At. Walden et Ml. Williams, DECONVOLUTION, BANDWIDTH, AND THE TRISPECTRUM, Journal of the American Statistical Association, 88(424), 1993, pp. 1323-1329
In the largest application area of time series analysis-geophysical ex
ploration-the underlying innovations sequence is of primary interest a
nd must be estimated. This sequence is estimated by deconvolving the n
on-Gaussian, noninvertible time series. This involves estimation of a
phase-shift correction from the time series, which can be carried out
by maximizing the kurtosis of the series. Unfortunately, the method is
hampered by the fact that the time series is typically deficient in p
ower in certain bands of frequencies (''band-limited''). The consequen
ces of this can be analyzed by studying the trispectrum-the third of t
he polyspectra-of the series. This reveals two important results. Firs
t, we are able to easily appreciate why for certain types of band-limi
tation, kurtosis cannot be used to determine a phase correction. Secon
d, by looking at the inner and outer subvolumes of the support volume
for the discrete-parameter trispectrum, we see that for the standard l
inear model the trispectrum is non-0 in both the inner and outer volum
es, whereas the trispectrum of a series sampled finely enough to avoid
aliasing from a continuous fourth-order stationary process is equal t
o the same continuous-parameter trispectrum value in the inner volumes
, but always 0 in the outer volumes. Hence, when looking at higher-ord
er structure, the standard linear model need not give results that acc
ord with physical reality.