DEVELOPMENTS OF CONTROL THEORY FOR ECONOMIC ANALYSIS

ADVANCED STUDIES IN THEORETICAL AND APPLIED ECONOMETRICS

VOLUME 7

Managing Editors:
J.P. Ancot, Netherlands Economic Institute, Rotterdam, The Netherlands
A.J. Hughes Hallett, University of Newcastle, U.K.

Editorial Board:
F.G. Adams, University of Pennsylvania, Philadelphia, U.S.A.
P. Balestra, University of Geneva, Switzerland
M.G. Dagenais, University of Montreal, Canada
D. Kendrick, University of Texas, Austin, U.S.A.
J.H.P. Paelinck, Netherlands Economic Institute, Rotterdam, The Netherlands
R.S. Pindyck, Sloane School of Management, M.I.T., U.S.A.
H. Theil, University of Florida, Gainsville, U.S.A.
W. Welfe, University of Lodz, Poland

For a complete list of volumes in this series see final page of this volume.
Developments of Control Theory
for Economic Analysis

edited by

Carlo Carrara and Domenico Sartore

(University of Venice)

1987
KLUWER ACADEMIC PUBLISHERS .1.
DORDRECHT I BOSTON I LANCASTER ,~
IV

Distributors

jor the United States and Canada: Kluwer Academic Publishers, P.O. Box 358.
Accord Station, Hingham, MA 02018-0358, USA
jor the UK and Ireland: Kluwer Academic Publishers, MTP Press Limited.
Falcon House, Queen Square, Lancaster LAI lRN, UK
jor all other countries: Kluwer Academic Publishers Group, Distribution Center.
P.O. Box 322, 3300 AH Dordrecht, The Netherlands

Library of Congress Cataloging in Publication Data

Developments of control theory for economic analysis.

(Advanced studies in theoretical and applied

econometrics; v. 7)
!'Proceedings of the Conference on "EconomIc Policy
and Control Theory" which was held at the University of
Venice (Italy) on 27 January-l February 1985"--Pref.
1. EconomIC pol1cy--Econometric models--Congresses.
2. Control theory--Econometric models--Congresses.
3. Macroeconomlcs--Econometric models--Congressۤ.
I. Carrara, Carlo. II. Sartore, Domenico. III. Confer-
ence on "Economic Policy and Control Th,'ory" (1985
University of Venice) IV. Series.
HD73.048 1986 338.9'007,4 86-8662

ISBN-13: 978-94-010-8063-7 e-ISBN-13: 978-94-009-3495-5

DOl: 10.1007/978-94-009-3495-5

Copyright

© 1987 by Martinus Nijhoff Publishers, Dordrecht.

Softcover reprint of the hardcover 2nd edition 1987
All rights reserved. No part of this publication may be reproduced, stored in a
retrieval system, or transmitted in any form or by any means, mechanical
photocopying, recording, or otherwise, without the prior written permission of
the publishers,
Martinus Nijhoff Publishers, P.O. Box 163, 3300 AD Dordrecht,
The Netherlands.
 v

PREFACE
Giovanni Castellani
Rector of the University of Venice

This book contains the Proceedings of the Conference on "Economic

Policy and Control Theory" which was held at the University of
Venice (Italy) on 27 January-l February 1985.
The goal of the Conference was to survey the main developments
of control theory in economics, by emphasizing particularly new
achievements in the analysis of dynamic economic models by con-
trol methods.
The development of control theory is strictly related to the
development of science and technology in the last forty years.
Control theory was indeed applied mainly in engineering, and only
in the sixties economists started using control methods for analys-
ing economic problems, even if some preliminary economic applica-
tions of calculus of variations, from which control theory was then
developed, date back to the twenties.
Applications of control theory in economics also had to solve
new, complicated, problems, like those encountered in optimal
growth models, or like the determination of the appropriate inter-
temporal social welfare function, of the policy horizon and the
relative final state of the system, of the appropriate discount factor.
Furthermore, the uncertainty characterizing economic models had
to be taken into account, thus giving rise to the development of
stochastic control theory in economics.
The Conference not only tried to survey the state of art of control
theory in economics, but also provided original contributions on
the mathematical aspects of control theory and on their macro-
economic applications, on new solution algorithms of nonlinear
control problems and on the relative computer software. In par-
ticular, the development of new software for performing reliable
applications of control methods in economics has to be considered
VI

a relevant achievement for its implications on a wider utilization of

control theory.
This book contains therefore the most important results presented
at the Venice symposium and can be considered a useful tool both
for applied and theoretical economists.
As Rector of the University of Venice, I would like to thank the
authors of the articles published in this volume and all the persons
who attended the Conference, the editors and the publisher of the
book and the Centro Nazionale delle Ricerche whose financial
contribution is gratefully acknowledged. I also thank the Comune
di Venezia, IBM, the Banco S. Marco, the Bank of Italy and the
Comitato per gli Studi e la Progammazione whose financial support
made possible the organization of the Conference.
 VII

TABLE OF CONTENTS

Preface V
G. Castellani

In trod uction XI
C. Carraro and D. Sartore

PART I: INTRODUCTION TO CONTROL THEORY:

METHODS AND ALGORITHMS

1. Developments of Control Theory in Macroeconomics 3

G. C. Chow
2. Linear Controllability: Results and Open Questions 21
R. Conti
3. A System Theoretic Approach to the Theory of
Economic Policy 31
M. L. Petit
4. Software for Economic Optimal Control Models 47
D. Kendrick

PART II: RECENT DEVELOPMENTS OF

CONTROL THEORY: OBJECTIVE
FUNCTION SPECIFICATION

5. Interactive Vector Optimization as a Complement to

Optimal Control in Econometric Decision Models 63
J. Gruber
VIII

6. Risk Reduction and the Robustness of Economic

Policies 83
A. S. Brandsma
7. Optimal Economic Policies under a Crawling-Peg
Exchange 105
H. M. Amman and H. Jager
8. Some Remarks on Forward Programming 127
J. Eppers and M. Leserer
9. Utility and Uncertainty in Intertemporal Choice 135
A . Montesano
10. Gradient Methods for FIML Estimation of
Econometric Models 143
G. Calzolari and L. Panattoni

PART III: RECENT DEVELOPMENTS OF

CONTROL THEORY: A GAME
THEORETIC APPROACH

11. Methods for the Simultaneous Use of Multiple

Models in Optimal Policy Design 157
B. Rustem
12. Optimal Policy Design in Interdependent Economies 187
A. J. Hughes Hallett
13. Hierarchical Games for Macroeconomic Policy
Analysis 215
C. Carraro
14. A Dynamic and Stochastic Model of Price
Leadership 239
P. Fanchon, E. Rifkin and J. K. Sengupta
15. Quality and Reputation Policies of Duopolists under
Asymmetric Information 261
K. Conrad

PART IV: ECONOMIC AND ECONOMETRIC

ANALYSIS BY CONTROL METHODS

16. A System Approach to Insurance Company

Management 279
C. Tapiero
 IX

17. Capital Theoretics, Business Cycles and Feedback

Policy: An Experiment in Macroeconomic Control 305
J. H. Gapinski
18. Exogeneity and Control 327
J. F. Richard
 XI

INTRODUCTION

Carlo Carraro and Domenico Sartore

This book contains a selection of the papers presented at the

symposium on "Economic Policy and Control Theory" which was
held in Venice (Italy) on 27 January-l February 1985 under the
auspices of the University of Venice.
The choice of the topics of the symposium was guided by three
principles. First, significant results and advances in the area have
taken place in the past ten years and the area is subject to continued
extensive research effort at present. Second, some emphasis was
placed on theoretical developments, in particular inasmuch as they
have contributed improvements to current research in more applied
areas of control theory and economic policy, or are likely to do so
in the future. Third, the advantage of using sophisticated meth-
odologies, like control theory, for policy analysis was greatly
emphasized.
Therefore, several sessions of the symposium were devoted to
topics like adaptive control, strategic (game theoretic) control, risk
and uncertainty in policy analysis, applied control with particular
emphasis on rational expectations models and econometric tech-
niques for control. Furthermore, in order to understand and define
future directions of research, some papers were devoted to survey-
ing the state of art of control methods in economics under different
viewpoints.
The goal of the symposium was then to review the developments
of control theory for economic analysis in the last decade and to
propose new theoretical and applied contributions aimed at
generalizing and improving the methodology upon which economic
analysis should be based.
Since a large number of papers was presented at the conference,
it was not possible to publish the complete proceedings of the
symposium and a careful selection had to be made. Therefore, all

C. Carrara and D. Sartore (eds). Developments of Control Theory for Economic AnalysIS
© 1987 Martinus Nijhoff Publishers (Kluwer), Dordrecht·
XII

the papers went through a refereeing process and were then

screened by following two criteria: the quality of the results con-
tained in the paper and its importance for future developments of
research in the area. The four sections of the book reflect this choice
and emphasize the editors' viewpoints on different aspects of control
theory and economic policy.
The first applications of control theory in economics highlighted
the usefulness of this methodology for determining optimal economic
policy (see Chow, 1975, for example). Subsequently, control theory
has been used for studying the properties of actual and optimal
policies by analysing the structure of the policy reaction function
and of the econometric model describing the underlying economic
system (e.g. Chow, 1981; Preston-Pagan, 1982). When uncertainty
was introduced into the policymaker's perception of the economic
model or into the welfare function that the policymaker wants to
maximizes, control theory developed appropriate tools for dealing
with those mathematical and economic problems, making use of
sophisticated numerical techniques and computer software (a good
example is Kendrick, 1981).
The first section of the book is aimed at surveying all these
aspects of control theory and their relationships with policy analysis
by presenting the main features of control and controllability
theory, optimal decision making under certainty and uncertainty
conditions and software for optimal control models. The evolution
of economic applications of control methods is also considered and
the importance of optimal control for providing better understand-
ing of behaviour, relationships and effects of economic variables is
emphasized.
The survey papers included in the first section also defines some
directions for future research. In particular, there are suggestions
for generalizing control methods both from the theoretical and the
computational viewpoint by solving control problems where either
the model or the welfare function is assumed to be nonlinear in
order to capture the actual features of the economic system and the
actual behaviour of the policymaker.
The development of control theory in economics is indeed
characterized by an increasing demand for new analytical tools in
order to increase reliability and plausibility of control experiments.
For years, economists have simply adapted to their own problems
mathematical methods derived from other disciplines, in particular
 XIII

engineering. However, the new features introduced into economic

models (e.g. rational expectations), the new hypothesis on the
information available to economic agents (incomplete information
models) and on the behaviour of the policymaker (risk aversion,
strategic planning), have forced economists to derive control
methods explicitly conceived for solving the above problems.
Therefore, the essays included into the second section of the book
are devoted to determining optimal economic policy when:
(1) the policymaker has time-varying targets;
(2) the policymaker wants to minimize not only the mean value of
his loss function, but also its variability. The importance of this
second goal is measured by a risk aversion parameter.
(3) there is not enough a priori knowledge about the functional
form of the policymaker's loss;
(4) there is uncertainty about the parameters of the model;
(5) the model is large and efficient computational methods are
needed.
Another relevant problem is emphasized in Chow's survey
article: standard control methods are based on the assumption that
the policymaker determines optimal policy by maximizing his
welfare function given the economic system, without facing one or
more economic agents who rationally anticipate his decisions and
maximize their own welfare given those expectations. In other
words, the behavioural assumptions underlying standard control
methods are not satisfactory and should be revised in order to
capture the strategic aspects of policy decisions.
Furthermore, poor assumptions about the expectations of the
economic agents hide important methodological and economic
problems like:
(1) the time inconsistency of optimal control strategies and the
consequent sub-optimality of actual policy decisions (see
Kydland-Prescott, 1977);
(2) the possible variability of the parameters of the model as a
consequence of the reaction of economic agents to policy-
maker's decisions (see Lucas, 1976);
(3) the credibility of the announced control strategy;
(4) the inclusion between the policymaker's targets of non-economic
entities like the policymaker's reputation;
(5) the importance of a correct specification of the information
available to the policymaker and the economic agents;
XIV

Finally, strategic assumptions have to be introduced when the

policymaker has to determine his optimal decision without know-
ing which of several rival models is the true model.
The importance of the previous remarks led us to introduce into
the book a section completely devoted to strategic aspects of
economic policy and to game-theoretic solutions of policy pro-
blems. The third section of the book includes indeed articles
which try to solve the above theoretical problems by making
use of game theory results. In particular, time-consistency and
credibility problems are examined and new solutions are pro-
posed. Furthermore, macro and micro models analysed by game-
theoretic methods are presented in order to emphasize the greater
insight on economic problems that can be achieved by using game
theory.
The last area that has to be covered is the applications of control
methods to relevant economic and econometric problems. Appli-
cations of optimal control are indeed particularly difficult when the
structure of the model is complicated or when new theoretical
aspects like expectations or uncertainty are introduced into the
model. Therefore, section four presents articles which apply control
methods to dynamic macro and micro models, and determine the
economic agents' optimal decisions under general hypotheses on
the economic system. Important normative conclusions are then
derived. Furthermore, the econometric implications of control
methods are explored.
Needless to say, each area considered in the four sections of the
book is broad, and some problems were probably left uncovered.
However, we believe to have included into this book the most
relevant developments of control theory for economic analysis and
the amount of papers presented at the symposium is the true
support to our beliefs.
In this Introduction we try to provide a guide to the contents of
the book by taking up briefly some of the major issues considered
by the authors. In doing so we have relied on the discussion at the
symposium, and we try to point out at least some matters on which
further research would be valuable.
Given the nature of the material, this Introduction is divided into
four sub-sections with the same title as the four sections of the
book. A final sub-section will present our interpretation of some
open problems and topics for future research.
 xv

1. Introduction to Control Theory: Methods and Algorithms

The first paper of this section contains Gergory Chow's survey of

the development of stochastic control theory in macroeconomic
policy analysis. The development is separated into three periods.
The first is pre-1970 when the major ideas of policy analysis and of
optimization were formed. The second is the early and middle
1970's when formal stochastic control theory was rapidly developed
for and applied to the study of macroeconomic policy. The third
period, beginning in the late 1970's, was stimulated by the intro-
duction of the idea of rational expectations in economic analysis
and is characterized by the development of new analytical tools for
studying the effects of given macroeconomic policy rules on
economic targets. Gregory Chow's survey is also able to point out
topics where the results available at present are not satisfactory so
that more research effort would valuable.
Roberto Conti's article presents old and new results on control-
lability theory by providing a unifying framework to the theory of
policy effectiveness from a strictly mathematical viewpoint. Both
continuous and discrete time models are considered. Necessary and
sufficient conditons for point and path controllability are provided
and new definitions of system controllability are suggested.
Conti's mathematical approach is translated into economic terms
by Petit's article where the theory of policy effectiveness is reviewed
from Tinbergen's seminal work to the most recent results. In par-
ticular the relationship between Tinbergen's condition for static
controllability and dynamic controllability conditions is explored.
The last paper of the first section of the book is David Kendrick's
review of available software for optimal control models. The paper
contrasts the software of the past with the software of the present
and outlines the likely form of the software of the future. In
so doing it chronicles the evolution of input and output from
numbers to equations to graphics. The evolution of software will be
reviewed both for nonlinear models and for linear-quadratic
problems. The emphasis in put on deterministic models since the
evolution of stochastic software closely parallels that for deter-
ministic models.
Therefore, the four papers of this section cover all the recent
developments of control theory in economics both from a theoretical
and a computational viewpoint. They also clarify hidden problems
XVI

and point out the necessity of new achievements which will be more
extensively taken up in the second section of the book.

2. Recent Developments of Control Theory: Objective Function

Specification

The contributions of this section are mainly theoretical and

provide some interesting generalizations of control methods for
economic analysis. In particular, attention is paid to the specification
of the policymaker's objective function and functional forms more
general and plausible than the quadratic form are proposed. At the
same time, the simplicity and analytical elegance of the feedback
control solution is often preserved.
The first paper contains the description of a new method for
solving control problems. Gruber's paper describes indeed inter-
active vector optimization methods which are shown to avoid the
necessity of explicitly specifying a scalar-valued objective function.
This approach has important advantages in comparison with control
theoretical decision models in which an explicit objective function
must be specified. The advantages are particularly relevant for
applied economists whose main interest is not the structure of the
feedback policy rule but the optimal values of the policy variables
given the specified econometric models. Gruber's paper also describes
Rosinger's algorithm for interactive vector optimization by empha-
sizing those aspects that are likely to be important in applied work.
Brandsma's paper is instead aimed at increasing the robustness of
economic policies by using optimisation techniques which try to
reduce the impact of uncertainty on policy decisions. Unfortunately,
certainty equivalence applied to a quadratic welfare function sub-
ject to linear constraints implies decisions which are invariant to the
magnitude of the risk undertaken by the policymaker. So, while the
importance of risk reduction is widely recognised, economists have
seldom been able to compute empirical risk sensitive decisions for
the multivariable dynamic control problem which they face in
practice. Brandsma's paper is a first attempt in this direction since
it provides decision rules which minimizes both the mean and the
variance of the policymaker's loss function. The paper provides not
only a unifying framework for risk sensitive decision problems but
also numerical comparison of different solution methods.
 XVII

The third paper considers a different generalization of the control

problem. In the linear quadratic control model the target values are
assumed exogenous. In contrast, Amman and Jager determine the
optimal policy rule under the assumption that the target values are
a function of the outcome or state variables of the decision model.
In this way, the specified welfare (loss) function with time-varying
target values can provide a more realistic description of the actual
policymakers's behavior. An application to exchange rate policy
underline the relevance of the extension of optimal control methods
provided by Amman and Jager. The two authors revise indeed
Pindyck's algorithm in order to allow for time-varying endogenous
target values and then compute the optimal economic policy for a
model of the Dutch economy under an exchange rate system
characterized by a crawling peg.
In the fourth essay of the section, Eppers and Leserer consider
the probabilistic aspects of multi-period planning by comparing the
common backward perturbation analysis with forward program-
ming methods. The paper shows the usefulness of forward pro-
gramming in combination with backward programming by studying
the information structure of multi-period planning. The proposed
algorithm is a two-stage procedure: first an adaptive feedback
optimization is done by looking at past information growth and
then adaptive feedforward optimization is done by looking at
future uncertainty reduction. This two-stage two-system approach
can be shown to be very useful for constructing flexible economic
strategies. Some numerical experiments emphasize the characteristics
of the algorithm. Eppers and Leserer's paper can be considered an
important contribution to adaptive stochastic control theory.
Finally, Montesano's paper faces a strictly theoretical problem
by providing a solution to a general intertemporal choice model
under uncertainty. In particular, attention is paid to some strong
hypotheses that are usually implicitly assumed for the objective
function in order to use dynamic optimization methods. A coherent
specification of the objective function for intertemporal choice
problems under uncertainty is then derived.
All the previous papers provide interesting discussions and new
proposals on the specification of the objective function to be
optimized by the policymaker. As an explicit feedback solution to
the control problem often does not exist when the objective func-
tion is not quadratic, it is important to study the properties of
XVIII

available numerical optImIzation methods. This task is accom-

plished by Calzolari and Panattoni whose paper is aimed at
generalising the present knowledge on algorithms commonly used
in nonlinear applied control. When applying control methods to
economic models, the optimization problem is often solved
numerically by using methods based on the computation of the
Hessian matrix (e.g. Newton method). This paper therefore
presents several Montecarlo experiments that have been performed
in order to better understand the behaviour of the Hessian and of
two of its approximations currently used and quoted in the litera-
ture. The results show that this type of optimization procedure
(gradient method) converges rapidly when the values of the coef-
ficients are close to the optimum, while the use of a suitable
approximation to the Hessian strongly improves the algorithm
efficiency far from the optimum and hence its robustness to the
choice of the initial values. Furthermore, Calzolari and Panattoni
show that the Hessian behaves better in an interval around the
optimum which narrower than commonly thought. Even if the
numerical experiments are performed in the case of likelihood
maximization for FIML estimation of econometric models, the
results seem to be relevant for more general optimization problems.

3. Recent Developments of Control Theory: A Game Theoretic

Approach

A large part of economic theory is awesomely dependent on

game theory, borrowing from it the conceptual apparatus for the
analysis of various economic problems. The goal of this section is
to show that game theory is also a useful tool for analysing pro-
blems related to policymaking both at a macro and micro level. The
importance of game theory is due to its ability to facilitate com-
prehension of different phenomena by preserving at the same time
simplicity and mathematical elegance. This is also shown by the
papers of this section which are able to solve difficult problems like
the optimal policy time inconsistency problem or the rival models
problem by using a game-theoretic approach.
The first paper presents Berc Rustem's results on optimal policies
with rival models. The problem to be solved is the following: in
general, more than one model is claimed to represent the economic
 XIX

system. Each model is justified by a different theory and may

represent a different regime of the economy. How can the policy-
maker use the information provided by all the rival models?
Assuming that all the models are correctly specified, a policymaker
may utilize simultaneously more than one of these models to assess
their combined effects or may minimize the effect of the most
adverse bounded outcome due to uncertainty in a given model.
Rustem's paper explores alternative solutions to the rival models
problem and determines a robust policy rule by using min-max
solutions to the policymaker's optimization problem. This approach
involves the simultaneous minimization of the welfare function and
its sensitivity to the sources of uncertainty in the model. Further-
more, a Pareto optimal characterization of policy formulation with
rival models is given and an algorithm is presented for solving the
optimization problem with pooled models. Min-max representations
of the pooling problem are formulate both as discrete and as
continuous min-max problems. Nash abd Stackelberg strategies are
also discussed when multiple models are used by rival agencies
within a game theoretic framework.
Hughes-Hallett contribution concerns instead the determination
of optimal policy when two policymakers act simultaneously and
interdependently on the economic system. Two countries are indeed
considered and the optimal strategy of each country is determined
by computing the Nash solution of the dynamic game between the
two policymakers. It is thus possible to know how much the two
economies would gain if their policymakers cooperated and which
strategy is optimal under different assumption on the strategic
behaviour of the policymakers. Furthermore, the solution tech-
niques utilized in the paper guarantees the time consistency of the
optimal strategy.
The time consistency problem is also the main topic of Carlo
Carraro's paper which analyzes a new solution of the game between
the policymaker and economic agents, the Closed Loop Stackelberg
solution, in order to determine the optimal time consistent economic
policy. The main features of the Closed Loop Stackelberg solution
are examined and it is shown that only if a certain degree of
uncertainty is introduced into the model can the Closed Loop
Stackelberg solution be made credible. Furthermore, it is shown
that the Closed Loop Stackelberg solution of the game can be used
to determine the optimal announced economic policy and conditions
xx
are derived for this announcement to be credible and time-consistent.
In particular, it is shown that the policymaker can induce the
economic agents to behave as if they were acting in the policy-
maker's interest. The effectiveness of the optimal time consistent
strategy is also explored.
The last two papers of the section are microeconomic appli-
cations of game theory results which show the importance of using
the game-theoretic approach for economic analysis. The first paper,
written by Fanchon, Rifkin and Sengupta, presents a dynamic
model of price leadership for a dominant firm, where output rather
than price is the major decision variable. The relationship between
uncertainty, risk and price stability is investigated and a simulation
shows the optimal trajectories resulting from various initial states.
Conrad's paper analyses instead a duopoly model and determines
optimal price and quality strategies under asymmetric information
with respect to quality. Since the model is dynamic and information
is incomplete, the problem of credibility and reputation immediately
arises and is solved under diverse assumptions on the information
structure of the players.

4. Economic and Econometric Analysis by Control Methods

The three papers of this section provide other applications of

control methods to different aspects of economic theory. Their
importance is given by the new insight that can be achieved by
looking at the economic problem through control methods.
In particular, the first essay, written by Charles Tapiero, pro-
vides a system approach to insurance companies management.
The system approach in an insurance setting is broadly stated
and applications to a stock and to a mutual insurance firm are used
to highlight the potential benefits of this approach to insurance
firm's management.
Analogously, the second essay, written by James Gapinski,
computes the optimal feedback policy rule for a policymaker facing
a dynamic economic system and aiming at stabilizing the business
cycle. In particular, the paper studies how the nature of capital
influences the effectiveness of feedback control policy.
Both articles carefully describes the structure of the model and
then show how control theory can determine the optimal behaviour
 XXI

of the economic agents involved into the model. The advantage of

using control theory is shown by the precise normative conclusions
achieved in both papers.
Finally, Richard's paper explores the relationship between con-
trol theory and econometrics by discussing some of the problems to
be faced by an econometrician who wishes to simulate the impact
of new control rules on an economy which has been operating in a
different control environment. In particular, the identification
between control variables and econometrically exogenous variables
is questioned. It is indeed possible that the variables to be controlled
either were not exogenous over the sample period or were not
subject to control over the sample period. These situations, which
correspond to the so called "change of regime", require additional
manipulations (conditioning) of the estimated model before it can
be used for simulation of the new control rules. Richard's paper
provides conditions for properly defining the exogeneity assump-
tions that are necessary for simulating the new control rules
and also provides an application to a UK demand for money
equation.

5. Open Problems and Further Research Topics

The concluding section of this introduction is devoted to an

exploration of open problems in control theory and economic
policy and suggestions for future research.
The first research field to be emphasized is concerned with the
correct specification of the welfare (loss) function to be maximized
(minimized) when solving the control problem. Some proposals
have been presented at the conference and are contained into this
book. However, other theoretical results are necessary for deter-
mining feedback rules derived from the solution of control problems
where the function to be optimized is such to capture the actual
features of the policymaker's goals. In particular, asymmetric
objective functions containing at least two moments of the target
variables should be considered. It is obviously possible to derive
numerical solutions of those control problems; numerical solutions,
however, cannot be used for analysing the structure and the impli-
cations of the optimal feedback rule and are often computationally
expenSIve.
XXII

A second open problem is related to the evolution of stochastic

control methods when there is uncertainty on the values of the
parameters of the model. The available solutions of the adaptive
control problem (e.g. dual control methods) have only been used
for small economic models and seem to be inadequate for more
general models. Furthermore, adaptive solutions should also be
derived when two or more polieymakers are considered and the
control problem is transformed into a game problem.
The game-theoretic approach to economic policy constitutes
another important research field. New solutions of the game between
policymaker and economic agents or between different policy-
makers have still to be completely explored (e.g. incentive solutions,
conjectural vanatIOns solutions). Furthermore, three level
hierarchical games could be considered when more than two policy-
makers are assumed. The implications of incomplete information
games for policy analysis should also be examined in particular
when incentive compatibility problems arise (see Carraro, 1985, for
a first attempt to deal with incomplete information hierarchical
games).
Another important research topic deals with the relationships
between control and econometrics. It is always assumed that
econometric identification at most affects the solution of the control
problem by introducing a stochastic term that has to be treated by
using passive or active learning control solutions, thus excluding
the certainty equivalence solution. However, the goal of using some
variables for control purposes also affects the identifiability of the
model, whereas, on the other side, it is difficult to conceive reliable
control experiments when the parameters of the model are likely to
be policy dependent and are estimated under a different policy
regime. The control problem should therefore be solved by assuming
time varying econometric models where the relationship between
parameters and control variables has been estimated.
Econometrics can also be useful for objective function specifi-
cation. The solution of the inverse control problem can indeed be
transformed into a set of estimation equations for determining the
empirical objective function (see Chow, 1981; Carraro, 1985). How-
ever, "change of regime" problems and statistical testing of the
estimated objective function have still to be solved.
Viceversa, control theory can be useful for econometric analysis
since important information on the specification of the model
 XXIII

(trade-offs, transmission mechanisms, causality directions, policy

effectiveness) can be obtained by using control methods. Further-
more, if the model is nonlinear, controllability, stabilizability and
decouplability of target variables and instrument stability can be
studied by performing appropriate control experiments (some
preliminary results on the analysis of the structure of a large non-
linear model by using a trade-off approach based on control methods
are contained in Calliari, Carraro and Sartore, 1985).
Finally, theoretical and applied work is needed for solving control
problems that take into account the difference in timing between the
decisions of different policymakers or the impossibility for a single
policymakers to move all his instruments at the same time. Further-
more, time-optimal control problems should be considered in
applied work in order to study the best performance of economic
policy in the shortest time-horizon.

References

Calliari, S., Carraro, c., Sartore, D. (1985): "Instruments and Intermediate

Targets of Monetary Policy", Journal of Economic Dynamics and Control, 10
(1986) 175-184.
Carraro, C. (1985): New Methods for Macroeconomic Policy Analysis, Ph.D.
Dissertation, Princeton University.
Chow, G. C. (1975): Analysis and Control of Dynamic Economic Models, J. Wiley
and Sons, Inc., New York.
Chow, G. C. (1981): Econometric Analysis by Control Methods, J. Wiley and Sons,
Inc., New York.
Kendrick, D. (1981): Stochastic Controlfor Economic Models, McGraw Hill, New
York.
Kydland, F. and Prescott, E. C. (1977): "Rules rather than Discretion: The Incon-
sistency of Optimal Plans", Journal of Political Economy, 473--491.
Lucas, R. E. (1976): "Econometric Policy Evaluation: a Critique", in K. Brunner
and A. H. Meltzer (eds): The Phillips Curve and the Labor Market, North
Holland, Amsterdam.
Preston, A. J. and Pagan, A. R. (1982): The Theory of Economic Policy,
Cambridge University Press, Cambridge.
 PART I

INTRODUCTION TO CONTROL THEORY: METHODS

AND ALGORITHMS
 3

CHAPTER 1

DEVELOPMENT OF CONTROL THEORY IN

MACROECONOMICS

Gregory C. Chow
Princeton University, USA

In this paper I will survey the development of stochastic control

theory in macroeconomic policy analysis, The development can
conveniently be separated into three periods. The first is pre-1970
when the major ideas of policy analysis and of optimization were
formed. The second is the early and middle 1970s when formal
stochastic control theory was rapidly developed for and applied to
the study of macroeconomic policy. The third period, beginning in
the late 1970s, was stimulated by the introduction of the idea of
rational expectations in economic analysis. This introduction has a
significant impact on the way stochastic control theory is applied to
studying the effects of given macroeconomic policy rules. The main
ideas in the three periods will be set forth in turn.

1. The Pre-1970 Period

Jan Tinbergen and Ragnar Frisch received the first Nobel Prize
in economic science in 1969 for their contributions to the modelling
of dynamic economic relationships. Macroeconometric models
form the basis for the study of macroeconomic policy and for the
formulation of good policies. Tinbergen (1952, 1956) provided a
conceptual framework for quantitative policy analysis. He dis-
tinguished between economic policies which constitute basic
reforms of economic institutions and those which attempt to
achieve specified objectives given the economic institutions. Quan-
titative analysis was applied to determine the latter kind of policies.
An important ingredient in the analysis is an econometric model
which is used to evaluate the effects of alternative economic policies.

C. Carraro and D. Sartore (eds.) Developments of Control Theory for Economic AnalysIS
© 1987 Martinus NiJhoff Publishers (Kluwer), Dordrecht-
4

An econometric model is a set of quantitative economic relations,

estimated by using statistical data, which will determine the values
of the endogenous variables given the values of the exogenous
variables. The endogenous variables are thus said to be explained
by the model, whereas the values of the exogenous variables are
taken as given. Among the exogenous variables, some are subject to
the control of the government authority. These are known as
control variables or policy instruments in Tinbergen's terminology.
Among the endogenous variables, some are called target variables
by Tinbergen as they measure how well the state of the economy is
from the viewpoint of the economic policy maker. If the policy
maker wishes to achieve given values for n target variables, he needs
to manipulate the values of n policy variables or instruments. An
econometric model enables him to determine the required values of
the instruments. In this context an economic policy consists of a set
of values for the instruments. The question of whether a given set
of targets can be reached by manipulating the control variables is
discussed under the heading of controllability in control theory.
After Tinbergen's early works (1952, 1956), macroeconomic
policy analysis developed in two important directions. First, the
analysis became dynamic and stochastic. Second, an objective func-
tion of the target variables was explicitly introduced, and math-
ematical techniques were invented to maximize the expected values
of the objective function by choosing the values of the instruments.
A basic ingredient in policy analysis is a dynamic, stochastic econo-
metric model.
A dynamic econometric model in discrete time explains the
values of the endogenous variables at period t by the current and
past values of the exogenous variables and the past or lagged values
of the endogenous variables. The latter variables are called
predetermined variables. In addition to the predetermined vari-
ables, random disturbances also enter the econometric model,
making it stochastic. Policy analysis was concerned with the effects
of different time paths of the control variables on the time paths of
the endogenous variables and especially the target variables.
To evaluate the consequences of different macroeconomic
policies, an objective function (or loss function) is specified with the
target variables and possibly some control variables as arguments.
To determine a good policy, an optimization problem is formulated.
It is to maximize the expectation of the value of the objective
 5

function (or to minimize the expectation of the value of the loss

function) of the target and control variables for many periods
subject to the constraint of a dynamic stochastic econometric
model. A solution to this problem for the case of a quadratic
objective function and a linear model was given by H. Simon (1956)
in his Econometrica article "Dynamic Programming Under Uncer-
tainty with a Quadratic Criterion Function." In this article, Simon
proved that the optimal setting of the policy variables for the first
period in a multiperiod optimization problem involving a quadratic
objective function and a linear dynamic stochastic model with
additive disturbances is the same as the optimal setting of the policy
variables for the first period in a modified, deterministic problem
obtained by replacing all stochastic variables in the former problem
by their mathematical expectations.
In the case of a linear dynamic model where the additive random
disturbances have zero expectations, replacement of stochastic
variables by their expectations amounts to omitting the disturb-
ances and converting the model to a deterministic model or to its
"certainty equivalent." Since the variables are no longer stochastic,
the expectation of the objective function to be maximized becomes
the objective function itself. The stochastic optimization problem is
converted to a deterministic optimization problem, its certainty
equivalent. Simon's theorem asserts that the first-period solution to
the original stochastic optimization problem is the same as the
first-period solution to its certainty-equivalent problem. Since the
latter problem is easy to solve, a solution to the former problem is
thus provided by using what Simon calls the "certainty-equivalent"
method. If one desires the optimal solution for the second period,
he awaits the end of the first period and treats a multiperiod
stochastic optimization problem from period two onwards using
the method of certainty equivalence.
H. Theil (1958) recognized the importance of the method of
certainty equivalence in the formulation of macroeconomic policies
and expanded on this subject while the example used in Simon
(1956) was microeconomic in nature, dealing with a problem in
inventory and production control by a factory. Although Simon
(1956) knew of and in fact quoted the work of R. Bellman (1953)
on dynamic programming, Simon's method of solution was dif-
ferent from Bellman's method based on the principle of optimality.
Bellman's method has influenced much of the later work on
6

stochastic control in economics and elsewhere. Simon's method

presumably has also influenced later work in the control literature
and is related to the separation theorem by which an optimal
feedback control rule is applied to the expectation of the state
variables when the latter variables are unknown and can be esti-
mated by the Kalman filter. More on this subject in the next section.

2. Early and Middle 1970s

Partly stimulated by the rapid growth of macroeconometric

modelling activities, interest among economists in quantitative
macroeconomic policy analysis increased rapidly in the late 1960s.
By that time, the relevance of stochastic control theory for the
determination of stabilization policies has become known to econo-
mists. The important works in bringing the ideas of control theory
to the attention of economists include A. W. Phillips (1958) and
P. Whittle (1963). However, both authors treated only univariate
models and applied methods of the frequency domain, rather than
time domain, to finding an optimal feedback-rule in the steady
state. Mention should also be made of the work of Holt, Modig-
liani, Muth and Simon (1963) as one of the first applications of
some of the ideas in control theory to economics.
Policy analysis using multivariate stochastic models and opti-
mization method in the time domain flourished in the 1970s.
Representative of such works are Chow (1970, 1972a and 1972b),
Friedman (1973), and Pindyck (1973). Beginning in 1972, the
National Bureau of Economic Research, with financial support
from the National Science Foundation, sponsored annual work-
shops in control theory and economics where both economists and
control theorists presented papers and exchanged ideas. From 1972
to 1978, the workshops met successively at Princeton University,
the University of Chicago, the Federal Reserve Board, M.LT.,
Stanford, Yale, and The University of Texas. Michael Athans and
Gregory Chow organized the first three conferences. David Kendrick
and Edison Tse organized the remaining ones. Many of the papers
presented were published in a journal of the National Bureau of
Economic Research, Annals of Economic and Social Measurement
(October 1972; January 1974; Spring 1975; Spring 1976; Summer
1976; Spring 1978). In 1979, the Society for Economic Dynamics
 7

and Control was founded, and the Journal of Economic Dynamics

and Control was established. Further annual meetings have been
organized by the Society and many papers in the field have been
published in the Journal.
To continue our discussion of the activities in the 1970s, it is
necessary to specify a system of linear simultaneous structural
equations often used to represent an econometric model:
(1)
where Yt is a vector of endogenous variables, X t is a vector of control
variables, Yt is a vector combining the effects of the exogenous
variables not subject to control, and Gt is a vector of serially uncor-
related disturbances. A higher-order system involving Yt-2 is con-
verted into first-order by suitable definitions. These structural
equations are based on economic behavioral relations, technical
relations, institutional relations, and definitional relations or iden-
tities. The system is simultaneous because the endogenous variables
Yt are interdependent, with several endogenous variables appearing
in one structural equation. The number of simultaneous equations
equals the number of endogenous variables, the matrix B being a
square matrix. Furthermore, B is assumed to be nonsingular, so
that the structural equations solved for the endogenous variables,
yield a system of reduced-form equations:
Yt B-1r1Yt_l + B- 1r 2 x t + B-1Yt + B-1G t
(2)

which corresponds to a model in state-space form in the literature

on control theory. The simultaneous structural equations (1)
represent a special feature not generally present in the control
literature. However, the vector Yt is frequently assumed to be
measured without errors, whereas the corresponding vector of state
variables in the control literature is typically assumed to be
measured with errors.
Let the loss function be quadratic
T
W = L (Yt
t=1
- at)' Kt(Yt - at) (3)

where at is a vector of targets toward which the vector Yt is aimed,

K t is a positive semi-definite matrix, often diagonal, with diagonal
8

elements showing the penalty weights given to the squared devi-

ations of different endogenous variables from their targets, and T
is the planning horizon. When a diagonal element of K, is zero, the
corresponding endogenous variable is not a target variable in Tin-
bergen's terminology. To incorporate some control variables into
the loss function, one simply defines additional elements in Yt as
equal to these control variables. The linear-quadratic optimal con-
trol problem is to minimize the expectation of the quadratic loss
function (3) given the linear dynamic stochastic model (2).
As pointed out in Section 1, by the method of first-period cer-
tainty equivalence of Simon (1956), one can convert the stochastic
model (2) to its certainty equivalent by dropping the random
disturbance VI which has zero expectation. One then solves the
resulting deterministic optimization problem, the certainty equiv-
alent of the original problem. The first-period optimal policy XI is
optimal for the original problem.
By contrast, using the method of Bellman (1953), we define a
value function V(x" YI_I) = Ji; which is the expected loss from
period t to period T, given the initial condition Yt-I of the dynamic
model and the policy XI for period t, and assuming all future policies
X I + 1 , Xt+2' • . . , X T shall be optimal. We first solve the optimal
control problem for the last period T by choosing XT to minimize

(4)

where E T _ 1 denotes the conditional expectation given all infor-

mation up to T - 1, which in this case only includes YT- j , and we
have let KT = H T, KTa T = hT and CT = a~KTaT' After sub-
stitution of (2) for Yr in (4), the solution to this simple problem is

(5)

where

-(C'HTC)-IC'HTA (6)

-(C'HTC)-I(C'HTb T - C'h T) (7)

Equation (5) gives the optimal feedback rule x T for the last period.
 9

The corresponding minimum expected loss for the last period is

VT V(XT' YT-I)
Y~_I(A + CGT),HT(A + CGT)YT-I
+ 2Y~_1 (A + CGT),(HTb T - h T)
+ (b T + CgT),HT(b T + Cg T) - 2(b T + CgT),h T
(8)
To obtain the optimal policies for the last two periods, we observe
that x is already found to yield the minimum loss VT for the last
period and that, by the principle of optimality of Bellman, we need
only to find X T _ I to minimize
VT_I = E T- 2[(YT-I - aT-dKT-I(YT-I - aT-I) + VT
= ET-2(YT-IHT-IY~-1 - 2Y~_lhT_I + C T_ I ) (9)

where, using expression (8) for VT , we have defined

H T_ I K T_ I + (AT + CTGT),HT(A T + CTG T), (10)

hT_ I = KT_Ia T_ I - (AT + CTGT),(HTb T - hT), (11)

CT_ I a~_IKT_IaT_I + (b T + CTgT),HT(b T + CTg T)

-2(b T + CTgT),h T + CT + ET_IV~HTVT· (12)

Because the second line of (9) is identical with the last expression of
(4) with Treplaced by T - 1, the solution for XT_I is identical with
(5) with T replaced by T - 1, where GT_ I and gT-I are defined by
(6) and (7) respectively with a similar change in time subscripts.
Accordingly, VT _ I will be given by (8) with the subscripts T replaced
by T - 1.
When we attempt to solve the problem for the last three periods,
we observe that xT and XT _ I have been found that would yield the
minimum expected loss VT -I for the last two periods and that, by
the principle of optimality, we need only minimize
VT- 2 = E T- 3 [(YT-2 - aT- 2)'KT- 2(YT-2 - a T- 2) + VT_ I ]
with respect to X T - 2 , and so forth. At the end of this process, we find
XI = Glyo + gl as the optimal policy for the first period, and the
associated minimum expected loss VI for all periods (or from
10

period 1 onward). Computationally, we solve (6) and (10) with

t replacing T for GI and H t backward in time, for t = T, T - 1,
... , 1. We then solve (7) and (11) with t replacing T for gt and hi
backward in time, for t = T, T - 1, ... , 1. Finally, solution of
(12) with t replacing T backward in time yields C l which is used to
evaluate C; given by (8), with 1 replacing T.
The contrast between Simon's certainty equivalent method and
the method based on Bellman's principle of optimality should be
stressed. The former solves a deterministic optimization problem,
obtains a deterministic optimal time path Xl' ... , XT , which is an
open-loop policy, and applies only Xl to the original stochastic
control problem. The latter yields a set of optimal feedback control
equations XI = GIYI~1 + gl (t = I, ... , T) which are valid for all
T periods. Thus the solution is in a closed-loop feedback form.
When these feedback control equations are combined with the
original model (2), the combined equations determine the dynamic
behavior of all XI and YI through time, and the dynamic properties
of the system under optimal control can be conveniently studied.
Furthermore, since the same feedback control equations for XI
would result in the certainty equivalent model obtained by omitting
VI in (2), the solution in feedback forms obeys the principle of
multi period certainty equivalence. That is, the certainty equivalent
solution in feedback form is optimal for the original stochastic
control problem for all T periods, and not just the first period.
To relate the above analysis to the literature of the control
engineers, the typical presentation in the latter employs instead of
(3) a loss function of the form
T T
W* = L (YI - alt)' KII(YI - alJ + L (XI - a 2tY K 21 (x l - a2J
1=1 1=1
(13)
and requires that K21 be positive definite. An example of its adop-
tion in economics is Pindyck (1973). This presentation not only
complicates the algebra and requires a more complicated set of
equations than (6), (7), (10), (11) and (12) for the solution, but it
also entails the unnecessary assumption that K2t be positive definite.
If YI is defined to incorporate Xl' the matrix KI in (3) includes both
matrices KII and K21 in (13). A unique optimal solution can be
obtained when C' K( Cis nonsingular, which does not require that
K21 be nonsingular.
 11

In the engineering literature, the state vector YI is often assumed

to be measured with an error. A typical formulation is to assume
that the observed vector SI is related to the true vector YI by

(14)

where Yfl is a random vector with zero mean and covarance matrix
Q and is statistically independent of the vector VI in (2). By the
Kalman filter of Kalman (1960), one can estimate the mean and
covariance matrix of y, recursively from the mean and covariance
matrix of YI_I and the new observation SI' Given the mean and
covariance matrix of Yo at time 0, successive observations SI' S2, . . .
are used to estimate the means and covariance matrices of YI,
12, .... The optimal policy is given by the same feedback rule as
before, except that in the feedback control equation for XI' YI_I is
replaced by £I-IYI-I as estimated by the Kalman filter. This result
is known as the separation theorem since the estimation of £t-IYI-I
can be performed separately from the derivation of the optimal
feedback rule. It is also related to the optimality of the certainty-
equivalent method since the optimal feedback rule is the same as the
rule derived when the state vector YI is known for certain.
Kalman filtering has found many applications in econometrics
and in macroeconomic policy analysis. One of the most important
is the estimation of parameters which are time-varying. The
simplest example is a regression model

(15)

where the vector of regression coefficients PI satisfies

(16)

Equation (15) corresponds to the observation equation (14) in the

Kalman filtering setup, with the dependent variable yi correspond-
ing to the observation SI in (14). Equation (16) corresponds to the
dynamic model (2) with XI and b l absent and with the state vector
YI corresponding to the unobserved time-varying coefficient vector

PI'
In macroeconomic applications of optimal control methods, one
finds that frequently the simultaneous-equations models used
are nonlinear. Let us write the ith structural equation in such a
12

model as
(17)
where, as in equation (1), Yt and Xt denote vectors of endogenous
and control variables respectively; 8 11 is a random disturbance and
)'11 is a vector consisting of exogenous variables not subject to
control and parameters. A system of nonlinear structural equations
can be written as
(18)
with <I> denoting a vector function. Analytical solution to the prob-
lem of minimizing the expectation of (3) subject to the nonlinear
model (19) is difficult to obtain. In econometric practice, two
approximate solutions have been proposed, one taking a feedback
form and the other an open-loop form.
To obtain an approximately optimal feedback rule, one can work
with a certainty equivalent model by omitting the random term 8 t
in (18), linearize the resulting model, and find the optimal feedback
rule for the linearized model. Specifically, after dropping 8 t from
(18) and given the initial state Yo and the vector At for t,
t = 1, ... , T, one assumes a tentative policy path x~ and solves
the certainty-equivalent model
Yt = <I>(YI'Yt_"xI'At) (19)
for Yt. The solution l satisfies
y~ = <I>(y~, Y~_I' x~, )'t) (20)
A popular method to solve the nonlinear simultaneous equations
for YI is the Gauss-Siedel method. One then linearizes the certainty-
equivalent model about the path (y~, x?) to yield
YI - y~ = <l>lt( YI - y~) + <l>2t( Y{-I - Y~-I) + <l>3{(X - x?) + 81
(21)
to which we have added the random disturbance 8{. Solving the
linearized structural equations (21) one obtains a linearized reduced
form
(22)
As before, optimal feedback rules to minimize the expectation of (3)
 13

subject to (22) can be derived, which are the same as (5), (6), (7),
(10), (11) and (12) with suitable time subscripts added to A and C.
Once the optimal rule for XI is derived, a new tentative path for x~
and y~ can be calculated for the certainty-equivalent model. A
second optimal rule for XI can be derived, and the iterative process
goes on until convergence. The resulting feedback rule is optimal
for the nonlinear certainty-equivalent model (19), as it can be easily
shown (Chow, 1975, pp. 289-295). It is presumably nearly optimal
for the original model (18).
A second approach is open-loop. Using the nonlinear certainty-
equivalent model (19), a solution y~ corresponding to any open-
loop policy x~ (t = 1, ... , T) can be calculated by the Gauss-
Siedel or another numerical method. The loss associated with
(l, x~) can be evaluated using the loss function. Thus loss is a
function of the open-loop policy. This function can be minimized
with respect to XI' • . . , X T by a gradient method. This resulting
open-loop solution is optimal for the certainty-equivalent model. It
is identical to the solution obtained by the first approach in the
certainty case since both are optimal (provided that the solution is
unique). Computationally, the second approach is efficient if the
number of control variables and the number of periods T in the
planning horizon are small, the total number of variables in the
minimization problem being the product of the two. As the number
of periods T is increased, the number of variables increases propor-
tionally by the second approach and the computational time in-
creases more than proportionally, whereas, by the first approach,
the computational time increases only proportionally. Having
available the feedback control equations obtained by the first
approach, one can conveniently study the dynamic properties of the
model under control.
If uncertainty is introduced by retaining the random disturbance
ef' both the closed-loop feedback approach and the open-loop
approach have to be modified. The former is discussed in Chow
(1976) and the latter is discussed in Fair (1974). In the latter case, one
has to minimize the expectation of the loss function with respect to
XI' . . . , X T where the expectation may be evaluated by stochastic
simulations of the nonlinear stochastic model (18). If uncertainty is
further introduced by allowing for the errors in estimating the
parameters of the model (18), we have a dual control problem. The
term dual control represents two aspects of setting the optimal
14

control policy, one to steer the state vector Yt to its target and the
second to set x t to improve the future observations of X t and Yt for
the purpose of obtaining better estimates of the parameter, thus
improving future control of Yt. Much work on dual control has
been done in the control literature. In the economics literature, the
works include MacRae (1972), Tse (1974), Chow (1975, Chapter
11), Kendrick (1981) and Norman (1984), among others

3. Late 1970s Onward

One important idea in econometrics which attracted much atten-

tion from the late 1970s on is that of rational expectations. It has
long been recognized that expectations have important influences
on economic behavior. Therefore, in econometric models, variables
to measure expectations are important. One popular hypothesis
concerning the formation of expectations up to the middle 1970s is
the adaptive expectations hypothesis, originally applied by Cagan
(1956). It assumes that economic agents adjust their expectations by
a fraction of the difference between the actual value of the variable
observed and the expectation formed in the last period, implying
that expectations are weighted averages of past observed values. By
contrast, the rational expectations hypothesis of Muth (1961)
assumes that the expectations variables entering an econometric
model are the same as the mathematical expectations of the corre-
sponding variables evaluated by the same model. The hypothesis is
based on the assumptions that the economic agents and the econo-
metric model builder share the same model and that the former will
use this model to form their expectations. Although these assump-
tions have been questioned, they are very powerful and may serve
as a good approximation to economic behavior in the long run
when knowledge of the economic environment on the part of
economic agents can be taken for granted. The rational expec-
tations assumption was proposed by Muth (1961), but its popu-
larity in economic research increased rapidly only since the middle
1970s.
If behavioral relations in econometric models are assumed to
result from dynamic optimization by economic agents given their
stochastic environment, they can be interpreted as optimal decision
rules derived from an optimal control problem facing the agents.
 15

Let (2) be a linear stochastic model facing certain economic agents

and (3) be their objective function, equation (5) with t replacing T
is the optimal rule describing their behavior and is a structural
equation in a simultaneous-equations model - the decision vari-
ables XI would become a function of current state variables YI rather
than YI_I, if X I _ 1 instead of XI entered the dynamic model (2). As an
example, XI may represent investment expenditures by business
firms and YI may represent capital stock, the rate of interest and
other variables affecting investment, all these variables appearing in
the investment equation (3) with YI_I replaced by YI. As another
example, XI may represent consumption expenditures and YI
includes income and other explanatory variables for consumption
expenditures.
Lucas (1976) made the important point that government policy
rules are often equations describing the stochastic environment
facing economic agents, i.e., equations (2) in the dynamic optimiz-
ation problems of the economic agents. Therefore, if government
policy rules change, the decision rules (5) describing the behavior of
economic agents will change. It is in this sense that behavioral
equations in traditional econometric models will change as govern-
ment policy changes. Lucas (1976) criticized the traditional method
of evaluating the effects of alternative government policies by
assuming a univariant set of behavioral equations, arguing that
these equations will change as the policy rule changes.
If Lucas' criticism is accepted, a correct approach to evaluating
the effects of alternative government policy rules is to derive the
behavioral equations of the economic agents corresponding to the
different rules by the method of optimal control. The control vari-
ables in such optimization problems are the decision variables
of the economic agents - the investors, the consumers, etc., rather
than the government policy makers. Specifically, let the stochastic
environment facing the economic agents be represented by a linear
model.
(23)
where Xl! is the control variable subject' to the control of the
economic agents, and X 21 are the control variables of the govern-
ment. A government policy rule is represented by
(24)
16

Given this policy rule, the environment facing the economic agents
becomes
YI (A + C 2G2)YI_1 + Clx lt + bl + C2g 21 + VI

AIYI_I + Clx lt + bit + VI' (25)

If the objective function of the economic agents can be represented
by
T
L
1=1
(Yt - all)' KII(YI - all) (26)

the equations describing their behavior are obtained by minimizing

the expectation of (26) given the model (25), yielding
(27)
where the coefficient matrix GIl may become time invariant, under
circumstances described in Chow (1975, pp. 170-172). The behav-
ioral equations (27) are affected by the government policy rule (24).
Using this optimal control framework, one can evaluate the econo-
mic effects of alternative government policy rules, allowing for the
effects of these rules on the behavioral equations (27) in econo-
metric models. (See Chow, 1981, pp. 257 ff. for an example.) Such
framework has been applied by Sargent (1979), Hansen and Sar-
gent (1980), Chow (1980, 1983), Blanchard (1983), among others.
Once the above optimal-control framework can be applied to
evaluate the economic effects of alternative government policy
rules, it is possible for the government to choose an optimal rule to
minimize the expected value of its own loss function, anticipating
the effects of its rule on the behavior of the economic agents (see
Chow, 1981, Chapter 17). This is a problem in dynamic games with
the government as the dominant player. The solution is stable only
if the government adheres to its announced policy, and the private
sector finds the policy credible. The eagerness of an unemployed
worker to seek work depends on the amount of unemployment
compensation which the government pays. By adhering to a limited
amount and duration of unemployment compensation, the govern-
ment may encourage people to seek work. However, when a large
number of people are unemployed, it is tempting for the govern-
ment to change to a more lenient compensation policy. If workers
anticipate a change of policy in hard times, they will be less eager
 17

to seek work. Such a problem has been discussed by Kydland

and Prescott (1977) and Miller and Salmon (1984) as a "time-
inconsistency" problem - the optimal lenient policy in hard times
is not an optimal long-run policy for the government if the private
sector anticipates it. On the other hand, if the government adheres
to its announced policy and the private sector believes in it, it is
possible for the government to find an optimal policy.
From the viewpoint of estimation of econometric models, the
important parameters are in the loss function (26) and the stochas-
tic model (23). Given (26), (23) and a government decision rule (24),
an econometrician can evaluate the economic effects of government
policies. The parameters of (26) and (23) correspond to the struc-
tural parameters in the simultaneous-equations models proposed
by the Cowles Commission at the University of Chicago in the late
1940s and early 1950s. As Marschak (1953) pointed out, to evaluate
the effects of economic policies, an econometrician needs to know
the parameters in a system of structural equations, and not just the
reduced-form equations which will change as policy changes. In the
present context, the econometrician needs to know the "structural"
parameters in (26) and (23), and not just the parameters of the
"reduced-form" equation (27) which will change as policy changes.
In the simultaneous-equations case, the "reduced-form" equations
are derived from the structural equations by solving algebraic
equations. In the present case, the "reduced-form" equations are
derived by solving an optimal control problem, but the basic ideas
are identical.
The above approach to evaluating the effects of economic poli-
cies is by no means generally accepted by the economics profession.
Many economists feel that the underlying assumptions concerning
the rationality and knowledge of economic agents and the stability
of government policies are too strong and that the econometric
assumptions required to make the approach operational are too
severe. Examples of the skeptics are Blinder (1983) and Wan (1984).
As of the early 1980s the majority of practicing econometricians are
still applying the approaches discussed in Section 2 to evaluate
economic policies by assuming a fixed econometric model in the
form of a system of simultaneous equations. The policies in which
they are interested are specific government actions in the short-run,
rather than policy rules which can be regarded as permanent. The
research described in Section 2 continues through the 1980s as
18

reported in various economic journals, especially the Journal of

Economic Dynamics and Control.
In this paper I have reviewed certain major ideas in the develop-
ment of control theory in macroeconomic policy analysis using
econometric models in discrete time. The applications of control
theory to econometric models in continuous time, and to micro-
economic analysis, are beyond the scope of this paper.

References

Bellman, Richard (1953): An Introduction to the Theory of Dynamic Programming.

Santa Monica: The Rand Corporation.
Blanchard, Olivier J. (1983): "The Production and Inventory Behavior of the
American Automobile Industry," Journal of Political Economy, 91.
Blinder, Alan S. (1983): " Skeptical Note on the New Econometrics." Cambridge:
National Bureau of Economic Research, Working Paper No. 1092.
Chow, Gregory C. (1970): "Optimal Stochastic Control of Linear Economic Sys-
tems," Journal of Money, Credit and Banking, 2, 291-302.
Chow, Gregory, C. (1 972a): "Optimal Control of Linear Econometric Systems
with Finite Time Horizon," International Economic Review 13 (I), 16-25.
Chow, Gregory C. (1972b): "How much could be Gained by Optimal Stochastic
Control Policies?" Annals of Economic and Social Measurement, 1 (4), 391-406.
Chow, Gregory C. (1975): Analysis and Control of Dynamic Economic Systems.
New York: John Wiley & Sons, Inc.
Chow, Gregory, C. (1980: "Estimation of Rational Expectation Models," Journal
of Economic Dynamics and Control, 2, 241-256.
Chow, Gregory C. (1981): Econometric Analysis by Control Methods. New York:
John Wiley & Sons, Inc.
Chow, Gregory C. (1983): Econometrics. New York: McGraw-Hill Book Com-
pany.
Committee on Policy Optimisation, R. J. Ball, Chairman (1978): Report. London:
Her Majesty's Stationery Office.
Friedman, Benjamin M. (1973): Methods in Optimization for Economic Stabiliz-
ation Policy. Amsterdam: North-Holland Publishing Company.
Hansen, Lars P. and Thomas J. Sargent (1980): "Formulating and Estimating
Dynamic Linear Rational Expectations Models," Journal of Economic Dynam-
ics and Control, 2, 7-46.
Holly, Sean, Berc, Riistem and Martin B. Zarrop, eds. (1979): Optimal Control
for Econometric Models: An Approach to Economic Policy Formulation. London:
The Macmillan Press Ltd.
Holt, Charles c., Franco Modigliani, John F. Muth, and Herbert A. Simon
(1963): Planning Production, Inventories and Work Force. Englewood Cliffs:
Prentice-Hall.
Kalman, R. E. (1960): "A New Approach to Linear Filtering and Prediction
Problems," Journal of Basic Engineering Trans., ASME, 82D, 33-45.
 19

Kendrick, David (1981): Stochastic Control for Economic Models. New York:
McGraw-Hili Book Company.
Lucas, Robert 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 Series on Public Policy, Vol. 1. Amsterdam:
North-Holland Publishing Company.
MacRae, R. C. (1972): "Linear Decision with Experimentation." Annals of
Economic and Social Measurement, 1,437-448.
Marschak, J. (1953): "Economic Measurements for Policy and Prediction," in
W. C. Hood and T. C. Koopmans (eds.). Studies in Econometric Method,
Cowles Commission Monograph 14. New York: John Wiley & Sons, Inc.
Miller, M. and M. Salmon (1984): "Dynamic Games and Time Consistent Policy
in Open Economics." Mimeo.
Muth, J. F. (1961): "Rational Expectations and the Theory of Price Movements,"
Econometrica, 29, 315~335.
Norman, A. L. (1984): "Alternative Algorithms for the MacRae OLCV Strategy,"
Journal of Economic Dynamics and Control 7, 21~38.
Phillips, A. W. (1958): "The Relation Between Unemployment and the Rate of
Change of Money Wage Rates in the United Kingdom, 1861~1957," Economica,
25, 283~299.
Pindyck, Robert S. (1973): Optimal Planning for Economic Stabilization. Amster-
dam: North-Holland Publishing Company.
Sargent, T. J. (1979): Macroeconomic Theory. New York: Academic Press.
Simon, H. A. (1956): "Dynamic Programming Under Uncertainty with a Quad-
ratic Criterion Function," Econometrica, 24 (I), 74-81.
Theil, H. (1958): Economic Forecasts and Policy. Amsterdam: North-Holland
Publishing Company.
Tinbergen, Jan (1952): On the Theory of Economic Policy. Amsterdam: North-
Holland Publishing Company.
Tinbergen, Jan (1956): Economic Policy: Principles and Design. Amsterdam:
North-Holland Publishing Company.
Tinbergen, Jan (1969): "The Use of Models: Experience and Prospects," Noble
Prize Lecture delivered in Stockholm, December 1969; American Economic
Review, 71 (6), 17~22.
Tse, Edison (1974): "Adaptive Dual Control Methods," Annals of Economic and
Social Measurement, 3, 65-68.
Wan, Henry Y., Jr. (1983): "The New Classical Economics ~ A Game-Theoretic
Critique." Mimeo. Forthcoming in George Feiwel (ed.), Issues in Macroecon-
omics and Distribution. New York: The Macmillan Co.
Whittle, P. (1963): Prediction and Regulation by Linear Least Square Methods.
New York: D. Van Nostrand Co.
 21

CHAPTER 2

LINEAR CONTROLLABILITY RESULTS AND OPEN

QUESTIONS

R. Conti
University of Florence, Italy

o
Many models in Economics, as well as in other disciplines, can be
represented by a family of differential equations
dx/dt = Ax + c (A, c)
depending on a parameter c, called control.
The choice of the control within a given set determines the
evolution of the state x in time t, once the dynamics, i.e. the operator
A, and an initial state XO are given. Therefore (A, c) is called a
control process.
For a given XO the set of points x(t, xO, c) corresponding to all the
t > 0 and all the possible choices of c is a reachable set, or, more
precisely, the set of points reachable along the trajectories of (A, c)
issuing from XO at time t = O.
The study of some properties of the reachable sets, generally
known as controllability, is the object of the present talk.

We shall confine ourselves to the finite dimensional case.

More explicitly, the values of x and c will belong to the real
euclidean n-dimensional space [Rn and A will be represented by a
real n x n matrix, so that CA, c) represents a family of ordinary
differential equations in [Rn.
We shall assume that every c is a function t 1---+ c(t) of t ~ 0,
measurable and locally integrable (Lebesgue) so that the value

C. Carrara and D. Sartore (eds). Developments of Control Theory for Economic Analysis
© 1987 Martinus Nijhoff Publishers (Kluwer), Dordrecht-
22

x(t, xO, c) of the solution of (A, c) at time t for each XO and c, is

represented by the Lagrange formula

x(t, xO, c) = e A XO + f~ e(t-s)A c(s) ds (1.1)

where

I, identity,

as usual.
From (1.1) it is clear that without some restrictions on the choice
of c every reachable set would be the whole of IRn. So we shall
assume that every value c(t) has to belong to some proper subset f
of IRn . On the other hand it is also clear that f cannot be empty and,
in fact, it must consist of two points at least.
From now on we shall speak of the control process (A, c) in terms
of the pair (A, r).

According to the Lagrange formula the set of points which can

be reached at time t starting from 0 is the set

W(t, A, r) = {f~ e(t-s)A c(s) ds: c(s) E f}.

We shall say that (A, f) is completely controllable if
W(t, A, r) = IRn for some t > 0
holds.
The problem of determining all the pairs (A, r) for which (C()
holds was recently solved by R. M. Bianchini [I] who proved the
following

THEOREM 2.1: The pair (A, r) is completely controllable if and only

if for every A E IR there exists a positive integer meA) such that
co[f + (A - AI)r + ... + (A - Al)m(A)r] = IRn (2.1)
(where co S = convex hull of S).
 23

When (Cl ) becomes

W(t, A, r) = [Rn, '<:It > 0
then (A, r) can be said to be instantly controllable.
We have (R. M. Bianchini, lac cit.)

THEOREM 2.2: The pair (A, r) is instantly controllable if and only if

co[r + (A - AI)r + ... + (A - AIy-lr] = [Rn (2.2)

holds.

R. M. Bianchini [2] also gave different characterizations of

instantly controllable pairs.
In the particular case when

where B is a real n x m matrix, (2.2) reduces to the well known

"rank condition"
rank [B A B . . . An - I B] = n. (2.3)

Complete controllability means, actually, that it is possible to go

from any point XO (not only the origin 0) to any other point Xl along
a trajectory of (A, c) during an interval of time which does not
depend on the pair XO , Xl .
So (CI ) is a nice property. Unfortunately it requires that the set
r be unbounded, an assumption which is quite unrealistic for
applications, not only in Economics.
This accounts for the introduction of other types of controllability
which do not require from the "reservoir" r to be an unbounded
one.
To this effect let us replace Wet, A, r) by the union of all
Wet, A, r) with respect to t > 0, i.e., let us consider the set
W(A, r) = U Wet, A, r)
1>0

of points which can be reached from 0 "soon or later", i.e., at some

time which depends on the point itself.
24

When
W(A, r) = [Rn

holds then every point can be reached from O. Note that this does
not mean that every point can be reached from any other point like
in the case of complete controllability.
Property (R)o can be labelled as global reachability from O.
Changing A into - A we have

which is a different property from (R)o.

From the Lagrange formula it is easy to see that (T)o holds if and
only if the origin 0 is reachable from every point of [Rn, i.e., if and
only if every point of [Rn can be transferred into O. For this reason
(T)o can be termed as global transferability into O.
When both (R)o and (T)o hold, then we can go from any X O E [Rn
into any Xl E [Rn along a trajectory of (A, c), the time of transfer
depending on the pair X o, Xl . In this case, i.e., when
W(A, f') = We-A, r) = [Rn (C)
holds we can say that (A, r) is globally controllable.
Obviously (R)o and (T)o imply, respectively, that 0 E int W(A, r)
and 0 E int W( - A, r), where int S = interior of S.
A remarkable fact is that, although (R)o and (T)o are independent
each other, the two properties 0 E int W(A, f'), 0 E int W( - A, r)
are equivalent (see, for instance, R. Conti [3]) so that both (R)o and
(T)o imply
o E int W(A, r) n int W( - A, r). (C)O,loc
When (C)O,loc holds we can say that (A, r) is locally controllable
around the origin 0, meaning that there exists an open set S contain-
ing 0 such that we can go from any point of S into any point of S
along a trajectory of (A, c),
The question arises: what condition should be added to (C)O,loc to
obtain (R)o?
One such additional condition is represented (L. A. Kun [4]) by
Re A ~ 0, VA E a(A) (3.1)
(Re A = real part of A, a(A), the spectrum of A).
 25

Actually, (3.1) is easily shown to be also a necessary condition for

(R)o to hold when

r is bounded,
but (3.1) is no longer necessary when r is unbounded, as examples
show.
Therefore it remains an open question that of determining some
condition, let us call it (x), less restrictive than (3.1), such that
(C)O,loc plus (x) ~ (R)o,
independently of (HI).

The next problem is that of characterizing local controllability

around 0, i.e., that of determining all the pairs (A, r) for which
(C)O,loc holds.
To this purpose let us first remark that (see, for instance, R. Conti,
loco cit.)
(1) 0 E int W(A, r) (if and) only if 0 E int Wet, A, r) for some
t > 0;
(2) Wet, A, r) is a convex set whose dimension is independent of
t;
(3) the dimension of Wet, A, r) is = n if and only if we have
Y E en, y* A = AY, y*y = const. Vy E r ~ y = 0
(a)
where en is the n-dimensional complex euclidean space and y*
is the transpose of y.
It follows that (C)O,loc implies (a).
When r = BO., B a real n x m matrix again, n a subset of [Rm,
then (a) reduces to the rank condition (2.3).
If
o E intO (4.1)
holds then (2.3) is also sufficient for (C)O,loc to hold.
However it was noticed first by H. Saperstone and J. Yorke [5]
that the assumption (4.1) is too much restrictive for the applications
26

to Mechanics. Later on R. F. Brammer [6], by adapting an econ-

omic model of Paul A. Samuelson, pointed out that (4.1) is too
restrictive also for applications to Economics.
Since 1971 several papers appeared (H. Saperstone [7], V. I.
Korobov-A. P. Marinic-E. N. Podol'skii [8], M. Heymann-
R. 1. Stern [9], V. I. Korobov [10)) with the aim of characterising
local controllability under weaker assumptions than (4.1).
What is presently known (see, for instance, R. Conti, loco cit.) is
that if
o E clcor
(cl S = closure of S) holds then (C)o lac is equivalent to (a) plus the
condition
Y E [Rn, y*A = Ay, y*y ~ 0, 'r/y E r => y = O. (4.2)
However simple examples show that (H2 ) is by no means a
consequence of (C)O,loc' Therefore it is another open question that of
determining some condition (x) more general than (4.2), such that
(a) plus (x) <=> (C)Oloc

independently of (H2)'

As we remarked already (C)O,loc is equivalent to the existence of

some t > 0 such that
o E int W(t, A, r), 0 E int Wet, - A, r) (5.1 )

hold. It may happen, in particular, that (5.1) hold for all t > O.
This means that for each t > 0 there is an open set Qo,{ contain-
ing 0 such that we can go from any point XO E Qo,{ to any other point
Xl E 0 0,/ along a trajectory of (A, c) in time ~ 2t.
Some authors call O-autoaccessibility this special kind of local
controllability. A characterization of pairs (A, r) having this
property is the following, due to R. M. Bianchini [11]. Let con r
denote the conic hull of r. Then (A, r) is O-autoaccessible if and
only if the pair (A, con r) is instantly controllable, i.e., if and only
if (2.2) holds with r replaced by con r.
 27

According to the Lagrange formula the set of points which can be

reached from a given x E [Rn is
Wet, A, f, x) = elA x + Wet, A, f).
If x E int Wet, A, f, x) then it happens that there is an open set
QX,Icontaining x such that it is possible to reach any point of QX,I
from any other point of QX,I in time ~ 2t.
Let
C(t, A, f) = {x E [Rn: x E int W(t, A, f, x)}.
Then we can define
C(A, f) = U C(t, A, f)
1>0

as the controllability set of (A, f).

It can be shown (R. M. Bianchini [11]) that C(A, f) is a convex,
open set, possibly empty. When
C(A, f) =I ¢
holds we can say that (A, f) is locally controllable.
Obviously (C)O.loc means that 0 E C(A, f), so that
(C)O,loc => (C)loc
(C)loc implies that the dimension of Wet, A, f) is = n, so that (C)loc
implies (a).
On the other hand it is not difficult to show that (C)loc implies
also
there exist x E [Rn: - Ax E reI int co f (b)
where reI int S is the interior of the convex set S relative to the affine
hull of S.
Conversely (see, for instance, R. Conti, loco cit.), (a) plus (b)
imply (C)loC' so that we have a characterization of local control-
lability represented by
(C)loc -= (a) plus (b).
The set
C'(A, f) = n CCt,
1>0
A, f)
28

can be called the autoaccessibility set of (A, r). In fact,

x E C'(A, r) means that for every t > 0 there is an open set nx •t
containing x such that any point of nx•t can be reached from any
other point of nx t in time ~ 2t. If C'(A, r) f:. ¢ we can say that
(A, r) is autoaccessible.
Obviously C'(A, r) c C(A, r) and it is easy to show that
C(A, r) f:. ¢ ~ C'(A, r) f:. ¢
so that (A, r) is auto accessible if and only if it is locally controllable.

All the kinds of controllability examined so far have one feature

in common: the existence of some trajectory of (A, c) going "exactly"
from one point into another. It is of interest, however, to consider
the possibility of going "approximately" from one point into
another.
To this purpose let us go back to (R)o. Obviously (R)o implies
cl W(A, r) = [Rn (7.1)
but, conversely, (see, for instance, R. Conti, loco cit.) (7.1) implies
(R)o. This means that going from 0 to any point of [Rn is just the
same as going from 0 into an arbitrary neighborhood of any point
of [Rn.
On the other hand if (R)o holds this obviously means that it is
possible to go from points in an arbitrary neighborhood of 0 to any
point of [Rn. But the converse is not true.
In other words, let W(A, r) denote the set of points which can
be reached from points arbitrarily close to O. More precisely
WE W(A, r) if and only if there are sequences {td, {c k }, {vd such
that

Obviously
W(A, r) c W(A, r), VA, r
so that (R)o implies
W(A, r) = [Rn.
 29

Simple examples show that (R)o does not imply (R)o' Take for

instanceA = (-1001),r = {(y, O):y E IR}. Then W(A, r) = 1R2,

v

whereas W(A, r) = {(WI' 0): WI E IR}.

It is an open question that of determining some condition (x) such
that
(R)o plus (x) <=> (R)o
Obviously
(R)o => (R)O,loc
where
oE int W(A, r). (R)O.loc
The problems of characterizing (R)o, (R)o,loc have been solved by
V. I. Korobov-Nguyen Khoa Son [12] under the additional
assumptions (HI) and (H2 ) on r.
Characterizations without such assumptions are not yet known.
As a last remark we note that not only
We-A, n = IRn
is independent of (R)o, but also (R)O,loc and
o E int W( - A, n
are independent each other. This is shown for instance by the
previous example where W(A, n
= 1R2, but W( - A, = n
{(WI, O):WI E IR}.

References

1. R. M. Bianchini, Complete Controllability, to appear lOTA.

2. R. M. Bianchini, lOTA 39 (1983), 237-250.
3. R. Conti, Processi di controllo lineare in [R", to appear in the series "Quaderni
dell'Unione Matematica Italiana".
4. L. A. Kun, Avtom. i. Telem. 10 (1977),12-15, (Russian); Eng!. trans!' Automat.
Remote Control 1978, No.5, Part 1,720--725.
5. S. H. Saperstone and 1. Yorke, SIAM J. Control 9 (1971), 253-262.
6. R. F. Brammer, SIAM l. Control to (1972), 339-353.
7. S. H. Saperstone, SIAM J. Control 11 (1973),417-423.
8. V. I. Korobov, A. P. Marinic and E. N. Podol'skii, Differ. Uravn. 11 (1975),
1967-1979 (Russian).
30

9. M. Heymann and R. J. Stern, J. Math. Anal. Appl. 52 (1975), 36-41.

10. V. I. Korobov, Differ. Uravn. 15 (1979), 1592-1599 (Russian).
11. R. M. Bianchini, SIAM J. Control & Optim. 21 (1983), 714-720.
12. V. I. Korobov and Nguyen Khoa Son, Differ. Uravn. 16 (1980), 385-404
(Russian): Eng!. trans!' Differ. Equations 16 (1980), 242-248.
 31

CHAPTER 3

A SYSTEM-THEORETIC APPROACH TO THE THEORY

OF ECONOMIC POLICy l

Maria Luisa Petit

University of Rome, Italy

1. Introduction

The relationship between Tinbergen's conditions and dynamic con-

trollability conditions has been considered in a large number of
papers which have followed Preston's 1974 pioneering article. In
particular a recent paper by Wohltmann [25] thoroughly analyses
this relationship for the discrete time case. The continuous time
case, however, in our opinion, deserves further discussion.
The purpose of this article is to stimulate this discussion, showing
in particular the non-necessity of Tinbergen's static controllability
condition for dynamic path controllability of a continuous time-
system. In Section 2 we review the definitions of controllability, in
order to show-in Section 3-that the dynamic generalization of
Tinbergen's theory can be made using both the concept of point
controllability and the concept of path controllability 2. It is in fact,
a widely held opinion that only path controllability is the "true"
generalization of Tinbergen's theory and that point controllability
has a very limited interest for economic applications while, in our
view, the importance of the one concept does not exclude the
importance of the other. It is the dynamic context in which the
control problem is analysed that permits us to make use of different
concepts of controllability, which would not be possible in a static
framework: in the same way as, in dynamics, the stability property
of a system can be defined in many ways, all equally important in
economic applications.
In Section 5 we analyse the concept of (and derive the conditions
for) path controllability following the approach introduced by Sain
and Massey [19] which is based on the invertibility of dynamic

C. Carraro and D. Sartore (eds.) Developments of Control Theory for Economic Analysis
© 1987 Martinus Nijhoff Publishers (Kluwer), Dordrecht-
32

systems and which is more immediate in our opinion, that the

Brockett and Mesarovic's [4] better known approach.
We discuss then-in Sections 4 and 6-the relations between
Tinbergen's static controllability condition and dynamic control-
lability conditions (point controllability, in Section 4, and path
controllability, in Section 6). In order to be able to make a com-
parison between Tinbergen's condition and dynamic controllability
conditions, we have to consider Tinbergen's static system as the
static equilibrium solution of the dynamic system in question. It is
in fact possible, in this way, to define a clear relation between the
matrices which appear in the dynamic model and those which
appear in the static one. We shall then show that Tinbergen's static
controllability condition is not a necessary condition for dynamic
controllability (neither for path controllability nor for point con-
trollability). It is thus possible to control a given number of policy
objectives dynamically-both point objectives and path objectives
~ven when the number of linearly independent instruments is
inferior to the number of linearly independent targets. The fact that
Tinbergen's condition (i.e., that there should exist at least as many
independent instruments as there are independent targets) is not a
necessary condition for point controllability has been demonstrated
in [15] as we shall see in Section 4. As far as path controllability is
concerned, there is no general proof3 of the non-necessity of
Tinbergen's condition for path controllability of continuous time
dynamic systems; this proof we provide in Section 6.
In what follows we shall refer to the so called fixed objectives,
that is to the objectives defined by a given numerical value (in the
case of point objectives) or by a given function of time (in the case
of path objectives). We also wish to underline that, as generally
accepted, we mean that the instruments (or the targets) are linearly
independent when the corresponding matrix is of full rank. (See [16,
pp. 6-14].) The economic meaning of this property has been clearly
illustrated by Mundell [10, pp. 201-202]: "Consistency also requires
that targets and instruments be mutually independent. For example,
full employment and maximum output could not be considered
targets if there is a unique functional relationship between the level
of employment and of output, just as an adjustment of the exchange
rate and certain applications of tariffs and export subsidies could
produce equivalent effects on output and on the balance of
pa ymen ts" .
 33

2. Dynamic Controllability and the Generalization of Tinbergen's

Theory

Point controllability and path controllability are the most important

concepts of dynamic controllability. In order to introduce these
concepts let us consider the continuous, time-invariant linear
dynamic system
Y ep) + A I y(p-I) + ... + A p-I y(l) + A py
= B od p ) + BI d p - I) + ... + Bp-I dl) + B pu (1)
where yet) E R is the target vector (output of the system), u(t) E R r
m

is the control vector (input of the system), and Al E R mm , Bi E R mr

(i = 0, 1, ... , p) are matrices of constant elements. The super-
script in brackets indicates the order of the derivative (some of the
BI matrices might be the null matrix). We assume, for simplicity,
that all endogenous variables are target variables and that all
exogenous variables are control variables.
A given point y* (defined by a vector of target variables) is
controllable at a given time tl if there exists an unbounded control
vector u*(t), defined on the interval (to t l ), capable of moving the
target vector from an initial position Yo at time to to the target
position y* at time t l • The dynamic system (1) is said to be com-
pletely controllable if any point y is controllable. From now on we
will use the term point controllability to indicate the complete
controllability of the system.
A dynamic system is said to be path controllable if there exists an
unbounded control vector u*(t), defined on the interval (to t I),
capable of moving an initial vector of target variables yo(to) along
any pre-assigned trajectory y*(t), (to ~ t ~ tl)' The pre-assigned
trajectory y*(t) can be either a vector which moves in time between
to and t l , or a vector whose value remains constant on that interva1. 4
It can be easily seen that the concept of path controllability is very
restrictive since it requires that the system follows exactly (and not
as closely as possible) any given trajectory (so, also disequilibrium
trajectories) during a given interval of time.
The dynamic generalization of Tinhergen's theory presented in
[15] is based on the concept of point controllability; but more
recently some authors [1, 2, 12] have maintained that this
generalization should instead be based on the path controllability
concept. As we have just seen, the point controllability property
34

assures only that the system will "hit" a given target position in a
given moment of time, but it does not say anything about the
behaviour of the system after that moment. In economics it is not
sufficient to reach a desired position, it is also important to be able
to stay there. It would in fact be useless, for example, to bring
employment to a given desired level at a given moment ifit were not
possible to keep that level subsequently. Path controllability assures
instead that the system will remain in the position reached for a
given period of time (or that it will follow, in that period, a pre-
assigned trajectory). This is the reason why path controllability is
often considered as the only controllability concept that is relevant
for economic applications and as the dynamic generalization of
Tinbergen's theory.
It is nevertheless possible to show, as we shall do in the next
paragraph, that path controllability is only one of the ways of
generalizing Tinbergen's theory in a dynamic context, and that
point controllability can be also considered a dynamic extension of
the same theory. It is just the dynamic framework in which the
economic policy problem is analysed which allows to generalize
Tinbergen's theory in two different but equally important ways.

3. Point Controllability and Tinbergen's Theory

As is well known, Tinbergens model can be summarized by the

following equilibrium relation in matrix form
Ay = Bu (2)
where y and u are the vectors already defined in system (1) and
A E Rmm, B E R mr are matrices of constant elements.
Tinbergen's problem is to verify the existence of a control vector
u* such that relation (2) is satisfied for any given pre-assigned value
of the target vector y*. Such vector u* exists if and only if
rank [11] = m. (3)
Condition (3) is known as Tinbergen's condition, or as the static
controllability condition. Given the dimension of matrix B, this
means that a vector u* exists if and only if the number of linearly
independent instruments is greater than or equal to the number of
linearly independent targets. We assume for simplicity's sake that A
 35

is a non-singular m x m matrix so that the targets are linearly

independent (see Introduction). Given this assumption, the fact
that the number of linearly independent instruments should be
greater than or equal to the number of linearly independent targets
means that the matrix iJ should contain at least m linearly indepen-
dent columns, or, what is the same, that its rank should equal m.
It follows from condition (3) that a necessary condition for static
controllability of system (2) is
(4)
i.e., the number of instruments should be greater than or equal to
the number of targets. This condition is known as Tinbergen's
"counting rule".
Since, with very few exceptions, the target vectors y* considered
by Tinbergen are static equilibrium vectors, the dynamic general-
ization of Tinbergen's problem should therefore be seen as the
problem of reaching in time a given equilibrium target vector. This
means that, given a dynamic system in disequilibrium like the one
defined by (1 )-described by target variables and control variables
which are functions of time-the dynamic extension of Tinbergen's
theory should be stated as the problem of the existence of a control
vector u* capable of moving an initial target vector yo(to) in any
given equilibrium point y*(t,), in the time interval (to, t,).
Since the static equilibrium solution of system (1) is
(5)
we can assert that, if system (1) is point controllable, a path for the
control vector u*(t) (to ~ t ~ t,) exists which will steer an initial
disequilibrium vector yo(to) to a given equilibrium position y*(t,)
which satisfies equation (5), that is such that
(6)
It seems thus clear that, in this framework, the problem of what
happens to the system after the target has been reached is no
longer important. In fact, if the equilibrium target pointS y*(t,)
is stable, the spontaneous forces of the economy will successively
keep the system at the equilibrium position reached. Once this
equilibrium position has been attained at time t, , the control vector
will assume the value which is consistent with the equilibrium
relation (5).
36

At this stage it could be objected that, if the dynamic system is

stable, it will tend to go back spontaneously to an equilibrium
position without the need of economic policy actions. Although this
is undoubtly true, we notice however that the stability property
assures only a tendency towards equilibrium (for t -+ (0), while the
point controllability property assures the existence of at least one
control vector capable of guiding the system exactly at the equi-
librium position in a given and finite interval of time.
Point controllability assumes therefore a great relevance as far as
Tinbergen's dynamic generalization is concerned since it assures
that targets which are-as in Tinbergen's theory~quilibrium
points can be reached in a given time and kept automatically till the
appearance of a new perturbation.
When the target position which the policy maker wishes to reach
and keep is not an equilibrium position, we have instead to turn our
attention-if we want to analyse the Tinbergen problem in a dynamic
framework-to the path controllability property of dynamic
systems.
It thus follows that we cannot speak of a unique dynamic
generalization of Tinbergen's theory. The choice of one type of
objective or another-and of the correspondong dynamic con-
trollability analysis (point or path controllability)-depends in fact
each time on the goals pursued by macroeconomic policy makers.

4. Non-Necessity of Tinbergen's Condition for Dynamic Point

Controllability

Let us consider again the dynamic system

yip) + A1y(p-l) + ... + Ap_1y(l) + Apy

(1)
As we have just seen, the static equilibrium solution of this
system is given by equation (5), that is by
(5)
and, therefore, Tinbergen's static system (2) can be considered as
the equilibrium position of the dynamic system (1), such that
A = Ap and B = Bp. It is possible in this way to specify a direct
 37

relation between Tinbergen's static model and the dynamic model

described by (1), in order to analyse the relation between static
controllability and dynamic controllability conditions, based on the
matrices of the state space representation of system (1).
Assuming for simplicity that the term Bo u(p) = 0 in system (1),
such representation is given by6

x(l) = Ax + Bu
(7)
y = Cx

where x(t) E R nn is the state vector, and A E Rnn, B E Rnr, C E R mn

are matrices of constant elements; yet) and u(t) are the vectors
defined in (1).
A necessary and sufficient condition for point controllability of
system (1) is

rank [P] = m (8)

where P is the mxnr matrix

P = [CB: CAB: ... : CAn-1B]. (9)

It has been demonstrated in [15] that Tinbergen's condition (3) is

not a necessary condition for point controllability. It is in fact
possible to have rank [P] = m even ifrank [Bp] < m. In a dynamic
framework it is therefore possible to reach a target point with a
number of linearly independent instruments inferior to the number
of linearly independent targets.
In addition, since (8) can be satisfied even if r < m, it is poss-
ible to have point controllability even if the number of instru-
ments is (numerically) inferior to the number of targets. That
is, neither the "counting rule" (4) is necessary for dynamic point
controllability.
In what follows we shall show that also when path controllability
is considered, Tinbergen's conditions lose their general validity (in
fact, not even in this case condition (3) is necessary for path con-
trollability). To that purpose we shall first reformulate the problem
of path controllability introduced in [4], following the approach
presented in [19] which is based on the relation between path
controllability and invertibility of dynamic systems.
38

5. Path Controllability

Consider the dynamic system in state space form (7). Taking the
Laplace transforms of both members we have
sX(s) - x(O) = AX(s) + BU(s) (10)

Yes) = CX(s) (11 )

where Xes), U(s) and Yes) indicate the Laplace transforms of x(t),
u(t), and yet) respectively. Assuming thae x(O) = 0 we have, from
(10) and (11),
Xes) = (sl - A)-l BU(s) (12)
where I is the identity matrix in Rnn. Substituting eq. (12) in eq. (11)
we have
Yes) = G(s)U(s) (13)
where G(s) is a rational matrix in s and is the the transfer function
matrix of system (1) defined by
G(s) = C(sl - A)-l B. (14)
Equation (13) describes the relation between the Laplace trans-
form of the control vector U(s) and the Laplace transform of the
target vector Yes). The dimension of matrix G(s) is m x r where,
we recall, m is the number of target variables and r is the number
of control variables.
We now report some definitions of invertibility and of path
controllability of dynamic systems in the frequency domain. (See
[19], [20] and [21].)
The dynamic system described by eq. (1) is said to be left invertible
if the transfer matrix G(s) has a left inverse, while it is said to be
right invertible if the transfer matrix G(s) has a right inverse. If the
system is right invertible a vector U(s) always exists so that equation
(13) is satisfied for any given arbitrary vector Yes).
The dynamic system (1) is said to be path controllable (see [19])
if, given an arbitrary vector Yes), there always exists a vector U(s)
so that equation (13) is satisfied. It therefore follows that path
controllability and right invertibility are equivalent properties of
system (1). In fact, if the matrix G(s) has a right inverse it is always
possible to solve eq. (13) for U(s), given Yes).
 39

But to say that the system described by eq. (1) is right invertible
is the same as saying that the dual of system (1) is left invertible
since the transfer matrix of the dual system is defined as (See [9] and
[18])
(15)
where the prime denotes transposition, that is
G'(s) = B'(sI - A')-' C' (16)

5.1. Conditions for left invertibility

The following necessary and sufficient conditions for left invert-
ibility are stated in [19]:
(a) the system described by eq. (1) is left invertible if and only if
rank [G(s)] = r (17)
over the field of rational functions in s.
The same condition can be expressed only in terms of the matrices
A, Band C that characterize the state space representation of
system (1), that is
(b) the dynamic system (1) is left invertible if and only if (see [19]
for a proof)
rank [N] = (n + I)r (18)
where N is the 2nm x (n + I)r matrix

CB 0 .............. 0
CAB CB . . . . . . . . . . . . . O

N (19)

CB

CA 2n - 1 B
We notice that being G(s) an m x r matrix if follows, from
condition (17), that:
40

(c) the dynamic system (1) is left invertible only if

m ~ r. (20)

5.2. Conditions for path controllability

From conditions (17) and (18) it is now possible to state the
necessary and sufficient conditions for path controllability
(d) The system described by eq. (1) is path controllable (i.e., the
dual of system (1) is left invertible) if and only if
rank [GD(s)] = m (21)
over the field of rational functions in s. This condition, stated in
[19], is the same as condition (b) in [4, p. 559].
(e) The system described by eq. (1) is path controllable if and only
if
rank [ND ] = (n + l)m (22)
where ND is the (n + l)m x 2nr matrix
CB CAB ............. CA 2n - 1 B
o CB . ............. CA 2n - 2 B
(23)

o o. CB.
Let us now assume that Tinbergen's condition (3) is not satisfied
(i.e., rank [Bp] < m). It can be shown that the path controllability
condition (21) and, consequently, condition (22) can be satisfied
even if condition (3) is not. It is in fact possible to have rank
[C(s! - A)-I B) = m even if rank [Bp] < m; since rank [C) = m,
rank [(sf - A)-I] = n (n ~ m), and since we can have rank
[B) = m also when rank [Bp] < m, it is possible to have rank
[GD(s)] = m. 8
The same conclusion holds as far as condition (22) is concerned.
It is in fact possible to have rank [ND ] = (n + l)m even when rank
[Bp] < m; as we have just seen, we can have rank [B) = m even
when rank [Bp] < m. So we can also have rank [CB] = m even
when rank [Bp] < m. Given the form of the matrix ND this would
be sufficient to have rank [ND ] = (n + l)m.
Therefore,Tinbergen's condition (3) is not a necessary condition
for path controllability of continuous time dynamic systems.
 41

Note that when no derivatives of the control variables are con-

°
sidered (i.e., B, = 0(; = 0, 1, ... ,p - 1), Bp #- in eq. (1)), the
Tinbergen condition is a necessary condition for path control-
labilbility. (This result can be easily checked by the reader.) A
particular case is the dynamic system with state target variables.
We notice however that when p = 1 in eq. (1) (with A, = 0,
B, = 0, i = 2, ... ,p). that is when the dynamics of the target
vector is of the first order but there is at least one first order
derivative of one of the control variables, the Tinbergen condition
(3) is no longer necessary for path controllability.
We also notice that Tinbergen's "counting rule" (4) (i.e., number
of instruments ~ number of targets) is, as we have seen (condition
(24)), a necessary condition for path controllability. This condition,
however, says nothing about the linear dependence or indepen-
dence between the instruments; thus, while a static system might not
be controllable because of the linear dependence between the
instruments (though they are as many as the targets), in a dynamic
framework a path objective might instead be controllable in spite of
the linear dependence between the instruments. The dynamics of
the system, in fact, allows the control variables to move in time so
that it is not only their levels (instant values) which influence the
targets, but also their time variations. And it is this dynamic effect
of the instruments on the target variables that "cancels" the necessity
of linear independence between the instruments, which is instead
essential in a static framework.
An example might clarify this result. Let us assume that the path
objective y*(t) that we want to control dynamically is given by a
constant value y* (i.e., y*(t) = Y6, to ~ t ~ t l ) and that Y6 is a
vector which satisfies the static equilibrium relation (2)
(25)
Since we assume rank [.8] < m (so that rank [.8] < r, since
r ~ m) the linear dependence between the instruments does not
allow, in this static framework, to obtain the desired value Y6 (i.e.,
a vector u does not exist which solves eq. (25), given y6). In a
dynamic framework, however, if the path controllability condition
is satisfied, it is possible to reach a vector Y6 in the target space
and to keep it at this level during a given interval of time. This
is due to the possibility of varying, in that interval, the control
vector u(t).
42

7. Concluding Remarks

We have argued in this paper that Tinbergen's static approach to

the theory of economic policy can be generalized in a dynamic
framework making use either of the concept of point controllability
or of the concept of path controllability. The choice of the one
approach or the other depends on the type of objective (equilibrium
or disequilibrium objective) pursued by the policy marker.
Following the approach presented in [19] based on the invertibility
of continuous time dynamic systems, we have then shown that the
conditions for path controllability which are derived are the same
as the conditions stated in [4]. Making a comparison then between
Tinbergen's static controllability conditions and dynamic path con-
trollability conditions we have proved that, although Tinbergen's
"counting rule" «4) i.e., number of instruments ~ number of
targets) is a necessary condition for path controllability of continu-
ous dynamic systems, Tinbergen's rank condition (3) (i.e., number of
linearly independent instruments ~ number of linearly independent
targets) is not a necessary condition for path controllability of
general dynamic linear systems.
We can therefore conclude that when general dynamic economic
models are considered, the condition of linear independence
between the instruments is no longer necessary, provided that there
are as many instruments as targets.
This result is not surprising if we consider the dynamic aspects of
economic models: in fact, the linear dependence which might exist
between the static values of a given number of instruments, might
no longer exist between the static value of one instrument and the
derivatives (of the first or higher order) or the others.
The economic implications of this result are immediate if one
thinks how frequently dynamic economic models include deriva-
tives of instruments of economic policy, like, for instance, the rate
of growth of the money stock.

Notes

I. This paper has been prepared in the context of a research financed by the
Consiglio Nazionale delle Ricerche. I wish to thank G. Gandolfo and A.
Isidori for helpful comments.
 43

2. Point controllability is also known as output controllability or pointwise

reproducibility, while path controllability is also known as perfect control-
lability or as functional reproducibility.
3. Buiter [5] presents an economic model which is not statically controllable (or
control1able across steady states) while it is path control1able.
4. We follow here Aoki's definition of path control1ability, refering therefore to
the case in which either
(a) the target function y*(t) can be any function in the target space and the
admissible control space includes impulse control functions, or
(b) the admissible target space is restricted only to sufficiently smooth target
functions and the admissible control space does not contain impulse
functions.
If this is not the case the path controllability conditions might be different
from the conditions stated in this paper. See [26].
5. This objective is similar to the "stationary objective" described in [16, Ch. 7].
6. As it is wel1 known, when Bou lP ) =1= 0, the state space becomes
x(l) Ax + Bu

y Cx + Du
where D E RW. This assumption does not alter the results obtained in the text,
but it eventual1y strengthens them. (See note (8).)
7. This assumption could be dropped if we consider the target vector yes) as the
difference between the same vector Yes) and the Laplance transform of the free
response of the dynamic system (I) (i.e. the trajectory which is the solution of
the corresponding homogeneous system). This is the approach followed in [4].
°
8. Note that in the case in which the term BouP(t) =1= in system (I), the transfer
function matrix of that system is
G(s) = C(sl - A) - 1 B + D

Therefore the result just stated in the text is strengthened by the presence of
the matrix D. In that case, in fact, it IS possible to have rank [G(s)] = m even
If rank [C(sl - A) -I B] < m.

References

I. Aoki, M.: "On a Generalization of Tinbergen's Condition in the Theory

of Policy to Dynamic Models", Reviel~' of" Economic Studies 42 (1975).
293-296.
2. Aoki, M.: Optimal Control and System Theory in Dynamic Economic Analysis.
Amsterdam & New York: North-Holland. \976.
3. Aoki, M. and M. Canzoneri: "Sufficient Conditions for Control of Target
Variables and Assignment of Instruments in Dynamic Macroeconomic
Models". International Economic Review 3 (1979). 605-616.
44

4. Brockett, R. W., and M. D. Mesarovic: "The Reproducibility of Multi-

variable Systems", Journal of Mathematical Analysis and Applications 11
(1965), 548-563.
5. Buiter, W. H.: "Unemployment-Inflation Trade-offs with Rational Expec-
tation in an Open Economy", Journal of Economic Dynamics and Cantrall
(1979), 117-141.
6. Buiter, W. H. and M. Gersovitz: "Issues in Controllability and the Theory of
Economic Policy", Journal of Public Economics IS, (1981), 33-43.
7. Buiter, W. H. and M. Gersovitz: "Controllability and the Theory of Economic
Policy. A Further Note", Journal of Public Economics 24 (1984), 127-129.
8. Hughes Hallett, A. H. and H. Rees: Quantitative Economic Policies and
Interactive Planning. Cambridge: Cambridge University Press, 1983.
9. Kwakernaak, H. and R. Sivan: Linear Optimal Control Systems. New York:
Wiley-Interscience, 1972.
10. Mundell, R. A.: "The Appropriate Use of Monetary and Fiscal Policy for
Internal and External Stability", IMF Staff Papers 9 (1962), 70-79.
II. Mundell, R. A.: International Economics. New York: Macmillan, 1968.
12. Nyberg, L. and S. Viotti: "Controllability and the Theory of Economic
Policy. A Critical View", Journal of Public Economics, 9 (1978),73-81.
13. Ogata, K.: Modern Control Engineering. Englewood Cliffs, N.J.: Prentice-
Hall, 1970.
14. Petit, M. L.: "Sulla teoria dinamica della politica economica. Due approcci
alternativi", Note Economiche, 3 (1984), 24-42.
15. Preston, A. J.: "A Dynamic Generalization of Tinbergen's Theory of Policy",
Review of Economic Studies, 41 (1974), 65-74.
16. Preston, A. J. and A. R. Pagan: The Theory of Economic Policy. Cambridge:
Cambridge University Press, 1982.
17. Preston, A. J. and E. Sieper: "Policy Objectives and Instrument Require-
ments for a Dynamic Theory of Policy" in Applications of Control Theory to
Economic Analysis, cd. by J. D. Pitchford and S. J. Turnovsky. Amsterdam:
North-Holland, 1977.
18. Ruberti, A. and A. Isidori: Teoria dei Sistemi. Torino: Boringhieri. 1982.
19. Sain, M. K. and J. L. Massey: "Invertibility of Linear Time-Invarient Dynami-
cal Systems", IEEE Transactions on Automatic Control 2 (1969), 141-149.
20. Silverman, L. M.: "Inversion of Multivariable Linear Systems", IEEE
Transactions on Automatic Control 3 (1969), 270-276.
21. Silverman, L. M. and H. J. Payne: Input-Output Structure of Linear Systems
with Application to the Decoupling Problem", SIAM Journal of Control 9
(197\), 199-233.
22. Tinbergen, J.: On the Theory of Economic Policy, 2nd ed. Amsterdam: North
Holland, 1955.
23. Tondilll, G.: "Controllabilita e Teoria della politica economica", Giornale
deglt Economisti e Annali di Economia 5/6, (1983), 307-321.
24. Tondini, G.: "Further Discussion on Controllability and the Theory of
Economic Policy", Journal of Public Economics 24 (1984), 123-125.
25. Wohltmann, H. W.: "Complete, Perfect, and Maximal Controllability of
Discrete Economic Systems", Zeitschriftfur Nationalokonomie 41 (1981), 39-58.
 45

26. Wohltmann, H. W.: "A note on Aoki's Conditions for Path Controllability of
Continuous-Time Dynamic Economic Systems". Review of Economic Studies
51 (1984), 343-349.
27. Wohltmann, H. W.: "Target Path Controllability of Linear Time-Varying
Dynamical Systems" IEEE Transactions on Automatic Control 30 (1985),
84-87.
28. Wohltmann, H. W., and W. Kromer: "A Note of Ruiter's Sufficient Con-
dition for Perfect Output Controllability of a Rational Expectations Model":
Journal of Economic Dynamics and Control 6 (1983),201-205.
29. Wohltmann, H. W., and W. Kromer: "Sufficient Conditions for Dynamic
Path Controllability of Economic Systems", Journal of Economic Dynamics
and Control 7 (1984), 315-330.
 47

CHAPTER 4

SOFTWARE FOR ECONOMIC OPTIMAL CONTROL

MODELS

David Kendrick
University of Texas at Austin, USA

The revolution in microelectronics is offering new opportunities for

the development of software for economic optimal control models
that is easier to use. This paper contrasts the software of the past
with the software of the present and outlines the likely form of the
software of the future. In so doing it chronicles the evolution of
input and output from numbers to equations to graphics.
Figure 1 shows the change of input and output methods for
optimal control problems in the evolution of computing from batch
processing to time sharing to microcomputers. The response time
for batch processing was hours as one waited for cards to be
processed and printout produced which provided cryptic messages
describing the remaining bugs. The input was in the form of fixed
format numbers on punched cards and the output consisted of
pages of numbers.
The advent of time sharing systems reduced the response time to
minutes while one ran a job from a remote terminal and waited for
the output to be processed and communicated through 100 or 300
baud modems. The input was now in free format and consisted of
equations as well as of numbers. The output now included nicely
laid out tables and crude graphs.
Microcomputers have further reduced the response time to
seconds as the results of calculations are flashed to the screen
instantaneously. On the larger microcomputers, which are also
called workstations, bit mapped graphics screens and high speed 32
bit microprocessors are resulting in an evolution of input forms
toward mathematics and graphics. Also, output is now in the form
of graphs and figures which can be paged through like a magazine.

C. Carraro and D. Sartore (eds). Developments of Control Theory for Economic Analysis
© 1987 Martinus NiJhoff Publishers (Kluwer), Dordrecht-
 48

Response
Method Input Output
Time

EJ
fixed format
hours numbers
and numbers

Time free format nice tables

minutes
Sharing and equations and crude graphs

tv\icro- graphics figures

seconds
computer and mathematics and graphs

(point and cl ick)

Figure 1. Methods of input and output

In this paper the evolution of software will be reviewed first for

general non-linear models and then for quadratic linear models.
The emphasis will be on deterministic models since the evolution of
stochastic software closely parallels that for deterministic models.

1. General Non-linear Problems

One way to trace the development of this software is to contrast

the way a single problem was solved fifteen years ago, the way it is
solved today and the way it will probably be solved in the future.
The problem used here is the optimal growth model of Chakravarty
 49

1 1
.75 .9 .03 input

READ (5. 8) N. t1
FORt1AT(2I4)
8 program
READ(5. 10) BETA. ttJ. PI
10 FCRtIAT(3f5.2)
derivative of
F(l) = BETA * A(T)*K(T) **(BETA-l) function

Time BETA EXPONENT OF CAPITAL 1.751 ;

ttJ ELAS OF MARGINAL UTILITY 1.9/;
Sharing
PI DISCOUNT RATE 1.03/;

CRITERION.. J =E= SlJ1(T. (1+PI) .... (-ORD(T» .. (1/1-ttJ)"C(T-1) .... (1-ttJ);

PRODUCT(T) .. Y(T) =E= A(T) ~ K(T)**BETA;

Micro- B = .15 ;
1.1 = .90 ;
Computers I = .03 ;

III!Iximize
T -t (1-1.1)
(1) J = L [(1+.) --- c
t-1
t=l (1-1.1)

subject to
B
(2) a k
t t

Figure 2. Inputs for general nonlinear problem.

(1962). The problem was solved fifteen years ago on an IBM 7094
with a Fortran code base on the conjugate gradient algorithm of
Lasdon, Mitter and Warren (1967). Part of an input file and some
code of this type is shown at the top of Figure 2. The data were in
fixed format and were read in with Fortran statements like those
shown in the figure. The equations of the model were represented
by functions for the derivatives like the one shown for function F(1)
which is the derivative of a Cobb-Douglas production function.
The derivatives were easy enough to do for this small model
50

as described in Kendrick and Taylor (1971), but were exceeding

tedious to do in large models using algorithms which required
second derivatives. In these cases it was sometimes necessary to take
several hundred derivatives by hand and then to code them into
Fortran. Such methods were not only time consuming they also
provided ample opportunities for error.
The current way of solving the Chakravarty problem is on time
sharing systems like the CDC Dual Cyber 170/750 using the GAMS
software of Alexander Meeraus (1983). An example showing a part
of this input is provided in the middle of Figure 2 opposite the "time
sharing" block. Input is now in free format form with parameters
not only having names in the input file like "BETA" but also
with provision for parameter descriptions like "EXPONENT OF
CAPITAL". However, the big change is not in the data input but
rather in the model description, For example the Cobb-Douglas
production function is written in the easier to understand form
YeT) = E = A(T)* K(T)**BETA;
Moreover, it is no longer necessary to provide the derivatives since
the GAMS system has internal software to take the analytical
derivatives of the equations. So the input of the model to the
computer is not only less time consuming, there is also much less
chance of error.
The input for the software of the future is shown in the bottom
part of Figure 2. This input is labeled "Microcomputers" but it is
not so much for Apple II's and IBM PC's as for SUN Workstations
and the successors of the Apple Lisa and the IBM PC AT. These
machines have and will have bit mapped graphics screens measur-
ing one thousand by one thousand pixels so that a megabit of
memory will be required to hold a single screen image. In contrast
the Apple Macintosh and the IBM PC screens measure more like
300 by 500 pixels. The finer resolution on the "workstations" allows
one to display mathematics on the screen in subtle detail. Moreover
the high speed of the 32 bit microprocessor in these machines
permits graphic images to be used for both input and output.
One bridge from the world of the present to the world of the
future is shown in Figure 3 which outlines a system which operates
on a VAX 11/780 under the UNIX operating system. Figure 3
shows how this kind of software could be applied to the Chak-
ravarty problem.
 51

maximize
T -t
(1) J = L [(1+0) ---
MATH t=1 (1-~)

subject to
B
(2) Yt a k
t t

t NEON
IMximize
.DE
.EQ
(1)
~~~~~~~~~~mark J ~ ;

~~~ SIGMA from {t = 1} to T [ ( 1 + pi ) sup -t

--- 1 over { ( 1 - mu ) } c sub t-1 sup { ( 1 - mu ) } ]
.EN

subject to

.EQ
(2)
-----~~~Y sub t -;- a sub t k sub t sup beta;
.EN

+ MEL

CRlTERlOO.. J =E= SlJ1(T, (1+PI)·..·(-ORD(T» .. (1/1-ttJ)"C(T-1) .... (1-ItJ);

GAMS
PROOUCT(T) .. V(T) =E- A(T) • K(T)*~TA;

Figure 3. A Unix based system.

The top part of the diagram which is labeled "MATH" is the

future and the bottom part of the diagram labeled 'GAMS' is the
present. The box in the middle of Figure 3 contains the same
equations as in the top and bottom parts of the figure but written
in the neqn language which was developed at Bell Telephone Lab-
oratories by Kernighan and Cherry (1977). Neqn is a formatting
language for mathematics. Another language of this same type is
TEX which was developed by Knuth (1979). Neqn is now used
regularly by many economists to produce mathematical economic
52

papers. For example at the University of Texas it is used along with

a Diablo ECS (Extended Character Set) printer to produce papers
that use both the Greek and the English alphabets. Thus the arrow
from the box in the middle of Figure 3 pointing upwards is to
mathematics via neqn.
The arrow pointing downward is via MEL (Mathematical
Economics Language) which translates the input into the GAMS
language. MEL is not yet really a language yet but rather only a
proof of principal that his sort of transformations can be done on
linear programming problems as outlined in Kendrick (1984). The
strength of this approach is that a mathematical model can be
translated by a computer into the GAMS language so that it can be
solved and can also be translated into mathematics so that it is easy
to read. The weakness of this approach is that the neqn language
is difficult to use as an input medium since it is hard to read. The
essential problem is that neqn is a one dimensional language that is
used to produce input which is essentially two dimensional. For
example
SIGMA from {t = I} to T
to produce
T

L
(=1

Rather we need a language which can read two dimensional math-

ematics and convert the mathematics to GAMS or some other
language which can then be used to solve the optimal control
problem. One possible solution to this problem is discussed at the
end of the next section. However, consider first the software for
solving quadratic linear control theory problems. Just as for the
general nonlinear models first the past, then the present and finally
the future software will be discussed.

2. Quadratic-Linear Models

Before discussing the software it is useful to review the math-

ematics of the quadratic linear control problem in order to establish
the notation. This is shown in Figure 4 which shows the criterion
 53

Maximize

J = (x - x ) 'N (x - x )
N N N N

N-l
+ L [ (x - x ) Wk (x - x ) + (u - ~ ) • 1\ (u - ~ )
k=O k k k k k k k k k

subject to
x
k+l

..here

x = state vector (n x 1)
u = control vector (m x 1)
Figure 4. Mathematics for quadratic linear problem (QLP).

and system equations for the standard quadratic-linear tracking

problem.
The x's are state variables and the u's are control variables. Also,
the - over a variable indicates that it is a desired path. The W's and
lambdas are penalty matrices on the state and control variables
respectively.
The input for the batch software of the past is shown in Figure 5.
Only the part in the bottom left hand quadrant is actually input to
the computer. The mathematics in the bottom right hand quadrant
is added in order to make the input easier to understand. The input
is almost entirely numbers and is in fixed format.
The input of the present is not from a time sharing system but
from a microcomputer, namely the IBM Pc. It is from a code
which was recently translated by the author from Fortran into
Pascal and installed on an IBM PC XT with an 8087 numerical
co-processor. The input is no longer in fixed format but rather in
the form of an interactive dialog with the computer as is shown in
Figure 6 where part of the data is input.
Then Figure 7 shows part of the dialog with this same program
after the data is input and the program is being run again. The code
ask for the file name on which the data is stored and then displays
the data so that the investigator can check it. Software with this
 54

C~uter Input Hath

DETERMINISTIC QlP PROOLEH IN CH 4 (J= KENlRICK (1981)

2 1
n m nt
20

print controls
1 1
1. 014 .002
.093 .153 A
-.004
-. 100 B
-1.312
.448 c
1. O.
O. 1.
100. O.
1I
O. 100. N
1.
460. 1
113.1
463.551 461.021 410.530414.059 411.615 481.191484.806
113.948 114.803 115.664 116.531 111.405 118.286 119.113 )(

153.644 154.196 155.951 151.121 158.305 159.492 160.689 u

460. 1
113.1

Figure 5. QLP input in batch mode.

type of interface has been available for many years on the TROLL
system of the Centre for Computational Research in Economics
and Management Science (1978) and is coming in MODULECO by
Neopmiastchy and Rechenmann (1983). Moreover, it is a standard
kind of interface for microcomputer software.
One possible input form for the software of the future is shown
in Figure 8. The problem here is a dynamic control problem for the
college education financing plan of a family. There are two state
variables, namely the amount of money loaned to the family and
the amount of money in their bank account. Also, there are four
control variables, namely borrowing, savings, payments on the loan
and withdrawals for college expenditures.
 55

Welcome to QLP.

Do you want to use an existing input file (YIN)?

n

On what file do you wish to put the data?

cM.dat

Input the number of state variables.

2

Input the number of control var iables.

1

Input the number of time periods.

7
Input the A matrix.
element 1. 1
1.014

element 1.2
.002

element 2. 1
.093
element 2. 2
.753
Input the B matr ix.
element 1. 1
-.004

Figure 6. QLP input in time sharing mode.

In Figure 8 regular rectangles are used to represent state variable

and rectangles with rounded corners are used to represent control
variables. There is one equation for each state variable. The small
circles and arrows are used to show which state and control vari-
ables are included in each equation. An arrow pointing into a circle
indicates that the variable enters the equation with a plus sign and
an arrow pointing out of circle indicates that the variable enters the
equation with a minus sign.
Now how can this graphical input be used to generate input to
an optimization system like GAMS? One promising avenue is to use
software like Lisa Draw. This is similar to the Mac Draw software
for the Apple Macintosh. A screen dump of the use of Lisa Draw
is shown in Figure 9.
56

Welcome to QlP.

Do you want to use an existing input file (YIN)?

y

What IS the filename?

cM.dat

a
1.014 0.002
0.093 0.753
b
-0.004
- O. 100

Do you want to change any of the print or control OptiOns?

y

Which option do you want to change?

12

Input the desired value for this option.

o
Do you wish to alter another option?
n

Figure 7. Input for time sharing mode with an existing file.

In order to draw a rectangle with this software one maneuvers the

mouse which in turn causes the cursor (the arrow head in the left
centre of Figure 9) to move. The cursor would be moved over the
rectangle in the column of shapes on the left side of the figure. The
mouse would then be clicked once to select the rectangle. The cursor
would then be moved to the location on the screen where the top
left corner of the rectangle is to be placed. Then the mouse button
would be clicked and held down while the cursor is moved to the
desired location for the bottom right corner of the rectangle and
then released. Similar procedures are used to draw circles and
rectangles with rounded corners as well as arrows. Then the text
selection at the top of the left hand column is selected in order to
write the names of the state and control variables in the appropriate
rectangles.
 57

This Period Next Period

loan (l+r) .()~_........I:I

~.,,-----..,./ LJ

bank bank
account account

withdrawals
for
college

Figure 8. College education financing problem.

The software for this system records the location of each rec-
tangle and circle. This information potentially can be used to create
state equations which can then be written into the GAMS language.
It is the aspect of Lisa Draw and Mac Draw of keeping up with the
location of each object which is important. In contrast the popular
MacPaint software for the Macintosh does not keep up with each
object in the drawing and thus could not be used for this kind of
software.
Software which uses graphic input for optimal control problems
is thus a realistic possibility for the not too distant future. Similarly,
the output will not be the tables of numbers which we now use, but
rather graphs, charts and figures.
 v.
00

This Period· t..Jed Period

I"E
0, ·1 loan (1+r) . / ' ..... loa"
c:J .
,~ ~ ~
r borrow!n, )'' ;
~
?t01 (
\ \\
~~.,;~~~ ') "\
'"
.~,
......... '-, ~

bank· ~~
acco::l · " . . . .r:,ank
account

Figure 9. Lisa Draw screen dump,

 59

3. Summary and Conclusions

The tremendous increase in the speed and processing power of

computer chips is bringing rapid changes to the software for solving
economic control theory problems. Interactive times have dropped
from hours to minutes to seconds and the interface with computer
software is changing accordingly. This has resulted in a change
from the input of numbers to equations and will soon result in input
in the form of mathematics and figures.

References

Centre for Computational Research in Economics and Management Science

(1978): "TROLL: An Introduction and Demonstration", M.LT., Cambridge,
MA.02139.
Chakravarty, S. (1962): "Optimum Savings with a Finite Planning Horizon",
International Economic Review, Volume 3, September, pp. 338-355.
Kendrick, David A. (1984): "A Mathematical-Computer Language for Linear
Programming Problems", in T. Basar and L. F. Pau, Dynamic Modelling and
Control of National Economies 1983, Pergamon Press, Oxford, England,
pp. 233-240.
Kendrick, David A. and Lance Taylor (1971): "Numerical Methods and Non-
linear Optimizing Models for Economic Planning", Ch. I in Hollis B. Chenery
(ed.), Studies in Development Planning, Harvard University Press, Cambridge,
MA.
Kernighan, Brian W. and Lorinda L. Cherry (1977), "Typesetting Mathematics-
Users Guide (Second Edition)", Chapter T.7 in T. A. Dolotta, R. C. Haight,
and E. M. Piskorik (eds.) (1977), Documentsfor the PWB/UNIX Time-Sharing
System, Edition 1.0, Bell Telephone Laboratories, Murray Hill, New Jersey
07974.
Knuth, Donald E. (1979), TEX and META FONT: New Directions in Typesetting,
Digital Press, Bedford, MA.
Lasdon, E. S., S. K. Mitter, and A. D. Warren (1967), "The Conjugate Gradient
Method for Optimal Control Problems", IEEE Transactions on Automatic
Control, Volume 12, April, pp. 132-138.
Meeraus, Alexander (1983), "An Algebraic Approach to Modeling", Journal of
Economic Dynamics and Control, Volume 5, No. I, February, pp. 81-108.
Nepomiastchy, Pierre and Francois Rechenmann (1983), "The Equation Writing
External Language of the MODULECO Software", Journal of Economic
Dynamics and Control, Volume 5, No. I, February, pp. 37-58.
 PART II

RECENT DEVELOPMENTS OF CONTROL THEORY:

OBJECTIVE FUNCTION SPECIFICATION
 63

CHAPTER 5

INTERACTIVE VECTOR OPTIMIZATION AS A

COMPLEMENT TO OPTIMAL CONTROL IN
ECONOMETRIC MODELS*

Josef Gruber
University of Hagen, FRG

1. Introduction: Towards Observed Preferences in Econometric

Decision Models

The usefulness of econometric forecasting can be increased by

taking these models as part of econometric decision models ( = opti-
mization models). The two most important approaches to econo-
metric decision models are control theory and interactive vector
optimization.
The control theory approach to econometric decision models has
the following major advantages: Given a numerically specified
model (i.e., mainly a scalar-valued objective function in which the
usually heterogeneous individual target and instrument variables
are "aggregated" into a single measure of welfare, etc., and con-
straints which consist of or are derived from an econometric
equation system), the corresponding optimal solution can be cal-
culated "uniquely". Therefore, the control theory approach is com-
paratively well suited for simulation studies in which the effect on
the optimal solution of changes in the parameters of the model or
of exogenous factors is evaluated under the ceteris paribus con-
dition. There is also a plenty of theories and solution techniques
from optimal control that may be applicable to modelling economic
policy problems and to solving policy decision models.

*Gratefully acknowledged are many helpful discussions with Elke Petersen and
Michael Olbrisch who are now working on the research projects briefly described
in Sections 3 and 4 of this paper.

C. Carraro and D. Sartore (eds.) Developments of Control Theory for Economic Analysis
© 1987 Martinus Nijhoff Publishers (Kluwer), Dordrecht-Printed in the Netherlands
64

Some major disadvantages of the control theory approach to

econometric decision models are the following: The amount of
information about the preferences of the decision maker which is
required for specifying numerically the scalar-valued objective
function of the econometric decision model is so large that it can
hardly be obtained in practice. There is a lack of well-developed
methods for constructing such objective functions. (For an earlier
review, see Gruber, 1979.) Therefore, "theoretically assumed"
preferences are used frequently instead of observed preferences.
The standard version of the control theoretic decision model (see,
e.g., Chow, 1975, 1981, and Kendrick, 1981) is not well adapted to
some important aspects of the economic decision problem. For
example, in the model consisting of a quadratic objective function
and linear equality constraints (the "standard" case), positive
deviations from desired values of target and instrument variables
are treated in the same way as negative deviations. The introduction
of inequality constraints (which are, as a rule, a matter of course for
an economist) makes some theories (e.g., certainty equivalence)
inapplicable; it also leads to increased complexities of the com-
putations (e.g., solution of the model by quadratic programming;
see e.g., Friedman, 1975, and Murata, 1982).
The interactive vector optimization approach to econometric
decision models works always in the mathematical programming
framework. Therefore, inequality constraints ("policy constraints"
in the terminology of quantitive economic policy (see Fox et ai.,
1966)) and nonnegativity conditions for the solution can be taken
into account. Most importantly, this approach avoids problems
connected with explicitly specifying a scalar-valued objective func-
tion. There is no explicitly specified scalar-valued ("overall") objec-
tive function in an econometric vector optimization model. Instead,
the preference structure of the decision maker (or of someone close
to him, e.g., the econometric analyst) is revealed in part (locally)
by an interactive question-answer procedure. The reduced form
equations of the econometric model are used not only as constraints
but also as "individual" objective functions. For example, the
reduced form equation explaining the number of unemployed per-
sons may be such an "individual" objective function. Therefore, in
the econometric vector optimization model there is no need for
incorporating observed or hypothetical information about the type
of the scalar-valued objective function and the numerical
 65

values of the parameters of this function. This is of advantage in

applications (see Leontief, 1981).
In all interactive vector optimization procedures the "optimal-
ity" of the solution is directly determined by the decision maker.
Therefore, the optimum is "subjective". The decision maker may
change his preferences from one solution session to the next (e.g.,
because of learning about the feasible solutions available to him).
Therefore, the econometric vector optimization model is not as well
suited for simulation studies as the decision model with an explicitly
specified scalar-valued objective function. Since simulation studies
often serve the important purpose of exploring decision alternatives
and to "check" the econometric equation system, the vector opti-
mization approach can be no substitute for the control theoretical
approach; it is rather a complement of it. Both approaches have to
be applied if better use is to be made of the potential in econometric
models for improved decision making.
In several interactive vector optimization procedures, a feasible
combination of instrument and target variables ("local point",
"alternative", "scenario", "menu") is presented to the decision
maker for evaluation. He then makes his choice. In this respect
vector optimization is similar to decision making on the basis of
econometric forecasting systems (without any objective function).
This property of interactive vector optimization algorithms enhances
the likelihood that they are used almost as frequently as econo-
metric forecasting models.

2. Economically Important Aspects of Rosinger's Algorithm for

Interactive Vector Optimization

2.1. Rosinger's algorithm as a promising alternative

There are several algorithms for interactively solving the vector
optimization model. These algorithms differ considerably in their
properties (e.g., in the amount and quality of information to be
supplied by the decision maker). Practically all algorithms used in
connection with econometric models and available in the literature
are dealt with or are at least referred to in Gruber (1983, Part 3).
A more general survey may be obtained in Hansen (1983).
66

The interactive method for vector optimization by Rosinger (see

Rosinger, 1981, 1982) is relatively well adapted to the needs of
decision makers working with econometric/economic models. The
decision maker has great flexibility in forming groups of objectives
(target and instrument variables) and in giving piecewise infor-
mation on his local preferences. All comparisons of marginal utili-
ties may be made in groups of only two objectives. Also, the
preference statements of the decision maker may be contradictory
to a certain extent. Therefore, the algorithm of Rosinger is an
important improvement over the algorithm by Geoffrion, Dyer and
Feinberg (1972).
In a research project at the University of Hagen, Rosinger's
algorithm has been programmed on an IBM 3031 computer and
has been combined with the 15 reduced form equations of the
macroeconometric model of the West German economy by Beck-
mann and Uebe (see Section 3 of this paper). The experiences
gained in this research project are rather favorable; therefore, the
project is continued, also in cooperation with actual decision
makers (or persons close to them, e.g., persons working in the
econometric divisions of Federal Ministries and in economic
research institutes). Reports on it are available in Streuff, 1983, and
Streuff and Gruber, 1983. Recently, the project has been extended:
It now includes also the reference-point approach to interactive
vector optimization which might prove as advantageous in econo-
metric decision models as Rosinger's algorithm (see Section 4 of
this paper).

2.2. The vector of local references to be approximated from

pieces of information
Rosinger's algorithm is based on the Franke/Wolfe algorithm for
the mathematical optimization of a scalar-valued objective func-
tion. In each iteration, two problems have to be solved: (a) the
direction problem and (b) the step-length problem.
The direction problem can be solved (e.g., by linear program-
ming) when the following vector is available (for details, see e.g.,
Streuff and Gruber, 1983, pp. 347 ff.):

d = (a;, ... ,a;, ... ,d )':= (aa~"·"~~'···'aa~)/.

m
Ylt uY;t Ymt
(2.1)
 67

Here, the symbols have the following meaning:

U denotes the scalar-valued objective function. This function
needs not be specified explicitly and filled out numerically in vector
optimization. (If it were available in this form, straightforward
optimization could be applied, i.e., there would be no need for
interaction with the decision maker and for vector optimization.)
Y~t denotes the jth objective (target variable) of the econometric
decision model. This variable is the dependent variable of the jth
reduced-form equation or an instrument variable that in the
opinion of the decision maker plays the same role as a target
variable.
t denotes time. The planning horizon may comprise one or more
time periods.
The superscript s in (2.1) denotes formally the number of iter-
ations (s = 1, 2, ... ). It also means that the partial derivatives are
evaluated at the sth combination of target and instrument variables
("local point", "scenario", "menu").
In multiple criteria decision making, the vector a' in (2.1) cannot
be obtained by simple partial differentiation of U in each iteration,
as is done in classical optimization (e.g., by the Franke/Wolfe
method). Instead, a' needs to be approximated in each iteration. Let
us denote an approximation of a' by
S
Q. == ( S ,
a*l,···,a*J, -.s )'
... ,u*m· (2.2)
The j th element of this vector is
s au
a.) appr -a. (2.3)
Y;t
i.e., a!J approximates the unknown partial derivative of the scalar-
valued objective function U with respect to the jth objective at the
sth local point. a~J approximates the slope of the hyperplane to U
in the jth direction at the sth local point. This means that «'J
approximates the "tangent" to the utility hill (whose form is
unknown) in the direction ofthejth objective (target or instrument)
at the sth combination of target and instrument variables.
Depending on the interpretation of U, a!J may be called the
"marginal utility ofthejth objective", "marginal (welfare) cost of
the j th objective" or the like (always at the sth local point).
To solve the direction problem, only the direction of the vector
a; in (2.2) is of interest. Therefore, a! may be normed arbitrarily.
68

2.3. Major difference between the Geoffrion algorithm and the

Rosinger algorithm
In Geoffrion's interactive algorithm the decision maker is assumed
to be able and willing to supply in each iteration the complete vector
a! in (2.2). (See Geoffrion, Dyer, Feinberg, 1972.) This may be an
unrealistic assumption: It may be difficult for the decision maker to
state all elements a!J of the vector do in a "consistent" manner,
whereby one objective has to be taken as the reference objective.
Usually, the problems increase with the heterogeneity (and the
number m) of the objectives.
In Rosinger's algorithm, the decision maker has to supply neither
the complete vector do in (2.2) nor subvectors of it directly. Instead,
the decision maker in each iteration supplies only information on
the direction of sub vectors of a' in (2.1).
Each subvector corresponds to a group of objectives. Such a
group consists of at least two objectives; one of these objectives
in each group is taken as the reference objective. Therefore, in
Rosinger's algorithm there are as many reference objectives as the
decision maker likes to form groups of objectives. Each group of
objectives contains those objectives for which the decision maker is
able and willing to state marginal utilities according to (2.3).
Preference statements are to be most easily made if each group
consists of only two objectives. The decision maker is only asked to
state in each iteration whether an additional unit of an objective
increases or decreases his utility. Furthermore, he is asked whether
it increases or decreases his utility more than a marginal unit of the
other objective (reference objective) in the same group of objectives.
Thus, in each group ofn objectives, information on the direction of
an n-dimension subvector is collected, with 2 ~ n ~ m. Some
examples given below will illustrate this point.
The information on the direction of subvectors supplied by the
decision maker in each iteration is in Rosinger's algorithm taken as
an input for a quadratic optimization problem (see, e.g., Streuff and
Gruber, 1983, pp. 345 ff). By solving this optimization problem, the
computer calculates in the sth iteration a vector a! which approxi-
mates the unknown vector a! in (2.1). In this way, contradictory
information given by the decision maker is reconciled.
On the basis of a! (and other information, e.g., the reduced form
of the econometric model) the direction problem can be solved.
 69

2.4. Great flexibility of the decision maker in forming groups of

objectives in Rosinger's algorithm
In order to be able to compute the vector a! in (2.2) by Rosinger's
method, the decision maker has to supply in each interation pieces
of information on his local preferences such that
(1) an inquiry pattern can be formed and
(2) the corresponding answer matrix can be filled in.
Let the inquiry pattern in the sth iteration be denoted by PS. This
matrix consists of m columns (one for each objective in the model)
and of pS rows (one for each group of objectives). The superscript
s denotes that the groups and their number may be altered in each
iteration. The groups are formed such that the decision maker is
able and willing to supply information on the local marginal utili-
ties (or trade-offs) for the objectives within each group. In forming
the groups of objectives, the following three conditions have to be
fulfilled:
(1) Each group consists of two or more objectives. In the i th group
of objectives, each objective belonging to this group is denoted
by 1. All objectives not in the i th group are denoted by zeros.
Therefore, the ith row of the inquiry pattern ps contains as
many 1's as there are objectives in the i th group.
(2) Each of the m objectives in the vector maximization model has
to be contained in at least one group, i.e., each column of the
inquiry pattern P' has to contain at least one I-element.
(3) The groups of objectives have to be formed such that each
inquiry pattern is "connected". This means some type of "over-
lapping" of the groups and is illustrated by means of the
examples below. (The mathematical details are, e.g., given in
Streuff and Gruber, 1983, pp. 343 f, and in Streuff, 1983, pp.
146 ff.)

EXAMPLES: In an economic policy problem let there be in the tth

period 5 objectives, namely three target variables:
Yr gross national product, GNP
AI number of persons employed
PI price level,
and two instrument variables:
GI government expenditures
RI the interest rate.
70

Let the planning horizon consist of two periods: t = 1, 2. Then the

set of objectives will contain m = 10 elements. They are shown in
the head-row of Table 1.

EXAMPLE I OF AN INQUIRY PATTERN: Row I of Table I results when

the decision maker is willing and able to compare the marginal
"utility" of the five objectives in period 1 which form group 1.
Similarly row 2 refers to the objectives in period 2. In each of these
two groups, no intertemporal comparisons are required. There is,
so to speak, no overlapping of groups 1 and 2.

Table I. Example I of inquiry pattern P'

Group r; AI PI GI RI r; A2 P2 G2 R2
Group I I I I 0 0 0 0 0
Group 2 0 0 0 0 0 I I
Group 3 0 0 0 0 0 0 0 0

If the decision maker tried to form the inquiry pattern Ps only

from these two groups he would be asked by the computer to
specify at least one additional group such that connectedness is
reached. Group 3 serves as an example of such an additional group.
Thus, the lack of overlapping of the objectives in groups 1 and 2 is
overcome. Group 3 states that intertemporal comparisons are to be
made for objective 1'; (GNP) in two periods. The Rosinger
algorithm is flexible enough to allow the decision maker to state his
preference also on additional groups of objectives.

EXAMPLE 2 OF AN INQUIRY PATTERN: Table 2 shows an inquiry

pattern Ps which results when the decision maker is willing and able
to make, first of all, intertemporal comparisons in groups of size 2:
In each of the groups 1 to 5 of this inquiry pattern, comparisons of
one objective in two periods are called for. In order to reach
connectedness of the inquiry pattern, at least one additional group
has to be specified, for example group 6 which requires comparisons
of the five objectives in period I.
The preference statements are easiest to make if each group of
objectives consists of only two elements, one of which is taken as the
reference objective. In each group of two objectives, a different
reference objective may be chosen. For example, in group 6 of
 71

Table 2. Example 2 of inquiry pattern P'

Group }! Al PI GI RI 12 A2 P2 G2 R2

Group 1 0 0 0 0 I 0 0 0 0
Group 2 0 I 0 0 0 0 I 0 0 0
Group 3 0 0 I 0 0 0 0 I 0 0
Group 4 0 0 0 I 0 0 0 0 I 0
Group 5 0 0 0 0 0 0 0 0 I
Group 6 I I I 0 0 0 0 0
Group 6' 0 0 0 0 0 0 0 0
Group 7 0 I 0 0 0 0 0 0 0
Group 8 0 0 I 0 0 0 0 0 0
Group 9 0 0 0 I 0 0 0 0 0

Table 2 the marginal utilities of five objectives are to be compared,

with one of these five objectives taken as the reference objective.
Preference statements in groups of this size may be difficult to
make. These difficulties may more than proportionately increase,
when the increase in the number of objectives coincides with an
increase in the heterogeneity of the objectives. This explains why in
Geoffrion's algorithm (in which only one group consisting of all m
objectives is formed) the problems may become unsolvable. In
Rosinger's algorithm all groups may principally consist of only two
objectives. This leads to an increase in the number of groups
required. For example, instead of group 6 in Table 2 (consisting of
5 objectives) the four groups 6', 7, 8 and 9 can be formed.
These two examples of the inquiry pattern Ps illustrate the great
flexibility of the decision maker in Rosinger's algorithm to state his
preferences in any iteration. They contain more columns than
needed in applications on the basis of dynamic equation systems. In
this case it suffices to include in the list of objectives the instrument
variables of all periods and the target variables of the last period of
the planning horizon (see, e.g., Streuff and Gruber, 1983, p. 360,
and Streuff, 1983, pp. 103 ff).

2.5. Great flexibility of the decision maker in filling in the

answer matrix in Rosinger's algorithm
After the decision maker has formed the inquiry pattern p s , he
is asked by the computer to fill in the nonzero elements of the
72

corresponding answer matrix D S :

DS = (d~), i = 1, ... ,ps, j = 1, ... , m. (2.6)
This answer matrix is of the same size as the inquiry pattern. All
elements that correspond to zero elements of the inquiry pattern ps
are zero.
The (i, j )th nonzero element d~ states the local preference of the
decision maker: d~ approximates the unknown au/aY;t in the ith
group and at the sth local point (see Section 2.2).
We illustrate the great flexibility with which the decision maker
may fill in the answer matrix D S by means of example 1 given in the
inquiry pattern in Table 1.
- An additional unit of ~ (GNP) increases his utility, i.e.
aUla ~ > O. Let us take ~ as the reference objective in group
1. Then we are free to assign d{1 approx au/a~ any positive
value, for example d{1 = 1
- An incremental unit of Al (number of persons employed)
increases, in the opinion of the decision maker, his utility
ceteris paribus twice as much as an additional unit of ~. In
this case we have d{2 = 2 appr aU/aA I .
- A marginal increase in PI (price level) diminishes the utility by
the same amount as an additional unit of ~ increases it. Now
we have d{3 = - 1 appr au/aP I .
- A marginal unit of G I (government spending) decreases the
utility of the decision maker ceteris paribus about half as much
as an additional unit of ~ increases it. This preference state-
ment of the decision maker leads to d{4 = - 0.5 appr ·au/aG I .
- A incremental increase of RI (interest rate) is supposed to
decrease the decision maker's utility about 0.1 times as much
as an additional unit of ~ increases it. Here we have
d{5 = - 0.1 appr aU/aRI.
Collecting the values assumed in this hypothetical example, the
following vector d{ is obtained:

( au aA
au au au au)'.
appr a~' I ' aPI' aG I ' aR I
This vector can be seen in row I of Table 3. Since only the direction
 73

of this vector matters, the elements of this vector may be multiplied

by any positive number.
A vector like df in (2.8) for the first group of objectives is specified
by the decision maker for all groups of objectives in the inquiry
pattern P" i.e., df for i = 1, ... , pS.
In our hypothetical example 1 let the result of inverviewing the
decision maker for his local preferences in the sth iteration be the
answer matrix D in Table 3.
S

Table 3. One version of an answer matrix D' for Example I of the inquiry pattern.
Group Y; Al PI GI RI l2 A2 P2 G2 R2
Group 1 2 -1 -0.5 -0.1 0 0 0 0 0
Group 2 0 0 0 0 0 I 2 -1 -0.5 -0.1
Group 3 I 0 0 0 0 1 0 0 0 0
Group 4 0 I 0 -0.5 0 0 1.5 0 -I 0
Group 5 2 2 0 -1 0 2 2 0 -I 0

This table needs some further explanation. The non-zero figures

of groups 1 and 2 reflect that the decision maker has expressed the
same preferences for periods 1 and 2. This, of course, needs not be
so. For example, if period 2 is an election period, the policy maker
could in this period have given geteris paribus much more "weight"
to the growth of employment and to the stability of the price level
than in period 1. In this case the corresponding vector d~ could, e.g.,
be d~ = 1,4, -2, -0.5, -0.1)'.
Group 3 calls for intertemporal comparisons. If the decision
maker's preferences are constant in time, the vector d3contains two
elements of the same size, e.g., dt, = (1, 1)' appr (8U/8~, 8U/8Y;)'
(as shown in Table 3). The values of these elements coincide with
the values of the corresponding first element of df and d~. But this
equality of nonzero elements in any column of D S is not required by
Rosinger's algorithm. Instead, the decision maker may "con-
tradict" himself up to a certain extent. For example, d3 could be
dt, = (2, 2)' in case of preferences which are constant in time and
(dt, = (2, 5)' in case of special emphasis on the growth of Y2 (GNP
in period 2).
A look at Table 3 reveals that the 4 columns headed by PI and R(>
t = I, 2, contains only one non-zero element; in these columns
there can be no contradictory statements on the local preferences of
the decision maker. But all other columns of the answer matrix DS
74

in Table 3 contain three non-zero elements: Not all non-zero elements

of one column are of the same magnitude, and this is permitted by
Rosinger's algorithm. But all non-zero elements of one column are
of the same sign, as required by Rosinger's algorithm.
This means that the decision maker's statements on his local
preferences (i.e., on the effect of a marginal unit of an objective on
his utility at the sth point) in different groups of objectives may be
contradictory as far as the magnitude of the non-zero entries in any
column is concerned. But these contradictions may not go so far
that the non-zero entries in any column of D are of different sign.
S

Contradictory preference statements of the type outlined so far

are reconciled by quadratic optimization. Herein the vectors dis,
i = 1, ... , pS, are taken as an input. Its output is the vector a! in
(2.2) which approximates the unknown vector as in (2.1). With do,
the direction problem of the sth iteration can be solved. For details
see Streuff, 1983, pp. 30 ff.
We now very briefly address the following questions: What are
the consequences of preference statements such that the direction of
diS (i = 1, ... , p') is a poor approximation of the direction of the
corresponding subvectors of a' in (2.1)? What are the consequences
if the solution vector do of the quadratic optimization problem is
also a "less than ideal" approximation of a' in (2.1)?
No general and definite answer to these questions can be given.
It is reasonable to suppose that the quality of the approximation of
a' in (2.1) affects the number of iterations required for solving the
multiple-criteria decision problem: The better a' is approximated by
do, the smaller is ceteris paribus the number of iterations needed for
obtaining the optimal solution.
If the structure of the decision problem is such that the decision
maker feels uncomfortable in stating numerical values of the non-
zero elements of D', it should suffice in Rosinger's algorithm to
proceed as follows:
(a) In specifying the inquiry patterns PS, form groups that contain
only two objectives.
(b) In specifying any non-zero d; -element of the answer matrix D",
ask the decision maker whether at the current point an
additional unit of the jth objective
(b 1) increases his utility or
(b2) decreases his utility or
(b3) does not changes his utility perceptibly.
 75

Then the person in charge of the computer program sets numeri-

cal values, for example
d~ = I in case of (bI),
d~ = - I in case of (b2) and
d~ = 0 in case of (b3), provided that consistency holds.
In such an interactive scheme, the number of rows in ps
and D will be relatively large. Also the number of iterations
S

(s = I, 2, ... ) may be considerably larger than in case that - which

should be the rule - more informative numerical values for the
non-zero d~-elements are given directly by the decision maker. It is
unknown how poor the information supplied by the decision maker
may be without destroying the feasibility of Rosinger's algorithm.
In summary, Rosinger's algorithm is so flexible in forming the
inquiry patter Ps and the answer matrix D that it can easily be
S

adjusted to the structure of the particular decision problem and to

the preferences of the decision maker.

2.6. Solving the step-length problem by interaction with the

decision maker
When the direction problem has been solved (e.g., by linear
programming), then the step-length problem is in Rosinger's
algorithm solved exactly in the same way as in the algorithm of
Geoffrion, Dyer and Feinberg:
The decision maker is asked by the computer "how far" he wants
to go in the direction of the steepest ascent. The decision maker may
be shown a summary like the one in Table 4 or a corresponding
graphical presentation: Each column of Table 4 shows the values of
the objectives that are forecast by the model if the decision maker
goes )'1 percent of the maximally available step-lt::ngth in the direc-
tion of the steepest ascent. The decision maker will select that
column of Table 4 which he prefers, i.e., which maximizes his
"utility". (He may also select any other step-length between 0
percent and 100 percent.)
If the decision maker considers the selected column of Table 4 as
the "globally" optimal combination of target and instrument vari-
ables, then the interactive vector optimization is completed. If the
decision maker is not yet satisfied with the selected column and if
there might exist a preferred combination of target and instrument
variables, another iteration of the Rosinger algorithm is started (see
Section 2.2).
76

Table 4. Information for solving the step-length problem.

Objective Value of objectives, if A, percent of the maximally available
step-length are gone in the sth iteration
Al = 0 A2 = 20 A] = 40 A4 = 60 As = 80 ,.1.6 = 100

Y1 Yi (0) Yi (20) Yi(40) y;(60) Yi (80) yi (100)

YJ y5(0) yj(20) y;(40) yj(60) yj(80) yj(lOO)

Ym )/",(0) Ym(20) Ym(40) )/",(60) Ym(80) y:"(lOO)

The decision maker most likely is already well acquainted with

this type of procedure for solving the step-length problem from a
different context: If econometric forecasting systems are used as an
aid in preparing policy decisions, the decision maker usually is
presented several sets of instrument values and the resulting sets of
values of target variables.
This enhances the probability of widespread applications of
Rosinger'salgorithm in conjunction with econometric equation
systems.

3. Research on the Application of Rosinger's Algorithm

A research project on Rosinger's interactive method was initiated

at the University of Hagen in 1979. The aim of this project was to
evaluate the suitability of Rosinger's algorithm for macroeconomic
programming on the basis of an econometric model. The following
four phases of research may be distinguished:

Phase 1: Verifying the theoretical requirements and conditions

used in the algorithm
On the one hand it had to be guaranteed that those requirements
of Rosinger's algorithm that are needed for its convergence and for
the existence of solutions are fulfilled in the framework of an
econometric decision model. On the other hand some inconsis-
tencies in Rosinger's proof of the convergence had to be clarified.
Especially the proof of Lemma 5.2 in Rosinger (1981) had to be
modified.
 77

A detailed description of the algorithm involving the modified

proof of its convergence as well as examples to illustrate the defi-
nitions and theorems included may be found in Streuff (1983).

Phase 2: Formation of a time-sharing computer program

In the next phase of the research project an interactive computer
program for the Rosinger algorithm was developed. The program
was written in PL/l using standard FORTRAN routines for the
minimization of the quadratic function that results in the vector a!
in (2.2), namely a quasi-Newton method, and for the linear pro-
gram that solves the direction problem, namely a modified simplex
method. The program is modularly structured and may easily be
adapted to various decision problems. Input and output are man-
aged via the CRT-screen only, but a complete documentation of
each session is stored in a separate file.
A detailed description of the computer program, its subroutines
and program parameters is presented in Streuff (1983). More infor-
mation about the possibilities of applying the program to other
decision problems may also be found there.

Phase 3: Evaluation of the computer program

To perform applications of Rosinger's algorithm in the frame-
work of an econometric decision model, an econometric model of
the West German economy was combined with the computer pro-
gram. The Beckmann-Uebe model was chosen for several reasons:
It consists of 15 equations; therefore it is of a size suitable for the
purposes of the study. It is linear in the parameters as well as in the
variables (growth rates, annual data). Its dynamic structure is
simple: Only first order lagged endogenous variables appear in the
equations. It has been used in former studies in connection with the
control theoretical approach, thus making it possible to compare
both methods in a later study.
As target variables were selected:
y gross national product,
A number of persons employed,
py = price index of the gross national product.
78

The instrument variables are:

G = government expenditures,
R = discount rate,
TY = indirect taxes,
The planning horizon was set to be 1965 to 1968, i.e., four
periods.
The program evaluation was done on the one hand by simulating
realistic decision situations and on the other hand by testing the
algorithm in extreme situations, i.e., when the inquiry pattern ps
(and thus the answer matrix D contains little information. The
S
)

latter happens if
- the inquiry pattern consists only of a small number of groups of
objectives,
- the number of active objectives is small in comparison with the
number of objectives specified (i.e., if the inquiry pattern contains
many zero-columns).
- the answer matrix DS (which corresponds to a "well-specified"
inquiry pattern) contains extremely contradictory information
(without being inconsistent).
The results of numerous applications of the interactive computer
program combined with the Beckmann-Uebe-model can be sum-
marized as follows:
1. Rosinger's interactive algorithm for multiobjective optimization
is a suitable tool for macroeconomic programming - from the
theoretical point of view as well as in realistic decision situations.
2. In realistic decision situations the speed and safety of con-
vergence proved to be satisfactory: In most cases 10 to 12
iterations were sufficient to reach an/the optimal solution ("opti-
mal" in terms of the decision maker's preferences).
3. Problems arise in case of "extreme" situations (in the above
sense): The standard optimizing routine used for the solution of
the quadratic minimization problem that yields the vector a! in
(2.2) often did not reach a minimal value in the given (and
unalterable) number of iterations. It is planned to replace the
quasi-Newton-procedure by a procedure with a higher speed of
convergence.
 79

Phase 4: Present and future work with Rosinger's algorithm

The encouraging results of the previous phases of research,
especially of phase 3, are the basis for the continuation of the
research effort in the statistics and econometrics unit of the Univer-
sity of Hagen. Presently (February, 1985), mainly the following
goals are pursued:
- Work with Rosinger's algorithm on the basis of an updated
version of the Beckman-Uebe model and/or a similarly "handy"
model.
- Adapting the computer program developed by H. Streuff to the
presently available IBM 3031 computer with operating system
VM.
- Replacing program subroutines that caused problems in extreme
situations (see above) by more suitable routines.
- Test runs with university staff members and students.
- Test runs with econometricians from institutions and institutes
which are regularly preparing policy decisions on the basis of
econometric equation systems (e.g., from Federal Ministries in
Bonn, from the Federal Reserve Bank (Deutsche Bundesbank) in
Frankfurt, from West Germany's five "big" Institutes of Econo-
mic Research).
- Comparing the performance of Rosinger's algorithm for vector
optimization and of the reference point approach to interactive
vector optimization (see the following section).
- Application of Rosinger's algorithm to a large econometric
model of the West German economy (most likely only in close
cooperation with the producer and/or user of such a model).

4. Research on the Reference Point Approach to Interactive

Vector Optimization

A promising alternative to Rosinger's algorithm for interactive

vector optimization is the reference point approach. Its theoretical
foundations have been developed mainly by Wierzbicki (see Wierz-
bicki, 1980, 1981). At IIASA in Laxenburg/Austria, an interactive
algorithm has been programmed (see Grauer, Lewandowski and
Wierzbicki, 1982, Lewandowski and Grauer, 1982, and Grauer,
1983). The program package DIDASS (see Grauer, 1983) for multi-
criteria decision analysis is based on the reference point approach
80

(aspiration level approach). During the interactive solution process,

the decision maker can change his aspiration levels. He may also
change parameters of an "achievement scalarizing function" so that
the computation of feasible solutions is steered to the optimal
solution in a relatively small number of iterations.
In a research project at the University of Hagen that was started
in 1983, the applicability of the reference point approach to econo-
metric decision models is investigated in a similar way as Rosinger's
algorithm. Both approaches are to be compared.

5. Concluding Remarks

According to my present knowledge, the control theory approach

and the vector optimization approach to (micro- and macro-)
econometric decision models are complementary tools of economic
analysis. Neither one alone canI serve all important purposes of
decision analysis equally well. The control theory approach
presently seems to be better suited for "theoretical" studies. The
vector optimization approach may be of advantage in "practical"
studies. There is considerable interest in and work on implementing
both approaches in some of the large economic research institutes
and in some Federal agencies in West Germany. Both approaches
need to be further investigated. Direct and indirect gains from such
investigations may be substantial.
The applicability of the control theory approach to econometric
decision models will be improved especially with the development
of improved methods for numerically determining the scalar-valued
objective function. If only a reasonable fraction of the effort for
constructing econometric equation systems were devoted to this
field, there should be noticeable progress - in spite of problems like
changes of preferences due to changes in the government or the
business cycle or the election cycle.
The control theory approach would also gain considerably from
the development of procedures and software for easily incorporat-
ing mathematical programming techniques into the control theor-
etical model.
 81

References

Ancot, J. P and Hughes Hallet, A. J. (1983): The determination of collective

preferences in economic decision models: With an application to Soviet econo-
mic policy. In: Gruber, J. (ed.) 1983, 263~304.
Chow, G. (1975): Analysis and Control of Dynamic Economic Systems. John Wiley
and Sons, New York, London.
Chow, G. (1981): Econometric Analysis by Control Methods. John Wiley and Sons,
New York, a.e.
Feichtinger, G. (ed.) (1982): Optimal control theory and economic analysis. First
Viennese workshop on economic applications of control theory held in Vienna,
October 28~30, 1981. North-Holland, Amsterdam, New York, Oxford.
Fox, K. A. Sengupta, J. K., and Thorbecke, E. (1966): The Theory of Quan-
titative Economic Policy with Applications to Economic Growth and Stabilization.
North-Holland Publishing Company, Amsterdam (second edition 1973).
Friedman, B. M. (1975): Economic Stabilization Policy. (Methods in optimization).
North-Holland, Amsterdam.
Geoffrion, A. M., Dyer, J. S. and Feinberg, A. (1972): An interactive approach
for multi-criteria optimization, with an application to the operation of an
academic department, Management Science 19, 357~368.
Grauer, M. (1983): A dynamic interactive decision analysis and support system
(DIDASS). User's guide (May 1983). System and Decision Sciences Program,
International Institute for Applied Systems Analysis (IIASA), Laxenburg,
Austria.
Grauer, M., Lewandowski, A., and Wierzbicki, A. P. (eds.) (1982): Multiobjec-
tive and stochastic optimization. International Institute for Applied Systems
Analysis Collaborative Proceedings Series 12, Laxenburg, Austria.
Grauer, M., Lewandowski, A., and Wierzbicki, A. P. (eds.) (1982): Multiobjec-
tive and stochastic optimization. International Institute for Applied Systems
Analysis Collaborative Proceedings Series 12, Laxenburg, Austria.
Gruber, J. (1979): Approaches to determining the weights in the objective function
of econometric decision models. Discussion Paper No. 35, Department of
Economics, University of Hagen.
Gruber, J. (ed.) (1983): Econometric decision models. Proceedings of a Conference
at the University of Hagen, June 19~20, 1981. Lecture Notes in Economics and
Mathematical Systems, Vol. 208, Springer-Verlag, Berlin, Heidelberg, New
York.
Gruber, J. (l983a): Flexible Inputgestaltung in Rosingers interaktivem Verfahren
zur Entscheidung bei mehreren Zielen (Vektoroptimierung), Agrarwirtschaft
(= Zeitschrift fiir Betriebswirtschaft, Marktforschung und Agrarpolitik), Vol.
32, 305~312.
Hansen, P. (ed) (1983): Essays and survey on multiple criteria decision making.
Lecture Notes in Economics and Mathematical Systems, Vol. 209, Springer
Verlag, Berlin, Heidelberg, New York.
Johansen, L. (1979): The report of the committee on policy optimization, UK.
Journal of Economic Dynamics and Control 1, 101~109.
82

Kendrick, D. (1981) Stochastic Control for Economic Models. McGraw-Hill, New

York.
Kuhbier, P. (1981): Grundlagen der quantitativen Wirtschaftspolitik. Walter de
Gruyter, Berlin, New York.
Leontief, W. (1971): Theoretical assumptions and nonobserved facts. The Ameri-
can Economic Review 61, 1-7.
Lewandowski, A. and Grauer, M. (1982): The reference point optimization
approach - methods of efficient implementation. Working Paper WP-82-26,
International Institute for Applied Systems Analysis, Laxenburg/Austria.
Loon, P. van (1982): A dynamic theory of the firm: Production, finance and
investment. Dissertation. Tilburg.
Merkies, A. H. Q. M. and Nijman, Th. E. (1983): The measurement of quadratic
preference functions with small samples. In: Gruber, 1. (ed.) 1983, 242-262.
Murata, Y. (1982): Optimal Control Methods for Linear Discrete-Time Economic
Systems. Springer-Verlag, New York, Heidelberg, Berlin.
Pau, L. F. (1979): Research on optimal control adapted to macro- and micro-
economics. Journal of Economic Dynamics and Control 1, 243-269.
Rosinger, E. E. (1981): Interactive algorithm for multiobjective optimization.
Journal of Optimization Theory and Applications 35, 339-365.
Rosinger, E. E. (1982): Erratum. Journal of Optimization Theory and Applications
38, 147-148.
Shupp, F. R. (1979): Control theory and stabilization policy. A review of the
report of the Committee on Policy Optimization. Journal of Economic Dynamics
and Control 1, 111-1 16.
Streuff, H. (1983): Ein interaktives Verfahren der Vektoroptimierung angewandt
auf okonometrische Entscheidungsmodelle - Der Rosinger-Algorithmus: zur
Theorie, seine programmtechnische Realisation, Aspekte seiner Anwendung.
Haag und Herchen, Frankfurt.
Streuff, H. and Gruber, 1. (I983): The interactive multi objective optimization
method by Elemer E. Rosinger: a computer program and aspects of appli-
cations. In: Gruber, 1. (ed.) 1983, 334-364.
Szego, G. P. (1982): Mathematical methods for economic analysis: a biased
review. In: Szego, G. P. (ed.) New Quantitative Techniquesfor Economic Analy-
sis. Academic Press, New York, 3-17.
Wierzbicki, A. P. (1980): The use of reference objectives in multi objective
optimization. In: Fandel, G. and Gal, T. (eds.), Multiple Criteria Decision
Making, Theory and Application. (Proceedings, Hagen/Konigswinter 1979).
Lecture Notes in Economics and Mathematical Systems. Vol. 177. Springer-
Verlag, Berlin, Heidelberg, 468-486.
 83

CHAPTER 6

RISK REDUCTION AND THE ROBUSTNESS OF

ECONOMIC POLICIES

Andries S. Brandsma
Erasmus University, Rotterdam, The Netherlands

Introduction

Increasing the robustness of economic policies by using optimis-

ation techniques which try to reduce the impact of uncertainty is an
important desire of policy makers. Unfortunately, certainty equival-
ence (Theil, 1964) applied to a quadratic objective (subject to linear
constraints) implies decisions which are invariant to the degree of
uncertainty involved and hence invariant to the magnitude of the
risks undertaken by the decision maker (Johansen, 1980). So, while
the importance of risk reduction is widely recognised, economists
have seldom been able to compute empirical risk sensitive decisions
for the multivariable dynamic control problems which they face in
practice. In Hughes Hallett (1984a) decisions were proposed which
aim to reduce the risks involved in the realisation of a certainty
equivalence strategy by including higher moments of the objective
function in the criterion to be optimised. Hughes Hallett (1984b)
considers regression analogies which try to reduce the mean square
errors with respect to the ex post optimal decisions. This paper
contrasts the theoretical structure and properites of both types of
decisions, and investigates their comparative performance when
applied to a given empirical model in an optimal control context.
The first approach, representative of that advocated by economic
theory (Arrow, 1970; Malinvaud, 1972; Johansen, 1978), is to take
a weighted sum of the moments of the underlying stochastic objec-
tive function as the optimisation criterion. In Brandsma and Hughes
Hallett (1984) we have argued that, since those moments are
expectations of different characteristics of the underlying "true"
objectives and since, taken together, they completely define those

C. Carraro and D. Sarlore (eds). Developments of Control Theory for Economic Analysis
© 1987 Martinus NijhoJf Publishers (Kluwer), Dordrecht·
84

objectives, this amounts to specifying a von Neumann-Morgenstern

utility function for the decision problem. In practice we operate
with the first two moments only, so that our criterion may be
interpreted as a second-order approximation to the von Neumann-
Morgenstern function about its risk neutral component (cf. New-
bery and Stiglitz, 1981). On the other hand, our approach
generalises on the "exponential approach" of Whittle (1982) in that
it is distribution free and explicitly recognises the intertemporal
stochastic dependence of risk sensitive decisions.
The other approach is to examine the mean square errors of the
optimised decision variables under uncertainty. The standard
statistical concept of risk is employed, and the result is used that
reducing mean square errors in a linear-quadratic selection problem
is equivalent to reducing the usual statistical loss or risk function
E(x - xoy C(x - xo) for that problem, where C is any positive
semi-definite matrix and Xo is the "true" value of the decision
variables under complete and perfect information. Moreover, the
expected losses due to uncertainty are generated when a particular
value of C is used. So this approach just requires the minimisation
of the mean square errors of x for a given model and objective
function.
The decision rules which follow from either approach can be
expressed in terms of reparameterisations of the original preference
structure (Section 4). However, in three respects reparameterisation
is more far reaching in the first approach than in the second; more
components of the preferences are changed, asymmetries are intro-
duced in the preference structure (cf. Waud, 1976), and more
characteristics of the underlying distributions are involved. These
results are illustrated empirically by application to a standard
example of Dutch macroeconomic planning (Section 5). Risk sen-
sitive decisions which fit into the theory of risk bearing are described
in Section 1. The regression analogy, established in Section 2, is
exploited to produce "optimal" ridge regression decisions in Sec-
tion 3. Both types of decisions are compared numerically, according
to a number of criteria which reflect their respective aims and
properties; the results are in Section 6. Finally, Section 7 investi-
gates some other decision rules from the class of ridge estimators.
 85

1. Risk Sensitive Decisions

Suppose the decision problem involves a vector of n instruments

Xt stacked as x' = (x; ... x~) over the planning period. Similarly
let y' = (y; ... y~) be a pT vector of targets and s a vector of
uncontrollable variables of the same length. They are connected by
a linearised model of the form y = Rx + s.
Consider furthermore an objective function which is additively
separable between targets and instruments:
w = !(y'By + x'Ax) + q'y, (1)

where B and A are positive semi-definite matrices and B has rank

at least nT. Transformation of the model into y = Rx + b where
b = s - yd + Rxd , and substituting it into (1), yields optimal
decision values as (Brandsma, Hughes Hallett and Van der Windt;
1983, 1984)
x* = -(R'BR + A)-IR'(Bb + q). (2)

Let M = R' BR + A and m = R'(Bb +q), then x* - M- 1 m.

In response to the uncertainties, w is replaced by the expectations
Et(w) as the criterion at each t and the certainty equivalence
theorem establishes the optimal decisions as x* = - M- 1 Et(m)
with Et(m) = R'(BEt(b) + q).
In Brandsma and Hughes Hallett (1984) we derived the risk
averse decisions, which minimise the variance of the objective
function, as
(3)

where 'I' = Et«(J(J') and v = tEt«(J(J'B(J). It is possible to write

x** = - K- 1 Et(k) with K = R' B'I' BR and k = R' B('I' Bb + v).
These decisions depend on the first, second and third moments of
the density function of the random "final form" variables in b. In
contrast, the certainty equivalent decisions, x*, depend only on the
first moment of the density function of b.
The risk averse decisions, x**, can be elaborated in various ways.
One possibility is to determine the dynamic mean-variance decisions
defined by
min {taV;(w)
x
+ (1 - a)Et(w)}, (4)
86

where rx is a risk aversion parameter. The solution to (4) is

Xl = - (rxK + (1 - rx)M)-1 E(rxk + (1 - rx)m). (5)

A comparison with (2) and (3) reveals that rx = 1 implies Xl = x**,

and that rx = 0 implies Xl = x*; intermediate values trade security
-indexed by V;(w)-against ambition-indexed by EI(w). It is
interesting to notice that (5) just involves a transformation of
the original preference parameters (B, A, q = 0) to (rxB'I' B +
(1 - rx)B, (1 - rx)A, rxBv).
An obvious simplification is to construct scalar linear combina-
tions, say

Xl = px* + (l - p)x**, (6)

where 0 ~ p ~ 1. Let Wo denote the increase in EICw) when some

X =I=- x* is chosen, and Vo be the increase in ~Cw) when x* =I=- x**
is used. Then, by Theil (1964, p. 145), Wo = Hx - x*)' M(x - x*).
Similarly the transformation of (Q, q) permits x** to be treated as
if it were a certainty equivalent decision under this reformulated
penalty scheme. That implies Vo = Hx - x**)' K(x - x**).
Moreover, since (6) implies x* - Xl = (1 - p)(x* - x**), we
have Wo = t(l - p)2(X - x**)' M(x* - x**) and similarly Vo =
tp2(X* - x**)' K(x* - x**). Therefore the value of p which mini-
mises (rxvo + (1 - rx)w o) subject to (6) is
(7)

where WI = (x* - x**)' M(x* - x**) and VI = (x* - x**)' x

K(x* - x**) are known. This formulation is convenient because rx
then explicitly trades a loss in ambition (w o) against a loss in security
(v o), and (6) with (7) supply combinations of x* and x** which
minimise those losses. In combining the decisions we assume that K
and M have similar scales. It is interesting to note that p* = t if
rxK = (l - rx)M. When A = 0 and rx = t, it depends on the posi-
tive or negative definiteness of B('I' B-1), so among other things
on the scale of the covariance matrix, whether p* tends to steer (6)
towards the risk averse or the certainty equivalent decisions.
 87

2. Risk Management in Statistical Decisions

In the statistical literature the problem of risk is handled in a rather

more straightforward manner. It is conventional to set up a risk
function which measures the losses to be expected from implement-
ing decisions using incomplete information, relative to those
decisions computed with complete and perfect information. If we
take a quadratic loss function measured in the metric of C, where
C is any positive (semi) definite matrix, then the appropriate risk
function would be E(x* - xO)'C(x* - xo). For the present prob-
lem we have takenx* = - M- 1 EtCm) from (2) andxo = - M- 1 m;
but in general x* will represent the (ex ante) decision rule to be
implemented and XO the (ex post) decisions which would have been
implemented under perfect information. The objective then is to
pick x* in such a way as to minimise these expected losses, given C
and what should best be done in the absence of any uncertainty.
Consider now the model in the "regression mode"
(8)
which is set to minimise the fitting errors y jointly with the prior
information that the decision errors x should be as small as poss-

ible, all in the metric of Q = [: :J. The Theil-Goldberger mixed

regression estimator is the standard solution, which, with prior

information in the form of 0 = x + e, yields the "estimator"
(9)
which of course corresponds with (2), given q = O.
This regression analogy implies that A can be used to generate
risk sensitive decisions, in the sense of having reduced mean square
errors and expected losses compared to certainty equivalent solu-
tions when it turns out that the realised values of h do not coincide
with their expected values. This is done by treating A as the matrix
of additive ridge factors in the (weighted) least squares mixed
regression. It then follows by standard results that the second
moment (mean square error) matrix Et(x* - XO)(x* - xOy is
smaller when certain nonzero values of A are chosen compared to
the case where A = O. Smaller here means the difference between
these two second moment matrices is negative semi-definite.
88

The decisions using such values of A are then also superior to

those using A = 0 in that the implied stochastic loss or risk func-
tion 1/1 = Et(x* - xO)'C(x* - xo) is smaller for any symmetric
and positive (semi)definite matrix C; see Theobald (1974). However
we already noted that the loss due to uncertainty in period 1 (i.e.,
because at t = 1 only x*, rather than xo, can be implemented)
is given by W o = !(x* - xO)'(R' BR + A)(x* - xo). Hence, to
reduce the general risk function, 1/1, is to reduce the expected losses
due to uncertainty, Et(wo), in particular. In this analysis, the ex post
decisions, xO, are what in retrospect should have been done; but the
ex ante decisions, x*, are what one should best decide to do now on
the basis of an information set which is incomplete with respect to
the future.
The use of a risk function has some practical advantages over the
formulation of risk aversion usually found in economic theory. The
latter is constructed from some implicit utility function satisfying
the von Neumann-Morgenstern axioms of choice, and may lead to
problems in handling multivariate and intertemporal measures of
risk aversion. In practice the main difficulty will be to specify an
appropriate measure of risk aversion, in order to calculate optimal
decisions numerically, while the underlying utility function remains
implicit.
In contrast, the risks considered here are simply: (i) the risk of
getting the decision values "wrong" because the "true" optimal
values, XO , depend on random variables which are unknown at the
moment of calculation; and (ii) the risk of large losses in the
objective due to the realisations of its stochastic components,
w - Et(w), which depend on the random fluctuations in the target
variables. While the latter is clearly related to the economic concept
of risk aversion, which aims to reduce the variability of the targets,
it is not exactly the same thing. The advantage of the present
approach is that, for an arbitrary risk function, it yields a straight-
forward way of determining the appropriate objective function, and
consequently the optimal decision values, explicitly. This is
achieved because picking a strategy which reduces the mean square
errors associated with the instrument values (lowering the risk of
instruments being "wrong") simultaneously reduces the expected
losses associated with the corresponding random target realis-
ations, when measured in any metric or utility which can be defined
through the matrix C.
 89

3. Optimal Ridge Regression Decisions

At this point it is helpful to introduce a canonical form of the

model:
y = p ~ + b, (10)
where P = RG and ~ G'x. Here G is a matrix of eigenvectors of
R' BR, and A is a diagonal matrix of eigenvalues, such that G'G =
GG' = I and R' BR = GAG'. The benchmark optimal decisions,
against which reductions in mean square errors and expected losses
would be measured, minimise (1) with A = 0 and subject to (8):
(11 )
We have taken A = 0 in (1) for the moment to emphasise the
introduction of ridge factors as a form of specifying instrument
penalties, like we took q = 0 before introducing skewness coef-
ficients (cf. Section 1). So, (11) reflects a free hand with the instru-
ments and no allowance for risk. The corresponding ex post opti-
mal decision would be
(12)
which has mean Et(~O) = ~* and covariance matrix ~(~O) =
A-leA-I, where e = G'R'B'PBRG is a known, positive definite
and symmetric matrix and where 'P is the prior conditional covari-
ance of the targets. The (generalised) ridge regression version of the
decisions in C11) and (12) implies
~* = (A + D)-IG'R'BEtCb) (13)
and
~O = (A + D)-IG'R'Bb (14)
respectively, where D is a diagonal matrix. The "bias" in (13) is
[(A + D)-I - A-I]G'R'BEtCb)

[(A + D)-I - A- I ]{3, (15)

where {3 = G' R' BEt (b) is a known vector. The covariance matrix of
~o about ~* is

(16)
90

The "ridge regression" decisions in (13) may be reformulated as (9)

where, in terms of the original notation, A = GDG'. In this context
"unbiased" means that the ex ante (implemented) decisions will, on
average, coincide exactly with what will appear ex post (with full
certainty) to be the optimal decisions. And, in the same way,
minimum variance implies that the associated ex ante decisions are
such that, on the basis of current information, the corresponding ex
post decisions are distributed (in the sense of variances) most tightly
about them.
Now the mean square errors associated with the benchmark
decisions (11) appear as the diagonal elements of A-I SA-I; and the
mean square errors associated with (13) can be constructed from
(15) and (16). Hence the mean square error associated with (13) less
that of (11) is

./, = d,2(fJ; - f),,) - ~A,f)"d, i = 1, ... , nT. (17)

'1'1 A2(A + d) ,
" 1

The value of D which uniquely minimises the mean square error of

~* about its ex post optimal value ~o is given by D* = diag (d,*)
where

d,* = A,f),,/fJ; > 0, i = 1, ... , nT (18)

(see Hughes Hallett, 1984b). The corresponding value of A* =

GD*G' leads to a minimal mean square error for the implemented
decisions x* about their ex post optimal values xo. Hence this choice
of A * also minimises the risk function, measured in any metric,
incurred by having to implement x* before the realisations of the
stochastic variables (upon which the ex post "true" optimum XO
depends) can be known. Hence A * also minimises the expected
losses due to uncertainty.

4. Structural Properties and Similarities

Earlier we saw that if the probability distribution of the target

variables is symmetric, then optimising a combination of the means
and variances of the target variables yields a decision rule of the
form (9), where B is replaced by rxB'P B + (1 - rx)B. The scalar 11.
is a risk aversion parameter such that 11. = 0 defines risk neutrality,
 91

and the coefficient of relative risk aversion is defined as -IXVI (w)/

«(1 - IX)EI (W)2) for 0 ~ IX < 1. In the extreme case of IX = 1, the
effect or risk sensitivity is to change the target penalties to B\f1 Band
remove the instrument penalties. For IX =I=- 0 this change will have
an analogous effect to that on the decisions which minimise the risk
function by introducing A * into (9), if B(\f1 B - l) is negative
definite. That is, the net effect in both cases is a relative reduction
in target penalties provided the scaled target variances are not too
large. But that similarity ends when \f1 B becomes large. Where the
important targets are also high risk variables, the approach of the
last section, which picks robust decision rules to offset the impacts
of random shocks on the objective function and thus to limit its
potential losses will differ from the mean-variance approach, which
is to pick decision rules which hold the high risk variables closer to
their ideal paths than would otherwise be the case so that given
shocks will cause smaller variations about those paths. The
approach of the last section will, in that case, be characterised by
flexible targets and relatively "sticky" instruments; whereas the
usual ecomomists' risk aversion will produce stable targets and
flexible instrument settings for the same problem. That result is, of
course, precisely opposite to the optimal response to increasing
uncertainty over policy impacts (of which there is none here), which
calls for smaller rather than larger interventions in terms of
xi - x1 (see Brainard, 1967; Hughes Hallett and Rees, 1983).
In practice the instrument penalty matrix, which was set to zero
in Section 3 to represent the most ambitious strategy, will contain
positive elements to prevent infeasible or unacceptable policy
adjustments. This will have implications for both the economic and
statistical risk sensitive decision rules. Writing k = R' B\f1(BE{(b) +
\f1-1 v) in which the elements of \f1-1 v represent slewness coefficients
measured in the same units as the elements of b, it is immediately
clear that the economic decisions in (3) are scale invariant. But
minimising ~(w) + X' Ax gives K = R' B\f1 BR + A in (3) and this
ruins the invariance property. It can easily be retraced in Section I
that such risk sensitive decisions are also a reparameterisation of
the certainty equivalent decisions in (2). In principle each new
preference structure B might call for a different specification of A to
keep the instrument changes within acceptable bounds. In Brandsma,
Hughes Hallett and Van der Windt (1983) it was suggested that
keeping the ratio between the traces of B and A constant might
92

to be a convenient device for comparing the results of different

diagonal Qs. Implementing the scaling factor proposed at the end
of Section 1 means that such a measure is applied here to the
elements of B'P B.
A conceptual problem arises if A is introduced in the statistical
decisions to control the deviations from x d . We saw in Section 3
that the optimal ridge factors d[*, i = 1, ... , nT, themselves lead
to a penalty matrix A*. But, in general, the implied penalities will
not keep the policy adjustments within aceptable bounds since they
are minimising mean square errors given a free hand with the
instruments. In fact, most ridge factors proposed in regression
analysis (Vinod, 1978) get smaller if the degree of uncertainty is
smaller or if the quadratic sum of the benchmark decisions, ~*'~*,
is larger. Therefore we prefer to reformulate (18) in terms of the
eigenvalues and eigenvectors of R' BR + A, where A is the original
penalty matrix specified in (1). Several alternative ridge factors will
be considered and the results in terms of objective function evalu-
ations and mean square errors are compared to the mean-variance
decisions.

5. An Application to Macroeconomic Planning

The features of certainty equivalent, mean-variance and statisti-

cal decisions are investigated here for the planning example set out
in Brandsma, Hughes Hallet and Van der Windt (1983). The exer-
cise takes 7 targets (B = balance of trade, X = production,
P = prices, S = labour's income share, U = unemployment,
D = budget deficit, R = real wages) and 5 instruments
(G = government expenditure, I = investment subsidies,
W = incomes policy, M = money supply, T = direct taxes). Esti-
mates for 'P and s* were obtained by averaging the corresponding
quantities constructed from the structural residuals and their
numerical evaluation is the same for all risk sensitive decisions.
Table 1 presents the dynamically reoptimised decisions over the
planning period 1976-1980. That means the policy formulated in
the first year (t = 1) is based on the information made available up
to that year. For t ~ 2 the information set at is updated for the
optimal decisions and the realisations of the past year(s), implying
also that the expectations Et(s) for the current and future periods
 93

Table I. Risk sensitive policies from economic and statistical theory.

1976 1977 1978 1979 1980 Total change
Instruments:"
G -0.82 -1.10 0.10 0.62 -0.52 -1.7
I 0.43 0.22 -0.54 -1.22 -1.63 -2.7
W -5.26 -5.38 -1.11 0.94 -0.27 -11.1
M 0.18 -0.03 -0.82 -0.80 0.00 -1.5
T 0.49 0.38 -0.32 -0.72 0.13 -0.0
Targets: b
B 0.33 1.56 2.55 1.91 0.18
X 0.71 0.99 1.85 3.00 3.45
p -3.15 -4.64 -2.67 -0.31 -0.09
S -2.46 -2.85 -1.57 -0.98 -0.87
U -1.00 -2.73 -4.81 -6.54 -7.72
D -0.52 -1.27 -2.20 -3.44 -5.00
R -3.95 -5.65 - 3.61 -1.18 -0.36
Risk averse decisions
G -0.46 -0.84 0.10 -0.12 -1.27 -2.6
I 0.34 -0.35 -0.53 -1.74 -2.46 -4.7
W -5.74 -5.35 -2.22 2.63 1.72 -9.0
M 0.03 0.26 - 1.41 -1.00 0.01 -2.1
T 0.22 0.34 -0.62 -0.24 0.69 0.4
B 0.32 1.68 2.70 2.23 0.12
X 0.81 0.91 2.16 2.72 3.45
p - 3.42 -4.75 - 3.35 0.39 1.44
S -2.72 -2.83 -2.21 -0.16 0.02
U -1.10 -2.84 -5.27 -6.76 -7.59
D -0.50 -1.20 -2.26 -3.52 - 5.19
R -4.34 -5.77 -4.59 -0.20 1.41
Optimal ridge decisions
G -0.61 -0.71 -0.17 0.21 -0.03 -1.3
I 0.41 0.28 -0.03 -0.24 -0.05 0.4
W -6.40 -5.70 -3.69 -1.10 -0.14 -17.0
M 0.08 0.06 -0.16 -0.01 0.00 -0.0
T 0.19 0.27 -0.16 -0.22 0.00 0.1
B 0.36 1.79 3.00 2.69 0.91
X 0.91 1.08 2.53 3.94 4.95
P -3.82 - 5.12 -4.40 -2.23 -0.84
S -3.03 -3.07 -2.93 -2.28 -1.40
U -1.23 -3.17 -6.05 -8.75 - 11.01
D -0.57 -1.32 -2.64 -4.44 -6.62
R -4.84 -6.22 -5.87 -3.61 -1.39
"policy adjustments, variables measured in percentage changes.
bdeviations from the central projection in percentage changes except Band D
(% national income) and U (% dependent labour force).
94

are revised. The optimisation technique applied in this paper does

not produce decisions in the form of an explicit dynamic (feedback)
rule, like in Chow (1975). Instead, an updating procedure ensures
that the certainty equivalent decisions remain optimal for each t.
The lack of a feedback decision rule must be counted against the
fact that optimisation techniques which do produce such rules
would generate suboptimal decisions in the face of risk sensitivity,
because it implies stochastic inseparability of the objectives over
time. Given that the approach adopted in this paper can handle
such situations, the updating procedure is greatly facilitated by
recognising that all decision rules considered follow from a repar-
ameterisation of the original preference structure. The instruments
in Table I are discretionary adjustments to the implemented poli-
cies in those years, the expected target outcomes represent devi-
ations from a central projection which is generated by the model
each year given the preassigned values for the instrument variables.
For a start, we took the preference matrices B and A equal to the
identity matrix. The certainty equivalent decisions are then given
by (2). Allowing for the amendments made in the beginning of this
section, the risk averse decisions follow from (3) and the "ridge"
decisions from (13) with (18) on the diagonal of D.
A deflationary scheme of wage moderation and retrenchment on
government expenditure characterises all policies. In this, they
follow the line of government's declared policy of the second half
of the 70s. Reducing wage rises has a favourable effect on most
indicators of economic performance in the Netherlands, but some
have argued that this feature is typically overemphasised by the use
of the Vintaf model in macroeconomic planning. Lower labour
costs induce export-led growth, and it takes rather a long time
before the investment demand created by the accelerated scrapping
of older vintages causes inflation to return to its old level. By using
the characteristic roots or R'R, ridge "regression" emphasises such
dominant features of a model. On average the downwards wage
adjustments are 1% larger in the statistical compared to the
economic decisions, while the other instruments are used more
sparingly. This indeed suggests that wages are a crucial variable in
the model. On the other hand, a comparison between the risk averse
and the certainty equivalent policies reveals that, in order to reduce
the variation of the performance index, it may be expedient to
prevent consumption demand from falling by raising the wage level
 95

again towards the end of the planning period. This feature of the
risk sensitive decisions derived from the economic theory of risk
bearing aims to balance the portfolio of consumption and invest-
ment shares of national income. Its intention is supported by the
relative tightening of monetary conditions and the abandonment of
direct investment stimulation by the government. The results con-
firm the observation of Section 4 that the economists' approach to
risk aversion would lead to target stabilisation by flexible instru-
ment settings, whereas the statistical approach opts for relatively
sticky instruments.

6. The Gains and Losses

We will now confront the different decisions with their objectives

and investigate some other features with the help of additional
criteria. Column (1) of Table 2 contains the expected values of the
objective function w, that is the criterion minimised at each
t = 1, ... , T by the certainty equivalent (CE) decisions given the
realisations of former years. So, at t = 1, any other policy must
have a higher expected value. Since the decisions are conditional on
the instrument values already implemented within their own strategy,
other policies than the CE one might enabe the policymakers to
procure a better expected position for t ~ 2, but this occurs in none
of the risk sensitive schemes considered here (RA represents the risk
averse decisions (3); RC the combination by (6) and (7) with IX = t;
and RR the optimal ridge decisions of (13) and (18)). Also, the
corresponding development of ex post figures (column 2), in which
each year's expectation is replaced by its actual value, does not
support the idea that sequentially updated risk sensitive decisions
might produce a better outcome than their CE counterpart.
Although the RA, RC and RR entries for t = 1 point to a potential
improvement of w - E\ (w), this promise is never fulfilled in the
end. Curiously, the RR strategy at the start does even better than
the mean-variance decisions but it finishes much worse.
Of course, this comparison is not completely fair with regard to
the statistical decisions, which are trying to reduce mean square
errors. Since the ex post optimal decisions XO are unknown, they
take the CE estimates as a benchmark and use the ridge factors to
minimise the expected loss due to uncertainty; column (4) contains
96

Table 2. Objective function and mean square error evaluations.

(1) (2) (3) (4)
CE
I 77.5 66.1 34.1 0.0
2 100.8 101.9 37.2 3.2
3 102.7 104.4 44.2 5.4
4 135.6 135.8 52.4 15.3
5 153.6 126.8 53.3 18.8
RA
I 82.3 72.0 43.6 1.7
2 110.8 112.6 49.0 9.2
3 111.6 115.2 60.1 7.7
4 158.4 161.2 79.9 25.1
5 185.2 155.1 82.8 36.2
RC
79.8 69.2 40.4 0.8
2 105.2 106.9 44.6 6.4
3 107.7 110.5 54.0 6.2
4 149.2 151.0 68.2 19.9
5 171.6 142.6 70.0 27.3
RR
I 81.5 72.7 40.7 1.0
2 105.2 110.1 44.1 4.5
3 136.0 143.4 45.4 4.8
4 218.8 219.5 46.4 7.8
5 244.9 211.4 46.5 9.4
(I) cost to go; (2) cost after having gone; (3) initial cost of instrument changes, plus
subsequent revisions; and (4) expected mean square errors.

the values of (x* - x°),(x* - xo). The revised decisions at each t

are actually new and probably better estimates for the remaining
part of the planning period. Therefore one might suppose that the
sum of squared revisions is reduced by the ridge decisions and this
is indeed confirmed in column (3) of Table 2. It displays the sum of
the squared deviations x*' x* for t = 1 plus the squared revisions
of the original plan in each sequential year. The figures show that
the costs of using instrument values which deviate from desired
values are higher for the RR than for the CE policy but the
revisions cost much less. Contrarily, the mean-variance decisions
which combine RA and CE decisions under the heading RC do
worse in both respects. A comparison between the RC and RR
 97

Table 3. Open loop and ex post optimal decisions.

1976 1977 1978 1979 1980 Total change
CE
G -0.82 -0.52 -0.24 -0.49 0.\3 -1.9
I 0.43 0.32 0.43 -0.16 0.29 1.3
W -5.26 -3.60 -2.99 -2.41 - 3.17 -17.4
M 0.18 0.30 -0.\3 0.22 -0.01 0.6
T 0.49 0.37 0.19 0.42 -0.30 1.2
RA
G -0.46 -0.38 0.23 -0.78 0.20 -1.2
I 0.34 0.03 0.76 -0.50 0.36 1.0
W -5.74 -4.23 - 3.49 -1.92 -4.13 -19.5
M 0.03 0.51 -0.42 0.26 -0.01 0.4
T 0.22 0.22 -0.29 0.69 -0.43 0.4
RR
G -0.61 -0.49 -0.41 -0.31 -0.00 -1.8
I 0.41 0.20 0.27 0.07 0.23 1.2
W - 6.40 -3.82 -2.97 -2.73 -3.02 -18.9
M 0.08 0.17 0.04 0.16 -0.01 0.4
T 0.19 0.29 0.37 0.25 -0.22 0.9
Ex post:
G 0.34 0.39 0.15 -0.06 -0.09 0.7
I -0.35 -0.20 -0.29 -0.24 -0.10 -1.3
W -1.91 -1.57 -1.38 -0.78 -0.56 -6.2
M -0.21 -0.25 -0.23 -0.12 0.00 -0.8
T -0.44 -0.51 -0.37 -0.\3 -0.03 -1.5

figures suggests that their respective performances match quite well

in the first part of the planning period but diverge later on. This
may be explained by the fact that the ridge regression policy is
rather sticky in the use of the instruments, thereby reducing its
costs, while the mean-variance decisions are more flexible in order
to take advantage of the information about the forecasting errors
of the model.
We know that the (mechanical) projections of the Vintaf model
were actually on the pessimistic side. Since we also saw that the
economic decisions are cautious in the beginning, it is interesting to
investigate what the results of sticking to the first year policies
would be. For that purpose we will employ the ex post optimal
decisions which can now be calculated with the help of the historical
values over the period 1976--80. Table 3 shows the ex post optimal
decisions and the open-loop component of the different ex ante
98

Table 4. Comparative losses and ex post mean square errors.

Decisions E(w) wo Vo w t/I
RA 82.3 4.8 0.0 184.6 23.3
RR 81.5 4.0 6.2 180.9 21.3
RC (ex = t) 79.8 2.3 1.0 172.0 21.1
RC eu = t) 78.7 1.2 2.8 164.3 19.8
CE 77.5 0.0 11.0 146.4 17.2

policies considered in this paper, that is the sequence of decisions

based on the information available at t = 1. The ex post decisions
reveal that, given the preference structure, less vigorous policies
than those considered would have sufficed to reach the goals, had
one known the actual values of the random variables. It is possible
to attribute the differences between predictions and realisations to
the favourable effects of the actually implemented policy, but
in that case we have to suppose that there are additional (non-
modelled) channels through which the policymakers exert their
influence since we already accounted for the actual values of the
instruments used in this exercise. Within the set of 5 instruments
wage cuts are still dominating but, on average, they are more than
2% smaller in the ex post optimal decisions. But, apart from this
less severe wage moderation, the signs of the majority of the policy
adjustments is in the direction opposite to that implied by the
certainty equivalent or risk sensitive decisions. This is also reflected
in Table 4 which sets out the mean square errors and objective
function evaluations of these decisions with respect to the ex post
optimal decisions. Among them, the CE policy receives both the
smallest error and the lowest value of w, the objective function to
be minimised. Next are the combined policies from economic
theory; direct combination by 11 = t gives better results than the
optimal trade-off between losses at (X = t. The open-loop RR
decisions lead to a slightly lower ex post outcome than their RA
counterpart, which is worst of all, but it also is only second to last
in terms of mean square errors. This is remarkable since the latter
is the criterion to be minimised by ridge regression. Moreover, the
updating implied by the closed-loop policies does not alter this
conclusion in favour of the RR decisions. On the contrary, while
the CE, RA and RC policies are all improving upon their expected
outcomes at t = 5 (cf. Table 2), the ex post RR mean square
error gets even worse. The only proviso to be made is that no
 99

reestimations are involved, and so no updates of iO have been

considered.

7. Other Ridge Decision Rules

In the ridge regression literature, reviewed in Vinod (1978) and

Judge, Griffiths, Hill and Lee (1980), several suggestions have been
made to use ridge estimators which are most robust to data per-
turbations. The class of shrink estimators, in our case those that are
expected to improve upon certainty equivalence by shrinking the
policy adjustments towards zero, include single ridge estimators.
They compress the information about the distribution of the ran-
dom variables in s to an overall ridge factor d = des), which gives
the design matrix R'R + dI for the problem of reducing the loss
function E(x - x°),C(x - xO). Another member of this clas is the
Stein rule. It gives a shrinkage factor b+ which then determines
x = b+ _~*. We will consider three of such proposals:
(i) the Hoerl-Kennard-Baldwin estimator,

d _ nT
HKB - I (lIA,)({3; 1( 11 ) '
,
which is the harmonic mean of the optimal generalised ridge
factors d, in (26);
(ii) the Lawless-Wang estimator,

d _ nT
LW - I ({3; 18,J ,
,
in which the denominator of the harmonic mean is replaced by
1:, A, (1 Id,*); and
(iii) the James-Stein rule, where
b+ = max {O, 1 - a*(1 - R2)1 R2}
for the optimal a* = (nT - 2)j(mT - nT + 2) and R2 is
analogous to the well-known regression measure of the same
name, applied to (8).
It can easily be seen that AI • d LW !'( dHKB !'( AnT· d LW , where )'1
and AnT are, respectively, the smallest and the largest eigenvalues
100

Table 5. Other ridge strategies.

Decision rules w

Lawless-Wang (d = 1.58)
Hoerl-Kennard (d = 2.06)
123.8
109.6
'"
13.1
10.7
d = 4.00 75.0 5.5
Optimal ridge regression 66.5 4.8
James-Stein rule (<5+ = 0.40) 45.0 1.8
d = 14.00 43.7 l.l
d, = 60.d,* 43.7 1.0
Ex post optimal decisions 42.4 0.0

of R'BR + A. So A = I implies dHKB ~ d LW ' and the inequality

yields if R has rank nT. Therefore we have two strictly ascending
ridge factors, and it depends on the relative impact of uncertainties
in the variation coefficients (j,J {3; and on the eigenvalues A, what
their values are. Given the structure of our problem, both esti-
mators turn out to be larger than the pre specified value (d = 1.0).
Both the ex post values of wand the mean square errors are reduced
considerably compared to the risk sensitive policies calculated
before, and the outcomes are even better than those of the updated
certainty equivalent decisions. Much higher values of the ridge
factor would have been needed to come close to the ex post optimal
results. The optimal d in the design matrix R'R + dI is about 14.0.
That in R' BR + dA * is closer to 60.0, which suggests that, by
construction of the A * matrix the RR decisions are less sensitive to
multiplications of the instrument penalties, all by the same factor.
Without needing such manipulations, the James-Stein rule does
very well on its own, making it an attractive device to be applied to
this kind of exercises. An interesting feature of that rule is that the
shrinkage factor uses a measure for the expected effectiveness of the
benchmark policy.

8. Conclusion

The two approaches of deriving decisions which aim to reduce

the impact of uncertainty in certainty equivalent policies have quite
distinct characteristics. The results show that the economists'
approach is to hold high risk variables closer to their ideal values
so that given shocks cause smaller deviations from those paths and
 101

the deviations are mutually compensating for the impacts of ran-

dom shocks, whereas its statistical counterpart is to pick robust
rules designed to produce decisions close to the ex post optimal
path. Implementing the rules derived from economic theory will
give stable target variables but a flexible use of the instruments,
while the statisticians' approach has more flexible targets and rela-
tively sticky instruments for the same problem.
The empirical results show that it is very important to be able to
update the decisions sequentially. The attempt to account for the
distribution of stochastic shocks by incorporating the moments of
past forecasting errors in the computation of (open loop) risk
sensitive policies did not lead to smaller ex post values of the
objective function.
COST OF USiNG INSTRUMENTS
100,- -
i 1
90":

RA

- - + - - - - - - t CE

--.-----~~~----~----~~

20-;

10 ..;

o +------·---r-----.- ----._._-,..--------,------.. -----~

2 3 4 5
Year
Figure I.
102

COST TO GO

1tKl-

160 ••

120 -

60 _.
,

20-:

o+----------- -~r-------~--,-----~-r--------__j
2 3 4 5
Year

Figure 2_

while allowing for revisions of risk averse decisions revealed the

need for sequential updating in exaggerated form. On the other
hand, the policies calculated by analogy to ridge regression stick to
their original design and this results in substantial losses in terms of
the objective function.
Figures 1 and 2 illustrate these points by displaying, on the one
hand, the initial costs of using the instruments at values different
from the benchmark path, augmented by further revisions in later
years, and, on the other, the development of the costs which have
been realised, including those which are expected for the remaining
part of the planning period. The risk averse policy clearly incurs the
 103

largest costs of adjustment (Figure I), but they do bring the ex post
outcome closer to the certainty equivalence one in comparison to
the ridge regression results (Figure 2).
Of course, both minimum variance and mean square error
decisions rely on the estimation of moments of the stochastic distri-
bution and could be improved if more accurate estimates were
available. The ridge analogy moreover faces the fundamental dif-
ficulty of assessing the true (ex post optimal) values of the instru-
ments. For that reason, a great number of ridge factors have been
proposed in the regression literature, which try to improve the
stability of the estimators. Some of these have been applied in this
paper, and they show there is scope for improvement by biasing the
instruments towards their desired values. Given that it is possible to
construct policies with smaller mean square errors than the cer-
tainty equivalent decisions, the important question remains whether
we want decision rules which are such that the overall performance
of the economy over time is expected to be robust against stochas-
tic shocks, or whether the policies themselves should be more stable
in the sense that they do not have to be changed each time the
information set is updated.

References

Arrow, K. J.: Essays in the Theory of Risk Bearing. Amsterdam: North Holland,
1970.
Brainard. W.: "Uncertainty and the Effectiveness of Policy", American Economic
Revle~\' (Papers and Proceedings) 57, 411-425.
Brandsma, A. S. and A. J. Hughes Hallett: "Dynamic Risk Sensitive Optimiz-
ation and von Neumann-Morgenstern Decision Theory", Discussion paper
8401/G, Erasmus University, Rotterdam, 1984.
Brandsma, A. S., A. J. Hughes Hallett, and N. van der Windt: "Optimal Con-
trol of Large Nonlinear Models: An Efficient Method of Policy Search Applied
to the Dutch Economy", Journal of Policy Modeling 5 (1983),253-270.
Brandsma, A. S., A. J. Hughes Hallett, and N. van der Windt: "Optimal
Economic Policies and Uncertainty: The Case against Policy Selection by Non-
linear Programming", Computers and Operations Research 11 (1984), 179-197.
Chow, C. G.: Analysis and Control of Dynamic Economic Systems. New York:
John Wiley & Sons, 1975.
Hughes Hallett, A. J.: "On Alternative Methods of Generating Risk Sensitive
Decision Rules", Economics Letters 16 (l984a), 37-44.
104

Hughes Hallett, A. J.: "The Use of Ridge Regression Techniques for Generating
Risk Sensitive Decision Rules", Communications in Statistics 13 (1984b),
127-138.
Hughes Hallett A. J. and H. J. B. Rees: Quantitative Economic Policies and Inter-
active Planning. Cambridge: Cambridge University Press, 1983.
Johansen, L.: "Parametric Certainty Equivalent Procedures in Decision Making
under Uncertainty", Zeitschriftfor Nationalokonomie 40 (1980),257-279.
Johansen, L.: LeCtures on Macroeconomic Planning, Part 2. Amsterdam: North
Holland, 1978.
Judge, G. J., W. E. Griffiths, R. C. Hill, and T. C. Lee: The Theory and Practice
of Econometrics. New York: John Wiley & Sons, 1980.
Malinvaud, E.: Lectures on Microeconomic Theory. Amsterdam: North Holland,
1972.
Newbery, D. M. G. and J. E. Stiglitz: The Theory of Commodity Price Stabiliz-
ation. Oxford: Oxford University Press, 1981.
Theil, H.: Optimal Decision Rules for Government and Industry. Amsterdam:
North Holland, 1964.
Theobald, C. M.: "Generalizations of Mean Square Error Applied to Ridge
Regression", Journal of the Royal Statistical Society, Series B, 36 (1974),
103-106.
Vinod, H. D.: "A Survey of Ridge Regression and Related Techniques for
Improvements over Ordinary Least Squares", Review of Economics and Statis-
tics 60 (1978), 121-13\.
Waud, R. M.: "Asymmetric Policymaker Utility Functions and Optimal Policy
under Uncertainty," Econometrica 44 (1976), 53-66.
Whittle, P.: Optimization over Time, Volume \. New York: John Wiley & Sons,
1982.
 105

CHAPTER 7

OPTIMAL ECONOMIC POLICIES UNDER A

CRAWLING-PEG EXCHANGE-RATE SYSTEM
AN EMPIRICAL APPROACH
Hans M. Amman and H. Jager
University of Amsterdam, The Netherlands

1. Introduction

During the last 13 years the world economy was faced with floating
exchange rates of the major currencies. After a short initial period
where enthousiasm prevailed, negative effects of floating exchange
rate:>, like short-term overshooting and long-term deviations of
rates from their equilibrium levels, became visible and, as a conse-
quence, doubts about the choice of favor of floating arose. It is,
therefore, not astonishing that a renewed interest in the positive
effects of stable exchange rates can be perceived. In theoretical
studies this recent shift was witnessed by an intensified attention to
mixtures of fixed and flexible exchange rates.!
Such combinations of flexible and fixed exchange rates can be
roughly divided into three variants. These variants are charac-
terized, successively, by a crawling peg, the use of a reference rate
and an asymmetrical peg-which may be allowed to crawl. In case
of a crawling peg, the parity-provided with a fluctuation margin-
is periodically changed a little when necessary. In the meantime the
exchange rate's value has to stay within this fluctuation margin by
means of official interventions in the foreign-exchange market. To
what extent the peg or parity will be changed, depends on either a
formula chosen for this purpose or on just a policy decision. 2,3 On
the other hand, in a system based on a reference rate monetary
authorities have no commitment to intervene in the exchange mar-
ket. By contrast, they are restricted in conducting foreign-exchange
market interventions: the only policy rule is that in the event the
exchange rate is outside the fluctuation margin around the parity,

C. Carraro and D. Sartore (eds.) Developments of Control Theory for Economic Analysis
©1987 Martinus NijhofJ Publishers (Kluwer), Dordrecht
106

interventions are only allowed when they are addressed to a

return of the exchange rate within the margin. 4 Further, the third
variant-the asymmetric parity-is characterized by a one-sided
fluctuation margin. It is obvious that this system can be unified with
a pure crawling-peg system. 5
Naturally, the question arises which variant is to be preferred.
The answer to this question is thwarted by the diversity of con-
tent that has been given to the exchange-rate variants distin-
guished. This is due to the choices which have to be made with
respect to: the way of changing the parity value; the intensity
and size which is allowed for this change; and the width of the
fluctuation margin. These degrees of freedom in applying the
exchange-rate variants make that we have to restrict ourselves in
our purpose.
Our analysis is only directed to exchange-rate systems based on
a crawling-peg variant. Optimal control techniques will be utilized
to look for the preferred or optimal rule of parity determination.
This method will be applied to the Dutch economy in the period
1970-1976. The crawling-peg variant has been chosen for the fol-
lowing reasons. First, of the variants distinguished here, both in
economic theory and policy-making most attention has been paid
to the crawling peg. Second, for this variant a variety of competing
formulae for parity adjustment has been developed; optimal con-
trol techniques lend themselves admirably to weigh one formula
against the others. Third, the intervention obligations of the
crawling peg-namely interventions in order to keep the exchange
rate within its fluctuation margin-fit in well with the design of
optimal control techniques.
The analysis will be elaborated as follows. In the next section
the arguments in favor of the use of control techniques are given.
Section 3 presents an outline of both an econometric model and
a welfare-loss function of the Dutch economy. Next, in Section 4,
the required extension of the optimal control algorithm is for-
mulated, while Section 5 contains the results that arise from
the application of the optimal control method to the model and loss
function derived in the preceding sections. Finally, Section 6 gives
the main conclusions.
 107

2. Choice of the Method

One way to determine the most favorable degree of exchange-

rate management, which seems to be the most obvious one, is
to compare the outcome of the economic process in periods
with the different exchange-rate arrangements being examined
in force. Such an approach, however, has the disadvantage that
some of the forms under discussion were never put into practice.
It will, therefore, be virtually impossible to find a single country
which applied all the exchange-rate systems concerned in the
past. This, however, is the least which is needed in order to avoid the
situation where distortions which are caused by inter-country
differences in economic structure, thwart the effects of the exchange-
rate systems.
Even if the analysis is confined to just one country, however, the
danger of distorting influences remains considerable. The mere
design of economic institutions, for example, habitually presents
changes in time in many more aspects than only the exchange-rate
regime. Furthermore, it is wrong to decide in favor of one of the
regimes on the basis of events that only occurred during the period in
which the regime was employed. One must, for example keep in
mind that the problems which were created under the Bretton-
Woods system of fixed exchange rates and led to its abandonment
in 1973, still burdened the period after that year. For this reason a
comparison in time will lead too easily to completely ascribing
economic developments to the functioning of the exchange-rate
system of the time. This danger is continuously present in empirical
investigations of the effects of exchange-rate regimes, like the
studies contained in Williamson (1981). They likewise examine
which echange-rate system is to be preferred. It must be emphasized
that, in general, the authors of the latter studies were conscious of
this flaw in their analyses.
The remaining ways to gain insight into the desired exchange-
rate system employ an analytical or empirical economic model.
Their drawback is that the economic structure may change when
an alternative exchange rate system is put into practice. Nor-
mally, information is absent about the direction and extent to
which the structure will be influenced by this alteration. In terms of
an econometric model this means that the influence on the para-
meter values and the specification of the economic relationships
108

is lacking. 6 This influence can, however, be of significance, espec-

ially, that regarding the structure of exchange-rate expectations. 7
In a second, often employed, approach to the desired exchange-
rate system an economic objective is pursued within the confines of
an analytical economic model while simultaneously introducing a
disturbance to the economy. The consequences of various exchange-
rate regimes for the objective are then studied. Frequently used
objectives in this approach are the maximizing of national income
and of national consumption. s The necessity to arrive at an
analytical solution, however, imposes rather drastic restrictions on
the size of the economic model. Consequently, the results of such
exercises are conditional upon the rigorous assumptions used
regarding the economic structure and economic objectives.
The objections against the analytical approach can be overcome
by choosing one of the two approaches to the desired exchange-rate
system which will be described below. One of them is the use of
simulation studies. It is then possible to employ an econometric
model which, in principle, is unrestricted in size and, therefore,
much more differentiated compared with an analytical model. This
offers scope for a substantially more realistic outcome. Since the
relevance of an econometric model is limited to both the time period
and country for which it has been estimated, this improvement is at
the cost of the general validity of the results. An additional draw-
back, which a simulation study usually has in common with the
analytical approach, is that the functioning of the exchange-rate
systems is studied in isolation from other macroeconomic policy.
This latter is, in other words, exogenously determined, whereas it is
likely that, in practice, a change in the nature of the exchange-rate
regime is coupled with changes in the actual time paths of instru-
ments of economic policy. This lack of concomitant adjustment of
the rest of macroeconomic policy is sometimes, only partly, elimi-
nated by introducing a policy rule for one of the policy instruments.
This restriction connected with the method in question is percep-
tible in, for example, the extensive simulation study in Kenen
(1975).
The required policy interdependence finds expression in the
approach which uses optimal control techniques. It allows one to
determine the optimal time paths of the different economic instru-
ments simultaneously. This important advantage in comparison
with other methods is the main reason why it has been chosen for
 109

the present study. Nevertheless, the application of optimal control

techniques has its restrictions too. When the approach uses an
econometric model-as is done here-it has the drawback in com-
mon with simulation studies that the relevance of the outcome is
restricted to both the period and country for which the model has
been estimated.
An obvious handicap that is specific for the application of opti-
mal control techniques is the necessity to use a quantified macro-
economic objective function. In implementing the method this
function is maximized-or minimized in the case of a loss function,
as in the present study-under the constraints of the econometric
model. Quantification of a macroeconomic objective function
requires information about policy makers' preferences and target
values. It appears to be difficult to gain insight into these desir-
abilities, because they are often of an implicit nature. Once con-
structed, however, such an objective function has the advantage
that now the simultaneous realization of a number of economic
objectives can be pursued and controlled. This is an important step
towards greater realism of the results of theoretical studies.

3. An Outline of the Decision Model

In this study optimal control techniques are applied to an econo-

metric model of the Netherlands' economy combined with an objec-
tive function of the Netherlands' policy makers. The model is linear
and has been estimated using quarterly data for the period
1967(I)-1976(IV) by means of two-stage least squares. The model,
in principle Keynesian in nature, stresses the international economic
relations of the Netherlands' economy. Since in the international
trade component of an economic model, in particular, the adjust-
ment lags are apt to be long, the model is very dynamic. The model
includes a Phillips curve, price indexation of wages and its feedback
to prices, exogenously determined production, and budget con-
straints for the government and the money market.
The part of the model consisting of the international economic
relations contains behavioral equations for the value as well as
the price of both imports and exports of goods and services. The
balance-of-payments identity determines the exchange rate. Apart
from this feature of the so-called balance-of-payments approach to
110

the exchange rate, the model possesses elements of portfolio-

balance theory through the specification of a behavioral equation
for the international capital flows. Because of its relevance for the
explanation of the exchange rate, the model's international econ-
omic relations will be outlined here. 9
Three explanatory variables appear to determine the speculative
part of private capital flows. First, the difference between the actual
exchange rate and its par value makes a statistically significant
contribution to the explanation of the Netherlands' speculative
capital inflows. Since the effective exchange rate is used in the
model, a positive difference between actual and par value of this
rate expresses a strong guilder. Theoretically, this would have to
stimulate the Netherlands' capital imports by means of its specula-
tion component. This appears to be in agreement with the facts. The
empirical outcome also shows that speculative capital inflows
depend on, second, the trend in the exchange rate and, third, the
tendency in official foreign-exchange market interventions. Since a
rise in the exchange rate in the preceding two quarters contributes
positively to capital inflows, speculators, apparently, have extra-
polative exchange-rate expectations of the elastic type. These
inflows are also promoted when the monetary authorities purchase
foreign exchange and, thus apparently pursue a stronger guilder.
The expectations scheme for the exchange rate that is found here,
supports the so-called bandwagon behavior in the foreign-exchange
market. The practical relevance of this behavior has also been
defended by, amongst others, Mayer (1982, pp. 24 and 41) and the
Group of Thirty (1982, p. 14). This viewpoint is prompted by the
consideration that exchange-rate expectations tend to be loosely
held, especially over a period of months or more because of the
relatively large white noise. For that reason, the expectations are to
a great extent influenced by current events and speculators tend to
take a short-term view, concentrating on making short-term gains
(see Group of Thirty (1982, p. 14)). Another plea in favor of
bandwagon behavior is that, once the exchange rate starts to move,
there are influences like the J-curve effect and the repercussions on
domestic inflation that give rise to a self-reinforcing tendency in the
exchange-rate movement.
The decision model includes seven objective variables, viz. the
growth of real gross national product (GNP), the stability of this
growth, the rate of inflation, the balance on current account, the
 111

exchange-rate level (in relation to the currency's parity), and the

stability of two terms of trade (or real exchange rates). One of these
terms of trade is defined as the Netherlands' export price level
related to that of competitors on foreign markets (or competitive-
ness on the export side), and the other as the price deflator of the
Netherlands' GNP related to the Netherlands' import price level (or
competitiveness on the import side). The econometric model lacks
a sub-model of the labor market, so that full employment could not
be included as an objective variable. This gap is partly filled by the
presence of the growth of GNP and the close ties which are nor-
mally supposed to exist between the two variables. For the design
of the loss function, the desired time paths of the objective variables
are needed. These paths were derived from official publications.
The same holds for the desired time paths of the instrumental
variables, or control variables, in the mode1. lo
The decision model contains six of these instruments of economic
policy. They are: the short-run interest rate of the Netherlands
Bank (the rate of advances), the government's nominal tax receipts,
the government's net borrowing, the government's liquidity short-
age, and official interventions in the foreign-exchange market.
Moreover, the model allows the foreign assets obtained through
these interventions to be converted to such long-term foreign assets
that they lose the high liquidity which is necessary for reserve assets.
The advantage of the existence of such a possibility is that the
opportunity costs of monetary reserves can be partly circumvented
through the higher rates of return of longer-term investments. II In
the analysis presented in the next section the exchange-rate arrange-
ments may differ in the design of both the time path of the parity
and the size of the fluctuation margin around the parity. For this
reason they also obtain the character of instruments.
The macroeconomic objective function which is employed is,
as already mentioned, a loss function for the Netherlands' policy
makers. The weights of this, quadratic, function have been derived
by means of a method based on revealed preference. This method
was, in essence, borrowed from Friedlaender (1973). With the aid
of the econometric model, the weights of the objective variables
have been obtained from the Netherlands' economic policy as it was
actually conducted in the period 1967-1970. 12 As the Netherlands'
economy showed a stable and, therefore, relatively predictable
development in that period, it has been assumed that this policy is
112

a good reflection of the policy preferred by the policy makers.

Likewise, the weights of the instrumental variables are needed in the
loss function. Apart from the information obtained from the
method of revealed preference, the weights of various instruments
and of the exchange rate have also been influenced by the necessity
to keep their optimal time paths within politically acceptable limits
and within the fluctuation margin around the parity, respectively,
in the application of optimal control to the decision model.
The optimal policy for the decision model can be computed
by means of the algorithm developed by Pindyck (1973). This
algorithm, however, does not allow desired or target values for the
objective and instrumental variables with an endogenous character.
An exchange-rate regime characterized by a crawling peg, however,
has an endogenous parity. This means that the parity in a certain
period is determined by the optimal time paths of the model's state
variables until then. The extension of the algorithm, required for
including endogenous target values, is the subject of the next
section.

4. Endogenizing the Target Values

Generally, economic policy problems can be expressed in the

form of a system of simultaneous linear difference equations. For
this reason, in applications of optimal control techniques, a linear
model is usually used-supplemented with a quadratic loss func-
tion. The present analysis forms no exception to this rule. The
linear-quadratic control model can be written in the form:

min: W = ~ ttl {(Yt - Yt)'Qt(Yt - yJ

(1)
under the constraint
Yt = A? Yt + A: Yt-l + BtX t _ 1 + CtZ t - 1, Yo gIven. (2)
In the above formulation (cf. Pindyck (1973, pp. 27 and 92)) three
types of variables are distinguished, viz. state variables YI' control
variables Xt and exogenous variables Zt. The character is used as I

a transposed sign. The vectors Yo X t and Zt have dimensions n, r

 113

and s, respectively. In addition, W is the loss and Yt and xt exo-

genous target values. The elements of the diagonal matrices Qt and
R, contain the welfare loss caused by a unit difference between the
actual and target value of, successively, the state and control vari-
ables. With the algorithms available for this problem, it is possible
to derive the optimal values of the control variables for each period.
Stoppler and Stein (1982) incorporated endogenous target revis-
ion by adding to the decision model the equation:

Yt = MtYt-1 + (/ - M t ) Yt (3)

where / is the unity matrix and M t a diagonal matrix. By means of

the matrix M t the target values are endogenized. They become a
function of the state vector lagged one period. The vector Y is the
exogenous part of the target value.
Unfortunately, the formulation of Stoppler and Stein has a
considerable shortcoming. 13 Since the matrix M t has to be a diag-
onal matrix, equation (3) has the disadvantage that only the state
variable concerned and, moreover, only lagged one period can be
of influence on its own target value. Furthermore, the introduction
of equation (3) will make it necessary to write new software.
The procedure which will be described now, preludes these draw-
backs. To endogenize the ith target value, that is Y"t' the state vector
is expanded with one element:

Yn+ I,t = Y"t - Y"t in which is substituted for Y"t: (4)

n

Y"t = L
j= I
a~+I,jYj,t + a~+I,n+IYn+l,t+1

+ L bn+l,jxj,t_1
j=1

s
L
+ j=l cn+l,jZj,t_1 + Cn+l,s+IZs+l,l-l' (5)

In the above formulation we have used the (n + l)th rows of the

A?, A~, BI and C t matrices to include the marginal impacts of the
state, control and exogenous variables. (For convenience, in the
114

elements of these matrices time indices were deleted.) It is even

possible that the target value is influenced by itself, since Yn+1 is
also included in the right-hand side of equation (5). The vector ZI
is expanded with one element to incorporate the target's exo-
genous part, which is not contained in the original model of
equations.
In fact, the state variable Yn+ I,t is now able to take over the role
of the objective variable Y"t in the loss function and is built up of
Y"I itself, but now already confronted with its endogenous target
value. For that reason, the next step is to move, in equation (1), the
value of the diagonal element q", of the matrix Qt to the element
qn+ I,n+ I' It is evident that in the loss function the element q", as
well as Yn+l.t have to be set equal to zero. In this way, the term
q",(Y"t - Y"J 2 has been replaced by qn+I,n+IY~+I'/' Consequently,
all possible flexibility is obtained to endogenize one-or more-of
the target values. Finally, we must give the element Yn+ 1,0 a starting
value as required by the algorithm.
In fact, the above procedure consists of a generalization and exten-
sion of Theil's carry-over-amendment (cf. Theil (1964, p. 265».
The procedure stays within the formulation of the decision model
so the necessary conditions for an optimum, in the form of a
minimum of the loss function, are not affected. In order to prove
that this is indeed the case, we will reformulate the control model
as follows:
(6)
-
Yt YI - DOI YI - DII Yt-I - F IOYI
- - FIt Yt-I
-

(7)
Vectors with a indicate the extended part of the original vectors.
Thus, the vector Yt is the vector of endogenized target variables.
Then, the control problem becomes:

min: W = 1 ttl {(VI - vJ'QJv I - VI)

(8)

under the constraint

(9)
 115

with

vt =
[;:l Vt
~ [~J Qt' -- [Q, Qt~J 0 Pt [;:J
[ A't
°"J _OF,,]
nOt [ A" nI -
I _tD? -Ft t - -D)

n 2t =
[~~J .,~[C,
t -Kt
_OJ
Further, it can be noted that when all target variables are endogen-
ized the vector Yt is equal to the null vector and the matrix Qt to the
null matrix. By making use of equations (8) and (9) we can con-
struct the augmented Lagrange function (cf. Chow (1975, p. 149»:

L(Vt, XI' AJ = l tto {(Vt - vt)'{2t(vt - Vt)

T
+ (x t - xt),Rt(x t - Xt)} - L A;(v t - n?v t
t=O

Differentiation with respect to its variables leads to:

(10)

(11)

oL (12)
OAt
By setting equations (10), (11) and (12) equal to zero and imposing
the transversality condition AT + I = 0, we can derive the optimal
values of the control variables, which we will not pursue here.
The second-order conditions can be checked easily. The bordered
Hessian matrix HI is:
116

Sufficient conditions for a minimum of the function L(.) are that the
following principal minors of the determinant H; are negative
(Chiang, 1974). This leads to the condition:

By means of Laplace expansion (Hadley, 1974), eventually this

implies the condition:
Vt {11t~ - 112 > 0 /\ 11t~ - 1121Rti > O}
--. Vt {IA~ - 11211 + FtOl 21R t i > O}
We can conclude that the matrices (A~ - I), (l + F;0) and R t must
be non-singular, which holds evidently by definition of the control
model and by the fact that the matrix R t is positive definite
(Pindyck, 1973, p. 27) as well. So the necessary conditions for a
minimum are satisfied.

5. Optimal Parity Rule and Economic Policy

The decision model which was described in Section 3 and the

extended algorithm developed in the preceding section, will now be
utilized to determine the optimal parity rule as well as the optimal
economic policy under an exchange-rate system with a crawling
peg. As mentioned earlier, there is a variety of possibilities to model
a crawling peg. As a starting point we can use the proposals put
forward at an earlier stage. In a survey on the crawling peg,
Williamson (1981a) distinguishes four formula variants for deter-
mining the size and direction of the crawl of the exchange rate's par
value of parity. They are, successively, based on: the change in the
actual exchange rate, the level of official interventions in the
foreign-exchange market, the difference between the actual and
desired level of monetary reserves, and, finally, the change in the
 117

relative purchasing power of the currency in question. Of course,

the parity adjustment does not need to take place instantaneously,
but may occur with a certain time lag.
The version of a crawling-peg system in which adjustment is
based on the difference between the actual and desired level of
monetary reserves, will be left out of consideration because there is
hardly any information available about the desired holdings of
reserves. The versions which contain the actual exchange rate and
official interventions in the foreign-exchange market are built on
the behavior of the participants in the foreign-exchange market. By
choosing the version based on the purchasing power of a currency,
attention is primarily directed towards longer-term tendencies of
the actual exchange rate.
When the parity change is determined by recent changes in the
actual exchange rate, the underlying motive is that the parity must
gradually be forced to follow the market rate. A longer-term resist-
ance to market tendencies is considered to be untenable. The same
holds for the official-interventions version: developments in the
foreign-exchange market can only partly be counteracted-and
even then only temporarily. As a consequence, a complementary
action in the form of a somewhat higher price in the exchange
market of the currency which has been sold by the monetary author-
ities, is still required. This implies a higher parity of the currency
concerned. The behavior of the monetary authorities with respect
to the parity which appears from each of these three parity-rule
versions will give support to the pattern of expectations of specula-
tors that was discovered in the empirical examination, in Section 3,
of international capital flows in the Netherlands. It is, therefore,
plausible to suppose that the pattern in question will only show
slight changes, if any, when one of these crawling-peg versions
would be put into practice.
Optimal policy will not only be determined for crawling-peg
versions, but also for a system with a parity that corresponds to
the actual time path of the parity of the Dutch guilder during the
period under study. In fact, this version boils down to the system
of the adjustable peg, but with the proviso, however, that the
parity's rate of adjustment is historically determined and thus
exogenous. The optimal policy under this system will be called
version 1.0 and will be set as a standard for the optimal policies
which are determined for the crawling-peg versions. In all versions
118

the bounds of the fluctuation margin are set at - 2.5 and + 2.5 per
cent of the par value. This bears resemblance to the fluctuation
margin of the exchange-rate arrangements of, successively, Bretton
Woods (until March 1973), the so-called Snake arrangement
(between April 1971 and March 1979) and the European Monetary
System (after March 1979).
The first crawling-peg version which we will introduce, version
2.0, is characterized by an endogenous parity that crawls according
to the relationship:
EI* - El*--I = E I - 1 - E I - 2 (version 2.0)
where E1' and EI are the exchange rate's parity and actual (or
optimal) level, respectively, in period t. Here the exchange rate is the
price of a Dutch guilder expressed in foreign currencies. In version
2.0 a change in the parity is induced by an identical change in the
actual exchange rate in the preceding period. This seems to be the
most natural exchange-rate rule. In order to emphasize that, in
practice, a change of the parity generally will be carried out by
taking into account longer-run tendencies, in the next experiment
the parity is adjusted on the basis of exchange-rate developments in
two foregoing periods. In the decision model this is identical to two
state variables. It is obvious that the most recent exchange-rate
change must dominate the effect of the exchange rate. This is
materialized by a weight for (EI _1 - E'_2) which is about twice as
large as that for (E'_2 - E. 3):
E,* - E'*--1 = 0.7(E'_1 - E'_2) + 0.3(Et _2 - Et - 3 )
(version 2.1)
Likewise, two formulae for the volume of official foreign-
exchange market interventions as a measure of parity changes are
introduced. Again one with the influence of the last two quarters:
Et* - El*--I = 0.7(RESI _ 1 - RESI _2)
+ 0.3(RESt _2 - RESt _2) (version 3.0)
and one with interventions in only the preceding quarter as the
determining variable:
(version 3.1)
where RES, stands for the level of monetary reserves at the end of
period t. As a consequence of the different dimensions of the
 119

exchange rate and the volume of interventions, the size of the

coefficients in these official-interventions versions is not self-
evident. Utilizing a trial-and-error process, the versions with the
sum of weights equal to one showed the lowest welfare losses.
A currency's purchasing power shows, by nature, only small
short-run fluctuations. In modeling the purchasing-power version
of the crawling peg, it is, therefore, plausible that only one version
comes into consideration, namely:

Et-2PXCt-2/PXt-2
(version 4.0)
where P XCt and P X t represent the price level of competitive exports
and the Dutch export-price level, respectively. Both price levels are
denominated in Dutch guilders.
For all the exchange-rate systems introduced in the preceding
part of this section, optimal policies have been determined. Table 1
presents for each of these policies the total costs-or loss-and,
moreover, the contributions of the time paths of the individual
objective variables to the total costs.
In the total costs reported, the de facto costs of the exchange rate
have been left out of consideration. This contributes to the achieve-
ment of a fair-minded appraisal of the systems, since it is not
relevant what exactly the time path of the exchange rate is, even if
a system has a fluctuation margin. The only purpose of the
exchange rate's weight in the loss function is to ensure that the
optimal time path of the rate stays within the fluctuation margin
around the parity. As long as this goal is pursued successfully, a real
loss in terms of economic costs is completely absent. Moreover, the
attainment of this goal can easily be judged by investigating
whether the optimal exchange rate lies outside this margin. When
this was the case, the exchange rate's weight was increased to such
a level that the optimal rate no longer crosses the margin. The costs
in economic terms of this higher weight will appear in the form of
a poorer realization of the desired time paths of the other means
and ends because of the greater emphasis on a small difference
between the optimal and the desired time paths of the exchange
rate.
In order to increase the comparability of the optimal policies,
they are subjected to some conditions pertaining to the outcome in
120

the last quarter of the period. First, it has been required that the
level of monetary reserves at the end of the period be roughly equal
to that at the beginning. Thus, it is made impossible for an
exchange-rate system to reduce total loss or costs by a net consump-
tion of monetary reserves, and in so doing burden the policy that
will be conducted in the next period. Despite this condition, the
stock of the country's international liquidity appears to grow under
all exchange-rate systems because a specified increase of the, offic-
ially held, longer-term foreign financial assets has been allowed. A
second requirement is that substantial divergences in the use of
sources for government expenditures over the entire period must be
prevented by assigning proper weights in the loss function to the
three instruments of financing government expenditures.
In Table 1 the disappointing results of the versions 3.0 and 3.1
are striking. The total costs of these versions, which employ official
interventions as the measure for par-value changes, are much
higher compared to the other exchange-rate arrangements pre-
sented in Table 1. Especially, bad records are observable with
respect to the level of real GNP and the rate of inflation. According
to Figure 1, the latter disadvantage may mainly be attributed
to the high degree of imported inflation due to the structural
depreciation of the guilder. In this respect the official-interventions
version distinguishes itself from the other versions in the diagram.
The effect suggested, is supported by a comparison of the costs
of inflation in Table 1 with the tendencies in the par values in
Figure 1. It emerges that the costs concerned rise as the trend in the
par value declines.
The remaining crawling-peg versions 2.0, 2.1 and 4.0 all lead to
lower total costs than the exchange-rate system that actually func-
tioned in the time period under examination, viz. version l.0. The
crawling-peg versions in question are based on an exchange-rate
and a purchasing-power formula. Compared to version 1.0, they
appear to ensure a high and stable growth of GNP - especially the
purchasing-power version, low inflation, and a stable terms of trade
for the export sector of the economy. A negative aspect of these
versions is the divergence between the realized and desired balances
of current international payments. In all three crawling-peg ver-
sions 2.0, 2.1 and 4.0 the guilder shows in Figure 1 a higher
appreciation when compared with version 1.0. As the relatively
strong appreciations are accompanied by relatively low total costs,
Table I. Varying parity rules and optimal-policy costs.
Exchange-rate Total costs* Contributions of the target variables
version
Real GNP Inflation Current Change terms of trade
account
Level Change Exports Imports
\.0 7549 5090 52 353 47 1 48 164
2.0 7507 5116 42 314 51 24t 144t
2.1 7370 1 5092 42 276 t 53 32 147
3.0 8234 5194 47 434 64 35 174
3.1 8246 5208 47 441 63 37 176
4.0 7403 5067 t 37t 278 52 27 194
* The total costs include both costs of the target variables and the costs of the instrumental variables.
1 Best record in the column.

tv
122

parity
150

115
exchange- rate version

liO
______ 1.0
2.0
2 1
3 0
/
135
_____ 4.0

/
130
/
125 /,"'/' /

,- -,---- ----- ---

120
/////
115

/
110

105

11 16 21 28
time period

Figure I. Exchange-rate parities.

the impression is created that-in agreement with the viewpoint of

the Netherlands' monetary authorities-a strong guilder is to be
preferred. The accompanying appreciation, however, does not
necessarily lead to a lower production growth, as the versions 2.1
and 4.0 illustrate. From Figure 1 can be seen that the crawling-peg
versions show a smooth time path for the parity. Nevertheless, in
the longer run they also display fluctuations.
From a comparison of the total costs presented in Table 1 it
appears that the crawling-peg version 2.1 has the lowest costs and
is, as a consequence, the optimal exchange-rate system in the set of
systems investigated. It is the version with a parity rule which
contains the exchange-rate values in the two preceding quarters.
The advantage with respect to the second rating system, the
 123

purchasing-power version of the crawling peg, is only small, how-

ever. Compared to optimal policy under the then existing adjustable-
peg system, putting into practice of the optimal crawling-peg sys-
tem would have resulted in an optimal policy characterized by good
performances with respect to the stability of economic growth, the
rate of inflation and the stability of the two terms of trade. These
attractive aspects would have been realized, however, at the expense
of pursuing the desired current-account balance.
The costs per objective variable indicate clearly that decidedly
the optimal exchange-rate system does not display the best record
for each objective. This observation underlines the partial character
of an approach in which the competing exchange-rate systems
are appreciated by means of the measure of realization of a
sole objective. This feature is often present in the approaches,
described in Section 2, which use simulations or lead to an analyti-
cal solution.

6. Summary and Conclusions

Various crawling-peg systems were evaluated by means of opti-

mal control techniques. The choice of these techniques was
prompted by the fact that changes in the exchange rate's parity,
which is one of the characteristics of a crawling peg, used to be
accompanied by induced changes in the time paths of other objec-
tives and instruments of economic policy. Only optimal control
techniques allow this interaction between policy instruments. Since
such flexibility in policy is common practice, the outcome for
optimal policies will become substantially more realistic. Before we
could apply these techniques to exchange-rate systems charac-
terized by a crawling-peg, the existing algorithm had to be refor-
mulated in such a way that endogenous exchange-rate parities-or,
generally, endogenous target values-are allowed.
Optimal control techniques were then applied to a model of the
Dutch economy for the period 1970-1976. This is a period distin-
guished by unexpected and, therefore, unpredictable economic
shocks. For this period the Dutch economy was modeled by means
of a linear, quarterly, dynamic econometric model, which highlights
the international relations. Elsewhere, the preferences of the policy
makers were derived by means of revealed preference from data for
124

the period 1967-1970; a period marked by a stable and, therefore,

predictable economic development.
From the evaluation of various crawling-peg versions, it turned
out that the crawling-peg system where changes in the parity
are based on changes in the exchange-rate lagged two time periods
(quarters), is the optimal exchange-rate system. When the crawl
is related to a different lag structure of the exchange rate or
to the purchasing power of the currency, the results are also
quite satisfactory. In any case, these crawling-peg versions give.
better results than the system of a fixed, but adjustable, par value
which functioned in the Netherlands during the time period
examined. When the crawling-peg is determined by official inter-
ventions in the foreign-exchange market, the results appear to be
very disappointing.

Notes

1. See, for instance, Helpman and Razin (1979) and (1980), Lapan and Enders
(1980), Helpman (1981), Frenkel and Aizenman (1982).
2. The authors who proposed a crawling-peg exchange-rate system were in favor
of the second variant. See Williamson (198Ia), Table 1.1., p. 6.
3. As already mentioned by Swoboda (1983), from a policy point of view,
a crawling-peg exchange-rate system is familiar to a system of managed
floating.
4. Ethier and Bloomfield (1975) and (1978) deal with this system.
5. This integration plays a central role in the so-called OPTIeA proposal. See
Basevi and De Grauwe (1978), p. 145.
6. In discussions on the application of optimal control techniques, for example,
this objection has frequently been raised. See, amongst others, Prescott (1977).
7. This drawback is counteracted to a certain extent in the present study by the
fact that the model's estimation period partly concerns years characterized by
an adjustable peg, while in the rest of the period examined the guilder did not
float against all other currencies. In addition, currencies are also susceptible to
speculative attacks once parity changes become an accepted way of relieving
exchange-market pressure built up under fixed exchange rates. Both the final
years of Bretton Woods and the so-called Snake arrangement show evidence
in defense of this argument.
8. For example, in the studies mentioned in note I.
9. The complete model, including its stability characteristics, has been presented
in Jager (1981, pp. 140-156 and 180-193) and Jager (1982, pp. 233~242).
10. For details concerning the desired time paths used for the economic objectives
and instruments, see Jager (1981, 156-169 and pp. 197~206) and Jager (1982,
pp. 243~245).
 125

11. The rate of return on monetary reserves and the oppertunity costs of holding
reserves are determined by the application of portfolio analysis to short-term
and long-term investments in reserve assets, respectively. Apart from interest
rates, the gains from exchange-rate of price changes have also been included
in the rates of return.
12. The following weights have been obtained. Growth of GNP: 19; stability of
GNP growth: 15; balance on current account: 7; the two terms of trade: 5. The
weight of the exchange rate was allowed to vary in order to realize an exchange
rate within the fluctuation margin. See Section 5 for further details.
13. The algorithm developed by Stoppler and Stein contains, in addition, some
minor errors due to a wrong sign in their equation (5), p. 224.

References

Basevi, G. and P. de Grauwe (1978): Vicious and virtuous circles and the OPTICA
proposal: A two-country analysis, in: M. Fratianni and T. Peeters (eds.), One
Money for Europe. Macmillan, London and Basingstoke, 144-157.
Chiang, A.-C. (1974): Fundamental Methods of Mathematical Economics.
McGraw-Hill, New York.
Chow, G. C. (1975): Analysis and Control of Dynamic Economic Systems. John
Wiley, New York.
Ethier, W. and A. I. Bloomfield (1975): Managing the Managed Float, Essays in
International Finance, 112, Princeton (N.J.).
Ethier, W. and A. I. Bloomfield (1978): The reference rate proposal and the recent
experience, Banca Nazionale del Lavoro Quarterly Review, 126, 211-232.
Frenkel, J. A. and J. Aizenman (1982): Aspects of the optimal management of
exchange rates, Journal of International Economics, 13, 231-256.
Friedlaender, A. F. (1973): Macro policy goals in the postwar period: A study in
revealed preference, Quarterly Journal of Economics, 87, 25-43.
Group of Thirty (1982): The Problem of Exchange Rates: a Policy Statement.
Group of Thirty, New York.
Hadley, G. (1974): Linear Algebra. Addison-Wesley, Massachussets.
Helpman, E. (1981): An exploration in the theory of exchange rate regimes,
Journal of Political Economy, 89, 865-890.
Helpman, E. and A. Razin (1979): Towards a consistent comparison of alternative
exchange rate regimes, Canadian Journal of Economics, 12, 394-409.
Helpman, E. and A. Razin (1980): A comparison of exchange rate regimes in the
presence of imperfect capital markets. Institute for International Studies Seminar
Paper 156. Institute for International Studies, Stockholm.
Jager, H. (1981): De behoefte aan internationale monetaire reserves als uitvloeisei
van optimaie economische politiek (The need for international reserves conse-
quent upon optimal economic policy), dissertation. Groningen, the Nether-
lands.
Jager, H. (1982): Optimal exchange-rate policy in an open economy, The Econom-
ist, 130, 228-263.
126

Kenen, P. B. (1975): Floats, glides and indicators: A comparison of methods for

changing exchange rates, Journal of International Economics, 5, 107-151.
Lapan, H. E. and W. Enders (1980): Random disturbances and the choice of'
exchange rate regimes in an intergenerational model, Journal of International
Economics, 10, 263-283.
Mayer, H. (1982): The theory and practice of floating exchange rates and the role
of official exchange-market intervention. BIS Economic Papers 5, Bank of
International Settlements, Basle.
Pindyck, R. S. (1973): Optimal Planning for Economic Stabilization. North-Holland,
Amsterdam.
Prescott, E. C. (1977): Should control theory be used for economic stabilization?,
in: K. Brunner and A. H. Meltzer (eds.), Optimal Policies, Control Theory, and
Technology Exports, Carnegie-Rochester Conference Series on Public Policy 7.
North-Holland, Amsterdam.
Stoppler, S. and 1. P. Stein (\ 982): A study of adaptive revision of target values in
an econometric decision model, in: 1. Gruber (ed.), Econometric Decision
Models. Springer-Verlag, Berlin.
Swoboda, A. (\981): Comment on Williamson, in: 1. Williamson (ed.), Exchange
Rate Rules. Macmillan, London and Basingstoke, 35-37.
Theil, H. (1964): Optimal Decision Rules for Government and Industry. North-
Holland, Amsterdam.
Williamson, 1. (ed.) (1981): Exchange Rate Rules. North-Holland, Amsterdam.
Williamson, 1. (I981a): The crawling peg in historical perspective, in: 1. William-
son (ed.), Exchange Rate Rules. North-Holland, Amsterdam, 3-30.
 127

CHAPTER 8

SOME REMARKS ON FORWARD PROGRAMMING

1. Eppers and M. Leserer

University of Gottingen, FRG

1. Introduction and Summary

Planning may be seen as simulating the future. This will be done by

extending experience from the past to future situations. Thus
planning has an essentially probabilistic aspect, because the gen-
esis of life is irreversible in time. One can only expect what will
occur in the future and the usual strategy is to compare data
continuously with expectations in order to adapt 'likelihoods' to
reality.
Against this background the purpose of our paper is to demon-
strate the usefulness of forward programming methodology
[5] in combination with backward programming. This is done
by some basic considerations on the information structure of
multi-period planning and by a worked numerical example.
After seeing how forward tuning works on the basis of a backward
nominal trajectory we conclude that forward programming may be
an operational alternative to the common backward perturbation
analysis.

2. Information and Forward Tuning

In multi-period decision problems one way to handle adaptive

programming is along the lines of the well-known nominal per-
turbation paradigm; see for example [4]. A special perturbation
approach is the forward tuning-scheme proposed in [6]. As prep-
aration we shall write the basic recursion of dynamic programming
obeying information-processing as mentioned above. This may
be done in the framework of conditional expectation theory. For

C. Carraro and D. Sartore (eds.) Developments of Control Theory for Economic AnalysIS
© 1987 Martinus NiJhoff Publishers (Kluwer), Dordrecht
128

periods t and t + 1, say, one has

opttE[ff; + Optt+,E(ff;+,IJt+dIJtb]

= Optt,t+' E( ff; + ff;+, IJIb), a.e"

if Jt c Jt+" and if it is feasible to interchange expectation and
minimization operation. This may generally be done whenever the
functions ff; and ff;+, are of such type that the described conditional
expectations exist [7].
Here optt means the minimization or maximization of the objec-
tive function ff; with respect to the instrument variables. The
symbol J tb stands for the information available at (in the beginning
of) period t. Formally J tb is considered to be an increasing sequence
of a-fields induced by the data-generating random variables in order
to simplify iteration in conditioning. If one assumes that infor-
mation processing is viewed from the last stage of a (finite horizon)
planning problem, the phrase 'increasing' may have a seman tical
identification: When backward induction begins one hopes that
information would grow as time progresses. So in each period new
random variables are added to those of former periods which
implies new cross-partitions of the former information a-field. * In
this sense J tb+, is finer than JIb.
In contrast to backward optimization which works against time,
forward programming operates from the first period in the direc-
tion of time. So an analogous recursion can be written in which only
the order of optimization is changed
optl+,E[optrE(ff; IJ{) + ff;+,IJ{+,]

= Optt,t+' E( ff; + ff;+, IJ{+,), a.e.,

if J{+, c J{. But this yields an overall optimization conditioned on
J{+ l ' Here a somewhat different semantic holds: The information
sets will be considered at the first stage when forward programming
starts. Then one hopes that uncertainty will decrease as time pro-
gresses. This is so when random variables are realized. Therefore
cross-partitions for this period vanish which implies that J{+, is
coarser than J{. Thus the phenomenon of formal increasing or
decreasing of information sets may be seen semantically as a

*We restrict our considerations to denumerable valued random variables. Thus we

can use the term cross-partition [3] for sake of clarity.
 129

problem of choosing one's viewpoint. It depends on the decision-

maker's position as being retrospective or prospective. So the same
information sequence has different formal properties. We conclude
that adaptive feedback optimization is done by looking at past
information growth whereas adaptive feed forward optimization is
done by looking at future uncertainty reduction. Although sym-
metric in theory the forward point of view will have some advan-
tages in practice as will be demonstrated in Section 3.
N ow turning to real dynamic planning these ideas will be imple-
mented by model specification. Whilst backward recursion formu-
lates lag systems, a forward recursion needs lead systems. To be con-
crete we assume that all observed variables occurring in the planning
model are split into deterministic and stochastic parts. We write
YI = YI + L\YI for the (vector of) state variables and XI = XI + L\x l
for the (vector of) instrument variables. Let now YI = h(YI-I' XI)
be an affine planning model. A deterministic nominal trajectory is
established by an affine feedback rule x~Pt = gJ YI_I) applying
Bellman's principle of optimality to some quadratic objective
function ~(YI' xJ [2].
The idea of forward tuning is then to construct a perturbation
trajectory from the lead relationL\YI = AIL\YI+I + BIL\x l + VIand
the objective function L\ W; = (L\YI - L\yi)' QI(L\YI - L\yi) by a
feedforward rule
L\x~Pt = AIL\Yt+ I + 111'
where
AI -(B;HIBI)-I B;HIAt,
111 (B;HIBJ- I B;hp
Ht+1 QI+ I + A;HI(AI + BIAJ,
hl +1 Qt+IL\yi+1 - A;HIBII1I + A;hp
HI QI and hI = QIL\yf.
The lead relation above is a reversed Markov model where VI is
white noise.

3. Operational Aspects-A Numerical Example

The proposed methodology of evaluating a multi-stage control
rule will be now demonstrated on a model of environmental
130

pollution. For that purpose we generated fairly realistic data for a

fictitious business company. In this sense we use the term real data
in the discussion below. Following [1], our model consists of the
two estimated equations
Pt 133.689 + 0.307 PH + 0.765Kt - 17.308IPt
Kt K t_ , + IPt + IPt - At·
The firm's produced pollution Pt is influenced by its lagged value
~-1' the capital assets K t and the investments in pollution control
IPt • The second equation is an identity. K t is explained by its lagged
value K t_, and the balance of all investments in the firm (IPt + IPt )
less the depreciation At.
A ten periods optimal control using a deterministic feedback rule
computes the feedback-fitted values shown in the first parts of
Figures 1 and 2. Starting with initial values 1050, 660, 10 for PI' K t
and IPt respectively a 5% decrease (increase) per period for Pt (Kt)
is assumed. The control variable IPt is held constant over the
planning periods. The matrix of welfare weights Qt is diagonal with
entries 1.0,0.1 and 1000 for the two state variables and the deviations
of IPt •
In concluding the first planning chapter we analyse the deviations
of the real and controlled values to formulate and estimate the lead
model which is in state space representation

[ 11.~
I1.Kt
]
=
[0.168
° -2.297
0.951
0] [I1.Pt+l]
° I1.Kt+1 +
[-14.529]
-0.711 I1.IPt.
I1.IPt ° ° ° I1.IPt+I 1
This representation is of Chow type [2] in which the instrument
variables are incorporated in the endogenous vector in order to
simplify the objective function.
The exogenous non-controllable variables IPt and At may be
forecast in a suitable manner and equal their real values, and can
therefore be neglected in the further discussion for convenience.
The second planning sequence of periods 11 to 20 is started with
a deterministic feedback rule solving the optimization problem
backwards in time as in the first ten periods. Starting with values of
the tenth period we desire a 2.5% and a 0.5% decrease for ~ and K t
whereas IPt is held constant. For correcting the nominal trajectory
 131
: 179~

1 : 241

1069~
1014j
959
"-
"-
904 ~

84 9~

794

739 real values

feedback target values

684 feedback-fltted values

feedforward-fltted values
629
, , , ~,--~-"--~--,,-----,,--~--,,--~-,,

o 2 4 6 8 I0 I2 I4 I6 18 20
1150

1102

1054

1006

958

910

862

814

~-~~i~
/~ / / / - - - ___ ~ F

I
/
,I

r--"T".-.----,------~,--~ , .-."---r----,'-
6 8 10 I4 16 18

Figure I. Results of a ten periods feedforward proceeding.

132

::::1
1069

10 14
!
959 1
904
849

794

739 real values

- - - feedback taqet values

684

629 - - - - -- - - feedforward-fltted values

--,-~
I I I I I I I I
0 2 4 6 8 10 12 14 16 18
I 120

1075

1030
/
985 /
/
940 /
/
895 /
850 /
/
805 /

760
715

670

o 2 6 8 10 12 14 16 18 20
26
24

22
20
18

16
14
12
10
8

e-~-~-~~-~ ~I-~--T -~---'-~-~--'--~I-~--~I-~-~I-~-'

o 2 6 8 10 12 14 16 18 20
Figure 2. Results of a two periods feedforward proceeding.
 133

a lead perturbation is added and determines the final recommen-

dations of the planning problem. The required information about
the future in our example consists in the deviations of the feedback-
fitted values and the real values of the periods 11 to 20. Figures 1
and 2 show two selected feedforward control rules based on the lead
system introduced above, in which in the first case the optimization
problem is solved as a ten periods planning problem. Figure 2
represents a two periods procedure.
The superiority of the two-stage tuning scheme is evident.
Indeed, the feedforward optimization over the whole planning
period leads to a quite good adaptation of the fitted values of I1Pt
to their real values. As well as being able to fit the variable I1Pt in
a more exact way to its targets we can adapt the other endogenous
variable 11K, or the instrument variable I1IPt to theirs if we alter the
welfare weights given in the matrices Q, which are taken as Q, = I,
Vt in the second planning sections of the proposed example.
The special efficiency of the demonstrated multi-stage decision
process is shown when we divide the planning sequence into five
two-period optimizations. For the most part there is a very good
solution of the given problem. There is no accumulation of errors
and deviations as in the undivided optimization for the planning
periods. By using the feedforward control rule we can interrupt the
optimization procedure at that very point where the dimension of
uncertainty gets to such a size that there is no point in a further
decision. Nevertheless, Bellman's principle is preserved and, ceteris
paribus, we do not lose optimality of the whole decision problem.

References

I. Bensoussan, A., E. G. Hurst, Jr., and B. Naslund (1974): Management

Applications of Modern Control Theory, Studies in Mathematical and
Managerial Economics Vol. 18. Amsterdam-Oxford.
2. Chow, G. C. (1975): Analysis and Control of Dynamic Economic Systems. New
York-London.
3. Kemeny, J. G., J. L. Snell, and A. W. Knapp (1966): Denumerable Markov
Chains. Toronto-New York-London.
4. Kendrick, D. (1981). Stochastic Control for Economic Models. Economics
Handbook Series. New York.
5. Larson, R. E. (1968): State Increment Dynamic Programming, Modern
Analyflc and Computational Methods in Science and Mathematics, No. 12,
New York.
134

6. Leserer, M. (1983): A Fine-Tuning Scheme for Economic Decision Rules. In:

1. Gruber (ed.), Econometric Decision Models, Lecture Notes in Economics and
Mathematical Systems, Vol. 208. Heidelberg.
7. Meier, L., R. E. Larson, and A. 1. Tether (1971): Dynamic Programming for
Stochastic Control of Discrete Systems. IEEE Transactions on Automatic
Control, Vol. AC-16, No.6, December, pp. 767-775.
 135

CHAPTER 9

UTILITY AND UNCERTAINTY IN INTERTEMPORAL

CHOICE

Aida Montesano
University of Milan, Italy

1. Introduction

Dynamic optimization is widely used in economics for the analysis

of problems of inter temporal choice. Although the first work in this
field dates back to Ramsey (1928), it is only in the last twenty years,
after the publication of the Pontryagin et al. book (1962), that these
analyses have been used currently with results, theorems and inter-
pretations of great interest, mainly in the theory of growth.'
The aim of this paper is twofold. On the one hand, attention is
directed at some strong hypotheses which are usually implicitly
assumed on the objective functional in order to employ the methods
of dynamic optimization (in the forms, nearly equivalent, of the
maximum principle, of the calculus of variations, and of dynamic
programming). On the other hand, a specification is determined for
the utility functional which results, in the case of uncertainty, when
those strong hypotheses are assumed. A special consideration will
be given to the specification of the utility functional in case of
randomness of the terminal time, which is a realistic assumption,
albeit not often adopted, for many problems of intertemporal
choice.

2. The Utility Functional in Problems of Intertemporal Choice

The problem of intertemporal choice, which I shall first consider

in conditions of certainty, presumes that the agent under examination
can choose among actions, represented by functions of time,
about which he has a preference ordering. The choice consists of

C. Carraro and D. Sartore (eds). Developments of Control Theory for Economic Analysis
© 1987 Martinus NlJhoff Publishers (Kluwer), Dordrecht
136

determining the action, or actions, which are preferable. In rather

more formal terms, by indicating with E the set of the functions of
time and with En the product E x Ex . .. x E; with a the actions
of the agent under examination (which are represented by a vector
of n functions of time); with A the set of the feasible actions
(therefore with A c En); and with <A, R) the preference ordering
on A given by a binary relationship R of the type "not less preferred
to", the problem of intertemporal choice consists of determining a
subset of A maximal with respect to the preference ordering. If the
preference ordering is regular, i.e., total, reflexive and transitive,
then a utility functional (in general, non-standard) exists,2 i.e., a
functional A -+ Re, where Re is the set of real (non-standard)
numbers, for which U {a,} ? U {a2} if and only if a, R a2for any
pair a" a2 E A. This intertemporal utility functional is ordinal, i.e.,
if F is an increasing function from Re to Re and V = F( U) then
V{a,}? V{a 2 } if and only if U{a,} ? U{a 2 }.
With these definitions and assumptions the problem of inter-
temporal choice becomes one of finding the maximum of the inter-
temporal utility functional on the set of feasible actions, i.e.,
maxaEA U{a}. There are, in mathematics, methods for solving the
above problem in some particular cases: when the objective func-
tional U{ a} is a "sum" of instantaneous elementary values of a
function U E E depending on actions a E A, i.e., of the following kind

U{a} = ito
tl
u(t; a)dt. (1)

For instance, the calculus of variations considers an objective

functional

U{ a} = i u(a(t), aCt), t) dt.

to
tl

Analogously, the maximum principle-by distinguishing among

state variables and control variables, the first depending on the
second through constraints represented by differential equations-
considers the objective functional

U{a} = Ito
tI
u(a(t), x(t), t)dt

with constraints x(t) = f(a(t), x(t), t) on the state variables x (the

control variables a are, in the language of the choice theory, the
actions).3
 137

This hypothesis on the utility functional can be interpreted by

considering the elementary values u(t) as instantaneous utilities, of
which the intertemporal utility U is the sum.
The hypothesis represented by relationship (1) however is, on the
one hand, restrictive since the intertemporal utility of the agent is
not always representable as in relationship (1): for instance, agent's
preferences may be those represented for AcE by an inter-
temporal utility functional of the type U = maXto<:;t<:;t] aCt) (as it
happens if the objective consists of obtaining, sooner or later, a
sporting record). On the other hand, the hypothesis introduces a
kind of additive utility: consequently, the instantaneous utilities u(t)
are no longer ordinal, but cardinal, defined only up to any linear
increasing transformation, while the intertemporal utility U is still
ordinal, i.e., defined up to any increasing transformation. In other
words, the preference ordering represented by the utility functional
(1) is likewise represented by the functional

with {3 > 0 and F being any increasing function, while it does not
happen in the case of non-linear transformations of u(t).

3. Intertemporal Choice in Conditions of Uncertainty

The relation between the utility functional (1) and the preference
ordering requires a deeper analysis in case the choice is made in
conditions of uncertainty. In general terms the intertemJ'oral choice
in conditions of uncertainty concerns actions which can be indi-
cated by means of distributions of probability on functions of time
which represent their consequences. In other words, by indicating
the set of consequences with C c En, actions are distributions of
probability on C, i.e., functionals a: C ~ I where I is the unitary
real interval, with
a{c} ~ 0 and I a{c} = 1.
CEe

By indicating with A the set of feasible actions (which is a set of

distributions of probability on C), if the preference ordering
<A, R) satisfies the von Neumann-Morgenstern, or other
138

analogous, axioms, then a utility functional on A (defined up to any

linear increasing function) exists, for which the Bernoullian principle
of expected utility holds, i.e., such that
Unm{a} = L
CEC
a{c}Unm{c}
Consequently, the problem of intertemporal choice becomes the
maximization of the expected utility, i.e., the problem
max L a{ c} Unm {c}.
a{c}EA CEC

If the preference ordering of the agent under examination is

represented with a utility index different from the von Neumann-
Morgenstern index (for which Unm is one of its increasing trans-
formations Unm = F(U)), then the utility of an action is the
particular associative mean of the utility of its consequences which
is defined by function F through the relationship 4

U{a} = F- 1 C~c a{c}F(U{C}))

Considering the problem of intertemporal choice in conditions of
uncertainty, the objective functional U {a} can be represented as
relationship (1) requires-in order to employ the usual mathematical
methods for solving the problem of dynamic optimization-only by
assuming a strong hypothesis. We must assume that the von
Neumann-Morgenstern utility index is such that
Unm{c} = Itto
I u(t; c)dt (2)

since only in this case can the 0 bjecti ve functional U{ a} be of the

kind (1).
On the contrary, if the von Neumann-Morgenstern index is a
concave transformation-as is more realistic to assume-of an inten-
sive utility index obtained by summing up instantaneous indices, i.e.,

Unm{c} = F(r: u(t; C)dt)

with F a strictly concave increasing function, then the objective
functional becomes

C~C a{ c} F (t u(t; c) dt)

and none of its increasing transformations are, in general, like type
(1).
 139

4. The Specification of the Utility Functional in Conditions of

Uncertainty

We can now determine the specification of the objective functional

in conditions of uncertainty assuming that both the von Neumann-
Morgenstern axioms and the strong hypothesis of relationship (2)
hold. Before determining the objective functional, set C of conse-
quences, set A of actions and the other required hypotheses must be
specified.
Let set BeEn be composed of vectors b of functions of time
with to ~ t < CIJ and interval T, of numbers t, ~ to. Set C of
consequences is their product C = T, x B: a consequence is deter-
mined by a terminal time t, and by a vector b of functions of time.
Let a function u be associated to every element of C. This function
represents the instantaneous utility of vector b in period to ~ t ~ t,
and determines, as already shown, the intertemporal utility through
the relationship

so that the objective functional becomes

We will consider first the case when only the functions bare
random, second the case when only the terminal time t, is random,
and third the general case when both terminal time t1 and functions
b are random.
When only the functions b are random we have the objective
functional

L
hE B
a{b; td I
tI

to
u(t; t" b)dt

(3)

where a{ b; t 1 } is the probability of having the consequence repre-

sented by vector b, given the terminal time t,. This intertemporal
utility is of the type (1).
140

When only the terminal time II is random the objective functional

IS

where a(/l; b) is the probability that the terminal time is II. In this
scheme, set A of feasible actions, which is composed of the feasible
distributions of probability on C = TI X B, associates a distri-
bution of probability on interval TI with some points of B. Let us
now indicate the distribution of probability on TI by means of a
function of density of probability J(/I; b): i.e., the probability that
the time period finishes in the interval between II and tl + dt l is
given by J(tl) d/ l . Consequently

Unm {a} = 1~ J(tl; b) (f~ u(t; II, b) dl) dt l ·

This utility functional can be written in the form required by

relationship (1). In fact, it is

Unm {a} = 1~ (r; J(tl; b)u(t; t l , b) dtl) dt. (4)

If function u(t; II' b) does not depend on II, i.e., it is of the kind
u(t; b), then, by introducing function

get; b) = r~ J(/I; b) dt l ,

we obtain

Unm{a} = foo get; b)u(l; t l , b) dt,

10

moreover, where g(t) is a non-increasing function with value 1 for

t = to and value 0 for t sufficiently high.
Let us consider now the general case where both terminal time II
and the functions b are random. Objective functional is then of the
type

where a{tl' b} is the probability of having the consequence repre-

sented by II and b, i.e., by vector b with 10 ~ I ~ II.
 141

Set A of feasible actions is a set of distributions of probability on

C = TI X B. By indicating the conditional distribution of tl given
b with a(tl: b), it is
a{t l , b} = a(tl: b)a{b}
where a{ b} is the marginal distribution of b. Thus the utility func-
tional can be written as

so that, by introducing the same functions J(tl: b) already con-

sidered for the case when only the terminal time is random, we
obtain

i.e.

Unm{a} = fCJ)
10
(L a{b} foc J(tl: b)u(t; t l, b)dt l ) dt,
bEB 1
(5)

which is again a utility functional of type (1). Ifit is u(t; b), then we
have

Unm{a} = f~
10
L
bE B
a{b}g(t: b)u(t; b)dt,

5. Conclusions

The usual methods of dynamic optimization require an additive

objective functional, as represented by relationship (1). This requisite
implies some crucial hypotheses, which are to be taken into account
in specifying problems of intertemporal choice.
Assuming that the intertemporal choice can be represented as a
maximum problem of a utility functional, the first crucial hypothesis
requires that the intertemporal utility functional be of type (1), i.e.,
a sum of instantaneous utilities. (Moreover, of course, the utility
functional must satisfy also the other mathematical hypotheses
required for applying the maximum principle, the calculus of
variations, or dynamic programming methods).
142

If the intertemporal choice is made in conditions of uncertainty,

once the von Neumann-Morgenstern (or other analogous) axioms
are assumed to hold, the second crucial hypothesis requires that the
von Neumann-Morgenstern utility index must be an additive func-
tional as indicated by relationship (2).
Having assumed these hypotheses, the intertemporal utility func-
tional can be represented as a relationship of type (1): in the case
that only the functions of time which represent the possible con-
sequences of actions are random we have relationship (3); in the
case that only the terminal time is random we have relationship (4);
and in the general case where both those functions of time and the
terminal time are random we have relationship (5).

Notes

1. See, for instance, the volumes edited by Shell (1967) and by Cass and Shell
(1976). Analyses of problems of stochastic control have also been used
frequently in recent times: an introduction to them is given by Tintner and
Sengupta (1972).
2. See Richter (1971), p. 43.
3. An introduction to the mathematical theory of dynamic optimization with
reference to economic problems is given by Intriligator (1971), pp. 291--448.
4. See Montesano (1982).

References
Casso David and Karl Shell (eds), The Hamiltonian Approach to Dynamic Econ-
omics, New York: Academic Press, 1976.
Intriligator, Michael D., Mathematical Optimization and Economic Theory,
Englewood Cliffs, N.J.: Prentice-Hall, 1971.
Montesano, Aldo, "The Ordinal Utility under Uncertainty", Rivista Internazionale
di Scienze Economiche e Commerciali, 1982, 29, 442--446.
Pontryagin, L. S., V. G. Boltyanskii, R. V. Gamkrelidze and E. F. Mishchenko,
The Mathematical Theory o/Optimal Processes, New York: Interscience Pub!.,
Wiley, 1962.
Ramsey, Frank P., "A Mathematical Theory of Saving", Economic Journal, 1928,
38, 543-559.
Richter, Marcel K., "Rational Choice", in John S. Chipman, Leonid Hurwicz,
Marcel K. Richter and Hugo F. Sonnenschein (eds), Preferences, Utility, and
Demand, New York: Harcourt Brace Iovanovich, 1971, 29-58.
Shell, Karl (ed.), Essays on the Theory of Optimal Economic Growth, Cambridge:
MIT Press, 1967.
Tintner, Gerhard and Jati K. Sengupta, Stochastic Economics (Stochastic Processes,
Control, and Programming), New York: Academic Press, 1972.
 143

CHAPTER 10

GRADIENT METHODS IN FIML ESTIMATION OF

ECONOMETRIC MODELS

Giorgio Calzolari and Lorenzo Panattoni

IBM Scientific Center, Pisa, Italy

1. Introduction

Efficient computational algorithms, to produce full information

maximum likelihood estimates of the structural form coefficients in
a system of simultaneous equations, have worried for a long time
and are still worrying econometricians.
Several optimization techniques have been proposed in the last
few years and experimented with on linear and nonlinear models of
increasing size. While some techniques are search algorithms which
do not make use of information on first and second derivatives (e.g.,
Parke, 1982), it is generally acknowledged that gradient methods,
and more specifically Newton-like methods, which make use of such
information, should be superior to the others, at least near the
optimum. The drawback of Newton-like methods, as well pointed
out in Belsley (1980, p. 222), lies in the excessive cost required in the
calculation of the Hessian matrix. Therefore, methods have been
proposed in the literature which replace the Hessian matrix with
other matrices, like those adopted in Berndt, Hall, Hall and Haus-
man (1974), Amemiya (1977), or in Dagenais (1978).
Belsley's findings, after comparing the computational optimiz-
ation performances of different matrices, placed the algorithm
which uses the "exact" Hessian in a dominant position for optimiz-
ation of the FIML objective function. On the other hand Dagenais'
experiments showed that a gradient method in which the Hessian is
replaced by a suitable approximation can be computationally more
efficient that a Newton-like algorithm, at least as long as the robust-
ness with respect to the initial guess of the coefficients is concerned.

C. Carraro and D. Sartore (eds.) Developments of Control Theory for EconomIc Analysis
© 1987 Martinus Nijhoff Publishers (Kluwer), Dordrecht
144

In our Monte Carlo study, the performances of the Newton-like

method are compared with the performances of gradient algorithms
in which more easily obtainable matrices are used: the outer product
matrix proposed in Berndt et al. (1974), and the generalized least
squares type matrix discussed in Amemiya (1977, p. 963) and exper-
imented with in Dagenais (1978).
A large set of Monte Carlo experiments is performed on models
of different size and with different sample period lengths. A sys-
tematic average behavior is derived from the Monte Carlo exper-
iments and evidenced in the paper.
Convergence with the outer product matrix is usually slow, at
least as the number of iterations is concerned; simplicity in the
computation of the matrix provides only a partial compensation in
terms of computation time. This result is in agreement with Belsley
(1980).
The convergence with the Hessian is usually faster near the
optimum, again in agreement with Belsley (1980), while the
generalized least squares type matrix works better far from it, and
this is not only in terms of "robustness" (less chance of false con-
vergence to a saddle point rather than a maximum), as already
Dagenais (1978) noticed, but also in terms of "gain" inside the
iterative gradient procedures. This result, which motivated this
paper, was observed and measured across a large set of Monte
Carlo replications on models with short sample period lengths (like
models with annual data) and might be approximately quantified as
follows. Whichever "good" starting point of the iterative maximiz-
ation process was adopted, such as the point obtained from single
equation estimation (least squares or instrumental variables), only
when most of the distance (99% or more, on the average) between
the initial point and the optimum had been covered, the convergence
became faster using the Hessian.
This suggests first of all that, although the Hessian as expected
performs better near the optimum, this "near the optimum" should
be interpreted in a much mor restrictive sense than usually believed
in practical applications.
On the other hand, the fact that the generalized least squares type
matrix "gains" usually more than the Hessian near the starting
point and less near the optimum might be quite useful for
implementing FIML procedures. A good improvement of the com-
putational efficiency has in fact been obtained by using a mixed
 145

gradient algorithm based on the generalized least squares type

matrix in the first iterations and on Hessian in the last iterations.

2. Three Gradient Methods

Let the system of simultaneous equations be represented as

h(Y" x" a,) = Uti, i = 1,2, ... , m; t = 1,2, ... , T, (1)
where Yt is the m x I vector of endogenous variables at time t, x,
is the vector of predetermined variables at time t and ai is the vector
of unknown structural coefficients in the ith equation. The m x I
vector of random error terms at time t, Ut = (Ul" U2" . . . , um ,)', is
assumed to be independently and identically distributed as N(O, L),
with L completely unknown, apart from being symmetric and
positive definite. The complete n x I vector of unknown structural
coefficients of the system will be indicated as a = (a~, a;, ... ,
a~)'.
The concentrated log-likelihood function is

IT = ~ log 118J;/8y; II - T/2 log IT- 1 ~J;J;/I· (2)

whereJ; = (/1,,/21> ... ,1m')' = Ut and the Jacobian determinant

18J;/8y;1 is taken in absolute value.
A gradient iterative procedure to maximize the log-likelihood
function can be represented by the formula:
a(k) = a(k-l) + AQ8IT /8a, (3)
where a(k-l) is the estimate of the coefficients vector obtained after
k - I iterations, Q is some n x n matrix, and A is a real number
(scalar).
Gradient methods differ in the way in which the matrix Q and the
scalar}... are selected at each iteration. The selection of the matrix Q
determines the choice of the direction along which the search for the
maximization of the log-likelihood function will be made. The
choice of A determines the step size in this direction to obtain the
new values of the coefficients.
As long as the choice of Q is concerned, three different approach-
es have been tried.
146

2.1
The matrix Q is given by the inverse of the Hessian of the
log-likelihood function. The analytical expression of the i, jth
block of the Hessian is given in Amemiya (1977, eq. 3.5)

-o2IT/oa/Jai = - ~ ogg//oull + T(~ gifJ,')(~;;;;'rl

+ [ ~ (agll/au}J(agitlau,t) ] + T (~};};' Xl (~gll git)
(4)
_ T(~gll!t')(~};};'rl (~};;;,}-l (~};g;t)
_ T (~};fr,)~l (~gll};') (~};};'rl (~};g;),
where gil = a};tloa, (in practice the vector gil contains the values of
the explanatory variables appearing in the ith equation, if the
model is linear in the coefficients), aglllau}t = (ogil/oy;)(a};lay;)}-I,
ogljt/OU,t = (og'}t/ay;)(a};/ay;),-l, and a single subscript i represents
the ith column of the matrix. In this case the gradient method
becomes a Newton-like algorithm.

2.2
The matrix Q is given by the inverse of the generalized least
squares type matrix introduced in Amemiya (1977, p. 963) and
experimented with in Dagenais (1978). Such a matrix is obtained as
follows. We first introduce the T x m matrix F, whose t, ith
element is};(Yt, XI> a) = U u (the matrix of residuals) and the matrix
G" whose t th row is g;t (in practice, for models linear in the
coefficients, the matrix with the values of the explanatory variables
appearing in the ith equation). We define, now,
G, = G, - T- 1 F L (oga/ou;)' (5)

and build the block diagonal matrix G, whose m diagonal blocks are
C,. The generalized least squares type matrix used in the gradient
procedure is the inverse of the matrix
(6)
 147

(for linear models, G has the form of the matrix used in Aitken-
Zellner estimation, containing the values of the explanatory vari-
ables, but with the historical values of the endogenous variables
replaced by the computed values).

2.3
The matrix Q is given by the inverse of the outer product matrix
proposed in Berndt et al. (1974) whose i, j th block is

T- 1~ [aga/aua - T(g,J/) (~J;J;'rl ]

x [agJI/auJI - T(gJ,J;') (~J;J;'rll (7)

The choice of the step size A. has been performed following an

optimality criterion, i.e., trying to maximize the log-likelihood
function by means of an univariate search in the selected direction
(see also Eisenpress and Greenstadt, 1966, or Dagenais, 1978). Of
course, the procedure is only based on heuristic considerations and
there is no assurance that such a strategy for the selection of the
value of A. is an optimal one; however, it appeared in practice to
accelerate the calculations and to assure the convergence in most
cases, and, therefore, it gave a good common basis for performing
comparisons of the gradient algorithms using the three matrices.
For the univariate search we used a part of Powell's algorithm,
as described in Pierre (1969, pp. 277-280), which does not involve
the use of derivatives, but is quadratic convergent all the same.
Particular care had to be used in the choice of the tolerance for the
convergence in this univariate search because, although the maxi-
mization process improved the computational efficiency of the
whole algorithm, this implied the evaluation of several values of the
log-likelihood function. These computations, for medium and large
size models, are rather time consuming and it can happen that with
a too tight tolerance the algorithm requires a high number of such
computations without a corresponding improvement in the efficiency
of the whole algorithm. For the experimented models we found that
values 0.01-0.001 of the relative tolerance on A. are usually good
values for the overall computational efficiency of the maximization
algorithm.
148

3. Experimental Comparison

Monte Carlo Experiments have been performed on four models

of small medium size. Two models are linear, and two are nonlinear
in variables.
(1) A multiplier-accelerator model, with three linear equations, two
of which stochastic, and 6 unknown structural coefficients; the
equations and empirical data can be found in Dhrymes (1970,
pp. 533-534).
(2) A model for the Italian economy proposed in Sitzia and Tivegna
(1975), consisting of 7 linear equations, 5 of which stochastic,
and 19 unknown structural coefficients.
(3) A mildly nonlinear version of Klein-I model (six equations,
three of which stochastic, and 12 unknown coefficients),
obtained by replacing the linear equation for consumption with
a log-linear equation (see Belsley, 1980, model 3B).
(4) The Klein-Goldberger model (Klein, 1969), which is nonlinear
in variables and consists of 20 equations, 16 of which stochastic,
with 54 unknown structural coefficients.
Monte Carlo experiments on all models are based on a few
hundred replications, each of which has been performed as follows.
Starting from the model with a given set of parameters ("true"
coefficients and covariance matrix of the structural disturbances,
held fixed in all replications), random values of the endogenous
variables over the sample period are generated by means of stochas-
tic simulation and are used for FIML estimation with the three
methods.
To reproduce as much as possible the conditions under which
FIML estimation is performed in practice, we choose a "good"
starting point for each estimation by getting a preliminary single
equation estimate (least squares or instrumental variables).
Several convergence criteria (on coefficients, on the likelihood
and on the gradient) have been experimented with. While some
differences have been encountered in several cases, the overall
behavior did not change very much with the different criteria, apart
from the obvious lengthening of convergence "tails" when adopting
a very tight tolerance. The same can be said about the choice of the
sample period length; the overall behavior did not change, apart
from the obvious shortening of convergence "tails" with all methods
when the sample period becomes longer. Again the overall
 149

behavior did not change with the different choice of the predeter-
mined variables in the sample period (exogenous variables have
been either kept fixed in all experiments, or randomly generated
with given means and covariance matrix, and lagged endogenous
variables have also been kept fixed in all experiments, or randomly
generated using dynamic stochastic simulation), and with the dif-
ferent choice of the "true" parameters of the model, on which
Monte Carlo generations are based.
The simple computation of the number of iterations required to
get convergence with the three matrices is not particularly illumi-
nating (some more details can be found in Calzolari and Panattoni,
1983). The only sure indications which were obtained are the fol-
lowing.
(1) The use of the Hessian never requires very long tails for the
convergence, while the other two matrices (the outer product
matrix, in particular) often do.
(2) The Hessian, apart from the computational burden, rises more
often than the other problems of false convergence to saddle
points when it is used for the estimation of rather complex
models (about one out of five cases with the Klein-Goldberger
model with less than 50 observations).
Much more interesting considerations are obtained if we have a
better insight in the convergence process. For each Monte Carlo
replication, we first compute the maximum with a very high pre-
cision, then we measure the fraction of the distance between the
starting point and the maximum covered at each iteration, with the
three methods. The distance is measured both on the values of the
log-likelihood and as length of the difference between the cur-
rent and the final coefficient vectors. As before, in some cases the
two measures give different results, but the overall behavior is
practically the same. In Figure 1 results related to the distances
measured on the values of the log-likelihood function are displayed
on a log-scale. Ifwe call D(k) the distance which, after k iterations,
still remains to get to the maximum, the value which is calculated
IS

d(k) = -log [D(k)/ D(O)]. (8)

The value of this variable is equal to zero at the starting point,
increases at any new (kth) iteration, as we move monotonically
"uphill", and would be infinite at the optimum (in practice it
150

Mu Itlpller-accelerator Linear Italian

model model

8 • 8
• •
• •
6
6 •
• • •
•• 4
• ••
4 j •
••
'I ~ *
I
:I< :I< * *
d(k) , * *
*

k= 2 4 8 10 k= 5 10 15 20
Log-linear Klein-I Klein-Goldberger
model model

8
•
B •
•
•
••
• • 6 •
6
• •
• .* ••
•• 4
•
4
•
• • •
2,.. • *
• :I<
• • :I< 2
• :I<
•• *:1< * *
:1<:1<:1<
•
••
• *:1<*:1<:1<*
**
d(k)··* *,* d(kl~_.~:1<_**~*__~__~__________
k= 5 10 15 20 5 10 15 20 25

•••
.. . Hessian
Generalized least squares type matrix
Outer product matrix
* **
Figure 1. Average rate of convergence of the three gradient algorithms.

assumes a value of a few units, depending on the choice of the

tolerance in the convergence criterion). For example, a value 4
means that the distance from the maximum of the point obtained
after k iterations is 10- 4 of the distance between the starting point
and the maximum.
For each model, and for each iteration number (k), the value
which is displayed in Figure 1 is the average value of all d(k), across
a few hundred Monte Carlo replications, obtained from using the
three matrices.
An interesting systematic behavior of the three methods can be
 151

observed for the models in Figure 1, where the length of the sample
periods are those of the historical data originally proposed for the
models themselves (only for the Klein-Goldberger model the
sample had to be enlarged of a few observations). The gradient
algorithm, which makes use of the generalized least squares type
matrix is considerably faster in the first iterations and, on average,
it allows to cover a good deal of the distance from a "good" starting
point up to the maximum (more than 99.9% for these experiments
based on rather short samples) in a smaller number of iterations
that the same algorithm which makes use of the other two matrices.
The dominance of the Hessian matrix becomes effective only in a
very tight neighborhood of the optimum, where it allows a con-
siderable reduction of the number of iterations.

8
• •
• •
6
• •
• "

• " "
• "
" "

• •
•.
4 •
••
"
"

"."
••
•
>\'

2
• •• >\' *
." • • •
>\'

•• >\' *
"
•"

.>\,. •
>\' *
>\' >\'
>\'
*
>\'
>\'

d( kJ

5 10 15 20 25

• • • • •• "Hessian

" " •• " • " Generalized least squares type matrix

>\' ***** Outer product matrix

• • • • • " Two-step mixed algorithm

Figure 2. Average rate of convergence of the three gradient algorithms and of the
mixed gradient algorithm on the Klein-Goldberger model.
152

This average behavior, which systematically occurs with minor

variations in all the models and under all the different conditions
experimented with, might be interesting for improving the com-
putational efficiency of FIML algorithms. The use of the
generalized least squares type matrix seems recommendable in the
first iterations (it becomes even more recommendable when con-
sidering that its computation is rather simple and fast even for
medium-large size models and is, in any case, considerably simpler
and faster than computation of the Hessian). After a few iterations,
the use of the Hessian should be preferred.
For example, since the slope of the curve related to the Hessian
in Figure 1 becomes the highest when, on the average, d(k) = 2
(10- 2 of the total distance still remains to get the maximum), a
two-step mixed iterative algorithm would produce a good improve-
ment of the computational efficiency. We first adopt a convergence
criterion with a wide tolerance (for example a relative tolerance
10- 2 on coefficients). Starting from a "good" initial value of the
coefficients vector, we first apply iteratively the gradient method
using the generalized least squares type matrix, until convergence is
reached. We then adopt a tighter tolerance for the convergence
criterion and apply iteratively the gradient method using the Hess-
ian. A mixed gradient method of this kind, applied to the Klein-
Goldberger model with a sample period of 50 observations, gave,
on the average, the improvement of the computational efficiency
evidenced in Figure 2.

References

Amemiya, T. (1977): "The Maximum Likelihood and the Nonlinear Three-Stage

Least Squares in the General Nonlinear Simultaneous Equation Model",
Econometrica 45, 955-968.
Belsley, D. A. (1980): "On the Efficient Computation of the Nonlinear Full-
Information Maximum-Likelihood Estimator", Journal of Econometrics 14,
203-225.
Berndt, E. K., B. H. Hall, R. E. Hall, and J. A. Hausman (1974): "Estimation
and Inference in Nonlinear Structural Models", Annals of Economic and Social
Measurements 3, 653--665.
Calzolari, G. and L. Panattoni (1983): "Hessian and Approximated Hessian
Matrices in Maximum Likelihood Estimation: A Monte Carlo Study". Pisa:
Centro Scientifico IBM, paper presented at the European Meeting of the Econo-
metric Society, August 29-September 2.
 153

Dagenais, M. G. (1978): "The Computation of FIML Estimates as Iterative

Generalized Least Squares Estimates in Linear and Nonlinear Simultaneous
Equations Models", Econometrica 46, 1351-1362.
Dhrymes, P. J. (1970): Econometrics: Statistical Foundations and Applications.
New York: Harper & Row.
Eisenpress, H. and J. Greenstadt (1966): "The Estimation of Nonlinear Econo-
metric Systems", Econometrica 34,851-861.
Klein, L. R. (1969): "Estimation of Interdependent Systems in Macroecono-
metrics", Econometrica 37, 171-192.
Parke, W. R. (1982): "An Algorithm for FIML and 3SLS Estimation of Large
Nonlinear Models", Econometrica 50,81-95.
Pierre, D. A. (1969): Optimization Theory with Applications. New York: John
Wiley & Sons.
Sitzia, B. and M. Tivegna (1975): "Un Modello Aggregato dell'Economia ltaliana
1952-1971", in Contributi alia Ricerca Economica No.4. Roma: Banca d'Italia,
195-223.
 PART III

RECENT DEVELOPMENTS OF CONTROL THEORY: A

GAME THEORETIC APPROACH
 157

CHAPTER II

METHODS FOR THE SIMULTANEOUS USE OF

MULTIPLE MODELS IN OPTIMAL POLICY DESIGN

Berc Rustem
Imperial College of Science and Technology, London, UK

1. Introduction

In economics, as in other disciplines, there usually is more than one

theory purporting to explain any particular phenomena of interest.
This has given rise to alternative models of the same economic
system. The existence of such rival models has given rise to two
different approaches. The first is the general econometric testing
and evaluation of rival models (see, e.g., Davidson and Mackinnon,
1981; Hendry, 1983; Hoel, 1947; Mizon, 1984). The second is the
simultaneous use of these models in generating pooled forecasts.
The latter approach accepts, in the absence of absolute econometric
judgement in favour of one of the models, that each model embodies
some useful information that is not present in the other models. The
forecast of each model is then pooled, for example, to minimize the
overall forecast error (see Granger and Newbold, 1977). Both
approaches are, nevertheless, interlinked since the former may use
the forecast pooling operation as a means to test the hypothesis for
the validity of a model against its rivals and the latter approach may
use econometric judgement in defining the pooling operation. In
this paper we discuss ways in which the latter can be extended to
policy optimization.
One obvious way in which rival models can be used in policy
optimization is when the decision maker collaborates in defining
the pooling process. This can be done either by defining a joint, or
pooled, objective function for both models or, as in forecasting, by
specifying the way in which the endogenous variables of rival
models are to be pooled. These aspects are discussed in Sections 2
and 3. One possible suggestion for defining the pooling operation

C. Carraro and D. Sartore (eds). Developments of Control Theory for Economic Analysis
© 1987 Martinus NijhofJ Publishers (Kluwer), Dordrecht
158

is based on an interactive approach to the definition of the weighting

matrices ofajoint quadratic objective function. This approach is an
extension of the method discussed in Rustem and Velupillai (1984)
and is explained in Section 5. A second approach to rival models is
to view these as policy advisers defending alternative theories. In
this case, a game between the advisers can be envisaged and a
Pareto optimal solution can be imposed by the decision maker. This
is discussed in Section 3. An alternative view of rival models is when
more than one decision maker is able to affect the way in which the
economy is run. Each decision maker may have a different model
reflecting his own view of the economy. Nash and Stackelberg's
strategies for these are discussed in Section 4. The final approach is
intended to free the pooling of model from any prior specification
of the pooling process. This can be done by adopting min-max
strategies, discussed in Sections 6 and 7, where the policy is designed
to be optimum under the worst possible model choice. As shown in
Proposition 7.6 and Corollary 7.10, these min-max strategies define
robust policies which cannot do worse if the real world turns out to
be based on any of the rival models.
The choice of computational techniques for the above strategies
is restricted by the size and nonlinearity of most forecasting models.
An important simplification is the use of a deterministic frame-
work. Any disturbances in individual models are therefore treated
as possible errors. The optimal solution can be designed to be
robust to these errors using the approach of Karakitsos, Rustem
and Zarrop (1981). In a dynamic framework, only open-loop optimal
policies are considered. These policies are sequentially updated as
time progresses, to account for incoming new information about
assumptions concerning the economic system-such as revised
exogenous assumptions (see Athans et al., 1976).

2. An Algorithm for the Simultaneous Use of Rival Models

A consequence of the well known divergence of opinion in

macroeconometric model building is the desire to incorporate sim-
ultaneously the rival alternative theories represented by these
models in policy optimization. Chow (1979) has illustrated one way
in which rival models may be utilized in policy analysis. Chow's
approach consists of the application of the optimal strategy, derived
 159

using each individual model, to the other rival models and choosing
the strategy that causes least damage in the real world happens to
be reflected by one other than that on which the optimal policy is
based. In this section we consider the alternative when the optimal
policy is based on a combination of rival models. The way in which
this combination is defined is discussed in later sections. It is shown
that the algorithm below is general enough to apply for all such a
priori combinations of models discussed below.
Throughout this section and Sections 3 and 5 only two rival
models are assumed to exist. This is to ensure simplicity. However,
the arguments below can be trivially generalised to more than two
models. Thus, let
(2.1)
denote the first model with Y I as the vector of endogenous variables
of Fl. FI is a vector valued, possibly nonlinear, function and V is
the vector of policy instruments. Similarly, the second model is
written as
(2.2)
In general, V may be regarded as an augmented vector of the policy
instruments of FI and F2 with V ~ [vi, vrf where VI and V 2 are
respectively the policy instruments of F I and F 2. The case in which
FI is also a function ofY 2 and F2 is also a function ofY I is discussed
in Sections 6 and 7. The vector of policy instruments is assumed to
be common to all models, although the functional structure of F I
or F 2 may exclude some of these instruments. The exogenous values
that are not subject to optimization are assumed to have been
substituted into both models. Finally, the apparent static formulation
in (2.1) and (2.2) also includes dynamic models as discussed in
Rustem (1981), Rustem and Zarrop (1979, 1981)*.
Consider the optimization problem with an objective function
l(Y I , Y2' V) and the constraints (2.1) and (2.2). It is shown below
that such an objective function is sufficient to characterize all
possible a priori model combinations considered in this paper.

*The variables Yl> Y2, U, Fl and F2 WOUld, in this case, be stacked vectors of
dynamic variables. As a particular structure has not been assumed for F 1 , F 2, this
representation also covers models in which a causal dynamic structure does not
necessarily hold (e.g., rational expectations type models).
160

Thus,
min {l(Y" Y2 , U) IF, (Y" U) = 0, F 2 (Y2 , U) = O}, (2.3)
and if the objective function is a quadratic, separable in Y" Y2
and U,
l(Y, , Y2 , U) = ![Y, - YfYQYl[Y' - Yf]
+ ![Y2 - Y~Y QY2[Y 2 - Y~]
+ ![U - UdYQu[U - U d ] (2.4)
then the Gauss-Newton type optimization algorithms discussed in
Rustem and Zarrop (1979, 1981) and Rustem (1981) may easily be
modified to solve (2.3). The superscript d denotes desired values,
Qy 1 , Qy2 , Qu are {I,.symmetric matrices with Qu > 0 and Qy ,Qy2 ~ o.
In general, U d = [Uf , U~ Y where Uf and U~ refer to the desired
T T l

values of the policy instruments of F, and F 2'

ASSUMPTION 2.5: Given the policy instrument values U, the set of

equations F, (Y" U) = 0 may be solved for VI' Thus there exists a
mapping Y, = g, (U). Similarly for F 2 (Y2 , U) = 0 there exists a
mapping Y2 = g2(U),

Assumption 2.5 implies that both econometric models

F,(Y" U) = 0 and F 2 (Y 2 , U) = 0 have model solution programs
which compute Y, and Y2 respectively, given U. Clearly, the map-
pings g, and gz are provided by the solution programs of (2.1) and
(2.2) respectively. This assumption may be used to eliminate Y, and
Y2 from (2.4) and reduce the constrained optimization problem
(2.3) to unconstrained minimization over U
min {G(U) IU E EnK} (2.6)
where E nK is the nK dimensional Euclidean space with n as the
number of the policy instruments and K as the number of time
periods over which policy optimization is being considered. The
function G(U) is given by
G(U) l(g, (U), g2(U), U)
![g, (U) - YfY QYI [g, (U) - YfJ
+ ![g2 (U) - Y~Y QyJg2 (U) - Y~]
+ ![U - UdYQu[U - U d ] (2.7)
 161

where the first two terms on the right hand side may be interpreted
as weighted least squares terms.
The Gauss-Newton method for solving (2.6) involves the iteration
(2.8)
starting from a specified initial value U o. The scalar IY..k ~ 0 is
chosen to ensure that G(Uk+I) ~ G(U k ) and the direction dk is
given by
dk = -Hk-1VG(U k) (2.9)

where Hk is obtained by ignoring the second derivative contri-

butions of gl and g2 with respect to U from the Hessian of G(U),
thus
Hk = N0cQYI N lk + N~QY2N2k + Qu (2.10)
where for i = 1, 2

Nk = ag, I = aYi I (2.11 )

' au U=Uk au U=Uk

As QYl' QY2 ~ 0 and Qu > 0 it can be concluded that Hk > o. With

VG(U k ) denoting the gradient of G(U) at Uk
VG(Uk ) = N0cQYl(gl(U) - Yf) + N~QY2(giu) - Y~)

+ Qu(U - U d ) (2.12)

and Hk > 0, it can be verified using (2.9) that dk is a descent

direction.
The significance of (2.10) and (2.12) is that the first terms on the
right of both these expressions may be computed in an overlay
structure independently from the second terms. Thus, each model,
its solution program, as well as the corresponding terms in (2.10)
and (2.12) can be located in different overlays thereby reducing the
amount of space required for solving (2.4) using (2.8).

3. Pooling and "Games" Between Rival Models

In policy optimization, information from rival models can be

pooled in two ways which, in spirit, are equivalent. Suppose, using
assumption (2.5), the endogenous variables from the rival models
162

are used to define a general endogenous variable

(3.1)
where

I«(l.J ,. 1, 2, for an n-dimensional

endogenous variable vector Y, with ex" denoting the weigth assigned

to the importance (relevance) attached to the j the member of Y,.
Let

= 1,2. (3.2)
An immediate choice is (I., = lex, where ex" i = 1, 2 are determined
using Granger and Newbold's (1977) criterion of minimizing the
pooled forecast error. The way in which (I., are chosen is not really
important in this section. They are, however, specified prior to any
evaluation of Yin (3.1). The choice of (I., may depend on econometric
considerations. Note that "the need to pool forecasts (i.e.,
(I., E (0, 1), i = 1, 2) is prima facie evidence of a failure (of each
model) to encompass"* its rivals (see Hendry, 1983). It is shown in
Section 7 (Corollary 7.10) that, even if a model encompasses its rival
in a narrow deterministic sense, a min-max approach to choosing
an alternative pooling operation (see (7.5), (7.9) below) is only
biased towards the model that encompasses its rival. t ex, can also be
specified to reflect the policy maker's confidence in each model. A
way that avoids the difficulty of specifying (I., in this way is depen-
dent on the iterative specification of the objective function. This is
discussed in Section 5 and involves the tailoring of the objective
function to the requirements of the policy maker. In the above

*"The encompassing principle is concerned with the ability of a model to be able

to account for the behaviour of relevant characteristics of others, or less
ambitiously, to explain the behaviour of relevant characteristics of other
models". (Mizon, 1984.)
t A min-max approach on its own would not, therefore, completely eliminate the
use of model 2. Econometric evidence would be decisive in ruling out model 2.
 163

characterisation a, is unrestricted. In reality, a, ?: 0 is imposed with

Ci l , + Ci 2, = a" where a, is some scalar which can clearly be nor-
malized to unity by redefining Ci; ~ Ci,/a. Since imposing these
restrictions is the more general case, we shall assume that
I a, = 1, a,?: 0, Vi. (3.3)

The pooling given by (3.1) suggests that, in effect, Y is the vector of

endogenous instruments and the policy optimization problem can
be written as
min {J(Y, U) IY = J(al)Y I + J(a 2 )Y 2 , F,(Y" U) = 0, i = 1,2}.
(3.4)
Using (3.1), the objective function can be written as
J(Y, U) H(Y - yd)T Qy(Y - y d ) + (U - Ud)T Qu(U - U d )]
t[(J(a l )(Y I - y d) + J(a 2 )(Y 2 _ yd))T

x Q/J(al)(Y I - y d) + J(a 2 )(Y 2 _ y d ))

+ (U - Ud)T Qu(U - U d )]
~ J(YI' Y 2 , U). (3.5)
The constrained minimization (3.4) can thus be rewritten as
min {J(YI' Y 2 , U) I F,(Y" U) = 0, i = 1, 2}, (3.6)
which is equivalent to (2.3), or as the unconstrained problem
min {G(U)}. The gradient and Hessian approximation, equivalent
to (2.1 0), of G(U) at Uk are given by

VG(U k ) = ,tI N,[/(a,)Qy Ctl J(a)(g;(U k ) - Yd))

+ Qu(U k - Ud)

ilk ItI N,[/(a,)Qy Ctl J(a)N;k) + Qu·

Clearly, VG and il are not as easily computable as (2.12) and (2.10)

respectively. Nevertheless, VI, NI can be evaluated independently
from Y2 , N2 and this results in considerable gain in storage on the
computer and the algorithm based on (2.8)-(2.10) can be utilized to
solve (3.4) be replacing va in (2.9) with va and H in (2.9)-(2.10)
with H.
164

The main difficulty with (3.6) is that (3.5) involves 1'; - }2 cross
terms (compare with (2.4». If these cross terms can be tolerated, an
even more general method for computing compromise solutions
which also does not require the specification of IXI and 1X2 is discussed
in Section 5. A simpler alternative to (3.4) is to consider the pooling
of performance measures (or objective functions) of the economy
under models I and 2. In a policy optimization framework, such as
a pooling has a number of interpretations.
Consider the vector minimization problem
min {J(YI' Y2 , U) I FI (Y I , U) = 0, F 2 (Y2 , U) O} (3.7)
where J is the two dimensional vector given by
J(Y" Y2, U) ~ [J, (Y" U), J 2 (Y 2 , U)f. (3.8)
Clearly, even if J, and J 2 are the same functions, the use of each
model produces Y, and Y2 which are different and this leads to
different objective functions (see (3.10) below). Thus, we consider
the general case with J" J 2 different. The minimization problem
(3.7) is a generalisation of the original minimization (2.3) with a
scalar objective function. Each element of (3.8) reflects a "pure"
strategy. J, (Y" U) stands for using Y, and U as the relevant policy
optimization variables. This implies that F, (Y" U) = 0 is the correct
model that reflects the behaviour of the economic system and that
it should be used in making policy decisions. Similarly, J 2 (Y 2 , U)
stands for using Y2 and U and implies that F 2 (Y2 , U) = 0 is correct.
For computational purposes, J, is assumed to be the quadratic J in
(2.4) with QY2 == 0 and Qu = QUI and J 2 is assumed to be (2.4) with
QYI == 0 and Qu == QU2·
Since the policy instrument (control) vector U has been assumed
to be common to both objective functions and models in (3.7), the
concept of a Nash solution (see Ho, 1970) is ruled out for (3.7). By
invoking Assumption (2.5), Y, and Y2 may be eliminated from
(3.7)-(3.8) to yield th~ vector minimization problem
min {G(U)IU E E nk }, (3.9)
where G(U) is given by

G(U) = [G,(U)] = J(g,(U), g2(U), U)

G2 (U)
(3.10)
 165

Thus the original vector minimization problem (3.7) is reduced to

(3.9) for which the concept of Pareto optimal solutions is defined as
follows:

DEFINITION: U* is defined to be a Pareto optimal or efficient solution

of (3.9) iff for any other U
(3.12)
for i = 1,2. Condition (3.12) can also be seen as the noninferiority
of U* over all other possible U values. D

A Pareto optimal solution for (3.9) may be obtained by solving

the scalar minimization problem
(3.13)
under the normality restriction (3.3) for the constants C{j, C{2 ~ 0;
C{j+ C{2 = 1. Clearly C{j and C{2 are "chosen before solving (3.13)
and different Pareto optimal solution values are obtained by vary-
ing C{j and C{2 over their permissible range
(3.14)
(see, e.g., Ho, 1970). Thus, if anyone of the objectives deviates from
Pareto optimality, then at least one objective, not necessarily the
one which deviated, does worse than the Pareto solution or all do
the same: they cannot all do better.
An important aspect of (3.13) is that it can be solved using the
Gauss-Newton algorithm described in Section 2 by simply setting
(3.15)
A solution of (3.13) is of interest in the context of the chosen
values of C{j and C{2. These values can be seen as weights indicating
the relative importance attached by the policy maker to each
objective function. In turn, the relative importance of the objective
function is a measure of the importance of the corresponding
model. The resulting Pareto optimal strategy can be considered as
the solution imposed by a decision maker upon his rival policy
advisers expounding the alternative theories on the economic system
offered by the models. In the next section a slightly different approach
that leads to a Nash strategy will be discussed. An alternative to the
above approach which eliminates the need for specifying C{j, C{2
166

is formulated in Section 7 as a min-max problem and in Section 5

a different approach to compromise strategies is discussed.
The difference between the policy optimization problems (3.6)
and (3.l3) is essentially the relative simplicity of the objective
function of(3.l3) which does not involve Y 1 - Y2 cross terms. This
can be an advantage in computation. Also, because of the special
structure of the quadratic form (3.5), it is difficult to see what would
be gained by adopting (3.6) instead of (3.l3). Nevertheless, a
quadratic form of general structure instead of (3.5), such as the
objective function resulting from (5.10) below, may have properties
which (3.15) might not. It can be shown, using the arguments of
Rustem and Velupillai (1984), that the set of constrained optima
(i.e., optimal policies) that can be characterised by a general quadratic
function is larger that the corresponding set for a quadratic function
such as (3.15) which does not possess Y 1 - Y 2 cross terms. This, in
turn, implies that the requirements of a policy maker might, in
certain circumstances, be better accommodated by using a general
quadratic form. However, an evaluation of (3.13) for various values
of ex have shown that the simpler objective function (3.15) is
reasonably adequate. Numerical results, using the H.M. Treasury
and the National Institute of Economic and Social Research
models of the U.K. economy as rival models, are discussed in
Becker et al. (1986) for (3.l3) evaluated with various values
of ex.

4. Nash and Stackelberg Strategies for Games Between Rival

Sectors or Countries

The recognition that in an economic system there may be more

than one agent trying to influence the general direction in which the
system ought to move, leads to various game theoretic strategies.
Each agent may have a separate aim, reflected mainly by his
objective function, and a separate view of the economic system
reflected by a model. Such situations may arise when the system
consists of the economies of several competing countries, or the
various competing sectors of a national economy (see, e.g., Miller,
1984; Pau, 1975). One such strategy, in which the overall policy is
improved by a Pareto optimal solution is discussed in Section 3
above. In this section other strategies are considered.
 167

In the following discussion, each sector or country is represented

by a model and an objective function. The game is considered to
occur between the models representing them. Assume that there are
L competing sectors or countries in an economic system. The
models for each sector may have common endogenous variables
(outputs) plus some additional endogenous variables particular to
the model of that sector or country. In this case the endogenous
variable of each sector cannot be assumed to be different from those
of the other sectors. We therefore define Yas the vector of endogen-
ous variables of all the sectors. Clearly, some sectors may not use
some elements of Y. The policy instruments (controls) of each
sector are assumed to be different for each sector. Thus, the aggregate
objective function for sector i is denoted by
J,(Y, U, .... , U" .... ,UL ), i = 1, .... , L. (4.1)
where U, is the vector of policy instruments (controls) of i. Let
F(Y, U], ... , U" ... ,UL ) = 0 (4.2)
denote the model equations of all the sectors. Two points about
(4.2) need to be clarified. The first is that the endogenous values of
one sector may be used in another sector which would otherwise
have used exogenous assumptions instead of these values. The
second is the possibility of an element of Y appearing as the
endogenous variable of more than one model. There are no general
solutions to these problems. Whether a new variable is to be defined
for each occurrence of this variable or some equations need to be
altered is a matter to be decided for each particular problem. In the
former case, however, the elimination of an exogenisation would
lead to a more realistic representation.

ASSUMPTION 4.3: Given the policy instrument vectors U" i = I,

... , L, there exists a mapping g such that

o (4.4)

Clearly, assumption (4.3) may be used to eliminate Y from each

objective function in (4.1) to yield
G,(U] , ... ,UL ) = J,(g(U], ... , U L ), U b ... , U L );
i = I, ... ,L (4.5)
168

Each sector or country can minimize its objective function with

respect to the policy instruments (controls) allowed to it. This is
basically the Nash equilibrium solution defined as follows:

DEFINITION 4.4: A Nash solution V*, i = 1, ... , L is defined by

G,(V*, ... , V;t', ... , Vi)
:::; G;(Vj", ... , V;t'-b V" Vi+], ... , Vi) "tV,
for i = 1, ... , L. D
Thus, the Nash strategy has the property that if all but one player
use their Nash strategies, the deviating player could not decrease his
or her objective function. We consider open-loop strategies where,
in the dynamic case, only the initial state of the system and, at each
instance t, time t is known to the players. In Section 8 it is argued
that a sequentially updated open-loop strategy is an acceptable
approach in applications with large nonlinear econometric models.
The open-loop Nash solution is obtained by minimizing G;(.)
with respect to Vi' jointly for all i, i = 1, ... , L. This leads to the
first order necessary conditions for optimality
aG,(v], ... ,V" ... ,VL ) = 0
i = 1, ... , L. (4.5)
av, '
An algorithm for computing Nash solutions satisfying (4.5) is
discussed in Rustem and Zarrop (1981). This framework can also
be extended to closed-loop solutions using the parameterised feed-
back approach in Karakitsos and Rustem (1984, 1985).
The concept of a Stackelberg strategy arises when some i in
(4.1)-(4.2) may have dominant (i.e., leading) positions. The rest of
the player(s) follow the leader(s). The leader is aware of the objective
function mapping of the follower but the follower may not know
the objective function mapping of the leader, however, the follower
knows the control strategy of the leader and takes this into account
(see, e.g., Cruz, 1975). In the two player case (i.e., one leader and
one follower) let 0/1],0/12 be two sets over which the control strategies
of the leader (1) and the follower (2) are defined and let G] (V]' V 2 )
be the objective function of the leader and GlV] , V 2 ) be the objective
function of the follower. Consider the set valued mapping T
 169

such that
TV I = {V21V2 = arg inf [G 2(V I , O2) I O2 E 0Ji2]}·
and TV, = ¢ if the infimum is not achieved. The leader's decision
is determined by the optimization problem
inf {GI(V I , V 2 ) IV, E 0Ji1, V 2 E 0Ji2 } (4.6)
where GI (VI' V 2 ) = + 00 if V 2 E TV, = ¢. (4.7)
DEFINITION 4.7: A pair (Ur, Vn E 0Ji x 1 0Ji2 is a Stackelberg
equilibrium pair if (Vr, Vn solves (4.6). 0
Thus, in a Stackelberg equilibrium the leader chooses V I E 0Ji1 with
cost GI and the follower chooses V 2 E 0Ji2 with cost G2 •
As in the case of Nash strategies, we consider open-loop Stackel-
berg strategies which are sequentially updated with incoming new
information. A number of particular cases of interest arise accord-
ing to the relationship between the leader(s) and the follower(s) and
these are summarised below.

4.1. One leader and two followers

Consider the case with one leader and two followers. The followers
play an open-loop Nash strategy between themselves. For i = 1 the
leader and i = 2, 3 the followers, the necessary conditions for an
open-loop Stackelberg-Nash strategy can be derived considering
that G 2 (V I , V z, V 3) is minimized with respect to V 2 by player 2,
given VI' V 3, G 3(V I , V z, V 3) is minimized by player 3 with respect
to V 3, given VI> V 2, and the leader minimizes G I (VI' V 2 , V 3) with
respect to VI' V 2 , V3 subject to the knowledge that GI and G2 are
also being minimized. This leads to the following conditions which
have to be solved simultaneously:
aG 2 (V I , V 2 , V 3)
av z o (4.8)

aG 3(V I , V 2, V 3 ) = 0
(4.9)
aV 3

(4.10)
170

The equalities in (4.10) are the necessary conditions for the optimality
of the problem

. lfG (V], V
mIll I 2,
aG 2 = 0, aV
V 3 ) I aV 2
aG3 = 0 }
3
(4.11)

and ),1' ),2 are the Lagrange multipliers associated with (4.8) and
(4.9) respectively.

4.2. Two leaders and one follower

Consider the mixed strategy when there are two leaders playing
an open-loop Nash strategy between themselves and jointly an
open-loop Stackelberg with one follower. For i = 1, 2 the leaders
and i = 3 the follower, the necessary conditions for this Stackelberg-
Nash strategy can be written as:
aG3 (V I , V 2 , V 3 ) = 0
aV3

~
av
[G'(V I,
V
2,
V)
3
),T aG
+,
3 (V I ,
av
V 2 , V 3 )] 0
) 3

for i = 1, 2 and j = i or 3.

4.3. Player 1 leading player 2 leading player 3

The necessary conditions for the hierarchical open-loop case in
which players 1 leads player 2 who, in turn, leads player 3 can be
written as:

= 0; i = 2, 3

a {
av, G I
~
+ )'='2 T
),)
a [
av) G 2 + ),1
T aG3 ]}
aV 3 0; i = 1, 2, 3.

An important computational difficulty in all the Stackelberg

strategies discussed above is the requirement of second derivatives
of the follower(s)' objective functions even in the first order necessary
 171

conditions. Furthermore, the last case shows that the deeper the
hierarchy of the leaders gets, the higher is the required order of
derivatives of the followers' objective functions.

5. The Relative Importance of Rival Models:

Compromise Solutions

Returning to the discussion of Section 3, in this section the

importance of selecting 0:\ and 0: 2 in (3.5) or (3.13) is considered. It
is clear that, in the absence of any help from econometric analysis,
the specification of 0:\ and 0: 2 is a difficult problem. In this section,
a method is suggested that avoids the specification of 0:\, 0: 2 by
tailoring the joint objective function (2.4) so that the optimal policy
solution of (2.3) is acceptable on all models.
Consider first the optimal policy decision based on the assumption
that the first model is correct (i.e., represents the economic system)
and that the second model is wrong. In this case the policy opti-
mization problem (2.3) reduces to
(5.1)
where, as in (3.8), J\ is the quadratic function (2.4) with Qu =
Q\u, QI'2 == o. Similarly, if the second model is correct and the first
model is wrong, the policy optimization problem reduces to
(5.2)
where, as in (3.8), J 2 is the quadratic function (2.4) with Qu = Q2u
and QYl == o. Allowing both J\ and J 2 to have the same desired
policy, U d , values is justified because Qlu and Q2u may be adjusted
to express the relative importance of the policy instruments in (5.1)
and (5.2).
Using the method described in Rustem and Velupillai (1984), the
objective functions J 1 (Y\, U) and J 2 (Y 2 , U) may be selected to
provide optimal solutions for (5.1) and (5.2) that are acceptable for
the policy maker. In each case the acceptability of the optimal
solutions is decided on the assumption that only the model used in
the optimization is correct. This implies complete acceptance of the
theory underlying the specification of each model. Thus, the solutions
of (5.1) and (5.2) reflect pure strategies (based on complete accept-
ance of a particular theory) for controlling the economy. However,
172

these strategies are nonunique because of the lack of knowledge

about the true objective functions of the policy maker. The selection
of , I and '2 with the method in Rustem and Velupillai (1984) gives
quadratic functions that are approximations to the true objective
functions. These approximations are selected to yield optimal
solutions for (5.1) and (5.2) that are also acceptable to the policy
maker. Neither 'Iand '2
are unique approximations, nor the
solutions of (5.1) and (5.2) are unique in their acceptability. How-
ever, it is assumed that the policy maker is indifferent between
acceptable solutions.
Consider next the original policy optimization problem (2.3).
(2.4) may be specified initially using ' I (Y I , U), '2(Y 2, U) selected
above and two nonnegative scalars (XI' (X2. Thus

Using Assumption (2.5), (5.3) may be reduced to

G(U) = (XI G I (U) + (XzG 2 (U); (XI' (X2 ~ o. (5.5)

The fact that (XI' (X2 are no longer restricted by (XI + (X2 = 1, is not
a problem since once the values for (XI and (X2 are known G(U)'(XI
and (X2 may be redefined as

G'(U) = (5.6)

Clearly the position of the optimum solution is not altered by this

redefinition.
It can be seen from (5.4) that increasing (XI relative to (X2 will
increase the importance of YI attaining the desired values Yf relative
to Y2 attaining Y~. The importance of U attaining U d will also be
reduced (unless (X2 = 0 in which case' = (XI'I). The same argument
also applies to (X2.
 173

The scalars 11\ and 112 in (5.3) are selected such that the solution
of (2.3) is acceptable to the policy maker. The acceptability of the
optimal solution is now based on the broader assumption that both
models may be correct in varying degrees. Hence, the optimal
solution may capture the combined effects of both models. Thus, by
adjusting 11\ and 112 an acceptable solution to (2.3) may be sought.
An alternative approach which does not involve 11\, 112 is to seek
an acceptable solution to (2.3) by determining a joint objective
function like (2.4) for both models. This approach determines the
weighting matrices QYI' QY2' Qu in (2.4). The corresponding matrices
in (5.3) are 11\ QYI' 11 2 QY2' (11\ + 112 )Qu respectively. The scalars 11\,
112 in (5.3) can be considered to have been absorbed in the matrices
in (2.4). The approach, discussed below, determines the weighting
matrices in (2.4) that yield an acceptable solution of (2.3) on both
models and is an extension of the method discussed in Rustem and
Velupillai (1984).

STEP 0: Given Yf, Y~, U d , start with a basic set of initial weights for
Qv. I , Qv"
.
Qu' In the absence of any information we can assume
~

QYI = J, QY2 = J, Qu = I. Set the current weighting matrix to be

QYI I 0 0
- - 1 !

QC~ 0 0 (5.7)
I I

QV2
- - i i
0 I 0 Qu

STEP 1: Solve the optimization problem (2.3) to obtain a "current"

optimal solution,

r Y~'l
Yc2
U C
• (5.8)

STEP 2: Ask the policy maker whether (5.8) is acceptable. Stop if it

is acceptable since QC represents the weights that reflect the policy
maker's relative emphasis on either model. If, on the other hand,
the current solution is not acceptable, then the policy maker is
174

asked to specify the correction, b, to the current solution that would

make it acceptable. Thus,

[ ~:l
UC
+ b
(5.9)

is, by definition, acceptable but not necessarily feasible with respect

to F, and F 2 •

STEP 3: Given b, compute the new weighting matrix

QCbbT QC
Qn = QC + J1 bTQcb (5.10)

where J1 ~ 0 is some scalar reflecting the emphasis to be given to

the rank-one term in (5.10). Replace QC by Qn and go to Step l.

It should be noted that (5.10) does not preserve the block diagonal
structure of (5.7) and this leads to a slightly more complicated
formulation of the Guass-Newton algorithm in Section 2. Thus, the
matrices obtained using (5.10) include general ~ - 1'; cross terms.
This yields a policy optimization formulation which is more general
than (3.5) and at no particularly greater computational complexity.
It can be shown that the above method ensures the acceptability of
the solution of (2.3). It is not possible to measure iX" iX2 in this
approach. The objective function is tailored to bring out the com-
bined effects of both models in an acceptable optimal solution.
It can be shown that the above method ensures the acceptability
of the solution (2.3). The following proposition indicates a desirable
characteristic of the method towards this end. Let, for simplicity,
the two models be linear and be given by

where b is a constant vector and N is the Jacobian matrix of the

model. Assume, again for simplicity, that the columns of N are
linearly independent.
 175

PROPOSITION 5.11: Let QC be positive definite. Then for b i= 0 and

Il :;:: 0, Qn given by (5.10) is positive definite and the new optimal
solution obtained using Qn in (2.3) is given by

where

furthermore,

PROOF: Rustem and Velupillai (1984, Theorem 1).

o
Extensions of this result to positive semidefinite Qc and nonlinear
models, along with the desirable characteristics of the method
arising from this result, are also given in the above reference. The
matrix P is an operator that projects general vectors onto the
feasible space defined by the models. This indicates that each stage
of the above method yields a solution that is the best feasible
alternative, given the models, to the policy maker's unrestricted
requirements in (5.9).
As IXI and 1X2 may also be seen as the expected probabilities of FI
and F2 respectively representing the economy, the policy maker is
allowed to insure himself against policy failure by specifying (2.4)
to obtain a solution of (2.3) that steers a mid-course between the
two models. This is illustrated in the following example:

EXAMPLE: Let model 1 be given by (5.12)

YI = U + 81
176

and model 2 be given by

Y2 = U + 82
where u is the policy instrument, y, ,Y2 and 8" 82 are respectively the
endogenous variables and parameters of the models. Let the prob-
ability of model 1 representing the economy be p and let the
probability of model 2 be (1 - p). Let the objective function be
given by
G(u, p) = p(u + 8,)2 + (1 - p)(u + ( 2)2.
The policy instrument that minimizes G, given p, is
u* = - (p8, + (1 - P)82)
with objective function value
G(u*, p) = p(1 - p)(8, - ( 2)2.
When the models are nonlinear, the optimal solution is more
complex. Nevertheless, u* remains a function of p. Clearly, p
reflects the relative emphasis on models 1 and 2 and the problem is
well defined if p is known. An alternative approach that avoids the
specification of this relative emphasis is formulated in Section 7 as
a min-max problem.

6. Deterministic Solutions to Uncertainty in a Single Model

Given a stochastic model of the economic system Fr(Y, U, e;) = 0,

where ei is the vector of random disturbances, a straightforward
formulation of the policy optimization problem is
min EPr(Y, U) I Fi(Y, U, e,) = O} (6.1)
where E denotes the expectation operator. When F is large and
nonlinear, the heavy computational requirements of algorithms for
solving (6.1) (see Kushner and Clark, 1978) lead to alternative
formulations of the problem of uncertainty in a system.
The first such formulation is
min max {Jr(Y, U) I Fr(Y, U, e;) = O} (6.2)
Y,U e,El:,

where er is taken to be a deterministic variable and ~i the space over

which er is defined and {Jr(Y, U) IFi(Y, U, eJ = 0 is bounded from
 177

above on ~. In this formulation ", is takem to be a deterministic

vector of variables and the worst case design problem (6.3) is the
minimization of the maximal effect of this vector. The vector ", is
no longer taken to be random but as just a vector of possible sources
of error. The optimal policy is thus determined assuming ", will take
its most adverse value. In a stochastic setting, (6.2) would be
formulated as a maximization over the probability distribution of
"" rather than the values of e,. This would eliminate the need to
specify these distributions which would, for example, be needed to
solve (6.1). Given the size and nonlinearity of the models, however,
(6.2) provides a convenient deterministic formulation.
In the presence of rival models, the minimization of the maxi-
mum damage that a policy based on the wrong model may cause,
can be formulated as a discrete min-max problem. Consider L rival
models
F/(Y, U, Il,) = 0, i = 1, ... , L (6.3)
where Y, U can be considered respectively as augmented vectors of
the endogenous variables and the policy instruments in all the
models and ", is the vector of disturbances in the i th rival model.
The augmentation of the endogenous variables yields a more
general structure in which the endogenous variables of each model
may affect another model. If the endogenous variables of each
model do not affect other models, the problem of simultaneity
between models implied in (6.3) becomes slightly simplified in that
each model may be solved individually, given U and "" This for-
mulation is considered in Section 2. In the min-max problems
considered in Section 7, it is proposed to resolve the problem of
uncertainty in each model with an approximate-robust policy-
formulation and the uncertainty between models with a min-max
strategy. The objective function corresponding to model i is given
by ',(Y, U).
A second alternative solution to the problem of uncertainty in
each model is the deterministic "robust" policy approach minimizing
the sensitivity of " to departures of ", from its assumed mean value
", = O. This can be formulated as

min {Q,J,(Y, U) + (1 - Q,)J~T A,J: I F,(Y, U, 0) = O} (6.4)

where the scalar Q" 0 ::;;; Q, ::;;; 1, is chosen to reflect the emphasis to
be given to either term in the objective function in (7.2) and is
178

specified prior to the minimization. The vector J: is defined as

J" ~ aJ,(Y, U) I = ay aJ, I
' at, 8,~O at, ay 8,~O

is the sensitivity of J, to t" evaluated at t, = 0 and the weighting

matrix A, reflects the relative emphasis given to each sensitivity. In
Karakitsos, Rustem and Zarrop (1981) it is shown that the value of
A, has a natural choice when the problem is re-examined from a risk
aversion point of view.

7. Rival Models: Continuous and Discrete Min-Max Strategies

The robust control approach of Section 6 has a straightforward

application in the presence of rival models. Thus, (6.4) can be used
to account for the uncertainties in each model in the worst case
design problem
min max {Q,J,(Y, U)
Y,U ,
+ (1 - Q,)J: T A,J:

IF,(Y, U, 0) = 0, i = 1, ... , L, 0 ~ £I, ~ I}. (7.1)

The discrete min-max problem (7.1) can be solved using nonlinear
optimization techniques and minimizes the maximum damage a
policy can cause on any model of the economy. Hence the damage
of the policy to the economic system is minimized in ignorance of
the particular model representing the economy. However, the
economic system is assumed to be represented by one of the L
models. This is useful in cases when the particular regime under
which the economy is operating is not known but models of all
possible regimes are available.
As an alternative min-max approach, consider the case discussed
in Sections 3 and 5 where the relative emphasis to be placed on each
model has to be specified. One way of avoiding the specification of
this relative emphasis is to evaluate the maximum damage a wrong
emphasis could cause. This is illustrated in the following problem.

EXAMPLE: Consider example (5.12). At the optimal solution,

G(u*, p) = p(1 - p)(8, - 82)2 is maximum for p = t. Whenp is
given, u* is the optimal solution. If p is not available, consider the
 179

strategy that makes G independent of p. Namely, consider

u = -H8 1 + 82 )
and G(u, 0) = G(u, 1) = H8 1 - 82 )2. Thus, for p = t, u minimizes
the worst possible case, i.e., the maximal G value.
A straightforward extension of this approach is given by the
problem

~Ln mx~x ttl IX,J,(Y, U) IF,(Y, U) 0, IX, ~ 0,

i = 1, ... , L, ± I}.
,~l
IX, = (7.3)

This approach is similar to the selection of "randomized" strategies

(see Luenberger, 1969). Problem (7.3) minimizes L IX,J, with respect
to Y, U while choosing the emphasis on models (i.e., IX,) which cause
the maxinal loss in the objective functions. This is a worst case
design problem which ensures that the economy is not going to be
any worse (and, in general, better) than the solution of (7.3)
provided all the L models used form an exhaustive set of possible
descriptions of the system.
The stochastic case is not considered for (7.3) because of the
computational difficulties involved. In the presence of uncertainty
in each model, the "robust" objective function formulation in (6.4)
may be used instead of J, in (7.3) to generate robust policies.
Finally, in the robust case, (7.3) can be extended to minimizing the
maximum robust objective over Qi. This leads to

~Ln ~~~ {,~ IX,[Q,J,(Y, U) + (1 - Q,)J~T A,J~] IF,(Y, U, 0) = 0,

IX, ~ 0, ~ IX, = 1, 1 ~ Q, ~ 0, i = I, ... , L} . (7.4)

Thus the worst case design problem is extended to the choice of Q,

that maximizes the contribution of either term in the robust objective
function in each model. Hence the problem of a priori choice of Q,
in the original robust policy formulation (6.4) may be resolved by
employing a min-max strategy. In the simple case when there is only
one model, (7.4) implies a minimization of the maximum contribu-
tion from the objective function and its sensitivity to 1:.
180

The min-max formulation (7.3) has a particularly simple structure

in the variables ai' chosen to maximize "i.i aJ,. It is also interesting
in that the optimal values of a, ensure that the policy maker does not
have to worry which model represents the economy. The following
Proposition illuminates these points. Suppose that, using Assump-
tion (2.5) we have eliminated the endogenous variables, and hence
the models, from (7.3) and that, for simplicity, we have two models.
The discussion below extends trivially to more than two models.
Problem (7.3) can thus be written as

min max {aG] (U) + (1 - a)G 2 (U) I a ~ 0, 1 - a ~ O}. (7.5)

U 7

PROPOSITION 7.6: Suppose Gi(U) are twice differentiable with

respect to U at a solution of (7.5) and that strict complementarity
holds for 1 ~ a ~ 0 at this solution. Let a*, U* solve (7.5)
then a* E (0, 1) iff G] (U*) = G 2 (U*), a* = 1(1 - a* = 0) iff
G] (U*) > Gz(U*) and a* = 0(1 - a* = 1) iff G] (U*) < Gz(U*).

PROOF: The necessary conditions of optimality for (7.5) are

a*VG](U*) + (1 - a*)VG 2 (U*) 0 (7.7a)

G] (U*) - G 2 (U*) + A] - A2 0 (7.7b)

a* ~ 0, (1 - a*) ~ 0 (7.7c)

a* A] 0 (7.7d)

(1 - a*)22 0 (7.7e)

A], A2 ~ 0 (7.7f)

where A], A2 are the Lagrange multipliers of the constraints. When

a* = 0(1 - a* = 1) we have, by (7.7d), A] > o. Multiplying
(7.7b) by 1 - a*( = 1) and using (7.7e) yields

G] (U*) - G2 (U*) = - A] < 0

hence
 181

To show the only if condition, let G 1 (U*) < G2 (U*). Mutliplying

(7.7b) by a* and using (7.7d) and (7.7c) yields
a*[G 1 (U*) - G 2 (U*)] ~* A2
A2
~ 0
Since G1 (U*) < G 2 (U*), we have A2 = 0, a* = 0(1 - a* = 1).
The proof for a* = 1 (1 - a* = 0) is analogous to that of
a* = 0 (1 - a* = 1) given above.
To show that a*, (l - a*) E (0, 1) iff G 1(U*) = GiU*) consider
(7.7b). We have
Al = A2
and using (7.7d), (7.7e)
o = a*Al = a*A2 = A2.
As A2 = )'1 = 0, strict complementarity implies that a*, (1 - a*) E
(0, 1). To show the only if part let a*, (1 - a*) E (0, 1). By com-
plementarity and (7.7d) and (7.7e),
a* Al = (1 - a*)A2 = o.
We have )'1 = A2 = O. Hence, by (7.7b)
G1 (U*) G2 (U*). o
Let
R, ~ {(Y, U) I F,(Y, U) = O}
denote the feasible region denoted by model i. Assuming that the
endogenous variable vector is common to all models (but each
model, i, defines a different value for y), a narrow, deterministic,
definition of encompassing can be given as
(7.8)
This implies that the feasible region of model 1 has at least all the
characterisations of the economic system that model 2 has. this is
also a narrow, deterministic, description of nesting of models within
each other. The following corollary states that if model 1 encom-
passes model 2, then the pooling of models in the sense of (3.13)
182

is biased towards model 1 in a min max strategy. The argument can

again be trivially extended to more than two models.

COROLLARY 7.10: Let J[(Y, U) in (7.3) be such that J 1 (Y, U) =

J 2 (Y, U). Ifmodell encompasses model 2 in the sense of(7.8), then
a* E (0, 1]. (7.10)

PROOF: Using Assumption (2.5) and (7.8) we have

min {G 1(U)} ~ min {G2 (U)}. (7.11)
U U

Note that, although J 1 and J 2 are the same functions, each model
implies a different value for the corresponding reduced cost function
°
Gp i = 1, 2. Assume that a* = then (7. 7a) represents the necessary
condition for a minimum of G2 (U) at U*. Thus, we have
min {G2 (U)} = G2 (U*) > G1(U*) ~ min {G 1(U)}
U U

which contradicts (7.11). hence a* -=I- 0. o

Only when a* = 1, the solution of the min-max problem, as
might intuitively be expected, occurs at the minimum of G1(U). This
can be verified by setting a* = 1 in (7.7a) and using (7.11) to yield
min {G1(U)} = G1(U*) > G2 (U*) ~ min {Giu)}.
u u
As model I encompasses model 2 in the sense of (7.8), one might
thus expect the solution to depend only on model 1. However,
values of a* which might yield G1(U*) = G2 (U*) may exist with
a* E (0, 1). Corollary (7.10) only establishes that the min-max
strategy will never be based only on model 2. In the case of
a* E (0, 1), the combined objective function in (3.13) clearly satisfies
a*G1(U*) + (1 - a*)G 2 (U*) ~ min {G1(U)}
u

since, otherwise a* = l. From Proposition (7.6) we have, for

a* E (0, 1), G, (U*) = G2 (U*). This yields
G1(U*) = G2 (U*) ~ min {G1(U)}.
u
Thus, the min-max strategy is able to find a* which makes the
individual G[(U*), i = 1, 2, take more adverse values than the
minimum of G1(U).
 183

Finally, the above discussion shows that the min max strategy is
not a simple pooling process, despite earlier suggestions in the
paper. If 0:* = 1 (or 0), model 1 (model 2) is taken to be the basis
of policy optimization. If 0:* E (0, l), then the optimal policy, U*,
is selected such that the policy maker's cost (or objective) function
is the same whether model 1 or model 2 turns out to represent the
economy.
In Becker et al. (1986) an approximate method for solving (7.3)
is discussed. This consists of evaluating (3.13) for various values of
0: (i.e., 0: = 0, 0.25, 0.5, 0.75, 1.0). As 0: changes, so do the values
of G1 (U) and G2 (U) at the corresponding optimal solution of(3.13).
Using curve fitting to determine the behaviour of G1 and G2 as a
function of 0:, at the solutions of (3.13), the rx corresponding to
G1 = G2 is obtained by invoking Proposition 7.6. The two models
involved being the National Institute of Economic and Social
Research and the H.M. Treasury models of the U.K. economy, the
value of 0: solving the min-max problem (7.3) was found to be 0.6
(where 0: = 1 implies total belief in the Treasury Model). In the
context of the above Corollary, the Treasury Model, which is
somewhat more complex, is built along views similar to those
reflected by the NIESR model. Thus, even if the former model can
be considered to encompass the latter in the sense of (7.8) (i.e., R1
denotes the Treasury Model), the optimal solution of 0: indicates
that the solution of (7.3), as predicted by the Corollary, is in the
range (0, 1) and not exactly at 0: = 1.

8. Feedback Laws and Sequential Updating of Open-Loop

Optimal Policies

If (2.3) is viewed as an optimization problem over a period of

time, at the beginning of this period J, F1 and F2 are fixed and (2.3)
is solved given the exogenous assumptions for this period. How-
ever, as time passes, the computed optimal values may differ from
the actual values observed from the economic system. This may be
caused (i) by changes in the exogenous assumptions, (ii) by the
policy maker emphasizing the wrong model in selecting his objective
function, (iii) by the system itself having shifted from the regime
represented by F 2, or vice versa. The effects of the first possibility
may be accounted for using the sequential updating technique
184

described by Athans et al. (1976). This involves the updating of the

exogenous assumptions and recomputing the optimal solution for
the remaining period. By this we mean that optimal open-loop
policies are calculated for the whole optimization period on the
basis of the information available at the beginning of the opti-
mization. Policy is applied for the first time period and then
calculations are repeated for the second time period onwards after
integrating the updated information which has become available.
This process is repeated. At each revision period the optimal policy
is recalculated over a rolling horizon and thus the long-run dynamics
of the system can be taken into account. The sequential updating of
open-loop optimal policies arises, in part, because the question of
selecting an optimal feedback law for nonlinear models is, in
general, not answerable. Further, given the need, in practical policy-
making to frequently update exogenous information and reassess
priorities, it can be argued that a sequential open-loop approach is
preferable to any feedback law. In Westcott et at. (1981) this
approach has been used to test policies when Monte Carlo simu-
lations are utilized to simulate the errors in the incoming infor-
mation. The expectations of the random variables discussed in
Section 6 can also be updated using incoming new information.
The second possibility, involving the wrong emphasis on a rival
model, can be resolved by modifying the objective function after
updating the exogenous assumptions. This also applies to the third
possibility. A detailed discussion of detecting changes in the regime
of the economy and sequential likelihood ratio tests for structural
change are given in Rustem and Velupillai (1979).
Further to the sequential updating of optimal strategies, in the
case of the various game strategies discussed in Sections 3 and 4, it
may be desirable for each agent to follow a prescribed feedback
strategy. Indeed, following such strategies may be a condition of a
bargaining process preceding the game. In these cases, a par-
ameterized feedback law such as

I = 1, ... , L
can be obtained within the game. The parameter vector (J" i = 1,
... , L can be determined to make the game strategy optimal.
In the single agent case; the determination of such optimal
 185

parameterized feedback laws are discussed m Karakitsos and

Rustem (1984, 1985).

9. Concluding Remarks

The optimization framework developed for policy optimization

problems with rival models is computationally feasible. Further-
more, (2.8)-(2.10) can be efficiently computed. Such an optimization
approach seems to be equivalent to a Pareto optimal solution given
by (3.13). The relative importance of the rival models may also be
measured by the weights attached to attaining the desired endogen-
ous values in each model. The solutions to Nash and Stackelberg
strategies, on the other hand, provide a framework for games
between sectors in an economy or between countries. The relative
importance of each sector or country may, in this case, be decided
by the models and the adoption of Nash or Stackelberg strategies.
The joint optimization framework (2.3) may be considered to be
too demanding on the optimal policy instrument values. The key
justification of (2.3) is the correct specification of the objective
function. The objective function is specified using the method in
Rustem and Velupillai (1984) such that the solution of (2.3) is
acceptable to the policymaker.
The computational difficulties in solving min max problems are
considerable. The min max problems considered in Sections 6 and
7 are designed to be as simple as possible in order to be applicable
to large nonlinear macro-economic models.

Acknowledgement

The author's research has been supported by a SERC Advanced Fellowship.

References
Athans, M et al. (1976): "Sequential Open-Loop Control of a Nonlinear Macro-
economic Model", in M. D. Intriligator (ed.), Frontiers of Quantitative Econ-
omics, North-Holland, Amsterdam.
Becker, R. G., Dwolatzky, B., Karakitsos, E., and Rustem, B. (1986): "The
Simultaneous use of Rival Models in Policy Optimization", The Economic
Journal, 96, 425-448.
Chow, G. C. (1979): "Effective Use of Economic Models in Macroeconomic
Policy Formulation", in S. Holly, B. Rustem, and M. Zarrop (eds.), Optimal
Control for Econometric Models, Macmillan, London.
186

Cruz, Jr., J. b. (1975): "A Survey of Nash and Stackelberg Strategies in Dynamic
Games", Annals of Social and Economic Measurement, 4(2, pp. 339-344.
Davidson, R. and MacKinnon, G. (1981): "Several Tests for Model Specification
in the Presence of Alternative Hypotheses", Econometrica 49, 781-793.
Granger, C. W. J. and Newbold, P. (1977): Forecasting Economic Time Series,
Academic Press, New York.
Hendry, D. (1983): "Econometric Evaluation of Linear Macro-Econometric
Models", Nuffield College, Oxford.
Hoel, P. G. (1947): "On the Choice of Forecasting Formulas", JASA, 605--611.
Ho, Y. C. (1970): "Differential Games, Dynamic Optimization and Generalized
Control Theory", JOTA 6 179-209.
Karakitsos, E. and Rustem, B. (1985): Optimal Fixed Rules and Simple Feedback
Laws in the Design of Economic Policy. Automatica, 21, 169-180.
Karakitsos, E. and Rustem, B. (1984): "Optimally Derived Fixed Rules and
Indicators", Journal of Economic Dynamics and Control, 8, 33-64.
Kushner, H. and Clark, D. S. (1978): Stochastic Approximation Methodsfor Con-
strained and Unconstrained Systems, Springer-Verlag, New York.
Luenberger, D. (1969): Optimization by Vector Space Methods, J. Wiley, New
York.
Miller, M. (1984): "Interdependence and Policy Coordination", paper presented
at the Thirteenth Money Study Group Conference, Brasenose College, Oxford.
Mizon, G. E. (1984): "The Encompassing Approach in Econometrics", University
of Southampton, paper presented at CEPR Workshop on the Evaluation of
Econometric Models.
Pau, L. F. (1975: "A Differential Game Among Sectors in a Macroeconomy",
Automatica 11,473-485.
Rustem, B. (1981): Projection Methods in Constrained Optimization and Appli-
cations to Optimal Policy Decisions, Springer-Verlag, Berlin.
Rustem, B. and Zarrop, M. (1981): "A Newton-type Algorithm for a Class of
N-Player Dynamic Games Using Nonlinear Econometric Models", in Janssen,
Pau, Straszak (eds.), Dynamic Modelling and Control of National Economies,
Pergamon Press, Oxford.
Rustem, B. and Zarrop, M. B. (1979): "A Newton-Type Method for the Control
of Nonlinear Econometric Models", JEDC 1, 283-300.
Rustem, B. and Zarrop, M. B., (1981): "A Class of Quasi-Newton Algorithms for
the Control of Nonlinear Econometric Models", Large Scale Systems 2,
105-111.
Rustem, B. and Velupillai, K. (1984): "On the Formalization of Political Prefer-
ences: A Contribution to the Frischian Scheme", European University Institute,
Discussion Paper 69, Florence.
Rustem, B. and Velupillai (1979): "On the Definition and Detection of Structural
Change", in M. Kohlmann and W. Vogel (eds.), Stochastic Control and Theory
and Stochastic Differential Systems, Springer-Verlag, Berlin.
Westcott, J. H., Karakitsos, E., Rustem, B., and Becker, R. (1981): "Memorandum
of Evidence on Monetary Policy to the Select Committee on the Treasury and
Civil Service", House of Commons, Third Report from the Treasury and Civil
Service Committee (/980-81): Monetary Policy, HMSO.
 187

CHAPTER 12

OPTIMAL POLICY DESIGN IN INTERDEPENDENT

ECONOMIES

A. J. Hughes Hallett*
University of Newcastle, UK

1. Introduction

In a world of interdependent economies, rational policy makers

in one country must condition their actions on the policies to be
expected in other countries and expect others to do the same. Policy
selection inevitably acquires the form of a dynamic game. The
persistence of world recession since 1974, and the debt crisis which
appeared in its wake, has certainly made policy makers more con-
scious of the links between their economies and that mutual depen-
dence through trade and capital movements means that the avail-
able policy choices are also interdependent. In fact, as the current
uneven recovery shows, de centralised control of the world economy
may effectively limit our ability to steer individual economies.
Economists and policy makers often call for concerted action to
overcome economic difficulties at both the world and individual
country levels; but there is, as yet, virtually no empirical evidence
on the costs of uncoordinated policy changes or the potential gains
from explicit cooperation. This paper reports such results for two
major economies-the US and the EEC-which are interdepen-
dent. It asks, given optimal selections, how much would these
economies gain if their policy makers cooperated or if they allowed
for interactions in a rational but noncooperative manner? It is
perhaps natural to suppose policy makers would take a nationalist
view, since most difficulties will appear to come from abroad. For
*This research was undertaken, in part, on behalf of the EEC Commission in
Brussels, and their support is gratefully acknowledged. However the opinions
expressed here are my own, and do not necesarily reflect the policy of the EEC
Commission.

C. Carraro and D. Sartore (eds). Developments of Control Theory for Economic Analysis
© 1987 Martinus Nijhoff Publishers (Kluwer), Dordrecht
188

example, the average OEeD country faced falling exports in

1981-2. Many countries have found their monetary policies largely
dictated from abroad through foreign monetary and interest rate
policies. Budget reductions abroad may conflict with domestic
reflation plans, whereas budget deficits abroad can crowd domestic
investment out; and so on.
In practice policy makers have tried to take a broader view.
Following the 1973/74 oil crisis, finance ministers of the major
economies agreed that they should avoid competitive deflations
which would pass current account deficits between partners. In
1977 the US administration called for joint action to expand the
major economies as a "locomotive" for world recovery. Policy
makers again called for joint reflation in the 1980-2 recession,
although coordinated policies were by then opposed by the US
government. More recently it has been argued that European
governments should accept greater fiscal expansion in return for
reduced US deficits, and that coordinated policies would have
helped by reducing exchange rate fluctuations. I
However advocating cooperation has led to very little action,
chiefly because there are no estimates of the costs of uncoordinated
policies, nor of the gain (or risks) from cooperation. Optimal
noncooperative policies are examined here as a benchmark against
which the benefits of cooperation can be measured for each econ-
omy in terms of its own goals. In each case, policy makers must
determine their strategies jointly with rational expectations of their
opponents' decisions which are simultaneously chosen in the same
way. The solution technique has often been used for the case where
multiple decision makers compete for control over the same target
variables in one economy.2 The purpose of this paper is to show
how that technique can be extended to the case where national
policy makers attempt to steer two separate but mutually depen-
dent economies, each having their own policy targets, operating
within one "world" economic system. A form of multiperiod con-
sistent conjectural variations solution is involved, and Section 2
indicates some of the characteristics of this solution. Moreover
several simpler solution concepts can be generated as special cases
by imposing prior restrictions on policy makers' reaction functions;
uncertainty and heterogenous information can similarly be intro-
duced; and the full solution shows how a rational expectations
equilibrium is built up by anticipating an adjustment process of the
 189

Lucas critique type. Section 3 then specifies the corresponding set

of cooperative policies.
An empirical illustration is then given using a macro-economic
model, typical of those employed by the policy makers themselves
(Section 4), applied to the recession of 1974-78 (Section 5). In these
exercises the gains from correctly anticipating the policy inter-
actions outweigh the gains from subsequent cooperation (Sections
6 and 7). The tendency of the US to monetary policy because of
budget financing difficulties, and of the EEC to fiscal measures,
shows that these gains arise mainly from coordinating the timing of
fiscal and monetary impacts, and that Europe would gain more
than the US through cooperation.

2. Methodology

2.1. Noncooperative decisions

Noncooperation implies that each decision maker maximises his
self-interest, subject to his perception of the constraints on his
decision variables and conditioned on rational expectations about
the decisions taken by others. Suppose that m targets, y/, are
involved in this game. Their outcomes may be influenced by the
decisions of several independent policymakers, i = 1, ... ,p. For
simplicity we will restrict attention to 2-player games. The i th
player has n policy instruments, X;I) and attempts to reach m of the
l l

targets, y;I). Thus y;l) is the subvector of y/ containing the goals of

the ith decision maker. All noncontrolled and uncertain variables
are gathered together in one (random) variable s/.
Let the planning period contain discrete decision intervals and a
finite horizon, t = 1 ... T. The decision variables can be written
as y(I)' = (y\I)', . . . , y~)'); X(I)' = (X\I)', . . . , x~)') for i = 1, 2;
and s' = (s;, ... ', s~). Suppose each policy maker has ideal values
/I)d and X(I)d for his own decision variables, so that y(l) = /1) _ /I)d
and i(t) = X(I) - X(I)d define his policy "failures". The interests of
each country can then be represented by the quadratic loss func-
tions:
W(I) = 1-[y(l) , BJ)y<I) + i(I)' A(I) i(r)], i = 1, 2 (1)

where B I) and A(I) are positive definite and symmetric matrices.

190

The decision makers face the dynamic economic system

Yt = f(YI'Yt_I,X~I),x~2),eJ (2)
where et represents any noncontrollable variables. For our purposes
the constraints on minimising each objective W(/) can be condensed to
(3)
where R(/,}), j = 1,2, are (miT x niT) matrices contammg the
dynamic multipliers R~k) = oy~/) /ox~) if t ;;::: k, and zeros else-
where; and where s(1) is the subvector of s associated with /i) from
y, The causal nature of (3) is reflected in the block triangularity of
the multiplier matrices R(/,}), which would in practice be evaluated
numerically along suitable reference paths for X~I), X~2) and e(. 3 Thus
R(/,}) describes the responses of player i's targets to its own instru-

ments, and R(/,)) his responses to player j's decisions.

Now each decision maker will find his own variables constrained
by
ji') = R(I,1) X(l) + b(l), (4)
The vector N') = R(l,J) xU) + C(l) , where C(l) = sCi) _ /,)d +
~} R(/,}) X(;)d, defines the information required by player i in order to
make his decision. Evidently each player's optimal strategy depends
on, and must be determined simultaneously with, the optimal
decisions to be expected from the other player. In the absence of
cooperation, the optimal decisions (x(l) *, X(2) *) will satisfy
W(I)(X(I)*, x(})*) ~ W(Il(X(I), x(})*), for i = I and 2 (5)
for all feasible X(I) =1= X(I)*. If (5) is satisfied then neither player has
any incentive to deviate from his choice.

2.2. Open loop Nash equilibria

Now if player i can take xl}) as fixed with respect to his own
decision, as in a Cournot-Nash equilibrium, then the necessary
conditions for an optimal value of X(/) are
OW(/) /OX(I) + (oy(l) /OX(/)Y OW(I) /0/ ' ) = 0 (6)
yielding
X(I)* = _[R(,,')IH'}R("'} + A(I)]-IR(I,I)IH')(cl ' ) + R(l,}}x(;)) (7)

Player i would be correct to use (7) if it is true that his opponent will
choose some value x~) irrespective of the value eventually chosen
 191

for x(I); i.e., if 8x{J) j8X(I) = 0 although 8X(I) j8x(}) =1= 0. 4 But x~) will
generally be unknown because player j is meanwhile attempting to
optimise x(;) conditional on X(l). In that case x(i) must be determined
jointly with x(;) by solving (7) simultaneously with the correspond-
ing expression for x(;)*, to yield the open loop Nash equilibrium
decisions. However this solution continues to impose the restric-
tions 8x(;) j8X(I) = 0 on the rule for x(l)*, and 8X(I) j8x(;) = 0 on the
rule for x(;)*. Suboptimal decisions will result if those restrictions
are not valid. In fact these restrictions must be illegitimate since it
is inconsistent to operate a rule for X(l) * which recognises
8X(I) j8x(;) =1= 0 but assumes 8x(;) j8X(I) = 0, while at the same time
using a rule for x(;)* which allows 8x{J) j8X(I) =1= 0 but assumes
8X(I) j8x{J) = O.
There are therefore good reasons to reject the naive nature of an
open loop Nash equilibrium, and instead it would be sensible to
allow each player to anticipate the opponent's reactions to any
decision he (the first player) might make and to plan counter-
reactions accordingly. This is often done by introducing a Stackel-
berg game, in which the anticipated reactions of only one of the
player's reactions are restricted in this fashion while the other
player's reactions are determined as part of the optimisation. Our
proposal is to go one step further and leave both player's reactions
unrestricted. If both players recognise both reactions to be unrestric-
ted, then neither player attempts to dominate the other and a
Stackelberg war game is avoided.
Of course none of the restrictions, in either Nash or Stackelberg
games, need be zero restrictions. Each player might take his
opponent's policy rule (as opposed to decision values) as given, in
which case a feedback Nash equilibrium will emerge; or the policy
rule of one player could be fixed in this way (while the other rule is
determined within the optimisation) to yield a feedback Stackelberg
solution. In these cases the reaction functions of one or both players
are restricted to preassigned values, and cannot vary with the level
of the anticipated policy interventions. So the next step is to relax
that invariance restriction; and to do that, both reaction functions
must be determined as part of the decision process. Indeed, unless
a player is obliged for institutional reasons (or by lack of infor-
mation) to treat his opponent's responses as given whatever he
himself may decide, it will be necessary to remove any prior restric-
tions in order to deduce rational expectations of his opponent's
decisions and thus to avoid systematic errors in his own decisions.
192

2.3. Conjectural variations and closed loop Nash equilibria

Given a suitable starting point, conjectural variations5 can be
used to deduce appropriate reaction terms within the optimisation
process. Like in chess, each player recognises that his decisions
affect all targets and hence the policy choices of his opponent, and
that variations in the latter then influence the targets again and
hence what the first player should decide to do. Instead of optimis-
ing conditional on fixed reactions by the opponent ("given he does
that, what should I doT'), each player must now anticipate his
opponent's reactions C:if I do this, how will he react and hQw
should I best accommodate that? etc."). Hence these conjectural
variations are captured by the axel) /ox Ci ) and oxCJ ) /oxU) terms. By
optimising the decisions associated with each variation in the con-
jectured responses to the currently envisaged choices, and then
constraining those variations so that the new pair of decisions
represent an improvement over the old pair, we reach the closed
loop (anticipations) Nash equilibrium. That equilibrium holds only
when both players perceive that no further gains can be made by
varying their reactions to the decisions currently expected from
their opponent, because to do so would trigger counter-reactions
(in the opponent's interest) which more than offset the gains of
unilateral action.
The necessary conditions for optimal decisions with conjectural
variations are

OW(I) /OX(I) + [(Oyc') /ox(}»ox(;) /OX(I) + oyCi) /ox(d], ow(i) /OyC') 0

(8)

for i = 1 and 2 (i =I- j). Excepting the reaction matrices ox CJ ) /OX(I),

all the partial derivatives required to evaluate (8) are given by (1)
and (4). If ox(j) /OX(I) = DU) were known for j = 1 and 2, we could
solve (8) for i = 1 and 2 simultaneously to obtain

_ [G~I)!.Ji~) ~(l.~) :- .A(.I). : . .G~I)! ~~) ~(I'~) . .J-I

G(2)! Ji2) R(2,1) . G(2)! Ji2) R(2,2) + A(2)

(9)
 193

where G(I) = R(I,!} + R(I,J) D(J). One obvious way of evaluating (9),
when D(l) and D(2) are unknown, is to construct a fixed point
between (D(l), D(2») and (X(l), X(2») satisfying (8). Indeed an iterative
procedure is implied; inserting trial values (D~l), D~2») into (8) auto-
matically generates new values (D~~ I, D~~ I) via (9):
- D(l)
s+ I ] [x(l) s+ I c(l)]
[F(I)
~(2)
s+ I ]
(10)
[ _ I X(2) F(2) C(2)
s+l s+ I s+ I

(11 )
Moreover a pair (D~), D~») and (X(l)*, X(2)*) satisfying (8) exists
since a Nash equilibrium exists for every nonzero sum game with
convex objectives and strategy sets. 6 If a solution exists for the open
loop Nash equilibrium, then one exists for the closed loop equi-
librium since the latter derives from objective function evaluations
which are better for at least one of the players, and probably for
both, A sensible way to search for this closed loop equilibrium
position is to start from the open loop solution (i.e., (10) with
D~) = 0, i = 1 and 2) and constrain the iterations in (10) and (11)
to generate improvements in one or both objective function evalu-
ations at each step. That ensures that the search will terminate. But,
because of its nonlinearity, (11) is not guaranteed to converge. We
therefore modify (11) by replacing D~'ll with yID~'ll + (1 - y.)D~'),
where 0 ~ Y, ~ I is a scalar chosen to force
hill" (i .e. , such that w(I)(X(I) X(2») ~
xn
W(I)(X(1)
I and x~~ I "down-
x(2)) for i = 1 and
s+ I, s+ I "'" s , s
2). Thus we search among the improvements available from the
reaction matrices which are generated by resolving (8) at each step.
An exhaustive search will ultimately lead to (X(l) *, X(2) *).
Thus, this modification of (11) may be used for various different
purposes. First, it helps overcome any convergence and uniqueness
problems in (10). To establish an equilibrium position, we must
pick that solution which yields the best objective function values for
both players simultaneously from the multiple solutions which (II)
may generate. Second it avoids problems of complex valued reac-
tions in the asymmetric case. Third it recognises that it is only
sensible for a player to alter his conjectures in a way which is Pareto
improving for both players, since otherwise the opponent simply
194

will not react in the way conjectured. One cannot expect any player
to react contrary to his own interests because that would amount
to imposing interpersonal comparisons without any compensating
bargain or side payments. Thus the conjectures must be incentive
compatible with (at least) the zero conjectures solution. In contrast,
removing incentive compatibility altogether, as van der Ploeg and
de Zeeuw (1986) do, automatically generates multiple solutions -
some of which are inferior to the zero conjectures case.

2.4. Uncertainty and the Lucas critique

Another expression for each x~'ll' can be obtained by inverting
the partitioned matrix in (10). Then multiplying out and rearrang-
ing terms yields
x~'ll -[G!i)/H'){G~') + R(I·J)(D~Jll - D~J»)} + A(I)]-l

X G~l)/ H')(C(') + R(I,J) F~l}l c(}») (12)

for i =I, 2 and i =1= j. Thus in view of the existence of values
(D~), D~») at the end of a "downhill" search, the optimal decisions
X(l)* and X(2)* will actually satisfy the model

YI) = (R(I.l) - R(I.j) D~»)X(l) + d~) (13)

for i = 1,2 and i =1= j; where d~) = C(I) + R(l,J) F¥)c(}) and where
F¥) is defined following (10) but is evaluated using D~). The addi-
tive nature of the uncertainty means that certainty equivalence may
be applied directly to (13) to yield the optimal decisions (at t = 1)
X(l)* = - [G~)/ H')G~) + A(I)]-l G~)/ HI) Edd~») (14)

for i = 1 and 2 and G~)/ = (R(I.l) + R(l,J) D~»), where Ell ( .) =

E('I01J denotes an expectation conditioned on player i's infor-
mation at time t. The optimal decisions can then be reoptimised
conditional on each information set 011 (i = either 1 or 2, and
t = 2, ... , T) as it becomes available. Notice that the pair of
decisions are always conditioned on a single information set
representing either one or the other player's view of the past and
future. Thus the term "closed loop" means closed loop with respect
to the past and anticipated decisions of the opponent based on the
current information. If each player then reoptimises his decisions
for each new information set, his rule will also become closed loop
with respect to nature. That ensures X(l) * and X(2) * remain optimal
with respect to EII(w(l)) for t = 2, ... , T, and it is achieved by
successive recalculations of (14) involving various adjustments to
 195

G~), and d~) at each t. The necessary steps, which are rather
Ifl), A(l)

complicated, have been summarised elsewhere. 7

In practice one seldom has accurate knowledge about the con-
tents of the information set employed by one's rivals, and player i's
perception of both optimal decisions may not coincide with player
j's evaluation. To allow for heterogeneous information, we have to
insert Ell(s) into d~) for j i= i as well as into d~) for player i's
determination of the decisions. Uncertainty has thus been reduced
to just two terms in (14); d~) contains both the direct uncertainty
over s within C(l) and the indirect uncertainty over the evaluation of
the other players' decisions (i.e., over s within c(j»). Different esti-
mates of the preferences in Q(I) and the multipliers in the R(l)
matrices can be similarly handled. Of course, the difference between
the ex ante expectation EIl(d¥») and its realisation will depend on
errors in player i's perception of player j's information (Ol) for
j i= i) as well as past information errors in (Ol~k" for k ? 1). But
any policy revisions will depend solely on 0 1" and hence on the
innovations in the policy maker's expectations of the noncontroll-
able variables which determine his anticipations of the future
behaviour of the other participants.
Finally reoptimisation with respect to EI,(W(I») maintains optimal
and time-consistent decisions (i.e., the revised decisions will be
identical to those expected earlier if there is no change in the
information sets). Time consistency holds because the optimisation
is (a) not dependent on a time-recursive objective function
structure, and (b) the reoptimisations using Or!' t ? 2, preserve
the necessary Markov property for the forward steps of the
revision process. 8 The trick here is that the order of operations
has been reversed; the decisions are determined simultaneously
for all periods and then the dependence between players is
solved out, rather than the other way round. The wider ques-
tion of time-consistency (or credibility) when the information
available is known to be incomplete or faulty is not considered
here.
A common objection to the conjectural variations approach is
that the conjectures which agents are presumed to hold are usually
wrong. We saw that the open loop Nash solution, for instance,
requires policy makers to base their actions on the assumption that
their rivals will not react to any policy adjustments. Yet that
assumption generates nonzero reactions (i.e., D\I), D\2) in (9) given
D\I) = 0, D&2) = 0) which falsify the original conjectures. Indeed
196

the iteration at (11) shows that this is true for all conjectures except
those at the fixed (convergence) point satisfying (8) where the
conjectures turn out to equal the optimal reactions. These inconsis-
tencies are of course just a form of the Lucas critique of policy
making. The policies of one player, based on certain conjectured
responses by others, will actually generate reactions different from
those conjectured; and that will invalidate the original policy selec-
tion since the spillovers from those new reactions change the par-
ameters controlling the responses of the first player's targets to its
own instruments. For instance, the open loop Nash solution
generates policies for player 1 assuming that R(I,I) describes the
responses to domestic policy changes, whereas in fact, by (10) and
(11), R(I,I) + R(I) D\2) will determine those responses, Player 1 is
then obliged to modify its proposed action since decisions based on
R(I,I) will be suboptimal for the changed dynamics due to the

opponent's reactions, More generally the responses R(I,I) +

R(l,J) D(}) change to R(I,) + R(I,) DU) Only at termination of the
s s+ I '
"downhill" search is the Lucas critique satisfied because there is no
advantage to either player in changing their response again.

2.5. A hierarchy of solutions

We noted above that the open loop Nash solution is given by (9)
with D(I) = 0 and D(2) = O. That solution is a special case of (10)
generated by setting s = 0 and D(I) = D(2) = O. An open loop
Stackelberg equilibrium assumes that one player (the leader, player
1 here) takes his opponent's responses as given and anticipates no
further reactions, while the follower optimises relative to the
leader's optimal decision; i.e., D(I) = 0 but D(2) "# 0, Thus the
leader expects X(2) * = D(2) x(l) * + i2) where
D(2) = _ (R(2,2) , If2) R(2,2) + A (2») - I R(2,2) , If2) R(2, I) = D\2)

and e(2) = Ff2) EI/(c(2)J. The leader's targets would then behave as
jil) = (R(I,I) + R(I,2) D\2»)x(l) + c(l) + R(I,2)e(2)

implying the optimal decisions

,XII) * = - [G\I), IfI)G\I) + A(I)]-I GP)' Ifl) EJC II ) + R(I,2)e(2))

where GI = R(I,I) + R(I,2) D\2), Using this equation to determine

 197

X(2) *, we get

which is a special case of (10) at s = 1 with Db' ) = mIl

= 0 and
Db ) = O.
2

On the other hand, the leader might take the follower's policy
rule as given, say X~2) = KtYt-' + k t . The follower's reaction
matrix can be obtained by stacking and substituting out the Yt-l
terms by (3):

(15)

KT''lT + kT

orx(2) = [J - KR(2)r ' [KR(')x(') + k]whereR(l) = (R(d»)fori =

R(l.2)
1,2. Hence D(I) = 0 again, but now D(2) = (J - KR(2»)-1 KR(l).
This feedback Stackelberg equilibrium is one step of (l0) from
Db' ) = 0 and D&2) = 0 = (J - KR(2 l )-1 KR(l). Finally, a feedback
Nash equilibrium holds when each player takes his opponent's
policy rule, rather than policy values, to be fixed. Player 2's reaction
matrix, D(2), is the same as in (15). Player 1, who uses the rule
X~I) = L,Yt_' + II' will have the corresponding reaction matrix
D(I) = (J - LR(I»)-1 LR(2). Thus the feedback Nash solution is the
first step of (10) from Db' ) = (J - LR(l»)-' LR(2) and D&2) =
(J - KR(2 l ) -I KR(l).
Conventional solutions are therefore special cases of the general
solution (9), in which certain prior restrictions are placed on the
conjectural variations or reaction matrices D(l). The generalisations
offered by a conjectural variations solution are twofold: (i) the
reaction matrices are not restricted to any preassigned values, and
(ii) the reaction matrices contain no implicit zero restrictions across
time intervals. The latter allows current decisions to react fully to
anticipated future decisions, and in that they represent more sophis-
ticated behaviour than Nash strategies which feedback only from
past economic states or past policies.
198

3. Cooperative Decision Making

It is a well known result that, in the absence of side-payments, the

outcomes of a noncooperative game are socially inefficient in that
alternative decisions exist which would make all parties better off.
Hughes Hallett and Rees (1983), for example, show that a necessary
condition for any noncooperative decision rule to be Pareto effici-
ent is that each player's reaction matrix, oxU )/ox(l) , must vanish and
that the aims and priorities for the union of targets in the game must
be identical for both players. If these conditions are not fulfilled,
and the fact that y(l) i= l2) (i.e., to the extent countries pursue
national rather than international goals) means that they can never
be fulfilled, then another decision rule will dominate because at
least one country can be made better off without the other being
worse off. Da Cuhna and Polak (1967) have shown that the set of
nondominated decisions can be generated by minimising the "col-
lective" objective function
w = aw(l) + (1 - a)w(2) where 0::::; a ::::; 1, (16)
subject to (3), with respect to X(I) and X(2) together. That constitutes
a single player's problem.
Hence, once a collective objective (or a value) is agreed, all coun-
tries can gain by following a Pareto efficient cooperative solution
but a redistribution of benefits may be implied. But, by the same
token, the cooperation solution is not optimal for country i in terms
of W(l) alone, given that it has now acquired a measure of control
over xU). There is therefore an incentive to cheat while the other
countries maintain their cooperatively optimal decisions, and the
existence of this kind of risk may explain why politically sovereign
policy makers have not been persuaded of the superiority of cooper-
ative solutions in practice. To do that one needs to establish that the
losses risked, if the other country does cheat, are acceptably small.
If those losses are not small, we will need to design suitable punish-
ment schemes: see Hughes Hallett (1986).

4. The Model

The mutual dependence between the US and EEC economies,

and its implications for policy making in those economies, has been
 199

explored empirically using a model derived from the COMET

multicountry econometric model (Barten, d' Alcantara, Carrin;
(1976)). The COMET model is typical of the empirical models
which are used for analysing economic interdependence, and the
potential for policy cooperation, in practice. It describes the econo-
mies of the EEC countries, the US, Japan, plus the rest of the
OECD, Socialist Countries, OPEC and developing countries. The
version used here has been respecified, and reestimated, to describe
the US, EEC and Rest of the World economies and their trade
flows. The real sectors have been augmented with financial sectors
which include international capital movements. These financial and
real sectors are combined in such a way that the Rest of the World's
accounts are forced to balance. A more detailed description will be
found in Van der Windt and Siebrand (1984), but the main changes
are: (a) aggregation into three economic blocks, to give a systematic
specification across the US and EEC economies; (b) consistent
accounts for the Rest of the World; (c) the introduction of financial
sectors, international capital movements, and endogenous exchange
rates.
The US and EEC blocks, some 200 equations in total, form the
backbone of the model. The two real sectors have similar structures,
based on a conventional Keynesian demand system covering con-
sumption, investment, and foreign trade. Cobb-Douglas produc-
tion functions determine labour demand and investment. The sup-
ply sides of both blocks are modelled by similarly specified potential
output functions; potential and actual outputs are reconciled by a
capacity utilisation index. In each sector prices are related to import
prices and the GDP deflator. The GDP deflator in its turn depends
on unit labour and capital costs and the utilisation index, while
wages are a function of consumption prices, productivity, unem-
ployment, taxes and social security contributions. The Rest of the
World is modelled by a rudimentary macro system explaining how
export revenues are devoted to increasing international reserves,
lent to one of the other two regions, or spent on imports.
Each monetary sector describes the financial relations between
the central bank, commercial banks, the government, the private
sector, and the foreign sector. Private and government financial
surpluses are the main input from the real sector. Interest rates,
which in turn affect both real spending and international borrowing,
are determined by those surpluses and by the behaviour of the banks.
200

There is an important difference between the US and the EEe here.

Government expenditure in both blocks are financed to a large
extent by loans from the private sector. In the EEe, this involves
selling bonds to the banks which adjust their portfolios and offset
any upward pressure on interest rates. Meanwhile the extra govern-
ment expenditures pass through the banking system and, both
directly and via credit creation, tend to reduce interest rates and
increase activity domestically and abroad until inflation and wage
increases set in to reverse that some periods later. Government
expenditures in the US, however, are mostly financed by borrowing
savings directly from the private sector or by tax measures. With a
comparatively low propensity to save and less opportunity for the
banks to maintain the term structure of their portfolios, coupled
with interest sensitive investment schedules in the US, this leads to
crowding private investment out plus higher interest rates and
falling activity levels. At the same time prices rise, mainly due to
higher capital costs. This crowding out process, and falling activity
levels, sets in after one year; and it is sustained because there is no
offsetting expansionary pressure or credit creation to the extent that
government expenditures just displace investment and that new
savings go into new government programmes.
Thus American and European policies may ultimately have the
same effects, but the US economy reaches the new situation much
quicker. Moreover, policy responses in the US are usually more
powerful than those in the EEe. That appears to be due partly to
the differences in reaction speeds just mentioned, and partly
because the impacts of those differences are exaggerated by the
strength of international capital movements to and from the US.
These capital movements depend on relative interest rates, domestic
savings (public and private), changes in the trade balances, and
relative growth rates. For instance, both economies react similarly
to a rise in the domestic interest rate, but the US balance of
payments alone is affected significantly (and positively). Meanwhile
the impact multipliers of a US interest rate rise on European targets
are roughly equal to those for an equivalent rise in the EEe; but the
EEe rate does not have the equivalent effects on US targets. Hence
internal US policies tend to attract capital. Expansionary fiscal
policies go together with relatively tight monetary policy in the US,
tend to act faster, and to be reinforced by their international
consequences. These opportunities do not arise in the EEe where
 201

fiscal and monetary policy appear to be somewhat offsetting and

without strong international implications. This greater responsive-
ness explains why the US has pursued a more active monetary
policy; and also why the European governments fear higher interest
rates which, if induced by the need to protect their balance of
payments and exchange rates in the face of high US rates, then lack
of any compensating credit creation effects.
In fact most government policies lead to quicker and stronger
responses in the US than in the EEC. This stems principally from
reactions of the real quantities to price and interest changes, which
are both larger and faster in the US structural equations. The effects
of these reactions then spread rapidly to other variables because the
real and monetary sectors are more closely connected. The upshot
is more powerful short run policy multipliers (which can be seen in
the diagonal blocks of R(I,I») but less powerful longer run multipliers
(which appear in the bottom left submatrices of R(I,I») in the US
final form equations (4).

5. The Decision Problem

The planning period 1974-78 was chosen in order to study policy

options open to the US and the EEC to the oil price shock of
1973-74. In practice Western Europe experienced a large rise in
unemployment, which later proved to be a permanent increase.
Aggregate demand policies resulted in high inflation, supported by
real wage rigidity. At the same time investment collapsed and
remained low throughout the 1970s. The oil importing countries
saw the balance of their current account become negative, although
the deterioration of the balance of payments was shortlived. While
the US deliberately let their current account deteriorate to shield
their economy from supply effects, investment did not slow down
as dramatically as in Western Europe. Unemployment reached a
postwar peak, but responded quickly to tax cuts in the US.
In line with these observations, the exercises which follow assume
the US and the EEC to have the same target variables, but the
priorities which the policy makers attach to these targets are dif-
ferent in accordance with the policy stance just described. The
planners of both economies are also assumed to be interested in
governing the supply conditions in the final year and thereafter. To
 tv
o
tv

Table 1. The objectives of economic policy 1974--78.

US EEC
Ideal Priority Ideal Priority Symbol Units
Targets
Inflation -3% 4 -3% P % annual growth
Output 3% 4 3% 4 X % annual growth
Balance of trade 1% 1% 4 B % of GNP
Employment It% It% E % of labour force
Investment ratio 3% 3% IR % of GNP
Profit ratio 3% 3% PR % of GNP
Instruments Historical Ave Historical Ave
Government expenditure 2% 4% G % annual growth
Social Security Outlays 15% 16% S % annual growth
Direct tax, households 10% 17% TH % annual growth
Direct tax, companies 9% 14% TC % annual growth
Discount rate 0.4% -t% R Level
Central bank loans 5% 19% LC % annual growth
Commercial bank loans 5% 19% LB % annual growth
 203

that effect, the targets combine the conventional stabilisation objec-

tives (1.5% less unemployment, 3% less inflation, 3% more produc-
tion growth, 1% improvement in the balance of trade) with indica-
tors of potential supply (a 3% increase in the real investment/
production ratio, and a similar increase in the ratio of gross operat-
ing surplus of companies to net national income). A representation
of the difference in policy stance is provided by giving inflation a
relatively high weight in the US objective function and weighing the
position of the current account more heavily in the case of the EEC.
The policy goals of this exercise are set out in Table 1. The policy
instruments have been selected from those used during the period.
Social security premiums do not constitute an independent instru-
ment since they are tied to the outlays in order to balance the social
security budget. Instruments are treated here as discretionary
adjustments to the historically chosen polcies. The instruments are
all expressed in annual growth rates, except the discount rate for
which the multipliers were based on level changes. Following the
procedures developed in Brandsma, Hughes Hallett, and Van der
Windt (1983), we found the penalities on deviations of the instru-
ments from their ideal values should be equally distributed. Thus,
in terms of (1) A(I) = I and H') is a diagonal matrix with the
elements indicated in Table 1.

6. The Results

Tables 2 and 3 present optimal noncooperative and cooperative

policies for the US and EEC over the years 1974-1978. The Cournot-
Nash solution here represents the case where each player assumes
that his opponent will choose the historical values, which are also
ideal values here, for his instruments. 9 The optimal decisions are
then given by (7) with x~) = X(;)d, even though the assumed x~)
value is inconsistent with the fact that player j is also optimising by
(7). Thus the Cournot-Nash solution takes the opponent's strategy
as known, rather than just taking the opponent's reactions to be
fixed at zero. The Cournot-Nash solution is therefore included as
a benchmark representing the result of optimising behaviour which
ignores the game aspects of interdependent policy making.
The conjectural variations solution allows unrestricted policy
reactions by each player, including anticipation terms, and estimates
204

Table 2. Optimal Cournot-Nash, non-cooperative and cooperative policy adjust-

ments for the US and EEC economies (1974 information set)
Conjectural
Cournot-Nash variations Cooperative
variables US EEC US EEC US EEC
1974: G -1.33 1.88 0.31 1.99 0.78 2.04
S -2.83 0.29 -1.20 -0.86 -0.62 0.28
Th -0.44 -1.74 -0.62 -2.91 -0.53 -1.60
TC -0.87 -0.21 -0.70 -0.50 -0.37 -0.19
R -1.44 -1.01 -1.33 0.60 -0.39 -0.57
LC 0.39 -0.02 0.22 0.18 0.10 -0.01
LB 0.76 -0.53 0.35 0.88 0.28 -0.53
1975: G -1.78 0.84 -0.07 0.87 0.37 1.44
S -2.80 0.39 -1.22 0.75 -0.55 0.42
TH -0.56 -0.94 -0.59 0.00 -0.42 -1.05
TC -0.62 -0.03 -0.35 0.01 -0.26 -0.06
R 0.37 -1.41 0.99 -2.22 -0.29 -0.55
LC 0.19 0.03 0.11 -0.12 0.04 0.01
LB 0.89 0.09 -0.20 -1.61 -0.18 -0.20
1976: G -0.88 0.31 0.06 0.81 -0.24 1.00
S -2.20 0.39 -0.95 0.28 -0.56 0.39
TH -0.78 -0.08 -0.63 -1.14 -0.24 -0.47
TC -0.36 0.03 -0.15 -0.02 -0.08 -0.03
R 0.67 -1.32 -0.27 -0.44 -0.13 -0.38
LC 0.09 -0.01 0.07 0.09 0.07 -0.01
LB 0.19 -0.33 -0.92 0.71 -0.20 -0.20
1977: G -0.32 -0.19 -0.25 -0.03 -0.14 0.38
S -2.13 0.11 -0.09 0.54 -0.61 0.31
TH -0.97 0.24 -0.58 0.63 -0.29 0.07
TC -0.26 0.03 0.02 0.07 -0.01 0.00
R 0.04 -0.28 -0.35 -1.47 -0.78 0.29
LC - 0.11 0.02 -0.07 -0.02 -0.01 0.01
LB -1.64 0.02 -1.43 -0.23 -0.53 -0.00
1978: G -0.49 -0.26 0.86 -0.22 0.38 0.20
S -1.96 0.10 -1.08 -0.06 -0.72 0.22
TH -0.91 0.46 -0.93 0.32 -0.53 0.10
TC -0.11 0.02 0.03 -0.04 0.01 -0.01
R -0.69 -0.01 -1.49 0.97 -0.58 0.44
LC -0.09 0.00 -0.10 0.00 -0.05 -0.00
LB -1.39 -0.02 -1.40 0.03 -0.71 -0.02
Legend and units: see Table I.
TABLE 3. Mean and mean absolute interventions under Cournot-Nash non-cooperative, and cooperative decisions (1974 information
set). RelatIve to the historical values.
Cournot-Nash Conjectural variations Cooperative
Instruments US EEC US EEC US EEC
G - 0.96 (0.96) +0.52 (0.70) +0.18 (0.31) + 0.68 (0.78) + 0.23 (0.38) + 1.01 (1.01)
S - 2.38 (2.38) + 0.26 (0.26) - 1.11 (1.11) + O. I3 (0.50) -0.61 (0.61) + 0.32 (0.32)
TH -0.73 (0.73) +0.41 (0.69) - 0.67 (0.67) - 0.62 (0.96) - 0.40 (0.40) - 0.59 (0.66)
TC - 0.44 (0.44) - 0.03 (0.06) - 0.23 (0.25) -0.10 (0.11) -0.14 (0.14) - 0.06 (0.06)
R - 0.21 (0.64) -0.81 (0.81) - 0.49 (0.89) -0.51 (1.14) - 0.43 (0.43) - 0.15 (0.46)
LC + 0.10 (0.17) + 0.01 (0.02) -0.05 (0.10) + 0.03 (0.08) + 0.03 (0.05) 0.00 (0.00)
LB - 0.24 (0.97) - 0.18 (0.25) - 0.72 (0.86) - 0.04 (0.69) - 0.27 (0.38) -0.19 (0.19)
Targets
P -1.6 -2.2 -2.2 -2.3 -2.3 -2.1
X + 1.9 +2.4 +2.3 +2.3 +2.7 +2.4
B -2.1 -0.1 -0.5 +0.3 -0.2 +0.4
E +0.9 +0.8 +1.2 +0.8 + 1.2 +0.9
IR +1.5 +3.8 + 1.8 +3.3 +2.0 +3.7
PR + 1.2 +3.4 + 1.5 +3.6 +1.3 +3.6
Note: The bracketed figures give the mean absolute interventions.

tv
o
Vl
206

the conclusion of the conjectural variations process (10) and (11).

The cooperative solution is the result of optimising the sum of US
and EEC objectives with respect to x(l) and X(2) jointly: i.e., (16) with
C( = t and subject to all the constraints in (3). That is a convenient

point of reference because it weights the national objectives by the

relative GNP shares of the period. These three solutions represent
the three fundamental approaches to interdependence which policy
makers might wish to adopt: to ignore the game aspects of inter-
dependence; to play an unrestricted game; or to pick an equitable
compromise by cooperation. However, all the policies here are open
loop with respect to nature. That is, they use only the initial (1974)
information set, and represent the options which would have con-
fronted policy makers when they had to decide on their fundamental
strategy.
Consider first optimal policies when the mutual economic
dependence is ignored. Inspection of the Cournot-Nash policies in
Table 2 reveals a cut in government expenditure and social security
benefits in the US, while there is a rise in both quantities in the EEe.
That contrast reflects the speed of policy responses in the US. The
financing of US public expenditures induces contradictory pressure
on output, plus upward pressure on prices and interest, within a
year. Hence reductions of expenditure (compared to the historical
path) are necessary to start with in order to reach the output,
employment, inflation and investment targets; but later on those
reductions can be eased. Social security benefits, however, which
depend on demographic and unemployment factors, lack the
required short run flexibility to adjust period by period. This fiscal
disengagement is matched by tax cuts (mainly favouring the cor-
porate sector to start with, but ultimately benefiting households by
more) which goes some way to improving the budget deficit. In
contrast, financing a fiscal expansion in the EEC has expansionary
effects on output and inflation in the first few years, and adverse
effects only later. So expenditures are increased to start with, while
taxes (almost completely on the household sector) are sharply
reduced. But this pattern is reversed in the final two years (expen-
ditures being cut and taxes increased) in order to offset the gradual
build up of adverse effects of the initial expansion. Substantial
financing of the budget would be necessary in the EEC as, overall,
tax cuts complement the extra expenditure.
 207

The monetary variables show that the change in European

government budgets is not financed by loans from the central banks
(LC) and that the loans from commercial banks (LB) cause no extra
money creation, while the main instrument, the discount rate, is
lowered throughout the period. After a few years, however, balance
of payments problems and the undesirability of monetary financing
lead to the switch in fiscal policy already noted. In the US the
dependence of budget financing on savings requires a more soph-
isticated strategy. First a lowering of the discount rate has to induce
a fall in interest charges to stimulate economic activity. A rise of the
discount rate (R) is then used to acquire savings until the loans from
the banks to the government start to decline, after which the
discount rate is lowered again.
The base policy position is therefore one of standard Keynesian
deficit financing in Europe, to stimulate aggregate demand and
output, with monetary control and falling interest rates to reduce
inflation. Meanwhile policies which have come to be associated
with "supply side" economics operate in the US; a reduction in
government intervention, especially in social security outlays and
taxes, is coupled with an activist monetary policy. Table 3 supplies
mean intervention values and gives an idea of the vigour of policy
adjustments by comparison with mean absolute interventions.
Policy switches are important only in the US monetary instruments.
Turning to the optimal noncooperative policies both economies
would now experience rising government expenditures, although
this is only marked in the US. The US social security cuts are
halved, and in Europe they turn positive in 1975-77. Once again tax
cuts appear in both economies, although they are more variable
than before in Europe and they are relatively more beneficial for
households in the US. The changes in monetary policies are more
complicated. An initial rise in European discount rates 'attracts
American capital, as witnessed by lower loans to the US govern-
ment. The latter necessitate the Federal Reserve Board to resort
earlier to negative discount rate adjustments. As a result of the
policy interaction process described in Section 2, rational expec-
tations together with the destabilising interim effects of US discount
rate changes then lead to larger adjustments in both monetary and
fiscal policy in Europe. Indeed Table 3 shows that the main changes
are in the monetary variables. The US follows a more restrictive
policy in 1977-8, but reduces interest rates; the EEC intervenes
208

more actively both with loans and the discount rate, and now has
a more activist monetary policy than the US.
In summary, recognising policy interactions had led to some
convergence in the national policies. The US had dropped most
of its "supply side" stance in favour of some demand creating
measures; and the EEC is partly able to overcome the inertia of its
own policy responses as indicated by more flexible interventions,
using particularly the discount rate and money supply. Finally,
excepting the discount rate, US interventions are smaller and more
consistent while European interventions run at the same level but
have become more active.
Cooperation between the US and the EEC takes the optimal
policies a step further towards convergence. Government expen-
ditures rise faster in both economies, and the US social security cuts
are halved once again. Tax cuts are still in evidence, although less
prominent in the US. The activism of monetary policy has vanished
-particularly the sharp restrictions of money supply and interest
rates which appeared in the non-cooperative solutions. In fact both
loans to the government and the discount rates now follow constant
or steadily changing paths. This suggests that it may be important
to coordinate monetary policies. However, the convergence pattern
does show that cooperation may require individual policies to be
surrendered for the sake of concerted action, as in the expenditure
variables here.
Overall the cooperative policies call for reduced intervention in
the US, and for more consistent policies in both economies, com-
pared to the noncooperative strategies. The exceptions to this
pattern are the rise in government expenditure in the first two years,
followed by a drop in 1976-1978. European countries benefit from
this initial rise in government expenditures and from the continuous
US discount rate reductions. The induced stability provides the
opportunity to follow more consistent aggregate demand policies in
the EEC. They are much like the Cournot-Nash solution for the
first two years; they are similar to the non-cooperative solution in
1976 except for a cut in the loans to the government, and only show
an increasing discount rate in the last two years of the planning
period. This break in the concerted action over discount rate
changes in the EEC and the US is necessary in order for Europe to
attract capital imports which then compensate for the balance of
trade deterioration induced by the growth in government outlays.
 209

In their turn, higher government expenditures in the EEC are

beneficial for the American economy.
Table 3 shows the average expected changes in the target vari-
ables, compared to those values which could have been expected by
.simulating the historical instrument choices. These expected out-
comes are better for all variables in both economies, excepting the
US balance of trade. The EEC does relatively better in the Cournot-
Nash solution, and in raising investment and profits in all solutions.
But it does badly with the employment target, indicating that
European unemployment will evidently be very persistent. In con-
trast to the associated instrument changes, the US target changes
become more favourable as we move from the Cournot-Nash to
the cooperative solution (especially for production, balance of
trade, and investment). The EEC targets however are not much
affected by the type of strategy chosen. Once again we have evi-
dence of the asymmetry in US-EEC economic relations. The gains
to cooperation for the US take the form of smaller interventions
and improved expected target outcomes. But in the EEC those gains
arise from reductions (or redistribution) of intervention effort.

7. The Incentives to Cooperate

The objective function evaluations for the US and the EEC

individually under the three basic strategies are set out in Table 4.
These are expected values based on the initial (1974) information
set relevant for choosing between these strategies.
The diagonal entries of both tables indicate that recognising and
accommodating economic interdependence in the policies adopted
brings larger gains to the US than the EEC, but that the gains to
explicit cooperation are significantly smaller and mainly benefit the
EEC. Since they contain separate normalisations, w(ll* and w(2l*
values are not directly comparable. Nevertheless the US has
proportionally nearly twice as much as the EEC to gain from
"playing the game" (72% vs. 39%), whereas the further gains from
cooperation are much smaller (9% vs. 33% respectively) and
favour the EEC by a factor of 3 to 1. To place these results in
context, the gains to cooperation can be expressed in units of
equivalent average GNP growth per year. That is we compute the
average rate of GNP growth, all other decision variables fixed,
210

Table 4 .. US objectives: W(I) *

"(BEC play, CN CV Coop
USPla~
CN 188.9 160.9 92.8
CV 65.4 52.2 75.2
Coop 91.8 96.0 47.5

EEC objectives: W(2) *

VECPlay, CN CV Coop
USpla~
CN 66.9 153.1 27.4
CV 57.7 40.7 120.2
Coop 62.2 146.2 27.2
Note: The simulated historical policies yielded objective values of 233.1 and 173.1
for the US and EEC respectively.
CN = Cournot-Nash, CV = Conjectural variations; Coop = Cooperative
«(t = t) solutions.

which would yield the same objective function gains for each econ-
omy. Here the gains to cooperation for the US are equivalent to an
extra 0.47% annual GNP growth for 5 years, while for the EEC it
is "worth" an extra 1.35% annual GNP growth for 5 years.
On the face of it, therefore, the incentives to cooperate are not
very large for the US; and they are quite a lot smaller than the
incentives to exploit the game fully, or at least to allow normal
competitive forces to guide policy making. This may partly explain
the reluctance of the US to engage in cooperative policy making.
Thus, on the basis of these results, we can expect the US to make
very little effort to coordinate its policies with the EEC, while the
EEC would stress the importance of such coordination between the
industrialised economies.

8. Conclusions

In this paper we have shown how the solution to dynamic games,

incorporating conjectural variations and "noncausal" anticipations
effects, can be extended to handle interdependent (open) econo-
mies, each having its own national target and instrument sets,
 211

within a "world" economic system. This solution was applied to the

problem of designing optimal policies for the US and EEC econo-
mies following the oil price shock of 1973/4. The following results
were obtained:
(a) Analysis of the model, and of the optimal policy sequences,
show inertia in EEC's policy responses compared to those in the
US. This leads to asymmetries in the interdependence of the
two economies in the short term. The figures in Table 4 show
that the EEC is also more sensitive to the strategy selected in the
US than vice versa. Knowledge of likely US policy choices is
more important to the EEC, than knowledge of EEC policies is
to the US.
(b) Policy isolation leads to a typical "supply side" stance of
reduced fiscal interventions and a tight but active monetary
policy in the US. Left to itself the EEC would tend to follow
standard Keynesian deficit financing and an accommodating
monetary policy. However recognising policy interactions leads
to some convergence in the policies adopted, and provides
larger gains for the US than for the EEC. With its quicker
policy responses, the US can more easily absorb external
shocks and the consequences of policy changes in the EEC than
vice versa. In particular it can induce capital flows to finance
larger government expenditures and can thereby go some way
to deficit financing without incurring the adverse effects of
having to do that domestically.
(c) The gains from cooperation favour the EEC more than the US,
but are relatively small compared to the gains from recognising
policy interactions. This implies that the EEC should try to
exploit the US as a "locomotive"; improvements in the US
economy would enable it to reach the same target values with
less effort. Thus for the US, competitive decision making is the
main issue, but for the EEC mutual cooperation is probably the
more important factor. These results may help explain why the
Americans have not been sympathetic to European arguments
for coordinating economic policies, and perhaps even why the
Europeans have found the US position hard to understand.
212

Notes

I. These views are noted in Oudiz and Sachs (1984), and summarise the views
expressed by Helmut Schmidt, V. Giscard d'Estaing, Martin Feldstein, and
Laurence Klein, in the Economist between February and June 1983. See also
Basevi et at (1984), Begg (1983).
2. See Van der Ploeg (1982), Hughes Hallett and Brandsma (1983), Brandsma
and Hughes Hallett (l984a, b), Hughes Hallett (1984).
3. The problem of how to keep the evaluation of the multiplier matrices consis-
tent with the selection of instrument values X(l) and X(2) is examined by
Brandsma et al. (1984). For the purposes of this paper we will treat those
matrices as known, in order to abstract from further complications of par-
ameter uncertainty.
4. This is a passive policy according to Fair (1978).
5. See Bresnahan (1981), Holt (1985) and references therein. This method has
been extended here to a multi-target, multi-period form under uncertainty.
6. See Aubin (1979) or Friedman (1977). Bresnahan (1981) shows that a "rational"
conjectural equilibrium exists in linear-quadratic problems, although, if the
player's policy responses are asymmetric, there is no proof that the corre-
sponding reaction functions are real valued. If that should happen, it seems
reasonable to stop the iteration (11) at the real valued solution nearest the
consistent conjectures equilibrium such that no player is worse off than he
would be with any other conjectures (including zero conjectures). That prob-
lem apart, the purpose of (II) is to ensure that the computed policy responses
ax(l) /ax(J) actually equal the same values conjectured for them. Uniqueness is
more difficult. The open loop Nash equilibrium is unique in the linear-
quadratic case (Aubin (1979»; and it seems that the closed loop Nash equi-
librium may be unique for the same reasons since there is only one information
set here, covering all elements in c(l) and C(2). However, the "downhill" directed
search, introduced next, will in any case locate the joint minimum. This is not
to suggest that the convergence of (II) will be unique.
7. See Hughes Hallett (1984).
8. Bellman (1961), pp. 54-57. This time consistency property is a particular case
of the time consistency of optimal muItiperiod decisions under full informa-
tion and perfect rational expectations established by Tesfatsion (1984).
9. See Neese and Pindyck (1984).

References

Aubin, J. P. (1979): Mathematical Methods of Game and Economic Theory, North

Holland, Amsterdam.
Barten, A. P., G. d'Alcantara, and C. J. Carrin (1976): "COMET: A medium-
term macroeconomic model for the European Economic Community", Euro-
pean Economic Review 7, 63-115.
 213

Basevi, G., O. Blanchard, W. Buiter, R. Dornbusch, and R. Layard (1984):

"Europe: the case for unsustainable growth", Centre for European Policy
Studies, Brussels.
Begg, D. (1983): "The Economics of Floating Exchange Rates: the lessons of the
70s and the research programme for the 80s", Memorandum in International
Monetary Arrangements: Vol. III (p. 4-56), HMSO, London.
Brandsma, A. S. and A. J. Hughes Hallett (1984a): "Noncausalities and time
inconsistency in dynamic noncooperative games: the problem revisited",
Economics Letters 14, 123-130.
Brandsma, A. S. and A. J. Hughes Hallett (1984b) "Economic Conflict and the
Solution of Dynamic Games", European Economic Review 26, 13-32.
Brandsma, A. S., A. J. Hughes Hallett, and N. van der Windt (1984): "Optimal
Economic Policies and Uncertainty: the case against policy selection by non-
linear Programming", Computers and Operations Research 11, 179-197.
Bresnahan, T. F. (1981): "Duopoly Models with Consistent Conjectures" Ameri-
can Economic Review 71, 934-945.
Canzoneri, M. and J. A. Gray (1983): Monetary Policy Games and the Conse-
quences of Noncooperative Behaviour, International Economic Review 26,
547-64.
Da Cuhna, N. and E. Polak (1967): "Constrained Minimisation of Vector-Valued
Criteria in Finite Dimensional Spaces", Journal of Mathematical Analysis and
Applications 19, 103-24.
Fair, R. C. (1978): "The Sensitivity of Fiscal Policy Effects to Assumptions about
the Behaviour of the Federal Reserve", Econometrica 46,1165-1179.
Friedman, J. W. (1977): Oligopoly and the Theory of Games, North Holland,
Amsterdam.
Holt, C. A.: "An Experimental Test of the Consistent Conjectures Hypothesis"
American Economic Review 75, 314-325.
Hughes Hallett, A. J. (1984): "Noncooperative Strategies for Dynamic Policy
Games and the Problem of Time Inconsistency", Oxford Economic Papers, 36,
381-399.
Hughes Hallett A. J. (1986): "International Policy Design and the Sustain ability
of Policy Bargains" Journal of Economic Dynamics and Control, 10 (forthcoming).
Hughes Hallett, A. J. and A. S. Brandsma (1983): "How Effective Could Sanc-
tions against the Soviet Union Be?", Weltwirtschaftliches Archiv, 119,498-522.
Hughes Hallett, A. J. and H. J. B. Rees (1983): "Quantitative Economic Policies
and Interactive Planning", Cambridge University Press, Cambridge and New
York.
Miller, M. and M. Salmon (1985): "Dynamic Games and the Time Inconsistency
of Optimal Policy in Open Economies", Economic Journal (Supplement,
124-135).
Neese, J. W. and R. S. Pindyck (1984): "Behavioural Assumptions in Decen-
tralised Stabilisation Policies", in A. J. Hughes Hallett (ed.) Applied Decision
Analysis and Economic Behaviour, Nijhoff, Boston and the Hague.
Oudiz, G. and J. Sachs (1984): "Policy Coordination in Industrialised Countries",
Brookings Economic Papers (I), 1-64.
Tesfatsion, L. (1984): "The Consistency of Benevolent Government Economies",
214

paper presented to the Society of Economic Dynamics & Control, Nice, June
1984.
Van der Ploeg (1982) "Government Policy, Real Wage Resistance and the Reso-
lution of Conflict". European Economic Review 19,181-212.
Van der Ploeg, F. and A. J. de Zeeuw (1986): "Noncooperative Strategies for
Dynamic Policy Games and the Problem of Time Inconsistency: a Comment"
unpublished manuscript.
Van der Windt, N. and J. C. Siebrand (1984): "An Annual Model of the US
economy-Rasmus 2a" and "An Annual Model of the EC-Rasmus 3a", Dis-
cussion papers 8408/G and 84l1/G, Institute for Economic Research, Erasmus
University Rotterdam.
 215

CHAPTER 13

HIERARCHICAL GAMES FOR MACROECONOMIC

POLICY ANALYSIS

Carlo Carraro
University of Venice, Italy

1. Introduction
In recent years, a new mathematical problem has been proposed in
the en'6ineering literature. The main features of this problem, called
closed-loop Stackelberg (CLS) problem, can be described in the
following way.
Suppose the control problem can be described as a game between
two decision-makers and suppose that one of the two players, called
the leader, has the power to announce his strategy first and to make
his strategy conditional on the other player's strategy. Can the
leader announce a strategy such that the follower is induced to
behave as if he were acting in the leader's interest? How can this
strategy be computed? What are the properties of the solution of the
game when such a strategy is announced by the leader? What are
the properties of the leader's optimal strategy? These questions can
be formalized into a mathematical problem and many papers have
recently tried to provide a solution to this (CLS) problem. I
The features of the CLS problem are not completely new in
economics. In fact, Chow (1981, Ch. 17) provides an algorithm
which can be used to compute the steady-state CLS solution of the
game, if the follower's strategy is included into the state vector of
the dynamic equation describing the economic system. However,
several aspects of the CLS problem, not explored by Chow (1981),
were recently emphasized in the engineering literature, where other
solutions of the problem have also been proposed. In particular, the
existence, uniqueness, and time-consistency of the CLS solution
have been analysed.
The author is grateful to Gregory Chow for helpful comments on a previous
version of this paper.

C. Carrara and D. Sartore (eds.) Developments of Control Theory for Economic Analysis
©1987 Martinus Nijhoff Publishers (Kluwer), Dordrecht·
216

Furthermore, no application of the CLS solution to games

representing economic problems has been provided, even if the
typical behaviour of the policy-maker can easily be fitted into the
CLS framework. The macroeconomic policymaker can indeed be
assumed to be the leader of the game and to announce the optimal
strategy by taking into account not only the follower's rational
reaction set (as in Kydland, 1975, 1977, where the feedback Stackel-
berg solution is considered), but also the actual follower's decision.
Therefore, the CLS solution determines that particular leader's
strategy which induces the follower to behave as if he were
cooperating. This provides a great advantage to the policy-maker.
If indeed the monetary authority or the government can
announce a strategy which implies the minimization of their loss
function with respect not only to their own decision variables, but
also with respect to the private sector's decision variables, then it is
obvious that, in general, a lower loss can be achieved, i.e., it
becomes easier for the policy-maker to attain the desired targets.
Does such a strategy exist? The goal of this paper is to answer this
question and to study the properties of different CLS strategies.
Two main problems will be examined: first, under what conditions
is a CLS strategy really effective, that is, when does it imply a loss
for the policy-maker which is lower than the loss implied by an other
policy? Second, under what conditions is the CLS strategy credible?
The effectiveness of a CLS strategy depends indeed on a set of
threats (or incentives) that the leader announces he will carry out
any time the follower does not comply with the strategy selected for
him by the leader. 2
A well-defined CLS strategy must therefore be based upon a set
of credible threats (or incentives) so that the leader does not have
any incentive to depart from the announced strategy, whenever the
follower does not act in accordance with his will. An effective and
credible CLS strategy can easily be shown to be the policy-maker's
best policy.
Suppose, for example, that the monetary authority controls the
money stock and the private sector controls the price level. If the
monetary authority can induce the private sector to keep the price
level constant, then the money stock can be used to stimulate
economic activity. In contrast, the standard optimal policy faces the
typical trade-off between output and inflation and both targets
cannot be attained.
 217

This paper will therefore consider the effectiveness and credibility

issues for general CLS strategies and will provide conditions for the
existence of credible, effective CLS strategies. Section 2 will deal
with static games, Section 3 with repeated games, and Section 4
with dynamic games. An increasing degree of complexity will be
introduced and the new problems arising when repeated and
dynamic games are considered will be analyzed.
The structure of the CLS problem and its solution for linear
quadratic static and dynamic games can be found in the engineering
literature. However, this paper provides a critical evaluation of a
relevant bulk of literature and an original analysis of the credibility
problem. It will be shown that only if a certain degree of uncertainty
is introduced into the model, can the CLS strategy be made cred-
ible. Furthermore, the relationship between the CLS strategy and
other solutions of the game will be explored, and conditions for the
optimality and time-consistency of a CLS strategy will be provided.
A list of remaining open problems will conclude the paper.

2. Static Games

Let us start the analysis by showing how the closed-loop Stackel-

berg (CLS) solution can be applied to static games. In this way, the
main features of the CLS solution can be presented without resort-
ing to the complex mathematical proofs which are necessary when
dynamic games are considered. A simple example will also be used
to clarify the relationship between the CLS optimal policy and
other solutions of the game between the policy-maker and econ-
omic agents.

2.1. Closed-loop Stackelberg strategy

Let us assume that there are only two decision-makers: the
leader (player 1) and the follower (player 2). Let x, and V, =
E[W;Cs{, S2, 8)], i = 1,2, be, respectively, the decision variables and
the expected loss of the two players, where s, E 8" i = 1, 2, is the
strategy of each player and e is a vector of random variables with
given distribution representing the uncertainties introduced into the
problem. The normal form of the game is therefore defined by the
strategy sets 8, and the loss functions EJ¥" i = 1, 2. The decision
218

variables are related to the solution of the game in its normal form
by XI = sl(0J, i = 1, 2, where 0 1 represents the information set
available to each player. Furthermore, let us define player-i's
rational reaction set RI = R;(s), i, j = 1, 2, i =#; j, as
RJs) = {S, E SI: E[U-:CS" SJ' s)] ~ E[U-:(s;, Sf'S)] (2.1)
for all SI E SI}
Therefore R,(s) defines player-i's optimal reaction to player-j's
strategy and can be determined by solving the following problem:
min E[U-:(s" Sf' s)] (2.2)
s,

where S; E S; and Sf E ~, i, j = 1, 2.3

The optimal policy which is obtained by solving a standard
control problem can be interpreted as the solution of the following
problem
min (2.3)
s, E[W;(SI' R 2 (SI), s)]
where the policy-maker is the leader of the game and the reaction
function of the follower describes the economic system (see Chow,
1981, for a similar interpretation).
Before defining the closed-loop Stackelberg solution of the game,
we have to determine the team solution. If the leader of the game can
control both decision variable sets Xl and X 2 , he can achieve the
absolute minimum of his cost function by solving:
(2.4)
SI, S2

Under standard assumptions, the solution of this control problem

exists and is defined by (s\, sD.
The CLS problem can be described in the following way: find a
strategy S~ls such that
S~ = arg min E[ U2 (S~I" S2, s)] (2.5.1)
S2

(2.5.2)
where it is important to stress that S2 is included into the leader's
information set. In other words, the CLS problem is solved if the
leader can determine a strategy S~ls such that the follower is induced
to behave as if he were minimizing the leader's loss function.
 219

The solution of the follower's problem given Sl = S~ls, must indeed

be S2 = si, the leader's desired strategy: Moreover, when the
follower plays s~, the leader's rational reaction is s\ E RI (sD so that
the team solution is the outcome of the game. Therefore, the
solution of the CLS problem is a strategy S~ls such that the follower
is induced to adopt in his own interest a strategy s~ which is the most
desirable from the leader's point of view.
For example, let
(2.6)
where the function h(·) is defined as

h(e], S2, sD = {o if S2 = s~
(2.7)
sf if S2 =I- s~
where sf must be such that
arg min E[U-;(S~IS, S2, 8)] = s~ (2.8)
S2

If (2.8) is satisfied, we have S~ls = s~ and (2.6)-(2.7) imply S~ls = s\.

The function h(·) is often called threat function since it defines a set
of threats (or incentives) that are used by the leader in order to
induce the follower to choose the strategy s~. Another particular
example is given by the linear function.
S~ls = s\ + P(S2 - s~),

where the matrix P penalizes any deviation of S2 from s~. In this last
case, the CLS problem is solved by choosing a matrix P such that
(2.8) holds. In Section 4 we will discuss the advantage of choosing
a nonlinear threat function with respect to the credibility of the
leader's strategy.
A comparison between the control problem (2.3) and the CLS
problem (2.5) is straightforward. If both problems are solvable,
the solution provided by the CLS strategy is preferred by the
leader since he achieves the absolute minimum of his loss func-
tion. However, the superiority of the CLS solution is based on a
larger information set. In fact, in the control problem, the leader
announces his decision first, given his knowledge of the follower's
reaction set (i.e., the follower's loss function and the initial condit-
ions). In contrast, in the CLS problem the leader again announces
his strategy first, but he also knows the follower's actual strategy.
220

Therefore, the follower is supposed either to act before the leader

or to declare his decision to the leader before the game starts. In this
last case, the sequence of actions during the course of the game
becomes irrelevant. The particular information assumption upon
which the CLS solution is based might seem restrictive in a static
game setting, but becomes very plausible when dynamic games are
considered.

2.2. Inducibility and credibility

The closed-loop Stackelberg strategy previously described is not
well defined. Two important problems affect the existence of a CLS
strategy:
(i) the follower may prefer to be penalized by the leader rather
than adopt his desired strategy if the follower's loss is lower under
the punitive strategy, i.e., if
E[U';(sL sL B)] > E[U';(S~]s(S2)' 82 , B)] (2.9)
for some 82 E S2 and 82 =1= s~.
(ii) the threats announced by the leader and defined by the
punitive strategy S~]S(S2) for S2 =1= s~, may not be credibile if the
leader finds it advantageous to follow a strategy which differs from
the announced one when S2 =1= s~, i.e., if
(2.10)
for some 82 E S2 and 82 =1= s~, where R] (S2) defines the leader's
rational reaction to 82.
In order to study the first problem we assume that the leader
commits himself to carrying out the declared strategy and we define
the most punitive strategy s'{'P as the leader's strategy which maxi-
mizes the follower's loss function any time he does not comply with
the policy selected for him by the leader. In other words, s'{'P is the
solution of:
max E[ U'; (Sl> S2, B)] (2.11 )
S1

The follower's rational reaction to s'{'P belongs to R2(s'{'P) so that we

can conclude that, whatever strategy the leader announces, the
follower can always unilaterally guarantee himself
Bmp E[U';(s7'P, R2(s'{'P), B)]
min max E[U';(SI> S2, B)]. (2.12)
S2 SI
 221

Suppose the leader tries to achieve (sf, sf). The most powerful CLS
strategy he can announce is:

sf if S2 = sf
Scls _ { (2.13)
1 - sr'P if S2 =f- sf.
Then, the following proposition can be proved:

PROPOSITION 1: If the desired strategies (sf, sf) are such that

E[UI2(sf, sf, e)] > Bmp, the leader cannot induce the follower to
choose sf by using the CLS strategy defined by (2.13).

PROOF: From (2.12), the follower can guarantee himself a loss

which is lower than the loss he attains when he adopts the leader's
desired strategy.

Therefore, the following definition of inducible region is implied

by the above proposition:
(2.14)
In other words, the inducible region defines all points in the strategy
space that the follower prefers to the conflict with the leader,
because they imply lower losses than the disagreement loss Bmp.
The definition of sr'P also implies: -

PROPOSITION 2: If (sf, sf) cannot be induced by S~ls as defined by

(2.13), it cannot be induced by any other CLS strategy.

PROOF: Obvious, since sr'P defines the most punitive (and effective)
strategy.

This proposition also implies that IRmp is the largest inducible

region and that the minimum loss the leader can achieve without
conflict with the other player is
(2.15)

The previous conclusions hold if we assume either that the leader

commits himself to carrying out his declared threat strategy or that
the leader's threats are credible (the credibility issue will be dis-
cussed later). Under this assumption, it is also possible to prove:
222

PROPOSITION 3: If the inducible region contains the team solution

(s~, sD, the leader can achieve the global minimum of his loss
function and his announced strategy is time-consistent.

PROOF: If (s~, sD E IRmp, the solution of (2.15) is the team solution,

l.e.,

which, by definition, provides the absolute minimum of the leader's

loss function. This implies that there exists no other strategy sf
such that the leader can attain a lower loss after having observed
the follower's decision. Therefore, the actual strategy coincides
with the announced strategy and no time-inconsistency problem
anses.
As stated above, these results depend largely on the assumption
either that the leader commits himself to carrying out the
announced strategy or that the leader's threats are credible. Sup-
pose they are not. Then, the follower knows that any time he
chooses S2 =I=- s~, the leader's optimal strategy will belong to the
rational reaction set RI (S2) so that the follower's optimal strategy
when the leader's threats are not credible is defined by:

52 = arg min E[Jt;(R I (S2)S2'

S2
t:)]. (2.16)

Consequently, the leader's optimal reaction will be 51 RI (52)' so

that E[ff'; (5 1 ,52, t:)] represents, by definition of rational reaction set,
the leader's minimum loss when the follower adopts the strategy
S2 = 52' Therefore, the leader will carry out the announced strat-
egy, i.e., will adopt S~IS(52)' if and only if
(2.17)
If we assume that the minimum problem (2.1) has a unique solu-
tion, and if the leader is not committed to carrying out his threats,
so that the announced strategy must be credible, then (2.17) implies
that the only CLS outcome of the game is (51, 52)'
The following proposition summarizes the previous analysis:

PROPOSITION 4: If the leader of the game is not committed to

carrying out his threats, the only strategy he can induce the follower
 223

to adopt is S2 = S2 where

S2 = arg min E[Jt;(R I (S2)' S2, e)]. (2.18)

S2

Therefore, the leader can achieve the absolute minimum of his loss
function if and only if s~ = S2.
Notice that the solution (SI' S2) is nothing more than the standard
Stackelberg solution of the game when the follower becomes the
leader and the leader becomes the follower.
Proposition 4 also implies that the interior of the inducible region
determined by a set of credible threats is empty. Define indeed the
follower's maximum loss when the CLS strategy is credible as B,
where

Then the inducible region is redefined as

IR = {Sl E S2' S2 E S2: E[Jt;(sl' S2, e)] ~ B}.
However, by definition S2 is the best follower's strategy when the
leader adopts RJS2) so that B is also his minimum loss and no
rational follower can be induced to adopt a strategy S2 =I- S2.
Therefore, the interior of IR is empty and IR = (SI' S2).
The conclusion that can be derived from the previous analysis is
that a credible CLS solution for a static game either does not exist
or coincides with the Stackelberg solution with a reversed role for the
two players. However, this disappointing conclusion will be shown
not to hold when repeated and dynamic games are considered.
In fact, under suitable assumption, credible CLS strategies which
do not belong to the leader's rational reaction set will be deter-
mined, and it will be shown that the leader can achieve the absolute
minimum of his loss function when the inducible region contains
the team solution.
We want to emphasize that the credibility problem that we
have discussed in this section is slightly different from the credi-
bility problem which arises when the optimal policy is time incon-
sistent. Suppose indeed that the leader is committed to carrying out
his threats and that the follower knows that. Then, the follower
will choose S2 = si, where si is the strategy desired by the leader.
If sf = s~, the leader's rational reaction is s\ and the leader
achieves the minimum of his loss function (Proposition 3). Suppose,
224

however, that the team solution does not belong to the inducible
regIOn, l.e.,
E[JtS(s\, s~, e)] > BmP.
If there exists a strategy sf such that
E[JtS(sf, s~, e)] ~ Bmp
the leader's best CLS solution will be
*'f
Sl 1 S2 = S2I
Scls _ {
I - s'{'P if S2 =1= s~
The leader is committed to carrying out his threats so that the
follower will choose S2 = s~. However, the leader, after having
observed S2 = s~, will choose Sl = s\ instead of the announced
· s I = S *I . 4
po1ICY
This is the time-consistency problem as it is presented in the
traditional control and rational expectations literature. 5
As proved by Proposition 3, the time-consistency problem (and
the following credibility problem) affects the CLS solution of the
game only if the inducible region does not contain the team solu-
tion. In contrast, we will show that the standard optimal control
policy is often time-inconsistent, even when the CLS policy is
time-consistent.

2.3. An example
Let the game be described by the following matrix:
Player 2
(0, 5) (2, 2) (3, 4)
Player 1 (1, 0) (2, 5) (4, 0)
(5, 3) (1, 1) (5, 4)
When the leader follows the most punitive strategy, the lowest loss
the follower can secure for himself is Bmp = 4, so that the inducible
regIOn IS:
IRmp = {(2, 1) (3, 1) (1, 2) (2, 3) (3, 2) (1, 3) (3, 3)}.
In contrast, if the leader's threats are credible, the follower can
loose only jj = 1, so that IR might be IR = {(2, 1) (3, 2) (2, 3)}.
 225

However, the choice U I = 2 is not the leader's rational choice when

either U 2 = 1 or U 2 = 3. Therefore, IR contains only (3, 2). In
contrast, IRmp includes several possibilities. In particular, the leader
can achieve WI = 1 by inducing U 2 = 1 through the following
strategy:
if U2 = 1
= if U2 = 2 (2.15)
{;
Scls
• 1

if U2 = 3
However, this policy is not time-consistent since when the follower
has chosen U2 = 1, the leader has an incentive to pick Ul = 1 in
order to achieve U'I = 0, the minimum loss.
This is not the case if the team solution belongs to the inducible
region. Suppose that the element (1.1) of the matrix is replaced by
(1, 5) and the element (2, 1) by (0, 0). Then the strategy (2.15)
becomes time-consistent. Therefore, when (s\, sD E IRmp, the CLS
strategy is time consistent but, being defined by the inducible region
IRmp, it is based on a set of non-credible threats. We can conclude
that two conditions must be satisfied for a closed-loop Stackelberg
strategy to be credible:
(i) the inducible region must contain the team solution;
(ii) the leader is committed to his declared strategy. This could be
the result of a binding contract, an institutional arrangement or
the minimization of a long-term loss.
Of course, this last possibility (which might be introduced
through a reputation mechanism) cannot be explored by using
static games. Therefore, the next sections will discuss the CLS
solution for repeated and dynamic games.

3. Repeated Games

The simplest way of introducing a multi-stage control problem is

to assume that the game between the policy-maker and economic
agents is repeated a finite (or infinite) number of times. Each stage
depends on the previous ones only as far as the players are not
memoryless, i.e., the information set at time t, where 1 ~ t ~ N
and N ~ 00 defines the length of the game, contains the decisions
of the players in the previous stages.
226

A multi-stage framework gives us the possibility of providing

new insight into the credibility problem. In the previous section we
have shown that the ability to raise threats can greatly reduce the
leader's loss, provided that the follower is convinced that the leader
is really committed to his threat if the circumstances arise in which
he claims he would use it. However, we have seen that the action
following from the execution of threats is generally not optimal
with respect to the leader's loss function at the time of their realiza-
tion. If the game is repeated N times, however, the leader may find
it advantageous to carry out his threats in the first stages of the
game in order to induce the follower to adopt the desired strategy
in the following periods. In other words, the punitive strategy,
though irrational in a single play of a game, may well be rational
in repeated play. The reason is that a carried-out threat enhances
the leader's credibility in doing the apparently irrational thing in a
single play so that, over the long run, the leader may develop a
sufficiently fearsome reputation to deter future undesired actions by
the follower. Thereby, while losing in the short run, the leader can
gain over time. This argument, however, can be shown to be correct
only under special assumptions.
Let V; = E[W,(s], S2' 8)], i = 1,2, be the players' loss and let
(st, si) be the leader's desired strategies. Suppose (st, si) E IRmP.
The inducibility of (st , si) implies
(3.l)
where V;* , i = 1, 2, is player-i's loss when the solution of the game
is (st, si) and Vi, i = 1,2, are the losses when the leader's threats
are actually carried out. Furthermore, (.~\, 82 ) is the solution of the
game when the follower assumes that the leader's reaction will
belong to his rational reaction set at any stage of the game, i.e.,
s] E R] (S2) for any S2 E S2' The relative losses are (V;, Vz).
The normal form of the game can be described in the following
way. The follower can choose between the leader's desired strategy
sf and his optimal strategy 82 when he does not think the leader is
committed to carrying out the announced threats. The follower will
adopt sf if he believes the leader's threats. He will choose 82 ,
otherwise.
In contrast, the leader's reaction is st whenever S2 = st, but the
leader can choose between his punitive strategy s'{'P and his single-
play rational strategy 8], when S2 = 82 , The following matrix
 227

describes the outcomes of the game

Follower
si 52
Leader 5] V;*, T-';* V;, V2
s'{'P / Vi, w2
where

Equation (3.1) and the definition of 52 imply:

(3.3)
Furthermore, in the previous section we have shown that the
leader's threats are not credible if
v; < Vi· (3.4.1 )
Therefore, (3.4.1) and the definition of desired strategies (team
solution, if possible) imply:
V;* :( V; < Vi (3.4.2)
Inequalities (3.3)-(3.4) imply that the dominant strategy for the
leader is 5] so that the solution of the game is (5], 52). Indeed, when
the leader plays 5], the follower's dominant strategy becomes 52.
This solution, obtained by recursively eliminating any dominated
strategy, is called the d-solution by Moulin (1981).
Therefore, we have proved again the main result of the previous
section. However, when the game is repeated N times, it can be
argued that it pays to the leader to loose V;P - V; in some early
stage of the game in order to get V; - V;* in the following periods.
This argument is based upon the assumption that the leader estab-
lishes his reputation by punishing the follower so that in the follow-
ing periods the follower will never choose a strategy which differs
from sf. However, if the game is deterministic, this argument is not
correct. The structure of the game is indeed equivalent to Selten's
Chain Store Paradox. It was proved by Selten (1978) that the
only perfect equilibrium of a game described by the normal form
(3.2) is (5], 52) at each stage of the game. The proof starts from
228

the last period by showing that at t = N the leader has no incentive

to punish the folower when S2N =1= sX since no remaining period
exists where the leader can get V; - ~*. Therefore, at t = N we
surely have S2N = S2N' But then, at t = N - 1, the leader has no
effect on the last stage (S2N = S2N)' Therefore, at t = N - 1, we
surely have S2N _ I = S2N _ I ' This argument can be repeated at each
stage, thus proving that {(SIP S2/); t = 1, ... , N} is the solution of
the repeated game. This is the unique perfect Nash (and Stackel-
berg; see Tirole, 1983) equilibrium of the game.
However, recent papers by Rosenthal (1981), Kreps-Wilson
(1982a), Milgrom-Roberts (1982), have shown that other solutions
of the game can be determined when some uncertainty is introduced
into the model.
Let us rewrite the normal form of the game in the following way:

Follower
sf S2

V; - ~*, 0 o ~*- is
Leader 81
Vi - ~ , Vi - V2

smp
V; - Vi ~* - Vi
I / Vi - ~' Vi - ~

(3.5)

where inequalities (3.3) and (3.4) have been used to transform the
normal form (3.2) defined by the players' losses into the normal
form (3.5) defined by the players' payoffs. The positive quantity
Vi - is > 0 has been used to normalize the payoffs of each
player.
Furthermore, let us assume:

A.I. The follower is uncertain whether the punitive action will be

carried out at stage t of the game. Since the leader will punish the
follower only when ~p ~ V;, an equivalent assumption is that the
follower is not certain about the payoffs of the leader.6
 229

A.2. The CLS strategy does not satisfy condition (2.17). In other
words, the strategy that the leader wants the follower to adopt does
not coincide with .52 ,
These assumptions and the normal form (3.5) imply that the
results derived in Kreps-Wilson (1982a) can be applied to deter-
mine under what conditions the CLS solution is actually a possible
solution of the game. The following inequalities are indeed assumed
to hold:

II. ~ - Vl* ? 0 by definition of desired (team) solution;

12. VI' - T';* > 0 by definition of inducible region;

V* - V
B. 1 > 2 2 == b > 0 by assumption A.2 and the definition
VI' - V2
of inducible region;

14.
-
~
<
-
>
Vr by assumptIOn
.
A.l.

By defining a = (~ - ~* )/( ~p - ~), the game can also be

described in the following way:

Follower
si .52
Leader 51 a, 0 0, b
s~P I -1, b - 1

which coincides with the normal form of the game analysed by

Kreps-Wilson (l982a) where a > 0 if A.2 hold and b ~ I if the
inducible region contains (sf, si).
The follower's uncertainty about the leader's payoff implies that,
at each stage t, the follower assesses a probability PI that the leader's
loss function is such that Vi ~ ~,so that the leader will carry out
his threats. At stage t + 1, PI will be revised on the base of the
leader's decision at time t. The standard Bayes' rule is assumed to
be used to compute PI+ I'
Furthermore, let b be the initial probability that the leader
finds it profitable to punish the follower when S2 # si at t = I
230

(i.e., PI = <5) and assume that both players remember the moves of
the game, as the game progresses. Therefore, we are dealing with a
game with imperfect information and perfect recall (see Kreps-
Wilson, I982a).
An equilibrium concept which is analogous to Selten's perfect
equilibrium but which takes into account the uncertainty intro-
duced into the model is the sequential equilibrium described by
Kreps-Wilson (l982b).
Therefore, we want to determine the sequential equilibrium of
the game (3.5). The function Pt is defined by the following four
conditions:

(i) If S2t = si" at stage t, then Pt+ I = Pt·

(ii) If S2t of-

Pt+1 = max (b
st, Pt > °and the leader's reaction is
N - t , Pt)·
SIt = s7'f. Then

(iii) If S2t of- st and either Sit = SIt or Pt 0, then Pt+1 0.

(iv) PI = <5

Given this recursive definition of Pt, the following proposition can

be proved:

PROPOSITION 5: Suppose the game is characterized by assumptions

A.I and A.2 and the length of the game is finite. Then, the CLS
strategy
st if S2 = si
Scls - {
I - s'{'P otherwise

where (st, si) E IRmp can induce the follower to adopt S2 = si at

any stage of the game if and only if <5 > b. Furthermore, if <5 > b
the sequence {s~ = st, sit = si; t = 1, ... , N} is a sequential
equilibrium of the game and the CLS strategy is credible.

PROOF: From Kreps-Wilson (1982a), Proposition 1, we have that

Pt > bN - t + I implies S2t = st. Furthermore, 1.2 implies b ~ I
and(j> bisnecessaryand sufficient for PI > bN-I+latanyt = 1,
2, ... , N. Therefore, S21 = sit for t = 1, 2, ... , N. The leader's
 231

consequent rational reaction is SIt = sft at any t = 1,2, ... , N so

that the sequence {S0, s~; t = I, ... , N} constitutes a sequential
equilibrium which, by definition, cannot be based on non-credible
threats (see Kreps-Wilson, 1982b).

An immediate implication of Proposition 5 is the following:

PROPOSITION 6: Suppose A.I and A.2 hold. Then, if c5 > band

the inducible region contains the team solution of the static
game:

(i) the leader can achieve the absolute minimum of his multi-stage
loss function;
(ii) the leader's CLS strategy is time-consistent.

PROOF: Define sf = s\, si = si and apply Proposition 5. Then at

any stage of the game the leader achieves the minimum of his loss
function. The time consistency of the CLS strategy follows from
Proposition 3.

For any c5 ~ b other sequential equilibria can be determined by

using the leader's strategy and the follower's strategy described in
Kreps-Wilson (l982a).7
However, these other sequential equilibria may be characterized
by S2t =I siat some t, so that it is impossible for the leader to achieve
the desired solution at any stage of the game. This implies that the
CLS strategy loses its most appealing property, i.e., the absolute
minimization of the leader's loss function. Consequently, either the
CLS strategy becomes time-inconsistent or it may not be the
leader's optimal strategy. In contrast, when Proposition 5 holds, the
CLS strategy defines the leader's optimal policy with respect to any
other possible strategy, since it provides the absolute minimum of
the leader's loss function. Therefore the concept of CLS solution of
the control problem is shown to be the best way of computing the
leader's optimal policy if the uncertainty introduced into the game
is large enough to imply an initial probability of the leader's com-
mitment to his threats greater than b. In other words, if the leader's
reputation is good enough (c5 > b), then his announced policy will
be credible, Pareto optimal and time-consistent.
232

Let us examine more carefully the condition (j > h. It can be

written as

(j > V;* - is. (3.6)

Vi' - V;
so that it will be more easily satisfied when V;* is. is small and
Vi' - is. is large, i.e., when the follower's relative loss from accept-
ing the leader's desired strategy is small and the follower's relative
loss when the leader's threats are actually carried out is large.
~ Finally, we want to emphasize that (3.6) is necessary for the
leader to achieve with certainty his absolute minimum loss only if
N < 00. It is indeed possible to prove (see Kreps-Wilson, 1982a;
Milgrom-Roberts, 1982) that if N = 00, for any (j > 0 the sequen-
tial equilibrium of the game (3.5) is determined by {(sG, sir);
t = 1, ... , N)}.
Therefore, when the game is played an infinite number of times,
the uncertainty which must be introduced into the game for the
leader's CLS strategy to be credible can be very small. In contrast,
if N < 00, condition (3.6) must hold. These conclusions will be
generalized to dynamic games in the next section.

4. Dynamic Games

The previous theoretical framework can easily be generalized to

dynamic games. Therefore, the ith player is supposed to minimize:

E[~(1, Yl' Xl' X 2)] = E [~tll g;(Yt> XII' X 2/) + g'T(YT)],

i = 1,2. (4.1)
where the first argument of the function ~ indicates the first of the
planning periods, subject to the dynamic system:
(4.2)

where x; = [Xii' . . . , X iT - d, i = 1,2, and 8/ is a vector of serially

uncorrelated random variables.
Again we assume that the leader declares his strategy first, but
he acts only after having known the follower's action (or the effects
of this action). This assumption is particularly plausible when
 233

dynamic games are considered. The CLS strategy may indeed imply
a punishment from time t + lon, any time the follower does not
adopt the leader's desired strategy at time t.
The leader will therefore try to achieve the absolute minimum of
his loss function by using the optimal strategy
xt if X2t - 1 = Xit_1
Scls - { (4.3)
It - •.mp
.Alt 'f *
1 X2t-1 i=- X2t - 1

where {xt, xit; t = 1, ... , T - I} is the team solution of the

game and {x:'f; t = 1, ... , T - I} is the punitive strategy deter-
mined by solving
min max E[US(1, YI, XI' x 2)], t = 1, ... , T
X2t XII

s.t. Yt+ I = f(Yt> Xlt, x 2t ) + St· (4.4)

The solution of this problem is a function P(1, YI) so that the
inducible region can be defined as
IRmp {(Xl> x 2): E[US(t, Yo Xl> X2)] ~ B';'P
for t = 0, 1, ... , T - 2} (4.5)
where B';'P is defined as
B';'P = min E[g;(Yt> xlt> X21 )
X21
+ P(t + 1, Yt+I)]' (4.6)

In other words, the follower will verify at any time t if his loss can
be reduced by choosing X2t i=- xit. If this is the case, the leader will
use his punitive strategy from time t + Ion. Therefore, the
sequence {B';'P; t = 0, 1, ... , T - 2} defines the inducible region
for the dynamic game (see Tolwinski, 1983). It must be emphasized
that in a deterministic setting the follower's decision at the last stage
of the game cannot be influenced by any threat, so that at the last
stage of the game no policy can be induced.
A common assumption is to exclude any follower's action at the
last stage of the game (Basar-Selbuz, 1979)8 or to impose some
restrictions on the leader's loss function (see Tolwinski, 1981).
However, these assumptions affect the effectiveness and not the
credibility of the CLS strategy. Indeed, they can be used to show
that the leader's CLS strategy is effective even in the last stage of the
game, so that the leader can achieve the absolute minimum of his
loss function. However, Selten's argument can again be used to
234

show that no threat will be carried out in the last period, so that
in all the other stages of the game the follower will choose a strat-
egy which differs from the leader's desired strategy. Furthermore,
the credibility of a CLS strategy for dynamic games is related to
the type of strategy (linear, nonlinear, continuous, etc.) which is
adopted by the leader. Let us consider, for example, the solution of
the CLS problem provided by Basar-Selbuz (1979) and Tolwinski
(1981). The Basar-Selbuz CLS strategy is defined by:
(4.7)

where Yr = f(Yr-" XG-I' Xir-,), i.e., YI is the state at time t if both

decision- makers used the desired strategies at t - 1. The solution
of the CLS problem is therefore a sequence {PI' P2 , . . . , PT - d
such that (xi, xi) E IRmP. Basar-Selbuz (1979) provide the solution
for general linear quadratic control problems. However, (4.7)
implies that if X21 =I xi, at any t, then in general we have YJ =I YJ for
j > t. Therefore the follower will be punished forever once a devia-
tion, however unintentional, is observed, even ifhe returns to xi, for
i > t. This type of CLS strategy is not likely to be credible unless
x~lrs belongs to the leader's rational reaction set for any X2r and any t. 9
Indeed, if this condition is not satisfied, when the payoff from
establishing a reputation is high (the first stages of the game), the
cost of carrying out the announced threats is also very high (the
punitive strategy lasts for all future periods). In contrast, when this
cost is low, the advantage of establishing a reputation is also very
low (few periods remain for the leader to get his desired solution).
Let us consider now Tolwinski's solution. His CLS strategy is
defined by
Xlr
cis -
- Xl!* + hr ( Yr - YI
- ), t = 1, ... , T - 1, (4.8)

where hr is a nonlinear function with hi (0) = 0 and Yr is defined as

YI = f( Yr-I, Xlr_I' Xi,_I)· In this case, as long as X21 _ 1 = Xi,-I'
Yr = YI regardless of whether or not XII _ I = XG_I. Thus, if the
follower acted improperly for whatever reason at t - 2 but
resumes the desired decision at t - 1, then the leader will only
punish at t - I for one stage of the game.
Therefore, Tolwinski's strategy is more likely to be credible since
at any t < T the leader can compensate his punishment loss with
the payoffs he can obtain, in all future periods, from establishing
the credibility of his threats. 10
 235

Furthermore, by using Tolwinski's strategy, the analysis of the

previous section can be repeated simply by adding a time index to
the losses V;*, v" V;p. Therefore, if the conditions given by Kreps-
Wilson (1982a) are satisfied for any t, where bN - t is substituted by
I1~= t+ I b" and b, is defined as b, = (V2; = V21 )/( Vi, - V21 ), then the
CLS strategy is credible and attains the absolute minimum of the
leader's loss function. However, the deterministic structure of the
game cannot be maintained. The conclusions derived from Propo-
sition 5 can be applied to dynamic games only if some uncertainty
about the leader's payoff is introduced into the model. How this
uncertainty affects the solution of stochastic dynamic games is a
matter to be investigated. The general solutions of the CLS problem
provided by Basar-Selbuz (1979) and Tolwinski (1981) can be
applied only to linear quadratic deterministic dynamic games and
few attempts to solve stochastic dynamic games have appeared in
the literature (see Ho-Luh-Muralidharan, 1981; Chang-Ho, 1981
and Chow, 1981). Summing up, we can conclude that three major
ingredients are necessary to determine a credible and effective CLS
strategy:
(i) The team solution (xt, xT) must belong to the inducible
region, otherwise the CLS strategy is time-inconsistent.
(ii) Either the leader has a reputation such that the probability of
his commitment to his announced threats is high or, ifit is low,
the time-horizon is infinite.
(iii) The punishment for any follower's deviation from the desired
strategy must last a finite number of periods, and the loss for
the leader must be finite.

6. Conclusion

This paper has tried to achieve several goals: first, a new interest-
ing solution of the control problem has been presented and its main
features have been discussed. This solution, called Closed-Loop
Stackelberg, is based on an optimal announcement strategy so that
a credibility problem arises. Therefore, this paper has also shown
under what conditions the optimal announcement is credible.
Static, repeated and dynamic games have been considered.
However, several extensions of the results contained in this paper
should be provided. For instance, a general CLS solution for
236

stochastic games has not been proposed (see Chang-Ho, 1981, for
a first attempt) and the new problems arising when multi-level
games are considered have not been examined (see Luh-Chang-
Ning, 1984). Furthermore, more effective CLS strategies can be
determined when two ore more followers are introduced into the
game, so that the leader can exploit their interaction in order to
achieve his team solution (see Chang-Ho, 1983). Finally, several
problems related to the information structure of the two players
have not been considered. If, for example, the follower's strategy is
not observable by the leader, who must therefore induce the fol-
lower to reveal his actual decision, then the CLS strategy becomes
more complex and a two-sided credibility problem must be solved
(see Ho-Luh-Olsder, 1982).

Notes

I. See, for example, Basaz-Selbuz (1979), Tolwinski (1981, 1983), Chang-Ho

(1983), Chang-Luh (1984), Basar (1979), Ho-Luh-Muralidharan (1981), and
Luh-Chang-Ning (1984). It must be stressed that the CLS problem for
dynamic games does not have a simple solution. See Simaan-Cruz (1973).
2. This is not a new idea in the economic literature. See, for example, the issue
of the Review of Economic Studies (1979) devoted to the "incentive comptabil-
ity" problem and the book by Green-Laffont (1979).
3. For the sake of simplicity, we assume that the minimum problem (2.2) has a
unique solution.
4. This cheating solution, which can be considered a particular, time-inconsistent
version of a CLS strategy, has been studied by Hamalainen (1981).
5. Luh-Chang-Chang (1984) define a policy as time-inconsistent when it does
not satisfy the principle of optimality along the equilibrium desired path
(s~, s~) and define a policy as not credible when it does not satisfy the principle
of optimality off the optimal path. These definitions are consistent with our
analysis.
6. A similar assumption is used by Kreps-Wilson (l982a) in order to provide a
solution of the Chain-Store Paradox.
7. The multiplicity of sequential equilibria that can be determined may be
considered a limit of this solution concept, if no other criterium is provided
that enables us to choose between different equilibria.
8. Basar-Selbuz (1979) also provide the CLS solution without assuming the
follower does not act at the last stage of the game, but, in this case, the team
solution is not attained.
9. This condition is equivalent to the conditions required by Luh-Chang-Chang
(1984) for a CLS strategy to be credible. See also note 4.
10. However, Tolwinski's solution is highly nonlinear as will be shown later on.
 237

II. n is the dimension of the state vector y,.

12. However, an explicit discussion of the existence of a non-empty inducible
region is not provided by Basar-Selbuz (1979) and Tolwinski (1981).

References

Basar, T. (1979): "Information Structures and Equilibria in Dynamic Games", in

New Trends in Dynamic System Theory and Economics, M. Aoki and A. Mar-
zollo, eds., Academic Press, New York.
Basar, T. and Selbuz, H. (1979): "Closed-Loop Stackelberg Strategies with Appli-
cations in the Optimal Control of Multilevel Systems", IEEE Transactions on
Automatic Control 166--179.
Chang, T. S. and Ho, Y. C. (1981): "Incentive Problems: A Class of Stochastic
Stackelberg Closed-Loop Dynamic Games", Systems and Control Letters,
16--21.
Chang, T. S. and Luh, P. B. (1984): "Deriviation of Necessary and Sufficient
Conditions for Single-Stage Stackelberg Games via the Inducible Region Con-
cept", IEEE Transactions on Automatic Control, 63--66.
Chow, G. C. (1981): Econometric Analysis by Control Methods. John Wiley &
Sons, New York.
Green, R. and Laffont, J. J. (1979): Incentives in Public Decision-Making. North-
Holland Publishing Company, Amsterdam.
Hamalainen, R. P. (1981): "On the Cheating Problem in Stackelberg Games",
International Journal of Systems Science, 753-770.
Ho, Y. c., Luh, P. B., and Muralidharan, R, (1981): "Information Structure,
Stackelberg Games and Incentive Controllability", IEEE Transactions on Auto-
matic Control, 454--460.
Ho, Y. c., Luh, P. B., and Olsder, G. J. (1982): "A Control Theoretic View on
Incentives", Automatica 18, 167-179.
Kydland, F. (1975): "Noncooperative and Dominant Player Solutions in Discrete
Dynamic Games", International Econometric Review, 321-335.
Kydland, F. and Prescott, E. C. (1977): "Rules Rather than Discretion: The
Inconsistency of Optimal Plans", 473-491.
Kreps, D. M. and Wilson, R. (1982a): "Reputation and Imperfect Information",
Journal of Econometric Theory 25, 253-279.
Luh, P. B., Chang, S. c., and Chang, T. S. (1984): "Solutions and Properties of
Multi-Stage Stackelberg Games", Automatica 20, 251-256.
Luh, P. B., Chang, T. S., and Ning, T. (1984), "Three Level Stackelberg Decision
Problems", IEEE Transactions on Automatic Control, 280--282.
Milgrom, P. and Roberts, T. (1982): "Predation, Reputation and Entry Deter-
rence", Journal of Econometric Theory, 27, 280-312.
Moulin, H. (1981): Theorie desjeux pour !'economie et la politique. Hermann, Paris.
Rosenthal, R. W. (1981): "Games of Perfect Information, Predatory Pricing and
the Chain-Store Paradox", Journal of Econometric Theory 25,92-100.
Selten, R. (1978): "The Chain-Store Paradox", Theory and Decision 9, 127-159
238

Simaan, M. and Cruz, J. 8., Jr. (1973): "Aditional Aspects of the Stackelberg
Strategy in Nonzero Sum Games", Journal o/Optimization Theory and Applica-
tion, 613-626.
Tirole, J. (1983): "Jeux dinamiques: un guide pour I'utilisateur", Revue d'Economie
Politique 4, 550-575.
Tolwinski, B. (1981): "Closed-Loop Stackelberg Solution to a Multistage Linear-
Quadratic Game", Journal of Optimization Theory and Application, 485-501.
Tolwinski,8. (1983): "A Stackelberg Solution of Dynamic Games", IEEE Trans-
actions on Automatic Control, 85-93.
 239

CHAPTER 14

A DYNAMIC AND STOCHASTIC MODEL OF PRICE

LEADERSHIP

B. Fanchon,' E. Rifkin 2 and J. K. Sengupta 3

, University of Nevada-Reno, U.S.A.; 2 Federal Trade Commission,
Washington, U.S.A.; 3University of California at Santa Barbara, U.S.A.

1. Introduction

This paper examines the optimal output strategy of a dominant

firm or a cartel attempting to maximize the present value of its
stream of profits from a market with potential entry. The rate of
entry into the industry is assumed to be a function of the current
price only and entry is defined as an increase in output from
cOE;>etitors which might or might not be already in the market. In
the dominant firm model, the dominant firm quotes the market
price and competitors supply all they want at that price. The
dominant firm then supplies the demand not met by rivals at the
given price. If the barriers to entry are great, the dominant firm can
produce the short run profit maximizing level of output with little
fear of losing its market share. However if entry is relatively easy,
the firm can increase its output to the point where the price corre-
sponding to aggregate industry supply induces no entry. This price
has been defined as the limit price. Progress in optimal control
theory has revived interest in limit pricing theory and numerous
new dynamic models have been formulated [1, 3-7, 11] where
different limit price paths correspond to various assumptions about
the behaviour of the entrants and about dominant firm's knowledge
of the parameters of the aggregate demand and of the cost functions
of existing and potential rivals. However if the demand or the
production of rivals is not known precisely, the dominant firm must
adjust its output in order to maintain the quoted market price.
While this classic formulation leads to high uncertainty about the
planned level of output, and no uncertainty about the price, such a

C. Carraro and D. Sartore (eds). Developments of Control Theory for Economic Analysis
© 1987 Martinus NijhofJ Publishers (Kluwer), Dordrecht
240

formulation can be very costly to the firm if it operates in a region

of steep marginal cost or if the production process takes a signifi-
cant amount of time. Since the price is ultimately determined by the
market, and not by the firm, price is the wrong control variable.
If the market allows small fluctuations about the quoted price,
the dominant firm has more room for error in determining the
optimal price and the most relevant demand function is then of the
form: q = j{p, e, t) where e is an error term. The dominant firm
can then quote its expected price and plan output at the same time,
hoping that after the market makes the final price adjustment,
actual price and expected price will coincide.
This paper departs from the now traditional formulation of
dynamic limit pricing on two points: First; it is assumed that the
dominant firm has imperfect knowledge of the parameters of the
market and determines its planned output as a function of expected
price. Second; it is assumed that the inverse demand function
contains a term reflecting expected market growth. The inverse
demand function also contains an additive error term, with known
stochastic properties and independent of time, representing uncer-
tainty about the aggregate demand at any price.

2. Stochastic Market Growth Model

The suppliers of a commodity are divided into two groups; the

dominant firm or a cartel (also called the "dominant" firm in this
paper), and other suppliers called the rivals (or "fringe"). The
equilibrium market price of the product at time t is determined by
the level of output of the firm and of rivals or entrants, and grows
exponentially. The market demand function is assumed to be of the
form:
pet) = a ent - b l (q(t) + x(t» + u, (1)
where pet) is the price of the commodity, q(t) is the output of the
dominant firms, x(t) is the output of the "fringe", a and hI are posi-
tive constants, n is the growth rate of demand, and u is an error term
independently distributed with known mean zero, and variance v.
The dominant firm estimates that the rate of entry (or exit) of new
firms is defined by:
x(t) = k(E[p(t)] - p), x(O) = X o, (2)
 241

where k is a positive response coefficient which reflects in some

sense the willingness of the dominant firm to give the newcomers
any part of the market (the smaller the value of k, the greater the
risk of failure for a potential entrant), p is the (fixed) limit price
defined as the price at which net entry equals 0, and E['] is the
expectation operator.
The objective function (J) of the dominant firms is to maximize
the present value of the utility of expected profits given by;
U(E[n(t)]) = E[n(t)] - m var (n(t)), (3)
where m is a risk aversion parameter, c is the average total cost of
production of the dominant firm, assumed to be constant over time,
and
E[n(t)] = (E[p(t)] - c)q(t) (4)
Var net) = vq2(t). (5)
Note that since the dominant firm is producing in a region of con-
stant average cost, the firm can capture most of the increase in
demand with no change in price. The limit price in this model is there-
fore assumed constant (the results derived are no longer true if the
dominant firm is operating in a region of non-constant average
production cost. Such a case will be treated in a separate paper).
The optimal output strategy of the dominant firm, and the corre-
sponding price path, is thereby obtained by maximizing the expected
utility of the net payoff J over a infinite horizon. Hence the objective
function of the dominant firm over the infinite horizon is:

J roo e-rl[(aenl - bJq(t) - bJx(t) - c)q(t) - mvl(t)]dt

Jo (6)
J fooo e- rl [a e l - bJx(t) - b2 q(t) - c]q(t) dt

where r is the discount rate of the dominant firm (assumed exogen-

ous) and
(7)
The Hamiltonian of the problem and the corresponding optimiz-
ing conditions are:
H = e- rl (a ent - bJx(t) - b2 q(t) - c)q(t)
+ ky(t)(a enl - b J x(t) - b J q(t) - p), (8)
242

where yet) is the costate variable which represents the "optimal"

shadow price of entry to the dominant firm.
After we make the transformation yet) = z(t) e- rt , an optimal
output strategy is defined by:
x*(t) k(a ent - b] x*(t) - b] q*«t) - p), (9)

oH(x*, z*, q*) *( )

i*(t) - ox*(t) + rz t, (10)

b]q*(t) + (b]k + r)z*(t) (11)

lim z*(t) = K, where K is a constant (12)
t-+oo

oH(x*, z*, q*)

oq*(t)
- 2b 2q(t) - c) - kb] yet) O. (13)
If we make the following transformations
X(t) = x(t) e- nt , Q(t) = q(t) e- nt , z(t) = z(t) e- nt •
(14)
Substitution of (13) into (9) and (10) yields the system:
X*(t) - (kb] + n - kAo)X*(t) + k 2AoZ*(t) + A] (t) (15)
Z*(t) - AoX*(t) + (kb] + r - n - kAo)Z*(t) + A 2(t),
(16)
where Ao = bi!2b2, A] (t) = k(a - aAolb] - (p + cAo/b]) e- nt ),
and A2(t) = (Aolb])(a - c e- nt ). The characteristic equation of
the system and its roots are:
f(A) = A2 + (2n - r)A
+ (n 2 - nr - b]kr - b]2~ + rkAo + 2~b]2Ao) (17)

A], A2 = - (1/2)(2n - r ± d]/2) (18)

where
d = 4(b]kr + bik2 - rkAo - 2b]k2A o) + r.
PROPOSITION 1: The optimal level of output q*(t) is always above
the short run payoff maximizing level of output qO(t) at every point
along the optimal trajectory.
 243

(i.e. if the dominant firm were to abandon the optimal policy and
revert to short run profit maximizing, it would always reduce
output, and price and entry would increase).

PROOF: Assuming that the dominant firm has been moving along
the optimal trajectory, at time t the short-run payoff is
h(q) = [a en! - blx*(t) - b2 q(t) - c]q(t), (19)
where x*(t) denotes that x(t) is on the optimal path. Maximizing
this short-run payoff h(q) leads to the myopic optimal output rule
aCt):
(20)
The long run optimal level of output q*(t), must also maximize
the Hamiltonian H at every point on the optimal path (equation
(13)). This leads to
(21)
where z*(t), the shadow price of entry, is necessarily a negative
quantity. Since k and b l are positive parameters, comparison of (18)
and (19) implies that q*(t) > qO(t). QED

PROPOSITION 2: An increase in risk aversion leads to a decrease in

myopic and long run output, with a corresponding increasing in
price.

PROOF: From (18) and (19), we have:

oqO(t)jom - qO(t)(vj2b 2 ) < 0
(22)
oq*(t)jom - q*(t)(vj2b 2 ) < 0

where the risk aversion parameter (m) and the variance of the error
term (v) are defined in (3) and (4), and by b2 = b l + mv. Since
prices are inversely related to output, the myopic price pOet)
is higher than the long run optimal price p*(t), and we have
opo(t)jom > 0 and op*(t)jom > O. QED.

PROPOSITION 3: Under suitable conditions, an increase in risk

aversion or greater uncertainty about the demand accelerates the
convergence of the price-output vector to the steady state.
244

PROOF: The general solution of (15) and (16) is;

x*(t) C] e)qt + C2 e A2t + (J(o(t) (23.1)

Z*(t) (J(] C] e A1t + (J(2C2 e ht + (J(3(t), (23.2)

where the (J(,'S are known functions of the given parameters of the
problem, and C,'s are suitable functions determined by the bound-
ary conditions. From (18) we have;
lim Ai = -1/2(2n - r + (2b] k ± r» (24)
br+oo

and the largest root, say A], approaches (b] k + r - n) as t - 00.

Hence whenever the response coefficient k satisfies kb] > (n - r)
increasing b2 will eventually yield a positive root. In this case the
system has a stable solution only if the initial conditions guarantee
that C] = 0 and (J(3 is bounded. (The transversality condition (12)
is met only if C] is zero and (J(3(t) is bounded for t - (0). The
preceding system reduces to:

x*(t) C2 e A2t + (J(o(t) (25.1)

Z*(t) 2C2 eA21 + (J(3 (t) (25.2)

where (J(o(t), (J(3(t) tend to suitable constants as t - 00. Since the

discriminant of the characteristic equation (17) is positive, an
increase in b2 leads to a higher absolute value of 12 (which is
negative) and therefore increases the rate of convergence to the
steady state. QED.

3. Stability Conditions and Long Run Analysis

After the change of variables (14), the reduced form of the system
(9}-(l3) is:
X*(t) - (kb] + n)X*(t) - kb] Q*(t) + k(a - p e- nt ) (26)
Q*(t) (2b 2 )-] (2kbf + rb])X*(t) + [b]k + r - n]Q*(t) - A(t)
(27)
where
 245

These equations may be interpreted as the optimal reaction curves

of the dominant firm and the potential rivals in differential form.
Note that in contrast with Gaskins' model, an optimal path for
outputs can be derived from equations (26) and (27) (several cases
are considered in the next section).
More insight about the long run properties of the model is
provided by equations (26) and (27). At the steady state (X* =
0* = 0), the long run market share of the dominant firm is given by:
S = n(kb l - n + r)[rkb l - n2 + nr + 2k(kb l - n + r)mv]-l
(29)
which explains why the difference between the discount rate of the
dominant firm (r) and the growth rate of the market (n) has also a
crucial impact on the long run market share of the firms. It is easily
shown that whenever r < n, a decrease in long run market share
will result from an increase in risk aversion (m) or an increase in
uncertainty (v) or an increase in the response coefficient (k) or a
decrease in the growth rate (n).
These results are consistent with Sylos' and Gaskins' models. In
contrast with earlier models, the price behaviour is not always
clearly established (Gaskins' model predicts a positive relationship
between Sand k; his assumption that the response coefficient takes
the time-varying form k = ko enl reduces the resulting system of
differential equations corresponding to equations (26) and (27) to
an autonomous system, which in turn allows specific conclusions
about long run price behaviour).
It can also be proved that changes in the market share of the
dominant firm is positively related to changes in the price elasticity
of demand. This result is consistent with Pashingian's model which
predicts that the rate of increase in the market share of the domi-
nant firm is negatively related to changes in the estimated price
elasticity of demand.

4. Simulation of Convergent Paths

The general solution of the system (26)-(27), when entrepreneurs
are risk neutral (b l = b2 = b), is given by:
X*(t) = Bl e A1t + B2 e ht + e-nt(bkc + rc - bkp - 2rp)(rb)-1
+ ak(r - n)(bkr + 2nr - 2n2)-1 (30)
246

z*(t) (l/bk 2)[(bk + r + (r2 + 2bkr)I/2)BI ei l l

+ (bk + r - (r + 2bkr)I/2)]B2 eJ. 2 1
+ (l/r)(c - p) e- nt - an/(bkr + 2nr - 2n2), (31)
where B, and B2 are determined by the initial values X*(O) and
Z*(O). It can be shown that stability of the solution without oscilla-
tions is guaranteed by the condition
12n - rl > I(? + 2bkr) 1/2 I. (32)
Four cases of simulated trajectory are presented below. In all
these cases, the parameter values are so chosen as to satisfy the
necessary condition (32) of stability without oscillation. In all these
cases, the marginal cost (c), initial output (X(O», limit price (p) and
static demand parameters (a and b) were kept fixed, and only the
response coefficient (k), the discount rate (r) and market growth
rate (n) were changed. In all cases, the initial output of rivals (Z(O»
was set at zero. The simulations reveal the following characteristics
of the model:
(a) The time to converge to the steady state depends upon X*(O),
Z*(O), and upon the extent by which (2n - r) exceeds the
absolute value of (r2 + 2bkr)1/2. An increase of Z *(0) or of n
will speed the rate of convergence to the steady state.
(b) The output and pricing policy is found to be very sensitive to
variations in k (the strength of potential entry parameter) but
not so much to the variations in the discount parameter (r) or
market growth (n). As plots 1 and 3 illustrate, a change in the
growth rate of demand and the discount rate does not change
the price and output paths significantly when the response
coefficient is large (k was set at 0.25 for these two simulations).
As shown in the second simulation, a firm with a low response
coefficient will rapidly loose market share and the market price
will increase at nearly the same rate as the growth rate of
demand. In such a case, the position of price leadership of the
firm will quickly be challenged by rivals and the initial assump-
tions of the model might no longer hold.
In all simulations, the firm initially loses market share. In the
long run, the firm can completely loose its market (simulation 2),
maintain a fairly constant share (simulation 4), or slowly regain the
initial loss after the entry of rivals (simulations 1 and 3).
 247

4. Conclusions

In contrast with most existing limit price models, output rather

than price is the major decision variable. In this model, price lies
somewhere between the short-run monopoly price and the com-
petitive price, the exact positioning depends on the barriers to entry,
risk aversion and other behavioural attributes of the market. This
model helps to explain why excessive profits can exist over time, but
its predictions about specific dominant firm's behavior are not
explicit. Unlike Gaskin's model, no definite relationship can be
established between the quality of information about the demand
parameters and price.
It was also found that an increase in risk aversion by the domi-
nant firm leads to a lower short and long run output level, leads to
a lower price in the long run, and accelerates the rate of convergence
to the steady state. The simulated profiles show that in cases when
there is a positive market growth. A variety of phases are possible
in the optimal strategy space of the dominant firm which can in the
long run increase, maintain or reduce its output.
248

Table I
Time e -nl x*( t) e- ni z*(t) e- nl q*(t) p*(t)
0.0 1.0 5.1 140.0 5.51
10.0 10.3 68.2 144.0 5.36
20.0 15.9 111.0 149.0 5.34
30.0 19.3 158.0 134.0 5.41
40.0 21.5 204.0 129.0 5.50
50.0 23.0 249.0 123.0 5.62
60.0 24.2 282.0 117.0 5.75
70.0 25.1 335.0 112.0 5.89
80.0 26.0 377.0 106.0 6.03
90.0 26.7 418.0 101.0 6.17
100.0 27.4 458.0 95.6 6.31
110.0 28.1 498.0 90.3 6.45
120.0 28.8 538.0 85.1 6.59
130.0 29.4 576.0 79.9 6.72
140.0 30.0 614.0 74.8 6.85
150.0 30.6 652.0 69.8 6.99
160.0 31.2 689.0 64.9 7.11
170.0 31.8 725.0 60.1 7.24
180.0 32.4 761.0 55.3 7.37
190.0 33.0 797.0 50.6 7.49
200.0 33.6 831.0 45.9 7.61
Note: The parameters are: x*(O) = 1.00, b = 0.03, k = 0.25, c = 1.0, r = 0.15,
p = 2.0, a = 10.0, n = 0.055. Time step size is 10 units, number of steps is 21, and
z*(t) is adjusted for sign.

Table 2
Time e- nl x*(t) e- nl z*(t) e- nl q*(t) p*(t)
0.0 1.00 0.5 150.0 5.48
10.0 0.601 58.0 156.0 5.29
20.0 0.361 112.0 160.0 5.17
30.0 0.217 163.0 163.0 5.11
40.0 0.130 213.0 164.0 5.06
50.0 0.782E-Dl 262.0 165.0 5.04
60.0 0.470E-Dl 318.0 166.0 5.02
70.0 0.283E-Dl 357.0 166.0 5.01
80.0 0.170E-Dl 483.0 166.0 5.01
90.0 0.102E-Dl 449.0 166.0 5.00
100.0 0.619E-D2 494.0 167.0 5.00
110.0 0.376E-D2 539.0 167.0 5.00
120.0 0.320E-D2 584.0 167.0 5.00
130.0 0.142E-D2 627.0 167.0 5.00
140.0 0.891E-D3 617.0 167.0 5.00
150.0 0.574E-D3 614.0 167.0 5.00
160.0 1.384E-D3 757.0 167.0 5.00
170.0 0.170E-D3 799.0 167.0 5.00
180.0 0.201E-D3 841.0 167.0 5.00
190.0 0.160E-D3 882.0 167.0 5.00
200.0 0.135E-D3 923.0 167.0 5.00
Note: The parameters are: x*(O) = 1.00, b = 0.03, k = 0.01, c = 1.0, r = 0.05,
P = 2.0, a = 10.0, n = 0.051. Time step size is 10 units, number of steps is 21, and
z*(t) is adjusted for sign.
 249

Table 3
Time e- nt x*(t) e- nl z*(t) e- nl q*(t) p*(t)
0.0 1.0 1.7 149.0 5.49
10.0 8.2 86.4 150.0 5.27
20.0 10.0 107.0 146.0 5.30
30.0 12.2 155.0 140.0 5.47
40.0 12.8 203.0 135.0 5.58
50.0 13.4 251.0 129.0 5.74
60.0 13.4 298.0 122.0 5.91
70.0 14.2 345.0 110.0 6.25
80.0 14.6 439.0 104.0 6.42
90.0 15.0 439.0 104.0 6.42
100.0 15.4 485.0 98.3 6.59
110.0 15.9 532.0 92.3 6.76
120.0 16.3 578.0 86.3 6.92
130.0 16.7 624.0 80.3 7.09
140.0 17.1 670.0 74.4 7.26
150.0 17.5 716.0 68.5 7.42
160.0 17.9 761.0 62.6 7.59
170.0 18.3 806.0 56.7 7.75
180.0 18.6 852.0 50.9 7.91
190.0 19.0 807.0 45.1 8.08
200.0 19.4 941.0 39.3 8.24
Note: The parameters are: x*(O) = 1.00, b = 0.03, k = 0.25, c = 1.0, r = 0.10,
p = 2.0, a = 10.0, n = 0.104. Time step size is 10 units, number of steps is 21, and
z*(t) is adjusted for sign.

Table 4
Time e- nl x*(t) e- nl z*(t) e- nl q*(t) p*(t)
0.0 1.00 2.2 149.0 5.49
10.0 4.23 55.0 153.0 5.28
20.0 5.97 98.1 154.0 5.20
30.0 6.89 134.0 154.0 5.17
40.0 7.41 165.0 153.0 5.18
50.0 7.71 192.0 153.0 5.19
60.0 7.90 216.0 152.0 5.22
70.0 8.01 236.0 151.0 5.24
80.0 8.10 254.0 150.0 5.26
90.0 8.16 270.0 149.0 5.28
100.0 8.21 284.0 148.0 5.30
110.0 8.24 296.0 148.0 5.32
120.0 8.28 307.0 147.0 5.34
130.0 8.32 316.0 147.0 5.35
140.0 8.33 325.0 146.0 5.36
150.0 8.35 332.0 146.0 5.37
160.0 8.37 338.0 146.0 5.38
170.0 8.39 344.0 145.0 5.39
180.0 8.40 349.0 145.0 5.40
190.0 8.41 354.0 145.0 5.40
200.0 8.42 358.0 145.0 5.41
Note: The parameters are: x*(O) = 1.00, b = 0.03, k = 0.10, c = 1.0, r = 0.05,
P = 2.0, a = 10.0, n = 0.064. Time step size is 10 units, number of steps is 21, and
z*(t) is adjusted for sign.
250

References

1. Alberts, W. W. (I 984): "Do OIigopolists Earn "Noncompetitive" rates of

return?," American Economic Review, 74.
2. Arrow, K. J. and M. Kurz (1970): Public Investment, the Rate of Return and
Optimal Fiscal Policy, Baltimore: John Hopkins Press.
3. Gaskins, D. (1971): "Dynamic Limit Pricing: Optimal Pricing Under Threat
of Entry", Journal of Economic Theory 3, 306-322.
4. Kamien, M. and Schwartz, N. (1981): Dynamic Optimization, North Holland.
5. 5.0kuguchi, K. (1976): Expectations and Stability in Oligopoly Models, Sprin-
ger Verlag.
6. Osborne, D. K. (1964): "The Role of Entry in Oligopoly Theory," Journal of
Political Economy, 396-402.
7. Pashigian, B. P. (1958): "Limit Price and the Market Share of the Leading
Firm", Journal of Industrial Economics, 165-177.
8. Rifkin, E. J. (1978): "Dynamic Limit Price Theory and Market Growth".
Unpublished Ph.D. Dissertation, University of California, Santa Barbara.
9. Sengupta, J. K. (1978): "Stabilization Policy with Stochastic and Nonlinear
Elements", Zeitschrift fur Nationalokonomie 38, 119-144.
10. Sengupta, J. K. (1978): "Noncooperative equilibria in Monopolistic Com-
petition under Uncertainty", Zeitschrift fur Nationalokonomie 38, 193-208.
II. Shupp, F. R.: "Dynamic Limit Pricing in a Mature Market", in Mathematical
Economics and Game Theory, ed. by R. Henn and O. Moeschlin, 435-445.
12. Sylos-Labini, P. (1962): Oligopoly and Technical Progress (translated from the
Italian by E. Henderson); Cambridge: Harvard University Press.

Appendix: Derivation of Formulae

Derivation of 15 and 16
From (14);
x (0.1)

(X + nX) e r (0.2)

Use (0.2) and (14) in (11);

(X + nX) e l = k(a ent - b] X e'lr - bzQ ent - p) (0.3)
X = ka - (kb] + n)X - kb]Q - kp e- nl • (0.4)

From (13):
o. (0.5)
 251

Using (14);
(a - 2b 2Q - b]X - c e- nt - b]kZ) e(n-r)t o (0.6)
Q = (2b z)-](a - b]X - b]kZ - ce- nt ). (0.7)
Substitution of (A7) into (A4) yields:
X = ka - (kb] + n)X - (kb] /2b z)
x (a - b]X - b]kZ - c e- nt ) - kp e- nt (0.8)
X - (kb] + n - k(bf/2b 2 ))X + e(bf/2b z)Z
+ k(a - (ab] /2b 2 ) - (p + b] c/2b 2 )) e- nt • (0.9)
Using Ao and A] as described below (16) yields equation (15).
Using (14) in (11);
(2 + nZ) e nt = b] Q e nt + (b] k + r)Z e nt (0.10)
2 = (b] k + r- n)Z + b] Q. (0.11 )
Using (0.7) in (0.11);
2 = (b] k + r - n - (b] k/2b 2 ))Z - (bf/2b z )X
(0.12)
+ (b] /2b 2 )(a - c e- nt ).
Using Ao and A2 as described below (16) yields equation (16).

Derivation of 17, 18 and 32

-(kb] +n- kAo) -), k2AO ]
f(),) = det [
- Ao kb] +r- n - kAo - ),
(0.13)
f(),) = ),2 - (kb] + r - n - kAo - kb] - n + kAo)
- (kb] - kAo + n)(kb] - kAo - n - r) + k2A~
(0.14)
.A,z + ),(2n - r) - (kb] - kAo)2 + n2 - rkb]
- nr + rkA o + k 2A~
),2 + ),(2n - r) - k2b~ - k 2A~ + 2eb]Ao + n2
- rkb] - nr + rkA o + e A~
252

A? + A(2n - r)

+ (2k 2 b,Ao + n2 + rkA o - k 2bf - rkb, - nr). (17)

The discriminant d of equation (17) is:

d (2n - r)2 - 4(2k2b,Ao + n2 + rkA o - k 2bT - rkb, - nr)

r2 + 4n2 - 4nr + 4(-2k2b,Ao - n2 - rkA o

+ ~bT + rkb, + nr)

(0.15)
Hence equation (18). If b, = b2 , then Ao = (lj2)b, and using this
result in (18) leads to;

A" A2 = - (1j2)(2n - r± (r2 + 4~bT + 4rkb l

- 2rkb, - 4k2 bD'/2

(0.16)
A, A2 = - (1j2)(2n - r± (r2 + 2rkbd'/2).

Hence (32).

Derivation of 26 and 27
From (13), we have 8Hj8q = 0, hence
(0.17)

(0.18)
and taking the derivative with respect to time yields
z = (b, k)-' (na ent - bl X - 2bA). (0.19)
Using (14) in (2);
(X + nX) ent = k(a ent - bl Q enl - b, X e t - p) (0.20)

X = - kb, Q - (kb, + n)X + k(a - p e- nt ). (26)

Substitution of (11) in (0.19) gives;

b,q + (b1k + r)z = (na ent - blx - 2b 2iJ)(b l k)-' (0.21)
 253

and using (0.18);

b~kq + (b]k + r)(a enl - b]x - 2b 2 q - c)
naenl - 2b 2 q - b]k(ae nt - b]q - b]x - p) (0.22)
2b 2 q (- bi k + 2b] b2 k + 2rb 2 + bi k)q
+ (bTk + b]r + bik)x - (b]k + r)(ae nl - c)
(0.23)
q (b]k + r)q + (2b 2 )-](2bik + b]r)x
+ (2b 2 )-][a e/(n - 2b]k - r) + b]k(p + c) + rc] (0.24)
and using (14) leads to;
Q = (b]k + r - n)Q + (2b 2 )-](2bik + b]r)X
+ (2b 2 )-][a(n - 2b]k - r) + (b]k(p + c) + rc) e- nl ].
(27)

- kb] ) (X\ (k(a - pe- nt )

(kb] + r - n) Q} + -A(t)
(0.25)
where A(t) is defined in equation (30). As t becomes infinitely large,
the system reduces to:

(
-(kb] + n) -kb] )(X\
(2kbi + rbd 2b 2 (kb] + r - n) Q}

+ (- a(2kb]k~ n + r») = (~) (0.26)

b kbi(2kb] + r) - 2b 2 (kb] + n)(kb] + r - n) (0.27)

X [2kab 2 (kb] + r - n) - akb] (2kb] - n + r)](-I) (0.28)
Q [a(kb] + n)(2kb] - n + r) - ka(2kbi + rb])]C I ) (0.29)
S Q(X + Q)-I = [(kb] + n)(2kb l - n + r) - k(2kbi + rb l )]

x [2kb 2 (kb] + r - n) - kb l (2kb l - n + r)r l (0.30)

254

- 2nkb2 + 2rkb2 - 2k2 hi - rkb]-' (0.31)

S n(kb , - n + r)[rkb , - n2 + nr + 2k(kb, - n + r)mv]-I
(29)

Derivation of (30) and (31)

With b = b, = b2, the system (15}--(16) can be written:

(o1 0)1 (~)

Z
+ (kb/2) + n
b/2
2
k b/2 ) (X) (AI)
- (kb/2) + r - n Z = A2
(0.32)
where Al = ak/2 + (e/2 - p)k e- nt , and A2 = (1/2)(a - e) e- nt •
The characteristic equation of the associated homogeneous system
is given by det (M) = 0 where

M __ (kb/2 + n - A k 2b/2 )
(0.33)
b/2 -(kb/2) + r - n) - A
After some manipulations;
A2 + (2n - r)A + «- bkr/2) - nr + n2) o (0.34)
which has roots:
- n + (r/2) + (1/2)(r2 + 2bkr)I/2 (0.35)
- n + (r/2) - (1/2)(r2 + 2bkr)I/2. (0.36)
Assuming a solution of the form X = C eAt, Z = KC e)·t, a non-
trivial particular solution must satisfy

(0.37)

The constants K, (i = 1, 2) associated with the eigenvalues A,

(i = 1, 2), solution of the above equation are:
(bk +r + (r2 + 2bkr)I/2)/(bk2) (0.38)
(bk +r - (r2 + 2bkr)I/2)/(bk2). (0.39)
 255

The particular solutions of the homogeneous system are:

Klc l e-).II (0.40)
K z C2 e- A21 (0.41)
and the general solution is:
X C l e- A1I + Cz e- A21
(0.42)
z
Solving for a particular solution of the inhomogeneous system by
the method of variation of constants;
C; e A1I + C~ e A21
= Al (0.43)
Kl C; e A1I + KzC~ e A21 = A z· (0.44)
Using Cramer's rule, it comes:
C; = (AIKI - A 2)(K2 - Kl)-l e- A1I (0.45)
C~ (AIKI - A z)(K2 - Kl)-l e- A21 • (0.46)
After substitution of AI' A 2 , K l , K 2 , the two constants can be
integrated, which yields:
Cl = ak(r - d)(2(r2 + 2bkr + rd - 2nd) - I e(2n - r - d)I/2

+ «c - 2p)(bk + r - d) + bkc)(2(r2 + 2bkr + rd))-lk

x e(d - r)I/2 + Bl (0.47)
C2 ak(r + + 2bkr - rd + 2nd)-1 e(2n-r+d)I/2
d)(2(r2
+ «c - 2p)(bk + r + d) + bkc)(2(r2 + 2bkr - rd))-lk
x e(d-r)t/2 + B2 (0.48)
where d = (r2 + 2bkr)I/2 and B l , B2 are constants of integration.
Combining the general and particular solutions for X and Z yields,
after simplification:
X = Bl e(r-2n+d)I/2 + B2 e(r-2n-d)t/2 + (bkc + cr - pbk - 2pr)
x (br)-l e- nl + ak(r - n)(bkr +2nr - 2n2)-1 (30)
Z (bk ++ d)(bk2)-1 Bl e(r-2n+d)t/2
r
+ (bk + r - d)(bk 2)-IB2 e(r-2n-d)I/2
+ (c - p)r- 1 e- nl - an(bkr + 2nr - 2n2)-I. (31)
256

The constants Bl and B2 can be obtained from the initial conditions;

X(O) = Bl + Bl + B2 + (bkc + cr - pbk - 2pr)(br)-1
+ ak(r - n)bkr + 2nr - 2n2)-1 (0.49)

Z(O) (bk + r + d)bk2Bl + (bk + r - d)(bk2)-1 B2

+ (c - p)r- 1 - an(bkn + 2nr - 2n2)-1 (0.50)

and using Cramer's rule yields;

Bl = (-2d)-I(X(0) - (bkc + cr - pbk - 2pr)(br)-1
- ak(r - n)(bkr + 2nr - 2n2)-I)(bk + r - d)

- b~(Z(O) + (c - p)r- 1 - an(bkn + 2nr - 2n2)-I)

(0.51)
B2 = (2d- 1)« - bk2)(Z(0) - (c - p)r- 1
+ an(bkn + 2nr - 2n2)-1

+ (X(O) - (bkc + cr - pbk - 2pr)(br)-1

- ak(r - n) (bkr + 2nr - 2n2)-I)(bk + r + d). (0.52)
 257

1000

100

10

20 40 60 80 100 120 140 160 180

Plot 1
258

100

10

Plot 2
 259

1000

100

10

Plot 3
260

1000

100

10

20 40 60 80

Plot 4
 261

CHAPTER 15

QUALITY AND REPUTATION POLICIES OF

DUOPOLISTS UNDER ASYMMETRIC INFORMATION

Klaus Conrad
University of Mannheim. FRG

1. Introduction

In this paper we consider a duopolistic industry the product of

which is characterized by asymmetric information. The product of
each producer is differentiated by quality which can be evaluated by
consumers only after purchase (experienced goods). Each duopolist
maximizes his profit with respect to price and quality, the variables
under his control. Under asymmetric information two special
features have to be taken into account in finding the equilibrium
situation. First of all, as higher quality is more costly to produce but
can not be incorporated into the price as consumers can not observe
the better quality before purchase, a special process of strategy has
to be added to the standard Launhard-Hotelling model to explain
the phenomenon that low quality does not drive out good quality.
Secondly, the choice of quality is dynamic and not independent of
the reputation or goodwill of the firm, a stock variable. Consumers
learn about quality of individual producers over time and goodwill
first has to be accumulated before the benefits of it can increase
profits. To get a reward for high quality products in terms of a
higher profit the firm aquires goodwill which will create expec-
tations regarding product quality. This will shift the demand func-
tion of the firm as consumers observe goodwill as a signal for the
unobservable quality, estimate quality from the stock of goodwill
and make their purchase decision. If the consumer experiences ex
post a quality level which coincides with the quality level expected
on base of the signal, he will consider the price charged as infor-
mationally consistent or justified. If the quality turns out to be
lower than expected on base of the ex ante level of the stock of

C. Carrara and D. Sartore (eds.) Developments of Control Theory for EconomIc AnalysIS
© 1987 Martmus Nijhoff Publishers (Kluwer). Dordrecht·
262

goodwill, the producer will loose goodwill on the market and his
demand function will shift downwards. The objective of the paper
is to show that reputation as an endogenous signal of quality can
prevent the kind of market failure characterized by Akerlof (1970)
for the case of asymmetric information on quality.
In the next section we will begin with a monopolistic industry
where the monopolist chooses price and quality under asymmetric
information starting with a given goodwill. We will characterize the
optimal paths of price and quality towards the steady state solution
and will investigate the direction of the goodwill formation process.
In the third section we will analyse the duopolistic market by
looking for a consistent solution on that market compatible with
the actions of both firms. In the fourth section we compare prices
and quality levels in the duopolistic market under perfect and under
asymmetric information.

2. Monopolistic Price and Quality Decisions Under Asymmetric

Information

We consider a monopolist who maximizes the present value of

profits choosing price p(t) and quality n(t) as control variables. The
demand function x(p, G) conjectured by the monopolist depends
positively on goodwill G which can be changed by an appropriate
quality policy.
The model is:

max
p(t),n(t)
rx e-
Jo
rt {px(p, G) - C(x(p, G), n)} dt (1)

subject to the goodwill formation condition

G(t) (n - ne(G))x(p, G)
(2)
G(O) Go and n ~ n.
n is the lowest quality level which means that for n < n consumers
can detect the poor quality and will not buy the product (see
Shapiro (1982)). The meaning of the variables is: e- rt is the discount
factor; pet) is the price of the good; C(x, n) is the cost of production
with Cn > 0, Cx > 0; net) its quality; G(t) is goodwill, and ne(G(t))
is the quality expected by the consumers having observed goodwill
G(t). Expected quality increases in goodwill (n~ > 0) and with
 263

higher goodwill the increase in expected quality becomes larger

(n~G > 0). This convexity assumption implies a kind of self-enforcing
effect of expectation formation under goodwill accumulation. ne (-)
is known to the producer and is derived from the evaluation of the
consumer knowing about the reputation G(t) of the firm. If the
consumer has experienced the quality after purchase, both the
producer and the consumer consider the quality to be equal to net).
If goodwill promises a higher quality, ne(G(t)) > net), it will be
reduced according to condition (2). The effect of consumers dis-
appointment, or pleasant surprise if n > ne(G), on goodwill is
proportional to the quantity sold in that period. A more concrete
and operational description of the impact of quality on reputation
can be given by interpreting n as the probability of the product not
to fail, G as consumers experience regarding expected performance
of quality and ne (G) as the expected probability of non-failure
based on experience G. o
In the same setting of asymmetric information Shapiro (1982)
uses a similar model to explore the properties of the steady state. In
a former paper (Conrad (1985)) we looked at similar questions
employing a monopolistic model with quantity and quality as
control variables and a reputation adjustment equation similar to
(2). In that paper we also dealt with the case of perfect competition
and furthermore introduced advertising as an active marketing
strategy.
The Hamiltonian of our model (1) and (2) is:
H = e- rt {px(p, G) - C(x(p, G), n)} + .Ie(n - ne(G))x(p, G).

We define p = e,t.le, and require (p, n, G, p) to satisfy the first order

conditions associated with the Hamiltonian; i.e. (2) and
o => x + xp(p - Cx) + p(n - ne(G))xp = 0 (3)

o => - Cn + px = 0 (4)

A =>fl- rp = -xG(p - cJ
+ n~xp - p(n - ne)xG' (5)
From (4) we see that the current (non-discounted) shadow price p
of goodwill is equal to the marginal cost of quality per unit sold.
A marginal increase in quality raises production costs per unit
which is the price for an additional unit of goodwill. If the quality
264

produced equals the quality expected, (3) postulates that marginal

revenue with respect to price should be equal to marginal cost. If the
quality produced exceeds the quality expected, the addition to
goodwill, evaluated by fl can be added to the price P as a premium.
It is then profitable for the producer to expand x by reducing p as
quantity also transmits the deviation of the produced quality from
the expected one.
Our objective is to characterize the optimal path of the control
variables p(t) and net) and of the state variable G(t) qualitatively.
The steady state, if it exists, can be studied by using f1 = 0, G = 0
in (2)-(5). We obtain:
x + pXp = Cxxp (6)

xGp = C~xG + Cnn~ + rfl· (7)

According to (7), marginal revenue of goodwill is equal to marginal

cost of goodwill, consisting of marginal cost of production due to
the shift of the demand function, marginal cost of quality necessary
to achieve higher goodwill, and the opportunity costs of an
additional unit in goodwill. System (2), (4), (6) and (7) can be solved
for the steady state solution (p, n, G, {l). Next we carry out a
diagrammatic analysis based on the differential equation system in
the neighbourhood of the steady state in order to characterize the
optimal path qualitatively. From (3) and (4) we obtain the solutions
p* and n* as functions of fl and G. We want to find the partial
derivatives of P*(fl' G) and n*(fl, G) with respect to fl and G. For
this purpose we differentiate (3) and (4) totally:
Hpp dp + Hpn dn - HpG dG - (n - ne)xp dfl

Hpn dp - Cnn dn -HnGdG - xdfl·

To facilitate the assessment of the signs of the partial derivatives we

make the following separability assumption:
C(x, n) = K(x) + Q(n)x. (8)
This assumption implies:)

Using Cramer's rule we obtain the following signs of the partial

derivatives if we assume the revenue function to be concave in p and
 265

G and the cost function to be convex in x and n: 2

8p* {> 0 if n < ne (9)

8J1 < 0 if n > ne

(10)

8n*
-> 0 (11 )
8J1
8n*
8G = o. (12)

Next, we substitute p* = P*(J1, G) and n* = n*(J1, G) into (2)

and (5) and obtain a pair of autonomous first order differential
equations:
[n*(J1, G) - ne(G)] . x(P*(J1, G), G) (13)
{1 J1[r + n~x(.)] - xGC)[p*O - eJ·)]
- J1[n*(-) - ne(G)]xGO· (14)
In order to find the properties of the dynamic path, we use the phase
diagram method of solution. From (13) we obtain: 3
8C 8C
8G < 0 and 8J1 > o. (15)

Next we differentiate (14) with respect to J1 and G: 4

8{1 8{1
8J1 > 0 and 8G > o. (16)

Finally we consider the slope of the locus of {1 = 0 and C = O. The

locus C = 0 has the slope:

dJ11 8Cj8G
dG G~O = - 8Gj8ji > 0 (see (15)). (17)

The locus of {1 = 0 has the slope:

dJ11 8{1j8G < 0 (see (16)). (18)

dG f1~O = - 8{1!8J1
266

G=O

jl=O

L-______________ ~ ____________ ~G

G*

Figure I. The neighbourhood of the steady state.

Figure I illustrates a phase diagram for (13) and (14) in the neigh-
bourhood of the steady state. Under the assumption made there is
one equilibrium which is a saddlepoint.
If Go, the initial stock of goodwill is less than G, then the optimal
path is characterized by a monotonically increasing G and a
monotonically decreasing p. The firm accumulates goodwill over
time by giving the consumers a pleasant surprise. As the state
variable G is a monotone function of time along a path converging
·to a particular steady state, G does not oscillate and producers
either permanently accumulate or decumulate goodwill.
We finally study the monotonicity of the optimal paths of p, n
and x. We find:

jJ
op*.
-G+-p
op*. {>
0 if Go < G
(19)
oG op ~ 0 if Go > G
(+) :>

on*. on*
-G+-/1
0 if Go {< <G
(20)
oG op > 0 if Go >G
(0) (+)
Let us consider the more realistic case that at the beginning Go is
lower than the steady state goodwill G. Then G increases (G > 0)
and p decreases (/1 < 0) which implies jJ > 0, price increases, and
Ii < 0, quality decreases. Thus, the monopolist offers high quality
at low prices at the beginning to accumulate goodwill. As time goes
on he reduces quality and increases the price. The quality is still
 267

worth its money but one has also to pay for the reputation. With
respect to quantity, we get:
x = xpfJ + xGG.
(-) (+)
The price increase reduces the quantity but the accumulation of
goodwill can more than offset this reduction in quantity.

3. Price and Quality Policy for a Duopolist

It is well-known that the duopolistic market is in equilibrium if

for differentiated products the values of their prices are such that
each duopolist maximizes his profit, given the price of the other,
and neither desires to change his price. However, this equilibrium
solution need not to be stable if we introduce quality as a control
variable and goodwill as a state variable. Given goodwill and
quality, in the standard model, each duopolist maximizes profit
with respect to price, treating the price of the rival as a parameter:
maxpjxj(Pj, P2; Gj, ( 2) - Cj(x j(.); iij)
PI

max P2X2(P2, Pj; Gj, ( 2) - C2(X 2(·); ii2)'

P2

In the equilibrium, price policy will not increase profit, however

quality policy can. As better quality increases the cost of pro-
duction, profit will be lower at least in the first period, because the
unobservability does not permit to increase price or quantity in a
profitable way. Again, as in the proceeding section, a shift of the
demand function can be achieved by accumulating goodwill. For
demand x,(Pj' P2, G j, G2) the following standard assumptions are
made (e.g., Thepot (1983». The function x, is decreasing in p, and
GJ and increasing in PJ and G,:

Furthermore, efficiency of the goodwill decreases when goodwill

becomes higher and higher: x,G,G, < O.
As the duopolistic price equilibrium need not be one if other
instruments are feasible we will look for an equilibrium involving
additional instruments like quality and goodwill formation. We
state the price-quality problem of duopolists as a differential game.
268

We assume that the game is one of perfect information in that the

duopolists know the value of the current state variable, goodwill G.
There is, however, asymmetric information with respect to quality
on part of the consumers and of the competitors. Each firm is
assumed to maximize its own present value profit by choosing the
time paths of its own control variables price and quality. The profit
of each firm depends on the control trajectories chosen by both
firms. Then the problem can be stated as the non-zero sum differ-
ential game:

max
p"n,
f: e-r,1 {PIXI(PI, G I ; P2, G2) (21)

max
P2,n2
r' e- r21 {P2 X2(P2, G2; PI' G I ) - Cix 2(·), n2)} dt (22)

(nl - ne(GI))XI(PI, GI ; P2, G2) (23)

(n2 - ne (G 2))X 2(P2, G2; PI' G I )· (24)
Each duopolist i determines his control variables P, and n, at any
time as a function of the state variables G I and G 2 . In this c1osed-
loop control the competitors are able to observe at any time the
state of the system. For sake of simplicity we consider only the
open-loop solution of our non-cooperative equilibrium solution of
the differential game. s Accordingly, the control variables are deter-
mined ex ante and their values at time t are thus chosen before the
beginning of the game. In the open-loop formulation, the problem
(21)-(24) is merely considered as an optimal control problem for
each firm where the rival's control variables p,C) and n/.) are
assumed to be fixed over all the horizon. 6 The current value Hamil-
tonian H, of such a problem may be written as
H, = p,X, - C,(x" n,) + 11,(n, - n"(G,))x, + Y/,(n j - nC(G))xj
where 11, and 1], are the costate variables associated to goodwill G,
and goodwill G" respectively, measuring their marginal values
assessed by firm i. First-order conditions which determine the open
loop-equilibrium are (23), (24) and
H p, = 0 ~ x, + (p, - C,x)x,p, + 11,(n, - nC(G,»)x,p,
+ 1],(nj - ne(G))xjp , = 0 (25)
(26)
 269

-HG,

(27)

x,G,[p, + pJn, - n"(GJ)] - C,x,x,G,

- 1J,n~,xJ + 1JJnJ - ne(GJ))xJG ,. (28)

According to the open-loop control each firm i determines an ex

ante value of the competitors quality nJ •
The qualitative characterization of the optimal paths of p" n, and
G, could be done in a similar way as in Section 2. Duopolist 1 has
to solve (25) and (26) for pf(p" G,; P2' G2 ) and nf(p" G,; P2' G 2 )·
Then pfO and nfO have to be substituted into (23), (27) and (28)
to become a system of autonomous first order differential equations.
We do not concentrate on an analysis of the dynamic game which
in general is very difficult, but resort to a consideration of consistent
price and quality decisions of each duopolist within a static duopoly
context. This will be done by making compatible the variables Pl,
G2 exogenous for duopolist 1 with the optimal values of P2' G2 of
duopolist 2, and the variables p" G, exogenous for duopolist 2 with
the optimal values of p" G, of duopolist 1. For doing this we
analyze a steady state equilibrium for the duopoly market. We have
to find a set of solutions {p" P2' 11 1 , 11 2 , GI' G2 } which satisfy the
following six equations (i = 1, 2) of the steady state version of the
open-loop control system (23)-(27):
(29)

n, = n"(G,) (30)

(31 )

with p, = Cu, ix, where PJ, are the opportunity costs of goodwill in
terms of producing the quality expected to sustain goodwill. As the
steady state system (25)-(28) is recursive with respect to 11" the
steady state version of (28) has no feedback on the solution set.
From (28) we obtain
(32)
270

which implies 1J, < 0 as X ,GJ < 0 by assumption; a marginal value

of the rival's goodwill has a negative impact on the firms profit.
For a description of the market process we introduce "steady
state reaction functions" which express price, quality and goodwill
as function of his rival's price, quality and goodwill. Of course, to
proceed with a static reaction function analysis around these steady
state equilibrium points is not quite consistent with the preceding
analysis. We solve (29)-(31) for i = 1 for duopolist 1 and for i = 2
for duopolist 2 and obtain:
p, RHp2' G2), n, R7 (P2' G2) (33)
RHp" Gd, n2 R¥(p" G,) (34)
It turns out that the rival's j quality nJ does not enter into the
reaction function R,O of duopolist i. This is a desirable property of
the solution of (29)-(31) as in the context of our model the quality
of the competitor is not observable anyway.
The following figure illustrates the equilibrium by a set of inter-
section points for the reaction curves. The figures are based on
double-logarithmic demand functions, on cost functions of the
type:
c =
I
(e, a constant) (35)
and on expected quality functions of the type:
ne ( G,) = JG,.
Our dynamic analysis based on movements along reaction func-
tions in the neighbourhood of the steady state is consistent with the
dynamic behavior of a firm acting as a monopolist. In section 2 we
found that in a situation of low goodwill firm 2 will increase price
and goodwill given the variables offirm 1 as parameters. This might
result in point A in Figure 2. If for firm I parameters change, it will
follow its strategy of increasing goodwill by first producing better
quality at low prices and later on quality that will deteriorate at
higher prices. The straight line from point A to point B in Figure 2
does not say anything about the path from A to B. One can think
of a sort of overshooting of quality before point B is reached. This
kind of analysis is, however not quite appropriate to represent the
two dynamic adjustments. One dynamic strategy emerges from the
optimal control problem of the firm, and the other one from the
 271

~----~--------~I----~--------------. P,
P,2

n,
n, n, n,
L---',:--------....J2'::-----:f----------.

Figure 2. Equilibrium as intersection of reaction functions.

readjustment to the variables chosen in each period by the com-

petitive firm and treated as exogenous by the other firm.

4. Quality Under Asymmetric and Perfect Information

In case of perfect information our problem of Section 2 can be

stated in the following way:
272

max px(p, G) - C(x(p, G), n) (36)

p,n

where G = ne-' (n); i.e., goodwill is derived from the inverse of the
former steady state conditions n = ne(G). Necessary conditions
are:
x(p, G) + pXp - Cxxp = 0 (37)
pXc - C,x c - Cnn~ = 0 (38)
as
dG
dn nec
If r = 0 in the steady state condition (7), the perfect information
solution (p, ii) and the steady state solution (ft, n) are the same. If
we assume r > 0 and compare the perfect information solution ii
of (38) with the asymmetric information solution n of (7),
(7)
we know from the second order conditions that the left hand side
of (7) is monotone decreasing in G so that G in (38) must be higher
than {; in (7) and therefore ii > n. This line of reasoning is correct
if the left of (7) is increasing in p; if it is decreasing in p, a price
increase with a lowering in goodwill is also a strategy to converge
for r ~ 0 from the solution (n, p) of (6) and (7) to the solution
(ii, p) of (37) and (38) (see also Shapiro (1982».

STATEMENT I: If the steady state characterized in Section 2 exists

then ii > n, that is the quality under perfect information is greater
than the quality under asymmetric information.

PROOF: In the Appendix we used Hpc > 0 to characterize the

steady state. H Gp is the partial derivative of (7) with respect to p and
its positive sign excludes the strategy mentioned before. QED

By a similar reasoning we conclude:

STATEMENT 2: If the steady state characterized in Section 2 exists

then p > p, that is the price under perfect information is higher
than the price under asymmetric information.
 273

PROOF: The left-hand side of (37) increases if n -+ n(n > n). In

order to satisfy (37), the price has to be raised from p to p as
Hpp < O. QED

We finally look at the duopoly market under perfect information.

If the duopolists maximize their profits with respect to p and n,
necessary conditions for an equilibrium on the market are equivalent
to (29)-(31) with " = 0 and the inverse G, = ne - (n,). 1

If we combine statement 1 and 2 we can conclude:

STATEMENT 3: A duopoly market for a different experience good is

characterized under perfect information by higher qualities and
prices than under asymmetric information.
As in the equilibrium state prices are worth their money, higher
prices under perfect information are informationally consistent
with the higher quality obtained. A consumer protection policy
with standards, warranties, and money back refunds to approximate
a world of perfect information can help to improve the quality on
the market. However, without knowing the price-quality ratio it is
difficult to evaluate the higher price and quality levels of (p, n)
under perfect information with the lower levels of the pair (p, n)
under asymmetric information as either prices are informationally
consistent. Our guess that the price-quality ratio is lower in a world
of perfect information can be enforced by the following statement:

STATEMENT 4: If a firm under asymmetric information has to

charge the perfect information price p, its quality level Fz will be
lower than the perfect information quality level n.

For a proof we have to show that the quality Fz which solves the
steady state first-order condition of the problem

m:x f e- rl
{px(p, G) - C(xC), n)} dt

subject to G= (n - ne(G))x, is less than n.7

6. Conclusion

The objective of this paper was to analyse price and quality

policies under imperfect information. We used the monopoly case to
274

introduce a dynamic price-quality policy with goodwill (or repu-

tation) accumulation as a signal for the unobservable quality.
The dynamic approach has been introduced to the Launhard-
Hotelling duopoly to analyse price, quality and goodwill policies
over time in a heterogeneous duopoly. Under asymmetric infor-
mation with respect to quality the model takes into account the
aspect that pricing decisions have an instantaneous effect on demand
and profits while the impact of quality decisions are displayed
over time. As the quality under perfect information turned out
to be higher than under asymmetric information consumer pro-
tection policies towards better information can help to improve
the quality on the market. Such a policy will also lower the price-
quality ratio.

Appendix

Proof of Hpn = 0 and HnG = 0

Hpn = xp (~n - cnx) = 0

as Cn = Q'x and Cnx = Q' according to (8).

Similarly,

HnG = XG ( - Cnx + ~n) = O. QED

Proof of the signs of (9)-(12)

We find:
Hpp = 2xp + (p - CJxpP - Cxxx; + J1(n - ne)xpp < 0
under the assumption made with respect to the concavity and
convexity assumptions for x(·) and CO, and n '" ne in the neigh-
bourhood of the steady state. Similarly,
HGG = (p - CJxGG - Cxxx~ - J1xn~G - 2J1n~xG

+ J1nxGG < 0,
 275

and HpG > 0 if xpG < 0 does not offset all positive terms in the
following expression for HpG:
HpG = XG + xpG(p - CJ - CxxxGxp
+ Il(n - ne)xpG - Iln~xp.

For the signs of derivatives we find according to Cramer's rule:

ap* 1 e
all = D (Cnn(n - n )xp)

where D = Hpp( - Cnn ) > O.

ap* I
aG D (CnnHpG ) > 0

an* I
-D (-x)Hpp > 0
all
an*
O. QED
aG

Proof of (15)
ac aG - nGe ) X
( an*
p*
+ (n* - ne) (xpaaG + XG ) < 0
aG
in the neighbourhood of the steady state and due to (12).
aG an* ap*
all = all x + (n - ne)xp all > 0 (see (11)).

afJ, e ( * e) H ap* H an* 0

all r + nGx - n - n XG - Gp all - Gn all >

in the neighbourhood of the steady state where ap*/all ~ 0, and

H Gn = 0 holds anyway.
Finally,
afJ, ap* an*
aG -HGG - HGp aG - HGn aG
(- ) (+ )( + ) (0)

_ H GG _ H Gp (C nn H pG ) =
-Hpp Cnn

as H is concave in p and G by assumption.

276

Proof of statement 4
The first-order conditions is equation (7), i.e., with p = p and
n = n"(G):
pXG(p, G) - C.(x(·), n) - CnC)n~ = rCn/x. (7')
With G = ne - 1(n) the left-hand side is strictly monotone decreasing
in n. Under perfect information the right-hand side of (7') is zero
(see (38». Thus a solution of (7') for n implies Ii < n.

Notes

O. lowe this suggestion to Ch. Tapiero.

I. See the Appendix.
2. See the Appendix.
3. See the Appendix.
4. See the Appendix.
5. See also Thepot (1983) for a model of duopoly as a dynamic game.
6. This assumption implies that the firms will not buy the product of its com-
petitors at any time. As they can not observe the quality, they will never learn
to know it.
7. The proof is given in the Appendix.

References

Akerlof, G. (1970): "The Market for Lemons: Qualitative Uncertainty and the
Market Mechanism", Quarterly Journal of Economics, 84, 488-500.
Conrad, K. (1982): "Advertising, Quality and Informationally Consistent Prices",
Zeitschriji fur die gesamte Staatswissenschaft, 138, 680-694.
Conrad, K. (1985): "Quality, Advertising and the Formation of Goodwill under
Dynamic Conditions", in G. Feichtinger (ed.), Optimal Control Theory and
Economic Analysis 2. North-Holland, Amsterdam, 215-234.
Nerlove, M. and K. H. Arrow (1962): "Optimal Advertising Policy under Dynamic
Conditions", Economica, 124-142.
Shapiro, D. (1982): "Consumer Information, Product Quality and Sellers Repu-
tation", The Bell Journal of Economics, 13, 20-35.
Thepot, J. (1983): "Marketing and Investment Policies of Duopolists in a Growing
Industry, Journal of Economic Dynamics and Control, 5, 325-358.
 PART IV

ECONOMIC AND ECONOMETRIC ANALYSIS BY

CONTROL METHODS
 279

CHAPTER 16

A SYSTEMS APPROACH TO INSURANCE COMPANY

MANAGEMENT

Charles S. Tapiero
Hebrew University, Israel and Case Western Reserve University

1. Introduction

The systems approach to risk management consists broadly in

altering, in a desirable manner, the probability distributions and the
states a system may reach at various times. Alteration of such states
and their probabilities can be reached by various instruments risk
managers may have at their disposal. These include, for example,
insurance, loss prevention and technological change as is sum-
marized in Figure 1. Insurance, in particular, can be viewed as a
medium, or a market for risk, substituting an exchange of payments
now for potential damages later. The size and the timing of such
payments and their effects on persons, firms, governments, etc. lead
to widely distributed preferences, to the possibility of exchanges
and of course to the realization of a market for insurance. Insurance
firms have recognized these distributed preferences and the basic
desires of persons and firms to manage risk, and as a result have
devised mechanisms for pooling and redistributing risk by using the
"willingness to pay to avoid losses." It is such a rationality, com-
bined with goals of personal gains, social welfare and economic
efficiency that markets for fire and theft insurance, sickness, unem-
ployment, accidents, and life insurance, etc. have come to be as
important as they are today. It is persons or institutions' desires to
avoid too great a loss (even with small probabilities) and face these
losses alone that markets for re-insurance (i.e., sub-selling portions
of insurance contracts) and mutual insurance (based on the pooling
of individual risks) have come into being. The purpose of this paper
will be to use this basic notion of insurance in developing a systems
approach framework which can be used, hand in hand, with control

C. Carraro and D. Sartore (eds.) Developments of Control Theory for Economic Analysis
© 1987 Martinus Nijhoff Publishers (Kluwer), Dordrecht
 280

Active Passive
Risk Managarent Risk Managarent

Technology States Future markets

loss
prevention Probabilities Insurance
reinsurance
risk sharing

Information

Figure l. Risk management-a simplified view.

and stochastic control concepts in formulating essential insurance

firm problems and in obtaining insights regarding some of the
policies they might follow in managing their risks. Issues in loss
prevention and technological change for buying or developing
better and safer equipment, more efficient managerial procedures,
etc. to alter the probabilities and the effects of undesirable damag-
ing states, will not be considered within the confines of this paper.
Further, issues such as partial and asymmetric information distri-
buted between insured and insurer and their effects on insurance
related behaviors, the effects of markets structure and government
regulation and intervention (extremely important in the economic
and insurance literature) will also be ignored. As a result, the scope
of the paper is restricted to the use of and the introduction of
systems notions that have been widely used in engineering, in econo-
mics, and in management. In insurance there is an emerging litera-
ture such as Balzer [1] and Balzer and Benjamin [2] which use
classical notions of control (the so-called frequency approach) to
manage the insurance firm's surplus funds, simulation of dynamical
models of insurance (as in Pentikanen [28]), general insurance firm
models (as in Bohman [1-6], and various dynamical models of
insurance firms (as in Borch [12, 13, 14]). Although these models are
not reviewed in detail, this paper is clearly written in their spirit and
provide another linkage between the classical approaches to classi-
cal insurance risk management and dynamical models of insurance
companies management.
 281

2. The Systems Approach to Actuarial Science

Actuarial science is in effect one of the first applications of probabil-

ity theory and statistics (Borch [11]). Tetens [38] and Barrois [3]
already in 1786 and 1834 respectively, were attempting to charac-
terize the "risk" oflife annuities and fire insurance and on that basis
establish a foundation for present day insurance. It is, however, due
to Lundberg [26] in 1909 and to a group of Scandinavian actuaries
[13] that we owe much of the current mathematical theories of
insurance. In particular, Lundberg provided a foundation for
collective risk theory which is very similar to the systems concepts
of inventory and dams management. In this sense, Lundberg's
approach, which we summarize in Figure 2, can be thought of as a
forerunner of the systems approach in insurance.
Here, we note that "controls," "states" and "disturbances," can
be represented by the "premium payments" required from insured,
"wealth" or the firm liquidity and "claims" which are of course
random. In its simplest form, the "control problem" of actuarial
science is to establish exchange terms between the insured which
pays a premium, allowing him to claim a certain amount from the
firm (in case of an accident). These terms are reflected in the
insurance contract which provides legally the "conditional right to
claim." Much of insurance literature has concentrated on the defi-
nition of the rules to be used in order to establish in a just and
efficient manner, the terms of such a contract. In this vein, "Premium
Principles," expected utility theory and a wide range of operational
rules have been devised (for an extensive survey see Lemaire [25],

Control Insurance States

finn,
Premiums Wealth
system
(production) (inventory)
dynamics
ruin
(s tock out)
Disturbance
Claims
(demands)

Figure 2.
282

as well as Buhlmann [15], Gerber [20)). This problem is of course

extremely complex, with philosophical and social undertones, seek-
ing to reconcile individual with collective risk, through the use of
market mechanisms. In its proper time setting (recognizing that
insurance contracts are a reflection of insured attitudes towards
time and uncertainty), this problem is of course conceptually and
quantitatively much more complicated. For this reason the approach
we outline below is necessarily a simplification of some of the
fundamental issues that insurance deals with.
In a time setting, the system (control) approach to insurance
necessitates that
(1) we define the processes of change affecting the insurance firms
(such as inflation, rates of returns, market conditions, govern-
ments, etc.)
(2) we specify the time and risk preference of the insurance firm,
translated into decision criteria reflecting the profit motives, the
avoidance of costs as well as the inherent costs of ruin (or
default risk)
(3) we establish the types of strategies and decision making
approaches (and their constraints) that could be applied to
manage the insurance firm. These may include the premiums
payments required from insured, assets indexation, investment,
reinsurance, co-insurance treaties, etc.
(4) we combine the means for managing the firm, specified in (3)
together with the "process of change" defined in (1) to obtain
a representation of the insurance firm process of change over
time and its realizable outcomes.
(5) we seek these policies (defined in (3)) which will optimize the
criteria which were specified in (2), so that what is intended can
likely occur.
In other words, a conception of the future states the insurance firm
may reach, the conditional (and probabilistic) realizations of these
futures in terms of policies and how they might be attained, is
required. This framework is implicit in most managerial decisions
evaluating the various desired outcomes and the implicit and
explicit costs incurred in reaching these outcomes at a particular
time. In insurance this is even more acute since, basically, insurance
seeks to substitute certain for uncertain payments at different times.
A dynamic approach to this problem, however, requires that these
be made explicit rather than implicit. To do so, we shall consider a
 283

simple insurance decision framework and elaborate on the key

elements of the dynamic problem. Subsequently, the problems to be
dealt with will be more specific, requiring the solution of control
problems.
Say that an insurance firm has assets At and liabilities L t at time
t. These assets consist of cash, cash equivalents, bonds, mortgages,
stock, real estate and other investments. Both assets and liabilities
can be managed. For example, by altering the premium policy,
reducing investments (or desinvesting by the selling of assets), by
delaying benefits payments, reducing expenses, etc. the firm can
manage its assets and liability balances. Define by V t the (controls)
period's intervention and let the ~t be the periods disturbances. The
relationship between these elements over time, define a state-space
model of the insurance firm. In essence, the states can be thought
of as summarizing the past history of the insurance firm with respect
to well-defined variables. This history is defined as a function of the
actions taken by the firm, the unforeseen "disturbances" and uncer-
tainties the firm has been subjected to and, of course, the mechanism
for combining these into economic indicators of the firm's health,
valuation and assets accounting.
The state-space representation is thus a model of the insurance
firm which represents in a unique way the firm's computation of
specific economic variables. Of course, there is not only one way to
compute the economic performance and the value of the firm, and
as a result there is more than one state-space representation if the
insurance firm (as a function of the degree of specificity of the
models, the choice of economic variables, etc.). As a result, the
construction of a dynamic system for the insurance firm is not only
a technique, but it is essentially an art in blending those relevant
dimensions of the insurance firm with manageable techniques that
can lead to useful and applicable analyses.
Generally, we can state that such a state-space representation is
given by the following type of system:

Dynamic system
{At} -+ {Ar<1' LI' VI' ~t}
current past periods periods periods
state states lia bili ties intervention disturbance
284

In our framework we can state that the following variables might

be constructed as representing some aspects of the insurance firm.

A L U ~
bonds -premium assets conversions benefit payments
cash -income dividends unforseen inflation
investments benefits premium rates transfer costs
real estate expenses reim bursemen ts

When such a model, representing realistically the mechanisms of

the insurance firm's functions has been constructed it becomes an
instrument of management, providing information regarding the
firm's economic state (such as economic viability, financial position,
worth, etc.), an instrument for testing alternative policies and
finally an instrument for designing actuarial and financial policies.
To do so, however, it is required that a valuation approach, for the
firm be defined.
In insurance, valuation is essentially a function of future interest
rates specifying the firm's earning capacity and the required capacity
to meet claims (solvency). Thus, the maintenance over time of
liquidity to meet claims and earning capacity of the assets portfolio
are prime goals.
In the pursuit of these goals, the policies the insurance firm may
use include (among others);
- the choice of investment assets
- the use of learning mechanism, a function of and in terms of the
various implicit costs of uncertainty sustained by the firm
the adoption of risk absorption approaches used by the firm, in
so far as portfolio, diversification strategies, etc. are followed
the selection of risk sharing approaches such as indexed and
linked premium formulae, co-insurance, re-insurance, etc.
In addition, the effects of time and risk may lead to the need of
contingencies to ensure the solvency and economic viability of the
firm. This would include as an example
- contingencies for under-valuing inflation effects on assets
pricing
contingencies for under-reserving (i.e., mis-assessing payments
and benefits over time)
 285

- contingencies to meet possible interest rate changes, particularly

reflected in lower incomes
- possible assets destruction, etc.
The choice of such contingencies and their quantities are explicit
(and if not, then implicit) outcomes of the insurance firm's states
and temporal preference, and not the opposite.
Theoretically, the quantitative definition of the temporal
preference is a cumbersome matter, a function of the insurance firm
investigated (i.e., whether it is a stock, or mutual insurance firm for
example). Generally, we can define it in the following terms:

E {f~ (benefit-costs) e- rl dt - (cost of insolvency) e- rr }.

which is the integral (the sum) of the measured (or expected util-
ities) benefits obtained by the insurance firm as long as it is solvent
less its costs (or disutility), less the cost of insolvency of the firm,
discounted at an appropriate rate of discount and "," is the firm's
random time reached when the firm becomes insolvent. Basically,
this time is a function of the firm position regarding its assets and
liabilities holdings. Alternatively, a function G(A, L) could be
defined such that insolvency occurs at the first instant of time
reached when.,
, E G(A, L)
In other words, when the firm reaches certain levels in the asset-
liabilities plan it may be dissolved and costs, a function of the firm's
penalities of involvency, are incurred.
In summary, then, the insurance firm would seek to
Optimize the
Temporal Objective
Including Bankruptcy Cost
Subject to
A State-Space Representation
of the Insurance Firm
+
Solvency & Regulation
Constraint
+
Other Constraints
286

To Obtain
Management Strategies for Investment
and Liquidity Maintenance
+
Premium Rates Policies Sensitive or
Insensitive to Change and
the Firm's States
+
Reserves for
Unforeseen Contingencies
and Assets Depreciations
+
Learning Mechanism for
Adaptation to Change

Within such strategies (many) questions such as the following

might be answered:
- What is the optimal surplus required to protect the insurance
firm against bankruptcy (insolvency), and assets depreciation
and generally determine the surplus management policy?
- What is the surplus policy required for protection against pric-
ing inadequacy resulting from uncertainty (or partial infor-
mation)
- What should the premium policy (feedback versus open loop)
be?
- What is the economic value of learning more precisely con-
sumers risk classification?
and so on. To reach such solutions, however, the problems defined
will necessitate a specification of the insurance firm's structure and
processes and the definition of the non-stationarities and uncertain-
ties it is subjected to. It will also lead to quantitative problems that
are of a high order of difficulty, such as stochastic dynamic pro-
gramming and control (e.g., see [5], [6], [9], [19], [23], and [39] for
references).
The insurance company management approach formulated
above follows traditional dynamical theory of the firms approach
as stated for example by Jorgenson [22] and extended to a stochastic
framework with bankruptcy by Bensoussan and Lesourne [7, 8] (see
also Borch [13, 14], Leland [24] Tapiero [33, 34], Van Loon [40].
This relates to the "modern" theory of financial economice (e.g.,
[16], [27], [29]) if the insurance firm's valuation function is expressed
 287

in terms of the market value of the firm's assets and insurance

portfolio's liabilities. Such an approach has been used in particular
by Merton [27] in pricing insurance contacts and loan guarantees.
An options pricing approach is then used by reinterpreting insur-
ance contacts as European put options on the insured asset with an
exercise price set at the insured value of the asset. Using additional
(and restrictive assumptions), a Black-Scholes put pricing solution
yields a general equilibrium price for the insurance contact (see also
Smith [29], p. 106]. While this approach clearly opens important
avenues for future research in insurance company management, the
market imperfections within which insurance firms operate and the
probabilistic structure of claims might preclude the validity of the
Black-Scholes pricing formula. Nevertheless, this approach, from a
company's viewpoint, may establish an important relationship
between riskless market rates, the insurance company portfolio of
contracts and the valuations of premiums. Practically, the solution
of these problems, (whether an options pricing or a classical
dynamical theory of the firm's approach is followed) are difficult to
resolve. Simulation may be used as an initial methodological frame-
work (as in Pentikanen [28]). Recent and important advances in the
numerical techniques available for solving such problems renders
their prospective use an increasing possibility. To demonstrate the
potentialities of the systems approach, however, we shall consider
two applications on a mutual and a stock insurance firm. For
simplicity, intuitive notions and developments are emphasized and
results are stated without proofs (these can be found in [32]-[37].

3. Selected Applications

Below, we shall consider several examples and applications which

highlight the systems approach to the management of insurance
firms. Extensive discussion of the models presented below as well as
additional models can be found in [1], [2], [4], [11], [18], [32], [33],
[34], [36] and [39]. To simplify the presentation, essential results are
summarized with numerical and theoretical analyses used to
explain the underpinnings and the implications of the systems
approach to insurance.
 288

3.1. A stock insurance firm problem

In this application (se [34]) we shall view the insurance firm in the
tradition of the general theory of the dynamic firm. We assume a
dividend maximizing insurance firm which controls four essential
policy instruments: investments, premium requirements, desinvest-
ments and dividends. The insurance firm collects also an income
from capital assets, and pays claims to insured as these occur
following some random process. The insurance firm's states, sum-
marizing the past history of policies and external effects (such as
claims and income from assets), are given by the capital assets and
liquid assets (cash) holdings. These are given in Figure 3 where each
of the variables is also defined. The dynamic relationships between
those variables are given in detail in Figure 4 where in addition, the
number of contracts N and the size of the insurance sales force (L)
are explicitly accounted for in the model. The model implied in
Figure 4 is stochastic, however, since the insured claims are stochas-
tic. Although other variables can be stochastic (such as the rate of
return on assets r, the number of insurance contracts, etc.) we shall
limit ourselves to claims' uncertainty.
A quantitative formulation of this problem written as a stochas-
tic control problem is briefly given by the following:
Maximize J(Ko, M o) = E f~ e- II D(t) dt - Be-II
subject to a control constraint set
o ~ I(t)/M(t) ~ lco 0 ~ Ie ~ I
o ~ g(t)/K(t) ~ g" 0 ~ gc ~ I
o ~(t) ~ Dmax
Controls States

Investment 1 - _...
1---...... K Capital Assets
Desinves tmen t g The Insurance

Firm
~
Premium
Dividend P ---tL..--.,:-_ _ _ _ _r---Jt------ M Liquid Assets
D --001. Cash

t
Income from
f
Claims
Assets rK F;(dt).

Figure 3. The insurance firm model: a state space representation.

 289

Hiring Quit
H sL rate

! t
L wages
Labor/
Agents
~L + aNP - Dividends DC t)
Outlays

Claims
~(dt)

Income
PremiUlll I Assets
Cash M
Contracts ~

lnves f'.
NP K
N

t Returns rK
I Income

Assets Conversion g

K = Assets ($)
rK return on assets ($/time)
g ~ assets conversion ($/time)
M .. Cash ($)
N - Contracts (Number) f; ~ claims (random)($/time)

L - Labor (Agents)
w = wage per agent ($/agent/time)
a a Commissions rates
P - Premium per contract
I a Investments ($/time)
5 = natural quit rate (a percentage)

D a Dividends ($/time)
H = Hiring or firing (Agents/time)

Figure 4. The insurance firm (model without borrowing).

an absorption (bankruptcy) time

T = Inf {t0, M(t) ~ O}
~
and the state equations
dK (l - g - bK) dt; Ko > 0
dM {rK + PN - PI + (1 y)g - D - wL} dt - ~(dt)};
P> 1, 0 ~ y ~ 1; Mo > O.
290

In this formulation, the costs of bankruptcy at time 'l: (the first time,
no more funds are available to meet claims) are given by B, dis-
counted at the rate i. The policy set states that the investment rate
can be at most equal to a fixed proportion Ie of cash holdings, that
desinvestment can at most be equal to a fixed proportion ge of
capital assets K and that the dividends rate is constrained by
[0, Dmax]. The cost of investments and desinvestments ar deducted
directly from cash with (13 - 1)1 equalling the investment trans-
action cost. Similarly, if g are desinvested then only (1 - y)g
dollars are actually added to cash holdings with g equalling the
transaction cost. Additional administrative costs, drawn out of
cash include the agents cost wL where w is the wage rate and L is
the number of agents hours per unit time. It is possible to replace
(or complement) these costs by commission costs which equal a
proportion of premium income. In this case, if rx is the commission
rate, then instead of the premium income rate PN we will have
(1 - rx)PN. The relationship between the number of insurance
contracts and the agents work force L is assumed to be given by a
Cobb-Douglas production function, such that N(t) = Nop-a L b ,
13 > 0, b > O. For simplicity, L is assumed constant (with Hiring
Rate = Quit Rate in Figure 5).
We also let bK in the "dK" equation be capital assets depreci-
ation, the rate of return r on capital assets be constant and finally
the claim rate be defined either as Weiner Process [20], [31], or by
a Jump process ([9], [32]). If the claim rate can be represented by a
Weiner process, then this means that

~(dt) == d~ = J1N dt + (JJN dw; ~(O) = 0

where J1 is the mean claim per contract and (J2 its variance (with all
contracts assumed to be statistically independent), and "dw" is a
normal distribution with mean zero and variance dt and uncor-
related in time. When the claim rate is a compound Poisson process,
(as often assumed in insurance) then the claim rate ((dt) is written
(in its integrated form) as follows; (see Bensoussand and Tapiero
[9], Tapiero [31];

~(t) = ~o + f~ R f J1(dv, dz)

with the function J1(t, A) denoting the number of jumps of the
 291

process in (0, t) and

fl(dv, A) = fl(t + d W, A) - fl(t, A).
Alternatively, fl(dv, A) is the rate at which claims occur of size A.
If the claim rate is Poisson, then the probability of such claims
occurring is A dv. When such a claim occurs however, its size is a
random variable A, subset of the integrand R. The solution of the
insurance firm's problem will of course depend on the kind of
assumptions made regarding the claims stochastic process. To
obtain such a solution we use Bellman's principle of optimality for
stochastic control problems. Broadly, this principle implies that the
value function (at a time t) J(K, M, t) defined as follows:

J(K, M, s) = Max E f e- ll D(t) dt - e-l! B

can, by definition be written in a recursive form as follows:

(a) If no bankruptcy occurs during two subsequent instants of
time, (s, s + ds);
J(K, M, s)
= Max {D(t) + (l - r ds)J(K + dK, M + dM, s + ds)}
subject to
policy instruments
constraint sets
(b) if bankruptcy occurs, at M = 0 and s = r and the cost B is
incurred.
J(K, 0, r) = - B.
The development of J(K + dK, M + dM, s + ds) in particular,
leads to what is commonly known as Bellman's equation for the
stochastic control insurance firm problem.
A direct intuitive observation of the above equations (which can
be proved rigorously [6], [34] indicates that the value function is
only dependent on the states (K, M), since regardless of the instant
of time we might be in, and presuming all parameters to be station-
ary (time independent), bankruptcy occurs if there is no more
cash (M = 0). As a result, we write J(K, M, S) = J(K, M)
and J(K + dK, M + dM, s + ds) = J(K + dK, M + dM), and
J(K, 0, r) = J(K, 0). If dK and dM are then inserted the value
function and the claim's rate is assumed to be of the Wiener type,
292

then two terms Taylor series approximation, leads at the limit

dt - t 0;

aJ aJ
J(K + dK, M + dM) J(K, M) + aK dK + aM dM

+ 1/2 ~; (dM)2
Note that (dK) is of order dt and therefore all terms of order (dK)2
are of the negligible order 0(dt 2). However, (dM) is of order dt and
(dM)2 has a term of order dt since (dw)2- t he Weiner component
has an expectation of order dt. With this information in hand, it is
easily proved that an equivalent formulation of our problem can be
written (using Bellman's Principle and its functional development)
as follows:

iJ(K, M) = Max {D + (l - g - (jK) ~~

+ [rK + PN - fJI + (1 - y)g - D - liL - flN]

aJ (J2Na 2 J}
x aM + 2aM2

subject to

and at bankruptcy, when the insurance firm has depleted its assets,

J(K,O) = - B.
The solution of such (partial differential equations) is very difficult.
Nevertheless, for some cases, analytical results might be found,
while in others the application of numerical optimization tech-
niques (as in [34]) can be used to obtain practical results. In our
case, due to the linearity of the policy instruments I, g and D it has
been shown in [34] that a solution can be characterized by "barrier
policies" in the (K, M) plan, as shown in Figure 5. Thus, given
Premium P and any (K, M) combination, the insurance firm can
reach an optimal action regarding its policy instruments (I, g, D) by
inspection of the Figure 5.
 293

.. ~3
"-
,
~

, (v)
" . '.

" :O~
" (ii)
'"
>
>
, , .~
-
I ,
-

,\
- °2
j
, "' , ). , ,
} , ) ~ (iii) \

:----v. ----1f
1°1
4 :. . 6
I
(i) (iv) I
.. I

M* M*2 M*
3 1

Policies:

(i) + (iv) Do not Distribute Dividends

(v) Invest
(1) + (11) Dc~lnve8t

Figure 5. The optimal plane separation for the firm's policies.

When the claims rate is a compound Poisson process with mean

claim rate A and claims size z, with density function dF(.), the
Bellman equation becomes:

iJ(K, M) = Max {D + (l - g - (jK), + ~~

+ (rK + PN - fJI + (l - y)g - D - ~iL)'

x !l!...
aM + A r" [J(K, M + z) - J(K, M)] dFtz)}

subject to
o~ I ~ ~ M, 0 ~ g ~ g, K, 0 ~ D ~ Dmax.
and at bankruptcy we have the condition J(K, 0) - B stated
earlier.
294

In this formulation, note that certain cash transactions (namely

claims) occur at the Poisson rate A dt. Given that a claim occurs,
cash is depleted by Z, (or M - Z becomes the new cash level with
Z stochastic). As a result E[J(K, M - Z) - J(K, M)] is the
expected change in the insurance firm value function due to a claim,
(given in the above equation by the integral over dF(z)-the claim's
density function).
The usefulness of insurance risk management instruments such as
reinsurance become now evident. Specifically, reinsurance will seek
to depreciate the detrimental effects of economic claims (occurring
with small probability) on the insurance firm value function. To do
so, reinsurance in a systems approach context will alter (albeit at a
certain cost) the probability distribution of such claims. For
example, if an excess loss insurance formula is used (e.g., [15], [20],
[35]) within the claims loss given by:

: if z < Q
z = {
.: if z ~ Q
then a claim's expected cost will be given by:

A f: [J(K, M - z) - J(K, M)] dF(z)

+ A[J(K, M - Q) - J(K, M)] Prob [z > Q].

The cost of such a reinsurance will be given by the premium rate the
ceding firm (in the case, our insurance firm) will pay to the reinsurer
which takes on the risk whose expected cost is

A foX! (z - Q) dF(z).

Numerical analyses may then be applied to evaluate the optimal

level of excess loss reinsurance Q, as a function of the required
premium payment to the reinsurer. Such analyses are clearly com-
plex. They have however, the distinguished advantage of relating
the insurance firm actions and risk management instruments
(as reinsurance) in terms of the insurance firm value function.
Although, we have concentrated our attention on a dividend maxi-
mizing firm, the financial theory of the firm makes it possible to
consider instead the firm's share price or some other objective. In
addition, factors such as borrowing, government regulation (in
particular regarding assets to liabilities ratio), inflation and rates
 295

of return uncertainty could be evaluated by extensions of the basic

systems framework presented above. For practical purposes, numeri-
cal and simulation might be the only means available for obtaining
meaningful results. When the problem formulated is very simple,
analytical results can be used in providing insights regarding special
aspects of the insurance firm's problems.

3.2. A mutual insurance problem

Mutual insurance firms are firms whose stockholders are the
bearers of insurance contracts. Insurance is then viewed as a collec-
tive risk process, of say N persons, each paying a fixed (or contin-
gent, variable) premium and seeking protection against claims that
may occur to anyone of them. Examples for such firms in agrkul-
ture, home and health insurance abounds. Risk reduction for each
insured is then exercised through the aggregation of individual
risks. For our purposes here, we shall assume a mutual insurer
which has no assets, except for the cash premiums collected from
the insured. Further, all insured are assumed to have similar (and
statistically independent) "normal risks", and seek to minimize the
discounted cost of premium payments, plus an (implicit) cost which
they incur in case of the mutual default. This latter cost will be
called the cost of bankruptcy and is designed by the first instant at
which the mutual cash balances are null. Both the variables and the
system equations of the insurance model we use are defined in
Figure 6.
Specifically, the objective is the minimization of future (dis-
counted) premiums by insured, plus an implicit cost K of bank-
ruptcy, at which time the mutual insurance firm is no longer cover-
ing its premium paying members. Given that there are N insured,
revenues per unit time equal NP dt while disbursements due to
claims are given by a Weiner process given by (Nil dt + (5JN dw)
where Il is the average claim rate and (52 its variance. Expense rates
are given by (hx + C) relecting the cost of money held (h) and the
fixed administrative costs of maintaining the mutual insurance firm.
These expenses are deducted from cash and their costs will be
reflected in the expected cost of bankruptcy. An application of the
Bellman principle to this problem is, as in the previous example,
straightforward. Let J(x) be the optimal expected discounted future
cost objective when mutual insurance wealth is x. Thus, as long as
296

Claims

2
o dt)

Premiums
NPdt
-x Wealth

r discount rate
~ mean claim rate
0 2 claim rate variance
N number of insured
P premium payment
x firm cash-assets
C administration costs
h opportunity cost of money
K bankruptcy cost
dw a standard Wiener process

T
Min J ex) E{fe -rt pdt + Ke- rT }
PEP 0

subject to

'I: = Inf{ t > 0, X < o}

dx = NP L - ~N - hx - CJdt Nodw

xo > 0

Figure 6.

there is no bankruptcy;
J(x) = Min E{P dt + (1 - r dt)J(x + dx)}
PEf!J>
while at bankruptcy; x o and J(O) K.
 297

Insert into J(x + dx) the stochastic wealth equation for dx and
approximate it by a two terms Taylor series expansion, or
aJ a2 J
EJ(x + dx) '" J(x) + ax E(dx) + 1/2 ar E(dr)
Since E(dx) = [NP - IlN - hx - C] dt and var (dx) = (Nu 2 /2) dt,
insertion of these equalities into Bellman's definition of optimality,
yields:
. { dJ Nu 2 d2 J
Mm - r J(x) - [NP - IlN - hx - C] dx + 2 dx2 + P
}
= 0

P E fllJ, x E [0, 00], J(O) = K.

A solution of this equation will provide the optimal premium
policy pO = P(x), which is a function of x-the mutual insurance
firm wealth. In other words, the premium policy is now state
dependent expressed as a function of the mutual cash balances. It
can be shown (see [33]) that such a premium policy, in the confines
of the model formulated above is a barrier policy with;
Max P E fllJ if x ::::; x*
pO = {
o if x > x*.
In other words, there is a wealth level x*, below which premium are
collected at the maximal rate, while above this level, no premium is
collected. In a mutual insurer's term, x* is the optimal "capitaliz-
ation" of the firm. Of course, this x* will be a function of the
problem's parameters. While an analytical solution can be obtained
for x*, the assumption of a fixed premium (which is far more
manageable from the point of view of mutual insurers) provide
additional insights due to the problem's simplicity.
Say that h = 0, (the cost of holding money) and let P be a
constant. The cost J(x) is then found by a solution of Bellman's
definition of optimality, stated by the above equation. This is an
ordinary second order differential equation with one boundary at
x = 0 (l(0) = K) and an additional boundary ad x = infinity at
which point we require J(x) -+ converges to a finite value-as
x -+ infinity. A solution in this particular case is straightforward
and for simplicity is written by

J(x) = Kn(x) + (f) [1 - n(x)]

298

where
n(x) = exp [ - Rx]

R = - [n(P - Jl) - C] + [ n(p - Jl) - CJ2 + ~.

na2 na 2 na2
The interpretation of the result is that n(x) is a risk discounting, or
an "actualized" probability of ruin associated to the firm by the
insured, and which is a function of the number of insured, claim
parameters, the time preference discounting factor r, premium
payments and administrative costs. Further, analysis of the above
equations, in terms of K, the implicit cost of bankruptcy, provides
a unique way to compute this cost, based upon insured utility.
Specifically,ifu(.), u'(.) > 0, u"(.) < O,istheutilityofariskaverse
insured, with absolute risk aversion index Q = - u"(.)ju'(.), then by
the expected utility premium principle, we have in a small time
interval dt and when the firm's wealth is x;
u(w - rJ(x) dt) = u(w - dz)
where "dz" is the uncertain loss in dt and rJ(x) is the actualized
value rate of participating in mutual insurance. Taking Taylor
series approximations, we find by solution of this equation that
K = {u + Ij2Qa 2 - P[1 - n(x)]}j[n(x)r]}
(where n(x) was defined earlier) which is the implicit cost of bank-
ruptcy, rendered specific for a specific insured person.
From such a formulation, it is clear that actuarial fair premiums
(defined by expected utility valuation principles [15], [25] are related
to the cost of bankruptcy an insured is willing to sustain. Since, the
above equation is a function of x, we may state that as long as K
is (for a particular insured) smaller than the right-hand equation, he
will join into mutual insurance; otherwise, he will not. At the
margin, this bankruptcy cost equals the difference between the
expected utility losses due to damages (equaling Jl + Ij2n( 2 ) less
the odds of paying "foreover" a premium P per unit "time" and no
bankruptcy occurring.
When the claim rate is a point process, and the mutual insurer
engages in re-insurance by ceding some of its risk, the wealth
process is then (as pointed out in the previous section);

x(t) = x(O) + f~ a(x, t) dt - f~ fR Jl(dv, dz)

 299

a(x, I) = PN - hN - C - Q(()
where Q(() is the premium paid by the mutual firm for a (-excess
loss reinsurance contract. This means that our mutual insurance
firm covers all claims up to e while the excess claim is covered by the
reinsurer (a government for example). As stated previously, the
function Jl(/, A) is a function denoting the number of jumps of the
process X(/) in (0, I] and Jl(dv, dz) is a measure on R+ x R defined
as follows
Jl(t1., A) = Jl(t + M, A) - Jl(/, A)
and A is a Borel subset on R. The measure Jl(t1., A) is called the jump
measure of the process {X(/), I ~ O}.
For a Poisson stationary claim rate qN and claim (severity)
distribution function F(z), an application of Bellman's Principle
yields

o = - r J(x) + a(x) ~: + qN {f~ [J(x - z) - J(x)] dF(z)

+ [J(x - () - J(x)] Prob [z > (]} + P

J(O) = K, x E [0, (0).

Optimization with respect to P and ( will provide the mutual
insurance policies regarding premiums and reinsurance. We can
prove again that the optimum premium policy is of the barrier type
but, with a smaller barrier x* (due to reinsurance). An analytical
expression for the optimal level of reinsurance, a function of (, is
much more difficult to obtain, however. For P and ( constants, an
application of Laplace transform techniques yields a transform L(s)
for J(x) which is given by (where h = 0);

L(s) aK4>(s)
.
+p 4>(s)
s

1
4>(s)
as + qN { 1 - f: e- ZS dF(z) + feoo e-~ dFtz) } + r

a = PN - C _ QO«().
Evidently, at the limit x --+ 0, lims oo sL(s) = K while at the limit
-+
300

x -+ 00, lims-+oo sL(s) = Plr. Say that the inverse transform of ¢ is

8(x), then a transformed expression for the expected cost J(x) is

J(x) = aK8(x) + P f: 8(z) dz.

°
Since the integral of 8(x) from x = to infinity equals Ilr, we may
interpret 8(x) as a risk (adjusting) discounting mechanism, weighing
the "likelihood" of bankruptcy as a function of the mutual's
wealth. Note that this is a decreasing function of the mutual insur-
ance wealth x. The computation of this "discounting mechanism",
requiring the Laplace inversion of D(s) is not simple. For example,
if ~(z) has a Pareto distribution, with
dF(z) = (klz)" dz, k > 0, IX > 0, z ~ k.
Then using integral formulas from Gradshtein and Ryzhik [21],
p. 317, we have
1
¢(s) = r + as + qN{1 - k"'[S",-l y(1 -IX;S~) - e-l;s~I-O:/(1 - IX)]}

where y is an incomplete gamma integral.

When F(z) is an exponential distribution with parameter 11,
dF(z) = 11 e-{tZ

then
1
¢(z)
as + r + qN{s[I - e-(S+Jl)I;]/(11 + s)}

which is also difficult to inverse analytically. If ~ is small (i.e., there

is much use of reinsurance),

and so
¢(s) = [l/(qN~ + a)][r/(qN~ + a)) + sri
which yields

1 e(rlNql; + a)x
qN~ + a
a = PN - C - Q - Q(~)
 301

when ~ is very large (so there is no reinsurance), then exp [- (s +

J-l)~ '" 0 and

J-l +s
cf>(s) = (as + r)(J-l + s) + qNs
which yields by analytical inversion
(s\ - J-l) e -SIX - (S2 - J-l) e -S2 X

eoo(x) = 2 [J-l + ria + gNla] 2 I

a 2 - rJ-l a

S, = [J-l + rla2+ qNla] ± l

r-----------------
[J-l + rl a2+ qNla r:.

Evidently, when (Nt, rt, pi) then et and when ~t then e may
increase or decrease, depending upon the cost Q(~) of reinsurance
to the mutual insurer.

4. Conclusion and Discussion

The approach used in this paper in dealing with selected prob-

lems in actuarial science is not new. Nevertheless, its application to
insurance and actuarial problems has attracted relatively little
attention. Although there are many important problems acturaries
deal with where the system approach might not be useful, it can
provide in some cases (such as those demonstrated here), a concep-
tual and quantitative approach for facing the problems of risks, not
only in their static setting, but in their dynamic-time dependent
setting as well. The problems we have used have emphasized two
variables of foremost interest in insurance. These are reserves (or
assets) and the probability of ruin. In the later case, we have shown
how risk discounting (even in case of risk neutrality regarding
benefits and costs) the insurance firm's policies are intimately
related, while in the former case, a risk neutrality asssumptions has
led to completely tractables optimum premiums and reserves
policies. For practical insurance firms management purposes, the
systems approach is important as it provides a framework for
assessing the mutual dependence of an insurance firm's various
assets and liabilities in a time and stochastic setting. Even though
analytical results might be difficult to obtain, the use of numerical
302

optimization techniques and simualtion can be beneficial. The

quantitative problems to be faced are well compensated however by
our ability to integrate through the systems approach, the various
issues related to managing the firm's assets and the "actuarial risks"
the insurance firm has to sustain. Although we have used relatively
simple means of risk management (such as excess loss reinsurance),
other means, such as options contracts (Merton [27], [29] Cox and
Ross [16] and the current effervescence of contingent claims con-
tracts with various clauses [29] can be considered more specifically
in terms of the "value function" of an insurance firm. Such exten-
sions and approaches, although numerically cumbersome, provide
a broad approach to a linkage of modern financial theory, the
dynamic theory of the firm and the traditional approach to insur-
ance theory and practice.

Acknowledgements

The author is grateful for the comments of an anonymous reviewer as wen as

to the participants of the Venice Conference on Control Theory and Economic
Policies held in Venice, January 29-Feb. L 1985, and at which an earlier version
of this paper was presented. The Kruger center at the Hebrew University is also
thanked for its support.

References

1. Balzer, L. A., Control ofInsurance Systems with Delayed Profit/Loss Sharing

Feedback and Persisting Un predicted Claims, Journal of the Institute of Actua-
ries 109, 1982, Part III, No. 442, pp. 285-316.
2. Balzer, L. A. and S. Benjamin, 1980, Dynamic Response of Insurance Sys-
tems with Delayed Profit/Loss Feedback to Isolated Unpredicted Claims,
Journal of the Institute of Actuaries 107, 1980, 513-528.
3. Barrois, T., "Essai sur I'application du caJcul des probabilites aux assurances
contre l'incendie," Mem. Soc. Sci. de Lille, 1834 85-282.
4. Bather, J. A., "Contribution to K. Borch's Paper: The Theory of Risk,"
Journal of the Royal Statistical Society B29, 1967,432-468.
5. Bellman, R., Dynamic Programming, Academic Press, New York, 1957.
6. Bensoussan, A. Stochastic Control by Functional Analysis Methods, North
Holland, Amsterdam, 1982.
7. Bensoussan, A. and J. Lesourne, "Growth of Firms: A Stochastic Control
Theory Approach," in Unternehmensplanung, Springer Verlag, 1981, pp.
101-116.
 303

8. Bensoussan, A. and J. Lesourne, "Optimal Growth of a Self-Financing Firm

in an Uncertain Environment," in A. Bensoussan, P. Kleindorfer and
C. S. Tapiero (eds), Applied Stochastic Control in Econometrics and Manage-
ment Science, North Holland, 1980.
9. Bensoussan, A. and C. S. Tapiero, "Impulsive Control in Management: Pros-
pect and Applications," Journal of Optimization Theory and Applications,
1982.
10. Bohman, H., "Insurance Business Described by a Mathematical Model,"
Scandinavian Actuarial Journal, 1973,70-79.
11. Borch, K., "The Theory of Risk," J. of the Royal Stat. Soc., Series B29, 1963,
432-467.
12. Borch, K., "Dynamic Decision Problems in an Insurance Company," Astin
Bull., 1967.
13. Borch, K. B., The Economics of Uncertainty, Princeton University Press,
Princeton, N.J., 1968.
14. Borch, K. B., The Mathematical Theory of Insurance, Lexington Books,
Lexington Mass, 1974.
15. Buhlmann, H. Mathematical Methods in Risk Theory, Springer-Verlag, Ber-
lin, 1970.
16. Cox, J. c., and S. A. Ross, The Valuation of Options for Alternative Stochas-
tic Processes, Journal of Financial Economics 3, 1976, 145-66.
17. Cramer, H., Collective Risk Theory, Jubilee Volume, Skandia Insurance Co.,
1955.
18. Dayanenda, P. A. W., "Optimal Reinsurance," Journal of Applied Probability
7, 1970, 134-156.
19. Fleming, W. H., and R. Rishel, Deterministic and Stochastic Control Theory,
Berlin, Springer-Verlag, 1975.
20. Gerber, H. U., An Introduction to Mathematical Risk Theory, Monograph
No.8, S. S. Huebner Foundation, The University of Pennsylvania, 1979.
21. Gradsteyn, 1. S., and 1. M. Ryshik, Tables of Integrals, Series and Products,
New York, Academic Press, 1965.
22. Jorgenson, D. W., "Capital Theory and Investment Behavior," American
Economic Review, 52, 1963.
23. Kushner, H., Introduction to Stochastic Control, New York, Holt, Rinehart
and Winston, 1970.
24. Leland, H. E., Theory of the Firm Facing Uncertain Demand, American
Economic Review 62, 1972, 278-291.
25. Lemaire, J., "Functions d'Utilite et Assurance," Revue Beige de Statistique,
d'Informatique et Recherche Operationnelle 21, no. 4, 1982, 43-xx.
26. Lundberg, F., Zur Theorie der ruckversicherung Verdandlungskongress fur
Versicherungsmathematik Wien, 1909.
27. Merton, R. c., An Analytic Derivation of Cost of Deposit Insurance and
Loan Guarantees: An Application of Modern Option Pricing theory, Journal
of Banking and Finance 1, 1977, 3-12.
28. Pentikainen, t., "A Model of Stochastic Dynamic Prognosis," Scandinavian
Actuarial Journal, 1975, 29-53.
304

29. Smith, C. W., Jr., Applications of Options Pricing Analysis, in J. L. Bicksler

(Ed.) Handbook of Financial Economics, North Holland Pub. Co., Amster-
dam, 1979, 79-12l.
30. Tapiero, C. S., Managerial Planning: An Optimal and Stochastic Control
Approach, New York, Gordon Breach Sci. Pub!., 1977.
31. Tapiero, C. S., Applied Stochastic Models and Control in Management, North
Holland, forthcoming 1986.
32. Tapiero, C. S., "The Optimal Control of a Jump Mutual Insurance Process,"
Astin Bull. 13, 1982, 13-2l.
33. Tapiero, C. S., "A Mutual Insurance Diffusion Stochastic Control Problem,"
Journal of Economic Dynamics and Control 7, 1984, 241-260.
34. Tapiero, C. S., "A Dynamic Stock Insurance Firm Model and Dividend
Optimization," Journal of Large Scale Systems, 1985.
35. Tapiero, C. S., and D. Zuckerman, "Optimum Excess-Loss Reinsurance: A
Dynamic Framework," Stochastic Processes and Applications 2, 1982,85-96.
36. Tapiero, C. S., and D. Zuckerman, "Optimal Investment Policy of an Insur-
ance Firm," Insurance: Mathematics and Economics 2, 1983, 103-112.
37. Tapiero, C. S., D. Zuckerman, and Y. Kahane, "Optimal Investment-
Dividend Policy of an Insurance Firm Under Regulation," Scandinavian
Actuarial Journal 198, 65-86.
38. Tetens, J. N., Einleitung zur Berechnung der Leibrenten und Anwartschaften,
Leipzig, 1786.
39. Vajda, S. (ed.), Non-Proportional Reinsurance, E. J. Brill, Leyden, 1955.
40. Van Loon, P., A Dynamic Theory of the Firm: Production, Finance and Invest-
ment, Berlin, Lecture Notes in Economics and Mathematical Systems.
Springer-Verlag, 1983.
 305

CHAPTER 17

CAPITAL THEORETICS, BUSINESS CYCLES, AND

FEEDBACK POLICY: AN EXPERIMENT IN
MACROECONOMIC CONTROL

James H. Gapinski
Florida State University, USA

1. Introduction

Do feedback controls have favorable countercyclical effects? Can

they make matters worse? How responsive is the macro system to
the control parameters? These questions, which delineate the
present effort, are hardly new. They have been asked and reasked;
they have been answered and reanswered. Along the way they have
given rise to a host of precedents including the seminal pieces by
Friedman (1948; 1959, pp. 89-99) and Phillips (1954; 1957); the
challenging endeavors by Pack (1968), Fischer and Cooper (1973),
Sargent and Wallace (1976), Phelps and Taylor (1977), Howitt
(1981), and Gapinski (1982, pp. 228-237); and the strict optimal-
control works by Chow (1975; 1981) and Garbade (1975; 1977).
What is new here is not the questions but rather the context within
which they are asked. It is capital theoretic. I
Because the precedents treat lightly the nature of the capital that
underlies the macro system, they suggest that capital is incon-
sequential from the standpoint of control. Yet the pioneering
research on capital heterogeneity by Johansen (1959) and Salter
(1960) along with its elaborations by Phelps (1963), Park (1966),
Gordon (1973), and Gapinski (1981; 1985) suggest just the
opposite. They suggest that the nature of capital bears heavily on
control.
To ascertain more precisely capital's involvement in feedback
control, the present inquiry appeals to the macro model by Gapin-
ski (1985). That paradigm takes capital to be putty-tin: Capital and
labor are substitutable before (ex ante) and after (ex post) capital

C. Carrara and D. Sartore (eds). Developments of Control Theory for Economic Analysis
© 1987 Martinus Nijhoff Publishers (Kluwer), Dordrecht-
306

is built, but they may be less substitutable ex post. It allows the

grade of tin to be selected from a continuum having putty at one
extreme and virtual clay at the other, and consequently it embraces
capital types that span much of the range from putty-putty, where
substitution is equally easy ex ante and ex post, to putty-clay, where
it is possible only ex ante. Furthermore, the model prescribes
proportional feedbacks of the Phillips sort for government pur-
chases and the money supply; that is, both fiscal policy and monet-
ary policy are control oriented.
Section 2 reproduces the model, albeit in a somewhat abbreviated
fashion, to make the discussion complete. Section 3 reports on the
method used for determining its temporal motion, and Section 4
looks at the effectiveness of feedback maneuvers under different
capital configurations. It also examines the sensitivity of the Phillips
curve to feedbacks in that capital environment. Section 5 offers
concluding comment.

2. A Macro Model

The model can be conveniently viewed as having six parts. Sec-

tion 2.1 addresses the production functions applicable to capital
ex ante and ex post. Section 2.2 describes the procedures followed
by entrepreneurs in specifying new capital and in operating the
capital already installed. Section 2.3 considers the components of
commodity supply and demand whereas Section 2.4 contemplates
prices. Section 2.5 notes measures of unemployment and capacity
utilization: Section 2.6 sketches the tax structure and gives income
definitions.

2.1. Production functions

In designing the capital to be built at, say, time v, entrepreneurs
refer to a blueprint which indicates for any quantity of capital Iv the
quantity of output Yv* that would be forthcoming if the quantity of
labor were L:. This ex ante production function observes the CES
postulate:
(1)
Technical progress is entirely embodied in new capital. It is Hicks
neutral, and it occurs at the exponential rate f1.*. Decreasing returns
 307

to scale prevail, 0 < f3* < 1, and the possibilities for factor sub-
stitution exist within bounds: 0 < 0"* < 00, where 0"*, the ex ante
elasticity of substitution, equals 1/(1 + Q*). Thus, - 1 < Q* < 00.
Moreover, 0 < y* < 00 while ~i > 0, ~L > 0, and ~i + ~L = 1.
The Yv* and L! may be termed respectively the capacity output and
the capacity employment of vintage v capital.
Once extant vintage v ordinarily encounters reduced substitution
possibilities, and those which remain are summarized by the ex post
production function
v
.1 vI -- Y e /l'V[J:'01 I-a
vI + J: L-a]-P'/a
'0 Lvt . (2)

In this CES rendering Yvt denotes the quantity of output generated

from vintage v at time t, LVI represents the quantity of labor
employed on vintage v at t, and IVI signifies the quantity of vintage
v used at t. Capital gestation in discrete time implies v ~ t - 1.
Operated without slack, machines remain intact for e periods but
then instantly disintegrate without trace. Therefore, IVI = Iv for
t - e ~ v ~ t - 1, but I vt = 0 for v < t - e. Since capital
ex post may be putty but cannot be clay, 0 < 0" ~ 0"*. Para-
meter 0", the ex post elasticity of substitution, equals 1/(1 + Q) and
translates the substitution inequalities into 00 > Q ~ Q*. Para-
meters y, ~/' and ~L obey the restrictions of their ex ante counter-
parts.
Ex post function (2) is linked to ex ante relation (1) in two
respects. First, when vintage v operates at capacity, YVI must equal
Yv*. Second, in the same circumstance, the marginal rate of techni-
cal substitution from (2) must equal that from (1). These two
linkages enable y and ~I to be expressed in terms of ex ante mag-
nitudes, and they lead to a restatement of the ex post production
function:
1':vI = Y* e/l'V[J:* I-a'
'01 v
+ J:* L'-a']-P*/a*
'oL v q,P'/a , (3)

where q, = [(~1I~!)k:Q-a' + k:Q]/[(~r/~!) k:a- a* + k~l] with k! =

Iv/L ! and k vt = Iv/L vt .
Despite its clumsiness, expression (3) is merely equation (2) in
disguise, and thus it exhibits normal CES properties. Furthermore,
it manifests the intuitive relationship with function (1). Let Q > Q*.
Then in general YVI = Yv* when LVI = L!, and Yvt < Yv* when
LVI =I L!. For Q = Q*, function (3) becomes identical to function
(1). In what follows 00 > (! ~ Q* > o.
308

2.2. Optimization procedures

Entrepreneurs design new capital to maximize its net discounted
quasi-rents after taxes, R t • This tactic yields for vintage t the objec-
tive function
u=t+O
Rt = L CY;u - w~ f tu )(1 + r~Y-u - Ir
u=t+1
u=t+O
L !nCYtu - w~ f tu - au-Jt)(1 + rn t - u, (4)
u=t+1
in which ~u and f tu signify respectively the quantities of output and
labor expected for vintage t at time u. The au - t represent accounting
fractions allowed in calculating depreciation deductions for tax
purposes; nonnegative, they sum to unity.
w~ and r~ are respectively the decision wage rate and the decision
interest rate. To calculate w~ entrepreneurs take into account the
payroll tax rate which they face, ow' and they presume that workers
always negotiate money wage settlements to prevent any decline in
the real wage. Although entrepreneurs recognize that the real wage
may depart from constancy in either direction during the life of
vintage t capital, they feel that over that lifetime the real wage will
"average out" to the level expected for the present. No change
in !w, or in the profit tax rate !n' is anticipated making
w~ = wt (1 + !w), wt notating the real wage expected to be paid to
labor currently. In a similar fashion, entrepreneurs suppose the real
interest rate to remain constant on balance over the life of capital.
Thus r~ = ~ - Pt,
~ and Pt
denoting respectively the nominal
interest rate and the inflation rate expected to prevail at t. Wage,
interest, and inflation expectations are all formed from past experi-
ence and adhere to the rule
j=~

ht = [0.511(11 + l)r l L (11 +

J=I
1 - })b t - j , (5)

where b r- i symbolizes the actual value of bat t - j. For the case of

inflation, b r - J becomes Pt-j, with Pt-j being (Pt-j - Pt-j-I )!Pt-J-I·
Of course, Pr-J signifies the price level at t - j.
Vintage t is intended for use at capacity, and thus R t simplifies to
Rt = (1 - !n)(Yt - w~Ln L (1 + rt)t-u
u

(6)
 309

When maximized with respect to Li and II subject to production

function (1), equation (6) gives
oY*/oL*
t I wi, (7)

o Yr*/oII qi, (8)

qi = (1 - T,,)-l [1 - T" ~ au- (1

t + riY-u ] I~ (1 + rn l
-
u.

(9)
Price variable qi in equation (9) is real user cost.
Optimum marginal products (7) and (8) and production function
(1) can be used to express capacity labor, investment, and capacity
output in terms of the prices wi and qi. Although awkward, the
resulting expressions can be written neatly as
L*I L*(wi, qn, (10)
I(wi, qn, (11)

Y*(wi, qn· (12)

On the ex post side of the ledger, each of the einstalled vintages
is manned to maximize its current after-tax quasi-rent R vl , a
criterion analogous to the one adopted in deciding the capacity level
of new capital. Thus, for any installed vintage entrepreneurs choose
labor LVI under real wave WI to maximize
R vt = [Yvt - wl(1 + Tw)Lv,] - T,,[ Yvt - wt(1 + Tw)Lvl - at-Jv]·
(13)
Optimization renders
o~,/oLv' = wl(1 + Tw). (14)
Each vintage is used to the point where its marginal product of
labor equals the real wage inclusive of tax.
Equi-marginal-product rule (14) implies that LVI and Yvt are
related to WI and Tw. For simplicity those relationships are cast as
the implicit functions
0, (15)
0, (16)
where the v subscripts in the function symbols serve as reminders
310

that the functions differ by vintage. For any vintage, however, L vt

and 1:t vary inversely with WI' The L vt ranges from zero to infinity
as W t declines, and the 1:t does so from zero to a finite maximum.
These patterns for each individual vintage translate into similar
ones for all () vintages combined, and aggregate functions obtain:
0, (17)
N(Yr, WI' Tw) O. (18)
Et represents total employment L~:::=~ LVI while Yr represents total
output L~:::=~ Yvt ' From equation (18) it can be shown that, if all
installed vintages begin at capacity, the inverse relationship
between WI and Yr is more pronounced for slight changes in Yr the
smaller is (J.

2.3. Commodity supply and demand

Each period entrepreneurs gear production to meet anticipated
sales Dt , expectations obeying formula (5). More precisely,
(19)
Actual sales D t , in turn, consist of real consumption expenditures
Cn real investment, and real government purchases G t ; succinctly,
Dt = CI + II + GI . (20)
Consumption follows the schedule
(21)
Ydt signifies real disposable income while rrt symbolizes the expected
real after-tax interest rate: rrt = ~(1 - Ty) - Pt,
Ty notating the
personal income tax rate. Government purchases have a counter-
cyclical orientation inasmuch as they are based on proportional
feedback controls; namely
GI = ¢o + ¢l (UI _ 1 - 0)+ + ¢2(Pt-l - p)+. (22)
~_I is last period's unemployment rate. 0 and p identify the
respective target rates of unemployment and inflation, and the +
superscript of each parenthetic expression indicates that a deviation
from target triggers policy only if it is positive. Negative misses are
treated as zero to prevent government action from disturbing
acceptable economic performance. Countercyclical fiscal policy is
asymmetric.
 311

2.4. Prices
As argued in conjunction with equation (4), new capital is designed
under the presumption that workers bargain the money wage to
avert any erosion of the real wage. Workers do try to preserve reals,
and therefore
Xt = Pt, (23)
Xt standing for the money wage at t. By virtue of the equality
wt = xt - Pt( wt/w t_ I ), it follows that
(24)
An equation for the real wage was already introduced. Relation
(18) ties Y, and W t together for any T w , but with Y, set by statement
(19), it yields W t • Formula (18), then, is the real wage equation.
Nominal interest rate it emerges from equilibrium in the money
market. The demand for real balances is
Mdt/pt = Va + VI 1'; + v2i t(1 - Ty) + v3Mdt _ l /pt_l, (25)
where Mdt symbolizes the nominal demand for money at t and
where the nominal after-tax interest rate enters to measure the
opportunity cost of holding money. Nominal money supply at t,
Ms t , follows the feedback rule
MS t = t/to + t/t1(Ut- 1 - 0)+ + t/t2(Pt-1 - p)+. (26)

Monetary policy, like fiscal policy, is countercyclical and asym-

metric. In equilibrium
it = -[v 2(1 - Ty)]-I[VO + VI Y, + v3Ms t-dpt-1 - Mst/pt]·
(27)
Nominal interest varies directly with real income and withcontem-
poraneous price; it varies inversely with contemporaneous money.
Restrictions in the money market subject nominal interest to a floor
of if.

2.S. Factor utilization

Labor supply LS t depends upon the real after-tax wage rate, and
it moves through time in step with population. Specifically,
LS t = AD + AI w/l - TJ +A2t. (28)
312

Labor supply combines with total employment to determine the

unemployment rate:

(29)
Forces controlling labor supply keep this rate from falling below
a value ~.
Machine utilization is reckoned from the capacity outputs of the
installed vintages; namely,

C Vt = Yt :' L Yv*, (30)

v

with CVt denoting the rate of capacity utilization. Elementary

algebra shows CVt to be a convex combination of the utilization
rates Yvt / Yv* of the () vintages.

2.6. Taxes and incomes

Rounding out the model are several tax variables and their
requisite income definitions.
Real payroll tax TW t depends proportionally upon real earnings
Tt;, real profit tax Tnt depends proportionally upon real profits Of'
and real personal income tax TYt depends proportionally upon real
personal income YPt less real transfer payments Tr t • Transfers, in
turn, depend linearly upon the unemployment rate. In equation
form these statutes appear as
TW t (1 + W)TwW;, (31)

Tnt TnO I, (32)

TYI T..{YPI - Tr t ), (33)

Trt Xu + XlVI' (34)

Fraction W in schedule (31) expresses the payroll tax contribution

by employees relative to that by employers. Furthermore,
Tv, = H'{E{,Ot = l"t - w{(1 + Tw)Et - Lv a{-Jv, and Yp{ = l"t -
o I 'LJv - O{ - Tw{ + Tr{.
From these relationships total tax net of transfers becomes T, =
Tw{ + Tnt + TYt - Trt making the budget deficit H{ = G{ - Tt .
Personal disposable income is Yd t = Yp{ - Ty{. Of course, TI' HI'
and Yd{ are all reals.
 313

3. Simulation and Business Cycles

Because the model does not lend itself to analytical investigation,

study of its temporal nature proceeds numerically through simu-
lation. In structuring the simulation work, the economy is posited
to begin from a stationary state. A stationary state is convenient
because it is independent of (J and thereby allows the economy to
have the same origin and the same baseline solution regardless of
capital's grade of tin. With the initial conditions of the model
programmed to end in 1982, the simulation apparatus begins its
iterations at 1983, runs through 2015, and yields along the way a
faithful reproduction of the stationary state. Although it produces
a drift of variables away from their true stationary-state values, the
approximation errors are slight amounting to less than 0.50 per-
cent, and most often less than 0.05 percent, even after 33 iterations.
The simulation horizon is set at 33 periods to include by a handy
date two full generations of capital, e being fixed at 15. A complete
list of parameter assignments appears in Table 1, which also lists for
the stationary state the magnitudes and dimensions of the variables.
Some parameter assignments, such as those for the tax coef-
ficients, are made by using ordinary least squares to fit the relevant
functions described in Section 2 to annual data covering the US
experience from 1953 to 1982. Other assignments, such as those for
the money demand and feedback control coefficients, are made by
adjusting regression estimates likewise obtained to theoretically
prudent levels. Still others, notably those for the ex ante par-
ameters, are set somewhat arbitrarily. The malleability regimes,
defined by a (J* of 0.90 and the seven (J alternatives 0.90,
0.80, ... , 0.30, allow for putty-putty and for an approach to
putty-clay. (J values less than 0.30 were tried, but for them the
simulation program had difficulty generating solutions. Appar-
ently, parameters in the separate equations become interdependent
through the working of the entire model, and assigning values to
some automatically constrains the values of others. For consistency
with the properties of stationary state, the time variable represent-
ing in equation (l) the technology level and in equation (28) the
population count is held constant-at 82 to capture 1982 events
-in all simulations.
Stationary-state conditions persist until 1985, when autonomous
real government purchases, <Po, rise by 10 billion dollars to 6lO.414
314

Table. 1. Numerical framework of the simulation apparatus."

Assignments for the parameters
Parameter Value Parameter Value Parameter Value
al 0.160 Ao -49.510 u 0.900,
a2 0.133 AI 0.110 0.800,
a3 0.120 .le 2 1.940 0.700,
a4 0.107 /1* 0.020 0.600,
as 0.093 Vo 62.690 0.500,
a6 0.080 VI 0.040 0.400,
a7 0.080 V2 - 1200.000 or 0.300
as 0.080 V3 0.620 u* 0.900
0.450
a9 0.080 (j) 0.800 'n
aJO 0.067 p 0.050 '. 0.050
all 0.000 ¢o 600.414 'y{} 0.150
15.000
aJ2-a I S 0.000 ¢I 85.000
fJ* 0.800 ¢2 - 39.000 a 0.400
Xo - 38.970 t/lo 500.000 (L 0.600
XI 2228.210 t/I, 111.840 (0 125.000
IJ 5.000 t/l2 - 160.500 (I 0.500
y* 1.500 Vf 0.018 (2 0.400
if 0.005 0 0.050 (3 - 1970.590
Characteristics of the stationary state
Variable Value Dimension Variable Value Dimension
C 638.186 Billion Dollars r, 0.045 Decimal
CV 1.000 Decimal T 303.982 Billion Dollars
E 104.735 Million Persons Tn 198.889 Billion Dollars
G 600.414 Billion Dollars Tr 72.441 Billion Dollars
H 296.432 Billion Dollars Tw 68.282 Billion Dollars
I 211.600 Billion Dollars Ty 109.251 Billion Dollars
i 0.052 Decimal V 0.050 Decimal
L* 6.982 Million Persons W 758.690 Billion Dollars
Ls 110.247 Million Persons w 7.244 Thousd. Dollars
Ms 500.000 Billion Dollars w* 7.606 Thousd. Dollars
p 2.827 Index x 20.482 Thousd. Dollars
p 0.000 Decimal Y 1450.200 Billion Dollars
IT 441.976 Billion Dollars Y* 96.680 Billion Dollars
q* 0.115 Decimal Yd 691.531 Billion Dollars
r* 0.052 Decimal Yp 800.782 Billion Dollars
"With minor exceptions required by style, entries appear in alphabetical order.
 315

Y o· = .90
1550

1500

1450

1400

1350 0=.30

1985 1995 2005 2015 Year

Figure 1.

and remain at that level forever. According to the model, this

action prompts traditional responses at least initially for any par-
ticular (0'*, 0') pair. Output increases and, through a decline in the
real wage, accelerates inflation. Nominal interest then rises in
"crowding out" fashion. Implied in this scenario is a trade-off
Phillips curve: More output means less unemployment and more
inflation.
These responses, although traditional for any given elasticity
pair, differ in strength and character across pairings. A glimpse of
those differences is given by Figure 1, which tracks the reaction of
output Y to the purchase policy under selected 0' assignments.
When capital is putty-putty (0'* = 0' = 0.90), .y follows a path
that oscillates mildly. Business cycles become more severe as 0' falls
from the putty mark, and by the time that 0' reaches the 0.30 norm
 w
0\
-

Table 2. Variability measures by types of capital.a

Measure (J = 0.90 (J = 0.80 (J = 0.70 (J = 0.60 (J = 0.50 (J = 0.40 (J = 0.30
Variance of Y 3.406 4.919 8.016 16.772 47.128 202.974 2019.100
Variability relative for Y 1.000 1.444 2.353 4.924 13.836 59.591 592.789
Variability composite 1 1.000 1.302 1.858 3.515 8.646 3 I.I80 355.790
Variability composite 2 1.000 1.330 1.957 3.797 9.684 36.862 403.190
d(J* = 0.90.
 317

for virtual clay, they take the shape of decided explosions? This
propensity of smaller (J to generate greater volatility reflects the
structural relationships noted in connection with equation (18). A
slight shock to Y has greater impact on w for smaller (J, and that
greater impact reverberates throughout the entire system.
Table 2 looks more closely at the volatility issue. Corroborating
the movements displayed in Figure 1, it indicates for a (J* of 0.90
that the variance of Y rises as (J begins its fall from putty and that
it soars as (J declines past the middle grade of tin, 0.60. The varia-
bility relative for Y, simply the Y variance expressed relative to its
(J = 0.90 level, drives home the point: For (J = 0.30 the variance

is almost 600 times its value for (J = 0.90. Unruliness becomes

acute in the vicinity of clay.
This characteristic holds widely. Variability composite 1 reported
in Table 2 is an index fashioned from the variability relatives for the
four series E, w, p, and i, those choices being made without loss of
generality. Each posting for the composite represents the arithmetic
average of the four relatives calculated under the corresponding (J.
Clearly, the index increases as (J falls, and the increase becomes
bolder as (J approaches clay. Variability composite 2 is offered for
completeness. Similar to composite I but defined to include Y as
a fifth series, it leads to similar conclusions. 3

4. Control from Feedbacks

Consensus wisdom maintains that feedback controls can improve

the performance of the macro system. Although such policies might
make matters worse, modest action tends to produce favorable
countercyclical effects. This consensus position, drawn from the
precedents cited in Section 1, extends to the present circumstance of
putty-tin capital; the evidence comes from simulation.
A first step in constructing the experiment is to decide how to
limit its scope. Seven elasticity pairs and four feedback coefficients
-the latter perhaps being treated singly, doubly, triply and quad-
ruply-imply that even a few alternative values for the control par-
ameters would necessitate a large number of simulations. For
instance, only three parametric sets could require more than three-
hundred runs. Guides for keeping the exercise manageable are
needed, and they come from the purchase-policy simulations
318

already in hand. Those seven scenarios indicate both that the

inflation response to the shock is uniformly weak across a and that
the unemployment responses becomes strong only below the middle
grade of tin: only for a ~ 0.50. From this information it seems
reasonable to narrow the feedback experiment to the two control
parameters 4>1 and I{I1 and to conduct it solely for the three (J values
0.50, 0.40, and 0.30.
Each feedback coefficient begins a simulation with its value at the
stationary-state setting, 4>1 = 85.00 and I{I1 = 111.84. But when 4>0
assumes a new level in 1985, 4>1 alone, I{I1 alone, or each together
assumes a new level and stays there forever. Six alternatives for 4>1
are permitted; namely, 185.00, 285.00, ... , 685.00. The six for
I{I1 are 2ll.84, 31l.84, ... , 7ll.84. Counting the three initial
maneuvers conducted with 4>1 and I{I1 permanently unchanged at the
starting values, the feedback experiment encompasses a total of 57
simulation trials.
Table 3 reports some of the findings. In particular, it gives the
variability composite 2, the broadest variability measure exam-
ined, along with its discounted counterpart. The discounted version
is constructed exactly as the original except that the "variance" of
any component series is computed from squared deviations dis-
counted at the stationary-state rate of interest. 4 Discounting ought
to provide a judicious second measure since, as Figure 1 attests,
most of the variation in a series develops late.
Because 4>1 impacts aggregate demand directly whereas I{I1 does
so indirectly through the interest rate, 4>1 policy has a more potent
countercyclical effect than does I{I1 policy. For example, when
(J = 0.40 a boost of 4>1 by 600 units to 685.00 drops variability

composite 2 to 68 percent of its initial level while a 600-unit

increment in I{I1 drops it to only 91 percent, almost the same
percentage achievable by 4>1 with merely a 100-unit hike. Counter-
cyclical monetary policy, operating through the interest rate, means
hightened interest rate variability, and this well-known trait contri-
butes to the poorer showing of I{I1: '" 1 increments work partly
against themselves when the gauge of success includes the interest
rate. With a = 0.50 composite 2 actually rises past unity as I{I1
gathers strength solo implying that overall I{I1 worsens macro per-
formance then. Discounted composite 2 is only slightly more for-
giving in its assessment of that case.
Table 3. Effectiveness profiles of feedback controls:

4>1 t/I[ Variability composite 2 Discounted variability composite 2

(J (J (J (J (J (J
= 0.50 = 0040 = 0.30 . = 0.50 = 0040 = 0.30
85.00 111.84 1.000 1.000 1.000 1.000 1.000 1.000
185.00 111.84 0.973 0.929 0.899 0.979 0.938 0.908
285.00 I I 1.84 0.948 0.877 0.823 0.960 0.894 0.840
385.00 111.84 0.923 0.828 0.756 0.941 0.853 0.778
485.00 111.84 0.900 0.765 0.696 0.922 0.797 0.722
585.00 111.84 0.877 0.720 0.642 0.905 0.758 0.672
685.00 I I 1.84 0.856 0.679 0.589 0.888 0.722 0.622

85.00 111.84 1.000 1.000 1.000 1.000 1.000 1.000

85.00 211.84 0.999 0.997 0.977 0.997 0.980 0.983
85.00 311.84 0.999 0.970 0.931 0.996 0.975 0.942
85.00 411.84 1.001 0.963 0.907 0.995 0.971 0.922
85.00 511.84 1.003 0.934 0.886 0.996 0.947 0.906
85.00 6 11.84 1.006 0.923 0.865 0.997 0.939 0.890
85.00 711.84 1.010 0.912 0.850 0.999 0.931 0.879

85.00 111.84 1.000 1.000 1.000 1.000 1.000 1.000

185.00 211.84 0.972 0.921 0.874 0.977 0.932 0.887
285.00 3 11.84 0.947 0.841 0.782 0.956 0.863 0.805
385.00 411.84 0.924 0.777 0.703 0.936 0.810 0.734
485.00 511.84 0.903 0.719 0.606 0.919 0.761 0.643
585.00 611.84 0.884 0.665 0.551 0.903 0.714 0.590
685.00 711.84 0.837 0.619 0.504 0.863 0.674 0.545
w
a(J* = 0.90. '-0
320

Y
0°= 90
1550
0=.30

::\\
.. \
\\
1500
~\
\ ,
:1
:1
:\
'.\
\\
1450 \\
\1
~\
\\
\\ +, = 685.00
'1', = 711.84
1400
+, = 685.00
,-
:'1', = 111.84
\. / +, = 85.00
-- '1', = 711.84

+, = 85.00
1350 '1', = 111.84

1985 1995 2005 2015 Year

Figure 2.

As (J falls toward clay, the feedback policies operate more fre-

quently and more forcefully and therefore have more pronounced
countercyclical repercussions. Given declining (J maximum <P1
reduces variability from about 85 percent to about 60 percent
according to either composite. Maximum 1/11 does so from 100
percent to 85 percent approximately. Applied in tandem the feed-
backs occasion results that generally agree with intuition. For
(J = 0.50 the stabilizing effect of <P1 ameliorates the perverse conse-

quence of 1/11 and presses both composites progressively below unity

across the parameter grid. For (J = 0.30 the joint action drives
variability, measured either way, to roughly half of its original self.
Macro performance under the controls can be summarized pic-
torially. Figure 2, like Figure 1, focuses on output Y though it
 321

1jI1
0
. =.90

711 84 \ " "-

\ - - 86 ~= .93 " ~ = .86
\ 0 = 40 0=.50 I 0 = .50
\ I
\ I
\ I
\ I
\ -= = 86 \ I
\ - \ I
\ 0 = .30 \ I
51184 \/ \ I
\ \ I
\ \ I
\ \ I
\ \ I
\ \
I
\ \
~ =.93 I
\ \
0=.40 \ \ I
\ \ I
31184 \ \
\ \
\ \
\ \
~=.93 \ \
0=.30 \ \
\ \
\ \
\ I
\ I
111.84
85.00 285.00 485.00 685.00 .1

Figure 3.

confines itself to the case of virtual clay, Since the feedbacks remain
dormant over the years immediately following the C/>O shock, output
then exhibits the same temporal motion for any c/>] and 1/11' Trigger-
ing occurs soon after 1995, and the tempering effect begins to take
hold. It becomes especially noticeable after 2005. To avoid clutter,
Figure 2 suppresses for the midterm the curves corresponding to
single policies; if drawn, they would lie between the two illustrated.
Figure 3 provides a fuller portrayal of macro behavior under the
policies. Specifically it sketches isovars, each being a curve that
shows the various combinations of the feedback coefficients necess-
ary to achieve a given degree of countercyclical control. Control is
reckoned in terms of variability composite 2, which is labeled S for
convenience. 5 At the origin S = 1.00 independently of (T,
Apart from indicating that isovars are hardly simple displace-
ments of each other, Figure 3 confirms several properties noted
322

previously. First, <PI is stronger medicine than is 1/11: For any (J the
dosage needed to work a prescribed amount of cure is less than that
ofl/l l . With (J = 0.40, for instance, as of 0.93 requires that <PI rise
by 100 units to 185.00, but it obliges that 1/1 I rise by 400 units to
511.84. Second, 1/11 has a destabilizing effect near the middle grade
of tin. When (J = 0.50 increasing 1/1 I first evokes a weakly favorable
response, but later the exacerbating tendency of 1/11 becomes domi-
nant. Maintaining the same S in that situation requires a furthering
of action from <PI and imparts a positive slope to the corresponding
isovars.6 For (J closer to clay, however, 1/11 always has a tempering
influence, and hence it and <PI are substitutes as the four negatively
sloped contours attest. Third, smaller (J mean stronger countercycli-
cal effects. As (J declines the macro system becomes more responsive
to the feedback parameters, and therefore a particular level of
control can be attained through less aggressive policy action. The
leftward shift in isovars for falling (J illustrates the case.
Section 3 reported that purchase policy is inclined to produce a
trade-off between inflation and unemployment. To see how feed-
back policy influences this relationship, the expression Pr = (Xo +
(XI / Ur is estimated within the simulation program by ordinary least
squares using the 31 observations generated from 1985 to 2015.
Intercept (xo represents the minimum rate of inflation: the asymp-
totic floor of the Phillips curve. Ratio - (XI /(Xo sites the price stability
point: the rate of unemployment at which price remains unchanged.
Both quantities, along with (XI' are scale-adjusted in Table 4, which
lists the magnitude of feedback policy as incremental from the
stationary-state values. A 300.00 in the second column means
<PI = 385.00 for <PI policy alone, 1/11 = 411.84 for 1/11 policy alone,
and <PI = 385.00 and 1/11 = 411.84 for both policies together.
Given the rudimentary nature of the regression, Table 4 remains
silent about the econometric properties of the fits.
On balance intensifying countercyclical policy rotates the Phillips
curve counterclockwise; that is, increased policy magnitude tends to
raise the minimum inflation rate, flatten the slope, and advance the
price stability point. In accord with the variability results, the most
pronounced rotation occurs by combining a (J of 0.30 with the
policies applied jointly. Then, as the control coefficients expand
over their grid of values, (Xo rises from - 0.0196 to - 0.0079, (XI
declines from 0.000490 to 0.000255, and - (XI /(Xo increases from
0.0249 to 0.0324. By contrast, and again in accord with the earlier
Table 4. Phillips curve characteristics under different types and magnitudes of feedback policy:
(J Policy Policy type
magnitude:
<PI policy alone 1/11 policy alone <PI and 1/11 policies jointly
base plus
lOao 10001X 1 - lOa l /ao lOao 1000a l - lOal/ao 101X0 1000al - lOa,/ao
0.50 0.00 -0.169 0.762 0.450 -0.169 0.762 0.450 -0.169 0.762 0.450
0.50 100.00 -0.168 0.758 0.450 -0.170 0.764 0.450 -0.168 0.759 0.451
0.50 200.00 - 0.167 0.753 0.451 -0.170 0.765 0.450 -0.167 0.755 0.452
0.50 300.00 -0.166 0.749 0.452 -0.170 0.765 0.451 -0.166 0.751 0.453
0.50 400.00 -0.165 0.744 0.452 -0.170 0.766 0.451 -0.165 0.748 0.453
0.50 500.00 -0.163 0.740 0.453 -0.170 0.767 0.451 -0.164 0.745 0.454
0.50 600.00 -0.162 0.737 0.454 -0.170 0.767 0.452 -0.166 0.757 0.455
0.40 0.00 -0.141 0.573 0.407 -0.141 0.573 0.407 -0.141 0.573 0.407
0.40 100.00 -0.138 0.568 0.411 -0.142 0.580 0.409 -0.138 0.568 0.412
0.40 200.00 -0.135 0.557 0.413 -0.141 0.580 0.411 -0.135 0.562 0.418
0.40 300.00 -0.131 0.546 0.416 -0.140 0.579 0.412 - 0.131 0.553 0.422
0.40 400.00 -0.129 0.542 0.420 -0.141 0.585 0.415 -0.128 0.544 0.427
0.40 500.00 - 0.126 0.534 0.423 -0.140 0.585 0.417 -0.125 0.538 0.431
0.40 600.00 -0.123 0.526 0.426 -0.140 0.585 0.419 -0.123 0.533 0.435
0.30 0.00 -0.196 0.490 0.249 -0.196 0.490 0.249 -0.196 0.490 0.249
0.30 100.00 -0.189 0.493 0.261 -0.198 0.508 0.257 -0.180 0.476 0.264
0.30 200.00 -0.172 0.455 0.265 -0.189 0.497 0.262 -0.151 0.399 0.265
0.30 300.00 -0.153 0.407 0.267 -0.180 0.473 0.263 -0.119 0.314 0.264
0.30 400.00 -0.133 0.355 0.267 -0.169 0.445 0.263 -0.108 0.311 0.289
0.30 500.00 -0.113 0.303 0.268 -0.158 0.414 0.262 -0.092 0.282 0.305
0.30 600.00 -0.105 0.295 0.281 -0.146 0.381 0.261 -0.079 0.255 0.324 w
N
w
.0'* = 0.90.
324

findings, the smallest rotation occurs by matching a (5 of 0.50 with

'" 1 policy alone. There the Phillips curve remains almost stationary
throughout the entire range of control values.
The Phillips curve differs by type and magnitude of control.
Those differences, however, must be viewed as slight even in the
most dramatic case because even there 600-unit increases in the
control coefficients only move the price stability position forward
by a single percentage point. In other words, the structural shifts
brought about through feedback policy have negligible adverse
stagflation side effects.

5. Concluding Comment
Capital bears on feedback control. The more dissimilar is capital
ex post versus ex ante, then the more unruly is business cycle
activity, and the more frequently and forcefully is a feedback
triggered. Capital akin to putty-clay causes marked oscillations
and therefore provides a solid opportunity for feedback success.
Conversely, putty-putty generates mild fluctuations, for which
controls are not particularly relevant.
In the capital-theoretic context, feedbacks can improve macro
behavior. They can worsen it. Nevertheless, behavior becomes
more responsive to control as capital ex post tends toward clay, and
near clay the prospect of improving performance and of doing so
appreciably is quite strong while that of worsening it is rather weak.
In fact, near clay feedback policy, including the monetary initiative,
neither deteriorates the overall behavior of the system nor provokes
serious stagflation movements even when the magnitude of policy
increases six- or eightfold. By virtue of these findings, the familiar
propositions, advanced with little regard to the nature of capital,
that active countercyclical maneuvers can make matters worse and
that policy should be applied lockstep without feedbacks may need
to be substantially revised.

Notes
l. It should be recalled at the outset that rational expectations need not disable
feedback policy. On the contrary, feedbacks can flourish under rationality as
Phelps and Taylor (1977), Chow (1981, pp. 225-238), and Howitt (1981) show.
Sargent and Wallace (1976) describe the standard rational-expectations
position.
 325

2. In keeping with one convention, business cycles are depicted as fluctuations in

output. A broader interpretation, one more closely aligned with the view held
by the National Bureau of Economic Research, is implicit in the composite
measures to be considered shortly.
3. Several tests were conducted with another choice for 0"*, this one being the
approximate Cobb-Douglas 0"* of 0.99. Applicable mutatis mutandis it led to
conclusions regarding unruliness that resembled those reached under the 0"* of
0.90, but it caused all variability measures to rise even faster as 0" fell toward
0.30. Making capital ex post more unlike capital ex ante, either by lowering
0" or by raising 0"*, increases the system's volatility.
4. For any series Z with mean 2, discounting leaves
[~Z:fm (Zh - 2)2 1.052 198H l/31.
5. Most points used to map the curves were located by running additional
simulations over the parameter grid. Upwards of 125 new runs were involved
in the mapping process.
6. The backward bend at the "top" of the :=: = 0.86 contour for 0" = 0.50 is a
curiosum that, although strictly inexplicable, happens to coincide with a very
steep region of the :=: function.

References

Chow, Gregory C. Analysis and Control of Dynamic Economic Systems. New

York: John Wiley and Sons, 1975.
Chow, Gregory C. Econometric Analysis by Control Methods. New York: John
Wiley and Sons, 1981.
Fischer, Stanley, and Cooper, J. Phillip, "Stabilization Policy and Lags." Journal
of Political Economy 81 (July/August 1973): 847-877.
Friedman, Milton. "A Monetary and Fiscal Framework for Economic Stability."
American Economic Review 38 (June 1948): 245-264.
Friedman, Milton. A Program for Monetary Stability. New York: Fordham
University Press, 1959.
Gapinski, James H. "Steady Growth, Policy Shocks, and Speed of Adjustment
under Embodiment and Putty-Clay." Journal of Macroeconomics 3 (Spring
1981): 147-176.
Gapinski, James H. Macroeconomic Theory: Statics, Dynamics, and Policy. New
York: McGraw-Hill Book Company, 1982.
Gapinski, James H. "Capital Malleability, Macro Performance, and Policy Effec-
tiveness." Southern Economic Journal 52 (July 1985).
Garbade, Kenneth D. Discretionary Control of Aggregate Economic Activity.
Lexington, MA: D. C. Heath and Company, 1975.
Garbade, Kenneth D. "Economic Stabilization in the Presence of Limited Dis-
cretion." Southern Economic Journal 43 (January 1977): 1243-1259.
Gordon, Nancy M. "Ex Ante and Ex Post Substitutability in Economic Growth."
International Economic Review 14 (June 1973): 497-510.
Howitt, Peter. "Activist Monetary Policy under Rational Expectations." Journal
of Political Economy 89 (April 1981): 249-269.
326

Johansen, Leif. "Substitution versus Fixed Production Coefficients in the Theory

of Economic Growth: A Synthesis." Econometrica 27 (April 1959): 157-176.
Pack, Howard. "Formula Flexibility: A Quantitative Appraisal." In Studies in
Economic Stabilization, pp. 5-40. Edited by Albert Ando, E. Cary Brown, and
Ann F. Friedlaender. Washington: The Brookings Institution, 1968.
Park, Seong Yawng. "Bounded Substitution, Fixed Proportions, and Economic
Growth." Yale Economic Essays 6 (Fall 1966): 343-414.
Phelps, Edmund S. "Substitution, Fixed Proportions, Growth and Distribution."
International Economic Review 4 (September 1963): 265-288.
Phelps, Edmund S., and Taylor, John B. "Stabilizing Powers of Monetary Policy
under Rational Expectations." Journal of Political Economy 85 (February 1977):
163-190.
Phillips, A. W. "Stabilization Policy in a Closed Economy." Economic Journal 64
(June 1954): 29~323.
Phillips, A. W. "Stabilization Policy and the Time-Forms of Lagged Responses."
Economic Journal 67 (June 1957): 265-277.
Salter, W. E. G. Productivity and Technical Change. London: Cambridge Univer-
si ty Press, 1960.
Sargent, Thomas J., and Wallace, Neil. "Rational Expectations and the Theory
of Economic Policy." Journal of Monetary Economics 2 (1976): 169-183.
 327

CHAPTER 18

EXOGENEITY AND CONTROL IN ECONOMETRIC TIME

SERIES MODELLING

1. F. Richard
e.O.R.E, Louvain, Belgium

Introduction

The object of the present note is to discuss some of the problems to

be faced by an econometrician who wishes to simulate the impact
of new control rules on an economy which has been operating in a
different control environment. This discussion heavily relies on
earlier work either on the specific issue of exogeneity, as analysed,
e.g., in Engle et aZ. (1983), or on the more general issue of econo-
metric modelling, as discussed, e.g., in Hendry and Richard (1982,
1983).
Throughout the discussion I will assume that the variables which
are the object of the control rules coincide with the so-called
"exogenous" variables in the model. In practice, however, control
variables and exogenous variables need not coincide with each
other in two important ways:
(i) Some control variables may not have been (statistically)
exogenous over the sample period and have, therefore, to be
included in the list of endogenous variables for the purpose of
estimation at the cost of having to model the corresponding
control rules;
(ii) the new control rules may operate on variables that were not
subject to control over the sample period.
These situations, which correspond to so-called "changes of
regimes", can be handled in the spirit of the discussion in Richard
(1980) and require additional manipulations (conditioning) of the
estimated model before it can be used for simulation of the new
control rules. They will not be discussed in the present paper mainly
for expository purposes.

C. Carraro and D. Sartore (eds). Developments of Control Theory for Economic Analysis
© 1987 Martinus Nijhoff Publishers (Kluwer), Dordrecht
328

The paper is organised as follows: the theory framework is

outlined in Section 1 with special emphasis on the relevant exogeneity
assumptions for estimation and control. An application to a UK
demand for money equation is presented in Section 2.

1. Econometric Models and Exogeneity Assumptions

Two time intervals will be considered. Let Is = [1, T] and

~. = [T + 1, T + P] respectively denote the sample and the control
simulation periods. For notational convenience, the k-dimensional
sequential stochastic process which generates the observation
matrices xi = (x], ... ,x T )' and xi:) = (xT+ I, . . . ,xT+P) is
assumed to be continuous with respect to an appropriate measure
and is, therefore, represented by the density functions
T

D(Xi I X o, Bs) = n D(x/ I X;-I' Bs)

/=1
(1)

P
D(Xi:) I X T , Be) = n D(xT+p I XT+p-l, BJ
p=1
(2)

where es E Bs and ee E Be are finite-dimensional, identifiable and

sufficient parametrisations, Xo is a matrix of known initial con-
ditions and X;' = (X~ X;I'). The links between Bs and Be are central
to the argument and are discussed below.
Let us now partition x; into (y; z;) where Z/ E IRkJ regroups the
control variables. The density functions in (1) and (2) are then
sequentially factored as follows
D(x t I X;-I, e) = D( y/ I ZI' X;-I' e,)D(ztlX;_" ( 2) (3)
with t: 1 -+ T for j = sand t: T + 1 -+ T + P for j = c. The
parameters ell E Bll and e 2l E B2l are implicitly reparametrisations
of the joint processes (1) and (2). Examples of such factorisations
are found, e.g., in Engle et al. (1983).
A number of critical issues are to be discussed:
(1) Insofar as D(z/ J X;-I, e 2s ) describes control rules which are
replaced by new (hypothetical) rules over the simulation period, B2s
is a "nuisance" parameter whose estimation is not required for
conducting the control experiment. The question, therefore, arises
of whether or not inference on the "parameters of interest" e,s can
 329

can be based solely on the "partial" likelihood function

LI«(}ls; XT) IX nT

t=1
D(Yt I zl' Xr-I' (}Is) (4)

i.e., of whether or not Zt is "weakly exogenous" for (}Is' This issue

is discussed further in Section 1.1 below.
(2) Imposing new control rules would definitely modify the
environment under which economic agents have been acting over
the sample period. Yet it is often (implicitly) assumed that their
future behaviour might be unaffected by such a change, i.e., that (}Is
and (}Ic might coincide. This issue has received a lot of attention in
the econometric literature-see, inter alia, Lucas (1976)-and
motivates the concept of "super exogeneity" which is discussed in
Section 1.2 below.
(3) (}2c will typically be assigned by the analyst so as to charac-
terise the new (hypothetical) control environment, though some of
its components might be explicitly related to (}2s and would,
therefore, require estimation. In the spirit of our discussion of weak
exogeneity we shall implicitly assume that (}Is and (}2s can be estimated
independently of each other. (Obviously estimation of (}2s requires
modelling the control process D(zt I Xr-I' (}2s) which would other-
wise be irrelevant to our analysis). Once (}2c is assigned the simulation
experiment can easily be conducted in "closed form" whereby
{(ZT+p I XT+p-l, (}IC>, (YT+p I zT+p' X T+p_ l , ()2c)} are drawn jointly
sequentially. Alternatively, one might wish to preassign a matrix
zF:~, avoiding thereby the need to describe explicitly the control
process, and to conduct the simulation experiment in "open form"
whereby {(YT+pIZF:~, XT+P-I)} are drawn sequentially from the
conditional submodel D(YT+p I zT+p' XT+p-l, (}IJ. The validity of
this exercise requires, however, an additional assumption of "non-
causality" which is discussed in Section 1.3 below.

1.2. Inference and weak exogeneity

According to the factorisation (3), the likelihood function can be
factored into the product of the "partial" likelihood functions
LI «(}Is; X T), as defined in (4), and
T

L 2«(}2s; Xr) = n D(zt I Xr-I' (}2s)'

t=1
(5)
330

The obvious question is that of whether or not the partial likelihood

function LI (Ols; X T ) contains all the sample information of rel-
evance for inference on 0ls. Sufficient conditions for this to occur
are provided by the concept of weak exogeneity introduced in
Richard (1980) and analysed further in Engle et at. (1983)-see also
Florens and Mouchart (1984) for extensions within a Bayesian
framework.

DEFINITION 1: z( is weakly exogenous for Ols (or functions thereof)

over the sample period if and only if Ols and 02s in the factorisation
(3) are "variation free" in the sense that
(6)

Note that Definition 1 makes sense only if 0" 0 1s and 0 2s are

sets of (a priori) "admissible" values for the corresponding par-
ameters so that under condition (6), 0ls and 02s are effectively free
of "cross-restrictions" (Bayesian extensions require prior indepen-
dence between Ols and 02s and are, therefore, unambiguous in that
respect).
°
When z( is weakly exogenous for 1" Maximum Likelihood (ML)
estimation of Ols or functions thereof is achieved by maximisation
of the partial likelihood function (4) and, therefore, the control
process D(z( I X;-I' 02s) need not be modelled-at least at this stage
of the analysis. Furthermore, under the (implicit) assumption that
the joint process {(x t I X;-I)} satisfies the usual assumptions under-
lying asymptotic theory, as described inter alia in White (1984), the
asymptotic covariance matrix of the ML estimator of 0ls can
be derived from the Hessian matrix associated with LI (Ols; X T ),
since the ML estimators of Ols and 02s are then asymptotically
independent.
Parametric tests for weak exogeneity assumptions are available
in the literature for specific classes of models, essentially linear
Gaussian models, see, e.g., Engle (1984), Holly (1983) or Richard
(1984). Note, however, that these tests are extremely unattractive
from a practical point of view since they all require modelling first
the joint process D(x( I X;-I, Os), admittedly within an Instrumental
Variable framework, and then analysing whether or not it can be
factorised as in (3) in such a way that condition (6) holds. Sufficient
overidentifying conditions for the validity of such a factorisation
 331

provide the basis for parametric weak exogeneity tests. Misspeci-

fication of the control process is likely to bias the test and practice
suggests that control rules are often difficult to model adequately if
only because their rational is generally less than fully obvious and,
moreover, subject to changes over the sample period.
The discussion in Section 1.2 will suggest indirect tests for weak
exogeneity assumptions in the form of tests for parameter constancy
across regimes.

1.2. Parameter invariance and super exogeneity

Simulation experiments typically assume that elc == els ' i.e., that
the decision rules of the economic agents under control would not
be affected by the policy change which is envisaged. This strong
assumption is often validated on the grounds of (weak) exogeneity
assumptions though it is essential to realise that an apparent lack
of relationships between els and e2s might be highly specific to
particular (and stable) control environments. This is precisely the
object of Lucas (1976) criticism. In fact weak exogeneity of Zr
"within regimes" does not imply weak exogeneity of Zr "across
regimes" in that a lack of relationships between els and e2s within
any given regime does not contradict the possibility that changes in
(}2s might induce changes in els (e.g., because economic agents
unsatisfied with the new control policy which is enforced upon them
might adjust accordingly their decision strategies). One is,
therefore, naturally lead to distinguishing explicitly between weak
exogeneity ("within regimes") and super exogeneity ("across
regimes") as in Engle et al. (1983).

DEFINITION 2: Zr is super exogenous for els over the sample period

with respect to a class of interventions affecting e2s if and only if: (i)
Zr is weakly exogenous for (}Is over that period; and (ii) (}Is is
(structurally) "invariant", i.e., remains constant under these
interventions.

As rightly pointed out by Deitsler in its discussion of Hendry and

Richard (1983), "super exogeneity can be understood as weak
exogeneity in a suitable parameter space which takes into account
parameter changes". Drawing a distinction between the two con-
cepts helps, however, distinguishing the problem of estimating
332

efficiently the parameters of a conditional submodel such as (4)

from that of using it for exploring the impact of policy changes.
Obviously super exogeneity can only be tested against policy
interventions that have been effectively observed over the sample
period and is, therefore, bound to remain conjectural within the
context of simulating new control policies. The best one can hope
is that superexogeneity across a set of observed policy interventions
might reflect superexogeneity against a wider set of interventions.
Note finally that insofar as the existence of restrictions across Ols
and 02s (some of which the analyst might be unaware of !) is
indicative of a setup whereby agents would naturally adjust Ols in
response to interventions affecting 02s (as, for example, in the con-
text of "rational expectations" model) weak exogeneity-within
regimes-is de facto necessary for super exogeneity-across regimes.
Therefore, the empirical finding that Ols' estimated from a partial
likelihood function such as (4), remains constant over a sample
period where a range of policy changes have occurred, yields as
strong evidence as one can think of in favour of the weak and super
exogeneity assumptions built in the corresponding conditional sub-
model. Econometricians should, therefore, pay special attention to
tests for parameter constancy across observed changes of the
exogenous process. (Note that recent economic history provides us
with plenty of opportunities for constructing such tests!). These
tests yield operational alternative to direct parametric exogeneity
tests and are, furthermore, of paramount relevance within the
context of policy simulation.

1.3. Simulation and non-causality

On top of the exogeneity assumptions whcih are required for
validating the estimation and the use of a conditional submodel
D(YT+p I zT+p' X T + p_ l , Ole) for policy analysis, open form simu-
lation whereby Zj::~ would be preassigned and kept fixed requires
the additional assumption that

D(YT+p I zT+p' XT+p-l, OIJ

== D(YT+p I Zj::~, X T + p_l , 0lc)P: I - P. (7)

Up to technical conditions that are discussed, e.g., in Florens and

Mouchart (1982, 1985) the relevant concept is that of non-causality
 333

in the sense of Granger (1969) which can be adapted as follows to

the context of our analysis.

DEFINITION 3: y does not (Granger-) cause z over the time interval

Ie = [T + 1, T + P] (conditionally on X T ) if and only if
D(ZT+p I XT+p-I' 82J == D(ZT+p I ZJ:~_I' X T, 82e )p: 1 ~ p. (8)

Obviously, though it is of no direct concern to us in the present

discussion, non causality can also be investigated over the sample
period and leads to the concept of strong exogeneity as defined in
Engle et al. (1983).

DEFINITION 4: Z is strongly exogenous for 8 1s (or functions thereof)

over a time interval [1, T] (conditionally on Xo) if and only if: over
that interval (i) Z is weakly exogenous for 81s ; and (ii) y does not
cause z.

It should now be obvious that simulation exercises have to be

conducted in closed form with respect to the exogenous variables
which are caused by the y's in order to be ofpractival relevance (It
might be worth mentioning at this stage that modern economic
policy-making often heavily relies on signals provided by the per-
formance of the economy under control). This requirement raises
no problems when 02e is preassigned for the purpose of the exercise.
It is usually quite demanding when 02e consists of a mix of pre-
assigned parameters together with parameters that are explicitly
related to the control rules that have been operative over the
sample period since it would then necessitate preliminary esti-
mation of the corresponding elements in 02s (even though such an
estimation would not affect inference on Ols under weak exogeneity
assumptions).

1.4. Econometric modelling

In conclusion of this brief discussion we note that the object of
the so-called "econometric modelling" approach is precisely to lead
to such factorisations as (3) whereby one tries to separate relatively
"invariant" parameters characterising a conditional submodel
from "transient" ones characterising a marginal submodel. This is
334

the sort of setup which seems to be appropriate for the analysis and
simulation of policy measures. This approach is often contrasted
with that known as "multivariate time series modelling" whereby it
is the joint process D(x t I X;-I' .) which is modelled directly in
terms of its "reduced" or "final" forms parametrisations. In the
present context the corresponding parameters are functions of both
() Is and ()2s'
Up to the use and testing of theory based prior information, the
choices between these two modelling approaches might not prove
critical in worlds where parameters do not change. In particular, the
time series modelling approach then provides us with a parametri-
sation which is convenient for investigating the Granger causality
structure of a multivariate process as in Sims (1980). In contrast
time series procedures seem to be less operational in worlds where
policy interventions modify ()2s since it is easily shown that such
interventions will typically induce parameter instability in the
reduced form and, thereby, complicate the task of separating tran-
sient parameters from invariant ones (if any!). Simple theoretical
examples supporting this argumentation are found, e.g., in Engle
et al. (1983). Its empirical relevance within a given context is subject
to testing using the sort of data modern economies have managed
to provide us with!

2. An Application to a UK Money Demand Equation

The monetary sector in the United Kindom provides an interesting

example of a major policy switch with the introduction in October
1971 of the measures known as Competition and Credit Control
(CCC). Detailed accounts of the British monetary policy may be
found in Goodhart (1975, 1978). A summary description of the
institutional background and an empirical investigation of the
exogeneity of short term interest rates are found in Lubrano et al.
(1986), hereafter LPR.
In short, before 1971 the primary object of the Bank of England
monetary policy was the so-called short term Bank Rate. Other
short term interest rates moved closely in line with the Bank Rate
through the operation of a clearing bank cartel. In 1971, the clearing
bank cartel was abolished and banks were encouraged to compete
for funds by offering competitive rates. The stated aim of the CCC
 335

regime was to achieve control of the monetary aggregates, with

special emphasis on M3, through portfolio adjustments to changes
in relative interest rates in a competitive market. The success of the
new mechanism of control heavily depended on the stability of the
demand for money funciton. This stability was immediately
•
brought into question as the result of a complete breakdown of the
Bank of England own forewasting equation for M3. Altogether the
CCC regime proved difficult to operate and was abandoned in
August 1981. In the light of our discussion in section 1, these are
precisely the sort of circumstances under which a correct appraisal
of the exogeneity structure of the money market is critical. It is, in
particular, important to discover whether the apparent instability
of the demand for money equation reflects a genuine lack of struc-
tural invariance against (major) policy changes or whether it might
result from estimating its coefficients under invalid weak exogeneity
assumptions. The policy implications of a correct interpretation of
this phenomenon are obvious.
The exogeneity of an interest rate variable has been analysed in
LPR by means of classical and Bayesian bivariate "Instrumental
Variables" procedures whereby the demand for money equation is
estimated jointly with an interest rate equation. The data consist of
35 seasonally unadjusted quarterly observations for each regime.
The reader is referre to LPR for details and results. A few results
are reported in the Appendix below. The key empirical findings in
LPR can be summarized as follows:
(1) The process generating the interest rate variable proves
extremely difficult to model and differs significantly across the two
regimes. Even within regimes the stability of the interest rate equation
can be questioned. In each regime, this equation is expressed solely
in terms of logarithmic differences and has, therefore, no steady
state long term solution. As argued above, these difficulties are
probably inherent to the modelling of many control processes and,
therefore, to the implementation of direct parameter tests for weak
exogeneity. Money does not seem to cause interest rate in the first
regime. In contrast the second regime exhibits a significant feed-
back of money on interest rate.
(2) Whether estimated by Ordinary Least Squares (OLS) under
the assumption that interest rate is weakly exogenous or by Full
Information Maximum Likelihood (FIML) the coefficients of the
demand for money equation are reasonably invariant except for the
336

interest rate short term coefficients. The interest rate impact coef-
ficient even changes sign across the two regimes. In contrast the
steady state long term solution of the demand for money equation
is invariant over the complete sample period though the policy
relevance of this empirical finding is limited since the equation
displays a slow adjustment speed towards its long term solution.
The evidence on the weak exogeneity of the interest rate variable
over the first regime is mostly inconclusive. In contrast, and some-
what counterintuitively, the weak exogeneity of interest rate over
the second regime cannot be rejected. (The weak exogeneity of
money over that regime had been rejected at an early stage of the
analysis). These conclusions are to be appraised in the light of the
difficulties encountered in the modelling of the interest equation but
it is worth noting that the OLS and FIML estimated coefficients
hardly differ from each other with the sole exception of the interest
rate impact coefficient.
(3) All together, it is clear that the predictive failure of the
demand for money equation across 1971 was not the result of an
invalid exogeneity assumption. The banking sector appears to have
modified substantially its behaviour in response to the introduction
of CCC in 1971 and one should account for the fact that it might
again do so if confronted to major changes in the monetary policy
of the Bank of England.

Appendix

All data are quarterly and seasonally adjusted. The following

abbreviations are used:

M: Personal Sector M3
R: Local Authority 3 month deposit rate
Y: Personal disposable income
P: Implied deflator for personal disposable income
U: Unemployment rate
B: Real value of UK official reserves.

The data sources are given in Lubrano et al. (1986) wherefrom the
results which follow are taken.
 337

The interest rate equations, estimated by FIML are

~ In ( 1 + RI ) = - 0.25 ~ In (1 + Rt- 2)
(0.09)

- O.IS ~ In (1 + R'_4) + 0.32 ~2 In PI - 2

(0.10) (0.10)

- 0.02 ~2 In (1 + VI) - 0.04 ~2 In (1 + VI- 3)

(0.00) (0.01)

- 0.02 ~4ln BI - 0.01 ~2 In B'_I + 0.06 ~ In M I _2

(0.00) (0.01) (O.OS)

for the period 1963(i)-197l(iii), and

~ In (1 + R I) = - 0.13 ~4In(l + Rt _ l) - 0.20 In (1 + R'-2)
(0.05) (0.05)

+ 0.29 ~ In (1 + R I _3 ) - 0.42 ~2 In PI - 0.23 ~ In PI- 2

(O.OS) (O.OS) (0.11)

+ 0.70 ~ In PI- 4 + 0.62 ~ In M I _ I - 0.03 ~ In BI

(0.14) (0.11) (0.00)

- 0.01 ~ In BI_4 - 0.06 ~2 In (1 + VI)

(0.00) (0.01)

for the period 1971(iv)-19S0(ii). Number in parentheses are asymp-

totic standard deviations. The money demand equation estimated
by FIML over the complete sample period with a multiplicative
dummy variable for interest rate coefficients (DI = 1 for 1963(i)-
1971(iii) and DI = 0 for 1971(iv)-19S0(ii)) is given by

~ In (~)I = 0.17 ~ In (~) - 0.39 ~ In PI

(0.06) I-I (O.OS)

- O.lS ~21n P I- 2 - 0.04 In (~) + 0.11 ~41n YI

(O.OS) (0.01) 1-5 (0.03)

[ - (0.12)
0.41 DI + 0.23 (1 - DI)] A I (1 R)
+ (0.09) L1 n + ,
- 0.32 DI~2 In (1 + R'_3) - 0.17 In (1 + R'_5)'
(0.09) (0.04)
338

A description of the modelling procedure that lead to these results

together with additional results and details is found in Lubrano
et al. (1985).

References

1. Engle, R. F. (1984): "Wald, Likelihood Ratio and Lagrange Multiplier Tests

in Econometrics", Chapt. Bin: Z. Griliches and M. D. Intriligator (eds.),
Handbook of Econometrics, Vol. 2.
2. Engle, R. F., D. F. Hendry, and J. F. Richard (1983): "Exogeneity", Econo-
metrica 51, 277-304.
3. Florens, 1. P. and M. Mouchart (1982): "A Note on Non-causality", Econo-
metrica 50, 583-592.
4. Florens, 1. P. and M. Mouchart (1958a): "Conditioning in Dynamic
Models", Journal of Time Series Analysis 6,15-34.
5. Florens, 1. P. and M. Mouchart (l985b): "A Linear Theory for Non-causality",
Econometrica 53,157-175.
6. Goodhart, C. A. E. (1973): "Monetary Policy in the United Kingdom",
Chap. 12 in: K. Holbick (ed.), Monetary Policy in Twelve Industrial Countries,
Boston, Mass.; Federal Reserve Bank of Boston.
7. Goodhart, C. A. E. (1978): "Problems in Monetary Management: The U.K.
Experience", in: A. S. Courakis (ed.), Inflation, Depression and Economic
Policy in the West: Lessons from the 1970's, Oxford; Basil Blackwell.
8. Granger, C. N. 1. (1969): "Investigating Causal Relations by Econometric
Models and Cross-spectral Methods", Econometrica 37, 424-438.
9. Hendry, D. F. and 1. F. Richard (1982): "On the Formulation of Empirical
Models in Dynamic Econometric", Journal of Econometrics 20, 3-33.
10. Hendry, D. F. and 1. F. Richard (1983): "The Econometric Analysis of
Economic Time Series", International Statistical Review 51, 111-163.
II. Holly, A. (1983): "A Simple Procedure for Testing Whether a Subset of
Endogenous Variables is Independent of the Disturbance Term in a Structural
Equation", Cahiers de Recherches Economiques No. 836, Universite de
Lausanne, Switzerland.
12. Lubrano, M., R. G. Pierse and 1. F. Richard (1986): "Stabihty of a UK
Money Demand Equation: A Bayesian Approach to Testing Exogeneity", The
Review of' Economic Studies 53, 603-637.
13. Lucas, R. E. (1976): "Econometric Policy Evaluation: A Critique", in:
K. Brunner and A. H. Melter (eds.), The Phillips Curve and Labor Markets,
Carnegie-Rochester Conference Series on Public Policy I, 19-46. North-
Holland, Amsterdam.
14. Richard, 1. F. (1980): "Models with Several Regimes and Changes in Exogen-
eity", The Review of Economic Studies 47, 1-20.
 339

15. Richard, J. F. (1984: "Classical and Bayesian Inference in Incomplete Simul-

taneous Equation Models", in D. F. Hendry and K. F. Wallis (eds.), Econo-
metrics and Quantitative Economics, 61-102, Basil Blackwell, Oxford.
16. Sims, C. A. (1980): "Macroeconomics and Reality", Econometrica 48, 1-48.
 ADVANCED STUDIES IN THEORETICAL AND APPLIED ECONOMETRICS
VOLUME 7

1. Paelinck J.H.P. (ed.): Qualitative and Quantitative Mathematical Economics, 1982.

ISBN 90 247 2623 9.
2. Ancot J.P. (ed.): Analysing the Structure of Economic Models, 1984.
ISBN 90 247 2894 o.
3. Hughes Hallett A.J. (ed.): Applied Decision Analysis and Economic Behaviour,
1984.
ISBN 90 247 2968 8.
4. Sengupta J.K.: Information and Efficiency in Economic Decision, 1985.
ISBN 90 247 3072 4.
5. Artus P. and Guvenen 0., in collaboration with Gagey F. (eds.): International Macro-
economic Modelling for Policy Decisions, 1986.
ISBN 90 247 3201 8.
6. Vilares M.J.: Structural Change in Macroeconomic Models, 1986.
ISBN 90 247 3277 8.
7. Carraro C. and Sartore D. (eds.): Developments of Control Theory for Economic
Analysis. 1987.
ISBN 90 247 3345 6.
8. Broer D.P. (ed.): Neoclassical Theory and Empirical Models of Aggregate Firm
Behaviour. 1987.
ISBN 90-247-3412-6.
9. Italianer A. (ed.): Theory and Practice of International Trade Linkage Models. 1986.
ISBN 90-247-3407-X.