IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. COM-20, NO.

5 , OCTOBER 1972 857

Dynamic Programming TrainingPeriod

for an MSE Adaptive Equalizer

Abstract-This paper concerns itself with the design of an algorithm n(t)

that wilI shotten the training period and adaptation timeof an adaptive p(t) Time Dispersive
equalizer. Time-invariant or slowly varying channels with white additive Channel
Gaussian noise are considered. An adaptive equalizer in the form of a
nonrecursivetransversal filter reduces the intersymbol interference.
The training period consists of the transmission of isolated pulses be-
tween which the equalizer is adjusted. mk
. . .
The algorithm uses a minimum mean-squared error criterion with a TC Calculator
variable step size on each iteration. A fixed number of iterations is fk Switches close every
allowed for the error to be minimized. A constraint related to the ( L + N - I ) T seconds

average excess mean-squared error is included, and the set of step sizes
isdetermined by invoking the principle of optimality in dynamic
programming.
I B
The resulting algorithm iscompared to the popular fixed step-size
algorithm. Predicted and experimental results are given. A fairly well (b)
conditioned and a poorly conditioned channel are considered. Results Fig. 1. Adaptiveequalizer for slowlyvaryingchannels. (a) Channel
show that the new algorithm has a faster adaptation time. It is more model. (b) Adaptive equalizer.
complex than the fixed step-size algorithm, but for longtransversal
filters requires little additionalkmputation time. training period algorithm that will shorten the training period.
For the .two channels considered, the training period has been
I. INTRODUCTION found to be considerably shorter than that of the commonly
used training period algorithm.
T. HEPROBLEMof transmitting information over an un-
known channel in the presence of intersymbol interference
and additive noise has been investigated for many years [ 11-
The channels t h a t we consider here are either time invariant
or areslowlyvarying with respect to the signaling rate. The
[3]. Recently,aconsiderable amount of attention has been channelthereforeexhibitsadelay spread butno Doppler
given to the adaptiveequalizer as a solution to the problem spread. In addition, white additive Gaussian noise is assumed.
[4]-[6], [lo], [12], [13], [lS]-[18]. The basic operation An adaptive equalizer in the form of a nonrecursive trans-
versal filter is used to reducetheintersymbolinterference
of the equalizers is the same for many proposed systems. A
nonrecursive digital filterwith variable tap gains learnsthe caused by the time dispersion. Each tap of the filter is delayed
T seconds from the previous tap, The tap weights { C i } areN
characteristic of thechannel. Thetap gains are adjusted to
in numbkr arid are variable. Fig. 1showsthechanneland
minimize some error criterion. When the filter has reduced the
Adaptive equalizer that we consider here. The channel output
error criterion sufficiently, this period of operation, called the
samples at t = kT, ?nk = m(kT), are assumed to be the sum of
learningortrainingperiod, is terminatedand reliable com-
the signal samples xk and noisesamples nk. The samples
munications may take place. The equalizer either then tracks
y k = y ( k T ) are the equalizer output samples. These samples
slow changes in the channel characteristics or remains fixed at
make up vectors m and y. If we assume that the channel is
the last tap gain settings, and periodically enters a short train-
dispersive by L T seconds, then mk = 0 for k < 1 and k > L ,
ing period to update the tap gains.
a n d y k = O f o r k < l a n d k > L + N - 1.
Clearly, there is a delay in data transmission proportional to
The training period consists of the transmission of isolated
the length of the, training period, and a decrease in this delay
pulses p ( t ) of duration T seconds.Each pulse is received
is desirable. Thispaperconcernsitselfwiththe design ofa
smeared out in time and is passed through the equalizer. Be-
fore thenext puiseis transmitted,theequalizer is adjusted.
Pa' er approved by the Communication Theory Committee of the The transmitted level vector corresponding to the L + N - 1
IEEB Communications Society for publication after partial presentation
at the 1971 IEEE Decision and Control Conference, Miami Beach, Fla., equalizer output samples is a. Clearly, this vector is composed
December 15-17, 1971.Manuscript.receivedSeptember 20, 1971; re- of zero elements, except for one element that corresponds to
vised February 29, 1972. It is based in part on a portion of a disserta-
tion by S. H. Richman in patial fulfillment of the requirements for the the transmitted pulse amplitude.
Ph.D. degree, at the Polytechnic Institute of Brooklyn, Brooklyn, N.Y., When the equalizer output is approximately a, the training
and was supportedinpart byNASA Grant NGR-33-006-040and
Signatron, Inc., Lexington, Mass. period ends. Two measures of the difference betweeny and u
s. H. Richman is with Signatron, Inc., Lexington, Mass. 02173. have beenpopularlyused:absoluteandsquarederror. We
M. Schwartz is with the Department of Electrical Engineering, Poly-
technic Ibstitute of Brooklyn, Brooklyn, N.Y. 11201. choose squared erroras our performance criterion.

Authorized licensed use limited to: BEIJING UNIVERSITY OF POST AND TELECOM. Downloaded on March 22,2023 at 03:53:56 UTC from IEEE Xplore. Restrictions apply.
 858 IEEE TRANSACTIONS ON COMMUNICATIONS, OCTOBER 1972

If matrix M of dimension (L t N - 1) X has the form vert directly. The equivalent set of equations

... 0
QC opt = g (8)
0.. . 0 is therefore iteratively solved for C opt.
The iteration algorithm that we propose to use to solve (8)
is of the form

... c.+ I = c . -ai- vci (e2 >i

I (9)
ml I 2

where .the subscript i indicates the ith iteration, Vci is the

M=
(') gradientoperatorwith respect to thetap settings ontheith
iteration, e; is the mean-squared error on the ith iteration, and
ai is a variable step size. In most applications, the step size is
... mL-N+2 heldconstant.Thisequalizer is called thefixed step-size
equalizer (FSE). Having a variable step size speeds upthe
training period [ 101 , [ 151 , [ 161. However, the equalizer be-
comes more complex.
It is standardprocedure to substitutethegradientofthe
squared error for that of the mean-squared error in the algo-
. * e 0 mL . rithm since e; is not available [6], [8].
By straightforward means, the gradient is found to be
then the equalizer output error for anisolated pulse is given by
e=MC- a (2)
and the algorithm that we shall use has the form
where the components { ck }f= are the weights of the trans.
versal filter. The squared error is therefore
€2 = CTMTMC - i C T M T a t a' (3) The problem thatwe now turn to is the selection of the aion
each iteration.
and the mean-squared error becomes
11. DYNAMIC PROGRAMMING EQUALIZER
e2 E
{ eE2 } = C T Q2CC- T g t a 2 (4) We shall use dynamic programming [9] as a tool for select-
where ing thestep sizes { a i } . Ratherthan having an unrestricted
training period, we shall allow only P iterations for the channel
Q=E{M~M}=R+Z to be equalized. It is required that after P iterations, the mean-
squared error be minimized. In addition, the constraint
R = XTX
P-1
.Z = E { N T N } a:<E
r= 0
g= XTa. (5)
is imposed. The constraint was originally introduced for ease
Both X and N have the same form as M,except that they are of solution of the dynamic programming relationships. How-
filled withsamples xk and nk, respectively. Matrices R and ever, it has been found that E is a constant proportional to the
2, and hence Q , are real, symmetric, and normal. It has been average excess mean-squared error.' Thus, the constraint may
shown [6] that Q is alwayspositive definite if the input se- be interpreted to be a constraint on the average excess mean-
quence has finite energy. In fact, Q is a Toeplitz matrix [7]. squared error.
The mean-squared equalizer output error is therefore a convex Let us define the cost functionJ
function of the tap weights, and there exists a unique mini-
mum error

chin = a' - gTC opt (6)

which we wish to minimize with respect to the selection of the
at the point where the gradient of e' is zero. It is well known {ai}:::. In (13) we observe that the Pth squared error is to
[6] , [8] that this occurs when the optimum tap weights C opt beminimizedsubject to the constraint and that X p is a La-
are chosen according to grange multiplier.
If we apply the principle of optimality in dynamic program-
C opt = Q-lg. (7)
'The average excess mean-squared error is defined to be the average,
In practice, C opt is not obtained from (7). Q ordinarily has over P iterations, of the difference between the actual squared error
a large dimension and requires too much computation to in- and the minimum possible mean-squared error.

Authorized licensed use limited to: BEIJING UNIVERSITY OF POST AND TELECOM. Downloaded on March 22,2023 at 03:53:56 UTC from IEEE Xplore. Restrictions apply.
RICHMAN
PERIOD
PROGRAMMING
TRAINING
ANDDYNAMIC
SCHWARTZ: 859
y *‘

ming [9] , [ 1 11, and we define the functionals problem by using the average step size on each iteration:

b f = E (ai}. (22)
The simulation results show that this assumption is valid. The
use of bf provides an upper bound on thesystem performance.
The true performance results when b: = ai for all i. Since we
where we have taken use the average stepsize,theactual convergence should be
P-1 much better than the boundswill indicate.
a; 0, It is found [lo] that the convergence of the tap gains of the
r=P DPE is given by2
we findthe recursive relationship forthefunctionals to be

In Appendix A we show that the “optimum” step size on the

rth iteration is given by
a, = ETA,EJAp (16)
with
A , =M ~ M $ .
In Appendix B we show that the Lagrange multiplier may be
determined from
respectively.
AP = [;max (E,TAo~ o ) I [e&in/(e:)avI
~ I (17)
Two typical communications channelswere used to compare
A
where pmax is an estimate of the maximum eigenvalueof Q the FS and DP equalizers. One channel was fairly well condi-
[defined in (5)] and (e:),, is the average excess mean-squared tioned (i.e., the condition number R, = pmax/pmin was small)
equalizer error. while the second channel was moderately poorly conditioned.
The training algorithmof the dynamic programming equalizer The channels had a characteristic parabolic delay and ampli-
(DPE) is therefore specified by (1 l), (16), and (17). tude ripple a,. The isolated training pulse was raised cosine in
shape. The input time sequence to the equalizer is therefore
111. PREDICTED RESULTS the sampled inverse Fourier transform of the product of the
spectra of thepulseand chdnnel plus the additivechannel
For ease of comparison, the predicted equalizer behavior for
noise. The product of the spectra, [ 101 is given by
the DPE and FSE equalizer is based on the tap gain error. This
error is defined to be T { 1 +a, cos ( 2 o T ) ) exp - j(ot- 03T3/7r2),
qiECi-E(C,} (18) O < o < n/2T
where E { C, } = C opt. For a fixed step size of ai = A, if we T
define - {l + sin ( U T ) } 11 + a r cos ( 2 w T ) )
2
PA I1 (1- AQ) II (19) 37r
. exp- j(wt- 03T3/r2), n/2T< u < -. (26)
where 11 11 implies the norm, the convergence of the FSE as a 2T
function of P is [6] The degree of intersymbol interference and thus the condition
of the channel is controlledbytheamplituderipplepa-
rameter a,.
The two channels we consider have conditionnumbers of
For the fastest convergence, Gersho [6] shows that R, = 3.28 and R, = 17.8. Thetime-dispersed pulse contains
almost all of its energy in only seven samples. Therefore a 17-
A = 2/(Pmax + Pmin) (21) tap equalizer has been chosen.
Predicted convergence andnormalizedstandarddeviation
where pmax and &,,in are the maximum and minimum eigen- curves are shown in Figs. 2-5 for both channels. Also shown
values of Q. in Figs. 2 and 3 are convergence simulation results. These will
The evaluation of a bound on the convergence of the DPE is
an extremelydifficulttask, since thestep sizes arerandom
2These bounds were evaluated by forming the indicated matricesand
variables thatdependuponthe channelnoise,intersymbol determining the norms using a direct iterative [ 141 norm approach. It
interference,andthestateoftheequalizer. We bypassthis was found using test matrices that this methodwas quite accurate.

Authorized licensed use limited to: BEIJING UNIVERSITY OF POST AND TELECOM. Downloaded on March 22,2023 at 03:53:56 UTC from IEEE Xplore. Restrictions apply.
 860 IEEE TRANSACTIONS ON COMMUNICATIONS,OCTOBER 1972

t
2 4 8 10
Moxlrnum No. of Iterotions, P
Maxlmum No. of I t e r a l l o n g P
Fig. 2. Tap gain convergence. Fig. 3. Tapgain convergence.

be discussed in the next section. As can be seen, it is predicted

thatthe DPE converges fasterthan the FSE;however, for
small P, the tap gainvariance may be slightly larger for the
DPE. The same result has been found in [ 151 and [ 161 for a
Chebyschevpolynomialequalizer. For the' two channels,a
30-dB SNR [12] was assumed, and 2 was determined using
(21) with estimates of pmax and pmin.

RESULTS
IV. SIMULATION
A computer program was written to compare the DP and FS
equalizers. An IBM 1130 computer wasused with at least
P < 10 iterations. Figs. 2 and 3 show the simulation results of
tap gain error convergence for both channels, while Figs. 6 and
7 showthestandard d e ~ i a t i o n . ~In all cases, thepredicted
trends are evident in the simulation results. In fact, the DPE
converges faster than predicted by theaverage step-size assump-
tion. The simulated tap gain standard deviation appears to be
hlgher for the DPE. This was also found in [ 151 and [ 161 for
the Chebyschev equalizer. Jt turns out that the equalizer out-
put error is a shallow function of the tap gains, and has the
same standard deviation for the DP and FS equalizers.
Since the physical quantitythat is mostimportantwith
respect to detection and probability of error is the equalizer
2
10-20 5
IO 15 20 25 30
Maximum No. of lteratlons, P
output error, we present these simulation results [13]. In all
cases, wehave constrainedthe averageexcess mean-squared Fig. 4. Predicted standard deviation bound.

error to be approximately 2egin. In addition, wehave indi-

31t is not possible to superimpose the predicted tap gain standard de-
viation curves on the simulationresults. The predictedcurves have been cated ernin by dashed lines.
normalized by the factor E { hTh ). where h = QTE
{c, } - - NT~ For the well-conditioned channel, we see in Fig. 8 that the
hasreached the convergence region afteronly five
Since E {hTh } is unknown, we cannot plot the ratio of the predicted
and simulated tap gain standard deviation. iterations, while afterteniterationstheFSE still hasnot

Authorized licensed use limited to: BEIJING UNIVERSITY OF POST AND TELECOM. Downloaded on March 22,2023 at 03:53:56 UTC from IEEE Xplore. Restrictions apply.
RICHMAN AND SCHWARTZ:DYNAMICPROGRAMMINGTRAININGPERIOD 861

-
h? 17.8; = 164
9- Pmin Pmln
hx17.8; = 164.
Pmin mln SNR = 30 dB
8-
. SNR = 30 dB 8= 0.0974
7- = 0.0974

6- -

5-
002 -

4-
- -------

o o '2 l l I '6 l lE l l10 l J

M a n m u m No, of Ilerotlons. P ,

Fig. 7. Simulation-SD { I q I }.

DP Dynamlc Programmlng
A FixedStepSize
I

) - l I l I I I I I l I I I
0 2 4 6 8 10
Moxlmum No. of lterottons. P

Fig. 5. Predicted standard deviation bound.

A _2_

0 0 ' "2 " " " 8 10

Maximum N o . of Iterations. P

Fig. 6. Simulation-SD { I q I }. Fig. 8. Comparison of FSEDPE

and performance.

reachedthis region. The vertical markson each curve mark it is expected that the DPE will also outperform the FSE. The
off *1 standard deviation of output error. The results for the standard deviation of the output error was found to be rela-
DPE aresignificantly different from those for the FSE, even tively equal for bothequalizers and both channels.
when the standard deviation is included.
Forthepoorlyconditionedchannel,the DPE hasentered V . REMARKS
the convergence region after four iterations, as shown in Fig. 9. With the inclusionofadditivenoise,aconstraint on the
However, after 11 iterations, both equalizers are close in per- averageexcess mean-squarederror,andlimitingthetraining
formance, and as shown by the vertical marks, are not signifi- period to P iterations, we have found a dynamic programming
cantlydifferent. Still, however,thetrainingperiodwould algorithm for adjusting the tap weights of an adaptive equalizer
have been terminated after five iterations had the DPE been during its training p e ~ i o d . ~As compared to the FSE, a con-
used. Forthetwochannelsconsidered,itappearsthatthe
DPE can be used to speed up the training period. For channels 4The equalizer may then continue to adapt to andtrack channel
whose condition falls between the two extremes we have used, changes that are slow compared t o the convergence time.

Authorized licensed use limited to: BEIJING UNIVERSITY OF POST AND TELECOM. Downloaded on March 22,2023 at 03:53:56 UTC from IEEE Xplore. Restrictions apply.
 862 IEEE TRANSACTIONS ON COMMUNICATIONS, OCTOBER 1972

clL-L+Lu
00 2 6 8 IO .
Fig. 10. Tap controller and DPE step-size calculator. (a) Tap controller.
(b) DPE stepsize calculator.

equalizer was, in general, higher than that for theFS equalizer

for small iteration numbers. In addition, the simulationresults
Maximum No. of Iterations, P
showed that the tap gain variance was lowest for the FSE, but
Fig. 9. Comparison of FSE and DPE performance. that when the variance of the equalizer output error was con-,
sidered, both the new equalizers and the FSE equalizers had
siderableimprovement in convergence of equaliieroutput comparable variances. Therefore,theseequalizersneedonly
error was possible. The price that we must pay, however, is a be compared with respect to the convergence of their output
more complex equalizer, able to determine the required step error. In the work presented here, the equalizer output error
sizes given by (16). We discuss the complexity in the follow- has been normalized to an initial error of unity. This has not
ing. It was also found that the variance of the equalizer error been done in [ 151 and [ 161. However, ifwe do normalize
was the same foreachequalizer.Thepredictedresults also this error, we find that for the two channels considered, both
showed that the average convergence of the DP equalizer was equalizers converge at about the same rate. The Chebyschev
better than that of the FS equalizer. It was observed that the equalizer, however, is slightly faster than the DPE. We would
use of the average step size in the DP equalizer gave a very expect the DPE to be slower since a constraint has been in-
loose upper bound to actual system performance. cluded.Theoretically,theChebyschevequalizerpredictsex-
Referring to (17), we observe thattheconstraintonthe tremelyfast convergence. Accurate eigenvalue estimates are
average excess mean-squared error affects the magnitude of the required to achieve this rate. This accuracy is not achieved in
Lagrange multiplierandhencethestep size. The DPE con- [ 151 and [ 161 and convergence was slowed considerably.
vergence cantherefore be changedbychoosing a different Inarecentstudyofthefeedbackequalizersdescribedin
constraint. We have foundthattheconstraintchoice is not [13] and [ 171, it was found that for a particular two-path flat
very critical.Althoughthereshould be an optimumchoice, fading channel, the equalizer output error was lower for the
. this seems to beadifficultproblem to solve. feedback equalizer than for the FSE. Convergencetime was
The two channels considered are typical telephone channels. muchgreater [18]. However, withjudiciousplacementof
Since theperformanceresultsshowimprovementfor both forward and feedback taps, the convergence time for the feed-
well- and poorly conditioned channel states, it is anticipated back equalizer could be reduced to less than that for theFSE.
that the DPE will show improvement for intermediately condi- In this study, it was decided that the speedup techniques for
tioned channels. forward equalizers
could be
easily applied to feedback
In the very recent literature, several other authors have di- equalizers to achieve low equalizer output errors after a short
rectlyandindirectly suggested methods of decreasing the training period.
length of the training period [ 151 -[ 181. It is not the purpose As a final comment, we indicate the increased complexity of
of this paper to compare all of these methods. We do, how- the DPE. For the FSE, all that was required to adjust the tap
ever, compareourresults to that for the Chebyschevpoly- gains was the correlation of the equalizer output error and the
nomial equalizer [15], [16]. Both this and the work on the signal at each tap of the transversal filter. This is the gradient
Chebyschevequalizer have used the same channelswitha in (20). In the DPE, we mustdeterminetheinnerproduct
mean-squarederrorcriterion.In both of the studies, it was eTA,e,. Thisinnerproductmayberewritten as [$Mr] *
foundthat the predicted variance boundfortheproposed [MTe,], which is not any more complicated since it is in terms

Authorized licensed use limited to: BEIJING UNIVERSITY OF POST AND TELECOM. Downloaded on March 22,2023 at 03:53:56 UTC from IEEE Xplore. Restrictions apply.
RICHMAN AND SCHWARTZ: DYNAMIC PROGRAMMING TRAINING PERIOD 863

of the gradient that is readily available. In the FSE, the gradi- APPENDIXA
ent was multiplied by a fixed constant. In the DPE, the gradi- DYNAMIC PROGRAMMING STEP SIZES
entmust be multipliedbya constant times itssquared The recursive dynamic programming relationship that we use
magnitude. Theadditionalhardwarerequired is aset of N to determinethe step-size relationship is given by (15) and
squarers, an accumulator, and a multiplier. Evidently, if the repeated here:
FSE can operate in real time, so can the DPE.
fp-s(fs)= min [ A+pf 4p q s + l ) )I.
The tap controller of Fig. l(b) is shown in Fig. lO(a) for the { “ r 1% \
rth tap and is the same fortheDPandFS equalizers. The
Starting at the last iteration s = P - 1, we begin to determine
present equalizer output error multiplies the contents of the
the step sizes. Let
rth tap cell. Thisprocedure is continuedfor (L + N - 1)T
secondsuntiltheisolatedpulsehasclearedtheline.Each Fp-1 = Ap&$-l + ~2 p ; (27)
product is accumulated in G,. After (L + N - l ) T seconds, then
the switch closes and G,, now equal to the rth component~,of f l ( ~ p -=~min
) Fp-l. (28)
V e 2 , entersthe step-size calculatorandthemultiplier. T% “P-1
step size is calculated as described in the following, and also
Using the approximation’
entersthemultiplier. The resultant product is subtracted in
the tap accumulator to form the newest update of the 7th tap € i + l = (I- ( ~ i A i€i) (29)
gain. At the start of equalization, the tap gain accumulator is where
set with an initial value of C,. Every (L + N - l ) T seconds the AiWiM?, (30)
gradient accumulator is reset to zero.
The above is requiredin boththeFS andDPequalizers. we find by substituting (29) into (27) and by differentiating
Only the step-size calculator is different. For the FSE, a fixed FPbl with respect to ap-p-lthat
step size equal to A is placed in the step-size calculator. It is T T 2
ap-p-1 = f p - 1 A p - 1 € p - l / ( A p + ep-1 AP-1 ep-1). (31)
this constant value that is outputed every (L + N - l ) T
This is the optimum step size to use for the last iteration, as-
seconds. In effect, the step-size calculator is a memory with suming that P - 1 iterations have already been performed. In
value equal to the chosenstep size. Thestep-sizecalculator order to proceed to determine the remaining step sizes, it is
for the DPE is shown in Fig. 10(b). In this case, each of the N necessary to write f l ( e p - l ) as afunction of theoptimum
gradientcomponents is squared,thenaddedtogether,and ~ . matrix (I - ap-l A P e 1 )becomes
o ~ p - The
scaled by l/Ap where Ap has been predetermined. The result is
the variable step size a . Theincreasedcomplexitytherefore
consists o f N squarers, an accumulator, and a multiplier.
Per sample interval, both equalizers require N multiplies and
adds to accomplish the convolution of input signal and filter
response. In addition, N multiplies and adds are required to
formthegradientoftheequalizererrorandoneadd is re-
quired to form the output error. After (L + N - l ) T seconds,
2 N ( L + N - 1) multiplies and ( 2 N + 1) (L + N -‘1) adds have
have beenperformed.Bothequalizersalsorequire N multi-
plies to scale the gradient by the step size and N adds to up-
date the tap gains, The DPE requires N additional multiplies
and adds to form the squared magnitude of the gradient andN
multiplies’ to scale thestep size bythe Lagrange multiplier.
Thetotal
number of
equalizer
operations is therefore
2N(L + N - 1/2) multiplies and 2N(L + N - 1/2) - 1/2 adds
for the FSE, and 2N(L + N + 1 / 2 ) multiplies and ( 2 N + 1) .
(L + N ) - 1 adds for the DPE. In terms of multiplies and adds,
the total number of operations for the FSE is approximately
2N(L + N - 1 / 2 ) and for the DPE 2N(L + N + 1/2). Actually,
the DPE equalizer requires ( N - 1) fewer adds. Ordinarily,N
is on the order of L , so that both equalizers require approxi-
mately 4 N 2 operations. The difference in the number of DPE
operations as compared to the FSE is 1/(L + N - 1/2), which
goes tozerofor large N. Fortheequalizers that wehave
simulated, L + N - 1/2 is 33.5, which meansthe DPE re-
quired 3 percentmoreoperationsthantheFSE.This is a T
SThe exactrelationship is given by ~ i =+ (Mi+,Ci
~ - a ) - c~iMi+~M q .i
small price to pay for
thespeedup of equalizer error I t has been found [ l o ] , however, that for signal-to-noise ratios of 10 dB
convergence. or more, (29) is a good approximation.

Authorized licensed use limited to: BEIJING UNIVERSITY OF POST AND TELECOM. Downloaded on March 22,2023 at 03:53:56 UTC from IEEE Xplore. Restrictions apply.
 864 IEEETRANSACTIONS ON COMMUNICATIONS,OCTOBER 1972

and that the stepsize to use on the final iterationis R. Bellman, Dynamic Programming. Princeton, N. J.: Princeton
T Univ. Press, 1957.
CXp-1 = Ep-1 A p - 1 E p - l l h p . (39) S. H. Richman, “Dynamic programming in adaptive equalization
of digital communication channels,”Ph. D. dissertation, Dep.
Clearly, since f l ( f p - , ) is of the same form as fO(+), the ex- Elec.Eng., Polytechnic Inst. Brooklyn,Brooklyn, N. Y., June
pression for the step size at the(P- 1)st iteration must be 1971.
J. T.Tou, Optimum Design of Digital ControlSystems. New
T T 2 York: Academic Press, 1963.
qP-2 = EP-2 AP-2 €P-Zl(XP + EP-2 A P - 2 EP-2). D. C. Coll, “A system for the optimum utilization of pulse com-
munication channels,” Defense Res. Telecommun. Establ., Rep.
The same assumptions are applied at the (P - 1)st iteration, 1168, Dec. 1968.
andfinallyforallprecedingiterations. A pointisreached P. Mmsen,“Feedbackequalizationforfading dispersive chan-
where the assumptions are no longer valid. This is the point nels, IEEE Trans. Inform. Theory, vol. IT-17,pp.
56-64,
where the set of step sizes deviates from a true dynamic pro- Jan. 1971.
L. Fox, An Introduction toNumerical LinearAlgebra. New York:
gramming solution. It appears, however, that the assumptions Oxford Univ. Press, 1965.
are true after a very small number of iterations. The DPE re- T. J. Schonfeld and M. Schwartz. “A raDidlv converging first-
sultsshowthatevenwiththeseassumptions,the DPE out- order training algorithm for an adaptive equalizer,” IEkE-Trans.
performs the FSE with respect to convergence. Inform. Theory, vol. IT-17, pp. 431-439, July 1971.
-, “Rapidly conver$ng second-order tracking algorithmsfor
The strategy of choosing the step sizes will be adaptiveequalization, IEEE Trans. Inform.Theory, vol. IT-17,
pp. 572-579, Sept. 1971.
= ETA , er/Xp. (40) D. George, R. Bowen, and J. Storey, “An adaptive decision feed-
back equalizer,” IEEE Trans. Commun. Technol., vol.COM-19,
pp. 281-293, June 1971.
APPENDIX
B R. W. Chang, “A new equalizer structure for fast start-up digital
communication,” Bell Sysf. Tech. J., vol. 50, July/Aug. 1971.
LAGRANGE MULTIPLIER S. H. Richman, Signatron,Inc., Lexington, Mass., Internal Memo.
The constant E in the constraint equationis f o u n d t o b e[ 101

Steven H. Richman (S’67-M’69)was bornin

where is t h e average
excess
mean-squared
error in ex- Brooklyn, N. Y., on April 15, 1945. Here-
8 Substituting (16 ) into (1 2), we find that
cess of emin. ceived the B.E. degree in electrical engineering
from the City College of New York, New York,
in1967, and the M.S. and Ph.D. degrees in
electrical engineering from the Polytechnic In-
r= 0 stitute of Brooldyn,Brooklyn, N.Y., in 1968
and 1971, respectively.
F o r a more conservative estimate o f . h p , we approximate this He has been employedatSignatron, Inc.,
sum b y P ( EA~. e o ) ’ , which is also obtainable before the first Lexington, Mass., as a Consulting Scientistsince
iteration is made. Therefore, Ap is larger and the resultipg step 1969. At Signatron he has pursued his interests
sizes are slightly smaller in magnitude. The estimate of A p that in equalizers and statistical receivers. He has spent a year and ahalf in
we shall use is the study of techniques for characterizing and measuring the under-
water acoustic channel and has written software for the underwater
channelmeasurement. He has designed filters forchannelprober
projects, evaluated new digital FM techniques, and has analyzed other
As a final note, p,, = T r [ Q ] , and to determine Dmin we use modulation techniques such as multichannel ASK/PSK and FSK/PSK.
the infimum of the sampled spectrumof the channel (squared) Dr. Richman is a member of Eta Kappa Nu, Tau Beta Pi, and Sigma Xi.
plus additive noise [ l o ] .

REFERENCES Mischa Schwartz (S’46-A’49-M’54-SM’54-

F’66) received the B.E.E. degree from Cooper
D. A. George,“Matched filters forinterfering signals,” IEEE Union, New York, N. Y., in 1947, the M.E.E.
Trans. Inform. Theory (Corresp.),
vol. IT-11,pp.
153-154, degree from the Polytechnic Instituteof Brook-
Jan. 1965. lyn,Brooklyn, N.Y.,in 1949, and the Ph.D.
D. W. Tufts, “Nyquist’s problem-The joint optimizationof trans- degree in applied physics from Harvard Univer-
mitter and receiver in pulse amplitude modulation,” Proc. IEEE, sity, Cambridge, Mass., in 1951.
vol. 53, pp. 248-259, Mar. 1965. From 1947 to 1952 he was a Project Engineer
M. R. Aaron and D. W. Tufts,“Intersymbolinterference and
errorprobability,” IEEE Trans. Inform.Theory, vol. IT-12, with the Sperry Gyroscope Company, working
pp. 26-34, Jan. 1966. in the fields of statistical communication theory,
R. W. Lucky,“Automaticequalizationfor digital communica- radar detection, and radar design, as a Member
tion,” Bell Syst. Tech. J., vol. 44, Apr. 1965. of the Basic Systems Study Group of the Radar Engineering Depart-
D.
C. Coll, “Adaptive receivers formultipathchannels,”in ment. In 1952 he joined the teaching staff of the Polytechnic Institute
IonosphericRadioCommunications. New York: Plenum,1968. of Brooklyn, where he is currently Professor of Electrical Engineering.
A. Gersho, “Adaptive equalization of highly dispersive channels He served as Head of the ElectricalEngineering Department from 1961
for data transmission: I,” Bell Syst. Tech. J., Jan. 1969. to 1965. During the year 1965 to 1966 he was a NSF Science Faculty
W. F. Trench, “An algorithm for the inversion of finite Toeplitz
matrices,” J. SOC. Ind. Appl. Math., vol. 12, Sept. 1964. Fellow at theLaboratoiredePhysique, EcoleNormaleSupBrieure,
B. Widrow, “Adaptive filters I: Fundamentals,”Stanford Elec- Paris, France. He is the author of various technical papers and several
tron. Lab., Stanford Univ., Stanford, Calif., Tech. Rep. 6764-6, books in the field of communications, including the recently published
Dec. 1966. second edition of Information Transmission, Modulation, and Noise.

Authorized licensed use limited to: BEIJING UNIVERSITY OF POST AND TELECOM. Downloaded on March 22,2023 at 03:53:56 UTC from IEEE Xplore. Restrictions apply.