649

Harald Uhlig*

ECONOMIC RISK

Approximate Solutions to
Dynamic Models
Linear Methods

BERLIN

SFB 649 Discussion Paper 2006-030

SFB

* Humboldt-Universitt zu Berlin, Germany

This research was supported by the Deutsche

Forschungsgemeinschaft through the SFB 649 "Economic Risk".
http://sfb649.wiwi.hu-berlin.de
ISSN 1860-5664
SFB 649, Humboldt-Universitt zu Berlin
Spandauer Strae 1, D-10178 Berlin

Approximate Solutions to Dynamic Models - Linear

Methods
by Harald Uhlig, Humboldt Universitt zu Berlin, uhlig@wiwi.hu-berlin.de, February 2006.
This research was supported by the Deutsche Forschungsgemeinschaft through the SFB 649,
"Economic Risk". Keywords: numerical methods, linear solution method, loglinearization,
dynamic stochastic general equilibrium methods, recursive law of motion. JEL codes: C60,
C61, C63, E32

Abstract:
Linear Methods are often used to compute approximate solutions to dynamic models, as these
models often cannot be solved analytically. Linear methods are very popular, as they can
easily be implemented. Also, they provide a useful starting point for understanding more
elaborate numerical methods. It shall be described here first for the example of a simple real
business cycle model, including how to easily generate the log-linearized equations needed
before solving the linear system. For a general framework, formulas are provided for
calculating the recursive law of motion. The algorithm described here is implemented with
the "toolkit" programs available per http://www.wiwi.hu-berlin.de/wpol/html/toolkit.htm .

0. Introduction
Linear Methods are often used to compute approximate solutions to dynamic models, as these
models often cannot be solved analytically. While a plethora of advanced numerical methods
exist, the most popular "bread-and-butter'' method for solving them is linearization. It shall
be described here first for the example of a simple real business cycle model, including how
to easily generate the log-linearized equations needed before solving the linear system. The
classic reference for solving linear difference models under rational expectations is Blanchard
and Kahn (1980), while Kydland and Prescott (1982) is the origin of the modern approach of
calculating numerically approximate solutions to dynamic stochastic models in order to obtain
quantitative results. Much of the material here is taken from Uhlig (1999), which builds on the
method of undetermined coefficients in King, Plosser and Rebelo (2002).

1. A basic example
As a basic example, consider a version of the real business cycle model of Hansen (1985). A
social planner or representative agent chooses ct, kt, yt, lt and nt to maximize the utility
function

U = E u (ct , lt )
t =0

for some twice differentiable utility function u(.), satisfying the usual conditions, subject to
the constraints

ct + kt = yt + (1 )kt 1
yt = t f (kt 1 , nt )
1 = nt + lt

as well as a given initial capital stock k-1,where ct denotes consumption, kt denotes capital, yt
denotes output, lt denotes leisure, nt denotes labor, f(k,n) denotes a twice differentiable
production function, typically assumed to obey constant returns to scale, is the discount
factor and t is total factor productivity, with

zt = log ( t ) log ( * )

evolving according to

zt = zt 1 + t

where

Et [ t +1 ] = 0

for some values and , with -1 < < 1. A solution is a stochastic sequence (ct, kt, yt, lt,
nt),t0 where all variables dated t are independent of all s for s>t and satisfies all constraints,
and which maximizes the utility function given above within the set of all such sequences.
The necessary first-order conditions for this problem are given by

uc ( ct , lt ) = t

ul ( ct , lt ) = f n ( kt 1 , nt )

t = Et [ t +1Rt +1 ]
Rt = f k ( kt 1 , nt ) + 1
2. Linearization
The first step towards solving the model by linear approximation is to linearize all the
constraints and necessary equations (possibly after substituting out some variables, if so
desired). Linearization amounts to finding a first-order approximation to all equations.
Formally, linearization amounts to replacing a set of equations

0 = g ( xt )
in a vector xt of variables with its linearized counterpart around some point of approximation
x*,

0 = g ( x* ) + g ' ( x* ) x%t

where

x%t = xt x*

is the deviation of xt from the approximation point x* and where G'(x*) is the matrix of first
derivatives of G(.). As point of approximation x*, the nonstochastic steady state is often
chosen, i.e. one solves the equations

0 = g ( x* )

under the assumption that all exogenous stochastic variables are constant (here: t= and all
s=0). Then, the remaining linearized system consists of

0 = g ' ( x ) x%t
*

Since many economic variables are constrained to be positive, it is often more attractive to
log-linearize the equations, rather than to linearize them. The difference between linearization
and log-linearization is that entries in xt denote the original variable (e.g. consumption ct) in
the case of linearization and the log of these variables (e.g. log(ct)) in the case of loglinearization. There is no need to choose either linearization or log-linearization for all entries
in xt. One may choose to linearize some and log-linearize others or take other
transformations. Indeed, for variables such as trade balances, it is better to use linearization
rather than log-linearization, if they can take negative values. Also, e.g. tax rates are often
more appropriately linearized rather than log-linearized to provide a more useful
interpretation.
This makes no difference as far as the linearized solution is concerned. More generally,
differentiable and differentiable invertible transformations (i.e. homeomorphisms) of the
variables (e.g. taking ratios of variables, etc.) make no difference to the properties of the
linearized solution. The differences only always lies in the recalculation of the original
variables, where one may want to take into account the nonlinearities originally inherent in
the model. To see more generally, that any homeomorphism (i.e. differentiable and
differentiably invertible transformation)

yt = h( xt )

of the variables makes no difference to remaining calculations, note that the equations can be
restated as

0 = g ( h 1 ( yt ) )

The linearized version is now

0 = g h 1 ( y* ) + g ' ( x* )( f 1 ) ' ( y* ) y%t

which coincides with the previous linearization if y*=F(x*), noting that

y = f ' ( x* ) xt

as well as

I = f '( x

)( f ) '( y )
1

While linearization can be performed numerically or with the usual rules of calculus, one can
often "read" the log-linearized version of an equation from its original form, exploiting

xt = exp ( yt ) x* + x* y%t
t Instead of y%t for the loglinear
where now y = log( x ) . Write x
t
t

deviation.

For log-linearization, the following useful "rules" can easily be derived. Let at, bt, ct be three
variables, with ct=h(at) for some monotone and differentiable function h(.), and let B be some
constant. Then,

)
Ba b ( Ba b ) + ( Ba b ) ( a + b )

at + Bbt ( a* + Bb* ) + a*at + Bb*bt

* *

t t

* *

h '(a* )a*
ct
at
*
h( a )
Either with these rules or directly, the equations in the example log-linearize to

c*ct + k *kt = y* yt + (1 )k *kt 1

fk k *
f n n*
yt = zt +
kt 1 +
nt
f
f
0 = n*n + (1 n* ) l
t

*
*
u
c
u
l
t = cc ct + cl lt
uc
uc

ucl l *
ull l * f nk k *
f nn n*
ct +
lt =
kt 1 +
nt
ul
ul
fn
fn

t = Et t +1 + Rt +1
*
*
f
k
f
n
*
R Rt = kk kt 1 + kn nt
fk
fk

3. Solving for the recursive law of motion

With some further algebra, one can turn this system into a second-order one-dimensional
difference equation,

0 = Et [ Fxt +1 + Lzt +1 ] + Gxt + Mzt + Hxt 1

plus the evolution of the exogenous state,

zt = Nzt 1 + O t

where xt=kt is the capital stock, and F, L, G, M, H, N and O are real numbers (here, with N=
and O=1). Alternatively, use the system of equations above directly (or with some variables
substituted out) and stack all variables into a vector xt to reformulate it in this form, where
now F, L, G, M and H are matrices of coefficients. Indeed, if there is more than one
predetermined variable like kt-1 in the system of equations, one will need to use such a matrix
restatement of the equations anyways. More generally, zt may also be a vector, and N and O
matrices.
Anderson et al (1996) as well as Binder and Pesaran (1997) contain detailed and general
results for solving linearized systems. In most cases, the system has a solution in the form of
a recursive law of motion,

xt = Pxt 1 + Qzt

for some coefficient matrices P and Q. Most models require the solution to be stable, i.e. all
eigenvalues of P to be less than unity in absolute value. Often, one also allows for roots equal
to unity in absolute value, as this arises easily e.g. in models of international trade or with
multiple agents: one may then want to think of the linear approximation as a local solution. In
many models, this uniquely determines the matrix P and usually also Q.
The solutions can be found by substituting the recursive law of motion in for xt+1 and again
for all xt into the second-order difference equation above, exploiting

Nzt = Et [ zt +1 ]

so that only xt-1 and zt and some coefficient matrices remain.

Examine first the equation by matching coefficients on xt-1. One obtains the equation

0 = FP 2 + GP + H

for P. In case of a one-dimensional difference equation (as can be obtained for the example
above and xt=kt), this is a quadratic equation in the feedback coefficient P, which has two
solutions. The system is said to be saddle-path stable, if only one of the two roots is smaller
than unity in absolute value. Thus, if a stable solution is desired, this is the unique solution
for P.
Generally, the equation above is a matrix quadratic equation, which can be solved per
computing generalized eigenvalues or by QZ-decomposition as follows. Let m be the
dimensionality of xt. Define the matrices

G H
F 0m
A=
,B =

0
0
I
I
m
m
m
m
where I
m is the m-by-m identity matrix and 0 m the m-by-m matrices of only zeros.

Recall that a generalized eigenvector s with eigenvalue for the matrices A and B is defined
as satisfying

Bs = As

The generalized eigenvector problem reduces to the standard eigenvector problem of B-1A, if
B is invertible. If s is a generalized eigenvector with eigenvalue for the matrices A and B
above, it can be written as s'=[ x',x'] for some m-dimensional vector x. If there are m
generalized eigenvalues 1,..., m together with generalized eigenvectors

si = [ i xi ', xi ']

such that

C = [ x1 ,..., xm ]

is of full rank, then

P = C C 1

is a solution to the matrix quadratic equation, where

1 0
0
2

=
M M
0 0

K 0

O M

K m
K

is the diagonal matrix of the eigenvalues for the generalized eigenvectors used as well as of P.
The system is said to be saddle-path stable, if there are exactly m generalized eigenvalues
smaller than unity in absolute value. In that case, the matrix P is unique, if one requires all
eigenvalues of P to be stable. If there are fewer than m eigenvalues smaller than (or equal to)
unity in absolute value, then there is no solution, such that the difference equation xt=Pxt-1
remains bounded for all x0. In that case, the set of bounded solution is characterized by e'x0 =
0 as well as e'Qzt = 0 for all t for all eigenvectors e of P corresponding to explosive
eigenvalues. The second of these two constraints may impose restrictions on the exogenous
shock process. If there are more than m eigenvalues smaller than (or equal to) unity in
absolute value, then sunspot solutions may arise, i.e., there are additional solutions. In the
one-dimensional case and if F is nonzero, the general solution is now given by the original
equation, i.e. as

xt = F Gxt 1 F Hxt 2 F ( LN + M ) zt 1 + t
1

where

t is any stochastic process with Et [ t +1 ] = 0 and which is

independent of all s for s>t, but not necessarily independent of t. Note that the recursive law
of motion now includes an additional lag of the state variable, as well as the possibility for
additional random influences ("sunpots") via

t , which are not part of the original system

of equations. Farmer (1999) provides a detailed treatment of sunspots in linearized solutions.

Equivalently, consider the stacked variable st'=[ xt', xt-1'], and note that the second half of this
vector is "predetermined", i.e. must be independent of all s for s>t-1. The linearized system
can be rewritten as

M LN
BEt [ st +1 ] = Ast +
zt

If B is invertible, the solutions can now be characterized in terms of the eigenvalues and
eigenvectors of B-1A. This is the approach taken in the classic reference of Blanchard and
Kahn (1980).
Alternatively, find the QZ-decomposition (or generalized Schur decomposition) of A and B,
see Sims (2002), i.e., find unitary matrices U and V as well as upper triangular matrices K and
L such that

A = U ' LV
B
U
KV
=
'
(and recall that a matrix is unitary, if the product with its complex conjugate transpose is the
identity matrix). Such a Schur decomposition always exists, although it may not be unique.
Partition U and V into m-by-m submatrices,

U11 U12
V11 V12
U =
,V =

U
U
V
V
21
21 22
22
If U
21 and V21 are invertible, then
P = V211V22

solves the matrix quadratic equation. Suppose furthermore, that the QZ-decomposition has

Lii / K ii are in ascending order. Furthermore, suppose

< 1. Then P is stable.

been chosen so that the ratios

Lmm / K mm

To solve for Q, given a solution to P, compare the coefficients on zt to find

V vec(Q) = vec( LN + M )
where vec(.) denotes columnwise vectorization and where

V = N ' F + I k ( FP + G )

with k the dimensionality of zt. If V is invertible, the solution is unique.

Many links for codes for solving dynamic stochastic models are available per
http://dge.repec.org/codes.html. The procedure outlined above has been used in particular in
the toolkit programs of H. Uhlig, see http://www.wiwi.hu-berlin.de/wpol/html/toolkit.htm.
For a discussion of the accuracy of linearized solution, see e.g. Taylor and Uhlig (1990) and
Aruoba et al (2005).

4. References
Anderson, Evan W. , Ellen R. McGrattan, Lars Peter Hansen and Thomas J. Sargent,
"Mechanics of forming and estimating dynamic linear economies", Chapter 04 in Handbook
of Computational Economics, 1996, vol. 1, pages 171-252, Elsevier, Amsterdam
Aruoba, Boraan, Jess Fernndez-Villaverde and Juan F. Rubio-Ramrez, "Comparing
Solution Methods for Dynamic Equilibrium Economies", Journal of Economic Dynamics and
Control, forthcoming (2006).
Binder, Michael & Pesaran, M Hashem, 1997. "Multivariate Linear Rational Expectations
Models: Characterization of the Nature of the Solutions and Their Fully Recursive
Computation," Econometric Theory, Cambridge University Press, vol. 13(6), pages 877-88.
Blanchard, Olivier Jean and Charles M. Kahn, "The Solution of Linear Difference Models
under Rational Expectations," Econometrica, vol. 48, no. 5 (July 1980), 1305-1312.
Farmer, Roger E.A., Macroeconomics of Self-fulfilling Prophecies, MIT Press, 1999,
Cambridge, MA.
King, Robert G & Plosser, Charles I & Rebelo, Sergio T, 2002. "Production, Growth and
Business Cycles: Technical Appendix," Computational Economics, Springer, vol. 20(1-2),
pages 87-116.
Kydland, Finn E. and Edward C. Prescott, "Time to Build and Aggregate Fluctuations,"
Econometrica, vol. 50, no. 6 (Nov. 1982), 1345-1370.
Sims, Christopher A, 2002. "Solving Linear Rational Expectations Models," Computational
Economics, Springer, vol. 20(1-2), pages 1-20.
Uhlig, Harald, "A toolkit for analysing nonlinear dynamic stochastic models easily," in
Ramon Marimon and Andrew Scott, eds, Computational Methods for the Study of Dynamic
Economies, Oxford University Press, Oxford (1999), 30-61.
Taylor, John B. and Harald Uhlig, "Solving nonlinear stochastic growth models: a comparison
of alternative solution methods," Journal of Business Economics and Statistics 8, 1-17 (1990).

SFB 649 Discussion Paper Series 2006

For a complete list of Discussion Papers published by the SFB 649,
please visit http://sfb649.wiwi.hu-berlin.de.
001
002
003
004
005
006
007
008
009
010
011
012
013
014
015
016
017
018
019
020
021
022

"Calibration Risk for Exotic Options" by Kai Detlefsen and Wolfgang K.

Hrdle, January 2006.
"Calibration Design of Implied Volatility Surfaces" by Kai Detlefsen and
Wolfgang K. Hrdle, January 2006.
"On the Appropriateness of Inappropriate VaR Models" by Wolfgang
Hrdle, Zdenk Hlvka and Gerhard Stahl, January 2006.
"Regional Labor Markets, Network Externalities and Migration: The Case
of German Reunification" by Harald Uhlig, January/February 2006.
"British Interest Rate Convergence between the US and Europe: A
Recursive Cointegration Analysis" by Enzo Weber, January 2006.
"A Combined Approach for Segment-Specific Analysis of Market Basket
Data" by Yasemin Boztu and Thomas Reutterer, January 2006.
"Robust utility maximization in a stochastic factor model" by Daniel
HernndezHernndez and Alexander Schied, January 2006.
"Economic Growth of Agglomerations and Geographic Concentration of
Industries - Evidence for Germany" by Kurt Geppert, Martin Gornig and
Axel Werwatz, January 2006.
"Institutions, Bargaining Power and Labor Shares" by Benjamin Bental
and Dominique Demougin, January 2006.
"Common Functional Principal Components" by Michal Benko, Wolfgang
Hrdle and Alois Kneip, Jauary 2006.
"VAR Modeling for Dynamic Semiparametric Factors of Volatility Strings"
by Ralf Brggemann, Wolfgang Hrdle, Julius Mungo and Carsten
Trenkler, February 2006.
"Bootstrapping Systems Cointegration Tests with a Prior Adjustment for
Deterministic Terms" by Carsten Trenkler, February 2006.
"Penalties and Optimality in Financial Contracts: Taking Stock" by
Michel A. Robe, Eva-Maria Steiger and Pierre-Armand Michel, February
2006.
"Core Labour Standards and FDI: Friends or Foes? The Case of Child
Labour" by Sebastian Braun, February 2006.
"Graphical Data Representation in Bankruptcy Analysis" by Wolfgang
Hrdle, Rouslan Moro and Dorothea Schfer, February 2006.
"Fiscal Policy Effects in the European Union" by Andreas Thams,
February 2006.
"Estimation with the Nested Logit Model: Specifications and Software
Particularities" by Nadja Silberhorn, Yasemin Boztu and Lutz
Hildebrandt, March 2006.
"The Bologna Process: How student mobility affects multi-cultural skills
and educational quality" by Lydia Mechtenberg and Roland Strausz,
March 2006.
"Cheap Talk in the Classroom" by Lydia Mechtenberg, March 2006.
"Time Dependent Relative Risk Aversion" by Enzo Giacomini, Michael
Handel and Wolfgang Hrdle, March 2006.
"Finite Sample Properties of Impulse Response Intervals in SVECMs with
Long-Run Identifying Restrictions" by Ralf Brggemann, March 2006.
"Barrier Option Hedging under Constraints: A Viscosity Approach" by
Imen Bentahar and Bruno Bouchard, March 2006.
SFB 649, Spandauer Strae 1, D-10178 Berlin
http://sfb649.wiwi.hu-berlin.de

This research was supported by the Deutsche

Forschungsgemeinschaft through the SFB 649 "Economic Risk".

023
024
025
026
027
028
029
030

"How Far Are We From The Slippery Slope? The Laffer Curve Revisited"
by Mathias Trabandt and Harald Uhlig, April 2006.
"e-Learning Statistics A Selective Review" by Wolfgang Hrdle, Sigbert
Klinke and Uwe Ziegenhagen, April 2006.
"Macroeconomic Regime Switches and Speculative Attacks" by Bartosz
Makowiak, April 2006.
"External Shocks, U.S. Monetary Policy and Macroeconomic Fluctuations
in Emerging Markets" by Bartosz Makowiak, April 2006.
"Institutional Competition, Political Process and Holdup" by Bruno
Deffains and Dominique Demougin, April 2006.
"Technological Choice under Organizational Diseconomies of Scale" by
Dominique Demougin and Anja Schttner, April 2006.
"Tail Conditional Expectation for vector-valued Risks" by Imen Bentahar,
April 2006.
"Approximate Solutions to Dynamic Models Linear Methods" by Harald
Uhlig, April 2006.

SFB 649, Spandauer Strae 1, D-10178 Berlin

http://sfb649.wiwi.hu-berlin.de
This research was supported by the Deutsche
Forschungsgemeinschaft through the SFB 649 "Economic Risk".