7.7 Estimates ofthe Power Spectrum
BO ETB
which gives
Da. D xa & x(n - i) = — J x(a) x(o~ i) fori 1,2....p) 778.181
Quiz 7.11 Derive [7.73.15]
Note that k = 0, 1,.. . ,p. When these equations are derived using expectation they
‘are Known as the normal equations, and are also sometimes referred to as the
Yule—Walker equations and the Wiener-Hopf equations. We will refer fo these equa
tions by the same names for the least square case as well. The equations can be x=
pressed in a more compact form using the autocorrelation estimates as follows:
Rix —K) = Rafi) 1 = 12... 773.161
scat
A 1
Ru) = yD (0) x(n + II) (7.7.3.171
Note thatthe 1/' term is not needed, as we mentioned previously. This is only a scale
factor and does not affect the solution of the equations forthe i, Equation [773.16] isa
Setof p equations wth p unknowns and can be solved using matrix techniques as follows:
Ryx(0) Rex) Rex(-(@ - )] [ar] Rx)
Reef) Rex) FRx“ - 9) Ru] yz ar0
[ue - 0) ex - 2) x — | La, Ral)
These equations can be solved by matrix inversion, but this is not as efficient as the
Levinson-Durbin algorithm, which solves the equations for the & in a recursive manner.
Algorithms for doing this are available in MATLAB and take advantage of the fact that
we autocorrelation matrix is Toeplitz, MATLAB has.a function called pe that returns
she 4, coefficients, given the data. Note that since the autocorrelation function is real and.
even, then the terms above the main diagonal in the correlation matrix in [7.7.3.18] ean
be replaced with terms with a positive argument.
The autocorrelation method is guaranteed to give a stable LP filter, namely, one
sach that the poles of 1/ A(z) are all within the unit cirele in the z-plane. Since we are as-
seming real data, the 4, coefficients are real, implying that the poles of 1/A(z) appear in
complex conjugate pairs. Thus, for example, if p is even, then we will have an even
saber of coefficients and an even number of poles, with one-half of the poles being in
se upper half of the unit cirelein the 2-plane and one-half ofthe poles being in the lower
£. Of course, one or more pairs of poles could fall on the real line in the z-plane. If p
odd, then at least one root will be on the real line.