Passive array imaging in random media

Liliana Borcea and Ilker Kocyigit

December 15, 2017
arXiv:1712.04934v1 [math.NA] 12 Dec 2017

Abstract
We present a novel algorithm for high resolution coherent imaging of sound sources
in random scattering media using time resolved measurements of the acoustic pressure
at an array of receivers. The sound waves travel a long distance between the sources and
receivers so that they are significantly affected by scattering in the random medium.
We model the scattering effects by large random wavefront distortions, but the results
extend to stronger effects, as long as the waves retain some coherence i.e., before the
onset of wave diffusion. It is known that scattering in random media can be mitigated
in imaging using coherent interferometry (CINT). This method introduces a statistical
stabilization in the image formation, at the cost of image blur. We show how to modify
the CINT method in order to image wave sources that are too close to each other to
be distinguished by CINT alone. We introduce the algorithm from first principles and
demonstrate its performance with numerical simulations.
Keywords Wave scattering in random media, coherent interferometric imaging, array
imaging.

1 Introduction
Coherent array imaging is an important technology in radar [10], sonar [15], seismic imaging
[2], photoacoustic imaging [16], medical imaging with ultrasound [18], and so on. We focus
attention on passive array imaging, where a collection of Nr receivers record waves generated
by Ns unknown sources. The receivers are located at points x ~ r A, for r = 1, . . . , Nr , where
A is the array aperture, assumed for convenience to be planar and square, of side a, as shown
in Figure 1. The unknown sources are located at points y ~ s D, where D is the imaging
region, a bounded set with center at distance L from the array, in the direction orthogonal
to the aperture, called the range direction. The coordinates in the plane orthogonal to this
direction are called cross-range coordinates. We let D be a rectangular prism with square
cross-section of side D in the cross-range plane, satisfying L  a > D, and side D3 in the
range direction, satisfying L  D3 . These scaling relations are typical in most imaging
applications.

L. Borcea is with the Department of Mathematics, University of Michigan, Ann Arbor, MI 48109 USA,
e-mail: borcea@umich.edu.

I. Kocyigit is with the Department of Mathematics, Dartmouth College, Hanover, NH 03755 USA,
e-mail: ilker.kocyigit@dartmouth.edu

1
Figure 1: Imaging setup with an array of receivers that is planar square of side a. The
range direction is orthogonal to the array aperture. The unknown sources are in the imaging
region, a rectangular prism with size D3 in range and D in cross-range.

~ ), the solution of the equation

The waves are modeled by the pressure p(t, x
Ns
1 2
X
~ ) p(t, x
p(t, x ~) = xy
fs (t)(~ ~ s ), (1)
x) t
c2 (~ s=1

~ R3 and t R, with zero initial conditions on p(t, x

for x ~ ) and t p(t, x
~ ) at time t prior to
the source excitation. The variable wave speed
  1/2
~
x
c(~
x) = co 1 + (2)
`

models the heterogeneous medium consisting of a homogeneous background with constant

wave speed co and numerous weak inhomogeneities of size O(`), commonly referred to as
clutter. Because these inhomogeneities are unknown in imaging, they introduce uncertainty
in the wave propagation, modeled in equation (2) by the dimensionless random process of
dimensionless argument. We assume that is stationary, and bounded almost surely. It has
zero mean, and autocorrelation

~ 0 |2
 
0 0 |~
xx
R (~xx~ ) = E[(~ x )] = exp
x)(~ , (3)
2

where E[] denotes expectation with respect to the distribution of . The Gaussian expression
of R is chosen for convenience, but the results extend to any integrable autocorrelation
function. The scale ` is called the correlation length and  1 quantifies the small amplitude
of the fluctuations of c(~ x).
The imaging problem is to determine the source locations {~ ys }1sNs from the measure-
ments {p(t, x~ r )}r=1,...,Nr .

2
1.1 Related work
Each inhomogeneity in clutter is a weak scatterer by itself since  1, but cumulative
scattering builds up over long ranges. Mathematically, this manifests in the exponential
decay of the coherent wave E[p(t, x ~ ) E[p(t, x
~ )] and the increase of the fluctuations p(t, x ~ )].
The range scale S of decay of the coherent wave is the scattering mean free path [19].
When L < S, the cumulative scattering effects are negligible. Much of the imaging
literature considers this case, and coherent methods known as reverse time migration [2],
matched filtering or backprojection [10, 13] work well. They are based on the data model

~ r ) = po (t, x
p(t, x ~ r ) + W (t, x
~ r ), r = 1, . . . , Nr , (4)

~ r ) is the solution of equation (1) with constant wave speed co , and W (t, x
where po (t, x ~ r ) is
additive noise with some statistics, assumed uncorrelated over the receivers. In the simplest
form, the imaging function is
Nr
X
J (~
y) = p ( (~ ~ ), x
xr , y ~ r) , ~ D,
y (5)
r=1

where
(~ ~ ) = |~
xr , y xr y
~ |/co (6)
is the travel time from the imaging point y ~ to the receiver at x
~ r . The image is robust to
additive noise, and it peaks in the vicinity of the source locations, with cross-range resolution
O(o L/a) and range resolution O(co /B), where o = 2co /o is the central wavelength of the
waves, calculated in terms of the central frequency o , and B is the bandwidth of the source
signals. Better resolution can be achieved using convex, sparsity promoting optimization, if
the noise is not too strong, and the sources are separated by more than o L/a in cross-range
and co /B in range, see e.g., [12, 8, 6].
The imaging problem is much more difficult when L  S, because the array measure-
ments are significantly affected by scattering in clutter and are no longer approximated by
the model (4). If the range L is so large that it exceeds the transport mean free path T , which
is the scale that marks the onset of wave diffusion [19], coherent imaging cannot succeed.
We assume an intermediate regime T > L  S, where coherent imaging is still possible.
Such imaging must involve statistical stabilization with respect to the uncertainty of clutter,
so that the estimates of the source locations are insensitive to the particular realization of
the random medium (clutter) in which the imaging takes place.
Statistical stability can be obtained with the coherent interferometric (CINT) approach
[4, 3], which forms images using the cross-correlations
Z
 
C(t, e ~ r, x
t, x ~ r0 ) = ds (t s)

 t
e   t
e 
~ r p? s +
p s ,x ~ r0 ,
,x (7)
2 2
where the star denotes complex conjugate. These are calculated around the time t, in a
time window modeled by the function of dimensionless argument and O(1) support. The

3
parameter is adjustable and it should be similar to d , the frequency offset over which the
waves decorrelate in the random medium [19, 4]. Because the waves also decorrelate over
directions of arrival, only cross-correlations at nearby receivers are useful. Thus, CINT uses
a spatial windowing function of dimensionless argument and O(1) support to ensure that
the receivers in (7) are at distance |~
xr x
~ r0 | X, with adjustable parameter that optimally
equals the decorrelation length [4]. This scale is proportional to the wavelength, but typically
B  o , so the decorrelation length Xd is approximately constant in the bandwidth. The
CINT imaging function is
Nr  |~
X xr x
~ r0 | 
J (~
y) =
0
r,r =1
X
C( (~ ~ r0 , y
xr , x ~ ), e(~ ~ r0 , y
xr , x ~ ), x
~ r, x
~ r0 ), (8)

where
1
(~ ~ r0 , y
xr , x ~) = [ (~ ~ ) + (~
xr , y ~ )] ,
xr 0 , y
2
e(~ ~ r0 , y
xr , x ~ ) = (~ ~ ) (~
xr , y ~ ).
xr0 , y

It is statistically stable with respect to the realizations of the random medium if d  B

and X Xd  a, meaning that its expectation near the peaks is much larger than the
standard deviation [3]. Moreover, these peaks are in the vicinity of the source locations
with resolution O(o L/X) in cross-range and O(co /) in range. These resolution limits are
similar to those in homogeneous media, except that the aperture size a and bandwidth B
are replaced by the windowing parameters X and . These are necessarily smaller than a
and B to have statistical stability, so the images are blurrier.
It is shown in [7] that the resolution of CINT images may be improved using convex
optimization. However, this requires detailed knowledge of the blurring kernel i.e., prior
calibration. The deblurring also works best when the sources are separated by distances
larger than o L/X in cross-range and co / in range, in the sense that if this is not so, there
is no guarantee of unique recovery of the source locations.

1.2 Contributions
In this paper we show how, by slightly modifying the CINT imaging function (8), it is
possible to recover the unknown sources almost as well as in homogeneous media. Explicitly,
we show that a collection of sources that are within a blurred peak of the CINT image and are
separated by distances O(o L/a) in cross-range and O(co /B) in range, can be estimated up to
a translation in the support of the CINT peak. We introduce an algorithm that achieves this
result, motivate it from first principles and assess its performance with numerical simulations.
Note that although we restrict our study to passive array imaging, the results generalize
easily to active arrays that probe the medium with waves and record the echoes, in order
to determine point-like scatterers with much larger reflectivity than , the reflectivity of the
clutter inhomogeneities. These scatterers are secondary sources of waves, which emit signals
proportional to the incident wave, so they can be viewed as the unknown sources considered

4
here. This is obvious in the single scattering (Born) approximation, but it extends to multiple
scattering as well, as explained in [9]. Point-like scatterers play an important role in sonar and
radar imaging because corners of targets create stronger echoes than other features. Thus,
the images are often a constellation of peaks from which target features are to be extracted
[14]. Our algorithm can be used for this purpose in random media, because it recovers the
relative location of the corner reflectors with the same resolution as in homogeneous media.
The paper is organized as follows: In section 2 we introduce and analyze the CINT-like
imaging function. In section 3 we use this function in the source reconstruction algorithm.
We illustrate the performance of this algorithm with numerical simulations in section 4, and
conclude with a summary in section 5.

2 The CINT-like imaging function

~ D,
We propose a simple modification of (8), where instead of searching for a single point y
~ 0 D. The CINT-like imaging function
~, y
we have two search points y
Nr  |~
X xr x
~ r0 | 
I(~ ~ 0) =
y, y
0
r,r =1
X
C( (~ ~ r0 , y
xr , x ~ 0 ), e(~
~, y ~ r0 , y
xr , x ~ 0 ), x
~, y ~ r, x
~ r0 ), (9)

~ 0 dependent travel times

~, y
superposes the cross-correlations (7) evaluated at the y
1
(~ ~ r0 , y
xr , x ~ 0) =
~, y [ (~ ~ ) + (~
xr , y xr 0 , y~ 0 )] , (10)
2
e(~ ~ r0 , y
xr , x ~ 0 ) = (~
~, y ~ ) (~
xr , y ~ 0 ).
xr 0 , y (11)

To explain why this is beneficial, we derive below the expression of I(~ ~ 0 ) using a random
y, y
travel time model [17] that accounts for large, random wavefront distortions in random
media, as assumed in adaptive optics [1]. This model is convenient for the calculations and
has been used in the analysis of CINT imaging in random media in [3, 7].
As shown in the appendix A, the calculations are based on the expression of the second
statistical moments of the pressure waves, which are qualitatively similar to those in stronger
scattering regimes [17, 5]. Thus, the results extend verbatim to such regimes.

2.1 Setup and the random travel time model

We introduce here a few assumptions that simplify the calculations and lead to an explicit
expression of (9).
The first assumption is that the sources emit the same pulse
 1/4
2
fs (t) = f (t) = B exp[io t tB 2 ], (12)

5
modulated at central frequency o , with Gaussian envelope normalized so that kf k2 = 1.
The Fourier transform of this pulse
!1/2
2 ( o )2 i
h
fb() = exp , (13)
B 4B 2

is also a Gaussian, centered at the frequency o and with standard deviation proportional to
B, called in an abuse of terminology the bandwidth. We assume that B  o . In general,
the sources will not emit the same signal and moreover, the signal may not be a pulse. The
imaging algorithm in this paper applies to arbitrary fs (t), that may even be noise-like, as
long as for nearby sources these signals are statistically correlated. This holds for example
in active array imaging, where unknown scatterers act as secondary sources of waves and
fs (t) are given by the convolution of the probing signal emitted by the array and the Greens
function that propagate the waves in the random medium to the scatterer locations.
The second assumption is that at any given frequency in the support of (13), the wave
propagation can be modeled by the Greens function
exp {i [ (~ ~ ) + (~
x, y ~ )]}
x, y
G(,
b x ~,y
~) = (14)
4|~xy ~|
of Helmholtzs equation, where
1
xy ~|  (1 u)~
Z
|~ y + u~
x
(~ ~) =
x, y du . (15)
2co 0 `
This is the random travel time model and we refer to [3, 17] for a detailed discussion of
its range of validity. Here it suffices to say that it holds when o  `  a  L and
is sufficiently small. We also recall from [3, Lemma 3.1] that the random
process (15) is
approximately Gaussian, with mean zero and standard deviation O( `L/o ).
The third assumption is that the size of the imaging region satisfies the scaling relations
o L co
 D a,  D3 . (16)
X
The lower bounds in these relations are the CINT resolution limits, so this assumption en-
sures that the search domain is large enough to observe the focusing of the CINT imaging
function (8). We also suppose that the aperture size a and the the size of the imaging region
are sufficiently small so that the rays connecting the sources and receivers are contained
within a narrow cone of small opening angle and axis along the range direction. This as-
sumption is described in more detail in the appendix A and in technical terms it means that
the waves are in a paraxial propagation regime.
Finally, to carry out explicit calculations, we take the Gaussian window functions
h (t)2 i
(t) = exp , (17)
2
 |~
xr x
~ r0 |  h |~
xr x~ r0 |2 i
= exp , (18)
X 2X 2

6
in definitions (7) and (9). We also suppose that the receivers are spaced at O(o ) distances,
so that Nr = O(a2 /2o )  1. This allows us to approximate the sums over the receivers by
integrals over the aperture A. To avoid specifying the aperture size in these integrals, we
use the Gaussian apodization

x|2 /(2(a/6)2 ) , x
 
exp |~ ~ A,

which is negligible outside the disk of radius a/2.

2.2 The imaging kernel

For simplicity, we neglect additive noise in the calculations in this section, although noise is
considered in the numerical simulations in section 4. As shown in the appendix A, in the
setup described in section 2.1, the imaging function (9) can be written as
Ns
X
I(~ ~ 0) =
y, y K(~ ~ 0, y
y, y ~s, y
~ s0 ), (19)
s,s0 =1

with kernel K that depends on the search points y ~ 0 and the locations y
~, y ~ s and y
~ s0 of pairs
of the unknown sources. To describe this kernel, consider the system of coordinates with
~ = (y, y3 ), with two dimensional vector y in the
origin at the center of the array and write y
cross-range plane and range coordinate y3 . Define also the center and difference cross-range
vectors
y + y0
y= , ye = y y 0 , (20)
2
and the center and difference range coordinates

y3 + y30
y3 = , ye3 = y3 y30 . (21)
2
~ s = (ys , ys,3 ) and define
Similarly, we let y
ys + ys0
y ss0 = , yess0 = ys ys0 , (22)
2
ys,3 + ys0 ,3
y ss0 ,3 = , yess0 ,3 = ys,3 ys0 ,3 , (23)
2
for all s, s0 = 1, . . . , Ns .
The kernel satisfies
yss0 |2 |e yss0 ,3 ye3 |2

0 |e
|K(~ ~ ,y
y, y ys0 )| exp
~ s ,~
2Xd2 2(co /B)2
yss0 ye|2
|e |y ss0 y|2

2[1 L/(ko a)]2 2[L/(ko Xe )]2
(|(y ss0 , y ss0 ,3 )| |(y, y 3 )|)2

, (24)
2(co /e )2

7
where means of the order of, up to a multiplicative constant. We refer to the appendix
A for the detailed expression of K, not just its absolute value. In (24) we introduced the
wavenumber
ko = o /co = 2/o
and the positive coefficients , 1 O(1). We also denote by e and Xe the frequency and
length scales defined by
1 1 1 1
2
= 2+ 2+ , (25)
e d 4B 2
1 1 1 1
2
= 2+ 2+ . (26)
Xe X Xd 4(a/6)2
Since B  d and a  Xd , the optimal windowing choice = O(d ) and X = O(Xd ) gives
e = O(d )  B, Xe = O(Xd )  a. (27)
The expression (24) says that the imaging function peaks when the center 12 (~ y+y ~ ) of
1 0
the imaging points is in the vicinity of 2 (~ ~ s0 ), for some pair s, s of source indexes.
ys + y
The radius of this vicinity is O(o L/Xe ) in the cross-range plane and O(co /e ) in the range
direction. This is the same as the focusing of the CINT image (8). The new observation is
that we can get much better estimates of the source offsets y ~s y
~ s0 , with the same resolution
as in the homogeneous medium i.e., O(o L/a) in cross-range and O(co /B) in range.

3 The imaging algorithm

The imaging algorithms consists of the following steps:

Step 1: Calculate the CINT image (8), which is the same as I(~ ~ ), and identify its
y, y
peaks in the search domain D. The unknown sources lie in the support of these peaks, but
they cannot be identified due to the poor resolution: O(o L/Xe ) in cross-range and O(co /e )
in range. To reduce the computations, it suffices to form the CINT image on a coarse mesh
with pixel size similar to these resolution limits.

Step 2: Let z ~o be the location of the center of a CINT peak. Suppose that there are
ns Ns sources within this peak and denote by Y the set of their locations. We can only
expect to determine these locations up to an overall translation, so we set z ~o as one point
~o , calculate
in the constellation of ns sources. To estimate the other locations, relative to z
I(~ ~ ) for y
z0 , y ~ in the support of the CINT peak, on a refined imaging mesh with pixel size
O(o L/a) in cross-range and O(co /B) in range. These are the resolution limits for the source
offsets in equation (24).

Step 3: Identify the peaks of I(~ ~ ), which are the points z

z0 , y ~j satisfying
~j z
z ~o E(Y), (28)
where
E(Y) = {~
ys y
~ s0 : y ~ s0 Y, y
~s, y 6 y
~s = ~ s0 } . (29)

8
The set E(Y) has cardinality ns (ns 1) in most cases, where the offset vectors y ~s y~ s0 are
0 0 0
distinct for different pairs (s, s ), with s, s = 1, . . . , ns and s 6= s . Thus, I(~ ~ ) is expected
z0 , y
to have Nz = ns (ns 1) peaks. This count reflects that if ~e E(Y), then ~e E(Y), as
well. However, there are special, unlikely cases, where different source pairs give the same
offset vectors. Thus, in general,
Nz ns (ns 1).

Step 4: From the Nz peaks of I(~ ~ ) estimate the set (29) by

z0 , y

E est = {~
zj z
~o , j = 1, . . . , Nz }. (30)

Use this set to determine the constellation of sources. For our purpose, it suffices to use the
exhaustive search algorithm given below, which is not optimal in terms of computational
cost. The output of this algorithm is a set Y est of points y
~ sest , so that

E(Y est ) = E est . (31)

Here E(Y est ) is defined as in (29), with Y replaced by Y est . The vectors in Y est are the
estimates of the source locations, up to the translation defined by fixing one source at z ~o
and the reflection about z~o .
Algorithm(E, Y )
Input: The sets E, and Y .
Output: Empty set or a non-empty set Y est .
1. If E (Y ) = E est then
2. return Y est = Y
3. End-If
4. While E 6= do
5. Select the first vector ~e and set E = E\ {~e}
6. Let y~=z ~0 + ~e
7. If { (Y y ~ )} E est then
8. Y 0 = Algorithm(E, Y {~ y })
0
9. If Y 6= then
10. return Y est = Y 0
11. End-If
12. End-If
13. End-While
14. return

3.1 Discussion
The first call of this recursive algorithm is made with the inputs E = E est and Y = {~ zo }.
When the algorithm outputs the empty set , the search has failed. We show in the appendix
B that, when noise is not an issue i.e., the offset vectors in the set E est are the same as those
in E(Y), the output of the algorithm is necessarily a non empty set Y est satisfying (31). In
practice, the testing of the equalities and the inclusion at lines 1 and 7 can be done up to

9
 4
14
2
10
0
6
-2
2 -4
50 100 150 200 250 300 350 400

Figure 2: Display of the realization of the fluctuations used in the numerical results. The
abscissa is range in units of ` and the ordinate is cross-range in units of `.

some tolerance. In our numerical simulations we consider two vectors to be same if their
difference has cross-range and range components that are smaller than the pixel size, in
absolute value.
Note that the expression (24) of the imaging kernel indicates that only sources at cross-

range offsets |ys ys0 | O( Xd ) contribute to the image I(~ ~ ) calculated at step 2. This
z0 , y
distance is at least O(o L/Xe ) in our case, so all the sources supported in the CINT peak
should contribute to I(~ ~ ). In other scattering regimes it may be that Xd  o L/Xe , so
z0 , y
the support of the CINT peak may be divided in smaller sets at step 2. The remainder of
the algorithm above can be used separately for each such set.
We already stated that we can only hope to determine the set Y of source locations up
to an overall translation, fixed by the starting point z ~o and up to a reflection. This is due to
ref
the fact that the set Y , defined as the reflection of Y with respect to a fixed point, satisfies
E(Y) = E(Y ref ).

3.2 Relation to localization

The estimation of the source locations from the set E est of offset vectors is somewhat related to
the network localization problem [20], [21]: Let N = {n1 , n2 , . . . , nm } denote the unknown
set of nodes of a network. Determine N from knowledge of a non-empty subset B N of
so-called beacon nodes and the distance map : N N R+ , defined by (i, j) = kni nj k,
for i, j = 1, . . . , m.
Our problem is different as follows: (1) The data are the offset vectors y ~j y
~ j 0 and not
just their norm. (2) We do not have access to the distance map , we only know its image.
That is to say, for any offset vector ~e E est , we do not know the pair (i, j) of sources that
give y~i y
~ j = ~e. (3) As explained at Step 3 in section 3, in general, we do not know the
number of sources. Only under the additional assumption that there is a unique pair of
sources that gives an offset vector ~e E est , we can determine the number of sources from
the cardinality of the set E est .

4 Numerical results
To minimize the computational cost, we present imaging results in two dimensions, in the
plane define by the range axis and one cross-range direction. The array cross-section in this
plane is the line segment [a/2, a/2].

10
4.1 Simulations setup
The data are obtained using the model
Z
d it h i
~ r) =
p(t, x e ~ r ) + W (, x
pb(, x c ~ r) , (32)
2

where W
c denotes additive noise and
Ns
X
~ r ) = fb()
pb(, x G(,
b x ~ r, y
~s) (33)
s=1

is the solution of Helmholtzs equation in the cluttered medium. The noise W c is complex
Gaussian, uncorrelated over the receivers and frequencies, with mean zero and standard
deviation 5% of the maximum absolute value of (33). The Greens function G b is calculated
using definitions (14)-(15), in the realization of the random medium displayed in Figure 2.
This realization is generated using random Fourier series [11], for the autocorrelation (3).
All the length scales are relative to the correlation length `. The central wavelength is
o = 1.1 105 `, the array aperture size is a = 16` and the range scale is L = 800`. The
frequencies are scaled with respect to o and the bandwidth is B = o /5. The strength of
the fluctuations is = 2 106 and the decorrelation frequency and length defined in (43)
are d = 0.039o and Xd = 0.068`. The window parameters are X = Xd /3 and = d /3,
so definitions (25)(26) give Xe = 0.0214` and e = 0.0124o and the decay scales in the
expression of the kernel are
L co
= 0.0654`, = 1.4 104 `,
ko Xe e
L co
= 8.8 105 `, = 8.8 106 `.
ko a B
The coefficient in (24) is
4Xd2
= = 1.025
4Xd2 Xe2

and 1 6 2.
The images displayed in the next section are calculated on an imaging mesh with pixel
size 3.92L/(ko a) in cross-range and 0.537co /B in range. The axes in the plots are in units
of the CINT resolution limits.

4.2 Source reconstructions

We present in Figure 3 the results for three nearby sources. As expected, the CINT image
I(~ ~ ) displayed in the top left plot has a blurry peak, and cannot distinguish the sources.
y, y
The conventional image displayed in the top right plot is calculated using definition (5). It
has many spurious peaks and our simulations show that these peaks change unpredictably
from one realization of the random process to another. This statistical instability is
expected, because the data are incoherent in our regime.

11
 100 100

50 50

0 0

-50 -50

-100 -100

-13.6 -6.8 0 6.8 13.6 -13.6 -6.8 0 6.8 13.6

100

25
50

0
0

-50

-25
-100

-13.6 -6.8 0 6.8 13.6 -3.4 0 3.4

100

25
50

0
0

-50

-25
-100

-3.4 0 3.4 -13.6 -6.8 0 6.8 13.6

Figure 3: Top row: CINT image I(~ ~ ) of three sources shown in red (left plot) and the
y, y
conventional image (5) (right plot). Middle row: The image I(~ ~ ) (left plot) and the offset
z0 , y
vectors that define the set (30) (right plot). Bottom row: Reconstruction of the three sources
(left plot) and reconstruction by deblurring the CINT image using convex optimization (right
plot). The abscissa is range offset with respect to the center location z ~o of the CINT peak,
in units of L/(ko a). The ordinate is cross-range offset with respect to z ~o , in units of co /B.

In the left plot of the middle row we display the image I(~ ~ ), with z
zo , y ~o at the center of
the CINT peak. Because we have ns = 3 sources supported in this peak, we observe Nz = 6
peaks z~j , for j = 1, . . . , 6. These define the set E est defined in (30), with offset vectors
displayed in the right plot of the middle row. Note that for each vector ~e in E est we also have
the vector ~e. The reconstruction of the sources using the algorithm described in section 3 is

12
shown in the bottom left plot. The reconstruction is exact up to the translation by the vector
~o , where by exact we mean with error that is smaller than the pixel size. For comparison, we
z
also show in the bottom right plot the reconstruction obtained with the debluring algorithm
introduced in [7]. This algorithm is guaranteed to give a good reconstruction of the sources
when they are further apart than the CINT resolution limits. In this simulation the sources
are much closer to each other so the results are worse than those in the bottom left plot. As
predicted by the theory in [7], the reconstruction in the bottom right plot is peaked near the
source locations, but it does not show three distinct sources.
In Figure 4 we show the reconstruction of 4 sources. As in the previous example, the
sources are located in the support of the CINT peak. The image I(~ ~ ) shown in the
zo , y
top right plot has one spurious peak, due to the noise. However, this can be easily filtered
out because the offset vector ~e E est corresponding to it does not have the property that
~e E est . The set of remaining offset vectors is displayed in the left bottom plot and the
reconstruction of the four sources is shown in the bottom right plot. The reconstruction is
exact, up to the translation by z~o .

84 84

50 50

17 17

-17 -17

-50 -50

-84 -84
-11.3 -5.9 -0.5 5 10.4 15.8 -11.3 -5.9 -0.5 5 10.4 15.8

17 17

-17 -17

-50 -50
-5.9 -0.5 5 -5.9 -0.5 5

Figure 4: Top row: Left: CINT image I(~ ~ ) of four sources shown in red. Right: The
y, y
image I(~ ~ ). Bottom row: Left: The offset vectors that define the set (30). Right:
z0 , y
Reconstruction of the four sources. The axes are as in Figure 3.

13
5 Summary
We introduced a novel algorithm for array imaging in cluttered media modeled by a random
sound speed. The algorithm is designed to work in the presence of strong scattering effects
in clutter, where the sound waves recorded at the array are incoherent i.e., their statistical
expectation is close to zero. Physically, this means that the range offset between the unknown
sources and the array is larger than the scattering mean free path in clutter. The algorithm
uses an imaging approach that is similar to the coherent interferometric (CINT) method.
CINT is known to be robust to clutter scattering effects, as long as the waves are not in
a diffusion regime i.e., for ranges less than a transport mean free path. The robustness
comes at the cost of image blur. This impedes imaging of sources at nearby locations y ~s,
for s = 1, . . . , Ns . The algorithm introduced in this paper uses the observation that the
blur affects only the estimation of the center locations (~ ~ s0 )/2 of pairs (s, s0 ) of sources,
ys + y
whereas the offset vectors y ~s y
~ s0 can be estimated with the same resolution as in the
absence of clutter. Thus, it is possible to determine constellations of nearby sources, up to
a translation within the support of a peak of the CINT image.
We motivated the algorithm from first principles, starting with the wave equation in
random media and assessed its performance with numerical simulations. To simplify the
presentation, we considered a high frequency scattering regime defined by large, random
wavefront distortions of the waves received at the array, although as explained in the paper,
the results extend verbatim to stronger scattering regimes.

Appendix A
In this appendix we derive the expression (24). We begin with the solution of the wave
~ r = (xr , 0),
equation (1) evaluated at the receiver location x
Z Ns
d it b X
~ r) =
p(t, x e f () G(,
b x ~ r, y
~ s ), (34)
2 s=1

where we used the assumption (12) and G

b is the Greens function modeled by (14). This
model holds under the assumptions

p o `
`  o L  o ,  , (35)
L
as explained in [17, 3, 7]. These ensure that the waves propagate along straight rays, as
in geometrical optics, and that the random fluctuations of the amplitude of the Greens
function are negligible. The first bound on ` is to have consistent assumptions on ,
2/3
o `1/6 o `
o   , (36)
`L L5/6 L

chosen large enough to give large random travel time fluctuations, as explained in section
2.1.

14
 It is shown in [7, Proposition 3.1] that the expectation of the Greens function is
h i ei (~xr ,~ys ) |~xr ~ys |
E G(, x
b ~s)
~ r, y e S . (37)
xr y
4|~ ~s|
The decaying exponential is due to the random phase, which is approximately Gaussian, and
the scattering mean free path is defined by
82o
S= . (38)
(2)5/2 2 `
The lower bound on in (35) implies that

S  L = O(|~
xr y
~ s |), (39)

so the wave recorded at the array is incoherent i.e.,

~ r )] 0.
E[p(t, x (40)

The kernel of the imaging function (9) is obtained from equations (7), (19) and (34),
Nr
X  |x x 0 |  Z d
0 0 r r
K(~ ~ ,z
y, y ~, z
~)=
0
r,r =1
X 2
Z
de
b  e b
 e  b? 
e b
f + f G(, x ~ r, z
~ )
2 2 2



? 0 i + e
(~ y )+i
xr ,~ e
(~ y0 )
xr0 ,~
G (, x
b ~r ,z
0 ~ )e 2 2 , (41)

where we replaced y ~ s by z
~ and y ~ 0 , to avoid carrying over the source indexes. We use
~ s0 by z
the definitions of fb, and b given in section 2.1, and replace the sum over the receivers by
the integral over the aperture, with Gaussian apodization,
Nr
|x|2
Z
X Nr
2(a/6)2
dx e .
r=1
a2 R2

A.1 Wave decorrelation and the paraxial approximation

Essentially the same calculation as in [3, Section 4] shows that the kernel is statistically
stable i.e., it is approximated by its expectation, when

X = O(Xd )  a.

The random travel time model accounts only for wavefront distortion and does not take into
consideration delay spread due to scattering. Thus, the bandwidth does not play a big role
in the statistical stability. However, in stronger scattering regimes the bandwidth is very
important [5] and statistical stability is achieved if

= O(d )  B,

15
as we assume here.
The second moment formula is derived in [7, Appendix B]
  

e  
?
b ,x
e 0 0
E G b + ,x ~,z ~ G ~ ,z ~
2 2
0 0 0 0
eie (~x,~x ,~z,~z )+ie (~x,~x ,~z,~z ) 1 ||~x~z||~x0 ~z0 ||
e S
(4)2 |~ xz x0 z
~ ||~ ~ 0|
1 e2
(|z 0 z|2 +(z 0 z)(x0 x)+|x0 x|2 )
2X 2 22
e d d , (42)

with and e defined in (10)(11). It decays with the frequency and cross-range offsets, due
to the decorrelation of the waves in the random medium. The decorrelation frequency and
length are
d
 
2o o
d =  o , X d = 3`  `, (43)
(2)5/4 `L o
where the inequalities are implied by (35).
As stated in section 2.1, we consider a paraxial wave propagation regime, where
 |x z|2 
o (~ ~ ) = ko |~
x, z xz
~ | ko z3 + , (44)
2L
with negligible residual
 a4   a2 D 
3
O 3
+O 2
 1. (45)
o L o L
Here we used the scaling relation (16). We also approximate the amplitude of the Greens
functions by
1 1
. (46)
4|~xz~| 4L
Since the expression (42) is large when the cross-range offsets are O(Xd ), we estimate
from (38), (35) and (45) that

||~
xz x0 z
~ | |~ ~ 0 ||
 1, (47)
S
so the decaying exponential in the second line of (42) is approximately equal to 1.

A.2 Calculation of the imaging kernel

Because of the decorrelation of the waves over the scale Xd , it suffices to consider imaging
points at cross-range offsets |y y 0 | O(Xd ). Substituting the results in (41), and using e

16
and Xe defined in (25)(26), which are similar to d and Xd , we obtain that
2
0 0 Nr2 2
|z|
e

K(~ ~ ,z
y, y ~)
~, z e 2X
d
2(4L)2 Ba4
|x|2 e 2
|x|
Z Z
2 x e z
e
2X 2
d
2 2Xe
dx e (a/6) de
xe
R2 R2
Z
(o )2
 x(z
e y) x(zy)

zzyy
d e 2B2 +i co ze3 ey3 L L + L
e e e e


Z e2
 x(zy) |z|2 |y|2


2 +i co z 3 y 3
e
L
+ 2L
de
e e 2 , (48)

with center vectors (y, y 3 ) and difference vectors (ye, ye3 ) defined in (20)(21). The vectors
(z, z 3 ) and (e
z , ze3 ) are defined the same way, by replacing y ~ 0 with z
~ and y ~ 0 in (20)(21).
~ and z
Note that in (48) we neglect the phase terms
 2
y |2 e Xd2
  
e |e z | |e
=O 1
co 4L co L
and
e Xd2
 

exe (ez ye)
=O  1,
co L co L
with the inequalities implied by (27), (35) and (43).
Carrying out the Gaussian integrals in (48), we obtain
0 0
n z |2
|e 2
K(~ ~ ,~
y, y ~ ) H exp
z, z
2Xd2 2 1 + ||22


2
2 + 2 + 2i 2 B/o o o
2 2
+i , (49)
2 1 + B2 B
o

where the amplitude factor is

Nr2 Xe2 e
H= 2
1/2
,
2 B 2 1/2
288a2 L2 1 + ||


2 2 1+ o2

and we used the positive constants and defined by

o Xe 1 X2 3
= , = 1 e2 > .
e (a/6) 4Xd 4
The inequality on is implied by definition (26). We also introduced the notation
zy ze ye
= , e = ,
L/(ko Xe ) 6 2L/(ko a)
2 2
z 3 y 3 + |z| 2L
|y|
|(z, z 3 )| |(y, y 3 )|
= ,
co /e co /e
e2
|| ||2 ||
e 2 | | e2
2 = ||2 + 2
+   ,
1 + ||
22
2
2 1 + 22 ||2

17
and
Xe2 ze (z y)

B z ze y ye
= ze3 ye3 + +
co L 2Xd2 L
#
e
2

.
2ko 1 + ||
2 2

The expression (49) simplifies, because it is large only when ||, ||, || = O(1). Since
B  o , we can write
2B 2
1+ 1,
o2
and neglect the 2 B/o phase, with absolute value much less than 1. We also have that ||
and ||
e are O(1), because = O(1), and therefore
   
B  z ze y ye  B  z (ez ye) + ye (z y) 
co  L  = co  L 
    
B Dko |e z ye| Xko |z y|
= O +O
o L L
    
B D X
= O +O  1.
o a Xe
Similarly, we obtain that
 2 
B  Xe ze (z y) 
   1,
co  2X 2 L 
d

and
   
B  e  B
=O .

co  ko 1 + ||22
 o

2

The definitions of , Xe and e give

 
`
=O  1.
a
We assume that this number is not too small, so that B/(o )  1, and we can use the
results above to approximate
B
(e z3 ye3 ).
co
We can now rewrite the kernel (49) as
n
o z |2
|e (ez3 ye3 )2
K(~ ~ 0, z
y, y ~ 0 ) H exp i
~, z
B 2Xd2 2(co /B)2
 2
|(z, z 3 )| |(y, y 3 | 2 o
2

. (50)
2(co /e )2 1 + ||2 2
2

18
Taking absolute value and using that
e2
||
2 ||2 + ,
||2
1+ 22
r  
||2
we obtain the result (24) with 1 = 6 2 1 + 22
O(1).

Appendix B
Suppose that there exists a constellation Y0 of sources such that E est = E (Y0 ). We show
here that
Y est = Algorithm(E est , {~
z0 }) (51)
returns a set Y est such that
E Y est = E est .


(52)
Once we show that Y est 6= , it is straightforward to see from the definition of Algorithm
that (52) must holds. It remains to show that Y est is not an empty set.
For a proof by contradiction, suppose that (51) returns Y est = . This means explicitly
that each call of Algorithm results in executing line 14. Since the set E(Y) of offsets is
translation invariant, let us replace the set Y = {~ ~ ns } of source locations by the set
y1 , . . . , y
Yo = {~ ~1 , z
z0 , z ~ns 1 } of translated source locations, with translation defined by z
~2 , . . . , , z ~o .
Define the vectors
~ejk = z~k z
~0 , k = 1, . . . , m, m = ns 1, (53)
which belong to E est . Note that E est contains other offset vectors, as well. Without loss of
generality, we can assume that the vectors (53) are enumerated in order in E est , meaning
that ~ejk comes before ~ejk+1 for k {1, . . . , m 1}.

1st -Recursion: In the first call of Algorithm, the arguments are E = E est and Y = {~ z0 }.
Since we assume (51) returned , the line 14 was executed. Thus in the while loop,
each element of E is selected at line 5. In particular, at line 5, the vector ~e = ~ej1 is
removed from the set E, and at line 6 we have

~=z
y ~0 + ~e = z
~0 + ~ej1 = z
~1 .

But then
{ ({~ ~ )} = {~ej1 } E est .
z0 } y
Therefore line 8 has to be executed. This will take us to the next recursion.

2nd Recursion: At this recursion level, E contains the offset vectors ~ej2 , . . . , ~ejm , and Y =
{~ ~1 }. We note that Y Y0 at every recursion level.
z0 , z
Once again, by assumption, line 14 was executed. Therefore, at some iteration of the
while loop, at line 5, the vector ~e = ~ej2 is removed from the set E, and in line 6,

~=z
y ~0 + ~e = z
~0 + ~ej2 = z
~2 .

19
 We have
{ (Y y ~2 )} E est
~ )} { (Y0 z
since Y Y0 , so line 8 must be executed. This takes us to the next recursion, with E
containing the offset vectors ~ej3 , . . . , ~ejm , and Y {~
y } = {~ ~2 }.
~1 , z
z0 , z

mth Recursion: At this recursion level, E contains the element ~ejm and Y = {~ ~m1 }.
~1 , . . . , z
z0 , z
Similar to above, we know that at some iteration of the while loop, at line 5, the vector
~e = ~ejm is removed from the E , and in line 6,

~=z
y ~0 + ~e = z
~0 + ~ejm = z
~m .

Moreover
{ (Y y ~m )} E est
~ )} { (Y0 z
since Y Y0 . Therefore line 8 has to be executed, and this take us to next recursion
with all the vectors (53) removed from E and Y {~ y } = {~ ~1 , . . . , z
z0 , z ~m }.
~m1 , z

(m + 1)th Recursion: At this recursion level since

Y = {~ ~ m } = Yo
~1 , . . . , z
z0 , z

and since E est = E (Yo ), line 2 is executed and the recursion returns Y est = Y , which
is a non-empty set.

This contradicts our assumption Y est = . Thus, we conclude (51) returns a non-empty Y est .

Acknowledgment
This work is supported in part by the NSF grant DMS1510429 and by the NSF Grant
DMS-1439786, while Liliana Borcea was in residence at the Institute for Computational and
Experimental Research in Mathematics in Providence, RI, during the Fall 2017 semester.

References
[1] J. M. Beckers, Adaptive Optics for Astronomy: Principles, Performance, and Applica-
tions, Annual Review of Astronomy and Astrophysics. 31 (1), 1993, pp. 13-62.

[2] N. Bleistein, J. K. Cohen and W. John Jr. Mathematics of multidimensional seismic

imaging, migration, and inversion, Vol. 13. Springer Science & Business Media, 2013.

[3] L. Borcea, J. Garnier, G.C. Papanicolaou and C. Tsogka, Enhanced statistical stability
in coherent interferometric imaging, Inverse problems, 27, 2011, p. 085004.

[4] L. Borcea, G.C. Papanicolaou and C. Tsogka, Adaptive interferometric imaging in clutter
and optimal illumination, Inverse Problems, 22, 2006, p. 1405.

20
[5] L. Borcea, G.C. Papanicolaou and C., Tsogka, Asymptotics for the space-time Wigner
transform with applications to imaging, Stochastic Differential Equations: Theory and
Applications. P.H. Baxendale and S.V. Lototsky, Editors, Interdisciplinary Mathematical
Sciences, vol 2, World Scientific, 2007.

[6] L. Borcea and I. Kocyigit, Resolution analysis of imaging with `1 optimization, SIAM
Journal on Imaging Sciences, 8, 2015, pp. 3015-3050.

[7] L. Borcea and I. Kocyigit, Imaging in random media with convex optimization, SIAM
Journal on Imaging Sciences 10.1, 2017, pp. 147-190.

[8] A. Chai, M. Moscoso, and G. Papanicolaou, Robust imaging of localized scatterers using
the singular value decomposition and `1 minimization, Inverse Problems, 29, 2013, p.
025016.

[9] A. Chai, M. Moscoso, and G. Papanicolaou, Imaging strong localized scatterers with
sparsity promoting optimization, SIAM Journal on Imaging Sciences, 7, 2014, pp. 1358-
1387.

[10] J.C. Curlander and R.N. McDonough, Synthetic aperture radar, New York, NY, USA:
John Wiley & Sons, 1991.

[11] L. Devroye, Nonuniform random variate generation, Handbooks in operations research

and management science, 13 (2006), pp. 83-121.

[12] A. C. Fannjiang, T. Strohmer, and P. Yan, Compressed remote sensing of sparse objects,
SIAM Journal on Imaging Sciences, 3, 2010, pp. 595-618.

[13] L.A. Gorham, B.D. Rigling and E.G. Zelnio, A comparison between imaging radar and
medical imaging polar format algorithm implementations, Proceedings of SPIE, 6568,
65680K, 2007.

[14] J. A. Jackson and R. L. Moses and E. G. Zelnio and F. D. Garber, Feature extraction
algorithm for 3 D scene modeling and visualization using monostatic SAR, in Proceedings
of SPIE, Vol. 6237, No. 1, 2006, pp. 55-66.

[15] R. O. Nielsen, Sonar signal processing, Artech House, Inc., 1991.

[16] P. Kutchment and L. Kunyansky, Mathematics of thermoacoustic and photoacoustic

tomography, Handbook of Mathematical Methods in Imaging, O. Scherzer, ed., Springer-
Verlag, 2010, pp. 817-866.

[17] S.M. Rytov, Y.A. Kravtsov, and V.I. Tatarskii, Principles of statistical radiophysics 4.
Wave Propagation through random media, Springer Verlag, Berlin, 1989.

[18] T. L. Szabo, Diagnostic ultrasound imaging: inside out, Academic Press, 2004.

[19] M. C. W. van Rossum, T. M. Nieuwenhuizen, Multiple scattering of classical waves:

microscopy, mesoscopy, and diffusion, Biomedical Optics Express 71 (1999) 313-371.

21
[20] T. Eren and O. K. Goldenberg and W. Whiteley and Y. R. Yang and A. S. Morse and
B. D. O. Anderson and P. N. Belhumeur, Rigidity, computation, and randomization in
network localization, IEEE Infocom, 2004.

[21] A. Javanmard and A. Montanari, Localization from Incomplete Noisy Distance Mea-
surements Foundations of Computational Mathematics,2013.

22