** Comment on multiple parameter estimation

Most of the ideas can be extended from the single parameter case
Estimate more than one parameter

Read 2.4.3 Van Trees

The General Gaussian Problem

Cases where conditional density of R is Gaussian.

Definition : A set of random variables r1, r2, ..., rN are defined as jointly Gaussian
if all their linear combinations are Gaussian random variables.
Definition : A vector r is a Gaussian random vector when its components r1, r2,
..., rN are jointly Gaussian variables.

If z =

g i ri

GT r

i 1

is a Gaussian random variable for all finite GT, then r is a Gaussian vector.

E[r] = m
cov[r] = E[(r - m) (rT- mT)]

If is nonsingular
N
p(R) = 2 2

1
1

2
exp R T m T 1 R m

Definition : A hypothesis testing problem is called a General Gaussian if p(R|Hi)

is a Gaussian density on all hypotheses.
Similarly, estimation problem p(R|A) Gaussian density A
General Gaussian case.

Jan12

xxx
DCE

40

 r1
r
2
r = . , E[r|H1] =

.
rN

Er1 | H1 m11
Er | H m
2 1 12
.
.
m1

.
.
ErN | H1 m1N

The covariance matrix is

1 K11
K
1 21
T
T
K1 E[(r m1) (r - m1 )|H1] = .

.
under H1
1 K N1

1K 23 ... 1K 2 N

1K NN

1K 12

1K 13 ...

1K 22

.
.

1 K 1N

If Q1 K1-1 Q1K1 = K1Q1 = I

N
Thus, p(R|H1) = 2 2 K 1

Similarly,

1
1 T

T
2

m
Q
R

m
exp
1
1
1
2

1 1
N
1

p(R|H0) = 2 2 K 0 2 exp R T m T0

2
The LRT
1
1

K 0 2 exp R T m1T Q1 R m1
2

(R) =
1
1

K 1 2 exp R T m T0 Q 0 R m 0
2

R m 0

H1

H0

>
<

Taking logarithm

1 T
1
(R m T0 )Q 0 (R m 0 ) R T m1T Q1(R m1 )
2
2
H1
>
< ln + ln |K | - ln |K | *
1

H0
L.H.S difference of quadratic forms.
Jan12

xxx
DCE

41

Special Cases
1.

Equal Covariance Matrices

K1 = K0 K
m1 m0
-1
Q = K symmetric

The above simplifies to

m1T m T0 Q R

H1
>
<

ln +

H0

1 T
m1 Qm1 m T0 Qm 0 *,
2

Let m = m1 m0
l(R) mTQR
or
l(R) RTQm

H1
>
<

*,

H0
H1
>
<

*,

H0
Scalar Gaussian r.v a linear transformation by Gaussian r.v.s.
Thus, performance can be completely characterized by
2

d
Thus

El | H1 El | H 0 2

varl | H 0

Normalizing

E(l|H1) = mTQm1
E(l|H0) = mTQm0
var(l|H0) = E{[mTQ(R-m0)] [(R-m0T) Qm]}
= mTQm
(because E{(R-m0)(R-m0 T) } = K = Q-1)
Therefore
d2 = mTQm
Jan12

xxx
DCE

42

If K = 2 I Q =

I
2
1
1
d2 = mT
I m =
mTI m
2
2

1
1
=
|m|T =
[(m11 m01)2 + (m12 m02)2 + ]
2
2
l(R) =

mTR

The sufficient statistic is just the dot (scalar) product between R and
the mean difference vector m.

Equal mean vectors

m1 = m0 m 0 without loss of generality
Q = Q0 Q1
l(R) RQR

H1
>
<
H0

R , QR l(R) is not a Gaussian r.v.

Jan12

xxx
DCE

43