Skip to content
Prev 132084 / 398506 Next

discrepancy between periodogram implementations ? per and spec.pgram

Prof Brian Ripley wrote:

            
Dear Prof. Ripley,

thanks very much for a quick and helpful response. In the last question 
I wanted to hint at
specARIMA which I am using to get the theoretical spectral density of an 
ARMA process.

This works very well in general, however, in a simple example

X_t=0.7*X_{t-1}+epsilon_t

I obtain a value 1.768253 for funscaled[1] ( the first Fourier frequency 
0.003141593)

using

str(f<-specARIMA(eta=c(H=0.5,phi=c(0.7),psi=c()),p=1,q=0,m=2000))
funscaled<-numeric(length(f$freq))
funscaled<-f$spec*f$theta1

where the theoretical value should be 0.901878 with

b<-0.7
omega<-0.003141593
1/(2*pi)*(1-b^2)/(1+b^2-2*b*cos(omega))
[1] 0.9018088


using the formula (2.40) in Fan and Yao, Nonlinear Time Series ( 
Springer 2003 ), page 54-55

Is there also a simple explanation for this ? am I overlooking something ?

Thanks and best regards,

Lieven Desmet,

maths dept - KULeuven - Belgium

Disclaimer: http://www.kuleuven.be/cwis/email_disclaimer.htm