CRAMER-RAO

BOUND ANALYSIS ON MULTIPLE SCATTERING IN MULTISTATIC

POINT SCATTERER ESTIMATION
Gang Shi and Arye Nehorai
Department of Electrical and Systems Engineering
Washington University in St. Louis
St. Louis, MO 63130
ABSTRACT
The resolution improvements of time reversal methods through exploiting nonhomogeneous media have attracted much interest recently with broad applications, including the destruction of kidney
stones, underwater acoustics, radar, detection of defects in metals,
communications, and mine detection. In this paper, we analyze the
effect of inhomogeneity generated by multiple scattering among
point scatterers under a multistatic sensing setup. We derive the
Cramer-Rao bounds (CRBs) on parameters of the scatterers and
compare the CRBs for multiple scattering using the Foldy-Lax
model with the reference case without multiple scattering using
the Born approximation. We nd that multiple scattering could
signicantly improve the estimation performance of the system.
For the case where multiple scattering is not possible, e.g., where
only a single target scatterer exists in the illuminated scenario, we
propose the use of articial scatterers, which could effectively improve the estimation performance of the target despite a decrease
in the degrees of freedom of the estimation problem due to the introduced unknown parameters of the articial scatterers. Numerical examples demonstrate the advantage of the articial scatterer.
1. INTRODUCTION
The time reversal approach and its super resolution [1] have attracted increasing interest recently with broad applications, including the destruction of kidney stones, underwater acoustics, radar,
detection of defects in metals, communications, and mine detection. The idea behind so-called physical time reversal is to record
a signal emitted by sources or reected by targets using an array of
transducers and then transmit the time-reversed and complex conjugated version of the measurements back into the medium. In a
reciprocal medium, the back-propagated wave will then retrace the
original trajectory and focus around the original source locations.
Experimental and theoretical evidence shows [2], [3] that the refocusing in a non-homogeneous or random medium is much tighter
than in homogeneous, which is referred to as the super-resolution
of the time reversal. This super-resolution is intuitively interpreted
as taking advantage of the inhomogeneity to distribute the wave
over a larger part of the medium and therefore carry more information about the source, and it is quantitatively measured as the
improved effective aperture [4], which corresponds to the equivalent aperture that produces the same refocusing resolution in a
reference homogeneous medium.
This work was supported by the DARPA/AFOSR grant FA9550-04-10187 and AFOSR grant FA9550-05-1-0018.

142440469X/06/$20.002006IEEE

Encouraged by results in the situation of inhomogeneity, we

investigate possible advantages of multiple scattering in point scattering estimation, in which the inhomogeneity is induced by interactions of the scatterers. Multiple scattering exists in many physical systems involving wave propagation, including electrons, ultrasound, electromagnetic, and seismic waves, and can be analyzed
using very much the same concepts [5]. Though modeling and understanding of multiple scattering has been of interest in various
domains ranging from solid-state physics to optics to seismology
[6], it is still widely ignored in the signal processing literature, to
the best of the authors knowledge. Recently, MUSIC and maximum likelihood (ML) algorithms for estimating the locations of
point scatterers with multiple scattering were proposed in [7]-[10].
Cramer-Rao Bounds (CRBs) are computed in [11] to evaluate the
performance of reectivity estimation in an unknown environment
using physical time reversal. In this paper, we continue our work
in [9], [10] by evaluating the effect of multiple scattering on the
performance of point scatter estimation. We will evaluate multiple
scattering under an active multistatic sensing setup without physical time reversal, and measure the performance in terms of the
well-known performance benchmark CRB.
This paper is organized as follows: In Section 2, we present
two physical models used in later comparisons: one takes into account the multiple scattering among the scatterers using the FoldyLax model [12], and the other ignores it by using the Born approximation [12], [13]. In Section 3, we compute the CRBs based on
these two physical models, and then present numerical evaluations
of the CRBs under randomized setups in Section 4. In Section
5, we propose the use of articial scatterers to improve the system performance and present conclusions in Section 6. Due to the
space limitation, interested readers are referred to the journal version of this paper [14] for more detailed discussions and proofs of
the results.
2. PHYSICAL AND MEASUREMENT MODELS
We consider a transmit antenna array of Nt isotropic point antennas centered at known positions 1 , 2 , . . . , Nt , and a receive
array of Nr elements at 1 , 2 , . . . , Nr . The so-called Nr Nt
multistatic response matrix [15], [16] K = [Kj,k ()], represented
in the frequency domain is a function of the frequency whose
element Kj,k () coincides with the received signal at the j-th
receive antenna due to an impulse excitation applied by the k-th
transmit antenna, where j = 1, 2, . . . , Nr and k = 1, 2, . . . , Nt .
We assume the scenario under probe is stationary during the period
of sensing and there is no direct link from any transmit antenna to

IV1117

ICASSP2006

the receive antennas, i.e., the measured elds at the receive array
are due solely to the scattering of the illuminated scenario. Since
the information about the scenario is fully embedded in the multistatic matrix, the inference on the probed scenario will be based
on the measured multistatic matrix directly. We consider the scenario that consists of M discrete point scatterers in a background
medium with known Green function G(r, r
, ). The unknown
locations and scattering potentials of the scatterers are denoted
by x1 , x2 , . . . , xM and 1 (), 2 (), . . . , M (), respectively.
The time harmonic Green function G(r, r
, ) of the background,
which represents the propagator at from location r
to r, satises the reduced wave equation [15]. In the rest of this paper,
we will drop the dependence on in all notations for the sake of
simplicity.
Adopting the Foldy-Lax multiple scattering equations [12], we
formulate the multistatic matrix in a closed matrix form [9], [10]
KFL (x, ) = Ar (x)[T 1 ( ) S(x)]1 ATt (x),

(1)

where T stands for a matrix transpose, x = [xT1 , xT2 , . . . , xTM ]T

R3M representing the unknown location parameters, = [1 , 2 ,
. . . , M ]T CM unknown scattering parameters, and T ( ) =
diag{ }. The matrix S(x) is dened as S(x) =

2
0
66 G(x2 , x1 )
66
..
.
64
G(xM 1 , x1 )
G(xM , x1 )

G(x1 , x2 )
0

..

.
0
G(xM , xM 1 )

3
G(x1 , xM )
G(x2 , xM ) 7
77
..
77 ,
.
G(xM 1 , xM )5
0

and Ar (x) = [g r (x1 ), g r (x2 ) g r (xM )], At (x) = [g t (x1 ),

g t (x2 ) g t (xM )]; the receive Green function vector g r (x
)
CNr as a function of arbitrary location x
R3 is g r (x
) =
[G( 1 , x
), G( 2 , x
), . . . , G( Nr , x
)]T and the transmit Green
function vector g t (x
) CNt is g t (x
) = [G(x
, 1 ), G(x
, 2 ),
. . . , G(x
, Nt )]T . Note that the closed-form matrix representation (1) is a function of the background Green function only.
Using the identity (I A)1 = I + A + A2 + , the
multistatic matrix (1) can be expanded into the power series as

by additive independent, identically distributed (i.i.d.) complex

circularly symmetric Gaussian noise, i.e.,
Y = K(x, ) + W,

(4)

where W is Nr Nt noise matrix whose elements vec(W )

CN (0, 2 INr Nt ) and vec() stacks the rst to the last columns of
the matrix one under another to form a long vector. It worth mentioning that though it is technically feasible to derive the CRBs
under a more general correlated noise model, the simple noise assumption is for gaining insight into the differences between the
two scattering processes with and without multiple scattering, respectively.

3. CRAMER-RAO
BOUND RESULTS
We start deriving the CRBs [17] from the expression for the Fisher
information matrix (FIM) in [18]. Reparameterizing the unknown
scattering parameters into real parameters = [Re{1 }, Im{1 },
. . . , Re{M }, Im{M }]T R2M , where Re{} and Im{} denote the real and imaginary parts of a complex number, respectively. Dene = [
T , xT ]T , and the FIM for is found as

iH h

nh vec K(x, )
2
Re
2

T
 H

2
= 2 Re D ()D() ,

I() =

io

vec K(x, )
T

(5)

transpose. The Jacobian mawhere H represents the conjugate

trix D() = vec K(x, ) / T could be partitioned as D() =


[D , Dx ], where D = vec K(x, ) / T and Dx =

T
vec K(x, ) /x . Following the same partition, the FIM is
partitioned accordingly as
I() =

T
I Ix
.
Ix
Ixx

(6)

The CRB matrix for location parameters x is

1
1 T
CRB(x; ) = (Ixx Ix
I
) ,

Ix

(7)

KFL (x, ) = Ar (x)T ( )ATt (x) + Ar (x)T ( )S(x)T ( )ATt (x)

+Ar (x)T ( )S(x)T ( )S(x)T ( )ATt (x) + . (2)

and the CRB matrix for scattering parameter is

This series form is actually a generalization of the Neumann series or Born series [12], [13] under the multistatic context, and its
convergence is guaranteed when the spectral radius of T ( )S(x)
is less than one. The leading term of (2), known as the Born approximation, represents the rst-order scattering, i.e., the scattering without taking into account the multiple scattering. The second
term represents the second-order scattering contribution, namely
the portion of the scattering that is reected by the scatterers exactly twice. The rest of the series are higher order scattering terms.
We will employ the Born-approximated model

For the case with multiple scattering, the Jacobian matrices

DFL and DFLx are found as

KB (x, ) = Ar (x)T ( )ATt (x)

(3)

T
1
1
CRB(
; x) = (I Ix
) .
Ixx Ix

DFL =

At (x)[IM T ( )S(x)]1

Ar (x)[IM T ( )S(x)]1 (1, i),

DFLx =
At (x)[T

(8)

At (x)[T
1

( ) S(x)]

( ) S(x)]

(9)

1n  Br (x)

Ar (x)[T 1 ( ) S(x)]1 C(x)

+Bt (x)  Ar (x)[T 1 ( ) S(x)]1 1Tn , (10)

where

as the reference model for studying the effect of multiple scattering

on the estimation performance. It is easy to see that the model (3)
is a special case of (1) when the S(x) matrix is set to be zero
matrix.
Before computing the CRBs, we further assume that Y , the
measured the multistatic matrix, deviates from the model K(x, )

IV1118

 stands for the Khatri-Rao product [19],

represents the Kronecker product [19],
i is the imaginary unit,
Bt (x) = [bt (x1 ), bt (x2 ) bt (xM )],
bt (xm ) = g t (xm )/xTm , m = 1, 2, . . . , M ,

 Br (x) = [br (x1 ), br (x2 ) br (xM )],

br (xm ) = g r (xm )/xTm , m = 1, 2, . . . , M ,
1n : n-dimension column vector with each element as 1,
n is the dimension of one location parameter xm ,
C(x) = [cT (x1 ), cT (x2 ) cT (xM )]T ,
c(xm ) = g(xm )/xT , m = 1, 2, . . . , M ,
g(xm ) = [G(x1 , xm ), G(x2 , xm ), . . . , G(xM , xm )]T ,
in which G(xm , xm )
0, m = 1, 2, . . . , M .
For the reference case with no multiple scattering, recalling
that multistatic model (3) is the special case of (1), we can nd the
Jacobian matrices DB and DBx by substituting S(x) as a zero
matrix into (9) and (10), which are
DB = At (x)  Ar (x) (1, i),

(11)

DBx = [At (x)T ( ) 1n ]  Br (x)

+Bt (x)  [Ar (x)T ( ) 1Tn ].

(12)

4. NUMERICAL COMPARISONS
We employ a homogeneous two-dimensional setup, i.e., n = 2, the
inhomogeneity is then due solely to the multiple scattering among
the scatterers. Dropping the unessential constant, we obtain back

ground Green function G(r, r
) = ei2|r r |/ / 2|r r
|/
under usual approximation [20]. Assume collocated transmit and
receive arrays are uniform linear arrays (ULAs) located between
(-20,0) and (20,0) with spacing of 5, i.e., Nt = Nr = 9. All the
coordinates are in the unit of wavelength, i.e., we assume the narrowband signal with = 1. Detailed discussions on the spatial
ambiguity of this sparse setup could be found in [10], whereas we
use the 5 antenna spacing here and [14] is solely for computational simplicity, it is easy to verify that all the results hold for
transmit and receive arrays with /2 spacing. Due to the space
limitation, more numerical examples are presented in the journal
version of this paper [14].
We computed the CRBs on the location and scattering parameters under randomized scatterer setup assuming that three scatterers (M = 3) are located randomly and uniformly over a rectangular area centered at (0, 30) with dimension 50 40. Moduli
of the scattering potentials were assumed to be independently and
uniformly distributed over [0.5, 1], and the phases uniformly distributed over [0, 2]. We ran 251 Monte Carlo runs and plot each
run as one point in Figure 1 with tr CRBB (x; )/ tr CRBFL (x; )
; x)/ tr CRBFL (
; x) as the y
as its x coordinate and tr CRBB (
coordinate, where tr represents the trace of a matrix. If we set
point (1, 1) as the coordinate origin, we can see in Figure 1 that all
the 251 Monte-Carlo runs are in the rst quadrant, which means
that multiple scattering improves the CRBs in the randomized setups. The median of tr CRBB (x; )/ tr CRBFL (x; ) is 7.1254,
; x)/ tr CRBFL (
; x). In addition, a
and 8.6120 for tr CRBB (
strong positive correlation between the two CRB ratios of x and
may be seen in Figure 1, which indicates that the estimation accuracy on the location parameters greatly affects that on the scattering parameters.
It worth mentioning that the favorable effect of multiple scattering on the estimation performance depends on the system setup.
For example, if we increase the scattering potentials until the interactions among the scatterers are large enough, the multiple scattering actually turns out to deteriorate the estimation performance.

Fig. 1. tr CRBB (
; x)/ tr CRBFL (
; x) versus tr CRBB (x; )
/ tr CRBFL (x; ) in the 251 Monte-Carlo runs.

Identifying conditions under which multiple scattering is benecial in terms of improving the CRBs is an interesting challenge,
which we will consider in our future work.
5. ARTIFICIAL SCATTERERS
One straightforward way of exploiting the advantage of the multiple scattering is simply to introduce it into the modeling. We further propose the use of articial scatterers [21] to create multiple
scattering. The articial scatterers can either be active or passive:
active articial scatterers could be relays that simply amplify and
retransmit the incident wave; passive articial scatterers could be
scatterers that reect efciently. The articial scatterers could be
deployed in the scenario of interest in a planned manner or happen to be nearby by opportunity. We use the following numerical
examples to demonstrate the efcacy of the articial scatterers assuming the location and scattering potentials of the deployed articial scatterers are unknown.
The array setup is the same as in Section 4 in which we consider one target scatterer and two articial scatterers. We compute
the traces of the CRBs on the target location x1 and scattering parameters 1 for the case without articial scatterers, denoted by
tr CRB1 (x1 ; 1 ) and tr CRB1 (
1 ; x1 ), respectively, where 1 =
[Re{1 }, Im{1 }]T , then compute the traces of the CRBs of the
target scatterer after deploying two articial scatterers, denoted by
1 ; x1 ). We plot 251 Montetr CRB1+2 (x1 ; 1 ) and tr CRB1+2 (
Carlo runs with tr CRB1 (x1 ; 1 )/ tr CRB1+2 (x1 ; 1 ) as x co 1 ; x1 )/ tr CRB1+2 (
1 ; x1 ) as y coordiordinates and tr CRB1 (
nates.
In Figure 2 the computation is performed based on Born approximation model (3), as expected all the runs appear in the third
quadrant which means that the added articial scatterers decrease
the estimation performance of the single scatterer in the case without multiple scattering. This is the direct consequence of the increased unknown parameters, thus reducing the degrees of freedom of the estimation problem. In Figure 3, the CRBs are computed using the Foldy-Lax model (1). It is interesting to see that
most of the Monte-Carlo runs lie in the rst quadrant in this case,
meaning that the two randomly deployed articial scatterers improved the estimation performance on the target scatterer via the

IV1119

created multiple scattering. The median of tr CRB1 (x1 ; 1 )

1 ; x1 )
/ tr CRB1+2 (x1 ; 1 ) is 6.6418, and 7.9206 for tr CRB1 (
1 ; x1 ). The multiple scattering surprisingly not only
/ tr CRB1+2 (
offsets the degradation of the CRBs on the target scatterers due the
lowered degrees of freedom, but also improves the estimation performance further in most of the randomized scatterer setups.

Fig. 2.
tr CRB1 (
1 ; x1 )/ tr CRB1+2 (
1 ; x1 ) versus
tr CRB1 (x1 ; 1 )/ tr CRB1+2 (x1 ; 1 ) in the absence of multiple
scattering.

Fig. 3.
tr CRB1 (
1 ; x1 )/ tr CRB1+2 (
1 ; x1 ) versus
tr CRB1 (x1 ; 1 )/ tr CRB1+2 (x1 ; 1 ) in the presence of multiple
scattering.

6. CONCLUSIONS
We derived the Cramer-Rao bounds on the location and scattering parameters of point scatterers under a multistatic sensing setup
for the cases in which multiple scattering either exist or not. We
demonstrated that the inhomogeneity induced by the multiple scattering could greatly improve the estimation performance in terms
of the CRBs. We then proposed the use of articial scatterers in
the absence of natural multiple scattering. Analytically comparing

the CRBs in terms of their trace, determinant, or any matrix norm

is challenging due to the nonlinear dependence of the CRB matrices on the system parameters. We will investigate this point in our
future work.
7. REFERENCES
[1] M. Fink, Time-reversed acoustics, Scientic American, Nov. 1999.
[2] M. Fink, D. Cassereau, A. Derode, C. Prada, P. Roux, M. Tanter, J.
L. Thomas, and F. Wu, Time-reversed acoustics, Rep. Progr. Phys.,
vol. 63, pp. 1933-1995, 2000.
[3] M. Fink and C. Prada. Acoustic time-reversal mirrors, topical review, Inverse Problems, 17:R1-R38, 2001.
[4] L. Borcea, G. Papanicolaou, and C. Tsogka, Theory and applications
of time reversal and interferometric imaging, Inverse Problems, vol.
19, pp. 5139-5164, 2003.
[5] A. Derode, A. Tourin, and M. Fink, Random multiple scattering of
ultrasound. I. Coherent and ballistic waves, Physical Review E, vol.
64, 036605, 2001.
[6] A. Derode, A. Tourin, and M. Fink, Random multiple scattering of
ultrasound. II. Is time reversal a self-averaging process? Physical
Review E, vol. 64, 036606, 2001.
[7] F. K. Gruber, E. A. Marengo, and A. J. Devaney, Time-reversal imaging with multiple signal classication considering multiple scattering
between the targets, J. Acoust. Soc. Am. 115, pp. 3042-3047, June
2004.
[8] A. J. Devaney, E. A. Marengo, and F. K. Gruber, Time-reversal based
imaging and inverse scattering of multiple scattering point targets, J.
Acoust. Soc. Am. , to appear.
[9] G. Shi, A. Nehorai, Maximum likelihood estimation of point scatterers for computational time-reversal imaging, Proc. 13th Annu. Workshop on Adaptive Sensor Array Processing, Lincoln Laboratory, MIT,
Lexington, MA, June 2005.
[10] G. Shi, A. Nehorai, Maximum likelihood estimation of point scatterers for computational time-reversal imaging, Communications in
Information and Systems, vol. 5, no. 2, pp. 227-256, 2005.
[11] R. Rangarajan and A. O. Hero III, Analysis of a multistatic adaptive target illumination and detection approach (MATILDA) to time
reversal imaging, IMA Hot Topics Workshop: Adaptive Sensing
and Multimode Data Inversion, June 27-30, 2004.
[12] J. H. Taylor, Scattering Theory, New York, Wiley, 1972.
[13] R. G. Newton, Scattering Theory of Waves and Particles, New York,
McGraw-Hill, 1966.
[14] G. Shi and A. Nehorai, Cramer-Rao Bound Analysis on Multiple
Scattering in Multistatic Point Scatterer Estimation, in preparation
for IEEE Trans. on Signal Processing.
[15] L. Borcea, C. Tsogka, G. Papanicolaou, and J. Berryman, Imaging
and time reversal in random media, Inverse Problems, vol. 18, pp.
1247-1279, 2002.
[16] A. J. Devaney, Super-resolution processing of multi-static data using
time reversal and MUSIC, J. Acoust. Soc. Am., to appear.
[17] P. Stoica and A. Nehorai, MUSIC, maximum likelihood and
Cramer-Rao bound, IEEE Trans. Acoust., Speech, Signal Processing,
vol. ASSP-37, pp. 720-741, May 1989.
[18] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation Theory, Englewood Cliffs, NJ: Prentice Hall, 1993.
[19] J. W. Brewer, Kronecker products and matrix calculus in system
theory, IEEE Transactions on Circuits and Systems, vol. cas-25, no.
9 pp. 772-781, Sep. 1978.
[20] P. Sheng, Introduction to Wave Scattering, Localization, and Mesoscopic Phenomena, San Diego: Academic Press, Inc., 1995.
[21] G. Shi, A. Nehorai, Macrocell multiple-input multiple-output system analysis, to appear in IEEE Trans. Wireless Communications.

IV1120