1

Cramer-Rao Lower Bound Computation Via

the Characteristic Function
Steven Kay, Fellow, IEEE, and Cuichun Xu

Abstract

The Cramer-Rao Lower Bound is widely used in statistical signal processing as a benchmark to
evaluate unbiased estimators. However, for some random variables, the probability density function has
no closed analytical form. Therefore, it is very hard or impossible to evaluate the Cramer-Rao Lower
Bound directly. In these cases the characteristic function may still have a closed and even simple form. In
this paper, we propose a method to evaluate the Cramer-Rao Lower Bound via the characteristic function.
As an example, the Cramer-Rao Lower Bound of the scale parameter and the shape parameter of the
K-distribution is accurately evaluated with the proposed method. Finally it is shown that for probability
density functions with a scale parameter, the Cramer-Rao Lower Bound for the remaining parameters do
not depend on the scale parameter.

I. I NTRODUCTION

In statistical parameter estimation theory, the Cramer-Rao Lower Bound (CRLB) is a lower bound on
the variance of any unbiased estimator [1]. Let p(x; θ) denotes the probability density function (PDF) of
a random variable X parameterized by θ , where θ = [θ1 , θ2 , · · · , θL ]T . Let θ̂ be an unbiased estimator
of θ . If p(x; θ) satisfies the “regularity” conditions, or
∂ ln p (x; θ)
 
E =0
∂θ

This work was supported by the Office of Naval Research under contracts N00014-04-10022 and N00014-04-M-0331 and by
the Air Force Research Laboratory, Sensors Directorate, Hanscom AFB, MA.
Steven Kay is with the Department of Electrical Computuer and Biomedical Engineering, University of Rhode Island, Kingston,
RI 02881. (Phone: 401-874-5804, Fax: 401-782-6422, e-mail: kay@ele.uri.edu)
Cuichun Xu is with the Department of Electrical Computuer and Biomedical Engineering, University of Rhode Island,
Kingston, RI 02881. (Phone: 401-782-7207, Fax: 401-782-6422, e-mail: ccxu@ele.uri.edu)
 2

where E [·] denotes the expectation, then

  h i
var θ̂i ≥ I−1 (θ) = CRLBθi (1)
ii
 
where var θ̂i denotes the variance of the ith element of the estimator, or the estimator for θi , I−1 (θ)
 
ii

is the (i, i) element of the inverse Fisher information matrix, and is the CRLB for θi . The Fisher
information matrix is defined as
" #
∂ ln p(x; θ) ∂ ln p(x; θ)
[I(θ)]ij = E . (2)
∂θi ∂θj
It is seen that the key step to obtain the CRLB is the evaluation of [I (θ)]ij . Compared to other variance
bounds [2], [3], the CRLB is usually easier to compute. Therefore it is extensively used in the signal
processing literature as a benchmark to evaluate the performance of an unbiased estimator. However, there
are cases in which p (x; θ) has no closed form. As a result, the CRLB is hard or impossible to evaluate
analytically via (2). However, in some cases in which p (x; θ) has no closed form, the characteristic
function (CF) of p (x; θ) does have a closed and even simple form, for example the K-distribution [4],
the α-stable distribution [5] and the Class A distribution [6].
The CF of p (x; θ) is defined as
h i
φX (ω; θ) = E ejωX . (3)

This is just the Fourier transform of p (x; θ). Therefore p (x; θ) can be thought of as a time domain
function and φX (ω; θ) as a frequency domain function. Since (2) is essentially an integration in the time
domain, it is possible to evaluate I (θ) equivalently in the frequency domain. This idea could be applied
to both univariate and multivariate distributions. In this paper, we focus on parameter CRLB evaluation
of univariate distributions. The extension to multivariate distributions will be explored in the future work.
The paper is organized as follows. In Section II, the method to evaluate the CRLB via the CF is
introduced. In Section III, an example is given and Section IV offers some conclusions.

II. CRLB C OMPUTATION VIA THE C HARACTERISTIC F UNCTION

Equation (2) can be rewritten as

Z ∞ ∂ ln p (x; θ) ∂ ln p (x; θ)
[I (θ)]ij = p (x; θ) dx
−∞ ∂θi ∂θj
Z∞ ∂p (x; θ) ∂p (x; θ) 1
= dx. (4)
−∞ ∂θi ∂θj p (x; θ)
L
Assume there is a real positive number L, such that p (x; θ) = 0 for |x| > 2, then
Z L/2 ∂p (x; θ) ∂p (x; θ) 1
[I (θ)]ij = dx. (5)
−L/2 ∂θi ∂θj p (x; θ)
 3

x
Let u = L, Z 1/2 ∂p (Lu; θ) ∂p (Lu; θ) 1
[I (θ)]ij = Ldu.
−1/2 ∂θi ∂θj p (Lu; θ)
Let G (u; θ) = p (Lu; θ)
Z 1/2 ∂G (u; θ) ∂G (u; θ) 1
[I (θ)]ij = L du. (6)
−1/2 ∂θi ∂θj G (u; θ)
From the property of a PDF, p (x; θ) and equivalently G (u; θ) ≥ 0, so we can consider G (u; θ) to be a
discrete-time power spectral density (PSD). Therefore, G (u; θ) can be approximated by an autoregressive
(AR) model for a sufficiently large model order, i.e. [7]
σu2
G (u; θ) ≈ (7)
|A (u)|2
where σu2 is the excitation noise variance, and

A (u) = 1 + a [1] e−j2πu + a [2] e−j2πu2 + · · · + a [p] e−j2πup

where a[1], a[2], · · · , a[p] are the AR coefficients with p being the order of the AR model. Let F −1 denote
the inverse Fourier transform. It is readily seen that
Z 1/2
F −1 {A (u)} = A (u)ej2πun du
−1/2

= a[n]

σu2
The proper model order p can be determined by comparing the estimated AR PSD |A(u)|2
with the true
G (u; θ) or numerically calculated G (u; θ). If the fitting error is too large, then a larger p should be
used. Plugging (7) into (6), we have
1/2 " #
L ∂G (u; θ) ∂G (u; θ) ∗
Z 
[I (θ)]ij = A (u) A (u) du
σu2 −1/2 ∂θi ∂θj
" #∗
L
Z 1/2 
∂G (u; θ) ∂G (u; θ)
= A (u) A (u) du. (8)
σu2 −1/2 ∂θi ∂θj
The last equation relies on the property that p (x; θ), and therefore G (u; θ) is real. By Parseval’s theorem,
∞
L X
[I (θ)]ij = x [n]y ∗ [n] (9)
σu2 n=−∞
where
∂G (u, θ)
 
−1
x [n] = F A (u)
∂θi
and ( )
∂G (u, θ)
y [n] = F −1 A (u)
∂θj
 4

Let g [n] = F −1 {G (u; θ)}, so that g[n] can be thought of as the autocorrelation sequence associated
with the PSD G (u; θ) [7]. It is given by
Z 1/2
g [n] = G (u; θ)ej2πun du
−1/2
Z 1/2
= p (Lu; θ)ej2πun du. (10)
−1/2

Letting x = uL in the above equation yields

Z 1/2 x x
g[n] = p(x; θ)ej2π L n d
−1/2 L
1 h n
i
= EX ej2π L x
L
1 2πn
 
= φX ;θ . (11)
L L
Since the CF is assumed to have a closed analytical form, g[n] can be easily obtained from (11).
With g[0], g[1], · · · , g[p] known, the Yule-Walker method can be used to determine the AR parameters
a[1], a[2], · · · , a[p], σu2 [7].


Next, since the inverse Fourier transform is linear, we have

∂G (u; θ) ∂F −1 {G (u; θ)}
 
−1
F = . (12)
∂θi ∂θi
Plugging (11) into (12), we have
 

∂G (u; θ)
 ∂ L1 φX 2πn
L ;θ
F −1 =
∂θi ∂θi
1 ∂ 2πn
 
= φX ;θ . (13)
L ∂θi L
Let
∂ 2πn
 
hi [n] = φX ;θ (14)
∂θi L
so that from (9)
∞
1 X
[I (θ)]ij = (a [n] ⋆ hi [n]) (a [n] ⋆ hj [n])∗ (15)
Lσu2 n=−∞
where ⋆ denotes the convolution.
If there is a real positive integer M , such that hi [n] ≈ 0, and hj [n] ≈ 0 for |n| > M , then (15) can
be approximated as
M
1 X
[I (θ)]ij = (a [n] ∗ hi [n]) (a [n] ∗ hj [n])∗ . (16)
Lσu2 n=−M

After the analytical form of hi [n] and hj [n] are obtained, M can be obtained.
 5

L
For an arbitrary p (x; θ), we can always find the positive number L, such that p (x; θ) ≈ 0 for |x| > 2.

This may be done directly from the expression of p (x; θ). However, since we are considering the case
when p (x; θ) has no closed form, it is necessary to be able to determine L from φX (ω; θ). To do so
note that the moments of X can be obtained by [8]
1 dn φX (ω) 

n
E [X ] = n .
j dω n ω=0
From the moments of X and φX (ω; θ), L can be found using probability bounds. In the example of the
next section, the Chernoff bound is used to determine L [9]. The steps to evaluate [I (θ)]ij via the CF
φX (ω; θ) are summarized below,
L
1) From φX (ω; θ), obtain L such that p (x; θ) ≈ 0 for |x| > 2, by using a bound.
2) Select an AR model order p.
3) Calculate g[0], g[1], · · · , g[p] from (11).
4) Determine a[1], a[2], · · · a[p], σu2 by the Yule-Walker method.
5) Obtain the expression of hi [n] and hj [n] from (14).
6) Select M from the expression of hi [n] and hj [n].
7) Evaluate I(θ) from (16).
We note that in the above procedure the PDF p (x; θ) need not be known.

III. CRLB C OMPUTATION FOR THE K- DISTRIBUTION

The K-distribution is widely used to model sea surface clutter in high-resolution radar systems and to
model sea floor reverberation in high-resolution active sonar systems [4], [10], [11], [12].
Let X = Xr + jXi be a zero mean, circularly symmetric complex Gaussian random variable, i.e.
[Xr , Xi ]T ∼ N 0, σ 2 I , where I is a 2 × 2 identity matrix. Let Y be a Gamma random variable with

the shape parameter being ν + 1 and the scale parameter being 1/2, or

1 ν+1
 ( 2 ) y ν exp − 1 y
 
y≥0

Γ(ν+1) 2
pY (y) = (17)
 0

y<0
√
where Γ (·) denotes the Gamma function and ν > −1. Then Y X is distributed according to a complex
K-distribution [12]. Denote this variable as R = Z + jT , where Z is the real part of R and T is the
imaginary part. The joint CF of (Z, T ) is given by [10]
h i
φZ,T (ω, γ) = E ejωZ+jγT
1
= . (18)
(1 + σ 2 (ω 2 + γ 2 ))ν+1
 6

Hence, the CF of Z is given by

1
φZ (ω) = φZ,T (ω, 0) = (19)
(1 + σ 2 ω 2 )ν+1
and the PDF of Z can be shown to be
ν+1/2
1 |z| |z|
  
pZ (z) = √ Kν+1/2 (20)
πσ 2 Γ (ν + 1) 2σ σ
where Kα (·) is the modified Bessel function of the second kind.
From (20) it is seen that σ is the scale parameter and ν is the shape parameter. The moments of Z
can be obtained from the CF as
1 ∂φZ (ω) 

E [Z] = =0 (21)
j ∂ω ω=0
and 
h
2
i ∂ 2 φZ (ω) 
E Z =− = 2 (ν + 1) σ 2 . (22)
∂ω 2 ω=0


Therefore the variance of Z is

h i
var (Z) = E Z 2 − E 2 [Z] = 2 (ν + 1) σ 2 .

It is seen that as ν decreases, the variance of Z decreases. When ν → −1, var (Z) → 0, which means
Z approaches the deterministic value 0.
In [10] an approximate CRLB for ν of the PDF of |R| is derived under the condition that σ 2 is known
and ν ≫ 0. In the rest of this paper, we will focus on computing the CRLBs for both ν and σ 2 of
pZ (z) without any additional constrain on the range of ν . Considering the expression of (20), it would
be a formidable task to evaluate [I (θ)]ij using (2). Therefore we evaluate [I (θ)]ij via the CF using the
method introduced in the last section.

A. Selection of L

L is determined from the Chernoff bound, which is given as [9]

prob (z ≥ τ ) ≤ exp (−sτ ) E [exp (sz)]

= exp (−sτ + log (E [exp (sz)])) (23)

for any s ≥ 0. With this bound, an L is found as

 √

 2 15 (ν + 1) σ for ν ≥ 4
L= √ (24)
 10 15σ

for ν < 4
The details are given in Appendix I.
In general, a simple procedure to determine L is to increase it, until the results converge.
 7

B. Selection of the AR Model Order p

It can be seen that from (20) as ν → −1 , pZ (z) becomes narrower, requiring p to be large. The
proper value of p is also determined by L. This is because as L increases, the PSD G(u; θ) becomes
narrower, requiring a larger p. For the selection of L given by (24), an empirical formula is obtained
from simulations, and is given as

 ⌈1200 exp (−7ν) +70⌉

for −1 < ν < 0.9
p=
 50

for ν ≥ 0.9

where ⌈·⌉ denotes “the smallest integer greater than”.

In general, a simple procedure to determine p is to increase it, until the results converge.

C. Expression of g[n] and hi [n]

Plugging (19) into (11) we have


1 
g[n] =

2 2 ν+1 
L (1 + σ ω ) ω= 2πn L

Plugging (19) into (14) yields

" #
∂ 1 
hσ2 [n] =

∂σ (1 + σ 2 ω 2 )ν+1
2


ω= 2πn
L

ω 2 (1 + ν) 
= (25)
(1 + σ 2 ω 2 )ν+2 

ω= 2πn
L

and
" #
∂ 1 
hν [n] =

∂ν (1 + σ 2 ω 2 )ν+1


ω= 2πn
L


ln 1 + σ 2 ω 2 
= (26)
(1 + σ 2 ω 2 )ν+1 

ω= 2πn
L

D. Selection of M

From the derivation in Appendix II,

M = min (M1 , M2 )

where & 1 '

L 1+ν
 
2ν+2 2ν+4
− 2ν+2
M1 = σ .
2π 10−4
 8

and  l m    2 
 min L L 1 2ν+1
2πσ 341.6 , 2πσ 10−4 ν≥0

M2 =
L √
l m
x −1<ν < 0


2πσ ν

where xν is the greatest root of the equation

log (1 + x)
= 10−4
(1 + x)1+ν
and it can be easily found by a numerical method.
In general, a simple procedure to determine M is to increase it, until the results converge.

E. A Corroboration Test

Now we have everything to calculate I (θ) from (16). As a corroboration test of the proposed method,
a special case in which ν = 0 is considered first. When ν = 0, it is readily shown that
1 |z|
 
pZ (z) = exp − . (27)
2σ σ
After some simple algebra, we have

!2 
∂ ln p x; σ 2 1
[I (θ)]σ2 σ2 =E = . (28)
∂σ 2 4σ4

The theoretical value given by (28) and the value computed from the proposed method are shown in
Fig. 1 as circles and stars respectively. It can be seen that the match is very good and the results are
precise enough for most applications. However, if higher precision is required, we can always increase
L, p and M .

F. CRLB Evaluation of ν and σ 2 for the K-PDF of (20)

Even though there are two parameters in pZ (z), it is shown in Appendix III that the CRLBν is
independent of σ 2 , and for a fixed ν , CRLBσ2 is a linear function of σ 4 . This is because σ is a scale
parameter of pZ (z). Therefore, there is no need to plot the 2 dimensional figures of CRLBσ2 and
CRLBν versus σ 2 and ν . In Fig. 2 the computed CRLBν is plotted versus ν and in Fig. 3 the computed
CRLBσ2
σ4 is plotted versus ν . In both figures the computed values are connected with straight line for
better viewing.
 9

ν=0
7
Theoretic
Calculated
6

4
[I(θ)]σ2σ2

0
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
2
σ

Fig. 1. Theoretical and computed [I(θ)]σ2 σ2 versus σ 2 for the K-PDF of (20) with ν = 0.

1800

1600

1400

1200

1000
CRLBν

800

600

400

200

0
0 0.5 1 1.5 2 2.5 3
ν

Fig. 2. Evaluated CRLBν versus ν for all σ 2

IV. C ONCLUSION

We have proposed a method to evaluate the CRLB via the CF. When the PDF has no closed form,
it can be very hard or impossible to evaluate the CRLB directly. With the proposed method, the CRLB
 10

120

100

80
CRLBσ2/σ4

60

40

20

0
0 0.5 1 1.5 2 2.5 3
ν

CRLBσ2
Fig. 3. Evaluated σ4
versus ν

can be accurately evaluated when the CF is known. As an example, the CRLBs of the scale parameter
and the shape parameter of the K-distribution have been computed with the proposed approach. It has
also been shown that for a PDF with a scale parameter, the CRLB for the remaining parameters do not
depend on the scale parameter.

A PPENDIX I
S ELECTION OF L

From (19), the moment generation function of Z is given as

1
E [ζ z ] = .
(1 − σ 2 ζ 2 )υ+1
Letting ζ = s and minimizing logarithm of (23)
 
F (s) = −sτ − (ν + 1) log 1 − σ 2 s2

with respect to s, we have

 q 2 
2 2
σ 2 (ν + 1) + τ2 − σ (ν + 1)
s
τ2  
F (ŝ) = − (ν + 1)2 + + ν + 1 − (ν + 1) log 1 − .
 
σ 2  τ2 
 11

q
If we let τ = 15 (ν + 1)2 σ 2 then
2 2
    
F (ŝ) = −3 − log − 3 + log ν.
5 5
It can be shown that for ν < 4

prob ( z ≥ τ | ν < 4) ≤ prob ( z ≥ τ | ν = 4) ≤ exp (F (ŝ; ν = 4)) < 2.987 × 10−5 .

Hence if we choose
√

 15 (ν + 1) σ

for ν ≥ 4
τ= √ (29)
 5 15σ

for ν < 4

then prob (z ≥ τ ) < 2.987 × 10−5 . Since pZ (z) is symmetric we have

prob (z ≤ −τ ) < 2.987 × 10−5 .

L
Letting L = 2τ , then pZ (z) ≈ 0 for |z| > 2.

A PPENDIX II
S ELECTION OF M

Differentiating (19) with respect to σ 2 and taking the absolute value yield
 d ω 2 (1 + ν)
 

 dσ 2 φX (ω) =
 
(1 + σ 2 ω 2 )ν+2
ω 2 (1 + ν)
<
(σ 2 ω 2 )ν+2
(1 + ν)
= . (30)
(σ )ν+2 ω 2ν+2
2
 
Hence, we have  dσd 2 φX (ω) < 10−4 , if
 

1 1
1+ν 1+ν
   
2ν+2 2ν+2 2ν+4
ω> 2ν+4
= σ − 2ν+2 .
σ 10−4 10−4
Equivalently, |hσ2 [n]| < 10−4 for all n ≥ M1 , if
& 1 '
L 1+ν
 
2ν+2 2ν+4
− 2ν+2
M1 = σ .
2π 10−4
 12

Differentiating (19) with respect to ν and taking the absolute value yield
ln 1 + σ 2 ω 2


 d
 

 φX (ω) =
(1 + σ 2 ω 2 )ν+1
 dν 

ln 1 + σ 2 ω 2


<
(σ 2 ω 2 )ν+1

1/2
σ2ω2
≤
(σ 2 )ν+2 ω 2ν+2
1
= . (31)
(σ 2 ω 2 )ν+1/2
 
d
If ν > − 21 then  dν φX (ω) < 10−4 for
 

2
1 1
 
2ν+1
ω≥ .
σ 10−4
Or equivalently |hν [n]| < 10−4 for all n ≥ M2 , if
& 2 '
L 1
 
2ν+1
M2 ≥ .
2πσ 10−4
However, when ν → 0, M2 could be very large and this bound is not tight enough. Therefore, another
bound is derived in the follows.
For ν > 0
log 1 + σ 2 ω 2 log 1 + σ 2 ω 2



ν+1 <
(1 + σ 2 ω 2 ) (1 + σ 2 ω 2 )
log(1+x)
solving the equation 1+x = 10−4 we have x = 1.1667 × 105 , so
log 1 + σ 2 ω 2


< 10−4
(1 + σ 2 ω 2 )ν+1
if ω ≥ σ1 341.6. Or equivalently |hν [n]| < 10−4 for all n ≥ M2 , if
L
 
M2 = 341.6 .
2πσ
Therefore for ν > 0 we have
 & 2 '!
L L 1
  
2ν+1
M2 = min 341.6 , .
2πσ 2πσ 10−4
For −1 < ν ≤ 0, the equation
log (1 + x)
= 10−4
(1 + x)1+ν
is first solved numerically. Let xν be the greatest root, then
log 1 + σ 2 ω 2


−4
2 2 ν+1 < 10
(1 + σ ω )
 13

√
xν
if ω > σ . Or equivalently |hν [n]| < 10−4 for all n ≥ M2 , if
L √
 
M2 = xν .
2πσ
So finally,  l m    2 
 min L L 1 2ν+1
2πσ 341.6 , 2πσ 10−4 ν≥0

M2 =
L √
l m
x −1<ν < 0


2πσ ν

and
M = min (M1 , M2 ) .

A PPENDIX III
I NDEPENDENCE OF THE CRLB OF A S CALE PARAMETER

Theorem 1: Let pZ (z; α, λ) be a PDF parameterized by α and λ. If pZ (z; α, λ) satisfies the “regularity
condition” and λ is a positive scale parameter, or pZ (z; α, λ) = λf (λz; α), then CRLBα is independent
of the value of λ, and for a fixed α, CRLBλ is a linear function of λ2 .
We note that the CRLB for α will be smaller when λ is known. Also, in the Theorem 1 there is only
one remaining parameter, i.e. α. When there are more than one remaining parameters the same conclusion
still holds, and it can be proved similarly.
Proof: By definition, the (1, 1) element of the Fisher information matrix is
" 2 #
∂ ln (pZ (z; α, λ))
Iαα = E
∂α
∞ 2
∂ (pZ (z; α, λ)) 1
Z 
= dz
−∞ ∂α pZ (z; α, λ)
∞ 2
∂λf (λz; α) 1
Z 
= dz. (32)
−∞ ∂α λf (λz; α)
It is assumed that the range of the PDF is from −∞ to ∞. For PDFs with other ranges, the theorem
can be proved similarly. In the rest of the proof the limits of the integral are omitted for simplicity. Let
x = λz , after straightforward algebra, we have from (32)
2
∂f (x; α) 1
Z 
Iαα = dx. (33)
∂α f (x; α)
This is a function of α only, so it can be denoted as g1 (α) .
 14

By definition
2
∂ (pZ (z; α, λ)) 1
Z 
Iλλ = dz
∂λ pZ (z; α, λ)
2
∂λf (λz; α) 1
Z 
= dz
∂λ λf (λz; α)
2
∂f (u; α)  1
Z  
= f (λz; α) + λz dz. (34)
∂u u=λz λf (λz; α)
Let x = λz , after straightforward algebra, we have from (34)
2
∂f (u; α)  1
Z  
Iλλ = f (x; α) + x dx. (35)
∂u u=x λ2 f (x; α)
1
This can be denoted as λ2 g2 (α), where g2 (α) is a function of α only.
Similarly
∂ (pZ (z; α, λ)) ∂ (pZ (z; α, λ)) 1
Z   
Iαλ = dz
∂α ∂λ pZ (z; α, λ)
∂ (λf (λz; α)) ∂f (u; α)  1
Z    
= f (λz; α) + λz dz. (36)
∂α ∂u 
u=λz λf (λz; α)

Let x = λz , after straightforward algebra, we have from (36)

∂ (f (x; α)) ∂f (u; α)  1
Z    
Iαλ = f (x; α) + x dx. (37)
∂α ∂u 
u=x λf (x; α)
1
This can be denoted as λ g3 (α), where g3 (α) is a function of α only.
Therefore, the Fisher information matrix can be expressed as
 
1
 g1 (α) λ g3 (α)
I (α, λ) =  .

1 1
λ g3 (α) λ2 g2 (α)

As a result,  
1  g2 (α) −λg3 (α) 
I−1 (α, λ) = 2 .
g1 (α) g2 (α) − g3 (α) −λg3 (α) λ2 g1 (α)


This expression proves the theorem.

Considering the K-distribution PDF given by (20), according to Theorem 1, CRLBν is independent
1 1
of σ. Hence CRLBν is independent of σ 2 . Also CRLB 1 is a linear function of σ2 . After parameter
σ

transformation it can be shown [1], CRLBσ2 is a linear function of σ 4 .

 15

R EFERENCES

[1] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation Theory, Englewood Cliffs, NJ: Prentice-Hall, 1993.
[2] R. J. McAulay and E. M. Hofstetter, “Barankin Bounds on Parameter Estimation,” IEEE Trans. Inform. Theory, vol. 17,
pp. 669–676, Nov. 1971.
[3] M. G. Kendall and A. Stuart, The Advanced Theory of Statistics, vol. 2, Macmillan, New York: 1979.
[4] E. Jakeman and P. N. Pusey, “A Model for Non-Rayleigh Sea Echo,” IEEE Trans. Antennas Propagation, vol. Ap-24, pp.
806–814, Nov. 1976.
[5] M. Shao and C. L. Nikias, “Signal processing with fractional lower order moments: stable processes and their applications,”
Proc. IEEE, vol. 81, pp. 986–1010, July. 1993.
[6] D. Middleton, “Canonical and Quasi-Canonical Probability Models of Class A Interference,” IEEE Trans. Electromagnetic
Compat., EMC-25, pp. 76–106, May. 1983.
[7] S. M. Kay, Modern Spectral Estimation: Theory and Application, Englewood Cliffs, NJ: Prentice-Hall, 1988.
[8] S. M. Kay, Intuitive Probability and Random Processes using MATLAB, New York: Springer, 2006.
[9] H. L. V. Trees, Detection Estimation and Modulation Theory, Part I, New York: Wiley, 1968.
[10] D. A. Abraham and A. P. Lyons, “Novel physical interpretations of K-distributed reverberation,” IEEE Journal of Oceanic
Engineering, vol. 27, no.4, pp. 800–813, Oct. 2002.
[11] A. P. Lyons and D. A. Abraham, “Statistical characterization of high frequency shallow-water sea-floor backscatter,” J.
Acoust. Soc. Amer., vol. 106, no 3, pp. 1307–1315, 1999.
[12] K. D. Ward, “Compound representation of high resolution sea clutter,” Electron. Lett., vol. 17, no.16, pp. 561–563, Oct.
2002.