Tommaso Proietti 391

yearly and monthly time series is discussed in Ravn and Uhlig (2002) and Maravall
and del Rio (2007).
We now show that the Leser filter is the optimal signal extraction filter for the

LLTM (9.2) withση^2 =0 andλ=σψ^2 /σζ^2. In fact, apart from an additive term
which does not depend onμ,PLSis proportional to lnf(y,μ)=lnf(y|μ)+lnf(μ),
wheref(y,μ),f(y|μ)denote, respectively, the Gaussian joint density of the ran-
dom vectorsyandμ, and the conditional density ofygivenμ, whereasf(μ)
is the joint density ofμt,t =1,...,n. Now, lnf(y|μ)depends onμonly via

( 1 /σψ^2 )
∑n
t= 1 (yt−μt)

(^2) , whereas lnf(μ)=lnf(μ
3 ,...,μn|μ 1 ,μ 2 )+lnf(μ 1 ,μ 2 ).
The first term depends onμt,t>2, only via( 1 /σζ^2 )
∑n
t= 3 (
(^2) μ
t)
(^2). The contri-
bution of the initial values vanishes under fixed initial conditions or diffuse initial
conditions.^1 In conclusion,μ ̃maximizes, with respect toμ, the joint log-density
lnf(y,μ)and thus the posterior log-density lnf(μ|y)=lnf(y,μ)−lnf(y). A con-
sequence of this result is that the components can be efficiently computed using
the Kalman filter and smoother (see Appendix C). The latter computes the mean
of the conditional distributionμ|y. As this distribution is Gaussian, the posterior
mean is equal to the posterior mode. Hence, the smoother computes the mode of
f(μ|y), which is also the minimizer of thePLScriterion.
The equivalenceλ=σψ^2 /σζ^2 makes clear that the roughness penalty measures
the variability of the cyclical (noise) component relative to that of the trend
disturbance, and regulates the smoothness of the long-term component. Asσζ^2
approaches zero,λtends to infinity, and the limiting representation of the trend
is a straight line. The Leser–HP detrended or cyclical component is the smoothed
estimate of the componentψtin (9.2) and, although the maintained representa-
tion for the deviations from the trend is a WN component, the filter has been one
of the most widely employed tools in macroeconomics to extract a measure of the
business cycle. For the US GDP series (logarithms) this component is plotted in the
top right-hand panel of Figure 9.4 (p. 404).
In terms of the reduced form of model (9.2), the IMA(2,2) model^2 yt=( 1 +
θ 1 L+θ 2 L^2 )ξt,ξt∼NID(0,σ^2 ), it can be shown that the restrictionση^2 =0 implies
[( 1 +θ 2 )θ 2 ]/( 1 −θ 2 )^2 =λandθ 1 =− 4 θ 2 /( 1 +θ 2 ). Therefore, forλ=1600, we
haveθ 1 =−1.778 andθ 2 =0.799, so thatθ( 1 )= 1 +θ 1 +θ 2 =0.021 and the MA
polynomial is close to noninvertibility at the zero frequency.
The theoretical properties of the Leser–HP filter are better understood by assum-
ing the availability of a doubly-infinite sample,yt+j,j =−∞,...,∞. In such a
setting, the Wiener–Kolmogorov filter (see Whittle, 1983, and Appendix B) pro-
vides the minimum mean square linear estimator (MMSLE) of the trend,μ ̃t|∞=
wμ(L)yt, where:
wμ(L)=
σζ^2
σζ^2 +| 1 −L|^4 σψ^2

1
1 +λ| 1 −L|^4

. (9.3)

The frequency response function of the trend filter (see Appendix A) is:

wμ(e−ıω)=

1

1 + 4 λ( 1 −cosω)^2

, ω∈[0,π];