Digital Signal Processing Reference
In-Depth Information
the point z located at infinity, z = z
≡∞. Then, the ensuing first L + K
expansion coe cients of the generated series for R (PA) (z −1 ) would exactly
coincide with F M (z −1 ) from (12.4)
P L (z −1 )
Q K (z −1 )
b n z −n
=
(12.13)
n=0
L+K
L+K
b n z −n =
c n z −n
(12.14)
n=0
n=0
b n
= c n
0≤n≤L + K.
(12.15)
Hence M = L + K (QED). Relationship (12.15) follows from the uniqueness
theorem of power series expansions. Thus, the PA from (12.2) and the original
infinite sum F(z −1 ) from (12.4) are in the best contact since they exhibit exact
agreement to within the first L+K terms of the Maclaurin expansion (12.13) of
P L (z −1 )/Q K (z −1 ). Moreover, the result M = L+K simultaneously provides
the Pade estimate of the difference between (12.4) and (12.2), as expressed
symbolically by
F(z −1 )− P L (z −1 )
Q K (z −1 )
=O (z −L−K−1 )
z−→∞ (12.16)
whereO (z −L−K−1 ) is the remainder of power series expansions around
z = z =∞. The functionO (z −L−K−1 ), as an explicit error of the ap
proximation F(z −1 )≈P L (z −1 )/Q K (z −1 ), itself represents a power series
with expansion coe cients{a n
}multiplied by z −L−K−m (m = 1, 2, 3,...,∞)
O (z −L−K−1 ) =
a n z −n
a n = c n −b n
(12.17)
n=L+K+1
where, in general, a n
= c n for L + K + 1≤n≤∞. In other words, the
mentioned “figure of merit” is explicitly given by the easily obtainable error
termO (z −L−K−1 ), which is an infinite sum with higherorder expansion
terms than those retained in the Maclaurin series (12.13) for the polynomial
quotient P L (z −1 )/Q K (z −1 ) from the PA.
The definition (12.16) of the PA is reminiscent of a variancetype estimate
for the difference between the input data (observed, measured), F(z −1 ), and
output (modeled) function, P L (z −1 )/Q K (z −1 )
Pade Variance =O (z −L−K−1 )
= F(z −1 ){input data (observed, measured)} (12.18)
P L (z −1 )
Q K (z −1 )
{output data (modeled, objective function)}.
Search WWH ::




Custom Search