Time-reversal-based imaging and inverse scattering of multiply

scattering point targets

Anthony J. Devaney, Edwin A. Marengo, and Fred K. Gruber
Department of Electrical and Computer Engineering, Northeastern University, Boston, Massachusetts 02115
共Received 17 September 2004; revised 17 July 2005; accepted 1 August 2005兲
The treatment of time-reversal imaging of multiply scattering point targets developed by the present
authors in Gruber et al. 关“Time-reversal imaging with multiple signal classification considering
multiple scattering between the targets,” J. Acoust. Soc. Am., 115, 3042–3047 共2004兲兴 is
reformulated and extended to the estimation of the target scattering strengths using the Foldy–Lax
multiple scattering model. It is shown that the time-reversal multiple signal classification 共MUSIC兲
pseudospectrum computed using the background Green function as the steering vector yields
accurate estimates of the target locations, even in the presence of strong multiple scattering between
the targets, and that the target scattering strengths are readily computed from the so-determined
target locations using a nonlinear iterative algorithm. The paper includes computer simulations
illustrating the theory and algorithms presented in the paper. © 2005 Acoustical Society of
America. 关DOI: 10.1121/1.2042987兴
PACS number共s兲: 43.60.Pt, 43.35.Zc 关DRD兴 Pages: 3129–3138

I. INTRODUCTION In the present paper, the treatment of time-reversal MU-

Wave field imaging methods based on the decomposi-
tion of the time-reversal operator 共“D.O.R.T.”兲 technique,1–3
usually referred to as “time-reversal imaging,” have been of
interest for a number of years in connection with sonar, un-
derwater communications, and medical and nondestructive
testing applications.4 One of these methods is the so-called
time-reversal imaging with multiple signal classification
共time-reversal MUSIC兲 method developed by Devaney5 and
Lev-Ari and Devaney6 and elaborated further in Lehman and
Devaney,7 Prada and Thomas,8 and Gruber et al.9 A similar
method has been investigated by Miwa and Arai.10 Time-
reversal MUSIC, like classical time-reversal imaging as per-
formed using the D.O.R.T. method,1–3 is designed to locate
M point targets 共representing scatterers that are small relative
to the relevant wavelength兲 embedded in a rather arbitrary
background medium from knowledge of the multistatic data
matrix K of the collection of targets as measured by an active
array of N ⬎ M transceivers. Unlike the D.O.R.T. method,
time-reversal MUSIC enables the super-resolved location of
closely spaced targets5–8 and is ideally suited for interroga-
tion with sparse transceiver arrays.5
The vast majority of these developments have been es-
tablished within the linear framework of the distorted wave
Born approximation 共DWBA兲,11,12 which ignores multiple
scattering between the targets, although the experimental re-
sults of Prada and Thomas8 suggest the practicality of the
technique, even for multiply scattering targets. The latter
finding has been investigated theoretically in a more recent
paper,9 which shows how the time-reversal imaging with the
MUSIC method can be generalized to incorporate multiple
scattering between the targets. In that paper we employ the
exact scattering formalism of the Neumann series to show
that the same method actually holds, even if there is signifi-
cant multiple scattering between the targets.
SIC for the location of multiply scattering point targets de-
veloped by Gruber et al.9 is reformulated and extended to the
estimation of the target scattering strengths using the Foldy–
Lax multiple scattering model.13–15 It is shown that even in
the presence of strong multiple scattering, the time-reversal
MUSIC pseudospectrum computed using the background
Green function as a steering vector yields accurate estimates
of the target locations, and that the target scattering strengths
are readily computed from the so-determined target locations
using a nonlinear iterative algorithm. The paper includes
computer simulations illustrating the theory and algorithms
presented in the paper. Unlike the numerical examples in
Gruber et al.,9 which focus on comparing the first-order
共Born-approximated兲 versus the second-order 共quadratic兲
scattering models, the present examples consider the full
multiple scattering interactions between the targets as incor-
porated numerically via the Foldy–Lax equations.
II. FOLDY–LAX FORMULATION OF THE MULTIPLE
SCATTERING PROBLEM
Consider a system of N acoustic or electromagnetic
wave transceivers located at spatial positions ␣k, k
= 1 , 2 , . . . , N that interrogate a set of M point targets 共scatter-
ers兲 located at xm, m = 1 , 2 , . . . , M. In the frequency domain,
the field ␺k generated by the interaction of the incident field
␺共in兲
k generated by the kth transceiver element with the M
point targets is given by the equation
M
␺k共r, ␻兲 = ␺共in兲
k 共r, ␻兲 + 兺 ␶m共␻兲G0共r,xm, ␻兲␺k共xm, ␻兲,
m=1
共1兲
where ␶m共␻兲 is the scattering strength of the mth target and
G0 is the Green function corresponding to the background
medium in which the targets are embedded. This Green

J. Acoust. Soc. Am. 118 共5兲, November 2005 0001-4966/2005/118共5兲/3129/10/$22.50 © 2005 Acoustical Society of America 3129

Downloaded 05 Oct 2012 to 136.159.235.223. Redistribution subject to ASA license or copyright; see http://asadl.org/terms
function will be referred to as the background Green func-
tion. Equation 共1兲 is readily derived from the Lippman-
Schwinger equation11 in the special case of a set of M point
scatterers. From this point on, we will work exclusively in
the frequency domain and will no longer explicitly display
the frequency ␻, with the understanding that the results ap-
ply equally well to both narrow band and wide band appli-
cations.
Equation 共1兲 applies at all space points except at the
actual target locations r = xm, where the background Green
function is singular. At the target locations the “self-field”
contribution is absorbed into the definition of the target scat-
tering strengths ␶m and Eq. 共1兲 is replaced by the so-called
Foldy–Lax equations,13–15
␺k共xm兲 = ␺共in兲
k 共xm兲 + 兺 ␶m⬘G0共xm,xm⬘兲␺k共xm⬘兲.
m⬘⫽m
If the incident field ␺共in兲
k , target locations xm, m = 1 , . . . , M
and target scattering strengths ␶m, m = 1 , . . . , M are known,
then the Foldy–Lax equations are a set of M coupled lin-
ear equations that can be solved for the field amplitudes
␺k共xm兲 at the target locations xm. Indeed, Eq. 共2兲 can be
expressed in the matrix form
H⌿k = ⌿共in兲
k ,
where
⌿k = 关␺k共x1兲, ␺k共x2兲, . . . , ␺k共x M 兲兴T ,
⌿共in兲 共in兲 共in兲 共in兲
k = 关␺k 共x1兲, ␺k 共x2兲, . . . , ␺k 共x M 兲兴 ,
T
are M-dimensional column vectors 共where the superscript
letter T denotes the transpose兲 whose elements are the total
and incident field amplitudes, respectively, at the target loca-
tions, and where H is the M ⫻ M matrix whose diagonal
elements are unity and whose off-diagonal elements are
given by the product of the target scattering strengths with
the background Green function evaluated at the target loca-
tions; in particular,
Hm,m⬘ = ␦m,m⬘ − 共1 − ␦m,m⬘兲␶m⬘G0共xm,xm⬘兲.
The matrix equation, Eq. 共3兲, clearly has a unique solution
for any given incident wave and set of targets. Once the field
strengths ␺k共xm兲 are computed at the various target locations
the field can be evaluated at an arbitrary field point by using
Eq. 共1兲.
III. IMAGING AND INVERSE SCATTERING FROM THE
MULTISTATIC DATA MATRIX
The imaging problem for point targets can be roughly
defined to be that of forming an “image” of the distribution
of target scattering centers xm, m = 1 , 2 , . . . , M from measure-
ments of the fields ␺k共r兲, k = 1 , 2 , . . . , N generated in a suite
of N scattering experiments that employ the set of incident
fields ␺共in兲
k . The inverse scattering problem goes one step
beyond the imaging problem and has as its goal the quanti-
tative determination of both the target positions xm, m
= 1 , 2 , . . . , M and the target scattering strengths ␶m, m
= 1 , 2 , . . . , M from the available scattered field data. Of par-
3130 J. Acoust. Soc. Am., Vol. 118, No. 5, November 2005
Downloaded 05 Oct 2012 to 136.159.235.223. Redistribution subject to ASA license or copyright; see http://asadl.org/terms
ticular interest in this paper is the inverse scattering problem,
but aspects of the associated imaging problem will be dis-
cussed as well. It is assumed that one knows the background
Green function G0 of the medium in which the targets are
embedded. In the simplest case this is free space and G0 is
the well-known free space Green function, but the back-
ground medium can also be heterogeneous with possible
sharp boundaries and reflecting surfaces.
In this work it is assumed that the field measurements
are performed at the same set of N transceivers that are used
to interrogate the targets; thus a “coincident” array is as-
sumed. It is also assumed that the transceivers are all identi-
cal and small compared with the wavelength so that they can
be approximated as point elements. Although these assump-
tions are not necessary, they simplify the discussion and do
共2兲 not severely limit the generality of the development. Under
these conditions the incident waves ␺共in兲 k 共r兲, k = 1 , 2 , . . . , N,
due to unit-amplitude excitations, are equal to the back-
ground Green function G0共r , r⬘兲 evaluated for a source point
r⬘ = ␣k at a transmitter location and the measured outputs
from the set of transceivers are equal to the full Green func-
tion G共r , r⬘兲 of the combined background plus the target
medium for a source point r⬘ = ␣k at the kth transmitter loca-
tion and a field point r = ␣ j at the jth receiver location.
共3兲
Within the present formulation the inverse scattering prob-
lem then reduces to that of deducing the target positions xm
and scattering strengths ␶m from knowledge of the full Green
共4a兲
function specified at all transmitter/receiver pairs as well as
of the background Green function G0共r , r⬘兲 at all pairs of
共4b兲
points 共r , r⬘兲 within the background medium.
Under our assumption of point transceivers, Eqs. 共1兲 and
共2兲 become
M
G共r, ␣k兲 = G0共r, ␣k兲 + 兺 ␶mG0共r,xm兲G共xm, ␣k兲,
m=1
共6a兲
G共xm, ␣k兲 = G0共xm, ␣k兲 + 兺 ␶m⬘G0共xm,xm⬘兲G共xm⬘, ␣k兲,
m⬘⫽m
共5兲 共6b兲
where G共r , r⬘兲 is the full Green function of the combined
background plus target medium. The data for the inverse
scattering problem are the outputs from the various trans-
ceivers, which, as discussed above, are simply the full Green
function G共r , ␣k兲 evaluated at the transceiver locations r
= ␣ j, j = 1 , 2 , . . . , N. However, since the background Green
function G0 is assumed known one can also regard the data
for the inverse scattering problem to be the scattered field
component of this Green function, which is the second term
in Eq. 共6a兲. This component of the Green function when
evaluated between various transceiver elements is known as
the multistatic data matrix and is given by the expression
K j,k = G共␣ j, ␣k兲 − G0共␣ j, ␣k兲
M
= 兺 ␶mG0共␣ j,xm兲G共xm, ␣k兲.
m=1
共7兲
The inverse scattering problem then reduces to solving for
the set of unknown target locations xm, m = 1 , 2 , . . . , M and
Devaney, Marengo, and Gruber: Time-reversal-based inverse scattering
target scattering strengths ␶m, m = 1 , 2 , . . . , M given the back-
ground Green function G0, the measured multistatic data ma-
trix K, and the Foldy–Lax model, Eqs. 共6兲.
Up to this point, care has been exercised in preserving
the order of the two arguments of the background and the
full Green functions. However, in the following, attention
will be restricted to the usual case of reciprocal media,
where the background and full Green functions obey the
standard reciprocity conditions:
G0共r,r⬘兲 = G0共r⬘,r兲, G共r,r⬘兲 = G共r⬘,r兲.
Reciprocity of the Green functions allows us to employ a
standard form where the transceiver coordinates will always
be placed second in the arguments of the Green functions.
Thus, with this standard notation, Eq. 共7兲 defining the mul-
tistatic data matrix can be expressed as
M strong multiple scattering will exist between the various tar-
K j,k = 兺 ␶mG0共xm, ␣ j兲G共xm, ␣k兲.
m=1
Note for future reference that because of the reciprocity of
the two Green functions the multistatic data matrix K is also
symmetric, as can be verified readily using Eq. 共7兲.
IV. IMAGING AND INVERSE SCATTERING WITHIN THE
DISTORTED WAVE BORN APPROXIMATION
A standard approximation that is used in time-reversal
studies is the DWBA. The DWBA ignores all multiple scat-
tering between the various targets so that the Foldy–Lax
equations 共6兲 return the approximate solution,
G共xm, ␣k兲 ⬇ G0共xm, ␣k兲,
which, when used in Eq. 共8兲, yields the following approxi-
mate expression for the multistatic data matrix within the
DWBA:
M
K j,k ⬇ Kbj,k = 兺 ␶mG0共xm, ␣ j兲G0共xm, ␣k兲,
m=1
where we have denoted the DWBA approximation of the
multistatic data matrix by Kbj,k. Within the DWBA the full
Green function does not appear and the only unknowns are
the target locations xm , m = 1 , 2 , . . . M and the target scatter-
ing strengths ␶m.
The DWBA approximation, Eq. 共9兲, for the multistatic
data matrix can be expressed in the matrix form
M tion 共SVD兲 of the K matrix that we write in the form
Kb = 兺
m=1
␶mg0共xm兲gT0 共xm兲,
where
g0共x兲 = 关G0共x, ␣1兲,G0共x, ␣2兲, . . . ,G0共x, ␣N兲兴T
is the background Green function vector and Kb = 兵Kbj,k其 is the
DWBA approximation to the multistatic data matrix.
It was shown some time ago in a now classic paper1 共see
also Prada et al.2 and Mordant et al.3兲 that within the DWBA
the target locations xm and target scattering strengths ␶m can
be estimated from the multistatic data matrix as long as the
number of targets is less than or equal to the number of
J. Acoust. Soc. Am., Vol. 118, No. 5, November 2005
Downloaded 05 Oct 2012 to 136.159.235.223. Redistribution subject to ASA license or copyright; see http://asadl.org/terms
transceiver elements 共M 艋 N兲 and the targets are “well re-
solved.” This condition requires that the targets are suffi-
ciently separated from each other that they can be “resolved”
by the transceiver array.1,5 In the case of well-resolved tar-
gets it was shown that within the DWBA there exists a one-
to-one relationship between the individual targets and the
singular vectors and singular values of the multistatic data
matrix. Moreover, in this case the locations of the individual
targets can be determined by forming “time-reversal images”
from the singular vectors associated to nonzero singular val-
ues using the background Green function and the target scat-
tering strengths can be determined from the set of singular
values. Unfortunately, this imaging approach fails 共in the
sense of not yielding selective focusing兲 if the targets are not
“well resolved,” as will be the case, for example, if the tar-
gets are not well separated from each other. In this case
共8兲 gets and the DWBA will not even be valid.
A. Time-reversal MUSIC within the DWBA
The “well-resolved” requirement of the original time-
reversal imaging method can be removed if instead of form-
ing estimates of the target locations using time-reversal im-
aging a generalized MUSIC algorithm is employed.5–8 In this
approach the singular vectors and singular values of the mul-
tistatic data matrix are still employed as in standard time-
reversal imaging,1 but the singular vectors associated to zero
singular values are now used in the image formation process
rather than the singular vectors associated with nonzero sin-
gular values, as are used in standard time-reversal imaging.
As mentioned earlier, the latter method requires that M 艋 N
共the number of targets be less than or equal to the number of
transceiver elements兲 and that the targets be well resolved.
The time-reversal MUSIC scheme does not require that the
targets be well resolved, but it does require that the number
of targets be less than the number of transceiver elements
共9兲 共M ⬍ N兲. In fact, as also mentioned earlier, this holds, even
in the presence of multiple scattering, as shown in Prada and
Thomas8 and in a contribution coauthored by the current
authors.9 These early results are explained next from an al-
ternative perspective based on the Foldy–Lax model, includ-
ing the associated estimation of the target scattering ampli-
tudes that was not considered in those papers.
Both the D.O.R.T. method1–3 as well as the time-reversal
MUSIC method are based on the singular value decomposi-
共10兲 K v p = ␴ pu p, K †u p = ␴ pv p , 共12a兲
N
K = 兺 ␴ pu pv†p . 共12b兲
共11兲 p=1
It is clear from Eq. 共10兲 that within the DWBA the multi-
static data matrix K maps CN, the vector space of complex
N-tuples, to the subspace S0 = Span兵g0共xm兲 , m
= 1 , 2 , . . . , M其 債 CN spanned by the background Green func-
tion vectors evaluated at the scatterer locations xm. It was
shown by Devaney5 that except in certain special and un-
usual remote sensing geometries and as long as M 艋 N, the
Devaney, Marengo, and Gruber: Time-reversal-based inverse scattering 3131
background Green function vectors form a linearly indepen-
dent set so that 兵g0共xm兲 , m = 1 , 2 , . . . , M其 forms a basis for S0.
On the other hand, it follows from Eq. 共12b兲 that S0
= Span兵u p , ␴ p ⬎ 0其. Moreover, since S0 = Span兵u p , ␴ p
⬎ 0其 ⬜ N0 = Span兵u p , ␴ p = 0其 and 兵g0共xm兲 , m = 1 , 2 , . . . , M其 is
also a basis for S0, it then follows that
u†pg0共xm兲 = 0, ␴ p = 0. 共13兲
The locations of the scatterers can then be determined from
the time-reversal MUSIC pseudo-spectrum,
1
⌽共x兲 = , 共14a兲
兺␴ =0 兩u†pg0共x兲兩2
p
which will peak 共ideally to infinity兲 when x = xm , m
= 1 , 2 , . . . , M. Thus, the target 共scatterer兲 location within the
DWBA reduces to finding the maxima of the pseudospec-
trum defined in Eq. 共14a兲, generated using the background
Green function vector g0共x兲 as a steering vector.
Since the multistatic data matrix is symmetric, it follows
that v p = u*p, so that the pseudospectrum can also be written in
the alternative form
1 shown in Gruber et al.9 the time-reversal MUSIC algorithm
⌽共x兲 = . 共14b兲
兺␴ =0 兩v†pg*0共xm兲兩2
p determine the target locations xm. In Gruber et al.9 this con-
Note also that time-reversal MUSIC will fail if the “noise
subspace” N0 = Span兵u p , ␴ p = 0其 is empty. This will occur if
the rank of the K matrix is equal to the number of trans-
ceivers N that will occur if M 艌 N. Thus, generally, time-
reversal MUSIC requires that M ⬍ N.
It is worth emphasizing that the above analysis has been
based on the DWBA and should not be expected to be valid
in the presence of strong multiple scattering, where the
DWBA approximation, Eq. 共9兲, will fail. In this latter case
there is no reason to believe from the above development
that the pseudospectra in Eqs. 共14兲, when computed using the
singular system 兵u p , v p , ␴ p其 of the exact 共e.g., measured兲
multistatic data matrix as defined in Eq. 共8兲 together with the
background Green function vector as a steering vector, will
peak at the correct scatterer locations. Amazingly, as sug-
gested in Prada and Thomas8 and as shown in Gruber et al.9
and in the following section, these pseudospectra actually do
apply even when computed using exact scattering data.
B. Scattering strength computation within the
DWBA
Equation 共9兲 for the multistatic data matrix within the
DWBA can be regarded as a matrix equation relating the
unknown scattering coefficients ␶m to the observed values of
K j,k expressed as an N2 long column vector as follows:
共15兲
where 关K j,k兴 = 关K1,1 , K1,2 , . . . , KN,N兴T, 关␶m兴 = 关␶1 , ␶2 , . . . , ␶ M 兴T
and
3132 J. Acoust. Soc. Am., Vol. 118, No. 5, November 2005
Downloaded 05 Oct 2012 to 136.159.235.223. Redistribution subject to ASA license or copyright; see http://asadl.org/terms
关G0共xm, ␣ j兲G0共xm, ␣k兲兴
冤 冥
G0共x1, ␣1兲G0共x1, ␣1兲 ¯ G0共x M , ␣1兲G0共x M , ␣1兲
G0共x1, ␣1兲G0共x1, ␣2兲 ¯ G0共x M , ␣1兲G0共x M , ␣2兲
= .
] ] ]
G0共x1, ␣N兲G0共x1, ␣N兲 ¯ G0共x M , ␣N兲G0共x M , ␣N兲
Since the values of the target locations xm and background
Green functions are known, using, for example, the time-
reversal MUSIC algorithm described in the previous section,
a least squares solution of Eq. 共15兲 will yield the M ⬍ N
scattering coefficients ␶m. Because the DWBA model is only
approximate and because there will always be some noise
sources in the measurement process, the fact that Eq. 共15兲 is
overdetermined and requires a least squares solution is a de-
sirable property of the inversion process.
V. IMAGING AND INVERSE SCATTERING IN THE
PRESENCE OF MULTIPLE SCATTERING
In the case where multiple scattering between the point
targets becomes important, the DWBA can no longer be em-
ployed and the coupled set of Eqs.共6兲 must be employed in
the imaging and inverse scattering problems. Remarkably, as
established in the preceding section can still be employed to
clusion was reached within the exact scattering formalism of
the Neumann series using the fact that the multistatic data
matrix K is symmetric even in the presence of multiple scat-
tering between the targets. Here, a different approach is em-
ployed to establishing this result that has the advantage of
adding insight into the estimation process. This approach
also has the advantage that it can be generalized to cases of
non-coincident transceiver arrays for which the multistatic
data matrix is not symmetric so that the treatment employed
in Gruber et al.9 is not applicable.
A. Time-reversal MUSIC in the presence of multiple
scattering
The exact expression, Eq. 共8兲, for the multistatic data
matrix, including multiple scattering between the set of point
scatterers, can be expressed in the form
M
K= 兺
m=1
␶mg0共xm兲gT共xm兲, 共16兲
where g0 is the background Green function vector defined in
Eq. 共11兲 and g is the total Green function vector, defined as
g共x兲 = 关G共x, ␣1兲,G共x, ␣2兲, . . . ,G共x, ␣N兲兴T . 共17兲
It is clear from Eq. 共16兲 that, as was the case within the
DWBA, the exact scattering model for K maps CN to the
M-dimensional subspace S0 = Span兵g0共xm兲 , m = 1 , 2 , . . . , M
⬍ N其 傺 CN spanned by the background Green function vec-
tors evaluated at the scatterer locations and that this set of
Green function vectors forms a basis as long as M ⬍ N. On
the other-hand, unlike the DWBA approximation, the adjoint
K† maps CN to the M-dimensional subspace S
= Span兵g*共xm兲 , m = 1 , 2 , . . . , M其 傺 CN spanned by the full
Devaney, Marengo, and Gruber: Time-reversal-based inverse scattering
Green function vectors again evaluated at the scatterer loca-
tions rather than the background Green function vectors
evaluated at these locations. If one again makes use of the
SVD of the K matrix via Eqs. 共12兲, then it is apparent that
S0 = Span兵u p , ␴ p ⬎ 0其 and that S = Span兵v p , ␴ p ⬎ 0其. More-
over, since S0 = Span兵u p , ␴ p ⬎ 0其 ⬜ N0 = Span兵u p , ␴ p = 0其 and
S = Span兵v p , ␴ p ⬎ 0其 ⬜ N = Span兵v p , ␴ p = 0其 it then follows
that
u†pg0共xm兲 = 0, if ␴ p = 0,
v†pg*共xm兲 = 0, if ␴ p = 0.
Equations 共18兲 then lead to the two time-reversal MUSIC
pseudospectra
1
⌽1共x兲 = , 共19a兲
兺␴ =0 兩u†pg0共x兲兩2
p
1 关G0共xm, ␣ j兲G共xm, ␣k兲兴
⌽2共x兲 = , 共19b兲
兺␴ =0 兩v†pg*共x兲兩2
p G0共x1, ␣1兲G共x1, ␣1兲 ¯ G0共x M , ␣1兲G共x M , ␣1兲
whose peaks 共pole兲 locations theoretically yield the sought
after target locations.
The pseudospectrum ⌽1 is quite remarkable, in that it is
functionally identical to the DWBA pseudospectrum defined
in Eq. 共14a兲. This, of course, means that even in the presence
of multiple scattering, the target locations can still be deter-
mined knowing only the background Green function rather
than the full 共unknown兲 Green function. This is precisely the
result obtained in Gruber et al.9 by using the fact that the
multistatic data matrix is symmetric, even in the presence of
multiple scattering due to the reciprocity of the full and
background Green functions. Also, as mentioned by Gruber
et al.,9 this result is independent of the specific multiple scat-
tering model and depends only on the assumption of a point
scattering model of the form given in Eq. 共6a兲 and the as-
sumed reciprocity of the full and background Green func-
tions. By this we mean that the result does not depend on the
use of the Foldy–Lax model. Indeed, in Gruber et al.9 the
Neumann expansion was employed as the multiple scattering
model.
The pseudospectrum ⌽2 uses the full Green function
vector as a steering vector and although it will peak at the
same target locations as ⌽1, the two pseudospectra will gen-
erally be different; e.g., will have a different lobe structure.
Also, we note that since the K matrix is symmetric, even in
the presence of multiple scattering, it follows that v p = u*p, so
that 兩u†pg0兩 = 兩vTp g0兩 = 兩v†pg*0兩, so that the pseudospectrum ⌽1 is,
in fact, identical to the other form of the DWBA pseudospec-
trum given in Eq. 共14b兲. This alternate form for ⌽1 coincides
with ⌽2 under the replacement of the full Green function
vector with the background Green function vector.
It is interesting to note that the pseudospectrum ⌽1
⬅ ⌽ computed using either DWBA or exact 共multiple scat-
tering兲 data will be identical. To see this we represent the
singular system computed using the DWBA model, Eq. 共10兲,
by 兵ubp , vbp , ␴bp其 and the system generated using the exact mul-
tiple scattering model Eq. 共16兲 by the usual 兵u p , v p , ␴ p其. Now
J. Acoust. Soc. Am., Vol. 118, No. 5, November 2005
Downloaded 05 Oct 2012 to 136.159.235.223. Redistribution subject to ASA license or copyright; see http://asadl.org/terms
the subspaces spanned by both sets of singular vectors
ubp , ␴bp ⬎ 0 and u p , ␴ p ⬎ 0 are identical since they are both
equal to S0 = Span兵g0共xm兲 , m = 1 , 2 , . . . , M其 傺 CN. From this it
follows from Parseval’s theorem that
兺 兩ub†p g0共x兲兩2 = ␴兺=0 兩u†pg0共x兲兩2 ,
␴bp=0 p
from which it follows that the pseudospectra computed from
共18a兲 either DWBA or exact scattering data are precisely the same.
共18b兲
B. Scattering strength computation in the presence
of multiple scattering
In analogy with what was done in the case of the
DWBA, one can express the exact multistatic data matrix,
Eq. 共16兲, in the form of Eq. 共15兲 where, however, the matrix
关G0共xm , ␣ j兲G0共xm , ␣k兲兴 is replaced by the matrix
冤 冥
G0共x1, ␣1兲G共x1, ␣2兲 ¯ G0共x M , ␣1兲G共x M , ␣2兲
= .
] ] ]
G0共x1, ␣N兲G共x1, ␣N兲 ¯ G0共x M , ␣N兲G共x M , ␣N兲
In the case of the DWBA, the matrix equation, Eq. 共15兲, only
involved the 共known via MUSIC兲 target locations xm and
known background Green function G0 and, hence, was least
squares invertible. In the multiple scattering case this equa-
tion contains the full Green function G共xm , ␣k兲, which is un-
known and must also be determined in the inversion process.
Moreover, in this case the set of equations 共15兲 are no longer
linear, in that products of the two sets of unknowns 共full
Green function and scattering strengths兲 appear in the equa-
tions.
Besides being related by the modified form of Eq. 共15兲,
the scattering strengths and full Green function are coupled
by the Foldy–Lax equations, which are written in matrix
form for general transceivers in Eq. 共3兲. For the special case
of point tranceivers under consideration here, this set of
equations still applies with the following definitions of the
column vectors ⌿k and ⌿共in兲k :
⌿k = 关G共x1, ␣k兲,G共x2, ␣k兲, . . . ,G共x M , ␣k兲兴T , 共20a兲
⌿共in兲
k = 关G0共x1, ␣k兲,G0共x2, ␣k兲, . . . ,G0共x M , ␣k兲兴 ,
T
共20b兲
and where the M ⫻ M matrix H is still defined according to
Eq. 共5兲 that we reproduce here for clarity:
Hm,m⬘ = ␦m,m⬘ − 共1 − ␦m,m⬘兲␶m⬘G0共xm,xm⬘兲.
Equation 共3兲, with ⌿k and ⌿共in兲 defined as above, can be
k
expressed in the shortened form
H关G共xm, ␣k兲兴 = 关G0共xm, ␣k兲兴, k = 1,2, . . . ,N, 共21兲
where 关G共xm , ␣k兲兴 and 关G0共xm , ␣k兲兴 are the column vectors
defined in Eqs. 共20兲.
Devaney, Marengo, and Gruber: Time-reversal-based inverse scattering 3133
C. Iterative inversion
Equations 共15兲 共with 关G0共xm , ␣ j兲G0共xm , ␣k兲兴 replaced by
关G0共xm , ␣ j兲G共xm , ␣k兲兴兲 and 共21兲 are a set of coupled nonlin-
ear equations that can be solved via iteration for the scatter-
ing strengths ␶m and the full Green functions G共xm , ␣k兲. In
particular, consider next a sequence 兵G共n兲共xm , ␣k兲其 , n
= 0 , 1 , . . ., of estimated Green functions and a sequence of
共n兲
兵␶m 其 , n = 1 , 2 , . . ., of estimated scattering strengths, where the
target locations are first computed using the MUSIC pseu-
dospectrum ⌽ defined in Eqs. 共14a兲 or 共14b兲, and the itera-
tion begins with the DWBA estimate for the Green function;
i.e., G共0兲共xm , ␣k兲 = G0共xm , ␣k兲. The iteration then proceeds to
compute the sequence of scattering strength estimates and
Green function estimates using the following set of equa-
tions:
共n+1兲
关K j,k兴 = 关G0共xm, ␣ j兲G共n兲共xm, ␣k兲兴关␶m 兴,
共n+1兲
H共n+1兲 = ␦m,m⬘ − 共1 − ␦m,m⬘兲␶m⬘ G0共xm,xm⬘兲,
H共n+1兲关G共n+1兲共xm, ␣k兲兴 = 关G0共xm, ␣k兲兴.
The above iteration scheme has been implemented in
MATLAB and employed in simplified examples presented in
the following section. In the absence of noise and perfect
data it works perfectly, as expected, but experiences conver-
gence problems in cases of closely clustered targets sepa-
rated by less than a few wavelengths in the presence of ad-
ditive noise and/or error in the target location estimates. The
convergence properties of the iteration scheme will not be
discussed in this paper. The interested reader is instead re-
ferred to standard treatments of such schemes that are avail-
able in the literature.16–18
VI. COMPUTER SIMULATIONS
In this section we present the results of a MATLAB simu-
lation study of two-dimensional scattering from a set of M
= 4 point scatterers embedded in free space and being inter-
rogated by a set of N = 7 coincident point transceivers. A
basic image grid was employed involving quarter-
wavelength 共␭ / 4兲 spacing with 256 points along the horizon-
tal 共x兲 axis corresponding to spatial locations varying from
x = −32␭ to x = 32␭ and 128 points along the vertical 共z兲 axis
corresponding to spatial locations varying from z = 0 to z =
−32␭. The transceiver array was a uniform linear array hav-
ing six ␭ element-element separation and was centered along
the x axis at the top of the image grid 共at z = 0兲. The four
targets were located at positions 共−2.25, −14.75兲, 共−0.25,
−14.75兲, 共0.75, −15.75兲, and 共2.75, −15.75兲, all in units of a
wavelength. Unit-amplitude target scattering strengths were
chosen for all targets, i.e., ␶m = 1, m = 1 , 2 , 3 , 4, and the targets
were placed sufficiently close to each other so as to simulate
conditions involving strong multiple scattering between the
targets.
The scattering data in the form of the multistatic data
matrix K = 兵K j,k其, j , k = 1 , 2 , . . . , 7 were computed according
to Eq. 共8兲 with the background Green function,
3134 J. Acoust. Soc. Am., Vol. 118, No. 5, November 2005
Downloaded 05 Oct 2012 to 136.159.235.223. Redistribution subject to ASA license or copyright; see http://asadl.org/terms
FIG. 1. Plots of the magnitude and phase of the third column of the multi-
共22a兲 static data matrix K computed using the DWBA and the Foldy-Lax multiple
scattering models.
共22b兲
i
G0共xm, ␣ j兲 = H0共k0兩xm − ␣ j兩兲,
4
共22c兲
where H0 is the zeroth-order Hankel function of the first kind
and k0 = 2␲ / ␭ is the free space wave number. The full Green
function G共xm , ␣k兲 appearing in the expression for the mul-
tistatic data matrix was computed using the Foldy–Lax equa-
tions 共6b兲 for the specified set of scatterer locations xm and
scattering strengths ␶m with m = 1 , 2 , 3 , 4. The Foldy–Lax
equations were solved directly via the matrix equation 共21兲.
In the first simulation we also computed, for comparison
purposes, DWBA data, where the free space Green function
G0共xm , ␣k兲 was employed in place of the full Green function
in the computation of the multistatic data matrix.
A. Comparison of time-reversal MUSIC using noise
free DWBA and multiple scattering data
The first simulation examined the effect of multiple scat-
tering on scatterer location estimation and scatterer strength
estimation using noise-free 共perfect兲 simulated data and the
generalized MUSIC algorithm. Figure 1 shows plots of the
third column of the multistatic data matrices computed using
the DWBA approximation and the Foldy–Lax multiple scat-
tering model. It is apparent from the figure that the two data
matrices differ significantly, indicating the presence of strong
multiple scattering between the various point scatterers.
The SVDs of the two data matrices shown in Fig. 1 were
employed to generate two pseudospectra across the basic im-
age grid of 256⫻ 128 points using quarter-wavelength spac-
ing. These spectra were also interpolated 共using a standard
MATLAB command兲 over a small region centered within the
target region using ␭ / 40 spacing. The pseudospectra have
large amplitudes in the immediate vicinity of the transceiver
array so that we only employ and plot these quantities over a
256⫻ 122 grid that is displaced by two wavelengths in the
vertical 共z兲 direction from the transceiver array. Mesh plots
of the two pseudospectra are shown in Fig. 2, where in the
top of the figure we show the raw pseudospectra displayed
over the 256⫻ 122 grid with ␭ / 4 spacing and in the bottom
the interpolated quantities using ␭ / 40 spacing. It is clear
Devaney, Marengo, and Gruber: Time-reversal-based inverse scattering
FIG. 2. Mesh plots of the noise-free time-reversal MUSIC pseudospectra for the DWBA 共left兲 and Foldy–Lax model 共right兲 cases. The upper plots correspond
to the raw data while the lower plots correspond to the MATLAB-interpolated data. The positions of the peaks in these pseudospectra coincide exactly with the
target locations, as desired.
from the figure that the two pseudospectra are virtually iden-
tical, as required by the theory presented in Sec. V. An actual
numerical comparison of these two pseudospectra shows
them to be the same to within 11 decimal places.
The noise-free data described above was used to esti-
mate the scatterer locations xm and scattering strengths ␶m
using both the DWBA algorithm, Eq. 共15兲, and the iterative
algorithm described in Sec. V. The scatterer locations were
determined by a simple grid search for maxima of the pseu-
dospectra illustrated in Fig. 2 and were found to yield exact
target location estimates. These locations were then input
into both the DWBA and exact inversion algorithms to gen-
erate estimates ␶ˆ of the scattering strength vector ␶
= 关1 , 1 , 1 , 1兴. The accuracy of the estimates was quantified by
using a normalized percent error, defined by
储␶ˆ − ␶储
E = 100 ⴱ ,
储␶储
where 储x储 is the usual L2 norm of the vector x. The DWBA
inversion algorithm applied to DWBA data and the iterative
exact algorithm applied to multiple scattering data both re-
turned essentially exact estimates of the ␶ vector. On the
J. Acoust. Soc. Am., Vol. 118, No. 5, November 2005
Downloaded 05 Oct 2012 to 136.159.235.223. Redistribution subject to ASA license or copyright; see http://asadl.org/terms
other hand, the DWBA inversion algorithm applied to the
multiple scattering data returned an estimate having a 32%
error. Other runs of the simulation code using differing
choices of the scatterer locations and transceiver element
separations yielded similar results.
B. Time-reversal MUSIC performance in the presence
of additive noise
To determine the effect of additive noise on the scatterer
location and strength estimation algorithms we added com-
plex additive white Gaussian noise 共AWGN兲 to the multiple
scattering data and performed the same estimation proce-
dures employed in the above example. In this and the later
examples we employed only multiple scattering data since
the noise-free simulation results presented above clearly
demonstrate the inadequacy of the DWBA data model for
this particular scattering configuration. The signal model was
of the form
K̂ = K + N,
Devaney, Marengo, and Gruber: Time-reversal-based inverse scattering 3135

where K is the noise-free multiple scattering multistatic data
matrix and N is an uncorrelated, zero mean complex Gauss-
ian process 共AWGN兲. The noise level was quantified by the
usual signal-to-noise ratio in dB defined according to the
equation
储K储
S/N = 20 log10 . upper left-hand portion of the figure and the DWBA and
储N储
Figure 3 shows the pseudospectrum computed from a
single realization of noisy data having a signal-to-noise ratio
of roughly 25 dB. The plot on the top of the figure is dis-
played over the 256⫻ 122 image grid using ␭ / 4 spacing
while the plot displayed in the bottom is of the interpolated
pseudospectrum using ␭ / 40 spacing. The effect of the noise
is evident from the raw pseudospectrum in the form of spu-
rious maxima surrounding the maxima due to the scatterer
locations. The scatterer locations as estimated from the raw
pseudospectrum had an error of less than a tenth of one per-
cent, even in the presence of the additive noise. The DWBA
and multiple scattering-based iterative inversion algorithms
were employed to estimate the scattering amplitudes ␶m , m
= 1 , 2 , 3 , 4, yielding errors of 33.7% and 1.76%, respectively.
Different noise realizations having similar signal-to-noise ra-
tios yielded comparable results.
As a final example, a Monte Carlo simulation was ran
where the Foldy–Lax multiple scattering data was subjected
to zero mean AWGN. This noisy data was then used to esti-
mate the target locations using the raw 共noninterpolated兲
MUSIC pseudospectrum and these location estimates were
then employed in the DWBA and multiple scattering-based
iterative inversion algorithms to estimate the scattering am-
plitudes ␶m , m = 1 , 2 , 3 , 4. A thousand noise realizations were
3136 J. Acoust. Soc. Am., Vol. 118, No. 5, November 2005
Downloaded 05 Oct 2012 to 136.159.235.223. Redistribution subject to ASA license or copyright; see http://asadl.org/terms
FIG. 3. Mesh plots of the time-
reversal MUSIC pseudospectrum
computed from noisy multiple scatter-
ing data with AWGN yielding a
signal-to-noise ratio of about 25 dB.
The positions of the peaks in the pseu-
dospectrum coincide exactly with the
target locations, as desired, despite the
presence of additive noise.
considered for each of ten noise variances and, for each noise
variance, the resulting signal-to-noise ratios and target loca-
tion and scattering strength error estimates were averaged
over the number of noise realizations.
The results of the Monte Carlo simulation are shown in
Fig. 4. It is seen from the figure that there is a high degree of
correlation between the location error estimates shown in the
multiple scattering inversion algorithm errors shown in the
upper right- and lower left-hand portions of the figure, re-
spectively. The raw 共noninterpolated兲 pseudospectrum and a
simple grid search algorithm were used in obtaining the lo-
cation estimates. One should expect us to obtain better re-
sults using the interpolated pseudospectra and a more refined
search algorithm. Improved location error estimates translate
directly into improved coefficient estimates as is clear from
the figure.
Finally, we wish to mention that, unlike conventional
MUSIC that depends on having several statistical signal re-
alizations, the present approach can handle the worst-case
scenario of a single realization or snapshot 共single snapshot
data兲 as has been, in fact, the focus in the examples above.
The fact that we considered only the most restrictive case of
single snapshot data also explains why large errors occur at
relatively low noise levels. Clearly, the performance can be
better, in practice, if one manages to record more than a
single field realization 共sample兲 since then the different noise
contributions will statistically cancel out.
VII. CONCLUSION
This paper reformulated, by means of the Foldy–Lax
multiple scattering model, the treatment of a previous paper9
concerning time-reversal imaging with MUSIC for the loca-
Devaney, Marengo, and Gruber: Time-reversal-based inverse scattering
FIG. 4. Plots of the target location and target scattering strength estimation errors of the time-reversal-based inverse scattering method developed in the paper
as a function of the signal-to-noise ratio. For reference, these percentage errors are presented within the same plot at the bottom right, with the error in target
position appearing multiplied by a factor of 100.

tion of multiply scattering point targets embedded in a
known linear and reciprocal but otherwise arbitrary back-
ground medium. In the present paper we went a step beyond
both in theoretical insight and in the inversion algorithm,
addressing within a common theoretical footing based on the
Foldy–Lax multiple scattering model both the target location
problem and the associated target scattering strength estima-
tion problem. The theoretical results were illustrated with
computer simulations that validated the main theoretical ex-
pectations on both target location and scattering strength es-
timations under significant multiple scattering conditions.
The examples of this paper were based on numerical simu-
lations of exact scattering derived from the Foldy–Lax equa-
tions and illustrated also the role of noise in the data.
It was shown that the MUSIC pseudospectrum com-
puted using the background Green function as a steering vec-
tor yields accurate estimates of the target locations, and that
the target scattering strengths can be computed from the so-
determined target locations by means of a nonlinear iterative
routine. These developments corroborate via a detailed for-
mulation that complements the one in Gruber et al. 共which
was based on the Neumann series兲 the main finding that the
time-reversal imaging with a MUSIC method for the target
location can be used, not only in the usual DWBA frame-
work but also in more general nonlinear scattering frame-
works, where there is significant multiple scattering between
the targets. This result is important toward practical imple-
mentations of both time reversal and MUSIC methods in,
e.g., sonar, underwater communications, medicine, nonde-
structive evaluation, and imaging, in general.4,8,10 The non-
linear scattering strength estimation routine derived in the
present paper completes the full target inversion 共target loca-
tion plus scattering strength estimation兲 of multiply scatter-
ing point targets from multistatic array data and it also pro-
vides a new general framework applicable to other inverse
problems.
ACKNOWLEDGMENTS
The authors gratefully acknowledge support under Grant
No. FA9550-04-1-0187 from the Defense Advanced Re-
search Projects Agency 共DARPA兲 and the Air Force Office of
Scientific Research 共AFOSR兲 for the research reported here.
This work is affiliated with CenSSIS, the Center for Subsur-
face Sensing and Imaging Systems, under the Engineering
Research Centers Program of the National Science Founda-
tion 共Award No. EEC-9986821兲.

J. Acoust. Soc. Am., Vol. 118, No. 5, November 2005 Devaney, Marengo, and Gruber: Time-reversal-based inverse scattering 3137

Downloaded 05 Oct 2012 to 136.159.235.223. Redistribution subject to ASA license or copyright; see http://asadl.org/terms
1
C. Prada and M. Fink, “Eigenmodes of the time-reversal operator: A so-
lution to selective focusing in multiple target media,” Wave Motion 20,
151–163 共1994兲.
2
C. Prada, S. Manneville, D. Spoliansky, and M. Fink, “Decomposition of
the time-reversal operator: Application to detection and selective focusing
on two scatterers,” J. Acoust. Soc. Am. 99, 2067–2076 共1996兲.
3
M. Mordant, C. Prada, and M. Fink, “Highly resolved detection and se-
lective focusing in a waveguide using the D.O.R.T. method,” J. Acoust.
Soc. Am. 105, 2634–2642 共1999兲.
4
M. Fink, D. Cassereau, A. Derode, C. Prada, P. Roux, M. Tanter, J. L.
Thomas, and F. Wu, “Time-reversed acoustics,” Rep. Prog. Phys. 63,
1933–1995 共2000兲.
5
A. J. Devaney, “Super-resolution processing of multi-static data using
time reversal and MUSIC,” unpublished manuscript available on the web
site www.ece.neu.edu/faculty/devaney, 2000.
6 15
H. Lev-Ari and A. J. Devaney, “The time-reversal technique re-
interpreted: Subspace-based signal processing for multi-static target loca-
tion,” IEEE Sensor Array and Multichannel Signal Processing Workshop,
Cambridge, MA, 2000, pp. 509–513.
7 17
S. Lehman and A. J. Devaney, “Transmission mode time-reversal super-
resolution imaging,” J. Acoust. Soc. Am. 113, 2742–2753 共2003兲.
8 18
C. Prada and J. L. Thomas, “Experimental subwavelength localization of
scatterers by decomposition of the time-reversal operator interpreted as a
3138 J. Acoust. Soc. Am., Vol. 118, No. 5, November 2005
Downloaded 05 Oct 2012 to 136.159.235.223. Redistribution subject to ASA license or copyright; see http://asadl.org/terms
covariance matrix,” J. Acoust. Soc. Am. 114, 235–243 共2003兲.
9
F. K. Gruber, E. A. Marengo, and A. J. Devaney, “Time-reversal imaging
with multiple signal classification considering multiple scattering between
the targets,” J. Acoust. Soc. Am. 115, 3042–3047 共2004兲.
10
T. Miwa and I. Arai, “Super-resolution imaging for point reflectors near
transmitting and receiving array,” IEEE Trans. Antennas Propag. 52, 220–
229 共2004兲.
11
J. H. Taylor, Scattering Theory 共Wiley, New York, 1972兲.
12
A. J. Devaney and M. Oristaglio, “Inversion procedure for inverse scat-
tering within the distorted wave Born approximation,” Phys. Rev. Lett.
51, 237–240 共1983兲.
13
L. L. Foldy, “The multiple scattering of waves,” Phys. Rev. 67, 107–119
共1945兲.
14
R. K. Snieder and J. A. Scales, “Time-reversed imaging as a diagnostic of
wave and particle chaos,” Phys. Rev. E 58, 5668–5675 共1998兲.
L. Tsang, J. A. Kong, K.-H. Ding, and C. O. Ao, Scattering of Electro-
magnetic Waves: Numerical Simulations 共Wiley, New York, 2001兲.
16
J. M. Ortega and W. C. Rheinboldt, Iterative Solution of Nonlinear Equa-
tions in Several Variables 共Academic, New York, 1970兲.
M. N. Vrahatis, G. D. Magoulas, and V. P. Plagianakos, “From linear to
nonlinear iterative methods,” Appl. Numer. Math. 45, 59–77 共2003兲.
M. S. Bazaraa and C. M. Shetty, Nonlinear Programming: Theory and
Algorithms 共Wiley, New York, 1979兲.
Devaney, Marengo, and Gruber: Time-reversal-based inverse scattering