Blind adaptive equalizers for quadrature amplitude modulated communication systems based on convex cost functions

Australian National University Department of Systems Engineering GPO Box 4, Canberra, ACT 2601 Australia

Rodney A. Kennedy

Abstract. Blind adaptive channel equalizers are important devices to

remove channel distortion in high data-rate, bandlimited digital communication systems when the transmission of a training sequence is impractical or very costly. Traditional blind equalization algorithms adapt the equalizer parameters to minimize some specially designed non-MSE cost functions. These algorithms can experience local convergence prob-

Zhi Ding
Auburn University Department of Electrical Engineering 200 Broun Hall Auburn, Alabama 36849-5201

lems and can thereby result in insufficient or no removal of channel

distortion. We present a new quadrature amplitude modulated blind equal-

ization scheme that is globally convergent in the equalizer parameter

space to a compact set containing the desired ideal equalizer parameter

setting. Our new algorithm is based on a convex cost function and a linear constraint on the equalizer parameters. For a generic class of
channels, this new algorithm results in the equalizer parameter convergence to a unique global minimum achieving intersymbol interference suppression and carrier phase error removal. Different implementation approaches are assessed and simulation results are shown to confirm the theoretical global convergence of the new algorithm.
Subject terms: adaptive signal processing; intersymbol interference; equalizers; bllnd equalization; blind deconvolution; global convergence.

Optical Engineering

31(6), 1 189 1 199 (June 1992).

1 Introduction
Blind equalization of an unknown, nonminimum phase channel is an important problem in data communication systems. Conventional adaptive equalizers remove the in-

of some

special non-MSE cost functions that do not require 16 the use of a reference They are sometimes known

as the Bussgang algorithms.7 However, the lack of uni-

tersymbol interference (ISI) caused by nonideal channels of limited bandwidth with the aid of a training session, which provides the needed reference signal for parameter adaptation. In situations where a training session is impossible or very costly, blind equalizers are required to combat the ISI effect. A blind equalizer achieves parameter adaptation based on observation of the channel output and prior knowledge of some distributional or statistical properties of the input 14 sequence. By eliminating the need for training signals, the receiver can begin its self-adaptation without disrupting

modality of these cost functions , established in recent studies,814 can result in the undesirable (local) convergence of

the corresponding algorithms such that the equalizer fails to remove sufficient 151. While further studies and experiments are needed to determine the seriousness of the local convergence problem for actual communication channels, preliminary real-time illustration of local convergence by the constant modulus algorithms (CMA) has been presented in Ref. 15 . Hence, despite the apparent plethora of schemes, there remains a strong practical need for development of new algorithms with more reliable and provable (global) convergence properties.

the normal flow of data transmission. It can also recover from a system failure during which the equalizer may have lost track of the desired parameter settings. In this paper,
we focus on the problem of blind linear equalization, which, in essence, consists of using an adaptive inverse linear filter

From recent works'214 it is clear that to design blind

to cancel the channel ISI provided that the frequency response of the channel does not possess any spectral nulls.3 Many schemes for blind equalization of a quadrature
amplitude modulation (QAM) system exist in the literature. Typically, they are based on gradient descent minimization
Paper SP-005 received July 31, 1991; accepted for publication Mar. 9, 1992. 1992 Society of Photo-Optical Instrumentation Engineers. 0091-32861921$2.OO.

equalization algorithms that converge to a desirable parameter setting from arbitrary initializations (based on gradient descent minimization of some cost function), certain aspects of the performance objective should be relaxed. In designing our new algorithm, we did not attempt to identify directly the exact gain of the channel inverse and focus instead on the elimination of 151. 16 Furthermore, once the 151 is removed such that the equalizer output is a scaled version of the channel input, it is then straightforward to estimate in-

dependently the unknown scalar gain by, for example,

also be readily resolved by the utilization of differential
OPTICAL ENGINEERING / June 1992 / Vol. 31 No. 6 / 1189

matching the power of the channel input and equalizer output. Any additional constant complex phase ambiguity can

KENNEDY and DING

RECEIVER

Channel

Input
Channel

Channe Output
k
I

Receiver

Output
A

Equalizer

Decision

ak

e(j')
1

Device

_::[

Q(.)

Fig. 1 QAM channel equalization systems.

encoding if it is a multiple of 'rr/2 . Thus , as long as we can remove the phase error in carrier recovery, ISI elimination

without gain identification is a sufficient objective for the ultimate recovery of the channel input in QAM systems. Blind equalization of pulse amplitude modulation (PAM) systems without gain recovery has been proposed in Refs. 17 through 2 1 . In these works , convex cost functions are employed; then, to prevent the convergence to the trivial all-zero parameter setting, tap anchoring is used. However, generalizing these results directly to QAM systems is not straightforward. The difficulty arises because with QAM systems, both the complex system parameters and signals possess magnitude and phase. However, whereas the system parameters can take arbitrary amplitudes and phases, the
same does not apply to the complex data that are constrained to take values in the QAM constellation. This feature, particularly with respect to the phase, indicates a nontrivial departure from the real PAM case (which exploits the property that all the phases are effectively only 0 or 'rr). Phase also introduces a second important issue. A simple constraint such as fixing the center tap O to a real or complex constant (as would be suggested by a translation of the PAM results) does not ensure the recovery of the constellation in the equalizer output because of the unknown carrier phase errorthis problem is known as carrier phase recovery. Ideally the adaptation scheme should provide both equalization (removal of ISI except for a gain factor) and carrier phase recovery (to rotate the output constellation to the
correct orientation before quantization) simultaneously. Here,

plicitly rely on higher order moments, cumulant-based approaches largely seek explicit estimation of higher order cumulants at the expense of computational and algorithmic complexity.22 With the constant improvements in compu-

cumulant-based techniques are competitive with the technique we are presenting. The main part of this paper is divided into the following sections: In Sec. 2, we describe the fundamental principles of linear blind equalization. In Sec. 3, we present the new blind adaptive equalization algorithm by introducing a new cost function with a linear constraint of the complex equalizer parameters . We describe different implementation approaches of the new blind equalization algorithm in Sec. 4, where the simulation success of these implementations is established. Global convergence results are relegated to Sec. 6, Appendix.

tational complexity reflected in, say, Ref. 22, these

2 Fundamentals of QAM Blind Equalization

2.1 System Description Figure 1 shows a diagram of baseband representations of
the QAM channel equalization system. The original symbol

sequence {ak} takes values from a complex signal set s4 (known as the QAM constellation) in a manner such that all possible finite symbol subsequences occur with nonzero
probability. This statistical assumption on the input is weaker than the typical somewhat idealistic independent identically

we achieve this desirable objective of joint blind equalization and carrier phase recovery using a novel nonanchoring tap constraint. Unlike many blind equalization proposals, we furnish proof that the algorithm we are developing possesses desirable unique parameter convergence for ge-

distributed (i.i.d.) assumption in other works and it does not exclude the possible source coding. (Further weak assumptions regarding various symmetries of the constellation

will be described later.) The complex channel input {ak} is transmitted through a
nonideal channel, assumed to be linear, causal, and (boundedinput-bounded-output) stable. The channel transfer function

neric channels even when finite impulse response (FIR) equalizers that can only approximate the desired channel inverse are employed. The potential for nonuniqueness of the global minimum (where the minimum cost is obtained by all points in a compact set rather than at a single point), which may occur for a class of nongeneric channels, is examined and shown to be intrinsic to the particular linear constraint and the problem formulation. Finally, we mention that blind deconvolution methods
based on higher order moments and cumulants are becoming

can be written as

H(z1)=

hz' ,

hEC

(1)

where {h1} represents the channel impulse response sequence. When the channel [Eq. (1)] is such that there is
more than one nonzero element in {h1}, the channel output,

increasingly popular; see Refs. 22 and 23 and references therein. These techniques form a natural complement to Bussgang approaches. Whereas Bussgang algorithms im1190 / OPTICAL ENGINEERING / June 1992 / Vol. 31 No. 6

xk=H(z)ak= hak i=0

BLIND ADAPTIVE EQUALIZERS FOR clAM COMMUNICATION SYSTEMS

is said to suffer from IS! and the removal of this distortion is the process of equalization. A linear channel equalizer is a linear filter O(z 1) that is applied to the channel output Xk to eliminate the IS!, as shown in Fig. 1. Initially we may consider this filter stable and potentially noncausal (doubly infinite), and of the form

sically subjected to a phase ambiguity of mrr/2 and the best possible result would be

H(z 1)O(z 1) = exp(mir/2)z

mE{O, 1 ,2,3}

(6)

which relaxes the objective [Eq. (5)1 . The remaining phase ambiguity can be resolved through differential encoding of

O(k)z ,

the input data. A further relaxation of the ideal objective

O1EC

(2)

[Eq. (5)] arises because in practical applications only equalizers with a finite number of adjustable parameters can be

so as to deal with nonminimum phase channels. The time dependence of the equalizer parameters signifies that they are subject to adaptation via an algorithm to be described. The equalizer output can then be written as
Zk=

implemented rather than those of Eq. (2). Typically these finite parameters are arranged in the form of a causal transversa! filter,
N

Oj(k)xkj=X'kO(k)

(3)

O(z1)=z' 1= -N

(7)

by denoting the adaptive equalizer parameter vector as

O(k)={ 020_i 0001 02 "I

and the equalizer input signal regressor as
AkL" Xk+2Xk+iXkXkiXk2 ...j

(4)

Because an FIR equalizer can only approximate the desired impulse response [Eq. (6)1, a quantizer should be used to recover the original channel input from the equalizer output Zk. This is often regarded as the practical objective of blind equalization.

2.3 Gradient Descent Formulation

Blind adaptive equalization algorithms are often designed as stochastic gradient descent (SGD) schemes to update the parameter vector by minimizing some special (non-MSE) mean cost functions that do not involve the use of the origma! input ak, but still reflect the current level of IS! in the equalizer output. Such schemes implicitly involve higher order statistics of the channel and equalizer outputs and are conveniently characterized through the mean cost function,

The ideal objective of any channel equalizer O(z 1) is to achieve distortionless reception,

zk=O(z)xk=ak_ vEl ,
or simply

Vk
(5)

H(z1)O(z ')=z

J(0) E{I'(zk)} = E{'I'(X0}

(8)

where we can tolerate a finite fixed time delay v. (The strict

form of this objective [Eq. (5)] will be modified subsequently.) Such an objective translates into the identification of channel inverse by the equalizer, i.e.,

where I': C is a scalar cost function. The mean cost

function J(0) should be minimized when 0 is such that Eq. (6) holds. Such algorithms are sometimes referred to as the Bussgang algorithms.7 Figure 2 shows a typical QAM blind equalization system based on the minimization of non-MSE cost functions. Bussgang algorithms have been shown to exhibit local convergence.814 The lack of unimodality by these algorithms can result in local, undesirable convergence of the equalizer parameters especially for FIR equalizers. Hence, this motivates (1) the development of new and better blind equalization algorithms, (2) the need for theoretical demonstration of the convergence properties of any new proposed scheme, (3) the establishment that global minimization of IS! is a property of the new scheme under an ideal doubly infinite parameterization, and (4) the demonstration that the global convergence attributes are robust to finite dimensional equalizer parameterizations, e.g. , Eq. (7), as required in practice.

O(z_l)=z_viz(z_l)

where h(z 1) H 1(z ') is understood to be the stable,

(potentially) doubly infinite (noncausal) inverse of the chan-

nel H(z 1) Such a goal could be achieved by adapting the equalizer parameters to minimize the MSE between the channel input ak v and the equalizer output, when the original channel input signal is available to the receiver as a training sequence.

2.2 Blind Equalization in QAM Systems In blind equalization, the original sequence is unknown to
the receiver except for its probabilistic or statistical properties over the known alphabet .si . Usually this signal constellation siCC has symmetrical properties such that
exp(jmir/2)sf=.sLi
i.e.
,

mE{O,l,2,3}

2.4 Blind Equalization without Gain Identification

To design globally convergent blind equalization algorithms based on gradient descent minimization of some cost function, further aspects of the performance objective [Eq. (5)1 should be relaxed beyond those already indicated, i.e. , the mrrI2 phase ambiguity [Eq. (6)1 and the finite dimensional equalizer implementation, meaning approximation of Eqs.
OPTICAL ENGINEERING / June 1 992 / Vol. 31 No. 6 / 1 191

tistics of the input data reflect this same symmetry (which is typical) over . , then an m'rr/2 phase rotation does not cause any statistical changes in the channel output. Thus, the data recovered from blind equalization will be intrin-

, the constellation is mir/2-rotation-invariant. If the sta-

KENNEDY and DING

Fig. 2 A Bussgang blind equalization system.

(6) and (7) requiring a quantizer. In designing our new

algorithm, we did not attempt to identify the exact gain of channel inverse and focused instead on the elimination of 1ST, which is the primary objective of channel equaliza16 Consequently, it is not essential to recover the exact (complex) gain of the channel inverse because once the ISI is removed such that the equalizer output is Zk cak = cexp(j4c)ak V
,
C

3 Globally Convergent Algorithms with Convex Cost Functions

3.1 Convex Cost Function and Parameter Constraint We assume the QAM constellation is such that the real and imaginary components are independent and identical (e.g., the square type) with
M maxlRe{ak}I = maxIm{ak}I
(or by rotational can be transformed to such). We can denote the total system (channel and equalizer) as

*0

then it is straightforward to estimate the unknown gain c by comparing the power of zk with that of the ak. The can also be readily resolved constant phase ambiguity by the utilization of differential encoding, provided it is a multiple of ir/2.
Blind equalization ofPAM systems without gain recovery has been proposed in Refs. 17 through 21 . Their essential

T(z1)=H(z1)O(z1)= i=ctizl
where {t1} is the impulse response of the combined system

idea is to fix the center tap O as a constant (unity in most cases) to prevent the local convergence of respective adaptation algorithms. As shown in Refs. 17 and 21 , because

given by

of the absence of carrier phase error in PAM systems, ' ' 'anchoring' 16 the center tap can beneficially constrain the freedom of equalizer parameters such that a convex cost
function can be utilized for parameter adaptation.

t=hO= k=O hkOk E

and denotes convolution. The equalizer is considered doubly infinite and noncausal to permit complete removal
of ISI. This assumption will be relaxed (as discussed above) to deal with the truly implementable finite equalizer parameter space and avoid the potential weaknesses of this simplification as discussed in Ref. 11. The cost function to be minimized is designed as

In Sec. 3, we present a new algorithm for joint blind

equalization and carrier phase recovery based on specialized convex cost functions coupled with a constrained equalizer parameterization as a generalization of the algorithm in Ref. 1 8 . For a general class of channels with (independent) real and imaginary distortion, we obtain a unimodal mean cost function with which gradient descent algorithms exhibit parameter convergence that globally minimizes the mean cost. This ideal convergence property, in contrast to the classical Bussgang approach, is robust to equalizer truncation [Eq. (7)1

J(O) maxIRe{zk}
= --max
M {ak}

because of the convexity of the cost functions employed. The robustness lies in that (1) spurious local minima are not created by the truncation and (2) the global minimum in the strictly doubly infinite parameter space [Eq. (6)] degrades gracefully into a global minimum in the practical finite dimensional parameter space [Eq. (7)1, where by selecting the "dimension" of the equalizer N in Eq. (7) appropriately, one can ensure an arbitrarily small residual 151 term. This latter property is a trivial consequence of convexity; see Ref. 16.
1192 / OPTICAL ENGINEERING / June 1992 / Vol. 31 No. 6

Re{tl}Re{ak_ } Im{t}Im{ak_ }

= i [teti + IIm{tiH] = jmaxIm{zk}I

where
Zk

(9)

Etak_1

BLIND ADAPTIVE EQUALIZERS FOR QAM COMMUNICATION SYSTEMS

The convexity of J(O) with respect to 0 follows from the triangle inequality

such that h(z

') exists on the unit circle. Then we have

J[XO1 + (1 X)O21

= I(\
I

(2) IRe{h[XO' + (1 X)O, 1}

hh = i(i)
where (i) is the Kronecker delta function.
Our theoretical demonstration of global convergence pro-

+ Im{h[XO1 + (1 X)O2I}f )

xLRh1hi0i }I
1

\ I

(1)

+ Im{hO1} I i
F

ceeds in a number of steps. First we present a lemma that identifies the specific parameter setting from the countable class of ideal equalizers, i.e. , ones leading to zero residual ISI, which minimizes the linearly constrained cost [Eqs. (9) and (1 1)]. In what follows, it is more convenient to express the constraint [Eq. (11)1 in polar form:

o$i) exp(j)
.
.

cos+sin
.

+(1X)I

L i

Re{hO2}

Armed with this candidate, the second step (our main theo-

v (2) 1 +2. Im{hO } I i j

I
1

rem) demonstrates that it minimizes globally the given cost function over arbitrary equalizer parameter settings that are . . subject to the linear constraint [Eq. (1 1)1.
Lemma 3. 1. For ideal equalizer parameter settings of the form

= XJ[OwI + (1 X)J[O2I, O<X< 1 .

(10)

Notice that by fixing the real or the imaginary part of the

parameter vector 0, the cost function J(O) =J(Re{O}, Im{O}) is also convex in terms of Re{O} and Im{O}, an observation

' q, O( ) 4 hi+q
hq

'

that will be important later. From

parameterized by qEl, and , the convex cost [Eq. (9)1

1

subject to the linear constraint [Eq. (1 1)1 is minimized by

J( 0) =

IRe{

t} + Im{ t}l ] = J(O)

VO*O
. .
.

the following (q,) pair:

m = arg max fRe{kq} + Im{Iq}
q
with kE{O,1,2,3} st. O<*<rr/2,

we have by convexity
J(O)O.5J(O)+(1O.5)J(O)=J(O) ,
.
. . .

Thus if the range of 0 is unrestricted then the convex cost

4*

.I
I._4

for m{O,ir/2,ir,3irI2}
for any 4E[O rr/21
otherwise

function [Eq. (10)] achieves a trivial global minimum at 0 = 0, with equalizer outputzk= 0.
.

(12)

To make the cost function J(O) useful, the equalizer parameter vector needs to be constrained from giving an allzero solution. But at the same time, the constraints should not damage the ability of the equalizer to achieve ISI and phase error removal. To satisfy both objectives, we constrain the equalizer parameters with the following linear constraint:
Re{Oo} + Im{Oo} 1 ,
(1 1)

where hm fhmIexp(jm).
Proof. See Sec. 6, Appendix. We defer comments regarding this lemma until after the next result. Now we present the main theorem.
Theorem 3. 1. For a linear channel with inverse impulse
response sequence {h}, a doubly infinite noncausal equalizer with parameter setting

in the equalizer parameter vector [Eq. (4)1. Due to the linearity of the constraint, the convexity of the cost function (with respect to the real and imaginary parts of the equalizer coefficients) is maintained and global convergence is therefore assured. We now show that under this parameterization, simultaneous removal of ISI and carrier phase error globally minimizes the cost function.

Oj(m,*)

i+m exp(j

ir/2)

Re{hm} + Im{hm}

(13)

subject to the constraint

Re{Oo} + Im{Oo} = 1

(14)

where
m = arg max Re{hqH + IIm{q

3.2 Global Convergence Suppose the ideal noncausal channel inverse (as before) is
given by

q(

h(z i) =H i(z_ i) =

i=

h1z

hEC

and where kE{1,2,3,4} is chosen such that Eq. (14) is satisfied, minimizes the blind convex cost
OPTICAL ENGINEERING / June 1 992 / Vol. 31 No. 6 / 1 193

KENNEDY and DING

maxlRe{zk}f =
{ak}

j=c L

I IRe{} + IIm{tjH 1 = maxfIm{zk}.

J
{ak}

imate arbitrarily closely the performance of the ideal nonimplementable doubly infinite noncausal equalizer

parameter setting. All this is important because

earlier13'24 we have shown a mechanism whereby global

Proof: See Sec. 6, Appendix.

3.3 Remarks and Comments

Before proceeding to implementation issues related to developing an algorithm based on the above convex function, we first examine the theoretical limitations and other considerations of the above scheme.
1.

convergence proofs for nonconvex cost function based algorithms, which are valid for doubly infinite (noncausal) equalizers , break down when finite truncated versions are attempted.
5. The minimum [Eq. (13)] as well as achieving zero

IS! also results in near-ideal phase recovery, i.e., except for an ambiguity of k'n12 in the phase, the
output of the equalizer is equal to the input since the channel/equalizer convolution is given by

Despite the unimodality of the cost function with respect to the equalizer parameters , a potential problem exists of nonuniqueness of the (global) minimum, i.e. , the minimum might be achieved by all points in a compact set (actually a convex polytope) rather than at a single point. There are two manifestations of this nonuniqueness: (1) as in Lemma 3. 1 there can be a phase ambiguity over a 'rrI2 range of hm 5 purely real or purely imaginary, which is a nongeneric property

hjOj(m,*) (i + m)

exp(jkr/2)
fRe{hm}I + Im{hm}I

Hence our formulation achieves simultaneous channel

equalization and carrier phase recovery. Note again that the gain factor can be corrected by an automatic
gain control, which scales the equalizer output to match

and (2) the nonuniqueness is manifested as in Eq.

(1 2) when the maximum is realized by more than one

m, which is again a nongeneric property. That these are the only ways nonuniqueness arises can be inferred from the proof of Theorem 3 . 1 in Sec. 6. 2. The potential for a minimizing set rather than a mmimizing point is (necessarily) traceable to the lack of strict convexity of the selected cost function. The question arises as to the appropriateness of the above selection of the convex cost function in the sense that perhaps other choices avoid this nongeneric problem. In fact, the problem is intrinsic, i.e. , such nonuniqueness comes about because of the simultaneous demand

the output power with the known input power. 6. It is not directly feasible to combine the Bussgang algorithms such as the constant modulus algorithm
(CMA) directly with a linear tap constraining equalizer parameterizations. Although it can be lumped conceptually into the same class as the center-spike initialization strategy, it will in general destroy the ability of CMA to recover the input because the unknown gain factor may be incompatible with the selected constant modulus.

4 Implementation and Simulation

that (1) the minimum be always achievable for an

ideal equalizer parameter setting, (2) convexity holds, (3) global minimum cost is invariant under ir/2 phase rotation, and (4) there is a constraint on the equalizer

4.1 Ideal Cost Function

The adaptation of the equalizer parameter based on the mm-

imization of the cost function that is equivalent to

J(O) = maxlRe{zk}I = IIRe{zk}II
requires some work since this cost function cannot be exactly

parameters. This is by no means a straightforward

observation; see Ref. 17 for a similar result developed for real data systems.

3. The linear constraint [Eq. (11)1 can be changed to any of three other possibilities according to various sign combinations. However, less obviously, more
general but not necessarily arbitrary linear constraints on the equalizer parameters (distributed over a number of equalizer parameters, not just Re{Oo} and Im{Oo}) can be employed. This fact is highlighted in the context that the nongeneric channels, which suffer from the nonuniqueness problem indicated above, can be ' 'cured' ' (to yield a point minimum) by using these

evaluated in practice with finite data length. Here we propose two algorithms: a gradient descent approach and an iterative minimization approach.

4.2 Gradient Descent Algorithm Based on l,,

One

Approximation natural approximation to l is the l,, norm with large

p. Given large enough p, the global minimum of the norm

more general linear constraints. We will not investigate this variation further here. 4. Convexity is an essential property in the above formulation because it guarantees that when we use causal finite dimensional practical equalizers the global convergence tendencies of the algorithm, or unimodality, are preserved. This behavior is a consequence of con-

will be close to global minima of the l norm. (Simulation examples later provide indications as to how large p needs to be.) In other words, the minimization of maxlRe(zk)I can be approximated by minimizing
IIRe{zk}IIP

(E{IRe{zk}})11"

or equivalently by minimizing E{Re{zk}I'}for a fixed p. Thus in practical implementation we minimize

E{IRe{zk}I"}

vexity since truncation (setting many taps zero) is a form of linear constraint, which does not destroy convexity. Convexity also ensures that one can approx1194 / OPTICAL ENGINEERING / June 1992 / Vol. 31 No. 6

for which a stochastic gradient descent minimization scheme

can be derived as

BLIND ADAPTIVE EQUALIZERS FOR QAM COMMUNICATION SYSTEMS

O(k+1)=O(k)pi

0IRe{zk}i

0O(k)

' i*0
(15)

(n+1) = oi

(n)

______________ _____________ (8max[Re{zk}12 + 8max[!m{zk}12)

ao

(k+1)=(k)2 8IRe{zk}i (k)

i*O
+ 1)

where pi and J12 should be chosen small enough to maintain numerical stability. Similarly, the minimization of maxllm(zk)l can be approximated through the minimization of E{IIm(zk)"} for a fixed p. To speed up the convergence rate, we choose to minimize the sum of the two convex cost functions (which

(n)

a4(n) (amax[Re{zk}12 + amax[Im{zk}12)

(17)

is still convex):
.(o) E{IRe{zk}I} + E{IIm{zk}I'}

The new parameters (r + 1) and 41? + 1) are then used on the same block of data {xk} to generate a new output sequence { zk} used for the next update step.

4.4 Simulations
We demonstrate the feasibility of our algorithms through

which is twice as sensitive to parameter variations as before. In practice, we only implement finite-dimensional parameter

vector 0 for the equalizer [see Eq. (7)]

p = 12 in our approximation of the l

several simulation examples. In all of our examples, we use

norm and a 21-tap blind equalizer. The initial estimates O,(O) are set at zero except for O(O) = 1, which satisfies the constraint [Eq. (11)1. In our simulations, the severity of IS! (eye-closeness) can be measured by the following definition of the interference to signal ratio (ISR) similar to that given by Lucky,25
ISR

O=[ON"OllOlONI

exp(j) . cos + sin4

Once 4 is initialized to be within ( ir/4, 3'rr/4), it will remain there. Since the sum of two l, norms J is strictly convex, linear constraints such as truncation preserve convexity. Since large values ofp must be used for l, to approximate the ideal lci norm, the adaptation step-sizes ii and 2 need to be very small, resulting in very slow convergence. To speed up the convergence rate of the gradient descent algorithm, various normalization techniques can be used. We can modify the adaptation algorithm into

IRe(t) + Im(t,)I
max{maxRe(t), maxIIm(t)I}

) miniRe(a1

\ maxlRe(ak)l

The combined channel and equalizer is an open-eye system if the ISR is less than unity. In the first set of simulations, we assume a real channel with phase error

O(k+1)=O(k)i1

maxlzki

1
pi

8iRe{zk}I + !m{zk}I 30,(k)

O.7z'
exp(j'rr/4)1

O.7z

i*0
4(k+1)=(k)i2
1

8IRe{zk}i + IIm{zi}

This channel is a first-order all-pass filter. For an i.i.d. quadrature phase shift key (QPSK) input, Fig. 3(a) illus(16)

maxlzkl

trates the output of the channel prior to equalization. Using the gradient algorithm [Eq. (16)] with ii = f-'2 = 0.02, the decreasing ISR level is shown in Fig. 4 and the equalizer

where max IZkI is updated every 100 iterations (output symbols). Due to the use of this normalization factor, the step-

sizes I.Li and 12 can be chosen reasonably large to enable faster convergence. The performance of this algorithm is presented after we introduce our alternative scheme.

output is shown in Fig. 3(b) after 20,000 iterations. The output eye is open after 10,000 iterations and the equalization effect is shown in the resulting equalizer output. As expected, the gradient algorithm requires a number of data samples to converge. If we use the block minimization algorithm with ii = p2 = 0.001 instead, the same equalizer converges after only 500 iterations operating on
a block of 400 data samples. The effectiveness of the block algorithm is further pronounced by simulation under rectangular QAM-16 input signal. Given 1000 channel output symbols, the converged output of the equalizer for QAM16 input (under x = 0.001) is shown in Fig. 5 after merely 200 iterations. Clearly, most IS! has been eliminated and phase error has been corrected.

4.3 Iterative Algorithm Based on Block

Approximation
As with many other blind equalization algorithms, such as

CMA, a gradient descent approach can be very slow in

convergence. The main advantages of these algorithms are their potential tracking ability and computational simplicity. The disadvantage is that they do not fully utilize the available data, which results in a rather slow convergence rate. Here we present a simple block minimization scheme. The algorithm simply collects a sequence of data {Xk} and generates a sequence of outputs {zk} through 0. Then one finds the zk in the sequence that has the maximum IRe{zk}I and maximum IIm{zk}I, respectively. The equalizer parameters are then updated through

We now test a complex channel whose real part and

imaginary part of its impulse response are shown in Fig. 6 as in the example of Ref. 4. The complex channel causes severe IS! and cross-coupling interference, which is evident from the set of channel output signals shown in Fig. 7. A 21-tap blind equalizer based on the block algorithm is then applied to the channel output in an attempt to reduce the effect of interference. After 300 iterations with a block of
OPTICAL ENGINEERING / June 1 992 / Vol. 31 No. 6 / 1195

KENNEDY and DING

4
2

0 Oo 8 c0 o 0P

0p

be

0
-2

0?
I

0)0
%0 0cP

ci)

0c1p0oo e

8cc 00 %

-'4

-4

-2

0
Real

1
000

(a)
I 0

Data Samples

Fig. 4 Declining ISR under gradient descent algorithm.

Input

.6

00 00 0 cs 00
5

-1
-2

%
L00 0
0 .9

q8
I

0 2
0 0

'

c;& 5cP o
t 0950

D 00

-2

-1

I 1

-5
-10
-10

Real
(b)

te 0%
0
(a)

Fig. 3 (a) Channel output and (b) equalizer output symbols.

-5

10

2000 data samples, the overall system has converged to a

output
4 2 0 -2 -4
-5

nearly open-eye system. The corresponding equalizer output shown in Fig. 7 demonstrates the effective reduction of 1ST.

The untilted square constellation of the equalizer output

indicates that the carrier phase error has also been corrected by the algorithm. The total impulse response of the channel and equalizer combination shown in Fig. 8 is very close to unity. Evidently, the block minimization algorithm requires far fewer data samples than our gradient descent algorithm and all other LMS-type gradient algorithms. The improvement of convergence rate depends largely on the step-sizes and data block size. Proper normalization can also be helpful. Admittedly, the iterative block algorithm requires more time for each iteration and may take as much computing time as the stochastic descent algorithm to converge. Its actual ad-

.J0000

0
(b)

vantage lies in the number of data samples required for convergence. It is more effective when only a short string of data is available, which is more often the case in blind
deconvolution particularly if the channel is time varying.

Fig. 5 (a) QAM-16 input and (b) output symbols of the equalizer on covergence.

ing cost function under linear constraints, which guarantees the global convergence of the equalizer parameters regard-

5 Conclusion
This paper presents a new adaptive algorithm for the joint

less of the number of equalizer parameters used. The ro-

bustness to truncation of equalizer length is in direct contrast

with what are known as the Bussgang algorithms. Simulation results demonstrate the effectiveness of this algorithm in removing 151 and correcting the carrier phase error.

blind equalization and carrier recovery of QAM signals.

Theoretical analysis shows the convexity of the correspond1

1 96 / OPTICAL ENGINEERING / June 1 992 / Vol. 31 No. 6

BLIND ADAPTIVE EQUALIZERS FOR QAM COMMUNICATION SYSTEMS

real part
1

Input

5
0.5

1i
I
5

-5
10
-5

-0.5

0
(a)

imaginary part

Output

0.5

?j T
5 10

-5

-0.5 0

Fig. 6 Impulse response of the complex channel.

-5

0
(b)

6 Appendix
Lemma 3.1 Proof. The objective here is to look over the special family of parameter settings corresponding to zero IS! (which is parameterized by the decoding offset q and phase rotation ) and determine which minimizes the cost [Eq. (9)]. Determined first is the optimal phase corresponding to a given q and then the optimal q. Since for arbitrary (q,4) the channel/equalizer convolution is given by
h101(q,4) =

Fig. 7 (a) Complex channel output and (b) equalizer output after equalizer convergence.

from which Eq. (18) reduces to

1

IRe{q}I + Im{1q}j

and so the optimal value, denoted by m, follows directly Eq. (12). [Note, k denotes the appropriate value of 1 in Eq. (19) when q=m.]
Theorem 3. 1 Proof. Lemma 3 . 1 furnishes the zero IS! equalizer parameter setting that minimizes the cost [Eq. (9)].

(i + q) exp(j4)
hq

cos4 + sin

then the cost [Eq. (9)] reduces to

Icos( q)I + Isin( 4q)I

The objective now is to compare this cost J[O(m,*)] with the cost J(O) for an arbitrary equalizer parameter setting 0

and show J(0)_J[0(m,*)I0.

(18)

hq Icos + sin(
For arbitrary q, Eq. (18) achieves its minimum when

Let 0 be an arbitrary parameter vector satisfying the linear

constraint. We now define Zt0 to be a perturbation (not

necessarily small) away from our candidate parameter setting, t to be the channel equalizer convolution for the candidate (which shows no IS!), z.t1 as the perturbation reflected in the channel equalizer convolution space, and 'y =J[0(m,*)] the cost realized by the candidate. These are written as

44q+l'fli2, q{0,iT/2,ii,3ii/2}

(19)

with integer I such that 0<4<ir/2, by elementary methods. Otherwise, when hq is purely real or purely imaginary, then Eq. (18) is invariant to any 4 such that 04rr/2 and this range of values achieves the minimum. To determine the optimal value taken by the parameter q, fix 4 according to Eq. (19) and = 0 otherwise. Then

zX00_0(m,4*)

= h,01(m,*) = (i .+ m)
Lti

Re{hm}I + Im{hm}I

exp(jklT/2)-

Icos + sinI = cos4ql + sin4ql

OPTICAL ENGINEERING / June 1 992 / Vol. 31 No. 6 / 1197

KENNEDY and DING

real part

1.5

<b 1

cos(kir/2) sgn(Re{h})(Re{uz Im{h

Re{hm}f +

IIm{m}

Notice that am = cos(/crr/2). Thus we have

J(O) J[O(m,*)I

Re{zt1}a , + Im{zt1}b
cos(k'rr/2) IRe{mH + IIm{1m}I
sgn(Re{hm})

0.5

=
ooO6Gbooe

0
-0.5

c?

?? 0Q0OOpOC
20
30

x [Re{
10

iXtl - } + Im{

- i}]

= cos(kIrr/2)Iy sgn(Re{hm})[Re{zOo}
+ Im{Mo}] = 0

1.5

imaginary part

Thus, whenever k =0 or 2, this is the desired result. In the case sgn(Re{1m}) = sgn(Im{hm}), we have k = 1 and 3. Then

0.5

t= hO(m,) =

j sin(krr/2) (i + m) Re{hm}I + IIm{mH

=j sin(k'rr/2)(i + m)
'J OppOOp?01QWU&?pOO(
-0.5

Thus,
J(O) J [O(m,*)]

10

20

30

= f Re{1t + tfll + E IIm{z1t + tfll

= IRe{t}I + Im{tm} + sin(k'rrI2)'yf

E fIm{zt1}f
m

Fig. 8 Total system impulse response after equalizer convergence.

1

IRe{}I + IIm{m}I
Consider the cases when Re{uim} and Im{lim} have the same

Re{zt1}f + sin(krr/2)Im{ztm}

sign. Thus k=0, 2 and

cos(kir/2) t= h,O1(m,) = i(i + m) IRe{hm}l +

+
m

IIm{zt}I

IIm{hm}I

= cos(kir/2)(i +
As a result,

(The last inequality becomes an equality for sufficiently small perturbations t_m.) Here we need the following elementary inequalities:
L 1c= sin(kir/2) sgn(Re{hm})(Re{h i} + Im{h 1

m)

Re{hm} + Im{hm}

E1

J(O)_J[O(m,4*)]

Re{zt,+tfl + IIm{/.ti+tflI Re{tj}I+IRe{t_m} m

Im{tj}I

1 1 <d

sin(k'rr/2) sgn(Re{hm})(Re{h i}
IRe{hmH + IIm{m}I

Im{)z })

+ cos(krr/2) y +

Notice that dm sin(k'rr/2). Thus, we have J(O) _J[O(m,*)] Re{Lt}c+ Im{zt}d

i m

IRe{t1}I + cos(k1T/2)Re{zt_m}

=
x

sin(k7r/2)

Im{zt}

Re{hm} + Im{hm}

sgn(Re{hm})

(The last inequality becomes an equality for sufficiently small perturbations t_m.) Here we need the following elementary inequalities:
1<

= sin(klT/2)y sgn(Re{hm})[Re{Mo}
+Im{Mo}]=0
Thus, whenever k = 1 or 3, this is the desired result.

[Re{

+ Im{

- i}]

-1 cos(k'rr/2) sgn(Re{hm})(Re{h } + IRe{Im}I +

Im{)z

IIm{m}I

1198 / OPTICAL ENGINEERING / June 1992 / Vol. 31 No. 6

BLIND ADAPTIVE EQUALIZERS FOR QAM COMMUNICATION SYSTEMS

References
1 . Y. Sato, ' 'A method of self-recovering equalization for multi-level amplitude modulation,' ' IEEE Trans. Commun. COM-23, 679682

adaptive equalizers in data transmission systems,' ' in Proc. lASTED

2. D. N. Godard, ' 'Self-recovering equalization and carrier tracking in

(June 1975).

two-dimensional data communication systems,' ' IEEE Trans. Cornmun. COM-28, 18671875 (Nov. 1980). 3. A. Benveniste, M. Goursat, and G. Ruget, ' 'Robust identification of a nonminimum phase system: blind adjustment of a linear equalizer in data communications,' ' IEEE Trans. Automatic Control AC-25, 385399 (June 1980). 4. A Benveniste and M. Goursat. ''Blind equalizers,' ' IEEE Trans. Cornmun. COM-32, 871882 (Aug. 1982). 5. J. R. Treichler and M. G. Agee, "A new approach to multipath correction of constant modulus signals,' ' IEEE Trans. Acoust. , Speech, Signal Proc. ASSP-31, 349472 (Apr. 1983). 6. 0. Shalvi and E. Weinstein, "New criteria for blind deconvolution ofnon-minimum phase systems (channels),' ' IEEE Trans. Info. Theory
7.

Paris (1985). 20. W. T. Rupprecht, "Adaptive equalization of binary nrz-signals by means of peak value minimization,' ' in Proc. 7th Europ. Conf. on Circuit Theory and Design, pp. 352355, Prague (1985). 21 . W. A. Sethares, R. A. Kennedy, and Z. Gu, "An approach to blind equalization ofnon-minimum phase systems,' ' in Proc. IEEE ICASSP91, pp. 15291532, Toronto, Canada (1991). 22. D. Hatzinakos and C. L. Nikias, "Blind equalization using a tncepstrum based algorithm, IEEE Trans. Cornmun., pp. 669682 (May

Int. Symp. Applied Signal Processing and Signal Filtering, pp. 99102,

23. C. L. Nikias, "Arma bispectrum approach to nonminimum phase

system identification,' ' IEEE Trans. Acoust. , Speech, and Signal Proc.,

1991).

IT-36, 312321 (Mar. 1990). S. Bellini and F. Rocca, ' 'Blind deconvolution: polyspectra or Buss-

ASSP-36, 513525 (Apr. 1988). 24. Z. Ding, C. R. Johnson, Jr. , and R. A. Kennedy, ' 'Local convergence of globally convergent blind adaptive equalization algorithms and initialization tactics," in Proc. IEEE ICASSP-91, pp. 15331536, Toronto, Canada (1991). 25. R. W. Lucky, "Techniques for adaptive equalization of digital communication systems," Bell Syst. Tech. J. 45, 255286 (Feb. 1966).

10. Z. Ding, R. A. Kennedy, B. D. 0. Anderson, and C. R. Johnson, Jr. , ' 'Ill-convergence of Godard blind equalizers in data communications," IEEE Trans. Comrnun., COM-39, 13131328 (Sep.
11.

gang techniques, in Digital Communications, E. Biglieri and G. Prati, Eds. , pp. 251263, North-Holland, Elsevier Science Publishers By. (1986). 8. J. E. Mazo, ''Analysis of decision-directed equalizer convergence," Bell Syst. Tech. J. 59, 18571876 (Dec. 1980). 9. 0. Macchi and E. Eweda, ' 'Convergence analysis of self-adaptive equalizers," IEEE Trans. Info. Theory IT-30, 162176 (Mar. 1983).
1991).

Rodney A. Kennedy received the BE

(Hons.) degree in electrical engineering from

the University of New South Wales, Australia, in 1982. From 1983 to 1 986 he worked

R. A. Kennedy, B. D. 0. Anderson, Z. Ding, and C. R. Johnson,

Honolulu (1990).

Jr. , ' 'Local stable minima of the Sato recursive identification scheme," in Proc. 29th IEEE Conf. on Decision and Control, pp. 3 1943199, Jr. , ' 'On the failure of proposed blind recursive identification schemes

on antenna design and control problems for the CSIRO Division of Radiophysics in Sydney. Concurrently he received the ME degree in digital control theory from the
University of Newcastle, Australia, in 1986. He received his PhD degree in December

12. R. A. Kennedy, B. D. 0. Anderson, Z. Ding, and C. R. Johnson,

based on the Sato algorithm and its generalizations,' ' in International Symposium on Signal Processing and its Applications, pp. 267270, Gold Coast, Australia (Aug. 1990). 13. Z. Ding, C. R. Johnson, Jr. , and R. A. Kennedy, "Non-global convergence of blind recursive identifiers based on gradient descent of continuous cost functions,' ' in Proc. 29th IEEE Conf. on Decision and Control, pp. 225230, Honolulu (Dec. 1990). 14. Z. Ding, "Application aspects of blind adaptive equalizers in QAM data communications,' ' PhD Thesis, Cornell University (Aug. 1990). 15. J. R. Treichler, V. Wolff, and C. R. Johnson, Jr. , "Observed misconvergence in the constant modulus adaptive algorithm, ' ' in Proc. 25th Asilomar Conference on Signals, Systems and Computers, pp. 663667, Pacific Grove, Calif. (1991). 16. 5. VerdU, B. D. 0. Anderson, and R. A. Kennedy, "Anchored blind

1988 from the Depament of Systems Engineering,Austra.an NationalUniversity(ANU)in Canberra. Currently Kennedy is a OEM Fellow in the Department of Systems Engineering at ANU. His research activities are in the fields of digital communications and signal processing. He is also an editor ofthe International Journal of Adaptive Control and Signal Processing.

Zhi Ding received the BEng degree in 1982

from the Department of Radio Engineering,

Nanjing Institute of Technology, Nanjing, China, and the MASc degree from the Department of Electrical Engineering, University of Toronto, Canada, in 1987. He was

equalization," in Proc. 25th Conf. on Info. Sci. and Systems, pp.

774779, Baltimore, MD (Mar. 1991). 17. 5. Vembu, S. Verdfl, R. A. Kennedy, and W. A. Sethares, "Convex cost functions in blind equalization," in Proc. 25th Conf. Info. Sci. and Systems, pp. 792797, Baltimore, MD (Mar. 1991). 18. W. T. Rupprecht, "Zeitdiskrete berechnung von verzerrungsmassen
als funktionen der filterkoeffzienten adaptiver datenleitungsentzerrer," ntz Archiv, pp. 25 1258 (July 1985). 19. W. T. Rupprecht, "Analysis of the mean-absolute-deviation-error for

a research and teaching assistant in the School of Electrical Engineering, Cornell University, Ithaca, New York, from August 1987 to August 1990, where he received

his PhD degree. Since September 1990 he has been with Auburn University, Alabama, where he is an assistant professor in the Department of Electrical Engineering. His general research interests include adaptive systems theory, data communications, and signal processing.

OPTICAL ENGINEERING / June 1 992 / Vol. 31 No. 6 / 1199