IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 45, NO.

1, JANUARY 1997

173

A Subspace Approach to Blind Space-Time Signal

Processing for Wireless Communication Systems
Alle-Jan van der Veen, Member, IEEE, Shilpa Talwar, and Arogyaswami Paulraj, Fellow, IEEE

AbstractThe two key limiting factors facing wireless systems

today are multipath interference and multiuser interference. In
this context, a challenging signal processing problem is the joint
space-time equalization of multiple digital signals transmitted
over multipath channels. We propose a blind approach that does
not use training sets to estimate the transmitted signals and
the space-time channel. Instead, this approach takes advantage
of spatial and temporal oversampling techniques and the finite
alphabet property of digital signals to determine the user symbol
sequences. The problem of channels with largely differing and
ill-defined delay spreads is discussed. The proposed approach is
tested on actual channel data.

I. INTRODUCTION

challenging problem in signal processing is the blind

joint space-time equalization of multiple digital signals
transmitted over multipath channels. An important area where
such a problem arises is wireless (mobile) communications.
Consider a scenario where several users are trying to talk to a
central base station, which has several antennas (viz., Fig. 1).
A space-time equalizer at the base station combines two signal
processing aspects: equalization (or echo canceling) to combat
the intersymbol interference due to large-delay multipath and
source separation to combat cochannel interference (CCI). The
CCI might be interfering signals at the same frequency from
neighboring communication cells, or we might intentionally
allow multiple users at the same frequency in order to increase
the system capacity. The latter is known as space division
multiple access (SDMA) because it essentially separates users
based on differences in location.
Current communication systems such as IS-54 and GSM
require some amount of equalization (up to five symbol periods
in GSM and up to one symbol period in IS-54) but are
not designed to handle cochannel users. To assist classical
single-user channel identification algorithms, a fair number
of training symbols are incorporated in the data packets.
Manuscript received December 1, 1995; revised August 21, 1996. This
work was supported by the Department of the Army, Army Research Office,
under Grant DAAH04-95-1-0436. The views and conclusions contained in this
document are those of the authors and should not be interpreted as necessarily
representing the official policies or endorsements, either expressed or implied,
of the Army Research Office or the U.S. Government. Additional support was
provided by Ericsson Inc., Bell Northern Research Inc., Qualcomm Inc., and
Hughes Network Systems Inc.
A. J. van der Veen is with the Department of Electrical Engineering/DIMES,
Delft University of Technology, Delft, The Netherlands.
S. Talwar was with the Scientific Computing Program, Stanford University,
Stanford, CA 94305 USA. She is now with Stanford Telecom, Sunnyvale, CA
USA.
A. Paulraj is with the Information Systems Laboratory, Stanford University,
Stanford, CA 94305 USA.
Publisher Item Identifier S 1053-587X(97)00529-1.

Fig. 1. Multiray scenario in wireless communications.

However, in recent years, it became gradually known that

digital signals can also be separated and equalized blindly,
i.e., without the aid of training sequences, by exploiting
the underlying structure of the signals. Although the use
of training sequences is an inherently more robust way to
estimate the channel, there are several reasons for studying
blind algorithms, aside from the obvious academic and military
motivations. Most notably, adding unnecessary training bits is
a direct waste of the available bandwidth. In addition, training
is not efficient in rapidly time-varying channels or in protocols
with very small data packages, such as the uplink of wireless
teletypes in PCS or in distributed networks. Training requires
synchronization, which is not always available or feasible
in multiuser scenarios. Finally, the insights gained are also
applicable to other systems such as CDMA, where it might be
used to improve the near-far resistance.
Blind algorithms have now become a very active research
area, in particular in the context of digital communication
signals, where there are several leverages for solving the blind
finite impulse response, multiple inputs, multiple outputs (FIRMIMO) identification problem considered in this paper. For
example, the fixed symbol rate of digital signals in combination
with linear channels, multiple antennas, and oversampling
allows us to blindly synchronize and equalize (but not separate) such signals. Statistically, oversampling digital signals
gives rise to cyclostationarity of the spectrum [3]. Tong et
al. were the first to realize that cyclostationarity allows the
identification of nonminimum phase FIR-SISO channels from
second-order statistics [4], [5]. In a deterministic discrete-time
setting, the property leads to structured (Toeplitz) matrices
and has inspired several subspace-based algorithms [6][9].
A second useful property is the finite alphabet (FA) structure

1053587X/97$10.00 1997 IEEE

174

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 45, NO. 1, JANUARY 1997

of digital signals. For equalization, this has been exploited

in decision-directed adaptive algorithms [10][13] as well as
in joint channel estimation and sequence detection [14], [15].
Several iterative algorithms for the separation of instantaneous
superpositions of multiple finite alphabet signals (I-MIMO)
were originally proposed by Talwar et al. [16][18]; an algorithm based on expectation maximization appeared in [19].
The two properties are in fact readily combined into one
algorithm to solve the FIR-MIMO problem, as was discovered
independently by Liu and Xu [20], [21] and the present authors
[1], [2]. Related work on blind FIR-MIMO identification was
carried out in parallel by Abed-Meraim et al. as well [22].
Many other signal properties can also be used for blind
estimation, for example, source independence and high-order
statistical properties and constant modulus properties. In addition, the spatial properties of the receiving antenna array might
be known, which allows signal separation based on differences
in directions of arrival, provided the number of antennas is
large enough [23]. Assuming a multiray propagation scenario,
knowledge of both the pulse-shape function and the array
manifold allows a joint delay and angle estimation of all
propagation paths (viz. [24], [25]).
A. Contributions
In this paper, we consider the FIR-MIMO source separation
and equalization problem. We assume all sources transmit
digital communication signals with the same symbol rate and
alphabet, both of which are known a priori. The measured data
is obtained from a cluster of fractionally sampled antennas. We
do not assume knowledge of the antenna array mainly because
we do not attempt to resolve the individual directions of the
incoming rays.
The proposed methods are subspace-based block-algorithms
and attempt to provide a structured factorization of the data
matrix into a channel matrix times a symbol matrix. The
constant symbol rate translates to a block-Toeplitz structure of
the symbol matrix, which is sufficient for equalization. We rely
strongly on the finite alphabet property for the separation of the
individual signals and to some extent also for the equalization.
The outline of the paper is as follows. The data model for
the FIR-MIMO scenario is presented in Section II. Maximumlikelihood techniques for blind sequence estimation are not
computationally feasible. As an alternative, a subspace-based
approach is presented in Section III. A proof of identifiability
and some insight into the underlying subspace intersection
method is given in Section IV, as well as a comparison
among a few alternative methods via computer simulation.
In Section V, we indicate problems that occur with subspace
intersection techniques if the channel lengths are not welldefined and suggest a possible solution. Finally, in Section
VI, the proposed algorithm is tested on simulated data based
on actual wireless indoor channels.
The paper encompasses preliminary short versions [1], [2].
During review, the full version of Liu and Xus approach
[20], [21] appeared in [26]. Although the original methods
used are quite identical, the present paper extends beyond [26]
by providing a significantly more efficient implementation,

comparing the proposed approach to a similar technique in

which the channel is identified first and subsequently inverted
(cf. [7]) and addressing the case with differing and ill-defined
channel lengths.
B. Notation
Lower-case bold, as in , denotes vectors. For a matrix ,
is the transpose,
is the complex conjugate transpose,
is the MoorePenrose pseudo-inverse, col
and row
denote the column span and row span of , and
is
the Frobenius norm of vec
is the vectoring operation
is the
that stacks all columns of in a single vector, and
Kronecker product.
II. DATA MODEL
An array of
sensors, with outputs
,
receives digital signals
, each of which is
described as a sequence of dirac pulses
For convenience, we assume the symbol rate
is normalized to
, and the digital symbols
belong to a known finite alphabet
for real signals, or
for complex signals. The waveform received at the array consists of multiple paths per signal,
with echos arriving from different angles, with different delays
and attenuations. The impulse response of the channel from
the th source to the th sensor
is a convolution of the
pulse shaping filter
and the actual channel from
to
. We include any propagation delays and delays due
to asynchronous signals in
. The data model is written
compactly as the convolution
where
..
.

..
.

..
.

..
.
It is common to assume at this point that all
are FIR filters of length at most
N:

channels

The maximal channel length among all sources is denoted

by
. An immediate consequence of the FIR
assumption is that, at any given moment, at most
consecutive symbols of signal play a role in
. Indeed, for
, where
Z and
, the convolution
can be expressed as
(1)
For simplicity of the exposition, we initially assume that all
channels have the same length and generalize later on.
Suppose we sample each
at a rate
N , where is
the oversampling factor, and collect samples during symbol

VAN DER VEEN et al.: SUBSPACE APPROACH TO BLIND SPACE-TIME SIGNAL PROCESSING

periods; then, we can construct a data matrix

as

..
.

175

an equalizer length measured in symbol periods), we obtain

.
...
.
...
..
.
.
..
.. ..
.
..

..
.

This augmented data matrix

The th column
of contains the
spatial and temporal
samples taken during the th interval. Based on the model of
in (1), it follows that
has a factorization

0
. .
.. ..

..
.

..
.
..

..

..

..
..
..

..

..

..

..

..

..

(2)

.
.

..

..

..

..

The matrix represents the unknown space-time channel, and

the block-Toeplitz matrix
contains the transmitted
symbols.1 For generality, we have assumed that the measured
block of data starts while the transmission of each of the
signals was already in progress, i.e.,
is determined by
previous symbols
as well as . A similar
assumption is made on
. Note that if the channels do
not all have the same length , then certain columns of
are equal to zero. Given , our goal in blind estimation is to
find
and such that is a block-Toeplitz matrix, and the
symbols in satisfy the finite alphabet property.
If the source alphabet is real, then it is customary to work
with a real-valued data model by redefining (with some abuse
of notation)
real
imag

..

has a factorization

real
imag

(3)

while
This effectively doubles the number of observables
halving the noise power on each entry.
The algorithms we consider in this paper rely on the
existence of a filtering matrix
such that
.
This implies that the row span of is equal to (or contained
in) the row span of . For this to be true, it is necessary
that
has full column rank, which implies that
.
This may put undue requirements on the number of antennas
or oversampling rate. However, it is possible to ease this
condition by making use of time-invariance and the structure
of . Extending
to a block-Hankel matrix by left shifting
and stacking times (as discussed later,
can be viewed as
1 The subscript L denotes the number of block rows in
the subscript if this does not lead to confusion.

S . We usually omit

(4)
(the shifts of to the left are each over positions) and the
objective becomes, for given , to determine factors
and
of the indicated structure such that the entries of belong
to the finite alphabet. As we show in the sequel, identification
is possible if this is a minimal-rank factorization. Necessary
conditions for
to have a unique factorization
are
that is a tall matrix and that is a wide matrix, which
for
leads to
(5)

poses a fundamental
Given sufficient data, only
identifiability restriction.
Note that these conditions are not sufficient for and to
have full rank. One case where
does not have full rank is
when the channels do not have equal lengths, in which case
is at most
. Ill-conditioned
the rank of
cases might occur when the channels are bandwidth limited
so that sampling faster than the Nyquist rate does not provide
independent linear combinations of the same symbols. In
principle, the maximal effective is given by the ratio of the
Nyquist rate and the symbol rate [27]. (There may be other
practical reasons to select a larger , e.g., to correct for errors
in carrier recovery. This is not considered here.)
For SISO models, the condition that
is of full rank
is usually formulated in terms of common zeros; if the transforms
of the rows of
do not have a root in
common, then
has full column rank for at least all
(viz., e.g., [7], [26]). For arbitrary channels, this technical
condition holds almost surely. In the FIR-MIMO case, the
corresponding requirement is that
is irreducible and
column reduced (viz. [22]).
III. SUBSPACE-BASED APPROACHES
According to the previous section, the basic problem in
solving the blind FIR-MIMO problem is, for a given matrix

176

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 45, NO. 1, JANUARY 1997

Fig. 2. Multistage equalization/separation filter.

, to find a factorization
, where
is blockToeplitz with entries
. If we assume that the data
matrix
is corrupted by additive white Gaussian noise, then
the maximum likelihood criterion yields the nonlinear least
squares minimization problem
(6)

block-Toeplitz

To find an exact solution of this nonlinear optimization problem is computationally formidable. It is possible to approach
the optimum via iterative techniques that alternatingly estimate
and , starting from some initial estimate for
[28]. This
approach is still computationally expensive due to the repeated
enumeration of all possible sequences of length using the
Viterbi algorithm. In addition, the initial point has to be quite
accurate in order to converge to the global minimum, rather
than one of the numerous local minima.
The subspace-based approaches derived in this section simplify the problem by breaking it up into two subproblems.
Suppose that the channels have equal lengths and that the
conditions (5) are satisfied. Then,
full column rank
full row rank

row
col

row
col

To factor
into
, the strategy is to find either ,
which is a block-Toeplitz matrix with a specified row span, or
, which is a block-Hankel matrix with a specified column
span. In the scalar case (
signal), a number of algorithms
have been proposed for doing the latter, in particular, [7] and
[8], and it is straightforward to extend these algorithms to the
vector case (
), presuming the channel lengths are all
equal. However, for
, subspace information alone leads
to an ambiguity:
is a factorization with the
same subspaces for
diag
and any invertible
matrix. This ambiguity is resolved in a second step by
taking advantage of the finite-alphabet property of the signals.
We outline three approaches: one that directly estimates
from its row span, as was originally proposed in [20], [1],
and [2], then an entirely equivalent but computationally more
is
attractive version, and finally, an approach in which
estimated first. In the absence of noise, all approaches give
exact results. Note that none of these approaches provides a
factorization
in which both factors are forced to
have the required Toeplitz or Hankel structure so that they are
suboptimal in that respect.

A. Step 1: Estimating the Row Span of

We work with the extended matrix
in (4). The
first step in the direct algorithms is the estimation of an
from
(orthonormal) basis of the row span of
the row span of . Suppose, as before, that the channels
has full column rank.
have equal lengths, (5) holds, and
row
so that we can determine the row
Then, row
span of from that of . This requires the computation of
where
are unitary matrices,
an SVD of ,
and is a diagonal matrix containing the singular values in
of
is
nonincreasing order [29]. Without noise, the rank
equal to the number of nonzero singular values, and we can
, where consists of the first
columns of
write
, is a diagonal
matrix consisting of the nonzero
is the first
rows of , forming an
singular values, and
. For well-conditioned problems
orthonormal basis for row
.
with equal channel lengths , we expect
is corrupted by noise, then the numerical rank
of
If
is estimated by deciding how many singular values of
are
above the noise level. The estimated row span is given by
rows of .
the first
If the noise on is white and i.i.d., with covariance matrix
, then asymptotically converges to a basis for the column
. For i.i.d. signals,
span of the noise-free data
(or a multiple thereof) so that
asymptotically converges
, where
contains the singular values of .
to
does not converge to its noise-free value since
Note that
its dimension grows along with . However, we can write
so that each column of
is determined
by a linear combination of the corresponding column of .
Because of the Hankel structure of , this column contains
consecutive symbol periods. Hence, the
samples from
matrix multiplication can be viewed as an FIR filter, where
is the equalizer length, and the rows of can be viewed as a
new, filtered data set. This is depicted in Fig. 2, where we have
just covered the first stage (of three). The filter coefficients are
. The main purpose of such a
given by the entries of
subspace filter is dimensionality reduction, although we will
as well.
use the orthonormality of
matrix requires
Computing the SVD of an
operations [29]. It is
about
possible to replace the SVD by computationally more efficient
adaptive subspace tracking algorithms, which update the filter
coefficients as increasingly more columns of are taken into

VAN DER VEEN et al.: SUBSPACE APPROACH TO BLIND SPACE-TIME SIGNAL PROCESSING

account. Several updating algorithms are available; see, for

example, [30][32].
B. Step 2: Forcing the Toeplitz Property of
The next step in computing the structured factorization
is to find a description of all possible matrices
that have a block-Toeplitz structure with
block rows and are such that row
row
. The latter
condition can be true only if each row of is in the row span
of :
row
row
..
.

(7)
row

conditions can be aligned to apply to a single

These
block vector in several ways. We choose to work with

which is the generator of the Toeplitz matrix . Hence, is

in the intersection of the row span of and shifts of this row
span (suitably embedded with zeros). Alternatively, we can
say that is orthogonal to the union of the complements of
these row spans. This leads to a standard procedure to enforce
the Toeplitz property of and was originally used in [20] and
[1]. We will briefly describe the method for reference and then
show how row span intersections are computationally more
efficient in producing exactly the same result.
1) Null Space Union: Let
be a matrix whose columns
constitute a basis for
, i.e.,
is the complement of
and can be determined from the SVD of . If
has full
column rank, then has dimensions
. Moreover,
.
Using the fact that is block-Toeplitz, we obtain

177

satisfies
. Given
, we take to be a matrix
whose rows form a basis for
. Hence, the Toeplitz
is determined uniquely up to multiplication at the
matrix
right by
diag
. Now, to identify , we have to
find the factorization
, which, in the case of finite
alphabet signals, can be done using a suitable I-MIMO signal
separation algorithm, as outlined in Section III-C.
The computation of
calls for an SVD of
,
.
which is a matrix with dimensions of order
Hence, this approach requires order
operations,
which is not feasible for
or so. It is possible to
alleviate the computational requirements as we need only the
basis vectors in the null space, which does not require
a full SVD. For example, a spherical subspace updating
algorithm, if applicable, would yield a complexity of roughly
.
Row Span Intersections: We again consider (7) and let
be a basis for row
as determined in the first step. Define

(9)

where we take
consider other values for
realigned into

for now, although we will

later. The conditions in (7) can be

row
row
..
.
row
(10)

0
..

..

..

(8)
is equal to
,
The number of block columns of
where is a parameter chosen equal to the channel length
(or maybe smaller, as we will propose later). The blocks are
each shifted down over one position.
If
is a wide matrix (this gives additional conditions
on
and ), then
determines , but only up
matrix , because
also
to a left invertible

reflect the fact

Indeed, the identity matrices in each
that, at that point, there are no range conditions on the
corresponding columns of . Thus,
is in the intersection
of the row spans of
until
, and the problem
is one of determining a basis for the intersection of a set of
given subspaces. One approach, as we saw in the previous
subsection, is to compute the union of the complements of
the subspaces and take the complement again. However, it
is possible to compute subspace intersections without forming
complements. To this end, we use the fact that for orthonormal
bases
in (9), precisely the same subspace intersection is
obtained by computing the right singular vectors of a matrix
formed by stacking the basis vectors (see Appendix A), i.e.,
by computing an SVD of

..
.

(11)

178

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 45, NO. 1, JANUARY 1997

More conveniently, (for intersections,

can compute the SVD of where the copies of

), we
are each

-dimensional. This implies that

for some invertible
matrix .
With noise, we follow the same procedure. We would like
to solve
arg

block Toeplitz

dist row

row

where dist denotes the distance between two subspaces [29].

Directly finding a solution to this optimization problem is not
feasible. Instead, we find a generator for the Toeplitz matrix
by solving
arg

dist row

row

(14)

(12)
shifted over one entry and
..

..

The matrices
summarize the identity matrices present
in
, which is possible because we are only interested
in the singular values and right singular vectors of
, and
these do not change by replacing the stack of identity matrices
by
. This is immediately seen by looking at
and observing that it is the same as the square of (11).
The estimated basis for the intersection
is given by the
right singular vectors of
that correspond to the largest
:by Appendix A, those that are equal to
singular values of
(if there is no noise). As we will motivate later in Section
IV-D, the next largest singular values are close to
.
Thus, the ISI filtering process is based on distinguishing
and
. It is clear that for
singular values between
large , this becomes a delicate matter. This motivates us to
keep
small, i.e., not to make the stacking
parameter
larger than necessary.
and
in (8) is (cf. Appendix A)
The relation between
(13)
and
Hence, the right singular vectors of
are pairwise identical, except for a reversal in ordering. In
addition, their squared singular values pairwise add up to .
This is independent of any noise influence and is entirely
to be orthonormal
caused by the fact that we took and
bases of complementary subspaces. Hence, the null span union
method is just as delicate: The two methods give exactly the
same results and have the same robustness and sensitivity to
noise.
We let
denote the dimension of , which is the estimated
basis of the intersection. Under the (noise-free) conditions
intersections,
specified in Section IV-A, with
, i.e., the intersecting subspace is precisely
we obtain

which determines a matrix such that the row span of each

segment of
is as close to the row span of
as possible.
The two optimization problems are not precisely the same.
The solution of (14) is given in terms of the SVD of
and is equal to the right singular vectors corresponding to the
singular values of this matrix: those that are close to
largest
and larger than
. Thus, the proposed intersection
algorithm solves the second optimization problem.
In Fig. 2, the second stage indicates how
is formed
delays of and that the basis is obtained
from and
by linear combinations of the rows of
. The coefficients
of the filter are obtained from the - and -matrix of the SVD
of
, which is similar to the first stage. The matrices
are ignored in the figure as they only play a role in the first
and last few columns of a block of data and not during the
filtering process itself.
If we take
, then
has dimensions
. Using an SVD,
this gives the subspace intersection algorithm a complexity of
, which is linear in . Similar to stage 1, we
can consider an updating implementation of this stage as well,
although the required orthonormality of the input signals to this
stage gives rise to some interesting complications. A spherical
subspace tracker, if applicable, would yield a complexity of
order
. An investigation of the details is beyond
the scope of the paper.
If the complexity of the intersection step is too large, it may
be interesting to consider a multistage intersection approach.
Instead of computing the joint intersection of
subspaces, which requires a stack of
shifts
of , we may place, e.g., two intersection stages in cascade,
each consisting of a joint intersection of
subspaces. Likewise, it is ultimately possible to have
stages, each consisting of one pairwise intersection. This
reduces the complexity to
, and without
noise, the result is precisely the same independent of the
scheme. Numerically and with noise, however, the result is
suboptimal because it is sensitive to the order in which the
intersections are performed.
C. Step 3: Forcing the Finite Alphabet Property
At this point, we have only obtained a basis
dimensional subspace that contains

of a

VAN DER VEEN et al.: SUBSPACE APPROACH TO BLIND SPACE-TIME SIGNAL PROCESSING

179

To find , we have to determine which linear combinations of

the rows of give a finite alphabet structure. This problem is a
structured factorization of the form
, which
is interpreted as separating an instantaneous linear mixture of
finite alphabet signals. Several algorithms have been proposed
to solve such problems. In particular, a maximum-likelihood
(ML) formulation of the problem leads to

TABLE I
ILSF ALGORITHM

(15)
which is precisely the problem studied in [16] and [17]. In that
paper, two iterative block algorithms are introducedILSE
and ILSPwhich are summarized below. Starting from an
initial estimate
, the algorithms proceed as follows:
ILSE
for
a)
b)
ILSP
for
a)
b)

arg

Proj

The operator Proj denotes element-wise projection on to the

alphabet . The ILSE algorithm converges to the ML estimate
of the pair
, provided the initial estimate for is close
to the true value. Solving step in ILSE involves enumeration
over all possible combinations of symbols. The ILSP algorithm
avoids the enumeration by replacing it with a least-squares
solve for
followed by a projection onto the alphabet.
This is computationally cheaper but suboptimal. Unless the
alphabet is BPSK and the number of rows of
is small, it
is important to have a reasonably accurate initial estimate of
. Good initial points are obtained by the recently introduced
analytical constant modulus algorithm (ACMA) [33], which
is readily specialized to give closed-form eigenvalue-based
solutions for simple finite alphabets such as BPSK and MSK
[34]. Depending on the matrix dimensions and noise level,
ILSE and ILSP usually converge to a fixed point in less than
510 iterations [16]. As mentioned in the introduction, several
other I-MIMO source separation algorithms are described in
the literature, some based on different properties such as source
independence or constant modulus [18], [19], [35][38].
Alternatively, we can minimize the MMSE criterion

matrix since the rows of become orthogonal for large (as

the signals are uncorrelated). In addition, since is orthogonal
as well, is close to unitary (up to a scaling). It rotates one
orthogonal basis into another. Hence, is the solution to an
orthogonal Procrustes problem [29]. Forcing
to be close
to unitary provides one way of enforcing independent signals
in . Note that for a unitary matrix , the criteria (16) and
(15) are the same; therefore, the performance of ILSF is quite
similar to ILSP.
The ILSF step is the last stage of the filter in Fig. 2, with
(
is considered in Section V-C). The coefficients
of the filter in this stage are the entries of . Similar to ILSP
in [17], it is straightforward to replace ILSF by an updating
version, which operates in a decision directed feedback mode.
D. Alternative: Computation of

First

Instead of estimating directly, we can also first estimate

and invert the resulting channel to estimate . This is
potentially interesting since the dimensions of do not grow
with ; therefore, it can be estimated consistently. We briefly
describe the procedure, which is basically an extension of [7]
to multiple signals.
Let
be a basis of the left null space of
. Assuming
to be of full rank, we have
. Write

Then
full rank

(16)

which essentially fits the subspace

to a FA matrix . An
iterative algorithm to solve this problem is called iterative least
squares with subspace fitting (ILSF) and is listed in Table
I. It is very similar to ILSP but has the advantage that the
pseudoinverse of the th iterate
is avoided and replaced
by a pseudoinverse of , which is constant. Since
is an
orthonormal basis, this inverse is simply equal to the complex
conjugate transpose
. For a small number of sources,
each iteration requires
flops.
One aspect of the problem (16) that is different from (6) is
that we explicitly require to be full rank in order to guarantee
independent rows of . Indeed, should be a well-conditioned

0
..
.
..
.

..

..

..
.
..
.

..
.

0
If the matrix on the left is tall (this gives minimal conditions
on ), then generically its right null space specifies up to a
right block-diagonal factor diag
. For any solution
, the basis
is found from an inverse

180

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 45, NO. 1, JANUARY 1997

filter associated with

as
..
.

where
vec
is a stack of all input data. At
this point, we are back at the model
, and the
ILSE/P/F algorithm is employed to remove the ambiguity that
represents.
For the estimation of , it is only required that be of full
row rank, which is a mild condition. In particular, it is not
necessary that all channels have equal length, although certain
modifications are in order (see [22], which also contains some
identifiability results).
It is unclear whether a direct estimation of
is to be
preferred over an indirect estimation via . The former
initially forces only the structure of , neglecting that of ,
whereas the latter does the opposite. In general, estimating
is computationally easier for large
and can be done
consistently. Our experience with simulations, however, is that
estimating directly might be more accurate in the presence
of model mismatch (see Section IV-E). In addition, if the
channel lengths are not well defined (i.e., the FIR assumption
is only approximately true), row span methods can potentially
obtain a better model fit. This is because they do not force
zeros in the lower right block of
but have the freedom to
insert the actual (nonzero) coefficients instead. Finally, without
going into details, we mention that the row span methods
are almost immediately applicable to more general ARMA
(rational) channel models, in which a state space model is
assumed.
IV. ASPECTS

OF THE

ALGORITHM

A. Identifiability
Does the intersection/FA algorithm provide a unique estimate of ? This identifiability issue is the subject of the
following theorem. Similar results for
were presented
in [7] but from the point of view of estimating
from its
Hankel structure. An alternative proof appears in [26].
Theorem 1: Consider the FIR-MIMO scenario with
sources and channels of equal length
. Suppose that
the dimension conditions (5) are satisfied for some and that
the rank
and rank
.
Let
be a structured factorization of
.
Taking only the Toeplitz structure into account,
is
uniquely specified by the condition row
row
up to a left block-diagonal factor
diag
, where
is an invertible
matrix.
Taking also the FA property into account, under conditions of [17, Theorem 3.2],2
is unique up to
diag
, where
can take the form of a permutation
and a diagonal scaling by
.
We first derive the following lemma, where can be any
number of intersections between 1 and
.
2 This theorem basically requires that S contains all possible d-dimensional
columns that can be generated by the finite alphabet. This is a sufficient but
pessimistically large condition on N .

Lemma 1: For
, let be an orthonormal
basis of row
, and define
as in (9). Under the
row
is a
conditions of Theorem 1, row
subspace of dimension
and contains
(
).
Proof of the Lemma: The rank condition on
implies
has full row rank. In turn, this implies that
has
that
full row rank equal to
for
(since any subset
of the rows of
has full row rank as well).
. In investigating row
row
, we
Suppose
instead of since they span the
may as well look at
same space. Consider

..

..

..

..
.
..
.
..
.
..
.
..
.
..
.
..
.

..

..

..

..

..

..

..

..

..

..

..

where stands for an arbitrary extension as enabled by the

identity matrices in the
. The intersection removes all
rows that are not linearly dependent on the rows of the opposite
block. With suitable extensions, this means only the first and
last rows are candidates for removal. Note that they cannot be
linearly dependent on the other rows because the submatrix
of the above matrix obtained by removing the first and last
column has the same set of rows as
, which has full
row rank. Hence, both rows are removed, and the result of
the intersection is a space with precisely less rows and is
generated by the rows of
(since it is of full row
rank). The result for larger
is obtained by repeating the
same argument.3
Proof of the Theorem: Setting
in the above
lemma gives an intersection subspace of dimension , which
is spanned by the rows of
. Hence,
is unique
up to left multiplication by some invertible
matrix ;
consequently,
is unique up to left multiplication by
diag
.
Taking the FA property into account as well, [17, theorem
3.2] claims that for sufficiently diverse symbols,
is unique
up to permutation and scaling by
.
3 The rank condition on X
+1 is necessary to avoid pathological cases:
Consider, e.g., a periodic symbol matrix

s2

s0

s1

s2

s0

s1

s0

s1

s2

s0

s1

s2

S3 = s1 s2 s0 s1 s2 s0

In this case the intersections do not remove any row. Note that
3d; therefore, it is not of full rank.

S4 has rank

VAN DER VEEN et al.: SUBSPACE APPROACH TO BLIND SPACE-TIME SIGNAL PROCESSING

B. Detection of

and

If
and
have full column rank and row rank, respectively, then the rank of
is
. The number
of signals can be estimated by increasing
by one and
looking at the increase in rank of
. This property provides
a very effective detection mechanism even if the noise level
is quite high since it is independent of the actual (observable)
channel length. Furthermore, it still holds if all channels do
not have equal lengths (see Section IV-C). In case they do,
then can be determined from the estimated rank of
and the estimated number of signals by
.
It is interesting to note that there is an efficient updating
algorithm for estimating the rank of
and the corresponding
column span for all from 1 to
at once, without requiring
SVDs and using only the full-size
. The SSE-1 subspace
estimator derived in [32] is a technique for computing the
number of singular values of a matrix that are larger than a
given threshold , and a basis for a subspace that is -close to
the column span of the matrix in some norm. The algorithm
is such that at the same time, this information is produced
on all principal submatrices of
as well. Applied to
, it
produces the ranks of all
, with respect to a given
threshold at the complexity of a QR factorization.
C. Unequal Channel Lengths
For simplicity of presentation, we have only considered
channels with equal length up to now:
.
In general, however, the lengths
may be different. In that
case, it is perhaps more natural to write the factorization
with a rank-deficient
as
(17)
and
correspond to the channel and
where each
symbol matrix of source only. Generically, these factors are
of full rank
. The rank of
is thus expected to be

181

rank of the intersecting subspace is less than the rank of

.
The next intersection removes the second row and drops the
rank again by , etc. This continues until the rank of one or
more of
is exhausted, in which case, the drop in rank per
intersection is now less. The latter starts to happen once more
than
intersections are taken, and the drop in rank follows
the rank profile
. In general, after
intersections, the
rank of the resulting intersecting subspace is
rank

row
tot
tot

In principle, this allows one to determine the rank profile

and, hence, the individual channel lengths.
In an approach outlined by Liu and Xu in [21], a technique
for estimating source signals with unequal channel lengths is
presented. Essentially, the idea is to compute all intersecting
subspaces for
to
. Starting with the
smallest dimensional subspace (i.e.,
), first,
all the multiple signals that are in this subspace are separated
(by ILSF), which are precisely the
signals with channel
length
. With these signals known, the next higher
dimensional intersection (smaller ) is computed, and the
signals in it are separated, using the signals that were already
found (and their shifts) as partial initial conditions for in the
ILSF algorithm. In this way, it is in theory possible to unwind
the separation problem.
If, instead of the SVD, we apply the SSE-1 subspace
estimator [32] to the full-size
, we obtain rank
and subspace information of all principal submatrices of this
matrix as well. Since these principal submatrices are equal
to the smaller size
(ignoring
the effect of
), this gives sufficient information to find the
complete rank profile at once, as well as a way to reconstruct
all intersections.
D. Singular Value Model of Intersections

rank

tot

tot

assuming the terms in (17) are linearly independent.

To obtain row
equal to the linear envelope
row
row
, it is necessary that
, i.e.,
tot
tot

To describe the result of the subspace intersections, we need

to define a rank profile4

i.e.,
is equal to the number of sources with a channel
length
. Thus,
is monotonically decreasing from
(
) to (
), and
tot .
If we perform intersections step by step, then the first
intersection removes the top row of every
, and the
4 For

a set

E;

#( ) denotes the number of elements in

E

Under noise-free conditions, we already know (by Appendix

A) that the largest
singular values of
are
. What is the magnitude of
precisely equal to
other singular values? It is straightforward to give an answer
for
.
Since
is a basis of row , we have
for some
square matrix . Hence
can be factored as (18), shown
at the bottom of the next page, where denotes entries
that are not of interest. For large
and i.i.d. signals, the
rows of
are approximately orthogonal to each other, that
is,
, which implies that
is close to a
unitary matrix. In that case, it follows that the columns of
are asymptotically orthogonal to each other. Ignoring
the second term in the factorization (18) for the moment,
the factorization of the first term directly translates into the
SVD of
. In particular, the singular values of
are the
norms of the columns of
and, thus, are equal to
each repeated times. The
left singular vectors are just normalizations of the columns
of
, and the right singular vectors are normalizations of

182

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 45, NO. 1, JANUARY 1997

the rows of
. The latter normalization is in the order
of
. For large , it is clear that rows of the second
term (containing
) become orthogonal to
since
the inner product is proportional to
. Obviously, the
. Hence, the second
columns of this term are orthogonal to
term contributes additional singular values
each
repeated two times. Altogether, asymptotically and under
noise-free conditions,
for
has singular
values equal to
and groups of
singular values equal
to
. If we take
, then similarly,
we can show that there are
singular values equal to
, followed by the groups of
singular values equal to
. The right singular
vectors corresponding to the
largest singular values are
a basis for
; the intersections have removed
echos of each signal.
If
is not large and if there is noise, then obviously, the
singular values start to deviate from these asymptotic values,
and in particular, the gap between the singular values around
and
closes. The assessment of these deviations is
subject to future research. Such an analysis would give pointers
to suitable minimal values for (in relation to the noise level)
such that there still can be a gap.
E. Comparison by Simulation
To assess and compare the performance of the proposed
algorithms, we consider a simple but unrealistic scenario in
which all assumptions on the model are satisfied. A more
challenging test case is deferred to Section VI. We took
real-valued BPSK sources and a randomly selected
complex channel matrix with
observables and
equal channel lengths
. (
and
are equivalent
in this example because there is no modulation function
and no multiray model.) We added complex white Gaussian
i.i.d. noise with variance . The number of snapshots was
. The signal-to-noise ratio (SNR) is defined as
, which is the average SNR per signal
per observable. The relative power of both sources was set
equal.
The singular values of
are displayed in Fig. 3. Without
noise, the rank of
is expected to be
,
which turns out to be the case. The number of sources can be
identified from the graph by looking at the increase in rank of
as increases. In addition, can be estimated, assuming

Fig. 3. Singular values of X (noise-free) for a range of . The dashed

lines indicate which singular values will be masked by the noise.

equal channel lengths. The dashed lines in the graph indicate

will
at which noise level the small singular values of
, but the singular
be obscured. This level increases with
do not; therefore, it is advantageous to keep
values of
small. Below 10 dB, the true rank of
is no longer visible,
too low.
and in practice, we would estimate the rank of
are considered next (Fig. 4).
The singular values of
For convenience, they are converted into the singular values
by computing
sv
. (Recall from (13)
of
and
squared, add up to
that the singular values of
by construction.) After transformation, we expect for full
a total of
zero singular
intersections (
values corresponding to the sources and groups of
singular values around
(as indicated by the
dotted lines in the figure). For SNRs of 10 dB or more, this
is indeed the case, but for SNR 5 dB, the second source is no
longer present after intersections (Fig. 4(b)). If we take
and truncate the rank of
at 7, which is its observable rank,
then the rank equation produces an estimated channel length
. Setting
instead, we can only take
intersections, and we are left with
signals/echos. As seen in Fig. 4(c), this number of
remaining signals is still well visible, even in the SNR 5 dB
case. This indicates that for the row span intersection method,
.
there is an advantage in underestimating and

(18)

VAN DER VEEN et al.: SUBSPACE APPROACH TO BLIND SPACE-TIME SIGNAL PROCESSING

(a)

(b)

(c)
Fig. 4. Transformed singular values of

VT (n) ,

namely, (n

sv

(VT (n) )2 )1=2 . Small values indicate the number of remaining signals
after intersections. (a) SNR = 10 dB, full intersections. (b) SNR = 5 dB, full
intersections (second source not resolved). (c) SNR = 5 dB, underestimating
d and taking less intersections. After intersections, d^S = 3 signals remain.

This is confirmed by Fig. 5, which shows the bit error rates

(BER) for varying SNR for various choices of the parameters
and
. Here, we compare the directly estimating- method
(row span intersection) with the estimating- -first method
(column nullspace union). If the exact parameters are used
in the identification, the performance of both methods is
approximately the same.
In [26], a CramerRao lower bound (CRLB) is derived for
the blind equalization of one source if only the Toeplitz/Hankel
is taken into account but not the finitestructure of and
.
alphabet property. The result is readily generalized to
However, a correction to [26] by about a factor 2 is in order,
which is discussed in Appendix B.
As seen in Fig. 5(a), the methods do not reach the approximate blind CRLB (22) because they only force one of the
or ) but not both. For
factors to have structure (either
comparison, we also show the CRLB for estimation of if
is known (the performance for a zero-forcing equalizer), which
has a better performance, especially for the second signal. The
ILSE curves are obtained by running the ILSE algorithm on
, initialized by the exact
so that it gives the ML estimate
if its Toeplitz structure is ignored. As its BER
of the factor
is well above that of the blind CRLB, this indicates that using
the Toeplitz structure is relevant.

183

Fig. 5(b) shows the case where the rank of

is underestimated, which would happen in practice below 10 dB.
In that case, underestimating
as well (hence, taking less
) leads to
remaining signals after
intersections
intersections, which are separated by ILSE. As shown by the
dotted lines, this greatly improves the performance. We even
go below the blind CRLB for the second signal, which is
possible because the estimators are not necessarily unbiased
and because the FA structure is used more strongly now but
is not considered in the bound. This holds for the row span
method. Using similar techniques, we were not able to to
improve the performance of column span method. Instead, it
collapsed on rank-truncated data.
The conclusions of the simulation can be summarized as
follows:
The current and proposed blind equalization/separation
methods force only one structural property out of three:
the Hankel structure of , the Toeplitz structure of ,
and its finite alphabet structure. For the assumed model,
each of these properties by itself is approximately equally
strong. As shown by the theoretical bounds, significant
gains can still be obtained by simultaneously forcing more
than one property.
The performance of the row span method can be significantly improved by truncating the rank of
at the
noise level, underestimating the channel length , and
separating the remaining signals plus echos based on
the finite alphabet property. The column span method is
apparently not robust on truncated data.
V. ILL-DEFINED CHANNEL LENGTHS
In reality, channels do not have well-defined channel
lengths. Multipath echos with a long delay generally have a
smaller amplitude; therefore, the channel responses trail down
to zero rather than filling out a sharply defined interval in time.
In such cases, is ill conditioned, and subspace intersections
cannot be used to precisely cancel all the echos. Ill-conditioned
channel matrices are also expected for bandlimited signals
[27].
A. Effect on Intersections
To illustrate the effect of ill-conditioned channels on the
computation of the intersecting subspace, consider the impulse
response
shown in Fig. 6(a). This is the convolution
of an actual line-of-sight indoor channel at 2.4 GHz with a
raised cosine pulse (
ns, modulation index
,
oversampling rate
). The main peak has a width of
about two symbols, but there are several smaller peaks as
well. For this example, we consider the data obtained from
antennas, with
signal present, which is already
sufficient to make our point. The singular values of
are
is not clear; it is certainly
shown in Fig. 6(b). The rank of
not of low rank in a mathematical sense, and the numerical
rank depends on the truncation level we choose. To avoid an
excessively large inverse of , we would in this case decide a
or so, corresponding to an estimated channel
rank of
length of
.

184

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 45, NO. 1, JANUARY 1997

(a)

(b)

^ = L; d^ = d
2 BPSK signals. (a) Using exact values L
Fig. 5. BER performance for d
X
X = d(L + m 1): (b) Using approximate values. For
comparison, the CRLB for a zero-forcing equalizer ( yN ) is indicated, assuming perfect knowledge of H , the CRLB for the blind scenario (not using
the FA property), and the performance of ILSE initialized with the exact H .

For large , the signals are approximately orthogonal to

their shifts, and in that case, the singular values of are equal
to the singular values of
. In fact, let
be
an SVD of . Then

so that, for orthogonal

we approximate
then

and

by truncating its SVD to some rank

. If
,

Ideally,
is full rank, and
is square, but for ill-defined
, where is the
channel lengths, has size
actual (large and fuzzy) channel length. Fig. 6(c) shows the
magnitude of the entries of (up to the first 24 rows of ).
The first 10 rows of
have 11 large entries; thus, the first
10 rows of are a linear combination of 11 rows of , plus
some weaker ISI from other rows. The reason for this is that
contains a sharp peak, which is smeared by the
shifts over 11 symbols. The next few rows of
show the
influence of the smaller peaks in
: An increasing number
of rows of
get involved.
For large , we may write, as in (18),

neglecting the edge effects caused by

. Since the rows of
are close to orthogonal, the SVD of
can be written as
times
. Fig. 6(d) shows the singular values
the SVD of
when
. There is one singular value
of
and two around
, as expected.
close to
Hence, there is one vector in the intersection. This vector
is given by the product of the corresponding right singular
vector of
times
. The right singular vectors are shown
in Fig. 6(e). Since we expect the result to be basically one
symbol sequence out of
, the top row should have only

one large entry. However, it is seen that the top row has at
least eight large entries; therefore, the vector in the intersection
is still a linear combination of at least eight symbols. Thus,
the intersection did not produce the desired effect of removing
all ISI. The structure of this figure is very characteristic
and shows how the intersections work. Indeed, small singular
values of
(or
) correspond to the top and bottom rows
of
since these are repeated only a few times in
. The
large singular values correspond to rows in the middle of
, which are repeated up to times. The width of the legs
of is nearly constant. For well-defined channel lengths,
the width of the legs is expected to be 1 because the right
singular vectors corresponding to a singular value are specific
echos (rows of
; cf. (18)). The widening of the second
leg of the in our example shows the influence of the
structured noise that is introduced by truncating the rank of
at 12. Qualitatively, it can be attributed to the second peak
in
, which is partly (but not entirely) eliminated by the
truncation of
to rank 12. The truncated data matrix still
contains one or a few linear combinations of echos, but since
there are fewer combinations than symbols that play a role
after truncation, the echos cannot be removed by intersections.
The conclusion drawn from this experiment is that for actual
channels the SVD-based intersection scheme may not remove
all the ISI if the rank of
is ill-defined.
B. Effect of Taking Fewer Intersections
intersections?
What happens if we take less than
is the
We provide an intuitive analysis. Let us say that
true rank of
and that the resulting approximation error is
lumped into the noise term. Since
, it is seen
amplifies
that the noise on the rows of is not uniform:
the noise at the top rows of
less than at the later rows.
and
. If we take
Consider a simple example where
intersections, then the basis of singular vectors of

VAN DER VEEN et al.: SUBSPACE APPROACH TO BLIND SPACE-TIME SIGNAL PROCESSING

185

(a)

(d)

(b)

(e)

(c)

(f)

^ , i.e., right singular vectors of

Fig. 6. (a) Channel impulse response h(t): (b) Singular values of 10 : the numerical rank of 10 is about 12. (c) Q
(magnitude of entries). (d) Singular values of QT (12) , with d^X = 12. (e) Right singular vectors of QT (12) and (f) of QT (10) .

corresponding to singular values close to still contains

echos of each signal. A straightforward
generalization of the singular value model (18) in Section IVshows that, with
, each row of
D to
is an average of out of
rows of . Rows of
that
contain less
are formed by combining the top rows of
noise than others.
is some other combination of the symbols
In general,
and
). However, the effect that, with fewer inter(
sections, some rows of are less contaminated by noise than

10

others is still often observed. This is illustrated in Fig. 6(f),

where we have taken
intersections rather than 12.
The first two singular vectors are each a linear combination
of only three symbols, rather than eight, as we had before
with full row span intersections. has
rows, and the
third singular vector is indeed noisy; it is seen to be a linear
combination of nine symbols.
C. Multistage Intersections
Motivated by the preceding subsection, we propose a multistage intersection scheme. The first intersection stage only

186

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 45, NO. 1, JANUARY 1997

TABLE II
BLIND FIR-MIMO IDENTIFICATION ALGORITHM

(a)

(b)
Fig. 7. (a) Relative power and (b) response to a raised-cosine pulse (T
ns,  = 0:5) of two measured indoor channels.

takes the well-defined intersections: At most,

, where
is an underestimate of the channel
length of signal , but without prior knowledge of channel
lengths, perhaps even only
. This produces a basis
which is too large. In fact, it contains
rows and is a basis for
. The second intersection stage
has to remove the remaining ISI. Instead of using an SVD, we
combine this stage with the separation stage (ILSF or some
other I-MIMO algorithm), i.e., the finite alphabet property is
used to do the remaining equalization and the signal separation
as well.
In principle, we can apply ILSF directly on . We could
and select those rows that are
recover all rows of
not shifts of each other and that have the smallest deviation
are a basis
from the alphabet. However, since the rows of
, it is more general to prepare for
for the Toeplitz matrix
to a Toeplitz
a subspace intersection step, i.e., augment
, where is some small number. Similar to the
matrix
from , we construct
by stacking
construction of
shifted copies of
(omitting , ). However, instead of
applying an SVD to
, we apply ILSF so that signals and
echos are separated based on finite alphabet properties. The
resulting variance on the symbol estimates should be lower
since ILSF has the same degrees of freedom as the SVD but
is not blind to symbol variance. The value for could vary
. A larger will always result in
between and
symbol estimates with lower variance. In the latter extreme
case, we act on the same data that a secondary SVD-based
cannot
subspace intersection stage would use. However,
be too large because the complexity and the reliability of
convergence of ILSF to the global minimum deteriorates with
growing dimensions.

Fig. 8.

Singular values of

m for m = 2;

111

=6

; 10.

The resulting algorithm has the general structure of Fig. 2

and is listed in Table II. The significance of taking
will
be clear from the simulation results in Section VI.
VI. SIMULATION RESULTS
In this section, we report on a test of the algorithm in an
off-line experiment, in which we simulate the reception of a
number of BPSK signals through an indoor wireless channel
at 2.4 GHz. The channel impulse responses are derived from
experimental data measured in an office at FEL-TNO, which
is in The Hague, The Netherlands, in 1992 [39].5
5 We are grateful to G. J. M. Janssen (now at TU Delft) for sharing his
measurement data.

VAN DER VEEN et al.: SUBSPACE APPROACH TO BLIND SPACE-TIME SIGNAL PROCESSING

The office has dimensions 5.6 m

5.0 m and height 3.5
m. The actual measurement set-up had a transmit antenna in
the center of the room at a height of 3.0 m and a receiving
antenna cluster located at varying positions at a height of 1.5
m. The cluster consisted of six wideband antennas spaced
in a circular array.
Assuming reciprocity (not quite true), we can pretend to
simulate a central basestation antenna array of up to six
elements, receiving a superposition of signals from a number
of user locations. We have used data from two such locations:
one with a direct line of sight (RMS delay spread
7.3
ns) and one without LOS (RMS delay spread
16.7 ns).
The relative powers in the frequency domain are plotted
in Fig. 7(a). Fig. 7(b) shows the amplitude of the impulse
responses to a raised-cosine pulse (
ns,
demodulated to baseband from a carrier frequency of 2.4
GHz), each normalized to unit power. We do not have any
application in mind with these numbers; they are chosen to
provide an ambitious test case that uses all of the measured
bandwidth.
In the experiment, we took
BPSK sources, transmitted
over the above channels,
antennas,
times
oversampling, and
samples. The received power
of both signals was scaled to be equal, and we added complex
white Gaussian noise with variance
such that the signal-tonoise ratio SNR
15 dB per antenna
per sample per signal. The singular values of are plotted in
Fig. 8 for a range of values of . It is seen that the numerical
rank of
(
) cannot be estimated very well, but clearly,
, as deduced from the horizontal shifts for increasing .
, it seems reasonable to set
in the range 2030,
For
which makes the observable channel length equal to 49,
if the channels had equal lengths. As in the single-user case
(Fig. 6(a) and (b)), the actual channel lengths cannot really be
deduced from the data.
Fig. 9 shows the standard deviations of the symbol estimates
(before classification as
) for a range of parameter settings:
, number of intersections , and secondary
estimated rank
equalizer . ILSF initialized with
was used as finite
alphabet separation algorithm. The values of these parameters
have a deliberate impact on the performance, but precisely
how to find the best settings a priori is an open problem. As
, but
a general observation, it is possible to underestimate
in that case, it is essential that is taken small (
) and
that ILSF is used as an equalizer as well (
). However,
should not be taken too large because then, the matrices on
which ILSF acts become too big, leading to an abundance of
local minima. To put the graphs into perspective, note that at
this noise level, the standard deviations of the symbol estimates
in an ISI-free scenario, where
and the antennas
and oversampling produce
independent observations per
symbol would be
. (The factor 2 is due to
the transformation of
to a real matrix, as in (3).)
VII. CONCLUSION
We have presented an algorithm for blind sequence estimation of multiple digital sources in a general multipath

187

environment. The algorithm uses information from multiple

sensors, oversampling to exploit the constant symbol period,
and the finite alphabet property of digital signals. It is set
in a deterministic framework and uses subspace properties
of the underlying structured matrix factorization problem.
This approach is effective in situations where the channel
lengths are well determined. We have indicated some problems
that may arise in subspace intersections algorithms when
the channel lengths are not well defined and suggested a
modification that should give improvements for channels with
well-defined peaks.
APPENDIX A
INTERSECTION OF SUBSPACES
Let
ments

be subspaces in C with orthogonal comple. Then,

The computation of the intersection via this equation requires

the formation of three orthogonal complements. With
matrices whose columns form orthonormal bases for
,
we can obtain a basis for the intersection of
and
from
the kernel of
. With noisy data, this requires the
computation of an SVD of
: A basis of the estimated
kernel is given by the singular vectors that correspond to small
singular values.
In our application, the dimension of the
is independent
of
so that the dimension of the complements grows with
. This means that for large
, it is not attractive to
compute the intersection in this manner. We show in the
following proposition that precisely the same information may
be gleaned from the large singular values and corresponding
singular vectors of a matrix
, where
are
orthonormal bases of
.
Proposition 1: Let
be subspaces in C with orthonormal bases
, and let
be an SVD. Suppose that
are orthonormal bases
for
,
. Then,
has an SVD
for some unitary matrix
.
Proof: Since
,
we have

Substituting
above equation with

and multiplying the

, we obtain

Since both
and
are diagonal, this implies that there is
a unitary matrix
such that
is diagonal.
This, however, constitutes precisely an SVD for
This result is readily generalized to the joint intersection
of subspaces
. Likewise, we compute an
SVD of
but now look for singular values that
are close to
.

188

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 45, NO. 1, JANUARY 1997

(a)

(b)

(c)

(d)

Fig. 9. Standard deviations of signal estimates for

APPENDIX B
APPROXIMATE CRAMERRAO BOUNDS
Suppose
, where
is a white i.i.d.
complex Gaussian noise process with covariance matrix
.
For simplicity of future notation, let us specialize to the
case of real signals (e.g., BPSK signals
). Define
vectors of unknown parameters
vec
and
. We assume that the number of sources
is known and that the channels have equal known channel
length . If we do not take into account that the entries of
belong to a finite alphabet, then the concentrated Fisher
information matrix for
is derived in [26] as

= 7 and varying settings of ^X

; n; p

an invertible
matrix . Indeed, the dimension of the null
space of is observed to be
in generic examples. To fix
, one has to assume that certain symbols are known.
, knowing the value of one symbol of suffices,
For
and the variance of the remaining estimates is obtained by
deleting the corresponding column of
. Let
be equal
to
with the column corresponding to the known symbol
taken out, and define and
accordingly. Then, the CRLB
on the covariance of
is

(the subscript denotes the (1,1) block of

the partitioning of ) so that, in particular,
var

(originally for a single signal, but the results are readily

generalized for
and adapted here for a real data model).
The CRLB that describes the lower bound on the covariance of
any unbiased estimator for and is obtained by inverting .
However, as noted in [26], this matrix turns out to be singular
because there is ambiguity in the parameter values: Without
forcing the FA property, we can only identify
and up to

diag

(19)
following

[CRLB with training, no FA]

(20)
This is basically the result in [26], where it is also noted
and , its
that although the bound is dependent on both
dependence on is only weak in practice. However, a number
of remarks that go beyond [26] are in order.
1) It makes a difference which symbol is assumed to be
known. Not surprisingly, knowing one of the center symbols gives significantly lower variances than knowing

VAN DER VEEN et al.: SUBSPACE APPROACH TO BLIND SPACE-TIME SIGNAL PROCESSING

one of the first or last

symbols because these play
a less significant role in
. Additionally, the variances
of the symbols in the range
are usually
approximately equal to each other, but the first and last
symbols have a significantly larger variance. In
the computation of the expected bit error rates, we have
taken these tail symbols out of consideration.
2) The result (20) strictly speaking applies to a scenario in
which we have a training sequence of length . It is
readily generalized for training sequences longer than 1
by leaving out more columns of
.
3) The above remark implies that in the actual blind
algorithm, the above lower bound on the variance is
too large by about a factor 2. Indeed, (20) is valid for
estimates where the variance of one symbol,
say, is
made zero. This is conceptually done by estimating any
sequence and then normalizing the th entry by dividing
the estimated sequence by the estimated value of and
multiplication by its desired value. Assuming relatively
small variances, the division causes the variance of all
other symbol estimates to be enlarged by the variance of
the estimate of .6 In the actual blind scheme, we do not
normalize on a single symbol but normalize
. In
that case, the lower bound (20) is too high and formally
not applicable. To attempt to correct for this, we have to
estimate the variance of , e.g., as median diag
and subtract to get
var

diag
median diag
blind
[approx. blind CRLB, no FA,
]

(21)

(We take the median instead of the mean to avoid the

influence of outliers at the tails of the sequence.) This is
the (approximate) CRLB for a blind scheme that relies
on a structured factorization
, not taking the
finite alphabet into account other than for removing the
ambiguity. If all estimates have approximately equal
variance, the originally derived bound (20) is about a
factor 2 too high.
For
, roughly the same derivation holds, except that
we have to pretend that more symbols are known because the
ambiguity factor now has degrees of freedom. Hence, we
have to fix symbols of signals, i.e., a
submatrix
of
somewhere in the center of . An extra complication is that
this submatrix
has to be full rank or else some ambiguity in
remains. Hence, in computing the bound, we have to select
independent columns of , which are located somewhere in
the center, and delete the corresponding columns of
to
obtain
. After this, the bound (20) is derived as before.
The correction for the unnatural normalization as assumed in
that bound is somewhat more intricate now. Indeed, before
normalization, let us say we have symbol estimates
blind

and
are the exact symbols, and
and
represent the
noise on the estimates. Normalization to arrive at an estimate
6 Here, the first-order approximation s (s + e )01 (s + e)
r r
r
e is used, as well as the BPSK assumption si = 1.

j j

 s 0 er s0r 1 s +

189

in which the known symbols have zero variance gives the

modified estimates
for which the CRLB (20) holds as

Note that
can actually amplify the noise contribution by
. In estimating the correction on the bound, assume (not
entirely correctly) that the columns of
are independent, zero mean, and have equal distribution E
. Let
be the th column of , and then
E
The left-hand side of this expression is given by the uncorsubmatrix of
rected CRLB, namely, , which is the
in (20) corresponding to . It follows that an estimate of
and an approximate lower bound on var blind can be
obtained as
median
var

(22)

blind

For
BPSK signals,
expression reduces to (21).

always, and the above

REFERENCES
[1] 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. 99102, May 1995.
, Blind identification of FIR channels carrying multiple finite
[2]
alphabet signals, in Proc. IEEE ICASSP, 1995, pp. 12131216.
[3] W. A. Gardner, A new method of channel identification, IEEE Trans.
Commun., vol. 39, pp. 813817, June 1991.
[4] L. Tong, G. Xu, and T. Kailath, Blind identification and equalization
based on second-order statistics: A time domain approach, IEEE Trans.
Inform. Theory, vol. 40, pp. 340349, Mar. 1994.
[5] L. Tong, G. Xu, B. Hassibi, and T. Kailath, Blind channel identification
based on second-order statistics: a frequency-domain approach, IEEE
Trans. Inform. Theory, vol. 41, pp. 329334, Jan. 1995.
[6] L. Tong, G. Xu, and T. Kailath, Fast blind equalization via antenna
arrays, in Proc. IEEE ICASSP, 1993, pp. IV:272274.
[7] E. Moulines, P. Duhamel, J. Cardoso, and S. Mayrargue, Subspace
methods for the blind identification of multichannel FIR filters, IEEE
Trans. Signal Processing, vol. 43, pp. 516525, Feb. 1995.
[8] D. Slock, Blind fractionally-spaced equalization, perfect-reconstruction
filter banks and multichannel linear prediction, in Proc. IEEE ICASSP,
vol. IV, 1994, pp. 585588.
[9] G. Xu and H. Liu, A deterministic approach to blind identification of
multi-channel FIR systems, in Proc. IEEE ICASSP, 1994, vol. IV, pp.
581584,
[10] F. R. Magee and J. G. Proakis, Adaptive maximum-likelihood sequence
estimation for signaling in the presence of intersymbol interference,
IEEE Trans. Inform. Theory, vol. IT-19, pp. 120124, Jan. 1973.
[11] G. Ungerboeck, Adaptive maximum-likelihood receiver for carriermodulated data transmission systems, IEEE Trans. Commun., vol.
COM-22, pp. 624636, May 1974.
[12] G. Picchi and G. Prati, Blind equalization and carrier recovery using a
stop-and-go decision-directed algorithm, IEEE Trans. Commun., vol.
COM-35, pp. 877887, Sept. 1987.
[13] Z. Ding, Blind equalization based on joint minimum MSE criterion,
IEEE Trans. Commun., vol. 42, pp. 648654, Feb. 1994.
[14] N. Seshadri, Joint data and channel estimation using fast blind trellis
search techniques, in Proc. Globecom, 1991, pp. 16591653.
[15] D. Yellin and B. Porat, Blind identification of FIR systems excited by
discrete-alphabet inputs, IEEE Trans. Signal Processing, vol. 41, pp.
13311339, Mar. 1993.
[16] S. Talwar, M. Viberg, and A. Paulraj, Blind estimation of multiple cochannel digital signals using an antenna array, IEEE Signal Processing
Lett., vol. 1, pp. 2931, Feb. 1994.

190

[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]

[36]

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 45, NO. 1, JANUARY 1997

, Blind separation of synchronous co-channel digital signals

using an antenna array. Part I: Algorithms, IEEE Trans. Signal Processing, vol. 44, pp. 11841197, May 1996.
A. Swindlehurst, S. Daas, and J.Yang, Analysis of a decision directed
beamformer, IEEE Trans. Signal Processing, vol. 43, pp. 29202927,
Dec. 1995.
A. Belouchrani and J.-F. Cardoso, Maximum likelihood source separation for discrete sources, in Proc. Eusipco, 1994, pp. 768771.
H. Liu and G. Xu, A deterministic approach to blind symbol estimation, IEEE Signal Processing Lett., vol. 1, pp. 205207, Dec.
1994.
H. Liu and G. Xu, Multiuser blind channel estimation and
spatial channel pre-equalization, in Proc. IEEE ICASSP, 1995, pp.
17561759.
K. Abed-Meraim, P. Loubaton, and E. Moulines, A subspace algorithm
for certain blind identification problems, IEEE Trans. Inform. Theory,
vol. 42, July 1996.
S. Andersson, M. Millnert, M. Viberg, and B. Wahlberg, An adaptive
array for mobile communication systems, IEEE Trans. Vehic. Technol.,
vol. 40, no. 1, pp. 230236, 1991.
J. Gunther and A. L. Swindlehurst, Algorithms for blind equalization
with multiple antennas based on frequency domain subspaces, in Proc.
IEEE ICASSP, Atlanta, GA, 1996, vol. 5, pp. 24212424.
A. J. van der Veen, M. C. Vanderveen, and A. Paulraj, Joint angle and
delay estimation using shift-invariance properties, submitted to IEEE
Signal Processing Lett.
H. Liu and G. Xu, Closed-form blind symbol estimation in digital communications, IEEE Trans. Signal Processing, vol. 43, pp. 27142723,
Nov. 1995.
A. J. van der Veen, Resolution limits of blind multi-user multi-channel
identification schemes - the bandlimited case, in Proc. IEEE ICASSP,
Atlanta, GA, May 1996, pp. 27222725.
S. N. Diggavi, Y. Cho, and A. Paulraj, Blind estimation of multiple cochannel digital signals in vector FIR channels, in Proc. IEEE Globecom,
1995.
G. H. Golub and C. F. Van Loan, Matrix Computations. Baltimore,
MD: John Hopkins Univ. Press, 1984.
P. Comon and G. H. Golub, Tracking a few extreme singular values
and vectors in signal procesing, Proc. IEEE, vol. 78, pp. 13271343,
Aug. 1990.
B. Yang, Projection approximation subspace tracking, IEEE Trans.
Signal Processing, vol. 43, pp. 95107, Jan. 1995.
A. J. van der Veen, A Schur method for low-rank matrix approximation, SIAM J. Matrix Anal. Appl., vol. 17, pp. 139160, Jan. 1996.
A. J. van der Veen and A. Paulraj, An analytical constant modulus
algorithm, IEEE Trans. Signal Processing, vol. 44, pp. 11361155,
May 1996.
A. J. van der Veen, Analytical method for blind binary signal separation, accepted for publication in IEEE Trans. Signal Processing, Oct.
1996.
J.-F. Cardoso, Super-symmetric decomposition of the fourth-order
cumulant tensor. Blind identification of more sources than sensors,
in Proc. IEEE ICASSP, Toronto, Ont., Canada, 1991, vol. 5, pp.
31093112.
, Iterative techniques for blind source separation using only
fourth-order cumulants, in Signal Processing VI: Proc. EUSIPCO-92,
Brussels, J. Vandewalle, Ed. New York: Elsevier, 1992, vol. 2, pp.
739742.

[37] L. Tong, Y. Inouye, and R. Liu, Waveform-preserving blind estimation

of multiple independent sources, IEEE Trans. Signal Processing, vol.
41, pp. 24612470, July 1993.
[38] K. Anand, G. Mathew, and V. U. Reddy, Blind separation of multiple
co-channel BPSK signals arriving at an antenna array, IEEE Signal
Processing Lett., vol. 2, pp. 176178, Sept. 1995.
[39] G. J. M. Janssen and R. Prasad, Propagation measurements in an indoor
radio environment at 2.4 GHz, 4.75 GHz and 11.5 GHz, in Proc. 42nd
VTS Conf., Denver, vol. 2, 1992, pp. 617620.

Alle-Jan van der Veen (S87M94) was born in The Netherlands in 1966. He
graduated (cum laude) from the Department of Electrical Engineering, Delft
University of Technology, Delft, The Netherlands, in 1988, and received the
Ph.D. degree (cum laude) from the same institute in 1993.
Throughout 1994, he was a postdoctoral scholar at Stanford University,
Stanford, CA, in the Scientific Computing/Computational Mathematics Group
and in the Information Systems Laboratory. At present, he is a researcher in
the Signal Processing Group of DIMES, Delft University of Technology. His
research interests are in the general area of system theory applied to signal
processing, in particular, system identification, time-varying system theory,
and in numerical methods and parallel algorithms for linear algebra problems.
Dr. van der Veen is the recipient of a 1994 IEEE SP paper award.

Shilpa Talwar received the M.S. degree in electrical engineering and the
Ph.D. degree in scientific computing and computational mathematics from
Stanford University, Stanford, CA, in 1996.
She is currently employed by Stanford Telecom, Sunnyvale, CA. Her
research interests include wireless communications, array signal processing,
and numerical linear algebra.

Arogyaswami Paulraj (F91) was educated at the Naval Engineering College,

India, and at the Indian Institute of Technology, New Delhi, where he received
the Ph.D. degree in 1973.
A large part of his career to date has been spent in research laboratories
in India, where he supervised the development of several electronic systems.
His contributions include a sonar receiver in 19731974, a surface ship sonar
in 19761983, a parallel computer in 19881991, and telecommunications
systems. He has held visiting appointments at several universities: the Indian
Institute of Technology, Delhi, from 1973 to 1974, Loughborough University
of Technology, UK, from 1974 to 1975, and Stanford University, Stanford,
CA, from 1983 to 1986. His research has spanned several disciplines, emphasizing estimation theory, sensor signal processing, antenna array processing,
parallel computer architectures/algorithms and communication systems. He is
currently a Professor of Electrical Engineering at Stanford University, working
in the area of mobile communications. He is the author of about 90 research
papers and holds several patents.
Dr. Paulraj has won a number of national awards in India for his contributions to technology development.