Lecture Notes 2

Towards nonlinear methods

In the previous lectures, we dealt with linear economies, for which there exist
straightforward methods to solve the involved expectational difference equations. However, most of the models we encounter in economics are fundamentally characterized by nonlinear dynamical features. We therefore need
methods to solve such models. The aim of this chapter is to introduce you
to such methods, by first pointing out the possible drawbacks of simple linearization a method commonly used in the literature. We will present four
methods which may be used to solve RE models:
1. Perturbation methods (see Judd (1998), Judd and Gaspard (1997), Collard and Juillard (2001a), Collard and Juillard (2001b) or ?)), which
essentially amount to take higherorder Taylor approximation of the
model;
2. Value iteration (see Christiano (1990), Tauchen and Hussey (1991)),
which may be simply thought of as finding a fixed point on an operator;
3. Parameterized Expectations Algorithm (PEA) (see Marcet (1988), Den Haan
and Marcet (1990) or Marcet and Lorenzoni (1999) among others), which
may be thought of as a generalization of the method of undetermined
coefficients to the higher order relying on simulations;
4. Minimum weighted residual methods (see Judd (1992), McGrattan (1996),
1

McGrattan (1999)), which, as PEA, may be thought of as a generalization of the method of undetermined coefficients to the higher order
but exploits some orthogonality conditions rather than relying on simulations;
Each method is illustrated by an economic example, which is intended to show
you the potential and simplicity of the method. However, before going to such
methods, we shall now see why linearizing many not always be a good idea.
The big question is then
What are we missing?

2.1

Risk and the Certainty Equivalence Hypothesis

Taking a either linear or loglinear approximation to the decision rules of an

economic model, is actually equivalent to taking a quadratic approximation
to the optimization problem that lies behind the optimal behavior of agents.
In so doing we encounter the socalled Certainty Equivalence property. In
order to understand the certainty equivalence property, let us consider the
following problem. Let us consider that x is a random variable with probability
density g(x) and let y be a variable decided by a decision maker (this may be
consumption for an household, investment or labor for a firm, the price for a
monopolist. . . ). This decision maker has an objective function u(y, x) which
is concave and twice continuously differentiable. Its y plan is then chosen by
solving
max E(u(y, x))
y

u(y, x)g(x)dx

The first order condition1 for choosing y is then given by (applying Leibniz
rule)

u(y, x)g(x)dx = 0

u(y, x)g(x)dx = 0
y

(2.1)

1
Note that since u(.) is concave and g(.) is positive, this condition is necessary and
sufficient.

Now, let us assume that u(y, x) is a second order Taylor expansion of

another objective, such that
H
u(y, x) = (y x)J + (y x)
2

y
x

where H is a negativedefinite (2 2) matrix, such that

1
y
Jy
hxx hxy
+ (y x)
u(y, x) = (y x)
x
Jx
hyx hyy
2

1
= Jy y + Jx x +
hxx x2 + (hxy + hyx )xy + hyy y 2
2

(2.2)

In such a case, (2.1) rewrites

Z h
i
2Jy + (hxy + hyx )Ex
x
Jy + (hxy + hyx ) + hyy y g(x)dx = 0 y =
2
2hyy
Now let us consider a situation where the objective of the decision maker
is

max u(y, E(x)) u y, xg(x)dx

y

Note that in this case, we are not maximizing the expected value of the problem
but the value, taking into account the expected value of x. Given the functional
form (2.2), the first order condition is now
Jy + (hxy + hyx )

2Jy + (hxy + hyx )Ex

Ex
+ hyy y = 0 y =
2
2hyy

which is exactly the same as before. In other words, we have for the
quadratic formulation
Argmax E(u(y, x)) = Argmax u(y, E(x))
y

This is what is usually called the Certainty equivalence principle: risk does not
matter in decision making, the only thing that matters is the average value
of the random variable x, not its variability. But this is usually not a general
result. Let us consider, for example, the case of Burnsides [1998] asset pricing
model.
3

2.1.1

An assetpricing example

This model is a standard asset pricing model for which (i) the marginal intertemporal rate of substitution is an exponential function of the rate of growth
of consumption and (ii) the endowment is a Gaussian exogenous process. As
shown by Burnside (1998), this setting permits to obtain a closed form solution
to the problem. We consider a frictionless pure exchange economy a
` la Mehra
and Prescott (1985) and Rietz (1988) with a single household and a unique
perishable consumption good produced by a single tree. The household can
hold equity shares to transfer wealth from one period to another. The problem of a single agent is then to choose consumption and equity holdings to
maximize her expected discounted stream of utility, given by
Et

X
=0

ct+
with (, 0) (0, 1]

(2.3)

subject to the budget constraint

pt et+1 + ct = (pt + dt )et

(2.4)

where (0, 1) is the agents subjective discount factor, ct is households

consumption of a single perishable good at date t, pt denotes the price of the
equity in period t and et is the households equity holdings in period t. Finally,
dt is the trees dividend in period t. Dividends are assumed to grow at rate xt
such that :
dt = exp(xt )dt1

(2.5)

where xt , the rate of growth of dividends, is assumed to be a Gaussian stationary AR(1) process
xt = (1 )x + xt1 + t

(2.6)

where is i.i.d. N (0, 2 ) with || < 1. Market clearing requires that et = 1

so that ct = dt in equilibrium. Like in Burnside (1998), let yt denote the
4

pricedividend ratio, yt = pt /dt . Then, condition for the households problem

can be shown to rewrite as
yt = Et [exp(xt+1 )(1 + yt+1 )]

(2.7)

Burnside (1998) shows that the above equation admits an exact solution of
the form2
yt =

i exp [ai + bi (xt x)]

(2.8)

i=1

where

and

2 2
2(1 i ) 2 (1 2i )
ai = xi +
+
i
2(1 )2
1
1 2

(1 i )
1
As can be seen from the definition of ai , the volatility of the shock, directly
bi =

enters the decision rule,, therefore Burnsides [1998] model does not make the
certainty equivalent hypothesis: risk matters for asset holdings decisions.
What happens then, if we now obtain a solution relying on a first order
Taylor approximation of the model?
First of all let us determine the deterministic steady state of the economy:
y ? = exp(x? )(1 + y ? )
x? = x? + (1 )x
such that we get
exp(x? )
1 exp(x? )
= x

y? =
x?

(2.9)
(2.10)

The first order Taylor expansion of the Euler equation yields

ybt = exp(x? )Et (b

yt+1 ) + exp(x? )Et (b
xt+1 )

See appendix ?? for a detailed exposition of the solution.

(2.11)

We actually recognize the simplest RE model we have been dealing with in

chapter 2 (yt = aEt yt+1 +bxt ) such that we may use a undetermined coefficient
approach and guess a decision rule of the form
ybt = b
xt

Plugging the guess in (2.11), we get

b
xt = exp(x? )Et (b
xt+1 ) + exp(x? )Et (b
xt+1 )
taking expectations and identifying, we obtain
=

exp(x? )
(1 exp(x? ))(1 exp(x? ))

such that the approximate decision rule may be written as

yt = y ? + (xt x? )

We are now endowed to compute the approximation error we make using

linear approximation. As the model admits a closedform solution, the accuracy of the approximation method can be directly checked against the true
decision rule. This is undertaken relying on the two following criteria

N
1 X yt yet
E1 = 100
yt
N
t=1

and

yt yet

= 100 max
yt

where yt denotes the true solution to pricedividend ratio and yet is the ap-

proximation of the true solution by the method under study. E1 represents

the average relative error an agent makes using the approximation rather than

the true solution, while E is the maximal relative error made using the approximation rather than the true solution. These criteria are evaluated over
the interval xt [x x , x + x ] where is selected such that we explore
6

99.99% of the distribution of x. Table 2.1 reports E1 and E for the different cases. Our benchmark experiment amounts to considering the Mehra and
Prescotts [1985] parameterization of the asset pricing model. We therefore
set the mean of the rate of growth of dividend to x = 0.0179, its persistence to
= 0.139 and the volatility of the innovations to = 0.0348. These values
are consistent with the properties of consumption growth in annual data from
1889 to 1979. was set to -1.5, the value widely used in the literature, and
to 0.95, which is standard for annual frequency. We then investigate the implications of changes in these parameters in terms of accuracy. In particular, we
study the implications of larger and lower impatience, higher volatility, larger
curvature of the utility function and more persistence in the rate of growth of
dividends.
Table 2.1: Accuracy check

E1
E
E1
E

Benchmark
1.43
1.46
=0.5
0.29
0.29

=0.5
0.24
0.26
=0.001
0.01
0.03

=0.99
2.92
2.94
=0.1
11.70
11.72

=-10
23.53
24.47
=0
1.57
1.57

=-5
8.57
8.85
=0.5
5.52
6.76

=0
0.50
0.51
=0.9
37.50
118.94

Note: The series defining the true solution was truncated after 800 terms,
as no significant improvement was found adding additional terms at the
machine accuracy. When exploring variations in , the overall volatility
of the rate of growth of dividends was maintained to its benchmark level.

At a first glance at table 2.1, it appears that linear approximation can

only accommodate situations where the economy does not experiment high
volatility or large persistence of the growth of dividends, or where the utility
of individuals does not exhibit much curvature. This is for instance the case in
the Mehra and Prescotts [1985] parameterization (benchmark) case as both
the average and maximal approximation error lie around 1.5%. But, as is
nowadays wellknown, increases along one of the aforementioned dimension
yields lower accuracy of the linear approximation. For instance, increasing the
7

volatility of the innovations of the rate of growth of dividends to =0.1 yields

approximation errors of almost 12% both in average and at the maximum, thus
indicating that the approximation performs particularly badly in this case.
This is even worse when the persistence of the exogenous process increases,
as =0.9 yields an average approximation error of about 40% and a maximal
approximation of about 120%. This is also true for increases in the curvature
of the utility function (see row 4 and 5 of table 2.1).
Figure 2.1 sheds light on these results. It reports the exact solution to the
problem (grey line) and the linear approximation of the true solution (thin
black line). We consider a rather extreme situation where = 5, = 0.5
and the volatility of the shock is preserved. As can be seen from figure 2.1,
Figure 2.1: Decision rule
20

Exact (high )
Exact (low )
Linear App.

18
16
14
12

Certainty equivalence bias

10
8
6
4
2
0
0.15

0.1

0.05

xt

0.05

0.1

0.15

Note: This graph was obtained for = 5 and = 0.5.

using a linear approximation induces two types of error:

8

0.2

1. in terms of curvature,
2. in terms of level.
The first type of error is obvious, as the linear approximation is not intended
(as it cannot) to capture any curvature. The second type of error is related
to the fact that we are using a approximation about the deterministic steady
state. Therefore, the latter source of error is related to the risk component. In
fact, this may be understood in light of the ai terms in the exact solution which
include the volatility of the innovations. In order to be sure that this
error is related to this component, we also report the exact solution when we
cut the overall volatility by 25% (thick dashed line). As can be seen the level
error tends to diminish dramatically, which indicates that the risk component
plays a major role in this as the average error is cut by 20% then (5% as
far as the maximal error is concerned). Hence, this suggests that the linear
approximation may only be accurate for low enough variability and curvature,
which prevents its use for studying structural breaks.

2.2

NonLinear Dynamics and Asymmetries

We now consider another situation where the linear approximation may perform poorly. This situation is related to the existence of strong asymmetries
in decision rules or strong asymmetries in the objective functions the economic
agents have to optimize. In order to illustrate this situation, let us take the
problem of a firm that has to decide on employment and which faces asymmetric adjustment costs. Asymmetric adjustment costs may be justified on
institutional grounds. We may argue for example that there exist laws in the
economy that render firings more costly than hirings.
We consider the case of a firm that has to decide on its level of employment.
The firm is infinitely lived and produces a good relying on a decreasing returns
to scale technology that essentially uses labor another way to think of it
would be to assume that physical capital is a fixedfactor. This technology is
9

represented by the constant returnstoscale production function 3

Yt = Ant with A > 0.
Using labor incurs two sources of cost
1. The standard payment for labor services: wt nt where wt is the real
wage, which positive sequence {wt }
t=0 is taken as given by the firm and
is assumed to evolve as
wt = wt1 + (1 )w + t
with t ; U[w ,w ] and4 w < (1 )w.
2. A cost of adjusting labor, C(t ) which satisfies
C(0) = 0, C 0 (0) = 0, C 00 () > 0
but that displays some asymmetries. An example of such a function is
depicted in figure (2.2)
Labor is then determined by maximizing the expected intertemporal profit
s

X
1
(f0 nt+s wt+s nt+s C(t+s ))
max Et
1+r
{n , }
=0
s=0

subject to
nt = t + nt1

(2.12)

which yields the two first order conditions

t = C 0 (t )
t

1
= A wt +
Et t+1
1+r

(2.13)
(2.14)

This will enable us to obtain an analytical solution to the problem

This assumption is imposed in order to guaranty the positivity of the real wage. Indeed
assume the economy experiences the
P worst shock in each and every period, then we would
have wt+j = j wt + (1 j )w jk=0 j which in the limit yields limj wt+j = w
/(1 ). The positivity condition then corresponds to what we impose in the main text.
4

10

Figure 2.2: Asymmetric adjustment costs

60

50

40

30

20

10

0
2

1.5

0.5

0
t

0.5

1.5

where t is the lagrange multiplier associated to (2.12) The second equation

may be simply solved iterating to yield
i

X
1
Et (f0 wt+i )
t =
1+r
i=0

Note that
j

Et (wt+j ) = wt + (1 )w +

j1
X
i=0

therefore
t =

d
= j wt + (1 j )w
2

w
1+r
(1 + r)A

wt
r
1+r 1+r

Then, t is given by
t = (wt ) C

0 1

(1 + r)A
w
1+r

wt
r
1+r 1+r

and we have
nt = (wt ) + nt1
11

Figure 2.3: Decision rule for t

t

wt

Since C 0 (.) may exhibit strong asymmetries, the decision rule may be extremely
nonlinear too to yield a decision rule of the form we depict in figure (2.3). As
can be seen from the graph, the linear approximation would do a very poor
job, as any departure from the steady state level (? = 0) would create a large
error. In other words, and as should have been expected, strong nonlinearities
forbid the use of linear approximations.
Beyond this point that may appear quite peculiar, since such important
nonlinearities are barely encountered after all, there exists a large class of
models for which linear approximation would do a bad job: models with binding constraints that we now investigate.

2.3

Dealing with binding constraints

In this section, we will provide you with an example where linear approximation should not be used because the decision rules are not differentiable. This
12

is the case when the agent faces possibly binding constraints. To illustrate it
we will develop a model of a consumer who is constrained on its borrowing in
the financial market.
We consider the case of a household who determines her consumption/saving
plans in order to maximize her lifetime expected utility, which is characterized
by the function:

Et

=0

c1
t+ 1
1

with (0, 1) (1, )

(2.15)

where 0 < < 1 is a constant discount factor, c denotes the consumption

bundle. In each and every period, the household enters the period with a level
of asset at carried from the previous period, for which it receives an constant
real interest rate r. It also receives an endowment t , which may either be
thought of as something totally extrinsic to the economy or as wages. But
this is taken to be exogenous as we are not interested by its determination.
Therefore, this will be an exogenous stochastic process of the form
log(t ) = log(t1 ) + (1 ) log() + t

(2.16)

with t ; N (0, 2 ). These revenus are then used to consume and purchase
assets on the financial market. Therefore, the budget constraint in t is given
by
at+1 = (1 + r)at + t ct

(2.17)

In addition, the household is submitted to a borrowing contraint that states

that she cannot borrow
at+1 > 0

The first order conditions to this model may be obtained forming the
5
Et (.) denotes mathematical conditional expectations. Expectations are conditional on
information available at the beginning of period t.

13

Lagrangean to the system:

!

X
c1
t+ 1

+ t+ ((1 + r)at+ + t+ ct+ at+ +1 ) + t+ at+ +1

Lt = E t

1
=0

where t and t respectively denote the Lagrangean multipliers associated to

the budget and the borrowing constraints. The first order conditions associated to the system are then
c
= t
t

(2.18)

t = t + (1 + r)Et t+1

(2.19)

t ((1 + r)at + t ct at+1 ) = 0

(2.20)

t at+1 = 0

(2.21)

t > 0

(2.22)

t > 0

(2.23)

together with

manipulating the system, we see that consumption is given by

c
= min ((1 + r)at + t ) , (1 + r)Et c
t
t+1

The decision rule of consumption is not differentiable in the point where assets
holdings are not sufficient to guaranty a positive net position on asset holdings:
((1 + r)at + t ) = (1 + r)Et c
t+1
Just to give you an idea of this phenomenon, we reported in figure 2.4 the
consumption decision rule for two different values of t as a function of cash
onhand which is given by (1+r)at +t .6 This nondifferentiability implies
obviously that linear approximation cannot be useful in this case, as they are
not defined in the neighborhood of the point that makes the household switch
from the unconstrained to the constrained regime. Nevertheless, if we are
6

We will see later on how these decision rules where computed.

14

Figure 2.4: Consumption decision rule

140

120

Consumption (c )

100

80

60

40

20

0
0

50

100

150
200
cashonhand ((1+r)att)

15

250

300

to consider an economy with tiny shocks and where the steady state lies in
the unconstrained regime, the linear approximation may be sufficient as the
decision rule is particularly smooth in this region because of consumption
smoothing).
We therefore need to investigate alternative methods, which however require some preliminaries

16

Bibliography
Burnside, C., Solving asset pricing models with Gaussian shocks, Journal of
Economic Dynamics and Control, 1998, 22, 329340.
Christiano, L., Solving the Stochastic Growth Model by Linear Quadratic
Approximation and by Value Function Iteration, Journal of Business and
Economic Statistics, 1990, 8 (1), 2326.
Collard, F. and M. Juillard, Accuracy of Stochastic Perturbation Methods:
The Case of Asset Pricing Models, Journal of Economic Dynamics and
Control, 2001, 25 (6/7), 979999.
and

, Accuracy of Stochastic Perturbation Methods: The Case of

Asset Pricing Models, Computational Economics, 2001, 17 (2/3), 125

139.
Haan, W. J. Den and M. Marcet, Solving the Stochastic Growth Model by
parametrizing Expectations, Journal of Business and Economic statistics,
1990, 8, 3134.
Judd, K., Projection Methods for Solving Aggregate Growth Models, Journal
of Economic Theory, 1992, 58, 410452.
Judd, K.L., Numerical methods in economics, Cambridge, Massachussets:
MIT Press, 1998.
and J. Gaspard, Solving Large-Scale Rational-Expectations Models,
Macroeconomic Dynamics, 1997, 1, 4575.
17

Marcet, A., Solving Nonlinear Stochastic Models by Parametrizing Expectations, mimeo, CarnegieMellon University 1988.
and G. Lorenzoni, The Parameterized Expectations Approach: Some
Practical Issues, in R. Marimon and A. Scott, editors, Computational
methods for the study of dynamic economies, Oxford: Oxford University
Press, 1999, pp. 143171.
McGrattan, E.R., Solving the Stochastic Growth Models with a Finite Element Method, Journal of Economic Dynamics and Control, 1996, 20,
1942.
, Application of Weighted Residual Methods to Dynamic Economic Models, in R. Marimon and A. Scott, editors, Computational methods for
the study of dynamic economies, Oxford: Oxford University Press, 1999,
pp. 114142.
Mehra, R. and E.C. Prescott, The equity premium: a puzzle, Journal of
Monetary Economics, 1985, 15, 145161.
Rietz, T.A., The equity risk premium: a solution, Journal of Monetary Economics, 1988, 22, 117131.
Tauchen, G. and R. Hussey, Quadrature Based Methods for Obtaining Approximate Solutions to Nonlinear Asset Pricing Models, Econometrica,
1991, 59 (2), 371396.

18

Index
Asymmetries, 9
Binding constraints, 12
Certainty equivalence, 2
Nonlinearities, 9

19

20

Contents
2 Towards nonlinear methods
2.1

Risk and the Certainty Equivalence Hypothesis . . . . . . . . .

2.1.1

An assetpricing example . . . . . . . . . . . . . . . . .

2.2

NonLinear Dynamics and Asymmetries . . . . . . . . . . . . .

2.3

Dealing with binding constraints . . . . . . . . . . . . . . . . .

12

21

22

List of Figures
2.1

Decision rule . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.2

Asymmetric adjustment costs . . . . . . . . . . . . . . . . . . .

11

2.3

Decision rule for t

. . . . . . . . . . . . . . . . . . . . . . . .

12

2.4

Consumption decision rule . . . . . . . . . . . . . . . . . . . . .

15

23

24

List of Tables
2.1

Accuracy check . . . . . . . . . . . . . . . . . . . . . . . . . . .

25