INTERNATIONAL JOURNAL OF ADAPnVE CONTROL AND SIGNAL PROCESSING, VOL.

10,213-237 (1996)

ADAPTIVE BLIND SYSTEM IDENTIFICATION USING

WEIGHTED CUMULANT SLICES
JOSEP VIDAL AND JOSE A. R. FONOLLOSA
Department of Signal Theory and Communications. Universitat Politecnica de Catalunya, Campu.7 Nord,
MODUL-DS, E-08034,Barcelona, Spain

SUMMARY
Many linear methods have been proposed in the literature to blindly estimate the ARMA parameters of a
time series using HOS. Nevertheless, they are mainly off-line and not much has been done in the adaptive
case. The method proposed in this contribution is the adaptive version of the w-slice method. The
recursion is based on the inversion lemma when attempting the solution of an underdetermined matrix
equation. The system impulse response can be recovered regardless of the ARMA or MA character of the
system. The number of operations depends on the square of the system order and is considerably reduced
with respect to previous approaches. Application to channel deconvolution is shown.

KEY WORDS inversion lemma; cumulant slice; equalization; digital communications

1. INTRODUCTION
Time-varying linear systems appear in a wide range of real problems. Control of plants with
slowly varying parameters, equalization of non-stationary communication channels, adaptive
spectrum estimation, signal detection, echo cancelling and adaptive beamforming are, among
many others, applications that require the update of the estimates as new data become
available.' The updating algorithm departs from an initial state which assumes complete
ignorance of the environment. After successive operations the method should converge in some
statistical sense to a desired solution. Very often the desired solution is the optimum Wiener one
and the algorithms may converge in the mean (also called stochastic convergence) or exactly,
leading to the popular LMS and RLS algorithms respectively. However, the Wiener solution is
not the only one that can be used. In particular, this work is devoted to finding recursive
methods that converge to the w-slice solution.'
The way all adaptive methods are evaluated and finally chosen for a certain application is by
considering various factors: I rate of convergence, defined as the number of iterations required
to obtain a given error with respect to the exact solution; misadjustment, which measures the
amount of error between the final solution and the exact solution; robustness to ill-conditioning
of the data (i.e. contaminating noise); computational demands, namely the number of floating
point products per iteration and the amount of memory required to store the intermediate results:
numerical properties, conceining the stability of the algorithms to round-off errors. The in-
depth analysis of all these factors is beyond the scope of this work. However, considerations
about the rate of convergence will be drawn from extensive simulations. Computational
demands and performance when noise is present will also be considered.

CCC 0890-632719610302 13-25

0 1996 by John Wiley & Sons, Ltd.
2 14 J. VIDAL AND J. A. R. FONOLLOSA

Adaptive system identification has been a cumbersome topic in signal processing for some
years, especially if the system (possibly non-minimum phase) has to be estimated in a ‘blind’
fashion, i.e. with no knowledge of a reference signal. There are many cumulant-based linear
approaches proposed in the literature to estimate the ARMA parameters given the output.
Nevertheless, they are mainly off-line and suitable for stationary conditions and not much has
been done in the adaptive case. Among the few, Swami and Mende13proposed a ‘double-lattice’
algorithm suitable for adaptive estimation using the HOS Yule-Walker equations. However, the
instrumental variable is chosen in such a way that only one slice (instead of y + 1) ia
considered, thus preventing identifiability in a number of cases.4 To overcome this problem, we
proposed’ the use of another instrumental variable which guarantees the full rank of the matrix.
The philosophy of the method is similar to the w-slice approach in the sense that the
instrumental variable is constructed using a linear combination of the input samples, where the
weights are the estimated AR parameters. Since the equation matrix being solved is full-rank in
all cases, identifiability is assured. Simulations showed good performance, although theoretical
convergence has not been proven. The recursion is of the LMS lattice type, so the number of
operations per iteration is proportional to the order of the model.
On the other hand, Friedlander and Porat6 use the so-called ‘overdetermined recursive
instrumental variable’ (ORIV)’ to compute recursive estimates of the Giannakis-Mendel
(GM) method.’ The method solves an overdetermined system of equations in which second-
and higher-order (third- or fourth-order) cumulants are combined. The use of the inversion
lemma makes the computational burden similar to conventional RLS, which is proportional to
the square of the model order, The drawbacks of the method are twofold. Some arise from the
use of the GM method. The use of second- and third-order statistics prevents unbiased
estimation in the case of coloured Gaussian noise. Moreover, the method is valid only in the
MA case. Other difficulties come from the use of an RLS-like approach, i.e. the accumulation
of round-off errors causes the matrices involved to be ill-conditioned and the algorithm
unstable. This point has been successfully solved by Porat and Friedlander’ using so-called
square root RLS.’’ This issue is of great value to the development of the method shown in the
sequel.
The paper is structured as follows. In Section 1 the w-slice method is briefly sketched and the
inversion lemma is used to derive a recursive underdetermined algorithm. Section 2 is devoted to
deriving a ‘square root’ version of the method and is heavily based on previous work by Porat
and Friedlander.’ Premises for the application of the method to an adaptive blind equalizer are
stated in Section 3. Finally, Section 4 shows some simulations for third- and fourth-order
cumulants.

2. THE UNDERDETERMINED RECURSIVE LEAST SQUARES

It was shown in Reference 2 that the use of a linear combination of cumulant slices allows
identification of the impulse response of an MA linear system in two steps, first by solving the
underdetermined system of equations

s a w= 1 (1)
and then by using the resulting w-vector in the equation

h,=S,W (2)
 ADAPTIVE BLIND SYSTEM IDENTIFICATION 215

where the matrices S are defined (when using third-order cumulants) as

s=[;:]=

hois the estimated impulse response, the vector 1 is (0, ..., 0, and p is in general an upper
bound of the model order. Equation (2) was first derived by Giannakis" for ARMA models: if
the weight vector w contains the true AR parameters, then a linear combination of slices yields
the impulse response
P
a(i)C,,(q - i, n, 0, ...,0) = y,,b(q)h(n), h(0) = 1, k > 2
i=O

This result was used later in Reference 5 to derive a recursive AR estimation algorithm.
However, there is a fundamental need for the true AR parameters, which cannot be estimated
directly from equation (l), since the matrix S, is not always full-rank." Also in Reference 12
we have shown that the inherent limitation of the approach in Reference 5 can be overcome and
the same estimation scheme based on equations (1) and (2) can be used to obtain consistent and
unbiased estimates of the impulse response of a causal AR system. Note that this result is not
obtained under the assumption of a truncation using a large enough MA model. In the same
paper it was shown that the cumulants of the AR and MA parts of a non-causal ARMA model
could be obtained using the approach in Reference 13 and that the w-slice approach is very
useful (versus the C ( q ,k) approach) for recovering both the AR and MA coefficients. Later
0nI4 the result was generalized to any linear (ARMA) system. It was proven that the same
framework can yield consistent unbiased estimates of the impulse response of a causal ARMA
system provided that two conditions are met.
1. At least a basic set of proper slices (corresponding to the columns of S) must be included,
namely the slices 0, ..., - p for an AR(p) system and the slices q, ..., q - p for an
ARMA(p, 4 ) system.
2. The causality condition on the impulse response (as imposed by equation (1)) must hold
for at least p samples (corresponding to the number of rows of S,).
This more general result has not appeared yet in an easily accessible publication, so for the
moment let us assume that all the models we will use are either purely AR or purely MA. One
of the simulations will show, however, the behaviour of the approach with an ARMA model.
On the other hand, for an MA(q) system the necessary number of slices is one, which
leads to the popular C ( q , k) method. More generally, the w-slice approach may incorporate
as many cumulant slices as desired (in the columns of S). The number of rows of S , and S ,
'must be at least q + 1 to obtain non-biased estimates. As a conclusion it is worth noticing
that the same framework can be used to estimate the impulse response of any linear system
provided that an upper bound of the true orders is known. Moreover, it can cope with
216 J. VIDAL AND J. A. R.FONOLLOSA

different cumulant orders just by adding columns to the matrix S, so that good statistical
properties can be obtained. Notice that the number of columns that can be added to the
matrix S is unlimited. Therefore equation (1) represents an underdetermined system of
equations. Among the solutions the classical approach is to find the one with minimum
norm, which leads naturally to the concept of the pseudoinverse of a more-columns-than-
rows matrix:
w = s:1= s:(sosyl
Note that this would be the optimum solution if all the estimated cumulant samples appearing in1
S had the same variance. The w-slice approach was generalized to the estimation of MIMO
systems in Reference 15.
An implicit assumption appearing in equation (1) and in all output-only identification methods
is that the zero tap of the estimated impulse response is forced to be unity regardless of its true
value. This term conveys the carrier phase information when dealing with communication!;
signals. This results in a phase and amplitude ambiguity in the restored constellation, i.e. it may
be rotated depending on the value of h(0). Fortunately, if the constellation moments are known,
h(0) can be estimated up to a phase ambiguity using the method sketched in Reference 16 and
reproduced later on in Section 3.
To solve equations (1) and (2) recursively, it is useful to write them as
x,z,w= 1, ho=Y,Z,W
where the matrices are defined (when using third-order statistics) as

0 ...
0 ...
x: = ...
T
Y, =
XO

xI-21J ...
1
(3)

2, =

Recall that y is an upper bound of the true (AR or MA) system order. If the observed linear
process (x, 1 is complex, the cumulant sequence can be defined in various ways. Here, with 110
loss of generality, we consider the definition conveying no conjugation of the arguments. Smdl
modifications have to be made if conjugation is considered. This will be investigated in Section
3. Notice that the subscript p indicates the number of samples of the impulse response that is to
be recovered. That is, the number of columns of X and Y is p + 1. The number of slices to be
used is the number of columns in 2. Equation (3) has been written considering the use of
2p + 1 slices, but the number of columns can be reduced along the lines given in References 2,
 ADAPTIVE BLIND SYSTEM IDENTIFICATION 217

12 and 14. The solution of equation (1) has been achieved using the pseudoinverse:
w, = SWpl= s;(sos;)-9 = z;x,H(x,z,z,Hx,H)-'1
(4)
6, = SWpW, = Y,Z,W,"
= Y,zp;x;(x,z,z,Hx,H)
-I1
This algorithm has been shown in Reference 2 and 12 to be asymptotically consistent. By
increasing the number of rows of the matrices X,Y and Z,more samples are considered and
the accuracy of the estimates is improved.
The notation that follows indicates how the various matrices are updated as more data become
available:

where x,,! indicates the ( p + 1) x 1 column added to X,, at time t + 1 and A is an exponential
forgetting factor which enables the algorithm to track slowly varying parameters. The
derivation of the method is based mainly on how the inverse in equation (4) is calculated.
Let
P, = (X,Z(Z,HX,H) -1 (6)
By substituting the relations in ( 5 ) into (6),we find a recursion for the inverse of P:
Pf-+Il= x,+,z,+,z:+,x:+1
= (nxrzz,
+ x,+d+I>(nz,"x,"
+ z:+,x:+,>
= W'+ A@r+lX:+I +k+l@:+l+ ~ , + l ~ : + l ~ ~ + l ~ ~ + l (7)
where

The matrix P can be written more compactly as:

p;:I = n2pr-'+ @,+A-:,#:+l (9)
where

By using the inversion lemma' on (9), we obtain

Pt+i=[Pr - Pt#,+I (JzAr+i+ @:+IPt$r+I ) - ' #jl+IPil/n'
The weights can be updated recursively by using

and then the impulse response estimate is obtained as

218 J. VIDAL AND J. A. R. FONOLLOSA

where S = S , and L = S,. The algorithm may be initialized in two different ways, either by
computing initial estimates of the matrices L and S using the first 2y signal samples, i.e.
s, = x,z,, P/;=( s , s y , L, = Y,Z,
or by assuming the matrix P diagonal with a large norm, i.e.
1
so = P [ I , + , 101, Po = 7 I,+ 1, L,=P[I,+ilol, P& 1
P
The value of the parameter p allows some control over the initial convergence of the algorithm,
i.e. for the RLS algorithm a small positive value of p provokes a large norm of the P-matrix
and a fast initial update. The recursion is done by taking equations (8) and (10)- (13) and noting;
that the matrix P is Hermitian, according to equation (6). Table I summarizes the necessary
steps for the URLS method when N slices are used. This algorithm converges to the exact
solution as t tends to infinity, in a stationary environment, when 1 = 1. If the data have time-
varying parameters, some tracking on the coefficients is desired. This is achieved by choosing ii
value of 1 between zero and unity. In this case the algorithm no longer converges to the exact
solution, but there is a residual stochastic error in the coefficients and convergence is in the
mean. It is customary in such algorithms to make 1 time-dependent in such a way that when the

Table I. Summary of third-order URLS w-slice algorithm

Initialization
S"= P u p + 1 I01
1
P,= - I
P2 P + l

Multiplications

2x2 P+l
 ADAPTIVE BLIND SYSTEM IDENTIFICATION 219

-
c4.v(-p,-p, 0) -.* C4.A-p, 0,O) -.* C4J-p, p , 0)

c4.y(-19-p70) .,. c4.v(-1,0, 0 ) C4.,.(-1,p90)

[
.a.

= = c4x(o-p,
, 0) ... 0,o)
c4.v(0, ... c4.v(0,p,0 )

- c4.v(P, - 9 9 0) C 4 h , 090) *a. p , 0)

c4,v(p,

-
M4.v(-p, p , 0) - r(O)r(2p) - 2r(0l2

M4.,-(-1,p,0) - r(O)r(p+ 1) - 2r(l)r(p)

M4.S(O,P10) - 3r(O)r(p)

Wdp, P , 0) - 2 f W 2- r(O) -
220 J. VIDAL AND J. A. R. FONOLLOSA

into account the fourth-order moments:

z,=

The autocorrelation vector r must be computed recursively, so a possible choice is to use the
expression
+ xru:,
r:= h:+l ... x,_~x, x,-,... x,J
where u r = [x,-/, (15)
Extra normalization of r is done by taking into account the number of samples used up to the
present updating, i.e.

(
r, c r,/rz, where ti = min t , -
lli)

As was done in equation (6), all these matrices should be updated as more data become
available. In particular the following notation will be useful:

This notation allows us to re-express the equations in (13) by using

Hence equations (4) may be reutilized in the fourth-order case just by substituting the matrix Z
by the matrix N in equation (17). The cumulant definition involving no conjugation in the
arguments has been used again. Note that when using equation (15) as an estimation of the
autocorrelation in combination with equation (17), some bias is introduced. This is due to the
fact that not all the terms in (17) are equally affected by the number of samples used. That is,
while the term Xr+lZ,+I times t + 1 equals the fourth-order moment matrix, Xr+lrr+,(0)T,-l has
to be multiplied by ( t + 1)2so that it yields the product of two autocorrelation samples matrices.
 ADAPTIVE BLIND SYSTEM IDENTIFICATION 22 1

Therefore equation (15) has to be used after averaging it over the number of samples being
considered in its estimation. Remember that the inverse of 1 - 1 is, roughly speaking, a measure
of the memory of the algorithm, i.e. of the number of signal samples being used to compute
the estimates. When A = 1, the algorithm has infinite memory. After these considerations the
fourth-order method can be deduced from the one in Table I except for some slight
modifications. The total number of multiplications required here depends also on the number N
of slices used. When 2 p + 1 slices are taken into account, the total number of floating point
operations is not significantly higher than for the third-order version: 1lp2+ 24p + 16.

3. SQUARE ROOT UNDERDETERMINED RECURSIVE LEAST SQUARES

The presented algorithm is not free from the defect encountered in standard RLS, namely poor
numerical stability. In second-order RLS the inverse matrix updating is done through the
inversion lemma. Note that the inversion lemma is computed as the sum of two non-positive
definite matrices and hence the symmetric characteristic of the inverse matrix might be lost if
the numerical accuracy is not high enough. In our case the matrix P is updated using equation
(1 1 ) and symmetry, given by P,= (X,Z,Z,!‘XF)-’, is lost in many runs. Moreover, the matrix P
is the product of four matrices and this may cause the matrix to be ill-conditioned. This problem
is more evident here than for the second-order case where P is the product of two matrices.
Hence the global effect is the loss of positive definiteness of P and a rapid divergence. This
effect has also been experienced in the recursive version of the GM algorithm6 and a square root
algorithm was proposed in Reference 9 to ameliorate the above problems. This algorithm is
based on the guidelines found in References 10 and 18 and will be used here to derive a robust
square root version of the w-slice algorithm.
Square root methods depart from the update of the lower triangular square root of the
matrix,’ i.e. P,“’ instead of the matrix P, itself. This point is common to all square root
methods. However, the approach proposed here differs in that one of the auxiliary matrices
(namely A I c l in Table I) is indefinite, as was remarked in Reference 9. Because of this,
orthogonal transformations (such as Givens rotations) are not enough to update the matrices
and hyperbolic transforms are necessary. Let us clarify this point first. We can rewrite A , + l
as

where a = (z,”, ,z,+ Define the signature matrix of A,+,as being J and assume we can find a
matrix Q such that it fits the so-called J-orthogonal property
QJN+,Q” = JNc2 (19)
with the additional feature of being the triangularizing operator of the matrix F, i.e.

where L, and L2 are lower triangular matrices of dimensions 2 x and N x N respectively and J N
is defined as

1 OT
222 J. VIDAL AND J. A. R. FONOLLOSA

Then, by using (19), and right multiplying both sides of equation (20) by JN+2LH,
we car1
rewrite
FJN+,FH = LJN+&"
This trick can be done because the matrix Q is always invertible by the QR decomposition
theorem." Hence, by substituting terms in (20) and equating the two sides, we have
MAH
+ $Y+iP,$t+i= A2At+i+ $Y+iP,$t+i=LiJ&'i' (21)
P,$t+i =MJzL': (22)
P, = MJ2MH+ L,L; (23)
The decomposition in equation (21) always exists because the signature of the matrix
M2AH + $ ~ + I P , $ ,is+ J2.
, This property is derived in Appendix A of Reference 9 and will not.
be reproduced here. By taking (22), we get
M = pt$,+i(L'i')-'J2
Substituting this in (23) and by similarity with the inversion lemma in equation (1 l),
L2Ly=P,-MJ2MH=P,-P,$,+l(L~)-iJ2L;'$Y+iP,
= Pr - Pr Q, + i (At + i + CP Y+ ,PtCP
t +1 - Y+ ip, = I. 2P,+ i (24)
Hence, since L, is lower triangular, it is proportional to the square root of P,, i.e.
P;$ =A -' L2
Therefore the process of updating the square root matrix departs from the triangularization of
F. We have assumed the existence of the matrix Q. Let us see how it may be computed.
Considering that the matrix P,",',is lower triangular, F has the following structure of zeros:
[x 0 x x x x1
x x x x x x
A Q,,",lpj'2 0 0
F=[o pi', 0 0 1- o
lo
o
0
x
x
x
x
0
x
x
x
0
0
x
x
0
0
o
XJ

where the crosses denote possibly non-zero terms. The entries in the first row can be zeroed by
using successive hyperbolic rotations." A rotation H,,, puts to zero one term of the mth row by

-:][
pivoting it with the diagonal entry, in the following way:

[
FH,, = ... y;- ... iy][
*
"'

... -s
c '**

... C'
= ... r ... O]

The matrix H, is constructed in such a way that it is the identity, except for the entries c and s,
which are such that
IcI?-[sIZ= 1 (26a)
* *
C Ynm = S Ymm (27a)
Note that there is no solution to this equation if ynm= yrnm# 0. If they are different, it is possible
to compute the transformation as in Table 11.
 ADAPTIVE BLIND SYSTEM IDENTIFICATION 223

Table II. Hyperbolic rotation

Such rotations have to be applied to cancel the non-zero elements of the first row in equation
(25). Remember the J-orthogonality condition imposed to the triangularizing operator in
equation (19). It is easy to show that H,,,is a J-orthogonal matrix, but only in the case where
lynml< ly,,\. Suppose that in a particular iteration one of the elements of the first row of the
matrix F is greater in absolute value than the first element. In this case the rotation matrix is not
J-orthogonal. Then the whole triangularization process is achieved but equation (21) will not
hold, which is in contradiction with the fact that AJ2AH+ #r+,P,#,+, has the signature J,.
The rotations for the second row must be orthogonal so that the complete triangularizing
matrix is J-orthogonal according to equation (19). Orthogonal (also called Givens or Jacobi)
rotations have the identity matrix as signature. They are defined as

The matrix G,, is constructed in such a way that it is the identity, except for the entries c and s,
which are such that
Is/’= 1 (26b)
* *
C Y n m = -3 Y m m (27b)
The Givens rotation can be computed efficiently according to Table 111. Note that this rotation is
more robust than the hyperbolic rotation, since it always exists, no matter what the values of Y , , ~
and y,, are.
The complete triangularization of the matrix F is achieved by the algorithm of table IV. It is
given in MATLAB syntax. The functions Givens-rot(.) and Hyperbolic rot(.) are functions
which return a rotated matrix given the matrix F itself and the coordinates of the entry to be
zeroed. The final third-order URLS w-slice algorithm written in its square root version, is given

Table 111. Givens rotation

~~

if y,,, = 0
s=o; c = l ;
else
rl = abs(y,,,,); r2 = Wy,,,)/r,;
r = l/sqrt(rl(l + r;));
c = -r * conj(y,,,,); s = r conj(y,,,);
end
224 J. VIDAL AND J. A. R.FONOLLOSA

Table IV. J-orthogonal triangularization algorithm

function F = J-ort(F)
for ii = N: - 1:3
F = Hyperbolic-rot(F, 1 , ii);
end
forii=N:-1:3
F = Givens-rot(F, 2 , ii);
end
F = Givens-rot(F, 2 , 1 ) ;
end

in Table V. The function J-ort(.) performs the triangularization algorithm of Table IV. Its
extension to the fourth-order case is straightforward. The computational cost does not increase
much. For the third-order case using 2 p + 1 slices, the number of floating point products is
(23/2)p2+ ( 8 3 / 2 ) p+ 16.
Table V. Summary of square root third-order URLS w-slice algorithm

Initialization
so = P [I,,+1 I01
1
P0 -- - I * p+l
P

Multiplications

N
A = [ -A/a
a A/a'1 2x2

Q = J-ort(R) = L '
0
[
M API+l ] lop + 5
 ADAPTIVE BLIND SYSTEM IDENTIFICATION 225

4. ADAPTIVE EQUALIZATION
The w-slice approach is basically an impulse response estimator. Thus it is necessary to
construct an equalizer filter from knowledge of the channel. The most frequently used equalizer
is the transversal linear filter. Its popularity is indebted to the ease of analysis allowed by the
FIR structure. Once the channel impulse response or the poles and zeros are known, the
equalizer taps are computed depending on a certain criterion. This criterion allows the
approximation of the non-causal impulse response of the inverse filter by an FIR structure. The
most meaningful criterion is the minimization of the average probability of error of the
received signal. However, this criterion is a highly non-linear function of the taps c ( i ) and
becomes impractical. Other criteria have been used, namely zero forcing (ZF) and minimum
square error (MSE). The MSE criterion concerns the noise (often assumed to be Gaussian)
introduced in the signal during the transmission as well as the IS1 introduced by the channel.
This will be the criterion used to construct the equalizer once we have a good estimate of the
channel impulse response. Both algorithms render the same equalization filter when there is no
noise present. If this is the case, complete reduction of the IS1 may be achieved theoretically
with an infinite taps equalizer. Otherwise there is some residual both IS1 and noise at the output.
For its later use the expression of the MSE equalizer is given by
Av(z-')
C(z) =
Ayz)A*(z-')+ p
where p is the ratio between the power of the noise at the output of the channel and the power
of the transmitted sequence and H ( z ) is the estimated response of the channel. We can also
express this equation in terms of samples of the estimated impulse response. Let c denote the
2K + 1 impulse response of the equalizer corresponding to the estimated channel ho. Let e be a
column vector of dimension 2K + q + 1 with a one in the (K + q , + 1)th position and zeros
elsewhere, where q is the length of the channel impulse response and 4, is the position where
we assume the maximum value of the equalizer impulse response will be placed. Let H be the
( 2 K + q + 1) x ( 2 K + 1) matrix whose (i,j)th entry is & j - j - c , , . Therefore the taps of the equalizer
are given by 2o
c = (PI i- AHA)-'HHe (28)
Note that the parameter 4 , permits the distribution of the total length of the equalizer between
the anticausal part (those taps ranging from 0 to q , )and the causal part (those taps ranging from
q , + 2 to K). Depending on the channel characteristics, the position of this parameter may be
important.
The inherent drawback of this equalizer is its bad performance when the channel presents
zeros very close to the unit circle. In this case the number of taps needed to achieve a desired
amount of IS1 is extremely large. Then one must resort to Viterbi techniques.2' On the other
hand, Viterbi techniques have large computing requirements, especially when the impulse
response of the channel is large and the constellation has a lot of states, and hence MSE
equalizers represent a good trade-off in many practical cases.
The purpose of this section is to highlight the usefulness of the adaptive w-slice method
presented in the previous sections when applied to the case of noisy digital communication
signals. Two families of modulations will be used amplitude modulation (PAM) and
amplitude-phase modulation (QAM) signals. Symmetric signals have zero third-order
cumulant and hence the fourth-order cumulant sequence has to be used. Some remarks are in
order.
226 J. VIDAL AND J. A. R. FONOLLOSA

Remark 1
The definition of the fourth-order cumulant is ambiguous when dealing with complex (QAM,
PSK) signals. The ambiguity refer to the position and number of conjugations in the arguments.
Fortunately, the number of possibilities is reduced when considering the kind of signals we are
dealing with and focusing our attention on the two-dimensional zero-lag slice. It is easy to show
that the fourth-order cumulant of PSK and QAM signals is zero if there is an odd number of
conjugations in the arguments. Consequently, we can either conjugate none or two of the
arguments, resulting in four different cases:
CJ.,v(O,m , n) = E( x,x,x,+mx,+n I - E ( x,x,I EI x r + m x r + m I - 2E( x,xt+mIE(x,x,+nI (29a)
C42.x(O,m,n)=EIx~:x:+mx,+nI- EIx~:IE(~:+rnixr+nI
-~(~tx:+mI~(x:x,+nI

- EIx,xr+nIE( x:x:+m I (29b)

Ci,.k(O* 111, n) = E( ~ , ~ : ~ , + m x * t + -n IEI x,x:IEIx:+nxr+mI - E(x,x,+mIE(x:x:+nI
- EI xrx:+nI E ( x:x,+m I (29c)
C,4,.x(O,m , n = EI x,x,x:+mx: + n I - EI xrx, I E ( x: + n x : +m I - 2E( x t x : + m I EI x,x: + n I (29d)

Remark 2
Among the four different definitions, only equations (29a) and (29b) can be adapted to
the recursive framework presented above. Either of them can be employed or even both, by
using the first columns of the matrix S (equations (1) and (2)) for the definition (29a) and
the last columns for the definition (29b). Anyway, it seems more advisable to consider only
definition (29b). Conjugation in the arguments provides additional robustness to phase
jitter.

Remark 3
It should be noted that for L2-QAM signalling the correlation terms E ( x , ~ , , , ~are
) all zero and
hence
Cl..,(O,m , n)= Ml,.,(O, m, ti) = E { ~ , x , x , + , , , x , + ~ I
However, unlike the fourth-order cumulant, the fourth-order moment of Gaussian noise is not
zero unless it is circular. If this is not the case, cumulants have to be computed to obtain
effective noise robustness.

Remark 4
The w-slice algorithm forces the zero tap of the estimated impulse response to be unity
regardless of its true value, which is usually called the carrier phase. This results in a phase
ambiguity in the restored constellation, i.e. it may be rotated depending on the value of h(0).
Fortunately, h(O), and hence the true constellation phase, can be estimated using the following
method, already sketched in Reference 16.
Note that the fourth-order cumulant of the received sequence has been estimated and is
available in the entries of the matrices L and S in Tables I and V. In particular, the zero lag can
 ADAPTIVE BLIND SYSTEM IDENTIFICATION 227

be used since it fits the relation

ci.y(o, r2) = 74,s

r13 h(n)2h(n+- t l ) h ( n -I-7 2 ) = y4..vM:.h(O, TI, rz) (30)
nb 0

where y4,s= C:.,(O, 0,O) is the fourth-order cumulant of the transmitted sequence and is fixed
because the transmitted constellation is known. Since we have an estimation of Ci,,(O, r l ,t 2 )
from the entries of L and S , h (0) can be obtained as

We could even think of averaging different lags to get a more consistent estimation of the first
sample of the impulse response. In practice we have used all the estimated samples of the zero
slice of the fourth-order cumulant obtained from the matrix S for both the QAM and PAM
signalling

I
We remark that the sign cannot be identified. For the case of PAM signals h(0) can also be
identified, with an ambiguity of sign, using second-order statistics in the following way.'6
Second-order statistics satisfy the relation

c:,w =K..U c
nb 0
h ( n ) h ( n+ t) = Y2..rM:,h(r) (33)

Hence an expression similar to equation (31) can be derived

Once an estimation of h(0) is available, all estimated samples of the estimated impulse
response must be multiplied by h"(0). Equations (30)-(34) can be easily derived for the
cumulant definition of equation (29b). Although there is still an ambiguity of sign for the
estimated carrier phase, usually data transmission is done in a differential way. The final square
root fourth-order conjugate algorithm is given in Table VI.

Remark 5
Note that the number of operations needed per iteration is proportional to O ( y 2 ) ,while other
cumulant-based algorithms (TEA, POTEA, etc.)22 need a number of multiplications
proportional to 0 ( p 3 ) ,where p is the number of impulse response samples.

Remark 6
Since the algorithm is exact, there is no step size parameter affecting the convergence or the
stability of the method unless the channel is time-varying, in which case 1 has to be set to a
228 J. VIDAL AND J. A. R. FONOLLOSA

Table VI. Summary of square root fourth-order URLS w-slice algorithm using definition (29b)
Initialization
So=P"I,+,IOl

Matrix size Multiplications

Nxl
(Y + 1 ) x 1
(P+l)Xl
Nxl N+1
N x l Ni-1

N+l

2x2

( p+ 1) x ( p+ 1) lop + 5

(P+l)X1 p 2 / 2 + ( 2 N + 5)p + 2N + 1

At convergence
hO = Est-hO(L, h')
h r + =hO.h:+,
,
 ADAPTIVE BLIND SYSTEM IDENTIFICATION 229

value close to unity. This is not the case in Bussgang-type methods, where a bad choice of the
step size may lead to divergence of the algorithm and some type of trade-off between speed of
convergence and smoothness of the steady state.

5. SIMULATIONS
In the first set of simulations the performance of the ORIV-GM and URIV w-slice algorithms
has been tested in computer simulations for MA and ARMA processes. Some particular cases
are shown below. In all of them we have considered the same number of third-order cumulant
slices. Zero-mean, unit-variance, exponentially distributed random noise is used as the driving
process. In all cases we have initialized the algorithm with p = 0.0001. Time histories are
compared in all cases with the adaptive GM method of Reference 6. Since the square root
versions of the algorithms are more robust, they have been applied throughout. Figures show
the time evolution of the estimated coefficients averaged after 100 Monte Car10 runs. Means and
means plus/minus standard deviations are displayed and numerical values for Examples 1 and 2
at convergence can be found in Table VII.

Example 1
Processes have been generated using two different MA( 1) systems, namely maximum phase,
M A ( l ) = 1 - 1.25z-’, and minimum phase, M A ( l ) = 1 -0.8z-’, systems. Results are shown in
Figures l(a) and 1(b) respectively. Note that the proposed URLS w-slice approach is much less
sensitive to the character of the system. The maximum phase system has also been estimated
under white Gaussian noise condition of 5 dB SNR at the output process. Figure I(c) shows
that the ORIV method exhibits a more ‘regularized’ performance. However, the statistical
behaviour of the U R N w-slice method is still better.

Example 2
We examine here the tracking capabilities of both methods. The MA(2) linearly time-varying
model appearing in Reference 6 has been used: the initial model is MA(2) = 1 - z - ’ + 0-75z-*
and the final one (after 2000 samples) is MA(2) = 1 - 1.5z-I + 1.25z-’. The forgetting factor r3.
has been set this time to 0-9975. Figure 2 shows the results. Although both methods exhibit good
tracking properties, note the faster initial convergence of the w-slice approach. Note also that
the residual error in the estimated coefficients does not fall to zero because of the stochastic
behaviour caused by 2.

Table VII. Means plus/minus standard deviations of estimated parameters in

Examples 1 and 2 at convergence

True parameter W-slice ORIV

~

Example l(a) -1.25 - 1.239 It 0.087 -1.264 k0.154

Example l(b) -1.80 - 1,793 f 0.056 - 1.794 f0.082
Example 1(c) -1.25 - 1-228rt 0.114 *
- 1.236 0.169
Example 2 -1.50 - 1.424 f0.130 -1.419 f0.129
1a25 - 1.195 f 0.142 1.147 f 0 . 136
230 J. VIDAL AND J. A. R. FONOLLOSA

W-slice ORN

r I
1 200 400 1 200 400
Figure I(a). Time evolution of estimated parameter for system MA(1) = 1 - 1 . 2 5 ~ '

W-slice ORIV
4 4

2 2

b(1) 0 MI) 0

-2 -2

4 II-4 .-

W-slice O W

-4
I 200 400 600 Roo loo0 1 200 400 600 800 lo00
Figure l(c). Time evolution of estimated parameters for system MA(1) = 1 - 1.25:-' corrupted with white Gaussian
noise up to 5 dB SNR

Example 3
The third-order algorithm with a narrow band ARMA system is considered:
ARMA(3,1)= (1 - 1-5.~-')/(1 - 1.52232-l +0*77292-*- 0 * 1 3 5 0 ~ - ~ )
Results are shown in Figure 3. Note that there is an allpass factor present and hence second-
order approaches to estimating the AR coefficients do not work at all.
 ADAPTIVE BLIND SYSTEM IDENTIFICATION 23 1

W-slice ORlV
21 I

-2 J I
1 400 800 1200 1600 2000

W-slice ORN

-2
1 400 wx) 1200 1600 zoo0 1 400 800 1200 1600 2000
Figure 2. Time tracking of estimated parameters of linearly time-varying MA(2) system

W-slice W-slice

W-slice

-41 I
I 500 loo0 IS00 m 2500 3Ooo
Figure 3. Time evolution of estimated impulse response of narrowband ARMA(3.1) system
232 J. VIDAL AND J. A. R. FONOLLOSA

Example 4
The second set of simulations shown below demonstrates the performance of the recursive
w-slice algorithm in combination with the MSE linear equalizer. We assume no timing error at
the sampler and one sample per symbol to fit the white condition of the driving process. All of
them have been computed for 15 signal- and noise-independent realizations. The noise is
Gaussian additive and white (complex for the QAM signals). In all cases the SNR is considered
at the input of the equalizer. Regarding the kind of signals, we have tested the performance
with 16-PAM and 16-QAM constellation signals on one real and two complex FIR channels
previously used in References 16 and 23. They all are shown in Figure 4. Note that channel 4
exhibits severe fading, which makes it more difficult to equalize.
Performance is assessed by using measures of MSE and the discrete eye patterns. The
probability of error at convergence is not displayed, since the method proposed is not intended
to reach low error rates. Rather, it is viewed as a fast convergence initial step to a Bussgang
equalizer. The estimated impulse response has been averaged over 50 iterations to alleviate the
inherent variance of the cumulant estimates. The eye patterns have been built by using 1501
samples per realization, taken at the output of the equalizer, when the w-slice algorithm

- 1 + 0.6562-l)
Channel 1 : (1 - 0.18332)(1 0.1987~-~)(

1+ (0.158+j0.426)~-~]
ChanneI 2: (0.944-j0.87)[1- (0.0787+jO.O768)z][1 - (0.188+j0.075)~-~][
lo I

Rul partof impubc nspan& Imaginary part of impullu. mponsc Fnqucocy respoosc

Channel 3: (2-jO.4)Z2 + (1.5+J1.8)z + I + (1.2-j1.3)~-1+ (O.8+1.6j)r2

Real part of impults r e s p n %

-20 1
Imaginary IWIof impuls wpaw
.I5 - Frequency reiponsu

Figure 4. Time and frequency characteristics of real and complex channels

 ADAPTIVE BLIND SYSTEM IDENTIFICATION 233

reaches the steady state. Performance comparisons are laid out with respect to the popular
Bussgang-type Benveniste-Goursat (BG) method developed in Reference 24.
From observation of the results in Figures 4-9 we can conclude the following.
1. For the case of complex constellations we have found that the use of the complex
definition of the cumulant (equation (29b)) outperforms the non-conjugate definition
(equation (29a)).
2. The URLS w-slice algorithm exhibits a larger residual MSE than the BG methods, as
expected. It is well known that the variance of the estimated cumulants impairs the
variance of the estimates and that Bussgang techniques allow lower MSE residual errors.
3. The performance of both methods degrades as the channel presents a larger spectral
dynamic range. Note the worst performance of both methods for channel 5. The behaviour
is also quite dependent on the noise level.
4. The choice of parameters in the BG method is very critical and depends very much on the
type of constellation, the channel characteristic and the SNR. However, the MSE tracings
present a wilder behaviour for the URLS w-slice in comparison with the BG ones.

Unqualizcd eye URLS W-slice

3.0

2.0
1.5

1 .o
'
1.o 0.5

0.0 0.0

-1.0 -0.5

-2.0 -1.0

-3.0 4 i -1.5 I I
0 30 60 90 120 150 0 30 60 90 120 150
Samples Samples

RenveniaeGoursat

1.5 ,- I

-1.5-L
0 30 60 90 120 150
Samplcv

Figure 5. Eye patterns obtained for 16-PAM constellation perturbed by channel 1 under 40 dB SNR. Equalization
using URLS w-slice and BG algorithms after 6000 iterations
 URLS W -slice Channel 1 Beovenisto-Gouraat Channel 1

0
' O
-5

-10

;g -IS

1
-20

-25 -25
-20

-30 -30
0 4000 8000 12000 0 4000 8000 I2000
Iterations Iterations

Figure 6. Mean square error convergence versus number of iterations for N = 31 taps equalizer using channel 1 with
SNR = 40 dB for 16-PAM signalling. URLS w-slice algorithm and BG algorithm with parameters y = 5 x k, = 5 ,
k, = 3

unequalized eye

3.0 I .5
I
-->- -
-
2.0 1.o

1.o 0.5

0.0 0.0

-1.0 -0.5

-2.0 -1.0

-3.0 . ----I -1.5

-3.0 -2.0 -1.0 0.0 1.0 2.0 3.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5

BcnvcnisteGouruat

-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5

Figure 7. Eye patterns obtained for 16-QAM constellation perturbed by channel 3 under 30 dB SNR. Equalization
using URLS w-slice and BG algorithms after 15 000 iterations
 ADAPTIVE BLIND SYSTEM IDENTIFICATION 235

Unqualized eye URIS W - d i e

-10.0 4 -1.5 4
I
-10.0 -5.0 0.0 5.0 10.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5

-1.5 -, - -- : .--
- -
-a
I --- ,
-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5

Figure 8. Eye patterns obtained for 16-QAM constellation perturbed by channel 3 under 40 dB SNR. Equalization
using URLS w-slice and BG algorithms after 20 OOO iterations

5 . The speed of convergence of the URLS w-slice is faster than that of the BG method.
Results suggest the use of the URLS w-slice as an initializing device for the BG
algorithm. More extensive simulations demonstrate that the convergence speed to the
minimum mean square error could be improved by a factor of 10 or more depending on
the channel characteristics and the constellation.

6. CONCLUSIONS
The adaptive version of the w-slice algorithm has been shown to provide good behaviour in
front of the adaptive GM method using the ORIV algorithm. The presented approach does
not suffer from the impairment of the GM, namely its poor robustness to noise. This
drawback comes from the use of second-order statistics, which are not used at all in the
construction of the weighted slice. Moreover, the ORIV algorithm shares the inconsistency
of the GM method when the true order is unknown. The URIV w-slice algorithm overcomes
these limitations. General enough, it can be applied to identify the impulse response of any
236 J. VIDAL AND J. A. R. FONOLLOSA

URLS w-slice Channel 2 Benvenistacoursai Channel 2

-5 I._.
01

0 4OOO 8000 12000 16000 3,0000

-25 I-..,.
0 jooo so00
---,---
12000 IGOOO 2oooo
ltaations lleratiow:

URLS w-slim Channel 3 BenvenisteGoursnt Channel 3

5 t----- v

MSE -5
(dB)

-10 -10 1
-15 -1s I i
0 SO00 10000 15000 20000 25000 0 so00 loo00 15ooo 20000 '5000
Itmations Iterations
Figure 9. Mean square error convergence versus number of iterations for N = 31 (channel 2) and 61 (channel 3) taps
equalizers with SNR = 30 and 40 dB respectively; 16-QAM constellation. URLS w-slice al orithrn and BG algorithm
with parameters y = 2 x k , = 3 , k2 = 1 for channel 2 and y = 8 x k , = 3 , k g = 1 for channel 3

linear system. The order of the statistics (provided that it is greater than two) to be used is
not constrained. Moreover, different cumulant orders may be used in the same framework.
The number of slices is also a design parameter. Finally, the algorithm has been applied to
the equalization of digital communication channels with remarkable results. However, the
authors are aware of the wider applicability of fractionally spaced equalizers and further
work is aimed to this goal.

ACKNOWLEDGEMENTS

This work has been supported by the Spanish Government under grant TIC92-0800-C05-04.
The authors would like to thank the reviewers for their careful reading and valuable
suggestions.
 ADAPTIVE BLIND SYSTEM IDENTIFICATION 237
REFERENCES
1. Haykin, S . , Adaptive Filter Theory, 2nd edn, Prentice-Hall, Englewood Cliffs, NJ, 1990.
2. Fonollosa, J. A. R. and J. Vidal, ‘System identification using a linear combination of cumulant slices’, IEEE
Trans. Signal Process., SP-41, pp. 2405-2412 (1993).
3. Swami, A. and J. M. Mendel, ‘Adaptive system identification using cumulants’, Proc. IEEE ICASSP, Glasgow,
May 1989, IEEE, New York, 1989, pp. 2248-2251.
4. Swami, A. and J. M. Mendel, ‘AR identifiability using cumulants’, Proc. Workshop on Higher-Order Spectral
Analysis, Vail, CO, June 1989, pp. 13-18.
5 . Fonollosa, J. A. R., J. Vidal and E. Masgrau, ‘Adaptive system identification based on higher-order statistics’,
Proc. IEEE ICASSP, Toronto, Ont., May 1991, IEEE, New York, 1991, pp. 3437-3440.
6. Friedlander, B. and B. Porat, ‘Adaptive IIR algorithms based on higher-order statistics’, IEEE Trans. Acoust.,
Speech, Signal Process., ASSP-37,485-495 (1989).
7. Friedlander, B., ‘The overdetermined recursive instrumental variable method’, IEEE Trans. Automatic Control,
AC-29, pp. 353-356 (1984).
8. Giannakis, G. B. and J. M. Mendel, ‘Identification of non-minimum phase systems using higher-order statistics’,
IEEE Trans. Acoust., Speech, Signal Process., ASSP-37, 360-377 (1987).
9. Porat, B. and B. Friedlander, ‘The square-root overdetermined recursive instrumental variable algorithm’, IEEE
Trans. Automatic Control, AC-34, pp. 656-658 (1989).
10. Bierman, G. J., Factori-ation Methods for Discrete Sequential Estimation, Academic, New York, 1977.
11. Giannakis, G. B., ‘On the identifiability of non-Gaussian ARMA models using cumulants’, IEEE Trans.
Automatic Control, AC-35, (1990).
12. Vidal, J. and J. A. R. Fonollosa, ‘Causal AR modeling using a linear combination of cumulant slices’, in Signal
Process., 36, pp. 329-340 (1994).
13. Giannakis, G. B. and A. Swami, ‘Identifiability of general ARMA processes using linear cumulant-based
estimators’, Automatica, 28, 771-779 (1992).
14. Vidal, J., ‘Non-Gaussian ARMA estimation using higher-order statistics’ Ph.D. Dissertation, Universitat
Polittcnica de Catalunya, Barcelona, 1993.
15. Vidal, J., J. A. R. Fonollosa and A. Moreno, ‘Onthe use of the weighted cumulant slices in the multichannel MA
modeling problem’, Proc. Int. Conf. on Signal Processing, Beijing, International Academic Publishers, October
1993.
16. Hatzinakos, D. and C. L. Nikias, ‘Blind equalization using a tricepstrum-based algorithm’, IEEE Trans. Conzrnun.,
COM-39,669-682 (1991).
17. Ljung, L. and T. Soderstrom, Theory and Practice of Recursive Identification, MIT Press, Cambridge, MA, 1983.
18. Morf, M. and T. Kailath, ‘Square-root algorithms for least-square estimation’, IEEE Trans. Automatic Control,
AC-20,487-497 (1975).
19. Golub, G. H. and C. F. Van Loan, Matrix Computations, Johns Hopkins University Press, Baltimore, MD, 1983.
20. Clark, A. P. Equa1i;ers for Digital Modems, Pentech, London, 1985.
21. Proakis, J. G., Digital Communications, McGraw-Hill, Singapore, 1983.
22. Brooks, D. H. and C. L. Nikias, ’Cross-bicepstrum and cross-tricepstrum approaches to multichannel
deconvolution’, Proc. Int. Signal Processing Workshop on HOS,Chamrousse, July 1991, pp. 91-94.
23. Porat, B. and B. Friedlander, ‘Blind equalization of digital communication channels using higher-order moments’,
IEEE Trans. Signal Process., SP-39, 522-526 (1991).
24. Benveniste,A. and M. Goursat, ‘Blind equalizers’, IEEE Trans. Commun., COM-32, 1867-1875 (1990).
25. Friedlander, B. and B. Porat, ‘Asymptotically optimal estimation of MA and ARMA parameters of non-Gaussian
processes from high-order moments’, IEEE Trans. Automatic Control, AC-37, 27-35 (1990).
26. Porat, B. and B. Friedlander, ‘Performance analysis of parameter estimation algorithms based on higher-order
moments’, Int. j . adapt. control signal process., 3, 191-229 (1989).