IEEE TRANSACTIONS O N COMMUNICATION TECHNOLOGY, VOL. COM-19, NO.
5, OCTOBER
under laboratory conditions is repeatably within a few
tenths of a decibel of predictionsbasedontheoryand
simulation. On realchannels,thefragmentaryresults
available to date regularly show the same order of gain,
albeit without the same sort of repeatability. These re-
sults confirm that a sequential decoder of modest com-
plexity can provide of the order of 5-dB effective power
gain at data rates to 5 Mbit/s and that above threshold
very low error probabilities can be achieved.
ACKNOWLEDGMENT
Wearegratefultothepersonnel of the U. S. Army
Satellite
Communication Agency who supportedthis
work: R.M.Langelier, without whose initial enthusiasm
the work would never have begun; D. L. LaBanca, who
providedcontinuity of support;and T. F. Page, R.. J.
Stark, R. C. Gibson, and G. R.. Ash, who were by turns
responsiblefor the progress of theprogram.Thedata
suppliedbythesegentlemenandby D. Quagliato of
the U. S. ArmyElectronicsCommandarealsograte-
fullyacknowledged.Wealsowish tocite s. H. Loui
of Codexforhisindefatigableassistanceinthe con-
struction of the prototype.
REFERENCES
[11 K. L.Jordan,Jr.,The performance of sequential decoding
in
conjunction with efficient modulation, IEEE Trans.
Commun. Technol., vol. COM-14, June 1966, pp. 283-297.
Viterbi Decoding for Satellite and Space Communication
Abstract-Convolutional coding and Viterbi decoding, along with
binary phase-shift keyed modulation, is presented as anefficient
system for reliablecommunication on power limitedsatelliteand
spacechannels.Performanceresults,obtainedtheoreticallyand
through computer simulation, are given for optimum short constraint
length codes forarange of codeconstraintlengths and code
rates. System efficiency is compared for hard receiver quantiza-
tion and 4 and 8 level soft quantization. The effects on performance
of varying of certainparametersrelevanttodecodercomplexity
and cost is examined. Quantitative performance degradation due to
imperfect carrier phasecoherence is evaluatedand compared to
that of an uncoded system. As an example of decoder performance
versus complexity,arecentlyimplementedZ-Mbit/sconstraint
length 7 Viterbi decoder isdiscussed. Finallya comparison is
Paper approved by the Communication Theory Committee of
theJEEECommunication Technology Groupfor publication
without oral presentation.This work was supported in partby
NASA Ames Research Center under Contract NAS2-6024. Manu-
script received June 10, 1971.
Theauthors are with theLinkabitCorporation,San Diego,
Calif.
835 1971
c21 I. L. Lebow and P. G.McHugh, A sequential decoding
technique andits realization inthe Lincolnexperimental
terminal, IEEE Trans. ,Commun. Technol., vol. COM-15,
Aug. 1967, pp. 477-491.
131 I. M. Jacobs, Sequential decoding for efficient communi-
cationfromdeep space, I E E E Trans. Commun. Technol.,
VO~ COM-15,
. Aug. 1967, pp. 492-501.
141 C.R.Cahn, Binary decoding extended to nonbinary de-
modulation of phase shift keying, IEEE Trans. Commun.
Technol., vol. COM-17, Oct. 1969, pp. 583-588.
151 G. D. Forney,Jr.,and R. M. Langelier, liA high-speed
sequential decoder for satellite communications, Conf. Rec.,
1969 IEEE I n t . Conj. Communications, Boulder, Colo., June
1968. OD. 39/9-17.
[61 G. D: -Forney, Jr., Use of a se uential decoder to analyze
convolutlonal code structure, I j E E Trans. Inform.Theory
(Comrresp.), vol, IT-16, Nov. 1970, pp. 793-795.
[71 R. ,G. Gallager, Information Theory and Reliable Con~muni-
catzo?~. New York: Wiley, 1968, pp. 263-273.
181 D. Quagliato, Advanced coding technlques for
satellite
communications, U. S. Army Elect.ron. Command, Ft.
Monmouth, N. J., CADPLRep. 167. Sept. 1970.
G. David Forney, Jr. (S59-M61), for a photograph and biography
please see page 781 of this issue.
Edward K. Bower was born in Columbia,
Mo., on July 8, 1943. He attended the Uni-
versity of Missouri, Columbia, from which
he received the B.S., M.S., and Ph.D. degrees
in electrical engineering in 1964, 1965, and
1968, respectively. While a graduate student,
he held an NSF Cooperative %ollowship and
a NASA traineeship. His area of specializa-
tion was coding theory.
Since graduation, he has been employed by
the Codex Corporation, Newton, Mass. As a
Senior Member of the Technical Staff, he has been active in the
development of coding and modulation devices.
madebetween Viterbi and sequentialdecodinginterms of suit-
ability to various system requirements.
I. INTRODUCTION
HE SATELLITE and space communication chan-
nels are likely candidates for the cost-effective use
of coding to improve communicationefficiency. The
primaryadditivedisturbanceonthesechannelscan
usually be accurately modeled by Gaussian noise which
iswhite enough t o beessentiallyindependentfrom
one bit time interval to the next, and, particularly on the
spacechannelbutalsoinmanyinstances on satellite
channels, sufficient bandwidthisavailabletopermit
moderatebandwidthexpansion.Two effective decoding
algorithms for independent noise (memoryless) channels
have been developed and refined, namely sequential and
Viterbi decoding of convolutional codes. These theoretical
accomplishments, combinedwith
real
communication
836
needsandtheavailability of low-costcomplexdigital
integrated circuits, make possible practical and powerful
high-speed decoders for satellite and space communica-
tion.
Communication from a distant and isolated object in
space to a ground-based station presents certain system
problems which are not nearly as critical in earth-based
communication systems. The most obvious among these
is the high cost of space-platform power. It is desirable
to designasystemwhichisas efficient aspracticalin
ordertominimizethespacecraftweightnecessaryto
generate power.
The nlodulatedsignalpower a t agroundstation
ceiver front end P depends upon the transmitted power,
the transmitting and receiving antenna gains, and
agationpath losses. Primarilydue to thermalactivity
at the receiverfrontend,widebandnoiseisaddedto
the received signal, resulting in a received signal power-
to-noise ratio ( P I N , , ) ,where N,, is the single-sided noise
spectral density. The noise is usually accurately modeled
as beingbothwhite and Gaussian. Other perturbations
caused by uncertainty in carrier phase at the demodulator
and inaccuracies in receiver AGC are treated in Sections
IV andV.
The efficiency of acommunicationsystemisusefully
measured by the received energy per bit to noise ratio
(E,/N,,) required to achieve a specified system bit error
rate. The E b / N o is expressable in terms of the modulat-
ing signal power by the relationship
E ,-- -.-
- P 1
NnNn R future systems; however, the trend, especially in satel-
where R is theinformationrate inbits per second.
Alternatively, (1) can be written as
The payoff forusingmodulation and/orcodingtech-
niques which reduce the Eb/lVo required for a given bit
errorprobabilityisanincreaseinallowabledatarate
and/or a decrease in necessary received PJN,,.
As apoint of reference, it istraditionaltocompare
the efficiency of modulation-coding schemes with that of
ahypotheticalsystemoperating a t channelcapacity.
Channel capacity for an infinite bandwidth white Gaus-
siannoisechannelwithaveragepower P is [ l ]
bit/s.
c- = 2-
From (1),when R = C,,
Thus,the lowerbound on achievable E b / N o isabout
-1.6 dB.
Without coding, required E b / N o can be minimized by
selecting an efficient modulation technique. For example,
IEEE TRANSACTIONS O N COMMUNICATION TECHNOLOGY, OCTOBER 1971
180 binary phase-shift keying (BPSK) is more efficient
than binary frequency shift keying (BFSK). For a de-
siredbit error rate of an Eb,/NOof 9.6dBisre-
quiredusingBPSK(antipodal)modulation,whereas,
12.6 dB is required with BFSK (orthogonal) modulation.
Quadraphase-shift keying (QPSK) is often used t o con-
serve bandwith. Under the assumption of perfect phase
coherence, QPSK has the same performance as BPSK.
I n designing a communication system t o operate a t a
specified data rate, the improvement in efficiency tobe
realized using coding must be weighed against the rela-
tivecosts.Potentialalternativesincludeincreasingthe
re- transmitted power,increasingthetransmittingantenna
gain,and/orthereceivingantennaarea,andaccepting
prop- a higher proba.bility of bit error. In many applications, a
minimum P , isrequiredandthcincrementalcostper
decibelincreasein P / N , is now greater(oftenmuch
greater)thanthecost of reducingtheneeded Eb/No
throughcoding.Soft dccision Viterbiandharddecision
sequential clecoding can provide a relatively inexpensive
4-6-dB improvement in
required Eb/No (at a bit
error rate), even a t multimegahit data rates. Sequential
decoding is extensively discussed in [ 5 ] . I n Section VII,
we compare these techniques. SectionsI1 and 111 examine
variousaspects of Viterbidecoding andpresentcurves
permitting system tradeoffs. I n Section VI, a particular
implementation of a Viterbi decoder is discussed to pro-
videonebenchmarkforcost-complexitydiscussions.
Inthe discussion that follows, we assume thatthe
channel is power limited rather than bandwidth limited.
Thisassumptionisrealisticformanypresentdayand
(1)
lite repeaters, is to larger P;/N,, without a proportional
increaseinavailablebandwidth. For thisreason, we
will limit consideration to codes which involve a band-
(2) widthexpansion of 3 or less; that is, we assume that
from 1 to 3 binarysymbolscan be transmittedover
thechannelforeachbit of informationcommunicated
witllout appreciable intersymbol interference.
11. SYSTEM
A . Convolutional Encoder
Fig. 1 shows a generalbinary-inputbinary-output
convolutional coder. The encoder consists of a k K stage
binaryshiftregisterandmod-2adders.Each of the
mod-2 adders is connected to certain of the shift register
stages.Thepattern of connections specifies the code.
Information bits
areshifted
into
the
encoder
shift
register k bits a t a time. After each k bit shift, the out-
puts of the mod-2addersaresampledsequentially
yielding the code symbols. These code symbols are then
usedby the modulator to specify the waveforms to be
sent over the channel. Sincev code symbols are generated
for each set of k information bits, the code rate RN is
k J v information bits per code symbol, where k < v. The
constraint length of the code is K , since that is the num-
ber of k bit shifts over which a single information bit can
HELLER A N D J A C O B S : DECODING FOR SATELLITES 837

is 0 or 1. The funct)ion p ( t ) is a convenient unit energy

low-pass pulse waveform, f c is the carrier frequency, E ,
is the energy per pulse, and T , is the time between suc-
cessive code symbols. E , and T , are defined by the rela-
tionships

E, = RNE, = kEb/v (6)

and
T, = RN/R. (7)
Thereareseveralreasonsforrestrictingattentionto
I I
t COMMUTATMI
BPSIi nodulation. Three important ones are as follows.
BINARY CODE
SYMBOLS 1) I3PSK signals are convenient to generate and am-
plify. Traveling wave tube amplifiers operate most effi-
Fig. 1. Ratc k/?c convolutiond encodcr. ciently at or near saturation. This nonlinear amplific t
a1011
would degradeperformancewithmultilevelamplitude
influence the encoder output. The state of the convolu- modulated waveforms.
tional encoder is the contents of the first k ( K - 1) shift 2) It can be shown thatantipodal (BPSK) modula-
register stages. The cncoder statc together with the nest tion results in little increase in required E,,/N, compared
IC inputbitsuniquelyspecifythe u output symbols. to optimum signaling when E S / N , ,is low [4].
As anesample, a K = 3, k = 1, u = 2 encoderis 3) BP.SK modulation of quadrature carriers is equiva-
shown in Fig. 2 ( a ) . T h e . firsttwocoderstagesspecify lent to quadraphase (QPSK) modulation of one carrier.
the state of the encoder; thus, there are 4 possible states. Thus, QPSK need not be separatelytrcatedexceptfor
The code words, or sequences of code symbols, generated synchronization and phase error requirements.
by the encoder for various input information bit sequences
is shown in the code trellis [a] of Fig. 2 ( h ) . The code C . De)wodwlntion clnd &unntisatio?z
trellis is really just a state diagram for the encocler of
Fig.2(a).Thefourstatesare representedbycircled At the receiver, the signal s(t) of ( 5 ) , is observed added
binarynumberscorrespondingtothecontents of the to whiteGaussian noise. Whenthecarrierphase e is
first two stages of the encoder, The lines or branches known, the optimum demodulator consists of an integrate
joining states indicate state transitions due to the input anddump filtermatched to p(t) cos (2sf,t + e). At
of sirlgle informationbits.Dashcdand solidlinescor- time j T 8 , thedemodulator outputs data r i relevant to
respond to 1 and 0 inputinformationbits, respec- the jth code symbol. Normalizing thematched filter
tively.Thetrellisisdrawnundertheassumptionthat output by dividing byZ/N,/Z yields
the encoder is in state 00 at time 0. If the first informa-
tion bit were a 1, the encoder would go to state 10 and
ri = xi d 2 E , / N 0 ni + (8)

would outputthe code symbols 11. Codesymbolsgen- when nj is a zero-n1ean unit variance Gaussian random
erated are shown adjacent to the trellis branches. As an variable. Each n j is independent of all others.
example, the input data sequence101 generates the
- a - T o f:tcilitate digitalproccssingbythedecoder,the
codesymbolsequence 111000 . . Furtherinterpreta- continuous r j must be quantized. The simplest quantiza-
tions of the encoder state diagram and a discussion of tion is a hard decision with 0 output if r j is greater than
[(goodconvolutional codes is presented in [3]. zero and 1 output otherwise. Here, the received data are
represented by onlyonebitper code symbol.Without
B . h4odulation coding, the matched filter sampler hard quantizer is an
The binary symbols output by tllc encoder are used to optimumreceiver.
modulatean RF carriersinusoid.Here we restrictour Whencodingisused,hardquantization of the re-
attentiontothecase of 180 BPSK modulation.Each ceived data usually entailsa loss of about 2 dB in
code symbolresults in thetransmission of apulse of E,/N, compared withinfinitely finequantization [4],
carrier a t either of two 180separated
phases. A [5j. Much of this loss can be recouped hy quantizing ri
squence of code symbolsproducesauniformlyspaced to 4 or 8 levels instead of merely 2. Adding additional
sequence of biphnse pulscs. The signal component of the lcvcls of quantization necessitates a 2- or 3-bit represen-
receivedwaveform thushastheform tation of each rj. Fig. 3(a) and ( b ) shows two quantiza-
tion schemes with 4 and 8 levels, respectively. Here the
~(t= ) d2EB p ( t - ~ T COS J (2Tfct ~ +/ 2e) + quantization level thresholdsarespacedevenly.The
spacing is 1.0 for 4 levels and 0.5 for 8 levels. Uniform
= d5E COS (zTff.t +
e) -xip(t - ~ T J . (5) quantization threshold spacings of 1.0and 0.5 canbe
shownbyanalyticalmeansandthroughsimulationto
Here xi is + I deperlding on whether the ith code symbol be very close tooptimumfor 4- and 8-level quantiza-
838 IEEE TRANSACTIONS ON C O M M U N I C A T I O N TECHNOLOGY, OCTOBER 1971

-I CODE

(a)

@ 00 @ 00 @ 00 @ 00 @

0 1 2 3 4

TRELLIS DEPTH *
(b)
Fig. 2. (a) K = 3, RN = 1/2 convolutional encoder. (b) Code trellis diagram.

two inputs (0 or 1,the code symbols) and 2,4, or 8 outputs.

3 I 2 1 I 0
-1.0 1.0
*I' Thechanneltransitionprobabilitiesareafunctiononly
of the symbol signal-to-noise ratio E,/N,. For example,
(a) with 8-levelprobability
quantization,
the of receiving
interval 6 given that a 0-code symbol i s sent is the prob-
abilitythataunit-varianceGaussianrandomvariable
7 1 6 1 5 1 4 3 , 2 1 1 , 0 *
I
-0.5 -1.0 -1.5
I
0.5 l!O
I
1.5
'i with mean 4 2 E 8 / N , , lies between -1.0 and - 1.5 in
value.
111.VITERBIDECODING
Fig. 3. Receiverquantizationthresholdsandintervalsfor (a)
4-level and (b) %level quantization. A . Basic Algorithm
The maximum likelihood or Viterbi decoding algorithm
tion. Furthermore, 8-level quantization results in a loss wasdiscoveredandanalyzedbyViterbi [6] in 1967.
of less than 0.25 dB compared to infinitely fine quanti- Viterbidecodingwasfirstshown to be an efficient and
zation; therefore, quantization to more than 8 levels can practical decoding technique for short constraint length
yieldlittleperformanceimprovement.Weconfineour codesbyHeller [ 71, [8]. Forney [2] and Omura [ 121
attention to hard decision quantization and the 4 and 8 demonstrated that the algorithm was in fact maximum
level schemes shown in Fig. 3. likelihood.
Receiver quantization converts the modulator, Gaussian A thorough discussion of the Viterbi decoding algorithm
channel,anddemodulatorinto a discretechannelwith is presented by Viterbi [3]. Here, it will suffice to briefly
HELLER A N D JACOBS: DECODING FOR SATELLITES 839

review the algorithm and elaborate on those features and technique is limited to relatively short constraint length
parameters which bear on decoder performance and com- codes duetothe exponentialdependence of decoder
plexity on satellite and spacc communication channels. operations per bit decoded on K . Fortunately, as will be
Referring to the code trellisdiagram of Fig. 2(b), a shown, excellent decoder performance is possible with
brute-force maximum likelihood decoder would calculate good short constraint length codes.
the likelihood of the received data forcodesymbol se-
quences on all paths through the trellis. The path with B . Path M e m o ~ y
thelargest likelihoodwould then be selected,andthe
Inorder to maketheViterbialgorithm a practical
information bits corresponding to that path would form decoding technique, certain refinements on the basic al-
the decoder output. Unfortunately, the number of paths gorithmaredesirable.First of all,periodicallyforcing
for a n L bit information sequence is 2l); thus, this brute the encoder into a known state by using preset sequences
forcedecodingquickly becomes impracticalas L in- multiplexed into the data stream is neither operationally
creases. desirablenornecessary. It canbe shown [ 2 ] , [9] that
With Viterbi decoding, it is possihle to greatly reduce with high probability, the 2 k ( R - 1 )decoder selected paths
the effort required for masirnurn likelihood decoding by will not be mutually disjoint very far back from the pre-
takingadvantage of the specialstructure of the code sent decoding depth.All of the 2 k ( K - 1 paths ) tend to have
trellis. Referring to Fig. 2 ( h ) , it is clear that the trellis a common stem which eventuallybranches off tothe
assumesa fixed periodicstructureaftertrellisdepth 3 variousstates.Thissuggests that if the decoder stores
(in general, I<) is reached. After this point, each of the enough of the past information bit history of each of the
4 statescan be enteredfromeither of twopreceding 2J'(K-1)paths,thentheoldestbits on allpaths will be
states. At depth 3, for instance, there are 8 code paths, identical. If a fixed amount of pathhistorystorageis
2 enteringeachstate.For example., state 00 a t level 3
provided, the decodercan outputthe oldest bit on an
hasthetwopathsentering it corresponding tothein- arbitrary path eachtime it steps oneleveldeeper into
formationsequences 000 and 100. These paths are said thetrellis.Theamount of pathstoragerequired u is
to have diverged at state 00, depth 0 andremerged a t equal to the number of states, 2k(K-1)multiplied by the
state 00, depth 3. Paths remerge after2[ingeneral length of theinformationbitpathhistory per state h ,
k ( l ( - I ) ] consecutiveidenticalinformationbits. A
Viterbidecodercalculatesthe likelihood of each of the = h2k'K-l).
(9)
2k pathsenteringa given stateandeliminatesfrom
furtherconsiderationallbutthemostlikelypath that Since the path memoryrepresentsasignificantportion
leads to that state. This is done for each of the 27G'K-1) of the total cost of aViterbidecoder, it isdesirable to
states at a given trellis depth; after each decoding opera- minimize the required path history length h. One refine-
tiononlyone pathremainsleadingtoeachstate.The ment which allows for a smaller value of h is to use the
decoder then proceeds one level deeper into the trellis and oldest bit on the most likely of the 2k(K-1) paths as the
repeats the process. decoder output, rather than the oldest bit on an arbitrary
For the K = 3 code trellis of Fig. 2 ( b ) , there are 8 path. It has been demonstrated theoretically [2] and
paths at depth 3. Decoding at depth 3 eliminates 1 path through simulation 191 that a value of h of 4 or 5 times
entering each state. The result is t h a t 4 paths are left. the code constraint length is sufficient for negligible
Going on to depth 4, the decoder is again faced with 8 degradation from optimum decoder performance. Simula-
paths.Decodingagaineliminates 4 of these paths, and tionresults showing performancedegradationincurred
so on. Note that in eliminating the less likely paths en- with smaller path history lengths are presented and dis-
tering each state, the Viterbi decoder will not reject any cussed in Section IV.
path which would have been selected by the brute force
maximum likelihood decoder. C. State and Branch Metric Quantization
The decoder as described thus far never actually de- The path comparisonsmadeforpathsenteringeach
cides upon one most likely pat.h. It always retains a set staterequirethecalculation of the likelihood of each
of 2 k ' K - 1 )paths after each decoding step. Each retained path involvedfortheparticular received information.
path is the most likely path to have entered a given en- Since the channel is memoryless, thepath likelihood
coder state. One way of selectingasinglemostlikely function is the product of the likelihoods of the individual
path is to periodicallyforce the encoderintoaprear- code symbols [3]
ranged state by inputting a I< - k bit fixed information
sequence to the encoder after each set of L information P(r*/x') = p(r,*/xi')
bits. The decoder can then select that path leading to the 1

known encoder state as its (1 bit) output. where r* = (r1*, r2*, . . . , ri*, . .) is the vector of quan-
The great advantage of the Viterbimaximumlikeli- tizedreceiveroutputsand X' = (x1', x2', * . , xi1, * . e )

hooddecoder is that the number of decoder operations is the code symbolvectorforthelthtrellispath. In

performed in decoding L bits is only L 2 k ( K - 1 which), is order to avoid multiplication, the logarithm of the likeli-
linearin L. Of course,Viterbidecoding asapractical hood is a preferable path metric
840 IEEE TRANSACTIONS ON COMMUNICATION TECHNOLOGY, OCTOBER 1971
LEVEL
M, = log p(r*/x') OUANTlZATlON

= x
i
log p(r,*/ziz) A x miz
1
(1 1)
0
0

7
1

6
2

5
3

4
4

3
5

2
6

1
7

0
CODE
where M 2 is the metric of the lth path and miz is the SYMBOL
I 0 1 2 3 4 5 6 7
metric of the jth code symbol on the Zth path. With this
type of additive metric, when a path is extended by one Fig. 4. Integer code symbol metrics for %level receiver quanti-
branch, the metric of the new path is the sum of the new zation.
branch symbol metrics and the old path metric. To facili-
tate this calculation, the path metric for the best path IV. SIMULATION
AND ERROR
PROBABILITY
BOUNDRESULTS
leading to each state must be stored by the decoder as a
state metric. This is an addition to the path information A . Tradeoffs Between Bit Error Probability and Eb/NO
bit history storage required. for Rate 1 / 2 Codes
Viterbi decoder operation can then be summarized as Viterbihasderivedtightupperboundstobiterror
follows, taking the K = 3 case of Fig. 2 as an example. probability for Viterbi decoding based on the convolu-
1) Themetricforthe 2 pathsenteringstate 00 are tional code transferfunction [ 3 ] . Theseboundsare
calculated by adding the previous state metrics of states particularly tight for the white Gaussian noise channel
00 and 01 to the branch metrics of the upper and lower for error probabilities less than
about This bound
branches entering state 00, respectively. has been numerically evaluated over a range of E b / N o
2) The largest of the two new path nletrics is stored for a variety of codes. Theupperboundispresented
as the new state metric for state 00. The new path his- along with some of the 8-level receiver quantized simu-
tory for state 00 is the path history of the state on the lationresultsforcomparison.Theupperbound'also
winningpathaugmentedbya 0 or 1 dependingon providesperformance data at very-low bit error rates,
whether state 00 or 01 was on the winning path. where simulation results are not available due to exces-
3) This add-compare-select(ACS,) operationisper- sivecomputertimerequired. I n comparingtheupper
formed for the paths entering each of the other 3 states. bounds to the simulation results, it is important to keep
4) Theoldestbit on thepathwiththelargest new inmindthattheupperboundwasderivedforanin-
path metric forms the decoder output. finitely finely quantized receiver output.
Since the code symbol metrics must be represented in The convolutional codes used in the simulations were
digital form in the decoder, the effects of metric quan- foundthroughexhaustivecomputersearch [9], [ l o ] .
tization come into question. Simulation has shown that The search criterion was maximization of the minimum
decoder performance is
quite
insensitive to
symbol freedistancefora givencode constraintlength [3].
metric quantization. In fact,use of the integers as symbol Where two codes had the same minimum free distance,
metrics instead of log likelihoods results in a negligible the number of codewords a t t h a t distance and the higher
performance degradation with 2-, 4-, or 8-level receiver order free distances were used for code selection. Simula-
quantization [7], [SI. Fig. 4 shows such a set of metrics tionshaveconsistentlyshownthatthefreedistance
for the %level quantized channel. Use of these symbol criterionyields codes withtheminimumerrorprob-
nletricsimplies thatsymbolmetricsas well as the re- ability.Theprincipalresults of thesimulationsand
ceived symbols themselves may be represented by 1, 2, code transferfunctionboundsareshowninFigs. 5, 6,
or 3 bitsfor 2-, 4-, and%levelreceiverquantization, and 7. All of these figuresshow biterrorrateversus
respectively. E h / N , , for Viterbi decoders using optimum rate 1/2 con-
volutional codes. In all cases, the decoder path history
D.Unknown Starting State lengthwas 32 bits. In all simulation runs, at least 25
error events contributed to the compiled statistics.
It has been assumed thus far that a Viterbi decoder
hasknowledge of the encoder starting state beforede-
codingbegins. Thus, inFig. 2 ( b ) , the starting state is B . Performance Depending on Quantization, Path
assumed to be 00. A known starting state may be opera- History, and Receiver Automatic Gain Control
tionallyundesirablesince i t requiresthatthe decoder ThesimulationresultsinFigs. 5 and 6 are for soft
knowwhentransmission commences. In reality, it has (%level) receiver quantization. Equally spaced demodu-
beenfoundthroughsimulation thataViterbi decoder lationthresholdsare used as showninFig. 3 ( b ) . This
may start decoding at any arbitrary point in a transmis- choice of %level quantizer thresholds is within a broad
sion, if all state metrics are initially reset to zero. The range of nearoptimunlvalues,as willbeshownpre-
first 3-4 constraint lengths worth of data output by the sently.Thetransferfunctionboundisforinfinitely
decoderwill be more or less unreliablebecause of the finely quantized received data, although tight bounds for
unknownencoderstartingstate.However,afterabout any degree of quantization can be obtained. Allowing for
4 constraintlengths,thestatemetricswithhighprob- the 0.20-0.25 dB loss usuallyassociatedwith&level
abilityhavevaluesindependent of thestartingvalues receiverquantizationcomparedwithinfinitequantiza-
and steady-state reliable operation results. tion, the transfer function bound curves are in excellent
 -
HELLER AND JACOBS: DECODING FOR SATELLITES 84 I

10-2 I
1 .' I I
l -

SIMULATION

--- UPPER BOZRJD

Y
e:
Y
:: \
W
u
4
\
m

\
10-5 I I I I I I I I 1

4 5 6 7

EblNo in d b

3 4 5 6 7 Fig. 7 . Biterrorrate versus Eb/Na for rate 1/2 Viterbi decod-

ing. Hardquantized received data with32-bit paths; K = 3
Eb/No in db through 8.
Fig. 5. Biterrorrate versus E r / N o for rate 1J2 Viterbidecod-
ing. 8-level qua.ntized simulationswith 32-bit paths, and
infinitely finely quantized transfer function bound, K = 3, 5 , 7 .
agreementwithsimulationresultsinthe lo-' tobit
error rate range.
Since theaccuracy of thetransferfunctionbound
increases with E , / N o , decoder performance can be ascer-
tained
accurately
the
in
region
to eventhe
in
absence of simulation. The symbol metrics used in the
simulation were theequallyspacedintegersasshown
in Fig. 4.
Fig. 7 gives the simulation results for Vitcrbi decoding
with hard receiver quantization. The same optimum rate
1/2, I< = 3 through K = 8 codes were usedhere as in
the 8-level quantized simulations.
Y The following points are obvious from the performance
3 curves.
1) 2-level quantizationiseverywhere close to2-dB
Y
r(
inferior to 8-level quantization.
m
2) Each increment in K provides an ilnprovement in
efficiency of something less than 0.5 d B a t abiterror
rate of
3) Performance improvement versus K increascs with
10-6
decreasing bit error rate.
To observe the effects of varying receiver quantization
more closely, simulation performance data are presented
in Fig. 8 for the K = 5 , rate 1/2 code, with 2-, 4-, and
8-level receiver quantization. The Q = 8- and Q = 4-level
10- thresholds are those of Fig. 3.
Fig. 9 shows bit error rate pcrformance versus E0/NO
forthreevalues of pathhistorylength (8, 16,and 32)
using the rate 1/2, K = 5 code, for both 2- and 8-level
Fig. 6. Bit errorrateversus Eb/N,, for rate 112 Viterbidecoding received data quantization. (The length 32 path curve is
8-level quantizedsimulationswith32-bitpaths,andinfinitely
finely quantized transfer function bound, K = 4, 6, 8. identical to the I< = 5 curve in Fig. 5 . ) Performance with
842 IEEE TRANSACTIONS ON COMMUNICATION TECHNOLOGY. OCTOBER 1971

10-2

x
r(
-4
.3
.D
n
:
.b
1"-3
0
$4

w
u
d
m

1
8-LEVEL QUANTIZATION 2-LEVEL QUANTIZATION

3 4 5 6 7

%/No in db

Fig. 8. Performancecomparison of Viterbidecodingusingrate

1/2, K = 5 code with 2-, 4-, and 8-level quantization,Path 3 4 5 6 7 8
length = 32 bits.
EblNo i n db

length 32 paths is essentially identical to that of an in- Fig. 9. Performmcecomparison of Viterbidecoding using rate
finite path decoder. Even for a path length of only 16, 1/2, K = 5 code with 8-, 16-, and 32-bit path lengths and 2- and
%level quantization,
thereisonlyasmalldegradationinperformance. As
previously mentioned, other simulations have shown that
a path length of 4-5 constraint lengths is sufficient for
other constraint lengths as well.
Codedsystems thatmake use of receiver outputs
quantized to more than two levels require an analog-to-
digitalconverter at the modemmatchedfilter.output,
with thresholds that depend on correct measurement of
the noise variance. Since the level settings are effectively
controlled by the automatic gain control (AGC) circuitry
in the modem, it is of interest to investigate the sensi-
tivity of decoder performance to an inaccurate or drift-
ing AGC signal. Fig. 10 shows the decoder performance
variation as a function of A-D converter level threshold
spacing. I n all cases, the thresholds are uniformly spaced.
Thesesimulationsusethe K = 5 rate 1/2 codewith
E , / N , , = 3.5 dB. It is evident that Viterbi decoding per-
formance is quite insensitive to wide variations in AGC
gain. In fact, performance is essentially constant over a 0.3 0.4 0.5 0.6 0.7
range of spacingfrom 0.5 to 0.7. This allowsfora
variation in AGC gain of better than t 2 0 percent with Quantizer Threshold Spacing
no significant performance degradation.
Fig. 10. Viterbidecoderbiterrorrateperformanceasfunction
of quantizerthreshold level spacing; K 5, rate 1/2, Ea/No =
C. Performance of Codes of Other Rates 3.5 dB, 8-level quantization with equally spaced thresholds.
The preceding simulation results have concentrated on
Viterbidecoding of rate 1/2 convolutional codes. The
HELLER A N D JACOBS: DECODING FOR SATELLITES 843

UPPER BOUND

\
\
\

3 4 5 6 7

E IN n
i db (signal energy t o noise ratio)
b o

Fig. 11. Performance of rate 1/3, K = 4, 6 , and 8 codes with

Viterbi decoding. Fig. 13. Performance of rate 2,/3 K = 4 code with Viterbi decod-
ing. Numerical bound and simulation results.
10-2
results on performance fluctuation due to decoder param-
eter va,riation carry over to other code rates with minor
changes.
Coderates less than1/2buy improvedperformance
atthe expense of increasedbandwithexpansionand
more difficult symbol tracking due to decreased symbol
energy-to-noise ratios. Rates above 1/2 conserve band-
width but are less efficient in energy.
x
4
Fig. 11 shows bit error rate versus E,/No performance
rl
n obtainedfrom simulations of Viterbi
decoding with
n
e 10-4
optimumrate 1/3, K = 4, 6, and 8 codes, and 8-
I4
level quantization.Figs.12and 13 show numerical
L4 bound and simulation performance results for rate 2/3
P
4J
4
K = 3 and K = 4 codes, respectively. Simulation curves
m
are for 2- and 8-level quantization, while the numerical
bound curves are for infinitely fine receiver quantization.
\ Comparingtheperformancedataobtainedthrough
\ simulations of Viterbidecoderswith rate 112 (Figs. 5 ,
6 , and 7 ) , andrate 1/3 codes, it isapparentthatthe
latter offers a 0.3-to-0.5-dB improvement over the former
UPPER B O W for fixed K , .in the range reported. This is close to the
\
improvement in efficiency of a channel with capacity 1/3
\ comparedwithone of capacity 1/2, and is therefore
10-6 1 I 1 I I i l I I I I
expected.
3 4 5 6 7 8
Comparison of the higher rate codes with the rate 1/2
Eb/No in db codes may also be made over the range spanned by the
simulationandanalyticaldata.Thefairestcomparison
Fig.12.Performance of rate 2/3 K = 3 code withViterbi de-
coding. Numerical bound and simulation results. is probably betweendecoderswithsimilarnumber of
844 IEEE TRANSACTIONS ON COMMUNICATION TECHNOLOGY, OCTOBER 1971
states, and hence similar decoder complexity. Thus, the 10-~ I )

K = 3 rate 2/3 data should be compared with the K = 5,

rate 1J2 data.
Fig. 14 shows theunionbounds on performancesfor
the rate 2/3, K = 3, and rate 1/2, K = 5 codes. Both
encoders have 16 states.The freedistance d f equals 7
for the rate 1/2 code and 5 for the rate 2/3 codes. At
very high EbjNO, the rate 1/2 must be superior. This is
becauseasymptotically, a t high Eb/MO,the error prob-
ability varies as

P, - ne exp (-d,E,/NJ = n, exp ( - 4 RNEb/NO)

where ne is the number of bit errors contributed by code-
,x
r(

c
10-6

words a t distance df. This gives therate l / 2 code an 0

advantage of about 0.2 dB in the limit. w

Y

In Fig. 14, the differencebetweenthotwocurvesis IO-

about 0.1 dBintheerrorprobabilityrange of to
Thissmalldifferenceisduetothefactthatthe
rate 2/3 codeused happensto be a particularly good
code; the value of n, is smaller for it than for the rate
1/2 code and this difference is significant even for F , as 10-8
small as lo-.

V. IMPERFECT
CARRIERPHASE
COYERENCE
Thusfarithas been assumdthatcarrierphaseis
kuownexactly at the receiver. I n realsystemsthisis
3 4 5 6 7 8
usually not the case. Oscillator instabilities and uncom-
C b l l u i n db
pensateddopplershiftsnecessitate closedloop carrier
phasetracking atthe receiver.Since thecarrier loop
Fig. 14. Biterrorprobability bound forrate 112, K = 5 , and
tracksanoisyreceivedsignal,thephasereferenceit rate 2/3, K = 3 code.
provides for demodulation will not be perfect.
An inaccurate carrier phase reference at the demodu-
written formally as
lator will degradesystemperformance. Inparticular a
constanterror + inthedemodulatorphase willcause
the signal component of the matched filter output to
suppressed by the factor cos + (see [4, ch. 7 1 ) .
be P, = f(?)
for a given code, receiver quantization, and Viterbi de-
coder.Since thecarrierphase is beingtrackedinthe
presence of noise the phase error + will vary with time.
The effect of an imperfect carrier phase reference on To simplify analysis, assume that the data rate is large
performanceisalwaysworsefor coded than uncoded compared to the carrier loop bandwidth so that the phase
systems: This is because coded systems are characterized errordoesnotvarysignificantlyduringperhaps 20-30
bysteepererrorprobabilityversus E,/N,, curvesthan information. bit times. Viterbi decoder output errors are
uncodedsystems. An imperfectcarrierphasereference typically several bits in length and are very rarely longer
causes an apparent loss in received energy-to-noise ratio. than 10-20 bitswhentheoveralldecoderbiterror
Since the coded curve is steeper, the loss in Eb/No de- probability is less than Therefore,the
phaseerror
grades error probability to a greater extent. Furthermore, is assumed to be constant over the length of almost any
unless care is taken in the design of the phase-tracking decodererror.Thisbeingthecase,thebiterrorprob-
loop,thephaseerrormightbehigher for the coded ability for a constant phase error +, can be written as
system than for an uncoded system, sipce loop perform-
ancemaydependupon E8/1V,,, whichissignificantly
smaller for coded than uncoded systems.
For convolutional coding with phase coherent demodu-
lation and Viterbi decoding, exact analytical expressions from (12) and (13j. Thisresultusesthefactthat re-
for bit error rate P, vcrsus Eb/lVO are not attainable. The ceived signalenergyisdegraded by cos2 +, If + is a
simulation results of the preceding section, however, de- randomvariablewithdistribution p (+) , the resulting
fine arelationshipbetween P, and E b / N O thatcan be error probability averaged on + is
HELLER ANDJACOBS: DECODING FORSATELLITES 845

3 4 5 6 7 8 9 10 11 12 13 14 15
Eb/No in d b

Fig. 15. Performance curves for rate 1/2; K = 7 Viterbi decoder with 8-level quantization as a function of carrier phase
tracking loop signal-to-noise ratio a.

a parameter (the R = 7, rate 172 simulation curve of Fig.

5 was extrapolated t o get the high E6/NOresults shown in
For the second-order phase-locked loop this figure). These curves exhibit the same general shape
as those for uncoded binary PSK modulation with phase
_ a COR m
a >> 1 coherenceprovidedby acarriertracking loop. As ex-
pected, the losses due to imperfect coherence are some-
-
where I , ( ) is the zeroth order modified Bessel function whatgreaterwiththanwithout coding. Fig. 16 shows
and ,(Y is the loopsignal-to-noiseratio [ l l ] . Usingthis the
additional E b / N o required
maintain
to bit
a
distribution and the P, versus E b / N o curve for the K '= error rate as a function of loop signal to noise ratio .a.
7, rate 1/2 code of Fig. 5 , the P, integral of (15) has Curvesareshownforthecase of uncoded BPSKand
been evaluatedforseveralvalues of (Y.The results are rate l/i, K = 7 convolutional encoding-Viterbi de-
shown in Fig. 15 as curves of P,' versus E o / N owith as coding.
846 IEEE TRANSACTIONS ON COMMUNICATION TECHNOLOGY, OCTOBER 1971
tions of the decoder. For a rate l / w decoder, an ACS is
used to add the state metrics for two states to the appro-
priate branch metrics, to compare the resulting two sums,
and to select the larger. The decisionis transmitted to
the path memory section and the larger of the two sums
becomes a new state metric. One ACS function must be
performed for each of the 2K-1 states. In a fully parallel
very-high-speeddecoder, 2K-1 ACSunitsarerequired.
I n general,thespeed of the ACS unit places an upper
zo bound on the speed of the decoder. For slower decoders,
a

e.g., R less thanseveralmegabitspersecondfor T2L
c
.rl
logic, ACS units may be time shared, decreasing decoder
cost significantly. Complexity of the ACS unit is strongly
dependentuponrequireddecoderspeed. It shouldbe
noted that implementation of Viterbi decoders is greatly
simplified by the fact that all ACS units perform iden-
tical functions and can be realized by a set of identical
circuits.
The path memory section must store about a 4 con-
straintlengthhistory of decisionsforeach state.The
memoryrequirementsarethusnontrivial.Considerable
advantage can be taken of new integrated-circuits mem-
ories tokeeptheequipmentcostsmall.However,the
complexity of the path memory and the ACS units both
increaseby a factorslightlylargerthan 2 foreach
increaseinconstraintlength of 1. Thus, an increase in
10.0 11.0 12.0 13.0 14.0 15.0 system performance of about 0.4 dB at a bit error rate of
which can be achieved by increasing K by 1, comes
a a t a cost of slightlymorethandoublingdecoder com-
Fig. 16. Comparison to increase in Ea/No due to
imperfect plexity.
phase, coherence necessary tomaintain 10-5 bit error rate for Acompletedecoderalsomustincludeinterfacecir-
uncoded BPSK and K = 7; rate 1/2, Q = 8 Viterbi decoding.
cuits, synchronization circuits, timing circuits, and gen:
erally an encoder. A recent implementation1 of a K = 7,
VI. IMPLEMENTATION OF A VITERBIDECODER
rate1/2self-synchronizedViterbidecodercapable of
It isconvenient tobreakthebasicViterbidecoder operating at up to R = 2 Mbit/s with 2-, 4-,or 8-level
intofivefunctionalunits;aninput or branchmetric quantized data required a total of 356 TTL integrated
calculationsection, an ACS arithmeticsection,and a circuitsforallfunctions. As noted in Fig. 5, this rela-
pathmemoryandoutputsection.Informationcanbe tively simple decoder provides over 5-dB E b / N o advan-
thought of as passing successively from one section to the tageoveranuncodedBPSKsystem a t P, = and
next. 6-dBadvantageat PC = when softquantization-is
Thebranchmetriccalculationsectionacceptsthe used.
input data and calculates (or looks up)themetricfor
each distinct branch. For a rate 1/2 code, four branches VII.COMPARISON OF SEQUENTIAL AND VITERBIDECODING
are possible corresponding to transmission of 00, 01, 10,
and 11. For a rate 1/3 orrate2/3code,eightdistinct BothseqhentialandViterbidecoding offer practical
branchmetricsarepossible.Note that this is the only alternativestoacommunic,ationsengineerdesigning a
section of the decoder thatisdirectly concernedwith high-performance efficient communicationsystem. The
the number of bits of quantization of the received data, two decoders have significant differences which are noted
and hence, the only section whose complexity is directly below. Both are capable of very-high-speed operation.2
dependent on quantization. (The complexity of the ACS
section also depends on quantization indirectly, in that 1 TheLinkabit LV7026 decoder is designed for use withdif-
thenumber of bitsrequiredforstoringstatemetrics ferentially encoded BPSKor QPSK systems. It automatically
increases with the number of bits of quantization.) The chronization resolves demodulator phase ambiguities and establishes node syn-
withoutmanualintervention.
input section is generally not critical in terms of either 2 The Linkabit LS4157 sequential decoderiscapable of oper-
complexity or speedlimitations. Its complexitydoes ation at dataratesupto R = 50 Mbit/s. It uses aconstraint
length K = 41, .rate 1/2 code and acceptsonly hard quantized
double,however,foreachincrease of thedenominator data. The decoder is fully self-synchronizing. The coding advan-
of the rate R N by one. tage over uncoded data is 4.4 dB at P , = 10-5 at R = 50 Mbit/s
andgreaterthan6dB a t P , = 10-*. T h e coding advantage is
The ACS sections perform the basic arithmetic calcula- larger at, lower d a h rates.
HELLER A N D JACOBS: DECODING FOR SATELLITES 847
A. Error Probabilit!j and state metrics. No storageisrequired.Asequential
I t should be recalled that, since the complexity of se- decoder, on the other hand, must store several thousand
quential decoders is relatively independent of constraint branches of received data, eachbranchcontaining log?
length, the constraint length is typically made quite large &/RAT bitsfor rate R N and Q level quantization. Al-
to provide a very small probability of undetected error. though the possibility exists of gaining 0.4 dB by using
Usually the important contributor of errors is received rate 1/3 rather than rate 1/2 and of gaining2dBby
data bufferoverflow duetoacomputationaloverload. using soft decisions rather than hard, these advantages
Such an event causes a long burst of rather noisy output are bought in sequential decoding a t a formidable storage
data until the decoder reestablishes code synchronization. and processing cost. I n general, then, practical high-rate
During this burst, the probability of bit error is that of sequential decoders are limited torate1/2and hard
the raw channel, perhaps PC = 3 x decisions. (It is conceivable that this cost could be min-
Error from a Viterbi decoder occurs in short bursts of imized by operating the dec.oder a t a very high ratio of
length a t most10to 20. Systemsthataresensitiveto computation rate to average bit rate, thereby minimizing
longbursts of errorsshould thus useViterbidecoding. the number of branches required in the buffer.)
Systems that can tolerate occasional long bursts, with an A second argumentagainstsoftquantizationwith
error indication provided if desired by the decoder, should sequential decoding involves the sensitivity of the prob-
consider sequential decoding. ability of buffer overflow to channel variations. In Fig.
The curve of error probability versus E,/No tends to 10, it was demonstrated thatchangesinreceiverAGC
bemuchsteeperforasequentialdecoder than fora of *20 percent had negligible effect on the performance
Viterbi decoder because of the differencc in I<. Thus, the of a
Viterbi decoder. The degradation ismuchmore
sequential decoding advantage tends to increase as lower pronounced forsequential decoding, since the computa-
probabilities of bitcrroraredemanded,although,as tional load is very sensitive to changes in channel param-
before, manyerrorstendto come inwidelyseparated eters. Thus, part of the 2-dB gain anticipated for soft
noisy bursts. decisionsmight be lostunlessgreatcarewas exercised
in controllingreceiverAGCprecisely.
R. Decoder Delay I n comparingsequentialdecodingandViterbidecod-
Sequentialdecoderstendtorequirelong buffers of ing, it thus appears fair to consider soft decisions only
a t least 200 bits and as much as several thousand bits for theViterbidecoder.Undertheseconditions,the
tosmoothoutthevariations in computationalload. efficiency advantage of long a c,onstraint lengthse-
Viterbidecodersrequire a pathmemory of a t most 64 quential decoder is considerably diluted. Consequently,
bits. Thus the decoding delay differs by u p to two orders performance of arate 1/2, K = 41 sequentialdecoder
of magnitude. isnobetterthanarate1/2Viterbidecoder of con-
straint length 5 to 7 (depending on the speed factor, that
C. Long Tail Required to Terminate Sequences is, the ratio of computation rate to bit rate) at a P , of
I n time-divisionmultiplexedsystems,bursts of sepa- Thesequentialdecoder doesshow adistinctad-
rately encoded data may be received at the samedecoder vantage for P, of 1k8or smaller.
from different sources. I n these instances, it may be de- On the other hand, building a system without receiver
sirable to time share the decoder.As noted in Section 111, quantization lowers systemcosts,sinceaconsiderably
termination of encoding can be achieved by transmitting more crude AGC may be used.
aknownsequence of length K - 1, thuscausingthe
encoder toenteraknownstate. Since K is typically E. Sensitivity to Phase Error and Rtcrsty Conditions on
larger for sequentialdecoding,thetailing off of the the Channel
encoded sequence can cause a significant degradation. in Theperformance of Viterbidecodingunderslowly
system efficiency. The tailing off of the short Constraint fluctuatingphaseerrorwaspresentedinFig.15.A
length codes for Viterbi decoding causes a much smaller similar calculation would indicate much greater degrada-
degradation. tionin the case of sequentialdecoding,sincetheerror
If time and implementation permit the storage of the probabilitycurveismuchsteeper.Furthermore,this
decoder state without code termination, then the cost of estimate would be optimistic in the case of sequential
tailing off can be ignored.The design of suchatime- decoding, since the assumption that the phase varied so
sharedsequentialdecoderremainsforfuturework. slowly thaterrorsoccurredindependently would prob-
ably not hold for sequential decoding. Thus, more care-
D.Rates Other than 1/2 and soft Quantization ful design of the phase-tracking loop is indicated for a
Viterbi decoders for rate 1/3 and 8-level quantization system utilizing sequential decoding rather than Viterbi
arenotsignificantlymore complex thanthoseforrate decoding.
1/2 and 4- or 2-level quantization. The chief costs occur
in the input section of the decoder as discussed in Section VIII. CONCLUSIONS
VI. In particular, the soft quantized data are processed Viterbidecoding has been shown to be apractical
in the input section and then incorporated in the branch method for improving satellite and space communication
848
efficiency by 4-6 dB, a t a. biterrorrate of lo+. The
successful implementation of 2-Mbit/s constraint-length-7
Viterbi decoders effectively demonstrates that the tech-
niqueis well beyond thestage of being a theoretical
curiosity. In fact, a major effort has heen under u7ay for
thepast 2-3 yearswiththeaim of modifyingand
adapting the algorithm for minimum complexity imple-
mentation without sacrificing performance significantly.
I n addition, Viterbi decoding has been shown to de-
grade gracefully in the prescncc of advcrse channel or
receiverconditions. Inparticular,theerrorprobability
does not change precipitously with E b / N o as is the case
withcodingtechniques that uselonger codes and/or
requirevariabledecoding effort,such as sequential de-
coding. This ensures that performance degradation due
to an imperfect phase or bit timing reference, or a slight
correlation between noise samples, will be minimal. Re-
quirements on AGC accuracy, even : for soft decisions,
were shown to be quite loose.
Finally the results presented here should provide the
communication engineer with the information necessary
to evaluate the applicability of Viterbi decoding to space
and satellite communication systems with a wide range
of requirements and constraints.
REFERENCES
111 C . E . Shannon,Communication is the presence of noise,
Proc. IRE, vol, 37, Jan. 1949, pp. 10-21.
121 Codex Corp.,FinalRep. Codingsystem design for ad-
vancedsolar missions,? Contract NAS 2-3637, NASA Ames
Res. Cent., M0ffet.t Field, Calif.
[31 A. J. Viterbi,ConvolutioEal codes and their performance
in communicationsystems, this issue, pp. 751-772.
141 J. M. Wozencraft and I . M . Jacobs, Pritbciples of CowLmu-
7aication Engitreering. New York : Wiley, 1965.
[51 I. M. Jacobs. Sequential decoding for efficient communi-
cation from deepspace, I E E E Trans. Commun. Technol., program in association with the Cornell Aeronautical Laboratory,
val. COM-15, Aug. 1967, pp. 492-501.
[SI A. J. Viterbi, Error bounds for convolutional codes and an
asymptoticallyoptimum decoding algorithm. IEEE T ~ t n s .
Inform. Theory, vol. IT-13, Apr. 1967, pp. 260-269.
171 J. A. Heller, Short constraint lengthconvolutional codes, search Laboratory of Electronics. He was promoted to Associate
Jet Propulsion Lab., California Inst.Technol., Space Pro-
grams Summary 37-54, vol. 111, Oct./Nov.. 1968, pp. 171-
.I_
111.
181 -. Improved performance of short, constraint,
length
convolutionalcodes, Jet Propulsion Lab., California Inst. principally with coding for deep-space communications. I n 1966,
Technol.,Space ProgramsSummary 37-56, vol. 111, Fcb./
Mar. 1969. pp. 83-84.
[91 Linkabit Corp., Final Rep., Coding systems study for high San Diego. I n 1970 he was promoted to full Professor. I n 1968 he
datarate
telemetry links, Contract NAS2-6024, NASA
Ames Res.Ctr.Rep. CR-114278, MoffettField,Calif.
[lo] J. P. Odenwalder, Optimum decoding of convolutional
codes, Ph.D. dissertation, Syst. Sei. Dep., Univ. Californla,
Los Angeles, 1970.
1111 A. J . Viterbi, Principles of Coherent Communicnlion. New
York : McGraw-Hill, 1966.
IEEE: T R A N S A C T I O N S ON C O M M U N I C A T I O N T E C H N O L O G Y , OCTOBER 1971
[121 J . K. Omura,On the Viterhi decoding algorithm, ZEEE
Trans. Inform. Theory (Corresp.), vol. IT-15, Jan. 1969,
pp. 177-179.
Jerrold A. Heller (M68) was bornin New
York, N. Y., on June 30, 1941. He received
theB.E.E. degree in 1963 from Syracuse
University,Syracuse, N. Y., andthe M.S.
andPh.D. degrees in electrical engineering
from theMassachusettsInstitute of Tech-
nology, Cambridge, in 1964 and 1967, re-
spectively. During his firstyear at M.I.T.
he was a National Science Fonndation
Fellow.
I n 1965 he joined theM.I.T. Research
Laboratory of Electronics where he was a Xerox Fellow for the
following two years. He held summer positions in 1962 at the Bell
TelephoneLaboratories,New York,N. Y . , and in 1963 atthe
IBM Research Center, Yorktown Heights, N. Y., where he worked
on the logical design of digital systems. From 1967 to 1969 he was
with the Communications Research Section of the Jet Propulsion
Laboratory,Pasadena, Calif., where his workcentered on the
application of coding to deep-space communication. He is presently
Director of TechnicalOperations for theLinkabitCorporation,
San Diego, Calif. Currently, his work is concerned with coding for
space, IF, and communication satellite channels.
Dr. Heller is a member of Tau Beta Pi, Et,a Kappa Nu, and
Sigma Xi.
Irwin Mark Jacobs (S55-M60) was born in
New Bedford,Mass., on October 18, 1933.
He received the B.E.E. degree from Cornell
University, Ithaca, N. Y., in 1956, and the
S.M. and Sc.D. degrees from the Massachu-
settsInstitute of Technology,Cambridge,
in 1957 and 1959, respectively. He was the
recipient of a McMullin RegionalScholar-
shipand a GeneralElectricTeachers Con-
ference Scholarship a t Cornell and par-
ticipated
in
the engineering cooperative
Buffalo, N. Y. In graduateschool, he was a General Electric Fellow
and an Industrial Fellow of Electronics.
I n 1959, he was appointed Assistant Professor of Electrical
Engineering at M.I.T. and was a Member of the staff of the Re-
Professor in 1964. On leave from M.I.T., he spent the academic
year 1964-1965 as a NASA ResidentResearch Fellow at the Jet
Propulsion Laboratory, Pasadena, Calif., and was concerned
he accepted anappointment as Associat,e Professor of Applied
Physics and Information Science at the University of California,
cofounded Linkabit Corporation, of which he is now President.
He is presentlyonleave from the University of Californiaand
devoting full time t,o Linkabit Corporation. He is currently working
in the area of information and computer science.
Dr. Jacobs is a member of Phi Kappa Phi, Sigma Xi, Eta Kappa
Nu, Tau Beta Pi, and the Association for Computing Machinery.