A New Adaptive Algorithm for Joint Blind Equalization and

Carrier Recovery

Z. D i n g R. A K e n n e d y
D e p a r t m e n t of Electrical Engineering D e p a r t m e n t of S y s t e m s Engineering
A u b u r n University A u s t r a l i a National University
Auburn, AL 36849, USA G.P.O. Box 4, C a n b e r r a , ACT 2601, A u s t r a l i a

Abstract a strong practical need for development of new algo-

rithms with more reliable and provable (global) cori-
vergence properties.
Traditional blind equalization algorithms based o n
m i n i m i z i n g s o m e specially designed n o n - M S E cost From recent works [9]it is clear that in order t o
f u n c t i o n s can ezperience local convergence problems design biind equalization algorithms that converge to
and thereby result in i n s u f i c i e n t o r n o removal o j a desirable parameter setting from arbitrary initializa-
channel distortion. I n t h i s paper, w e present a n e w tions (based on gradient descent minimization of some
QAM blind equalization s c h e m e t h a t i s globally conzier- cost function), certain aspects of the performance ob-
gent in t h e equalizer p a r a m e t e r space t o a compact set jective should be relaxed. In designing our new algo-
containing t h e desired ideal equalizer p a r a m e t e r set- rithm, we d o not attempt t o directly identify the exact
ting. F o r a generic class of channels, this n e w algo- gain of the channel inverse and focus instead on the
rithm results in t h e equalizer p a r a m e t e r convergence t g elimination of IS1 [lo]. So long as we can remove the
a unique global minimum achieving intersymbol inter- phase error in carrier recovery, IS1 elimination without
ference suppression and Carrie?. phase error removal. gain identification is a sufficient objective for the ulti-
Diflerent i m p l e m e n t a t i o n approaches are assessed and mate recovery of the channel input in Q A M systems.
simulatzon results are s h o w n t o confirm t h e theoretical Blind equalization of pulse amplitude modulation
global convergence o f t h e n e w algorithm. ( P A M ) systems without gain recovery has been pro-
posed in [11, 12, 131. However, generalizing these re-
sults directly to Q A M sysLerris is not straightforward.
1 Introduction The difficulty arises since in Q A M systems the input
signal is constrained to take values in the Q A M con-
Blind equalization of a n unknown, non-rniriirnum stellation. Phase also makes tap-fixing constraints un-
phase channel is a n important problem in d a t a corn- usable since they do not ensure the recovery of the
munication systems. Conventional adaptive equaliz- constellation in the equalizer output because of the
ers remove the intersymbol interference (ISI) caused unknown carrier phase error. I n the paper, we achieve
by non-ideal channels of limited bandwidth with the this desirable objective of joint blind equalization and
aid of a training session which provides the needed ref- carrier phase recovery using a novel non-anchoring t a p
erence signal for parameter adaptation. In situations constraint. Unlike many blind equalization propos-
where a training session is impossible or very costly, als we furnish a theorem that the algorithm we a r e
blind equalizers need to be used to combat the IS1 developing possesses desirable unique parameter con-
effect. vergence for generic channels even when finite impulse
Many schemes for blind equalization of quadra- response (FIR)equalizers which can only approximate
ture amplitude modulation ( Q A M ) system exist in the desired channel inverse are employed.
the literature. Typically, they are based on gradi- The major part of this paper is divided into the
ent descent minimization of some special non-MSE following sections. In section 2 , we describe the fun-
cost functions that do not require the use of refer- damental principles of linear blind equalization. Then
ence signal [l,2, 3 , 4, 51. They are sometimes known i.n section 3 , we present the new blind adaptive equal-
as the “Bussgang” algorithms [ 6 ] . However, the lack ization algorithm by introducing a new cost function
of unimodality of these cost functions, established in with a linear constraint of the complex equalizer pa-
[7, 8, 91, can result in the undesirable (local) con- rameters. We describe different implementation ap-
vergence of the corresponding algorithms such that proaches of the new blind equalization algorithm in
the equalizer fails to remove sufficient ISI. Ilence, de- section 4.
spite the apparent plethora of schemes, there remains

699
1058-6393/91 $01.00 0 1991 IEEE
2 Background Information In blind equalization, the original sequence is un-
known to the receiver except for its probabilistic
2.1 Blind Equalization in QAM Systems or statistical properties over the known alphabet A.
Usually this signal constellation is ms/2-rotation-
In QAM channel equalization system, the original invariant. If the statistics of the input d a t a reflects
symbol sequence {uk} takes values from a complex sig- this same symmetry (which is typical) over A, then
nal set A (known as the QAM constellation) in a man- a m s / 2 phase rotation does not cause any statisti-
ner such t h a t all possible symbol subsequences occur cal changes in the channel output. Thus, the d a t a
with non-zero probability. This statistical assumption recovered from blind equalization will be intrinsically
on the input is weaker than the typical i.i.d. assump- subjected t o a phase ambiguity of m s / 2 and the best
tion assumed in other works. (Further weak assump- possible result would be
tions regarding various symmetries of the constellation ~ ( z - ’ ) ~ ( z -=
l )e m n / 2 z - u m E (0, 1 , 2 , 3 } (2.6)
will be described later.)
T h e complex channel input {ak} is transmitted which relaxes the objective (2.5). T h e remaining
through a nonideal channel, assumed t o be linear, phase ambiguity can be resolved through differential
causal, and (bounded-input-bounded-output) stable. encoding of the input data. A further relaxation of
T h e channel has transfer function the ideal objective (2.5) arises since in practical ap-
00
plications only equalizers with a finite number of ad-
justable parameters can be implemented rather than
H(z-1) = p a . - ’ , hi E c, (2.1) those of (2.2). Typically these finite parameters are
i=O
arranged in the form of a causal F I R transversal filter.
where {hi} represents the channel impulse response Since an F I R equalizer can only approximate the de-
sequence. sired impulse response (2.6) a quantizer should be used
A linear channel equalizer is a linear filter O(2-l) t o recover the original channel input from the equal-
that is applied t o the channel output Zk in order t o izer output zk. This is often regarded as the practical
eliminate the ISI. Initially we may take this filter as objective of blind equalization.
being stable and potentially non-causal (doubly infi- Blind adaptive equalization algorithms are often de-
nite), and of the form: signed as stochastic gradient descent (SGD) schemes
t o update the parameter vector by minimizing some
m
special (non-MSE) mean cost functions t h a t do not
q Z - l )= ei(k)z-i, E c (2.2) involve the use of the original input ah but still reflect
a=-m the current level of IS1 in the equalizer output. Such
so as t o deal with non-minimum phase channels. T h e
schemes implicitly involve higher order statistics of the
time dependence of the equalizer parameters signifies channel and equalizer outputs. These algorithms are
that they are subject t o adaptation via a n algorithm sometimes referred to as the “Bussgang” algorithms
to be described. T h e equalizer output can then be [GI.
written as Bussgang algorithms have been shown to exhibit
00
local convergence [7, 8, 91. T h e lack of unimodality
by these algorithms can result in local, undesirable
Zk = ei(lc)zk-i = x ; e ( k ) (2.3) convergence of the equalizer parameters especially for
a=--00
F I R equalizers. Hence this motivates the development
by denoting the adaptive equalizer parameter vector of new and better blind equalization algorithms.
as
2.2 Constrained Blind Equalization
e(k) = [ ... 6-2 e-1 eo el 62 . . . 1’ (2.4)
In order t o design globally convergent blind equal-
and the equalizer input signal regressor as
ization algorithms further aspects of the performance
Xk = [ ... Zk+2 Zk+l 2k Zk-1 Zk-2 ...I/. objective (2.5) can be relaxed beyond those already in-
dicated. In designing our new adaptation rules, we do
T h e ideal objective of any channel equalizer O ( z - ’ ) not attempt t o identify the exact gain of channel in-
is t o achieve distortionless reception verse and focus instead on the elimination of IS1 which
is the primary objective of channel equalization [lo].
=~akPv,
zk = O ( Z - ~ ) Z v E Z+, V k , (2.5) Consequently, it is not essential t o recover the exact
(complex) gain of the channel inverse because once the
where we can tolerate a finite fixed time delay v. Such IS1 is removed such t h a t the equalizer output is
a n objective translates into the identification of chan-
nel inverse by the equalizer.

700
then it is straightforward t o estimate the unknown and @ denotes convolution. T h e equalizer is taken as
gain IcI by power-matching. T h e constant phase am- being doubly infinite and non-causal t o permit com-
biguity q ! ~can
~ also be readily resolved by the utiliza- plete removal of ISI. This assumption will be relaxed
tion of differential encoding, provided it is a multiple (as discussed above) in order t o deal the truly imple-
of n/2. mentable finite equalizer parameter space and avoid
Blind equalization of pulse amplitude modulation the potential weaknesses of this simplification as dis-
(PAM) systems without gain recovery has been pro- cussed in [14].
posed in [ll,12, 131. Their essential idea is parameter T h e cost function t o be minimized is designed as
“anchoring”. As shown in [11, 131, due t o the absence
A 1
of carrier phase error in PAM systems, “anchoring” M maxlRe{zk}/
J(0)= -
[IO] the center t a p can beneficially constrain the free-
dom of equalizer parameters such that a convex cost
function can be utilized for parameter adaptation.
03
In the next section, we present a new algorithm
for joint blind equalization and carrier phase recov-
ery based on specialized convex cost functions coupled ;=-m
with a constrained equalizer parametrization as a gen- 1
eralization of the algorithm in [12]. For a general class = m a x (Im{zk}( (3.1)
of channels with (independent) real and imaginary dis-
tortion, we obtain a unimodal mean cost function with
where z& = E,“=-, tiak-,.

which gradient descent algorithms exhibit parameter The convexity of J ( 8 ) with respect t o 6 follows from
convergence t h a t globally minimizes the mean cost. the triangle inequality. Notice that by fixing the real
This ideal convergence property, in contrast t o the or the imaginary part of the parameter vector 6 , the
classical “Bussgang” approach, is robust t o equalizer cost function J ( 6 ) = J(Re{e}, I m { e } ) is also convex
length truncation due t o the convexity of the cost func- in terms of Re{e} and Im{6}, a n observation which
tions employed. T h e robustness lies in that: (i) spu- will be important later. From
rious local minima are not created by the truncation;
and (ii) the global minimum in the strictly doubly in- ~ ( - 6 )= [lRe{-ti}/+ ~ ~ m { - t ; > l ]= ~ ( 6 ) .
finite parameter space (2.6) degrades gracefully into 1

a global minimum in the practical finite dimensional we have by convexity

parameter space where by selecting the length of the
equalizer appropriately one can ensure a n arbitrarily J ( O )5 0 . 5 ~ ( - 6 ) + (1 - ) qe),
0 . 5 ) ~ ( e= ve # 0.
small residual IS1 term. Thus if the range of 6 is unrestricted, then the convex
cost function (3.1) achieves a trivial global minimum
3 Globally Convergent Algorithms a t 6 = 0 , with equalizer output Z k = 0.
In order t o make the cost function J ( e ) useful, we
3.1 Convexity and Linear Constraint constrain the equalizer parameters with the following
linear constraint
We assume the QAM constellation is such that the Re{&} + Im(6,) =1 (3.2)
real and imaginary components are independent and
identical (e.g., the square type) with in the equalizer parameter vector (2.4). Due t o the
linearity of the constraint, the convexity of the cost
M 2 m a x JRe{ak}J = m a x IIm{ak}J function (with respect to the real and imaginary parts
of the equalizer coefficients) is maintained and global
(or by rotation can be transformed to such). We can convergence is therefore assured. We now show that
denote the total system (channel and equalizer) as under this parametrization, simultaneous removal of
IS1 and carrier phase error globally minimizes the cost
m function.
7(q-1) = H(q-1)qq-l) = tiq-a,
1=-m 3.2 Global Convergence
where {ti} is the impulse response of the combined Suppose the ideal non-causal channel inverse (as
system given by before) is given by
CO 00

t; = h, @6i = hk0i-k q-1) =f P ( 2 - 1 ) = i,z-i, x, E c

kt=O r=-w

701
such t h a t k ( z - ' ) exists on the unit circle. Then we and where IC E (1, 2 , 3 , 4 } i s c h o s e n s u c h t h a t ( 3 . 5 ) is
have satisfied, m i n i m i z e s t h e blind convez cost
hi @ii = 6 ( i ) ,
where 6 ( i ) is the Kronecker Delta function.
Our theoretical demonstration of global conver-
gence proceeds in a number of steps. First we present
a lemma which identifies the specific parameter set- 3.3 Remarks and Comments
ting from the countable class of ideal equalizers-- ~ e . ,
ones leading t o zero residual ISI-that minimizes the Despite the unimodality of the cost function with
linearly constrained cost ( 3 . 1 ) a n d ( 3 . 2 ) . In what fol- respect t o the equalizer parameters there is a potential
lows it is more convenient t o express the constraint problem of non-uniqueness of the (global) minimum,
( 3 . 2 ) in polar form: i.e., the minimum might be achieved by all points in a
compact set jact,ualiy a convex polytope) rathei than
(J(4) e'@ x 3x
-- < $fJ < -. a t a single point. There are two manifestations of this
-+
cos $ sin 4' 4 4 non-uniqueness: ( i ) as in Lemma 3.1 there can be a
Armed with this candidate the second step (our main phase ambiguity over a 7r/2 range if h, is purely real
theorem) demonstrates t h a t it minimizes globally the or purely imaginary which is a non-generic property;
given cost function over arbitrary equalizer param- and (ii) as in ( 3 . 3 ) when the maximum is realized by
eter settings which are subject t o the linear con- more t h a n one m which is again a non-generic prop-
straint ( 3 . 2 ) . T h e proofs are omitted here due t o space erty. T h a t these are the only ways non-uniqueness
limitation. arises can be inferred from the proof of Theorem 3.1.
T h e linear constraint ( 3 . 2 ) can be changed to any
Lemma 3.1 F o r ideal equalizer p a r a m e t e r settings of of three other possibilities according t o various sign
the f o r m possibilities. However, less obviously, more general
but not necessarily arbitrary linear constraints on the
equalizer parameters (distributed over a number of
equalizer parameters. This fact is highlighted in the
context t h a t the non-generic channels which suffer
parametrized by q E E ,and 4, the convez cost ( 3 . 1 ) from the non-uniqueness problem indicated above can
subject t o the linear constraint ( 3 . 2 ) is m i n i m i z e d b y be "cured" ( t o yield a point minimum) by using these
t h e following ( q , d ) - p a z r : more general linear constraints. We will not further
investigate this variation here.
m = argmax{lRe{i,)i + lIm{Lq}!) (3.3) Convexity is a n essential property in the above
formulation because it guarantees t h a t when we
0 < $m + <~ usc causal finite dimensional practical equalizers the
$* = { kr/2
q5 for any q5 E
/ 2 $m # m x / 2
io, x / 2 ] ,
otherwise global convergence tendencies of the algorithm, or uni-
modality, are preserved. This behavior is a conse-
where xm = /im1ej+". quence of convexity since truncation (setting many
taps zero) is a form of linear constraint which does
We defer comments regarding this lemma until after not destroy convexity. Convexity also ensures t h a t one
the next result. Now we present the main theorem. can approximate arbitrarily closely the performance
of the ideal non-implementable doubly infinite non-
Theorem 3.1 F o r a l i n e t r channel w i t h i n v e r s e im- causal equalizer parameter setting.
pulse response sequence {hi}, a doubly infinite n o n -
casual equalizer w i t h p a r a m e t e r setting 4 Implementation
~i+mefk~/2
4 ( m ,d*) = (3.4) T h e adaptation of the equalizer parameter based on
/Re{kmII +/~m(L)j the minimization of the cost function which is equiv-
alent t o
subject t o t h e constraint
J ( 6 ) = max/Re{zk}/ = IIRe{zk}/ja,
Re{Bo} + I m { e o } = 1, (3.5)
requires some work since this cost function can not be
where exactly evaluated in practice with finite d a t a length.
Here we propose two algorithms: a gradient descent
m = argmax{/Re{ilq}/ + /im{i,}l} approach a n d a n iterative minimization approach.

702
4.1 I, Approximation what are known as the “Bussgang” algorithms. Sim-
ulation results not shown here have demonstrated the
effectiveness of this algorithm in removing IS1 and cor-
One natural approximation t o I , is the 1, norm
recting the carrier phase error.
with large p . Given large enough p , the global min-
imum of the lp norm will be close t o global minima
of the I , norm. In other words, the minimization of References
m a x IRe(zk)[can approximated by minimizing
Y. Sato. IEEE Trans. on Communications,
COM-231679-682, June 1975.
D.N. Godard. IEEE Trans. on Communications,
or equivalently by minimizing E { l & ! { Z k } 1 ’ } for a fixed COM-28:1867-1875, November 1980.
p . Thus in practical implementation we minimize A. Benveniste, M. Goursat, and G . Ruget. IEEE
Trans. on Automatic Control, AC-251385-399,
E{ IRe{zk} 1
1’ June 1980.
for which stochastic gradient descent minimization J.R. Treichler and M.G. Agee. IEEE Trans. on
scheme can be derived. Similarly the minimization of Acoustics, Speech, and Signal Processings, ASSP-
m a x ( I m ( z k ) ( can be approximated through the mini- 31:349-472, April 1983.
mization of E{lIm(zt)I*}for a fixed p . To speed up
the convergence rate, we choose t o minimize the sum 0. Shalvi and E. Weinstein. IEEE Trans. o n Info.
of the two convex cost functions (which is still convex) Theory, IT-36:312-321, March 1990.
S. Bellini and F. Rocca. In E. Biglieri and
G. Prati, editors, Digital Communications, pages
251-263, North-Holland, 1986. Elsevier Science
In practice we only implement finite-dimensional pa-
Publishers B.V.
rameter vector B for the equalizer. Since the sum of
6

two Ip-norms J is strictly convex, linear constraints J.E. Mazo. Bell Syst. Tech. Journal, 59:1857-
such as truncation preserve convexity. 1876, December 1980.
Z. Ding, R.A. Kennedy, B.D.O. Anderson, and
4.2 Block Approximation C.R. Johnson, Jr. IEEE Trans. on Communica-
tions, September 1991.
As with many other blind equalization algorithms,
such as CMA, a gradient descent approach can be very Z. Ding, C.R. Johnson, Jr., and R.A. Kennedy. In
slow in convergence. T h e main advantages of these al- Proc. 29th IEEE Conf. o n Decision and Control,
gorithms are their potential tracking ability and com- pages 225-230, Honolulu, Dec 1990.
putational simplicity. T h e disadvantage is that they S. Verd6, B.D.O. Anderson, and R.A. Kennedy.
do not fully utilize the available d a t a , which results in In Proc. 25th Conf. on Info. Sci., and Systems,
rather slow convergence rate. Here we present a sim- pages 774-779, Baltimore, MD, March 1991.
ple block minimization scheme. T h e algorithm simply
collects a sequence of d a t a { z k } and generates a se- S . Vembu, S. Verd6, R.A. Kennedy, and W.A.
quence of outputs { Z k } through 8. Then one finds the Sethares. In Proc. 25th Conf. on Info. Sci., and
zk in the sequence that has the maximum ( R e { z k } l and Systems, pages 792-797, Baltimore, MD, March
maximum (Im{zk} 1, respectively. T h e equalizer pa- 1991.
rameters are then updated t o minimize m a ~ [ R e { z k } ] ~
W . T . Rupprecht. Zeitdiskrete berechnung von
and max[Im(zk}]’
verzerrungsmassen als funktionen der filterkoef-
fiaienten adaptiver datenleitungsentzerrer. ntz
5 Conclusion Archiv, pages 251-258, July 1985.
W.A. Sethares, R.A. Kennedy, and Z. Gu. In Pro-
This paper presents a new adaptive algorithm for
ceedings of IEEE ICASSP-91, pages 1529-1532,
the joint blind equalization and carrier recovery of
Toronto, Canada, 1991.
QAM signals. Theoretical analysis shows the con-
vexity of the corresponding cost function under linear R.A. Kennedy, B.D.O. Anderson, Z. Ding, and
constraints, which guarantees the global convergence C.R. Johnson, Jr. In Proc. 29th IEEE Confer-
of the equalizer parameters regardless of the number ence on Decision and Control, pages 3194-3199,
of equalizer parameters used. T h e robustness t o trun- Honolulu, HI, 1990.
cation of equalizer length is in direct contrast with

703