1592 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 47, NO.

6, JUNE 1999

Dithered Signed-Error CMA:

Robust, Computationally Efficient
Blind Adaptive Equalization
Philip Schniter and C. Richard Johnson, Jr.

Abstract—Adaptive blind equalization has gained widespread robustness properties. A sizeable body of theoretical analysis
use in communication systems that operate without training exists to support this claim [3], including, for example, studies
signals. In particular, the constant modulus algorithm (CMA) on CMA’s robustness to noise [8], channel undermodeling [9],
has become a favorite of practitioners due to its LMS-like
complexity and desirable robustness properties. The desire for and lack of disparity [10].
further reduction in computational complexity has motivated Low-cost consumer applications (e.g., HDTV) motivate
signed-error versions of CMA, which have been found to lack blind equalization techniques requiring minimum implemen-
the robustness properties of CMA. This paper presents a simple tation cost. Although it is noted for its LMS-like complexity,
modification of signed-error CMA, based on the judicious use of CMA may be further simplified by transforming the bulk of
dither, that results in an algorithm with robustness properties
closely resembling those of CMA. In this paper, we establish its update multiplications into sign operations [2]. A recent
the fundamental transient and steady-state properties of dithered study suggests, however, that straightforward implementations
signed-error CMA and compare them with those of CMA. of signed-error CMA (SE-CMA) do not inherit the desirable
Index Terms—Adaptive equalizers, adaptive signal processing, robustness properties of CMA [11]. In this paper, we present
blind equalization, constant modulus algorithm, deconvolution, a simple modification of SE-CMA based on the judicious
dither techniques, HDTV. incorporation of controlled noise (sometimes referred to as
“dither”) that results in an algorithm with robustness properties
closely resembling the standard (unsigned) CMA. In fact, we
I. INTRODUCTION
show that the mean behavior of dithered signed-error CMA

T HE CONSTANT modulus algorithm (CMA) [1]–[3] has

gained widespread practical use as a blind adaptive
equalization algorithm for digital communications systems
(DSE-CMA) is identical to CMA under realistically achievable
conditions. The anticipated drawback to this dithering is a
degradation in steady-state mean-square error (MSE) perfor-
operating over intersymbol interference channels. Modern mance. Hence, we derive an expression for the excess MSE
receiver implementations often realize the advantages offered (EMSE) of DSE-CMA and discuss implications on step-size
by a fractionally spaced equalizer (FSE), i.e., an equalizer and equalizer-length selection. We note in advance that the
operating at a rate higher than the baud rate [4] and/or EMSE expression for DSE-CMA bears close resemblance to
processing data from multiple sensors (see, e.g., [5]). Under a an analogous expression derived for CMA in [12].
set of perfect blind equalizability (PBE) conditions (listed in The paper is partitioned as follows. Section II presents
Section II-B), CMA-adaptation of a FSE is known to converge the fractionally spaced system model and reviews some fun-
in mean to an equalizer setting capable of perfect symbol damental properties of fractionally-spaced CMA. Section III
recovery [6], [7]. discusses computationally-efficient blind equalization and in-
Although assumptions of ideality are convenient for the troduces the new algorithm. The transient and steady-state
theoretical analysis of blind equalization schemes, they are properties of DSE-CMA are studied in Section IV and result
unconditionally violated in physical implementations of com- in the design guidelines of Section V. Simulation results based
munication systems. This fact motivates the consideration of on the Signal Processing Information Base (SPIB)1 microwave
algorithm performance under realistic (nonideal) conditions. channel models are presented in Section VI. Section VI also
We will use the term robust when referring to a blind algo- includes a comparison study with another robust computa-
rithm’s ability to perform “well” under violations of the PBE tionally constrained implementation of CMA. For simplicity,
conditions. It has been reasoned that the widespread practical we restrict the focus of this paper to the case of real-valued
use of fractionally spaced CMA bears testament to its superior quantities. As discussed in Section VII, however, extension to
the complex-valued case is straightforward.
Manuscript received February 18. 1998; revised August 14, 1998. This The following is a word on notation: We use lower-case
work was supported in part by the National Science Foundation under Grant
MIP-9509011, a Schlumberger Foundation Fellowship, and Applied Signal bold-face quantities (e.g., ) to denote vectors and upper-case
Technology, Inc. The associate editor coordinating the review of this paper bold-face quantities (e.g., ) to denote matrices. Conjugation
and approving it for publication was Prof. James A. Bucklew.
The authors are with the School of Electrical Engineering, Cornell Univer-
sity, Ithaca, NY 14853 USA. 1 The Rice University Signal Processing Information Base (SPIB) mi-
Publisher Item Identifier S 1053-587X(99)03688-0. crowave channel database resides at http://spib.rice.edu/spib/microwave.html.

1053–587X/99$10.00  1999 IEEE

SCHNITER AND JOHNSON: DITHERED SIGNED-ERROR CMA: ROBUST, COMPUTATIONALLY EFFICIENT BLIND ADAPTIVE EQUALIZATION 1593

Perfect symbol recovery (PSR) occurs when the channel

noise is absent, and when and are such that
for all , some fixed system delay , and some
fixed scalar such that The PSR system responses
Fig. 1. T =2-spaced baseband communication system model. are characterized by (where denotes a vector with
1 in the th position and zeros elsewhere). We refer to PSR-
inducing equalizers as zero-forcing,3 and denote a zero-forcing
is denoted by , transposition by , and the norm by equalizer associated with system delay by The goal of
Finally, the variable is reserved for the baud-rate time blind equalization can be considered the achievement (or near-
index. achievement4) of PSR based only on knowledge of the system
output and the (marginal) statistics of the source process
II. FRACTIONALLY SPACED CMA

B. The Constant Modulus Algorithm

A. The Fractionally Spaced System Model
The CMA is a stochastic gradient algorithm minimizing
In this section, we construct a received signal model based
the Godard criterion: The positive
on a single-sensor receiver operating at twice the baud rate.2
constant is referred to as the dispersion constant and is
Note that an equivalent model results from the use of two sen-
chosen in accordance with the source statistics. Conceived
sors and that generalization to multiple sensors/oversampling
independently in [1] and [2], the Godard criterion penalizes
is straightforward [5]. Consider a baseband communication
the dispersion of the squared output modulus away from
system operating at baud interval A -spaced symbol
As an FSE update algorithm, CMA takes the form
sequence is transmitted through a linear time-invariant
and finite impulse response channel characterized by , which (1)
is a length- vector of -spaced impulse response co-
efficients In addition to intersymbol interference, the
-spaced received signal is also corrupted by an
additive noise process The baseband receiver consists of where is a (small) positive step size. The function
a -spaced linear equalizer specified by the coefficients identified in (1) is referred to as the CMA error function and
in the vector At baud time index , the receiver forms the will appear many times throughout this paper.
symbol estimate from the previous received The following perfect blind equalizability (PBE) conditions
samples, as collected in the vector Fig. 1 shows the are known to be sufficient to guarantee that equalizers mini-
linear system relating transmitted symbols to the system mizing achieve perfect symbol recovery [3]:
outputs We assume that the source symbols are drawn A1) full column-rank ;
from a finite, zero-mean, symmetric alphabet with variance A2) no additive channel noise;
A3) sub-Gaussian source: the source’s normalized kurto-
Defining the fractionally spaced convolution sis must be less than that of a
matrix , we have the equation at the bottom of the page, Gaussian process;
which allows us to write the received vector as A4) i.i.d. zero-mean source (circularly symmetric in the
, where is a vector containing the pre- complex-valued case:
vious samples of the channel noise process. The baud- Note that A1) and A2) pertain to the channel-equalizer pair’s
rate linear system relating to is now compactly de- ability to achieve PSR, whereas A3) and A4) pertain specifi-
scribed by the -spaced impulse response vector cally to blind adaptive equalization using the Godard criterion.
so that with length- source vector
3 The terminology “zero-forcing” stems from the equalizer’s ability to force
The structure of implies
the symbol estimation error to zero.
that 4 When used for blind startup, i.e., those situations in which training is
not present and the system error rate is too high for decision-directed (DD)
2 A more tutorial (and more complete) development of the fractionally techniques to function reliably, the goal of the blind algorithm may be that of
spaced system model can be found in [3]. reducing error rate to a level adequate for successful decision-direction.

.. .. ..
. . .
.. ..
. .
.. .. ..
. . .
1594 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 47, NO. 6, JUNE 1999

Fig. 2. CMA, SE-CMA, and DSE-CMA error functions.

Fig. 3. SE-CMA trajectories for BPSK transmitted over noiseless channel
h = 0
(0:1; 0:3; 1; 0:1; 0:5; 0:2)t superimposed on Jcm cost contours.
III. COMPUTATIONALLY EFFICIENT CMA Dotted lines delineate SE-CMA constant-gradient facets.

Straightforward implementations of LMS-like adaptive al-

gorithms (such as CMA) require a multiplication between
the error function and every regressor element [see update
(1)]. Many practical applications benefit from eliminating
these regressor multiplies. Signed-error (SE) algorithms
present one method for doing so, whereby only the sign of
the error function is retained [13]. When a SE algorithm is Fig. 4. Quantization noise model (right) of the dithered quantizer (left).
combined with a power-of-two step size, it is possible to
construct a multiply-free fixed-point implementation of the
B. Dithered Signed-Error CMA
equalizer update algorithm. The subsections below discuss
two versions of SE-CMA. (For the remainder of the paper, “Gimme noise, noise, noise, noise ”—The Replacements,
we restrict our focus to the case where all quantities are real Stink, 1982.
valued. Extensions to the complex-valued case are discussed In this section, we describe a simple modification to SE-
in Section VII.) CMA that results in an algorithm whose mean behavior closely
matches that of standard (unsigned) CMA.
A. Signed-Error CMA Viewing the SE-CMA error function as a one-bit quantizer,
we might wonder whether a suitable dithering technique [14]
The real-valued SE-CMA algorithm [2] is specified as would help to remove the unwanted behavioral artifacts caused
sgn (2) by the sign operator.5 Dithering refers to the addition of a
random signal before quantization in an attempt to preserve the
information lost in the quantization process. From an additive
noise perspective, dithering is an attempt to make the so-
where sgn denotes the standard signum function. Equation
called quantization noise (see Fig. 4) white, zero-mean, and
(2) defines the SE-CMA error function depicted in Fig. 2.
independent of the signal being quantized. We might expect
A recent investigation into SE-CMA has shown that while
that such quantization noise could be “averaged out” by a
satisfaction of the PBE conditions and correct selection of
small step-size adaptive algorithm, yielding mean behavior
ensure mean convergence to PSR, violation of A1) can
identical to that of its unsigned counterpart. These ideas are
severely hinder SE-CMA convergence behavior [11]. Specif-
made precise in Section IV-B.
ically, there may exist vast yet highly suboptimal regions
The real-valued dithered signed-error constant modulus al-
in equalizer space in which the expected update in (2) is
gorithm (DSE-CMA) is defined [17] by the update
zero. Fig. 3 presents an example of such behavior, in which
the trajectory labeled “B” appears not to converge. (See
Fig. 7 for examples of CMA trajectories under identical con- (3)
ditions.) Thus, while computationally efficient, SE-CMA does
not inherit the desirable robustness properties of CMA. This
fact motivates the search for computationally efficient blind 5 The authors acknowledge a previous application of controlled noise to
algorithms that do inherit these robustness properties. The SE-LMS in the context of echo cancellation [15], [16]. However, both the
following section describes one such algorithm. analyzes and goals were substantially different than those in this paper.
SCHNITER AND JOHNSON: DITHERED SIGNED-ERROR CMA: ROBUST, COMPUTATIONALLY EFFICIENT BLIND ADAPTIVE EQUALIZATION 1595

where is an i.i.d. “dithering” process uniformly dis- B. DSE-CMA Transient Behavior

tributed on , both and are positive constants, and The average transient behavior of DSE-CMA is completely
is the DSE-CMA error function. The practical se- determined by the expected DSE-CMA error function
lection of the dispersion constant and the “dither amplitude” Equations (5)–(8) indicate that
are discussed in Section V. It should become clear in the is a “hard-limited” version of the CMA error function
next section why appears twice in (3). , i.e.,

IV. THE FUNDAMENTAL PROPERTIES OF DSE-CMA

(10)
Sections IV-B to IV-D utilize an additive noise model of
the dithered sign operation to characterize the transient and
steady-state behaviors of DSE-CMA. Before proceeding, we Fig. 2 plots the various error functions and for
present the details of this quantization noise model. comparison. In the theorems below, the implications of (10)
are formalized in terms of DSE-CMA behavior over specific
A. Quantization Noise Model of DSE-CMA ranges of
Lemma 1: Define
At first glance, the nonlinear sign operator in (3) appears to
complicate the behavioral analysis of DSE-CMA. Fortunately, (11)
the theory of dithered quantizers allows us to subsume the
sign operator by adopting a quantization-noise model of the The choice of dither amplitude ensures that
DSE-CMA error function (see Fig. 4). Appendix A collects for all equalizer outputs satisfying the output amplitude
the key results from classical quantization theory that allow constraint
us to formulate this model. Proof: By evaluating at the locations where , it
DSE-CMA can be connected to the quantization literature can be seen that the “humps” of the cubic CMA error function
with the observation that the operator sgn is identical to (see Fig. 2) occur at heights Thus, is
the two-level uniform quantizer specified by unique and well-defined for Since (10)
implies that such values of prevent these humps from being
(4) clipped in forming the expected DSE-CMA error function,
and are identical over the interval when
for quantizer spacing Furthermore, the specification
that be uniformly distributed on ensures that For values is determined by the unique
real-valued root of the cubic polynomial and
satisfies the requirements for a valid dither process
can be expressed as
outlined in Appendix A as long as is selected large enough
to satisfy

(5) (12)

for relevant values of the equalizer output Recall that From (12), it can be shown that
denotes the CMA error function, defined in (1). Writing the system output as for a (fixed) re-
Employing the model of Fig. 4, we write the DSE-CMA ceived vector and arbitrary equalizer allows the following
error function in terms of the quantization noise equalizer-space interpretation of Lemma 1.
Theorem 1: Denote the set of possible received vectors by
(6) , and define to be the convex hull formed by the set
which leads to the following DSE-CMA update expression: of hyperplanes for
Then, choice of dither amplitude ensures that the
(7) expected DSE-CMA update is identical to the CMA update
for equalizers within
When and satisfy (5), the properties of follow from Proof: Choose any two equalizers and that satisfy
(28), (29), and (31) in Appendix A. Specifically, we have the output constraint for all (Re-
that is an uncorrelated random process whose first moment call that is well defined for ) The triangle
obeys inequality implies that any convex combination of and
also satisfies this output constraint. Lemma 1 ensures that,
(8) for that satisfy the output amplitude constraint,
Hence, the two updates are identical within
and whose conditional second moment is given by

(9) For an -ary source, the set of possible source vectors

is of size Then, in the absence of channel noise, we
In (8) and (9), the expectation is taken over the dither process, expect at most equalizer input vectors Hence,
thus leaving a dependence on in this noiseless case, is the convex hull formed by the
1596 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 47, NO. 6, JUNE 1999

finite set of hyperplanes TABLE I

for In other words, is a polytope formed by the CRITICAL VALUES OF  FOR M -PAM
boundary set An illustrative example of and is
provided by Fig. 6.
Next, we concern ourselves with neighborhoods of the zero-
forcing (ZF) equalizers
Theorem 2: Define
(13)

Under satisfaction of the PBE conditions, choice of dither

amplitude ensures the existence of a neighborhood
around every ZF solution within which the expected DSE-
CMA update is identical to the CMA update.
Proof: When , the satisfaction of the PBE
conditions implies that for all In this case,
(10) and the definition of imply that
for In other words, guarantees that the
expected DSE-CMA update is identical to the CMA update at
the zero-forcing solutions.
Now, consider an open ball of radius centered at
Equalizers within can be parameterized as
for Then, there exists a finite constant for which
From the continuity of
the polynomial function , we claim the following: For any
and any , there exists a such that
implies Applying (10), we
conclude that for any equalizer within the ball Fig. 5. CMA error function and critical values of  for 4-PAM and 16-PAM
sources.
Note that the constant may be less than , in which
case, there would exist isolated “CMA-like” neighborhoods
around the ZF solutions—i.e., neighborhoods not contained in within which the expected DSE-CMA update is identical to
any “CMA-like” convex hull. the CMA update.
Theorem 2 is of limited practical use since it requires Proof: The proof is identical to that of Theorem 2 after
satisfaction of the PBE conditions. Fortunately, the concept replacing by
is easily extended to the set of “open-eye” equalizers In summary, is the lower limit of for which the
Denoting the minimum distance between any pair of adjacent convex set exists, whereas and are the lower
symbols in by , we define the set as6 limits of for which “CMA-like” local neighborhoods around
the zero-forcing and open-eye equalizers exist, respectively.
Table I quantifies the values of for -PAM
The corresponding set of open-eye equalizer outputs is defined alphabets, and Fig. 5 illustrates their relationship to the CMA
by error function. Note that the difference between and
narrows as the alphabet size increases. This can be attributed
to the fact that the open-eye neighborhoods shrink as the
constellation becomes more dense.
For -PAM, becomes the open interval
minus the set of points halfway between
C. DSE-CMA Cost Surface
adjacent elements of Here, and are used to denote
the minimum and maximum positive-valued elements of , Studies of the multimodal cost surface give substantial
respectively. insight into the transient behavior of CMA (see, e.g., [3]).
Theorem 3: Define Thus, we expect that an examination of , which is the
cost stochastically minimized by DSE-CMA, should also prove
(14) worthwhile. First, however, we need to construct Since
we know that a gradient descent algorithm minimizing has
Choice of dither amplitude ensures the existence of the general form , we conclude from
a neighborhood around every open-eye equalizer (3) that It is then possible
6 We acknowledge that the definition of FOE is overly strict in that to find (to within a constant) by integrating
it bounds the outermost decision region from both sides. In addition, the over -dimensional equalizer space.
definition of FOE only makes sense in the context of bounded inputs r :
Although the AWGN channel model does not ensure bounded r , all practical Fig. 6 shows an illustrative example of contours
implementations do. superimposed on contours in equalizer space for
SCHNITER AND JOHNSON: DITHERED SIGNED-ERROR CMA: ROBUST, COMPUTATIONALLY EFFICIENT BLIND ADAPTIVE EQUALIZATION 1597

Fig. 6. Superimposed DSE-CMA (solid) and CMA (dotted) cost con-

tours in equalizer space for BPSK, noiseless channel h = (0:1; Fig. 7. Trajectories of DSE-CMA (rough) overlaid on those of
0
0:3; 1; 0:1; 0:5; 0:2)t and  = 1: Dashed lines show the set of hyperplanes CMA (smooth) for BPSK, noiseless channel h = (0:1; 0:3; 1; 0:1; 0
B F
 whose convex hull  ensures expected DSE-CMA behavior identical 2
0:5; 0:2)t ;  = 5 1004 ; and  = 1: Solid lines are Jcm contours, and
to that of CMA. B
dashed lines form the boundary set  :

Note that the two sets of cost contours are identical error function can be written as
within the convex polytope formed by the hyperplanes
Outside , the CMA cost contours rise much quicker than
the DSE-CMA contours. This observation can be attributed to
the fact that for large is proportional to ,
For small output error (i.e., ), the error function can
whereas the hard limiting on makes proportional
be approximated by
to As a result, we expect that CMA exhibits much
faster convergence for initializations far outside of Unlike (15)
standard SE algorithms [13], however, DSE-CMA converges
as rapidly as its unsigned version within Fortunately, there In the absence of channel noise, we can write
is no need to initialize the adaptive algorithm with large ; using the parameter error vector
the “power constraint property” of CMA [8] ensures that the defined relative to the zero-forcing equalizer For
CMA minima lie in a hyperannulus that includes7 adequately small , (15) implies that the CMA error
(see, e.g., Fig. 9). Initialization of DSE-CMA is discussed in function has the approximate form
Section V.
Fig. 7 shows two low-dimensional examples of a DSE- (16)
CMA trajectory overlaid on a CMA trajectory. Note that the
DSE-CMA trajectories closely follow the CMA trajectories but Under the PBE assumptions and a reasonably small step-
exhibit more parameter “jitter.” The effect of this parameter size, we expect asymptotically small Thus, the small-error
variation on steady-state MSE performance is quantified in approximation (16) can be used to characterize the steady-state
the next section. behavior of DSE-CMA.
2) The Excess MSE of DSE-CMA: We define EMSE at
D. DSE-CMA Steady-State Behavior time index as the expected squared error above that
achieved by the (local) zero-forcing solution Since, under
The principle disadvantage of DSE-CMA concerns its satisfaction of the PBE conditions, achieves zero error
steady-state behavior: The addition of dither leads to an
increase in excess mean-squared error (EMSE). EMSE is (17)
typically defined as the steady-state MSE above the level
attained by the fixed locally minimum MSE solution. The We are interested in quantifying the steady-state EMSE:
subsections below quantify the EMSE of DSE-CMA under Our derivation of steady-state EMSE
the satisfaction (or near-satisfaction) of the PBE conditions. assumes the following:
1) Small-Error Approximation of the CMA Update: By B1) The equalizer parameter error vector is statisti-
writing the equalizer output in terms of the delayed source cally independent of the equalizer input
and defining the output error , the CMA B2) The dither amplitude is chosen sufficiently greater
7 Assuming that the equalizer input is power-normalized, as occurs in than so that for all under consid-
practice. eration.
1598 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 47, NO. 6, JUNE 1999

B3) The PBE conditions A1)–A4) are satisfied to the extent V. DSE-CMA DESIGN GUIDELINES
that the zero-forcing solution attains near-zero error,
i.e., A. Selection of Dispersion Constant
B4) The step size is chosen small enough for the small-
We take the “Bussgang” approach used in [1], whereby
error approximation (15) to hold asymptotically.
is selected to ensure that the mean equalizer update is zero
The classical assumption B1) implies that is independent when perfect equalization has been achieved. From (3), (10),
of the source process Assumption B2) is needed for and the system model in Section II-A, we can write the mean
the results of the quantization noise model in Section IV-A update term of DSE-CMA at (in the absence of noise)
to hold. as For an i.i.d. source, is
Using the facts that tr for any scalar and that independent of all but one element in , namely,
tr tr and tr tr for any Hence, we require that the value of in be chosen so that
matrix , the EMSE at time index can be written
(24)
tr
When , Theorem 2 ensures the existence of a
tr (18)
neighborhood around within which For
where the second step follows from B1). Defining the expected such , (24) implies that should be chosen as for CMA, i.e.,
[1]. When , closed-form expressions
equalizer outer product matrix and
for in the case of -PAM DSE-CMA are difficult to derive.
the source-power-normalized regressor autocorrelation matrix
However, satisfying (24) for these cases can be determined
, we can write the EMSE as
numerically.
tr (19)
B. Selection of Dither Amplitude
Note that since is i.i.d. and , we have Although Section IV-D demonstrated that EMSE is pro-
portional to , Section IV-B showed that larger values of
Appendix B uses the quantization noise model from increase the region within which DSE-CMA behaves like
Section IV-A and the error function approximation from (16) CMA. The selection of dither amplitude is therefore a design
to derive the following recursion for , which is valid for tradeoff between CMA-like robustness and steady-state MSE
equalizer lengths : performance.
Theorems 1 and 2 imply that the choice
ensures that the zero-forcing equalizers are contained in the
(20) convex polytope Thus, under near-satisfaction of the PBE
conditions, could be considered a useful
Using (19) and (20), Appendix C derives the following ap- design guideline since the CMA minima are expected to be in
proximation to the steady-state EMSE of DSE-CMA: close proximity to the zero-forcing solutions [3]. In fact, since
is convex and contains the origin, we expect that a small-
(21) norm initialization (see Section V-D) will lead to equalizer
trajectories completely contained within Such a strategy
is advantageous from the point of robustness.
where The approximation in (21) closely
In situations where the PBE conditions are more
matches the outcomes of experiments conducted using mi-
severely violated and CMA can do no better than “open
crowave channel models obtained from the SPIB database.
the eye,” selection of dither amplitude in the range
The simulation results are presented in Section VI.
is recommended
Equation (21) can be compared with an analogous expres-
to retain CMA-like robustness.
sion for the EMSE of CMA [12]:
Table I presents these critical values of for various -
PAM constellations. Note that the value of for BPSK
(22) appears unusually large because near-closed-eye operating
conditions for BPSK are quite severe.
It is apparent that the EMSE of CMA and DSE-CMA differ
by the multiplicative factor C. Selection of Step-Size
As in “classical” LMS theory, the selection of step size
(23) becomes a tradeoff between convergence rate and EMSE.
If convergence rate in noncritical, could be selected with
via Note the dependence on both the robustness in mind and selected to meet steady-state MSE
dither amplitude and the source distribution. Table II requirements.
presents values of for various -PAM sources and Say that the goal was to attain the same steady-state MSE
particular choices of (to be discussed in Section V-B). performance as CMA. Then, under satisfaction of the PBE
SCHNITER AND JOHNSON: DITHERED SIGNED-ERROR CMA: ROBUST, COMPUTATIONALLY EFFICIENT BLIND ADAPTIVE EQUALIZATION 1599

TABLE II TABLE III

STEADY-STATE MSE RELATIVE PERFORMANCE FACTOR: Jex = K;S Jexjcma Jex DEVIATION FROM PREDICTED LEVEL
FOR VARIOUS SPIB CHANNELS AND M -PAM

conditions, should be chosen times that of CMA,

where was defined in (23). Table II presents values of
over the recommended range of and can be used to
predict the typical range of CMA convergence speed relative
to DSE-CMA (for equal steady-state performance).
When neither convergence rate nor steady-state MSE per-
formance can be sacrificed, Table II suggests choosing
closer to In this case, CMA-like robustness
is sacrificed instead. For such , however, it becomes hard
to predict the effects of PBE violations on the transient and
steady-state performance of DSE-CMA. Loosely speaking, as
is decreased below , the performance of
DSE-CMA becomes more like that of SE-CMA.

D. Initialization of DSE-CMA
The single-spike initialization [1] has become a popular
initialization strategy for baud-spaced CMA, as has double-
spike initialization [3], which is its -spaced counterpart.
The similarities between DSE-CMA and CMA suggest that
these initialization strategies should work well for DSE-CMA Fig. 8. Averaged MSE trajectories for DSE-CMA and CMA initialized at
the same locations using 8-PAM and (normalized) SPIB channels 1, 2, 6, 8,
as well. and 13. For all simulations: SNR =40 dB, Nf = 32;  = 2 1005 ; and
2
In the interest of preserving CMA-like robustness, however,  = OE = 2:25:
it is suggested the norm of the DSE-CMA initialization be kept
small.8 Under proper selection of (i.e., ), this strategy
ensures that the parameter trajectories begin within the convex iterations. Overall, the simulation results closely match our
region (see Fig. 9). Extending this idea, Section IV-B approximation (21).
implies that large enough choices of (e.g.,
ensure that the entire mean trajectory will stay within B. Average Transient Behavior
(and for adequately small step-sizes, the actual trajectories Throughout the paper, we have emphasized the importance
should closely approximate the mean trajectory). To conclude, of performance evaluation in realistic (nonideal) environments.
proper choice of initialization norm and dither amplitude It is only proper to present a comparison of DSE-CMA
will guarantee that the mean behavior of DSE-CMA never to CMA in this context as well. Fig. 8 shows ensemble-
differs from that of CMA. averaged MSE trajectories of the two algorithms operated
under identical conditions and initialized at the same locations
VI. SIMULATION RESULTS using various SPIB microwave channels. Noise levels (SNR
40 dB) and equalizer lengths were selected
A. Excess MSE Under PBE Conditions to represent typical applications while providing open-eye
performance (for an 8-PAM source) at convergence. The
Table III presents simulation results verifying the approxi- following “double-spike” equalizer initialization was used in
mation of the excess MSE of DSE-CMA given in (21). The all simulations: taps 10 and 11 were set to 0.5, and all
simulations were conducted using length-64 MMSE approxi- others were set to zero. Although (purposely) sub-optimal,
mations of three (noiseless) SPIB microwave channels, length- this initialization represents a reasonable choice given the
62 -spaced FSE’s, and various i.i.d. -PAM sources. microwave channel profiles and the discussion in Section V-D.
In other words, PBE conditions A1) to A4) were satisfied. As evident in Fig. 8, the DSE-CMA trajectories track the CMA
The step sizes were chosen so that B4) was satisfied, and trajectories closely until the effects of EMSE take over. Fig. 8
the dither amplitude of satisfied B2). Table III gives also suggests that the EMSE approximation in (21) remains
percentage deviations from the EMSE levels predicted by (21), a useful guideline even under typical violations of the PBE
which were obtained by averaging the results of 2.5 108 conditions.
8 This is consistent with recent recommendations on the initialization of Although parameter trajectory comparisons are impractical
CMA in single-user applications [18]. with length-32 equalizers, it is easy to visualize two-tap
1600 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 47, NO. 6, JUNE 1999

Fig. 10. Averaged MSE trajectories for DSE-CMA and update-decimated

Fig. 9. Averaged DSE-CMA and CMA tap trajectories initialized at the
CMA initialized at the same locations using 8-PAM and SPIB channel 8.
Relevant parameters: SNR = 40 dB, Nf = 32;  = OE = 2:25; and
same locations and superimposed on CMA cost contours for channel
= 0
h (0:1; 1; 0:5; 0:1; 0:2; 0:1)t ; SNR = 30 dB, 4-PAM, and  = 2: update-decimation factor D = 16:
Dotted lines indicate CMA power constraint boundaries and dashed lines
B
indicate  :
factor (Recall that typical values of appear
in Table II.)
examples. Fig. 9 shows ensemble-averaged DSE-CMA tra-
As verification of our claim, Fig. 10 presents ensemble-
jectories overlaid on ensemble-averaged CMA trajectories
averaged MSE trajectories comparing DSE-CMA with UD-
for a noisy undermodeled channel and 4-PAM. The two
CMA for and The operating environment
trajectories in each pair correspond so closely that they are
and design quantities used were the same as those of Fig. 8,
nearly indistinguishable from one another. The trajectories
with the exception that for UD-CMA. This
were initialized from various locations on the inner CMA
UD-CMA step size was adjusted to equate steady-state per-
power constraint boundary and remain, for the most part, in
formance, and thus, the advantage of DSE-CMA appears in
Note that for trajectories that cross a single boundary
the form of increased convergence rate. Checking Table II,
plane in the set , the expected DSE-CMA update differs
we find that for dither amplitude and an 8-PAM source,
from CMA for only one element in the set of possible received
DSE-CMA is expected to “beat” UD-CMA whenever must
vectors In other words, loss of CMA-like behavior outside
be selected
occurs gradually.
VII. CONCLUSIONS
C. Comparison with Update-Decimated CMA This paper has derived the fundamental properties of the
One popular technique used to reduce the computation dithered signed-error constant modulus algorithm. In sum-
requirements of CMA involves updating the equalizer every mary, we have found that under proper selection of algorithmic
baud samples [rather than every sample, as (1) suggests]. design quantities, the expected transient behavior of DSE-
This is possible in situations where the channel time variations CMA is identical to that of CMA. Although the steady-state
are slow with respect to the equalizer adaptation speed. As an MSE of DSE-CMA is larger than that of CMA, its value is well
example, fixed-site microwave applications can often tolerate characterized and can be accounted for in the design procedure.
update decimations of and higher [19]. The funda- With the exception of computational complexity, the new
mental drawback to these decimated algorithms is that their algorithm has been designed to mimic CMA rather than
convergence rates decrease in proportion to “improve” on its performance. Our primary motivation for
Since DSE-CMA and update-decimated CMA (UD-CMA) this is twofold. First, CMA is well-regarded by practitioners.
both present strategies for computationally efficient CMA-like It has established itself over the last 20 years as the most
blind adaptation, a comparison is in order. In Section V-C, we popular practical blind equalization algorithm, due in large
discussed how DSE-CMA step size may be selected to achieve part to its robustness properties [3]. It is precisely these
steady-state MSE levels on par with CMA and argued that robustness properties that we have attempted to preserve.
the penalty is DSE-CMA convergence rate times slower Second, CMA has been extensively analyzed by theoreticians.
than CMA. Although, for a given step size, UD-CMA should The bulk of these analyses apply directly to DSE-CMA. As
achieve the same steady-state performance as CMA, we expect it is often the case that modifications of classic algorithms
a convergence rate that is times slower. Taken together, we have disadvantages that outweigh the proposed advantages,
anticipate advantages in using DSE-CMA in situations where the spirit of DSE-CMA is a computationally efficient algorithm
the implementation budget demands a UD-CMA decimation that “leaves well enough alone.”
SCHNITER AND JOHNSON: DITHERED SIGNED-ERROR CMA: ROBUST, COMPUTATIONALLY EFFICIENT BLIND ADAPTIVE EQUALIZATION 1601

Although we have restricted our focus to the real-valued the average noise power increases [14], but more
case, a straightforward complex-valued extension of DSE- importantly, the class of information signals satisfying (26)
CMA is obtained by replacing the real-valued sgn in (3) for a fixed shrinks. Take, for example, the case where
with the complex-valued operator csgn sgn Re so that has a triangular distribution on
sgn Im and by replacing the real-valued dither process In this case, (26) is only guaranteed when Worse yet,
with the complex-valued Here, choices of fail to meet (26) for any In other words,
and the processes and are real-valued, uniformly distributed on is the only
independent, and distributed identically to It can be dither process that yields a useful quantization noise model
shown that with minor modifications, the properties of real- for the two-level quantizer of (4).
valued DSE-CMA apply to its complex-valued counterpart We will now quantify for uniformly distributed
[20]. Hence, the design guidelines of Section V apply to both dither. Note that the quantization noise takes on the values
the real- and complex-valued cases. with conditional proba-
Finally, we mention a potentially useful modification to bilities , respectively. The
DSE-CMA. In the case of SE-LMS, the extension of the sign conditional expectation then becomes
operator to a multilevel quantizer has been shown to yield
significant performance improvements at the expense of a
modest increase in computational complexity [21]. Perhaps
multilevel quantization would yield similar advantages for
DSE-CMA: most importantly, a reduction in EMSE.
(31)
APPENDIX A
FUNDAMENTAL PROPERTIES OF
NONSUBTRACTIVELY DITHERED QUANTIZERS
APPENDIX B
In this appendix, we review the key results from the
DERIVATION OF
theory of dithered quantizers that allow us to formulate a
quantization-noise model for the DSE-CMA error function. This appendix derives a recursion for the DSE-
Fig. 4 illustrates the model described below. CMA parameter-error-vector expected-outer-product
We define the quantization noise arising from the non- We assume that B1)–B4), which
subtractively dithered quantization of information signal were stated in Section IV-D2, hold. In the sequel, the notation
as will be used to denote a matrix whose th entry
is specified by
(25) Under B2), subtracting from both sides of (7) yields
for a dither process and for defined in (4). When Thus, the expectation
the quantizer spacing is large enough to satisfy of the outer product of is

(26)
and the dither is the sum of i.i.d. random variables uniformly
distributed on (and statistically independent
of ), the quantization noise has the following properties [14]:
(27)
The quantization noise properties (8) and (9) can be applied
(28)
to simplify the previous expression.
In words, (27) and (28) state that the quantization noise
is an uncorrelated random process whose th moment is
uncorrelated with the information signal Note that for all
values of , we have the important property that quantization
noise is uncorrelated with the information signal : Applying the small-error approximation (16), the outer product
(29) recursion is well described, for small , by

For , however, we have the property that the quantization

noise power is correlated with the information signal
(30)
Although dither processes characterized by higher values
of make the quantization noise “more independent” of the (32)
information signal , it is not without penalty. For one, The individual terms in (32) are successively analyzed below.
1602 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 47, NO. 6, JUNE 1999

The second and third terms in (32) are transposes of one The diagonal nature of implies ,
another. For now, we concentrate on the first of the pair, for where and represent the th diagonal elements of
which we can use B1) and the fact that and , respectively.
to write The similarity transformation can be applied to (20) to
obtain a recursion in terms of

Since , we de-
fine the matrix with elements For the characterization of , we are interested in only the
steady-state values of the diagonal elements In terms
of the th element

Then in matrix notation,

Because as , the limit of the
(where is a vector with a one in the th position and zeros
previous equation becomes
elsewhere). Incorporating the definition of from A3), we
conclude that

where we have introduced the shorthand notation

For long equalizers (i.e., ), the second term in the We can now sum over to obtain
preceding equation is dominated by the first so that we can
approximate

Finally, since , these approximations yield Using the fact that tr

, we finalize our approximation for , which is
the asymptotic EMSE of DSE-CMA:

As for the fourth and fifth terms of (32), no- An alternate, although useful, form for can be obtained
tice that B1) implies using the relation tr for even
As we know from
Section V-A, the dispersion constant is selected to force
Thus, the fourth and fifth terms
of (32) vanish.
Rewriting the final term of (32), the approximated outer
product recursion (valid for small and ) becomes REFERENCES
[1] D. N. Godard, “Self-recovering equalization and carrier tracking in two-
dimensional data communication systems,” IEEE Trans. Commun., vol.
COMM-28, pp. 1867–1875, Nov. 1980.
[2] J. R. Treichler and B. G. Agee, “A new approach to multipath correction
of constant modulus signals,” IEEE Trans. Acoust., Speech, Signal
Processing, vol. ASSP-31, pp. 459–472, Apr. 1983.
APPENDIX C [3] C. R. Johnson, Jr. et al., “Blind equalization using the constant modulus
DERIVATION OF criterion: A review,” Proc. IEEE (Special Issue on Blind System
Identification and Estimation), vol. 86, pp. 1927–1950, Oct. 1998.
In this appendix, we use (20) to determine an expression for [4] R. D. Gitlin, J. F. Hayes, and S. B. Weinstein, Data Communications
the steady-state EMSE achieved by DSE-CMA. A similarity Principles. New York: Plenum, 1992.
[5] E. Moulines, P. Duhamel, J. Cardoso, and S. Mayrargue, “Subspace
transformation of the symmetric Toeplitz matrix is employed methods for blind identification of multichannel FIR filters,” IEEE
to simplify the derivation , where the matrix Trans. Signal Processing, vol. 43, pp. 516–525, Feb. 1995.
is diagonal, and the matrix is orthogonal. Applying this [6] Y. Li and Z. Ding, “Global convergence of fractionally spaced Godard
(CMA) adaptive equalizers,” IEEE Trans. Signal Processing, vol. 44,
transformation to yields , where pp. 818–826, Apr. 1996.
is, in general, not diagonal. Using the properties of the [7] I. Fijalkow, F. Lopez de Victoria, and C. R. Johnson, Jr., “Adaptive
trace operator and the fact that , we can express the fractionally spaced blind equalization,” in Proc. IEEE Signal Process.
Workshop, Yosemite Nat. Park, CA, Oct. 1994, pp. 257–260.
EMSE from (19) in terms of the transformed variables [8] H. H. Zeng, L. Tong, and C. R. Johnson, Jr., “Relationships between the
constant modulus and Wiener receivers,” IEEE Trans. Inform. Theory,
tr vol. 44, pp. 1523–1538, July 1998.
SCHNITER AND JOHNSON: DITHERED SIGNED-ERROR CMA: ROBUST, COMPUTATIONALLY EFFICIENT BLIND ADAPTIVE EQUALIZATION 1603

[9] T. J. Endres, B. D. O. Anderson, C. R. Johnson, Jr., and M. Green, Philip Schniter was born in Evanston, IL, in 1970.
“Robustness to fractionally-spaced equalizer length using the con- He received the B.S. and M.S. degrees in electrical
stant modulus criterion,” IEEE Trans. Signal Processing, vol. 47, pp. and computer engineering from the University of
544–549, Feb. 1999. Illinois, Urbana-Champaign, in 1992 and 1993, re-
[10] I. Fijalkow, A. Touzni, and J. R. Treichler, “Fractionally spaced equal- spectively. Since 1996, he has been pursuing the
ization using CMA: Robustness to channel noise and lack of disparity,” Ph.D. degree in electrical engineering at Cornell
IEEE Trans. Signal Processing, vol. 45, pp. 56–66, Jan. 1997. University, Ithaca, NY, where he has received the
[11] D. R. Brown, P. Schniter, and C. R. Johnson, Jr., “Computationally 1998 Schlumberger Fellowship and the 1998–1999
efficient blind equalization,” in Proc. 35th Allerton Conf. Commun., Intel Foundation Fellowship.
Contr., Comput., Monticello, IL, Sept. 1997, pp. 54–63. From 1993 to 1996, he was employed by Tek-
[12] I. Fijalkow, C. E. Manlove, and C. R. Johnson, Jr., “Adaptive fraction- tronix, Inc., Beaverton, OR, as a Systems Engineer,
ally spaced blind CMA equalization: Excess MSE,” IEEE Trans. Signal where he worked on signal processing aspects of video and communications
Processing, vol. 46, pp. 227–231, Jan. 1998. instrumentation design, including algorithms, software, and hardware archi-
[13] O. Macchi, Adaptive Processing. New York: Wiley, 1995. tectures. His research interest is signal processing for communication systems,
[14] R. M. Gray and T. G. Stockham, Jr., “Dithered quantizers,” IEEE Trans. especially blind adaptive equalization.
Inform. Theory, vol. 39, pp. 805–812, May 1993.
[15] N. Holte and S. Stueflotten, “A new digital echo canceller for two-
wire subscriber lines,” IEEE Trans. Commun., vol. COMM-29, pp.
1573–1580, Nov. 1981.
[16] M. Bonnet and O. Macchi, “An echo canceller having reduced size word
taps and using the sign algorithm with extra controlled noise,” in Proc.
Int. Conf. Acoust., Speech, Signal Process., San Diego, CA, Mar. 1984, C. Richard Johnson, Jr. was born in Macon, GA,
pp. 30.2.1–30.2.4. in 1950. He received the Ph.D. degree in electrical
[17] P. Schniter and C. R. Johnson, Jr., “The dithered signed-error constant engineering with minors in engineering-economic
modulus algorithm,” in Proc. Int. Conf. Acoust., Speech, Signal Process., systems and art history from Stanford University,
Seattle, WA, May 1998, pp. 3353–3356. Stanford, CA, in 1977.
[18] W. Chung and C. R. Johnson, Jr., “Characterization of the regions of He is currently a Professor of electrical engi-
convergence of CMA adaptive blind fractionally spaced equalizers,” in neering and a Member of the Graduate Field of
Proc. Asilomar Conf. Signals, Syst., Comput., Pacific Grove, CA, Nov. Applied Mathematics at Cornell University, Ithaca,
1998. NY. His research in adaptive parameter estima-
[19] J. R. Treichler,, M. G. Larimore, and J. C. Harp, “Practical blind tion theory with applications in signal processing,
demodulators for high-order QAM signals,” Proc. IEEE, Special Issue communication systems, system identification, and
on Blind System Identification and Equalization, vol. 86, pp. 1907–1926, digital control has been supported by the National Science Foundation, the
Oct. 1998. Engineering Foundation, the National Aeronautics and Space Administration,
[20] P. Schniter and C. R. Johnson, Jr., “Dithered signed-error CMA: The Tellabs Research Laboratory, MOOG Technology Center, United Technolo-
complex-valued case,” in Proc. Asilomar Conf. Signals, Syst., Comput., gies Research Center, Lucent Technologies, and Applied Signal Technology.
Pacific Grove, CA, Nov. 1998. His current research interest is in adaptive parameter estimation theory useful
[21] D. L. Duttweiler, “Adaptive filter performance with nonlinearities in the in applications of digital signal processing to telecommunication systems. His
correlation multiplier,” IEEE Trans. Acoust., Speech, Signal Processing, principal focus in the 1990’s has been blind linear equalization for intersymbol
vol. ASSP-30, pp. 578–586, Aug. 1982. interference removal from received QAM sources.