Log Linearization

Graduate Macro II, Spring 2010

The University of Notre Dame
Professor Sims
The solutions to many discrete time dynamic economic problems take the form of a
system of non-linear dierence equations. There generally exists no closed-form solution for
such problems. As such, we must result to numerical and/or approximation techniques.
One particularly easy and very common approximation technique is that of log lineariza-
tion. We rst take natural logs of the system of non-linear dierence equations. We then
linearize the logged dierence equations about a particular point (usually a steady state),
and simplify until we have a system of linear dierence equations where the variables of in-
terest are percentage deviations about a point (again, usually a steady state). Linearization
is nice because we know how to work with linear dierence equations. Putting things in
percentage terms (thats the log part) is nice because it provides natural interpretations
of the units (i.e. everything is in percentage terms).
First consider some arbitrary univariate function, ,(r). Taylors theorem tells us that
this can be expressed as a power series about a particular point r

, where r

belongs to the
set of possible r values:
,(r) = ,(r

) +
,
0
(r

)
1!
(r r

) +
,
00
(r

)
2!
(r r

)
2
+
,
(3)
(r

)
3!
(r r

)
3
+ ...
Here ,
0
(r

) is the rst derivative of , with respect to r evaluated at the point r

, ,
00
(r

) is
the second derivative evaluated at the same point, ,
(3)
is the third derivative, and so on. :!
reads : factorial and is equal to :! = :(:1)(:2) ... 1. In words, the factorial of : is
the product of all non-negative integers less than or equal to :. Hence 1! = 1, 2! = 2 1 = 2,
3! = 3 2 1 = 6, and so on.
For a function that is suciently smooth, the higher order derivatives will be small,
and the function can be well approximated (at least in the neighborhood of the point of
evaluation, r

) linearly as:
,(r) = ,(r

) + ,
0
(r

) (r r

)
Taylors theorem also applies equally well to multivariate functions. As an example,
suppose we have ,(r. ). The rst order approximation about the point (r

) is:
,(r. ) ,(r

) + ,
x
(r

) (r r

) + ,
y
(r

) (

)
Here ,
x
denotes the partial derivative of the function with respect to r and similarly for
.
Suppose that we have the following (non-linear) function:
,(r) =
q(r)
/(r)
1
To log-linearize it, rst take natural logs of both sides:
ln ,(r) = ln q(r) ln /(r)
Now use the rst order Taylor series expansions:
ln ,(r) ln ,(r

) +
,
0
(r

)
,(r

)
(r r

)
ln q(r) ln q(r

) +
q
0
(r

)
q(r

)
(r r

)
ln /(r) ln /(r

) +
/
0
(r

)
/(r

)
(r r

)
The above follows from the fact that
d ln f(x)
dx
=
f
0
(x)
f(x)
. Now put these all together:
ln ,(r

) +
,
0
(r

)
,(r

)
(r r

) = ln q(r

) +
q
0
(r

)
q(r

)
(r r

) ln /(r

)
/
0
(r

)
/(r

)
(r r

)
Group terms:
ln ,(r

) +
,
0
(r

)
,(r

)
(r r

) = ln q(r

) ln /(r

) +
q
0
(r

)
q(r

)
(r r

)
/
0
(r

)
/(r

)
(r r

)
But since ln ,(r

) = ln q(r

) ln /(r

), these terms cancel out, leaving:

,
0
(r

)
,(r

)
(r r

) =
q
0
(r

)
q(r

)
(r r

)
/
0
(r

)
/(r

)
(r r

)
To put everything in percentage terms, multiply and divide each term by r

:
r

,
0
(r

)
,(r

)
(r r

)
r

=
r

q
0
(r

)
q(r

)
(r r

)
r

/
0
(r

)
/(r

)
(r r

)
r

For notational ease, dene e r =

(xx

)
x

, or the percentage deviation of r about r

. Then
we have:
r

,
0
(r

)
,(r

)
e r =
r

q
0
(r

)
q(r

)
e r
r

/
0
(r

)
/(r

)
e r
The above discussion and general cookbook procedure applies equally well in multivariate
contexts. To summarize, the cookbook procedure for log-linearizing is:
1. Take logs
2. Do a rst order Taylor series expansion about a point (usually a steady state)
3. Simplify so that everything is expressed in percentage deviations from steady state
2
A number of examples arise in economics. I will log-linearize the following four examples:
(a) Cobb-Douglass production function; (b) accounting identity; (c) capital accumulation
equation; and (d) consumption Euler equation.
(a) Cobb-Douglass Production Function: Consider a Cobb-Douglas production func-
tion:

t
= c
t
/

t
:
1
t
First take logs:
ln
t
= ln c
t
+ cln /
t
+ (1 c) ln :
t
Now do the Taylor series expansion about the steady state values:
ln

+
1

(
t

) = ln c

+
1
c

(c
t
c

)+cln /

+
c
/

(/
t
/

)+(1c) ln :

+
(1 c)
:

(:
t
:

)
As above, note that ln

= ln c

+ cln /

+ (1 c) ln :

, so these terms cancel:

1

(
t

) =
1
c

(c
t
c

) +
c
/

(/
t
/

) +
(1 c)
:

(:
t
:

)
Now using our denition of tilde variables being percentage deviations from steady
state, we have:
e
t
= ec
t
+ c
e
/
t
+ (1 c)e :
t
(b) Accounting Identity: Consider the closed economy accounting identity:

t
= c
t
+ i
t
Take logs:
ln
t
= ln (c
t
+ i
t
)
Now do the rst order Taylor series expansion:
ln

+
1

(
t

) = ln (c

+ i

) +
1
(c

+ i

)
(c
t
c

) +
1
(c

+ i

)
(i
t
i

)
Now we have to ddle with this a bit more than we did for the production function case.
First, note that ln (c

+ i

) = ln

, so that these terms cancel out:

1

(
t

) =
1
(c

+ i

)
(c
t
c

) +
1
(c

+ i

)
(i
t
i

)
3
Now multiply and divide (so as to leave the expression unchanged) each of the two terms
on the right hand side by c

and i

, respectively:
1

(
t

) =
c

(c

+ i

)
(c
t
c

)
c

+
i

(c

+ i

)
(i
t
i

)
i

Now simplify and use our tilde notation:

e
t
=
c

ec
t
+
i

e
i
t
(c) Capital Accumulation Equation: Consider the standard capital accumulation equa-
tion:
/
t+1
= i
t
+ (1 o)/
t
Take logs:
ln /
t+1
= ln(i
t
+ (1 o)/
t
)
Do the rst order Taylor series expansion:
ln /

+
1
/

(/
t+1
/

) = ln(i

+(1o)/

)+
1
(i

+ (1 o)/

)
(i
t
i

)+
(1 o)
(i

+ (1 o)/

)
(/
t
/

)
Now simplify terms a bit, noting that ln(i

+ (1 o)/

) = ln /

, so that again terms

cancel:
1
/

(/
t+1
/

) =
1
/

(i
t
i

) +
(1 o)
/

(/
t
/

)
Now multiply and divide the rst term on the right hand side by i

:
1
/

(/
t+1
/

) =
i

(i
t
i

)
i

+
(1 o)
/

(/
t
/

)
Using our tilde notation:
e
/
t+1
=
i

e
i
t
+ (1 o)
e
/
t
(d) Consumption Euler equation: Consider the standard consumption Euler equation
that emerges from household optimization problems with CRRA utility:

c
t+1
c
t

= ,(1 + :
t
)
o 0 is the coecient of relative risk aversion. Take logs:
o ln c
t+1
o ln c
t
= ln , + ln(1 + :
t
)
Now do the rst order Taylor series expansion:
4
o ln c

+
o
c

(c
t+1
c

) o ln c

o
c

(c
t
c

) = ln , + ln(1 + :

) +
1
1 + :

(:
t
:

)
Some terms on the left hand side obviously cancel:
o
c

(c
t+1
c

)
o
c

(c
t
c

) = ln , + ln(1 + :

) +
1
1 + :

(:
t
:

)
Note that, in the steady state, 1 + :

=
1

, hence ln(1 + :

) = ln ,. Using this, we
have:
o
c

(c
t+1
c

)
o
c

(c
t
c

) =
1
1 + :

(:
t
:

)
There are two semi-standard things to do with the right hand side. First, since :
t
is
already a percent, it is common to leave it in absolute (as opposed to percentage) deviations.
Hence, we can dene e :
t
= (:
t
:

), while, for all other variables, like consumption, we use

the tilde notation to denote percentage deviations, so ec
t
=
(ctc

)
c

, as before. Secondly, we
approximate the term
1
1+r

= 1. If the discount factor is suciently high, this will be a

good approximation. Then, simplifying, we can write:
ec
t+1
ec
t
=
1
o
e :
t
This says that the growth rate of consumption is approximately proportional to the
deviation of the real interest rate from steady state, with
1

interpreted as the elasticity of

intertemporal substitution.
5