Digital Signal Processing Reference
In-Depth Information
Proof
The cost function given in (
6.29
) can be rewritten as
E
tr
x(k)
k)
Z
k
−
1
2
J
(k,P
FA
)
=
−ˆ
x(k
|
k)
Z
k
−
1
E
tr
x(k)
k)
T
x(k)
=
−ˆ
x(k
|
−ˆ
x(k
|
k)
T
Z
k
−
1
E
tr
x(k)
k)
x(k)
=
−ˆ
x(k
|
−ˆ
x(k
|
k)
T
Z
k
−
1
tr
E
x(k)
k)
x(k)
=
−ˆ
x(k
|
−ˆ
x(k
|
tr
E
E
x(k)
k)
x(k)
k)
T
Z
k
Z
k
−
1
=
−ˆ
x(k
|
−ˆ
x(k
|
tr
E
P(k
Z
k
−
1
=
|
k)
|
tr
P
MRE
(k
|
k)
≈
tr
P(k
1
)
−
q
2
λ(k)V(k),P
D
tr
W(k)S(k)W
T
(k)
,
=
|
k
−
where the first equality is due to the property that the trace of a scalar is itself, the
third one is due the property that tr
{
AB
}=
tr
{
BA
}
, the fourth one is due to linearity
of tr
operators, and the fifth one follows from the
smoothing property
[
9
]
of expectations. Note that
W(k)S(k)W
T
(k)
{·}
and E
[·]
W(k)S(k)W
T
(k)
0,
and
q
2
(λ(k)V(k),P
D
)
is the only term that depends on
P
FA
. Hence the minimiza-
tion of
≥
0 implies tr
{
}≥
(k,P
FA
)
can be achieved by maximizing
q
2
(λ(k)V(k),P
D
)
over
P
FA
,
which completes the proof.
J
Remark 6.2
We experimentally observe that choosing any other scalar mea-
sures for the function
f
S
[·]
results
in the same optimization problem given in (
6.30
) where the elements of the
set are the determinant, 1-norm (the largest column sum), 2-norm (the largest
singular value),
from the set
{|·|
,
·
1
,
·
2
,
·
∞
,
·
F
}
∞
-norm (the largest row sum) and Frobenius-norm of a ma-
trix, respectively.
Due to mathematical intractability, the problem given in (
6.30
) was solved by uti-
lizing some line search algorithms that require only the evaluation of the cost func-
tion (e.g., Golden-Section or Fibonacci Search methods) [
22
]. We call this scheme
as
DYNAMIC-MRE-LS
in Fig.
6.2
.
Lemma 6.2
(A Closed-Form Solution [
5
])
An approximate closed-form solution
for the MRE-based dynamic threshold optimization can be found for a special type
Search WWH ::
Custom Search