Some Notes on Log-linearization

Katrin Rabitsch, Central European University

October 9, 2008

The log-deviation of a variable from its steady state value is dened as:

xt xt x xt x
bt = log
x = log (xt ) log (x) = log 1 + ' (1)
x x x
With this notation, a variable is at steady state when its log-deviation is zero,
bt = 0. The popularity stems largely from the fact that this approximation
x
is unit free. Log-deviations are approximate percentage deviations from steady
state and the coe cients of log-linear approximations are elasticities:

xt xt x xt x
bt = log
x ' log 1 + '
x x x
where the last step has used the rst order approximation:
1
log (1 + z) ' log (1 + z0 ) +(z z0 ) = z
1 + z0
around the point z0 = 0, so it is an approximation that is valid when the
deviation is small.

There are serveral ways to obtain the log-linearization of a function. Let us,
in particular, go through three methods to do so. We will use two sample
equations and apply each method on them. We will denote the left hand side
of Example equation 1 and 2 by LHS, and, respectively, the right hand side by
RHS. Our two example equations are:

Kt = It + (1 ) Kt 1 (Example equation 1)

Yt = Zt Kt Nt1 (Example equation 2)

which have steady states:

K = I + (1 )K =) I= K

1
Y = ZK N

1
Method 1:
Method 1 consists of taking a rst-order Taylor approximation of the log
of a function, around the log steady state of the variables.

Suppose we have an equation of the following form:

f (xt ; yt ) = g (zt ) (2)

where xt ; yt and zt are stricly positive variables. This equation is clearly also
valid at steady state:

f (x; y) = g (z) (3)

To nd the log-linearized version of eq. (2), rewrite the variables using identity
xt = exp (log (xt )) and then take logs on both sides:

log f elog(xt ) ; elog(yt ) = log g elog(zt )

Now take a rst order Taylor approximation around the steady state
(NOTE: the approximation is around log (x) ; log (y) and log (z), and NOT
x, y and z). After some calculations, we can write the left hand side as:

1
log (f (x; y))+ [f1 (x; y) (log (xt ) log (x)) + f2 (x; y) (log (yt ) log (y))]
f (x; y)
(4)
Similarly, the right hand side can be written as:
1
log (g (z)) + [g 0 (z) (log (zt ) log (z))] (5)
g (z)
Equating eq. (4) and eq. (5) and using the steady state eq. (3), yields the
following log-linearized equation:

bt + f2 (x; y) ybt ] = g 0 (z) zbt

[f1 (x; y) x (6)
Notice that this is a linear approximation in the deviations!

Example equation 1:

Lets go to our example equations. Rewriting Kt = elog(Kt ) , It = elog(It ) and

taking logs of both sides of Example equation 1 gives:

log(It )
log elog(Kt ) = log et + (1 ) elog(Kt 1)

The (log) LHS can be approximated as (applying formula eq. (5) with g (Kt ) =
Kt ):

2
 1
log (LHS) = log elog(Kt ) ' log K + elog K log Kt log K
elog K
= log K + log Kt log K
= log K + b
kt
Notice that this can be written as log (LHS) log K = log (LHS) log LHS =
d =b
LHS kt . The (log) RHS can be approximated as (applying formula eq. (4)
with f (It ; Kt ) = It + (1 ) Kt 1 ):

log(It )
log RHS = log et + (1 ) elog(Kt 1)

log(I )
' log et + (1 ) elog(K ) +
1
+ elog I log It log I +
log(I ) log(K )
et + (1 )e
1
+ (1 ) elog K log Kt 1 log K
log(I ) log(K )
et + (1 )e

I bit + (1 )K b
= log I + (1 )K + kt 1
I + (1 )K I + (1 )K
Notice that this can be written as log (RHS) log I + (1 ) K = log (RHS)
d I b (1 )K b
log RHS = RHS = I+(1 )K it + I+(1 )K kt 1 . Joining the (log) LHS and
the RHS and cancelling out the steady state relationship log K = log I + (1 )K :
Method 1 delivers the following loglinear version of Example equation 1:

b I bit + (1 )K b
kt = kt 1
I + (1 )K I + (1 )K
b
kt = bit + (1 )b
kt 1 (7)

Example equation 2:

Rewriting again Kt = elog(Kt ) , Zt = elog(Zt ) , Zt = elog(Zt ) , Nt = elog(Nt ) ,

Yt = elog(Yt ) , and taking logs of both sides of Example equation 2 gives:

1
log (LHS) = log elog(Yt ) ' log Y + elog Y log Yt log Y
elog Y
= log Y + log Yt log Y
= log Y + ybt

3
The RHS of equation (Example equation 2) gives:

log(Z ) log(K ) (1 ) log(N )

log (RHS) ' log et + log et log et

1 log(Z )
+ e log Zt log Z +
log(Z ) t
et
1 log(K )
+ e log Kt log K +
log(K ) t
et
1 (1 ) log(N )
+ e (1 ) log Nt log N
(1 ) log(N ) t
et
1
h i
= log ZK N + zbt + b
kt + (1 bt
)n

Join the (log) LHS and the RHS and cancel out the steady state relationship
1
log Y = log ZK N .
Method 1 delivers the following loglinear version of Example equation 2:

ybt = zbt + b
kt + (1 bt
)n (8)

Method 2:
Method 2 is a simple trick and consists of rewriting the variable in such
a way that we can directly approximate around the point where per-
centage deviations from steady state are zero.

In most cases, there is no need for explicit dierentiation of the function f and
g. Instead, the log-linearized equation can usually be obtained with a simpler
method. Notice, that we can write:
xt xt
xt = x= xelog( x ) = xexbt
x
bt = 0,
Taking a rst order Taylor approximation around the steady state, x
yields

xexbt ' xe0 + xe0 (bxt 0)

bt )
' x (1 + x

Lets apply this simple trick method to our example equations:

Example equation 1:

4
 Kt
Rewrite the LHS Example equation 1 as LHS = Kt = K K
K
t
= Kelog( K ) =
b
Kekt and take a rst order Taylor approximation around the steady state,
i.e. around b
kt = 0.

b
LHS = Kekt
' Ke0 + Ke0 b
kt 0

= K 1+b
kt

Rewrite the RHS of Example equation 1 by writing in a similar fashion Kt 1 =

b b
Kekt 1 and It = Ieit :
b b
RHS = Ieit + (1 ) Kekt 1

' Ie0 + Ie0 bit 0 + (1 ) Ke0 + (1 ) Ke0 b

kt 1 0

= I 1 + bit + (1 )K 1 + b
kt 1

Join the LHS with the RHS and eliminate the steady state expressions.
Method 2 delivers the following loglinear version of Example equation 1:

b
kt = bit + (1 )b
kt 1 (9)

Example equation 2:

LHS ' Y (1 + ybt )

1
RHS ' ZK N 1 + zbt + b
kt + (1 bt + zbt b
)n kt + zbt (1 bt + b
)n kt (1 bt
)n
1
ZK N 1 + zbt + b
kt + (1 bt
)n

where the last step follows from the fact that for an approximation close to
bt ; b
zbt ; n kt = 0 the cross terms can be ignored.
Join LHS with RHS.
Method 2 delivers the following loglinear version of Example equation 2:

ybt = zbt + b
kt + (1 bt
)n (10)

Other examples:

by the same logic, we can approximate xt as:

5
 xt xt
xt = x = x e log( x ) = x e bt
x
x
x exbt ' xe0 + xe0 (b xt 0)
' x (1 + xbt )

taking log-linear approximations of a multiplicative function xt yt :

xt yt bt ) y (1 + ybt )
' x (1 + x
bt + ybt + x
' xy (1 + x bt ybt )

an additive function, xt + yt , can be approximated as:

xt + yt ' xb
xt + yb
yt
bt ybt ' 0, since x
where x bt and ybt are numbers close to zero.

Method 3:
Alternatively, we might want to take a rst order Taylor approximation
directly around the level x, and transform the approximation such that
we carry percentage deviations. By dividing (and multiplying) the term
(xt x) (which is the absolute deviation of variable xt from its steady state
level x) by x, we also obtain a linear approximation in percentage deviations,
xt x
x (which, approximately, equals log xt log x):

f (xt ) ' f (x) + f 0 (x) (xt x)

xt x
' f (x) + f 0 (x) x
x
f 0 (x) x
' f (x) + f (x) bt
x
f (x)
' f (x) (1 + xbt )
f 0 (x)x
where f (x) is the elasiticity.

Example equation 1:

6
First order Taylor approximation around the steady state, i.e. around K
of the LHS eq. (Example equation 1) gives:

LHS = Kt
' K + Kt K
Kt K
= K +K
K
' K 1+b
kt

the RHS gives:

RHS = It + (1 ) Kt 1
' I + (1 ) K + It I + (1 ) Kt 1 K
It I Kt 1 K
= I + (1 )K + I + (1 )K =
I K
= I + (1 b
) K + I it + (1 ) Kb
kt 1

Join LHS with RHS and eliminating steady state.

Method 3 delivers the following loglinear version of Example equation 1:

b
kt = bit + (1 )b
kt 1 (11)

Example equation 2:

LHS = Yt
Y
' Y + Yt Y
Y
= Y (1 + ybt )

1 1
RHS ' ZK N +K N Zt Z +
1 1
+ZK N Kt K + ZK N (1 ) Nt N
1
ZK N 1 + zbt + b
kt + (1 bt
)n

Join LHS with RHS.

Method 3 delivers the following loglinear version of Example equation 2:

ybt = zbt + b
kt + (1 bt
)n (12)