IEEE TRANSACTIONS ON SIGNAL PROCESSING 1

Reconstruction and Estimation of Overspread

Scattering Functions
Onur Oktay, G otz Pfander, Pavel Zheltov
Abstract
In many radar scenarios, the radar target or the medium is assumed to possess randomly varying
parts. The properties of a target are described by a random process known as the spreading function.
Its second order statistics under the WSSUS assumption are given by the scattering function. Recent
developments in the operator identication theory suggest a channel sounding procedure that allows to
determine the spreading function given complete statistical knowledge of the operator echo. We show that
in a continuous model it is indeed theoretically possible to identify a scattering function of an overspread
target given full statistics of a received echo from a single sounding by a custom weighted delta train.
Our results apply whenever the scattering function is supported on a set of area less than one. Absent
such complete statistics, we construct and analyze an estimator that can be used as a replacement of the
averaged periodogram estimator in case of poor geometry of the support set of the scattering function.
Index Terms
radar, channel estimation, spreading function, scattering function, time-varying operators
I. INTRODUCTION
In the classical delay-Doppler radar system, the echo y(t) that is reected from a target can be expressed
as a superposition of time-frequency shifts of the transmitted waveform x(t), that is,
y(t) =
__
(, )M

x(t) d d, (1)
O. Oktay, G. E. Pfander and P. Zheltov are with Jacobs University Bremen
Emails:{o.oktay, g.pfander, p.zheltov}@jacobs-university.de
G. E. Pfander and P. Zheltov acknowledge funding by the Germany Science Foundation (DFG) under Grant 50292 DFG PF-4,
Sampling Operators.
November 17, 2011 DRAFT
IEEE TRANSACTIONS ON SIGNAL PROCESSING 2
where T

x(t) = x(t ) is a time shift operator, M

x(t) = e
2it
x(t) is a frequency shift operator and
is the spreading function of the target. If the target has randomly varying components, then (, ) is
assumed to be a random process. It is often assumed that the radar environment satises the wide-sense
stationarity with uncorrelated scattering (WSSUS) assumption [1], [2]. In this case, (, ) satises the
relation
E
_
(, )

)
_
= (

) (

) C(, ), (2)
where C(t, ) is called the scattering function of the target.
A problem in radar is to determine C(t, ) from the echo y [3], [4]. A standard approach is to view
the scattering function C(t, ) as power spectral density of the 2-D stationary random process known
as the time-variant transfer function. This gives rise to the averaged periodogram estimator of C(t, ).
One of the signicant restrictions of this approach is the requirement to assume an underspread target,
that is, the support of C(t, ) must be contained in a rectangle of area less than one. Building up on the
recent results in operator identication [5], [6], [7], [8], [9], we show that it is the area of the support
of C(t, ) itself that matters, and not the area of the bounding box. This fact calls for reclassication of
the channels that would previously be considered overspread due to the non-rectangular geometry of the
set supp C(t, ).
In Section III we prove the following theorem.
Theorem 1: Let C(t, ) be the scattering function of a radar target. Let the support of C(t, ) have area
M < 1. There exists T > 0, a natural number J and a xed J-periodic sequence c
k
, such that we can
identify the scattering function C(t, ) from the received echo y(t) = Hx(t) where x(t) =

kZ
c
k

kT
(t)
is the sounding signal.
In the tradition of [10], we provide an explicit formula (12) for the reconstruction of the scattering
function that is supported on a possibly overspread domain of the time-frequency plane using a specially
constructed delta train as the sounding signal. A procedure for the estimation of the scattering function
is naturally derived from this reconstruction formula.
The paper is organized as follows. Section II gives an overview of classical delay-Doppler radar. In
Section III we describe the target identication problem and prove the main result, Theorem 1. We suggest
and analyze an estimator in Section IV, followed by our conclusions in Section V.
II. OVERVIEW OF RADAR
The classical scenario in a delay-Doppler radar system is that a testing signal x(t) is transmitted. The
signal x(t) might be a short pulse, as well as a wideband linear chirp, coded waveform, pseudonoise
November 17, 2011 DRAFT
IEEE TRANSACTIONS ON SIGNAL PROCESSING 3
sequence, etc. [11], [12].
The echo from a point target at distance d and traveling at constant speed v, is considered to be of
the form
y =
0
x(t t
0
) e
2it0
,
where t
0
= 2d/c, c is the speed of light,
0
is the Doppler shift , and
0
is the reection coefcient that
depends on the distance and the speed of the point target.
As is customary and indicated in (1), we model the echo realized in a complete environment as a
continuous superposition of time-frequency shifts of the transmitted signal x(t) [3]. Equivalently to (1),
from the time-varying impulse response
h(t, ) =
_
(, ) e
2it
d (3)
we obtain the linear time-varying representation of the channel
y(t) =
_
h(t, )x(t )d.
Certain characteristic features of a target or a radar scene, for example, sea clutter [13], [2], [14], [15],
are often modelled as random processes. In this model, h and are assumed to be a pair of random
processes related via the one-sided Fourier transform (3).
Let the quantity
R
h
(t, s; , ) = E{h(t, )

h(s, )}
denote the autocorrelation function of h, where E{} denotes expected value. The usual WSSUS assump-
tion in radar applications is that h(t, ) is a) zero-mean in both variables, b) wide sense stationary (WSS)
in t, and c) has uncorrelated scatterers , which reads
R
h
(t, t + ; , ) = P
h
(, )( ).
Then the scattering function C(t, ) dened in (2) satises
C(, ) =
_
P
h
(t, ) e
2it
dt.
C(t, ) is the previously mentioned scattering function associated with the channel, and P
h
(t, ) is the
channels autocorrelation function, or ACF. The computations above and in the remainder of this paper
involve delta distributions, thus the convergence and equality of the integrals must be understood in a
weak sense. We preserve current notation for clarity and relegate a rigorous treatment of this matter to
[16].
November 17, 2011 DRAFT
IEEE TRANSACTIONS ON SIGNAL PROCESSING 4
In radar applications, it is usually assumed that the scattering function C(t, ) of a target has its support
on a time-frequency rectangle
[0, ] [/2, /2].
The degree of dispersion of the echo is quantied by the spreading factor . A target is said to be
underspread if < 1 and overspread if > 1. In recent years it has become increasingly clear
that the correct criterion is that the area of the support of the scattering function must be less than one
[17]. We show that it is indeed correct, and require only that suppC(t, ) is compactly supported within
some rectangle [0, ] [/2, /2] without demanding < 1. A classical identication problem
in radar and communications is to estimate the scattering function of a given randomly varying target,
or, equivalently, an linear time variant random channel. The general approach is to transmit a signal
x(t), once or multiple times, and to construct an estimator

C(t, ) for the scattering function C(t, )
using the received echoes. Obtaining measurements in this way is referred to as channel sounding in
communications.
III. SCATTERING FUNCTION IDENTIFICATION
Let be the spreading function of a radar target such that supp C(t, ) [0, ] [/2, /2],
with possibly larger than 1. Let the set supp C(t, ) be covered by the translations of the prototype
rectangle with some J prime number, as follows:
supp C(t, )
J
_
j=1
R + (a
j
T, b
j
B)
where a
j
, b
j
Z index the cover, T = /, B = /J. This implies JBT = 1. Note that other
bounding boxes, for example, a rst quadrant box [0, ] [0, ], can be accommodated by a simple
translation; the case JBT < 1 can be padded with empty boxes, if necessary. The theory of Jordan
domain shows that for any compact set M of area less than one the prime number J and the covering
{a
j
, b
j
}
J
j=1
can always be found. Given , we dene patches of to be

j
(t, ) =
R
(t, ) (t + a
j
T, + b
j
B),
and their Fourier transforms
h
j
(x, t) =
_

j
(t, ) e
2ix
d.
We can now reconstruct
(t, ) =
J

j=1

j
(t a
j
T, b
j
B),
November 17, 2011 DRAFT
IEEE TRANSACTIONS ON SIGNAL PROCESSING 5
and
h(x, t) =
J

j=1
_

j
(t a
j
T, b
j
B) e
2ix
d
=
J

j=1
_

j
(t a
j
T, ) e
2i(+bjB)x
d
=
J

j=1
e
2ixbjB
_

j
(t a
j
T, ) e
2ix
d
=
J

j=1
e
2ixbjB
h
j
(x, t a
j
T).
For each t xed,
j
(, t) is supported on [0, B). Therefore, for any we have the orthonormal expansion

j
(t, ) =
1
B

jZ
__
B
0

j
(t, s) e
2is(+
n
B
)
ds
_
e
2i(+
n
B
)
=
1
B

jZ
h
j
_
+
n
B
, t
_
e
2i(+
n
B
)
.
(4)
In particular, setting =
j
(t) = t + a
j
T, we obtain
(t, ) =
J

j=1

j
(t a
j
T, b
j
B)
=
1
B

nZ
J

j=1
h
j
_
t +
n
B
, t a
j
T
_
e
2i(bjB)(t+
n
B
)
=
1
B

nZ
h
_
t +
n
B
, t
_
e
2i(t+
n
B
)
.
For the random
j
s, we have
E {
j
(t, )

(s, )}
= (t s + (a
j
t

)T) ( + (b
j

)B)
R
(t, )
R
(s, )C(t + a
j
T, + b
j
B)
= (t s + (a
j
t

)T) ( + (b
j

)B)
R
(t, )
R
(t + (a
j
t

)T, + (b
j

)B)
C(t + a
j
T, + b
j
B).
But this is zero unless a
j
= t

and b
j
=

, that is j = , in which case we have

E{
j
(t, )

j
(s, )} =
R
(t, )(t s)( )C (t + a
j
T, + b
j
B) .
In particular, dening C
j
(t, ) =
R
(t, )C(t + a
j
T, + b
j
B), we have the expansion
C(t, ) =
J

j=1
C
j
(t a
j
T, b
j
B). (5)
November 17, 2011 DRAFT
IEEE TRANSACTIONS ON SIGNAL PROCESSING 6
Similarly, for the autocorrelation of hs we have whenever j =
E{h
j
(, t)

(, s)} =
__
e
2i()
E{
j
(t, )

(s, )} d d = 0,
and, when j = ,
E{h
j
(, t)

h
j
(, s)} =
__
e
2i()
C
j
(t, )(t s)( ) d d
=
_
e
2i()
C
j
(t, ) d(t s)
= P
hj
( , t)(t s),
(6)
where P
hj
(, t) =
_
e
2i
C
j
(t, ) d. Moreover, as a result of (5), we have
P
h
(, t) =
_
e
2i
C(t, )d
=
J

j=1
_
e
2i
C
j
(t a
j
T, b
j
B)d
=
J

j=1
e
2ibjB
P
hj
(, t a
j
T),
(7)
Next, we derive a formula similar to (4) for the scattering functions C
j
(t, ).
Lemma 2: With
j
, h
j
, P
hj
and C
j
dened as above, we have
C
j
(t, ) =
1
B

nZ
e
2i(+n/B)
P
hj
_
+
n
B
, t
_
,
where is arbitrary.
Proof: By equations (4) and (6), we have
E{
j
(t, )

j
(s, )} =
1
B
2

n,mZ
e
2i(+
n
B

m
B
)
E
_
h
j
( +
n
B
, t)

h
j
(
m
B
, s)
_
=
1
B
2

n,mZ
e
2i(+
n
B

m
B
)
P
hj
_
+
n m
B
, t
_
(t s)
=
1
B
2

n,mZ
e
2i(+
n
B

m
B
)
P
hj
_
+
n
B
, t
_
(t s)
=
1
B

nZ
e
2i(+
n
B
)
P
hj
_
+
n
B
, t
_
(t s)( ).
Since E{
j
(t, )

j
(s, )} = (t s)( )C
j
(t, ), the result follows.
We are now ready to prove the main result of this paper.
November 17, 2011 DRAFT
IEEE TRANSACTIONS ON SIGNAL PROCESSING 7
Proof of Theorem 1: Let c
k
be a J-periodic complex sequence. Transmit x(t) =

kZ
c
k

kT
(t)
and observe echo y(t) = Hx(t). Also, dene
y
n
(t) = (Hx)(t + nT) = H
_

k
c
k

kT
_
(t + nT) =

kZ
c
k
h(t + nT, t + (n k)T), (8)
for 0 t < T and n Z. Clearly, if h(t, ) is a stochastic process, so are y
n
(t). In fact, for 0 t, < T
and n Z, we have
E{y
n
(t)

y()} =

k,lZ
c

k
c

E{h(t + nT, t + (n k)T)

h(, T)}
=

k,lZ
c

k
c

P
h
(nT + t , t + (n k)T)(t (n k + )T).
But for 0 t, < T,
(t (n k + )T) =
_

_
(t ), if n k + = 0,
0, if n k + = 0.
Thus
E{y
n
(t)

y()} =

kZ
c

k
c
kn
P
h
(nT, t + (n k)T)(t )
=

kZ
c

k+n
c
k
P
h
(nT, t kT)(t )
=
n
(t)(t ),
(9)
where we dened

n
(t) =

kZ
c

k+n
c
k
P
h
(nT, t kT).
Using (7) and (9), we obtain

r+mJ
(t) =

kZ
c

k+r+mJ
c
k
P
h
_
m
B
+ rT, t kT
_
=

kZ
c

k+r
c
k
J

j=1
e
2i(
m
B
+rT)bjB
P
hj
_
m
B
+ rT, t (k + a
j
)T
_
=
J

j=1
e
2irbj/J

kZ
c

k+r
c
k
P
hj
_
m
B
+ rT, t (k + a
j
)T
_
.
Recall JBT = 1 and denote for r = 1, . . . , J
S
r
(t, ) =
1
B

mZ
e
2iT(mJ+r)

mJ+r
(t).
November 17, 2011 DRAFT
IEEE TRANSACTIONS ON SIGNAL PROCESSING 8
By Lemma 2 we compute
S
r
(t, ) =
1
B

mZ
e
2iT(mJ+r)

mJ+r
(t)
=
J

j=1
e
2irbj/J

kZ
c

k+r
c
k
1
B

mZ
e
2iT(mJ+r)
P
hj
((mJ + r)T, t (k + a
j
)T)
=
J

j=1
e
2irbj/J

kZ
c

k+r
c
k
C
j
(t (k + a
j
)T, ).
(10)
But, C
j
(, ) 0 only if 0 < T. Then, C
j
(t kT a
j
T, ) = 0 if and only if
t T < (k + a
j
)T t.
Since 0 t < T, we must have that C
j
(t (k + a
j
)T, ) = 0 only if k + a
j
= 0. Therefore, by (10),
S
r
(t, ) =
J

j=1
e
2irbj/J
c

raj
c
aj
C
j
(t, ). (11)
Let {c
k,l
}
J1
k,l=0
be the Weyl-Heisenberg frame for C
J
, where
c
k,l
(r) = e
2irl/J
c
rk
, r = 1, . . . , J.
The coefcient sequence {c
k
}
J
k=0
can be chosen so that any J element subset of the so-called Gabor frame
vectors is linearly independent [18], a condition of the frame known as the Haar property. Moreover, the
selection procedure allows us to choose c to be unimodular. In fact, choosing the entries of c randomly
from a uniform distribution on a unit circle guarantees the Haar property will hold with probability one.
In particular, this means that a matrix of our interest, determined by the geometry of the support set of
C(t, ), is invertible. Lets denote by U the J J matrix whose j-th column is c
aj,bj
, that is,
U =
_
c
a1,b1

c
a2,b2

c
aj,bj

.
Then, (11) can be expressed in the matrix form as
[S
r
(t, )]
J
r=1
= U

[c
aj
C
j
(t, )]
J
j=1
.
so that we can rewrite the above equation as
[S
r
(t, )]
J
r=1
= U

D[C
j
(t, )]
J
j=1
with D C
JJ
a non-singular diagonal matrix with elements of c on the diagonal, that is, D
ij
=
c
aj

ij
. Let V = [V
jr
]
J
r,j=1
be the inverse of U

D
c
. We can now recover C
j
(t, ) pointwise from the
November 17, 2011 DRAFT
IEEE TRANSACTIONS ON SIGNAL PROCESSING 9
autocorrelation of the output signal with an explicit reconstruction formula derived from (8), (9) and (10),
that is, for any j [1, J] and t [0, T), [0, B)
C
j
(t, ) =
J

r=1
V
jr
S
r
(t, ) =
J

r=1
V
jr
1
B

mZ
e
2iT(mJ+r)

mJ+r
(t). (12)
where
n
(t)(t ) = E{y(t + nT)

y()} . This completes the proof of Theorem 1.

IV. SCATTERING FUNCTION ESTIMATION
Since h(t, ) is a stochastic process, so is the returned echo y(t). Although Theorem 1 enables us to
reconstruct the scattering function perfectly from the stochastic process y(t), in practice we do not know
the values of E{y
n
(t)

y()}. Secondly, the delta trains that we use for identication of the channel have
innite energy, hence, so do their echoes. To address the rst problem, we sound the channel repeatedly
with the same input signal x(t) =

kZ
c
k

kT
(t) to obtain y
(1)
(t), . . . , y
(L)
(t), L independent identically
distributed samples of the channel output. As the channel is WSSUS, repeated soundings do not pose an
insurmountable problem in general. Since the primary purpose of this paper is a proof of concept for the
reconstruction of the scattering function with non-rectangular support, we do not explore the effects of
time-gating the delta train.
Regarding the second obstacle of innite energy responses, we consider as data for our estimator the
functions
n
(t) dened in (9) for t, [0, T] as follows.

mJ+r
(t)(t ) =
1
L
L

l=1
y
(l)
(t (mJ + r)T)

y
(l)
().
Clearly, plugging the above into (12) this induces an unbiased estimator

C(t, ) of the scattering
function C(t, ).

C
j
(t, ) =
J

r=1
V
jr
1
B

mZ
e
2iT(mJ+r)

mJ+r
(t).
We need the following trivial observation. We renumber the delta train
x(t) =

k[J]
c
r
k

nZ
(t (nJ mJ + k r)
and observe
y(t (mJ + r)T)
=
__
(, ) e
2i(t(mJ+r)T)
x(t (mJ + r)T ) d d
=

k[J]
c
r
k

nZ
__
(, ) e
2i(t(mJ+r)T)
(t (mJ + r)T (nJ mJ k r)T ) d d
November 17, 2011 DRAFT
IEEE TRANSACTIONS ON SIGNAL PROCESSING 10
=

k[J]
c
r
k

nZ
_
(t (nJ k)T, ) e
2i(t((mJ+nJk+r)T)
d
=

k[J]
c
r
k
_
(t + kT, ) e
2i(t(mJk+r)T)
d,
where n now depends on k, m and r, and k, as well as the sequence c
r
k
, depend on r.
Let us pre-compute the covariance between

m1J+r
(t) and

m2J+r
(t). As different measurements y
(l)
and y
(l

)
are independent, it is enough to consider
E

m1J+r
(t
1
)

m2J+r
(t
2
)

(t
1
t
3
)(t
2
t
4
)
=
1
L
E
_
y(t
1
(m
1
J + r)T)

y(t
3
)y(t
2
(m
2
J + r)T)y(t
4
)

_
=
1
L

k1,k2,k3,k4[J]
c

k1
c
k2
c
k3
c

k4

n1,n2,n3,n4Z
____
E
_
(t
1
(n
1
J k
1
)T,
1
)

(t
2
(n
2
J k
2
)T,
2
)
(t
3
(n
3
J k
3
)T,
3
) (t
4
(n
4
J k
4
)T,
4
)

_
e
2i[(t1(m1J+n1Jk1+r)T)1+(t4(n4Jk4)T)4]
e
2i[(t3(n3Jk3)T)3+(t2(m2J+n2Jk2+r)T)2]
d
1
d
2
d
3
d
4
To estimate the variance, we assume that all (t, ) are jointly proper Gaussian processes. With the help
of the Isserlis moment theorem, we simplify the expectation
E
_
(t
1
(n
1
J k
1
)T,
1
)

(t
2
(n
2
J k
2
)T,
2
)
(t
3
(n
3
J k
3
)T,
3
) (t
4
(n
4
J k
4
)T,
4
)

_
= C
_
t
1
(n
1
J k
1
)T,
1
_

_
t
1
(n
1
J k
1
)T (t
2
(n
2
J k
2
)T)
_
(
1

2
)
C
_
t
3
(n
3
J k
3
)T,
3
_

_
t
3
(n
3
J k
3
)T (t
4
(n
4
J k
4
)T)
_
(
3

4
)
+ C
_
t
1
(n
1
J k
1
)T,
_

_
t
1
(n
1
J k
1
)T (t
3
(n
3
J k
3
)T)
_
(
1

3
)
C
_
t
2
(n
2
J k
2
)T,
2
_

_
t
2
(n
2
J k
2
)T (t
4
(n
4
J k
4
)T)
_
(
2

4
)
+ C
_
t
1
(n
1
J k
1
)T,
1
_

_
t
1
(n
1
J k
1
)T (t
4
(n
4
J k
4
)T)
_
(
1

4
)
C
_
t
2
(n
2
J k
2
)T,
2
_

_
t
2
(n
2
J k
2
)T (t
3
(n
3
J k
3
)T)
_
(
2

3
).
All three summands can be estimated similarly. We provide detail for the rst summand only. Con-
siderations of delta functions and support constraints of the scattering function cause cancellations. In
November 17, 2011 DRAFT
IEEE TRANSACTIONS ON SIGNAL PROCESSING 11
particular, we have n
1
= n
2
= 0, k
1
= k
2
, n
3
= n
4
= 0, k
3
= k
4
,
I
1
(t) =

k1,k2,k3,k4[J]
c

k1
c
k2
c
k3
c

k4

n1,n2,n3,n4Z
____
d
1
d
2
d
3
d
4
C
_
t
1
(n
1
J k
1
)T,
1
_

_
t
1
(n
1
J k
1
)T (t
2
(n
2
J k
2
)T)
_
(
1

2
)
C
_
t
3
(n
3
J k
3
)T,
3
_

_
t
3
(n
3
J k
3
)T (t
4
(n
4
J k
4
)T)
_
(
3

4
)
e
2i[(t1(m1J+n1Jk1+r)T)1+(t2(m2J+n2Jk2+r)T)2+(t3(n3Jk3)T)3(t4(n4Jk4)T)4]
=

k1,k3[J]
c

k1
c
k1
c
k3
c

k3

n1,n3Z
__
C
_
t
1
+ k
1
T,
1
_
e
2i(t1t2)1
C
_
t
3
+ k
3
T,
3
_
e
2i(m1m2)JT1
d
1
d
3
.
Hence, we have for the covariance
cov
_

m1J+r
(t),

m2J+r
(t)
_
= I
1
(t) + I
2
(t) + I
3
(t).
We can now estimate the variance on (t, ) [0, T)[0, B).

var

S
r
(t, )

var
1
B

m
e
2iT(mJ+r)

mJ+r
(t)

=
1
B
2

m1,m2
e
2iTJ(m1m2)
cov

m1J+r
(t)

m2J+r
(t)

=
1
B
2

m1,m2
e
2iTJ(m1m2)
(I
1
(t) + I
2
(t) + I
3
(t))

=
1
B
2

k1[J]
|c
k1
|
2
_
C
_
t + k
1
T,
1
_

m1
e
2i(1)TJm1
d
1

k3[J]
|c
k3
|
2
_
C
_
t + k
3
T,
3
_

m2
e
2i(3)TJm2
d
3
+

k1[J]
|c
k1
|
2
_
C
_
t + k
1
T,
1
_

m1
e
2i[(+1)m1JT+rT1]
d
1

k2[J]
|c
k2
|
2
_
C
_
t + k
2
T,
2
_

m2
e
2i[(+2)m2JT+rT2]
d
2
+

k1[J]
(c

k1
)
2
_
C
_
t + k
1
T,
1
_

m1
e
2i(21)TJm1
e
4i1(t+k1TrT)
d
1

k2[J]
c
2
k2
_
C
_
t + k
2
T,
2
_

m2
e
2i(2)TJm2
e
4i2(tk2TrT)
d
2

November 17, 2011 DRAFT

IEEE TRANSACTIONS ON SIGNAL PROCESSING 12
=
1
B
2

k[J]
|c
k
|
2
C(t + kT, )
_
2
+
_

k[J]
|c
k
|
2
C(t + kT, B )
_
2
+

k[J]
(c
k
)
2
_
C(t + kT, /2) + C(t + kT, (B + )/2)
_

J
B
2

k[J]
|C(t + kT, )|
2
+|C(t + kT, B )|
2
+ 2|C(t + kT, /2)|
2
+ 2|C(t + kT(B + )/2)|
2
.
This gives us a pointwise estimate. We now average to obtain the following estimate for the variance
of

C
j
,

var

C
j

=
1
BT
__

V var

S
r
(t, )V

dt d
4V
2
J
2
LB
2
C
2
2
.
This completes the bias-variance analysis.
V. CONCLUSION
In this paper, we have shown that it is possible to recover the scattering function of a target using the
second order statistics of the returned echoes from sounding by a custom weighted delta train, provided
that supp C(t, ) is contained in a union of rectangles of total area 1. This includes channels commonly
considered overspread under the classic criterion that the area of the bounding box is less than 1. In the
abstract setting of a continuous WSSUS channel sounded by an innite delta train, being given complete
second-order statistics of the received signal, we guarantee an exact recovery of the scattering function.
It is interesting to note that a single signal sufces for recovery of an entire collection of channels which
share the gridding of the support set of their scattering functions, but not the support itself. We suggest
that the proposed technique to reshufe the patches of the support of the scattering function can and
should be seen as a replacement of the Zak transform in the efcient implementation of the averaged
periodogram estimator procedure offered in [19].
For completeness, we give an explicit recipe of a simple unbiased estimator, with a predictable behavior
of the variance given repeated soundings under the assumption of pointwise jointly proper Gaussianity
of the spreading function. It would be interesting to explore the stability of the technique under the
time-gating of the input signal. We currently do not pursue this direction, since the purpose of this paper
is to establish the possibility of the scattering function reconstruction and to demonstrate the usefulness
of the weighted delta train technique in the eld of radar and sonar acquisition.
November 17, 2011 DRAFT
IEEE TRANSACTIONS ON SIGNAL PROCESSING 13
REFERENCES
[1] P. Bello, Characterization of randomly time-variant linear channels, Communications Systems, IEEE Transactions on,
vol. 11, no. 4, pp. 360 393, december 1963.
[2] H. Van Trees, Detection, Estimation, and Modulation Theory, vol.3. Wiley, New York, 2001.
[3] P. Green, Radar measurements of target scattering properties. New York, NY: McGraw-Hill, 1968.
[4] J. Harmon, Planetary delay-Doppler radar and the long-code method, Geoscience and Remote Sensing, IEEE Transactions
on, vol. 40, no. 9, pp. 1904 1916, sep 2002.
[5] G. E. Pfander and D. F. Walnut, Measurement of time-variant linear channels, Information Theory, IEEE Transactions
on, vol. 52, no. 11, pp. 4808 4820, nov. 2006.
[6] W. Kozek and G. E. Pfander, Identication of operators with bandlimited symbols, SIAM J. Math. Anal., vol. 37, no. 3,
pp. 867888, 2005. [Online]. Available: http://dx.doi.org/10.1137/S0036141003433437
[7] G. Pfander, H. Rauhut, and J. Tanner, Identication of matrices having a sparse representation, IEEE Trans. Signal
Process., vol. 56, no. 11, pp. 53765388, 2008. [Online]. Available: http://dx.doi.org/10.1109/TSP.2008.928503
[8] G. E. Pfander, Measurement of time-varying multiple-input multiple-output channels, Appl. Comput. Harmon. Anal.,
vol. 24, no. 3, pp. 4393 401, 2008.
[9] N. Grip and G. Pfander, A discrete model for the efcient analysis of time-varying narrowband communications channels,
Multidimens. Syst. Signal Process., vol. 19, no. 1, pp. 340, 2008.
[10] G. E. Pfander and D. F. Walnut, Sampling and reconstruction of operators, Preprint.
[11] M. Skolnik, Introduction to Radar Systems. New York: McGraw-Hill Book Company, 1980.
[12] J. Taylor, Ultra-Wideband Radar Technology. Boca Raton: CRC Press, 2001.
[13] N. Gaarder, Scattering function estimation, Information Theory, IEEE Transactions on, vol. 14, no. 5, pp. 684 693,
sep 1968.
[14] U. Doyuran and Y. Tanik, Multi-range and multi-pulse radar detection in correlated non-Gaussian clutter, in Microwave
Conference, 2007. European, oct. 2007, pp. 1469 1472.
[15] D. Fuhrmann, L. Boggio, J. Maschmeyer, and R. Chamberlain, Clutter scattering function estimation and ground moving
target detection from multiple STAP datacubes, in Acoustics, Speech, and Signal Processing, 2005. Proceedings. (ICASSP
05). IEEE International Conference on, vol. 5, march 2005, pp. v/593 v/596 Vol. 5.
[16] G. E. Pfander and P. Zheltov, Stochastic operator sampling, Preprint.
[17] G. Durisi, U. G. Schuster, H. B olcskei, and S. Shamai, Noncoherent capacity of underspread fading channels, IEEE
Trans. Inform. Theory, vol. 56, no. 1, pp. 367395, 2010. [Online]. Available: http://dx.doi.org/10.1109/TIT.2009.2034807
[18] F. Krahmer, G. E. Pfander, and P. Rashkov, Uncertainty in time-frequency representations on nite abelian
groups and applications, Appl. Comput. Harmon. Anal., vol. 25, no. 2, pp. 209225, 2008. [Online]. Available:
http://dx.doi.org/10.1016/j.acha.2007.09.008
[19] H. Artes, G. Matz, and F. Hlawatsch, Unbiased scattering function estimation during data transmission, in Vehicular
Technology Conference, 1999. VTC 1999 - Fall. IEEE VTS 50th, vol. 3, 1999, pp. 1535 1539 vol.3.
November 17, 2011 DRAFT