# Derivation of CRLBs for the correlated interferer

models
Sean O’Rourke
March 31st, 2010
1 Correlated Interference Model
Assume that an arbitrary m element array recieves a single scaled and delayed
replica of a known signal s(t), corrupted by an interferer with an unknown (and
possibly complex) signal y(t). The sampled baseband array output is modeled
as the m×1 complex vector
x[n] = as(nT
s
−τ
0
) +by(nT
s
) +e[n] , (1)
where T
s
is the sampling period; a, b ∈ CI
m
are the spatial signatures of the
direct-path arrival and the interferer; τ
0
is the direct path time of arrival; and
e[n] comprises additive (white Gaussian) noise. Let us further suppose that
the interferer’s waveform is correlated with the delayed replica of s(t) to some
extent; that is, we can decompose y
u
(t) as
y(t) = s(t −τ
o
) + y
u
(t) , (2)
where E[y
u
(t)y

u
(t)] = σ
2
b
and E[y
u
(t)s

(t−τ
o
)] = 0 We note that σ
2
b
is a measure
of how uncorrelated y(t) is with the desired signal, and in general is dependent
on τ
0
.
Using this, we can now form Equation 1 as
x[n] = (a +b) s(nT
s
−τ
0
) +by
u
(nT
s
) +e[n] (3)
x[n] = (a +b) s(nT
s
−τ
0
) +z[n] (4)
x[n] = α
u
s(nT
s
−τ
0
) +z[n] (5)
where the covariance matrix of the new “noise and uncorrelated interference”
term z[n] is Q = E

z[n]z
H
[n]

= σ
2
n
I
m
+ σ
2
b
bb
H
. We refer to Equation 4 as
the unstructured model. If we assume that the array is calibrated such that
the response to the direct-path signal, say a
0
, is known to within a complex
constant α
0
, we obtain the form
x[n] = (α
0
a
0
+b) s(nT
s
−τ
0
) +z[n] (6)
1
where the covariance matrix of z[n] is identical to the previous case. We shall
refer to Equation 6 as the structured model.
If we assume that z[n] ∼ CI N(0, Q), then x[n] ∼ CI N(µ, Q) where µ[n] =
(a +b) s(nT
s
−τ
0
) for the unstructured model, and µ[n] = (α
0
a
0
+b) s(nT
s

τ
0
) for the structured model. Since our model is Gaussian, we can use the
Slepian-Bangs formulation [?] to calculate the Fisher information matrix, and
thus the Cramer-Rao bounds for each of our models. The formulation states
that, for a given parameter vector η = [η
1
· · · η
p
] ∈ CI
p
, η[n] ∼ CI N(µ[n], Q), the
(i, j)th element of the associated Fisher information matrix is:
FIM
ij
= 2Re

N
¸
n=0

∂µ[n]
∂η
i

H
Q
−1

∂µ[n]
∂η
j

¸
+Tr

Re

Q
−1

∂Q
∂η
i

Q
−1

∂Q
∂η
j

(7)
Then, the Cramer-Rao bound for each parameter in η can be found along the
diagonal of FIM(η).
2 CRB Derivation for the unstructured model
For this case, our parameter vector is η
u
= [Re{a}, Im{a}, τ
0
, Re{b}, Im{b}, σ
2
b
, σ
2
n
].
Note that for this parameter vector, the structure of our FIM is:
FIM =

F
aa
f

.
.
. F
ab
f

2
b
f

2
n
f
τa
F
ττ
.
.
. f
τb
F
τσ
2
b
F
τσ
2
n
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
F
ba
f

.
.
. F
bb
f

2
b
f

2
n
f
σ
2
b
a
F
σ
2
b
τ
.
.
. f
σ
2
b
b
F
σ
2
b
σ
2
b
F
σ
2
b
σ
2
n
f
σ
2
n
a
F
σ
2
n
τ
.
.
. f
σ
2
n
b
F
σ
2
n
σ
2
b
F
σ
2
n
σ
2
n
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
(8)
where, for example,
F
aa
=
¸
F
a
R
a
R
F
a
R
a
I
F
a
I
a
R
F
a
I
a
R

f
τa
=

f
τ0a
R
f
τ0a
I

with a
R
= Re{a}, a
I
= Im{a}. Each of these submatrices contain the appro-
priate elements from the Slepian-Bangs formulation above.
With this structure in mind, we form the derivatives of the mean and co-
variance matrix needed for the formulation. From inspection, we see that µ[n]
2
has no dependence σ
2
b
or σ
2
n
, nor does Q have any dependence on a or τ
0
. Thus:
∂µ[n]
∂σ
2
b
=
∂µ[n]
∂σ
2
n
= 0
(2m+2)×1
∂Q
∂a
R,i
=
∂Q
∂a
I,i
= 0
m×1
∂Q
∂τ
0
= 0
m×m
where a
R,i
(a
I,i
) is the ith element of the real (imaginary) part of the vector
a. Note that the derivative of Q with respect to a is element-wise, because we
choose not to deﬁne a “3-D” structure to handle the derivative of a matrix with
respect to a vector. These mean that the blocks of the FIM that have derivatives
with respect to a, τ
0
contain only the ﬁrst term in Equation 7; similarly, blocks
with derivatives w.r.t. σ
2
b
, σ
2
n
contain only the second term.
The derivatives of µ[n] are:
∂µ[n]
∂a
R
=
∂µ[n]
∂b
R
= s(nT
s
−τ
0
)I
m×m
(9)
∂µ[n]
∂a
I
=
∂µ[n]
∂b
I
=  s(nT
s
−τ
0
)I
m×m
(10)
∂µ[n]
∂τ
0
= (a +b)d(nT
s
−τ
0
) (11)
where d(t) =
ds(t)
dt
. Likewise, the derivatives of Q are:
∂Q
∂b
R,k
= σ
2
b

be
T
k
+e
k
b
H

(12)
∂Q
∂b
I,k
= − σ
2
b

be
T
k
−e
k
b
H

(13)
∂Q
∂σ
2
n
= I
m×m
(14)
∂Q
∂σ
2
b
= bb
H
(15)
where e
k
is the kth m-dimensional unit vector (e.g., a 1 at position k and zeros
elsewhere). Given these derivatives, we can form the constitutent blocks of the
FIM. We note that any block whose subscripts are reversed (say, τ
0
a
R
) is the
transpose of the original block (say, a
R
τ
0
).
2.1 Direct path FIM components (F
aa
, f

, f
τa
, F
ττ
)
For the upper diagonal blocks in Equation 8 (listed in this subsection’s title),
the zero derivatives discussed above lead to these blocks requiring only the ﬁrst
3
term of Equation 7. The blocks are:
F
aa
=
¸
Re {F
1
} Im{F
1
}
−Im{F
1
} Re {F
1
}

(16a)
f

=
¸
Re {f
3
}
Im{f
3
}

(16b)
f
τa
=

Re {f
3
}
T
Im{f
3
}
T

(16c)
F
ττ
= Re{F
2
} = F
2
(16d)
where
F
1
= 2Q
−1
s(τ
0
)
2
(17a)
F
2
= 2(a +b)
H
Q
−1
(a +b)d(τ
0
)
2
(17b)
f
3
= 2Q
−1
(a +b)s
H

0
)d(τ
0
) (17c)
are used for notational ease. In Equation 16d, note that Q (and thus Q
−1
) is a
positive semideﬁnite matrix, thus any quadratic form x
H
Q
−1
x, x ∈ CI
m
, x = 0
is necessarily real and non-negative. Hence, Re{F
2
} = F
2
.
2.2 “Cross term” FIM components (F
ab
, f

2
b
, f

2
n
, f
τb
, F
τσ
2
b
, F
τσ
2
n
)
The blocks for this section of the FIM are:
F
ab
= F
aa
(18a)
f

2
b
= f

2
n
= 0
2m×1
(18b)
f
τb
= f
τa
(18c)
F
τσ
2
b
= F
τσ
2
n
= 0 (18d)
We obtain zero (a zero vector) for F
τσ
2
b
, F
τσ
2
n
(f

2
b
, f

2
n
) because each of the
parameters zeroes out one term in the Slepian-Bangs formulation. The cor-
respondence of the other blocks with blocks from the previous subsection is a
result of the mean depending on both a and b. As mentioned previously, the
blocks F
ba
, f
σ
2
b
a
, f
σ
2
n
a
, f

, F
σ
2
b
τ
, and F
σ
2
n
τ
are the transposes of the blocks listed
in Equation 18, so they are not listed explicitly.
2.3 Interference and noise FIM components (F
bb
, f

2
b
, f

2
n
,
F
σ
2
b
σ
2
b
, F
σ
2
b
σ
2
n
, F
σ
2
n
σ
2
n
, etc.)
So far, calculation of the FIM blocks has only required the ﬁrst term of the
Slepian-Bangs formulation, because of terms that only existed in the mean;
hence, the derivative of Q with respect to these terms was always zero. With
these blocks, we require either both terms (in the case of F
bb
) or the second
term only (for the rest of the blocks) of Equation 7. This requires more careful
derivations, so we shall address certain blocks in their own subsections.
4
2.3.1 Calculation of F
bb
First, we consider F
bb
. As described above, this block uses both terms of the
Slepian-Bangs formulation. For ease of notation, let us write this block as the
sum of two matrices, one formed from each term of Equation 7:
F
bb
= M+C (19)
where M is the 2m× 2m matrix from the term of mean derivatives, and C is
the 2m×2m matrix from the term of covariance derivatives. Each of these are
further subdivided into 4 blocks as follows:
F
bb
=
¸
F
b
R
b
R
F
b
R
b
I
F
b
I
b
R
F
b
I
b
R

(20a)
M =
¸
M
b
R
b
R
M
b
R
b
I
M
b
I
b
R
M
b
I
b
R

(20b)
C =
¸
C
b
R
b
R
C
b
R
b
I
C
b
I
b
R
C
b
I
b
R

(20c)
where the subscripts R and I indicate the real and imaginary parts of a vector,
respectively.
Since the derivatives of the mean with respect to the elements of b are the
same as those with respect to a, it is clear that
M =
¸
Re {F
1
} Im{F
1
}
−Im{F
1
} Re {F
1
}

(21)
with F
1
is deﬁned as in Equation 16a.
In order to derive the form for C, we require the use of several properties
from matrix algebra and calculus. First, we use the following property of the
trace operator [?]:
Tr(ABCD) = vec(D)
T
(A⊗C
T
) vec(B
T
) (22)
where A, B, C, D are conformable matrices, vec(A) stacks the columns of A
into a single vector, and ⊗ is the Kr¨onecker product. Furthermore, note that
for a rank-one matrix, say uv
H
, its vectorization is:
vec(uv
H
) = v

⊗u (23)
Given this, we can write Equations 121315 Using the second term of the Slepian-
Bangs formulation, we can write the (k, l)th element of C
b
R
b
R
as:
[C
b
R
b
R
]
k,l
= Tr

Q
−1
∂Q
∂b
R,k
Q
−1
∂Q
∂b
R,l

(24)
= vec

∂Q
∂b
R,l

T
(Q
−1
⊗Q
−1
T
) vec

∂Q
∂b
R,k
T

(25)
3 CRB Derivation for the structured model
Here, our parameter vector is η
s
= [Re{α
0
}, Im{α
0
}, τ
0
, Re{b}, Im{b}, σ
2
b
, σ
2
n
]
5