Sei sulla pagina 1di 13

IEEE TRANSACTIONS ONCOMMUNICATIONTECHNOLOGY,VOL.

COM-19,

NO.

3,

JUNE

1971

281

An Adaptive Decision Feedback Equalizer

Abstract-An adaptive decisioxi feedback equalizer to detect digital information transmitted by pulse-amplitude modulation (PAM) through a noisy dispersive linear channel is described, and its performancethrough several channels is evaluated by means of analysis, computer simulation, and hardware simulation. For the channels considered, the performance of both the fixed and the adaptive decision feedback equalizers are found to be notably better than that obtained with a similar linear equalizer. The fixed equalizer, which may be used when the channel characteristics are known, exhibits performance which is close to that of the optimum, but impractical, Bayesian receiver and is considerably superior to that of the linear equalizer. The adaptive decision feedback equalizer, which is used when the channel impulse response is unknown or time varying, has a better transient and steady-state performance than the adaptive linear equalizer. The sensitivity of the receiver structure to adjustment and quantization errors is not pronounced.

I. INTRODUCTION ADAPTIVE decision feedback equalizer is described, and it,s performance is discussed in this paper. The equalizer is used to recover a sequence of digits that has been transmitted at a high rate overa noisy dispersive linear communications channel by some linear modulation process. The channelis used efficiently by sending the digital information a t such a high rate that there is intersymbolinterference at the receiver input between several successive digits. The receiver is able to combat both t,he additivenoise and the intersymbol interference, and also to adapt itself to an unknown or slowly varyingchannelwithout the aid of a training digit sequence. Thus it can track a continual slow drift in channel characteristics without interrupting themessage transmission. Past decisions about the digits are used in minimizing the intersymbolinterference by coherentlysubtracting the interference from previously detected digits, and also are used in adapting the equalizer parameters to a change in channel characteristics. It is shown that this receiver is insensitive to quantization of the input signal and quantization and adjustment of its own parameters, and SO can be constructed a t reasonable cost.
Paper approved by the Data Communication Systems Committee of the IEEE Communication Technology Group for publication after presentation at the 1970 IEEE International Conference on Communications, San Francisco, Calif., June 8-10. Manuscript received September 25, 1970; revised December 23, 1970. D. A. George is with the Department of Electrical Engineering, Carleton University, Ottawa, Ont., Canada. R. R. Bowen and J. R. Storey are with Communications Research Center, Department of Communications, Ottawa, Ont., Canada.

AN

The adapt,ive linear transversal equalizer [1]-[3] has been developed in recentyears to accomplish the task outlined previously. With that receiver it has been possible to utilize unknown or slowly varying dispersive channels much more effectively than was possible with fixed lumped-parameter equalizers. Concurrently, however, it has been shown [4]-[7] that the statistically optimum receiver for the recovery of the digit sequence, when the dispersivechannel is known, is nonlinear. At high data rates the performance of this receiver is much better than that possible with the transversal equalizer, which is the opt,imum linear receiver. Unfortunately, the statistically optimum receiver is very complex when there is a large amount of intersymbol interference, and is not practical withtodays technology. This suggests that one seek a statisticallysuboptimum receiver that is practicaland has a performance that is significantly better than that of any receiver that is constrained to be linear. A decision feedbackequalizer, described by Austin [SI, is sucha receiver. It is shown in Fig. 1. This equalizer is not adaptive but an adaptive version may readily be obtained, as is shown in this paper. The decision feedback equalizer is similar to the transversalequalizer in that bot,h haveafiltermatched to the isolated received pulse, followed by a baud-rate tapped delay line. However, it makes use of the fact that at the transversalequalizeroutputthere is intersymbolinterference caused by both undetected digits and previously detected digits. If the previous decisions are correct, they can be used to coherently substract the intersymbol interference caused by t,he previously detected digits. This is done by passing the past decisions through the feedback t,apped delay line. The feedback delay line tap values are chosen on the assumption that these past decisions are all correct. The matchedfilter and the forward tapped delay line are used to minimize the effects of the additive noise and t.he intersymbolinterference from undetected digits. Errorsattheoutput of this equalizer occur in bursts, of course, because a decision error in the feedback delay line tendsto cause yet more incorrect decisions. However, the equalizer is able to recover spontaneously fromthis condition.Simulationstudies show thatthe performance of the decision feedbackequalizercanbe considerably better than that of the linear equalizer even though it,s output errors occur in bursts. I n Section I1 of this paper the decision feedback equalizer is described, and its performance is compared with the performance of a number of other receivers. This com-

282

IEEE TRANSACTIONS ON COMMUNICATION TECHNOLOGY, JUNE

1971

! !
T

fGf.(llD
T
TAPPED DELAY LINE T

Fig. 1. Decision feedbackequalizer.

parison is done both analyt,ically and by digital computer simulat,ion. It is shown by examplethat the decision feedback equalizer is an attractive compromise between what is theoret,ically possible and what is now in use. Next, in Section 1 1 it is demonstrated that the decision feedback 1, equalizer may be made adaptive to an unknown channel. Twodifferent adaptationalgorithmsare described and comparedbyhardwaresimulationstudies. In the cases considered, the adaptive decision feedback equalizer significantly outperformed the adaptive linear equalizer, and a training sequence was not required for adapt,ation. Rat,her, the decisions can be used for adaptation as well as to coherent.ly substract the intersymbol interference. Finally, inSect.ion IV, the practical nature the equalizer of is demonstratedbyshowingthatthenumber of delay line taps that it requires is modest, and that its digital implementat.ionrequires no finer quanbization than the linear equalizer. 11. FIXED EQUALIZER

calculated and manually adjusted after the channel characterist.ics are determined. The equalizer makes the estimate
N
Y

e(j) =
k-0

a(k)y(j

+ k) -

b(z)J(j - I )
z=1

(1)

about e( j ) , the digit t h a t is sent at time t = j T , and then converts this estimate to a final decision e^( j ) with a nonlinear memoryless circuit,. (If the digits { e ( j ) are binary this circuit is a clipping circuit with zero bias. If the digits are m-ary the circuit is an m-output quantizer.) In ( 1 ) y( j k ) is the output of the initial filter a t time t = (j k ) T.One met.hod of choosing the tap values is to adjust for minimization the of the probability that j ) # e( j ) . However, this direct opt,imization difficult is because an analytical expression for the error probability interms of the equalizer tap values is notknown. A practicabie way to optimize the tap values is to choose them such t.hat the output mean-square error E[e2(j)] is minimized, where

+ +

e(

In this section t.he fixed nonadaptive decision feedback e ( j )~ - e (- e ( j > . j> (2) equalizer for a known dispersive channelwill be examined. The error rate of this equalizer is a lower bound on the As shown later, t.his leads to a set of linear equations that error rate of an equalizer t,hat must also adapt to random specify the tap values. Thismethod of optimization is changes in the channei characteristics. The basic assump- also attractive because it canbe used to make theequalizer tion made in deriving the receiver is that the decisions self-optimizingor adaptivetoanunknownora slowly made by t,he receiver as to the transmitted signal samples varying channel. While this optimization does not miniare essentially correct. Given the bit error rate require- mize t.he digit error probability directly, computer simument.s in modern communication systems, this is a valid lation studies [SI have shown that the probability density assumpt,ion. It is furthermoreassumedthattheanalog function of the error e ( j ) is close to Gaussian, and so signal from thecomhunicationschannelhas been de- the two performance criteria are similar. modulated, filtered, and sampled at the digitbaud rate The process of determining the tap values starts with with t,he appropriate phase. Previous work [l], [SI has the evaluation of the mean-square error: indicated the desirability of a matched filter before the E C ~ ~ ( O= E C $ ( ~ - e ( j > ~ ~ l I ) sampler, as shown in Fig. 1. In the approach taken in this N paper, any suitable band-limiting filt,er may be used, at = EC{C a ( k ) y ( j k) the price of some loss in performance. It remains to k-0 determine tap the gains { a ( k ) ; = 0,1, - , N } and { b(m) ; m = 1,2,3, .,M) , as illustrated in Fig. 1. In the M adaptive formof the equalizer thesetaps are automatically - b ( o J ( j- I ) - e(j)121 (3) 1-1 adjusted; for the fixed equalizer the tap gains must be

GEORGE

et al.: ADAPTIVE

DECISION FEEDBACK EQUALIZER

283

where the signal sample


M

In general, it is quite difficult to calculate the digit error probability at the equalizer output.The calculation is (j +k) = e( j + k - qZ(i)+ n( j + k ) (4) particularly difficult because the assumption that all past iO = correctdigits strictly are will, of be violated, course, and the errors may tend to occur in bursts. Nonetheless, some and x ( i ) is t.he value of the impulse response of the linear id& of the improved performance of the decision feedback modulator, the channel, and the initialreceiver filter after equalizer over the transversal equalizer can be obtained a delay iT,and n( j k) is the additive noise at theoutby assuming an ideal equalizer with an infinite number put of the initial filter a t time t = ( j k ) T. The forof taps and a matched filter in an environment of white ward tap a ( n ) is optimum when additive noise with spectral density No. I n this case the equationsfor the optimum tap values of atransversal equalizer are

y(j

+ k ) is

for all k
=

(10)

0,

n = 0,1,..-,N.

(5)

where 6 ( - ) is the Kronecker delta and function

Similarly, the feedback tap b ( m ) is optimum when


N

a E C e 2 ( j ) 1= 2EC( a ( k ) y ( j ab (772) k=O


. M

+k)
- e( j ) )

1-1

6(z)8( j

- z)

.e( j

-m)]

the channelincascade.Themean-squareerror output of this equalizer is

of the

E [ e 2 ( j ) ] = (12) Noc(0).

= - 2 ~ [ e ( )e(j- m)] j = 0,

m = 1 , 2 , . . . , M.

(6)

In the same situation the forwardt,aps of the decision feedback equalizer are given by
01

The M N 1 ( 5 ) and (6) can written a be asset of M N 1 linear,equat,ions with the unknowns a ( i ), i = O , l , - . . , N and b ( k ) , k = 1 , 2 , . - - , M .This is done by order reversing t,he of t,he summation and the averaging for in (5) and ( 6 ) and by assuming that j ) = e( j ) and k ) 3 is zero if k # 0. The resulting that E[e( j ) -e( j equations are

+ + + +

a ( j ) ( + q ( j- 772)
j=O

+ NoS(j

- nz)) = ~ ( ~ ~ ~ >
772

e(

2 0. ( 1 3 )

As before, the feedback tap values are given by (8). If the past decisions (e^( j - 772) ) are all correct, the decision feedback equalizer output mean-square error is

c a ( i ){*(ilk) + dn(k - i).)


i=O

= Z(k), =

E [ e 2 (j ) ]

N o(a (4 )) . 10

k and
N

0,1, * . , N

(7)

b(nl) =
i=O

a ( i ) x ( m + i),

772

= 1,2,.

..,M

(8)

The performances of thetwo equalizerscan nowbe compared by comparing ( 1 2 ) with ( 1 4 ) through the medium of an equivalent received pulse which has samples thei ) at p( sampling instants t = iT. Use of sampled data notation the and z transform is convenient here. The transformed shape pulse P ( z ) is defined to be
w

where &(k - i) is the autocorrelationfunction of the noise n ( t ) a t delay T = ( k - i) T , and \k (ilk)is defined by the equation
i

P(z)
i=O

p(i)z-i.

(15)

\k(i,k) g
2 4

x(Z)x(Z

+ k - i).

(9)

With this notation, the equivalent problem at hand is given by

received pulse for the

In theparticular case where a matched filter used ahead is of the tapped delay lines, the x ( t ) is the autocorrelation function of the impulseresponse of the modulator and channel in cascade, and (7)-(9) become equivalent to those given by Austin [ S I .

where

284

IEEE TRANSACTIONS

ON COMMUNICATION TECHNOLOGY, JUNE

1971

The quant,ity of significance hereis the inverse of the equivalent received pulse, which is defined by

03

q ( r I = 7e-r

BAUD INTERVAL T = I O

and the r ( i ) can be determined by simple long division. Substituting the results of this series of definitions back into (12) and (14)) gives the mean-square error
P e
01

Ere2(j ) ] = No
i=O

r2(i)

(19)
3

for the transversal equalizer and

Ere2(j ) ] = Nor2(0)

(20)

for the decision feedback equalizer. It should. be noted that these results are for an idealized situation, and as such form a lower bound on the actual mean-square error; 20 4.0 6.0 80 0 0 12.0 140 . 160 however, they do allow ready comparisons. For example, EINo. dB for the simple case where p (i) = P exp ( -ir) , i 2 0, Fig. 2 . Comparison of receiver errorprobabilities as functions of signal t o background noise ratio. the advant.age of the decision feedback equalizer is limited to 3 dB, since
I

00

r 2 ( i )= P ( 1
is0

+ exp ( - 2 7 ) )
=

(21)

21 0
pe =
LINEAR EQUALIZER

and
r2(0)

P.

(22)

In contrast, if the background noise spectral density No is small and the actual pulse q ( r ) is a rect(angu1ar pulse of length L, where the int,erpulse period T is aL, 1/2 5 a _< 1, then t.he ratio of the mean-square errors is

5 F

DECISION FEEDBACK

12.0

EQUALIZER

s
0
BAYESIAN DEMODULATOR

00 -30

-20

-1.0
lo

00

10 .

20

30

40

50

6.0

70

TRANSMISSION RATE 3eB d+ -

where AI-a. 1 + p = The ratio becomes very large as a 4 1/2. Thus the advantage achieved by using the decision feedback equalizer depends on the channel impulse response, and can, in some cases, be quite large. The performance of the equalizers were compared also by Monte Carlosimulation of the twoequalizers on a digitalcomputer. I n addition, the statistically optimum or Bayesian demodulator [SI was simulated to determine how close to the optimumperformance were the performances of t.he statistically suboptimum, but much simpler, equalizers. I n one series of simulation tests the isolated received pulse was Are- andtheadditive noise was Gaussian and white with power spectral density No. The message was a sequence of independent binary digits. The measured error probabilit,ies at the outputof the decision feedback equalizer, the linear equalizer, the Bayesian

Fig. 3.

Effect of dat,a transmission rate change on receiver performances.

-1

P2

demodulator, and a matched filter with clipper are shown in Fig. 2, as a function of the signal to background noise level E/No. ( E is the isolat.ed pulse energy, equal to & ( O ) of (11) .) Also shown is t.he error function curve, the performance that one could achieve if there was nointersymbolinterference. I n this series of teststhechannel parameter a was unity.Afiltermatchedto Are- was used as part of both the linear and the decision feedback equalizer. Approximately one hundred errors were observed in making each error probability measurement. A sequence of tests was donewithdifferentnumbers of forward and feedback equalizer taps. It was found that the linear equalizer performance improved as the number of taps was increased to 5 , but a further increase did not appreciably improve the performance. In the same way, 3 forward taps and 6 feedback taps were found for the decision feedback equalizer. (The question of the number

GEORGE

et al.:

ADAPTIVE DECISION FEEDBACK EQUALIZER

285

1 2

1 6

x) 24 28 TIME, NANOSECONDS

32

36

40

44

48

Fig. 4.

Impulse response of coaxial cable PCM channel.

of taps is discussed in more detailinSectionIV.) As expected, it was observed that the errors a t t,he decision feedback equalizer output occurred in short, burst,s. At low signal to background noise ratios, 5 6 dB in thiscase, the bursts occur so frequently that the linear equalizer performance is better. However, a t higher signal-to-noise ratios (SNRs) the improved ability of the decision feedback equalizer to reduce the intersymbol interference is more important than the tendency t.0 produce errors in bursts. It was found t,hat at low error probabilities, wherever the digiterrorprobability was less thanthe performance of t,he decision feedback equalizer, the linear equalizer, and t,he Bayesian demodulator could all be accurately described by t,he empirical formula

COAXIAL PCM CHANNEL

ERROR PROBAEILITY : 1
I6 0

6 ~

I2 0 j 14 O

0 0
10 0

I 500

200

300

400

I 600

T(R)E P[e] = 0.5 1 - erf 2No

11

DATA RATE R , MEGAEITS/SECOND

(25)

Fig. 5.

Effect, of data transmission rate change through channel.

PCM

where R is the data transmissionrate, defined for this proexample to be equal to (aT)-l. q ( R ) may be considered evaluated with a computer Monte Carlo simulation to be the efficiency of the modem, and must, be in the gram. I n t.his series of tests the channel included the range 0 5 q ( R ) 5 1.0. I n all cases q -+ 0 as R -+ and transmitter, a solid coaxial cablewith an air dielectric, q -+ 1.0 as R -+ 0. q ( R ) for thethree demodulators, meas- and a fixed lumped-parameter equalizer. This channel has ured a t digit error probability, is shown in Fig. 3 as no dc response, a 100-MHz bandwidth at the -3.0-dB a funct,ion of R. At all transmission rates the efficiency points,a 240-MHz bandwidth at the -20.0-dB points, of the decision feedback equalizer was greater than that and a 70.0-dB per octave roll-off a t higher frequencies. previous example, where the of t,he linear equalizer and less than that of the Bayesian (Thiscontrastswiththe demodulator. At high transmission rates t.he efficiency of roll-off was 12.0 dB peroctave.) The channel impulse both nonlinear receivers decreased by 4.0 d B when the response is shown in Fig. 4. The nominal data rate through rate was doubled. The efficiency of the Bayesian demodu- this channel without further equalization is 225 Mbit/s. lator was 2.0 dB better than thatof the decision feedback Simulation tests showed that inclusion of a linear transequalizer a t all high rates. I n contrast, bhe efficiency of versal equalizer with a matched filter would allow one to the linear equalizer decreased by 9.0 d B when t,he rate increase the data rate to 400 Mbit/s, but not beyond. was doubled. This difference in rate of efficiency decrease I n contrast, the decision feedback equalizer with a becomes very important, of course, if t,he channel is used matched filter can be used a t 450 Mbit/s with only 6.0 a t very high rates and a t high SNRs. d B more signal strength than that required a t low data The usefulness of the decision feedback equalizer in a rates, and even higher rates if the signal strength is incoaxial cable pulse-code modulation (PCM) link was also creased further. The efficiency q ( R ) of the linear and the

286

TECHNOLOGY, COMMUNICATION ON IEEE TRANSACTIONS

JUNE

1971

decision feedback equalizer for t,his channel example, measured a t digit probability, error is shown in Fig. 5 . Thus both the performance analysis and the simulation results indicate that the decision feedback equalizer performance is considerably better t,han that of the linear equalizer. Moreover, in the examples in which the much more complex statistically optimum demodulator was also evaluated, the decision feedbackequalizerperformance was quite close, to this limiting performance. However, these results cannot be extended to other channels without either a simulation study in each case or the development of an appropriate analysis technique.

_--__-___ADAPTATION STEADY TIME STATE MSE KNOWN CHANNEL PERFORMANCE LOWER BOUND

TIME

Fig. 6. Adaptation to step change in channel impulse response.

tap value error is


d ( k ) 4 a ( l c ) - ao(k),
-b ( - k ) n -

1 1 ADAPTIVE 1. EQUALIZER
I n this section i t is shown that the decision feedback equalizercan be madeadaptiveto unknown or slowly varying channels, i.e., channels in which the impulse response does not change appreciably during the transmission of several hundred digits. The dynamic performance of the decision feedback equalizer, that is, the performance of the equalizer while it is adapting, is described. A method by which the decision feedback equalizer can bemadeadaptivecan be seen from ( 2 ) , ( 6 ) , and (7). For any set of tap values

0,1,...,N
(29)

bo(-k),

k = -l,..*,-M.

1,c.t 11s also define a set, of signals ( z ( j


z(

+ i) ] by
*

+ i)

= y(

+ i),

O,l,.

.,N

d ( j + i),

i = -l,.*.,--M. (30)

Then it, can be shown that if the tap values are in error t,he mean-square error is

+ c c d ( i ) d ( k ) E [ z ( j+ i ) z ( j + k)]
+--M
k=-M

and

(31)

where eo( j ) would be the error if the tap values were all correct.. (The assumption was made that e( j - 171) = e ( j - m ) , 111 = 1,2,. .,M, to derive (31).) Thusthe mean-square error is a quadratic function of the tap gain If e ( j ) of ( 2 ) is replaced by errors,,in the same way that the mean-square error of the of its tap gain Z(j) =7(j)-e(j) (28) linearequalizer is aquadrat,icfunction errors. Because of t,his, thereare no locally optimum hill-climbing adaptation can be by assuming that the decisions are correct, then all the tap gain settings, and a signals in (26) and (27) are available atthe receiver. made to readily converge close to the correct set of tap When t.he error probability is low this substitution does values given by (7) and (8). Of course, tJhecross correlations E[;( j ) y ( j k ) ] and not change thevalue of (26) or (27) appreciably. By Ere(j ) g ( j - n z ) ] cannot be measured exactly in a finite changing the forward tap values byamounts approximately proport,ional to -E[;( j ) y ( j k ) ] , and the feed- time, so any particular sequence of tap adjustments is a sample function of a random process. A Robbins-Monro I back tap values by amounts approximately proportional K to E[;( j ) i ( j - m ) ] , the taps are automatically adjusted procedure [9] would beapplicable if t,hechannel were t o near their optimum values. Thus the forward taps are unknown but t,ime invariant. However, if t,he channel is be caadjusted by means of measurement of the cross correla- slowly varying then the adaptive algorithm must pable of tracking a slowly varying channel and of tion between the error and the input signals, just as for learning the optimum tapvalues for a n unknown chanthe linear transversal equalizer. On the other hand, adjustment of the feedback taps makes use of the cross nel. In that,case a procedure such as the Robbins-Monro correlation between the error and the output signal, i.e., procedure is not applicable, and a compromise between t.he decisions. asmallersteady-statemean-squareerrorandashorter The potential of this type of algorithm can be seen by adaptation time to channel changes is necessary. There are a ,number of adaptation algorithms available, observing how t,he mean-square error E[e2(j ) ] depends on tap value errors. Let ( a ( k ) ; k = 0 , 1 , . - - , N ) and in which the exact details of the algorithm are somewhat {b(Z); Z = 1,2,-* .,MI be theactualtap values, and different,. A typical response of an adaptive receiver to a ( a o ( k ) ;k = O , l , - . . , N } and ( b o ( Z ) ;I = 1 , 2 , . . . , M ) be the step change in the channel characteristics when any of optimum tap values, specified by (7) and (8). Then the these nlgorithms is used is shown in Fig. 6. The mean-

EEDBACK

GEORGE et d.: ADAPTIVE DECISION

287

square error at the receiver output is plotted as a function of the number of baud intervals after a step change in the channel characterist,ics. It is a nonstationary random variable, of course, since it is a function of the equalizer tap values, which in turn are nonstat,ionary random variables. Also shown in Fig. 6 is the, average of many such adaptation curves. The twomost important characteristics of such a curve are the "steady-state" meansquare error, andthe "adaptation time," the timerequired to reach the steady-statemean-square error after a specific change in the channel characteristics. .Theadaptation curves that were obtained in a series of simulation studies are discussed later in the paper, but some general observations may be made now. First', since a cross correlat,ion is being measured, the variance of the square of the signals involved contribute significantly to the measurement error. Particularly when binary signals are involved, there is a notable difference between the input samples .(y( j ) ) and the output decisions j ) ] in this regard. I n particular,

{e(

The tap values are thereby shown to be equally sensitive to adjustment, and this point combined with that of the previous paragraph implies that the feedback taps can be adjusted more quickly and/or more accurately than the forward t,aps or the taps of a transversal equalizer. Thus the decision feedback equalizer would be expected have to a better adaptation performance thanthe transversal equalizer. This has also been observed experimentally, as will be described. The advantages of using the dross correlation between the error and the decisions suggest t.he use of this same measurement for adjustment of the t.ransversa1or forward tap values. Of course, the taps would not converge to the correct value to minimize the error due t.o the additive noise and the intersymbol interference, but in some cases at least the taps converge to a value close t,o the correct value. As these taps may be subject to less error due to the cross correlat.ion measurement., improved adaptive behavior can result,. Suppose we specify the taps { a ( i ) ] by

E [ { @ j( ) - E[@(j ) ] ) 2 ]
when
e ( j ) = AI

(32)

E[e(j)d(j

+ Z)]

0,

0,1,..-,N

(36)

rather t,han by ( 5 ) . Then the tap values are the solution to the equat'ions
N

and
-KiY2(j) -~

rather t.han to (7). Thus adjustment of the forward taps by cross correlation between t'he error and the decisions (The additive noise samples n ( j ) are assumed to be "forces zeros" in the overall transmission characteristic. Gaussian in the calculation,) Certainly then in the case If the equalizer had an infinite number of correctly spaced of strongintersymbolinterference, Cq2((i) largeand taps, specified by (37), the result would be an inverse is so @ (j ) is much less '!noisy" than y2( j ) . Thus an esti- filter. In the limiting situation of no additive noise ahd a ideal equalizer, an equalizer adjusted by the mate of E[e ( j ) e^( j - m ) ] would usually involve less error similarly minimummean-square errorapproach would alsoyield than an estimate of E[e( j ) y ( j f k ) ] . Whether t'his implies that the decision feedback equal- an inverse filt'er. Consequently, it is not surprising thatin izer can adapt more rapidly or more closely to the per- somehigh SNR situat.ionswhereeffectiveequalization formance possible froma fixed optimum equalizer than is being obt.ained, the two methods give similar results. A potential difficulty with this "zero-forcil)g" algorithm can the linear transversal equalizer, andthat t.he feedback taps can adapt more easily than theforward t'aps, depends is that only as many system impulse response zeros can beforcedast'here are taps in the delay line,withone on the sensitivity of thetapadjustments.Thepartial additionalt,apreservedtoforceaunitresponse at the derivatives of ( 5 ) and (6) indicate this sensitivity. Since desiredtime. The overallsystemimpulseresponsecan the absolutesignallevelsare, of course, arbitrary,the over decisions j ) are taken to be f1 and the average signal become large both before andafterthisinterval power E [ y 2 (j ) ] is t'aken as unity. This effectively means which the responseisforced to zero. I n contrast, when that the taps { a( k ) } and { b (1) ) are of the same order of the equalizer is adjusted by the minimum output meanof magnitude. This done, the sensit,ivityof the tap values squareerrorapproach,t.hemean-squarecontribution the tot.al system impulse response is minimized, not just can be evaluated, giving: the responseover anintervalas largeas the equalizer delay line. Note, however, that if the taps of the transversal filter of the decision feedback equalizerare adjusted by the zero-forcing algorithm, adjustment of the taps of = 2 (34) the feedbackfilter will automatical1.ycancel any large impulse response after the main pulse, without causing a large impulse response at an even greater delay. This is the basicideabehind the decisionfeedbackequalizer, = 2. (35) ba.sed on the assumption that the decisions in the feed-

~ ~ ~ 2 ~ ~ ~ 1 1 2 1 = 2{Cq"(i))Z + E"n"jj>I).

a ( i ) z ( i- 2)
i=O

= 6(Z),

1 = 0,1,- * , N

(37)

(33)

e(

288

IEEE TRANSACTIONS ON COMMUNICATION TECHNOLOGY, JUNE

1971

ER -CONTROLLED

CHANNEL

Fig. 7. Hardware simulator.

back delay line arecorrect. Thus only the system impulse under the control of a PDP-5 digital computer was used response before t.he area in which zeros are forced can to do bhis. A block diagram of the simulator is .shown in contribute to the output error. From this and from con- Fig. 7 . The transmitt,edmessage was the pseudorandom output of the m-sequence generator corresponding to the siderations of the errors associated measuring with 2 E[e( j ) 8 ( j k ) ] and E[e( j ) y ( j k ) ] , it is hypothe- poiynomiaIx31+ x28 x 7 x24 217 x16 x9 x* 1. sized that the decision feedbackequalizercan use bhe I n some cases a time-invariant analog filter was used t o cross correlation between the error and the decisions to simulate the channel. The filter output was sampled. at of advantage for adjustment of both forward and feedback the baud rat,e after the addition filtered Gaussian noise. taps. The experimentalresults t,hatare described later I n othertests, a 32 tap 12-bit baudrate nonrecursive digital filter was used to simulate the channel filter. I n substantiate t.his hypothesis. The act,ual adaptation algorithm t.hat was used in t,he this case the additive noise was sampled at the baud rate dispersed signal. I n both cases, experimental invest.igat,ion will now be discussed. AS andthenaddedtothe the composite sampled signal was processed with a 7-bit shown, the transversalorforward tap gain should be changed byanamountproportionalt0 a measure of baud rate digital filter, as shown in Fig. 7. The sum of -E[e( j ) y ( j k ) ] or -E[e( j ) 8 ( j k ) ] , and thefeed- the number of taps in the two nonrecursive filters could backt,ap byanamountproportional to a measure of be as great as eleven, with any division of taps between E[e ( j ) 8 ( j - nz) 1. Actually, rather than taking a fixed the two. These filter tap values were under direct computer control. finite time average of these products and then changing At the beginning of eachadapt,ationtest,, all taps of the tap values, the adaptatlion is done indirectly with an the last forward algorit,hmsimilar to that developed byLucky [ a ] for t,he decision feedbackequalizerexcept adaptation of thelinear equalizer. Thetap values are tap were set to zero, so that the output would be 0 if there were no noise or intersymbol interference. The adaptchanged in t.he following way. ive transversal equalizer was tested in a similar way. 1) An accumulator for the tap is set equal to zero. The digital computer was. used to change the tapvalues, 2) Each t.ime a digit is processed, the product and took the sequences ( y ( j ) } , (e^( j ) } , and (e^( j ) } e^(j)y(j+k)fortheforwardtapa(k),or--e^(j)8(j-m) directly as inputs. This method was used to avoid confor the feedback tap b ( m ), is added to the accumulator; Only t.he signs of e^( j ) , y ( j k ) , and e^( j - m ) are used st,ruction of adaptation circuitry for each tap. As a result, only a few of the binarydigits that were processed by in this calculation to simplify the equalizer synthesis. 3) I t.he accumulator contents exceed a threshold V , the equalizer were used for adaptation processes. The f then thet.ap value is decreased byand step1is repeated. digits that were used are called independentdigits, A because the time between successive observations is long I the contents become less than - V , the tap valueis f timesover which theautocorrelation increased by A and step I isrepeated. I the contents \comparedtothe f functions of e ( j ) and y ( j ) are significant. A specified remain between V and - V , then step 2 is repeated. independent digits, usually 100, were I n t,he alternate procedure previously discussed, the number of these e ^ ( j ) y ( j k ) of step 2 are replaced by e^(j)8( j k ) . processed according to the preceding algorithm to change Both the transversal and decision feedback equalizer were t.he taps. Then 2000 digits were used to estimateE[e2(j ) 1, without changing either the tap values or theaccumulat,or tested using each of, these adaptation algorithms. The adaptive equalizers were evaluated by observing contents. Then 100 more samples were used for adaptaanother measurement of their ability to adapt to an unknown but fixed channel, tion purposes, followed by rather than to a time-varying channel. This was done by E[e2( j ) ] . This sequence continues until it is evident that measuring the mean-square error at the equalizer output the equalizer has reached a steady-state mode of operaas a. funct,ion of adaptation t,ime. A hardwaresimulatortion where t.he trend in mean-squareerror is nolonger

+ + + + + + +

GEORGE

et d.: ADAPTIVE DECISION FEEDBACK EQUALIZER

289

TIME IN MULTIPLES OF l39psec BAUD INTERVAL

Fig. 8. Impulse response of simulated telephone cable channel.

changingwithtime. The results of 50such adaptation runs are then averaged to give a mean adaptation curve. Both the signal samples { y ( j ) } and the equalizer tap values were quantized with a maximum accuracy of 7 bits. (More will be said about quantization accuracy requirement,s in Section IV.) The least significant bit of the tap gain values was changed during adaptation each time the threshold V or - V was exceeded. Thus the adaptation parameter A in these tests is V 0.016. Tests were carried out t,o determine whether the decision feedback equalizer can adapt better than the linear equalizer to an unknown channel, whether the results are valid for a variety of channels, whether use of a learning sequence is necessary or even advantageous, and whether or not use of an estimate of E [ e ( j ) e ( j k ) ] results in better adapt.ation than an estimate of E[e ( j ) y ( j k)] for the forward taps { a ( k )1. The channels that were simulated in these tests included a channel with an exponential impulse response, a coaxial cable PCM channel, and an audio-loaded telephone line. One series of tests was made to compare the performances of the adaptive linear and decision feedback equalizers in an audio-loaded telephone cable system, and to determine the advantage that could be gained by using a training sequence. The channel filter was a lumped-parameter filter designed by Bell Canada to simulate a15 000 ft audio-loaded telephone cable and was terminated in 600 ohms. The impulse responseof the filter is shownin Fig. 8. Binary information was transmitted through this channel at 7200 bit/s. Neither any intentional additive Gaussian noise nor a filter matched to t,he isolated received pulse were used in this series of tests. Because of the resulting mismatch, choice of the third of 11 taps as the main tap minimized the linear equalizer output mean-square error. Similarly, it was found that the decision feedback e,qualizer with 4 forward taps and 7 feedback taps made the best use of the 11 available taps.Theadaptation thresholds were set a t f20 during this series of tests. It

0.3-

,
0.2- 8

UNKNOWN CHANNEL= SIMULATED 15,000 AUDIO-LOADED TELEPHONE L I N E DATA RATE = 7 2 0 0 BPS BINARY

5
00

X.
OoX.
X.

LINEAR EOUALIZER, NO TEST SIGNAL

0.1-

0
0

X* X=

0
P K
W

w 0.05.

/
LINEAR EWALIZER, WITH TEST
0

SIGNAL

a
2

5 : 5 0.03. W
I
0.02

nono

SJ
oooo

DECISIONFEEDBACK EOUALIZER, NOTEST SIGNAL

O 0 0 ~ O ooooooooo o~

DECISIONFEEDBACK EQUALIZER. WITH TEST

d
I

SIGNAL

0.01

500

IO00 1500 2000 NUMBER OF DIGITS PROCESSED

2500

Fig. 9.

Adaptation curves with and without test sequence.

was found that adaptation based on the measurement of E[e ( j ) e ( j k ) ] resulted in a better performance than measurement of E [ e ( j ) y ( j k ) 1. The mean-square errors a t the equalizer outputs, using the former measurement, are shown in Fig. 9 as a function of the number of independentdigits that were processed. As shown, the steady-state mean-square error of the decision feedback equalizer is 5 dB better than that of the linear equalizer. This is consistent with the fixed equalizer tests that were

290
0 180.17-

I E E E TRANSACTIONS ON COMMUNICATION TECHNOLOGY, J U N E

1971

0 160.150.14-

UNKNOWN CMANWCL

qI

X, OdE

E /N.

BAUD P l T E R V A L
YlNlYUY

1 = 10

0.130.12-

= 2

AV- 2

H :o

0.

\.

3
W

$ 0.08.f
t

Om-

007-

006LINE AR EQUALIZE R

- INPUT Y A S U R E Y E N l

til &

0.050.04LINEAR EOUALIZER OUTPUT YEASURYLIIT

-.

0.03-

a02-

0.aO f
r

MClSKm FEEOBILtK EQUALIZER


I

-OUTPUT Y E A S U R E Y E N 1
1
8
!

500

lo00

1 5 0 0

2000

2500 3000 3500 4000 4 m ADAPTIITION T I M E , NUMBER OF BAUD INTERVALS

5000

5503

6Ooo

Fig. 10. Adaptation performances at various threshold

yalues, high signal-to-background noise ratio example.

UNKNOWN CHANNEL

q(r1 *

EIN. BAUD INTERVAL


MINIMUM

--

1 0 dB 3

IO

AV. 2

LINEAREOUALIZER

- OUTPUTMEASUREMENT

LINEAR EWALIZER

- INPUTMEASVREMENT

0 I-

\
l

MCISION FEEDBACK -EWALIZER OUTPUT MEASUREMENT


I I l I l 1

\1500

m-0-

DECISION FEEDBACK EWALIZER INPUT MEASUREMENT


1 1 l 5000 I I

00
0

SC

lo o0

2000 2500 3000 3500 4000 4 5 0 0 ADAFTATION TIME, NUMBER BAUD OF INTERVALS

5500

6000

Fig. 11. Adaptation performances a t various threshold values, low signal to background noise ratio example.

described in the previous section. (The results shown in Fig. 9, and the results to follow, are normalized in amplitude such that I B ( j ) 1 = 1.) The decision feedback equalizer converged to i.ts steady-state performance after 1500 digits were processed, or in about 0.2 s if every digit was used for adaptation. It isnotablet.hat knowledge of the correctdigit sequence did not improve the adaptation of eit.her equalizer significantly; use of the decisions gave almost the same

performance. However, the-. digit. sequence was the output of a pseudorandom generator. It, is likely that the trans. ___ miGiZiGf%ertain redundant sequences could cause a significant increasein adaptationt.ime, especially if the eye at the equalizer output were initially closed. Two important characteristics that can be taken from a graph as such Fig. 9 are steady-state the meansquare error and t,he adaptation time. Unfortunately, i t is not possible to improveboththesecharacteristics simultaneously by changing the adaption parameters V and A. By increasing V , or decreasing A, or doing both, the mean-square error can be reduced, but the adaptation time is increased. A series of test,s was made to determine how t,he mean-square error and the adaptation time are dependent on the t.hreshold V . I n t,hese tests the channel impulse response was Ae-0.5r/T. A filter matched to this pulse was used as part, of the equalizer in this series of tests.Measurements were made a t high SNR, approximately 70 dB, and a t 13.0 dB. The effect of using both input measurements 2( j ) y ( j k) and output measurements e^( j ) 8 ( j k) were tested with both the decision feedback equalizer and linear the equalizer. The effect of changing V from 2 to 24 in increments of 2 is (17rshownin Figs. 10 and 11. As is shown, the performance w , . L r of the decision .feedback equalizer is better than that of the linear equalizer when either tap-error measurement is used. Also, the performance of both equalizers when ? adaptation is based on e^( j ) 8 ( j k) is better than when / it is based on e^( j ) y ( j k ) , except for the decision feed~

I_

&

I -

GEORGE

et al.:

ADAPTIVE DECISION FEEDBACK EQUALIZER

291
lo

back case when E / N o = 13 dB and a largethresholdwas used. I n allcasesamuchquicker adaptation could be achieved by using e^( j ) 6 ( j k ) . I n all cases a minimum product of adaptation time and mean-squareerrorwas achieved when V was 4. Thusboththe experimentalresults andthe analysis indicate that the decision feedback equalizer can be made adaptive, and that its adaptive performance is better than that of the linear equalizer. Estimates of either E [ e ^ ( j ) y ( j k ) ] or E[;( j ) 6 ( j k ) ] can be used to modify the forwardtaps. The experimentalresultsdescribedpreviously show that the latter measurement is \ better in many cases. However, the work of Hirsch and 1 Wolf [lo] indicates that when other channelsare used measurement of E[;( j)6( j k ) 3 cannot beused to make the linear equalizer adaptive. Further investigation is required to determine which is the better measurement to adapt the decision feedback equalizer to such channels.

\
\

\ \
\ o \

\ \
\

a
\
\

.\"
\

IV. PRACTICAL CONSIDERATIONS

\
\
\

\
\\

It has been shown previously that the performance of the decision feedback equalizer is considerably better than that of the linear equalizer, both when the channel impulse response is known and when it is not. This result is particularly important when one realizes that the decision feedbackequalizer is no more complex than the linear equalizer. For instance, in the telephonecableexample t.he decision feedback equalizer with 4 forward taps and 7 feedback taps outperformed the linearequalizer wit,h 11 taps. In both cases the input signal and the tap values were quantized to 7 bits. In fact, the decision feedback equalizer is potentially simpler, since the feedback delay line can be a single bit shift register if the data is a binary one. It was determinedfromcomputersimulationstudies that the decision feedback equalizer is no more sensitrive to signalquantizationerrors or tap quantizat.ionerrors thanthelinear equalizer, though even it achieves its superior performance by coherent subt,ractionof the intersymbol interference. This is consistent with the tap sensitivity analysis of the previous section, (34)and(35). Inthe computer simulation study matched the filter output g ( j ) and the equalizer tap values could be independently quantized any to specified accuracy. The results of a typical test are shown in Fig. 12. I n this test t,hechannelimpulse response was ATe-O.sr/T,the signalto-background noise ratio was 16.0 dB, the linear equalizer had .5 taps, and the decision feedback equalizer had 3 forward taps wit'h 8 feedback Two taps. curves of Fig. 12, one for each equalizer, show the error probability as a function of the number of quantization bits when all quantities quantized the are with same accuracy. As shown, the error probability starts to increase when the number of quantization bits is reduced to 8. I n contrast, when the quantization of the signal was held a t 10 bits, the tap gain quantization of both equalizers could be reduced t o 6bits before any significant increaseinerror probability was observed.

X LINEAR EQUALIZER
0

DECISION FEEOeACK

EQUALIZER

101 0 NUMBER OF QUANTIZATION BITS


1 4

Fig. 12.

Effect of signal andquantization tap performance.

on receiver

I n a separate experiment, quantization of only the signal values j y ( j ) ) resulted in a performance very similar tothat achieved when all quantities were quantized equally. It is believed that the reason for t,he 2-bit difference in quantization requirements is that the signal includes the additive noise andtheintersymbol interference, and so the quantization error is a larger percentage of the desired signal than of the total signal y (1). Similarresults were observed when the coaxial cable PCM channel was simulated. Over that channel a t 360 Mbit/s it was necessary to use 7-bit accuracy for the signal and 6-bit accuracy for the taps. At 450 Mbit/s the linear equalizer could not effectively equalize the channel, even with as many as 21 taps and with no quantization error in either the signal or the tap values. The decision feedbackequalizerwith 5 forward taps and 6feedback taps required 9-bit signal quant#ization accuracy and 7-bit tap accuracy. These results are directly applicable if a digital synthesis method is used. They show that the decision feedback equalizer is no more sensitive to quantization inaccuracies than the linear equalizer, and so is no more expensive t o construct. If an analog synthesis method is used, these results indicate that the decision feedback equalizer is n o more sensitive to componentinaccuracies or signal distortion than the linear equalizer.
I

292 TABLE I STEADY-STATE OUTPUT SNRs AS FUNCTION OF NUMBER TAPS OF

IEEE TRANSACTIONS ON COMMUNICATION TECHNOLOGY, JUNE

1971

of the linear equalizer. As well, at any specified low error probability the decision feedback equalizer allowed data transmission a t rates beyond that possible with the linear Number of Forward Taps equalizer. Two practical algorithms are described that Number of make the decision feedback equalizeradaptive to unknown 2 5 3 4 Feedback Taps or slowly varying channels. It was found experimentally 11.0 1 10.7 9.6 11.1 that thealgorithm based on t,he cross correlations between 14.2 14.9 13.3 11.4 2 theestimatederrorandtheestimateddigits gave the 15.4 16.7 3 14.5 11.9 16.0 15.9 14.3 11.9 better performance, and that a training sequence was not 14.7 16.0 16.1 11.8 necessary foradaptation. Also, i t was found thatthe 16.2 14.9 16.1 11.8 14.9 7 11.7 decision feedback equalizer is no more sensitive to quan14.8 11.7 8 tization errors than the linear equalizer. Because of t,hese advantages, and because its performance is close to that of the much more complex Bayesian demodulator when equalizer The required number of equalizer taps was also exam- the channel is known, the decision feedback ined. Computer simulation tests showed that; when the should be considered wheneverlinearmodulationtechchannel simple had a impulse response such as o r niques are used to transmit digital information over disboth equalizers equalized the channel as well with persive linear channels a t a high rate. This is additionally verified in the theoretical portions a few transversal taps as with many, but that the decision feedback equalizer required several feedback taps to real- of the paper where it is shown that the idealized decision ize its full potential, and as many as 10 or 12 a t very high feedback equalizer will always yield smaller mean-square rates. Note however, from Figs. 2 and 3, that the more error than the transversal equalizer. As well, theoretical of complex decision feedback equalizer could att,ain a per- considerations indicate that the adaptive properties the formance not possible wit,h alinearequalizer wit.h any decision feedback equalizer will t.end to be superior. number of taps. In the more complex channel examples REFERENCES the linear equalizer did not retain this advantage in simD. C. Coll and D. A. George, A receiver for time-dispersed p1icit.y. Computersimulationtests of the coaxial cable pulses, in Conf. Rec., 1965 IEEE Int. Conf. Communications, PCB1 channel a t 360 Mbit./s, wit.h a 20.0-dB signal to pp. 753-758. R.. W. Lucky, Techniques for adaptive equalization of digital background noise ratio, showed that the linear equalizer communication svstems. Bell Svst. Tech. J., vol. 45. Feb. required 9 taps to realize its full potenbial, and the deci1966, pp. 255-268. .I. G . Proakisand J. H. Miller. An adaDtive receiver for .. sion feedbackequalizerrequired4forward taps and 5 digital signaling through channels with iitersymbol interfeedback taps. ference, IEEE Trans. Inform. Theory, vol. IT-15, July 1969, pp. 484-497. A series of hardware simulation tests was carried out R. A. Gonsalves, Maximum-likelihood receiver for digital to determine the numberof taps required by the adaptive transmission, IEEE Trans. Commun. Technol., vol. COM-16, June 1968, pp. 392-398. decision feedback equalizer. I n thesetests 450-Mbit8/s R. R. Bowen, Bayesian decision procedure for interfering transmissionover thePCRI channel a t high signal to digital signals, ZEEE Trans. Inform. Theory (Corresp.), vol. IT-15, July 1969, pp. 506-507. background noise ratio was simulated. The threshold V - Bayesiandetection of noisy time-dispersed pulse se, was held a t 16 during these tests. The steady-state output quences, Ph.D. dissertation,CarletonUniv., Ottawa, Ont., Canada; Sept. 1969. SNR in d B is shown as a function of the number of taps K. Abend andB. D. Fritchman, Statistical detection for in Table I. As shown, 4 forward taps and 4 t.0 6 feedback communication channels with intersvmbol interference. Proc. ZEEE, vol. 58, May 1970, pp. 779-785. taps is a good compromise between better performance M. E. Austin, Decision feedback equalization digital for and highercost. Use of more than 6 feedback taps decommunication over dispersive channels, M.I.T./R.L.E. Tech. Rep. 461, Aug. 11, 1967. grades the performance, because the amount of quantizaH. Robbins and S. Monro, A stochastic approximation tion noise and adaptation noise that the t.ap introduces is method, Ann. Math. Statist., vol.,22, 1951, pp. 400-407. D. Hirsoh and ~ .J. Wolf. A slmole adaDtlve eauahzer for . . ~ W. more than the amount of intersymbol ,interference that is efficient data t,ransmission, IEEE krans. &rnmun. Technol., removed. vol. COM-18, Feb. 1970, pp. 5-12. From these experiment(a1results it, is concluded that the decision feedback equalizer is a practical as well as a very high performance receiver.
~ ~~ ~ ~ ~ ~~ ~

Donald A. George (SJ54-M59) was born inGalt, Ont., Canada,on April 24, 1932. It has been shown that the decision feedback equalizer He received the B.Eng. degree in engineering can be used to detect digital information that has been physics from McGill University, Montreal, P. Q., Canada, in 1955, the M.S. degree in sent a t high rates over an unknown or slowly varying electrical engineering from Stanford Unidispersive channel. The equalizers ability to combat interversity, Stanford, Calif., in 1956, andthe symbolinterferencecaused by severalchannels was inSc.D. degree in electrical engineering from Massachusetts Institute of Technology, vestigated experimentally. I n each of the examples that Cambridge, in 1959. were investigated the digit error probability of the deciFrom 1959 to 1962 hewas a n Assistant sion feedback equalizer was considerably less than that Professor of Electrical Engineering, University of New Brunswick,
I

V. CONCLUSIONS

IEEE TRANSACTIONS ON COMMUNICATION

TECHNOLOGY, cobs-19, NO. 3, VOL.

JUNE

1971

293

Fredericton, Canada. Since then he has been a member of the Facult,y of Engineering, Carleton University, Ottawa, Ont., Canada. While teaching, he has been a Consultant to a number of organizations principally the Defence Research Telecommunications Establishment (now the Communications Research Center) and Canadian Westinghouse Company, Ltd. Also, he three spent summers and a nine-month sabbatical period with the Communications Research Center. Carleton At University, while being concerned withthe development of the engineering programin general, he has been particularly involved in building up graduate activity in communications. His recent research activity has been in the area of optimum and adaptive PAM systems and in signal processing with small computers. At present, he is a Professor of Engineering and Dean of Engineering a t Carleton University. His research interests in are communication and information theory, cybernetics and systems, and signal processing.

has since become the Communications Research Center of the Department of Communications, Ottawa, Ont., and worked on radar signal processing problems. I n 1967 he returned to the Communications Research Center, where he continued his work on the transmission of data through dispersive media. More recently, he and two colleagues have developed a computerlanguagefor simulation of adaptive communication systems.

Robert R. Bowen (S'57-M'61) was born in Peterborough, Ont., Canada, on June 10, 1935. He received the B.Sc. degree in engineering physics and the M.Sc. degree in electrical engineering from Queen's University,Kingston, Ont., in 1958 and 1960, and t,he Ph.D. degree in electrical engineering from Carleton University, Ottawa, Ont., in 1970. I n 1960 he joined the Defence Research Telecommunications Establishment, which

John R. Storey ("68) was born in Trelewis, Wales, on September 19, 1926. He graduated from Lewis School, Pengam, Wales. After graduation he joined the Royal Navy for a period of three years. I n 1952 he joined Decca Radar, Ltd., Surrey, England, working primarily on the development of a baseband communication system. He emi1955 and joined the gratedtoCanadain Ferranti Packard Company where his main interest was in H F communications using meteor trial reflections. In 1957 he joined the Avro Aircraft Company,Malton,and was involved indata processing duringthe flight trials of the Avro Arrow aircraft. He is now with the Communications Research Center, Ottawa, Ont., Canada, (formerly the Defence Research Telecommunications Establishment) having with joined themin 1959. He has worked onprojectsinvolved pulse compression techniques for ionospheric sounding and on research in the field of adaptive receivers for data transmission. He now heads a communicationengineering group currently working on an automatedsystem to measure noise in the H F communication spectrum.

Multipath Intermodulation Associated with Operation o FM-FDM Radio Relays in f Heavily Built Areas
DECIO ONGARO

Abstract-Microwave FM radiolink paths concerning heavily built areas may be found to suffer from high amounts of intermodulation noise caused by multiple reflections on buildings and reflectingobstacles lying alongside the direction of propagation. Little direct information maybe gained on these effects using standard techniques for measuring transmission performance. The limitations of those techniques are discussed and some additional ones are proposed allowing a moredirectinsightinto the Paperap roved bythe Radio Communication Committee of the IEEE 8ommunication Technology Group for publication after presentation at the 1970 International Conference on Communications, San Francisco, Calif., June 8-10. Manuscript received July 1, 1970; revised December 7, 1970. The author is with the Radio Communications Division of Societd Italiana Telecomunicazioni, Siemens, Milan, Italy.

real phenomenon. Theresults obtained inmeasurements on a hop suffering from marked echo effects of this type are reported and discussed.

I. INTRODUCTION

AD10 RELAY terminals are normally located in densely populated areas where the conveyed informationmust beutilized. It is widely known t h a t R F interference is the main problem in such stations, especially when many routes converge to the same terminal. The R F channel allocation plans regulating RF spectrum utilization take this fact into full account. It is perhaps less known that additionaltroublesmay arise due to

Potrebbero piacerti anche