Perturbation PDF

Perturbation Methods
Jess Fernndez-Villaverde
University of Pennsylvania
June 13, 2010
Jess Fernndez-Villaverde (PENN)
June 13, 2010
1 / 92
Introduction
Remember that we want to solve a functional equations of the form:
H (d ) = 0
for an unknown decision rule d.
Perturbation solves the problem by specifying:
d n (x, ) =
i (x
x0 )i
i =0
We use implicit-function theorems to nd coe cients i s.

Inherently local approximation. However, often good global properties.
June 13, 2010
2 / 92
Motivation
Many complicated mathematical problems have:
1
either a particular case
or a related problem.
that is easy to solve.

Often, we can use the solution of the simpler problem as a building
block of the general solution.
Very successful in physics.
Sometimes perturbation is known as asymptotic methods.
June 13, 2010
3 / 92
Applications to Economics
Judd and Guu (1993) showed how to apply it to economic problems.
Recently, perturbation methods have been gaining much popularity.
In particular, second- and third-order approximations are easy to

compute and notably improve accuracy.
Perturbation theory is the generalization of the well-known

linearization strategy.
Hence, we can use much of what we already know about linearization.
June 13, 2010
4 / 92
Regular versus Singular Perturbations

Regular perturbation: a small change in the problem induces a small
change in the solution.
Singular perturbation: a small change in the problem induces a large
change in the solution.
Example: excess demand function.
Most problems in economics involve regular perturbations.
Sometimes, however, we can have singularities. Example: introducing
a new asset in an incomplete market model.
June 13, 2010
5 / 92
References
General:
1
A First Look at Perturbation Theory by James G. Simmonds and

James E. Mann Jr.
Advanced Mathematical Methods for Scientists and Engineers:

Asymptotic Methods and Perturbation Theory by Carl M. Bender,
Steven A. Orszag.
Economics:
1
Perturbation Methods for General Dynamic Stochastic Models by

Hehui Jin and Kenneth Judd.
Perturbation Methods with Nonlinear Changes of Variables by

Kenneth Judd.
A gentle introduction: Solving Dynamic General Equilibrium Models

Using a Second-Order Approximation to the Policy Function by
Martn Uribe and Stephanie Schmitt-Grohe.
June 13, 2010
6 / 92
A Baby Example: A Basic RBC

Model:
max E0
t log ct
t =0
s.t. ct + kt +1 = e zt kt + (1
zt = zt
+ t , t
) kt , 8 t > 0
N (0, 1)
Equilibrium conditions:
1
1
= Et
1 + e zt +1 kt+11
ct
ct + 1
ct + kt +1 = e zt kt + (1 ) kt
zt = zt
+ t
June 13, 2010
7 / 92
Computing the RBC

The previous problem does not have a known paper and pencil
solution.
One particular case the model has a closed form solution: = 1.
Why? Because, the income and the substitution eect from a
productivity shock cancel each other:
1
Labor is constant.
Consumption is a xed fraction of income.
Not very realistic, but we are trying to learn here.

June 13, 2010
8 / 92
Solution
Equilibrium conditions with = 1.
e zt +1 kt+11
1
= Et
ct
ct + 1
ct + kt +1 = e zt kt
zt = zt
+ t
By guess and verify

ct = ( 1
) e zt kt
kt +1 = e zt kt
How can you check? Plug the solution in the equilibrium conditions.
June 13, 2010
9 / 92
Another Way to Solve the Problem

Now let us suppose that you missed the lecture when guess and
verify was explained.
You need to compute the RBC.
What you are searching for? A decision rule for consumption:
ct = c (kt , zt )
and another one for capital:
kt +1 = k (kt , zt )
Note that our d is just the stack of c (kt , zt ) and k (kt , zt ).
June 13, 2010
10 / 92
Equilibrium Conditions
We substitute in the equilibrium conditions the budget constraint and
the law of motion for technology.
And we write the decision rules explicitly as function of the states.
Then, we have:
e zt +t +1 k (kt , zt ) 1
1
= Et
c (kt , zt )
c (k (kt , zt ) , zt + t +1 )
c (kt , zt ) + k (kt , zt ) = e zt kt
This is a system of functional equations.
June 13, 2010
11 / 92
Main Idea
Transform the problem rewriting it in terms of a small perturbation

parameter.
Solve the new problem for a particular choice of the perturbation
parameter.
This step is usually ambiguous since there are dierent ways to do so.
Use the previous solution to approximate the solution of original the
problem.
June 13, 2010
12 / 92
A Perturbation Approach
Hence, we want to transform the problem.
Which perturbation parameter? Standard deviation .
Why ? Discrete versus continuous time.
Set = 0 )deterministic model, zt = 0 and e zt = 1.

We know how to solve the deterministic steady state.
June 13, 2010
13 / 92
A Parametrized Decision Rule

We search for decision rule:
ct = c (kt , zt ; )
and
kt +1 = k (kt , zt ; )
Note new parameter .
We are building a local approximation around = 0.
June 13, 2010
14 / 92
Taylors Theorem
Equilibrium conditions:
Et
1
c (kt , zt ; )
e zt +t +1 k (kt , zt ; ) 1
c (k (kt , zt ; ) , zt + t +1 ; )
c (kt , zt ; ) + k (kt , zt ; )
=0
e zt kt = 0
We will take derivatives with respect to kt , zt , and .

Apply Taylors theorem to build solution around deterministic steady
state. How?
June 13, 2010
15 / 92
Asymptotic Expansion I
ct
= c (kt , zt ; )jk ,0,0 = c (k, 0; 0)

+ck (k, 0; 0) (kt k ) + cz (k, 0; 0) zt + c (k, 0; 0)
1
1
+ ckk (k, 0; 0) (kt k )2 + ckz (k, 0; 0) (kt k ) zt
2
2
1
1
+ ck (k, 0; 0) (kt k ) + czk (k, 0; 0) zt (kt k )
2
2
1
1
+ czz (k, 0; 0) zt2 + cz (k, 0; 0) zt
2
2
1
1
+ ck (k, 0; 0) (kt k ) + cz (k, 0; 0) zt
2
2
1
+ c2 (k, 0; 0) 2 + ...
2
June 13, 2010
16 / 92
Asymptotic Expansion II
kt +1 =
k (kt , zt ; )jk ,0,0 = k (k, 0; 0)
+kk (k, 0; 0) (kt k ) + kz (k, 0; 0) zt + k (k, 0; 0)

1
1
+ kkk (k, 0; 0) (kt k )2 + kkz (k, 0; 0) (kt k ) zt
2
2
1
1
+ kk (k, 0; 0) (kt k ) + kzk (k, 0; 0) zt (kt k )
2
2
1
1
+ kzz (k, 0; 0) zt2 + kz (k, 0; 0) zt
2
2
1
1
+ kk (k, 0; 0) (kt k ) + kz (k, 0; 0) zt
2
2
1
+ k2 (k, 0; 0) 2 + ...
2
June 13, 2010
17 / 92
Comment on Notation
From now on, to save on notation, I will write
#
"
e zt +t +1 k (k t ,zt ;) 1
1
c (k (kt ,zt ;),zt +t +1 ;)
c (k t ,zt ; )
=
F (kt , zt ; ) = Et
c (kt , zt ; ) + k (kt , zt ; ) e zt kt
0
0
Note that:
F (kt , zt ; ) = H (ct , ct +1 , kt , kt +1 , zt ; )
= H (c (kt , zt ; ) , c (k (kt , zt ; ) , zt +1 ; ) , kt , k (kt , zt ; ) , zt ; )

I will use Hi to represent the partial derivative of H with respect to
the i component and drop the evaluation at the steady state of the
functions when we do not need it.
June 13, 2010
18 / 92
Zero-Order Approximation
First, we evaluate = 0:
F (kt , 0; 0) = 0
Steady state:
1
k
=
c
c
1 = k
or,
Then:
c = c (k, 0; 0) = () 1
() 1
k = k (k, 0; 0) = () 1
June 13, 2010
19 / 92
First-Order Approximation
We take derivatives of F (kt , zt ; ) around k, 0, and 0.
With respect to kt :
Fk (k, 0; 0) = 0
With respect to zt :
Fz (k, 0; 0) = 0
With respect to :
F (k, 0; 0) = 0
June 13, 2010
20 / 92
Solving the System I

Remember that:
F (kt , zt ; )
= H (c (kt , zt ; ) , c (k (kt , zt ; ) , zt +1 ; ) , kt , k (kt , zt ; ) , zt ; ) = 0

Because F (kt , zt ; ) must be equal to zero for any possible values of
kt , zt , and , the derivatives of any order of F must also be zero.
Then:
Fk (k, 0; 0) = H1 ck + H2 ck kk + H3 + H4 kk = 0
Fz (k, 0; 0) = H1 cz + H2 (ck kz + ck ) + H4 kz + H5 = 0
F (k, 0; 0) = H1 c + H2 (ck k + c ) + H4 k + H6 = 0
June 13, 2010
21 / 92
Solving the System II

Note that:
Fk (k, 0; 0) = H1 ck + H2 ck kk + H3 + H4 kk = 0
Fz (k, 0; 0) = H1 cz + H2 (ck kz + ck ) + H4 kz + H5 = 0
is a quadratic system of four equations on four unknowns: ck , cz , kk ,

and kz .
Procedures to solve quadratic systems:
1
Blanchard and Kahn (1980) .
Uhlig (1999).
Sims (2000).
Klein (2000).
All of them equivalent.

June 13, 2010
22 / 92
Solving the System III

Also, note that:
F (k, 0; 0) = H1 c + H2 (ck k + c ) + H4 k + H6 = 0
is a linear, and homogeneous system in c and k .
Hence:
c = k = 0
This means the system is certainty equivalent.
Interpretation)no precautionary behavior.
Dierence between risk-aversion and precautionary behavior. Leland
(1968), Kimball (1990).
Risk-aversion depends on the second derivative (concave utility).
Precautionary behavior depends on the third derivative (convex
marginal utility).
June 13, 2010
23 / 92
Comparison with Linearization

After Kydland and Prescott (1982) a popular method to solve
economic models has been the use of a LQ approximation of the
objective function of the agents.
Close relative: linearization of equilibrium conditions.
When properly implemented linearization, LQ, and rst-order
perturbation are equivalent.
Advantages of linearization:
1
Theorems.
Higher order terms.
June 13, 2010
24 / 92
Some Further Comments
Note how we have used a version of the implicit-function theorem.

Important tool in economics.
Also, we are using the Taylor theorem to approximate the policy
function.
Alternatives?
June 13, 2010
25 / 92
Second-Order Approximation
We take second-order derivatives of F (kt , zt ; ) around k, 0, and 0:
Fkk (k, 0; 0) = 0
Fkz (k, 0; 0) = 0
Fk (k, 0; 0) = 0
Fzz (k, 0; 0) = 0
Fz (k, 0; 0) = 0
F (k, 0; 0) = 0
Remember Youngs theorem!
We substitute the coe cients that we already know.
A linear system of 12 equations on 12 unknowns. Why linear?
Cross-terms k and z are zero.
Conjecture on all the terms with odd powers of .
June 13, 2010
26 / 92
Correction for Risk
We have a term in 2 .
Captures precautionary behavior.
We do not have certainty equivalence any more!
Important advantage of second order approximation.
Changes ergodic distribution of states.
June 13, 2010
27 / 92
Higher Order Terms

We can continue the iteration for as long as we want.
Great advantage of procedure: it is recursive!
Often, a few iterations will be enough.
The level of accuracy depends on the goal of the exercise:
1
Welfare analysis: Kim and Kim (2001).
Empirical strategies: Fernndez-Villaverde, Rubio-Ramrez, and Santos

(2006).
June 13, 2010
28 / 92
A Numerical Example
Parameter
Value
0.99
0.33
0.95
0.01
Steady State:
c = 0.388069
k = 0.1883
First-order components:
ck (k, 0; 0) = 0.680101
cz (k, 0; 0) = 0.388069
kk (k, 0; 0) = 0.33
kz (k, 0; 0) = 0.1883
Second-order components:
ckk (k, 0; 0) = 2.41990
ckz (k, 0; 0) = 0.680099
czz (k, 0; 0) = 0.388064
c2 (k, 0; 0) ' 0
kkk (k, 0; 0) = 1.1742

kkz (k, 0; 0) = 0.33
kzz (k, 0; 0) = 0.1883
k2 (k, 0; 0) ' 0
c (k, 0; 0) = k (k, 0; 0) = ck (k, 0; 0) = kk (k, 0; 0) =

cz (k, 0; 0) = kz (k, 0; 0) = 0.
June 13, 2010
29 / 92
Comparison
ct = 0.6733e zt kt0.33
ct ' 0.388069 + 0.680101 (kt k ) + 0.388069zt
2.41990
0.388064 2
zt
(kt k )2 + 0.680099 (kt k ) zt +
2
2
and:
kt +1 = 0.3267e zt kt0.33
kt +1 ' 0.1883 + 0.33 (kt
1.1742
(kt k )2 + 0.33 (kt
2
k ) + 0.1883zt
0.1883 2
k ) zt +
zt
2
June 13, 2010
30 / 92
June 13, 2010
31 / 92
A Computer
In practice you do all this approximations with a computer:
1
First-,second-, and third- order: Matlab and Dynare.
Higher order: Mathematica, Dynare++, Fortran code by Jinn and

Judd.
Burden: analytical derivatives.

Why are numerical derivatives a bad idea?
Alternatives: automatic dierentiation?
June 13, 2010
32 / 92
Local Properties of the Solution

Perturbation is a local method.
It approximates the solution around the deterministic steady state of
the problem.
It is valid within a radius of convergence.
What is the radius of convergence of a power series around x? An
0
0
r 2 R
z j < r , the power series of x 0 will
+ such that 8x , jx
converge.
A Remarkable Result from Complex Analysis

The radius of convergence is always equal to the distance from the center
to the nearest point where the policy function has a (non-removable)
singularity. If no such point exists then the radius of convergence is innite.
Singularity here refers to poles, fractional powers, and other branch
powers or discontinuities of the functional or its derivatives.
June 13, 2010
33 / 92
June 13, 2010
34 / 92
Remarks
Intuition of the theorem: holomorphic functions are analytic.

Distance is in the complex plane.
Often, we can check numerically that perturbations have good non
local behavior.
However: problem with boundaries.
June 13, 2010
35 / 92
Non Local Accuracy Test

Proposed by Judd (1992) and Judd and Guu (1997).
Given the Euler equation:
1
= Et
i
c (kt , zt )
e zt +1 k i (kt , zt ) 1
c i (k i (kt , zt ), zt +1 )
we can dene:
EE i (kt , zt )
c i (kt , zt ) Et
e zt +1 k i (kt , zt ) 1
c i (k i (kt , zt ), zt +1 )
Units of reporting.
Interpretation.
June 13, 2010
36 / 92
June 13, 2010
37 / 92
The General Case

Most of previous argument can be easily generalized.
The set of equilibrium conditions of many DSGE models can be

written as (note recursive notation)
Et H(y , y 0 , x, x 0 ) = 0,
where yt is a ny
states.
1 vector of controls and xt is a nx
1 vector of
Dene n = nx + ny .
Then H maps R ny
R ny
R nx
R nx into R n .
June 13, 2010
38 / 92
Partitioning the State Vector
The state vector xt can be partitioned as x = [x1 ; x2 ]t .

x1 is a (nx
x2 is a ne
ne )
1 vector of endogenous state variables.
1 vector of exogenous state variables.
Why do we want to partition the state vector?
June 13, 2010
39 / 92
Exogenous Stochastic Process

x20 = x2 + e e0
Process with 3 parts:
1
The deterministic component x2 :

1
2
The scaled innovation e e0 where:

1
2
is a ne ne matrix, with all eigenvalues with modulus less than one.

More general: x20 = (x2 ) + e e0 , where is a non-linear function
satisfying that all eigenvalues of its rst derivative evaluated at the
non-stochastic steady state lie within the unit circle.
e is a known ne ne matrix.
e is a ne 1 i.i.d innovation with bounded support, zero mean, and
variance/covariance matrix I .
The perturbation parameter .
We can accommodate very general structures of x2 through changes

in the denition of the state space: i.e. stochastic volatility.
Note we do not impose gaussianity.
June 13, 2010
40 / 92
The Perturbation Parameter
The scalar
0 is the perturbation parameter.
If we set = 0 we have a deterministic model.
Important: there is only ONE perturbation parameter. The matrix e

takes account of relative sizes of dierent shocks.
Why bounded support? Samuelson (1970) and Jin and Judd (2002).
June 13, 2010
41 / 92
Solution of the Model

The solution to the model is of the form:
y = g (x; )
0
x = h (x; ) + e0
where g maps R nx
R + into R ny and h maps R nx
The matrix is of order nx
ne and is given by:

=
R + into R nx .
June 13, 2010
42 / 92
Perturbation
We wish to nd a perturbation approximation of the functions g and

h around the non-stochastic steady state, xt = x and = 0.
We dene the non-stochastic steady state as vectors (x,

y ) such that:
H(y , y , x,
x ) = 0.
Note that y = g (x;

0) and x = h (x;
0). This is because, if = 0,
then Et H = H.
June 13, 2010
43 / 92
Plugging-in the Proposed Solution

Substituting the proposed solution, we dene:
F (x; )
Et H(g (x; ), g (h (x; ) + e0 , ), x, h (x; ) + e0 ) = 0
Since F (x; ) = 0 for any values of x and , the derivatives of any

order of F must also be equal to zero.
Formally:
Fx k j (x; ) = 0
8x, , j, k,
where Fx k j (x, ) denotes the derivative of F with respect to x taken

k times and with respect to taken j times.
June 13, 2010
44 / 92
First-Order Approximation
We are looking for approximations to g and h around (x, ) = (x,
0)
of the form:
g (x; ) = g (x;
0) + gx (x;
0)(x
x ) + g (x;
0)
h (x; ) = h (x;
0) + hx (x;
0)(x
x ) + h (x;
0)
As explained earlier,
g (x;
0) = y
and
h (x;
0) = x.
The remaining four unknown coe cients of the rst-order

approximation to g and h are found by using the fact that:
Fx (x;
0) = 0
and
F (x;
0) = 0
Before doing so, I need to introduce the tensor notation.
June 13, 2010
45 / 92
Tensors
General trick from physics.
An nth -rank tensor in a m-dimensional space is an operator that has n
indices and mn components and obeys certain transformation rules.
[Hy ]i is the (i, ) element of the derivative of H with respect to y :

1
2
The derivative of H with respect to y is an n
ny matrix.
Thus, [Hy ]i is the element of this matrix located at the intersection of

the i-th row and -th column.
ny
n x H i g h
Thus, [Hy ]i [gx ] [hx ]j = =
1 =1 y x x j .
[Hy 0 y 0 ]i :
1
Hy 0 y 0 is a three dimensional array with n rows, ny columns, and ny

pages.
Then [Hy 0 y 0 ]i denotes the element of Hy 0 y 0 located at the
intersection of row i, column and page .
June 13, 2010
46 / 92

gx and hx can be found as the solution to the system:
[Fx (x;
0)]ij = [Hy 0 ]i [gx ] [hx ]j + [Hy ]i [gx ]j + [Hx 0 ]i [hx ]j + [Hx ]ij =
i = 1, . . . , n; j, = 1, . . . , nx ; = 1, . . . , ny
Note that the derivatives of H evaluated at (y , y 0 , x, x 0 ) = (y , y , x,

x )
are known.
Then, we have a system of n nx quadratic equations in the n
unknowns given by the elements of gx and hx .
nx
We can solve with a standard quadratic matrix equation solver.
June 13, 2010
47 / 92

g and h are identied as the solution to the following n equations:
[F (x;
0)]i =
Et f[Hy 0 ]i [gx ] [h ] + [Hy 0 ]i [gx ] [ ] [e0 ] + [Hy 0 ]i [g ]
+[Hy ]i [g ] + [Hx 0 ]i [h ] + [Hx 0 ]i [ ] [e0 ] g
i = 1, . . . , n;
= 1, . . . , ny ;
= 1, . . . , nx ;
= 1, . . . , ne .
Then:
[F (x;
0)]i
= [Hy 0 ]i [gx ] [h ] + [Hy 0 ]i [g ] + [Hy ]i [g ] + [fx 0 ]i [h ] = 0;
i = 1, . . . , n;
= 1, . . . , ny ;
= 1, . . . , nx ;
= 1, . . . , ne .
Certainty equivalence: this equation is linear and homogeneous in g

and h . Thus, if a unique solution exists, it must satisfy:
h = 0/
g = 0
June 13, 2010
48 / 92
Second-Order Approximation I
The second-order approximations to g around (x; ) = (x;
0) is
[g (x; )]i
= [g (x;
0)]i + [gx (x;
0)]ia [(x x )]a + [g (x;
0)]i []
1
+ [gxx (x;
0)]iab [(x x )]a [(x x )]b
2
1
+ [gx (x;
0)]ia [(x x )]a [ ]
2
1
+ [gx (x;
0)]ia [(x x )]a []
2
1
0)]i [][]
+ [g (x;
2
where i = 1, . . . , ny , a, b = 1, . . . , nx , and j = 1, . . . , nx .
June 13, 2010
49 / 92
Second-Order Approximation II
The second-order approximations to h around (x; ) = (x;
0) is
[h(x; )]j
= [h(x;
0)]j + [hx (x;
0)]ja [(x x )]a + [h (x;
0)]j []
1
+ [hxx (x;
0)]jab [(x x )]a [(x x )]b
2
1
+ [hx (x;
0)]ja [(x x )]a []
2
1
+ [hx (x;
0)]ja [(x x )]a []
2
1
0)]j [][],
+ [h (x;
2
where i = 1, . . . , ny , a, b = 1, . . . , nx , and j = 1, . . . , nx .
June 13, 2010
50 / 92
Second-order Approximation III
The unknowns of these expansions are [gxx ]iab , [gx ]ia , [gx ]ia , [g ]i ,
[hxx ]jab , [hx ]ja , [hx ]ja , [h ]j .
These coe cients can be identied by taking the derivative of F (x; )
with respect to x and twice and evaluating them at (x; ) = (x;
0).
By the arguments provided earlier, these derivatives must be zero.
June 13, 2010
51 / 92

We use Fxx (x;
0) to identify gxx (x;
0) and hxx (x;
0):
[Fxx (x;
0)]ijk =
[Hy 0 y 0 ]i [gx ] [hx ]k + [Hy 0 y ]i [gx ]k + [Hy 0 x 0 ]i [hx ]k + [Hy 0 x ]ik [gx ] [hx ]j
+[Hy 0 ]i [gxx ] [hx ]k [hx ]j + [Hy 0 ]i [gx ] [hxx ]jk
+ [Hyy 0 ]i [gx ] [hx ]k + [Hyy ]i [gx ]k + [Hyx 0 ]i [hx ]k + [Hyx ]ik [gx ]j
+[Hy ]i [gxx ]jk
+ [Hx 0 y 0 ]i [gx ] [hx ]k + [Hx 0 y ]i [gx ]k + [Hx 0 x 0 ]i [hx ]k + [Hx 0 x ]ik [hx ]j
+[Hx 0 ]i [hxx ]jk
+[Hxy 0 ]ij [gx ] [hx ]k + [Hxy ]ij [gx ]k + [Hxx 0 ]ij [hx ]k + [Hxx ]ijk = 0;
i = 1, . . . n,
j, k, , = 1, . . . nx ;
, = 1, . . . ny .
June 13, 2010
52 / 92
We know the derivatives of H.

We also know the rst derivatives of g and h evaluated at
(y , y 0 , x, x 0 ) = (y , y , x,
x ).
Hence, the above expression represents a system of n nx nx linear
equations in then n nx nx unknowns elements of gxx and hxx .
June 13, 2010
53 / 92
Solving the System III

Similarly, g and h can be obtained by solving:
[F (x;
0)]i
= [Hy 0 ]i [gx ] [h ]
+[Hy 0 y 0 ]i [gx ] [ ] [gx ] [ ] [I ]
+[Hy 0 x 0 ]i [ ] [gx ] [ ] [I ]
+[Hy 0 ]i [gxx ] [ ] [ ] [I ] + [Hy 0 ]i [g ]
+[Hy ]i [g ] + [Hx 0 ]i [h ]
+[Hx 0 y 0 ]i [gx ] [ ] [ ] [I ]
+[Hx 0 x 0 ]i [ ] [ ] [I ] = 0;
i = 1, . . . , n; , = 1, . . . , ny ; , = 1, . . . , nx ; , = 1, . . . , ne
a system of n linear equations in the n unknowns given by the elements of

g and h .
June 13, 2010
54 / 92
Cross Derivatives
The cross derivatives gx and hx are zero when evaluated at (x,
0).
Why? Write the system Fx (x;
0) = 0 taking into account that all
terms containing either g or h are zero at (x,
0).
Then:
[Fx (x;
0)]ij = [Hy 0 ]i [gx ] [hx ]j + [Hy 0 ]i [gx ] [hx ]j +
[Hy ]i [gx ]j + [Hx 0 ]i [hx ]j = 0;
i = 1, . . . n;
= 1, . . . , ny ;
, , j = 1, . . . , nx .
This is a system of n nx equations in the n

the elements of gx and hx .
nx unknowns given by
The system is homogeneous in the unknowns.

Thus, if a unique solution exists, it is given by:
gx
hx
= 0
= 0
June 13, 2010
55 / 92
Structure of the Solution

The perturbation solution of the model satises:
g (x;
0) = 0
h (x;
0) = 0
gx (x;
0) = 0
hx (x;
0) = 0
Standard deviation only appears in:
1
2
A constant term given by 12 g 2 for the control vector yt .

The rst nx ne elements of 21 h 2 .
Correction for risk.

Quadratic terms in endogenous state vector x1 .
Those terms capture non-linear behavior.
June 13, 2010
56 / 92
Higher-Order Approximations
We can iterate this procedure as many times as we want.
We can obtain n-th order approximations.
Problems:
1
Existence of higher order derivatives (Santos, 1992).
Numerical instabilities.
Computational costs.
June 13, 2010
57 / 92
Erik Eady
It is not the process of linearization that limits insight.
It is the nature of the state that we choose to linearize about.
June 13, 2010
58 / 92
Change of Variables
We approximated our solution in levels.
We could have done it in logs.
Why stop there? Why not in powers of the state variables?
Judd (2002) has provided methods for changes of variables.
We apply and extend ideas to the stochastic neoclassical growth
model.
June 13, 2010
59 / 92
A General Transformation
We look at solutions of the form:
c0
= a k
k0 + bz
k 0
k0
= c k
k0 + dz
Note that:
1
If , , and are 1, we get the linear representation.
As , and tend to zero, we get the loglinear approximation.
June 13, 2010
60 / 92
Theory
The rst order solution can be written as
f (x ) ' f (a ) + (x
a ) f 0 (a )
Expand g (y ) = h (f (X (y ))) around b = Y (a), where X (y ) is the

inverse of Y (x ).
Then:
g (y ) = h (f (X (y ))) = g (b ) + g (b ) (Y (x )
b )
where g = hA fi A Xi comes from the application of the chain rule.

From this expression it is easy to see that if we have computed the
values of fi A , then it is straightforward to nd the value of g .
June 13, 2010
61 / 92
Coe cients Relation
Remember that the linear solution is:

k0
= a1 ( k
l0 ) = c1 ( k
k0
(l
k0 ) + b1 z
k0 ) + d1 z
Then we show that:
a3 = k0
c3 =
a1
1 1
k0 c1
l0
b3 = k0
d3 =
b1
1
l0 d1
June 13, 2010
62 / 92
Finding the Parameters
Minimize over a grid the Euler Error.

Some optimal results
1
0.986534
Euler Equation Errors
SEE
1
1
0.0856279
0.991673 2.47856 0.0279944
June 13, 2010
63 / 92
Sensitivity Analysis
Dierent parameter values.
Most interesting nding is when we change :
Optimal Parameters for dierent s
0.014 0.98140 0.98766 2.47753

0.028 1.04804 1.05265 1.73209
0.056 1.23753 1.22394 0.77869
A rst order approximation corrects for changes in variance!
June 13, 2010
64 / 92
June 13, 2010
65 / 92
A Quasi-Optimal Approximation
Sensitivity analysis reveals that for dierent parametrizations
'
This suggests the quasi-optimal approximation:
k 0
k0
= a3 k
k0 + b3 z
l0
= c3 k
k0 + d3 z
Note that if dene kb = k
Linear system:
1
2
k0 and bl = l
l0 we get:
kb0 = a3 kb + b3 z
bl = c3 kb + d3 z
Use for analytical study.

Use for estimation with a Kalman Filter.
June 13, 2010
66 / 92
Perturbing the Value Function

We worked with the equilibrium conditions of the model.
Sometimes we may want to perform a perturbation on the value
function formulation of the problem.
Possible reasons:
1
Gain insight.
Di culty in using equilibrium conditions.
Evaluate welfare.
Initial guess for VFI.
Perturbation
Methods
June 13,
2010 we 67 / 92
More general point: we can
perturb
any operator problem
that
Basic Problem
Imagine that we have:
"
c
) t
+ Et V (kt +1 , zt +1 )
1
V (kt , zt ) = max (1
ct
s.t. ct + kt +1 = e zt kt + (1
zt = zt
+ t , t
) kt
N (0, 1)
Write it as:
"
V (kt , zt ; ) = max (1
ct
c
+ Et V (kt +1 , zt +1 ; )
) t
1
s.t. ct + kt +1 = e zt kt + (1
zt = zt
+ t , t
) kt
N (0, 1)
June 13, 2010
68 / 92
Alternative
Another way to write the value function is:
max
ct
"
V (kt , zt ; ) =
1
(1 ) c1t +
Et V e zt kt + (1 ) kt ct , zt + t +1 ;
This form makes the dependences in the next period states explicit.
The solution of this problem is value function V (kt , zt ; ) and a

policy function for consumption c (kt , zt ; ).
June 13, 2010
69 / 92
Expanding the Value Function

The second-order Taylor approximation of the value function around the
deterministic steady state (kss , 0; 0) is:
V (kt , zt ; ) '
Vss + V1,ss (kt kss ) + V2,ss zt + V3,ss

1
1
1
+ V11,ss (kt kss )2 + V12,ss (kt kss ) zt + V13,ss (kt
2
2
2
1
1
1
+ V21,ss zt (kt kss ) + V22,ss zt2 + V23,ss zt
2
2
2
1
1
1
+ V31,ss (kt kss ) + V32,ss zt + V33,ss 2
2
2
2
kss )
where
Vss
Vi ,ss
Vij ,ss
= V (kss , 0; 0)
= Vi (kss , 0; 0) for i = f1, 2, 3g
= Vij (kss , 0; 0) for i, j = f1, 2, 3g
June 13, 2010
70 / 92
Expanding the Value Function

By certainty equivalence, we will show below that:
V3,ss = V13,ss = V23,ss = 0
Taking advantage of the equality of cross-derivatives, and setting

= 1, which is just a normalization:
V (kt , zt ; 1) ' Vss + V1,ss (kt kss ) + V2,ss zt
1
1
+ V11,ss (kt kss )2 + V22,ss ztt2
2
2
1
+V12,ss (kt kss ) z + V33,ss
2
Note that V33,ss 6= 0, a dierence from the standard linear-quadratic
approximation to the utility functions.
June 13, 2010
71 / 92
Expanding the Consumption Function

The policy function for consumption can be expanded as:
ct = c (kt , zt ; ) ' css + c1,ss (kt
kss ) + c2,ss zt + c3,ss
where:
c1,ss
c2,ss
c3,ss
= c1 (kss , 0; 0)
= c2 (kss , 0; 0)
= c3 (kss , 0; 0)
Since the rst derivatives of the consumption function only depend on

the rst and second derivatives of the value function, we must have
that c3,ss = 0 (precautionary consumption depends on the third
derivative of the value function, Kimball, 1990).
June 13, 2010
72 / 92
Linear Components of the Value Function

To nd the linear approximation to the value function, we take
derivatives of the value function with respect to controls (ct ), states
(kt , zt ), and the perturbation parameter .
Notation:
1
Vi ,t : derivative of the value function with respect to its i-th argument,

evaluated in (kt , zt ; ) .
Vi ,ss : derivative evaluated in the steady state, (kss , 0; 0).
We follow the same notation for higher-order (cross-) derivatives.
June 13, 2010
73 / 92
Derivatives
Derivative with respect to ct :
(1
) ct
Et V1,t +1 = 0
Derivative with respect to kt :

V1,t = Et V1,t +1 e zt kt
+1
Derivative with respect to zt :

h
i
V2,t = Et V1,t +1 e zt kt + V2,t +1
Derivative with respect to :
V3,t = Et [V2,t +1 t +1 + V3,t +1 ]

In the last three derivatives, we apply the envelope theorem to
eliminate the derivatives of consumption with respect to kt , zt , and .
June 13, 2010
74 / 92
System of Equations I
Now, we have the system:
ct + kt +1 = e zt kt + (1
V (kt , zt ; ) = (1
(1
) ct
1
ct
) kt
+ Et V (kt +1 , zt +1 ; )
Et V1,t +1 = 0
V1,t = Et V1,t +1 e zt kt 1 + 1
h
i
V2,t = Et V1,t +1 e zt kt + V2,t +1
V3,t = Et [V2,t +1 t +1 + V3,t +1 ]
zt = zt
+ t
June 13, 2010
75 / 92
System of Equations II
If we set = 0 and compute the steady state, we get a system of six
equations on six unknowns, css , kss , Vss , V1,ss , V2,ss , and V3,ss :
css + kss = kss
1
Vss = (1
(1
) css
css
+ Vss
1
V1,ss = 0
V1,ss = V1,ss kss 1 + 1

h
i
V2,ss = V1,ss kss + V2,ss
V3,ss = V3,ss
From the last equation: V3,ss = 0.

From the second equation: Vss =
From the third equation: V1,ss =
css
1 .
1
css .
June 13, 2010
76 / 92
System of Equations III

After cancelling redundant terms:
css + kss = kss
Then:
1 = kss 1 + 1
h
i
V2,ss = V1,ss kss + V2,ss
1
kss =
1
1
1+
css = kss
kss
1
k c
V2,ss =
1 ss ss
V1,ss > 0 and V2,ss > 0, as predicted by theory.
June 13, 2010
77 / 92
Quadratic Components of the Value Function

From the previous derivations, we have:
(1
) c (kt , zt ; )
Et V1,t +1 = 0
V1,t = Et V1,t +1 e zt kt 1 + 1
h
i
zt
V2,t = Et V1,t +1 e kt + V2,t +1
V3,t = Et [V2,t +1 t +1 + V3,t +1 ]
where:
kt +1 = e zt kt + (1
zt
= zt
) kt
+ t , t
c (kt , zt ; )
N (0, 1)
We will now take derivatives of each of the four equations with

respect to kt , zt , and .
We will take advantage of the equality of cross derivatives.
The envelope theorem does not hold anymore (we are taking
derivatives of the derivatives of the value function).
June 13, 2010
78 / 92
First Equation I
We have:
(1
) c (kt , zt ; )
Et V1,t +1 = 0

h
(1
) c (kt , zt ; )
Et V11,t +1 e zt kt
In steady state:
V11,ss
(1
) css
or
c1,t
c1,t
h
c1,ss = V11,ss kss
=0
+1
V11,ss
c1,ss =
V11,ss
(1
where we have used that 1 = kss

+1
) css
+1
.
June 13, 2010
79 / 92
First Equation II
(1
) c (kt , zt ; )
Et V11,t +1 e zt kt
c2,t
c2,t + V12,t +1 = 0
In steady state:
(1
V11,ss
) css
c2,ss = V11,ss kt + V12,ss
or
c2,ss =
V11,ss
(1
) css
V11,ss kss + V12,ss
June 13, 2010
80 / 92
First Equation III

(1
) c (kt , zt ; )
c3,t
Et ( V11,t +1 c3,t + V12,t +1 t +1 + V13,t +1 ) = 0

In steady state:
V11,ss
(1
) css
c3,ss = V13,ss
or
c3,ss =
V11,ss
(1
) css
V13,ss
June 13, 2010
81 / 92
Second Equation I
We have:
V1,t = Et V1,t +1 e zt kt

"
V11,t +1 e zt kt 1 + 1
V11,t = Et
+V1,t +1 (
+1
e zt kt
c1,t
+1
1) e zt kt
In steady state:
V11,ss = V11,ss
c1,ss
+ V1,ss (
1) kss
or
V11,ss =
V1,ss (
+ c1,ss
1) kss
June 13, 2010
82 / 92
Second Equation II
2
V11,t +1 e zt kt c2,t e zt kt 1 + 1
V12,t = Et 4
+V12,t +1 e zt kt 1 + 1 + V1,t +1 e zt kt
In steady state:
V12,ss = V11,ss kss

or
V12,ss =
1
1
c2,ss + V12,ss + V1,ss kt
V11,ss kss
c2,ss + V1,ss kss
3
5
June 13, 2010
83 / 92
Second Equation III

V13,t = Et [ V11,t +1 c3,t + V12,t +1 t +1 + V13,t +1 ]
In steady state,
V13,ss
V13,ss
= [ V11,ss c3,ss + V13,ss ] )
V11,ss c3,ss
=
1
but since we know that:
c3,ss =
V11,ss
(1
) css
V13,ss
the two equations can only hold simultaneously if V13,ss = c3,ss = 0.

June 13, 2010
84 / 92
Third Equation I
We have
h
i
V2,t = Et V1,t +1 e zt kt + V2,t +1

V11,t +1 e zt kt c2,t e zt kt + V12,t +1 e zt kt
+V1,t +1 e zt kt + V21,t +1 e zt kt c2,t + V22,t +1 2
V22,t = Et
In steady state:
V22,t
V22,ss
=
=
V11,ss kt c2,ss kss + V12,ss kss + V1,ss kss

+V21,ss kss c2,ss + V22,ss 2
1 2
V11,ss kt c2,ss kss + 2V12,ss kss

+V1,ss kss V12,ss c2,ss
where we have used V12,ss = V21,ss .

June 13, 2010
85 / 92
Third Equation II

V23,t = Et
V11,t +1 e zt kt c3,t + V12,t +1 e zt kt t +1 + V13,t +1 e zt kt

V21,t +1 c3,t + V22,t +1 t +1 + V23,t +1
In steady state:
V23,ss = 0
June 13, 2010
86 / 92
Fourth Equation
We have
V3,t = Et [V2,t +1 t +1 + V3,t +1 ] .
V33,t = Et
V21,t +1 c3,t t +1 + V22,t +1 2 2t +1 + V23,t +1 t +1

V31,t +1 c3,t + V32,t +1 t +1 + V33,t +1
In steady state:
V33,ss =
V22,ss
June 13, 2010
87 / 92
System I
V11,ss
c1,ss =
V11,ss
(1
) css
V11,ss kss + V12,ss

1
) css
V11,ss =
V1,ss ( 1) kss 2
1
1 + c1,ss
i
1 h
V12,ss =
c2,ss + V1,ss kss 1
V11,ss kss
1
V11,ss kt c2,ss kss + 2V12,ss kss

V22,ss =
2
+V1,ss kss V12,ss c2,ss
1
2 V22,ss
V33,ss =
1
c2,ss =
V11,ss
(1
plus c3,ss = V13,ss = V23,ss = 0.

June 13, 2010
88 / 92
System II
This is a system of nonlinear equations.
However, it has a recursive structure.
By substituting variables that we already know, we can nd V11,ss .
Then, using this results and by plugging c2,ss , we have a system of
two equations, on two unknowns, V12,ss and V22,ss .
Once the system is solved, we can nd c1,ss , c2,ss , and V33,ss directly.
June 13, 2010
89 / 92
The Welfare Cost of the Business Cycle

An advantage of performing the perturbation on the value function is
that we have evaluation of welfare readily available.
Note that at the deterministic steady state, we have:
1
V (kss , 0; ) ' Vss + V33,ss
2
Hence 12 V33,ss is a measure of the welfare cost of the business cycle.
Note that this quantity is not necessarily negative. Indeed, it may well
be positive in many models, like in a RBC with leisure choice. See
Cho and Cooley (2000).
June 13, 2010
90 / 92
Our Example
We know that Vss =
css
1 .
Then, we can compute the decrease in consumption that will make

the household indierent between consuming (1 ) css units per
period with certainty or ct units with uncertainty.
To do so, note that:
1
(1
(css (1
1
css
1
+ V33,ss
1 2
)1
= (1
1 css1
=1
1+
css
1
V
2 33,ss
1
) V33,ss
2
or
1
))1
1
1
June 13, 2010
91 / 92
A Numerical Example
We pick standard parameter values by setting
= 0.99, = 2, = 0.0294, = 0.3, and = 0.95.
Then, we get:
V (kt , zt ; 1) '
0.54000 + 0.00295 (kt

0.00007 (kt
0.975082
kss )
kss ) + 0.11684zt
0.00985zt2
0.00225 (kt
c (kt , zt ; ) ' 1.85193 + 0.04220 (kt
kss ) zt
kss ) + 0.74318zt
DYNARE produces the same policy function by linearizing the

equilibrium conditions of the problem.
Also, the consumption equivalent of the welfare cost of the business
cycle is 8.8475e-005, even lower than Lucas(1987) original
computation because of the smoothing possibilities implied by the
presence of capital.
Use as an initial guess for VFI.
June 13, 2010
92 / 92

Perturbation PDF

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Perturbation PDF

Uploaded by

Copyright:

Available Formats

Perturbation Methods

June 13, 2010

Jess Fernndez-Villaverde (PENN)

June 13, 2010

We use implicit-function theorems to nd coe cients i s.

June 13, 2010

either a particular case

that is easy to solve.

June 13, 2010

Recently, perturbation methods have been gaining much popularity.

In particular, second- and third-order approximations are easy to

Perturbation theory is the generalization of the well-known

Hence, we can use much of what we already know about linearization.

Jess Fernndez-Villaverde (PENN)

June 13, 2010

Regular versus Singular Perturbations

Jess Fernndez-Villaverde (PENN)

June 13, 2010

A First Look at Perturbation Theory by James G. Simmonds and

Advanced Mathematical Methods for Scientists and Engineers:

Perturbation Methods for General Dynamic Stochastic Models by

Perturbation Methods with Nonlinear Changes of Variables by

A gentle introduction: Solving Dynamic General Equilibrium Models

Jess Fernndez-Villaverde (PENN)

June 13, 2010

A Baby Example: A Basic RBC

Jess Fernndez-Villaverde (PENN)

June 13, 2010

Computing the RBC

Consumption is a xed fraction of income.

Not very realistic, but we are trying to learn here.

June 13, 2010

By guess and verify

Jess Fernndez-Villaverde (PENN)

June 13, 2010

Another Way to Solve the Problem

June 13, 2010

And we write the decision rules explicitly as function of the states.

Jess Fernndez-Villaverde (PENN)

June 13, 2010

Transform the problem rewriting it in terms of a small perturbation

Jess Fernndez-Villaverde (PENN)

June 13, 2010

Which perturbation parameter? Standard deviation .

Why ? Discrete versus continuous time.

Set = 0 )deterministic model, zt = 0 and e zt = 1.

Jess Fernndez-Villaverde (PENN)

June 13, 2010

A Parametrized Decision Rule

Note new parameter .

We are building a local approximation around = 0.

Jess Fernndez-Villaverde (PENN)

June 13, 2010

We will take derivatives with respect to kt , zt , and .

Jess Fernndez-Villaverde (PENN)

June 13, 2010

= c (kt , zt ; )jk ,0,0 = c (k, 0; 0)

Jess Fernndez-Villaverde (PENN)

June 13, 2010

k (kt , zt ; )jk ,0,0 = k (k, 0; 0)

+kk (k, 0; 0) (kt k ) + kz (k, 0; 0) zt + k (k, 0; 0)

Jess Fernndez-Villaverde (PENN)