IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 44, NO.

5 , MAY 1996

I184

Blind Separation of Synchronous CO-Channel Digital

Signals Using an Antenna Array-Part I: Algorithms
Shilpa Talwar, Student Member, IEEE, Mats Viberg, Member, IEEE, and Arogyaswami Paulraj, Fellow, IEEE

Abstruct- We propose a maximum-likelihood (ML) approach

for separating and estimating multiple synchronous digital signals
arriving at an antenna array. The spatial response of the array
is assumed to be known imprecisely or unknown. We exploit
the finite alphabet property of digital signals to simultaneously
estimate the array response and the symbol sequence for each
signal. Uniqueness of the estimates is established for signals
with linear modulation formats. We introduce a signal detection
technique based on the finite alphabet property that is different
from a standard linear combiner. Computationally efficient algorithms for both block and recursive estimation of the signals are
presented. This new approach is applicable to an unknown array
geometry and propagation environment, which is particularly
useful in wireless communication systems. Simulation results
demonstrate its promising performance.

I. INTRODUCTION
IRELESS communication systems are witnessing rapid
advances in volume and range of services. A major
challenge for these systems today is the limited radio frequency spectrum available. Approaches that increase spectrum
efficiency are therefore of great interest. One promising approach is to use antenna arrays at cell sites. Array processing
techniques can then be used to receive and transmit multiple
signals that are separated in space. Hence, multiple co-channel
users can be supported per cell to increase capacity. In this paper, we study the problem of separating multiple synchronous
digital signals received at an antenna array [l]. The goal is
to reliably demodulate each signal in the presence of other
co-channel signals and noise. The complementary problem of
transmitting to multiple receivers with minimum interference
at each receiver has been studied in [2]-[4].
Several algorithms have been proposed in the array processing literature for separating co-channel signals based on
availability of prior spatial or temporal information. The traditional spatial algorithms combine high resolution directionfinding techniques such as MUSIC and ESPRIT [5], [6] with
optimum beamforming to estimate the signal waveforms [7],
[8]. However, these algorithms require that the number of
Manuscript received December 7, 1994; revised October 16, 1995. The
work of S. Talwar was suppoited by the Computational Science Graduate
Fellowship Program of the Office of Scientific Computing, U.S. Department
of Energy. The associate editor coordinating the review of this paper and
approving it for publication was Prof. Michael D. Zoltowski.
S. Talwar is with the Scientific Computing and Computational Mathematics
Program, Stanford Univerity, Stanford, CA 94305 USA.
M. Viberg is with the Department of Applied Electronics, Chalmers
University of Technology, S-41296 Gothenberg, Sweden.
A. Paulraj is with the Information Systems Laboratory, Stanford University.
Stanford, CA 94305 USA.
Publisher Item Identifier S 1053-587X(96)03072-3.

signal wavefronts including multipath reflections be less than

the number of sensors, which restricts their applicability in a
wireless setting. In the recent past, several property-restoral
techniques have been developed that exploit the temporal
structure of communication signals while assuming no prior
spatial knowledge. These techniques take advantage of signal
properties such as constant modulus (CM) [9], discrete alphabet [lo], [ l l ] , self-coherence [12], and high-order statistical
properties [13], [14]. In this paper, we propose a new propertyrestoral approach that takes advantage of the finite alphabet
(FA) property of digital signals.
Our approach is termed blind since it does not require any
training signals for signal demodulation. This is particularly
useful in situations where training signals are not available.
For example, in communications intelligence, training signals
are not accessible. In cellular applications, blind algorithms
can be used to reject interference from adjacent cells. In IS54, for example, adjacent cell interference appears only over a
partial burst and training signals do not help. Blind algorithms
are also bandwidth-efficient due to the elimination of training
sets. Moreover, the study of blind algorithms can be used to
complement existing non-blind techniques. For example, as a
result of our investigation of uniqueness for the blind problem,
we propose a minimal set of training signals that can be used
in a non-blind multi-user scenario. An important advantage of
our approach is that in non-blind scenarios, training sets can
easily be incorporated to initialize our algorithms.
The algorithms presented herein can be used to demodulate
multiple synchronous digital signals in a coherent multipath
environment. To guarantee unique signal estimates, we assume
that the number of signals does not exceed the number of sensors, and that the channel is constant over a sufficient number
of snapshots. The synchronous assumption is reasonable in
microcell-air interfaces where symbol timing can be effectively controlled. Extension of our approach to asynchronous
transmission and delay spread channels is relatively straightforward, the main differences in these scenarios being (i)
instead of matched filtering, the array output is oversampled;
and (ii) the signal estimation step is replaced by maximumlikelihood sequence estimation (MLSE) since the channel is
no longer memoryless. Recently, some non-ML techniques
have also been proposed that use subspace information to first
synchronize the signals and remove intersymbol interference,
and then use one of the synchronous FA algorithms presented
herein to separate the signals [15]-[17].
The outline of this paper is as follows. We introduce the data
model in Section 11. In Section 111, we consider the problem

1053-587)3/96$05.00 0 1996 IEEE

TALWAR et ul : SYNCHRONOUS CO-CHANNEL DIGITAL SIGNALS-PART I

of uniquely identifying the signals when the array response

structure is unknown. In Section IV, the ML estimator for
the array responses and symbol sequences is discussed. Two
efficient block algorithms are presented in Section V, and
their convergence is analyzed. Recursive extensions of these
algorithms are discussed in Section VI. In Section VII, we
present simulation results to demonstrate the performance of
these algorithms. Finally, we conclude with directions for
future work in Section VIII.

1185

For simplicity, we assume that the symbols belong to the

alphabet R = (k1,
&3, . . . , f ( L - 1)) for real signals, and
fl = {&1,&3,.. . ,&(L-I)j~B{ztj,*,j3... . . & j ( L - 1 ) } for
complex signals. These correspond to the important cases of
PAM and QAM modulation formats.
Now, assuming that the signals are symbol-synchronous, we
perform matched filtering on (2), and sample the filtered array
output at symbol rate. This yields the following equivalent
discrete representation of the data:
d

11. PROBLEMFORMULATION
Consider d narrowband signals impinging at an array of
m, sensors with arbitrary characteristics. The signal waveform
received at each sensor is demodulated with respect to the carrier frequency (assuming perfect carrier phase lock recovery).
The m x 1 vector of sensor outputs, x(t), in the absence of
multipath, is given by
d

x(t) = X P k a ( B k ) s k ( t )

+ v(t)

=CPkakbk(n)

X(.)

(1)

k=l

.(e,)

where l ) k is the amplitude of the kth signal,

is the
array response vector to a signal from direction O h . s k ( . ) is
the kth signal waveform, and V( .) is additive white noise with
covariance n21.
In a realistic communication scenario, however, there are
multiple reflected and diffracted paths from the source to the
array. These paths arrive from different angles and with different attenuations and time delays. The array output becomes

k=l1=1

where q k is number of subpaths for the kth signal, and

a k l and q~ are, respectively, the attenuation and time delay
corresponding to lth subpath. We assume that the propagation
delays associated with these paths is much smaller than the
inverse bandwidth of the signals. The delays can thus be
modeled as phase-shifts under the narrowband assumption.
The new data model becomes

+ v(n).

(3)

k=l

The noise term ~ ( nremains

)
white, and it is easily seen that
the output of the matched filter is a sufficient statistic for
determining the transmitted symbols [18]. We can rewrite (3)
in matrix form

~ ( n=)As(n) + ~ ( n )

(4)

where
x(n) is
the
filtered
data,
s(n)
=
[bl(n). . . b d ( n ) l T , v(n) is additive white noise, and
A is an m x d matrix of array responses scaled by the signal
amplitudes A = hlal . . .pdad].
Assuming that the channel is constant over the available N
symbol periods, we obtain the following block formulation of
the data

X(N) = AS(N) + V ( N )

(5)

where X(N) = [ x ( l ) . . . x ( N ) ] , S ( N=
) [ s ( l ) . . . s ( N ),]and
V ( N )= [ v ( l ) .. .v(N)].The matrix A represents the spatial
structure of the data, and the matrix S represents its temporal
structure. The problem addressed in this paper is the combined
estimation of the array response matrix A and the symbol
matrix S ( N ) ,given the array output X ( N ) . We assume that
the number of signals is known or has been estimated [19]. For
notational convenience, we denote X G X ( N )and S S ( N ) ,
from here on.
111. IDENTIFIABILITY

Before discussing the estimation problem, we consider the

d
problem
of uniqueness of signal estimates in the absence of
x(t) = Cpkaksk (t) ~ ( t )
(2)
noise. This problem can be viewed as a nonlinear factorization
k=l
~ factors A m x d and S d x ~ such
,
of the data matrix X m x into
where a k is now the total array response vector a k = that X = AS. In the case that columns of matrix A lie on
c v k l e - J w r T ~ a ( O k , ) , and w, is the carrier frequency. The
the array manifold (defined as the set {.(e): 0 E [ 0 , 2 ~ ] } )
spatial structure of the array response vector a k cannot be and S is an an arbitrary full-rank matrix, it is well known
exploited if the number of paths is larger than the number of that this factorization is unique provided (i) any set of m
sensors. However, we can exploit the temporal structure of vectors from the array manifold is linearly independent; and
digital signals with memoryless linear modulation formats
(ii) d < m [20].There is an ordering ambiguity in the signal
N
estimates, since

x
where N is the number of symbols in a data batch (burst),
{bk(.)) is the symbol sequence of the kth user, T is the
symbol period, and 9(.) is the symbol waveform. We let
9 ( . ) be a square-root raised cosine waveform so that matched
filtering yields a pulse that satisfies the Nyquist criterion.

AS = A P ~ P S
=As

where A and S is also a valid solution pair for any permutation

matrix P .
Our problem is the opposite of the standard problem. We
assume that A is an arbitrary full-rank matrix, but the elements
of S belong to a finite alphabet R. We first consider the

1186

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 44, NO. 5, MAY 1996

case of binary antipodal signals, i.e., R = {*1}, and then

generalize to larger alphabets. As before, the solution to the
nonlinear system of equations X = AS is not unique under
our assumptions. For example, A = AT and S = T-lS is a
solution for any non-singular matrix Tdxd that is a diagonal
with f 1 entries or a permutation matrix or a product of the
two. In this case, there is an additional ambiguity in the sign
of the estimated signals, beyond the usual ordering ambiguity.
However, the sign ambiguity can be easily removed once the
signals have been estimated by appropriate decoding of each
symbol sequence. Hence, the sign and ordering ambiguities do
not present a serious problem as the correct signals are still
received. We define the system X = AS to be identifiable if
all simultaneous solutions can be written as AT and T-lS,
where T is a non-singular matrix with exactly one non-zero
element (+1 or - 1) in each row and column. For convenience,
we call such a matrix an admissible transform matrix (ATM).
The following theorem gives a sufJicient condition for the
identifiability of binary signals.
Theorem 3.1: Let X = AS where Amxd is an arbitrary
~ a full-rank matrix
full-rank matrix with d 5 m, and S d x is
with i 1 elements. If the columns of S include all the 2d-1
possible distinct (up to a sign) d-vectors with f l elements,
then A and S can be uniquely identified up to a matrix T
with exactly one non-zero element { 1 - l} in each row and
column.
Pro03 It is easily seen that the condition on S also
implies that it is full rank. Suppose there exists another pair,
A and S, both full rank, such that X = AS. Then

Dropping the superscripts in (11), we have the following

equivalent system of equations:

where S is an arbitrary full-rank matrix of il elements, and

S is a normalized matrix with the first n = 2d-1 columns
distinct. We can partition S = [S, S,], where S, is the
submatrix of n distinct columns and S, is the submatrix of
remaining columns. Note that the columns in S , are repeated
from the first n, and thus provide no extra information in
determining T . Therefore, we consider only the equations
defined by S,. Each row t of T must satisfy t.TS, = S z ,
where S: is a subvector of the corresponding row of and can
be of any one of the 2" possible n-vectors. More explicitly, for
tT = [ t d . . tztl],we have a system of equations of the type
td

td -

td

(id-1

+ . .. + +tl)= i 1

(td-1 +

t2

' ' '

+t 2 +tl)=

+ (td-1 + . + t 2
(td-1 + . .. + t 2

td -

' '

t l ) = It1
t l ) = fl.

In the Appendix, we show by induction that the solution of

this system has the form t; = k l for some i and tj = 0
for j = 1:. . . d ; j # i. This implies that each row of S is
X = AS = AS.
(6) a multiple of a row of S. Since S has rank d , all rows of S,
up to a sign, are included in S. Therefore, the rows of T are
Solving for A, we get
distinct, and it is an ATM.
0
The identifiability of larger alphabets follows easily from the
A = ASSt
(7)
binary case. Consider the alphabet R of length L with symbols
where (.)t denotes the pseudo-inverse. We multiply (7) by S {fl, 1 3 ; . . . f(L - I)}. We assume that the columns of S
to obtain
include all the 1 = L d / 2 possible distinct (up to a sign) dvectors. In particular, they include n = 2d-1 distinct vectors
AS = ASStS
of il elements and n = 2d-1 distinct vectors of k ( L - I)
and together with (6), this implies A(S - SStS) = 0. Since elements. As before (without loss of generality), we can
A is full rank, we must have
normalize each column of S such that the first symbol is
positive, and permute the columns so that we can partition
s = SStS.
(8) S = [S, (L- l)S, S,], where S, is a submatrix of n distinct
Note StS is a projection matrix that projects the rows of S vectors with fl elements and S, is the remaining submatrix.
onto the row space of S. Thus, from (8), we see that S and S Each row t of T again satisfies
must share a common row space. That is
tTS = tT[S, (L- 1)s" S,] = ST
(12)
S = TS.
(9)
where ST is the corresponding row of S, and is partitioned
Now, it remains to show that T is an ATM.
likewise as ST = [Sr, Szn ST]. Although there are 1 distinct
We normalize the first element of each column of S to +l
equations defined by the columns of S, the subset of equations
by postmultiplying (9) by a diagonal matrix D with diagonal determined by the first 2n columns alone results in a trivial
elements D,, = SI, ( j = 1 . . . N )
solution for t, i.e., only one non-zero element. This follows
from (12)
SD = TSD jS(l) = TS(l).
(10)
tT s, = s;,
(13)
Next, we postmultiply the resulting system (10) by a N x N
tT ( L - l)S, = ST"
permutation matrix that reorders the columns of S(l) to make
(14)
;he first 2d-1 columns distinct
Since the elements of both
which implies (L - l)ST,, = "'.
S ( l ) p= TS(1)p jS(2)= TS(2).
(11) SI,, and % 2 , % belong to the alphabet Ci, the only possibilty for

TALWAR et al.: SYNCHRONOUS CO-CHANNEL DIGITAL SIGNALS-PART

the entries of SI.^ are f l s . Thus, from (13), we see that the
problem is reduced to one of identifiability of binary signals.
For this problem, we have shown previously that t is a trivial
solution, and consequently, T is an ATM.
Finally, the generalization to complex signals is straightforward. In this case, we see that the-signals can be identified
uniquely up to a factor { 1 - 1,+ j , - j } . As before, we have
the system of equations S = T S , where S and S are now
complex matrices with elements in R = { f l ,5 3 , . . . f ( L 1)) 69 {ij,f j 3 , . . . , f j ( L - 1)). We are interested in finding
a condition on the columns of S such that T is an ATM. For
complex signals, we extend the definition of an ATM to a nonsingular matrix with one non-zero element{ & 1,kj), in each
row and column.
We begin by noting that multiplication of complex matrices
is isomorphic to multiplication of real matrices with twice the
dimensions [21]. In particular, we have

Rc{S)
[Im{S}

R.e{T} -Im{T)
Rc{S}] = [Im{T}
Re{T}
Re{S} -Im{S}
[Im{S}
Re{S}

-Im{S}

and by denoting each of the real block matrices above

with a subscript (.)n, we get S R = TRSR. The matrix of received signals Sn now has dimensions d x
2 N , for n! = 2d. If- the first N columns of S R include all the 1 = L d / 2 possible distinct vectors, then
1 or a - 1,
there is exactly one non-zero element, a
in each of the d rows of the submatrix [Re{T} - Im{T}].
Furthermore, each row of this submatrix must be distinct since
S is full rank, and thus, T = Rc(T} + j I m { T } is an ATM. It
is important to note here that the identifiability of d complex
signals is equivalent to that of d = 2d real signals. We can
summarize the above discussion by the following theorem.
Theorem 3.2: Let X = AS where A,n,xdis an arbitrary
~ ~ matrix
full-rank matrix with d 5 m, and S ~ ~is xa lfull-rank
with elements in R.
1) Real Case: If the columns of S include all the L d / 2
possible distinct (up to a sign) d-vectors with elements
in fi = {*I, f 3 , . . . f ( L - I)}, then A and S can be
uniquely identified up to a matrix T with exactly one
non-zero element, { + l )-l}, in each row and column.
2) Complex Case: If the columns of S include all the L 2 d / 2
possible distinct (up to a sign) d-vectors with elements in
, i ( L - 1)) {f,j;f , j 3 , .. . , k j ( L I)}, then A and S can be uniquely identified up
to a matrix T with exactly one non-zero element,
{+l.-1, + j , - , j } , in each row and column.
In Theorem 3.2, we give a sufficient condition for identifiability of signals that belong to a finite alphabet 0. Now,
we show that if N is sufficiently large, the probability of
achieving identifiability approaches I. We consider the case
with real signals, noting that the following result also holds
for complex signals, with d replaced by d.
Theorem 3.3: Let p denote the probability of receiving
all the L d / 2 distinct (up to a sign) columns of S in N
independent snapshots, N >> L d / 2 .Assuming that the signals

1 I87

are independent, and the probability of receiving each symbol

in R is equal, p is bounded by

(16)
Proof See the Appendix A.2.
U
We see from (16) that for a fixed value of d, as N increases,
p approaches 1. The probability q = 1- p that one of the L d / 2
distinct vectors is not picked can then be bounded by

Hence, the probability of missing one of the distinct vectors

approaches zero exponentially fast, the rate of decay depending
on the ratio ( L d / 2 ) '
Nevertheless, for large values of d, the number of snapshots
required for identifiability seems quite large. But the condition
requiring L d / 2 distinct vectors is only sufficient, and far from
being necessary. In fact, we see in Appendix A.3 that if S
contains a specific set of d 1 columns for R = {fl},it
can be determined uniquely up to an ATM. Although this
result remains to be generalized, it shows that in practice
identifiability can be achieved with far fewer snapshots.

IV. MAXIMUM-LIKELIHOOD
ESTIMATION
In this section, we consider the problem of estimating
the digital signals in the presence of noise. From Section
11, we see that the signals can be modeled as unknown
deterministic sequences corrupted by white Gaussian noise:
~ ( n=
) As(n)
~ ( nwhere
)
[v(n)v(k)*] = 0 2 1 S n k .
The maximum-likelihood (ML) estimator yields the following
separable least-squares minimization problem:

in the variables A and S ( N ) , which are, respectively, continuous and discrete. We assume N is large enough to ensure
unique signal and array response estimates.
It is proved in [22] that the minimization can be carried out
in two steps. First, we minimize (17) with respect to A since
it is unconstrained

A = X ( N ) S ( N ) t= x(N)s(N)*(s(N)s(N)*)-l
Then, substituting A back into (17), we obtain a new criterion,
which is a function of S ( N ) only as follows:

where P&Ar)= IN - S(N)*(S(N)S(N)*)-'S(N).

The
global minimum of (18) can be obtained by enumerating over
all possible choices of S ( N ) . However, this search has an
exponential complexity in the number of symbols N and the
number of signals d (2 for complex signals), and can be
computationally prohibitive, even for modest-size problems.
In the next section, we consider two iterative block algorithms
that have a lower computational complexity.

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 44, NO. 5, MAY 1996

1188

= IIRe{X} - Re{A}SII$

V. BLOCKALGORITHMS
The block algorithms ILSP and ILSE were introduced in
[1]. These algorithms take advantage of the ML estimator
in (17) being separable in the variables A and S. The ML
criterion is minimized with respect to the two variables using
an alternating minimizations procedure. The idea is to visit the
received data iteratively until a best fit with the channel (array
response) and signal model is obtained. The ILSP algorithm
performs well with no prior estimate of the array responses,
and can be used to initialize ILSE. For a sufficiently good
initialization, the ILSE algorithm converges rapidly to the ML
estimate of the array responses and signal symbol sequences.
Apart from their efficiency, the algorithms are naturally parallelizable, and can be easily extended for recursive estimation
(see Section VI).
A. ILSP Algorithm

We begin by assuming that f ( A , S; X ) = IIX - AS/I; in

(17) is a function of unstructured continuous matrix variables
A and S. Given an initial estimate A of A, the minimization
of f ( A , S; X) with respect to continuous S is a least-squares
problem. Each element of the solution S is projected to its
closest discrete values, say S.AThena better estimate of A is
obtained by minimizing f ( A , S; X ) with respect to A, keeping
S fixed. This isAagaina*least-squaresproblem. We continue this
process until A and S converge.
Iterative Least-Squares with Projection (ILSP):
1) Given Ao, k = 0
2) k = k + l

Sk = (A;-~A~-~)-'A;-~X
* S k = proj[Sk]
e

Ak = XS;(SkS;)-'
3) Repeat 2 until (Ak, S k ) = (Ak-1; Sk-1).
In the above description of the algorithm, proj[.] implies
projection onto a discrete alphabet. However, this definition
may be extended to projection onto a constant modulus or
any other signal characteristic. Moreover, in scenarios where
the array manifold structure is applicable, A k can be projected
onto the manifold at each iteration. Hence, the ILSP algorithm
outlines a general approach for imposing known structure on
variables A and S in a minimization criterion of the form
/ / X- ASll$. The main advantage of the algorithm is its
low computational complexity. At each iteration, two leastsquare problems are solved, each requiring U ( N m d ) flops for
N >> m. In particular, ~ m +d2 d 2 ( ~ $1 + rnd2 flops are
required to solve for A, and Nmd+2d2(m- $ ) + N d 2 flops to
solve for S. Thus, the algorithm's complexity is polynomial
in N and d.
Real Signals: If the signals belong to a real alphabet, we
take advantage of this fact by constraining the imaginary part
of S to be zero in each step, and thereby reducing the number
of unknowns by half. Equivalently, we can minimize a slightly
modified criterion f ( A n , S; X,) by noting that
IIX - ASI/$
= II[Rc{X) - Re{A}S] +j[Im{X}

Im{A)S]/l;

= IIXR

+ llIm{X}

Im{A}SII$

ARSII$

where XR and AR are real augmented matrices as shown

above. Thus, the algorithm proceeds as before, the only
difference being that for real signals, we consider augmented
matrices and replace (.)* by (.)'.
Initialization: A common initialization strategy in optimization for nonlinear problems with mixed discrete and
continuous variables is to use the solution of the continuous
problem as an estimate for the mixed problem [23]. The
continuous solution for the ML criterion in (18) is Vd, the right
singular vectors corresponding to the largest d singular values
of of X . Hence, we can initialize ILSP with So = proj[Vd].
It is easily seen that this is equivalent to initializing with
A0 = Ud, the largest d left singular vectors of X . Since
it is expensive to compute the singular value decomposition
(SVD), we initialize instead with Xd, the first d rows of X
(equivalently A0 = Imx&which have the same rowspan as
Vd in the absence of noise. This approximation works well if
all signals have nearly equal powers. If some signals are much
stronger than the others, then So = proj(Xd) may be rank
deficient. In this case, we add a small diagonal perturbation
matrix to So in order to restore its rank. In general, if S k or
Ak becomes rank deficient at any iteration, we have found this
simple strategy of perturbing the diagonal to be quite effective
in extracting all the signals in difficult scenarios. One may also
use the sofi-orthogonalization strategy proposed in [9] for the
multiple-source CM problem.
The availability of prior information to initialize the ILSP
algorithm can improve the possibility of global convergence,
and help reduce the number of iterations. For example, in scenarios where imprecise knowledge of the spatial structure of
the array is available, estimates of the array responses obtained
from traditional techniques, such as MUSIC or ESPRIT, can
be used to construct Ao. Alternatively, if the users initially
transmit a short set of training signals, such as the sequence
of d symbols that define the matrix Sd given in (56) (see
the Appendix), then A0 can be estimated from the received
data by Ao = XS;'. The advantage of using this particular
training set is that Si', given in (59), is easily computed and is
sparse. Although the algorithm is no longer blind, revisiting the
data iteratively can significantly improve channel and signal
estimates in situations with low SNR's.
Comparison with Relevant Algorithms: The ILSP algorithm is similar to the LS-CMA and multitarget LS-CMA
algorithms proposed in the CM literature [9], [24]-[26]. There
are two key differences, however. First, the FA property is
stronger than CM for digital signals (with PSK modulation
format) since the signals are restricted to lie on discrete
points on a disk. Second, these algorithms use a MMSE
beamformer to estimate the signal waveforms. They minimize
the following performance criterion using the alternating
projections technique:
niin

W,SECM

//w*x
- sll$.

(19)

TALWAR et al.: SYNCHRONOUS CO-CHANNEL DIGITAL SIGNALS-PART

It is shown in [25] that for a single user and an unknown spatial

noise covariance, the above criterion yields the ML signal
estimate. However, when multiple signals are present, this criterion suffers from the drawback that only the strongest signals
may be captured. For example, consider the case where two
weight vectors have converged to the same (stronger) signal.
The MMSE residual, llW*X - Sll; with S = proj[WCX],
may be small in this case. Yet, the ML residual, IIX - ASllg
where A = X S + , will be large since a best fit of the array
response and signal model to the received data is not obtained.
Now, let us examine the ML and MMSE schemes to see
precisely how they differ. We can express ILSP algorithm
succinctly as
sk+1=

proj [(xs:)+x]

(20)

and a MMSE scheme such as multitarget LS-CMA [26]

without soft-orthogonalization as

S k + l = proj[SkX+X].

(21)

Since the pseudo-inverse does not satisfy ( X S i ) + = SkX+

in general [27], the two algorithms yield different signal
estimates. Note the matrix AI, = XS; in (20) is poorly
conditioned near the solution A, if the angular separation
between array response vectors is small. In contrast, the data
matrix X in (21) becomes poorly conditioned for high SNRs.
The MMSE algorithm is computationally less expensive. However, we have found the performance of the ML algorithm
to be more favorable in a blind multiple-signal scenario. In
[lo], Swindlehurst et al. have proposed a decision-directed
technique for digital signals using both the MMSE and ML
beamformers (similar to ILSP), assuming a rough estimate
of the signal of interest is available. They have shown that
the asymptotic symbol error rate for the MMSE beamformer
W* = Rsx.R&& is lower than the error rate for the ML
beamformer W* = A+. Hence, for large N , the converged
signal estimates obtained from ILSP may be improved by
applying the the MMSE approach.

B. ILSE Algorithm
A limitation of the ILSP algorithm is that its performance
is limited by that of the ML beamformer. This is easily seen
by considering the case where A is known, and the ML
criterion is to be minimized with respect to the variable S
only. In ILSP, S E R is not estimated directly, but in two
steps, (i) least-squares and (ii) projection. The least-squares
step causes noise enhancement if the array response vectors
are not well separated in angle, i.e., A is ill conditioned.
The optimal approach is to enumerate over all possible S
matrices with elements in R,and choose the S that minimizes
I IX - AS 11 $. However, this is computationally demanding,
since LdN matrices need to be considered. Fortunately, the
search can be reduced to enumerating Ld vectors in R ( N
times) by exploiting the following property of the Frobenius
norm:

minI/X-AS//$ = min I l x ( l ) - A s ( l ) l l $ + . . . f
SER

s(1)En

min llx(N) - As(N)IJ$.

s (IV)E R

(22)

I I89

Hence, minimization over the signal vectors s(1) . . s ( N ) can

be carried out independently. For each s ( n ) ,the ML estimate
6(n)is obtained by enumerating over all Ld possible vectors
s ( j ) E R,and choosing the one that minimizes

The ILSE algorithm proceeds in a similar fashion as ILSP.

Given an initial estimate of A (possibly from ILSP), we
iterate using an alternating minimization technique: minimize
f ( A , S; X ) with respect to S E R and then with respect to A,
at each step. The residual function f ( A , S; X) is decreased
at each iteration, and for a reasonably good initial Ao, convergence to the global minimum is rapid. The algorithm has
L ~per
) iteration: ~ m d2 d 2 ( -~
complexity o ( N ~ ~ flops
$) md2 flops to solve for A and NmLd(d 1) flops to
enumerate.

Iterative Least-Squares With Enumeration (ILSE)

Given Ao, k = 0
IC = k + l
Let A = An-1 in (22), and minimize for S I , (enumeration).
AI, = X S i ( S k S i ) -
Repeat 2 until (Ak)S k ) = (Ak-1, Sk-I).
C. Convergence

We have recently discovered that ILSE algorithm is very

similar to the segmental K-means algorithm used for estimating parameters of hidden Markov models [28]. The K-means
algorithm is an iterative scheme that alternates between two
key steps: (i) segmentation performed via generalized Viterbi
algorithm, and (ii) optimization. In [28], Juang and Rabiner
prove fixed-point convergence of the algorithm. Their proof is
based on Zangwills convergence theorem, which is a general
result for algorithms in nonlinear programming [29]. It is
shown that the K-means algorithm satisfies the conditions of
the theorem. Rather than taking the same approach, we present
a simple proof that shows that ILSE algorithm converges to a
fixed point in a finite number of iterations.
We first review some definitions that will be needed for the
proof [29]. Let 7 :V i V be a mapping from a point in space
V to a point in V . An algorithm is an iterative process that
generates a sequence {v~,};rP=~,
given a point VO in V , by
successive transformations VI, = 7(Vk-l). A point V* is a
jixedpoint of T if V* = 7 ( V * ) . Let A be the set of fixed
points of 7. A function f is a descent function for mapping
7 if it satisfies the conditions:
1) f : V + IR is nonnegative and continuous
2) f ( V ) < f ( V ) for V = 7 ( V ) and V 6 A
3) f (V) 5 f (V) for V = 7 ( V ) and V E A.
Many algorithms use the objective function to be minimized
as the descent function f . The conditions 1-3 require that
the algorithm reduce f at each iteration until a fixed point is
reached.
Within this framework, we can consider ILSE algorithm as
a mapping on the cross-product space V = A x S where A =
p X d and S = f l d x N Starting
.
with an arbitrary pair Vo =

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 44, NO. 5 , MAY 1996

1190

(Ao,So), the iterates VI, = (AI,,S I , ) = 7((Ak-1, Sk-1)) are

determined by the following two minimization steps: (i) S I , =
arg minsEn /IX- Ak-1Sll$, and (ii) AI, = arg minA I/XASI,~I'$.In order for the mapping 7 to be well defined, we
need to ensure that for each pair (Ak-1,Sk-l) E V , the
algorithm computes a unique pair (Ah, S I , ) E V . That is,
the algorithm computes a unique S k for a given Ak-1, and
a unique Ah for a given s k . For a given Ak-1, it may be
possible that S k in (i) is not unique, such as when Ak-1 is rank
deficient. In this case, we need to impose a rule that uniquely
defines S k . For example, in a computer implementation of the
algorithm, there is an ordering associated with enumeration,
and we choose S k to be the minimizer of (i) with the lowest
index. With respect to AI, in (ii), given a full rank S I , , the
least-squares estimate Ah is unique. However, if S k is rank
deficient, we need to compute the minimum-norm least-squares
solution for Ak; that is unique.
It is natural to consider the residual function f ( A : S: X) =
/ / X- ASll$ as a descent function of the algorithm. Clearly,
f (A, S; X) is nonnegative and continuous in A and S. Consider

AI,, S k ; X) = IIX - AlcSlcl/%

= rnin IIX - A S I , ~ / $
A

(24)
(25)

Theorem5.1: Let (Ao,So) E C m x d x f l d x N , and the

iterates (AI,;S k ) = I((Ak-1, SI,-^)), k 2 1 be determined
by ILSE algorithm. Let f ( A , S ; X ) be a descent function
of the algorithm. Then, there exists some j o such that for
k 2 jo;(Arc.S~,)= 7((Aj,,,Sj,)) and f ( A ~ , , s k ; X=
)
f (Aj, * s,, ).
Hence, a fixed point is reached in a finite number of steps,
and can be detected by a lack of change in the residual.
The global minimum is a fixed point of the iteration. This
follows from the fact that a transition to another point would
be possible only if the residual is reduced, but this cannot
occur since the global minimum is the point with the lowest
residual. Note that the theorem is valid for any initial guess
(Ao, SO).However, the sequence { (Ah, S k ) } and the fixed
point it converges to depends on the initial guess.
We have found in our simulations that for some initial
iterates, the algorithm converges to fixed points that are not
the global minima. This case can be detected by considering
the magnitude of the residual
IIX - ASll$, where A and
S are converged estimates from ILSE. If a global minimum
is reached, this residual is close to the noise power, since the
true residual

&

/IX - ASll$ = IlVll$ z Ntr(a21)= N m a 2 .

If the residual is not decreased to noise level, we have reached

a non-global solution. In this case, we restart the algorithm
= niin IIX - Ak-1S/I$
with another initial guess. We proceed in this fashion until the
SER
5 11x - Ak-1Sk-l//$
(28) global minimum is reached.
For a wide range of parameters (SNR and array response
= f(Ak-1, S k - 1 : X ) .
(29)
vector separation), ILSE algorithm converges to the global
Based on our assumptions, inequality (26) is strict unless solution when initialized with array response estimates from
Ak = Ah-1 and inequality (28) is strict unless S I , = S k - 1 .
ILSP. If the global solution is not reached, we re-initialize
Hence, at each iteration the residual is strictly decreased unless ILSP with a random guess, and then re-initialize ILSE with
(Ak-1 S k - 1 )
I((Ak-1 S h - 1 ) ) is a fixed point.
the estimate from ILSP. Usually one or two re-initializations
The convergence proof is based on the observation that each are sufficient to yield the global minimum. Since both ILSP
of the iterates S k E f l d x N , belongs to a finite set. Let us and ILSE converge very rapidly, re-initialization is not comassume there are P possible S I , matrices. Since there is a putationally expensive. Hence, the success of our approach is
unique least-squares estimate Ak associated with each S I , , not hindered by the presence of additional fixed points. We
there are only P pairs (Ah, S I , ) that can be generated by cannot apply the above convergence theory to ILSP since the
the ILSE algorithm (for IC 2 1). Consider the sequence of algorithm does not necessarily decrease the residual at each
pairs {(Ak,Sk)}f_+,' obtained from the algorithm. There are iteration. However, in our simulations, we have found that it
P 1 iterates in the sequence taken from a set of P possible also converges to a fixed point in a finite number of steps.
elements. Hence, at least two of the iterates must be the same. Moreover, in scenarios where the array response vectors are
1 are the lowest indexes for which well separated, it converges to the global minimum.
Let us assume j and j
(Aj, S j ) and (Aj+l,Sj+l) are the same. The residual for the
Cost Function: In Fig. 1, we symbolically depict the ML
two iterates is equal as follows:
cost function /IX - AS1I2 in order to understand the path
taken by ILSE algorithm. We consider a scenario with d = 2
f ( A3+1>
. S3-+ l , . X ) = f ( A j , S j ; X ) .
(30)
BPSK signals at 15 dB SNR, arriving from Q = [0, lo]", and
If I = 1, then (A,,Sj) is a fixed point of the algorithm, a block size of N = 3. We choose this simple case due to the
since (Aj+l,Sj+1) = T ( ( A j , S J ) ) = (Aj,Sj). Let us complexity involved in enumerating all possible S matrices.
assume (Aj, Sj) is not a fixed point, and I > 1. Then, Neverthless, this picture gives a good qualitative understanding
f(Aj+l,Sj+l; X) must be strictly less than f ( A j , Sj; X); of the path of the algorithm to a global minimum or some
and, hence
other fixed point.
In this figure, the x-axis corresponds to the 26 possible
(31) signal matrices S ( J ) , j = 1 . . .64. The y-axis corresponds to
f(Aj+l, sj+l;X) < f ( A ; ,s;:W .
However, since (30) and (31) are contradictory, (Aj, S j ) the least-squares estimates A(') associated with each S('));that
for i = 1 . . .64. The graph
must be a fixed point. To summarize, we have the following is, A(L)= argmina IjX - AS('))1I2
itself corresponds to a matrix of residuals, wherein the entry
convergence theorem for ILSE.

5 IIX - Ak-lSkll$

(26)
(27)

TALWAR et al. SYNCHRONOUS CO-CHANNEL DIGITAL SIGNALS-PART

1191

Note that in rows 1, 16, 22, 27 etc., there are multiple gray
points. These correspond to cases where A() is singular, thus,
multiple S ( J ) syield the same residual. We have shown a few
of the paths that may be taken by the algorithm, depending
on the initial pair (Ao, SO).Paths to the global minima are
indicated by solid lines, and paths to other fixed points by
dashed lines. We note that for this particular scenario, 56 paths
of the 64 possible paths lead to global solutions, and eight
paths lead to other fixed points.
VI. RECURSIVE
ALGORITHMS
In this section, we consider two classes of recursive algorithms for estimating the received signals. In recursive
estimation, we are interested in solving the following minimization problem at symbol period n

10

20

30

si?) 40

50

60

Fig. 1. Representation of a matrix of residuals, wherein the entry at

row i and column j corresponds to the the value of the residual
IIX - A()S(J)Il,i ; j = 1 . . . 6 4. Shown are paths of ILSE algorithm
that lead to global minima (solid lines) and to other fixed points (dashed
lines).

where X ( n ) = [ X ( n - 1) x ( n ) ] , S ( n=) [ S ( n- 1) ~ ( n ) ] ,
and B ( n ) = diag(a-l, an- i . . . ,1) is a diagonal weighting
matrix for some 0 < Q < 1. Our objective is to compute
A ( n ) and ~ ( n assuming
),
that a good estimate of S(n - 1)
(or equivalently A(. - 1))is available. This estimate may be
obtained blindly by using the block algorithms of the previous
section or by a short training set. The exponential weighting is
used to de-emphasize old data in a time-varying environment.
The fading memory least-squares solution for A ( n ) is given
by

at row i and column j corresponds to the the value of the

residual /IX - A(z)S(j)/I2, i , , j = 1 . . .64. In particular, the ith
row in the graph corresponds to the residuals associated with
A ( n )= X ( n ) B( n )S*( n )(S( n ) B( n )S* (n))-
matrix A(i) for all possible S ( j ) s . For simplicity, we have
(33)
chosen a three-color scheme to depict the residuals. A point
which can be updated recursively (see [30]) as follows:
( i , j ) is colored gray if the value of its associated residual
IIX - A(i)S(j)112is minimum with respect to other residuals
A ( n ) = A(. - 1)
in row i . Otherwise, it is colored white. If a point is a global
(4.1 - A(. - l ) s ( n ) ). s * ( n ) P ( n - 1). (34)
minimum with respect to the whole matrix, it is colored black.
Q
s*(n)P(n- l)s(n)
For a given Ao, ILSE generates an SO, which is the starting
l , also
point in our graph. From this So, the algorithm computes In the above equation, P ( n ) = ( S ( n ) B ( n ) S * ( n ) ) -can
an A l . This step is symbolically depicted by a vertical line be expressed recursively as
starting at column j such that S(J) = So, and moving up to
P( - l ) s ( n ) s * ( n ) P ( n - 1)
P ( n ) = - P ( n - 1) the diagonal. Now, the algorithm generates S2 by enumerating
a s * ( n ) P ( n- l ) s ( n )
Q
l(
over all possible Ss. We represent this step by a horizontal
P(0) = I.
move on the line associated with A ( j ) = A l . Since we seek
a unique minimum, the line segment from the diagonal ends In addition, we define H ( n ) = X ( n ) B ( n ) S * ( n ) ,so that we
at the leftmost gray point. This follows from the fact that in can rewrite A ( n ) = H ( n ) P ( n ) in (33). Using this framework,
our implementation, we always pick the S ( j ) with the lowest we are now ready to consider the two classes of recursive
index that minimizes IIX - A S ( j ) / I 2 ; j= 1 . . . L d N .Once SI algorithms.
is picked, the algorithm computes the corresponding Az, and
this once again corresponds to a move to the diagonal. We A. Class A
proceed in this fashion until a global minimum or a fixed
The first class includes algorithms that alternate between
point is reached.
, then updating A(n), at each symbol
estimating ~ ( n )and
In this graph, fixed points are located on the diagonal. It period. The recursive extensions of the two block algorithms
is clear that if the leftmost minimum with respect to S for a belong to this class. For each data vector ~ ( n )we, first
particlular A(;) is on the diagonal, the scheme described above estimate s ( n ) by minimizing
will not generate any more line segments, and the iteration will
remain fixed at that point. Since the global minima are fixed
points of the algorithm, they are also located on the diagonal.
There are multiple global solutions, since the signals can be using either the least-squares with projection approach or the
identified only up to an ATM, as described in Section 111. enumeration approach. This step requires O(md2) flops or
+

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 44, NO. 5 , MAY 1996

1192

C?(mdLd)flops, respectively. Then, A(n) is computed using

(34), which requires C?(md)flops. The enumeration approach
is more robust for low SNR's, and is thus recommended
whenever computationally feasible (for small values of d). The
recursive
has
flops
per snapshot.
Recursive Least Squares with Enumeration (RLSE):
1) Given A(n - 1)
2) n = n + l
* Minimize (35) for s(n) (enumeration).
Update A(n) using (34)
(RLS).
3) Continue with the next snapshot.

Likewise, we can partition

n - l ) X ( n - 1) X*(n - l)x(n)
x * ( n ) X ( n- 1)
x*(n)x(n)
(42)
Multiplying (40) and (42), and using properties of the trace,
we can rewrite (39)

X * ( n ) X ( n )= ["*(

max tr[a2H(n- l ) P ( n ) H * ( n- 1)
s(n)EQ

+ a H ( n l)P(n)s(n)x*(n)
+ aP(n)H*(n l)x(n)s*(n)
-

+ P ( n ) s (n)x*(n)x(n)s*( n ) .]

(43)

It is easily seen that H(n) can be updated recursively as

B. Class B

H(n) = aiH(n - 1)

In this class, we minimize (32) jointly over A(n) and

s(n).This is achieved by substituting the weighted leastsquares solution for A(n), given by (33), back into the original
minimization criterion to yield a new criterion

+ x(~)s(Tz)*

(44)

and thus, (43) can be simplified to

max tr [H(n)
P ( n )H*(n)]
.
s(n)ECl

Since A(n) = H ( n ) P ( n ) ,we estimate s ( n ) by maximizing

G(n)= arg max tr[A(n)H*(n)].

s(n)Ea

where PL
B(n)iS(n)'
In - ' B ( n ) t ~ ( n ) - ' and pB(il)fS(n)is a d-dimensional projection matrix defined by the rows of
S(n)B(n)'2 as follows:
pB(nj4s(n)*

= ~ ( n3 s(n)*
)
( ~ ( n ) ~ ( n ) ~ -(lns )(*~)) (n)
B + . (37)

Note that the key difference between the criterion in (36)

and the block ML criterion in (18) is that the minimization
in (36) is over s ( n ) only, since S ( n - 1) is assumed to be
known. Hence, recursive minimization of the ML criterion is
computationally tractable.
We can equivalently maximize
(38)
which can be expressed in terms of the trace operator

max t r [B( n )4PBCn,

3 s(n) B (n)x ( n )*X(n)].

s ( n ,) E-Q

(39)

Using (37), we see that

B(n)4PB(Ir)
4s(n)"B(n)3 = B(n)S(n)*P(n)S(n)B
(n)
which can be partitioned as in (40) at the bottom of the page.
The above follows by noting that the weighting matrix can be
expressed as

B(n) =

1 '

[ilZB
(

(41)

(45)

Substituting H*(n) = S(n)B(n)X*(n)in (45), it follows

that we choose s ( n ) to maximize the correlation between
A(n)S(n)B(n)i and X ( n ) B ( n ) i .
In ( 4 3 , we compute A(n) and H ( n ) for each of the
Ld possible vectors s ( n ) E R. This is done recursively
using (34) and (44), and hence the computation requires
C3(mdLd)flops. Next, we compute the diagonal entries of
A(n)H*(n) since we are only interested in the trace of the
product. Computing the trace for all possible vectors s ( n )also
requires O ( m d L d )flops. Hence, this recursive approach has
computational complexity O(mdLd),which is the same as
=SE. However, we obtain better signal estimates using this
approach. The algorithm is summarized below.
Recursive Projection Update (RPU):
1) Given A(" - 1) and H ( n - I)
2) n = n + l
Maximize (45) for s(n)
Update A(n) and H ( n )
3) Continue with the next snapshot

VII. SIMULATION
RESULTS
We present the results of three different sets of simulations
in this section. For simplicity, we assume a uniform linear
array of m = 4 sensors. In the first set of simulations, we
study the performance of the block algorithms for a block size
of N = 100. We consider d = 3 digitally modulated BPSK
signals arriving from [IO,16,251" relative to array broadside.
We assume all three signals have equal powers. Starting with

1-

- l ) S * ( n - l ) P ( n ) S ( n- I)B(n - I )
o B ( n - l)S*(n - l)P(n)s(n)
as*(n)P(n)S(n- I)B(n - I)
s* ( n ) P ( n ) s ( n )

(40)

TAI,WAR c t

a/

YYNCHRONOUS CO-CHANNbL DIGITAIL SIGNALS-PART

I193

14

-loo
- -

- _
12fn
0
K

10 -

.c

$
-

*.

--?K

..,mILSP
\

,
m\

%,

8-

,YC

6 6-

1o

;
0
2
3
4
5
6
0
1

-~
0

SNR (dB)
Fig. 2. Bit error rate for ILSP.

loo

1o

- I~
0

SNR (dB)
Fig 3

SNR (dB)
Fig. 4. Number of iterations for ILSP and ILSE.

Bit error rnlc for ILSE

A0 = I l r l x d , we first estimate A and S using ILSP. The

estimate A is then used to initialize ILSE for improved
array response and signal estimates. To ensure that the global
minimum is achieved, we check if the residual is close to the
noise power level. If this is not the case, we re-initialize ILSP
with a random initial guess. This blind estimation process is
repeated lo4 times, each time over a different noise realization.
Hence, a total of lo6 bits are estimated for each signal.
In Fig. 2, we show the bit error rates achieved using
ILSP with signal to noise ratios ranging from 0 to 6 dB.
Although the signals are closely spaced in their directions
of arrival (DOAs), this algorithm is successful in separating
and estimating the three signals. We observe that the bit error
rates (BERs) depend strongly on the separation between the
DOAs of the signals since sz, which is the closest to the
other two signals, has the highest BER, and s3, which is well
separated, has the lowest BER. This follows from the fact that
using least-squares to estimate the signals (S = At X) causes

noise enhancement, especially if the columns of A are nearly

dependent due to closely spaced signals. In ILSE, however,
we use enumeration to estimate the signals. As seen in Fig. 3,
this algorithm yields significantly lower BERs. We achieve
bit error rates lower than l o p 3 for SNRs greater than 5 dB.
In Fig. 4, we present the average number of iterations
required by both ILSP and ILSE. The number of iterations
for ILSP decrease moderately fast with increasing SNR. ILSP
requires more iterations than ILSE, but each iteration is computationally cheaper. Both algorithms converge fairly rapidly
to the global solution. In this difficult scenario, re-initialization
was needed in less than 4% of the runs.
Finally, in Fig. 5, we compare the use of training signals
with FA property for the same scenario. In our simulations
with training signals, we assume 8 out of 100 symbols in each
data block are known, which represents the proportion used in
IS-54. We first estimate A from the training data, and then use
the RLS algorithms described in Section VI-A to estimate the
signals, namely, RLS with projection (RLSP) and RLS with
enumeration (RLSE). The BERs obtained from RLSP (*) and
RLSE ( 0 ) for signals s2 and s3 are shown in Fig. 5. Also
shown for comparison are the BERs obtained blindly from
ILSP (dashed-dotted line) and ILSE (solid line) algorithms.
We observe that the signal estimates obtained blindly from
ILSP and ILSE algorithms are slightly better than those
obtained from training-based RLSP and RLSE algorithms.
Note, however, that RLSE already makes strong use of the
FA property. In order to make a fair comparison of FA with
training signals, we must compare the performance of ILSE
with RLSP. We see in Fig. 5 that ILSE performs significantly
better than RLSP. Hence, FA is a powerful property for
blind signal estimation. In practice, one may combine a short
training set with FA algorithms to obtain good signal estimates
at a low computational cost.
In the next set of simulations, we compare the performance
of the blind algorithms to the case where the array response
matrix is completely known. The simulation set up is the

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 44, NO. 5, MAY 1996

1194

s2

s3

-i o o 1

SNR (dB)

SNR (dB)
Fig. 5. Comparison of training-based RLSP (*) and RLSE

(0)algorithms

with FA-based ILSP (dashed-dotted line) and ILSE (solid line) algorithms.

same as before, except that two BPSK signals at 0 = [0.5]

are received at the array. Again, both ILSP and ILSE are
used to estimate the digital signals. We see from (22) that
in ILSE, we make a joint decision on all d bits in each
snapshot s ( n ) .n = 1 . . . N . Hence, it is convenient to consider
the snapshot error rate (SER), which is the probability that
s(n) does not equal ~ ( T L ) . In Fig. 6, the dashed-dotted and
the solid lines indicate the theoretical SER for the signal
detection approach used in ILSP and ILSE, respectively, with
a known A. More specifically, we detect the signals in ILSP by
obtaining a least-squares estimate of S and projecting on to the
discrete alphabet, and in ILSE we enumerate over all possible
S matrices. For the ML minimization criterion (17), the SER
curve for ILSE represents the lowest SER attainable by any
algorithm. The iand
0 indicate the SER performance of the
blind algorithms, initialized with A0 = I, d . Again, for less
than 0.5% of the runs, we needed to re-initialize with random
Aos. The plot shows that there is virtually no difference in the
performance of the blind algorithms as compared to the nonblind approach. A more detailed analysis of these algorithms
and their comparison is presented in [31] and [32].
In the final set of simulations, we study the performance
of recursive algorithms. We use d training signals, given in
(56), to estimate Ao. For simplicity, we choose Q = 1. As in
the previous experiment, we first consider the simple scenario
with two BPSK signals arriving from 0 and 5. At 3 dB
SNR, we estimate N = 250 snapshots using RLSE and RPU.
We average the results over lo4 such runs, so that a total of
2.5 x lo6 bits per signal are estimated. The BERs achieved
by both the algorithms are equal: 1.45 x lop2 for each signal,
although the number of bits in error is slightly larger for
RLSE. The SER for the two algorithms is also 1.45 x 10V2.
In Fig. 7, we plot the SER using RPU at each symbol period
n,n = 1. . . N . The SER plot using RLSE is virtually identical
for this scenario. We observe that the SER is high initially
since the estimate of A is poor. But as n increases, this
estimate improves and the SER decreases. After n = 120,
the SER converges to about 1.14 x lo-, the theoretical SER
for known A, as shown in Fig. 6. Finally, we make the current

ioo 1

cn

Io-~;

I
2

SNR (dB)
Fig. 6.

Snapshot error rate for ILSP and ILSE

scenario more difficult by adding another signal at lo, with 3

dB SNR. Also, we reduce N = 100 to ease the computational
burden. The BERs achieved by RLSE and RPU are given in
Table 1. We see that RPU yields lower BERs than RLSE.
VIII. CONCLUSION
We have presented a blind approach for the separation
of synchronous digital signals in a coherent multipath environment. We have shown that given a sufficient number of
snapshots, the signal estimates obtained by this approach are
unique. The block algorithms ILSP and ILSE take advantage
of the FA property of digital signals to simultaneously estimate
the array response and symbol sequence for each signal. We
have proved that the ILSE algorithm converges to a fixed point
in a finite number of iterations. A fixed point that is not a global
solution can be detected by considering the magnitude of the
residual. Hence, we are ensured that the global minimum will
be reached by re-initializing one or more times. We may note

TALWAR et al.: SYNCHRONOUS CO-CHANNEL DIGITAL SIGNALS-PART I

1195

Alg.

SZ

S1

s3

RLSE 2.88 x lo- 5.26 x

2.59 x
RPU 2.66 x loF2 4.92 x loF2 2.39 x loF2

0.04

where tT = [ t d . . . t z t l ], is such that ti = fl for some i , and

t j = 0 for j = 1, . . . ,d , j # i .
Proofi The proof is by induction. The case d = 1 is
trivial. Let us consider the case d = 2. Define n1;l = t l . We
have the following n = 2 equations:
0.02

-- +
t2

Q2,J

n 2 , 2 =: t 2

0.01

i,

100

50

150

200

I
250

Symbol Period (n)

Fig. 7.

Snapshot error rate profile for RPU.

that despite the difficulties associated with the existence of

fixed points that are not global solutions, these algorithms are
important in providing a computationally feasible alternative
to complete enumeration, which is intractable. Our simulation
experiments have shown that in scenarios of practical interest,
the converged performance of ILSP and ILSE is virtually
the same as predicted in the case where the array response
matrix is completely known. We have also described recursive
extensions of ILSP and ILSE, and proposed a new algorithm,
RPU, that minimizes the ML criterion recursively. In contrast
with block minimization, recursive minimization of the ML
criterion is computationally tractable.
Our approach can be extended to asynchronous transmission and to multipath channels with large delay spread, as
discussed in Section I. Since our algorithms require that the
number of signals be known or accurately estimated, we have
proposed in [33] a robust scheme for determining the number
of incident signals in a multipath propagation environment.
Determining the length of the impulse response associated
with each signal is a more difficult problem, and is currently
under investigation. Other directions for future work include
developing computationally efficient initialization strategies
in order to avoid restarts, estimating signals in the presence
of non-Gaussian interference, and combining coding schemes
with blind estimation.
APPENDIXA.l
IDENTIFIABILITY
Lemma I : Let S d ,be a matrix with f1 elements such that
the columns of S include all the n = 2dp1 possible distinct
d-vectors with the first element normalized to $1. Then, the
solution t for the system of equations of the form
tT[SlSZ

. . . s,]

= [H*l . . +1]

a1,1 =
Ql.1 =

(46)
(47)

f-1
41.

There are four possible right-hand sides, which we denote as

[xl 521. Then for [xl 221 equal to:
1) [l -11 or [-1 l ] ; t * = 0 and tl = + I or -1,
respectively.
2) [l 11 or [-1 -1l;tl = 0 and t2 = + I or -1,
respectively.
We assume that the above theorem holds for some d = k .
Then for d = k
1, we obtain n = 2k equations of the
following type:

Qk+l,l
Qk+l,2

ak+l.n-l

--

tk+l

+ Qk.1 = Zt1

tk+l

- Q k , l = 41

tk+l

ak+l,n

E tk+l

+ k,?
-

(48)
(49)
(50)
(51)
(52)

= *I

nk.2

It1

where a k , j , for j = 1 . . . 2k-1 is of the form tk & . . . t 2 & t l .

From (48) and (49), four possibilities exist for t k + l and a k , ~ :
1) t k + l = 0 and a k , ~= $1 or -1.
2) a k , l = 0 and t k + l = +1 or -1.
Now, if t k + l = 0, from (48)-(52), we get the reduced
system { Q ~ , J= 5 1 , .. . , a k , ? = fl} which by assumption
has the solution: ti = +1 for some i , and t j = 0 for
j = 1, . . . d ; j # i. If t k + l = +1 or -1, we must have
{ a k , ~ = 0 ; . . . , k , f = 0). The first two equations of this
system
ak:l
ak,2

= t k +@kpl,l = 0
tk

- ak-1,1 = 0

imply t k = a k - 1 , ~ = 0. Given t k = 0, we obtain the smaller

system { a k - l , l = 0 , . . . ak-1,:
= O}. Next, we consider
the first two equations of this system to show t k - 1 = 0, and
continuing in this fashion, it is easily seen that t k - 2 = . . . =
tl = 0. Thus, by induction, the theorem holds for all d .
0
APPENDIXA.2

Theorem 3.3: Let p denote the probability of receiving

all the Ld/2 distinct (up to a sign) columns of S in N
independent snapshots, N >> Ld/2. Assuming that the signals

1196

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 44, NO. 5, MAY 1996

are independent, and the probability of receiving each symbol

in R is equal, p is bounded by

where ST is the corresponding row of S, and can be likewise

partitioned as ST = [sz S d f l s:]. The first d columns of s
form a linearly independent set, and thus define t uniquely in
terms of elements of Sd as follows:

Ld

Ld-2

LpL1

(53)

Prooj? Our sample space S is the set of all d-vectors with

entries in I1 = {&I, + 3 , . . . , & ( L - 1)},S= {x:xk E Cl:IC =
1 . . . d ) . Let A be the event that all the I = L d / 2 distinct
vectors from S are picked in N >> 1 independent trials. We
are primarily interested in showing that given a fixed value of
d, the probability p of event A approaches 1 as N increases.
For this reason, it suffices to compute a lower bound for p .
Consider the complement of A. A is the event in which
at least one of the distinct vectors has not been picked. Let
d:,i = 1 . . . 1 be the event where neither a particular +x,
nor -xi has been picked in N trials. Then A = Uf=,X.
Furthermore

P ( k )= P

(1,U

A;

5 CP(A;)

(54)

i=l

since the events 4 are not disjoint. Now, the probability that
+xi and -xi is not picked in the kth trial is one minus the
probability +xi or -xi is picked, i.e., (1- &).Since the trials
Then from (54), we get
are independent, P ( A I ) is (1 -

6).

The result in (53) easily follows.

The inverse of S d can be computed easily by noting that S d =

eeT+2eleT-21 where e = [l 1. . . 1ITa n d e l = [l O...0lT,
and then applying the Sherman-Morrison-Woodbury formula
[341

3-d

1 +I

+1 +1

...

...

-11

Remark: In situations where blind identification is not

required, the rows of s d define a useful set of d training signals
sent by each of the users to learn the array response matrix A.
Hence, t can be expressed in closed form as
1
2

t T = -[(3

d)Sl

+ 5 2 + . . . + Sd

31 - 5 2

...

51 - S d ] .

(60)
Since S d is vector with 5 1 entries, there are a finite number of
possibilities for t. The question then becomes whether there
exists another vector of f l s in S such that only a trivial t
is possible. The answer to this question is the vector
We see from (57) that t must satisfy t T s d + l = Sd+l, which
0 yields the relation
(2 - d ) s l

+ 52 + S d

= sd+l.

(61)

If we let k denote the number of -1s in {Sz . . . S d } , then

(61) becomes
(2 - d)3l

+ IC(-1) + ( d

1 - k)(+1) = Sd+I

and solving for IC, we get

1

k = ?((a - d)s1 - Sd+l

2

sd=

-+1
+1
+1

+1
-1
$1

+1
$1
-1

. . . +1. . . +I

_+I

+I

+I

...

. . . +l
-1-

+ ( d - 1)).

(62)

We consider the four different possibilities for [SI Sd+l]. In

the case that [SIS d + l ] is equal to +[l 11 or -[1 I],IC equals 0
or d - 1, respectively. This implies sz = +[1 1 1 . . . 11, and
then from (60), we see that tT = +[1 0 0 . . . 01. Similarly, if
[SI S d + l ] = &[I: -11, then sz =
. . . 1 - 1 1 . . . 11 and
( 5 6 ) correspondingly tT = &[0 . . . 0 1 0 . . . 01, with a41 only in
the ith position, i = 2 . . . d. In all four cases, a trivial solution
is obtained. We have shown previously in Section I11 that each
row t of T must be distinct. Therefore, T is an ATM.
REFERENCES

S. Talwar, M. Viberg, and A. Paulraj, Blind estimation of multiple cochannel digital signals using an antenna array, IEEE Signal Processing
L e t t , vol. 2, no. 1, pp. 29-31, Feb. 1994.
D. Gerlach and A. Paulraj, Adaptive transmitting antenna arrays with
feedback, IEEE Signal Processing Lett., vol. 10, no. 1, pp. 150-152,
Oct. 1994.
G. Raleigh. S. Diggavi, V. Jones, and A. Paulraj, A blind adaptive
transmit antenna algorithm for wireless communication, in Proc. IEEE
ICC, 1995.

TALWAR et ul.: SYNCHRONOUS CO-CHANNEL DIGITAI. SIGNALS-PART

[4] G. Xu and Hui Liu, An effective transmission beamforming scheme

for frequency-division-duplex digital wireless communication systems,
in Proc. IEEE ICASSP, 1995, pp. 1729-1732.
[ 5 ] R. 0. Schmidt, A signal subspace approach to multiple source location and spectral cstimation, Ph.D. dissertation, Stanford University,
Stanford, CA, 1981.
[61 A. Paulraj, R. Roy, and T. Kailath, Estimation of signal parameters
via rotational invariance techniqucs-ESPRIT, in Proc. 19th Asilomar
Con$ Circuits, Syst., Comput., 1985.
[7] B. Otterstcn, R. Roy, and T. Kailath, Signal waveform estimation in
sensor array processing, in Proc. 23rd Asilomar Con$ Signals, Syst.,
Comput., 1989, pp. 787-791.
[SI S. Anderson, M. Millnert, M. Viberg, and B. Wahlberg, An adaptive
array for mobile communication systems, IEEE Trans. Veh. Technol.,
vol. 40, no. I , pp. 230-236, 1991.
[9] B. Agee, Blind separation and capture of communication signals using
a multitarget constant modulus beamformer, in Proc. IEEE MILCOM,
1989. nn. 340-346.
A. Swindlehurst, S. Daas, and J. Yang, Analysis of a decision directed
beamformer, IEEE Trans. Signal Processing, vol. 43, pp. 2920-2927,
Dec. 1995.
R. Gooch and B. Sublett, Joint spatial and temporal equalization in
a decision-directed adaDtive antenna system, in Proc. Asilomur Conf:,
1989, pp. 255-259.
B. G. Agee, A. V. Schell, and W. A. Gardner, Spectral self-coherence
restoral: A new approach to blind adaptive signal extraction using
antenna arrays, Proc. IEEE, vol. 78, pp. 753-767, Apr. 1990.
1131 J. Cardoso, Source separation using higher order moments, in Proc.
IEEE ICASSP, May 1989, pp. 2109-21 12.
1141
- - L. Tong, Y. Inouye, and R. Liu, Waveform-preserving blind estimation
of multiple independent sources, IEEE Trans. Signal Processing, vol.
41, pp. 2461-2470, July 1993.
[IS] H. Liu and G. Xu, A deterministic approach to blind symbol estimation, IEEE Signal Processing Lett., vol. I , pp. 205-207, Dec. 1994.
[I61 H. Liu and G. Xu, Multiuser blind channel estimation and spatial
channcl pre-equalization, in Proc. f E E E ICASSP, pp. 1756-1759, 1995,
[I71 A. J. van der Veen, S . Talwar, and A . Paulraj, Blind identification of
FIR channels carrying multiple finite alphabet signals, in Proc. IEEE
ICASSP, 1995, vol. 2, pp. 1213-1216.
[ 181 H. L. Van Trees, Delection, Estimation and Modulation Theory. New
York; Wiley, vol. I, 1968.
[I91 M. Wax and T. Kailath, Detection of signals by information theoretic
criteria, IEEE Trans. Acoust., Speech., Signal Processing, vol. ASSP33, pp. 387-392, Apr. 1985.
[20] M. Wax and 1. Ziskind, On unique localization of multiple sources by
passive sensor arrays, IEEE Trans. Acoust., Speech, Signal Processing,
vol. 37, pp. 996-1000, July 1989.
1211 N. V. Efimov and E. R. Rozendorn, Linear Algebra and MulliDimensional Geometry. Moscow, Russia: MIR, 1975.
1221 G. Golub and V. Pereyra, The differentiation of pseudo-inverses and
nonlinear least squares problems whose variables separate, SIAM J.
Num. Anal., vol. IO, pp. 413432, 1973.
[23] P. Gill, W. Murray, and M. Wright, Practical Optimization. San Diego,
CA: Academic, 1981.
1241 B. Agee, The least-squares CMA: A new technique for rapid correction
of constant modulus signals, in Proc. ICASSP, pp. 953-956, 1986.
[25] B. Agce, Maximum-likelihood approaches to blind adaptive signal
extraction using narrowband antenna anays, in Proc. Asilomar Con$,
1991, pp. 716-720.
1261 Y. Wang et al., A matrix factorization approach to signal copy of
constant modulus signals arriving at an antenna array, in Proc. CISS,
Princeton, NJ, 1994.
1271 S. Barnett, Matrices-Methods and Applications. Oxford, U.K.: Clarendon, 1990.
[28] B. Juang and L. R. Rabiner, The segmental K-means algorithm for
estimating parameters of hidden Markov models, IEEE Trans. Acoust.,
ng, vol. 38, no. 9, pp. 1639-1641, Sept. 1990.
~~

1197

[29] W. I. Zangwill, Nonlinear Programming: A Unified Approach. Englewood Cliffs, NJ: Prentice-Hall, 1969.
[30] S. S. Haykin, Introduction to Aduplive Filters. New York: Macmillan,
1984.
[31] S. Talwar and A. Paulraj, Performance analysis of blind digital signal
copy algorithms, in Proc. MZLCOM, 1994, vol. I, pp. 123-128.
[32] S. Talwar, A. Paulraj, and M. Viberg, Blind separation of synchronous
co-channel digital signals using an antenna array. Part 11. Performance
analysis, IEEE Trans. Signal Processing, submitted for publication.
[33] A. J. van der Veen, S . Talwar, and A. Paulraj, Blind estimation of
multiple digital signals transmitted over FIR channels, IEEE Signal
Processing Lett., vol. 2, pp. 99-102, May 1995.
G.
H. Golub and C. F. Van Loan, Matrix Computations. Baltimore,
[34]
MD: John Hoplans Univ. Press, 1984.

I I

Shilpa Talwar (S95) received the Bachelors de-

gree in mathematics and computer science from the

State University of New York, Geneseo, NY, in
1989. She is currently working toward the Ph.D.
degree in the scientific computing and computational mathematics program at Stanford University,
Stanford, CA.
Her research interests include numerical linear
algebra, array signal processing, and wireless communication systems.

Mats Viberg (S87-M90) was born in Linkoping,

Sweden, on December 21, 1961 He received the
M S degree in applied mathematics in 1985, the
Lic. Eng. degree in 1987, and the Ph I) degree in
electrical engineering in 1989, all from Linkoping
University, Sweden
He joined the Division of Automatic Control
at the Department of Electrical Engineering,
Linkoping University in 1984, and from November
1989 until August 1993 he was a research associate
From October 1988 to March 1989, he was on leave
at the Informations Systems Laboratory, Stanford University, Stanford, CA,
as a visiting scholar From Aug 1992 until Aug 1993, Dr Viberg held a
Fulbright-Hayes grant scholarship as a Visiting Researcher at the Department
of Electrical and Computer Engmeering, Brigham Young University, Provo,
UT, and at the Informations Systems Laboratory, Stanford University Since
September 1993, he has been a professor of signal processing at the
Department of Applied Electronics, Chalmers University of Technology,
Sweden. His research interests are in Statistical Signal Processing and
its application to sensor array signal processmg, system identification,
communication, and radar systems
Dr Viberg received the IEEE Signal Processing Societys 1993 Paper
Award (Statistical Signal and Array Processing Area), for the paper Sensor
array processing based on subspace fitting, coauthored with Bjom Ottersten
He is currently a member of the IEEE Signal Processing Societys technical
committee on Stdtistical Signal and Array Processing

Arogyaswami Paulraj (SM85-F9 I ) , for a photograph and biography, see

this issue, p. 1155.