628

TRANSACTIONS IEEE

O N COMMUNICATION TECHNOLOGY, VOL. COM-19, NO.

5 , OCTOBER 1971

Data Transmission by Frequency-Division Multiplexing Using the Discrete Fourier Transform

S. B. W E I N S T E I N ,
MEMBER, IEEE, AND

PAUL M. EBERT,

MEMBER, IEEE

AZisfracf-The lourief transform data commtdcation system i6 a realization of freqbency-division multiplexing (FDM) in which discrete Fourier transforms are computed as part of the modulation and demodulation processes. In addition to eliminating the banks of subcarrier oscillators and coherefit demodulators usually required in FDM systems, a completely digital implementation can be built around a special-purpose computer, performing the fast Fourier transform. In this paper, the system is described and the effects of linear channel distoriion are investigated. Signal design criteria and equalization algorithms are derived and explained. A differential phase modulation scheme is presented that obviates any equalization.

I. INTRODUCTION

tl\

A
W

r\
W

*
W

r : as COS(IOllt/T)
T

ATA ARE usually sent as a serial pulse train, but therehas longbeen interestinfrequency-division mult.iplexing with overlapping subchannels as a means of avoiding equalization, combating impulsive noise, and making fuller use of the available bandwidth. These parallel da.ta systems, in which emh member of a sequence of N digits modulates a subcarrier, have been +I 1 . a0 studied in [2] and [4]. Multitone systems are widely used I T and have proved to be effective in [3], [SI, and [9]. Fig. (C) 1 compares the transn~issions of a serial and a parallel Fig. 1 . Comparison of waveforms in serial and parallel dat.a system. transmission systems. (a) Serial stream of six binary digits. (b) Typical appearance of baseband serial transmission. For alargenumber of channels,the arrays, of sinuto (c) Typical appearance of waveforms that are summed soidal generators and coherent demodulators required in create para,llel data signals. parallel systems become unreasonably expensive and complex. However, it can be shown [ l ] t h a t a multitone data subchanneismaybeusedinsimpletransformations of signal is effect,ively the Fourier trans.form of the origina.1 the receiver output data .to produce, excellent estimates serial data train, and that the bank of coherent denioduof the original data. Further, a simple equalization algolators is effectively an inverse Fourier transform generarithm will minimize mean-square distortion on each subtor.Thispoint of view suggests a completely digital channel,anddifferentialencoding of theoriginaldata modem built around a special-purpose computer permay make it possible to avoidequalizationaltogether. forming the fast Fourier transform (FFT). Fourier transform techniques, although not necessarily the signal 11. FREQUENCY-DIVISION MULTIPLEXING AS A DISCRETE format described in this, paper, have been incorporated TRANSFORMATION into several military d a h communicationsystems [5]Considera data sequence (d,,, d l , . . ., d,where [71. each d, is a complex number.d,, = a,, + jb,. Because each subchannel covers only a small fraction I f a discreteFouriertransform (DFT) is performed of the original bandwidth, equalization is potentially onthevector {2dn}n,0N-1,theresultis a vector S = simpler than for a serial system. In particular, for very (S,,, S1, . . . , S,,- 1) of i V complex numbers, with narrow subchannels, soundings made at the centersof the

8,
Pa.per approvedbytheDataCommunicationsCommittee of the IEEE Communication Technology Group for publication after presentation a.t the 19871 IEEE International Conference on Communications, Montreal, Que., Canada, June 14-16. Manuscript received January 15,1971 ; revised March 29, 1971. The authors are with the Advanced Data Communicat.ions N. J . Department, Bell TelephoneLaboratories,Holmdel,

N-1

=
n=O

2dne-i(2*nm/N)

N-1
n-0

dne-j2=fntm,

m = 0 , 1,

. . . , N - 1,

(1)

where

AN5

WEINSTEIN

EBERT: TRANSMISSION BY FREQUENCY-DIVISION MULTIPLEXING

t, = m At

'

(3)

SOURCE

and A t is an arbitrarily chosen interval. The real part of the vectorX has components
N- 1

dn'an+i&. n=O,I, ..., N-l

629
:(REAL PART

DFT

-: ONLY)

.
CHANNEL

LOWy(t) PASS . : FILTER

DISTORTIONLESS

Y,

2
,,=O

(a, cos 2nj,t,

+ 6, sin 2nfntm),
Fig. 2. Fouriertransformcommunicationsystem channeldistortion. in absence of

m = 0, 1, , N - 1. (4) If thesecomponentsareapjdiedto a low-passfilter a t timeintervals, A t , asignal is obtained that closely approximates the frequency-division multiplexed signal
N- 1

y(t) = 2

%=n

(a, cos 2 ~ f , t

+ b, sin 2nf,t),

0 _< t _< N A t . (Fi) A hlock diagram of the communication system in which y ( t ) is the trammitted signal appears in Fig. 2. Demodulation at. t.he receiver is carried out via a discrete Fourier transformatid of a vector of samples of the received signal.Becauseonlythe rea.1 part of the Fourier transform has been transmitted, it is necessary to sample twice as fast as expected, i.e., a t intervals A t / 2 . When there is no channel distortion, the receiver DFT operates on the 2N samples

Y,

= y(k

9)

= 2

(u,, cos 2~

2mk

+ b,sin __ 2N
'n-2
!WI

0, 1, . * . , . 2 N - 1, (6) where definitions (2) a n i (3) have been substituted into (4). The DFT yields

'n

fn+ I

'n+2

Fig. 3. Power dtnsity spectra of subchannel components of y ( t )

2N--1

111. EQUALIZATION BY USE OF CHANNEL SOUNDINGS

l = O
1
=

J2%,

irrelevant, where the equality

=1

a, - j b , ,

I , 2, . . * , N - 1

(7)

1 > N - 1,

a[ and b, are has been employed. The original data available(exceptfor 1 = 0 ) as the real and imaginary conlponents, respectively, of z l , as indicated in Fig. 2. A synchronizing signal is reciuired, but one or several channels of the transnlitted signal can readily be utilized for this purpose. Because the sinusoidal components of the parallel data signal y ( t ) aretruncated in time,thepowerdensity spectrum of y ( t ) consis'ts of [sin ( f ) / f ] "shaped spectra, as sketched in Fig. 3. Nevertheless, the data on the differentsubchannelscan be completely separated by the DFT operation of (7). This will not be exactlytrue when linearchanneldistortionaffectsthereceived signal, hut it will be shown later that a modest reduction in transmission rate eliminatesmost. interferences.

Exceptfortheaddedlinearchanneldistortionand final equalizer, the Fourier transform data communication systeni shown in Fig. 4 is identical to that of Fig. 2. Ideally,thediscreteFouriertransformationinthereceiver should be replaced by another linear transformation,derivedin 1 1 1 1 , which minimizes theerrorinthe receiver output. However, it is preferable, if possible, t o retain the DFT with its "fast" implementations and carry out suboptimal but adequate correctional transformationsatth2receiverout<put.Thesystem of Fig. 4 performs this approximate equalization. Consider the waveform at the receiver input,
r(t) = Y ( t ) * W ,

(9)

wheretheasterisk deliotcs convolution.Thiswaveform is a collection of truncated sinusoids modified by a. linear filter. If the sinusoid cos 2Tf,,t were not truncated, then the result of passing i t through a channel with transfer function H ( f ) would be H , cos(2~f,t +,,) , where

y ( t ) [see (5) 3 The sinusoids in the transmitted signal are truncated to the interval ( 0 , N A t ) , so that the nth subchannelmustaccommodate a [sin N T ( ~ - fn)At]/

630
SOURCE dn*an+ibn. n=O,l, N-I

IEEE TRANSACTIONS O N COMMUNICATION TECHNOLOGY, OCTOBER

1971

...,

OFT

:(REAL PART :ONLY1

-Low-

Y(t)

e PASS
FILTER

I
SAMPLE AT INTERVALSkAt/2

.",.. i

DFT

EOUALIZATION * TRANSFORMATION

7bn

Fig. 4 . Fourier trmsform communicati.on system ' linear channel distortion and final equalization.

including

[ h T r (f f n ) A t ] spectrum instead of the impulse at fn, which would correspond to a pure sinusoid. However, if l / ( N A t ) is smallconlparedwiththetotaltransmission bandwidth, then H ( f ) does not change significantly over thesubchannelandanapproximate express,ion for the received signal r ( t ) is
N-1

r(t)

2
n=1

Hn[an COS (27rfat

+ 4n) + b, sin (W,t + 4Jl + 2H0an

+ b, sin &)
cos 27rf,t

Equations (13a-c) describe a 2 X 2 transformation to be performed on each of the DFT outputs z I , I = 1, 2, * , N - 1. Forareasonablylarge N andatypical communication channel, the approximation of H ( f ) by a constant over each subchannel, which leads-to (13a-c) , may be adequate. However, linear rather than constant approximations to the amplitude and phase of the channel transfer function as it affects each subchannel waveformaremuchclosertoreality.Thefollowingsection examines the consequences of theseapproximations. It is shown that the truncated subchannel sinusoids are delayed by differing amounts, and that distortion is concentra-ted at the on-off transitions of thesewaveforms. Further, the magnitude of the distortion is, proportional totheabruptness of thetransitions.Hence a "guard space," consisting of a modest increase in the signai duration together with a smoothing of the on-off transitions, will eliminate most interference among channels and between adjacent transmission blocks. The individual channels can then hc equalized in accord with (13a-c).

N- 1

2
n= 1

H,[(a, cos 4,

+ ( b , cos 4" - a, sin +J sin 27rfnt] + 2H0ao, 0 5 t 5 N A,?.

e ,

IV. APPROXIMATE ANALYSIS OF THE EFFECTS OF CHANNEL DISTORTION Thetransmittedsignal y ( t ) as givenby(5)exists only on the interval (0, NAt), so that each suhchannel must, as noted earlier, accommodate a sin f/f type spec(11) trum. As suggested in t,he last section, let this spectrum he narrowed by increasing the signal duration t o some T > N A t and requiring gradual rather than abrupt rolloffs of the transmitted waveform. S'pecifically, the transmitted signal will be redefined :IS

As indicated in Fig. 4, r ( t ) is sampled at times k: (At/2) , k = 0, 1, . . 2N - 1, and the samples { r k } are applied to a discrete Fourier transformer. The output of the DFT is

El
o .

2N-1

2Hoao,

=.o
=

H,[at cos 4, b, sin - a, sin +,I, 1

irreievant,

- H , [ b , cos 4, 1, 2, . . . , N - 1

(12)

1>N-1.

o l t < T

Estimates of nl and b, are obt,ained from the computa-

ions
6, =

1 [ R e (2,) cos 4, H,

+ Im (z,) sin
1
=

+1]

1 0 elsewhere.
(15)
, ( t ) is. sketched in Fig. 5. The "window function" g When y ( t ) is pa.ssed through.the channel filter with impulse response h ( t ), the received signal is

1, 2,

,N -

1.

(13a)

is Tn complex notation, the appropriate computation

&,- j6, where
=

w,z,,

(13b)

WEINSTEIN AND EBERT: TRANSMISSION

BY FREQUENCY-DIVISION MULTIPLEXING

63 1

--io1

TiZOT

Fig. 5 . Window g.(t)

multiplying all subchannel sinusoids in transmitted signal.

where

p ( t )

2h(t)

qb'n)(t)=

* g&) 2h(t) * g,(t)

cos 2af,t sin 2nfnt.

(17)

I n Appendix I, linear approximations to t.he amplitude and phase of H ( f ) around f = *f,, result in the following approximate expression for qa(n) ( t ).

qn(n)(t) E 2Hn COS Y27rfnt

+ +.]ga(t

- PJ

Fig. 6. Linearapproximations to amplitudeand phase of H ( f ) in relation to spectra G,,(f - fn) and Ga(f 1.n).

r Z N SAMPLES-

where ( H N , is the channel sounding a t frequency f n , and LY, and Pn aretheslopesat f = fn of thelinear approximations to amplitude and phase of H ( f ) , as shown in Fig. 6. A similar expression resulte for q b ( n()t ). The first term on the right-hand side of (18) is the 7zth cosine element in the transmitted signal (14), except that it is modified by a channel sounding ( H a , &) and subjectedtoanenvelopedelay Pa. Interblockinterferencecanresult if adelayedsinusoidfrom a previous block impinges on the current sampling period. The second term is distortion arising from the amplitude variations of H ( f ) , and it is a potential source of interchannel interference.Suppose,however,that T is largeenough so that

Fig. 7 . Shifted versions of g,(t) corresponding to subchannels wit.h minimumandmaximumdelayandlocations of samples taken by receiver. Here t,,,, = min, fin, L a x = rnax,% 6".
N-1

r(t) E 2
n=1

Hn[a, cos (2nfnt

to

+ 4,J + bn sin (27rfnt+ 4JI

to

+ 2Hoao,

At + ( 2 N - 1) 5.

(23)

>

( 2 N - 1)

At y + max (PJ
n

- min ( A ) ,
n

(19)

as pictured in Fig. 7. Then for all to > max / 3 , such that

- g,(t

n there exists a t,ime

d dt

- pn)

0,

to

5 t I (2N

- 1)

At

to.

(20)

Fig. 7 shows where the interval ( t o ,to ( 2 N - 1) ( A t ] 2 ) ) is locatedwithrespecttotheminimumandmaximum valuesof the time shift &. Thus,

q.'")(t) "= 2 H , cos (27rf"t

+ +"),

By a similar derivation,
~~("'(2)

Exceptfortheshifteddomain of definition, ( 2 3 ) is identical t,o ( l l ) ,which led to ( 1 3 ) for retrieval of the data. It can be shown that initiating sampling of r ( t ) a t t = t,)instead of a t t = 0 is equivalent t.o incrementing each phase +[ by f l t,, rad in the equalization equations (13). Intuitively, r ( t ) reduces to ( 2 3 ) because linear distortion delays different spectral components by different anlounts and responds to abrupt transitions with "ringing." If the subchannel waveforms, are examined during an interval when they are all present, and if all the onoff transitions lie well outsidethisinterval,thenthe waveforms will look like the collection of sinusoids described by ( 5 ) . The linear approximations to amplitude and phase of H ( f ) restrict the ringing from the transition periods t,o those periods themselves. Inpractice, the ringing can be expected to die out very rapidly outside of the transition periods. This use o f a guard space is a commontechnique,as,forexample,describedin

2H, sin (2af,t

to

+ 4,J,
to

VI.

5t i

+ -2--

Ti. ALGORITHM FOR MIXIMIZING MEAN-SQUARE

( 2 N - 1) A t

(22)

DISTORTION Under the assumption supported by the results of the last section that interchannel interference is negligible, a simple algorithm can be devised for determination of the

Therefore,substituting for n = O ) ,

( 2 1 ) and ( 2 2 ) into (16) (except

632

IEEETRANSACTIONS ON COMMUNICATION TECHNOLOGY, OCTOBER

1971

parameters cos +!/HI andsin + J H , on each channel, which minimize mean-square distortion when used in thetransformation of (13a-c).Because of propagation delay and the' presence of noise, these parameters may notexactlycorrespondto a sounding of the I t h subcarrierchannel.The w e of anautomaticequalization procedure based on the minimum mean-square distortion algorithm leads to accurate demodulation without having to make precise channel soundings. Further, the algorithmcanworkadaptivelyafteraninitialcoarse adjustment. The receiver produces estimates d l and 6, according to the formulas
d , = TI, R e 2 ,

+ T,, Im 2,
(24)

6,

T,, R e 2 , - T , , I m Z , ,

whichresemble(13a-c),except thatthe T coefficients are to be chosen t o nlinilnize the estimation err0r.l Mean-square distortion is defined by

implernent,ation of (30) is showninFig. S. Theinitial value of 56 is probably best obtained from crude channel soundings, or specified as some "typical" vector quantity. It is expected that the first round of adjustments, made at the end of the first block transmission, will suffice t o reduce the error to a low level, if it is not already low with the initial value of T. At At = 0.5 ms, the length of one blockbeforeaddition of a guardspace will varyfrom about 8 ms (16 subchannels) to about 64 ms (128 subchannels). This block length, plus the guard space necessary to minimize interference, is a transmission delay that cannot be avoided. Theequalizationalgorithmgivenhereonlyequalizes distortion due to cochannel interference (channels on thesamefrequency)andcompletelyignoresinterchanne1 interference. An unpublished analysis by the authors shows that for this equalizer, the interchannel interference becomes arbitrarilysmallasthenumber of subchannels increases. This is true even without any of the signal modification described in Section IV.

VI. TRANSMISSION WITHOUT EQUALIZATION

N- 1

2=1

[(TI,. R e 2,

+ T,, . I m Z , - a,)'
(25)

+ (T,,.Re 2, - T , , . I m 2 , - b,)'],
where

It is shown in Appendix I1 that the vector F , where

is a convex function of

We have shown that for narrow subchannels the channel can be equalized by multiplying zz, the receiver output for the Zth subcarrier channel, by a number w1[see (13c) 1 . This is simply compensation for attenuation and phase shift in that particular subchannel. The interference among subchannels is made small by using a guard time and smooth transitions between blocks. For binary transmission, the attenuation need not be compensated,and if differentialphasetransmission between subchannels is used, no phase equalization is needed. In order for this technique to, work, the difference in phase of the transmission channel transfer function H ( f ) between adjacent subchannels should be small. Assume this is the case and let
dn
=
(an

Thus a steepest descent algorithm is sure to converge to the vector yielding the minimum mean-square distortion. The l,,,th components of the gradient of E with respect t o T are

- jbJdn-1,

(30)

where a, and b,, are binary information digits on the nth subchannel. For the first block transmission, do is necessarily a fixed reference. At the output. of the DFT in the receiver, form the product
Z,Z,-~*

(VE),, = __ " - E[2e,,

R e 2, - 2elb I m Z,]

= h,d,h,-l*d~-l*
=

(a,

- jb,) -

lhnl' ldn-ll'
Idn-lI',

The steepest descent algorithm makes changes at the end of each block transmission in a direction opposite t o t h e gradient:

+ (a,

jbn)hn(hn-,*- hn*
n = 1, 2,

... , N

- 1,

(31)

ATll

= =

- k [ e , , R e 2, - e l b I m Z , ] --k[el, I m 2 ,

AT,,

+ elbR e Z,],

(29) of the
not he

where k controls the step

size. A blockdiagram

The vector T and its subscrbted comDonentsshould confused with the signal durat,ion ?' used eailier.

where h, = H ( f , ) is the complex channel transfer function at the center frequency of the nth subchannel. The last part of (31) is t.he information signal times an unknown amplitude term, plus, an error term depending on h, - h,, - 1. Forbinarytransmission,theinformation signalcan be reliablyrecovered if the second term is less than half of the first term. Thisis equivalent to saying that the phase of h does not change by more than

W E I N S T E I N A N D EBERT

TRANSMISSION BY FREQUENCY-DIVISION MULTIPLEXING

633

Fig. 8. AutomaticequalizerforFouriertransformreceiver. One such apparatus is requirctl for each of the ( 7 1 - 1) usable outpnts of the discrete Fourier transformer.

30"betweenthecenters of adjacentsubchanncls.Because a,, and b, are binary, (a7,- j b , ) may be recovered bydetermining which quadrant of the complex plane contains z , ~ , , even though h, is unknown. For the second and a11 subsequent block transmissions, do can carry information by comparing it with do from the previous T-second t.ransmission,i.e.,
do
current

(a,

- jbo)&

Drevious.

(32)
Thus
2 ~ cos , pTf,t

APPENDIX I

RECEIVED SIGNAL UNDER LINEAR APPROXIMATIONS TO na(nl(t) AMPLITUDE A N D PHASE OF THE CHANNEL TRANSFER FUNCTION
The received signal is given by (16) as
N-1

+ 4n]sa(1- p,)
[W,t

+ >sin

+ &I

d g,(t - P,).

(41)

r(t> =
where

*=O

bnqa(")(t)

+ b,dn)(t)l,

(33)

This approxinlation appears as (19) in the main body of thispaper.Asimilarderivationyieldsananalogousapto prosimation q b ( a()t ).

I1 APPENDIX
q a y t ) = 2h(t) * [ga(t) cos (2Tf"t)l

CONVEXITY OF MEAN-SQUARE ERROR (34) The mean-square error has been defined as
N- 1

qb(")(t) = 2h(t) * [g,(t) sin (27rfnt)].

Restrictingattentionforthetimebeingto

q,(") ( t ) ,

the Fourier transform of this function is

H(f)[Ga(f - fn) Ga(f fall, (35) where G a ( f ) is theFouriertransform of the "window function" ga ( t ). We now make linear approximations to the amplitude and phase of H ( f ) as they affect the separatespectra Ga(f - f n ) and G(f + f n ) in (35). For G,(f - f,!), the approximation is
Qa(n)(f)

E(?)

E
1=1

{ [ T IRe , z,

+ T,, Tm x , - all2
(42)
as

+ [T,, R e x , - T,, I m x, - b , ] ' } ,

where the vector

T is defined

(TI,, Tzl, T z z , Tzz, . ~ * 7 . Ijl(N-1)) T ~ ( N - I ) ) (43) .

Taking a single term of the sum, we have

H ( f ) s [H,
and for G, (f

+ a,(f

- f,~)]e'i"-B~cr-r~",

(36)

TIz2(Re z,),
-

+ T2,'(Irn x,)' + 2 R e 2, I m x l T l l T P l

fn)

the approximation is
fn)]e-i[+n+Pn(f+fn)l

2al[TlL R e zL

+ T,, I m x , J + a,' + T2,'(Re x J 2

+ b,'

H ( f ) e [H, - a,(f -

(37)

+ TzZ2(Im x , ) '
=

- 2 R e x , I m z,T,,T,,

Thevalidity of these approximations depends on the narrowness of G,(f) and the size of f n , as illustrated in Fig.6.Onecanalwaysselectawindowfunction ga(t) for which the approximations of (36) and (37) lead

2b,[T,, R e x , - T,, I m x , ]

+ 2T,,(b, I m x , - a, Re z l ) - 2Tz,(a, I m x , + 6, Re x , ) + a,' + b,',

(T,:

+ T,,')

Jz,~'

634

IEEE TRANSACTIONS O N COMMUNICATION

TECHNOLOGY, VOL.

COM-19, NO.

5, OCTOBER 1971

which is the sum of functions and is thus itself. Since E(T)is the sum of convex functions, it too is convex. ACKNOWLEDGMENT The authors are indebted to J. E. Mazo for comments onapproximationtechniques,andto .J. Salzforearlier work on this project.

[9l J. 1,. Holsinger, Digital communication over

fixed timecontinuouschannelswithmemory,withspecialapplications to telephone channels. M.I.T. Res. Lab. Electron., Cambridge, Rep. 430,1964.

REFERENCES
111 J. Salz anti S. B. Weinstein,FouriertransformcommunicaComput,. MaFhinery tion system? presented a t the Ass. Conf.ComputersandCommunication,PineMountain, Gn., Oct. 1969. 121 B. R . Sa,lzberg. Performance of an efficient, parallel data transmissionsystem, IEEE Tmns. C o r n m w ~ .Technol., vol. COM-15, Dec. 1967. pp. 505-811. C31 M. L. Doelz, E. T. Heald, and D. L. Martin, Binary data R o c . IRE, vol. transmissiontechniquesforlinearsystems, 45: May 1957. pp, 656-661. C4l R. W. Chang and R. A. Gibby, A theoretics1 study of perorthogonal multiplexing data t.ransmission formance of an COM-16, scheme, I E E E Trans. Co.mm,,un. Tecknol., vol. Aug. 1965, pp. 529-540. r5! TADIM-A digital - .~ E. N . Powersand M. S. Zimmermnn. implementation of a multichannel datamodem, presented at thc 1968 IEEE Int. Conf.Communicntions,Philadelphia, Pa. 161 W. W . Ahhott, R. C. Benoit, Jr.: and R. A . Northrup, An all-digital ndaptlve data modem, presenkd at, the IEEE Computers and Communication Conf., Rome, N . Y., 1969. 171 W. W . Ahhott, I,. W. Blocker,and G. A.Bailey,Adaptive data modem. . Page - Commun. Eng.., - . Inc., RADC-TR-69-296, Sept. 1069. [SI M. S. Zimmermsn and A . L. Kirsch. The AN/GSC-10 (KATHRYN) variable rate data modem for HF radio, IEEE T m n s . Cornmun. Technol., vol. COM-15, Apr. 1967, pp. 197-204.

S. B. Weinstein (S59-M66) was born in NewYork, N. Y.,onNovember25,1938. He received the B.S. degree from the Massachusetts Institnte of Technology, Cambridge, in 1960, the M S . degree from the University of Michigan,AnnArbor,in1962,and the Ph.D. degree from the University of California, Berkeley, in 1966. Upon finishing his graduate studies he worked for approximately one year with the Philips Research Laboratories in Eindhoven, the Netherlands, beforejoining the Advanced Data Communications Department, at Bell Telephone Laboratories, Holmdel, N. J., in 1968. He works in thy areas of st,atislical communication theory and data commnnications.

Paul M. Ebert (M60) was born in Madison, Wis., on December 30, 1935. He received from the University of the B.S. degree Wisconsin, Madison, in1958, and the M.S. and Sc D. degreesfrom theMassachusetts Institl1te of Technology, Cambridge, in 1962 and 1965, respectively. a member of From1958 to 1960hewas the Airborne Communications Division at the Radio Corporation of America, Camden, N. J. Since1965he hi5 been a member of the Advanced Data Communications Department, Bell Telephone Laboratories, Holmclel, N. J. He has worked in the fields of information theory, coding, and digital filtering.

Eye Pattern for the Binary Noncoherent Receiver

Abstract-The assessment of imperfect channels indatatransmission usually involves noise considerations. Channel imperfections are related to system performanceby determining the increase in carrier-to-noiseratiorequiredtomaintain a fixed error rate. The impairment is determined by examining the eye pattern, which shows the effective reduction in signal amplitude caused by intersymbol interference. This paper derives the eye function for the binary quadratic receiver. The binary instantaneous frequency discriminator, the differentialphase shiftreceiver, andthenoncoherentfrequencyshiftkeying (FSK) receiverimplementedusing thedifference of two envelopes are binary quadraticreceivers. The eye pattern is easily obtained by computation using the eye function. Pa.per n,pproved by the Data Communicat,ions CommitLee of the IEEE Communication Technology Group for puhlication after presentation a t the 1970 IEEE International Conference on Communicntions, San Francisco, Calif., June 8-10, Manuscript received September 22,19870. The author is with Bell-Northern Research, Ottawa, Ont., Cnnadn.

I. INTRODUCTION USEFUL design met.hod for determining the effect of channel distortions on a linearly moduhted and demodulated signal is to use the eye-opening criterion, which basically determines the dist.ance from the decision threshold for the worst data sequence. This distance allows the designer t o determine the increa.se incarrierpowernecessarytocorrectfor this distortion. With the noncoherent quadra.tic receiver i t is difficult to perform a conlparable analysis. Thus, in t.he preliminary treatment of such systems i t is conlmon t o assume t h a t t.he noisepoweris that obtained through a filterwithabandwidthequaltothesignalingspeed, whereas the signal is undist.orted. It is then necessary t o determinetheincreaseincarrier-to-noise ( C / N ) ra.tio that would be necessary withtheactualchannel. To