1427

A Stop-and-Go Dual-Mode Algorithm

for Blind Equalization
Ching-Hsiang Tseng and Cheng-Bin Lin
Department of Electrical Engineering
National Taiwan Ocean University
2 Pei-ning Rd., Keelung, Taiwan, R.O.C.

ABSTRACT A dual-mode algorithm which incorpo- when the equalizer tap weights are reasonably closed to
rates a stopand-go technique for blind equalization is their optimal values. However, one may have difficulty in
investigated in this paper. The constant modulus algo- determining the proper timing for making the switch. To
rithm (CMA) exhibits very slow convergence when a p deal with this problem, a dual-mode CMA was developed
plied to QAM signals. To speed up the convergence rate, [5]. The dual-mode CMA has the blind eqaulizer oper-
a dual-mode CMA was developed. We show that the con- ated in either the CMA or a radius directed equalization
vergence properties of the dual-mode CMA can be signif- (RDE) mode [6]. In addition, the transitions between the
icantly ‘improved by simply adding a stopand-go tech- two modes are automatically done by evaluating an error
nique. The operation of the stopand-go dual mode CMA function determined by the radius of the equalizer out-
is independent of the carrier offset and thus can be used put. If the equalizer output has a relatively large error
even when the carrier offset is unknown. level, the equalizer considers itself far from optimum and
thus adopts the CMA to adjust its tap weights. However,
I. Introduction if the error level of the equalizer output comes within a
predetermined range, the equalizer considers itself closed
Adaptive equalization is a widely used technique for
to optimum and thus updates its tap weights using the
combating intersymbol interference generated in a ban-
RDE. Such a dual-mode method has been shown to ex-
dlimited communication channel. Conventionally, the o p
hibit good performance [5].
timal tap weights of the equalizer are estimated by trans-
mitting a predetermined training data sequence. How- Although the aforementioned dual-mode algorithm
ever, for a multipoint network, the training technique is switches in two modes based on a radius-determined er-
unfavorable due to the fact that it significantly reduces ror function, it never stops adjusting the equalizer tap
the data throughput [l]. Another situation where train- weights, even when the adjustment is in the wrong di-
ing reduces transmission capacity is a mobile communica- rection. If we can tell whether a particular adjustment
tion channel subject to deep selective fading [a]. For such is correct (i.e., in the right direction) or not, we may
applications, a blind equalizer [I, 3, 41 which can adapt improve the convergence behavior by making only those
its tap weights without the aid of a training sequence is right adjustments but bypassing those wrong ones. In the
called for. blind equalization case, we do not know for sure whether
Among various blind equalization algorithms, the CMA a particular adjustment is correct or not. However, we
originated from Godard [l]has received much attention. may evaluate the reliability (in a probabilistic sense) of
One of the nice features of the CMA is that it oper- each adjustment, and allow only those adjustments which
ates independently of the carrier offset, which is generally are highly reliable. Such a concept has been applied to
unknown in many situations. However, one unfavorable a decision-directed blind equalizer and was termed “stop
feature of the CMA is that it converges very slowly for and-go” [7]. However, to the best of our knowledge, it has
QAM signal constellations. One of the reasons is that, not been applied to a CMA-type algorithm, whose oper-
for QAM signals, the error term of the CMA does not ation is independent of the carrier trachng. In this pa-
go to zero even when the equalizer tap weights have con- per, we develop a stopand-go dual-mode CMA for blind
verged to their optimal values. Therefore, a relatively equalizers. This algorithm operates in two modes as in
small step size is required for the CMA to achieve an ac- the conventional dual-mode CMA. But no matter which
ceptable steady-state error. The small step size leads to mode the equalizer is in, the proposed algorithm allows
slow convergence of the CMA. To increase the conver- adaptation of equalizer tap weights only when a simple
gence rate, one may have the equalizer operated in the flag shows the current adaptation is sufficiently reliable.
CMA mode with a large step size in the early stages, and If the flag suggests the current adaptation is likely to be
then switch the equalizer to the decision-directed mode incorrect, the current adaptation is stopped. We show

0-7803-3336-5196 $5.00 0 1996 IEEE

Authorized licensed use limited to: IC Plus Corp. Downloaded on April 22,2021 at 03:34:05 UTC from IEEE Xplore. Restrictions apply.
 1428

that, by avoiding certain incorrect adaptations, the pro-

posed algorithm achieves a better convergence rate than
the conventional dual-mode CMA does. The goodness
of the proposed algorithm is demonstrated via computer
simulation.

11. The Conventional Dual-Mode CMA

Consider a twedimensional communication channel
with a baseband equivalent channel impulse response of
h(n). Suppose the channel input, output, and noise are
denoted by a(n), u(n), and ~ ( n )respectively.
, We have
u ( n ) = U(.) * h(n) + v(n),where * denotes linear conve
lution. The channel output u ( n ) is sent to a tapdelay-
line equalizer with a tap length N = 2L 1. Denot- +
ing the input data vector of the equalizer by u(n) =
[u(n + L ) . . . u(n - L)IT and the equalizer tap weight
vector by w(n), we may express the equalizer output y ( n ) Figure 1: Partition regions of the 16-QAM constellation
as y ( n ) = wT(n)u(n). for the DMCMA.
The conventional dual-mode CMA (DMCMA) (which
was called dual-mode Godard algorithm (DMGA) in [5])
approaches convergence, the tap weights are good enough
divides the two-dimensional signal space into several an-
to generate equalizer outputs which are localized around
nular regions Dk, k = 1, 2, . . .. The annular region Dk
the QAM data points. Under such a circumstances, most
contains the QAM signal points which possess a common
of the equalizer outputs will be in UDk and (1) will be
amplitude Rk. DI, is also bounded by its inner radius
used to adjust the tap weights most of the time. This
+
RI, - d and outer radius R k d , where d is a predeter-
provides a good steady-state behavior for the DMCMA.
mined constant. For example, the annular regions for the
18QAM signal constellation are shown in Fig. 1. The
DMCMA can be described by the following equations:
111. The Stop-and-Go Dual-Mode CMA
w(n + 1) = N.1 -a [lWl2- G]y(n)u*(n),
k = 1,2,.. . (1)
y ( n ) E Dk, The DMCMA successfully makes judgement on
w ( n + 1) = w(n) - a [ly(n)I2 - R2] y(n)u*(n), whether the adaptation of the equalizer is in a tran-
sient state or a steady state. It uses CMA to adjust
y(n) GUDk (2) the tap weights if the equalizer is in a transient state,
where a denotes the step size, UDk denotes the union of and uses RDE otherwise. However, the DMCMA does
the annular regions DI,, k = 1, 2, . . ., and R2 is defined not make judgement on whether a particular adjustment
as follows: (no matter the CMA or the RDE is used) is correct or
not. Since the equalizer is in a blind equalization mode,
(3) many adjustments could be actually in wrong directions.
If we can somehow avoid some of those incorrect adjust-
ments, the convergence behavior of the equalizer will be
Note that the adaptations described by (1) and (2) are ac-
improved. Such an idea was first proposed in [7], where
tually the same as the ones performed in RDE and CMA,
a blind “stopand-go” decision-directed equalizer decided
respectively [6, 11. Therefore, whether the CMA (Eq. (2))
or the RDE (Eq. (1))will be used is determined by the whether adaptation of the current iteration should ‘(stop”
location of y ( n ) in the signal space. The regions for which or (‘go” by observing a simple flag. The flag suggests “go)’
the RDE and the CMA will be performed are indicated
if the self-determined output error i s sufficiently reliable
in Fig. 1. for adaptation, and suggests “stop” otherwise.
The DMCMA operates as follows. During the tran- In the following, we present a stopand-go DMCMA
sient stages, the equalizer tap weights are far from o p (SAG DMCMA) for blind equalization. This algorithm
timum and thus the equalizer outputs are scattered far inherits the stopand-go concept proposed in [7] but dif-
away from their corresponding QAM data points. There- fers in that it can be used independently of the carrier
fore, most of the equalizer outputs will not be in UDk tracking. Define two error terms 6(n)and E(n) as follows:
and (2) will be used to adjust the tap weights most of
the time. This provides a CMA-like initial convergence S(n) = I m I 2 - R;, (4)
behavior for the DMCMA. However, when the equalizer E(n) = Iy(n)I2 - R2, (5)

Authorized licensed use limited to: IC Plus Corp. Downloaded on April 22,2021 at 03:34:05 UTC from IEEE Xplore. Restrictions apply.
where

The SAG DMCMA then can be described by the following

equations:

w(n + 1) = w(n) - af(n)2(n)y(n)u*(n),

y ( n ) E Dk, IC = 1,2,. . . (7)
w(n + 1) = w(n) - af(n)E(n)y(n)u'(n),
y(n) 61 U D I , (8)

where

1 if sgn 2(n) = sgn E(n)

0 if sgn 2(n) # sgn E(n) (9)
Figure 2: Partition regions of the 16-QAM constellation
Note that the only difference between the stopand-go and for the SAG DMCMA.
the conventional DMCMA methods is that a flag f ( n ) is
added in the SAG DMCMA. and for the SAG DMCMA we have
The operation of the SAG DMCMA is as follows. No-
tice that the two error terms 2(n) and E(n) actually con-
trol the adaptation directions of the RDE and the CMA,
respectively. When the signs of 2(n) and E(n) match,
4n) = { C ( n ) y ( n ) y ( n ) E DI, and f ( n ) = 1
e"(n)y(n)
0 f(n)= 0
# U D k and f(n)= 1(12)
which means the RDE and the CMA suggest the same
adaptation direction, then we feel more confident on the If we vary y ( n ) in a radial direction, we may write y ( n ) in
correctness of the adaptation and thus may go ahead to polar coordinates as (ry,O),where ry = I y ( n ) l and O is the
make the adjustment. On the contrary, when the signs angle of the radial direction. From (11) and (12) we see
of 2(n) and E(n) disagree, which means the RDE and the that e ( n ) , 2(n),and .G(n)can also be written in polar coor-
CMA suggest opposite adaptation directions, we are not dinate forms as ( r e $ ) , (.,e), and (re$), respectively. For
sure which direction is the correct one and thus may stop comparison, we plot re as a function of ry of the 64-QAM
the current adjustment. For those adjustments which are case for both the conventional and the SAG DMCMA
granted, whether the RDE or the CMA is used is still methods with a d = 0.2. The result is shown in Fig. 3,
dependent of the region the equalizer output y ( n ) is 10- where rg and r g as functions of ry are also indicated. We
cated. That is, the RDE is used if y ( n ) E Dk,k = 1,2,. . ., see that, both methods have the error re set to zero when
and the CMA is used otherwise. Using the 16-QAM sig- rY is equal to one of the amplitudes of the 64-QAM sig-
nal constellation as an example, as shown in Fig. 2, we nal points. This guarantees that the equalizer will stop
see that the SAG DMCMA partitions the signal space adjusting the tap weights when the channel is perfectly
into several annular regions. These regions are catego- equalized. We may also see from Fig. 3(b) that, for those
rized into three sets: RDE, CMA, and STOP. The sets to regions where r$ and r,- suggest opposite signs, the SAG
which the annular regions belong are indicated in Fig. 2, DMCMA simply sets the error re to zero.
where the "dotted" circles are the decision boundaries of The effect of the stop-and-go strategy on the conver-
the RDE, and the "dashed" circles are the boundaries of gence behavior can be explained as follows. Consider
the Dk. Some DI, boundaries are omitted in Fig. 2 for using a step size which is acceptable for both the s t o p
simplicity. The difference between the stopand-go and and-go and the conventional DMCMA methods. In the
the conventional DMCMA methods is clearly shown by transient state, since the SAG DMCMA adjusts the tap
comparing Figs. 1 and 2. weights only when the flag suggests so, it may miss some
In fact, both the stop-and-go and the conventional DM- adjustments which are actually correct. Therefore, the
CMA methods can be reformulated as follows: initial convergence rate of the SAG DMCMA may seem
a little bit slower than the one of the conventional DM-
CMA. However, since the adjustments made by the SAG
w(n + 1) = w(n) - ae(n)u*(n) (10) DMCMA have a higher probability of being correct, the
where for the conventional DMCMA we have convergence rate of the SAG DMCMA will catch up and
approach the rate of the conventional DMCMA. When
the adaptation approaches the steady state, the equalizer
outputs are localized around the QAM signal points and

Authorized licensed use limited to: IC Plus Corp. Downloaded on April 22,2021 at 03:34:05 UTC from IEEE Xplore. Restrictions apply.
 1430

0.520
0.854

-
400

300
*** r-,
‘e - 0.W OM9 0.03 om
200
- -.....J. . ............ ..
100

real component imaginary component

Figure 4: The channel impulse response used in the sim-

ulation.

-5
-10
t 1
400

re 300
-15
200
100
-5% -20

0 -25
-100 E
I -30
-200; 2 4 6 8 10 12
-35
rY
-40 t I
2 4 6 8 10 12 14 16 18 20
3
NCJh3BER OF ITERATIONS 10
Figure 3: The re as a function of ry for (a) the DMCMA
and (b) the SAG DMCMA cases. Figure 5: The convergence curves of the tapweight vector
for the d = 0.1 case and for (a) the DMCMA with an
a = 8 x lo-?, (b) the SAG DMCMA with an Q = 8 x
thus the adjustment direction suggested by Z(n) is cor- and (c) the SAG DMCMA with an a = 2 x lop6.
rect most of the time. However, for an equalizer output
y(n) which is not in U D k , the conventional DMCMA ad-
justs the tap weights in the direction suggested by E(n)
IV.Computer Simulation
(see (ll)),whose sign may be different from Z ( n ) ’ s (for In the following, we demonstrate the goodness of the
example, see the point p (denoted by “0”)in Fig. 1). For SAG DMCMA by simulations. The communicationchan-
such equalizer outputs, the conventional DMCMA will nel used in the simulations has an impulse response shown
make adjustments which are in incorrect directions (and in Fig. 4 and has been used in [7, 51. A 11-tap ( N = 11
are perhaps large). On the contrary, the SAG DMCMA and L = 5) complex equalizer was used for the simula-
adjusts the tap weights only when 6 ( n ) = g ( n ) . For ex- tions. The signal constellation used was 64-&AM and
ample, the point p in Fig. 1 is now stopped as shown in the signal-tcmoise ratio (SNR) was set to 40 db. We
Fig. 2 due to the fact that 8(n) # Z ( n ) for this point. first utilized a step size of Q = 8 x lop7 for both the
No matter what mode (CMA or RDE) is eventually se- conventional and the SAG DMCMA methods. For each
lected for a granted adjustment, the adjustment direction iteration, the mean square error (MSE) of the tapweight
will be the one suggested by 8 ( n ) (which is correct most vector was calculated and is plotted in Figs. 5 and 6 for
of the time when the adaptation approaches the steady the d = 0.1 and d = 0.2 cases, respectively, where the
state). Therefore, we expect the SAG DMCMA to exhibit MSE was calculated by ensemble averaging 10 indepen-
a lower steady-state error. dent realizations of data. We see that, for both cases, the
From another point of view, to achieve a given steady- SAG DMCMA initially converges slower but catches up
state error, the conventional DMCMA requires a smaller at a certain point, and eventually achieves a much lower
step size than the SAG DMCMA does. Since increas- steady-state MSE.
ing the step size increases the convergence rate, the SAG We next increased the step size to a = 2 x lop6. The
DMCMA is able to achieve a better convergence rate if convergence curves for the d = 0.1 and d = 0.2 cases
both methods are requested to achieve the same level of are also shown in Figs. 5 and 6, respectively. Note that
steady-state error. the curves for the conventional DMCMA are not shown,

Authorized licensed use limited to: IC Plus Corp. Downloaded on April 22,2021 at 03:34:05 UTC from IEEE Xplore. Restrictions apply.
 1431

-5 1 1
-10

-15

-20

-25

-30

-35

-40

2 4 6 8 10 12 14 16 18 20
-5
NUMBER OF ITERATIONS 10
-10

Figure 6: The convergence curves of the tapweight vector -15

5
h

for the d = 0.2 case and for (a) the DMCMA with an

‘
-20

a = 8x lop7, (b) the SAG DMCMA with an a = 8 x u2. -25

and (c) the SAG DMCMA with an a = 2 x lop6. -30

_c
-22 ’ II. 1, ,‘a? 1 ’ . ‘T’ r I. , .
’,““ ,, I .
because for such an Q the conventional DMCMA does
not converge at all. These results clearly show that the
SAG DMCMA performs better. More specifically, we
see that the SAG DMCMA achieves a lower steady-state
MSE for a given step size, and converges faster for a re- Figure 7: The convergence curves of the equalizer output
quested steady-state MSE. The simulation results confirm for (a) the DMCMA case with an a = 8 x lop7 and (b)
the statement we made in Section 111. the SAG DMCMA case with an a = 2 x lop6.
We also checked the equalizer output by plotting the
variation of the output MSE with respect to the number system,” IEEE Trans. Commun., vol. COM-28, pp.
of iterations. The result for the d = 0.2 case is shown 1867-1875, NOV. 1980.
in Fig. 7. Again, we see the SAG DMCMA is better in
the sense of convergence rate as well as final steady-state [a] J. Karaoguz and P. Yegani, “Performance of a blind
MSE. clustering equalizer for digital mobile radio chan-
nels,” Proc. IEEE 44-th Vehicular Technology Con-
V. Conclusion ference, pp. 267-271, Mar. 1994.
In this paper, we have presented a stopand-go dual- [3] Y. Sato, “A method of self-recovering equalization
mode CMA for blind equalization. The proposed algo- for multilevel amplitude-modulation system,” IEEE
rithm checks a simple flag to see if the self-determined Trans. Commun., vol. COM-23, pp. 679-682, June
output error is sufficiently reliable for tapweight adapta- 1975.
tion. The tap weights are adjusted only when the flag sug-
[4] A. Benveniste and M. Goursat, “Blind equalizer,”
gests so. The proposed algorithm inherits the nice CMA
IEEE Trans. Commun., vol. COM-32, pp. 871-883,
property of being independent of carrier tracking. We
Aug. 1984.
have demonstrated via simulation that the convergence
property of the stopand-go dual mode CMA is better [5] V. Weerackody and S. A. Kassam, ‘(Dual-modetype
than the one of the conventional dual-mode CMA. algorithm for blind equalization,” IEEE Trans. Com-
mun., vol. COM-42, pp.22-28, Jan. 1994.
Acknowledgement
[6] M. Ready and R. Gooch, ”Blind equalization based
This work is supported in part by the National Science on radius directed adaptation,” Proc. ICASSP 90,
Council under Contract NSC 85-2611-E019-011. Albuquerque, pp. 1699-1702.
[7] G. Picchi and G. Prati, “Blind equalization and car-
References rier recovery using a stop-and-go decision- directed
algorithm,” IEEE Trans. Commun., vol. CQM-35,
[l] D. N. Godard, “Self-recoveringequalization and car- pp. 877-887, Sept. 1987.
rier tracking in two-dimensional data communication

Authorized licensed use limited to: IC Plus Corp. Downloaded on April 22,2021 at 03:34:05 UTC from IEEE Xplore. Restrictions apply.