Lecture Notes 9

Minimum Weighted Residual

Methods
The minimum weighted residual method has been introduced in economics
by Judd [1992]. As for the PEA algorithm, it may receive an interpretation in terms of a generalized method of undetermined coefficients. However,
here stops the comparison, as the philosophy of minimum weighted residual
methods differs fundamentally from that of PEA as it will become clear in a
moment. Nevertheless, as for the PEA algorithm, the basic idea of the method
is to approximate either the decision rule in Judds original implementation
or the expectation function of the individuals as proposed by Christiano
and Fisher [2000] by a smooth function, which in most of the cases will
combine orthogonal polynomials. The parameters of the function are then
revealed by imposing identifying restrictions dictated by economic theory.

9.1
9.1.1

The Basic idea

Stating the problem

The minimum weighted residual method may be implemented to solve a large

set of models that admit the following general representation
Et F (yt+1 , xt+1 , yt , xt , t+1 ) = 0
1

(9.1)

where F : Rny Rnx Rny Rnx Rne Rny +nx describes the model. t

is the set of innovations of the structural shocks that hit the economy. The
solution of the model is a set of decision rules for the control variables
yt = g(xt ; )
that define the next period state variables as
xt+1 h(xt , yt , t+1 ) = h(xt , g(xt ; ), t+1 )
such that the model can be rewritten
Et R(xt , t+1 ; g, ) = 0

(9.2)

The idea of the minimum weighted residual method is to replace the true
decision rule by a parametric approximation function, (xt; ), of the current
state variables xt and a vector of parameters . Therefore, in its original
implementation the minimum weighted residual method differs from the PEA
algorithm in that we are not seeking directly an expectation function but a
decision rule.1 The problem is then to find a vector of parameters such that
when the agents use the so defined rule of thumb, Et F (xt , t+1 ; g, ) can be
made as small as possible. But, What do we mean by small? In fact we
want to find a vector of parameters b such that
kEt R(xt , t+1 ; g, {}ni=0 )k = 0

which corresponds to
Z

Et R(xt , t+1 ; g, )i (x)dx = 0 for i = 0, . . . , n

(9.3)

where (x) is a weighting function, which operationalizes the notion of small.

1
Note however that we will present a variant of the method that attempts to reveal the
expectation instead.

9.1.2

Implementation

Choosing an implementation of the minimum weighting residual method basically amounts to make 3 choices
1. the choice of a family of approximating functions,
2. the choice of the weighting function,
3. the choice of the method to approximate the integral involved by (i) the
rational expectation and (ii) the identifying restriction (9.3).
Choosing a family of approximating functions:

The choice of a family

of approximating function defines the class of minimum weighting residual

method we use. It is commonly admitted to range the method in two classes:
Finite element methods,
Spectral methods.
The Finite element methods can be viewed as a piecewise application of
the minimum weighted residual method. This approach assumes that the
decision rule can be accurately represented by a collection of small elements of
finite dimensions called as Finite Elements. The original decision rule is then
considered as an assemblage of these elements. These elements are connected
to each other at joints called as nodes (or nodal points) to form the entire
decision rule. Therefore, the first step involved in solving the model using finite
elements is to divide the state space into disjoint intervals the elements.
Then the method will essentially amount to fit low order polynomials or spline
functions on each subintervals. Once an approximation is obtained over each
subinterval, all approximations are then pieced to form the decision rule over
the whole domain of the state space we
xx
i1

xi xi1
xi+1 x
(xi ) =
x
x

i+1 i
0
3

consider. Typical finite elements are

if x [xi1 ; xi ]
if x [xi ; xi+1 ]
otherwise.

we therefore recognize spline functions. McGrattan [1996] and McGrattan

[1999] discuss and present the implementation of finite element methods in
various economic models.
The spectral methods instead make use of higher order polynomials usually orthogonal polynomials but do not divide the domain of approximation
into subintervals. The approximation is therefore conducted over the whole
interval, which requires the decision rule to be continuous and differentiable.
Therefore, in this case, the typical approximation will look like
yt =

p
X

i i (xt )

i=0

where {i (.)}ni=0 is a family of orthogonal polynomials.

Choosing a weighting function:

The choice of the weighting function is

extremely important in that it will define the method we will use. Traditionally, we may define 3 broad classes of method, depending on the weighting
function we use:
1. The Least square method sets
i (x) =

R(xt , t+1 ; g, )
i

such that the problem amounts to find b such that

Z
b
Argmin
Et R(xt , t+1 ; )2 dxt

2. The collocation method uses the Dirac function as a weighting function,

such that
(x) = (x xi ) =

1 if x = xi
0 if x 6= xi

Therefore, by selecting this weighting function we set the residuals to

zero on the nodes x1 , . . . , xn the collocation points such that the
4

approximation is exact on these nodes. Note that nothing is imposed

outside these nodes. When orthogonal polynomials are selected, these
nodes are the roots of the polynomial of the highest order and the method
is called orthogonal collocation. Note that in this approximation method,
we only need n + 1 collocation points to identify the n + 1 elements of
, otherwise the system is overidentified.
3. The Galerkin method uses the basis functions as weighting function. For
instance, if the approximation is a linear combination of Chebychev polynomials of order 0 to n, the weighting functions are given by Chebychev
polynomials of order 0 to n:
i (x) = t (x)
In other words, this amounts to impose the orthogonality of the residuals
vis a vis the basis functions we are using. We are actually making use of
the fact that a continuous function is identically zero if it is orthogonal
to each element of a complete set of functions.
Choosing an integration method:

We shall consider two problems here:

1. The problem of the rational expectation: In this case, everything depends on the form of the process for the shocks. If we are to use Markov
chains, the integration problem is highly simplified as it only involves
a discrete sum. If we use continuous support for the shocks, the choice
of an integration method is dictated by the type of distribution we are
assuming. In most of the cases, we will use gaussian shocks such that
we will rely on GaussHermite quadrature methods described in lecture
notes # 4.
2. The problem of the inner product (equation (9.3)): In this case everything depends on the approximation method. If we are to use a
collocation method, no integration method is to be used as collocation
5

amounts to impose that the residuals are zero at each node, such that
(9.3) reduces to
Et R(xi , t+1 ; g, {}ni=0 ) = 0 for i = 1, . . . , n
When Least square methods are used, it is often the case that Legendre
quadrature will do the job. When a Galerkin method is selected, this will
often be the right choice too. However, when we are to use Chebychev
polynomials, a Chebychev quadrature is in order provided each weighting
function is defined as2
Ti ((x))
i (x) = p
1 (x)2

where (x) is a linear function mapping the domain of x into [-1;1].

9.2

Practical implementation

In the sequel, we will essentially discuss the collocation and Galerkin implementations of a spectral method using Chebychev polynomials as they seem
to be the most popular implementation of this method.3

9.2.1

The Collocation method

In this section, we present the collocation method in its simplest form. We will
start by presenting the general algorithm, when Chebychev polynomials are
used, and then present as an example the stochastic optimal growth model.
For the moment, let us assume that we want to solve a rational expectation
model that writes as
Et R(xt , t+1 ; g, ) = 0
2

Remember that Chebychev quadrature computes integrals of the form

Z 1
1
F (x)
dx
1 x2
1

3
The least square method can be straightforwardly implemented minimizing the implicit
objective function we discussed.

and we want to find an approximating function for the decision rule g(xt ) over
the domain [x, x]. Assume that we take as approximating function
(xt , )

n
X

i Ti ((xt ))

i=0

where Ti (.) is the Chebychev polynomial of order i = 0, . . . , n. The problem

is then to find the vector , such that
Et R(xi , t+1 ; , ) = 0 for i = 0, . . . , n
The algorithm is then as follows4
1. Choose an order of approximation n, compute the n + 1 roots of the
Chebychev polynomial of order n + 1 as

(2i 1)
zi = cos
for i = 1, . . . , n + 1
2(n + 1)
and formulate an initial guess for .
2. Compute xi as
xi = x + (zi + 1)

(x x)
for i = 1, . . . , n + 1
2

to map [-1;1] into [x; x].

3. Compute
Et R(xi , t+1 ; g, ) for i = 1, . . . , n + 1
4. if it is close enough to zero then stop and form
(xt , )

n
X

i Ti ((xt ))

i=0

else update and go back to 3.

4

We will discuss the one dimensional case. However, the multidimensional case will
be illustrated in an example. You are also referred to lecture notes #3 which presented
multidimensional approximation technics

The updating scheme for will typically be given by a nonlinear solver of the
type described in lecture notes #5. The computation of the integral involved
by the rational expectation will depend on the process we assumed for the
shocks. In order to better understand it, let us discuss the implementation of
the collocation method on the stochastic optimal growth model both in the
case of a Markov chain, and an AR(1) process.
The stochastic OGM (Markov chain)

: We will actually consider a

OGM in which the production function writes

yt = exp(at )kt
where at can take on two values [a, a] with a transition matrix

p
1p
=
1p
p
The Euler equation is given by

1
c
=0
t Et ct+1 exp(at+1 )kt+1 + 1

and the law of motion of capital

kt+1 = exp(at )kt ct + (1 )kt

But, as a matter of fact, by taking the definition of the Markov chain into
account, the model may be restated as the following system

ct (a) p ct+1 (a) exp(a)kt+1 (a)1 + 1

= 0
(1 p) ct+1 (a) exp(a)kt+1 (a)1 + 1

ct (a) (1 p) ct+1 (a) exp(a)kt+1 (a)1 + 1

= 0
p ct+1 (a) exp(a)kt+1 (a)1 + 1

kt+1 (a) exp(a)kt + ct (a) (1 )kt = 0

kt+1 (a) exp(a)kt + ct (a) (1 )kt = 0
8

where the notation ct (a) (resp. ct (a)) stands for the fact that the consumption
decision is contingent on the realization of the shock, a (resp. a), likewise for
kt+1 (a) (resp. kt+1 (a)).
It therefore appears that we actually have only two decision rules to compute, one for ct (a) and one for ct (a), each taking the form5

ct (a) ' (kt , (a)) exp

ct (a) ' (kt , (a)) exp

n
X
j=0

n
X
j=0

j (a)Tj ((log(kt )))

(9.4)

j (a)Tj ((log(kt )))

(9.5)

where the notation (a) accounts for the fact that the form of the decision
rule may differ depending on the realization of the shock, a. We may also
select different order of approximation for the two rules, but in order to keep
things simple we chose to impose the same order. The algorithm then works
as follows.
1. Choose an order of approximation n, compute the n + 1 roots of the
Chebychev polynomial of order n + 1 as
zi = cos

(2i 1)
2(n + 1)

for i = 1, . . . , n + 1

and formulate an initial guess for (a) and (a).

2. Compute ki as

log(k) log(k)
ki = exp log(k) + (zi + 1)
2

for i = 1, . . . , n + 1

to map [-1;1] into [k; k].

5

The exponential form was imposed in order to guarantee positivity of consumption.

3. Compute

n
X
j (a)Tj ((log(ki )))
ct (a) ' (kt , (a)) exp
j=0

n
X
ct (a) ' (kt , (a)) exp
j (a)Tj ((log(ki )))
j=0

and

kt+1 (ki , a) = exp(a)ki (ki , (a)) + (1 )ki

kt+1 (ki , a) = exp(a)ki (ki , (a)) + (1 )ki

at each node ki , i = 1, . . . , n + 1.
4. Then compute the possible levels of future consumption
at
a
a
a
a

at+1
a
a
a
a

Consumption
P
(kt+1 (ki , a), (a)) nj=0 j (a)Tj ((kt+1 (ki , a)))
P
(kt+1 (ki , a), (a)) nj=0 j (a)Tj ((kt+1 (ki , a)))
P
(kt+1 (ki , a), (a)) nj=0 j (a)Tj ((kt+1 (ki , a)))
P
(kt+1 (ki , a), (a)) nj=0 j (a)Tj ((kt+1 (ki , a)))

5. Evaluate the residuals

R(ki , a; ) = (ki , (a)) p(ki , a, a) (1 p)(ki , a, a)
R(ki , a; ) = (ki , (a)) (1 p)(ki , a, a) p(ki , a, a)

where

(ki , at , at+1 ) (kt+1 (ki , at ), (at+1 )) exp(at+1 )kt+1 (ki , at )1 + 1

for all i = 1, . . . , n.

6. if all residuals are close enough to zero then stop, else update and go
back to 3.
10

From a practical point of view, the last step is performed using a Newton
algorithm. Initial conditions can be obtained from a linear approximation of
the model (see the matlab codes in directory growth/collocmc).
Matlab Code: Collocation Method (OGM, Main Code)
clear all
global kmin ksup XX kt;
global nstate nbk ncoef XX XT PI;
nbk
nodes
nstate
ncoef

=
=
=
=

4;
nbk+1;
2;
nbk+1;

%
%
%
%

% need to set some parameters

% globally

Degree of polynomials
# of Nodes
# of possible states for technology shock
# of coefficients

delta
= 0.1;
% depreciation rate
beta
= 0.95;
% discount factor
alpha
= 0.3;
% capital elasticity
sigma
= 1.5;
% parameter of utility
%
% Steady state
%
ysk
=(1-beta*(1-delta))/(alpha*beta);
ksy
= 1/ysk;
ys
= ksy^(alpha/(1-alpha));
ks
= ys^(1/alpha);
is
= delta*ks;
cs
= ys-is;
%
% Markov Chain: technology shock (Tauchen Hussey)
%
rho
= 0.8;
se
= 0.2;
ma
= 0;
[agrid,wmat] = hernodes(nstate);
agrid
= agrid*sqrt(2)*se;
PI
= transprob(agrid,wmat,0,rho,se);
at
= agrid+ma;
%
% grid for the capital stock
%
kmin
= log(0.1);
ksup
= log(6);
rk
= rcheb(nodes);
% roots
kt
= exp(itransfo(rk,kmin,ksup)); % grid
XX
= cheb(rk,[0:nbk]);
% Polynomials
%
% Initial Conditions

11

%
a0=repmat([-0.2 0.65 0.04 0 0],nstate,1);
a0=a0(:);
%
% Main loop
%
param
= [alpha beta delta sigma];
th
= fcsolve(residuals,a0,[],param,at);
th
= reshape(th,ncoef,nstate);

Matlab Code: Residuals Function (OGM)

function res=residuals(theta,param,at);
global nbk kmin ksup XX kt PI nstate;
alpha
beta
delta
sigma
lt
theta

=
=
=
=
=
=

param(1);
param(2);
param(3);
param(4);
length(theta);
reshape(theta,lt/nstate,nstate);

RHS=[];
LHS=[];
for i=1:nstate;
ct
= exp(XX*theta(:,i));
% C(t)
k1
= exp(at(i))*kt.^alpha+(1-delta)*kt-ct;
% k(t+1)
rk1 = transfo(log(k1),kmin,ksup);
% k(t+1) in kmin,ksup
xk1 = cheb(rk1,[0:nbk]);
% polynomials
%
% Euler equation for all states
%
aux = 0;
for j=1:nstate;
c1
= exp(xk1*theta(:,j));
% c(t+1)
%
% Cumulates the left hand side of the Euler equation for all states
%
resid = beta*(alpha*exp(at(j))*k1.^(alpha-1)+1-delta).*c1.^(-sigma);
aux
= aux+PI(i,j)*resid;
end;
RHS = [RHS -sigma*log(ct)];
LHS = [LHS log(aux)];
end;
res=LHS-RHS;
res=res(:);

% Note that residuals are taken in logs (easier)

12

For the parameterization we use in the codes, and using 4th order Chebychev polynomials, we obtain the decision rules reported in table 9.1. Figure
Table 9.1: Decision rules (OGM, Collocation)
0
-0.2807
-0.4876

1
0.7407
0.8406

2
0.0521
0.0546

3
0.0034
0.0009

4
-0.0006
-0.0006

(9.1) reports the decision rules of consumption and next period capital stock.
As can be seen, they are (hopefully) alike those obtained from the value iteration method (see lecture notes #7).
Figure 9.1: Decision rules (OGM, Collocation)
6

Next period capital stock

1.5

4
3

0.5

1
0
0

Consumption

0
0

The AR(1) process : In this case, we assume that the technology shock is
modelled as
at+1 = at + t+1
where ; N (0, 2 ). The Euler equation is given by

1
c
=0
t Et ct+1 exp(at+1 )kt+1 + 1
13

and the law of motion of capital

kt+1 = exp(at )kt ct + (1 )kt
But, as a matter of fact, by taking the definition of AR(1) process, the model
may be restated as the following system
2
Z

1
1

ct p
ct+1 exp(at + t+1 )kt+1 + 1 exp t+1
dt+1 = 0
2
22
2
kt+1 exp(at )kt + ct (1 )kt = 0

We therefore have to compute the integral in the Euler equation. However,

since we have assumed a gaussian distribution for the shock we may use the
GaussHermite quadrature. Therefore, we make the following change of vari
ables: z = / 2 , such that the Euler equation rewrites as
Z

1
ct+1 exp(at + z 2 )kt+1
ct
+ 1 exp z 2 dz = 0

which makes more explicit the need for a GaussHermite quadrature. We then

just have to compute the nodes and weight for the quadrature, such that the
Euler equation rewrites
c
t

1 X
1

+1 =0
j ct+1 exp(at + zj 2 )kt+1

j=1

We now have to formulate a guess for the decision rule. Note that since we
have not specified particular values for the technology shock, we can introduce
it into the approximating rule, such that we will take

nk X
na
X
ct ' (kt , at , ) exp
jk ja Tjk ((log(kt )))Tja ((at ))
jk =0 ja =0

Note that since we use collocation, the numbers of nodes should be equal to
the total numbers of coefficients. It may then be much easier to work with
tensor basis rather than complete polynomials in this case. The algorithm
then works as follows.
14

1. Choose an order of approximation nk and na for each dimension, compute the nk + 1 and na + 1 roots of the Chebychev polynomial of order
nk + 1 and na + 1 as
zki
zai

(2i 1)
for i = 1, . . . , nk + 1
= cos
2(n + 1)

(2i 1)
= cos
for i = 1, . . . , na + 1
2(n + 1)

and formulate an initial guess for .

2. Compute ki as

log(k) log(k)
i
for i = 1, . . . , nk + 1
ki = exp log(k) + (zk + 1)
2
to map [-1;1] into [k; k] and
ai = log(a) + (zai + 1)

aa
for i = 1, . . . , na + 1
2

to map [-1;1] into [a; a] and

3. Compute

ct ' (ki , aj , ) = exp

nk X
na
X

jk =0 ja =0

and

jk ja Tjk ((log(ki )))Tja ((aj ))

kt+1 (ki , aj ) = exp(aj )ki (ki , aj , ) + (1 )ki

at each node (ki , aj ), i = 1, . . . , nk + 1 and j = 1, . . . , na + 1.
4. Then, for each state (ki , aj ), compute the possible levels of future consumption needed to compute the integral

ct+1 ' (kt+1 (ki , aj ), aj + z` 2 , )

nk X
na
X

jk ja Tjk ((log(kt+1 (ki , aj ))))Tja ((aj + z` 2 ))

' exp
jk =0 ja =0

for ` = 1, . . . , q.

15

5. Evaluate the residuals

R(ki , aj ; ) = (ki , aj , )

q
X

` (ki , aj , z` ; )

`=1

where

(ki , aj , z` ; ) (kt+1 (ki , aj ), aj +z` 2 ) exp(aj + z` )kt+1 (ki , aj )1 + 1

for all i = 1, . . . , nk + 1 and i = 1, . . . , na + 1.

6. if all residuals are close enough to zero then stop, else update and go
back to 3.
From a practical point of view, the last step is performed using a Newton
algorithm. Initial conditions can be obtained from a linear approximation of
the model (see the matlab codes in directory growth/collocg).
Matlab Code: Collocation Method (Stochastic OGM, Main Code)
clear all
global nbk kmin ksup XK kt;
global nba amin asup XA at;
global nstate nodea nodek wmat wgrid;
nbk=4;
nba=2;
nodek=nbk+1;
nodea=nba+1;
nstate=12;

%
%
%
%
%

% parameters that will be common to

% several subfunctions
%

Degree of polynomials (capital)

Degree of polynomials (technology shock)
# of nodes for capital stock
# of nodes for technology shock
# of nodes for Gauss-Hermite quadrature

delta
= 0.1;
% depreciation rate
beta
= 0.95;
% discount factor
alpha
= 0.3;
% capital elasticity
sigma
= 1.5;
% parameter of utility
rho
= 0.8;
% persistence of AR(1)
se
= 0.2;
% standard deviation of innovation
ma
= 0;
% average of the process
%
% deterministic steady state
%
ysk
=(1-beta*(1-delta))/(alpha*beta);
ksy
= 1/ysk;
ys
= ksy^(alpha/(1-alpha));
ks
= ys^(1/alpha);

16

is
= delta*ks;
cs
= ys-is;
ab
= 0;
%
% grid for the technology shocks
%
[wgrid,wmat]=hernodes(nstate);
% weights and nodes for quadrature
wgrid
= wgrid*sqrt(2)*se;
amin
= (ma+wgrid(nstate));
asup
= (ma+wgrid(1));
ra
= rcheb(nodea);
% roots
at
= itransfo(ra,amin,asup);
% grid
XA
= cheb(ra,[0:nba]);
% Polynomials
%
% grid for the capital stock
%
kmin
= log(.1);
ksup
= log(6);
rk
= rcheb(nodek);
% roots
kt
= exp(itransfo(rk,kmin,ksup)); % grid
XK
= cheb(rk,[0:nbk]);
% Polynomials
%
% Initial Conditions
%
a0
= [
-0.23759592487257
0.60814488103911
0.03677400318790
0.69025680170443
-0.21654209984197
0.00551243342828
0.03499834613714
-0.00341171507904
-0.00449139656933
0.00085302605779
0.00285737302122
-0.00002348542016
-0.00011606672164
-0.00003323351559
0.00018045618825];
a0
= a0(:);
%
% main part!
%
param
= [alpha beta delta sigma ma rho];
th
= fcsolve(residuals,a0,[],param);

17

Matlab Code: Residuals Function (Stochastic OGM)

function res=residuals(theta,param);
global nbk kmin ksup XK kt;
global nba amin asup XA at;
global nstate nodea nodek wmat wgrid;
alpha
= param(1);
beta
= param(2);
delta
= param(3);
sigma
= param(4);
ma
= param(5);
rho = param(6);
RHS=[];
LHS=[];
XX=[];
for i
= 1:nodek;
for j=1:nodea;
XX0 = makepoly(XA(j,:),XK(i,:));
ct = exp(XX0*theta);
% consumption
XX = [XX;XX0];
k1 = exp(at(j))*kt(i).^alpha+(1-delta)*kt(i)-ct;
% k(t+1)
rk1 = transfo(log(k1),kmin,ksup);
% log(k1) to [-1;1]
xk1 = cheb(rk1,[0:nbk]);
% Cheb. polynomials
a1 = rho*at(j)+(1-rho)*ma+wgrid;
% next period shock
ra1 = transfo(a1,amin,asup);
% a(t+1) to [-1;1]
XA1 = cheb(ra1,[0:nba]);
% Cheb. Polynomials
XX1 = makepoly(XA1,xk1);
c1 = exp(XX1*theta);
% c(t+1)
%
% computes the integral
%
tmp = beta*(alpha*exp(a1)*k1.^(alpha-1)+1-delta).*c1.^(-sigma);
aux = wmat*tmp/sqrt(pi);
RHS = [RHS;log(aux)];
LHS = [LHS;log(ct.^(-sigma))];
end
end;
res = LHS-RHS;
res = res(:);

For the parameterization we use in the codes, and using 4th order Chebychev polynomials for the capital stock and second order polynomials for the
technology shock, we obtain the decision rules reported in table 9.2.
18

Table 9.2: Decision rule (Consumption, OGM, Collocation, AR(1))

T0 (at )
T1 (at )
T2 (at )

T0 (kt )
-0.3609
0.6474
0.0338

T1 (kt )
0.7992
-0.2512
0.0103

T2 (kt )
0.0487
-0.0096
-0.0058

T3 (kt )
0.0019
0.0048
-0.0004

T4 (kt )
-0.0003
0.0001
0.0004

Figure (9.2) reports the decision rules of consumption and next period
capital stock.

Figure 9.2: Decision rules (OGM, Collocation, AR(1))

Next period capital stock

Consumption

10

3
2

5
1
0
10

5
k

9.2.2

0 2
t

0
10

5
k

0 2

The Galerkin method

In this section, we present the Galerkin method. As in the previous section, we

will start by presenting the general algorithm, when Chebychev polynomials
are used, and then present as an example the stochastic optimal growth model.
Assume that we want to solve a rational expectation model that writes as

Et R(xt , t+1 ; g, ) = 0
19

We therefore want to find an approximating function for the decision rule g(x t )
over the domain [x, x]. Assume that we take as approximating function
(xt , )

n
X

i Ti ((xt ))

i=0

where Ti (.) is the Chebychev polynomial of order i = 0, . . . , n. The problem

is then to find the vector , such that
Z 1
Et R(x, t+1 ; , {i }ni=0 )i ((x))d(x) = 0 for i = 0, . . . , n
1

where

Ti ((x))
i (x) = p
1 (x)2

In fact, the integral may be evaluated using a GaussChebychev quadrature,

such that, as the weights of the GaussChebychev quadrature are constant,
the problem amounts to solve
m
X
j=1

Et R(xj , t+1 ; , {i }ni=0 )Ti ((xj )) = 0 for i = 0, . . . , n

which rewrites
T ((x))R(x, t+1 ; , ) = 0
where

T0 ((x1 )) . . . T0 ((xm ))

..
..
..
T (x) =

.
.
.
Tn ((x1 )) . . . Tn ((xm ))

The algorithm then works as follows6

1. Choose an order of approximation n, compute the m > n roots of the

Chebychev polynomial of order m > n as

(2i 1)
for i = 1, . . . , m
zi = cos
2m
and formulate an initial guess for .
6

We will discuss the one dimensional case. However, the multidimensional case will
be illustrated in an example. You are also referred to lecture notes #3 which presented
multidimensional approximation technics

20

2. Compute the matrix T (x) as

T0 (z1 ) . . . T0 (zm )

..
..
..
T (x) =

.
.
.
Tn (z1 ) . . . Tn (zm )

3. Compute xi as
xi = x + (zi + 1)

(x x)
for i = 1, . . . , n + 1
2

to map [-1;1] into [x; x].

4. Compute
Et R(xi , t+1 ; g, ) for i = 1, . . . , n + 1
and evaluate
T ((x))R(x, t+1 ; , ) = 0
5. if it is close enough to zero then stop and form
(xt , )

n
X

i Ti ((xt ))

i=0

else update and go back to 3.

As before, the updated in steps 4 and 5 will typically be obtained from
a nonlinear solver of the type described in lecture notes #5. Likewise, the
computation of the integral involved by the rational expectation will depend
on the process we assumed for the shocks. In order to better understand
it, and as in the previous section, we will discuss the implementation of the
method on the stochastic optimal growth model both in the case of a Markov
chain, and an AR(1) process.
The stochastic OGM (Markov chain) : We consider the same OGM as
in the previous section, such that the production function writes
yt = exp(at )kt
21

where at can take on two values [a, a] with a transition matrix

p
1p
=
1p
p

The Euler equation is given by

1
=0
c
t Et ct+1 exp(at+1 )kt+1 + 1
and the law of motion of capital

kt+1 = exp(at )kt ct + (1 )kt

In fact, the algorithm is exactly the same up to some slight differences that I
will just mention:
1. Rather than evaluating the residuals on n + 1 nodes, we evaluate them
on m > n nodes.
2. The functions we want to set to zero are not the residuals anymore, as
we have to compute the integral. Therefore, the vector solves
T (z)R(k, a; ) = 0
T (z)R(k, a; ) = 0
The matlab codes associated to the Galerkin method are basically the same
as the ones for the collocation method, up to some minor differences. We
were using a greater number of nodes (20) such that in the main code, the
line nodes=nbk; now reads nodes=20;, and the residuals are projected on the
chebychev matrix, such that the codes for the residuals is slightly modified,
as the line res=LHS-RHS; is now replaced by res=XX*(LHS-RHS);. Applying
these modifications, we end up with the decision rules reported in table 9.3.
These decision rules do not differ much from those obtained from collocation
at the 4 digits precision, they actually differ if we report numbers with a
greater precision. This actually illustrates the remarkable capability of the
collocation method to capture the main features of the decision rule in this
particular example (this is actually related to the smoothness of the decision
rules). This would not be the case for other more sophisticated models.
22

Table 9.3: Decision rules (OGM, Galerkin)

0
-0.2807
-0.4876

1
0.7407
0.8406

2
0.0521
0.0546

3
0.0034
0.0009

4
-0.0005
-0.0006

The AR(1) process : Like in the case of a Markov chain, the matlab codes
associated to the Galerkin method are basically the same as the ones for the
collocation method, up to some minor differences. We were using a greater
number of nodes (20 for capital and 10 for the shock) such that in the main
code, the line nodek=nbk+1; and nodea=nba+1; now read nodek=20; and
nodea=10;, and the residuals are projected on the chebychev matrix, such
that the codes for the residuals is slightly modified, as the line res=LHS-RHS;
is now replaced by res=XX*(LHS-RHS);. Further, we used complete basis
rather than tensor basis (see the makepoly.m function)
Table 9.4: Decision rule parameters

T0 (a) -0.359970
T1 (a) 0.009890
T2 (a) 0.001764

T0 (k)
0.647039
0.048143
0.004738

T1 (k)
0.032769
-0.009799
-0.000247

23

T2 (k)
0.799127
-0.005171

T3 (k)
-0.252257

T4 (k)

9.3

The PEA modification

The PEA modification to minimum weighted residual method has recently

been proposed by Christiano and Fisher [2000]. The basic idea to this modification is basically to apply exactly the same technic as the one we just
described to the expectation function rather than the decision rule. One very
attractive feature of this modification is that it will enable us to handle, as
easily as in the case of PEA, problems in which there is a possibly binding
constraint. Since, this modification is only a change in the function we are
approximating, I will not elaborate any further on the method which is
the exact application of the previously described methods to the expectation
function but will rather provide detailed examples of application, focusing
on the Galerkin method.

9.3.1

The optimal growth model

Let us first recall the type of problem we have in hand. We are about to solve
the set of equations

1
t Et t+1 exp(at+1 )kt+1
+1
= 0
c
t t = 0

kt+1 exp(at )kt + ct (1 )kt = 0

at+1 at t+1 = 0
Our problem will therefore be to get an approximation for the expectation
function:

1
Et t+1 exp(at+1 )kt+1
+1

In this problem, we will deal with a continuous AR(1) process,7 such we have
2 state variables: kt and at , such that (.) should be a function of both kt and
7

See directory growth/peagalmc for the matlab code with a Markov chain.

24

at . Like in the standard Galerkin procedure, we will use a guess of the form

(kt , at , ) exp

jk =0,...,nk
ja =0,...,na
ja +jk 6max(nk ,na )

jk ja Tjk ((log(kt )))Tja ((at ))

therefore imposing a complete basis of polynomials. Then the algorithm is

pretty much the same as the one for standard Galerkin procedure with an
AR(1) process.
1. Choose an order of approximation nk and na for each dimension, compute the mk > nk (nk + 1 if you use collocation) and ma > na (na + 1
for collocation) roots of the Chebychev polynomial of order mk and ma
as
zki
zai

(2i 1)
= cos
for i = 1, . . . , mk
2(n + 1)

(2i 1)
= cos
for i = 1, . . . , ma
2(n + 1)

and formulate an initial guess for .

2. Compute ki as

log(k) log(k)
i
ki = exp log(k) + (zk + 1)
for i = 1, . . . , mk
2
to map [-1;1] into [k; k] and
ai = log(a) + (zai + 1)

aa
for i = 1, . . . , ma
2

to map [-1;1] into [a; a] and

3. Compute the approximating expectation function at each node (ki , aj ),
i = 1, . . . , mk and j = 1, . . . , ma :
(ki , aj , )
25

4. Deduce the level of consumption, which, from the Euler equation, is

given by
ct (ki , aj , ) = (ki , aj , )1/
and the capital stock
kt+1 (ki , aj ) = exp(aj )ki ct (ki , aj , ) + (1 )ki
5. Then, for each state (ki , aj ), compute the possible levels of future consumption needed to compute the integral by computing the expectation
function in t + 1

(kt+1 (ki , aj ), aj + z` 2 , )
and the consumption in t + 1

ct (kt+1 (ki , aj ), aj + z` 2 , ) = (kt+1 (ki , aj ), aj + z` 2 , )1/

for ` = 1, . . . , q
6. Evaluate the residuals ((ki , aj , ) minus the expectation function)
q
X
` (ki , aj , z` ; )
R(ki , aj ; ) = (ki , aj , )

`=1

where

(ki , aj , z` ; ) (kt+1 (ki , aj ), aj +z` 2 ) exp(aj + z` )kt+1 (ki , aj )1 + 1

for all i = 1, . . . , mk and i = 1, . . . , ma .

7. if all residuals (projected on polynomials in the case of standard Galerkin)

the are close enough to zero then stop, else update and go back to 3.
The implementation of this version of the problem is extremely close to the
standard Galerkin (collocation) procedure, such that you are left to the matlab
code (in /growth/peagalg and /growth/peagalmc directories).
26

At this point, it is not clear why the PEA modification to the minimum
weighted residual method could be interesting, as it does not differ so much
from its standard implementation. However, it has the attractive feature of
being able to deal with binding constraints extremely easily. As an example
of implementation, we shall now study the stochastic optimal growth model
with an irreversibility in investment decision.8

9.3.2

Irreversible investment

We now consider a variation to the previous model, in the sense that we restrict
gross investment to be positive in each and every period:
it > 0 kt+1 > (1 )kt

(9.6)

This assumption amounts to assume that there does not exist a second hand
market for capital. In such a case the problem of the central planner is to
determined consumption and capital accumulation, such that utility is maximum:
max E0

{ct ,kt+1 }t=0

s.t.

X
t=0

c1
1
t
1

kt+1 = exp(at )kt ct + (1 )kt

and
kt+1 > (1 )kt
Forming the Lagrangean associated to the previous problem, we have
"

c1 1

ct+ + (1 )kt+ kt+ +1

Lt = E t
t+
+ t+ exp(at+ )kt+
1
=0
#
+ t+ (kt+1 (1 )kt )

8
You will also find in the matlab codes a version of the problem of a consumer facing a
borrowing constraint in the directory borrow.

27

which leads to the following set of first order conditions

c
= t
t

(9.7)

1
t t = Et t+1 exp(at+1 )kt+1
+ 1 t+1 (1 )

kt+1 = exp(at )kt ct + (1 )kt

t (kt+1 (1 )kt )

(9.8)
(9.9)
(9.10)

The main difference with the previous example is that now the central planner
faces a constraint that may be binding in each and every period. Therefore,
this complicates a little bit the algorithm, and we have to find a rule for both
the expectation function
Et [t+1 ]
where

1
t+1 t+1 exp(at+1 )kt+1
+ 1 t+1 (1 )

and t . We then proceed as we did in the PEA algorithm, that is we will

compute an approximation to the expectation function and to the Lagrangean
multiplier, checking at each computed node whether the constraint is binding
or not. In order to keep things simple, we will discuss the case where the shock
follows a Markov Chain,9 Therefore, the rule of thumb will be such that

nk
X
Et [t+1 ] ' (kt , a` ; ) exp
(a` )Tj ((log(ki ))) for each a` , ` = 1, . . . , na
j=0

where na denotes the number of possible states for a.

The algorithm is very close to the one we used in the case of PEA and
closely follows ideas suggested in Marcet and Lorenzoni [1999] and Christiano
and Fisher [2000]. In the case of a Markov chain and for the Galerkin procedure, it can be sketched as follows.
9
You are left to study the matlab code located in /irrev/peagalg for the continuous
AR(1) case.

28

1. Choose an order of approximation nk and a number of possible states na

for the shock, compute the mk > nk roots of the Chebychev polynomial
of order mk as
zi = cos

(2i 1)
2(n + 1)

for i = 1, . . . , mk

Compute a parameters of the Markov chain (Transition matrix = [` ]

and formulate an initial guess for .
2. Compute the matrix T (z) as

T0 (z1 ) . . . T0 (zm )

..
..
..
T (z) =

.
.
.
Tn (z1 ) . . . Tn (zm )

3. Compute ki as

log(k) log(k)
for i = 1, . . . , nk + 1
ki = exp log(k) + (zi + 1)
2
to map [-1;1] into [k; k].
4. Compute the approximating expectation function at each node ki , i =
1, . . . , mk , and for each possible state a` , ` = 1, . . . , na :

nk
X
(ki , a` ; ) = exp
(a` )Tj ((log(ki )))
j=0

5. Deduce the level of consumption, which, from the Euler equation, is

given by
ct (ki , a` , ) = (ki , a` , )1/
the level of investment
it (ki , a` ; ) = exp(a` )ki ct (ki , a` , )
the level of next period capital stock
kt+1 (ki , a` ; ) = it (ki , a` ) + (1 )ki
29

6. Check whether the constraint is binding for each ki , i = 1, . . . , nk and

each a` , ` = 1, . . . , na
If the constraint is not binding then keep computed values for
ct (ki , a` , ) and kt+1 (ki , a` ; ) and set

t (ki , a` ; ) = 0
If it is binding then set
kt+1 (ki , a` ; ) = (1 )ki

ct (ki , a` , ) = exp(a` )ki

t (ki , a` ; ) = ct (ki , a` ; )
t (ki , a` ; ) = t (ki , a` ; ) (ki , a` ; )
7. Then, for each state (ki , a` ), i = 1, . . . , nk and ` = 1, . . . , na , compute
the possible levels of future consumption without taking the constraint
into account
ct+1 (kt+1 (ki , a` ; ), a ; ) = (kt+1 (ki , a` ; ), a ; )1/
for = 1, . . . , na , the level of next period investment
it (kt+1 (ki , a` ; ), a ; ) = exp(a )kt+1 (ki , a` ; ) ct (kt+1 (ki , a` ; ), a ; )
8. Check whether the constraint is binding at each position
If the constraint is not binding then keep computed values for
ct+1 (kt+1 (ki , a` ; ), a ; ) and set

t+1 (kt+1 (ki , a` ; ), a ; ) = 0

If it is binding then set
ct+1 (kt+1 (ki , a` ; ), a ; ) = exp(a )kt+1 (ki , a` ; )
t+1 (kt+1 (ki , a` ; ), a ; ) = ct+1 (kt+1 (ki , a` ; ), a ; )
t+1 (kt+1 (ki , a` ; ), a ; ) = t+1 (kt+1 (ki , a` ; ), a ; ) (kt+1 (ki , a` ; ), a ; )
30

9. Evaluate the residuals ((ki , a` ; ) minus the expectation function)

R(ki , a` ; ) = (ki , a` ; )

na
X

` (ki , a` , a ; )

=1

where
(ki , a` , a ; ) ct+1 (kt+1 (ki , a` ; ), a ; )

exp(a )kt+1 (ki , a` ; )1 + 1

t+1 (kt+1 (ki , a` ; ), a ; )(1 )

for all i = 1, . . . , mk and ` = 1, . . . , na .

10. Compute the inner product involved by the Galerkin procedure
T ((k))R(k, a; )
11. if all inner products are close enough to zero then stop, else update
and go back to 3.
Note that, as in the standard PEA algorithm, we treat t as an technical
variable which is just used to compute the residuals. We therefore do not need
to compute explicitly its interpolating function.
Matlab Code: PEA Galerkin (Irreversible Investment)
clear all
global kmin ksup XX kt;
% parameters that will be common to
global nstate nbk ncoef XX XT PI; % several subfunctions
nbk=10;
nodes=30;
nstate=5;
ncoef=nbk+1;
delta
beta
alpha
sigma
ysk
ksy

%
%
%
%

Degree of polynomials
# of Nodes
# of possible states for the shock
# of coefficients

= 0.1;
% depreciation rate
= 0.95;
% discount factor
= 0.3;
% capital elasticity
= 1.5;
% CRRA Utility
=(1-beta*(1-delta))/(alpha*beta);
= 1/ysk;

31

ys
= ksy^(alpha/(1-alpha));
ks
= ys^(1/alpha);
is
= delta*ks;
cs
= ys-is;
%
% Markov Chain: (Tauchen-Hussey 1991) technology shock
%
rho
= 0.8;
% persistence
se
= 0.075;
% volatility
ma
= 0;
[agrid,wmat]=hernodes(nstate);
agrid
= agrid*sqrt(2)*se;
PI
= transprob(agrid,wmat,0,rho,se);
at
= agrid+ma;
%
% grid for the capital stock
%
kmin
= log(1);
ksup
= log(7);
rk
= rcheb(nodes);
% roots
kt
= itransfo(rk,kmin,ksup);
% grid
XX
= cheb(rk,[0:nbk]);
% Polynomials
%
% Initial Conditions
%
a0=[
-0.32518704
-0.22798091
-0.15021991
-0.12387075
-0.61535041
-0.64879442
-0.70031744
-0.84056683
-0.02892469
-0.03369096
-0.05570433
-0.15063263
-0.00565378
-0.00941495
-0.02618621
-0.09526710
-0.00296230
-0.00547789
-0.01692339
-0.06102182
-0.00145337
-0.00274142
-0.00988611
-0.03461623
-0.00065225
-0.00115070
-0.00536657
-0.01776343
-0.00025245
-0.00038302
-0.00291504
-0.00838555
-0.00004845
-0.00001360
-0.00162292
-0.00365588
0.00001150
0.00013398
-0.00088759
-0.00124601
-0.00003248
0.00013595
-0.00055706
-0.00013116
];
%
% Solves the problem effectively
%
param
= [alpha beta delta sigma];
th
= fcsolve(residuals,a0,[],param,at);
th
= reshape(th,ncoef,nstate);
%
% Evaluate the approximation
%
lt
= length(th);
nb
= 100;

32

-0.23608269
-1.23998020
-0.45685868
-0.32031660
-0.21319594
-0.13635186
-0.09027963
-0.06431768
-0.04689460
-0.03294968
-0.02163670

kt
= [kmin:(ksup-kmin)/(nb-1):ksup];
rk
= transfo(kt,kmin,ksup);
XX
= cheb(rk(:),[0:nbk]);
kt
= exp(kt);
kp
= zeros(nb,nstate);
it
= zeros(nb,nstate);
ct
= zeros(nb,nstate);
mu
= zeros(nb,nstate);
for i=1:nstate;
Upsilon = exp(XX*th(:,i));
y
= exp(at(i))*kt.^alpha;
it(:,i) = max(y-Upsilon.^(-1/sigma),0);
ct(:,i) = y-it(:,i);
kp(:,i) = it(:,i)+(1-delta)*kt;
mu(:,i) = ct(:,i).^(-sigma)-Ephi;
end;
Matlab Code: Residuals Function (Irreversible Investment)
function res=residuals(theta,param,at);
global nbk kmin ksup XX kt PI nstate;
alpha
= param(1);
beta
= param(2);
delta
= param(3);
sigma
= param(4);
lt
= length(theta);
theta
= reshape(theta,lt/nstate,nstate);
RHS=[];LHS=[];
% lhs and rhs of Euler equation
for i=1:nstate;
Upsilon = exp(XX*theta(:,i));
% evaluate expectation
y
= exp(at(i))*exp(kt).^alpha;
% Output
iv
= max(y-Ephi.^(-1/sigma),0);
% takes care of the constraint
k1
= (1-delta)*exp(kt)+iv;
% next period capital stock
rk1
= transfo(log(k1),kmin,ksup);
% maps it into [-1;1]
xk1
= cheb(rk1,[0:nbk]);
% Computes the polynomials
aux
= 0;
for j=1:nstate;
Upsilon1 = exp(xk1*theta(:,j));
% used for the consumption
y1
= exp(at(j))*k1.^alpha;
% next period output
mu1
= max(y1.^(-sigma)-Ephi1,0); % mu>0 <=> lambda=(c=y)^(-sigma)
tmp
= beta*((alpha*y1./k1+1-delta).*Upsilon1-mu1*(1-delta));
aux
= aux+PI(i,j)*tmp;
end;
RHS = [RHS (aux)];
LHS = [LHS (Ephi)];
end;
res = XX*(LHS-RHS);
% Galerkins projection
res = res(:);

33

Once again an important issue in this problem is that of initial conditions.

The initial conditions reported in the matlab codes are actually extremely
close to the solution. They were obtained using a homotopy type of approach.
In praxis, we started from the deterministic optimal growth model without any
irreversibility, and get a decision rule using an approximation of order 10. We
then assign the decision rule we obtained in this case to a stochastic optimal
growth model, with 5 possible states but low volatility (0.01), and obtained a
new solution. Then we introduced the irreversibility constrained, and starting
from the set of initial conditions we had, we get an approximate solution for the
optimal growth model with irreversibility but with a low volatility. Starting
from this new solution, we solved the same problem with higher volatility
(0.02) and applied again the process until we reached the desired volatility
(se=0.075). At a first sight, this may seem quite a long and useless process,
but it saves a lot of time as it is much easier to go from a problem for which
we have a solution to the problem we actually want to solve than trying to
guess initial conditions.
Table 9.5: Decision rules parameters (Irreversible investment, PEAGalerkin)

T0 (log(kt ))
T1 (log(kt ))
T2 (log(kt ))
T3 (log(kt ))
T4 (log(kt ))
T5 (log(kt ))
T6 (log(kt ))
T7 (log(kt ))
T8 (log(kt ))
T9 (log(kt ))
T1 0(log(kt ))

a1
-0.271633
-0.617972
-0.020999
-0.000673
-0.000403
-0.000251
-0.000073
-0.000031
-0.000042
-0.000028
-0.000008

a2
-0.198195
-0.639585
-0.021617
-0.001298
-0.000913
-0.000670
-0.000312
-0.000103
-0.000104
-0.000093
-0.000032

a3
-0.131885
-0.662475
-0.024964
-0.004093
-0.002900
-0.002002
-0.001192
-0.000519
-0.000363
-0.000359
-0.000182

a4
-0.076585
-0.701972
-0.042518
-0.017748
-0.012222
-0.007493
-0.004215
-0.001893
-0.000923
-0.000890
-0.000557

a5
-0.044720
-0.786584
-0.094447
-0.053895
-0.032597
-0.015666
-0.005548
-0.000922
0.000202
-0.000995
-0.001279

Table 9.5 and Figure 9.3 report the approximate decision rules for an economy where = 0.3, = 0.95, = 0.1, = 1.5, the persistence of the shock
34

is = 0.8 and the volatility = 0.075. As can be seen, only low values
for the shock make the constraint binds. This is illustrated by the fact that
the Lagrange multiplier becomes positive, and investment becomes negative.
Therefore, the next period capital stock is given by what is left by the depreciation and output is fully consumed.
Figure 9.3: Decision rules (Irreversible investment, PEAGalerkin)
7

Next period capital stock

Consumption

6
5

1.5

4
3

2
1
0

4
k

0.5
0

4
k

Investment

0.5

0.3

0.4

Lagrange multiplier

0.2

0.3
0.1
0.2
0

0.1
0
0

4
k

0.1
0

4
k

Note that up to now, we have not address the important issue of the
evaluation of the rule we obtained using any of the method we have presented
so far. This is the object of the next lecture notes.

35

36

Bibliography
Christiano, L.J. and J.D.M. Fisher, Algorithms for solving dynamic models
with occasionally binding constraints, Journal of Economic Dynamics
and Control, 2000, 24 (8), 11791232.
Judd, K., Projection Methods for Solving Aggregate Growth Models, Journal
of Economic Theory, 1992, 58, 410452.
Marcet, A. and G. Lorenzoni, The Parameterized Expectations Approach:
Some Practical Issues, in M. Marimon and A. Scott, editors, Computational Methods for the Study of Dynamic Economies, Oxford: Oxford
University Press, 1999.
McGrattan, E., Solving the Stochastic Growth Model with a Finite Element
Method, Journal of Economic Dynamics and Control, 1996, 20, 1942.
, Application of Weighted Residual Methods to Dynamic Economic Models, in M. Marimon and A. Scott, editors, Computational Methods for the
Study of Dynamic Economies, Oxford: Oxford University Press, 1999.

37

38

Contents
9 Minimum Weighted Residual Methods
9.1

9.2

9.3

The Basic idea . . . . . . . . . . . . . . . . . . . . . . . . . . .

9.1.1

Stating the problem . . . . . . . . . . . . . . . . . . . .

9.1.2

Implementation . . . . . . . . . . . . . . . . . . . . . . .

Practical implementation . . . . . . . . . . . . . . . . . . . . .

9.2.1

The Collocation method . . . . . . . . . . . . . . . . . .

9.2.2

The Galerkin method . . . . . . . . . . . . . . . . . . .

19

The PEA modification . . . . . . . . . . . . . . . . . . . . . . .

24

9.3.1

The optimal growth model . . . . . . . . . . . . . . . .

24

9.3.2

Irreversible investment . . . . . . . . . . . . . . . . . . .

27

39

40

List of Figures
9.1

Decision rules (OGM, Collocation) . . . . . . . . . . . . . . . .

13

9.2

Decision rules (OGM, Collocation, AR(1)) . . . . . . . . . . . .

19

9.3

Decision rules (Irreversible investment, PEAGalerkin) . . . . .

35

41

42

List of Tables
9.1

Decision rules (OGM, Collocation) . . . . . . . . . . . . . . . .

13

9.2

Decision rule (Consumption, OGM, Collocation, AR(1)) . . . .

19

9.3

Decision rules (OGM, Galerkin) . . . . . . . . . . . . . . . . . .

23

9.4

Decision rule parameters . . . . . . . . . . . . . . . . . . . . . .

23

9.5

Decision rules parameters (Irreversible investment, PEAGalerkin) 34

43