7.7 Estimates ofthe Power Spectrum
P, the data record is N long. No window is applied to the
The mam {ssume we have suficient data to calculate the desired conclation noe
tmatrix equations have the same form, namely
[Oxc(1, 1) Cyx(1,2) Cox(1, p) Cx (1, 0)
] Cx (22) Cex2p)l] a} _fexx(2,0)
| 773.25)
[ Croxl. 1) Cxx(p,2) Gx fli} — [exxp.0)
ee correlation matrix is no longer Toeplitz, but hs the properties of a covariance
ot i te i tame for this method is derived. The solution for this mehad
sor te Levinson-Durbin algorithm, but instead uses Cholesky decomposition
her methods), which is also available in MATLAB.
VAG) de oe netiod may yield estimates ofthe i coeficients such that the LP
sic doesnot have all of is potes inside the unit cree thats theft may be
able
tum to [7.7.3.20] and note that all ofthe equations are prediction equations for
Ssocorrelaton function. In particular the last equation ean be written
Rux(p) + a Rex(p 1) + + apRux(0) =0 17.7.3.26]
P)=0 7327
Therefore, the linear prediction equations lead to a comrela-
treby the correlation function can be extended, This is the
Feat maximum entropy spectral estimator mentioned earlier, which fe ene
ther in the Problems
74 Power Srectrat, Estimation wir Linear PRrepicrion
re gion iS a data model, Consequently, itis frequently used to estimate the
vim of data that are modeled using linear prediction techniques,
Definition 7.19. The LP power spectral estimate isthe reciprocal ofthe square
{the absolute value ofthe transform ofthe LP coeficients, namely
1 L
ba = a =
1A@)| f+ Dae
Bp ferz=eP pany
least squares
as the number of
Polynomial in z, the roots of this polynomial are
ch in tum are the poles of 1/A(z),