A Decision-Directed Bayesian Equalizer

Michael E. Wilhoyte Tokunbo Ogunfunmi

Applied Signal Technology Santa Clara University

400 N. California Ave Dept. of Electrical Engineering
Sunnyvale, CA 94086 Santa Clara., CA 9505 1

coefficients. Thus, these type of blind equalizers

ABSTRACT implicitly invert the transfer function of the channel.
This paper investigates a new decision-directed The problems with the Bussgang equalizers are their
Bayesian equalizer based on the blind Bayesian slow convergence as well as their ill-convergence, that
equalizer introduced by Iltis, Shynk, and Giridhar. The is, since the cost function is non-quadratic, there is a
blind equalizer of Iltis suflers @om high complexity possibility that the algorithm will converge to an
making it impractical for sofhvare implementation of undesired local minima. In addition, channel inverting
QAM signals beyond 4 states and with bit rates greater equalizers tend to strike a compromise between
than 16 kbps. Complexity of the decison-directed achieving enough gain in the nulled regions and
Bayesian equalizer is compared to the blind version. It enhancing the noise in those regions. This makes
is shown that even though the decision directed equalization of channels with deep spectral nulls
implementation is less complex, implementation of it in dif€icult. In contrast, the BBE does not invert the
software still seems limited to BPSK and QPSK with bits channel. Instead, a worst case channel impulse length
rates less than 30 kbps. It is also demonstrated through is assumed and all MNb+lpossible symbol sequences
simulation results that even for initially poor channel are formed and fed through channel estimates, where M
estimates, the decision-directed equalizer still is the length of the symbol alphabet and N b + l is the
converges to a close approximation to the actual
worst case length of the channel. The output of the
channel impulse response.
channel estimates are then compared to the incoming
sample which is used to 1) compute likelihood
1. INTRODUCTION
probabilities which are then used to select the most
A decision-directed Bayesian equalizer is introduced. In probable sequence given particular channel estimates,
the following, decision-directed means that the weights and 2) to update the channel coefficients for each
of the adaptive filter are updated based on an internally sequence. The problem with this technique is that for
generated desired response or decision. The maximum every hypothesized sequence, there is a corresponding
a-posteriori (MAP) algorithm was first introduced by channel estimate. This means that the update
Abend and Fritchman 111. It was shown that the MAP computations for the channel estimates grow
algorithm is the optimal symbol by symbol detection exponentially with the channel length and the
scheme assuming that the channel is known apriori. modulation complexity. This paper investigates the
Howver, in most cases, the channel charmeristics are possibility of using one channel estimate for all
not known very accurately and are generally non- hypothesized sequences and then use a decision dlrected
stationary. Recently, Iltis, Shynk, and Giridhar approach to update the channel coefficients. Thus the
[2,3,4,5,6] introduced a new blind approximation to the new algorithm is a decision-directed Bayesian equalizer
optimal MAP algorithm based on Bayesian probabilities (DDBE). Two algorithms are investigated for coefficient
of the channel coefficients. Up until the development updating: the LMS algorithm, and the Kalman filter.
the blind Bayesian equalizer (BBE), there were Section 2 gives a brief description of the BBE. Section 3
generally two classes of blind equalizers: Bussgang-type gives a description of the DDBE. Section 4 presents
equalizers which minimize a non-quadratic simulation results, and conclusions are given in section
performance surface, and blind equalizers based on 5.
higher order statistics. The Bussgang-type equalizers,
such as CMA [7], are “property restoral” type
algorithms that use gradient descent to update the filter

058-6393/96 $5.000 1996 IEEE 325

Proceedings of ASILOMAR-29
 2. BLIND BAYESIAN EQUALIZATION constant so that the mm of the conditional probabilities
This section will outline the derivation of the BBE over all possible symbol sequences is exactly 1.
presented in [2]. A discrete time signal and channel Equation 1 are the conditional probabilities that give
model is assumed rise to the optimum MAP algorithm derived in [11. The
Nb -1 subopbmum Bayesian equalizer operates on a set of
r(k)= Eb,(k)d(k-n)+n(k), truncated symbol sequences denoted by
n=O
where d@) is the current transmitted symbol, r(k) is the
output of the matched filter, and {b,,(k)} are the time-
varying channel coefficients. The symbol r’ is used to which has the same number of terms as the impulse
denote a collection of all the matched filter outputs up to response of the channel. In [2], two versions of the
time k. Thus r = {r(k),r(k- l),...,r(O)). Likewise, algorithm are given; one version uses a Kalman filter to
update the channel caefficients, and the other version
d: = { d i ( k ) , d i ( k - l ) , ..., d j ( 0 ) }which denotes the ith uses the LMS algorithm which is much simpler to
of Adk+’ symbol sequence up to time k. The conditional implement. During this study, the LMS version of the
probabilities of the ith symbol sequence given a received BBE was considered. A biock diagram of the LMSBBE
matched filter output sequence are given by is given in Figure 1. Each element has an associated
channel estimate denoted as 6, ( k ) . Each of the N
possible subsequences are filtered by their
corresponding channel estimate and compared to the
input signal r@). The error signal at stagej is then used
to compute the conditional probabilities which gwe rise
to the current transmitted symbol and is used to update
the channel estimate

2.1 COMPUTATIONALCOMPLEXITY OF BBE

The complexity of the LMSBBE algorithm is given in
Figure 2. In the table below N = AdNb+’. Since a
complex multiply is the same as 4 real multiplies and 2
real adds, the computations are given in terms of total
real multiplies and real adds. Most DSP processors,
(TMS320C50 for example) can do a multiply and
accumulate with data move in one cycle. T h s allows
FIR filtering to be done in L clock cycles, where L =
filter length. In [2], a complexity analysis is performed
and it was determined that for N b = 2 , and employing
a QPsK modulation with a symbol rate of 8 kbaud
would take four 50 ns devices assuming little overhead.

Variable Real Adds RealMult Divides

~~

Figure I : Block Diagram of LMS BBE

Figure 2: Complexity Table For M S B B E

3. DECISION DIRECTED BAYESIAN

EQUALIZER
for i = 42,.. .,A4 ,and where p(d: Irk ) denotes the The motivation for tlus algorithm comes from the fact
probability of d: given r , and c is a normalization that for many digital signals, a channel estimate may be
calculated using the framing structure common in many

326
digital systems. For example, the digital A M P S (IS-54)
and the GSM cellular standards both transmit training Two methods were used to update the channel
sequences or sync bits which have very good coefficients, namely the LMS algorithm, and the
autocorrelation properties. That is, the correlation Kalman filter. The update equations for the LMS
function of these sequences is exactly a delta function algorithm are
for lags up to 5. Thus, ifx(k) is modulated transmitted
discrete time signal which is corrupted by a channel 6(k+1)= 6(k)+pedj”b&*(k), where
h(k), then the received signal is y(k) = x(k)*h(k). A
possible channel estimation algorithm is to employ a ~ ( k=)r(k)- Pj (k) ,
matched filter to the incoming sequence ye), where the and where j is the index which results in the maximum
matched filter coefficients are the conjugated time likelihood estimate of r(k). The Kalman filter
reversed modulated training sequence s(k). Thus when implementation is given by first defining the process
the incoming signal is the modulated training sequence, equation and the measurement equation. Assuming a
we rset random walk for the channel coefficients, we get

s*(-k)*y(k) = s*(-k)*s(k)*h(k) = 6(k)*h(k)= h(k) b(k + 1) = b ( k ) + ~ l ( k ) ,

where {v (k)} is a stationary zero mean noise process

This technique shall be referred to at the cross and whose correlation matrix is q I . Thus the above
correlation technique for channel estimation. The
decision-directed approach would thus initialize the equation may be viewed as the process equation of the
channel estimate with this initial estimate given above channel coefficients for the Kalman filter where the
(or some other estimate), and then further improve this state transition matrix @(k, k + 1) is equal the identity
estimate with an adaptive technique. matrix. The measurement equation may be written as

A block diagram of the decision-directed Bayesian r,(k)=djPNb .b(k)+v2(k), where

equalizer (DDBE)is given in Figure 3.

the measurement matrix C(k) = dj”’””, and the state

vector is b(k) . The Kalman filter algorithm for channel
estimation is summarized below.
b
ddJf”b i r k )

e, (k) 3 i, (k) produces the M L a(k)=r(k)-d;”bb(k-l)

estimatefor r@) 6 ( k ) = 6(k - 1)+ G(k)a * (k)
I
K ( k ) = K ( k , k - 1 ) - G ( k ) . ( d j v N b) H .K(k,k-1)
Figure 3: Decision-Directed Bayesian Equalizer K(k + 1, k) = K ( k )+ q I ,where

Note that there is only one channel estimate 6 ( k )that is

o2is the estimated noise variance, and q is the variance
used for all signal estimates. The error signal which
corresponds to the maximum likelihood estimate for
of the weight noise process.
r(k)is fed back to update the channel estimate. The
3.1 COMPUTATIONAL C O M P L E m Y OF DDBE
output symbols at time k are denoted by 6 ( k ) . In order
The computational complexity of the DDBE using the
for this algorithm to work, an initial channel estimate
LMS algorithm to update the channel coefficients are
must be supplied to bootstrap the system. If an initial
given in Figure 4.
channel estimate is too poor, then the channel
coefficients may not converge. However, it was shown
through simulation studies that even if the initial
channel estimate is poor, the channel coefficients m a y Variable Real Adds Real Mult Divides
still converge. Thus, there is good indication that this 4N(N& +1) 0
method is robust enough for real applications. YY 0 0

327
 AdFNblrk) N(ll+M-1) 33N 0 Two cases were studied: case 1 initialized the algorithm
70 with a fairly good channel estimate obtained from the
2Nb sNb

Figure 4: Complexity Table For DDBE cross correlation channel estimation technique discussed
in section 3.0,and case 2 initialized the algorithm with
To get a clear understanding for how these two a poor channel estimate. Figures 6 and 7 show the
algorithms compare in terms of their computational results for case 1, and figures 8 and 9 show the results
complexity, the tables in Figures 2 and 4 were used to for case 2. In Figures 6 and 8, the solid line is the actual
compute the number of 50 ns devices required assuming channel magnitude response and the dashed line shows
varylng conditions, It was assumed that a multiply takes the initial estimate. Figures 7 and 9 show the channel
one clock cycle as well as an addition. It was also coefficient learning curves using the LMS algorithm
assumed that a divide takes 30 cycles. An overhead of and the Kalman filter. Note that in both cases, the
20% was added to account for lines of code that are not Kdman filter converges faster than the LMS as
directly related to the computations given in the tables expected. Also note that even when the channel estimate
of Figures 2 and 4. Figure 5 gives a plot of the required is rather poor, both the LMS and the Kalman filter
number of 50 ns devices versus the symbol rate converge to a close approximation to the actual channel
assuming a QPSK modulation with a channel length, estimate. However in this case, it takes roughly twice
N + 1 ,equal to three.
~
the number of iterations to achieve convergence. This is
due to the initially poor symbol decisions being made
because of the poor channel estimate. More work should
be done to investigate the convergence properties of the
decision-directed Bayesian equalizer, but a general rule-
of-thumb for convergence is a 10% symbol error rate.

5. CONCLUSIONS
A decision directed algorithm was investigated for the
blind Bayesian equalizer. The advantage of the decision
directed approach over the blind approach is reduced
computational complexity. However, the complexity of
the DDBE still grows exponentially with the channel
length, so its implementation in firmware seems
limited to QPSK and BPSK modulation with multipath
channels of about 2 symbols, and with bit rates of about
Figure 5: Required Number of50 ns Devices Vs. 30 kbps or less. The disadvantage of this approach is
Symbol Rate For QPSK Modulation With Nb = 2 that an initial channel estimate must be provided to
'bootstrap" the algorithm. However, obtaining channel
estimates for certain signal standards is quite feasible,
and as was demonstrated through simulation results,
This graph pretty much corresponds to the calculation
even for initially poor channel estimates, the DDBE
made in [2] which reported 4 devices required for a
symbol rate of 8 kbaud. In the graph, this point is may still converge allowing demodulation of the data.
actually closer to 5 devices which is probably due to the
6. REFERENCES
20% overhead. This graph shows that approximately a
40% savings is acheved using the decision-directed [l] K. Abend, B.D. Fritchman, "Statistical detection for
approach. Note that both graphs are linear as expected communication channels with intersymbol
since the symbol rate is varylng and not the parameters interference", Proceedings of the IEEE, Vol 58, pp.
of the tables 779-785, May 1970.

4. SIMULATION RESULTS [2] R. A. Iltis, J. J. Shynk, and K. Giridhar, "Bayesian

Simulations of the DDBE were performed assuming a algorithms for blind equalization using parallel adaptive
filtering", Personal Copy.
QPSK modulation and with N b = 2 . The channel
model assumed was the following: [3] R. A. Iltis, "Bayesian maximum-likelihood sequence
estimation algorithm for a priori unknown channels and
H ( z ) = l + 0.99exp(jz).z-' +0.3exp(jz/4).z" symbol timing", IEEE Joumal on Selected Areas in
Communications, Vol 10, pp. 579-588, Apr. 1992.

328
[4] R. A. Iltis, J. J. Shynk, and K. Giridhar, “Recursive
Bayesian algorithms for blind equalization”, Proc.
Twenty-FiJh Asilomar Con$ on Signals, Systems, and
Computers, Pacific Grove, Calif.,pp. 710-715 (Nov.
1991).
i* -31
J -I
[5] K. Giridhar, J. J. Shynk, and R. A. Iltis,
‘Bayesiddecision-feedback algorithm for blind f *.
adaptive equalization”, Optical Engineering, pp. 1211-
1223, Vol. 31, June 1992.
8
[6] R. A. Iltis, J. J. Shynk, and K, Giridhar, “Bayesian i*.*.t1..

blind equalization for coded waveforms”, IEEE, pp

0211-0215,1992.
Figure 7: Channel Coeficient Leaming Curves For
[7] J. R. Treichler, and B. G. Agee, “A new approach to DDBE Initialized With Channel Estimate of Figure 6.
multipath correction of constant modulus signals,”
IEEE Transactions on Acoust. Speech and Sig. Proc.,
I. ,
Vol. ASSP-31, pp. 459-472, April 1983. 1
[8] S. Haylan, Adaptive Filter Eheory, 2nd ed.,
RenticeHall, Englewood Cliffs, NJ (1991).

-111.

-a- ’
-0.4 I
.- . 0 .Z e.4

-7:
-21

-a 4 . 4
;v:
-0.1 0 @.a 0.4
Figure 8: Magnitude Response of The Poor Channel
Estimate Compared to Actual

CL.....I CrrfI.l..* .”l....L f.l....

Figure 6:Magnitude Response ofActual Channel Filter

Compared to Estimate Obtained From Cross
Correlation Technique.

Figure 9: Channel Coe@cient Leaming Curves For

DDBE Initialized With The Channel Estimate of Figure
8.

329