of trees of type 2 is smaller the larger is the absolute value of g . Interpret this result. G. Give an argument to establish that a0i given by (4) would be the equilibrium price of type i trees if a market in trees (or equivalently, perpetual claims to the dividends from a tree) were opened at time 0. H. Suppose that d1 = d2 < b1 = b2 . Suppose that g = 0, so that the second tree has a perfectly sure yield. Compute the equilibrium consumption allocation. I. Suppose that d1 = d2 = 8, that b1 = b2 = 15, and that g = 1. Compute the equilibrium consumption allocation (for both types of consumer). i. Verify that the equilibrium consumption allocations satisfy a “sharing rule”.

282

Gorman Heterogeneous Households

ii. Which type of consumer consumes more in equilibrium? Why? J. For this economy, compute the price at time t of a perfectly sure claim to one unit of consumption at time t + 1.

12.8.2. Part two Now you are to use the computer. Use MATLAB to do the computations. Hint: We have written some programs that should be a big help in doing this problem. The main programs are and ; does the main calculations while reads in the parameter values for the economy. A program called does the asset pricing calculations. You can edit these files and run them to answer the question. A. For the economy described in Exercise 1 of Part One, compute the equilibrium consumption allocation and simulate it. Write down the consumption allocations and the endowment realizations for t = 0, . . . , 10. B. For the economy described in Exercise 1 of Part One, change the value of g from 1 to 0. Recompute the equilibrium consumption allocation, and simulate it. Write down the consumption allocations and endowment realizations for t = 0, . . . , 10. C. How do the results of A and B conform to your answers in Part One? D. Use the program to compute the value of trees at time 0 under the specification of parameter values given in Exercise 1 of Part One (in particular, set g = 1). Do the results conform with your reasoning in Part One?

Economic integration

283

12.9. Economic integration A. The world consists of two virtually identical but separated economies, j = 1, 2. The economies never trade with one another. Within each economy there exist Arrow-Debreu markets at time 0. Economy j has the following structure:

12.9.1. Preferences: ∞ 1 X t − E0 β [(cjt − bj )2 + ℓ2t ] 2 t=0

12.9.2. Technology cjt + ijt = γkjt−1 + djt ,

γ>0

φijt = gjt , φ > 0 kjt = δk kjt−1 + ijt , 0 < δk < 1 2 gjt = ℓ2jt

kj,−1 given

12.9.3. Information zt+1 = A22 zt + C2 wt+1 djt = Udj zt z0 given. Here cjt is consumption at t , ijt is investment at t , kjt is the capital stock at 2 is the square of labor absorbed in adjusting the capital the end of period t , gjt stock, djt is an endowment shock, and bj is a fixed (across time) preference parameter; all of these objects are scalars. The vector zt is a set of information variables common to the two economies, and {wt+1 } is a martingale difference ′ = I . The parameters γ, φ, β, δk , A22 , and C2 are sequence with Ewt+1 wt+1 common across the two economies. The two economies differ in their values for bj and Udj .

284

Gorman Heterogeneous Households

a. Define a competitive equilibrium for economy j . b. Describe how to compute a competitive equilibrium for economy j using dynamic programming. c. Let p0jt be the (scaled) time zero Arrow-Debreu price for consumption in economy j . Show that there exists a representation for the equilibrium for economy j of the form o kjt A11 Ao12,j kjt−1 0 = + wt+1 zt+1 0 A22 zt C2 (1)

cjt p0jt

=

Scj Mcj

kjt−1 . zt

In particular, argue that Ao11 is the same across the two economies, but that Ao12,j , Scj , and Mcj depend on j . Hint: Two approaches to this problem will bear fruit. First, one can obtain an Euler equation for capital and solve it, as in our treatment of Hall’s model in chapter 3. Second, one can use the method of “adding speed by partitioning the state vector” described in chapter 9. B. Consider a world consisting of two economies with preferences, technology, information, and initial capital stocks identical to the previous one. Now, however, the two economies are integrated, there being world-wide time zero ArrowDebreu markets. Residents of country j own the initial capital stock kj,−1 and the endowment process {dj,t }. a. Define a competitive equilibrium for the integrated economy. b. Argue that the integrated economy is a representative consumer economy, being careful to define what you mean by a representative consumer economy. c. Describe how to compute the equilibrium of the representative consumer economy. d. Let pˆ0t be the (scaled) time zero Arrow-Debreu price for consumption in the integrated economy, and let kˆt = kˆ1t + kˆ2t , and cˆt = cˆ1t +ˆ c2t , where (ˆ·) denotes equilibrium objects for the integrated economy. Show that for the integrated economy, the equilibrium has a representation ˆ o ˆ kt A11 Ao12 kt−1 0 = + wt+1 zt+1 0 A22 zt C2

Economic integration

cˆt pˆ0t

Sc = Mc

ˆ kt−1 , zt

where Ao11 and A22 and the same objects that appear in (1). e. Show that kˆt = k1t + k2t and cˆt = c1t + c2t . f. Find formulas for cˆ1t and cˆ2t . Show that cˆ1t 6= c1t and cˆ2t 6= c2t .

g. Show that p01t + p02t = α pˆ0t for some positive α .

285

Chapter 13 Permanent Income Models This chapter describes a class of permanent income models of consumption, which stress a connection between consumption and income implied by present value budget balance, and which generate interesting predictions about the responses of various components of consumption to identifiable shocks to the information sets of economic agents. The models allow us to characterize how consumption of durables act as a form of savings and how habit persistence alters consumption-savings profiles.

13.1. Technology To focus on dynamics induced by the household technology, it serves our purposes to adopt the following technology specification: φc · ct + it = γkt−1 + et kt = kt−1 + it

(13.1.1)

where φc is a vector of positive real numbers with nc elements, et is a scalar exogenous endowment of consumption and kt−1 is a scalar capital stock. We set δk = 1, thereby ignoring depreciation in capital so that it is net investment. Introducing depreciation in capital would add nothing to our analysis because we shall eliminate any additional input requirement for making new capital productive. With no intermediate inputs required for investment, even if there were depreciation in the capital stock, a version of the first equation of (13.1.1) would apply to net investment by suitably altering the marginal product of capital parameter γ . The empirical counterpart to the scalar endowment process {et } is typically labor income (e.g., see Flavin 1981, and Deaton 1992). Labor is supplied inelastically and produces et units of output independently of the level of capital. The absence of curvature in the technology has some troublesome implications for equilibrium prices that we will discuss later. Nevertheless, this technology provides a good laboratory for studying how the household technology alters consumption-savings profiles. Moreover, this specification has played

– 287 –

288

Permanent Income Models

a prominent role in the empirical literature on the permanent income theory of consumption. To make the model behave well, we impose the restriction that (1+γ)β = 1. Relaxing this restriction to make capital more or less productive has unpleasant implications for the solution to the model. Thus, reducing the marginal product of capital will produce a solution in which the capital stock must eventually become negative. 4 Increasing the marginal product of capital typically produces a solution with asymptotic satiation in a deterministic version of the model; stochastic versions yield a marginal utility vector with mean zero in a stochastic steady state. This razor’s edge linkage between the marginal product of capital and subjective discount factor is the price we pay for eliminating the role of intermediate goods in making new capital productive. To put this technology within the general specification of Chapter 3, we include an additional equation φi it − gt = 0,

(13.1.2)

where φi is a small positive number. Strictly speaking, this introduces a form of adjustment cost by requiring a household input be used to make capital productive. This small penalty makes capital satisfy the square summability constraint. When there are multiple consumption goods, to make the resulting matrix [Φc Φg ] nonsingular, we introduce nc − 1 additional investment goods equal to the last nc − 1 entries of ct Thus, combining these constraints, when nc is one, we form φc 1 0 Φc = Φi = , Φg = , (13.1.3) 0 φi −1 γ Γ= , ∆k = 1, Θk = 1; 0 and when nc is greater than one, φ′c 1 Φc = 0 Φi = φi 0 −I 0

0 0 0 , Φg = −1 , I 0

(13.1.4)

4 A more interesting solution of the model imposes a period-by-period nonnegativity constraint on capital. Instead we limit the terminal behavior of the capital stock by imposing square summability.

Two Implications

γ Γ = 0 , ∆k = 1, Θk = ( 1 0

289

0);

Thus, to embed this model into the general setup of Chapter 3, we have nc investment goods (the original good and nc − 1 additional ones introduced for technical purposes), and one intermediate good (used to enforce square summability in the capital stock).

13.2. Two Implications We extract two sharp implications of this class of permanent income models of consumption. We obtain the first by substituting kt − kt−1 for it in (13.1.1) and solving the resulting difference equation for kt−1 : kt−1 = β

∞ X

β j (φc · ct+j − et+j ).

(13.2.1)

β j E(φc · ct+j − et+j )|Jt .

(13.2.2)

j=0

From this formula, it follows that kt−1 = β

∞ X j=0

Relation (13.2.2) will help us to find recursive solutions to the model. Formula (13.2.1) also has important implications for how consumption and (endowment) income respond to the underlying shocks, as displayed by a set of dynamic multipliers - or impulse responses – {χj } and {ǫj } for {ct } and {et }, respectively, where χj wt gives the response of ct+j to wt and ǫj wt the response of et+j to wt . Since the capital stock kt−1 cannot depend on wt it follows from (13.2.1) that the shock must be present-value neutral. In other words, the impact of wt on current and future values of et must be offset in a present-value sense by its impact on current and future values of ct : ∞ X j=0

β j (φc )′ χj =

∞ X

β j ǫj .

(13.2.3)

j=0

Equality (13.2.3) is precisely the present-value relation studied by Flavin (1981), Hamilton and Flavin (1986), Sargent (1987), Hansen, Roberds and Sargent (1991), and Gali (1991).

290

Permanent Income Models

The second implication pertains to the martingale behavior of the shadow price vector for consumption and capital. To begin, note that the forward evolution equation for the shadow price of capital is Mkt = E(βMkt+1 |Jt ) + E(βγMet+1 |Jt ).

(13.2.4)

The first-order conditions for the first component of investment imply that Met + φi Mgt = Mkt ,

(13.2.5)

where the left side captures the cost of an additional unit of investment and right side the benefit. The second term on the left side reflects the adjustment cost, and is zero in the limiting φi = 0 case. By substituting (13.2.5) into (13.2.4) and using the fact that β(1 + γ) = 1, we obtain the martingale implication for the shadow price of capital: Mkt = E(Mkt+1 |Jt ),

(13.2.6)

and likewise for the multiplier process {Met }. Finally, the shadow price of consumption is given by: Mct = (Φc )′ Mdt = φc Met ,

(13.2.7)

since the last nc − 1 components of Mdt are zero because these are the multipliers on a set of bookkeeping identities. Hence the shadow price process for consumption depends on a single scalar multiplier process {Met }, a martingale that we call the marginal utility process for income. We shall pursue the presentvalue budget balance and martingale implications further, and use them to find and represent the solution of the model.

Solution

291

13.3. Solution To solve the model, we begin by deriving allocation rules for consumption and investment which can be represented in terms of the scalar martingale process {Met }. Then we use present-value relation (13.2.2) to compute Met . Our focus on the marginal utility of income imitates an aspect of the analysis in Bewley (1977). To accomplish the first step, we use the notion of a canonical household technology. Recall that the household technology determines the sequence of consumption services associated with a given sequence of consumption goods and an initial condition for the household capital stock. When the household technology is canonical, we can construct an inverse system which maps a given sequence of consumption services and an initial condition on the household capital stock uniquely into a sequence of consumption goods required to support that service sequence. For a household technology to be canonical, there must be the same number of services as goods, the matrix Π must be nonsingular, and the absolute values of the eigenvalues of the matrix (∆h − Θh Π−1 Λ) must be strictly less than β −1/2 . Under these restrictions, the inverted system can be represented recursively as: ct = Λ∗ ht−1 + Π∗ st

(13.3.1)

ht = ∆∗h ht−1 + Θ∗h st where Λ∗ ≡ −Π−1 Λ, Π∗ ≡ Π−1 ; and ∆∗h ≡ (∆h − Θh Π−1 Λ), Θ∗h ≡ Θh Π−1 . In chapter @[email protected], we showed that there always exists a representation of the induced preferences for consumption goods in terms of a canonical technology. Governing the multipliers is an analogous dual system: Mht = E[β(∆h )′ Mht+1 |Jt ] + E[βΛ′ Mst+1 |Jt ] Mct = (Θh )′ Mht + Π′ Mst ;

(13.3.2)

292

Permanent Income Models

and an associated inverse system: Mht = E[β(∆∗h )′ Mht+1 |Jt ] − E[β(Λ∗ )′ Mct+1 |Jt ]

(13.3.3)

Mst = −(Θ∗h )′ Mht + (Π∗ )′ Mct . Since {Mct } is a martingale sequence, it follows from the inverse dual system that {Mht } and {Mst } are both martingales. In fact they are linear combinations of the scalar martingale sequence Met . For instance, Mst = Ms Met

(13.3.4)

Ms ≡ {(Π∗ )′ + (Θ∗h )′ [I − β(∆∗h )′ ]−1 βΛ∗′ }φc

(13.3.5)

where

Consequently, we can solve for the service sequence in terms of the scalar martingale {Met } from the simple link between the vector of services and the corresponding marginal utility vector: st = −Ms Met + bt .

(13.3.6)

From this relation and from the inverse household technology (13.3.1) it follows that ct = Λ∗ ht−1 − Π∗ Ms Met + Π∗ bt (13.3.7) ht = ∆∗h ht−1 − Θ∗h Ms Met + Θ∗h bt . To characterize the decision rule for investment, we solve (13.1.1) for it and substitute the right-hand side of (13.3.7) for ct : it = γkt−1 + et − φc · ct

(13.3.8)

= γkt−1 + et − (φc )′ Λ∗ ht−1 − (φc )′ Π∗ (bt − Ms Met ). So far, we have derived a recursive representation for consumption, investment and household capital in terms of the scalar multiplier process {Met }. However, we have not shown how to initialize this sequence, and we do not yet have a formula relating the time t increment of this process to the underlying martingale difference sequence {wt }. We now derive formulas for both of these objects.

Solution

293

To find an expression for the marginal utility of income process, we exploit the present-value budget balance restriction (13.2.2). In light of the inverse system (13.3.1) for the household technology, we are led to compute the expected discounted sum of services and household capital: ∞ X j=0

β j E(st+j |Jt ) = −[1/(1 − β)]Ms Met +

∞ X j=0

β j E(bt+j |Jt ),

(13.3.9)

and β

∞ X j=0

β j E(ht+j |Jt ) = β

∞ X j=0

β j E(∆∗h ht+j−1 + Θ∗h st+j |Jt ).

(13.3.10)

Rewriting (13.3.10), we see that (I − β∆∗h ) or

∞ X j=0

∞ X j=0

β j E(ht+j−1 |Jt ) = ht−1 + βΘ∗h

∞ X j=0

β j E(st+j |Jt ),

β j E(ht+j−1 |Jt ) =

(I −

β∆∗h )−1 ht−1

+ β(I −

β∆∗h )−1 Θ∗h

∞ X j=0

(13.3.11) j

β E(st+j |Jt ).

Since φc · ct = (φc )′ (Λ∗ ht−1 + Π∗ st ), it follows from equation (13.3.1) and (13.3.9) that (1 + γ)kt−1 − (φc )′ Λ∗ (I − β∆∗h )−1 ht−1 = (13.3.12) − Ms′ Ms Met /(1 − β) +

∞ X j=0

β j E(Ms′ bt+j − et+j |Jt ).

Solving for {Met } results in Met = 1/(Ms′ Ms )[(1 − β)

∞ X j=0

β j E(Ms′ bt+j − et+j |Jt ) −

(13.3.13)

γkt−1 + (1 − β)(φc )′ Λ∗ (I − β∆∗h )−1 ht−1 ]. To interpret this solution, it is useful to decompose the right-hand side of (13.3.13) into three components. First, we follow the permanent income

294

Permanent Income Models

literature by defining permanent income to be that amount of income to be spent on consumption that can be expected to persist in the future and still satisfy (13.2.2): ytp ≡ γkt−1 + (1 − β)

∞ X j=0

E(β j et+j |Jt ).

(13.3.14)

Formally, this is obtained by letting {ytp } be a hypothetical expenditure process p for consumption, assuming it is a martingale, substituting yt+j for φc · ct+j in p equation (13.2.2), and solving for yt . Note that this measure of permanent income does not adjust for risk in the endowment sequence and hence even when divided by (1 − β) is distinct from equilibrium wealth. Nevertheless, it is an important component of the solution to the model. In fact, ct = ytp is the solution for consumption in the in the special case of single good, no preference shocks ( bt constant), and time separable preferences (Λ zero), which is Hall’s (1978) permanent income model of consumption. In this simple case, the marginal utility process for endowment income is formed by translating the negative of permanent income: Met = b−ytp . More generally, when the preference shock process is not expected to be constant, the term of interest is a ‘permanent’ measure of the preference shock sequence: ∞ X bpt ≡ (1 − β) E[β j (Ms )′ bt+j |Jt ]. (13.3.15) j=0

Finally, when preferences are not separable over time, the household capital stock also enters the solution for the marginal utility of endowment income. To interpret its coefficient, consider the sequence of consumption goods required to support zero consumption services from now into the future. To compute this sequence, simply feed a sequence of zeros into the inverse of the household technology. Discounting the resulting consumption sequence and premultiplying by (φc )′ results in: yth ≡ (1 − β)(φc )′ Λ∗ (I − β∆∗h )−1 ht−1 .

(13.3.16)

Hence yth adjusts the permanent income measure to account for implicit consumption associated with a “baseline” zero service sequence. To summarize, the marginal utility of endowment income can be represented as: Met = (1 − β)(1/Ms′ Ms )(bpt − ytp + yth ). (13.3.17)

Deterministic Steady States

295

Relation (13.3.17) gives a decomposition of marginal utility of income into three components: bpt , ytp and yth . Increases in bpt result in an outward shift in preference for consumption, which increases the marginal utility of income; increases in ept correspond to an outward shift of permanent income, which reduces the marginal utility of income; and alterations in yth reflect movements in the initial household capital stock. The measurement of this third component is the discounted endowment-equivalent consumption sequence associated with a baseline (zero) sequence of services. Increasing it has an opposite impact on the marginal of income from increasing permanent income. The final task of this section is to deduce a formula for the increment of Met of the form µwt for some µ. Note that yth depends on time t − 1 information. Hence only the bpt and ytp terms enter into consideration. Let {ψj } denote the sequence of matrices of dynamic multipliers for the preference shock process {bt }. It follows from (13.3.13) that µ = [(1 − β)/(Ms′ Ms )]

∞ X j=0

β j [(Ms )′ ψj − ǫj ].

(13.3.18)

The dynamic multipliers {χj } for consumption can then be computed recursively from (13.3.7), and by construction satisfy the present-value budget balance restriction (13.2.3).

13.4. Deterministic Steady States It is useful to study consumption in a deterministic steady state, partly to verify that there exist configurations of the model for which consumption of all goods is positive in this steady-state. Otherwise, the stochastic versions of the model would likely have some perverse implications. We consider cases in which {bt } and {et } are constants set at the values b and e, respectively. For a deterministic version of the model, the marginal utility of income is constant over time. Of course, we are only interested in initial conditions such that the initial marginal utility is positive and hence the entire sequence is positive. Thus from equation (13.3.17), we are lead to require that: Me0 = (1/Ms′ Ms )(Ms′ b − e − γk−1 + y0h ) > 0.

(13.4.1)

296

Permanent Income Models

Since 1/(Ms′ Ms ) is positive by construction, we only need to be concerned with requiring that (Ms′ b − e − γk−1 + y0h ) be positive. Any changes in (b, e, h−1 , k−1 ) that alter the right side of (13.4.1) will clearly change the marginal utility of income (in all time periods). Since the preference shifter sequence is fixed at a constant level, the constant marginal utility of income sequence implies a constant service sequence given by: s = b − Ms Me0 . (13.4.2) The corresponding sequences of consumption goods and household capital need not be constant, and we now investigate the limiting behavior of these objects. Armed with the consumption service sequence, the consumption and household capital sequences can be computed from the inverse household technology. The absolute values of the eigenvalues of ∆∗h are less than β −1/2 but can be greater than or equal to one. Without further restricting the eigenvalues of ∆∗h to have absolute values that are strictly less than one, the consumption sequence may not converge to a steady state. 5 With the additional restriction that the absolute values of eigenvalues are strictly less than one, the consumption and household capital sequences will converge with limits: h∞ = (I − ∆∗h )−1 Θ∗h s c∞ = Λ∗ h∞ + Π∗ s

(13.4.3)

where variables with subscript ∞ denote limit points. By combining (13.4.2) and (13.4.3), it can be checked whether the limiting consumption vector is strictly positive. As an illustration, consider a setting with a single consumption good, a single physical capital stock, and the following household technology: ht = δh ht−1 + (1 − δh )ct st = λht−1 + ct ,

(13.4.4)

where δ is a depreciation factor between zero and one. Notice that the household capital stock is constructed to be a weighted average of current and past consumption. The inverse system is given by: ht = [δh − λ(1 − δh )]ht−1 + (1 − δh )st ct = −λht−1 + st .

(13.4.5)

5 As emphasized by Becker and Murphy (1988), configurations with explosive eigenvalues can arise in models of ‘rational addiction’.

Cointegration

297

In this simple case, the eigenvalue of ∆∗h is simply the coefficient on the lagged capital stock in the evolution equation for household capital. This coefficient has an absolute value less than one if: −1 < λ < (1 + δ)/(1 − δ).

(13.4.6)

When these inequalities are satisfied consumption and the household capital stock both converge to s/(1 + λ). Negative values of λ that violate the inequalities in (13.4.6) display a form of ‘rational addiction’ as analyzed by Becker and Murphy (1988). For instance, when λ is −1, the coefficient on lagged capital is unity, and the consumption sequence required to support a constant service sequence must grow linearly over time. Lower values of λ (i.e., negative ones with larger absolute values) imply geometric growth in consumption. Simply requiring the limiting value of consumption to be positive guarantees that consumption will be positive for initial levels of household capital close to the steady state, but it would be nice to obtain a stronger result. One strategy would require entries of the matrices of the inverse household technology all to be positive. Unfortunately, this restriction that would eliminate some important examples in which there is substitutability across goods or over time. For instance, in (13.4.4), when λ is positive as in the case of a durable good, the inverse household technology has a negative coefficient. Nevertheless, starting from an initial level of household capital below the steady state will result in a positive consumption sequence.

13.5. Cointegration A key feature of our solution is that the marginal utility of income process is a martingale, which implies via (13.3.6) that the consumption service process is nonstationary. If in addition the preference shock process {bt } is asymptotically stationary, then the service process and consumption are each cointegrated. Suppose that the preference shock process {bt } is asymptotically stationary, but unobservable to the econometrician. This implies that there are ns − 1 linear combinations of consumption services that are asymptotically stationary. To show this, take any vector ψ that is orthogonal to Ms . It follows from

298

Permanent Income Models

(13.3.6) that ψ ′ st = ψ ′ bt .

(13.5.1)

Evidently there are ns − 1 linearly independent ψ ’s. Each such ψ is referred to as a cointegrating vector, in the terminology of Granger and Engle (19??). An extensive literature treats the efficient estimation of cointegrating vectors. However, what interest us are not the cointegrating vectors but rather the vector Ms that is orthogonal to all of the cointegrating vectors for consumption services. The cointegrating vectors for consumption services differ from the cointegrating vectors for consumption goods. To deduce the cointegrating vectors for the consumption flows, we shall build upon the deterministic steady-state calculations reported in (13.4.3). From (13.4.3), we know that for a deterministic steady state c∞ = [Λ∗ (I − ∆∗ )−1 Θ∗ + Π∗ ]s∞ . (13.5.2) The matrix on the right-hand side of (13.5.2) also gives the transformation mapping a date t shock to consumption services to the limiting response of consumption. Any vector ψ satisfying ψ ′ [Λ∗ (I − ∆∗ )−1 Θ∗ + Π∗ ]Ms = 0

(13.5.3)

is a cointegrating vector for consumption. Let Ψ denote an nc − 1 by nc matrix with rows that are linearly independent cointegrating vectors. Notice that the implied model for Ψct and c1,t − c1,t−1 contains only stationary endogenous variables, so that it can be estimated using methods that require asymptotic stationarity, like the frequency-domain methods of chapter 9. An estimation strategy based on recursive formulations of the Gaussian conditional likelihood function would not require such a transformation to a stationary set of endogenous variables, but would require confronting the nonexistence of an asymptotically stationary distribution of the state vector from which to draw an initial estimator of the state. In chapter 9, we described a method based on ideas of Kohn and Ansley (1985) designed to construct an initial estimator of the state in such a circumstance. If we were to assume that the preference shock process is itself nonstationary and that there does not exist any nontrivial cointegrating vector for this process, then it would follow that there exist no cointegrating vectors for either the service process or the consumption process. In this case, to utilize estimation methods

Constant Marginal Utility of Income

299

requiring stationary, we would base parameter estimation on the implications for the differenced processes for consumption and household capital.

13.6. Constant Marginal Utility of Income In the absence of uncertainty, the marginal utility of income process will be constant. In this section we introduce uncertainty in the endowment and preference shock processes, and ask: when will the marginal utility of income process remain constant through time? Constancy of the marginal utility of income is an extreme version of the prediction of permanent income theory that the ability to transfer consumption over time results in “smoothness of consumption” over time. In the absence of preference shocks, a constant marginal utility of income process implies that consumption services will also be constant through time. We address this question from two angles. Initially, we investigate the limiting behavior as the subjective discount factor β approaches unity, and provide conditions on the stochastic structure sufficient to imply constant marginal utilities of income in the limit. In taking this limit, we will not concern ourselves with interpreting directly the limit economy. Instead we will study the limiting behavior of the solutions to the optimal resource allocation problems along with the associated marginal utility of income processes. The second attack on the question is to characterize the specifications of uncertainty that imply a constant marginal utility of income process for a given β that is strictly less than one. The initial portion of our investigation will imitate and replicate features of Bewley’s (1977) study of the permanent income model of consumption. Our analysis is mechanically simpler than Bewley’s, but different. When we drive β to one, we maintain the link between the subjective discount factor and the marginal product of capital. Hence as β tends to one in our analysis, the marginal product of capital as measured by γ is simultaneously being driven to zero. In contrast, Bewley considered setups in which the counterpart to the marginal product of capital is always zero. Since capital is less productive, nonnegativity constraints are a central feature of his analysis of the economies with β strictly less than one. Suppose that {zt } is a stationary stochastic process and that zt has a finite PN −1 second moment. Then it is known that the time series average (1/N ) j=0 zt+j

300

Permanent Income Models

converges. 6 Moreover, the limit vector is invariant to the starting date t of the average. Consistent with our setup in previous chapters, we assume that both bt and et are linear functions of zt . Recall that the portion of the solution for the marginal utility of income that is not predetermined (the portion that can respond to a current-period shock) is a linear combination, say ν , of a P∞ conditional expectation of the geometric average (1 − β) j=0 β j zt+j where ν ≡ [1/(Ms′ Ms )](Ms Ub − Ue )

(13.6.1)

bt = Ub zt and et = Ue zt [see (13.3.13)]. While the simple time-series average and the geometric average will not typically agree, they can be made arbitrarily close by driving N to infinity and β to one. Under both limits, tail terms in the average become relatively more important as the limit point is approached. Formally, it follows from the theory of Cesaro and Abel summability that lim (1/N )

N →∞

N −1 X j=0

zt+j = lim (1 − β) β→1

∞ X

β j zt+j

(13.6.2)

j=0

(e.g. see Zygmond 1979, Theorem 1.33, page 80). Therefore, as the discount factor tends to one, the right side of (13.6.2) converges to a vector independent of t . Moreover, for the information structures we impose, the limit vector must be in the initial period information set. 7 The constancy of the marginal utility of income as β goes to unity follows immediately. The argument just provided relies on stationarity but does not require linearity in the evolution equation for {zt }. In fact, stationarity often can be replaced by a weaker notion of “asymptotic stationarity” as we now illustrate using the linear specification zt+1 = A22 zt + C2 wt+1

(13.6.3)

imposed elsewhere in this book. This specification can be exploited to obtain an alternative demonstration of the constancy of the marginal utility of income. 6 The convergence is both with probability one and in mean square, where mean-square convergence is defined using the square root of the second moment as a norm. We use meansquare convergence in our subsequent analysis. 7 This follows because the limiting random variable has a finite second moment. As long as the forecast error variance is independent of calendar time and information accumulates over time, the forecast error variance must be zero.

Constant Marginal Utility of Income

301

Recall that (1 − β)

∞ X j=0

β j E(zt+j |Jt ) = (1 − β)(I − βA22 )−1 zt .

(13.6.4)

To investigate the limit as β tends to one, it is convenient to uncouple the dynamics according to the eigenvalues. Let A22 = T DT −1

(13.6.5)

be the Jordan decomposition, and suppose that D can be partitioned as: I 0 D= (13.6.6) 0 D2 where the absolute values of the diagonal entries of D2 are all strictly less than one. Using the Jordan decomposition, it follows that I 0 −1 (1 − β)(I − βA22 ) = T T −1 . (13.6.7) 0 (1 − β)(I − βD2 )−1 Taking limits, we see that −1

lim (1 − β)(I − βA22 )

β→1

Therefore, (1 − β)

P∞

j=0

=T

I 0

0 0

T −1 .

(13.6.8)

β j E(zt+j |Jt ) depends only on zt∗ ≡ ( I

0 ) T −1 zt ,

(13.6.9)

where zt∗ has law of motion ∗ zt+1 = zt∗ + C ∗ wt+1 ,

(13.6.10)

C∗ ≡ ( I

(13.6.11)

where 0 ) T −1 C2 .

Sufficient conditions for the marginal utility of income to be constant are that the Jordan decomposition of A22 satisfies (13.6.6) and C ∗ be zero. When these restrictions are satisfied, the process {zt } will be asymptotically stationary because the process {zt∗ } will be constant over time and because {( 0 I ) T −1 zt } will converge to a stationary process. This latter result follows since the diagonal

302

Permanent Income Models

entries of D2 have absolute values that are strictly less than unity. Stationarity is only implied when the {zt } is initialized appropriately. The arguments just given cannot be extended to {zt } processes with more fundamental forms of nonstationarity. For instance, time trends or unit roots in the endowment process would suffice to overturn constancy of the marginal utility of income in the limit. In the case of time trends, the averages may diverge as the limits are taken. For unit root processes (without drifts) the limits are well defined, but the uncertainty in the marginal utility of income process does not vanish. We now change experiments and hold fixed the subjective discount factor and ask if it is still possible for the marginal utility of income to be constant. The answer to this question turns out to be yes. Assume the Jordan decomposition (13.6.5) and (13.6.6) along with restriction (13.6.11), except that D2 can now have eigenvalues with absolute values that are equal or even greater than one (but less than β −1/2 ). If ν(I − βA22 )−1 = ( ν ∗

0)

(13.6.12)

for some vector ν ∗ , then the marginal utility of income will be constant over time. While this clearly imposes a restriction on the matrix A22 , it is satisfied by some stationary and nonstationary specifications of the endowment and preference shock processes.

13.7. Consumption Externalities One of the prime motivations for intertemporal complementarities put forth by Ryder and Heal (1973) is that individual consumers are concerned in part about their consumption relative to the past community average. In other words, there is an externality in consumption. This motivation is in contrast to that given by Becker and Murphy (1988) in which the complementarities are purely private. The solution described previously is applicable even if this consumption externality is present as a solution to an optimal resource allocation problem. However, the link between optimal resource allocation and competitive equilibrium may vanish when there is a consumption externality. We now investigate this issue.

Consumption Externalities

303

To capture the externality, we endow the consumer with the household technology: Ht = ∆h Ht−1 + Θh Ct (13.7.1) st = ΛHt−1 + Πct where Ht , Ht−1 and Ct are interpreted as community-wide vectors that are beyond the control of the private consumers but are equal to their lower case counterparts in equilibrium. The previous argument justifying the martingale properties of the marginal utilities of income and consumption relied only on the technology specification and still applies when the externality is present. In light of the externality interpretation, the marginal utility of services now satisfies: Mst = (Π∗ )′ Mct

= (Π∗ )′ φc Met .

(13.7.2)

Recall that Π∗ is equal to Π−1 . Although the link between the marginal utility of services and marginal utility of consumption goods is altered, the marginal utility of service process remains a martingale. The previous solution method can now be imitated by substituting the matrix (Π∗ )′ φc for Ms given by (13.3.4). When there is a single consumption good and the household technology is canonical, the two solutions coincide. This alteration can be verified by taking the previous solution for the marginal utility of income and showing that all of the equilibrium conditions and first-order conditions remain satisfied. While the marginal utility of services is altered by a constant scale factor over time, this clearly has no impact on the implied marginal rates of substitution for consumption services and therefore the original quantity allocation remains intact with the externality interpretation. When there are multiple consumption goods, the quantity allocations can be altered. Also, when the original household technology is not ‘canonical,’ the quantity allocations can be altered even when there is a single consumption good. While there generally exists a canonical household technology that implies the same induced preferences for consumption goods, the externality version of the specification can give rise to a fundamentally different canonical technology, breaking the simple link between solution to the resource allocation problem and the decentralized economy. To elaborate on this last point, suppose the original household technology is not canonical. In Chapter ? we showed how to find the corresponding canonical

304

Permanent Income Models

technology to be used in solving the optimal resource allocation problem. In the presence of consumption externalities, we can find the analog to a canonical household technology by first noting that bt − st = Bt − Πct ,

(13.7.3)

where Bt ≡ bt − ΛHt−1 . Consequently, (bt − st ) · (bt − st ) = Bt · Bt − 2(Bt )′ Πct + (ct )′ Π′ Πct .

(13.7.4)

Suppose that Π′ Π is nonsingular, and obtain a factorization: ˆ ′Π ˆ = Π′ Π Π

(13.7.5)

ˆ is nonsingular. Also, define where Π ˆ ≡Π ˆ ′−1 Π′ Λ Λ ˆbt ≡ Π ˆ ′−1 Π′ bt

(13.7.6)

ˆ t−1 + Πc ˆ t. sˆt ≡ ΛH Then (bt − st ) · (bt − st ) and (ˆbt − sˆt ) · (ˆbt − sˆt ) agree except for a term that is not controllable by the individual consumer. Consequently, technology (13.7.6) and the implied preferences for the original household technology are the same. For this solution method to apply, we need this transformed version of ˆ is nonsingular the household technology to be canonical. Since the matrix Π ˆ to have eigenvalues ˆ −1 Λ by construction, it suffices for the matrix ∆h − Θh Π −1/2 with absolute values that are strictly less than β . If this restriction is not satisfied then there fails to exist a competitive equilibrium.

Tax Smoothing Models

305

13.8. Tax Smoothing Models By reinterpreting variables, our model can represent a class of linear models of optimal taxation, versions of which were studied by Barro (1979) and Judd (1990). Let τt be a vector of taxes collected from various sources (e.g., capital, labor, imports, etc.); Gt a scalar stochastic process of government expenditures; Bt−1 the stock of risk-free one-period government debt bearing net one-period interest rate of γ = β1 − 1; and def t the gross-of-interest 0 deficit. Match up variables as follows: ct ∼ τt , et ∼ Gt , kt−1 ∼ Bt−1 , it ∼ def t . Set φc to a vector of ones, so that φ′c τt measures total time t government tax revenues. With these associations, (13.1.1) become def t = γBt−1 + Gt − φ′c τt Bt = Bt−1 + def t .

The preference ordering is interpreted as minus the loss function associated with taxation, and measures the dynamics of tax distortions. Consider three special cases of this model, each of which sets bt to a vector of zeroes. 1. Random walk taxes. To capture Robert Barro’s specification, set φc = 1 (so there is only one kind of tax revenues), Λ = 0, Π = 1, Θh = ∆h = 0. This version makes taxes follow a random walk. It is a relabelling of Hall’s model of consumption. 2. White noise taxes. To capture a one-tax version of Judd’s specification, again set φc = 1, but now set Π = ∆h = 1, Θh = Λ = −1. With these P∞ Pt settings, the government’s objective function is −.5E0 t=0 β t ( j=0 τt−j )2 . This specification is intended to capture the long-lived adverse effects of taxation on capital. The optimal policy makes taxes a white noise process, a feature that characterizes the asymptotic behavior of capital taxation in the model of Chari, Christiano, and Kehoe (1994). To deduce the white noise property for this model, use (13.3.7) and the relations defining Λ∗ , Π∗ , ∆∗h , Θ∗h under (13.3.1). In particular, we obtain τt = −Ms (Met − Met−1 ). 3. Two taxes. Set φc = [ 1 1 ], and specify two taxes whose ‘distortion technology’ is obtained by stacking the two technologies described in examples 1 and 2. This is the kind of setup advocated by Judd (1990), and makes one tax a random walk, the other a white noise.

Chapter 14 Non-Gorman Heterogeneity Among Households

14.1. Introduction The previous chapter studied a setting in which households have heterogeneous endowments and preference shock, but otherwise have identical preferences and household technologies, implying that all Engle curves are linear with the same slopes. The property of identically sloped linear Engle curves delivers a tidy and tractable theory of aggregation assuring the existence of a representative household. This theory applies when different households share the same household technology (Λ, Π, ∆h , Θh ). In this chapter, we maintain the linearity of households’ Engle curves, but permit their slopes to vary across classes of households. In particular, we now allow households’ technology parameters (Λi , Πi , ∆hi , Θhi ) to differ across classes of households indexed by i. This alteration causes the existence of a representative household, in the sense of there being a preference ordering over stochastic processes for aggregate consumption that is independent of the initial wealth distribution, to vanish. Nevertheless, the structure still fits within a class that readily yields to linear quadratic dynamic programming algorithms. Competitive equilibria can be calculated using an algorithm based on Negishi’s idea of finding a fixed point within a class of Pareto problems, where the fixed point is a list of Pareto weights that deliver budget balance at candidate equilibrium prices. In this chapter, we describe how the algorithms can be applied very efficiently within our class of economies. We also describe how a more limited form of aggregation than Gorman’s can be carried out for this economy. In particular, implementation of the Negishi algorithm enables us to uncover a ‘mongrel’ preference ordering over aggregate consumption streams, where the preference ordering depends on the initial distribution of wealth, as do the parameters of any ‘household technology’ for representing those preferences.

– 307 –

308

Non-Gorman Heterogeneity Among Households

14.2. Households’ Preferences There are equal numbers of two types of households, indexed by i = 1, 2. Households of type i have preferences ordered by ∞ 1 X t − E β [(sit − bit ) · (sit − bit ) + ℓ2it ] | J0 . 2 t=0

(14.2.1)

Here sit is a consumption service vector for consumer i, bit is a preference shock process, and ℓit is labor supplied by consumer i. Services sit are produced via the technology sit = Λi hit−1 + Πi cit

(14.2.2)

hit = ∆hi hit−1 + Θhi cit , i = 1, 2

(14.2.3)

Here hit is consumer i’s stock of household durables at the end of period t , and cit is consumer i’s rate of consumption. The preference shock process bit is governed by bit = Ubi zt (14.2.4) where zt continues to be governed by zt+1 = A22 zt + C2 wt+1 Notice that this specification permits each class of households to have its own list of matrices (Λi , Πi , ∆hi , Θhi ) that determine a technology for converting consumption goods into services.

A Pareto Problem

309

14.2.1. Technology Consumption goods (c1t , c2t ) are produced via the technology Φc (c1t + c2t ) + Φg gt + Φi it = Γkt−1 + d1t + d2t

(14.2.5)

kt = ∆k kt−1 + Θk it

(14.2.6)

gt · gt = ℓ2t ,

ℓt = ℓ1t + ℓ2t .

(14.2.7)

As before, gt denotes the quantity of labor-using intermediate production activities; dit is the amount of the endowment vector of household i used in the production process. We assume that dit = Udi zt

, i = 1, 2.

(14.2.8)

14.3. A Pareto Problem The social welfare function is a weighted average of the utilities of the two households, with weight on household 1’s utility being λ, 0 < λ < 1. For fixed λ , we want to find an allocation that maximizes ∞ X 1 − λE0 β t [(s1t − b1t ) · (s1t − b1t ) + ℓ21t ] 2 t=0

∞ X 1 − (1 − λ)E0 β t [(s2t − b2t ) · (s2t − b2t ) + ℓ22t ] 2 t=0

subject to the constraints that describe the household and production technologies. By way of fitting it into an optimal linear regulator, it is convenient to note a property of the solution to the problem that permits us to avoid carrying along ℓ1t and ℓ2t as variables, and to replace them by functions of ℓt . The solution of the social planning problem implies a pair of simple ‘sharing rules’ for labor. We deduce these sharing rules before solving the full problem in order to economize the number of control variables.

310

Non-Gorman Heterogeneity Among Households

Let Mtℓ be the stochastic Lagrange multiplier associated with the constraint ℓ1t + ℓ2t = ℓt . With respect to ℓ1t and ℓ2t , the first order conditions are Mtℓ = λℓ1t and Mtℓ = (1 − λ)ℓ2t . These conditions imply that ℓt = ℓ1t + ℓ2t = Mtℓ /(λ(1 − λ)), or Mtℓ = λ(1 − λ)ℓt . Substituting this last equality for Mtℓ into the marginal conditions for ℓ1t and ℓ2t gives the ‘sharing rules’ ℓ1t = (1 − λ)ℓt , ℓ2t = λℓt . Use these two equations to represent the terms in ℓ1t and ℓ2t in the social planning criterion as λℓ21t + (1 − λ)ℓ22t = λ(1 − λ)ℓ2t . Substituting in the constraint gt · gt = ℓ2t , we can represent the social planning criterion as ∞ X 1 − E0 β t [λ(s1t − b1t ) · (s1t − b1t ) + (1 − λ)(s2t − b2t ) · (s2t − b2t ) 2 (14.3.1) t=0

+ λ(1 − λ)gt · gt ].

The objective function (14.3.1) is to be maximized subject to the following constraints: (12.2)

sit = Λi hit−1 + Πi cit

, i = 1, 2

(12.3)

hit = ∆hi hit−1 + Θhi cit

, i = 1, 2

(12.5)

Φc (c1t + c2t ) + Φg gt + Φi it = Γkt−1 + d1t + d2t

(12.6)

kt = ∆k kt−1 + Θk it

(12.4)

dit = Udi zt ,

bit = Ubi zt ,

zt+1 = A22 zt + C2 wt+1

i = 1, 2

A Pareto Problem

311

This problem can be set up as an optimal linear regulator problem by following steps paralleling those for the single-household economy described in Chapter 4. Define the state and controls of the system as h1t−1 h2t−1 xt = kt−1 zt

,

ut =

it c1t

.

Notice that from (14.2.5), (c2t , gt ) can be expressed as functions of the state and controls at t : n o c2t = [Φc Φg ]−1 Γ kt−1 + (Ud1 + Ud2 ) zt − Φc c1t − Φi it . (14.3.2) gt Substitution from the above equation into (14.2.3) for i = 2 shows that the law of motion for xt+1 can be represented

h1t h2t = k t zt+1

∆h1 0 0 0

0 ∆h2 0 0

0 Θh2 Uc [Φc Φg ]−1 Γ ∆k 0

0 −Θh2 Uc [Φc Φg ]−1 Φi + Θk 0

or

0 0 + 0 C2

Θh1 −Θh2 0 0

0 Θh2 Uc [Φc Φg ]−1 (Ud1 + Ud2 ) 0 A22

it c1t

h1t−1 h2t−1 k t−1 zt

wt+1

xt+1 = Axt + But + C wt+1 .

(14.3.3)

Here Uc is a matrix that selects the first nc rows of the right hand side of (14.3.2), the rows corresponding to c2t , and Ug is a matrix that selects the rows of (14.3.2) corresponding to g1t . Here and below, we use the equalities

312

Non-Gorman Heterogeneity Among Households

Uc [Φc Φg ]−1 Φc = I, Ug [Φc Φg ]−1 Φg = I, Uc [Φc Φg ]−1 Φg = 0, Ug [Φc Φg ]−1 Φc = 0. Now substitute from (14.3.2) into (14.2.2) for i = 2 to get n o s2t = Λ2 h2t−1 + Π2 Uc [Φc Φg ]−1 Γkt−1 + (Ud1 + Ud2 )zt − Φc c1t − Φi it . Use this equation and (14.2.2) for i = 1 to deduce ′ Λ1 0 0 (s1t − b1t ) = −Ub1 0 Π1

h1t−1 h2t−1 k t−1 zt it

c1t

′ 0 Λ2 −1 Π2 Uc [Φc Φg ] Γ (s2t − b2t ) = Π2 Uc [Φc Φg ]−1 (Ud1 + Ud2 ) − Ub2 −Π2 Uc [Φc Φg ]−1 Φi

−Π2

or

(s1t − b1t ) = H1 (s2t − b2t ) = H2

c1t

(14.3.4)

(14.3.5)

xt ut

xt ut

h1t−1 h2t−1 k t−1 zt it

.

Similarly, we have ′ 0 0 Γ −1 gt = Ug [Φc Φg ] (Ud1 + Ud2 ) −Φi 0

or

gt = G

xt ut

.

h1t−1 h2t−1 k t−1 zt it

c1t

(14.3.6)

A Pareto Problem

313

Now notice that the current return function in (14.3.1) can be represented as

λ(s1t − b1t ) · (s1t − b1t )

+ (1 − λ)(s2t − b2t ) · (s2t − b2t ) + λ(1 − λ)gt · gt ′ xt xt = S ut ut

(14.3.7)

where S = λH1′ H1 + (1 − λ)H2′ H2 + λ(1 − λ)G′ G.

(14.3.8)

The analysis on pages 79–80 now applies to the present system. In particular, let x′t Sxt = x′t Rxt + u′t Qut + 2x′t W ut , and write the law of motion in the form (14.3.3). This makes the Pareto problem with weight λ into a discounted optimal linear regulator problem. The solution of the Pareto problem is a law of motion xt+1 = A0 (λ)xt + CWt+1

(14.3.9)

and a list of matrices Sj (λ) such that optimal allocations are given by cit = Sci (λ)xt

, i = 1, 2

it = Si (λ)xt hit = Shi (λ)xt

, i = 1, 2

sit = Ssi (λ)xt

, i = 1, 2

(14.3.10)

The value function for the Pareto problem has the form V (xt ) = x′t V1 (λ)xt + V2 (λ).

(14.3.11)

Associated with the solution of the Pareto problem for a given λ are a set of Lagrange multiplier processes given by

314

Non-Gorman Heterogeneity Among Households

Mth1 (λ) = 2β[I 0 0 0] V1 (λ)A0 (λ)xt Mth2 (λ) = 2β[0 I 0 0] V1 (λ)A0 (λ)xt Mtk (λ) = 2β[0 0 I 0]V1 (λ)A0 (λ)xt Mts1 (λ) = λ(Sb1 − Ss1 (λ))xt

≡ Mk (λ)xt

Mts2 (λ) = (1 − λ)(Sb2 − Ss2 (λ))xt

Mtc1 (λ) = Θ′h1 Mth1 (λ) + Π′1 Mts1 (λ)

(14.3.12)

Mtc2 (λ) = Θ′h2 M h2 (λ) + Π′2 Mts2 (λ) Mti (λ) = Mi (λ)xt , Mi (λ) = Θ′t Mk (λ) ′ −1 ′ Φc Θhi Mthi (λ) + Π′i Mtsi (λ) d Mt (λ) = , Φ′g −λ(1 − λ)gt From the structure of the Pareto problem and the fact that c1t , c2t appear additively in the technology (14.2.5), it follows that Mtc1 (λ) = Mtc2 (λ).

14.4. Competitive Equilibrium We will use the following Definition: A price system is a list of stochastic processes [{p0it , wt0 ; qt0 , rt0 , αt0 }∞ t=0 , v0 ], 2 each element of which belongs to L0 . Definition: An allocation is a list of stochastic processes {cit , sit , hit , ℓit , i = 2 1, 2; kt }∞ t=0 each element of which is in L0 .

Competitive Equilibrium

315

14.4.1. Households Households of type i face the problem of maximizing −

(12.1)

∞ 1 X t E β (sit − bit ) · (sit − bit ) + ℓ2it |J0 2 t=0

subject to the budget constraint E

∞ X t=0

β t p0t · cit | J0 = E

∞ X t=0

i β t [wt0 ℓit + αt0 · dit ] | J0 + v0 k−1 ,

(14.4.1)

the household technology sit = Λi hit−1 + Πi cit

(12.2)

hit = ∆h hit−1 + Θhi cit ,

(12.3)

and the initial conditions hi,−1 , ki,−1 .

14.4.2. Firms of type I and II Firms of type I and II face the problems described in chapter 6, with ct = c1t +c2t and dt = d1t + d2t .

14.4.3. Definition of competitive equilibrium We use the following standard definition: Definition: A competitive equilibrium is an allocation and a price system such that, given the price system, the allocation solves the optimum problem of households of each type and firms of each type.

316

Non-Gorman Heterogeneity Among Households

14.5. Computation of Equilibrium To compute an equilibrium, we use an iterative algorithm based on ideas of Negishi (1960). For each given value of the Pareto weight λ , we know that there exists a competitive equilibrium, though it will typically be associated with some distribution of wealth other than the one associated with the allocation of ownership of capital and endowment processes that we have assigned. An algorithm for computing an equilibrium with a pre-assigned distribution of ownership is to search for a λ ∈ (0, 1) that delivers budget balance for each household.

14.5.1. Candidate equilibrium prices For a given λ , candidate equilibrium prices can be computed from the Lagrange multipliers associated with the solution of the Pareto problem for that value of λ . By pursuing arguments paralleling those of Chapter 6, we find p0t = Mtc1 (λ)/µw 0 rt0 = Γ′ Mtd (λ)/µw 0 qt0 = Θ′k Mtk (λ) αt0 = Mtd (λ)/µw 0

(14.5.1)

w k ′ v0 = Γ′ M0d (λ)/µw 0 + ∆k M0 (λ)/µ0

wt0 = λ(1 − λ) | Sg (λ)xt | /µw 0 These prices and the associated allocations are the ingredients used in the Negishi algorithm.

Computation of Equilibrium

317

14.5.2. A Negishi algorithm The algorithm consists of the following steps. 1. For a given λ ∈ (0, 1), solve the Pareto problem. Compute the Lagrange multipliers from (14.3.12) and use them to compute the candidate competitive equilibrium prices and quantities via (14.5.1). 2. At the candidate equilibrium prices and quantities, compute the left and right side of each household’s budget constraint (14.4.1). In particular, compute Gi = E

∞ X t=0

β t [wt0 ℓit + αt0 · dit ] | J0 + v0 · ki,−1 − E

∞ X t=0

β t p0t · cit | J0

Use the method described in chapter 10 to compute this. For our twohousehold economy, G1 and G2 will either be of opposite signs, or both will equal zero. 3. If G1 > 0, increase λ and return to step 1. If G1 = G2 = 0, terminate the search and accept the allocation and price system associated with the current value of λ as equilibrium objects. 1 In practice, one can improve on this algorithm by using any of a number of root finders to find the zero of the function Gi (λ) defined in step 2. We have found it efficient to use a ‘secant method.’

1 The Negishi algorithm is implemented in the MATLAB program solvehet.m. Some trial inputs are contained in the file clex11h.m, which inputs a two agent version of the economy in clex11.m, the one good stochastic growth model. As a benchmark, clex11h.m has the economy start out with identical endowments for the two households, and has them share identical household technologies. With these inputs, solvehet.m should find Pareto weight (which the program calls ‘alpha’) equal to .5, and should recover the same solution that solvea.m does with inputs clex11.m. Modify the inputs to get a non-Gorman aggregatable example. The program simulhet.m simulates the equilibrium computed by solvehet.m.

318

Non-Gorman Heterogeneity Among Households

14.6. Mongrel Aggregation Except in the special case that Λ1 = Λ2 , Π1 = Π2 , ∆h1 = ∆h2 , Θh1 = Θh2 , the specification of household technologies (14.2.2) – (14.2.3) violates the Gorman conditions for aggregation, so that there does not exist a representative household in the sense described in Chapter 7. However, for each Pareto weight λ , there does exist a representative household in the sense of a mongrel preference ordering over total consumption (c1t + c2t ). This mongrel preference ordering depends on the distribution of wealth, i.e., the value of initial endowments and capital stocks evaluated at equilibrium prices.

14.6.1. Static demand Mongrel aggregation of preferences is easiest to analyze in the special case that the demand curve is ‘static’ in the sense that time t demand is a function only of the current price p0t . Let the household technology be determined by a nonsingular square matrix Π , where each of Λ, ∆h , Θh are matrices of zeros of the appropriate dimensions. For this specification, the (canonical) demand curve is ct = Π−1 bt − µ0 Π−1 Π−1′ pt , (14.6.1) where µ0 is the Lagrange multiplier on the household’s budget constraint. The inverse demand curve is −1 ′ ′ pt = µ−1 0 Π bt − µ0 Π Πct .

(14.6.2)

In equations (14.6.1) and (14.6.2), the price vector pt can be interpreted as the marginal utility vector of the consumption vector pt . Integrating the marginal utility vector shows that preferences can be taken to be (−2µ0 )−1 (Πct − bt ) · (Πct − bt ).

(14.6.3)

From (14.6.2) or (14.6.3) it is evident that the preference ordering is determined only up to multiplication of Π, bt by a common scalar. We are free to normalize preferences by setting µ0 = 1. Now suppose that we have two consumers, i=1,2, with demand curves −1 −1′ cit = Π−1 i bit − µ0i Πi Πi pt .

Mongrel Aggregation

319

Adding these gives the total demand −1 −1 −1′ c1t + c2t = (Π−1 + µ02 Π2 Π−1′ 1 b1t + Π2 b2t ) − (µ01 Π1 Π1 2 )pt .

(14.6.4)

Setting c1t + c2t = ct and solving (14.6.4) for pt gives −1′ −1′ −1 −1 pt = (µ01 Π−1 + µ02 Π−1 (Π−1 1 Π1 2 Π2 ) 1 b1t + Π2 b2t ) −1′ −1′ −1 − (µ01 Π−1 + µ02 Π−1 ct . 1 Π1 2 Π2 )

(14.6.5)

We want to interpret (14.6.5) as an aggregate preference ordering associated with an aggregate demand curve of the form (14.6.2). To do this, we shall evidently have to choose the Π associated with the aggregate ordering to satisfy −1 −1′ −1′ −1 ′ µ−1 + µ02 Π−1 . 0 Π Π = (µ01 Π1 Π2 2 Π2 )

(14.6.6)

To find a matrix Π determining an aggregate preference ordering, we have to form and then factor the matrix on the right side of (14.6.6). This matrix looks like the inverse of a weighted sum of two moment matrices. Even after normalizing Π by setting µ0 = 1, a solution Π will in general depend on the ratio µ01 /µ02 , which functions like a Pareto weight on the two types of consumers. There is a special case for which the aggregate or mongrel preference matrix Π is independent of µ01 /µ02 , namely: Π1 = kΠ2 for scalar k > 0

(14.6.7)

Notice that when Π1 and Π are scalars, condition (14.6.7) is automatically satisfied. So for the one consumption good case with this special specification (i.e., with Λ being zero), Gorman aggregation obtains. In the more general case that demand curves are dynamic (quantities demanded at t depending on future prices), attaining a mongrel preference ordering becomes more difficult. In place of the problem of factoring a moment matrix as required in (14.6.6), we have to factor something that resembles a spectral density matrix, frequency by frequency. For studying mongrel preference orderings in the general case, it is convenient to work with a frequency domain representation of preferences.

320

Non-Gorman Heterogeneity Among Households

14.6.2. Frequency domain representation of preferences From chapter 9, recall the decomposition of services st = smt + sit , where smt are services resulting from market purchases of consumption and sit are services flowing from the initial household capital stock. Let (∆h , Θh , Λ, Π) correspond to a canonical household service technology, and recall that smt = σ(L)ct where σ(L) = [Π + ΛL[I − ∆h L]−1 Θh ]

σ(L)−1 = Π−1 − Π−1 Λ[I − (∆h − Θh Π−1 Λ)L]−1 Θh Π−1 L, and sit = Λ∆th h−1 . We use the transform methods described in the appendix to chapter 9. For P∞ t ′ any matrix sequence {yt } satisfying t=0 β yt yt < +∞ , define T (y) (ζ) = P∞ t/2 t .5 yt ζ . We define S(ζ) = σ(β ζ). Evidently, the transforms obey t=0 β T (sm ) (ζ) = S(ζ)T (c)(ζ)

T (sit )(ζ) = Λ[I − β 1/2 ∆h ζ]−1 h−1 . As in chapter 9, express the one-period return as (st − bt ) · (st − bt ) = smt · smt + 2smt · sit − 2smt · bt + (bt − sit ) · (bt − sit )

(14.6.8)

The term (bt − sit ) · (bt − sit ) is beyond control and therefore influences no decisions. So it can be ignored in describing a preference ordering. In terms of Fourier transforms, we have ∞ X

1 = 2π

Z

π

1 2π

Z

π

1 β smt · bt = 2π t=0

Z

π

t=0

t

β smt · smt

∞ X t=0

∞ X

β t smt · sit = t

T (c)′ S ′ ST (c) dθ

(14.6.9)

−π

T (c)′ S ′ T (si ) dθ

(14.6.10)

T (c)′ S ′ T (b) dθ,

(14.6.11)

π

−π

A Programming Problem for Mongrel Aggregation

321

where it is understood that S = S(ζ), T (c) = T (c)(ζ), T (b) = T (b)(ζ), and ζ = e−iθ . Here ( ′ ) denotes transposition and complex conjugation.

14.7. A Programming Problem for Mongrel Aggregation To find a preference ordering over aggregate consumption, we can pose a nonstochastic optimization problem. 2 We rely on a certainty equivalence result to assert that the preference ordering over random consumption streams is given by the conditional expectation of the optimized value of this nonstochastic problem. Thus, our strategy in deducing the mongrel preference ordering over ct = c1t +c2t is to solve the programming problem: maximize over {c1t , c2t } the criterion ∞ X t=0

β t [λ(s1t − b1t ) · (s1t − b1t ) + (1 − λ)(s2t − b2t ) · (s2t − b2t )]

(14.7.1)

subject to hjt = ∆hj hjt−1 + Θhj cjt sjt = ∆j hjt−1 + Πj cjt ,

j = 1, 2 j = 1, 2

c1t + c2t = ct subject to (h1,−1 , h2,−1 ) given, and {b1t }, {b2t }, {ct } being known and fixed sequences. Substituting the {c1t , c2t } sequences that solve this problem as functions of {b1t , b2t , ct } into the objective (14.7.1) will determine the mongrel preference ordering over {ct }. In solving this problem, it is convenient to proceed by using Fourier transforms. Using versions of (14.6.8), (14.6.9), (14.6.10), and (14.6.11) for households 1 and 2, in terms of transforms we can represent the Pareto-weighted average of discounted utility from consumption as

2 These calculations are done by the MATLAB program .

322

Non-Gorman Heterogeneity Among Households

−

∞ X t=0

β t [λ(s1t − b1t ) · (s1t − b1t ) + (1 − λ)(s2t − b2t ) · (s2t − b2t )] =−

1 2π

Z

π

−π

n [λT (c1 )′ S1′ S1 T (c1 ) + (1 − λ)T (c2 )′ S2′ S2 T (c2 )]

+ 2[λT (c1 )′ S1′ T (s1i ) + (1 − λ)T (c2 )′ S2′ T (s2i )] o − 2[λT (c1 )′ S1′ T (b1 ) + (1 − λ)T (c2 )′ S2′ T (b2 )] dθ

(14.7.2)

+ terms not involving T (c1 ) or T (c2 ). where it is understood that each transform on the right side is to be evaluated at ζ = e−iθ . We want to maximize the right side of (14.7.2) over choice of {c1t , c2t }∞ t=0 or equivalently T (c1 ), T (c2 ), subject to the constraint c1t +c2t = ct , or equivalently the restriction T (c1 ) + T (c2 ) = T (c). To do this optimization, we form the Lagrangian J =− +

1 2π

Z

π

[λT (c1 )′ S1′ S1 T (c1 ) + (1 − λ)T (c2 )′ S2′ S2 T (c2 )]

−π 2[λT (c1 )′ S1′ T (s1i )

+ (1 − λ)T (c2 )′ S2′ T (s2i )]

− 2[λT (c1 )′ S1′ T (b1 ) + (1 − λ)T (c2 )′ S2′ T (b2 )] + µ[T (c) − T (c1 ) − T (c2 )] dθ

(14.7.3)

where it is understood that there is a Lagrange multiplier µ = µ(e−iθ ) for each frequency θ ∈ [−π, π]. We can perform this maximization “frequency by frequency” (i.e., pointwise for each θ ∈ [−π, π]). The first-order necessary conditions with respect to T (c1 ) and T (c2 ), respectively, are λS1′ S1 T (c1 ) + λS1′ T (s1i ) − λS1′ T (b1 ) = µ/2

(1 − λ)S2′ S2 T (c2 ) + (1 − λ)S2′ T (s2i ) − (1 − λ)S2′ T (b2 ) = µ/2

Using these two equations and the constraint T (c1 ) + T (c2 ) = T (c) to solve for µ gives h1 i−1 1 µ=2 (S1′ S1 )−1 + (S2′ S2 )−1 λ 1−λ (14.7.4) n o × T (c) + S1−1 T (s1i ) − T (b1 ) + S2−1 T (s2i ) − T (b2 ) .

A Programming Problem for Mongrel Aggregation

Let S′S =

h1

λ

(S1′ S1 )−1 +

i−1 1 (S2′ S2 )−1 1−λ

323

(14.7.5)

where S is a matrix Fourier transform that satisfies the condition: det S(ζo ) = 0 implies |ζo | > 1. Equation (14.7.5) is the dynamic counterpart of (14.6.6), and collapses to (14.6.6) in the special case in which Sj (L) = Πj . Recall that in the static case, constructing the aggregate preference ordering required factoring −1′ the inverse of a ‘moment’ matrix formed as a weighted sum of Π−1 . j Πj 1 The matrix [ λ1 (S1′ S1 )−1 + 1−λ (S2′ S2 )−1 ]−1 can be regarded as a spectral density matrix. Equation (14.7.5) expresses S ′ S as a spectral factorization 1 of [ λ1 (S1′ S1 )−1 + 1−λ (S2′ S2 )−1 ]−1 . Later, we shall show how to achieve the factorization expressed in (14.7.5). For now, we just assume that we have an S that satisfies (14.7.5) and the condition that the zeros of det S(ζ) all exceed unity in modulus. Using (14.7.5) in (14.7.4) gives µ = 2S ′ S T (c)+S1−1 T (s1i ) − T (b1 ) + S2−1 T (s2i ) − T (b2 ) .

The fact that the Lagrange multiplier is the derivative of the return function with respect to T (c) implies that the mongrel return function has the representation

1 2π

′ T (c) S′S S ′ SS1−1 S ′ SS2−1 T (s1i ) − T (b1 ) S1−1′ S ′ S − − −π −1′ ′ T (s2i ) − T (b2 ) S2 S S − − T (c) T (s1i ) − T (b1 ) dθ T (s2i − T (b2 )

Z

π

where the blank terms do not involve T (c), and do not affect the choice of T (c). Therefore, we can represent the mongrel preference ordering over T (c) by Z π 1 T (c)S ′ ST (c) + (2T (c)′ S ′ )SS1−1 (T (s1i ) − T (b1 )) 2π −π (14.7.6) + (2T (c)S ′ )SS2−1 (T (s2i ) − T (b2 )) dθ.

324

Non-Gorman Heterogeneity Among Households

Compare this with the single agent case of chapter 9, in which the preference ordering was shown to be Z πn o 1 ¯ (¯ T (¯ c)′ S¯′ ST c) + 2T (¯ c)′ S¯′ (T (¯ si ) − T (¯b)) dθ, (14.7.7) 2π −π where we have put bars ( ¯ ) over the objects in (14.7.7) to represent the corresponding single agent-objects. Evidently, for the mongrel preference ordering (14.7.6) to match up with a single agent ordering (14.7.7), we can match objects up as T (c) ∼ T (¯ c) S ∼ S¯ SS1−1 (T (s1i ) − T (b1 ))+

SS2−1 (T (s2i ) − T (b2 )) ∼ T (¯ si ) − T (¯b)

(14.7.8)

where the object on the right of the right of the ∼ corresponds in each case to the single-agent 4. We have two major tasks to complete our work. First, we have to show how to achieve the factorization (14.7.5). Second, we have to show how to interpret and to implement the correspondence given in (14.7.8).

14.7.1. Factoring S ′ S To achieve the spectral factorization (14.7.5), we notice that 1 [ λ1 (S1′ S1 )−1 + 1−λ (S2′ S2 )−1 ] can be regarded as the spectral density matrix of a stochastic process cλt that is generated by the state space system −1 h1t = β .5 (∆h1 − Θh1 Π−1 1 Λ1 )h1t−1 + Θh1 Π1 s1t

−1 h2t = β .5 (∆h2 − Θh2 Π−1 2 Λ2 )h2t−1 + Θh2 Π2 s2t h1t−1 1 1 −1 −1 λ .5 Π2 Λ2 ] ct = β [− √ Π1 Λ1 , − √ h2t−1 1−λ λ s1t √1 + ( √1λ Π−1 Π−1 ) 1 , 1−λ 2 s2t

(14.7.9)

where (∆hi , Θhi , Λi , Πi ) are each associated with a canonical representation, s1t and where st = is a white noise with covariance Est s′t = I . Write this s2t

A Programming Problem for Mongrel Aggregation

system compactly as

325

˜ h ht−1 + Hst ht = ∆ cλt = Gλ ht−1 + Mλ st .

1 The spectral density of cλt can be directly computed to be [ λ1 (S1′ S1 )−1 + 1−λ (S2′ S2 )−1 ]. 3 We can factor the inverse of

1 1 [ (S1′ S1 )−1 + (S ′ S2 )−1 ] λ 1−λ 2 by obtaining an innovations representation for the system (14.7.9), then using it to form a ‘whitener’. The innovations representation is ˆt = ∆ ˆ t−1 + Kat ˜ hh h ˆ t−1 + at cλ = Gλ h

(14.7.10)

t

˜ h , Gλ , HH ′ , Mλ M ′ , where Eat a′t = Ω = Gλ ΣG′λ +Mλ Mλ′ , and [K, Σ] = kfilter (∆ λ ′ HMλ ), where kfilter is the matrix valued function defined in chapter 7 . 1 (S2′ S2 )−1 ], let r′ r = Ω be the To get the inverse of [ λ1 (S1′ S1 )−1 + 1−λ Cholesky decomposition of Ω, and define sˆt by sˆt = r′ −1 at . Then use (14.7.10) to get the ‘whitener’ ˆ t = (∆ ˆ t−1 + Kcλ ˜ h − KGλ )h h t ′−1 ′−1 λ ˆ sˆt = −r Gλ ht−1 + r c t

or

ˆt = ∆ ˆ t−1 + Θ ˆ hh ˆ h cλ h t λ ˆ ˆ ˆ sˆt = Λht−1 + Πc

(14.7.11)

t

where

ˆ h = (∆ ˜ h − KGλ ) ∆ ˆh = K Θ ˆ = −r′−1 Gλ Λ ˆ = r′−1 Π

3 To see this, we make use of the fact that −1 −1 −1 Θ Π−1 ζ. σj (ζ)−1 = Π−1 hj j j − Πj Λj [I − (∆hj − Θhj Πj Λj )ζ]

(14.7.12)

326

Non-Gorman Heterogeneity Among Households

As a consequence of the factorization identity and associated matrix identities described in chapter 7, 4 we have that S ′ S satisfies (14.7.5) where h i ˆ + Λζ[I ˆ ˆ h ζ]−1 Θ ˆh S(ζ) = Π −∆ i h = Π + Λβ .5 ζ[I − ∆h β .5 ζ]−1 Θh . It follows that a (canonical) version of the mongrel household technology is ht = ∆h ht−1 + Θh (c1t + c2t ) st = Λht−1 + Π(c1t + c2t ), where

ˆ h, ∆h = β −.5 ∆ ˆ Λ = β −.5 Λ,

ˆh Θh = Θ ˆ Π = Π.

(14.7.13a)

(14.7.13b)

Collecting results, we have that the canonical household technology is determined by the matrices ˜ h − KGλ ) ∆h = β −.5 (∆ Θh = K

Λ = −r′

Π = r′

−1

−1

Gλ β −.5

.

These equalities imply that ∆h − Θh Π−1 Λ =

∆h1 − Θh1 Π−1 1 Λ1 0

0 . ∆h2 − Θh2 Π−1 2 Λ2

It follows that for the mongrel household technology, the counterpart to (BLANK.1) is 1 1 ′ ct = [− √ Π−1 Π−1 2 Λ2 ]ht−1 + r st 1 Λ1 , − √ 1−λ λ ∆h1 − Θh1 Π−1 0 1 Λ1 ht = ht−1 + Kr′ st . Λ 0 ∆h2 − Θh2 Π−1 2 2 4 In effect, we are using the factorization identity (7.33) and the matrix inversion identity (7.36). We are expressing (S ′ S)−1 from ( 14.7.5 ) first in a form like (7.31a), then via the factorization identity in a form like (7.31b). Then we apply the inversion formula (7.36) to the two factors of (7.31b), replacing Ω with its Cholesky factorization, to construct a factored version of S ′ S .

The Mongrel Preference Shock Process

327

Notice how the weight λ influences this representation: λ appears in the ‘observer’ matrix multiplying ht−1 in the first equation, and it appears indirectly through its influence on the matrices [r′ , K]. However, the state transition matrix ∆h − Θh Π−1 Λ is independent of λ .

14.8. Summary of Findings We have found that the operator σj (L)−1 is implemented by the state space −1 −1 −1 system defined by the four matrices 5 [∆hj − Θhj Π−1 j Λj , Θhj Πj , Πj Λj , Πj ]. The operator σ(L) associated with the mongrel household technology is realized by the state space system [∆h , Θh , Λ, Π] determined by (14.7.12) and (14.7.6). We can use these state space systems to derive a state space system for the mongrel preference shock.

14.9. The Mongrel Preference Shock Process Our next goal is to construct a mongrel preference shock process that achieves the match up given in (14.7.8). Evidently, from (14.7.6), we have to operate on T (s1i ) − T (b1 ) with the “filter” SS1−1 , operate on T (s2i ) − T (b2 ) with the “filter” SS2−1 , then add the results to get a process that we can interpret as the mongrel T (si ) − T (b). The following cascading of state space systems evidently implements the required filtering and adding:

B

(

zt+1 b 1t b2t A hjt sjt

= A22 zt + C2 wt+1 = Ub1 zt = Ub2 zt = ∆hj hjt−1 = Λj hjt−1

j = 1, 2 , j = 1, 2

−1 xjt = (∆hj − Θhj Π−1 j Λj )xjt−1 + Θhj Πj (bjt − sjt ) −1 yjt = −Π−1 j Λj xjt−1 + Πj (bjt − sjt )

, j = 1, 2

5 Defined as usual as the matrices in the state equation followed by the matrices in the observation equation.

328

Non-Gorman Heterogeneity Among Households

C

(

gt = ∆h gt−1 + Θh (y1t + y2t ) (bt − sˆt ) = Λgt−1 + Π(y1t + y2t ).

System A generates the “inputs” (b1t , b2t ), (si1t , si2t ). System B operates on P −1 i (bjt − sijt ) with σj−1 . System C operates on j σj (bjt − sjt ) with σ , as required by (14.7.8). In C , we are free to set sˆt = 0, and to regard the resulting bt as our mongrel preference shock process. A recursive representation of bt − st is attained by linking the three systems in a series. 6 It is evident how to use similar methods to break out the processes bt and sˆt separately.

14.9.1. Interpretation of sˆt component The term SS1−1 T (s1i ) has the following interpretation. T (s1i ) is (the transform of) the contribution of services flowing to the household from the initial household capital stock h−1 . Then S1−1 T (s1i ) is the (transform of the) equivalent amount of consumption that it would have taken to generate those services had they been acquired through new market purchases. The term S1−1 T (s1i ) amounts to a consumption goods equivalent of the transient component of services flowing to the first household.

14.10. Choice of Initial Conditions Our calculations do not tell us how to choose the correct initial condition at time 0 for the mongrel household capital stock vector ht−1 . Here is the reason. The first-order necessary conditions leading to (14.7.4)–(14.7.5) imply the following solution for the transform T (c1 ): 1 1 ′ (S S1 )−1 (S ′ S)T (c) + (S1′ S1 )−1 (S ′ S)S2−1 (T (s2i − T (b2 )) λ 1 λ 1 + ( (S1′ S1 )−1 (S ′ S) − I)S1−1 (T (b1 ) − T (s1i )). λ (14.10.1)

T (c1 ) =

6 The MATLAB command series can be used to link the systems.

Choice of Initial Conditions

329

A similar expression holds for T (c2 ). Our calculations assure that T (c1 ) + T (c2 ) = T (c). We assume that T (c) is the transform of a sequence that is onesided (i.e., ct = 0 ∀t < 0), but this does not guarantee that T (c1 ) and T (c2 ) are each transforms of one-sided sequences, only that their sum is. When Sj′ Sj is not a constant times S ′ S for j = 1, 2, as will generally be the case when the two household technologies are not identical, then T (c1 ) given by (14.10.1) will be the transform of a sequence that is nonzero for t < 0. 7 Thus, our frequency domain programming problem allows the ‘mongrel planner’ to reallocate past consumptions between the two types of households, subject to the restriction c1t + c2t = 0 for t < 0. These choices of cjs for s < 0 translate into choices of initial conditions for h−1 , the vector of mongrel household capital stocks at date t = −1. We will not pursue calculations of the initial conditions here, because they are intricate and only effect the transient responses of services. Our main interest is not in the selection of the initial conditions but in the ‘nontransient’ part of the mapping from total consumption ct = c1t +c2t to the mongrel service vector st , which is given by (14.7.6).

7 The operator (S ′ S )−1 (S ′ S) is two-sided except when it is proportional to the identity 1 1 operator.

Part III Extensions

Chapter 15 Equilibria with Distortions

15.1. Introduction In earlier chapters, we often used the fact that the competitive equilibrium allocations solve a Pareto problem. This chapter describes classes of economies for which the connection between Pareto optimality and equilibrium breaks down, because distortions cause competitive equilibrium allocations not to be Pareto optimal. The distortions occur in the form of externalities in preferences and production technologies, and distorting taxes. These features can be accommodated using the invariant subspace methods described in chapter 8, which provide an approach to studying the existence and uniqueness of an equilibrium, and if one exists, to computing it. This chapter describes ways of adapting our earlier methods to compute, represent, and manipulate equilibria. 1 We describe two classes of economies with distortions. The first class is designed partly as a warmup for the second, but is also interesting in its own right. In this first class of models, an equilibrium can be computed by one pass through the invariant subspace algorithm. This is possible possible because of two sorts of simplifying assumptions: first, that there is a representative agent (which permits Gorman-heterogeneity);and

1 The invariant subspace computational approach follows the tradition of Blanchard and Khan (1980), Whiteman (1983), Anderson and Moore (1985), King, Plosser, and Rebelo (1989), and McGrattan (1991) in expressing the equilibrium conditions as a system of linear difference equations, then adapting the method of Vaughan to solve it. The principal caveat to keep in mind in applying these methods is that in distorted economies, there is no guarantee that the eigenvalues of the matrix ‘characteristic p equation’ of the system of equilibrium conditions will ‘split’ half into those exceeding 1/ β in modulus, and half into those falling

p

short of 1/

β in modulus. Failure of the eigenvalues to divide in that way signals either an

existence or a uniqueness problem.

– 333 –

334

Equilibria with Distortions

second, that the government has the ability to levy lump sum taxes or transfers. 2 Thus, the first setup is a representative agent setting with consumption and production externalities and distorting taxes. 3 . In the second setup, equilibria must be computed with multiple passes through the invariant subspace algorithm. What necessitates this is that there are (non-Gorman) heterogeneous agents with consumption externalities and distorting taxes, but no lump sum taxes or transfers. In this setting, equilibrium computation means finding a set of taxes and households’ marginal utilities of wealth that assure present value budget balance for households and the government. We find the equilibrium taxes and marginal utilities of wealth by repeatedly resorting to the invariant subspace algorithm, constructing candidate prices and quantities at each pass of the invariant subspace algorithm, then checking every agent’s budget constraint at each pass. It is convenient to describe these two classes of models sequentially, partly because the first one is complicated enough, and because it reveals about half of the difficulties involved in calculating models of the second class. We shall use our machinery to compute equilibria of some simple models embodying preference and technology specifications that have occurred in the recent literature.

2 The MATLAB program solvdist.m computes equilibria for the first type of economy, while disthet.m computes equilibria for the second type of economy. The MATLAB program compare.m is useful for comparing the results of disthet.m with related undistorted economies whose equilibria have been computed with solvea.m. 3 The first class of models are linear-quadratic relatives of the ones studied by Braun (1991) and McGrattan (1991b). In particular, both Braun and McGrattan use lump sum transfers to balance the government’s budget.

A Representative Agent Economy with Distortions

335

15.2. A Representative Agent Economy with Distortions We alter the model discussed in chapters 3 and 5 by adding two sorts of distortions: consumption and production externalities, and distorting taxes. We add these distortions in a way that is designed to preserve the applicability of the forms of equilibrium prices and quantities: we want the equilibrium law of motion to be linear in the state, quantities and (scaled Arrow-Debreu) prices to be linear functions of the state, and asset prices to be quadratic functions of the state. We add the following distortions to the Chapter 3–5 model.

15.2.1. a. Consumption externalities The technology for producing household services is now taken to be st = Λh ht−1 + ΛH Ht−1 + Πct + ΠC Ct ht = ∆h ht−1 + ∆H Ht−1 + Θh ct + ΘH Ct , where Ht−1 is the vector of aggregate household capital stocks, and Ct is the vector of aggregate consumption rates. In equilibrium, ct = Ct and ht = Ht , but the representative household is assumed to take {Ht , Ct } as given and beyond control when allocating its resources.

15.2.2. b. Production externalities Firms produce subject to the linear technology Φc (ct + Gt ) + Φi it + Φg gt = Γk kt−1 + ΓK Kt−1 + dt where Kt−1 is the vector of aggregate capital stocks, and Gt is government purchases of consumption goods. In equilibrium, kt−1 = Kt−1 , but the representative firm is assumed to regard {Kt−1 } as beyond its control when choosing its inputs and outputs. The presence of Kt−1 on the right side of the technology constraint is designed to accommodate externalities of the sort analyzed by Paul Romer (1985). Below, we shall assign ownership of the ‘technology’ so that households sell the joint process ΓK Kt−1 + dt to firms. This assignment will confront the firm with a constant returns to scale technology. We assume that the government purchase vector process is exogenous and satisfies Gt = UG zt .

336

Equilibria with Distortions

15.2.3. c. Taxes We add to the model of chapters 3–5 a government that taxes consumption goods, investment goods, capital goods, and the intermediate labor activity. The government’s budget constraint is ∞ X 0′ 0′ 0′ 0′ t β pt τc ct + qt τi it + rt τk kt−1 + wt τℓ gt |J0 E t=0

=E

∞ X t=0

β t p0t · Gt |J0 + T0

where Gt = UG zt . Here τc is a diagonal matrix of tax rates on consumption goods, τi is a diagonal matrix of tax rates on investment goods, τk is a diagonal matrix of tax rates on rentals of capital, and τℓ is a diagonal matrix of tax rates on the labor-using activity vector gt ; T0 is lump sum transfers to the household. The government sets the tax rates τc , τi , τk , τℓ and the expenditure process UG zt = Gt . For this economy, a price system is a collection of stochastic processes {p0t , qt0 , rt0 , 2 wt0 , αt0 }∞ t=0 , each element of which belongs to L0 . A tax system is a list of diagonal matrices [τc , τi , τk , τℓ ]. An allocation is a collection of stochastic processes 2 {st , ct , ht , kt , gt , it }∞ t=0 each element of which belongs to L0 . We now describe the choices faced by households and firms.

15.3. Households Households own an ‘endowment stream’ {dt + ΓK Kt−1 }∞ t=0 and the initial capital stocks, all of which they take as given. Households take the price system, the tax system, and {Ct , Ht−1 , Kt−1 }∞ t=0 as given, and choose contingency plans ∞ for {ct , it , st , ht , kt , gt }t=0 to maximize −

∞

1 X t E β [(st − bt ) · (st − bt ) + gt · gt ] | J0 2 t=0

(15.3.1)

subject to the household technology,

st = Λh ht−1 + ΛH Ht−1 + Πct + ΠC Ct

(15.3.2)

ht = ∆h ht−1 + ∆H Ht−1 + Θh ct + ΘH Ct

(15.3.3)

Firms

337

the law of accumulation for physical capital, kt = ∆k kt−1 + ∆K Kt−1 + Θk it ,

(15.3.4)

and the budget constraint E

∞ X t=0

′

′

β t [p0t (I + τc )ct + qt0 (I + τi )it ] | J0 =E

∞ X t=0

′

β t [wt0 (I − τℓ )gt + αt0 · (dt + ΓK Kt−1 )

(15.3.5)

′

+ rt0 (I − τk )kt−1 ] | J0 + T0

15.4. Firms There is one kind of firm, a production firm that takes the price system 2 and {Kt−1 }∞ t=0 as fixed, and chooses an allocation and a process {ζt } ∈ L0 that maximize E

∞ X t=0

β t {p0t · (ct + Gt ) + qt0 · it − rt0 · kt−1 − αt0 · ζt − wt0 · gt |J0 }

(15.4.1)

subject to the technology Φc (ct + Gt ) + Φg gt + Φi it = Γk kt−1 + ζt · ζt .

(15.4.2)

338

Equilibria with Distortions

15.5. Information We assume that zt+1 = A22 zt + C2 wt+1 , where {wt } is a martingale difference sequence with identity covariance matrix. The stochastic process {zt } drives the exogenous processes bit = Ubi zt dit = Udi zt Gt = UG zt .

15.6. Equilibrium An equilibrium is defined as an allocation, a price system, and a tax system such that (i) Given the price system, the tax system, and {Ht−1 , Ct−1 , Kt−1 }∞ t=0 , the allocation solves the household’s problem; (ii) Given the price system and {Kt−1 }∞ t=0 , the allocation solves the firm’s problem with ζt ≡ ΓK Kt−1 + dt ; (iii) The representative household is representative, Ht−1 = ht−1 , Ct = ct ; and the representative firm is representative, Kt−1 = kt−1 for all t ≥ 0. Given that conditions (i), (ii), (iii), are satisfied, the lump sum tax T0 , which is a present value, can be chosen to make the government budget constraint hold. We have set things up so that the equilibrium law of motion takes the form xt+1 = Ao xt + Cwt+1 , and equilibrium quantities and normalized Arrow-Debreu prices are linear func′ tions of the state x′t = [h′1t−1 h′2t−1 kt−1 zt′ ]. Appendix A describes how to compute the equilibrium of this model by manipulating agents’ first order conditions and the other equilibrium conditions

Heterogeneous Households with Distortions

339

into a system susceptible to the application of Vaughan’s algorithm. The MATLAB program solvdist.m computes an equilibrium. Equilibrium calculation is facilitated by the fact that there is a representative households (so that if there is heterogeneity among households, the features of Gorman aggregation isolate demand functions from effects of redistributions of wealth) and the permission that we give the government to levy lump sum taxes. In the next class of models, we give up both of these simplifying features.

15.7. Heterogeneous Households with Distortions In this section, we describe how to extend the above setup to enable us to compute the equilibrium of a model with externalities, government expenditures and taxes, and two classes of agents with non-Gorman aggregable preferences. To produce the model, we combine elements of the preference specification treated in chapter 8 with the specification and methods described earlier in this chapter. We do not let the government raise lump sum taxes or dispense lump sum transfers, but require the government to balance its budget by levying only flat rate taxes. These changes vis a vis the first class of models have the consequence that equilibrium must now be computed by finding values of tax rates and marginal utilities of wealth for each household that make households’ and the government’s budgets balance. 4

4 Between these two classes of models, we have changed the way we price labor or ‘intermediate goods’ gt . In the first class, the households own and sell to the firm the vector gt , which is priced by the vector wt0 . In the second model, households sell ℓit for wt0 to firms.

340

Equilibria with Distortions

15.7.1. Households There are two classes of households, indexed by i = 1, 2. Preferences of a household of type i are ordered by −.5E

∞ X t=0

β t [(sit − bit ) · (sit − bit ) + ℓ2it ]|J0 ,

(15.7.1)

where the household has access to the household technology sit = Λi1 hit−1 + Λi2 H1t−1 + Λi3 H2t−1 + Πi1 cit−1 + Πi2 C1t + Πi3 C2t hit = ∆hi hit−1 + ∆Hi1 Hit−1 + ∆Hi2 H2t−1 + Θhi cit + ΘH1i C1t + ΘH2i C2t

(15.7.2)

dit = Udi zt bit = Ubi zt . Here Cit , Hit are the aggregate consumption and stock of durables, respectively, of household type i, and cit , hit are the individual consumption and durables, respectively, of a household of type i. There are firms of two types, to be described in detail below. A household of type i is assumed to own a a fraction fi of the ‘technology’, which entitles it to sell fi (Γk + dt ) of an ‘endowment’ to a type I (production) firm, and to rent fi ∆K Kt−1 of capital to a type II (capital renting) firm at time t . Thus, the intertemporal budget constraint of a household of type i is E

∞ X t=0

′

β t {p0t (I + τc )cit − fi αt0 · (ΓK Kt−1 + dt ) −

fi rt0

· ∆K Kt−1 −

wt0 (1

(15.7.3)

+ τℓ )ℓit }|J0 − v0 k−1,i = 0

No lump sum transfers occur in (15.7.3), because we shall require the government to balance its budget using flat rate taxes only. Below, we will carry along a Lagrange multiplier µ0i on (15.7.3) for each type of household.

Heterogeneous Households with Distortions

341

15.7.2. Firms of type I Firms of type I are production firms. They maximize

E

∞ X t=0

β t {p0t · (ct + Gt ) + qt0 · it − rt0 · kt−1 − αt0 · ζt − wt0 ℓt }|J0

(15.7.4)

subject to

Φc (ct + Gt ) + Φi it + Φg gt = Γk kt−1 + ζt , gt · gt = ℓ2t

(15.7.5)

15.7.3. Firms of type II Firms of type II are capital renting firms. They maximize

E0

∞ X t=0

′

′

β t {rt0 (I − τk )kt−1 − qt0 (I + τi )it − νt rt · ∆K Kt−1 }|J0 − v0 k−1 (15.7.6)

subject to kt = ∆k kt−1 + νt ∆K Kt−1 + Θk it . Here νt is the amount of the ‘endowment’ ∆K Kt−1 purchased from households.

15.7.4. Government The government makes a flow of expenditures Gt on consumption goods, governed by Gt = UG zt . (15.7.7) The government’s budget constraint is ∞ X ′ ′ ′ β t {p0t · Gt − p0t τc · (c1t + c2t ) − qt0 τi it − rt0 τk kt−1 E t=0

−

τℓ wt0 (ℓ1t

(15.7.8)

+ ℓ2t )}|J0 = 0

A fiscal policy is a collection of diagonal matrices (τc , τi , τk , τℓ ) determining taxes and a matrix UG determining government expenditures.

342

Equilibria with Distortions

15.7.5. Definition of equilibrium An equilibrium is a price system [{p0t , qt0 , rt0 , αt0 , wt0 }∞ t=0 , v0 ], an individual allocation {c1t , c2t , h1t , h2t , kt , it , gt }, an aggregate allocation {C1t , C2t , H1t , H2t , Kt }, and a fiscal policy (τc , τi , τk , τℓ , UG ) that satisfy the following conditions: i. Given the aggregate allocation, the price system, and the fiscal policy, the individual allocation solves the optimum problems of households and the firms, with ζt ≡ 1 and νt ≡ 1. ii. The allocations satisfy cit = Cit , i = 1, 2 hit = Hit , i = 1, 2 kt = Kt ζ = (ΓK Kt−1 + dt ). iii. The government budget constraint is satisfied. As we have defined it, an equilibrium is typically not unique. In particular, there will usually be a set of taxes that serve to assure equilibrium. Below, we shall select one from among these equilibria partly by fixing enough of these taxes and ‘solving’ for others. 5

15.7.6. Equilibrium computation We compute an equilibrium by finding a fixed point in the parameters that index the tax system τc , τi , τk , τℓ and a pair of marginal utilities of wealth (multipliers) µ01 , µ02 for households. For fixed values of the tax system and multiplier parameters, we can use the modified Vaughan algorithm described in the preceding section to compute a ‘candidate’ 0 allocation and price system. For that allocation and price system, we evaluate all elements of the budget constraints of the government and the two types of households. We use a nonlinear search algorithm We use a secant algorithm. to find a tax system and pair of multipliers µ01 , µ02 that assure that the budget constraints are all satisfied. Thus, we compute an equilibrium by solving a fixed point problem in a space of tax rates and marginal utilities of wealth for the different household types. 5 Even when we fix the ‘right number’ of the taxes there can still be multiple equilibria for reason of ‘Laffer curves’.

Government Deficits and Debt

343

Appendix B describes how the first order conditions for households and firms together with the other equilibrium conditions can be arranged to obtain a system of equations of the form (9.25), which with the invariant subspace methods of chapter 9 permits us to compute our candidate equilibrium. The equilibrium is computed by disthet.m.

15.8. Government Deficits and Debt The government deficit at time t , measured in time t ‘spot’ prices, is Dtt = ptt · (Gt − τc (c1t + c2t )) − qtt · τi it − rtt · τk kt−1 − wtt τℓ (ℓ1t + ℓ2t ). Evidently, Dtt can be represented as Dtt = where

x′t QD xt , e1 Mc xt

(15.8.1)

QD = Mc′ [SG − τc (Sc1 + Sc2 )] − Mi′ τi Si τℓ − Mk′ τk Sk1 − S ′ Sg . (1 − τℓ )(µ01 + µ02 ) g

Let Vt denote the present value of the government deficit, which satisfies the difference equation Vt = −Dtt + βEt {pt1,t+1 Vt+1 }, (15.8.2) where pt1,t+1 = e1 Mc xt+1 /e1 Mc xt is the time t price of a state contingent claim to the first (numeraire) consumption good in time t + 1. Notice that βpt1,t+1 acts as a stochastic discount factor for evaluating government indebtedness next period from the standpoint of this period. This equation is the counterpart of the following version of a one period government budget constraint, which occurs in various nonstochastic macroeconomic models: −Dt + Vt+1 /Rt = Vt , where here Vt is interpreted as one-period debt falling due at time t . Equations (15.8.1) and (15.8.2) imply that Vt =

xt QV xt + σV , e1 Mc xt

where QV , σV are determined as follows. Define v˜ =

β doublej2(βAo , C, Ao , C). 1−β

(15.8.3)

344

Equilibria with Distortions

Then [QV ] = doublej2(βAo ′ , QD , Ao ′ , 1) and σV = trace(QV v˜). The matrix valued function doublej2 was used repeatedly in our asset pricing calculations. The forms of (15.8.2) and (15.8.3) mean that the econometric methods described in Hansen and Sargent (1993) for interpreting observations on asset prices can be used to model government budgets and bond holdings. 6

15.9. Examples

15.9.1. A production externality The following technology is designed to capture features of a specification of DeLong and Summers. There is one consumption good, but two capital goods (‘machines’ and ‘structures’) and two rates of investment. Machines generate a positive production externality, but not structures: ct +Gt + i1t + i2t = γ1 k1t−1 + γ2 k2t−1 + Γ1 K1t−1 + Γ2 K2t−1 + dt φ1 i1t = g1t φ2 i2t = g2t k1t = δk1 k1t−1 + i1t k2t = δk2 k2t−1 + i2t To capture DeLong and Summers’s idea, we set Γ1 = 0, Γ2 > 0, so that we interpret the first capital good as structures and the second as machines. To complete this example, we incorporate a single-agent version of a simple quadratic utility specification. We suppress heterogeneity among consumers and instead make the two types of consumers be identical in their preferences and endowment sequences. Preferences are ordered by −

∞

1X t β [(cit − bit )2 + ℓ2it ]. 2 t=0

6 The restrictions embedded in ( 15.8.2 ) and ( 15.8.3 ) should be compared with those studied in the literature on linear models of ‘present value budget balance.’ See Hansen, Roberds, and Sargent (1991) for a summary of this literature.

Examples

The information process satisfies 1 0 0 0 .95 0 A22 = 0 0 .8 0 0 0

0 0 0 .1

d1t = d2t = [ 3.5 Gt = [ 5

345

0

1√ 20 2 C2 = 0 0

0

1 0

.5

0 ] zt ,

0 0 .5 0

0 0 0 1

1 ] zt .

b1t = b2t = 30. These settings make endowments of the two types of households each the same first order autoregressive processes with mean 3.5 and serial correlation parameter .8, while government expenditures follows the process √ (20 2)−1 .5 Gt − 5 = w1t + w3t , 1 − .95L 1 − .1L where L is the lag operator. This process approximates a mixture of a ‘permanent shock’ (the moving average in w1t ) and a ‘transitory shock’ (the moving average in w3t . The mean of government expenditures is 5. We set γ1 = γ2 = .12, Γ1 = 0, Γ2 = .04, φ1 = φ2 = .5, δk1 = δk2 = .95, β = 1/1.05. These parameter settings make the two types of capital symmetric with respect to adjustment costs and ‘private productivity’ (γi ), but inject a positive production externality for the second capital good.

15.9.2. Consumption tax only With these parameter settings, we computed an equilibrium where the only tax is the scalar consumption tax τc on the single consumption good. The equilibrium tax τc = .2497. Figures 15.9.1.a and 15.9.1.b show a simulation of this equilibrium starting from the initial condition k−1,1 = k−1,2 = 100. Because we start from equal initial stocks of machines and structures, and because their private productivities are the same, the equilibrium must retain equality of structures and machines throughout time. Figures 15.9.1.a and 15.9.1.b embody this property, the rates of investment in machines and structures being identical. Figures 3 display realizations of government indebtedness Vt determined by (15.8.2). Figure 4 shows the government flow deficit.

346

Equilibria with Distortions

110

25

109 20

108

cons

107 15

106 105

10

104 inv

103

5

102 govt

0 0

10

20

30

101 40

50

60

70

80

90

100 0

100

Fig. 15.9.1.a. Total consumption, investment, and government expenditures in Delong-Summers economy with τc = .2497 and no investment subsidy.

10

20

30

40

50

60

70

80

90

100

Fig. 15.9.1.b. Machines and structures in Delong-Summers economy with τc = .2497 and no investment subsidy.

15.9.3. Machinery investment subsidy Figures 15.9.2 and 15.9.3 report the results of recomputing the equilibrium when we impose an investment subsidy τi2 = −.04. The only other tax that we permit the consumption tax, which must be set at τc = .2611 to induce equilibrium. The simulations show how the investment rates for machines and structures now diverge in the direction that we would expect: the economy moves to a path with more machines and fewer structures than the first (no machine subsidy) equilibrium. 7

7 Realizations of the exogenous stochastic processes are held constant across these simulations.

Examples

12

347

15

10 10 8 6

5

4 0

2 0

-5 -2 -4 0

10

20

30

40

50

60

70

80

90

100

Fig. 15.9.2.a. Present value of government surplus in Delong-Summers economy with τc = .2497 and no investment subsidy.

-10 0

10

20

30

40

50

60

70

80

90

100

Fig. 15.9.2.b. Government deficit in Delong-Summers economy with τc = .2497 and no investment subsidy.

15.9.4. ‘Personal’ habit persistence We consider an economy with one consumption good that is allocated to two types of households. The first type of household has preferences ordered by ∞ X β t [(c1t − 15)2 + ℓ21t ], E0 − .5 t=0

while the second has the habit-persistence indicated by the preferences ∞ X β t [(s2t − 15)2 + ℓ22t ] E0 − .5 t=0

where

s2t = −1h2t−1 + 2c2t

h2t = .8h2t−1 + .2c2t We set β = 1/1.05. The production technology has

ct + it = .11kt−1 + .0001Kt−1 + dt kt = .95kt−1 + it g1t = .5it

348

Equilibria with Distortions

125

25

120

cons

20

115 15

machines

110

105

10 machines

100

structures

5 structures

0 0

10

20

30

40

50

60

95

70

80

90

90 0

100

Fig. 15.9.3.a. Total consumption, investment, and government expenditures in Delong-Summers economy with τc = .2611 and investment subsidy, τi2 = −.04 .

10

20

30

40

50

60

70

80

90

100

Fig. 15.9.3.b. Machines and structures in Delong-Summers economy with τc = .2611 and investment subsidy, τi2 = −.04 .

which is a version of our one-good “growth, adjustment cost” technology with a small production externality. We set A22

Ud2

1 = 0 0

0 .8 0

0 0 .5

UG = [ 4 .1 0 ] 3.5 0 1 = Ud1 = 0 0 0 k−1,1 = 125, k−1,2 = 25.

Notice that we endow the first household with more capital and equal claim to the endowment stream, so the first household is richer. We do the following experiments with this economy. First, we set all taxes except τc equal to zero, and solve for the equilibrium values of (µ02 , τc ). They are (2.0765,.2117). A simulation of this economy is in figure 7. Then we reset

Examples

349

τi = .08, and solve again for (µ02 , τc ) = (2.0467, .27). A simulation of this economy is in figure 8. The simulations start from identical initial conditions. In this economy, the habit persistence of the second type of consumer is something of an “engine of growth”. The second type of household accumulates capital to support planned growth in its consumption. Taxing investment causes the household to cut back on these investments.

14

13 12

c1

12 c1

11

10

10

c2 9

8

8

c2 invest

6

7 6

4

invest

govt

5 2 4

govt

3 0

10

20

30

40

50

60

70

80

90

100

Fig. 15.9.4.a. Simulation of two-agent economy with τc = .2117, τi = 0 . Four series are plotted: consumption of type 1 consumer, consumption of type 2 consumer, investment, and government purchases. Equilibrium marginal utilities of wealth are µ01 = 1, µ02 = 2.0765 .

0 0

10

20

30

40

50

60

70

80

90

100

Fig. 15.9.4.b. Simulation of two-agent economy with τc = .1884, τi = .08 . Equilibrium marginal utilities of wealth are (1, 2.0467) .

350

Equilibria with Distortions

15.9.5. ‘Social’ habit persistence Our next example is identical to the previous one, except that we alter the equation generating ‘habits’ of the second type of household to h2t = .8h2t−1 + .2C2t . We computed equilibria of this economy with the same government expenditure process and the same permissible tax instruments as for the previous economy. Figures 15.9.4 and 15.9.5 report simulations of this economy starting from the same initial conditions as for the previous economy. Notice that slightly lower consumption tax rates provide equilibria in this economy, for the reason that due to the ‘social’ rather than ‘personal’ nature of habit persistence, they have slightly lower adverse demand effects on the second type of household.

14

12 c1

11

c1

12 10 9

10

8

c2 8

7 6

c2

6

invest

invest

5 4 4 3 0

govt

govt 10

20

30

40

50

60

70

80

90

100

Fig. @[email protected] Simulation of two-agent economy with τc = .2101, τi = 0 . The second type of agent has ‘keeping up with the Jones’ habit persistence with other agents of his type. Four series are plotted: consumption of type 1 consumer, consumption of type 2 consumer, investment, and government purchases.

2 0

10

20

30

40

50

60

70

80

90

100

Fig. 15.9.5.b. Simulation of two-agent economy with τc = .1883, τi = .08 .

Invariant subspace equations for first specification

351

15.10. Conclusions The models that we have described are rigged to make it practical to extend the data matching exercises performed by Braun (1991) and McGrattan (1991b). Our calculations make possible various extensions of Braun and McGrattan. The ease of computing equilibrium quantities, prices, and present values means that we can use data on quantities, interest rates, and asset prices to do method of moment estimation along the lines described by Hansen and Sargent (1993, chapter 11). In particular, we can use data on government indebtedness and deficits as well as asset prices to help estimate parameters. By accepting our linear-quadratic specifications, we purchase the ability to get our hands on the Arrow-Debreu prices (which is harder with the approximations used by Braun, McGrattan and others in the real business cycle literature), which makes it feasible for us to do without lump sum taxes in our second class of models. It is possible for us to have numbers of capital stocks and enough heterogeneity among households to generate interesting tax incidence effects. Finally, the machinery in this chapter is a useful one within which to revisit issues in the literature on time series implications of ‘present value budget balance’ (e.g., Hansen, Roberds, and Sargent (1991)).

A. Invariant subspace equations for first specification Our strategy is to use Lagrangian methods to obtain first order necessary conditions for households and firms. After obtaining those first order conditions, we substitute the equilibrium conditions (Ht−1 = ht−1 , Kt−1 = kt−1 , Ct = ct , and so on) into them. Then we rearrange the system into the form of (9.25) so that it is susceptible to application of the modified Vaughan method. We can exploit the certainty equivalence principle and solve a nonstochastic version of the model first, and later adjust the solution to accommodate randomness.

352

Equilibria with Distortions

15.A.1. Household’s Lagrangian A Lagrangian for a nonstochastic version of the household’s problem is ∞ X

1 L= β − [(st − Ub zt ) · (st − Ub zt ) + gt · gt ] 2 t=0 t

′

+ µ0 [wt0 (I − τℓ )gt + αt0 (Ud zt + ΓK Kt−1 ) ′

′

′

+ rt0 (I − τk )kt−1 + Tt − p0t (I + τc )ct − qt0 (I + τi )it ]

+ µs′ t [Λh ht−1 + ΛH Ht−1 + Πct + ΠC Ct − st ]

(15.A.1)

+ µh′ t [∆h ht−1 + ∆H Ht−1 + Θh ct + ΘH Ct − ht ]

+ µk′ t [∆k kt−1 + ∆K Kt−1 + Θk it − kt ] z′ + µt [A22 zt − zt+1 ]

We can set the multiplier on the budget constraint µ0 = 1, which will amount to selecting a numeraire. The first-order conditions for the household’s problem are µst = (bt − st )

(15.A.2)

(I + τc )p0t = Π′ µst + Θ′h µht

(15.A.3)

µht = βΛ′h µst+1 + β∆′h µht+1

(15.A.4)

0 µkt = β∆′k µkt+1 + β(I − τk )rt+1

(15.A.5)

gt = (I − τℓ )wt0

(15.A.6)

(I + τi )qt0 = Θ′k µkt

(15.A.7)

0 µzt = β[A′22 µzt+1 + Ub′ (st+1 − Ub zt+1 ) + Ud′ αt+1 ]

(15.A.8)

Invariant subspace equations for first specification

353

The law of motion for the state variables chosen by the household and by nature is ht ∆h 0 0 ht−1 Θh 0 ct kt = 0 ∆k 0 kt−1 + 0 Θk it zt+1 0 0 A22 zt 0 0 (15.A.9) ∆H 0 ΘH Ht−1 + 0 ∆K 0 Kt−1 0 0 0 Ct

15.A.2. Firm’s first order conditions The first-order conditions for the firm’s problem are p0t = Φ′c αt0

(15.A.10)

qt0 = Φ′i αt0

(15.A.11)

wt0 = −Φ′g αt0

(15.A.12)

rt0 = Γ′k αt0

(15.A.13)

The feasibility condition is

Φc (ct + Gt ) + Φi it + Φg gt = Γk kt−1 + ΓK Kt−1 + dt

(15.A.14)

354

Equilibria with Distortions

15.A.3. Representativeness conditions Additional equilibrium conditions are ht−1 = Ht−1 kt−1 = Kt−1

(15.A.15)

ct = Ct We describe in detail how to compute the equilibrium of the first type of model by arranging its equilibrium conditions into the form of equation (9.25). Substituting the equilibrium conditions (15.A.15) into (15.A.9) gives ¯ ht ∆h kt = 0 zt+1 0

0 ¯k ∆

0

¯ 0 ht−1 Θh 0 kt−1 + 0 A22 zt 0

0 ct Θk it 0

(15.A.16)

¯ h = ∆h + ∆H , ∆ ¯ k = ∆k + ∆K , Θ ¯ h = Θh + ΘH . Equations (15.A.10) where ∆ and (15.A.12) imply ′ −1 0 Φc pt . (15.A.17) αt0 = Φ′g −wt0 Substituting from (15.A.3) and (15.A.6) into (15.A.17) gives αt0

Φ′c = Φ′g

−1

(I + τc )−1 [Π′ µst + Θ′h µht ] −(I − τℓ )−1 gt

(15.A.18)

Using kt−1 = Kt−1 in (15.A.14) and solving for [c′t gt′ ]′

ct gt

¯ t−1 + Uf zt − Φi it } = [Φc Φg ]−1 {Γk

(15.A.19)

¯ = Γk + ΓK . where Uf = Ud − Φc UG and Γ Substituting ht−1 = Ht−1 and ct = Ct into equation (15.3.2) st gives ¯ t st = Λht−1 + Πc ¯ = Π + ΠC . where Λ = Λh + ΛH and Π Collecting our results to this point, we want to solve the following system of difference equations µst = (Ub zt − st ) (15.A.20)

Invariant subspace equations for first specification

355

¯ t st = Λht−1 + Πc

(15.A.21)

µht = βΛ′h µst+1 + β∆′h µht+1

(15.A.22)

0 µkt = β∆′k µkt+1 + β(I − τk )Γ′k αt+1

(15.A.23)

0 µzt = β[A′22 µzt+1 + Ub′ (st+1 − Ub zt+1 ) + Ud′ αt+1 ]

¯ ht ∆h kt = 0 zt+1 0

0 ¯k ∆

αt0 =

0

ct gt

Φ′c Φ′g

0 0 A22

¯ ht−1 Θh kt−1 + 0 zt 0

(15.A.24)

0 ct Θk (15.A.25) it 0

¯ t−1 + Uf zt − Φi it } = [Φc Φg ]−1 {Γk

−1

(I + τc )−1 [Π′ µst + Θ′h µht ] −(I − τℓ )−1 gt

(15.A.26)

(15.A.27)

Φ′i αt = (I + τi )−1 Θ′k µkt ,

(15.A.28)

where the last equation comes from combining (15.A.11) (among the firm’s firstorder necessary conditions) with (15.A.7) (among the household’s first-order necessary conditions). Our goal is to manipulate this system into the form (9.25) that is susceptible to the application of the invariant subspace algorithm of chapter 9. Our strategy will be successively to eliminate it , ct , gt , αt0 , st , and µst from the system, so that we are left with a difference equation in the “state” (ht−1 , kt−1 , zt ) and the “co-state” variables µht , µkt , µzt . We begin by substituting (15.A.27) and (15.A.26) into (15.A.28) to get

Φ′i

Φ′c Φ′g

−1

(I + τc )−1 [Π′ µst + Θ′h µht ] ¯ t−1 + Uf zt − Φi it } −(I − τℓ )−1 Ug [Φc Φg ]−1 {Γk

−1

= (I + τi )

Θ′k µkt

(15.A.29)

356

Equilibria with Distortions

where as usual Ug is a selection matrix that picks off the components of the right side of (15.A.26) corresponding to gt . We adopt the partition ˜= Φ

Φ′c Φ′g

−1

˜1 Φ ˜ 2] = [Φ

(15.A.30)

˜ 1 is (nd × nc ) and Φ ˜ 2 is (nd × ng ). Then equation (15.A.29) can be where Φ written as ˜ 1 (I + τc )−1 [Π′ µst + Θ′h µht ] Φ′i Φ ˜ 2 (I − τℓ )−1 Ug [Φc Φg ]−1 {Γk ¯ t−1 + Uf zt } − Φ′i Φ ˜ 2 (I − τℓ )−1 Ug [Φc Φg ]−1 Φi it − (I + τi )−1 Θ′ µk = −Φ′ Φ k t

i

Solving for it gives it = L1 µst + L2 µht + L3 µkt + L4 kt−1 + L5 zt

(15.A.31)

where ′ ˜ −1 ′ L1 = G−1 Π 1 Φi Φ1 (I + τc ) −1 ′ ′ ˜ Θh L2 = G−1 1 Φi Φ1 (I + τc ) −1 ′ L3 = −G−1 Θk 1 (I + τi ) −1 ′ ˜ ¯ L4 = −G1 Φi Φ2 (I − τℓ )−1 Ug [Φc Φg ]−1 Γ −1 ′ ˜ −1 −1 L5 = −G1 Φi Φ2 (I − τℓ ) Ug [Φc Φg ] Uf ˜ 2 (I − τℓ )−1 Ug [Φc Φg ]−1 Φi G1 = −Φ′ Φ

(15.A.32)

i

Substituting (15.A.31) into (15.A.26) and rearranging gives

where

ct = L6 kt−1 + L7 zt + L8 µst + L9 µht + L10 µkt

(15.A.33)

gt = L11 kt−1 + L12 zt + L13 µst + L14 µht + L15 µkt

(15.A.34)

Invariant subspace equations for first specification

357

¯ − Φi L4 ) L6 = Uc [Φc Φg ]−1 (Γ

L7 = Uc [Φc Φg ]−1 (Uf − Φi L5 )

L8 = −Uc [Φc Φg ]−1 Φi L1

L9 = −Uc [Φc Φg ]−1 Φi L2

L10 = −Uc [Φc Φg ]−1 Φi L3 ¯ − Φi L4 ) L11 = Ug [Φc Φg ]−1 (Γ

(15.A.35)

L12 = Ug [Φc Φg ]−1 (Uf − Φi L5 ) L13 = −Ug [Φc Φg ]−1 Φi L1

L14 = −Ug [Φc Φg ]−1 Φi L2

L15 = −Ug [Φc Φg ]−1 Φi L3

˜1 Φ ˜ 2] = Substituting (15.A.34) into (15.A.27) and using our partition [Φ gives αt0 = L16 µst + L17 µht + L18 µkt + L19 kt−1 + L20 zt

Φc Φg

−1

(15.A.36)

where ˜ 1 (I + τc )−1 Π′ + Φ ˜ 2 (I − τℓ )−1 Ug [Φc Φg ]−1 Φi L1 L16 = Φ ˜ 1 (I + τc )−1 Θ′ + Φ ˜ 2 (I − τℓ )−1 Ug [Φc Φg ]−1 Φi L2 L17 = Φ h ˜ 2 (I − τℓ )−1 Ug [Φc Φg ]−1 Φi L3 L18 = Φ

(15.A.37)

˜ 2 (I − τℓ ) Ug [Φc Φg ] {Γ ¯ − Φi L4 } L19 = −Φ ˜ 2 (I − τℓ )−1 Ug [Φc Φg ]−1 (Uf − Φi L5 ) L20 = −Φ −1

−1

Substituting (15.A.33) for ct into (15.A.20) and (15.A.21) gives µst = L21 ht−1 + L22 kt−1 + L23 zt + L24 µht + L25 µkt where

(15.A.38)

L21 = −A −1 Λ ¯ 6 L22 = −A −1 ΠL

¯ 7) L23 = A −1 (Ub − ΠL −1 ¯ L24 = −A ΠL9 ¯ 10 L25 = −A −1 ΠL ¯ 8) A = (I + ΠL

(15.A.39)

358

Equilibria with Distortions

Because st = Ub zt − µst , (15.A.38) implies st = L26 ht−1 + L27 kt−1 + L28 zt + L29 µht + L30 µkt

(15.A.40)

where L26 = −L21

L27 = −L22

L28 = Ub − L23

(15.A.41)

L29 = −L24

L30 = −L25

We can use (15.A.38) to eliminate µst from the right sides of (15.A.31), (15.A.33), (15.A.34), (15.A.36) to obtain it = N1 ht−1 + N2 kt−1 + N3 zt + N4 µht + N5 µkt

(15.A.42)

ct = N6 ht−1 + N7 kt−1 + N8 zt + N9 µht + N10 µkt

(15.A.43)

gt = N11 ht−1 + N12 kt−1 + N13 zt + N14 µht + N15 µkt

(15.A.44)

αt = N16 ht−1 + N17 kt−1 + N18 zt + N19 µht + N20 µkt

(15.A.45)

where N1 N 2 N3 N4 N5

= L1 L21 = L1 L22 + L4 = L1 L23 + L5

(15.A.46)

= L1 L24 + L2 = L1 L25 + L3

N6 = L8 L21 N = L8 L22 + L6 7 N8 = L8 L23 + L7 N9 = L8 L24 + L9 N10 = L8 L25 + L10

(15.A.47)

Invariant subspace equations for first specification

N11 N12 N13 N14 N15 N16 N17 N18 N19 N20

359

= L13 L21 = L13 L22 + L11 = L13 L23 + L12

(15.A.48)

= L13 L24 + L14 = L13 L25 + L15 = L16 L21 = L16 L22 + L19 = L16 L23 + L20

(15.A.49)

= L16 L24 + L17 = L16 L25 + L18

We now substitute (15.A.42), (15.A.43), (15.A.44), (15.A.45) into (15.A.22), (15.A.23), (15.A.24), (15.A.25) to obtain the following system of difference equations: µht = βΛ′h L21 ht + βΛ′h L22 kt + βΛ′h L23 zt+1

(15.A.50)

+ (βΛ′h L24 + β∆′h )µht+1 + βΛ′h L25 µkt+1 µkt = [β∆′k + β(I − τk )Γ′k N20 ]µkt+1

+ β(I − τk )Γ′k N16 ht + β(I − τk )Γ′k N17 kt

+ β(I −

τk )Γ′k N18 zt+1

+ β(I −

(15.A.51)

τk )Γ′k N19 µht+1

µzt = βA′22 µzt+1 + β[Ub′ L26 + Ud′ N16 ]ht + β[Ub′ L27 + Ud′ N17 ]kt + β[Ub′ L28 + Ud′ N18 − Ub′ Ub ] zt+1

(15.A.52)

+ β[Ub′ L29 + Ud′ N19 ]µht+1

+ β[Ub′ L30 + Ud′ N20 ]µkt+1 ¯h + Θ ¯ h N6 )ht−1 + Θ ¯ h N7 kt−1 ht = (∆ ¯ h N10 µk ¯ h N8 z t + Θ ¯ h N9 µh + Θ +Θ t

¯ k + Θk N2 )kt−1 + Θk N1 ht−1 kt = (∆ + Θk N3 zt + Θk N4 µht + Θk N5 µkt zt+1 = A22 zt

(15.A.53)

t

(15.A.54) (15.A.55)

360

Equilibria with Distortions

We can arrange these equations in the form of (9.25) as follows: m2 where m2 =

0 0 0 ¯ ¯ h N6 ) (∆h + Θ Θk N1 0

and

xt µt

= m1

0 0 0 ¯ h N7 Θ ¯ k + Θk N2 ∆ 0

xt+1 , µt+1

0 0 0 ¯ h N8 Θ Θk N3 A22

I 0 0 ¯ h N9 Θ Θk N4 0

0 I 0 ¯ Θh N10 Θk N5 0

0 0 I 0 0 0

m1 =

βΛ′h L21 βΛ′h L22 βΛ′h L23 ′ β(I − τk )Γ′ N16 β(I − τk )Γk N17 β(I − τk )Γ′k N18 k β(U ′ L + U ′ N ) β(U ′ L + U ′ N ) β(U ′ L + U ′ N − U ′ U ) b 26 d 16 b 27 d 17 b 28 d 18 b b I 0 0 0 I 0 0 0 I βΛ′h L24 + β∆′h βΛ′h L25 0 β(I − τk )Γ′k N19 β∆′k + β(I − τk )Γ′k N20 0 ′ ′ ′ ′ β(Ub L29 + Ud N19 ) β(Ub L30 + Ud N20 ) βA′22 0 0 0 0 0 0 0

0

0

′ k′ z′ where we set x′t = [h′t−1 , kt−1 , zt′ ] and µ′t = [µh′ t µt µt ]. Then the equilibrium law of motion for {xt } is given by

xt+1 = Ao xt

(15.A.56)

where −1 V21 ). Ao = W11 ∆1 (V11 − V12 V22

The shadow prices are determined via (12.5b), namely, µt = M xt

(15.A.57)

Invariant subspace equations for heterogeneous agent model

361

−1 where M = W21 W11 . Using (15.A.42), (15.A.43), (15.A.44), and (15.A.45) with (15.A.56) we can write

or

n o it = [N1 N2 N3 ] + [N4 N5 0] M xt n o ct = [N6 N7 N8 ] + [N9 N10 0] M xt n o gt = [N11 N12 N13 ] + [N14 N15 0] M xt n o αt0 = [N16 N17 N18 ] + [N19 N20 0] M xt

(15.A.58) (15.A.59) (15.A.60) (15.A.61)

it = Si xt ct = Sc xt gt = Sg xt

(15.A.62)

αt0 = Sα xt . Substituting (15.A.62) for αt0 into (15.A.10), (15.A.11), (15.A.12), (15.A.13) gives p0t = Sp xt qt0 = Sq xt wt0 = Sw xt

(15.A.63)

rt0 = Sr xt where

We also have

Sp = Φ′c Sα , Sq = Φ′i Sα Sw = −Φ′g Sα , Sr = Γ′k Sα ht = Sh xt , kt = Sk xt

(15.A.64)

(15.A.65)

where Sh = [I 0 0] Ao

,

Sk = [0 I 0] Ao .

(15.A.66)

These formulas express the law of motion for the state xt and all quantities and prices in forms identical to those described in chapters 3–5. With these formulas in hand, all subsequent features of our analysis proceed identically as with that for an undistorted economy.

362

Equilibria with Distortions

B. Invariant subspace equations for heterogeneous agent model

From the first-order conditions of the two types of households and firms the market clearing conditions we can deduce the following set of equations:

s1t = Λ1 h1t−1 + Λ13 h2t−1 + Π1 c1t + Π13 c2t

(15.B.1)

s2t = Λ2 h2t−1 + Λ22 h1t−1 + Π2 c2t + Π22 c1t (15.B.2) ˜ h1 h1t−1 + ∆H12 h2t−1 + Θ ˜ h1 c1t + ΘH12 c2t (15.B.3) h1t = ∆ ˜ h2 h2t−1 + ∆H21 h1t−1 + Θ ˜ h2 c2t + ΘH21 c1t (15.B.4) h2t = ∆ zt+1 = A22 zt ct = Γkt−1 + dt − Φi it − Φc UG zt (Φc Φg ) gt ˜ k kt−1 + Θk it kt = ∆

(15.B.5)

(I +

τc )µ0i p0t Mhi t

= =

Π′i1 (Ubi zt − sit ) + Θ′hi Mhi t ′ hi ′ β∆hi Mt+1 + βΛi1 (Ubi zt+1

(15.B.6) (15.B.7) i = 1, 2

− sit+1 ) i = 1, 2 (15.B.9)

′ zi ′ Mzi t = β[A22 Mt+1 + Ubi (sit+1 − Ubi zt+1 )

αt0 =

Φ′c Φ′g

−1

′ 0 + Udi αt+1 µ0i ] i = 1, 2 0 pt wt0 − lt gt

lit = (1 − τl )µ0i wt0

(I +

Mkt τi )Φ′i αt0

=

=

β∆′k Mkt+1 Θ′k Mkt

i = 1, 2

+ β(I −

(15.B.8)

0 τk )Γ′k αt+1

(15.B.10) (15.B.11) (15.B.12) (15.B.13) (15.B.14)

Invariant subspace equations for heterogeneous agent model

363

where Λ1 = Λ11 + Λ12 Π1 = Π11 + Π12 Λ2 = Λ21 + Λ23 Π2 = Π21 + Π23 ˜ h1 = ∆h1 + ∆H11 ∆ ˜ h2 = ∆h2 + ∆H22 ∆ ˜ h1 = Θh1 + ΘH11 Θ ˜ h2 = Θh2 + ΘH22 Θ ˜ k = ∆k + ∆K ∆ Γ = Γk + ΓK Now, use (15.B.8) twice for i = 1, 2 and (15.B.1), (15.B.2), we get ′ µ−1 01 (Π11 (Ub1 zt − Λ1 h1t−1 − Λ13 h2t−1 − Π1 c1t − Π13 c2t )

+ Θ′h1 Mh1 t )=

′ µ−1 02 (Π21 (Ub2 zt − Λ2 h2t−1 − Λ22 h1t−1 − Π2 c2t − Π22 c1t )

(15.B.15)

+ Θ′h2 Mh2 t )

From (15.B.6) we get c1t + c2t = Uc (Φc Φg )−1 [Γkt−1 − Φi it + Uf zt ] −1

gt = Ug (Φc Φg )

[Γkt−1 − Φi it + Uf zt ]

(15.B.16) (15.B.17)

Also, combining (15.B.14), (15.B.11), (15.B.12) and (15.B.8) gives us ′˜ −1 µ−1 [Π′11 (Ub1 zt − Λ1 h1t−1 01 Φi Φ1 (I + τc )

− Λ13 h2t−1 − Π1 c1t − Π13 c2t ) + Θ′h1 Mh1 t ] ′˜ −1 ′ − Φ Φ2 gt /(1 − τl )(µ01 + µ02 ) = (I + τi ) Θ Mk i

k

(15.B.18)

t

˜1 , Φ ˜ 2 , Uc and Ug are as defined in Chapter 10, Uf = Ud1 + Ud2 − Here Φ Φc UG . Equations (15.B.15) – (15.B.18) can be used to solve 4 variables c1t , k c2t , gt and it in terms of the state variables zt , hit−1 , Mhi t , kt−1 and Mt , i = 1, 2, as:

Equilibria with Distortions

364

zt zt h1t−1 c1t h2t−1 ˜ c2t = N ˜ kt−1 L Mz1 gt t it Mz2 t h1 Mt Mh2

(15.B.19)

t

Mkt

where

˜= L

and

µ01 Π′21 Π22 − µ02 Π′11 Π1 Ic 0 −1 Π′ Π ′˜ µ−1 11 1 01 Φi Φ1 (I + τc )

µ01 Π′21 Π2 − µ02 Π′11 Π13 Ic 0 −1 Π′ Π ′˜ µ−1 11 13 01 Φi Φ1 (I + τc )

0 0 Ig ˜2 (1 − τl )−1 (µ01 + µ02 )−1 Φ′i Φ

0 Uc (Φc Φg )−1 Φi Ug (Φc Φg )−1 Φi 0

Invariant subspace equations for heterogeneous agent model

365

˜ = N [−µ02 Π′11 Ub1 ]′ [µ01 Π′21 Ub2 ]′ [µ02 Π′11 Λ1 − µ01 Π′21 Λ22 ]′ [µ02 Π′11 Λ13 − µ01 Π′21 Λ2 ]′ 0 0 0 [−µ02 Θ′h1 ]′ [µ01 Θ′h2 ]′ 0

[Uc (Φc Φg )−1 Uf ]′ 0 0 0 [Uc (Φc Φg )−1 Γ]′ 0 0 0 0 0

[Ug (Φc Φg )−1 Uf ]′ 0 0 0 [Ug (Φc Φg )−1 Γ]′ 0 0 0 0 0

−1 Π′ U ]′ ′ ′˜ [µ−1 11 b1 01 Φi Φ1 (I + τc ) 0 −1 Π′ Λ ]′ ′˜ [−µ−1 11 1 01 Φi Φ1 (I + τc ) ˜ 1 (I + τc )−1 Π′ Λ13 ]′ [−µ−1 Φ′ Φ 01

11

i

0 0 0 ′Φ ˜ 1 (I + τc )−1 Θ′ ]′ [µ−1 Φ 01 i h1 0 [−(I + τi )−1 Θ′k ]′

From equation (15.B.19) c1t , c2t , gt , it can be expressed as a linear combination of state vector xt and costate vector Mt , as: c1t c2t = N xt gt Mt it

where ˜ ˜ −1 N N =L xt = ( zt′

zt′

h′1t−1

h′2t−1

′ kt−1 )

′

and Mt is the corresponding multipliers of each component of xt .

Divide N into blocks according to the dimensions of cit , gt , it , xt and Mt , such that:

Equilibria with Distortions

366

c1t = N11 xt + N12 Mt = N1 (x′t M′t )′

c2t = N21 xt + N22 Mt = N2 (x′t M′t )′

gt = N31 xt + N32 Mt = N3 (x′t M′t )′

(15.B.20)

it = N41 xt + N42 Mt = N4 (x′t M′t )′

As an intermediate step, we express s1t , s2t and αt0 in terms of xt and Mt . This is done by using (15.B.1), (15.B.2), (15.B.8), (15.B.11) and (15.B.12):

s1t = Ns1 (x′t M′t )′

s2t = Ns2 (x′t M′t )′

(15.B.21)

αt0 = Nα (x′t M′t )′

such that

Ns1 = Π1 N1 + Π13 N2 + (0 0 Λ1 Λ13 0 0 0 0 0 0) Ns2 = Π22 N1 + Π2 N2 + (0 0 Λ22 Λ2 0 0 0 0 0 0) −1 ˜ Nα = µ−1 [−Π′11 Ns1 + (Π′11Ub1 0 0 0 0 0 0 Θ′h1 0 0)] 01 Φ1 (I + τc ) ˜ 2 N3 − (1 + τl )−1 (µ01 + µ02 )−1 Φ

Now, it is possible to combine all equations (15.B.1)–(15.B.14) to write down the Vaughan’s linear equations:

˜1 M

where

xt+1 Mt+1

˜2 =M

xt Mt

(15.B.22)

Invariant subspace equations for heterogeneous agent model

I 0 0 0 0 ˜1 = M ′ U −βUb1 b1 0 βΛ′ Ub1 0 11 0 0

0 I 0 0 0 0 ′ U −βUb2 b2 0 βΛ′21 Ub2 0

0 0 I 0 0 0 0 0 0 0

0 0 0 I 0 0 0 0 0 0

0 0 0 0 I 0 0 0 0 0

0 0 0 0 0 βA′22 0 0 0 0

0 0 0 0 0 0 βA′22 β∆′h1 0 0

0 0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 β∆′h2 0

0 0 0 0 0 0 0 I 0 0

0 0 0 0 0 0 0 0 I 0

0 0 0 0 0 β ′ ′ N α Ub1 Ns1 + µ01 Ud1 U ′ N + µ U ′ Nα 02 d2 b2 s2 ′ N −Λ 11 s1 −Λ′21 Ns2 (I − τk )Γ′k Nα

and A22 ˜2 = M

0 0 0 0 0 0 0 0 0

0 A22 0 0 0 0 0 0 0 0

0 0 0 0 0 0 ˜ ∆ ∆H12 0 h1 ˜ ∆H21 ∆ 0 h2 ˜ 0 0 ∆ k 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 ˜ Θ N2 h1 N1 + ΘH12 ˜ N2 Θ N1 + Θ H21

+

h2

Θk N4 0 0 0 0 0

0 0 0 0 0 I 0 0 0 0

0 0 0 0 0 0 I 0 0 0

0 0 0 0 0 0 0 0 0 I

0 0 0 0 0 0 0 0 β∆′k

367

+

Chapter 16 Recursive Risk Sensitive Control

16.1. Introduction This chapter describes a class of preferences with a single additional parameter that permits us to represent altered attitudes toward risk vis a vis our earlier quadratic specification. We build on work by Jacobson (1973) and Whittle (1990), who proposed an exponential type of risk adjustment that, in linear– quadratic environments with Gaussian disturbances, preserves most of the computational conveniences of the standard undiscounted linear quadratic dynamic programming problem. But Whittle’s (1981, 1990) way of introducing discounting into this problem had the unpleasant feature of introducing time dependence into the optimal decision rules, with the time dependence causing the effects of the risk parameter σ to wear off as the planning horizon is extended. By using a recursive specification of utility along the lines of Koopmans, Lucas and Stokey, Epstein and Zin, and Weil, Hansen and Sargent (1992) altered Whittle’s specification to model discounting in a way that implies time invariance of optimal decision rules. In this chapter we describe this preference specification and how it can be implemented with versions of our earlier formulas for computing equilibria and asset prices. The principal features of our specification are that: (a.) the equations for solving the optimal linear regulator problem are modified by replacing the operator associated with the usual matrix Riccati equation with the composition of that operator with another operator that is easily computed and interpreted; (b.) optimal decision rules remain linear in the state; (c.) ‘certainty equivalence’ no longer holds, i.e., the decision rules, although linear in the state, depend on the conditional covariance matrix of the innovations to the state; (d.) asset pricing formulas have an additional layer of ‘risk adjustments,’ because the covariance conditional covariance matrix of the innovations to the state now enter the quadratic form in the state in our asset pricing formulas.

– 369 –

370

Recursive Risk Sensitive Control

16.2. A Control Problem This section describes a ‘discounted linear quadratic exponential Gaussian control problem that we shall eventually use for the social planning problem of our standard economy. This is an infinite horizon control problem associated with iterations to convergence on the following equation in the value functions Vj (x): Vj+1 (x) = max′ {u′ Qu + x′ Rx + u,x

2β log E(exp(σVj (x′ )/2)|J)}, σ

(16.2.1)

where the maximization is subject to x′ = Ax + Bu + Cw, where w is a Gaussian random vector with Eww′ = I , and where β ∈ (0, 1) is a discount factor. Associated with the solution of this control problem are three operators: D(V ) = V + σV C(I − σC ′ V C)−1 C ′ V

T (W ) = R + A′ (βW − β 2 W B(Q + βB ′ W B)−1 B ′ W )A β d(k, V ) = βk − log det(I − σC ′ V C). σ

(16.2.2)

The operator T is the usual one associated with the matrix Riccati difference equation for the discounted optimal linear regulator problem. We shall give an interpretation of the new operator D shortly. The value function associated with the solution of the infinite horizon control problem is V∞ (x) = x′ U1 x + U0 (16.2.3) where U1 = lim (T ◦ D)j (0) j→∞

U0 = lim dj (0, 0) j→∞

The optimal decision rule is time invariant ut = −F xt where F = β[Q + βB ′ D(U1 )B]−1 B ′ D(U1 )A.

(16.2.4)

These formulas can be derived by solving the maximization problem on the right side of (16.2.1) and using a Lemma stated by Jacobson (1973) on

Pessimistic Interpretation

371

a property of the Gaussian distribution. For details, see Hansen and Sargent (1992). 1 The optimal decision rule is linear in the state. When σ = 0, we get the standard linear–quadratic situation, because in that case D = I , so that the operator T ◦ D = T . When σ = 0, the feedback rule F is independent of the conditional covariance parameters in C , so that ‘certainty equivalence’ holds when σ = 0. When σ 6= 0, the decision rule F depends on the parameters in C.

16.3. Pessimistic Interpretation For some purposes, it is convenient to use an interpretation of the D operator due to Jacobson (1973) and Whittle (1990). Evaluate the following ‘aggregator function’ associated with the functional equation (16.2.1): A(−F x, x, y ′ V y + k|J) = x′ R∗ x +

2β log E{exp(σ(y ′ V y + k)/2)|J} (16.3.1) σ

where A∗ = A − BF, R∗ = F ′ QF + R, and y = A∗ x + Cw. We obtain A(−F x, x, y ′ V y + k|J) = x′ [R∗ + βA∗′ D(V )A∗ ]x + d(k). The piece x′ [R∗ + βA∗′ D(V )A∗ ]x has an interpretation in terms of ‘pessimism,’ to be seen as follows. Consider the following deterministic minimum problem β min{− w′ w + x′ R∗ x + βy ′ V y} w,y σ subject to y = A∗ x + Cw. 1 This specification of discounting differs from the one used by Whittle (1990) in a way designed to assure time-invariant decision rules.

372

Recursive Risk Sensitive Control

In this problem, w is treated as a control vector. As long as (I − σC ′ V C) is positive definite, this problem has a unique solution. The minimized value of the criterion is verified to be x′ [R∗ + βA∗′ D(V )A∗ ]x , which is the quadratic in x part of A(−F x, x, y ′ V y + k|J). The ‘solution’ for w is w = σ(I − σC ′ V C)−1 C ′ V A∗ x. Thus, the optimum problem gives an interpretation of the operator D and a piece σ(I − σC ′ V C)−1 C ′ V A∗ x in terms of a ‘pessimistic’ (when σ < 0) or ‘optimistic’ (when σ > 0) adjustment to the random term wt in the law of motion for the state.

16.4. Recursive Preferences The preceding generalization of linear–quadratic control theory can readily be applied in the context of representative agent versions of the class of economic models that we have been studying. We use the usual household technology st = Λht−1 + Πct ht = ∆h ht−1 + Θh ct . We define the information process zt as usual by zt+1 = A22 zt + C2 zt , and let the law of motion for the complete state xt be denoted xt+1 = Axt + Cwt+1 , where now wt is a Gaussian white noise. We define a utility index recursively by Ut = −(st − bt ) · (st − bt )/2 + βRt (Ut+1 ) where Rt (Ut+1 ) =

2 log E[exp(σUt+1 /2)|Jt ]. σ

(16.4.1)

Asset Pricing

373

Hansen, Sargent, and Tallarini (1993) have shown how to construct the prices induced by this preference specification. We describe these prices first in the context of an endowment economy. 2

16.4.1. Endowment economy Consider a pure endowment economy in which ct = Uc zt is exogenous. Infinite recursions on (16.4.1) lead to the quadratic form Ute = x′t Ωxt + ρ, where Ω, ρ are the limits as j → ∞ of recursions on Ωj+1 = −(Ss − Sb )′ (Ss − Sb )/2 + βA′ D(Ωj )A ρj+1 = d(ρj , Ωj ).

For endowment economies, we use Ω as one of the ingredients in constructing a probability measure appropriate for the inner-product representation for asset pricing.

16.5. Asset Pricing We want to represent asset prices in the usual way as a conditional mathematical expectation of an infinite discounted sum of an inner product of a ‘scaled ArrowDebreu price’ vector and the payout of the asset. Hansen, Sargent, and Tallarini (1993) show that when σ 6= 0, the required conditional expectations operator for representing asset prices corresponds to a probability distribution that is distorted relative to the ‘objective’ one. In particular, they show that to attain the inner product representation of asset prices it is appropriate to construct a conditional expectations operator Ft defined by Ft Ut+1 = E(Vt+1 Ut+1 |Jt )/E(Vt+1 |Jt ), 2 The Matlab program solvex.m computes equilibrium quantities and prices in one of our economies with this preference specification; solvex.m calls doublex.m, which implements a doubling algorithm to solve the control problem.

374

Recursive Risk Sensitive Control

where e Vt+1 = exp(σUt+1 /2).

The operator Ft behaves like a conditional expectation operator in that it is linear, monotone, and maps bounded random variables that are measurable with respect to Jt+1 into bounded random variables that are measurable with respect to Jt . 3 We can use βFt to value a contingent claim to time t + 1 utility, and Mst for the equilibrium valuation of services. Recursively define the sequence of expectations operators St,τ = Ft Ft+1 Ft+2 · · · Ft+τ −1 , where Ft,0 = I . Then the valuation of a stream of consumption services {st } is just ∞ X St,τ Mst+τ · st+τ . t=0

To compute equilibrium asset prices, we have to evaluate consumption rates ct . We compute a multiplier Mct from Mst in the usual way, except that we substitute the conditional expectation operators St,τ for the usual ones in equation (6.*) (refer to asset pricing chapter): Mct = Π′ Mst + Θ′h

∞ X

τ =1

β τ (∆h )τ ′ Λ′ St,τ (Mst+τ ).

3 V t+1 /EVt+1 |Jt is used as a Radon-Nikodym derivative in this construction.

Characterizing the Pricing Expectations Operator

375

16.6. Characterizing the Pricing Expectations Operator Hansen, Sargent, and Tallarini show how to evaluate the conditional expectations operators St,τ by constructing a distorted probability measure for the state xt . 4 In particular, consider the phony law of motion for xt ˆ t + Cw ˆ t+1 xt+1 = Ax where

Aˆ = [I + σC(I − σC ′ ΩC)−1 C ′ Ω]A Cˆ Cˆ ′ = C(I − σC ′ ΩC)−1 C ′ .

Then we can compute Mct = Mc xt where Mc = [Π′ + Θ′h

∞ X

τ =1

β τ ∆τh′ Λ′ Aˆτ ](Sb − Ss ).

We can evaluate a claim on a stream dt = Sd xt by ∞ X

τ =0

ˆt β τ St,τ Mct+τ · dt+τ = E = x′t

∞ X

τ =0 ∞ X

β τ Mct+τ · dt+τ β τ (Aˆ′ )τ Sd′ Mc Aˆτ xt

τ =0

+

β 1−β

trace[Sd′ Mc

∞ X

β τ Aˆτ Cˆ Cˆ ′ (A′ )τ ].

τ =0

4 These calculations are implemented in the program assetxq.m.

376

Recursive Risk Sensitive Control

16.7. Production Economies The preceding formulas also apply to asset pricing in production economies, with the understanding that the matrix Ω now corresponds to the ‘quadratic-form’ matrix in the value function for the social planning problem.

16.8. Risk-Sensitive Investment under Uncertainty For sake of illustration, we now consider a one consumption good, one capital good production economy that is a version of a Brock-Mirman stochastic growth model with adjustment costs. This delivers a linear-quadratic version of Lucas and Prescott’s (1971) theory of investment under uncertainty. Consumption and investment satisfy: ct + it = γkt−1 + dt (16.8.1) where the capital stock kt evolves according to: kt = δk kt−1 + it

(16.8.2)

and {dt } is an exogenous endowment process. Labor input is required to adjust the capital stock, reflected in a quadratic adjustment cost in the preferences of the fictitious social planner. Also, there is an exogenous ideal consumption level process {b1t }. The time t contribution to preferences is: −(ct − b1t )2 − φ2 it 2 = −(γkt−1 + dt − it − b1t )2 − φ2 it 2 which is quadratic in the control it , the endogenous state variable kt−1 and the exogenous states b1t and dt . We begin with a parameter specification that implies no investment in a deterministic steady state, an unrealistic but useful starting point that we adopt to compare two alternative ‘explanations’ for investment. One is that altering the risk parameter σ induces a precautionary savings motive into the preference ordering. The other is that capital is more productive than in the benchmark economy, increasing the physical return to investment. Benchmark Economy with No Steady-State Investment We set β = 1/1.05, γ = .1, δk = .95, φ = .5. With this parameter configuration and constant values of dt and b1t , the model implies steady state values

Risk-Sensitive Investment under Uncertainty

377

of zero for capital and investment. This follows from the fact that γ + δk = β −1 , which equates the physical rate of return to investment net of adjustment costs to the subjective rate of time discount. Since adjustment costs are also present, reflected by a positive value of φ , physical investment in new capital becomes ‘unattractive.’ As a consequence, asymptotically there are no savings in the deterministic version of this model. To see this, the Euler equation for capital is given by E[(1 − L−1 )(1 − β −1 L)kt + φ2 (1 − βδk L−1 )(1 − δk L)kt | Jt ] = E[(1 − L)(b1t+1 − dt+1 ) | Jt ]

(16.8.3)

where L is the lag operator. For future reference we have permitted there to be uncertainty in the forcing processes for {b1t } and {dt }. The solution to this stochastic difference equation has representation: kt = λkt−1 + ψ[(dt − b1t ) − (1 − βλ)

∞ X j=0

(βλ)j E(dt+j − b1t+j | Jt )].

(16.8.4)

where 0 < λ < 1, and 0 < ψ . In contrast to the familiar permanent income model in which λ and ψ are one, the presence of adjustment costs lowers λ . Notice that the term multiplying ψ is the difference between dt − b1t and a geometric average of current and future values of dt − b1t . This simple link to the forcing processes is a result of the γ + δk = β −1 restriction. For a model with b1t and dt constant, the term multiplying ψ will be zero and the capital stock sequence will converge to zero. An Economy with Risk Sensitivity As a precursor to illustrating investment induced by risk sensitivity, we now introduce a specific model of uncertainty in the endowment and preference shock processes. Let an exogenous state vector process {zt } follow a first-order vector autoregression: zt+1 = A22 zt + C2 wt+1 , and suppose that b1t = Sb zt and dt = Sd zt , so that both the endowment and the preference shock are linear functions of exogenous state vector. Construct 0 0 1 0 0 (16.8.5) A22 = 0 .8 0 C2 = 1 0 , 0 .1 0 0 .5

378

Recursive Risk Sensitive Control

and set Sb = [ 30 0 1 ] , Sd = [5 1 0]. We initialize the first component of zt to be one, which replicates itself over time. Consequently, the preference shock process has mean 30 and the endowment shock process has mean 5. For the σ = 0 version of this economy, ordinary certainty equivalence applies. Since {dt } and {b1t } are asymptotically stationary, so are the endogenous processes for consumption, capital, and investment. It again follows from (16.8.4) that investment and capital both have mean zero in the stochastic steady state, and consumption has a mean equal to the mean of the endowment shock dt . Because of certainty equivalence, no precautionary savings occur in this model. When σ is less than zero, ordinary certainty equivalence no longer holds, and mean investment is positive in a stochastic steady state. Table 2.1 reports the means and standard deviations computed with respect to both the objective probability distribution, and the appropriate distorted distribution for σ = −.005. TABLE 2.1: Means and Standard Deviations for the Adjustment Cost Economy with σ = −.005 True Process

c1t it b1t dt

Pessimistic Process

mean

standard deviation

mean

standard deviation

9.83 4.83 30.00 5.00

1.37 1.22 .12 1.67

2.07 0.00 30.01 2.07

1.57 1.32 .12 1.71

The mean of consumption is now notably higher than the mean endowment and the mean of investment is positive. In effect, there is a precautionary savings motive at work here. A version of the pessimistic certainty equivalence principle derived by Jacobson (1973) and Whittle (1981) applies here. Thus, an alternative way to obtain the same solution is to endow the fictitious social planner with a pessimistic view of the world. The stochastic difference equation (16.8.3) and its solution (16.8.4) still apply except that the distorted expectation operator, call ˆ , is used in place of E. As originally suggested by Jacobson (1973), we can it E

Risk-Sensitive Investment under Uncertainty

379

imagine there being a second agent, say distinct from the social planner, that picks future values of the shocks in a pessimistic fashion. The degree of pessimism is governed by the value of σ , and the difference equations associated with these pessimistic forecasts for b1t+j and dt+j used for the time t decisions are given by: (1 − β −1 L)kt+j − dt+j + b1t+j + (100/σ)(1 − .5βL−1 )(1 − .5L)(b1t+j − 30)] = 0 (16.8.6)

−(1 − β −1 L)kt+j + dt+j − b1t+j + (1/σ)(1 − .8βL−1 )(1 − .8L)(dt+j − 5)] = 0. for j = 1, 2, . . . , where kt−1 , b1t and dt are given initial conditions. In effect, we can think of these as the Euler equations for the second agent. Notice that the capital stock enters these difference equations, and hence the pessimistic forecasts of future values of dt and b1t depend on kt−1 . Certainty equivalence works here because the optimal choice for kt can be obtained by (i) shifting (16.8.3) forward j − 1 periods for j = 1, 2, . . . and eliminating the conditional expectation operator; (ii) combining the resulting difference equation with (16.8.6); and (iii) solving the composite system of difference equations for kt , b1t+1 , dt+1 as a function of the state vector kt−1 , b1t and dt , imposing the appropriate terminal conditions. The composite solution gives the evolution equation for the pessimistic forecasts of b1t and dt , and the solution for kt gives the optimal decision rule for kt . This latter equation, when combined with the actual law of motion for b1t and dt , governs the evolution of the optimal capital stock process. While λ is .9852 for this economy, once this forecasting dependence is incorporated, the feedback of kt onto kt−1 drops slightly to .9817. A more dramatic implication is that there is now a positive constant term (1.766) in the decision rule for capital, whereas in the σ = 0 economy the constant term is zero. Table 2.1 also reports the pessimistic means and standard deviations for forcing processes and for consumption and investment. Notice that the perceived means for dt and c1t are considerably less than their true means. The mean of investment for the perceived process is again zero because of the applicability of (16.8.4) with distorted expectations. Hence the perceived long run average of the capital stock is not altered by changing expectations operators. Not surprisingly, the perceived standard deviations are larger than the true ones.

380

Recursive Risk Sensitive Control

An Approximating Economy with σ = 0 Next we show that we can mimic the quantity implications of the preceding model by setting σ = 0 and modifying some of its parameters. The parameters that we alter are the mean of b1t , the mean of dt and the productivity parameter for capital γ . In light of our previous discussion, to make average investment positive it is necessary to relax the restriction that δk + γ = β −1 by making capital more productive. We used the Kullback-Leibler (1951) information criterion as a device for picking the three parameters because of its well known link to maximum likelihood estimation (e.g., see Akaike (1973), Ljung (1978), and White (1982)). 5 The information criterion was constructed in the same manner as in Hansen and Sargent (1993) using the consumption and investment processes implied by the original model and the approximating σ = 0 economy. The parameters that make the approximate (σ = 0) model as close as possible to the original model are .1032 for the productivity parameter γ and 47.0522 and 4.6812 for the means of b1t and dt . In addition to increasing γ to compensate for setting σ = 0, we were led to increase substantially the mean of b1t and decrease slightly the mean of dt . It turns out that these three parameter adjustments are sufficient to make the quantity implications for the two models to be extremely close. This is shown in Figure 16.8.1 which displays the ratios of the spectral densities of consumption and investment, respectively, for the approximating economy to the corresponding spectral densities for the original economy. Departures from 5 Let each model be a member of our class of models, with parameters of the first model being denoted by a vector δ and those of the second model being denoted by α . For the first model, let the mean vector for the observables be ν(δ) , and the spectral density matrix be S¯y (ω, δ) . For the second model let the spectral density matrix be Sy (ω, α) , and the mean vector be µ(α) . The parameters α that make the second model as close as possible to the first are those that minimize the criterion

n

−

1 2π

−

1 2π

Z

π

−π Z π −π

log det Sy (ω, α) dω trace Sy (ω, α)−1 S¯y (ω, δ) dω

− [ν − µ(α)] Sy (0, α)−1 [ν − µ(α)]′

o

.

Risk-Sensitive Investment under Uncertainty

381

unity are small, and confined to very low frequencies. Departures of this small magnitude and frequency-location would be virtually impossible to detect, via say a likelihood ratio test, without an extremely large time series sample. The adjustments in the means of b1t and dt are sufficient to make the means for consumption and investment the same for both economies. Hence, from the standpoint of data on consumption and investment, the precautionary savings version of the model is (almost) observationally equivalent to a specification in which savings are induced by making capital more productive.

1.3

1.25

1.2

1.15

1.1

1.05

1

0.95 0

0.5

1

1.5

2

2.5

3

3.5

Figure 16.8.1: Ratio of spectral density of consumption and investment (dotted line) in the approximating economy to the spectral density of consumption and investment in the true economy.

382

Recursive Risk Sensitive Control

16.9. Equilibrium Prices in the Adjustment Cost Economies We now return to the two adjustment cost economies with their competing explanations for investment. While the quantities for the σ = 0 ‘approximating’ economy are, by design, close to those for the ‘true’ risk-sensitive economy, behaviors of asset prices and rates of return differ markedly. To illustrate this phenomenon, we first consider the implications for the two components of equilibrium wealth: the value of the endowment process from the current period forward (a Lucas tree), and the value of the existing capital stock. Figures 16.9.1.a and 16.9.1.b depict realizations of these two components to wealth for a common realization of the vector of Gaussian noises driving the composite preference and endowment shock process. The value of the endowment process is always lower and the value of the capital stock higher for the original economy than for the approximating σ = 0 economy. Hence a bigger fraction of wealth is held in the form of capital in the original (risk-adjusted) economy. This is true even though the mean of the endowment process is higher for the original economy. The mean gross returns are (1.0444, 1.0786) for capital and the endowment process, respectively, in the true economy; and (1.0500, 1.0504) in the approximating economy. 6 These phenomena trace to the fact that the capital stock is a less risky investment than the endowment and that less risky investments are more highly valued in the original (risk-adjusted) economy than in the approximating economy. Figure 16.9.2 shows probability densities for risk-free gross returns for both economies. Notice that the density for the approximating economy is centered around β −1 , and the modal value is higher than for the original economy with a risk adjustment. The density for the original (risk-adjusted) economy displays more dispersion than its counterpart for the approximating economy, reflecting in part the smaller mean of b1t . Consider next the market prices of risk for the two economies. The probability densities for these prices are reported in Figures 16.9.3.a and 16.9.3.b. The market price of risk is considerably higher in the risk-sensitive (σ = −.005) economy. Finally, Figures 16.9.4.a and 16.9.4.b report spectral densities for the logarithmic one-period returns to holding a claim to the endowment process and 6 These means were computed from simulations of length 100,000.

Equilibrium Prices in the Adjustment Cost Economies

383

to holding capital, respectively. Both returns are more variable in the original (risk-adjusted) economy than in the approximating one. Moreover, the low frequency dip in the spectral density for the return to holding the endowment is substantially more pronounced. Returning to the decomposition of equilibrium wealth described earlier, it is clear from Figures 16.9.4.a and 16.9.4.b that holding capital is much less risky than holding the endowment for both economies. It is also evident from Figure 16.9.2 that a riskless security is more valued (commands a lower equilibrium rate of return) in the original economy. This apparently underlies the fact that the capital stock is more highly valued in the original (σ = −.005) economy as depicted in Figure 16.9.1a. Moreover, as illustrated in Figures 16.9.3.a and 16.9.3.b, the market prices of risk tend to be higher in the original economy, so that the equilibrium risk adjustments are more pronounced. This helps to explain why equilibrium values for claims to the endowment process are lower for the original economy as depicted in Figure 16.9.1.a even though the mean of the endowment process is higher.

125

120

110 120 100 115

90

80

110

70 105 60

50 0

10

20

30

40

50

60

70

80

90

100

Fig. 16.9.1.a. Realizations of price of Lucas tree in true economy ( σ = −.005 ) and approximating ( σ = 0 ) economy (dashed line).

100 0

10

20

30

40

50

60

70

80

90

100

Fig. 16.9.1.b. Realizations of value of capital stock in true economy( σ = −.005 ) and approximating ( σ = 0 ) economy (dashed line).

384

Recursive Risk Sensitive Control

200 180

approximation

160 140 120 100 true

80 60 40 20 0 1.02

1.025

1.03

1.035

1.04

1.045

1.05

1.055

1.06

Figure 16.9.2: Figure 5.3 Densities of risk-free interest rate for true and approximating economies.

1200

14

12

1000

10 800 8 600 6 400 4 200

2

0 0.35

0.4

0.45

0.5

0.55

0.6

Fig. 16.9.3.a. Density for market price of risk for true adjustment cost true economy.

0 0.01

0.01

0.011

0.011

0.012

0.012

0.013

0.013

Fig. 16.9.3.b. Density for market price of risk for approximating adjustment cost economy.

Equilibrium Prices in the Adjustment Cost Economies

385

10 -3

10 -2 true

true 10 -4

approximating approximating

10 -5

10 -3

0

0.5

1

1.5

2

2.5

3

3.5

Fig. 16.9.4.a. Spectral densities of windsorized logarithmic returns to endowment streams for true and approximating adjustment costs economies.

10 -6

0

0.5

1

1.5

2

2.5

3

3.5

Fig. 16.9.4.b. Spectral densities of logarithmic returns on capital in true and approximating adjustment costs economies.

Chapter 17 Periodic Models of Seasonality

17.1. Introduction Until now, each of the matrices defining the preferences, technology and information flows has been specified to be constant over time. In this chapter, we relax this assumption, and let the matrices be strictly periodic functions of time. Our interest is to apply and extend an idea of Denise Osborn (1988) and Richard Todd (1983, 1990) to arrive at a particular model of “seasonality”. Seasonality can be characterized in terms of the spectral density of a variable. A variable is said to “have a seasonal” if its spectral density displays peaks at or in the vicinity of the frequencies commonly associated with the seasons of the year, e.g., every twelve months for monthly data, every four quarters for quarterly data. Within one of our equilibria, it is possible to think of three ways of modelling seasonality. The first two ways can be represented within the time-invariant setup of our previous chapters, while the third way requires following Todd and departing from the assumption that the matrices that define our economies are time invariant. The first model of seasonality is created by specifying the matrices [A22 , C22 , Ub , Ud ] that determine the information structure in the economy. We can exogenously inject a seasonal preference shock into the model by specifying [A22 , Ub ] in such a way that components of the shock process bt have seasonals. Similarly, we can specify [A22 , Ud ] so that components of the endowment shock process dt have seasonals. The seasonality of these exogenous processes will be transmitted to the prices and quantities determined in equilibrium. The way in which this seasonality is transmitted can be subtle, determined as it is by the restrictions across the parameters of the {bt , dt } processes and the price and quantity processes that are determined by the equilibrium. 1 1 Sargent [1976, 1987,chap XI] described some of the ways in which the cross equation restrictions of linear rational expectations models determine the kind of seasonality in endogenous variables that is induced by imposing seasonality in the variables that agents within a model are implicitly forecasting.

– 387 –

388

Periodic Models of Seasonality

The second model of seasonality is created by specifying the matrices [∆h , Θh , Λ, Π] that determine preferences and the matrices [Φc , Φi , Φc , Γ, ∆k , Θk ] that determine the technology so that they make prices and quantities display seasonality even when the preference shocks bt and the endowment shocks dt do not display any seasonality. Seasonality can come either from the technology side or from the preference side. Notice that in the first kind of model the source of seasonality is imposed exogenously, while in this second kind of model the idea is that preferences and technology are such that the equilibrium of the economy creates a “propagation mechanism” that converts nonseasonal impulses into seasonal responses in prices and quantities. This chapter is devoted to studying a third model of seasonality, by following Todd. We now specify an economy in terms of matrices that are periodic functions of time. This specification captures the idea, for example, that the technology is different in Winter than it is in Spring. You will get less corn if you plant in Minnesota in January than if you plant in May. As we shall see, this model of seasonality has properties that contrast in interesting ways to the other two models of seasonality.

17.2. A Periodic Economy The social planner now faces the problem of maximizing −.5

∞ X t=0

subject to

β t (st − bt ) · (st − bt ) + lt2

(17.2.1)

Φc,s(t) ct + Φi,s(t) it + Φg,s(t) gt = Γs(t) kt−1 + dt kt = ∆k,s(t) kt−1 + Θk,s(t) it ht = ∆h,s(t) ht−1 + Θh,s(t) ct st = Λs(t) ht−1 + Πs(t) ct zt+1 = A22,s(t) zt + C22,s(t) wt+1 bt = Ub zt dt = Ud zt

(17.2.2)

A Periodic Economy

389

In (17.2.2), s(t) is a periodic function that assigns integers to integers. In particular, s : (. . . , −1, 0, 1, . . .) → [1, 2, . . . , p] s(t + p) = s(t) ∀ t

(17.2.3)

s(t) = t for t = 1, 2, . . . , p.

A consequence of (17.2.3) is that the constraints in (17.2.2) can be represented in the form Φc,j cp·t+j + Φi,j ip·t+j + Φg,j gp·t+j = Γj kp·t+j+1 + dp·t+j kp·t+j = ∆k,j kp·t+j+1 + Θk,j ip·t+j hp·t+j = ∆h,j hp·t+j+1 + Θh,j cp·t+j sp·t+j = Λj hp·t+j+1 + P e¯j cp·t+j

(17.2.4)

zp·t+j+1 = A22,j zp·t+j + C22,j wp·t+j bp·t+j = Ub zp·t+j dp·t+j = Ud zp·t+j where t = 0, 1, 2, . . . , and j = 1, 2, . . . , p. Notice for t = 0, as j goes from 1 to p , that p · t + j goes from 1 to p ; for t = 1, as j goes from 1 to p , that p · t + j goes from p + 1 to 2p , and so on. Thus, (17.2.4) describes a setting in which the matrices that represent preferences and the technology are periodic with period p . The social planning problem can be expressed in the form of a periodic optimal linear regulator problem. The social planner chooses a sequence of functions expressing ut as functions of xt , for all t ≥ 0, to maximize −E

∞ X t=0

β t {x′t Rs(t) xt + u′t Qs(t) ut + 2u′t Ws(t) xt }

(17.2.5)

subject to the constraints xt+1 = As(t) xt + Bs(t) + Cs(t) wt+1

(17.2.6)

′ , zt ]. In (17.2.5), (17.2.6), the matrices [Rs(t) , Qs(t) , Ws(t) , where x′t = [h′t−1 , kt−1 As(t) , Bs(t) , Cs(t) ] are the same functions of the matrices [Φc,s(t) , Φi,s(t) , Φg,s(t) , Γs(t) , ∆k,s(t) , Θk,s(t) , ∆h,s(t) , Θh,s(t) , Λs(t) , Πs(t) , A22,s(t) , C22,s(t) , Ub , Ud ] that the

390

Periodic Models of Seasonality

matrices [R, Q, W, A, B, C] are of the matrices [Φc , Φi , Φc , Γ, ∆k , Θk , ∆h , Θh , Λ , Π, A22 , C22 , Ub , Ud ] in the constant coefficient case. These functions were described in chapter 3. The Bellman equations for this problem are Vt (xt ) = max{x′t Rs(t) xt + u′t Qs(t) ut + 2u′t Ws(t) xt ut

(17.2.7)

+ βEt Vt+1 (xt )} where the maximization is subject to xt+1 = As(t) xt + Bs(t) + Cs(t) wt+1 .

(13.6)

In (17.2.7), Vt (xt ) is defined as the optimal value of the problem starting from state xt at time t . For the periodic optimal linear regulator problem, the optimal value function is quadratic but time varying: Vt (xt ) = x′t Pt xt + ρt ,

(17.2.8)

where the n × n matrix Pt satisfies the matrix Riccati difference equation ′ Pt = Rs(t) + βA′s(t) Pt+1 As(t) − (βA′s(t) Pt+1 Bs(t) + Ws(t) )

′ ′ Pt+1 As(t) + Ws(t) ), Pt+1 Bs(t) )−1 (βBs(t) × (Qs(t) + βBs(t)

(17.2.9)

while the scalar ρt satisfies ′ ). ρt = βρt+1 + β trace (Pt+1 Cs(t) Cs(t)

(17.2.10)

Now think of solving Bellman’s equation by iterating backwards on (17.2.9), (17.2.10), starting from some terminal values for P and ρ. Because the matrices [Rs(t) , Qs(t) , Ws(t) , As(t) , Bs(t) ] are all functions of time when p ≥ 2, it is too much to hope that {Pt , ρt} will converge in these iterations as t → −∞ to objects that are independent of time. What is reasonable to hope for, and what will indeed obtain under the assumptions made in our setup, is that iterations on (17.2.9) and (17.2.10) will each produce p convergent subsequences. In particular, backwards iterations on (17.2.9) and (17.2.10) will converge to a sequence that oscillates periodically among p value functions associated with

A Periodic Economy

391

the p seasons of the year. Thus, after many iterations, we will eventually have Vt (xt ) = Vs(t) (xt ), where Vs(t) (xt ) = x′t Ps(t) xt + ρs(t)

(17.2.11)

We can also represent these value functions as Vj (xp·t+j ) = x′p·t+j Pj xp·t+j + ρj ,

(17.2.12)

where t = 0, 1, 2, . . . and j = [1, 2, . . . , p]. Equation (17.2.12) summarizes the outcome that there are p value functions, one for each of the p seasons of the year. The optimal decision rules can be represented as ut = −Fs(t) xt

(17.2.13)

where ′ ′ Fs(t) = −(Qs(t) + βBs(t) Ps(t+1) Bs(t) )−1 βBs(t) Ps(t+1) As(t) .

(17.2.14)

The optimal decision rules are thus periodic with period p . Substituting (17.2.13) into the law of motion (17.2.6) gives the following “closed loop” representation of the solution of the social planning problem:

or

xt+1 = As(t) − Bs(t) Fs(t) xt + Cs(t) wt+1

(17.2.15)

xt+1 = Aos(t) xt + Cs(t) wt+1

(17.2.16)

where Aos(t) = As(t) − Bs(t) Fs(t) . We can also represent (17.2.16) in the form xp·t+j+1 = Aoj xp·t+j + Cj wp·t+j+1

(17.2.17)

for t = 0, 1, 2, . . . and j = [1, 2, . . . , p]. Thus the laws of motion are periodic with a periodicity p that is inherited from that of the matrices specifying preferences, technology, and information flows. The matrices [Aoj , Pj ] for j ∈ [1, 2, . . . , p] can be used to construct the quantities and prices associated with the equilibrium of our model. Formulas for the matrices determining our equilibrium, namely the M and S matrices,

392

Periodic Models of Seasonality

are given by the very same formulas described in chapters 3 and 5, with the proviso that in the periodic case s(t) or j subscripts appear on all objects in those formulas. Thus, we have that the quantities determined in our equilibrium are given by ht = Sh,s(t) xt dt = Sd,s(t) xt kt = Sk,s(t) xt

ct = Sc,s(t) xt

kt−1 = Sk1,s(t) xt it = Si,s(t) xt

gt = Sg,s(t) xt

(17.2.18)

st = Ss,s(t) xt

bt = Sb,s(t) xt where

Sh,s(t) Sk,s(t)

Ao11,s(t) = Ao12,s(t)

Sk1,s(t) = [0 I 0]

Si,s(t) = −Fs(t)

Sd,s(t) = [0 0 Ud ] Sb,s(t) = [0 0 Ub ]

(17.2.19)

Sc,s(t) = Uc,s(t) [Φc,s(t) Φg,s(t) ]−1 [−Φi,s(t) Si,s(t) + Γs(t) Sk1,s(t) + Sd,s(t) ] Sg,s(t) = Ug,s(t) [Φc,s(t) Φg,s(t) ]−1 [−Φi,s(t) Si,s(t) + Γs(t) Sk1,s(t) + Sd,s(t) ] Ss,s(t) = Λs(t) [I 0 0] + Πs(t) Sc,s(t) The Lagrange multipliers associated with the social planning problem are determined by the following counterparts of the formulas that we described in chapters 3 and 5: Mk,s(t) = 2β[0 I 0] Ps(t) Aos(t) Mh,s(t) = 2β[I 0 0] Ps(t) Aos(t) Ms,s(t) = Sb,s(t) − Ss,s(t) ′ −1 ′ ′ Φc,s(t) Θh,s(t) Mh,s(t) + Π′s(t) Ms,s(t) Md,s(t) = Φ′g,s(t) −Sg,s(t) ′ ′ Mc,s(t) = Θh,s(t) Mh,s(t) + Πs(t) Ms,s(t) Mi,s(t) = Θk,s(t) Mk,s(t)

(17.2.20)

Asset Pricing

393

Formulas for the equilibrium price system can be stated in terms of the objects defined in (17.2.20):

ptt′ = Mc,s(t′ ) xt′ /[¯ ej Mc,s(t) xt ] wtt′ = |Sg,s(t′ ) xt′ |/[¯ ej Mc,s(t) xt ]

rtt′ = Γ′s(t) Md,s(t′ ) xt′ /[¯ ej Mc,s(t) xt ] qtt′ = Mi,s(t′ ) xt′ /[¯ ej Mc,s(t) xt ]

(17.2.21)

αtt′ = Md,s(t′ ) xt′ /[¯ ej Mc,s(t) xt ] vt = [Γs(t′ ) Md,s(t′ ) + ∆′k,s(t′ ) ]xt′ /[¯ ej Mc,s(t) xt ] These formulas give the time t price system for pricing goods to be delivered at all t′ ≥ t .

17.3. Asset Pricing With the above formulas in hand, we can derive formulas for pricing assets. These formulas generalize those described in chapter 5 to the case in which the economy is strictly periodic. We begin by pricing an asset that entitles its owner to a stream of returns in the form of a vector of consumption goods described by yt = Ua,s(t) xt , where Ua,s(t) is a periodic sequence of matrices. We let at denote the price of this asset at time t . By the same reasoning applied in chapter 5, at satisfies at = Et

∞ X

ej Mc,s(t) xt ], β h x′t+h Za,s(t+h) xt+h /[¯

(17.3.1)

h=0

′ where Zaj = Ua,j Mc,j . We shall show that (17.2.5) can be represented as

at = [x′t µa,s(t) xt + σa,s(t) ]/[¯ ej Mc,s(t) xt ],

where µa,s(t) and σa,s(t) satisfy

(17.3.2)

394

Periodic Models of Seasonality

o 2 o′ o′ o o µa,1 = Za,1 + βAo′ 1 Za,2 A1 + β A1 A2 Za,3 A2 A1 + · · ·

o′ o′ o′ o o o o + β p−1 Ao′ 1 A2 · · · Ap−2 Ap−1 Za,p Ap−1 Ap−2 · · · A2 A1 o′ o′ o′ o o o o + β p Ao′ 1 A2 · · · Ap−1 Ap µa,1 Ap Ap−1 · · · A2 A1

o µa,p = Za,p + βAo′ p µa,1 Ap

(17.3.3)

o′ µa,p−1 = Za,p−1 + βAp−1 µa,p Aop−1 .. . o µa,2 = Za,2 + βAo′ 2 µa,3 Ap−2

and

σa,1 = β trace(µa,2 C1 C1′ ) + βσa,2 σa,2 = β trace(µa,3 C2 C2′ ) + βσa,3 .. .

(17.3.4)

σa,p = β trace(µa,1 Cp Cp′ ) + βσa,1 The matrix µa,1 can be computed from the first equation of (17.3.3) by using a doubling algorithm that is described in chapter 8. Then the remaining equations of (17.3.3) can be used to compute the remaining µa,j ’s. Given the µa,j ’s, (17.3.4) is a system of p equations that can be solved for the p σa,j ’s. To verify (17.3.3), (17.3.4), we can proceed as follows. Let the numerator of (17.3.1), (17.3.2) be denoted a ˜ t = Et =

∞ X

β h x′t+h Za,s(t+h) xt+h

h=0 ′ xt µa,s(t) xt

(17.3.5)

+ σa,s(t)

Recall the equilibrium transition laws (17.2.16): xt+1 = Aos(t) xt + Cs(t) wt+1 . Evidently, (17.3.5) and (17.2.16) imply that a ˜t = x′t Za,s(t) xt + β Et a ˜t+1 or

x′t µa,s(t) xt + σa,s(t) = x′t Za,s(t) xt +βEt (Aos(t) xt + Cs(t) wt+1 )′ µa,s(t+1) (Aos(t) xt + Cs(t) wt+1 ) + βσa,s(t+1)

(13.19)

Prediction Theory

395

The above equation implies that o µa,s(t) = Za,s(t) + βAo′ s(t) µa,s(t+1) As(t)

(17.3.6)

′ σa,s(t) = β σa,s(t+1) + β trace (µa,s(t+1) Cs(t) Cs(t) )

(17.3.7)

Equation (17.3.4) is equivalent with (17.3.7). The first equation of (17.3.3) is the result of recursions on (17.3.6) starting from s(t) = 1, while the remaining equations of (17.3.3) are simply (17.3.6) for s(t) = 2, 3, . . . , p. This completes the verification of (17.3.3), (17.3.4). We shall give a formula for the term structure of interest rates after we have described the prediction theory associated with (17.2.16).

17.4. Prediction Theory For a model with period p ≥ 2, there are two natural alternative ways of specifying the information sets upon which means, covariances, and linear least squares predictions are conditioned. First, we can calculate moments and forecasts by conditioning on the season. This amounts to computing different moments and different forecasting formulas for each of the p seasons. In the appendix to this chapter, we formally describe a sigma algebra, which we denote I p , that contains the information that corresponds to conditioning on the season. Second, we can calculate moments and forecasts by disregarding information about the season, which amounts to averaging data across seasons in a particular way. In the appendix, we formally describe a sigma algebra, denoted I , which corresponds to not conditioning on the season. In this section, we describe parts of the prediction theory for our periodic models that correspond to conditioning on the season. In this section,the notation Et ( · ) denotes a mathematical expectation conditioned on xt , under the assumption that we are also conditioning on the information in I p . We are assuming that the fictitious social planner uses this information to compute all relevant prices. Recursions on (17.2.16) can be used to deduce the linear least squares predictions of the state vector xt . There are p different sets of formulas for the j -step ahead predictions of xt+k conditioned on xt , one for each season of the

396

Periodic Models of Seasonality

year. Recursions on (17.2.16) lead directly to xt+k = Aos(t+k−1) Aos(t+k−2) · · · Aos(t) xt

+ Aos(t+k−1) Aos(t+k−2) · · · Aos(t+1) Cs(t) wt+1 + · · ·

(17.4.1)

+ Aos(t+k−1) Cs(t+k−2) wt+k−1 + Cs(t+k−1) wt+k Equation (17.4.1) implies Et xt+k = Aos(t+k−1) Aos(t+k−2) · · · Aos(t) xt

(17.4.2)

and E(xt+k − Et xt+k )(xt+k − Et xt+k )′ ≡ Σk,s(t)

′ = Aos(t+k−1) Aos(t+k−2) · · · Aos(t+1) Cs(t) Cs(t) Ao′ s(t+1) · · · o′ Ao′ s(t+k−2) As(t+k−1)

+

(17.4.3)

′ · · · + Aos(t+k−1) Cs(t+k−2) Cs(t+k−2) Ao′ s(t+k−1) ′ + Cs(t+k−1) Cs(t+k−1)

Recursive versions of (17.4.2) and (17.4.3) are available. Equation (17.4.2) implies Et xt+k = Aos(t+k−1) Et xt+k−1 . Equation (17.4.3) implies ′ Σk,s(t) = Aos(t+k−1) Σk−1,s(t) Ao′ s(t+k−1) + Cs(t+k−1) Cs(t+k−1) .

The prediction formulas (17.4.2), (17.4.3) are evidently predicated on the assumption that we know the matrices [Aoj , Cj ] for j = [1, . . . , p]. They also assume that xt is in the information set of the forecaster. Later in this chapter, we shall briefly describe how the Kalman filter can be used to compute the linear least squares forecast of yt , conditioned only on the history of observed y ′ s , and also on I p . We shall also describe a different theory of prediction, which assumes that we do not know the values of [Aoj , Cj ], and that we cannot condition on the season, so that all that we possess is a time invariant representation for the {xt , yt } process.

Conditional Covariograms

397

17.5. The Term Structure of Interest Rates In light of formula (17.4.2), the same logic that led to formula (5.65) for the reciprocal of the risk-free interest rate on j -period loans, Rjt , now leads to the following formula: Rjt = β j e¯1 Mc,s(t+j) Aos(t+j−1) Aos(t+j−2) · · · Aos(t) xt /[¯ ej Mc,s(t) xt ]

(17.5.1)

This formula gives the price at time t of a sure claim on the first consumption good j periods ahead.

17.6. Conditional Covariograms In this section, we present formulas for the covariance function of x and y , conditioned on season, i.e., conditioned on I p . The conditional covariogram of {xt } can be expressed in terms of the conditional contemporaneous covariance function cx,t (0) = Ext x′t |I p via the formulas cx,t (−k) ≡ Ext x′t+k |I p

o′ o′ o′ = Ext x′t |I p Ao′ s(t) As(t+1) · · · As(t+k−2) As(t+k−1) , k ≥ 1

or

o′ cx,t (−k) = cx,t (0)Ao′ s(t) As(t+1) · · ·

o′ Ao′ s(t+k−2) As(t+k−1) , k ≥ 1.

(17.6.1)

To solve for the matrices cx,t (0), we can solve the equations ′ Ext+1 x′t+1 |I p = Aos(t) Ext x′t |I p Ao′ s(t) + Cs(t) Cs(t)

or ′ cx,t+1 (0) = Aos(t) cx,t (0)Ao′ s(t) + Cs(t) Cs(t) .

(17.6.2)

By solving the system formed by (17.6.2) for t = 1, 2, . . . , p, we can determine the p contemporaneous covariance matrices cx,1 (0), cx,2 (0), . . . , cx,p (0). Here is a fast way of solving this system. Iterating on (17.6.2) p times yields cx,t+p (0) o′ o′ = Aos(t+p−1) Aos(t+p−2) · · · Aos(t) cx,t (0) Ao′ s(t) · · · As(t+p−2) As(t+p−1)

′ o′ + Aos(t+p+1) Aos(t+p−2) · · · Aos(t+1) Cs(t) Cs(t) Ao′ s(t+1) · · · As(t+p−1)

′ ′ · · · + Aos(t+p−1) Cs(t+p−2) Cs(t+p−2) Ao′ s(t+p−1) + Cs(t+p−1) Cs(t+p−1) .

(17.6.3)

398

Periodic Models of Seasonality

We compute cx,t (0) by setting cx,t+p (0) equal to cx,t (0) in (17.6.3). Equation (17.6.3) is a discrete Lyapunov equation that can be solved by a doubling algorithm that is described in chapter 8. Once (17.6.3) is solved for t = 1 to compute cx,1 (0), (17.6.2) can be used to compute cx,t (0) for t = 2, . . . , p. There is one covariance matrix cx,t (0) for each of the p seasons of the year. ′ Given cx,t (−k) for k ≥ 0, we can compute cy,t+k (−k) = Eyt yt+k |I p by using (17.6.1). We obtain ′ Eyt yt+k | I p = Gs(t) cx,k (−k)G′s(t+k) , k ≥ 0.

(17.6.4)

Although we are starting calendar time at t = 0, cx,t (k) and cy,t (k) are both defined for positive k so long as t ≥ k . For any such t , cx,t (k) = cx,t−k (−k)′ and cy,t (k) = cy,t−k (−k)′ , implying that {cy,t (k)} and {cy,t (k)} are both periodic starting from t = k . For notational convenience, we extend this construction for 0 ≤ t ≤ k by defining cx,t (k) = cx,t+ℓp (k) and cy,t (k) = cy,t+ℓp (k) for any ℓ such that t + ℓp ≥ k . This guarantees that the conditional covariograms are periodic for all values of k .

17.7. The Stacked and Skip-Sampled System The equilibrium has the system of periodic transition laws described in (17.2.16) or (17.2.17). The equilibrium stochastic process for xt is time-varying, albeit in a highly structured way. We have seen that conditional on knowledge of the season, there are p covariograms, and p sets of formulas for linear least squares predictions that apply in the p seasons of the year. Using these formulas requires knowledge of the set of matrices [Aoj , Cj ] for j = [1, . . . , p] that characterize the transition laws (17.2.16). In this section, we describe a time invariant representation that also characterizes the system. We shall use this representation for several purposes. We shall use it to deduce two kinds of impulse response functions or moving average representations that can be defined for periodic models. 2 We shall also use it to compute a population version of a time-invariant vector autoregression for xt . 2 Each of these impulse response functions conditions on knowledge of the season. Later we shall describe yet another moving average representation that does not condition on season.

The Stacked and Skip-Sampled System

399

The p distinct covariograms as described by equations (17.6.1) and (17.6.4) are conditional covariograms, meaning that they are computed by conditioning on the season of the year. Sample counterparts of these conditional covariograms are computed by creating p distinct averages, averaging each over observations p periods apart. Sample covariograms can also be computed ‘unconditionally’, i.e., in a way that ignores the seasonal structure of the transition laws. This amounts to computing sample moments in the standard way, simply by averagPT ′ ing over adjacent observations, namely, as T −1 t=1 yt yt−j . For a periodic model, such averages will converge as T → ∞ , and they will converge to well defined functions of the parameters of the model. In particular, as T → ∞ , PT ′ T −1 t=1 yt yt−k would converge to an average of the p covariograms, namely, −1 p [cy,1 (k) + cy,2 (k) + . . . + cy,p (k)]. The convergence of these sample autocovariances assures the existence of a time invariant vector autoregressive representation for yt . We begin by defining for t = 0, 1, . . . the vector Xt′ = [x′p·t−p+1 , x′p·t−p+2 , . . . , x′p·t ]′ .

(17.7.1)

′ Xt+1 = [x′p·t+1 , x′p·t+2 , . . . , x′p·t+p ].

(17.7.2)

Evidently, we have

To verify this, substitute (t + 1) for t everywhere that t appears on the right side of (17.7.1). We also define ′ ′ ′ ′ Wt+1 = [wp·t+1 , wp·t+2 , . . . , wp·t+p ].

(17.7.3)

It follows from (17.2.17) that DXt+1 = F Xt + GWt+1 , where

I −Ao1 D= 0 . .. 0

0 I −Ao2 .. . 0 F =

0 0 I .. .

··· ··· ··· .. .

0

· · · −Aop−1 Aop 0

0 0

0 0 0 .. .

(17.7.4) 0 0 0 .. . I

(17.7.5)

(17.7.6)

400

Periodic Models of Seasonality

Cp 0 G = .. .

0 C1 .. .

0 ··· 0 ··· .. . . . .

0 0 .. .

0

0 ···

Cp−1

0

Solving (17.7.4) for Xt+1 gives

ˆ t + CW ˆ t+1 Xt+1 = AX

(17.7.7)

(17.7.8)

′ ′ where Aˆ = D−1 F and Cˆ = D−1 G . We also define the vector Yt′ = [yp·t−p+1 , yp·t−p+2 ,..., ′ yp·t ]. Then we have that

Yt = HXt

where H =

G1 0 .. .

0 G2 .. .

··· ··· .. .

0 0 .. .

(17.7.9)

. Thus we have that {Yt } is governed by the

0 0 · · · Gp time invariant state space system

ˆ t + CW ˆ t+1 Xt+1 = AX

(17.7.10)

Yt = HXt

(17.7.11)

Notice that while {xt , yt } is governed by a time varying linear state space system, the stacked and skip sampled process {Xt , Yt } is governed by a time invariant system. 3 From representation (17.7.10) – (17.7.9), we can use standard formulas to deduce the moving average representation of Yt in terms of Wt Yt =

∞ X

C¯j Wt−j .

(17.7.12)

j=0

The moving average representation (17.7.12) implies the following representation for the components of Yt in terms of the components of Wt : ypt−p+k =

p ∞ X X

C¯j (k, h)wp(t−j)−p+h , k = 1, . . . , p,

(17.7.13)

j=0 h=1

3

In terms of the language introduced in the appendix, because S is of period p, S p is of period one.

The Stacked and Skip-Sampled System

401

where C¯j (k, h) denotes the (k, h)th (m × m) block of C¯j , where m is the dimension of yt . According to representation (17.7.13), there are two distinct concepts of a moving average representation, and p embodiments of each of these concepts. The first concept is a representation of ypt−p+k in terms of current and lagged wt ’s. The response of ypt−p+k to lagged w ’s is evidently given by the sequence 4 ¯ ¯ ¯ ¯ ¯ {dk,v }∞ v=0 = {C0 (k, k), C0 (k, k − 1), . . . C0 (k, 1)C1 (k, p), C1 (k, p − 1), . . . , C¯1 (k, 1), C¯2 (k, p), C¯2 (k, p − 1), . . . , C¯1 (k, 1), . . .}. (17.7.14) In particular, we have from (17.7.13) that ypt−p+k =

∞ X

dk,v wpt−p+k−v .

(17.7.15)

v=0

Notice that there is a different moving average of type (17.7.15) for each season k = 1, . . . , p. The second concept of a moving average is the response of the {yt } process to an innovation wpt−p+k in a particular season k . The response of {yt } to wpt−p+k is evidently given by the sequence ¯ ¯ ¯ {gk,v }∞ v=0 = {C0 (k, k), C0 (k + 1, k), . . . , C0 (p, k), C¯1 (1, k), C¯1 (2, k), . . . , C¯1 (p, k), . . .}.

(17.7.16)

In the special case in which the true periodicity is one, it is straightforward to verify that for any p > 1, the impulse functions constructed from the stacked system (17.7.10) – (17.7.9) satisfy the restrictions: C¯j (1, 1) = C¯j (2, 2) = . . . = C¯j (p − 1, p − 1) = C¯j (p, p) C¯j (2, 1) = C¯j (3, 2) = . . . = C¯j (p, p − 1) = C¯j+1 (1, p)

C¯j (3, 1) = C¯j (4, 2) = . . . = C¯j+1 (1, p − 1) = C¯j+1 (2, p) .. . ¯ Cj (p − 1, 1) = C¯j (p, 2) = . . . = C¯j+1 (p − 2, p − 1) = C¯j (p − 2, p) C¯j (p, 1) = C¯j+1 (1, 2) = . . . = C¯j+1 (p − 2, p − 1) = C¯j+1 (p − 1, p) 4 Notice that by construction C ¯0 (k, j) = 0 for k < j .

(4.14)

402

Periodic Models of Seasonality

Under these restrictions, it follows that gk,v = gj,v for all j, k, for all v dk,v = dj,v for all j, k, for all v dk,v = gk,v for all k,

for all v

Thus, in the case in which the hidden periodicity is truly one, all of the impulse response functions defined in (17.7.15) and (17.7.16) are equal, and are equal to each other. However, when the hidden periodicity is truly some p > 1, there are p distinct impulse response functions {dk,v } of ypt−p+k to lagged w ’s, and p distinct impulse responses {gk,v } of {yt } to wpt−pk , for k = 1, . . . p. In general the {dk,v } are different from one another and from the {gk,v }’s for k = 1, . . . , p. These differences provide a useful way of describing how the operating characteristics of a periodic model with p ≥ 2 differ from a period one model. 5 Later in this chapter, we compute the impulse response functions {dk,v } and {gk,v } for investment for a period 4 version of Hall’s model described above. These impulse response functions are depicted in figures 17.10.1.a and 17.10.1.b. The impulse responses are with respect to the one shock in the model, which is a white noise endowment process. Figure 17.10.2 depicts the impulse responses {dk,v } for k = 1, . . . , 4. Notice that they are smooth, but that they vary across quarters. Figure 17.10.3 shows the impulse response {gk,v } for k = 1, . . . , 4. They vary across quarter k , and have shapes that are jagged, in contrast to the smooth {dk,v }’s. Notice how the amplitude of the oscillations in {dk,v } grows as v increases from v = 0 to v about 30.

5 A MATLAB program simpulse performs these calculations.

Covariances of the Stacked, Skip Sampled Process

403

17.8. Covariances of the Stacked, Skip Sampled Process The stacked, skip-sampled process {Yr } is constructed to have periodicity one. We can compute for k ≥ 1, ′ Cr (−k) ≡ E(Yr Yr+k | Ip) =

cy,pr (−pk) cy,pr+1 (−pk + 1) .. . cy,pr+p−1 (−pk + p − 1)

cy,pr (−pk − 1) cy,pr+1 (−pk) .. . cy,pr+p−1 (−pk + p − 2)

··· ··· .. . ···

cy,pr (−pk − p + 1) cy,pr+1 (−pk − p + 2) . .. . cy,pr+p−1(−pk) (17.8.1)

An implication of the period 1 nature of {Yr } is that Cr (−k) is independent of r . This follows immediately from (17.8.1). In particular, we have for k ≥ 1, C(k) ≡ C0 (−k)

cy,0 (−pk) cy,1 (−pk + 1) = .. . cy,p−1 (−pk + p − 1)

cy,0 (−pk − 1) cy,1 (−pk) .. . cy,p−1 (−pk + p − 2)

... ... ..

.

...

cy,0 (−pk − p + 1) cy,1 (−pk − p + 2) (17.8.2) . .. . cy,p−1 (−pk)

The k = 0 term must be treated separately. It is given by Cr (0) ≡ E(Yr Yr′ | I p ) cy,pr (0) cy,pr (−1)′ = .. .

cy,pr (−p + 1)′

cy,pr (−1) cy,pr+1 (0) .. .

··· ··· .. .

cy,pr+1 (−p + 2)′

···

cy,pr (−p + 1) cy,pr+1 (−p + 2) (17.8.3) , .. . cy,pr+p−1 (0)

which can also be shown to be independent of r . The covariance generating function of the {Yr } process is given by S(z) ≡ C(0) +

∞ X

[C(−k)z −k + C(−k)′ z k ].

(17.8.4)

k=1

It is useful to calculate the covariance generating function S(z) of {Yr } by substituting (17.8.2) – (17.8.3) into (17.8.4). We obtain f s11 (z) sf12 (z) sf13 (z) · · · sf1p (z) sf (z) sf (z) sf (z) · · · sf (z) 21 22 23 2p f s31 (z) sf32 (z) sf33 (z) · · · sf3p (z) (17.8.5) S(z) = .. .. .. .. .. . . . . . f f f sp1 (z) sp2 (z) ··· · · · spp (z)

404

Periodic Models of Seasonality

where

sfj,j+ℓ (z) = cj−1 (−ℓ) +

∞ X

k=1

[cy,j−1 (−pk − ℓ)z −k

(17.8.6)

+ cy,j+ℓ−1 (−pk − ℓ)′ z k ], and where the lower triangular terms of S(z) are obtained from the upper by setting S(z) = S(z −1 )′ for z = e−iω . The hypothesis that {yt } is of period one places restrictions on S(z). Period one of {yt } implies that cy,j (k) = cy,1 (k) for all j . By using this equality in (17.8.6) it can be shown that sf11 (z) = sf22 (z) = · · · = sfpp (z)

sf12 (z) = sf23 (z) = · · · sfp−1,p (z) = z −1 sfp,1 (z)

sf13 (z) = sf24 (z) = · · · = sfp−2,p (z) = z −1 sfp−1,1 (z) = z −1 sfp,2 (z)

(17.8.7)

.. .

zsf1,p (z) = sf2,1 (z) = sf3,2 (z) · · · = sfp,p−1 (z) The first line of equalities in (17.8.7) asserts that the block of matrices along the diagonal of S(z) are equal to each other, and to a folded spectrum of the original unsampled {yt } process.

The Tiao-Grupe Formula

405

17.9. The Tiao-Grupe Formula Define ′ ry,t (−k) = E(yt yt+k | I).

It follows that ry,t (−k) = ry,1 (−k) ≡ ry (−k) for all t . Furthermore, by the law of iterated expectations ry (−k) = ry,t (−k) = E cy,t (−k) | I .

(17.9.1)

It follows from (17.9.1) that ry (−k) can be computed either by computing covariances without skip sampling, or by averaging across covariances that have been computed by skip sampling. That is, ry (−k) = p−1

p X

cy,j (−k),

j=1

and by a law of large numbers N 1 X ′ yt yt+k . N →∞ N t=1

ry (−k) = lim

It is useful to derive Tiao and Grupe’s (1980) formula for the covariance generating function of {yt }, not conditioned on season, as a function of the covariance generating function conditioned on season. Tiao and Grupe’s formula expresses the generating function for the covariances not conditioned on season in terms of the (conditional on season) covariance generating function of the stacked and skip sampled process Yt . 6 We define the generating function for the covariances not conditioned on season to be: sy (z) =

∞ X

r(k)z k

k=−∞

or

sy (z) = p−1

∞ X

k=−∞

zk

p X

(17.9.2) cy,j (k).

j=1

6 Gladysev (1960) states a formula restricting the Cramer representations for Y and y t t that has the same content as the Tiao-Grupe formula.

406

Periodic Models of Seasonality

To compute sy (z), define the operator Q(z) = [I zI . . . z p−1 I],

(17.9.3)

where each of the p identity matrices in (17.9.3) is (n×n). Note that for k ≥ 1, Q(z)C(k)Q(z −1 )′ = [cy,0 (−pk) + cy,0 (−pk + 1)z −1 + · · · + cy,0 (−pk − p + 1)z −p+1

+ cy,1 (−pk + 1)z + cy,1 (−pk) + · · · + cy,1 (−pk − p + 2)z −p+2 .. .

+ cy,p−1 (−pk + p − 1)z p−1 + cy,p−1 (−pk + p − 2)z p−2 + . . . + cy,p−1 (−pk)] (17.9.4a) Notice also that for k = 0, we have

Q(z)C(0)Q(z −1 )′ = [I z I . . . z p−1 I]C(0)

I zI −1 .. . z −p+1 I

= cy,0 (0) + cy,1 (0) + . . . + cy,p−1 (0)

+ z[cy,0 (−1)′ + cy,1 (−1)′ + . . . + cy,p−2 (−1)′ ]

(17.9.4b)

+ z −1 [cy,0 (−1) + cy,1 (−1) + . . . + cy,p−2 (−1)]+ .. . + z p−1 cy,0 (−p + 1)′ + z −p+1 cy,0 (−p + 1). Applying (13.67) to (17.9.2) gives sy (z) = p−1

∞ X

z ph Q(z)C(h)Q(z −1 )′

h=−∞

sy (z) = p−1 Q(z)[

∞ X

z ph C(h)]Q(z −1 )′

(17.9.5)

h=−∞

sy (z) = p−1 Q(z)S(z p )Q(z −1 )′ , where S(z) is the generating function for the {Yr } process, which is defined in (17.7.9) and (17.8.4). Equation (17.9.5) is the Tiao–Grupe formula.

The Tiao-Grupe Formula

407

Equation (17.9.5) shows how the generating function of the {yt } process can be obtained by transforming the generating function of the stacked, skip sampled process {Yr }. Equation (17.9.5) is helpful in displaying the types of fluctuations that will occur in a periodic process {yt }. Suppose that we were to take a realization {yt }Tt=1 of the {yt } process, compute the sample covariances as T 1 X rˆ(k) = yt yt−k (17.9.6) T t=k+1

and the sample spectrum as sˆ e−iωh =

T X

w(k)ˆ r(k)e−iωh j , ωh =

k=−T +1

2πh , h = 1, . . . , T T

(17.9.7)

where w(k) is one of the popular windows. Notice that in computing (17.9.6) and (17.9.7) we are ignoring the hidden periodicity. In large samples, rˆ(k) given by (17.9.6) will converge to r(k) defined in (17.9.1), and sˆ e−iωh will converge to sy e−iωh .

17.9.1. A state space realization of the Tiao-Grupe formulation We now return to representation (17.7.10) – (17.7.9). We will use this representation in conjunction with formula (17.9.3) to get a representation for the generating function sy (z) in terms of the parameters of our economic model. Then we shall describe how to use state space methods to factor this covariance generating function, thereby obtaining a Wold representation for yt . If the eigenvalues of Aˆ are bounded in modulus by unity, 7 then {Xt , Yt } will be asymptotically covariance stationary, with covariance generating matrices SX (z) and SY (z) given by SX (z) =

∞ X

RX (k)z k

k=−∞

or

ˆ −1 Cˆ Cˆ ′ (I − Az ˆ −1 )−1′ SX (z) = (I − Az) 7 This is the condition alluded to in section 1.

(17.9.8)

408

Periodic Models of Seasonality

and SY (z) =

∞ X

RY (k)z k

k=−∞

SY (z) = HSX (z)H

(17.9.9)

′

′ ′ where RX (k) = EXt Xt−k , RY (k) = EYt Yt−k . By substituting SX (z) or SY (z) for S(z) in formula (17.9.2), we can compute the covariance generating function for the process {yt } by averaging across covariograms for different periods. For the yt process under study here, we have 8

sy (z) = Q(z)SY (z p )Q(z −1 )′ (17.9.10)

or ˆ p )−1 Cˆ Cˆ ′ (I − Az ˆ −p )−1′ H ′ Q(z −1 )′ . sy (z) = Q(z)H(I − Az

We now show how to use (17.9.10) to deduce a state-space representation for {yt }. The first step involves recognizing that (17.9.10) is realized by the system ˆ t + CV ˆ t+p Zt+p = AZ

(17.9.11)

yt = Q(L)HZt

where L is the lag operator and {Vt } is a vector white noise with identity contemporaneous covariance matrix. It is convenient to stack (17.9.11) into the first order system Z 0 0 t+p Zt+p−1 I 0 Zt+p−2 0 I = .. .. .. . . . 0 0 Zt+1 Z

... ... ... .. . ...

t+p−1

yt+p−1

0 0 0 .. . I

Aˆ 0 0 .. . 0

ˆ C Zt+p−2 0 Zt+p−3 0 + Vt+p .. .. . . 0 Zt Z

t+p−1

(17.9.12)

Zt+p−2 ˜ = H Zt+p−3 . .. Zt

8 A MATLAB program spectrs implements ( 17.9.10 ) for the equilibrium of one of our periodic general equilibrium models.

The Tiao-Grupe Formula

409

where . . . ˜ = p−.5 [G1 0 0 . . . 0 .. 0 G2 0 . . . 0 .. . . . .. 0 0 0 . . . Gp ]. H ˜ takes this form, recall that the operator Q(L) is defined as To see why H p.5 Q(L) = [I IL . . . ILp−1 ] = [I 0 . . . 0] + [0 I . . . 0]L + . . . + [0 0 . . . I]Lp−1 . ˜ take the form This structure for Q(L) and the form of H dictates that H that it does and that the state in (17.9.12) takes the form that it does in order ˜ to map (17.9.11) into a first-order system. Notice that the structure of H implies that yt is formed by averaging over linear combinations of the first n rows of Zt+p−r , the second n rows of Zt+p−2 , . . . , and the pth n rows of Zt . Furthermore, notice that according to (17.9.12), the np×np process Zt consists of p completely uncoupled systems, each of which depends on its own past in exactly the same was as do the others. That is, (17.9.12) has the property that Zt is independent of Zt−1 , Zt−2 , . . . , Zt−p+1 for all t ; and that Zt is correlated with Zt−p in exactly the same way for all t . Thus, the “state equations” of (17.9.12) in effect describe p “parallel realizations” of the process Zt+p defined in (17.9.11). Running p parallel processes is a way of realizing in the time domain the randomization over laws of motion that is involved in adopting a description of {yt } in terms of a stationary probability distribution. As noted above, yt is formed by averaging across these p uncoupled realizations. We can use Kalman filtering methods to derive a Wold representation for {yt }. Modify and represent system (17.9.12) as Z˜t+1 = A˜Z˜t + C˜ V˜t+1 ˜ t + ǫ˜t yt = HZ

(17.9.13)

′ ′ where Z˜t′ = [Z˜t+p−1 , Z˜t+p−2 , . . . , Z˜t′ ] and

0 I A˜ = 0 .. .

0

0 0 I .. . 0

... ... ... .. . .. .

0 0 0 .. . I

Aˆ 0 0 , .. . 0

˜ C 0 C˜ = 0 , .. . 0

410

Periodic Models of Seasonality

where {V˜t } is a white noise, and where ǫt is a (potentially very small) measurement error which is a white noise process that is orthogonal to {V˜t } and satisfies E˜ ǫt ǫ˜′t = R . To obtain the Wold representation for yt which achieves the factorization of the spectral density matrix (17.9.10) for yt , we use the Kalman filter to obtain an innovations representation associated with system (17.9.13). The innovations representation is ˜ at Zˆt+1 = A˜Zˆt + K˜ ˜ Yˆt + a yt = H ˜t ,

(17.9.14)

where a ˜t = yt − E[yt | yt−1 , yt−2 , . . .], Zˆt = E[Zt | yt−1 , yt−2 , . . .], ˜ are the state covariance matrix Σ = E(Zt − Zt )(Zt − Zˆkt )′ and where Σ and K and the Kalman gain computed via the Kalman filter for system (17.9.13). The ˜ H ˜ ′ + R.9 covariance matrix of the innovations is given by E˜ at a ˜′t = HΣ

17.10. Some Calculations with a Periodic Hall Model We use a periodic version of Hall’s model as an example. The model is identical to the version of Hall’s model described in chapters 3 and 5, except that the productivity parameter γ now varies periodically. The social planner chooses contingency plans {ct , kt , it }∞ t=0 to maximize the utility functional ∞

1 X t − ( )E β [(ct − bt )2 + ℓ2t ] | J0 2 t=0 0 0

ℓ2t ,

s(t + p) = s(t),

∀t,

s(t) = t for t = 1, . . . , p

9 These calculations are performed by the MATLAB program .

Some Calculations with a Periodic Hall Model

411

and subject to the (exogenous) laws of motion

bt = 30 dt = .8dt−1 + w1t + 5 ∗ (1 − .8) We set p = 4, and γ1 = .13, γ2 = .1, γ2 = .1, γ4 = .08. We set φ1 = .3, δk = .95, β = 1/1.05. The only source of disturbance in the model is the endowment shock, which is a first order autoregression. The variance of the innovation w1t is unity. The following MATLAB programs can be used to analyze the model. solves.m: simuls.m: steadsts.m: assets.m:

assetss.m: seasla.m:

simpulse.m: spectrs.m: factors.m:

computes the equilibrium of a periodic model, simulates a periodic equilibrium; computes the means of variables from a periodic equilibrium, conditional on the season; computes the objects in the formulas for equilibrium assets prices and the term structure of interest rates for a periodic model; simulates the asset prices in a periodic equilibrium; computes the time invariant state-space representation for the stacked, skip sampled version of a periodic model; computes the two different concepts of perioddependent impulse response functions; computes the spectral density of a periodic model, using the Tiao-Grupe formula; factors a univariate spectral density computed via the Tiao-Grupe formula in order to obtain a univariate Wold representation for a single variable of a periodic equilibrium model.

We computed the equilibrium of the periodic version of Hall’s model using solves.m. Figures 17.10.1.a and 17.10.1.b report the spectral density of consumption and investment, computed by using the Tiao-Grupe formula. Both consumption and investment display seasonality, it being more pronounced in

412

Periodic Models of Seasonality

investment than in consumption. This is a reflection of the consumptionsmoothing property of the model. For the impulse response functions of investment with respect to the innovation in the endowment sequence, we used simpulse.m to compute the {dk,v } and {gk,v } sequences corresponding to the moving average representations defined in (17.7.15) and (17.7.16). Figure 17.10.3 reports {dk,v } . The coefficients for each quarter are smooth functions of the lag, but they vary across quarters. Figure 17.10.3 reports {dk,v }, which are each oscillatory functions of the lag. Recall that our periodic 1 of Hall’s model is a one-shock model, with the only stochastic source of disturbances coming from the white noise endowment process. It follows that if we were to shut down the periodic time variation in the productivity of capital, all of the impulse response functions displayed in figure 17.10.3 and 17.10.4 would be equal to one another. 10 The discrepancies across these impulse response functions is a convenient “window” for examining the hidden periodic structure present in investment in this model. Figure 17.10.4 reports the moving average coefficients associated with the univariate Wold representation for investment, which we have normalized by setting the innovation variance equal to unity (so that it is comparable in units with the impulse response functions in figures 17.10.2 and 17.10.3. The coefficient at zero lag in this moving average is .7528, while the coefficients at zero lag for the moving average kernels in figures 17.10.2 and 17.10.3 are (by quarters) .7075, .7069, .7062, .7002. 11 The squared values of each of these coefficients are the one-step ahead forecast error variances in investment, by quarter, when we condition on knowledge of the quarter. The squared value of the coefficient .7528 from the (time-invariant) Wold representation formed by not conditioning on quarter is larger, as we would expect.

10 For this statement to be true in general requires checking that the first of the “two difficulties” discussed by Hansen and Sargent [1990] is not present. 11 The zero lag coefficients are equal for both the {d } and the {g } sequences. k,v k,v

Periodic Innovations Representations for the Periodic Model

10 3

413

10 3

10 2 10 2

10 1 10 1 10 0

10 0 10 -1

10 -2

0

0.5

1

1.5

2

2.5

3

10 -1

3.5

Fig. 17.10.1.a. Spectral density of consumption for a periodic version of Hall’s model, calculated by applying the Tiao-Grupe formula.

0

0.5

1

1.5

2

2.5

0.7 0.6 0.5 0.4 0.3 0.2 0.1

0

10

20

3.5

Fig. 17.10.1.b2. Spectral density of investment for a periodic version of Hall’s model, calculated by applying the Tiao-Grupe formula.

0.8

0

3

30

40

50

60

Figure 17.10.2: The response of the investment component of yp·t−p+k to an innovation in the endowment shock in a periodic verions of Hall’s model.

414

Periodic Models of Seasonality

0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

0

10

20

30

40

50

60

Figure 17.10.3: The response of investment to wp·t−p+k in a periodic version of Hall’s model.

0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 -0.1 0

10

20

30

40

50

60

Figure 17.10.4: The moving average coefficients for a Wold moving average representation of investment, calculated by factoring the spectral density of investment given by the TiaoGrupe formula.

Periodic Innovations Representations for the Periodic Model

415

17.11. Periodic Innovations Representations for the Periodic Model An equilibrium can be represented as xt+1 = Aos(t) xt + Cs(t) wt+1

(17.11.1)

yt = Gs(t) xt + εyt

(17.11.2)

where yt is a vector of objects that are linear combinations of the state xt , plus a white noise measurement error εyt . The matrix Gs(t) is built up from components of the matrices S·,s(t) and M·,s(t) described above. We assume that the measurement error εyt is orthogonal to the wt+1 process, and that it is serially uncorrelated with contemporaneous covariance matrices ¯ s(t) . E εyt ε′yt = R

(17.11.3)

Associated with system (17.11.1) – (17.11.2) is a periodic innovations representation x ˆt+1 = Aos(t) x ˆt + Ks(t) at (17.11.4) yt = Gs(t) x ˆt + at where x ˆht = E[xt | yt−1 , . . . , y1 , x ˆ0 ], at = yt − E[yt | yt−1 , . . . , y1 , x ˆ0 ], and ′ Eat at = Σs(t) . In (17.11.4), Ks(t) is the periodic Kalman gain. The matrices {Σs(t) , Ks(t) } are the p limits of the p convergent subsequences of the Kalman filtering equations: ′ Σt+1 = Aos(t) Σt Ao′ s(t) + Cs(t) Cs(t)

¯ s(t) )−1 Go Σt A′ − Aos(t) Σt G′s(t) (Gs(t) Σt G′s(t) + R s(t) s(t) ¯ s(t) )−1 . Kt = Ao Σt G′ (Gs(t) Σt G′ + R s(t)

s(t)

(17.11.5)

s(t)

¯ s(t) ] are time-varying, system (17.11.5) Because the matrices [Aos(t) , Cs(t) , Gs(t) , R will not converge. But because the matrices [Aos(t) , Cs(t) , ¯ s(t) ] periodic, there is a prospect that {Σt , Kt }∞ will consist of p conGs(t) , R t=1 vergent subsequences. This prospect is realized under regularity conditions that typically obtain for our problems.

416

Periodic Models of Seasonality

The innovation covariance matrix associated with (17.11.4) is Eat a′t = Ωs(t) ¯ s(t) . = Gs(t) Σs(t) G′s(t) + R

(17.11.6)

Given a sample of observations for {yt }Tt=1 , the likelihood function conditioned in x ˆ0 can be expressed as L∗ = − T ln 2π − .5 − .5

t X

T X t=1

ln | Ωs(t) |

(17.11.7)

a′t Ω−1 s(t) at .

t=1

A. A Model of Disguised Periodicity This appendix characterizes a notion of hidden periodicity in a stationary time series, and describes a strategy for detecting its presence in a given vector time series. 12 The notion of hidden periodicity permits realizations of a stochastic process to be aperiodic, but requires that some particular functions of the tail of the stochastic process be periodic. As we shall see, these particular functions are time series averages of skip-sampled versions of the underlying process. It is averaging and skip sampling that causes the hidden periodicity to drop its 0. Because the apparatus introduced in this appendix is abstract, we begin in section A1 with a heuristic account that is designed to indicate the motivation behind the formal apparatus introduced in section A2.

12 Breiman [1968] is a useful background for the material presented in this section.

A1. Two Illustrations of Disguised Periodicity

417

17.13. A1. Two Illustrations of Disguised Periodicity Let {yt } be an n -dimensional stochastic process that is observed by an econometrician. We can use the Kolmogorov Extension Theorem to construct such a process on a sample space Ω = (Rn )∞ , which is the infinite product space formed by taking copies of n -dimensional Euclidean space. A sample point in Ω can be expressed as an infinite-dimensional vector (r0 , r1 , . . .) where rj is in Rn for each j . Probabilities are then defined over the product sigma algebra generated by taking products of the Borel sets of Rn . Armed with this construction, for any ω = (r0 , r1 , . . .), let yt (ω) = rt . Thus, yt (ω) is simply the tth component of the sample point ω = (r0 , r1 , . . .). An alternative way to represent the process {yt } is in terms of a shift operator S . First, define a random variable y : Ω → Rn as y(ω) = r0 . Define the shift transformation S via: S[(r0 , r1 , r2 , . . .)] = (r1 , r2 , r3 , . . .). Then because yt (ω) = rt , an alternative representation of yt is yt (ω) = y[S t (ω)], where S t is interpreted as applying S t times in succession. In thinking about hidden periodicity, the following example is of pedagogical interest. Example 1: Suppose that n is one and that all of the probability on Ω is concentrated onto two points, say a and b . Let a be a sequence of alternating ones and minus ones, beginning with a one. Let b be a similar sequence except that it begins with a minus one. Note that S(a) = b, and S(b) = a. There are many probability structures that we can impose on Ω in Example 1. We can assign any probability between zero and one to a and the remaining probability to b . This assignment amounts to initializing the process. Unless we assign probability one half to each point, the resulting process will not be stationary.

418

Periodic Models of Seasonality

In one sense, the initial assignment of the probability is quite irrelevant. The {yt } process is deterministic in the sense that given knowledge of the y0 , the entire future of process can be forecast perfectly. Since the initial condition tells the whole story, one might just as well condition on it. However, from the vantage point of interpreting time averages of the process, the initial assignment of probability one half to each point is convenient. Independently of how we initialize the stochastic process, it obeys a Law of Large Numbers. Thus, take any Borel measurable function φ mapping Ω into R and form the sequence {zt } where zt = φ(yt , yt+1 , . . .). Then lim (1/N )

N →∞

N −1 X

zt = (1/2)z(a) + (1/2)z(b)

(17.13.1)

(17.13.2)

t=0

where z0 ≡ z . The equality holds when the left-side of equation (17.13.2) is evaluated at either a or b . When probability one half is assigned to each point, the right side of (17.13.2) can be expressed as Ez , so that we have the usual characterization of the limit points of sample averages as mathematical expectations. With this assignment of probabilities, the process {yt } is both stationary and ergodic. For this particular example, realizations of both the original process {yt } and the constructed process {zt } are periodic sequences. While realizations of {yt } have period two, realizations of the constructed process {zt } can have period one for particular choices of zt . For instance, let zt = yt + yt+1 . Then for either a or b, {zt } is a sequence of zeroes and hence has period one. More generally, for this example the periodicity of {zt } can never exceed two. This follows from the fact that S 2 (a) = a and S 2 (b) = b implying that zt+2 (ω) = zt [S 2 (ω)] = zt (ω). Since the maximum periodicity of any constructed process {zt } is two, we will say that the periodicity of S is two. There is something very special about Example 1. Since realizations of the original {yt } process are periodic, every constructed process {zt } turns out to be periodic. In this paper, we are interested in more general circumstances in which the periodicity is disguised. We do not wish to confine attention to processes {yt }

A1. Two Illustrations of Disguised Periodicity

419

whose realizations have an exact periodicity. The following example embodies what we mean by a hidden periodicity. Example 2: Let {wt } be an nw -dimensional Gaussian white noise with covariance matrix I . Construct an nx -dimensional stochastic process {xt } recursively via xt+1 = At xt + Bt wt+1 where {(At , Bt )} is a periodic sequence with period two, where At is an (nx × nx ) matrix, and {Bt } is an (nx × nw ) matrix. Let {yt } be an n -dimensional process generated as a time-varying function of {xt } yt = ft (xt ) where {ft } is a sequence of Borel measurable functions mapping nx -dimensional Euclidean space into n -dimensional Euclidean space. Let ft be a sequence of period two. Realizations of {yt } will not be periodic, but will inherit a sort of disguised periodicity from {(At , Bt , ft )}. There are two aspects of the process {yt } that we have left unspecified, namely x0 and the periodic sequence {(At , Bt , ft )}. As in Example 1, there is flexibility in the probabilistic specification of {(At , Bt , ft )}. One possibility is, in effect, to condition on {(At , Bt , ft )}, in which case the resulting process {yt } will not, in general, be stationary. Alternatively, we can view {(At , Bt , ft )} as emerging from a random draw from two possible sequences indexed by, say, a and b where [At (b), Bt (b), ft (b)] = [At+1 (a), Bt+1 (a), ft+1 (a)] for all t . As in Example 1, if we assign probability one half to each of these outcomes, under a restriction 13 on a matrix that is a function of At (a) and At (b), we can find an initial specification of x0 under which {yt } is a stationary stochastic process. In this case, we can apply the Law of Large Numbers for stationary processes both to show that time series averages converge and to obtain a characterization of the limit points. Suppose that it is possible to complete the specification in Example 2 so that {yt } is stationary. Consider how the hidden periodicity can be characterized and detected. Let ψ be a Borel measurable function mapping Ω → R and form a 13 The restriction is that the matrix A ˆ in equation (4.7) below have eigenvalues that are bounded in modulus by unity.

420

Periodic Models of Seasonality

scalar stochastic process {zt∗ } via

zt∗ ≡ ψ(yt , yt+1 , . . .) or

zt∗

∗

(17.13.3)

t

= z (S (ω))

We assume that E | z ∗ (ω) |< +∞, where z ∗ (ω) = z0∗ . In contrast to Example 1, when the periodicity is hidden, there is no necessity that {zt∗ } form a periodic sequence. Hence we must have a weaker notion of periodicity, if the S implied by Example 2 is to be classified as periodic with period 2. A workable notion of hidden periodicity can be formulated in terms of a reduced class of constructed processes. Given an integer j ≥ 1 and given ψ , define φ : Ω → R, via φ(yt , yt+1 . . .) = zt where zt ≡ lim (1/N ) N →∞

N −1 X τ =0

∗ zt+τ ·j ,

(17.13.4)

and where the right side of (17.13.4) is defined as an almost sure limit. Note that the process {zt } is constructed by taking time series averages of skip samples of the process {zt∗ } with skip interval j . Notice that zt depends only on the tail of the stochastic process {yt }. It follows by construction that zt is a periodic process with a period not exceeding j . The time series averages of skip samples will reveal the hidden periodicity. The idea is to compute (17.13.4) for j = 2, 3, 4, . . . , and then to determine the period pˆ of this sequence for each j . Thus, in example 2, it will turn out that for j = 1, 3, 5, . . . the number pˆ is one. For j = 2, 4, 6, 8, . . . , the number pˆ will turn out to be 2. We shall define the hidden periodicity p as the maximum of these numbers pˆ over j = 1, 2, 3, . . . , where the maximum is also understood to be taken over a class of “test functions” ψ . Thus, the notion of hidden periodicity in a stochastic process that we shall use is the periodicity to be found in time series averages of skip sampled versions of the data. In the next subsection, we develop these ideas formally, and define hidden periodicity precisely in terms of the properties of the shift operator S and its iterates.

A2. Mathematical Formulation of Disguised Periodicity

421

17.14. A2. Mathematical Formulation of Disguised Periodicity We now use the familiar formalism for stationary stochastic processes. 14 As in the previous subsection, let (Ω, F, P r) denote the underlying probability space, and let S be a measurable, measure-preserving transformation mapping Ω into itself. Definition 1: A transformation S is measure preserving if P r(f ) = P r(S −1 f ) for all f ∈ F . Let I be the collection of invariant sets of the transformation S . Definition 2: f ∈ F is an invariant set of S if S −1 (f ) = f . The collection I turns out to be a sigma algebra of events (see Breiman), so expectations conditioned on I are well defined. The invariant events of the transformation S given in example 1 are the null set and any set containing {a, b}. Definition 3: S is ergodic if all invariant events have probability zero or one. Notice that S in example 1 is ergodic. Let L be the space of random variables with finite absolute first moments, and let M be the subspace of L consisting of the random variables that are I measurable. The expectation operator E(· | I) maps L into M. Throughout this section, we use the common convention that equality between random variables is interpreted formally as equality with probability one. Hence the equivalence class of random variables in L that are equal almost surely are treated as one element. Similarly, for a random variable to be in M, it suffices for it to be in L and to be equal almost surely to a random variable that is measurable with respect to I . When S is ergodic, M contains only random variables that are constant almost surely. A transformation S that is measure-preserving can be used to construct processes that are strictly stationary. Let z be a random variable in L, and construct zt (ω) ≡ z[S t (ω)]. 14 See Breiman [1968, chapter 6].

(17.14.1)

422

Periodic Models of Seasonality

Then {zt } is strictly stationary and hence obeys a Law of Large Numbers. The limit point of the time series averages is given by E(z|I). A z ∈ L has two interpretations. First, it indexes a stochastic process via (17.14.1); and second it denotes the time zero component of that stochastic process. Our purpose is to define a notion of periodicity for the transformation S . Suppose there exists a random variable z such that the realizations of the resulting process {zt } are periodic. That is, for some j the resulting process satisfies: zt+j = zt for all t ≥ 0.

(17.14.2)

The fact that (17.14.2) holds for a particular stochastic process is informative about the periodicity of S but falls short of determining the periodicity of S . Notice that one can always find a random variable z such that (17.14.2) is satisfied for j = 1. In particular, let z be constant over states of the world. Since S is measure-preserving, zt = z for all t . Heuristically, we shall define the periodicity of S by forming a large set of periodic stochastic processes defined as in (17.14.1) and satisfying (17.14.2) for some j , and then calling the periodicity S the maximum j over these processes. Notice that all transformations S have periodicity of at least one. To define formally the periodicity of S , we investigate the collection of invariant events of integer powers of the transformation S . Evidently, if S is measure-preserving, then S j is measure-preserving for any positive integer j . We can think of S j as corresponding to skip-sampling every j time periods. Let I j denote the collection of invariant events of S j , and let Mj denote the corresponding subspace of L of random variables that are I j measurable. Any invariant event of S is also an invariant event of S j . Consequently M ⊂ Mj . The converse is not true, however. Consider example 1. Note that S 2 (a) = a and S 2 (b) = b . Consequently, {a} and {b} are invariant events of S 2 but not of S . In this case I 2 = F . When M consists only of random variables that have the same values on a and b, M2 = L. Processes that are generated (indexed) by elements of M are constant over time and hence have period one. On the other hand, processes generated by elements of M2 can oscillate with period two. It of interest to obtain a characterization of Mj that applies more generally. Lemma 1: For any z ∈ Mj , zt+j = zt for all t ≥ 0. Conversely, for any z ∈ L such that zt = zt+j for all t ≥ 0, zt ∈ Mj for all t ≥ 0.

A2. Mathematical Formulation of Disguised Periodicity

423

Proof: Suppose that z ∈ Mj . Then S −j ({z ∈ b}) = {z ∈ b} for any Borel set b of R. Note that S −j ({z ∈ b}) = {zj ∈ b}. Consequently for any Borel set b , {z ∈ b} = {zj ∈ b}. Equivalently, zj = z . Repeating this same argument, it follows that z = zτ ·j for any positive integer τ . Recall that S is measure-preserving, as is S t . Consequently, for any Borel set b , P r({zt ∈ b} ∩ {zt+τ ·j ∈ b}) = P r({z ∈ b} ∩ {zτ ·j ∈ b}), P r({zt ∈ b}) = P r({z ∈ b}), and P r({zt+τ ·j ∈ b}) = P r({zτ ·j ∈ b}). Since z = zτ ·j , it follows that P r{zt = zt+τ ·j } = 1 for any positive integer τ . Next consider the converse. Suppose that z ∈ L such that zt = zt+j for all t ≥ 0. It remains to show that zt ∈ Mj . The sequence of time series averages {(1/N )

N −1 X τ =0

zt+τ ·j }

converges almost surely to zt as well as to E(zt | I j ). Therefore, P r{zt = E(zt | I)} = 1. In light of Lemma 1, processes generated by elements of Mj are periodic with a period that is no greater than j . We wish to use this insight to construct a formal definition of periodicity. Let Mcℓ be the closed linear space generated by {Mj }∞ j=1 where closure is defined using the standard norm on L, E(| · |). Definition 4: The transformation S is said to have periodicity p if p is the smallest integer such that Mp = Mcℓ . Under this definition, random variables in Mcℓ generate periodic processes with maximum period p . Applying this definition to the transformation S given in example 1, we verify that S has period 2. Next we describe an alternative way to deduce the periodicity of S . Mimicking the previous logic, we can show that for any positive integer j , Mj ⊂ Mτ ·j for τ = 1, 2, . . . .

(17.14.3)

It turns out that if ⊂ in (17.14.3) can be replaced by = , the period of S is no greater than j , and in fact j must be an integer multiple of the actual periodicity p . In other words, once skip-sampling reaches a point where further sampling fails to increase the collection of invariant events, this point is an integer multiple of the periodicity of S .

424

Periodic Models of Seasonality

Lemma 2: Let j be any positive integer such that Mj = Mj·τ for τ = 1, 2, . . . . Then Mcℓ = Mj and S has periodicity p where j = ℓ · p for some positive integer ℓ. Proof: First we show that Mcℓ = Mj . Suppose to the contrary that there is some random variable in Mcℓ that is not in Mj . Since Mj is closed and random variables in Mcℓ are limit points of sequences of random variables in S τ M , there exists a positive integer τ and a random variable z such that z is in Mτ but not in Mj . However, Mτ ·j = Mj by assumption, which is a contradiction. Therefore Mcℓ = Mj and p ≤ j . It remains to show that j = p·ℓ for some integer ℓ. Note that Mp = Mj = Mcℓ . In light of Lemma 1, random variables in Mj generate processes with period p and period j . Let ℓ be the smallest integer such that ℓ · p ≤ j and suppose that ℓ·p < j . Then p > k > 0 where k ≡ j−ℓ·p . For any z ∈ Mp , with probability one z = zp = zℓ·p = zℓ·p+k . Since S is measure-preserving, {zt } is periodic with period k . It follows from Lemma 1 that, z ∈ Mk . Consequently, Mk = Mp which is a contradiction. This in turn implies that the period of S is at least j − ℓ · p , which is a contradiction. Therefore j = ℓ · p . An implication of Lemma 2 is that processes generated by random variables in Mcℓ are periodic with a period equal to j , where j = ℓ · p for some integer ℓ. Note that if S has periodicity p , then S p has periodicity one. Definition 4 of periodicity can be applied to any S transformation that is measure-preserving. Our interest is in the case in which S is the shift transformation described in section 1a. This transformation is measure-preserving by construction as long as the probability measure induced on Ω comes from a process {yt } that is strictly stationary. When the shift transformation is periodic with period p , we say that the process {yt } has hidden periodicity p . Consider again constructions (17.13.3) and (17.13.4). The processes {zt } constructed via (17.13.4) are periodic by construction and hence it follows from Lemma 1 (or from the Law of Large Numbers for Stationary Processes) that the corresponding random variable z is in Mj . The periodicity of {zt } can, in fact, be less than j . By choosing a sufficiently rich collection of test functions, we can span Mj . Let pˆ(j) be the maximum periodicity over such a class of functions. The hidden periodicity p of {yt } is then the supremum of the sequence {ˆ p(j) : j = 1, 2, . . .}. Lemma 2 describes a particular feature of subsequences of {ˆ p(j) : j = 1, 2, . . .}. For instance, for any j = p · ℓ for some ℓ, the subsequence {ˆ p(τ · j) : τ = 1, 2, . . .} is constant. Turning this observation

A2. Mathematical Formulation of Disguised Periodicity

425

around, if one finds a constant subsequence of the form {p(τ · j) : τ = 1, 2, . . .}, then the hidden periodicity of {yt } must satisfy j = p · ℓ.

Part IV Economies as Objects

Chapter 18 Introduction to Objects This help manual is intended to help students use the MATLAB programs referenced in this book. To that end, it is divided into two main chapters, organized by increasing level of difficulty. This first chapter explains a little about object oriented programming (OOP), and why it’s so useful in this context. The second chapter applies these ideas to the construction of an economy object, and offers a more in-depth coverage of all the features of an economy and all the nifty things one can do with it. This second chapter also provides the user with the tools, via examples, to invent new economies to experiment with. Throughout, actual code and MATLAB file names will be in typewriter font, but references to an object will not be. For example, economy.m is in typewriter font, but when an economy is referred to, it is not. Also note that any actual MATLAB file has a short help section at the beginning, which can be accessed by typing help file name at the MATLAB prompt.

18.1. Matlab Objects For those users who have not encountered OOP before, this subsection goes over some definitions and some examples of what one can and cannot do with objects. Of course, all examples are in the context of MATLAB.

18.1.1. Definitions We start with some definitions of basic concepts in OOP. Class The relevant analogy here is that of a type. A class is a new data type that you define. It includes not only the actual structure of the type, but also the functions that operate on it. So, for example, suppose we define a new class called a slde (short for stochastic linear difference equation) which is a collection of two matrices, and we need a function on it that displays it

– 429 –

430

Introduction to Objects

in a nice way. Then both the constructor slde.m and the display function disp.m are in the class. Objects An object is a run-time value which belongs to some class. If a class is a type, an object is a variable. A variable of a given class is called an instance of that class. For example, suppose we define a class that is a matrix. Then the identity matrix would be an object, or an instance of the matrix class. Hierarchy Classes in OOP are arranged in a tree-like hierarchy. A class’ superclass is the class above it in the tree. The classes below it are subclasses. By convention, the root of the tree is called the “Object” class. The semantics of the hierarchy are that any class includes all the properties of its superclasses. In this way the hierarchy is general towards the root and specific towards its leaves. The hierarchy helps add logic to a collection of classes. It also enables similar classes to share properties through inheritance. Superclasses are often referred to as parent classes, and subclasses are often referred to as children. Inheritance A subclass inherits all of the data and functionality of its parent classes. In particular, a class inherits all of the methods. When an object receives a message, it checks for a corresponding method. If one is found, it is executed. Otherwise the search for a matching method travels up the tree to its parent and so on recursively. This means that a class automatically responds to all the messages of its superclasses. Most OOP languages include controls to limit how the data and methods are inherited. A subclass can also extend beyond the inherited functionality by adding data and defining new methods. Overriding and Overloading A class inherits all the methods of its superclasses, but a class can choose to respond to a message in a different way by re-defining a method. When an object receives a message, it checks its own methods before consulting it superclass. If the object’s class and its superclass both contain a method for a message, the object’s method is used. In other words, the first method found in the hierarchy takes precedence. When a subclass responds to a

Matlab Objects

431

message in a different way than its superclass does, the subclass is said to have overriden its superclass’s method—the class overrides and intercepts the message before it gets to the superclass. When a method has more than one definition depending on the context (i.e., it’s defined for both a class and its parent), it is said to be overloaded. In MATLAB, most functions, including addition and subtraction, for example, can be overloaded to be class specific. For example, the function disp.m is overloaded for each class, so that it displays each class properly. Fields A field is a component of a class which is itself a class, or predefined type. For example, if we define a class called foo that consists of an integer and a matrix, then foo has two fields. If we defined a child class called foo child that additionally has a vector, then foo child also has two fields. The first is the foo field that is inherited, and the second is the vector field. There are not three fields, as the first two are subsumed into the foo field that foo child inherits.

18.1.2. Matlab Specifics There are two main ways in which MATLAB departs from the structure laid out above: accessing fields, and overloading certain functions. Additionally, there is a specified way in which MATLAB checks for methods, which we discuss here, as it may sometimes cause confusion. Accessing Fields Generally, the most direct way to access a field of an object is to call it via foo.fieldname. When defining new objects, this method works immediately within the functions of a class. However, to be able to do so at the command line, two extra functions are needed in each class directory: subsref.m and subsasgn.m. The first allows you to reference the value in the field, so if you type foo object.fieldname on the command line, it returns the value contained in the fieldname field of object foo object. The latter function, subsasgn.m, allows you to reassign a field of an object to be a different value. These two functions should be overloaded for every class. They work by taking an object both as an argument and as the

432

Introduction to Objects

output. This will turn out to be especially handy for reassigning values in an object in the economy class, as we will see in the next chapter. Overloading Functions One must be wary about overloading certain functions in MATLAB. For example, we defined our own method sigma.m to extract the sigma field from an economy object. However, sometimes it seems to work, and sometimes it seems to not work. We suspect that this is because there is already a sigma function defined by MATLAB, which it does not like to override. In general, it may be wise to try and name a function a different name, such as sig.m, rather than overloading. One can always check for taken names by typing help function name; if MATLAB says it cannot find the function, it hasn’t been taken. Calling Functions MATLAB has a specified protocol for where to seach when a new name is called. In the following order it looks for a: variable, subfunction, private function, or function on the search path. A subfunction is a function that resides in the same file as the calling function; we almost never use these. A private function is one that is in a private directory, and hence is only accessible to files in the directory immediately above it. An example of this is the solve.m function in the @economy/private directory. The search path is the predefined path along which MATLAB searches, to which we have added paths for examples/econ and clex. For more information on paths in MATLAB, type help path.

18.1.3. How to Define a Matlab Class Suppose one wants to define a MATLAB class called foo. Then in a folder called @foo there must a constructor function called foo.m. If one defines a child class called foo child, then there must be another folder @foo child with a constructor function foo child.m. Note that when constructing a foo object, one must be in a directory such that @foo is a subdirectory. The folders @foo and @foo child may both be in the same directory. The constructor function takes inputs and assigns them to fields. Once all fields are assigned, there is a declaration of the class, and the new object is

Matlab Objects

433

returned. A short example for @foo/foo.m follows, where foo has two fields, an integer and a matrix. Note that it is a good idea to include some comments about what a function does in the function definition file. In Matlab, the first comments section is displayed when help function name is typed. The character for commenting out a line is the percent symbol: %. The help section should include definitions of all variables used, the first line of the function definition, and a short description of what the function does.

function f = foo(integer1, matrix2); % function f = foo(integer1, matrix2); % This function is the constructor called when a new foo object % is defined. It takes two arguments, an integer and a matrix % and assigns them to the two fields of a foo object. foo.integerfield = integer1; foo.matrixfield = matrix2; f = class(f, ‘foo’); To create an object called foo1 using the integer 5 and the matrix eye(4), one simply types at the command line: foo1 = foo(5, eye(4)). Now suppose one wants to define a class called foo child that is a child of foo. In the same directory as @foo create a folder called @foo child that contains the constructor function foo child.m. Suppose that in addition to the two fields in foo, one wants foo child to have a vector field. The file @foo child/foo child.m is given for reference..

function fc = foo child(foo1, vector2); % function fc = foo child(foo1, vector2); % This function is the constructor called when a new foo child % object is defined. It takes two arguments, a foo object % and a vector, and assigns them to the two fields of a foo child % object. foo child.vectorfield = vector2; fc = class(fc, ‘foo child’, foo1);

434

Introduction to Objects

Notice that, since foo1 is an object of the foo class, foo child automatically inherits from foo, and the inherited fields are filled with the values from foo1. For a more thorough introduction to defining MATLAB objects, see Chapter 14 of the MATLAB manual “Using MATLAB”.

18.2. Summary In this chapter we have defined some of the basic concepts in object oriented programming: classes, objects, hierarchy, inheritance, overriding and overloading, and fields. We then went on to point out some pitfalls of MATLAB objects: field access and overloading. Finally, very briefly, we went over how to define a MATLAB class. One can now see why object oriented programming would be a valuable tool in defining and using economies. An economy is defined to be a collection of matrices. These divide naturally into three categories, each of which define the information, technology and preferences structures. Thus one can create three classes representing these structures, and have an economy class inherit from all three. Also, most interesting operations are executed on an economy, and with an economy class, one can simply define functions that operate on an economy object. The exact definitions and functionality of these four classes are discussed in the next chapter.

Chapter 19 Economies as Matlab Objects

19.1. Introduction We describe in the first few sections the structure of the economy class starting with the structure of the parent classes: information, technology and preferences. Notice that the field names correspond as much as possible with the names of matrices and vectors in the main book. We also describe the functions available to manipulate the various objects. To get more information on any of these functions, type help function name in the MATLAB command window. Also, recall that the information, technology and preference divisions are introduced in Chapter 3 of Hansen and Sargent. The last section deals with three different ways of working with objects of this class: using the built-in economies, mixing and matching the built-in parent objects (information, technology and preferences) and building a customized economy.

19.2. Parent Classes: Information

19.2.1. Structure The first parent of the economy class is the information class. An information object contains matrices describing the laws of motion of taste and technology shocks. zt+1 = A22 zt + C2 wt+1 bt = Ub zt dt = Ud zt . An information object has four fields: a22:

The matrix A22

– 435 –

436

c2:

Economies as Matlab Objects

The matrix C2

ub: The selector matrix Ub , which transforms the process zt into taste shocks. ud: The selector matrix Ud , which transforms the process zt into technology shocks.

19.2.2. Functions construction: information, a constructor function that takes as argument the matrices a22, c2, ub, ud

19.3. Parent Classes: Technology

19.3.1. Structure The technology class contains matrices that describe the technology of the economy : Φc ct + Φg gt + Φi it = Γkt−1 + dt kt = ∆k kt−1 + Θk it . Recall that ct is consumption, gt a vector of intermediate goods, it investment, kt capital, and dt the production shock. A technology object has six fields : phic:

The matrix Φc

phig:

The matrix Φg

phii:

The matrix Φi

gamma:

The matrix Γ

deltak:

The matrix ∆k

thetak:

The matrix Θk

Parent Classes: Preferences

437

19.3.2. Functions construction: technology, a constructor function which takes as arguments the matrices phic, phig, phig, gamma, deltak, thetak

19.4. Parent Classes: Preferences

19.4.1. Structure The preferences object contains scalars β, σ and matrices ∆h , Θh , Λ, Π that describe the preferences of the representative agent. The household technology is: ht = ∆h ht−1 + Θh ct st = Λht−1 + Πh ct and preferences are ordered by E0

∞ X t=0

β t (st − bt )2 + ℓ2t .

Recall that ht is a household stock of durables, ct is a consumption, and st is services from the stock of durables. The vector bt is used in the household objective function. A preferences object has six fields : deltah:

The matrix ∆h

thetah:

The matrix Θh

lambda:

The matrix Λ

pihh:

The matrix Πh

beta:

The discount factor β

sigma:

The risk sensitivity σ

438

Economies as Matlab Objects

19.4.2. Functions construction: preferences, a constructor function that takes as argument deltah, thetah, lambda, phih, beta, sigma

19.5. Child Class: Economy

19.5.1. Structure The elements of an economy are contained in the parent fields: information, technology, and preferences. An economy as a child object inherits these three fields from its three parents (technology, preferences, and information). In addition, an economy object has several other fields that are not inherited, namely a set of matrices that characterize a competitive equilibrium. These are calculated automatically when an economy object is defined and become fields in the economy object. Furthermore, the function subsasgn.m for the economy class has been defined so that whenever one changes the value of a field of an object, the equilibrium is automatically recalculated. We shall illustrate this useful feature below. Chapters 4 and 6 showed that an equilibrium has the representation xt+1 =A0 xt + Cwt+1 yt =Gxt . ′ ht−1 0 A11 A12 . , C= where x′t = kt−1 , Ao = C2 0 A22 zt−1 The observables and the shadow prices are in the vector yt . They are all linear combinations of the state variables in xt . The coefficients of these linear combinations are in the matrix G , various rows of which were denoted Mj in Chapter 4 for a price of a quantity j .

The fields containing the solutions of the resource allocation problem are the following :

Child Class: Economy

439

ao : The matrix A0 endo: The eigenvalues of the block A11 of A0 exo: The eigenvalues of the block A22 of A0 nnc: The coordinate of the row of x containing a constant, if any c: The matrix C sj, mj: Quantities and shadow price are stored in sixteen different fields: sb, sc, sd, sg, sh, si, sk, skl, ss, mc, md, mg, mh, mi, mk, ms. These correspond with the matrices of analogous names. For example, sc is the Sc of Chapter 4, which multiplies xt to yield the optimal decision for ct . And mc multiplies xt to yield the Lagrange multiplier Mct (the shadow price of consumption). s space: An s space (state space) object containing the state space representation of the economy (@[email protected] ). To have more information about what a state space object is, type help s space at the MATLAB prompt. This field is useful for computations involving the MATLAB control toolbox.

19.5.2. Fields containing the history of the economy An economy object contains also the history of the economy, that is an initial condition x0 . and a sequence of shocks {wt }Tt=0 . (The sequence of shocks can be initialized at a null matrix, and has been in our sample economies. We include it as a potential field because it can be useful for generating simulations.) This information is stored in the following fields: hinitial: The initial condition for the household capital goods kinitial: The initial condition for the capital stock zinitial: The initial condition for the information process shocks : The sequence of shocks in the information process

440

Economies as Matlab Objects

19.5.3. Functions

19.5.4. Constructing the object and changing parameters This part will be detailed in section 3 where we will explain what sequence of commands is needed to construct an economy object. For the time being, we simply list the functions : construction: economy, a constructor function that takes as arguments an information object, a technology object and a preferences object. display: A function disp which display the structure of the economy

19.5.5. Analyzing the economy steady state: A function steadst to compute the steady state of the model. This uses the resource allocation solution fields. asset pricing: A function asset price to compute and simulate the price of an asset with payoffs that are linear combinations of the state variable of the economy A function riskprem to compute the risk premium on an asset A function sure4j to compute the prices of a j-period sure claim on consumption simulation: A function simulate to compute and graph time path for the observables and shadow prices. reopening markets: A function reopening to compute and graph the time path of prices in markets that reopen every period impulse response: A function impulse to compute and graph the impulse response of the observables and shadow price to the shocks hitting the economy arma representation: A function arma rep to compute the arma representation of the

Working with economies

441

spectral density: A function spect to compute and graph the spectral densities of the observables and shadow prices.

19.6. Working with economies

19.6.1. The built-in economies You will find in the directory examples/econ a series of script files which build standard economies – most of them described in the book. A typical script file is named economy name.m. MATLAB will build the economy when you type economy name.m at the command line. The program will create the following variable : eeconomy name: an economy object

19.6.2. Mixing and matching built-in parent objects By using the built-in economies, your freedom is very restricted: you cannot set any structural parameters, the only thing you can modify is the history of the economy. To give more freedom to your experiments, we have constructed some standard structures corresponding to the examples given in chapter 3. You can mix and match them and set some of their parameters. Technology You will find in the directory examples/tech functions which create the technologies given as examples in chapter 3. A typical function is named techj.m. By typing help techj you will get a description of the parameters you are free to set. Note that all these parameters have default values. You build a technology by typing name of tech = techj(parameters) Preferences You will find in the directory examples/pref functions which create the preferences given as examples in chapter 3. A typical function is named prefj.m. By typing help prefj you will get a description of the parameters you are free to set. Note that all these parameters have default values. You build a technology by typing name of pref = prefj(parameters)

442

Economies as Matlab Objects

Information You will find in the directory examples/info functions which create information processes, most of them following the examples given in chapter 2. A typical function is named infoj.m. You will have to specify parameters of an underlying stochastic linear difference equation and extractor matrices ub and ud. This is a delicate step: you have to make sure that those matrices are comformable in columns with the process zt and in rows with the consumption services vector (for ub) or with the technology shocks vector (for ud). Having a look at the script files in examples/econ can be useful, although the information constructor is designed to warn you when your matrices are not conformable. Economy After you have built the three parents objects info, tech and pref you are ready to create their child, the economy object. To do so you simply type econ name = economy(i,t,p) and a new economy will be born. You will probably immediately want to reset the initial conditions for the h−1 , k−1 , z0 , which the economy constructor sets at vectors of 1’s as their default values. The initial conditions can be assessed from the economy object by typing econ name.hinitial, econ name.kinitial, and econ name.zinitial, respectively. To reset z0 , for example, type econ name.zinitial = [ 5 2 0 ]’. You can also directly reset other objects of any of the three parent objects (preferences, technology, or information), which will then be automatically inherited by the child economy object. We’ll describe how to do this soon.

19.6.3. Building your own economy In the directory examples/econ there is a script file blank.m which may be useful in building your first economies. Just fill in the blanks (the null matrices) with conformable matrices and run the script file; you’ll be ready to experiment with your new economy. % Creates an economy with null matrices everywhere. % Required dimensions are given in comments. %%% Technology %%% deltak=[] ; % n k by n k thetak=[]; % n k by n i phic=[]; % m by n c phig=[] ; % m by n g phii=[]; % m by n i

Working with economies

443

gamma=[] ; % m by n k %%% Preferences %%% deltah=[]; % n h by n h lambda=[]; % n s by n h thetah=[]; % n h by n c pih=[] ; % n s by n c beta=; % scalar sigma = 0; % scalar %%% Information %%% a22=[]; % n z by n z c2=[]; % n z by n w ud=[]; % n d by n z ub=[]; % n b by n z %%% Construction %%% iblank = information (a22, c2, ub, ud); tblank = technology (phic, phig, phii, gamma, deltak, thetak); pblank = preferences (deltah, thetah, lambda, pih, beta, sigma); eblank1 = economy (iblank, tblank, pblank); eblank1.hinitial = []; eblank1.kinitial = []; eblank1.zinitial = []; clear iblank1 tblank1 pblank1; clear phi gam sigma beta; clear a22 c2 u* phi*; clear gamma del* the* lambda pih;

444

Economies as Matlab Objects

19.7. Tutorial Our object oriented programs are contained in a directory called hansar1 that contains various subdirectories. Say that you keep the directory hansar1 in the location c: \projects \hansar1. After you start MATLAB type addpath c: \projects \hansar1 then type startup. To read one of our existing economies, type for example clex11. The object eclex11 is then created. Type eclex11 to display it. Then to conserve notation rename the economy object simply e by typing e=eclex11. To access one of the fields of the economy e, a child object, type e.j where j is one of the economy fields described above, namely, ao, c, endo, exo, sv, mv, hinitial, kinitial, zinitial, where v denotes one of the variables c, i, h, k. To assess one of the parent objects, type either e.information or e.technology or e.preferences. To assess one of the fields of one of the parent objects, type either e.information.j where j=a22, c or e.preferences.j where j= lambda, deltah, pi, thetah,beta, sigma or e.technology.j where j = deltak, thetak, gamma, phig, phii, phic. To reset an element of a parent object, type for example e.information.a22(2,2) = .4, a command that sets A22 (2, 2) = .4, and that then recomputes all of the equilibrium objects in the child economy e.

Chapter 20 MATLAB Programs This chapter consists of a manual of MATLAB programs that implement the calculations described in earlier chapters. Many of the programs use programs in MATLAB’s Control Toolkit. You should load our programs into a subdirectory of MATLAB, and put this subdirectory on the matlabpath statement in your matlab.bat file. There is a demonstration facility for some of our programs, which supplies a small course on how to use many of our programs. To use this program, just get into MATLAB, type hsdemo, and choose one of the options that the menu offers you.

20.1. Matlab programs Our ordinary MATLAB programs are available via ftp at < ftp://zia.stanford.edu/pub/˜sargent/webdocs/matlab/hansar/hansarprograms.zip> . Our object oriented programs are available at < ftp://zia.stanford.edu/pub/˜sargent/webdocs/matlab/hansarobjects.zip>

– 445 –

446

MATLAB Programs

aarma

Purpose: Creates arma representation for a recursive linear equilibrium model. Synopsis: [num,den,p,z]=aarma(ao,c,sy,i)

Description: The equilibrium is computed by first running

solvea.

The equilibrium is

xt+1 = ao xt + c wt+1 A vector of observables is given by yt = sy xt , where sy is formed to pick off the described variables. For example, if we want yt = [c′t , i′t ], we set sy=[sc; si]. aarma creates num and den, which pertain to the representation den (F )yt = num (F )wit where F is the forward shift operator defined by F yt = yt+1 . This is an arma representation for the response of yt to the i-th component of wt . num(F) and den(F) are each stored with the coefficients being arranged in order of descending powers of F . The poles (zeros of den(F)) are returned in the vector p. The zeros of num(F) for each variable are returned in a column vector z, where each column corresponds to a variable. References: [1] Hansen, Lars Peter and T.J. Sargent, Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988.

Matlab programs

447

aggreg

Purpose: Computes state space representation of sampled (time aggregated) data. Synopsis: [ Ar , Cr , aa, bb, cc, dd, V1] = aggreg (A,C,G,D,R). Description: The underlying model is xt+1 = Axt + Cwt+1 yt = Gxt where wt+1 is a martingale difference sequence. Error ridden observations on y are available only every r periods. The state space model for the data is then xs+1 = Ar xs + Cr wrs+1 ys = Gxs + vs vt+1 = Dvs + us+1 ′ where s = t·r, Eut u′t = R, Ar = Ar , Cr = I, Ewrt wrt = Vr Vr = CC ′ +ACC ′ A′ + · · · Ar−1 CC ′ A′ r−1 . The program uses innov to create an innovations representation for the sampled process {yt , t = 0, r, 2r, 3r, . . .} = {ys , s = 0, 1, 2, . . .}. varma2 can be used to compute an arma representation for the sampled data.

References: [1] Hansen, Lars Peter and T.J. Sargent, Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988.

448

MATLAB Programs

aimpulse

Purpose: Computes impulse response function for a recursive linear equilibrium model Synopsis: [z]=aimpulse(ao,c,sy,ii,ni)

Description: The equilibrium is computed by first running

solvea.

The equilibrium is

xt+1 = ao xt + c wt+1 . A vector of observables is given by yt = sy xt where sy is formed to pick off the desired variables. For example, if we want yt = [c′t , i′t ]′ , we set sy=[sc;si]. aimpulse computes the impulse response of yt with respect to component ii of yt for ni periods. References: [1] Hansen, Lars Peter and T.J. Sargent, Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988.

Matlab programs

449

asimul

Purpose: Simulate a recursive linear equilibrium model Synopsis: asimul, a script file. The outputs of solvea must be in memory, as must the matrix sy and the integer t1 . Description: The equilibrium is xt+1 = Ao xt + Cwt+1 A vector of observables yt obeys yt = sy ∗ xt , is to be specified by the user. If we want yt = (c′t i′t )′ , we would set sy = [ sc; si]. asimul computes a simulation of y of length t1 and stores the output in the matrix y .

where

sy

References: [1] Hansen, Lars Peter and T.J. Sargent, Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988.

450

MATLAB Programs

asseta

Purpose: Computes and simulates asset prices for a recursive equilibrium model. Synopsis: asseta is a script file which requires that pay and nt, as well as the output of solvea, reside in memory. Description: Run solvea and

asimul

first. An asset pays out a stream of returns yt = pay ∗ xt

where pay is a vector and where xt is governed by the equilibrium law of motion xt+1 = Ao xt + Cwt+1 The asset is priced by asset price at t = Et

∞ X

β t ptt+j yt+j .

t=0

The program computes the intertemporal marginal rate of substitution, the payoff, the asset price, and the gross rate of return on the asset. A similation of these of length nt is stored in y . The program also calculates the prices of claims on sure j -period forward consumption for j = 1, 2, 5. A simulation of length nt of these for j = 1, 2, 5 are stored in R1, R2, and R5, respectively. References: [1] Hansen, Lars Peter and T.J. Sargent, Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988.

Matlab programs

451

assets

Purpose: Creates matrices and scalars needed to price an asset in a strictly periodic equilibrium model of period p . Synopsis: assets is a script file.

solves

must be run first and its output must be in memory.

Description: An asset with payoff payt = Ua ∗ xt is to be priced, where xt is the state vector for a dynamic linear equilibrium model that is periodic with period p . The asset price at is given by xt = [x′t µa,s(t) xt + σa,s(t) ]/[ij · Mc,s(t) xt ]. This program computes the matrices µa,s(t) and the scalars σa,s(t) for s(t) = 1, 2, . . . , p. These matrices and scalars are stored in memory. To simulate the asset price, use the program assetss. See also: simuls, assetss.

452

MATLAB Programs

assetss

Purpose: Simulates asset price and term structure of interest rates for a strictly periodic equilibrium model with period p . Synopsis: assetss is a script file. The programs solves, first and their outputs must reside in memory.

simuls,

and

assets

must be run

Description: A simulation is constructed for the asset priced in assets. The term structure of interest rates is also computed. The output of the simulation is returned in the vector y , which equals [mrs, pays, as, ret]. Here mrs is the marginal rate of substitution at time, pays is the payoff of the asset, as is the price of the asset and ret is the return on the asset. The prices of risk free claims on comsumption 1, 2, and 5 periods forward are returned in R1, R2, R5, respectively. See also: simuls, solves,

assets

Matlab programs

453

assetx

Purpose: Computes and simulates asset prices for a recursive equilibrium model with Gaussian Exponential Quadratic specification. Synopsis: assetx is a script file which requires that pay and nt , as well as the output of solvex, reside in memory. Description: Run solvex and

asimul

first. An asset pays out a stream of returns yt = pay ∗ xt

where pay is a vector and where xt is governed by the equilibrium law of motion xt+1 = Ao xt + Cwt+1 The asset is priced by asset price at t = Et

∞ X

β t ptt+j yt+j .

t=0

The program computes the intertemporal marginal rate of substitution, the payoff, the asset price, and the gross rate of return on the asset. A similation of these of length nt is stored in y . The program also calculates the prices of claims on sure j -period forward consumption for j = 1, 2, 5. A simulation of length nt of these for j = 1, 2, 5 are stored in R1, R2, and R5, respectively. See also: asseta, solvex

References: [1] Hansen, Lars Peter and T.J. Sargent, Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988.

454

MATLAB Programs

avg

Purpose: Prepares linear system for analysis of aggregation over time with “integrated” or “summed” data Synopsis: [AA, CC] = avg(A, C, m) Description: The state xt evolves according to xt+1 = Axt + Cwt+1 Let zt = [x′t , x′t−1 , . . . , x′t−m+1 ]′ . Then zt evolves according to zt+1 = AA ∗ zt + CC ∗ wt+1 where

A I AA = 0 .. .

0

0 0 I .. .

··· ··· ··· .. .

···

I

The program forms AA and CC . See also: aggreg

0 C 0 0 0 , CC = .. . . .. . 0 0

Matlab programs

455

canonpr

Purpose: Computes canonical representation of preferences. Synopsis: [lamh, pihh] = canonpr (beta, lamba, pih, deltah, thetah)

Description: The program computes a canonical representation of preferences by solving the auxiliary consumer choice problem, maximize ∞ 1 X t β st · st − E0 2 t=0

subject to ht = ∆h ht−1 + Θh ct st = Λht−1 + Πct , h1 given. The solution is a feedback rule ct = −F ht−1 where F = (Π′ Π + βΘ + h′ P Θh )−1 (βΘ′h P ∆h + Π′ Λ), and where P is the nonnegative definite P ˆ Π) ˆ that solves the algebraic Riccati equation for the problem. A canonical (Λ, is chosen for the equations ˆ −1 Λ ˆ =F Π ′ ˆ Π ˆ = (Π′ Π′ + βΘ′ P Θh ). Π h

456

MATLAB Programs

clex 10, 11, 13, 14, 18, 35, 101c, 101f

Purpose: Read in matrices defining an economy. Synopsis: clex*.m is always a script file. Description: Each clex*.m file creates a list of matrices Φc , Φg , Φi , Γ, ∆k , Φk , ∆h , Φh , Γ, Π, A22 , Ud , Ub , and Ud that define an economy. The economies are as follows: clex clex clex clex clex clex clex clex

10 11 13 14 18 35 101c 101f

The Jones-Manuelli examples of chapter 3. Hall’s model of chapter 3. Hall’s model with higher adjustment costs. Lucas’s model of chapter 3. The “seasonal preferences” model of chapter 7. The “heterogeneous agent” example of chapter 6. The hog model of chapter 8. The corn-hog model of chapter 8.

Matlab programs

457

compn

Purpose: compn creates companion matrix. Synopsis: [B] = compn[a] Description: The companion matrix B of the 1 × n row vector a is defined as a1 1 B= 0 .. .

0

a2 0 1 .. .

··· ··· ··· .. .

an−1 0 0 .. .

0

···

1

an 0 0 . .. . 0

458

MATLAB Programs

disthet

Purpose: Compute equilibrium of general equilibrium with two types of households, externalities, distorting taxes, and exogenous government expenditures. Synopsis: disthet is a script file. All matrices must be in memory. Description: disthet computes a competitive equilibrium of a distorted heterogeneous economy. Two types of agents live in an economy with a government. There are externalities. Type i agent’s problem is to maximize: ∞ X β t [(si (t) − bi (t)) · (si (t) − bi (t)) + ℓi (t)2 ]} E0 − 0.5{ t=0

subject to: si (t) = Λi1 hi (t − 1) + Λi2 ∗ H1 (t − 1) + Λi3 H2 (t − 1) + Πi1 ci (t) + Πi2 C1 (t) + Πi3 C2 (t)

hi (t) = ∆hi ∗ hi (t − 1) + ∆Hi1 H1 (t − 1) + ∆Hi2 H2 (t − 1)

+ Θhi ci (t) + ΘHi1 C1 (t) + ΘHi2 C2 (t) X ∞ E{ t=0 β t [(I + τc )p(t) · ci (t) − (1 − τℓ )w(t)li (t) − α(t) · di (t) − fi ∗ (P1 (t) + P2 (t) − Ti (t)]}|I0 − v0 ∗ k0i = 0

where si , hi , ℓi , ci , Ti , Pi are consumption services, household capital stock, labor, consumption, government transfer of type i agent and firms of type i’s profit at t, i=1,2. Capital letters denote aggregate variables. τj is tax on j, j=c,l,k,i. Firms of type 1’s problem is to maximize expected profit: E

∞ X t=0

{β t [p(t)[c(t) + E(t)] + q(t) · i(t) − r(t) · k(t − 1) − α(t) · d(t) − w(t)ℓ(t)]}

subject to: Φc (c(t) + E(t)) + Φi i(t) + Φg g(t) = Γk k(t − 1) + ΓK ∗ K(t − 1) + d(t) g(t) · g(t) = ℓ(t)2

Matlab programs

459

where c(t) = c1 (t)+c2 (t), and similarly for d(t), ℓ(t). E(t), g(t) are government spending and intermediate goods, respectively. Firms of type 2’s problem is to maximize expected profit: E

∞ X t=0

β t [(I − τk )r(t) · k(t − 1) − (I + τi )q(t) · i(t)] − v0 ∗ k0

subject to: k(t) = ∆k k(t − 1) + ∆K K(t − 1) + Θk i(t), where k0 = k01 +k02 . The state vector in this program is defined as [z(t); z(t); h1(t− 1); h2(t − 1); k(t − 1)].

460

MATLAB Programs

dog

Purpose: Compute ‘mongrel’ (i.e., non-Gorman) preference ordering over aggregate consumption for two households. Synopsis: function[Deltah,Thetah,Lambdah,Pih,Am3,Bm3,Cm3]=dog(alpha, beta,lambda1, pih1, deltah1,thetah1,lambda2,pih2,deltah2,thetah2, a22,c2,ub1,ub2) Description: Computes the canonical mongrel service technology for two households with parameter alpha (the Pareto weight on the first consumer). The mongrel household technology is H(t) = ∆h H(t − 1) + Θh c(t) s(t) = ΛH(t − 1) + Πc(t)

The mongrel preference shock is given by the series connection of the three state space systems (A1,B1,C1,D1), (AA,BB,GG,HH), (∆h , Θh , Λ, Π). We calculate a system representation (Am,Bm,Cm,Dm) for the mongrel shock. The mongrel shock is thus described by Z(t + 1) = Am2 Z(t) + Bm3 w(t + 1) bb(t) = Cm3 Z(t). In using this program, it is important to set the initial condition for the state appropriately. The given initial conditions for h01 and h02 are loaded into the SHOCK process, and the initial conditions for the MONGREL h01,h02 are set to zero. Type [Amm,Bmm,Cmm,Dmm]=minreal(Am,Bm,Cm,Dm) to find minimal realization for preference shock.

Matlab programs

461

doubleo

Purpose: Computes time invariant Kalman filter or time invariant linear optimal control. Synopsis: [K, S] = double (A, C, Q, R) Description: The program creates the Kalman filter for the following system: xt+1 = Axt + et+1 yt = Cxt + vt where Eet+1 e′t+1 = Q, Evt vt′ = R , and vt is orthogonal to et for all t and s . Here A is n × n , C is k × n , Q is n × n , and R is k × k . The program creates the observer system x ˆt+1 = Aˆ xt + Kat yt = C x ˆt + at , where K is the Kalman gain, and S = E(xt − x ˆt )(xt − x ˆt )′ where x ˆt = Ext | yt+1 yt−2 , . . . . Also, at = yt − Eyt | yt−1 , yt−2 , . . . . By using duality, the program can be used to solve optimal linear control problems. Let the control problem be to choose a feedback law ut = −F xt to maximize ∞ X {x′t Qxt + u′t Rut } − t=0

subject to

xt+1 = A′ xt + B ′ ut , with x0 given. The optimum control is then given by F = K ′ , where [K, S] = double (A, B, Q, R) and where the optimal value function is x′t Sxt . The doubling algorithm is used to compute the solution. See also: mult and double3. References: [1] Anderson, B.D.O., and J. Moore, Optimal Filtering, 1979, p. 160.

462

MATLAB Programs

doublex

Purpose: Solves recursive undiscounted Gaussian Quadratic Exponential control problem. Synopsis: [K, S, ST ] = doublex (A, C, Q, R, c, sig) Description: This program uses the “doubling algorithm” to solve the Riccati matrix difference equations associated with the undiscounted quadratic-Gauusian linear optimal control problems. The control problem has the form S(t) = max{x(t)′ Qx(t) + u(t)′ Ru(t) + (2/σ) log Et exp(σ/2)S(x(t + 1))}, u(t)

subject to x(t + 1) = A′ x(t) + C ′ u(t) + cw(t + 1), where w(t+1) is a Gaussian martingale difference sequence with unit covariance matrix. The program returns the steady state value function in S . The optimal control law is u(t) = −K ′ ∗ x(t) The program also returns ST , which is the quadratic form in Et exp(sig/2)S(x(t + 1)). See also: mult and double and

solvex.

References: [1] Anderson, B.D.O., and J. Moore, Optimal Filtering, 1979, p. 160. [2] Jacobson, D.H. “Optimal stochastic linear systems with exponential performance criteria and their relation to deterministic differential games.”IEEE Transactions on Automatic Control, AC-18, 124–31.

Matlab programs

463

doublej

Purpose: Computes infinite matrix sums of squares. Synopsis: V = double (a1 , b1 ) Description: The program computes the infinite sum V in V =

∞ X

aj1 b1 aj′ ,

j=0

where a1 and b1 are each n × n matrices. The program iterates to convergence on the following doubling algorithm, starting from V0 = 0: a1j = a1j−1 ∗ a1j−1

Vj = Vj−1 + a1j−1 ∗ Vj−1 ∗ a1j−1 .

The limiting value of Vj is returned in V . References: [1] Hansen, Lars P. and Thomas J. Sargent Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988.

464

MATLAB Programs

doublej2

Purpose: Computes infinite matrix sums of squares. Synopsis: [V ] = doublej2 (a1 , b1 , a2 , b2 ) Description: The program computes the infinite sum V in V =

∞ X

aj1 (b1 b2 )aj2

j=o

where a1 and a2 are each n × n matrices, b1 is n × k and b2 is k × n . The program iterates to convergence on the following doubling algorithm, starting from V0 = 0: a1j = a1j−1 ∗ a1j−1 a2j = a2j−1 ∗ a2j−1

Vj = Vj−1 + a1j−1 Vj−1 a2j−1 .

The limit point is returned in V . References: [1] Hansen, Lars P. and Thomas J. Sargent Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988.

Matlab programs

465

double3

Purpose: Raw doubling algorithm for raising a symplectic matrix to higher and higher powers. Synopsis: [aa, bb, gg] =

double3

(a, b, g)

Description: The algorithm iterates to convergence of gj in the following recursions: aj+1 = aj (I + bj gj )−1 gj gj+1 = gj + a′j gj (I + bj gj )−1 aj , bj+1 = bj + aj (I + bj gj )−1 bj a′j where aj , bj , gj are each n × n matrices. If we let Ej , be the symplectic matrix

a−1 j gj a−1 j

a−1 j bj ′ aj + gj a−1 j bj

j

then Ej = (E0 )2 . References: [1] Anderson, B.D.O., and J. Moore, Optimal Filtering, 1979, p. 160. [2] Hansen, Lars Peter and T.J. Sargent, Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988.

466

MATLAB Programs

heter

Purpose: Computes allocation to an individual who lives within a recursive linear equilibrium model. Synopsis: heter is a script file. The program solvea must be run first, and its inputs and outputs must be in memory. The matrices Udc and Ubi , and the scalars k0i, h0i, and tol > 0 must all be in memory. Description: The consumer maximizes ∞

1 X t i −( )E β [(st − bit ) · (sit − bit ) + ℓ2t ], 0 < β < 1 2 t=o subject to sit = Λhit−1 + Πcit hit = ∆h hit−1 + Θh cit ∞ ∞ X X β t p0t cit | Io = E β t (wt0 ℓit + αt0 dit ) | Io E t=0

bit dit

t=0

i + υ0 k−1 = Ubi zt = Udi zt

i where k−1 = k0i, hi−1 = h0i are parameters to be fed in. The matrices Udi = udi and Ubi = ubi must also be fed in. The parameter tol > 0 must be fed in. The program computes the optimal solution for consumer i in the form cit = Sci xt , hit = Shi xt , sit = Ssi xt , bit = Sbi xt , dit = Sdi xt , where xt is the state variable of the economy augmented by i the state variables kt−1 , hit−1 ideosyncratic to the individual. The program also computes the aggregate allocations ct = Sca xt , ht = Sha xt , and so on. The individual allocations are determined by the matrices sci, shi, . . . , which are placed in memory. The aggregate allocation are placed in the matrices sca, sha, . . . , which are placed in memory.

Algorithm:

Matlab programs

See Hansen and Sargent, Chapter 6 See also: simulh

467

468

MATLAB Programs

innov

Purpose: Compute the innovations representation for a recursive linear model whose observations are corrupted by first-order serially correlated measurement errors. Synopsis: [aa, bb, cc, dd, V1 ] = innov (ao, c, sy, D, R ) Description: The model is assumed to have the state space representation xt+1 = ao xt + cwt+1 yt = Sy xt + et+1 where wt is a white noise with Ewt wt ′ = I and et is a measurement error process governed by et+1 = Det + ηt+1 where ηt is a white noise with contemporaneous covariance matrix R . The matrices R and D must each be m × m where [m, n] = size(Sy ). The program forms the innovations representation for yt , zˆt+1 = aaˆ zt + bbut yt = ccˆ zt + ddut where ut = yt+1 − E[yt+1 | yt , yt−1 , . . .], and Eut u′t = V1 . Algorithm:

ao aa = GG

0 k1 , bb = D I

cc = [0 I], dd = [0], where k1 is the Kalman gain associated with the Kalman filter for the original system. References: [1] Sargent, Thomas, “Two Models of Measurements and the Investment Accelerator,” Journal of Political Economy, April 1989.

Matlab programs

mult

Purpose: Multiplies two symplectic matrices. Synopsis: [a, b, g] = mult (a1 , b1 , g1 , a2 , b2 , g2 ) Description: A symplectic matrix Ei is represented in the form (∗)

Ei =

a−1 i gi a−1 i

a−1 i bi . a′i + gi a−1 i bi

We desire to form E = E2 E1 . We can compute a = a2 (I + b1 g2 )−1 g1 g = g1 + a′1 g2 (I + b1 g2 )−1 a1 b = b2 + a2 (I + b1 g2 )−1 b1 a′2 ,

and represent E as in representation (∗). References: [1] Anderson, B.D.O., and J. Moore, Optimal Filtering, 1979, p. 160.

469

470

MATLAB Programs

seasla

Purpose: Creates a time invariant representation for a strictly periodic, time varying linear equilibrium model Synopsis: seasla is a script file, which requires that the output of

simuls

reside in memory.

Description: Let xt be the state vector for a strictly periodic seasonal process of period p . Let Xt′ = [x′pt−p+1 , x′pt−p+2 , . . . , x′pt ]. The law of motion for Xt is ˆ t + CW ˆ t+1 Xt+1 = AX where Wt+1 is a vector white noise and Aˆ and Cˆ are defined as simuls. The ˆ −1 Cˆ Cˆ ′ (I − spectral density matrix of the Xt process is given by S(z) = (I − AZ) ˆ −1′ . Embedded in the spectral density matrix of the stacked process Xt AZ) are the spectral density matrices s1 (z), s2 (z), . . . , sp (z) for the periodic process {xt }. The process xt whose spectral density is defined to be s(z) = Pp p−1 k=1 sk (z). It can be shown that (∗)

ˆ p )−1 Cˆ Cˆ ′ (I − Az ˆ −p )−1′ , s(z) = p−1 Q(z)(I − Az

where Q(z) = [I zI · · · z p−1 I]. A state space representation for a process xt with spectral density matrix (*) is (†)

ˆ t + CV ˆ t+p Yt+p = AY xt = p−.5 Q(L)Yt

where Vt is a vector white noise with identity covariance matrix. The program seasla creates the spectral density for a univariate process that is a linear function of the state. Let the process be ct = scs(t) xt . We form the time invariant, averaged process, as c˜t which is determined by the system (‡)

ˆ t + CV ˆ t+p Yt+p = AY ct = p−1 Qc (L)Yt

. . . where Qc (z) = [sc1 .. sc2 z .. · · · ..scp z p−1 ].

Matlab programs

471

The program maps into a first-order system, then deduces the impulse response of ct with respect to innovations Vt corresponding to representation (‡). This is stored in z1 . The program also uses the Kalman filter to obtain an innovations representation corresponding to (‡), and returns the impulse response of ct with respect to the innovation in ct in the vector z2 . See also: simuls, assets,

assetss

472

MATLAB Programs

seas1

Purpose: To aid in creating the matrices that define a periodic recursive linear equilibrium model. Synopsis: seas1.m is a script file. Description: seas1 creates matrices Φcs(t) , Φgs(t) , Φis(t) , Γs(t) , ∆ks(t) , ∆hs(t) , A22s(t) , Cs(t) , Φks(t) , Φhs(t) , Λs(t) , and Πs(t) that are needed to define a periodic linear recursive model. It creates time invariant versions of these matrices as follows. It first reads in Φc , Φi , Φy , Γ, ∆k , ∆h , A22 , C, Φk , Φh , Λ1 and Π for a time invariant economy. One of our clex*.m files can be used to read in such matrices. Then seas1 simply sets the matrices Φcs(t) = Φc , and so on. To create a periodic model, the user may find it useful to run seas1 first, and then to modify the resulting time invariant setup, rather than building up all of the matrices from scratch. In a typical periodic model, many of the matrices may in fact be time invariant. See also: solves.m

Matlab programs

473

simpulse

Purpose: Creates different impulse response functions for a periodic linear equilibrium model. Synopsis: simpulse is a script file. memory.

solves

must be run first, and its outputs must be in

Description: A stacked version of a periodic model has state space form (†)

ˆ t + CW ˆ t+1 Xt+1 = AX Yt = HXt ,

′ ′ ′ where Xt′ = [x′tp−p+1 , x′tp−p+2 , . . . , x′tp ], Yt′ = [ytp−p+1 , . . . , ytp ], Wt′ = [wtp−p+1 , ′ ˆ ˆ . . . , wtp ] and where A, C , are as defined in simuls.

This program first uses dimpulse to compute the impulse response function of the stacked system (†). From this impulse response function, it forms two impulse response functions for the periodic process yt . First, it computes {dk,υ } in the representation ypt−p+k =

∞ X

dk,υ wpt−p+k−υ .

υ=0

This is the response of ypt−p+k (i.e., yt in a particular season) to lagged w ’s. Second, the program computes the {hk,υ } that give the response of {yt } to wpt−p+k (i.e. an innovation in a particular season). The value of p must be in memory. The program prompts the user for the index of the innovation whose response functions are to be computed. See also: solves, simuls

474

MATLAB Programs

simulh

Purpose: Simulates allocation of individual i who lives within a recursive linear equilibrium model. Synopsis: simulh is a script file.

heter

must be run first and its output must be in memory.

Description: The user is asked to specify which series he wants to simulate; e.g., to simulate the consumption allocation to agent i and the aggregate consumption allocation, respond [sci; sca]. See also: heter

Matlab programs

475

simulhet

Purpose: Simulates heterogeneous agent economy. Synopsis: simulhet is a script file, not a function. Description: Simulates the prices and quantities for a recursive linear equilibrium model with non-Gorman heterogeneity. solvehet must be run first and its output must be in memory. To simulate the individual consumption allocations, set sy=[sc1;sc2] when asked what series you want to simulate. To simulate the individual consumption service allocations, set sy= [ss1;ss2]. See also: solvea and heter.

476

MATLAB Programs

simuls

Purpose: To simulate a strictly periodic recursive linear model. Synopsis: simuls is a script file, which requires that all of the outputs of solves be in memory. simuls prompts the user for the number of “years” to simulate. Description: simuls creates a simulation of the state vector xt for a strictly periodic model of period p . The stacked state vector Xt is formed, where Xt′ = [x′pt−p+1 , x′pt−p+2 , . . . , x′pt ]. The law of motion for Xt is ˆ t + CW ˆ t+1 Xt+1 = AX where Aˆ = D−1 F, Cˆ = D−1 G , where I −A01 D= 0 . .. 0

0 I −A02 .. .

0 0 I .. .

··· ··· ··· .. .

0 0 0 .. .

0

0

···

−A0p−1

0 F = 0

A0p 0

Cp 0 G = .. .

0

Wt′

1

0 C1 .. .

0 ··· 0 ··· .. . . . .

0 0 .. .

0

0 ···

Cp−1

′ ′ ′ ]′ . [wpt+1 , wpt+2 , . . . , wpt+p

0 0 0 .. .

,

= and where The output of simuls is returned in the matrix X . The matrix X is arranged as follows: x′1 x′2 ··· x′p x′p+1 x′p+2 · · · x′2p X= . .. .. .. .. . . . ′ ′ ′ xT p+1 xT p+2 · · · xT p+p

Matlab programs

477

where T is the number of “years” specified by the user. To simulate ct , it , etc., the user can write a program to put the relevant linear combinations off X . Alternatively, the user can edit the files simulc, simulk, simuli, simulg, simulb, or simuld.

478

MATLAB Programs

solvea

Purpose: Computes solution of recursive linear equilibrium model Synopsis: solvea is a script file. The matrices A22 , C2 , Ud , Ub , Φc , Φg , Φi , Γ, ∆k , Θk , ∆h , Θh , Λ, and Π and the scalar β must be in memory. Description: The social planning problem is to maximize ∞

1 X t −( )E β [(st − bt ) · (st − bt ) + ℓ2t ], 0 < β < 1 2 t=0 subject to Φc ct + Φg gt + Φi it = Γkt−1 + dt gt · gt = ℓ2t

kt = ∆k kt−1 + Θk it

ht = ∆h ht−1 + Θh ct st = Λht−1 + Πct zt+1 = A22 zt + C2 wt+1 bt = Ub zt , dt = Ud zt Here st is consumption services, bt a stochastic bliss process, ℓt is labor services, ct is consumption rates, gt is “intermediate goods”, it is investment goods, dt is an endowment shock process, kt is physical capital, ht is household capital, zt is a vector of exogenous information variables, and wt+1 is a martingale difference sequence. Each of these is a vector, except for ℓt , which is scalar. Let ′ xt = [h′t−1 , kt−1 , zt′ ]′ . The program computes the solution of the social planning problem in the form xt+1 = Ao xt + Cwt+1 kt = Sk xt , gt = Sq xt ht = Sh xt , it = Si xt ct = Sc xt , bt = Sb xt st = Ss xt , dt = Sd xt

Matlab programs

479

The program also computes Lagrange multipliers µjt = Mj xt for variable j = k, c, h, s, i. The program computes and leaves in memory Ao , C, Sj (for j = k, h, c, s, g, i, b, and d ) and Mj (for j = k, c, h, s, and i). Algorithm: The social planning problem is formulated and solved as an optimal linear regulator problem. References: [1] Hansen, Lars Peter and T.J. Sargent, Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988.

480

MATLAB Programs

solvdist

Purpose: Computes equilibrium of representative agent economy with distorting taxes, exogenous govenment expenditures, and externalities. Synopsis: solvdist is a script file. Description: solvdist, a script file (not a function), finds a competitive equilibrium for a representative agent economy with distortions. Households maximize: X E0 − .5 β t [(s(t) − b(t)) · (s(t) − b(t)) + ℓ(t)2 ] subject to g(t) · g(t) = ℓ(t)2

z(t + 1) = a22 ∗ z(t) + c2 ∗ w(t + 1) b(t) = ub ∗ z(t),

d(t) = ud ∗ z(t),

E(t) = ue ∗ z(t)

h(t) = ∆h h(t − 1) + ∆H H(t − 1) + Θh c(t) + θH C(t) s(t) = Λh h(t − 1) + ΛH H(t − 1) + Πc c(t) + ΠC C(t)

∞ X t=0

k(t) = ∆k k(t − 1) + Θk i(t)

β t [(I + τc )p(t) · c(t) + (I + τi )q(t) · i(t) − (1 − τℓ )w(t) · g(t)

−α(t) · (d(t) + γK K(t − 1)) − (I − τk )r(t) · k(t − 1) − T (t)] = 0 Firms maximize profits: ∞ X β t [p(t) · (c(t) + E(t)) + q(t) · i(t) − r(t) · k(t − 1) − α(t) · d(t) − w(t) ∗ g(t)] E0 t=0

subject to g(t) · g(t) = ℓ(t)2

Φc (c(t) + E(t)) + Φg g(t) + Φi i(t) = Γk k(t − 1) + ΓK K(t − 1) + d(t)

Where x(t) = [h(t − 1)′ , k(t − 1)′ , z(t)′ ]′ , the solution of the problem is x(t + 1) = ao ∗ x(t) + c ∗ w(t + 1)

j(t) = sj ∗ x(t), where j = k, h, c, e, s, g, i, b, d, p, q, w, r, α . The program also computes the household’s Lagrange multipliers µj = mj x(t) for j = k, h, s, z . (µ0 is set to 1.)

Matlab programs

481

solvehet

Purpose: Solves Pareto problem for two-agent, non-Gorman preferences. Description: solvehet solves the pareto problem for two agents with heterogeneous household production functions, i.e. to maximize E

∞ X t=0

β t − .5α([(s1 (t) − b1 (t)) · (s1 (t) − b1 (t)) + ℓ1 (t)2 ]) + (1 − α)([(s2 (t) − b2 (t)).(s2 (t) − b2 (t)) + ℓ2 (t)2 ])

subject to Φc c(t) + Φg g(t) + Φi i(t) = Γk(t − 1) + d(t) gi (t) · gi (t) = ℓi (t)2 ,

i = 1, 2

g1 (t) + g2 (t) = g(t)

k(t) = ∆k k(t − 1) + Θk i(t)

hi (t) = ∆hi hi (t − 1) + Θhi ci (t)

si (t) = Λi ∗ hi (t − 1) + Πhi ci (t),

i = 1, 2

c1 (t) + c2 (t) = c(t)

z(t + 1) = a22 ∗ z(t) + c2 ∗ w(t + 1) i(t) = ubi ∗ z(t),

i = 1, 2; d(t) = ud ∗ z(t)

The state vector is x(t) = [h1(t-1)’,h2(t-1)’,k(t-1)’,z(t)]’. The control vector is u(t) = [c1(t)’,i(t)’]. The solution of the problem is given by: x(t + 1) = ao ∗ x(t) + c ∗ w(t + 1) j(t) = sj x(t)

for j = k,c,g,i,d, and ci,bi,gi,hi,si, for i=1,2. The program also computes Lagrange multipliers.

482

MATLAB Programs

solvex

Purpose: Computes solution of recursive linear equilibrium model with Gaussian Quadratic Exponential preference specification. Synopsis: solvex is a script file. The matrices A22 , C2 , Ud , Ub , Φc , Φg , Φi , Γ, ∆k , Θk , ∆h , Θh , Λ, and Π and the scalars σ and β must be in memory. Description: ′ Let xt = [h′t−1 , kt−1 , zt′ ]′ , and let the law of motion for xt be xt+1 = Axt + But + Cwt+1 . The social planning problem is to find a value function V (x(t)) = max{−.5[(s(t) − b(t)).(s(t) − b(t)) + l(t)2 ] +β ∗ (2/σ) ∗ log Et exp(σ/2 ∗ (V (x(t + 1))}

subject to Φc ct + Φg gt + Φi it = Γkt−1 + dt gt · gt = ℓ2t

kt = ∆k kt−1 + Θk it

ht = ∆h ht−1 + Θh ct st = Λht−1 + Πct zt+1 = A22 zt + C2 wt+1 bt = Ub zt , dt = Ud zt Here st is consumption services, bt a stochastic bliss process, ℓt is labor services, ct is consumption rates, gt is “intermediate goods”, it is investment goods, dt is an endowment shock process, kt is physical capital, ht is household capital, zt is a vector of exogenous information variables, and wt+1 is a martingale difference sequence. Each of these is a vector, except for ℓt , which is scalar. The program computes the solution of the social planning problem in the form xt+1 = Ao xt + Cwt+1 kt = Sk xt , gt = Sq xt ht = Sh xt , it = Si xt ct = Sc xt , bt = Sb xt st = Ss xt , dt = Sd xt

Matlab programs

483

The program also computes Lagrange multipliers µjt = Mj xt for variable j = k, c, h, s, i. The program computes and leaves in memory Ao , C, Sj (for j = k, h, c, s, g, i, b, and d) and Mj (for j = k, c, h, s, and i). Algorithm: The social planning problem is formulated and solved using

doublex.

References: [1] Hansen, Lars Peter and T.J. Sargent, Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988.

484

MATLAB Programs

solves

Purpose: Computes the solution of recursive linear equilibrium model with periodic coefficients. Synopsis: solves is a script file. The matrices A22s(t) , C2s(t) , Ud , Ub , Φcs(t) , Φgs(t) , Φis(t) , Γs(t) , ∆ks(t) , Θks(t) , ∆hs(t) , Θhs(t) , Λs(t) , and Πs(t) and the scalar β must be in memory.

Description: The social planning problem is to maximize ∞

1 X t −( )E β [(st − bt ) · (st − bt ) + ℓ2t ], 0 < β < 1 2 t=0 subject to Φcs(t) ct + Φgs(t) gt + Φis(t) it = Γs(t) kt−1 + dt gt · gt = ℓ2t

kt = ∆ks(t) kt−1 + Φks(t) it ht = ∆hs(t) ht−1 + Φhs(t) ct st = Λs(t) ht−1 + Πs(t) ct zt+1 = A22s(t) zt + C2s(t) wt+1 bt = Ub zt , dt = zt . where s(t+p) = s(t), where p is the period of the model. Here st is consumption services, bt is a stochastic bliss process, ℓt is labor services, ct is a vector of consumption rates, gt is “intermediate goods”, it is investment goods, dt is an endowment shock process, kt is physical capital, ht is household capital, zt is a vector of exogenous information variables, and wt+1 is a martingale difference sequence. Each of these is a vector, except for ℓt , which is a scalar. ′ Let xt = [h′t−1 , kt−1 , zt′ ]′ . The program computes the solution of the social

Matlab programs

485

planning problem in the form xt+1 = Aos(t) xt + Cs(t) wt+1 kt = Sks(t) xt , gt = Sgs(t) xt ht = Shs(t) xt , it = Sis(t) xt ct = Scs(t) xt , bt = Sbs(t) xt st = Sss(t) xt , dt = Sds(t) xt The program also computes Lagrange multipliers µjt = Mjs(t) xt for variables j = k, c, h, s, i. The program computes and leaves in memory Aos(t) , Cs(t) , Sjs(t) for j = k, h, c, s, g, i, b and d , and Mj for j = k, c, h, s, and i. The user is advised to use the Matlab program seas1 an an aid in creating the matrices that must be fed into solves. The user must edit solves to set the period p . Also, it will vastly accelerate computations if the user will load either the file seas4.mat (in the case p = 4) or the file seas12.met (in the case p = 4). The lines to edit occur immediately after the information provided by the help command, i.e. the first lines without %. Algorithm: The social planning problem is formulated as a periodic optimal linear regulator problem and solved using doubling algorithms described by Hansen and Sargent. References: [1] Hansen, Lars Peter and T.J. Sargent, Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988.

486

MATLAB Programs

spectr1

Purpose: Computes spectral density of endogenous variables of a dynamic linear equilibrium model. Synopsis: spectr1 is a script file. The matrices ao, c, sy, R, D , and the scalar nnc must be in memory. Description: The equilibrium model is of the form xt+1 = ao xt + cwt+1 yt = sy xt + vt vt+1 = Dvt + ut+1 where Ewt wt′ = I, E, ut u′t = R . The constant corresponds to row number nnc of the state vector xt . The eigenvalues of D and the eigenvalues of A (except for the unit eigenvalue associated with the constant term) must be less than unity in modulus. spectr1 computes the spectral densities variables in yt . Algorithm: spectr1 deletes the nncth row and/or column of ao, c, and sy , which correspond to the constant term. Then spectral is used to compute the spectral density matrix of yt . See also: spectral

Matlab programs

487

spectral

Purpose: Computes spectral density matrix for a linear system. Synopsis: spectral, a script file. The inputs A, C, G, D, R and T must reside in memory. Description: Let the system be xt+1 = Axt + Cet+1 yt = Gxt + vt vt+1 = Dvt + ut Eet e′t

I, Eut u′t

where = = R , and where et and us are orthogonal for all t and s . The vector yt is rg × 1. The spectral density matrix for y is computed for ordinates ωj = 2πj/T, j = o, 1, . . . , T − 1. The spectral density matrix for ordinate j is stored in Syj, j = 0, 1, . . . , T −1. The spectral densities (diagonals of the spectral density matrices) are stored in the matrix S . The matrix S has rg rows and T columns, and S(k, j) = Syj(k, k). The eigenvalues of A and D must all be less than unity in modulus. Algorithm: Let Sy(ωj ) be the spectral density matrix at frequency wj . Then Sy(ωj ) = G(I − Ae−iω−j )−1 CC ′ (I − Aeiωj )−1 G′ + (I − De−iωj )−1 R(I − De+iωj )−1

.

488

MATLAB Programs

spectrs

Purpose: Computes spectral density matrix for set of variables determined by a periodic linear equilibrium model. Synopsis: spectrs is a script file. solves and simuls must be run first, and their outputs must reside in memory. The integer nnc (the index of the constant in the state vector) must be in memory. Description: The spectral density of a process yt with hidden periodicity p is given by the Tiao-Grupe formula ˆ p )−1 Cˆ C(I ˆ − Az ˆ −p )−1′ H ′ Q(z −1 )′ , Sy (z) = Q(z)H(I − Az ˆ Cˆ , and H ˆ are from the stacked state space system where z = e−iωj ; where A, ˆ t + CW ˆ t+1 Xt+1 = AX Yt = HXt , ′ ′ and where Xt′ = [x′pt−p+1 , x′pt−p+2 , . . . , x′pt ], Yt′ = [ypt−p+1 , ypt−p+2 , ′ ′ ′ ′ ′ . . . , ypt ], Wt = [wpt−p+1 , wpt−p+2 , . . . , wpt ] . The program returns the spectral density matrices for frequencies ωj = 2πj/T , for j = 0, 1, . . . , T − 1 in the matrices Sy0, Sy1, . . . , SyT − 1. The spectral densities of the individual series are returned in the matrix S . The user can edit the file to specify T and the particular series whose spectrum is computed.

See also: solves, simuls

Matlab programs

489

steadst

Purpose: steadst computes steady state values of observable variables determined by a recursive linear equilibrium model. Synopsis: steadst is a script file which requires that the scalar nnc and matrices ao, sc, ss, si, sd, sb, sk, sh reside in memory. Description: The equilibrium model is represented as xt+1 = ao ∗ xt + c ∗ wt+1 yt = Gxt

where G = [sc; ss; si; sd; sb; sk; sh]. The integer nnc gives the row in the state vector xt that corresponds to the constant term. steadst assumes that except for the eigenvalue associated with the constant term, all eigenvalues of ao are less than unity in modulus. The program calculates the steady state value of xt , putting its value in zs . Then the program successively calculates the steady state values of c, s, i, d, b, k, and h , which are the components of y . Algorithm: The steady state value of x is obtained as a basis vector for the null space of (I −ao), normalized so that the component corresponding to the constant equals unity. References: [1] Hansen, Lars Peter and T.J. Sargent, Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988. See also: null

490

MATLAB Programs

steadsts

Purpose: Computes seasonal steady states and seasonal means for a periodic recursive linear equilibrium model. Synopsis: steadsts is a script file. solves and simuls must be run first, and their outputs must be in memory. So must nnc, the index of the constant term in the state vector. Description: The equilibrium for the stacked version of a periodic model can be represented as ˆ t + CW ˆ t+1 Xt+1 = AX where Xt′ = [x′tp−p+1,..., x′tp ], Wt′ = [wpt−p+1,..., wpt ]. The program computes the ˆ which gives the steady for Xt = X ¯ . Then seasonal means null space of (I − A), ¯ by matrices formed for individual variables are formed by pre-multiplying X from appropriate seasonal decision rules. The user must edit the file to compute seasonal means of the particular variables he is interested in. See also: steadst, solves,

simuls.

Matlab programs

491

vardec

Purpose: Calculates variance of k -step ahead prediction errors in zt for k = 1, 2, . . . , N for an “innovations system”. Synopsis: [tab] = vardec (A, C, K, V, N ) Description: Consider the innovations system xt+1 = Axt + Kut zt = Cxt + ut Eut u′t = V vardec prepares a table of diagonal elements of the covariance matrices of k -step ahead errors in predicting zt , k = 1, . . . , N . The output is returned in tab, which has N rows and max(size( V )) columns. The (k, h) element of tab gives the variance of the k -step ahead prediction errors for the hth variable in zt .

Algorithm: Let the covariance matrix of k -step ahead prediction error in z be Vk . Then V1 = V V2 = CKV K ′ C ′ + V Vk = Vk−1 + CAk−1 KV K ′ Ak−1 C ′ . References: [1] Sims, Christopher “Macroeconomics and Reality,” Econometrica, 1980. [2] Hansen, Lars Peter and Thomas Sargent, Recursive Linear Models of Dynamic Economies, (manuscript), Dec. 1988.

492

MATLAB Programs

vardeci

Purpose: Compute decomposition of k -step ahead prediction error variances for an “innovations system”. Synopsis: [tab ] = vardeci (A, C, K, V, N, j) Description: Consider an innovations system xt+1 = Axt + Kut zt = Cxt + ut ′

where Eut ut = V. Let r′ r = V be a Cholesky decomposition of V . Form the innovations system with orthogonalized innovations xt+1 = Axt + Bvt zt = Cxt + Dvt where B = K · r′ , D = r′ , and Evt vt′ = I . The program prepares a table of the part of the diagonal elements of the covariance matrix of the k -step ahead prediction errors, k = 1, . . . , N , that is attributable to the j th innovation. The table is returned in tab , which has dimension N × max(size(V )). The (k, h) element of tab gives the variance in the k -step ahead variance in predicting the hth component of z due to the j th orthogonalized innovation in vt . Algorithm: Let Sj be a selector matrix for j , equal to an m × m matrix of zeros except of a one in the (j, j) element. Let Vk be the covariance of the k -step ahead prediction error in z due the jth orthogonalized innovation. The Vk are calculated using the ecursions V1 = DSj Sj′ D′ V2 = CBSj Sj′ B ′ C ′ + V1 Vk = Vk−1 + CAk−1 BSj Sj′ B ′ A′

k−1

C′

References: [1] Sims, Christopher “Macroeconomics and Reality,” Econometrica, 1980. [2] Hansen, Lars Peter and Thomas Sargent, Recursive Linear Models of Dynamic Economies.

Matlab programs

493

varma

Purpose: varma computes an innovations representation for a recursive linear model whose observations are corrupted by first-order serially correlated measurement errors. Synopsis: varma is a script file, which requires that the matrices ao, c, sy, D , and R reside in memory. Description: The model is assumed to have the state space representation xt+1 = ao ∗ xt + c ∗ wt+1 yt = Sy ∗ xt + et+1

where wt is a white noise with Ewt wt ′ = I, and et is a measurement error process governed by et+1 = Det + ηt+1 , where ηt+1 is a vector white noise with contemporaneous covariance matrix R . The matrices R and D must each be m × m , where [m, n] = size(sy). The program uses the Kalman filter to form the innovations representation x ˆt+1 = aoˆ xt + k1 ∗ ut y˜t+1 = GGxt + ut

where GG = [syao − DSy ], y˜t = yt+1 − Dyt , and ut is the innovation in yt+1 , ut = yt+1 − E[yt+1 | yt , yt−1 , . . .]. The program uses evardec to compute a decomposition of variance for the innovations system. References: [1] Hansen, Lars Peter and T.J. Sargent, Recursive Linear Models of Dynamic Economies, manuscript, Dec. 1988. [2] Sargent, Thomas, “Two Models of Measurements and the Investment Accelerator,” Journal of Political Economy, April 1989.

494

MATLAB Programs

varma2

Purpose: varma2 creates impulse response functions associated with an innovations representation. Synopsis: varma2 is a script file which requires that the matrices aa, bb, cc, dd, V 1 be in memory. Description: varma2 takes the output of innov and creates impulse response functions of y with respect to components of u . Impulse response functions with respect to −1 the orthogonalized innovations vt = r′ ut are also computed, where r′ r = V 1 is a Cholesky decomposition of V1. Algorithm: dimpulse is applied. References: [1] Sims, Christopher, “Macroeconomics and Reality,” Econometrica, 1980.

Matlab programs

495

varrep

Purpose: Computes a vector autoregressive representation from a state space model with serially correlated measurement errors. Synopsis: function [AA,V1]=varrep(ao,c,sy,D,R,nj,nnc) Description: Computes (an infinite order) vector autoregressive representation for a recursive linear model whose observations are corrupted by first-order serially correlated measurement errors. The model occurs in the state space form x(t + 1) = ao ∗ x(t) + c ∗ w(t + 1) y(t) = sy ∗ x(t) + e(t + 1)

where e(t) is a measurement error process e(t + 1) = D ∗ e(t) + ee(t + 1) and where ee(t+1) is a vector white noise with covariance matrix R. We assume that ee(t+1) and w(t+1) are orthogonal at all leads and lags. The program computes the autoregressive representation y(t) =

∞ X j=1

A(j)y(t − j) + a(t)

where a(t)=y(t) - E[y(t)— y(t-1),y(t-2),...], and the A(j) are square matrices. The program creates the covariance matrix of a, which it stores in V1. The program returns nj of the matrices A(j), stacked into the ((m times nj) by m) matrix AA, where m is the number of rows of y. A(j) occurs in rows ((j-1)*m+1) to row j*m of AA. nnc is the location of the constant term in the state vector.

496

MATLAB Programs

white1

Purpose: Creates a state space system [AA,BB,CC,DD] that accepts the innovation to the (information) state vector wt+1 as in input and puts out the innovation ut to yt as an output. Synopsis: function[AA,BB,CC,DD]=white1(ao,c,sy,D,R) Description: The program couples the systems x(t + 1) = ao ∗ x(t) + c ∗ w(t + 1) y(t) = sy ∗ x(t) + v(t)

v(t) = D ∗ v(t − 1) + η(t)

Eη(t)η(t)′ = R. and

xh(t + 1) = (ao − k1 ∗ GG) ∗ xh(t) + k1 ∗ y(t) u(t) = −GG ∗ xh(t) + u(t)

where w(t+1) is the innovation to agents’ information sets and where u(t) is the fundamental (Wold) representation innovation. A (minimum-realization) state space system [AA,BB,CC,DD] for the coupled system is returned. To compute the impulse response function, use dimpulse.

Matlab programs

497

white2

Purpose: Creates the state space system [AA,BB,CC,DD] that accepts the measurement error v(t + 1) as in input and puts out the innovation u(t) to y(t) as an output. Synopsis: function[AA,BB,CC,DD]=white2(ao,c,sy,D,R) Description: The program couples the systems x(t + 1) = ao ∗ x(t) + c ∗ w(t + 1) y(t) = sy ∗ x(t) + v(t)

v(t) = D ∗ v(t − 1) + η(t)

Eη(t)η(t)′ = R. and

xh(t + 1) = (ao − k1 ∗ GG) ∗ xh(t) + k1 ∗ y(t) u(t) = −GG ∗ xh(t) + u(t)

where w(t + 1) is the innovation to agents’ information sets, η(t) is the innovation to measurement error, and where u(t) is the fundamental (Wold) representation innovation. The (minimum-realization) state space system [AA,BB,CC,DD] for the coupled system is returned. To compute the impulse response function, use dimpulse.

References Akaike, H. (1973). ‘Information Theory and an Extension of the Likelihood Principle’. In B.N. Petrov and F. Csaki (eds.), Proceedings of the Second International Symposium of Information Theory. Budapest: Akademiai Kiado. Anderson, B.D.O. (1978). ‘Second-Order Convergent Algorithms for the Steady-State Riccati Equation’. International Journal Control, 28, 295–306. Anderson, Brian D.O. and John B. Moore (1979). Optimal Filtering. Englewood Cliffs: Prentice Hall. Anderson, Evan W., L.P. Hansen, Ellen R. McGrattan, and T.J. Sargent (1995). ‘Mechanics of Forming and Estimating Dynamic Linear Economies’. In John Rust (ed.), Handbook of Computational Economics, forthcoming. Anderson, Gary S. and George R. Moore (1985). ‘A Linear Algebraic Procedure for Solving Linear, Perfect Foresight Models’. Economic Letters, vol. 17, No. 3, 247–252. Ansley, C.F., and R. Kohn (1983). ‘Exact Likelihood of Vector Autoregressive-Moving Average Process with Missing or Aggregated Data’. Biometrika, 70, 275–278. Ansley, C.F., and R. Kohn (1985). ‘Estimation, Filtering, and Smoothing in State Space Models with Incompletely Specified Initial Conditions’. The Annals of Statistics, Vol. 13, No. 4, 1286–1316. Bell, William (1984). ‘Signal Extraction for Nonstationary Time Series’. The Annals of Statistics, Vol. 12, No. 2, pp. 646–664. Barro, R. J. (1977). ‘Unanticipated Money Growth and Unemployment in the United States’. American Economic Review, 67. Barro, R.J. (1979). ‘On the Determination of the Public Debt’. Journal of Political Economy, 87, 940–71. Becker, Gary and Kevin Murphy (1988). ‘A Model of Rational Addiction’. Journal of Political Economy, 96:4, 675–700. Beltrami, E.J. and M.R. Wohlers (1966). Distributions and the Boundary Values of Analytic Functions. Academic Press, New York. Bernanke, B. (1985). ‘Adjustment Costs, Durables, and Aggregate Consumption’. Journal of Monetary Economics, 15, 41–68. Bernanke, B. (1986). ‘Alternative Explorations of the Money-Income Correlation’. In Karl Brunner and Allan H. Meltzer (eds.), Real Business cycles, Real Exchange Rates, and Actual Policies. Carnegie-Rochester Conference Series on Public Policy. 25, 49–99, Amsterdam: North-Holland Bergstrom, A.R. (1976). Statistical Inference in Continuous Time Economic Models. Amsterdam: North-Holland. Bergstrom, A.R. (1983). ‘Gaussian Estimation of Structural Parameters in Higher Order Continuous Time Dynamic Models’. Econometrica, 51, 117–152. Birchenhall, C.R., R.C. Bladen-Hovell, A.P.L. Chui, D.R. Osborne, and J.P. Smith (1989). ‘A Seasonal Model of Consumption’. Economic Journal, Vol. 99, No. 397, September, pp. 837–843. Blanchard, Olivier J., and C.M. Kahn (1980). ‘The Solution of Linear Difference Models Under Rational Expectations’. Econometrica, 48: pp. 1305–11.

– 499 –

500

References

Blinder, A.S. and A. Deaton (1985). ‘The Time Series Consumption Function Revisited’. Brookings Paper on Economic Activity, 2, 465–511. Braun, R. Anton (1991). ‘The Dynamic Interaction of Distortionary Taxes and Aggregate Variables in Postwar U.S. Data’. Mimeo. University of Virginia. Breeden, D.T. (1979). ‘An Intertemporal Asset Pricing Model with Stochastic Consumption and Investment Opportunities’. Journal of Financial Economics, 7, 265–296. Breiman, Leo (1968). Probability. Addisson-Wesley Publishing Company, Reading, Massachusetts. Brock, William A. (1982). ‘Asset Prices in a Production Economy’. In The Economics of Information and Uncertainty (ed.), John J. McCall Chicago: University of Chicago Press, pp. 1–43. Brock, William A., and Leonard J. Mirman (1972). ‘Optimal Economic Growth and Uncertainty: the Discounted Case’. Journal of Economic Theory, 4: 479–513. Brock, W.A. and M.J.P. MaGill (1979). ‘Dynamics Under Uncertainty’. Econometrica, 47, 843–868. Burdick, Clark, Lars Hansen and T.J. Sargent (1990). ‘Statistics for Detecting Hidden Periodicity as a Source of Seasonality’. Mimeo. forthcoming. Burridge, Peter and Kenneth F. Wallis (1990). ‘Seasonal Adjustment and Kalman Filtering: Extension to Periodic Variances’. Journal of Forecasting, vol. 9, pp. 109-118. Campbell, J.Y. (1986). ‘Bond and Stock Returns in a Simple Exchange Model’. Quarterly Journal of Economics, 101, 785–803. Campbell, J.Y. (1987). ‘Does Saving Anticipate Declining Labor Income? An Alternative Test of the Permanent Income Hypothesis’. Econometrica, 55, 1249–1273. Campbell, J.Y. and R.J. Shiller (1987). ‘Cointegration and Tests of Present Value Models’. Journal of Political Economy, 95, 1062–1088. Campbell, J.Y. and R.J. Shiller (1988). ‘Dividend-Price Ratios and Expectations of Future Dividends and Discount Factors’. Review of Financial Studies, 1, 195–228. Cassing, Shirley and Tryphon Kollintzas (1991). ‘Recursive Factor of Production Interrelations and Endogenous Cycling’. International Economic Review, Vol. 32, No. 2, pp. 417–440. Christiano, L.J. (1980). ‘Notes on Factoring Continuous Time Rational Spectral Densities’. Mimeo. Federal Reserve Bank of Minneapolis. Christiano, L.J. (1984). ‘The Effects of Aggregation over Time on Tests of the Representative Agent Model of Consumption’. Mimeo. University of Chicago Working Papers in Economics and Econometrics, 84–15. December. Christiano, L.J. and M.S. Eichenbaum (1985). ‘A Continuous Time, General Equilibrium, Inventory-Sales Model’. Mimeo. COMPLETE REFERENCE. Christiano, L. J. and M.S. Eichenbaum (1986). ‘Temporal Aggregation and Structural Inference in Macroeconomics’. Carnegie-Rochester Conference on Public Policy, Vol. 26. Christiano, L.J., M.S. Eichenbaum, and D. Marshall (1991). ‘The Permanent Income Hypothesis Revisited’. Econometrica, Vol. 59, No. 2, pp. 397–423. Churchill, R.V., J.W. Brown and R.F. Verhey (1974). Complex Variables and Applications. 3rd edition, New York: McGraw Hill. ´ Cleveland, W.P., and G.C. Tiao (1979). ‘Modeling Seasonal Time Series’. Economie Appliqu´ ee, Tome-XXXII, No. 1, pp. 107–129.

References

501

Cochrane, J.H. (1988). ‘How Big is the Random Walk in GNP?’. Journal of Political Economy, 96, 893–920. Coddington, E. A. and N. Levinson (1955). Theory of Ordinary Differential Equations. New York: McGraw-Hill. Coleman, Wilbur J. II. (1990). ‘Solving the Stochastic Growth Model by Policy Function Iteration’. Journal of Business and Economic Statistics, January, Vol. 8, No. 1, pp. 27–29. Constantinides, G.M. (1990). ‘Habit Formation: A Resolution of the Equity Premium Puzzle’. Journal of Political Economy, 98, 519–543. Cox, John C., Jonathan E. Ingersoll, and Stephen A. Ross (1985). ‘An Intertemporal General Equilibrium Model of Asset Prices’. Econometrica, 53: 363–84. Cox, J.D., J.E. Ingersol, Jr., and S.A. Ross (1985). ‘A Theory of the Term Structure of Interest Rates’. Econometrica, 53, 129–151. Deistler, M., W. Dunsmuir and E. J. Hannan (1976). ‘Vector Linear Time Series Models: Corrections and Extensions’. Advanced in Applied Probability, 8, 360–372. De Jong, Piet (1991). ‘The Diffuse Kalman Filter’. The Annals of Statistics, Vol. 19, No. 2, pp. 1073–1083. De Jong, Piet (1991). ‘Stable Algorithms for the State Space Model’. Journal of Time Series Analysis, Vol. 12, No. 2, pp. 143–157. DeJong, D.N., J.C. Nankervis, N.E. Savin, and C.H. Whiteman (1988). ‘Integration versus Trend-Stationarity in Macroeconomic Time Series’. Mimeo. University of Iowa, Department of Economics. Working Paper #88-27, December. DeJong, D.N. and C.H. Whiteman (1989a). ‘The Temporal Stability of Dividends and Stock Prices: Evidence from the Likelihood Function’. Mimeo. University of Iowa, Department of Economics. Working Paper #89-03, February. DeJong, D.N. and C.H. Whiteman (1989b). ‘Trends and Random Walks in Macroeconomic Time Series: A Reconsideration Based on the Likelihood Principle’. Mimeo. University of Iowa, Department of Economics. Working Paper #89-04, February. De Long, J. Bradford and Lawrence H. Summers (1991). ‘Equipment Investment and Economic Growth’. Quarterly Journal of Economics, pp. 445–501, May. Denman, E.D. and A.N. Beavers (1976). ‘The Matrix Sign Function and Computations in Systems’. Applied Mathematics and Computation, 2, 63–94. Doan, Thomas, Robert Litterman, and Christopher Sims (1984). ‘Forecasting and Conditional Projection using Realistic Prior Distributions’. Econometric Reviews, Vol. 3, pp. 131–44. Dunsmuir, W. and E. J. Hannan (1978). ‘Vector Linear Time Series Models’. Advances in Applied Probability, 8, 339–364. Dunsmuir, W. (1978). ‘A Central Limit Theorem for Parameter Estimation in Stationary Linear Time Series and its Application to Models for Signal Observed with Noise’. Annals of Statistics, 7, 490-506. Eckstein, Z. (1984). ‘Rational Expectations Modeling of Agricultural Supply’. Journal of Political Economy, 93, 1–19. Eichenbaum, M.S. (1983). ‘A Rational Expectations Equilibrium Model of Finished Goods and Employment’. Journal of Monetary Economics, 12, 259–278.

502

References

Eichenbaum, M.S. and L.P. Hansen (1990). ‘Estimating Models with Intertemporal Substitution Using Aggregate Time Series Data’. Journal of Business and Economic Statistics, 8, 53–69. Emre, E. and O. Huseyin (1975). ‘Generalization of Leverrier’s Algorithm to Polynomial Matrices of Arbitrary Degree’. IEEE Transactions on Automatic Control, February. Engle, R.F., and C.W.J. Granger (1987). ‘Cointegration and Error-Correction: Representation, Estimation and Testing’. Econometrica, 55, 251–276. Engle, R. and M. Watson (1981). ‘A One Factor Multivariate Time Series Model of Metropolitan Wage Rates’. Journal of the American Statistical Association, 76, 774–781. Flavin, M.A. (1981). ‘The Adjustment of Consumption to Changing Expectations About Future Income’. Journal of Political Economy, 89, 974–1009. Futia, C.A. (1981). ‘Rational Expectations in Stationary Linear Models’. Econometrica, 49, 171–192. Gantmacher, F.R. (1959). The Theory of Matrices. Vol. 1, New York: Chelsea. Geweke, J.B. (1977a). ‘Wage and Price Dynamics in U.S. Manufacturing’. In C.A. Sims (ed.), New Methods in Business Cycle Research: Proceedings From a Conference 111–56, Minneapolis: Federal Reserve Bank of Minneapolis. Geweke, J.B. (1977b). ‘The Dynamic Factor Analysis of Economic Time Series’. In D.J. Aigner and A.S. Goldberger (eds.), Latent Variables in Socio-Economic Models. Ch. 19, 365–383, Amsterdam: North-Holland. Geweke, J.B. (1978). ‘Temporal Aggregation in the Multivariate Regression Model’. Econometrica, 46, 643–662. Geweke, J.B. and K.J. Singleton (1981a). ‘Maximum Likelihood ‘Confirmations’ Factor Analysis of Economic Time Series’. International Economic Review, 22, 37–54. Geweke, J.B. and K.J. Singleton (1981b). ‘Latent Variable Models for Time Series: A Frequency Domain Approach with an Application to the Permanent Income Hypothesis’. Journal of Economics, 17, 287–304. Gladysev, E.G. (1961). ‘Periodically Correlated Random Sequences’. Sov. Math, 2, 385–88. Gohberg, I., P. Lancaster and L. Rodman (1982). Matrix Polynomials. Academic Press. Gould, J.P. (1968). ‘Adjustment Costs in the Theory of Investment of the Firm’. Review of Economic Studies, 35, 47–56. Granger, C.W.J. (1969). ‘Investigating Causal Relations by Econometric Models and CrossSpectral Methods’. Econometrica, 37, 424–438. Grether, David and M. Nerlove (1970). ‘Some Properties of ‘Optimal’ Seasonal Adjustment’. Econometrica, vol. 38, pp. 682–703. Griliches, Z. (1967). ‘Distributed Lags: A Survey’. Econometrica, 35, 16–40. Grossman, S., A. Melino, and R.J. Shiller (1987). ‘Estimating the Continuous Time ConsumptionBased Asset Pricing Model’. Journal of Business and Economic Statistics, 6, 315–327. Hakkio, C.S. and M. Rush (1986). ‘Co-integration and the Government’s Budget Deficit’. Mimeo. Working Paper, Federal Reserve Bank of Kansas City. Hall, R.E. (1978). ‘Stochastic Implications of the Life Cycle-Permanent Income Hypothesis: Theory and Evidence’. Journal of Political Economy, 86, 971–987. Hall, R.E. (1988). ‘Intertemporal Substitution and Consumption’. Journal of Political Economy, 96, 339–357. Halmos, P.R. (1950). Measure Theory. New York, Van Nostrand.

References

503

Halmos, P.R. (1957). Introduction to Hilbert Space. New York, Chelsea Publishing Company. Hamilton, J.D. and M.A. Flavin (1986). ‘On the Limitations of Government Borrowing: a Framework for Empirical Testing’. American Economic Review, 76, 808–819. Hannan, E. J. (1970). Multiple Time Series. Wiley: New York. Hannan, E. J. (1973). ‘The Asympotic Theory of Linear Time Series Models’. Journal of Applied Probability, 10, 130–145. Hansen, Gary and Thomas J. Sargent (1988). ‘Straight time and Over time in Equilibrium’. Journal of Monetary Economics, 21: 281–308. Hansen, Lars P. (1980). ‘Large Sample Properties of Generalized Method of Moments Estimators’. Econometrica, 50 (July): pp. 1269–1286. Hansen, Lars P. (1987). ‘Calculating Asset Prices in Three Example Economics’. In T.F. Bewley (ed.), Advances in Econometrics Vol. 4, Cambridge University Press, 207–243. Hansen, L.P., D. Epple and W. Roberds (1985). ‘Linear Quadratic Models of Resource Depletion’. In T.J. Sargent (ed.), Energy, Foresight, and Strategy Resources for the Future, Washington, D.C., 101–142. Hansen, Lars Peter, John Heaton, and Thomas J. Sargent (1991). ‘Faster Methods for Solving Continuous Time Recursive Linear Models of Dynamic Economies’. In Lars P. Hansen and Thomas Sargent (eds.), Rational Expectations Econometrics. Westview Press. Hansen, L.P. and R.J. Hodrick (1980). ‘Forward Exchange Rates as Optimal Predictors of Future Spot Rates: An Econometric Analysis’. Journal of Political Economy, 88, 829–853. Hansen, L.P. and R.J. Hodrick (1983). ‘Exchange Rates and International Macroeconomics’. In J. Frenkel (ed.), NBER 113–152. Hansen, Lars Peter, William Roberds, and Thomas J. Sargent (1991). ‘Time Series Implications of Present Value Budget Balance and of Martingale Models of Consumption and Taxes’. In Lars P. Hansen and Thomas Sargent (eds.), Rational Expectations Econometrics. Westview Press. Hansen, L.P. and T.J. Sargent (1980a). ‘Formulating and Estimating Dynamic Linear Rational Expectations Models’. Journal of Economic Dynamics and Control, 2, 7–46. Hansen, L.P. and T.J. Sargent (1980b). ‘Methods for Estimating Continuous Time Rational Expectations Models from Discrete Time Data’. Staff Report 59, Federal Reserve Bank of Minneapolis. Hansen, L.P. and T.J. Sargent (1981a). ‘Linear Rational Expectations Models for Dynamically Interrelated Variables’. In R.E. Lucas, Jr., and T.J. Sargent (eds.), Rational Expectations and Econometric Practice. Minneapolis: University of Minnesota Press. Hansen, L.P. and T.J. Sargent (1981b). ‘A Note of Wiener-Kolmogorov Prediction Formulas for Rational Expectations Models’. Economic Letters, 8, 255–260. Hansen, L.P. and T.J. Sargent (1981c). ‘Identification of Continuous Time Rational Expectations Models from Discrete Time Data’. Mimeo. COMPLETE REFERENCE. Hansen, L.P. and T.J. Sargent (1981d). ‘Formulating and Estimating Continuous Time Rational Expectations Models’. Staff Report 75, Federal Reserve Bank of Minneapolis, October.

504

References

Hansen, L.P. and T.J. Sargent (1981e). ‘Exact Linear Rational Expectations Models: Specification and Estimation’. Staff Report 71, Research Department, Federal Reserve Bank of Minneapolis. Hansen, L.P. and T.J. Sargent (1982). ‘Instrumental Variables Procedures for Estimating Linear Rational Expectations Models’. Journal of Monetary Economics, 9, 263– 296. Hansen, L.P. and T.J. Sargent (1983a). ‘Aggregation over Time and the Inverse Optimal Predictor Problem for Adaptive Expectations in Continuous Time’. International Economic Review, 24, 1–20. Hansen, L.P. and T.J. Sargent (1983b). ‘The Dimensionality of the Aliasing Problem in Models with Rational Spectral Densities’. Econometrica, 51, 377–387. Hansen, L. P. and T.J. Sargent (1984). ‘Two Difficulties in Interpreting Vector Autoregressions’. Mimeo. Working Paper 227, Federal Reserve Bank of Minneapolis. Hansen, L.P. and T.J. Sargent (1990). ‘Recursive Linear Models of Dynamic Economies’. Mimeo. COMPLETE REFERENCE. Hansen, Lars Peter and and Thomas J. Sargent (1991). ‘Lecture Notes on Least Squares Prediction Theory’. In Lars P. Hansen and Thomas Sargent (eds.), Rational Expectations Econometrics. Westview Press. Hansen, Lars Peter, and Thomas J. Sargent (1991). ‘Identification of Continuous Time Rational Expectations Models from Discrete Time Data’. In Lars P. Hansen and Thomas Sargent (eds.), Rational Expectations Econometrics. Westview Press. Hansen, Lars P., and Thomas J. Sargent (1991). ‘Two Difficulties in Interpreting Economic Time Series’. In Lars P. Hansen and T.J. Sargent (eds.), Topics in Rational Expectations Econometrics. Westview Press. Hansen, Lars Peter and Thomas J. Sargent (1990b). ‘Disguised Periodicity as a Source of Seasonality’. Mimeo. June COMPLETE REFERENCE. Hansen, Lars Peter and Thomas J. Sargent (1993). Recursive Linear Models of Dynamic Economies. University of Chicago and Hoover Institution. Hansen, L.P. and K.J. Singleton (1982). ‘Generalized Instrumental Variables Estimators of Nonlinear Rational Expectations Models’. Econometrica, 50, 1269–1286. Hansen, L.P. and K.J. Singleton (1983). ‘Stochastic Consumption, Risk Aversion, and the Temporal Behavior of Asset Returns’. Journal of Political Economy, 91, 249–265. Harris, Milton (1987). Dynamic Economic Analysis. New York: Oxford University Press. Harrison, Michael and D. Kreps (1979). ‘Martingales and Arbitrage in Multiperiod Securities Markets’. Journal of Economic Theory, 20: 381–408. Harvey, A.C. (1981). Time Series Models. New York: Halsted. Harvey, A.C., and G.D.A. Phillips (1979). ‘Maximum Likelihood Estimation of Regression Models with Autoregressive-Moving Average Disturbances’. Biometrika, 66, 1, pp. 49–58. Harvey, A.C., and R.G. Pierse (1984). ‘Estimating Missing Observations in Economic Time Series’. Journal of the American Statistical Association, Vol. 79, No. 385, pp. 125–131. Harvey, A.C., and James H. Stock (1985). ‘The Estimation of Higher Order Continuous Time Autoregressive Models’. Econometric Theory, 1: 97–112.

References

505

Harvey, A.C., and James H. Stock (1987). ‘Estimating Integrated Higher Order Continuous Time Autoregressions with an Application to Money-income Causality’. Mimeo. Working Paper No. E 87-28. Stanford, CA: Hoover Institution Press. Harvey, C. (1988). ‘The Real Term Structure of Consumption Growth’. Journal of Financial Economics, 22, 305–333. Hatanaka, M. (1975). ‘On the Global Identification of the Dynamic Simultaneous Equations Model with Stationary Disturbances’. International Economic Review, 16, 545– 554. Hayashi, F. and C.A. Sims (1983). ‘Nearly Efficient Estimation of Time Series Models with Predetermined, but not Exogenous Instruments’. Econometrica, 51, 3, 783–798. Heaton, John (1988). ‘The Interaction Between Time-Nonseparable Preferences and Time Aggregation’. Mimeo. University of Chicago. Heaton, J. (1989). ‘The Interaction Between Time-Nonseparable Preferences and Time Aggregation’. Ph.D. dissertation, University of Chicago. Heaton, John C. (1990). ‘The Interaction Between Time Nonseparable Preferences and Time Aggregation’. Mimeo. Sloan Working Paper 3181-90-EFA, Massachusetts Institute of Technology. Hildenbrand, W. (1974). Core and Equilibria of a Large Economy. Princeton: Princeton University Press. Houthakker, H. and L. Taylor (1970). Consumer Demand in the United States: Analyses and Projections. Second Edition. Cambridge, MA: Harvard University Press. Jacobson, David H. (1973). ‘Optimal Stochastic Linear Systems with Exponential Performance Criteria and Their Relation to Deterministic Differential Games’. IEEE Transactions on Automatic Control, Vol. AC-18, No. 2, April, 124–131. Jacobson, David H. (1977). Extensions of Linear-Quadratic Control, Optimization and Matrix Theory. Academic Press, New York. Jacobson, D. H. (1973). ‘Optimal Stochastic Linear Systems with Exponential Performance Criteria and their Relation to Deterministic Differential Games’. IEEE Transactions on Automatic Control, AC-18, 124–31. Jacobson, D. H. (1973). Extensions of Linear-quadratic Control, Optimization and Matrix Theory. Academic Press, New York. Jones, Larry and Rodolfo Manuelli (1988). ‘A Model of Optimal Equilibrium Growth’. Mimeo. Stanford University. Jones, R.H. (1980). ‘Maximum Likelihood Fitting of ARMA Models to Time Series with Missing Observations’. Technometrics, 22, 389–395. Jones, R.H. and W.M. Brelsford (1967). ‘Time Series With Periodic Structure’. Biometrika, vol. 54, pp. 403–408. Kalman, R.E. (1960). ‘A New Approach to Linear Filtering and Prediction Problems’. Trans. ASME Ser. D. J. Basic Engrg., 82: 35–45. Kehoe, Timothy and David Levine (1984). ‘Empirical Implications of Complete Contingent Claims Markets’. Mimeo. University of California at Los Angeles. Kennan, J. (1979). ‘The Estimation of Partial Adjustment Models with Rational Expectations’. Econometrica, 47, 1441–1456. Kennan, John (1988). ‘An Econometric Analysis of Fluctuations in Aggregate Labor Supply and Demand’. Econometrica, 56, 2, 317–333.

506

References

King, Robert G., Charles I. Plosser, and Sergio T. Rebelo (1988). ‘Production, Growth, and Business Cycles: I. The Basic Neoclassical Model’. Journal of Money, Credit, and Banking, vol. 21, No. 2/3, March, May, 195–232. King, Robert G., Charles I. Plosser, and Sergio T. Rebelo (1988). ‘Production, Growth, and Business Cycles: II. New Directions’. Journal of Money, Credit, and Banking, vol. 21, No. 2/3, March, May, 309–342. Kohn, Robert and C.F. Ansley (1985). ‘Efficient Estimation and Prediction in Time Series Regression Models’. Biometrika, 72, pp. 694–7. Kohn, Robert and C.F. Ansley (1986). ‘Estimation, Prediction, and Interpolation for ARIMA Models with Missing Data’. Journal of the American Statistical Association, Vol. 81, No. 395, pp. 751–61. Kollintzas, T. (1985). ‘The Symmetric Linear Rational Expectations Model’. Econometrica, 53, 963–985. Kullback, S. and R.A. Leibler (1951). ‘On Information and Sufficiency’. Annals of Mathematical Statistics, 22, 79–86. Kwakernaak, H. and R. Sivan (1972). Linear Optimal Control Systems. New York, Wiley Interscience. Kydland, Finn and Edward C. Prescot (1982). ‘Time to Build and Aggregate Fluctuations’. Econometrica, 50: pp. 1345–70. LeRoy, S.F. (1973). ‘Risk Aversion and the Martingale Property of Stock Prices’. International Economic Review, 14, 436–446. LeRoy, S.F. and R. Porter (1981). ‘The Present Value Relation: Tests Based on Implied Variance Bounds’. Econometrica, 49, 97–113. Litterman, R. and L. Weiss (1985). ‘Money, Real Interest Rates and Output: A Reinterpretation of Post-war Data’. Econometrica, 53, 129–156. Ljung, Lennart (1987). System Identification: Theory for the User. Prentice Hall, Englewood Cliffs. Lo, A. (1988). ‘Maximum Likelihood Estimation of Generalized Ito Processes with Discretely Sampled Data’. Journal of Econometric Theory, 4, 231–247. Lucas, R.E., Jr. (1967). ‘Adjustment Costs and the Theory of Supply’. Journal of Political Economy, 75, 321–334. Lucas, R.E., Jr. (1972). ‘Econometric Testing of the Natural Rate Hypothesis’. In O. Eckstein (ed.), The Econometrics of Price Determination Conference. Washington: Board of Governors of the Federal Reserve System, pp. 50–59. Lucas, R.E., Jr. (1973). ‘Some International Evidence on Output-Inflation Tradeoffs’. American Economic Review, 63 (June): 326–34. Lucas, R.E., Jr. (1976). ‘Econometric Policy Evaluation: A Critique’. In K. Brunner and A.H. Meltzer (eds.), The Phillips Curve and Labor Markets. Carnegie-Rochester Conference on Public Policy 1: 19–46, Amsterdam: North Holland. Lucas, R.E., Jr. (1978). ‘Asset prices in an Exchange Economy’. Econometrica, 46, 1429– 1445. Lucas, R.E., Jr. (1981). ‘Optimal Investments with Rational Expectations’. Rational Expectations and Econometric Practice, Vol. 1, 127–156. Lucas, Robert E., Jr. (1987). ‘Money and Interest in a Cash-in-Advance Economy’. Econometrica, 55: 491–513.

References

507

Lucas, R.E., Jr. and E.C. Prescott (1971). ‘Investment Under Uncertainty’. Econometrica, 39, 659–681. Lucas, Robert E., Jr., and Edward C. Prescott (1974). ‘Equilibrium Search and Unemployment’. Journal of Economic Theory, 7: 188–209. Lucas, R.E., Jr. and T.J. Sargent (1981). Introductory Essay to Rational Expectations and Econometric Practice. Minneapolis: University of Minnesota Press. Luenberger, David G. (1966). ‘Observers for Multivariable System’. IEEE Transactions on Automatic Control, 11: 190–97. Luenberger, David G. (1967). ‘Cononical Forms for Multivariable System’. IEEE Transactions on Automatic Control, 12: 290–93. Luenberger, D.G. (1969). Optimization by Vector Space Methods. John Wiley & Sons. Mace, Barbara (1989). ‘Full Insurance in the Presence of Aggregate Uncertainty’. Mimeo. Rochester Center for Economic Research, Working Paper #197. Mankiw, N.G. (1982). ‘Hall’s Consumption Hypothesis and Durable Goods’. Journal of Monetary Economics, 10, 417–425. Marcet, A. (1987). ‘Approximation of the Continuous Wold Decomposition by Frequent Sampling’. Mimeo. Carnegie Mellon University. Marcet, Albert (1988). ‘Solving Nonlinear Models by Parameterizing Expectations’. Mimeo. Carnegie-Mellon University, G.S.I.A. Marcet, Albert (1991). ‘Temporal Aggregation of Economic Time Series’. In Lars P. Hansen and Thomas Sargent (eds.), Rational Expectations Econometrics. Westview Press. McGrattan, Ellen R. (1991a). ‘A Note on Computing Equilibria in Linear Models’. Mimeo. Duke University, June. McGrattan, Ellen R. (1991b). ‘Fiscal Policy and Technical Change in a Model of Business Cycles’. Mimeo. Duke University, December. Miller, P.J. and W. Roberds (1987). ‘The Quantitative Significance of the Lucas Critique’. Staff Report, Federal Reserve Bank of Minneapolis, #109. Monahan, J. (1984). ‘A Note on Enforcing Stationarity in Autoregressive-Moving Average Models’. Biometrika, 71, 403–404. Monin, A.S. (1963). ‘Stationary and Periodic Time Series in the General Circulation of the Atmosphere’. In M. Rosenblatt (ed.), Proceedings of the Symposium on Time Series Analysis. New York: Wiley. pp. 144–151. Mortenson, D.T. (1973). ‘Generalized Costs of Adjustment and Dynamic Factor Demand Theory’. Econometrica, 41, 657–665. Muth, John F. (1960). ‘Optimal Properties of Exponentially Weighted Forecasts’. Journal of the American Statistical Association, 55, 299–306. Muth, John F. (1961). ‘Rational Expectations and the Theory of Price Movements’. Econometrica, 29: 315–35. Negishi, T. (1960). ‘Welfare Economics and Existence of an Equilibrium for a Competitive Economy’. Metroeconomica, vol. 12: 92–97. Nerlove, M. (1967). ‘Distributed Lags and Unobserved Components in Economic Time Series’. In W. Fellner, et. al. (eds.), Ten Economic Studies in the Tradition of Irving Fisher. New York: Wiley. Nobel, Ben and James W. Daniel (1977). Applied Linear Algebra. 2nd edition, Prentice Hall: Englewood Cliffs, New Jersey.

508

References

Osborn, Denise R. (1991). ‘The Implication of Periodically Varying Coefficients for Time-series Processes’. Journal of Economics, Vol. 43, No. 3, June, pp. 373–384. Osborne, Denise (1990). ‘The Implications of Periodically Varying Coefficients for Seasonal Time Series Processes’. Journal of Econometrics, forthcoming. Osborn, Denise R. (1988). ‘Seasonality and Habit Persistence in a Life Cycle Model of Consumption’. Journal of Applied Econometrics, 3, October-December, pp. 255–266. Osborne, Denise R., and Jeremy P. Smith (1989). ‘The Performance of Periodic Autoregressive Models in Forecasting Seasonal U.K. Consumption’. Journal of Economic and Business Statistics, vol. 7, no. 1, January, pp. 117–128. Osborne, Denise, A.P.L. Chui, J.P. Smith, and C.R. Birchenhall (1988). ‘Seasonality and the Order of Integration for Consumption’. Oxford Bulletin of Economics and Statistics, 50, 4, pp. 361–377. Pagano, M. (1978). ‘On Periodic and Multiple Autoregressions’. The Annals of Statistics, 6, pp. 1310–1317. Park, Jong Ahn (1984). ‘Gestation Lags with Variable Plans: An Empirical Study of Aggregate Investment’. Ph.D. dissertation, Carnegie-Mellon University. Phillips, A.W. (1959). ‘The Estimation of Parameters in Systems of Stochastic Differential Equations’. Biometrika, 59, 67–76. Phillips, P.C.B. (1972). ‘The Structural Estimation of a Stochastic Differential Equation System’. Econometrica, 40, 1021–1041. Phillips, P.C.B. (1973). ‘The Problem of Identification in Finite Parameter Continuous Time Models’. Journal of Econometrics, 1, 351–362. Phillips, P.C.B. (1974). ‘The Estimation of Some Continuous Time Models’. Econometrica, 42, 803–823. Pollak, R.A. (1970). ‘Habit Formation and Dynamic Demand Functions’. Journal of Political Economy, 78, 745–763. Prescott, Edward C. and Robert E. Lucas, Jr. (1972). ‘Price Systems in Infinite Dimensional Space’. International Economic Review, 13: pp. 416–22. Prescott, Edward C., and Rajnish Mehra (1980). ‘Recursive Competitive Equilibrium: the Case of Homogeneous Households’. Econometrica, 48: pp. 1364–79. Roberds, William (1991). ‘Implications of Expected Present Value Budget Balance: Application to Postwar U.S. Data’. In Lars P. Hansen and Thomas Sargent (eds.), Rational Expectations Econometrics. Westview Press. Roberts, J.D. (1971). ‘Linear Model Reduction and Solution of Algebraic Riccati Equations by use of the Sign Function’. CUED/B-Control/TR-13 Report, Cambridge University. Robinson, P.M. (1977). ‘The Construction and Estimation of Continuous Time Models and Discrete Approximations in Econometrics’. Journal of Econometrics, 6, 173–197. Romer, Paul M. (1986). ‘Increasing Returns and Long-Run Growth’. Journal of Political Economy, vol. 94, pp. 1002–37. Rosen, S., K.M. Murphy, and J.A. Scheinkman (1994). ‘Cattle Cycles’. Journal of Political Economy, Vol. 102, No. 3, pp. 468–492. Royden, H.L. (1968). Real Analysis. New York, MacMillan Publishing Co., Inc. Rozanov, Y.A. (1967). Stationary Random Processes. Holden-Day: San Francisco. Rubenstein, Mark (1974). ‘An Aggregation Theorem for Security Markets’. Journal of Financial Economics, Vol. 1, No.3, pp. 225–244.

References

509

Rubinstein, M. (1976). ‘The Valuation of Uncertain Income Streams and the Pricing of Options’. Bell Journal of Economics, 7, 407–425. Rudin, W. (1974). Real and Complex Analysis. McGraw Hill: New York. Ryder, H.E. and G.M. Heal (1973). ‘Optimal Growth with Intertemporally Dependent Preferences’. Review of Economic Studies, 40, 1–31. Sargent, T.J. (1971). ‘A Note on the ‘Accelerationist’ Controversy’. Journal of Money, Credit, and Banking, 3, 721–725. Sargent, T.J. (1978). ‘Estimation of Dynamic Labor Demand Schedules Under Rational Expectations’. Journal of Political Economy, 86, 1009–1044. Sargent, Thomas J. (1978). Comments on. ‘Seasonal Adjustment and Multiple Time Series Analysis’. In A. Zellner (ed.), Seasonal Analysis of Economic Time Series, (by K.F. Wallis). Washington, D.C.: Bureau of the Census, pp. 361–363. Sargent, T.J. (1979). ‘A Note on Maximum Likelihood Estimation of the Rational Expectations of the Term Structure’. Journal of Monetary Economics, 5, 133–143. Sargent, T.J. (1981a). ‘Interpreting Economic Time Series’. Journal of Political Economy, 89, 213–248. Sargent, Thomas J. (1981b). ‘Lecture notes on Filtering, Control, and Rational Expectations’. Mimeo. University of Minnesota. Sargent, T.J. (1983). ‘Preliminary Introduction to Continuous Time Stochastic Processes’. Mimeo. COMPLETE REFERENCE. Sargent, T.J. (1987a). Dynamic Macroeconomic Theory. Cambridge: Harvard University Press. Sargent, Thomas J. (1987b). Macroeconomic Theory. Second Edition, Academic Press, New York, (pp. 336-342). Sargent, Thomas J. (1989). ‘Two Models of Measurements and the Investment Accelerator’. Journal of Political Economy, 97: 251–87. Sargent, T.J. and C.A. Sims (1977). ‘Business Cycle Modeling Without Pretending to Have too Much A Priori Economic Theory’. In C.A. Sims (ed.), New Methods for Business Cycle Research, Federal Reserve Bank of Minneapolis. Sargent, Thomas J. and Neil Wallace (1973). ‘Rational Expectations and the Dynamics of Hyperinflation’. International Economic Review, 63: 328–50. Scheinkman, Jose (1987). ‘Market Incompleteness and the Equilibrium Valuation of Assets’. In S. Bhattacharya and G. Constantinides (eds.), Theory of Valuation: Frontiers of Modern Financial Theory. pp. 45–51. Schwarz, Z.G. (1978). ‘Estimating the Dimension of a Model’. Annals of Statistics, 6, 461–464. Shiller, R.J. (1972). ‘Rational Expectations and the Structure of Interest Rates’. Ph.D. dissertation, (Massachusetts Institute of Technology, Cambridge, MA). Shiller, R.J. (1979). ‘The Volatility of Long-Term Interest Rates and Expectations Models of the Term Structure’. Journal of Political Economy, 87, 1190–1219. Shiller, R.J. (1981). ‘Do Stock Prices Move too much to be Justified by Subsequent Changes in Dividends?’. The American Economic Review, 71, 421–436. Shim, S.D. (1984). ‘Inflation and the Government Budget Constraint: International Evidence’. Ph.D. dissertation, University of Minnesota. Sims, C.A. (1971). ‘Discrete Approximations to Continuous Time Lag Distributions in Econometrics’. Econometrica, 39, 545–564.

510

References

Sims, C.A. (1972a). ‘Approximate Prior Restriction on Distributed Lag Estimation’. Journal of the American Statistical Association, 67, 169–175. Sims, C.A. (1972b). ‘Money, Income and Causality’. American Economic Review, 62, 540–552. Sims, C.A. (1974a). ‘Distributed Lags’. In M.D. Intrilligator and D.A. Kendrick (eds.), Frontiers of Quantitative Economics. Vol. II, Amsterdam: North-Holland, 289– 332. Sims, C.A. (1974b). ‘Seasonality in Regression’. Journal of the American Statistical Association, vol. 69, September pp. 618–626. Sims, Christopher (1976). ‘personal letter to Thomas Sargent’. U.S. mail, October. Sims, Christopher A. (1980). ‘Macroeconomics and Reality’. Econometrica, 48: 1–48. Sims, C.A. (1984). ‘Martingale-Like Behavior of Prices and Interest Rates’. Mimeo. Department of Economics, University of Minnesota, working paper, October. Sims, C.A. (1988). ‘Bayesian Skepticism on Unit Root Econometrics’. Journal of Economic Dynamics and Control, 12, 463–474. Sims, Christopher (1990). ‘Rational Expectations Modeling with Seasonally Adjusted Data’. Mimeo. December. Sims, C.A. and H. Uhlig (1991). ‘Understanding Unit Rooters: A Helicopter Tour’. Econometrica, Vol. 59, No. 6, pp. 1591–1599. Singleton, Kenneth J. (1988). ‘Econometric Issues in the Analysis of Equilibrium Business Cycle Models’. Journal of Monetary Economics, Vol. 21, No. 2/3, March/May; pp. 361–386. Siow, A. (1984). ‘Occupational Choice Under Uncertainty’. Econometrica, Vol. 52, No. 3, pp. 631–645. Stigler, G.J. and G.S. Becker (1977). ‘De Gustibus Non Est Disputandum’. American Economic Review, 67, 76–90. Stokey, Nancy L. and Robert E. Lucas, Jr., with Edward C. Prescott (1989). Recursive Methods in Economic Dynamics. Cambridge, MA: Harvard University Press. Sundaresan, M. (1989). ‘Intertemporally Dependent Preferences and the Volatility of Consumption and Wealth’. The Review of Financial Studies, 2, 73–89. Telser, L.G. and R.J. Graves (1972). Functional Analysis in Mathematical Economics. Chicago: University of Chicago Press. Tiao, G.C., and M.R. Grupe (1980). ‘Hidden Periodic Autoregressive-Moving Average Models in Time Series Data’. Biometrika, 67, 2, pp. 365–73. Todd, Richard (1983). ‘A Dynamic Equilibrium Model of Seasonal and Cyclical Fluctuations in the Corn-Soybean-Hog Sector’. Ph.D. dissertation, University of Minnesota. Todd, Richard (1990). ‘Periodic Linear Quadratic Periodic Models of Seasonality’. Journal of Economic Dynamics and Control, Vol. 14, No 3/4, July/Oct., pp. 763–796. Topel, Robert and S. Rosen (1988). ‘Housing Investment in the United States’. Journal of Political Economy, vol. 96, pp. 718–40. Townsend, R.M. (1983). ‘Forecasting the Forecasts for Others’. Journal of Political Economy, 91, 546–588. Townsend, Robert M. (1987). ‘Arrow-Debreu Programs as Microfoundations for Macroeconomics’. In Truman F. Bewley (ed.), Advances in Econometrica: Fifth World Congress, Cambridge, UK: Cambridge University Press, pp. 379–428. Treadway, A.B. (1969). ‘On Rational Entrepreneur Behavior and the Demand for Investment’. Review of Economic Studies, 367, 227–240.

References

511

Trehan, B. and C.E. Walsh (1988). ‘Common Trends, the Government Budget Constraints, and Revenue Smoothing’. Journal of Economic Dynamics and Control, 12, 425– 446. Vaughan, D.R. (1969). ‘A Negative-Exponential Solution to the Matrix Riccati Equation’. IEEE Transactions on Automatic Control, 14, 72–75. Vaughan, D.R. (1970). ‘ A Nonrecursive Algebraic Solution for the Discrete Riccati Equation’. IEEE Transactions on Automatic Control, Vol. AC-15, No. 5, October, pp. 597– 599. Wallis, Kenneth (1974). ‘Seasonal Adjustment and Relations Between Variables’. Journal of the American Statistical Association, vol. 69, March pp. 18–32. Watson, M. and R. Engle (1983). ‘Alternative Algorithms for the Estimation of Dynamic Factor, MIMIC and Varying Coefficient Regression Models’. Journal of Econometrics, 23, 385–400. West, K.D. (1988). ‘The Insensitivity of Consumption to News About Income’. Journal of Monetary Economics, 21, 17–33. Wheeden, R.L. and A. Zygmund (1977). Measure and Integral; An Introduction to Real Analysis. Marcel Dekker Inc.: New York. White, H. (1982). ‘Maximum Likelihood Estimation of Misspecified Models’. Econometrica, 50, 1-26. Whiteman, Charles (1983). Linear Rational Expectations Models: A User’s Guide. Minneapolis: University of Minnesota Press. Whittle, Peter (1990). Risk Sensitive Optimal Control. Wiley, New York. Wilcox, D.W. (1989). ‘What Do We Know about Consumption?’. Mimeo. Federal Reserve Board of Governors, Washington, D.C. Wilson, Robert, (1968). ‘The Theory of Syndicates’. Econometrica, Vol. 36, No. 1, pp. 119–132. Working, Holbrook (1960). ‘Note on the Correlation of First Differences of Averages in a Random Chain’. Econometrica, 28: 916–18.

Index

periodic model, 351 cross-products between states and controls removal of, 185

adjacent complementarity, 219 aggregation over time, 172 allocation rules, 238 approximation theory, 171 arma autoregressive, moving average processes, 18 Arrow-Debreu competitive equilibrium, 2 asset pricing periodic version, 347 via arbitrage on Arrow-Debreu prices, 114 autocovariance function, 23 autoregression, 27 Bellman equation, 64 Bellman’s “curse of dimensionality”, 2 Bellman’s functional equation, 182 canonical representation of household services computation, 214 canonical representation of service technology, 209 cattle, 134 certainty equivalence, 281 choice household’s, 106 commodity space, 98 competitive equilibrium, 104 definition, 107 computed example a Jones–Manuelli like economy, 72 an economy with a durable consumption good, 72 Brock and Mirman’s model, 71 Brock–Mirman model with adjustment costs, 81 Lucas’s pure exchange economy, 72 periodic Hall model, 364 controllability canonical form, 182, 195 covariance stationarity, 17 covariogram

demand schedules, 209 detectability, 151, 186 directional derivatives, 55 discounted control problem replacement by undiscounted problem, 186 disguised periodicity, 370 two illustrations, 370 doubling algorithm, 69, 116, 182, 192 periodic regulator, 198 dynamic programming linear quadratic, 58 eigenvalues endogenous, 61 exogenous, 61 Engel curve, 235 exogenous state vector, 36 factorization identity, 147 proof, 227 firms leasing, 107 producing, 106 type I, 106 type II, 107 Fourier transforms, 225 frequency domain estimation, 168 Frisch demand functions, 49 fundamental theorems of welfare economics, 104 Gateaux derivative, 55 Gorman aggregation, 232 growth condition Jones–Manuelli, 81 heterogeneous consumers, 237 household technology, 45 adjustment costs, 48 consumer durables, 46 habit persistence, 47 multiple consumption goods, 49 housing market, 133

– 513 –

514

Index

impulse response, 9 inner-product representation, 99 innovation, 148 innovation accounting, 11 innovations functions of model noises, 162 innovations representation, 146 as whitener, 162 periodic version, 368 serially correlated measurement errors, 152 invariant subspace method, 182 invariant subspace methods, 188 Kalman filter, 146, 152 Lagrange multipliers, 52, 53 expressed in terms of gradient of value function, 61 Lagrangian firm of type I, 109 firm of type II, 110 Lagrangian for household’s problem, 108 linear functional, 99 Lyapunov equation, 151 discrete, 17 Martingale difference sequence, 8 measurement errors serially correlated, 152 moving average representation, 9 Negishi weights for social planning problem, 231 occupational choice, 138 optimal linear regulator problem, 183 periodic version, 197 optimal resource allocation problem, 52 Parseval’s formula, 224 partial equilibrium, 128 partitioning state vector, 194 periodic models, 341 periodic optimal linear regulator problem, 182 policy functions dynamic programming, 59

prediction theory linear difference equations, 9 periodic version, 349 preferences, 45 examples, 46 time separable, 46 price system, 111 problem firms of type I, 106 firms of type II, 107 rational expectations hyperinflation – stock price model, 22 reciprocal pairs property, 206 recursive dynamic competitive economies, 1 representative consumer, 231 representative household, 45, 238 Riccati equation, 2, 69, 184, 215, 227 Riccati equations convergence conditions, 151 Riesz Representation Theorem, 103 risk premium, 115 risk sharing, 241 limited markets, 241 risk-sensitive control theory, 201 Schur decomposition, 189 seasonality deterministic, 24 indeterministic, 26 periodic, 341 sharing rules, 238 skip-sampled system, 352 social planning problem, 52 as optimal linear regulator problem, 63 periodic version, 342 variational approach, 53 spectral density matrix, 23 square summability, 50 stability conditions, 151 stabilizability, 151, 186 stacked system, 352 stochastic difference equation linear first order, 7 summation by parts, 95 Sylvester equation, 151, 196

Index

symplectic, 193 symplectic matrices, 189, 206 technology, 36 a depletable resource, 42 examples, 38 interrelated factor demand, 44 learning by doing, 43 multi-period adjustment costs, 39 pure endowment, 38 single-period adjustment costs, 39 time to build, 39 technology: growth, 41 term structure of interest rates, 115, 117 term structure of interest rates, periodic version, 350 Tiao–Grupe formula, 360 Tiao-Grupe formula, 359 state space realization, 361 two fund theorem, 242 uncontrollable stochastic process, 53 unit root, 26 value function dynamic programming, 59 iterations on, 184 Vaughan’s method, 187 vector autoregression, 1, 14 whitener, 162 whitening filter, 150 Wold representation, 157

515

Author Index

Altug, Sumru, 275 Anderson, Brian D. O., 3, 168, 182, 193 Ansley, Craig, 298 Arrow, Kenneth, 3 Barro, Robert, 305 Becker, Gary, 77, 239, 249 Bellman, Richard, 4, 65 Blanchard, Olivier, 166 Breiman, Leo, 416 Brock, William, 46, 95 Browning, Martin, 54

Jones, Larry, 46, 53, 87 Judd, Kenneth, 4, 305 Kahn, Charles, 166 Kehoe, Patrick, 305 Kohn, Robert, 298 Koopmans, T.C., 46 Kreps, David, 109 Kwakernaak, Huibert, 3, 170 Kydland, Finn, 45 Lancaster, Kelvin, 50 Le Roy, Stephen, 24 Liu, blank, 252 Lucas, Robert E., Jr., 3, 64, 78, 124, 142

Eichenbaum, Martin, 142 Engle, Robert, 298

Mace, Barbara, 275 MaCurdy, Thomas, 54 Mankiw,Gregory, 77 Manuelli, Rodolfo, 46, 53, 87 Marcet, Albert, 4 Miller, Robert, 275 Mirman, Leonard, 46, 95 Moore, James, 168 Moore, John, 182, 193 Moore, John B., 3 Mortensen, Dale, 49 Murphey, Kevin, 146 Murphy, Kevin, 77, 239, 249 Muth, John F., 23

Flavin, Marjorie, 46 Frisch, Ragnar, 54

Negishi, T., 265 Nerlove, Marc, 252

Gabel, R. A., 252 Gladysev, E. G., 405 Gorman, Terrance, 7, 50, 243, 265 Granger, C. W. J., 298 Grether, David, 252 Grupe, M. R., 405

Osborn, Denise, 387 Owens, Raymond, 154

Hall, Robert E., 46, 95 Harris, Milton, 3 Harrison, Michael, 109 Heal, Geoffrey, 53, 249 Heckman, James, 54 Houthakker, Hendrik, 53

Richard, Scott, 109 Roberds, William, 206 Roberts, R. A., 252 Rosen, Sherwin, 144, 146, 149 Rubinstein, Mark, 277 Ryder, Harl, 53, 249

Irish,

Scheinkman, Jose, 146

Cagan, Phillip, 24 Carvalho, Jose, 252 Cass, David, 46 Chari, V. V. , 305 Christiano, Lawrence, 305 Cochrane, John , 275 Coleman, Wilbur, 4 Deaton, Angus, 54 Debreu, Gerard, 3, 106

, 54

Park, Jong Ahn, 45 Pollak, Robert, 249 Porter, Richard, 24 Prescott, Edward C., 3, 45, 64, 142

– 517 –

518

Shiller, Robert, 24 Sims, Christopher A., 3, 11, 213 Siow, Aloyisius, 149 Sivan, Raphael, 3, 170 Slutsky, Eugen, 11 Stigler, George, 249 Stokey, Nancy, 3, 64 Tauchen, George, 4 Taylor, Lester, 53 Tiao, George, 405 Todd, Richard, 387 Topel, Robert, 144 Townsend, Robert, 275 Vaughan, D. R., 157 Whiteman, Charles, 166

Author Index

Matlab Index

Matlab Index

dog.m, 321 seasla.m, 410 aarma.m, 74 aggreg.m, 217 aimpulse.m, 74 asimul.m, 74 asseta.m, 75, 132 assets.m, 411 assetss.m, 411 avg.m, 215, 217 bewley4.m, 282 bewley5.m, 282 canonpr.m, 244 clex11.m, 72, 217, 317 clex11h.m, 317 clex14.m, 130, 134 clex35b.m, 282 compare.m, 334 dimpulse.m, 12, 217 disthet.m, 334, 343 dlsim.m, 26 dlyap.m, 19 dog.m, 321 double2.m, 343

519

double2j.m, 75, 172, 176, 343 doublej.m, 20 doublej2.m, 20 doubleo.m, 75 doublex.m, 373 evardec.m, 13 factors.m, 411 heter.m, 278 kfilter.m, 325 seasla.m, 410, 411 series.m, 328 simpulse.m, 411, 412 simulh.m, 278 simulhet.m, 317 simuls.m, 411 solvdist.m, 334, 339 solvea.m, 74, 75, 95, 217, 278, 317, 334 solvehet.m, 317 solves.m, 411 solvex.m, 373 spectral.m, 25 spectrs.m, 411 steadst.m, 74, 411 varrep.m, 203 white1.m, 204 white2.m, 204

Our partners will collect data and use cookies for ad personalization and measurement. Learn how we and our ad partner Google, collect and use data. Agree & close