Digital Signal Processing Reference
In-Depth Information
We note that these asymptotic biases are similar to those obtained in batch estimation
derived from a Taylor series expansion [77, p. 68] with expression (4.77) of
C
u
.
!
u
1
þo
and E[
l
(
k
)]
l
1
¼ o
:
X
n
1
k
l
1
l
i
2(
l
1
l
i
)
2
1
k
1
k
E[
w
(
k
)]
u
1
¼
i¼
2
Finally, we see that in adaptive and batch estimation, the square of these biases are an
order of magnitude smaller that the variances in
O
(
m
)or
O
(
k
).
This methodology has been applied to compare the theoretical asymptotic per-
formance of several adaptive algorithms for minor and principal component analysis
in [22, 26, 27]. For example, the asymptotic mean square error E(
kW
(
k
)
W
k
2
Fro
)of
the estimate
W
(
k
) given by the WSA algorithm (4.57) is shown in Figure 4.1, where
the stepsize
m
is chosen to provide the same value for
m
Tr(
C
u
). We clearly see in this
figure that the value
b
2
/
b
1
¼
0.6 optimizes the asymptotic mean square error
/
speed
of convergence tradeoff.
10
1
WSA algorithm
10
0
(1)
(2)
(6)
(5)
(3)
(4)
10
−1
(0)
10
−2
0
500
1000
1500
2000
2500
3000
3500
4000
Iteration number
2
Fro
) averaging 100
independent runs for the WSA algorithm, for different values of parameter b
2
/
b
1
¼ 0.96
(1), 0.9 (2), 0.1 (3), 0.2 (4), 0.4 (5), and 0.6 (6) compared with mTr(C
u
) (0) in the case n ¼ 4,
r ¼ 2, C
x
¼ Diag(1.75, 1.5, 0.5, 0.25), where the entries of W(0) are chosen randomly
uniformly in [0, 1].
Figure 4.1 Learning curves of the mean square error E(kW(k)W
k
Search WWH ::
Custom Search