632 IEEE JOURNAL ON SELECTEDAREAS IN COMMUNICATIONS, VOL. SAC-2, NO.

5 , SEPTEMBER 1984

Efficient Mo.duIation for Band-Limited

Channels
G. DAVID FORNEY, JR., FELLOW,
IEEE,ROBERT G. .GALLAGER, FELLOW,IEEE,GORDON R. LANG,
FRED M. LONGSTAFF, AND SHAHID U. QURESHI, SENIOR MEMBER,IEEE

Abstract-This paperattempts to presentacomprehensivetutorial decade. (It was not unknown ig this period to encounter
survey of the development of efficient modulationtechniques for band- users who thought that Nyquist or Shannon or someone
of advances in
limited channels, such as telephone channels. After a history
commercial high-speed modemsandadiscussionoftheoreticallimits, it else had proved that 2400 bits/s was about the maximum
reviews efforts to optimize two-dimensional signal constellations and pre- rate theoretically possible on phone lines.)
sents further elaborations of uncoded modulation. Its principal emphasis, The first commercially important 4800 bit/s modem was
however, is oncoded modulation twhniques, in which there is an explosion the Milgo 4400/48, introduced in about 1967,which in-
ofcurrent interest, both for researchand for practical application. Both cluded a manually adjustable equalizer to allowuseof a
block-coded and trellis-codedmodulation are covered, in a common frame-
work. A few new techniques are presented. nominal 1600 Hz bandwidth in conjunction with 8-phase
modulation to send 3 bits/Hz. The development of digital
adaptive equalization soon allowed expansion of the nomi-
nal bandwidth to 2400Hz,essentially the full telephone
I. HISTORICALINTRODUCTION line bandwidth. Following a first generation of single-
sideband 9600 bit/s modems in the late 1960's, which were
B AND-LIMITED channels (as opposed to power-
'limited) are those on which the signal-to-noise ratio is
high enough so that the channel can support a.number of
only marginally successful, broad success was achieved by
the Codex 9600C, introduced in 1971, which used quadra-
bits/Hz of bandwidth. The telephone channel(particularly ture amplitude modulation (QAM) with a 16-point signal
the dedicated private line) has historically been the scene constellation to send 4bits/Hz in a nominal 2400 Hz
of the earliest application of the most efficient modulation bandwidth. (16-point QAM,had been used at 4800 bits/s
techniques for band-limited channels. The reasons have to by ESE and ADS about 1970.) Modems with these char-
do both with the commercial importance of such channels acteristics remained the state of the art.for another decade,
and with the fact that they can be modeledto first order as and it was thought by many (including someof the present
a u t h o r s , who should have known better) that higher-speed
linear time-invariant channels,sharplyband-limitedbe-
tweentypically 300-3000 Hz,withhighsignal-to-noise modems would never be widely used commercially.
ratios, typically 28dB or greater.Theirrelatively low In 1980, a first generation of 14 400 bit/s modems was
bandwidth permits a great deal of signal processing per introduced by Paradyne (MP 14400), followed in 1981 by
transmission element, and thereforeearly application of the Codex/ESE SP14.4 andbyothers.Thesemodems
the most advancedtechniques, which have often then been simply extended 2400 Hz QAM modulation to 6 bits/Hz
applied several years later to broader-band channels ( e g , by using 64-point signal constellations, and by using ad-
radio). vanced implementation techniques and exploiting the grad-
The earliest commercially important telephone-line mod- ual upgrading of the telephone network, proved to operate
ems appeared in the 1950's and used frequency shift keying successfully over a high percentage of circuits. In a second
generation of 14.4 kbit/s modems that will begin appear-
to achieve speeds of300 bits/s (Bell 103), or 1200 bits/s
ing in 1984, coded QAM modulation is beingintroduced to
on private lines (Bell 202). The earliest commercially im-
portant synchronous modem was the Bell 201, introduced provide greater performance margins. The principal focus
in about 1962, which used 4-phase modulation in a nomi- of this paper willbe onthese new codedmodulation
nal 1200 Hz bandwidth to achieve 2400 bits/s on private techniques.
lines. This remained the state of the art for most of the This evolution of high-speed modems to ever higher bit
rates usingsuccessivelymorecomplicated modulation
schemes is summarized in Table I, along with the designa-
Manuscript received February 14, 1984; revised May 3,1984. tion and year of final adoption of CCITT international
G. D. Forney, Jr. iswith the Motorola Information Systems Group,
Mansfield, MA 02048. standards embodyingtheseschemes;.How far canthis
R. G. Gallager is with the Department of Electrical Engineering and evolution go? History would suggest caution in stipulating
Computer Science,Massachusetts Institute of Technology, Cambridge,
MA 02139. any ultimate ceiling. However, without any dramatic gen-
G.R. Langand F. M. Longstaff are with MotorolaInformation eral upgrading of the telephone network,we venture to say
Systems Ltd., (ESE) Rexdale, Ont., Canada M9V 1C1.
S. U. Qureshi is with the Codex Corporation, Canton, MA 02021. that 19.2 kbits/s is the maximum conceivable rate for a
0733-8716/84/0900-0632$01.00 01984 IEEE

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY TIRUCHIRAPALLI. Downloaded on November 14,2024 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 FORNEY et al. : MODULATION FOR BAND-LIMITED CHANNELS 633

TABLE I
MODEMMILESTONES

cos ,oct

x, , y i pulse
AF1-V
sequences; F pulses/sec. sin kt
LPF low-passfilter;bandwidth F / 2 Hz.
c: carrier frequency (radians/sec).
s(t):linesignal.

Fig. 1. Canonical QAM modulator.

Yt Y ‘I
c

( x , , y,):coordinates of 2dimensionalsignalpoint. 4n,

(na , nd: 2dimensional gaussian noise variable.
(x’~, y’J:2dimensionalreceivedpoint.

Fig. 2. QAM channel model.

telephone-line modem for general use, even withall-out use a regular rate of F .points/s where F Hz is the nominal
of the most powerful codedmodulation. We shall see. (Nyquist) bandwidth of the channel. A signal point is also
called a “symbol,” and the symbol interval is 1/Fs. The
11. CHANNEL
MODELAND BASIC LIMITS model indicates that the twosignal point coordinates
(xt,y,) are independently transmitted overdecoupled
A quadrature amplitude modulator can be used to gener- parallel channels and perturbed byGaussiannoise vari-
ate any standard linear double-sideband modulated carrier ables (n,,, n y , ) , eachwithzeromeanand variance N .
signal(including forms of phase modulation and
, Alternatively, we could regard the two-dimensional signal
phase/amplitude modulation), which includes all types of point as being perturbed by a two-dimensional Gaussian
modulationin generaluse in synchronousmodems.A noise variable. If the averageenergy (the mean squared
canonical QAM modulator’is shown in Fig. 1. value) of each signal coordinate is S , then the signal-to-noise
Assuming that the only channel impairment is Gaussian ratio is S / N .
noise and that the receiver achieves perfect carrier phase The simplest method of digital signaling through such a
tracking, the simple model of Fig, 2 applies.Signals are system is to use one-dimensional pulse amplitude modula-
sent in pairs (xt,y,); the channel is essentially two-dimen- tion (PAM) independently for each signal coordinate. (This
sional. We shall call such a pair a “signal point,” imagined is sometimes called narrow-sense QAM.) In PAM, to send
to lie on a two-dimensional plane. Signal points are sent at m bits/dimension, each signal point coordinate takes on

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY TIRUCHIRAPALLI. Downloaded on November 14,2024 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 634 IEEE JOURNAL ON SELECTED
AREAS IN COMMUNICATIONS, VOL.
NO.SAC-2, 5, SEPTEMBER 1984

one of 2” equally likely equispaced levels, conventionally In what follows we shall show that simple coding tech-
f1, k 3, f 5, -,f(2“ - 1). The average energy of each
+
niques gain about 3 dB or 1 bit/symbol, while the most
coordinate is then elaborate techniques described have theoretical gains the
of
order of 6 dB or about 2 bits/symbol.Thisisentirely
S, = (4” - 1)/3, consistent with the R , estimate given above, and suggests
and it follows that that little can begainedbyseekingstillmore elaborate
schemes. (In [6] the capacity of the telephone channel was
Sm+,=4Sm+1. estimated as of the order of 23500 bits/s, roughly con-
sistent with what we are saying here.)
That is, it takes approximately (asymptotically, exactly) 4
times as much energy (or 6 dB) to send an additional 1
bit/dimensionor 2 bits/symbol. The probability P ( E ) 111. UNCODEDMODULATIONSYSTEMS
that either x, or y, is in error is upperbounded and closely
approximated by the probability that the two-dimensional Digital QAM signaling schemes are conventionally and
Gaussian noise vector (n,,, n,,,) lies outside a circleof usefully represented by two-dimensional constellations of
radius 1, whichiseasilycalculated to be- P ( E ) = all possible signal points. A 2”-point constellation can be
exp( - 1/2N). A noise varianceN of about 1/24 therefore used to send n bits/symbol. A f a i r amount of effort has
yields anerror probability persymbol in the rangeof gone into finding “optimum” constellations. We shall
about 6 x shortly see that thepayoffforthis effort onpurely
Thechannel capacity of theGaussianchannel was Gaussian-noise channelsisrelativelyslight,althoughthe
- original papers [l].Subject to an
calculated in Shannon’s schemes found are helpful precursors for more elaborate
energy constraint x: Q S , the capacity is schemes.

C = (1/2)log, (1 + S / N ) bits/dimension, A . Rectangular Constellations

achieved when x, is a zero-mean Gaussian variable with
A brief flurry of ,theoretical papers in the early 1960’s
-variance S. Notethat when S becomeslargewith N
[7]-[lo]developedtwo-dimensionalsignal constellations
constant, it takes approximately (asymptotically, exactly)4
from various viewpoints.Themost interesting for our
times as much power to increase the capacity by an ad-
present purposes are the family of constellations developed
ditional 1 bit/dimension. Therefore the ratio of bits/&-
by Campopiano and Glazer [9], reproduced in Fig. 3. (We
mension achieved by PAM to channel capacity approaches
have taken the liberty of substituting a “cross constellation”
1 as S becomes large. This fact was used in [2] to argue
for theirs at n = 7; the two are equally good.) For even
that coding has little to offer on highly band-limited chan-
integer numbers of bits/symbol, the constellations are
nels.
simply representations of two independent PAM channels,
We can make a more quantitative estimate of the poten-
so the constellations are squareandhave points drawn
tial gains from codingas follows..Usingnarrow-sense
from the rectang4ar lattice of points withodd-integer
QAM (two-dimensional PAM) to send m bits/dimension
coordinates. It takes about 6 dBmorepowertosend 2
or n = 2m bits/symbol at an acceptable error rate (of the
more bits/symbol, as expected. For odd integer numbers
order of 10-5-10-6) requires an average energy in each of bits/symbol, the constellations lie withinan envelope in
dimension of
the form of a cross’ (and havehencecome to be called
S = (2“ - 1)/3 “cross constellations”) and the points are drawn from the
when N = 1/24, or S / N = 8 X 2” for n moderately large.If same rectangular lattice (except for the 8-point constella-
channel capacity could-be achieved, we could send about tion, where the outerpoints are put on the axes for
+
n’ = log, ( S / N ) bits/symbol, or about n 3 bits/symbol symmetry
that the
and
minimum
energy savings).
distance
With
between
the figures scaled so
any twopoints is equal
at the same signal-to-noiseratio. Thus, the potential gain is
about 3 bits/symbol or, alternatively, about a factor of 8 to 2, the average signal energy in absolute terms and in dB
(9 dB) of.power savings. is as given in Table 11. We see that the “cross constella-
Many authors (see, e g , [3], [4]) regard the parameter R, tions” require about 3 dB more or less than the next lower
as a better estimate than C of the maximum rate that is or higher square constellation, respectively, as we would
practically achievable using coding. Onthe Gaussian chan- expect.
nel, R, is [5] The Campopiano-Glazer construction. can be gen-
eralized as follows: from an infinite array of points closely
+
R, = (1/2)log, (1 S/2N) bits/dimension. packed in a regular array or lattice, select a closely packed
subset of 2“ points as a signal constellation.This important
It thus takes a factor of 2 (3 dB) more power to ‘signal at principle is at theroot of muchrecentwork. We shall
R , than to signal at C . The maximum practical improve- explore applications of this principle, working up from the
ment obtainable by coding might thereforebe estimated as simpler to the more sophisticated.
of the order of 6 dB, or 2 bits/symbol (although the R, When constellations are drawnfrom a regular lattice
estimate is not universally accepted). within some enclosing boundary, the following asymptotic

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY TIRUCHIRAPALLI. Downloaded on November 14,2024 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 FORNEY et al. : MODULATION FOR BAND-LIMITED CHANNELS 635

+ n=2

.. .. .. .. .. .. .. ..
.. .. .. .. .. .. .. ..
.. .. .. .. .. .. .. ..
.. .. .. .. .. .. .. ..
n-6
n-3 n-4

.. .. .. .. .. .. .. ..
.............
...........
............
............
............
........
........
n=7
. .. :I.. .. .

.. .. .. .. .. .. .. .. .. .. .. ..
............

... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
n=5

[a= P ~ ~ f Campopiano/Glazer

:I;
.f ::
...
o r constellations, n odd].
Fig. 4. Crossconstellationboundary.

.. .. .. .. .. .. .. .. .. .. .. .. .. .. .. ..
.. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. n=4cross(S=10)

................
................
.................
............... . .I .
................
................
.................
...............
......
........
1
n = 6 cross (S 41)

................. ........
........
n=8
........
........
........
Fig. 3. Rectangular signal constellations (after Campopiano and Glazer
[91).

TABLE I1
CAMPOPIANO-GLAZER CONSELLATIONS
...... t
n = 6 circular (S =40.9)
Fig. 5. Improvedrectangular constellations.

Comparisons between S and 9 are given in Table 11; the

approximations are good ones. Furthermore, note that the
approximation-isuseful and remarkably accurate for con- cross is slightly more efficient than the square, by a factor
stellations of even moderate size (16 points and up). Let A of 31/32 or 0.14 dB (because it is more like a circle); this
be the areaof the region consisting of points that are closer suggests that thecross would be the better shape even for' n
to a given lattice point than to any other (i.e., the decision even, and indeed this is the case and can be easily achieved
region belonging to that point, or the Voronoi (or Dirich- by taking alternate points from the nexthigher cross
let) region [ll]). If we take some boundary that encloses a constellation, as shown in Fig. 5 for n = 4 and n = 6. The
region in space of area B, then the number of lattice points n = 4 cross constellation is as good as the conventional
in that region will be approximately B / A , and the average 4 X 4 square constellation, and the n = 6 cross constellation
energy of these lattice points will be approximately equal is 0.1 dB better than the 8 X 8 square constellation, about
+
to the average energy ( x 2 y ') of all points within the as predicted.
boundary. Of course, the best enclosing boundary would bea circle,
For example, the Campopiano-Glazer constellations are the geometrical figure of least average energy for a given
drawnfrom a rectangular lattice with A = 4, andare area. A circle with radius R has area n R 2 andaverage
bounded by a square of side 2 X 2"12 and area B, = 4 X 2" energy equal to R2/2; setting R 2 = 4 X 2", we find that the
for n even, and by a cross of area B, = 4 X 2" for n odd, average energy for a 2"-point circular constellation ought
with dimensions as shown in Fig. 4. The energy calculated to be about
by the integral approximation'is S = (2/m)2" (circle)
= (2/3)2" (square) which would be only about n/3 or 0.20 dB better than the
= (31/48)2" (cross). square, or0.06 dB better than the cross. Fig. 5 also shows a

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY TIRUCHIRAPALLI. Downloaded on November 14,2024 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 636 IEEE JOURNAL ON SELECTEDAREAS IN COMMUNICATIONS, VOL. SAC-2, NO. 5, SEPTEMBER 1984

t
t .:j:. 4:::il ::-

.
n=2
S = 2 (3.0dB) n=3
..
S=4.3125 (6.3dB) n=4
S = 8.75 (9.4dB)

- - . .... . .1
m .
.......
. . . . f. . . . n=5
S = 17.59 (12.5dB)

.........
~ ~ ~ . .

.,....*,..
....
e . . 4

n=6
. m . n=3
SI 4.5 (6.5dB) ..I..
.......
n=6
S = 35.25 (15.5dB) S=35.4375(15.5dB)
Fig. 6 . Optimum hexagonal constellations.

more circular constellation for n = 6, with the outer points was actually used in a 4800 bit/s Hycom modem in the
moved tothe axesas in Campopiano-Glazer’s 8-point mid-1970’s. There was a great deal of attention to n = 4
constellation; this constellation has been used in the Para- structures in the early 1970’s because.of their importance
dyne 14.4 kbit/s modem and, like the n = 6 cross, is about in 9600 bit/s modems;the rather strange-lookingone
0.1 dB better than the n = 6 square. shown here was apparently first discovered by Foschini et
al. [14], and is still the best 16-point constellation known.
B. Hexagonal Constellations The n = 6 suboptimal structure is used in the Codex/ESE
SP14.4 modem.)
As the densest lattice in two dimensions isthe hexagonal
lattice (try pennypacking), constellations usingpoints from
a hexagonal lattice ought to be the most efficient. Indeed, Iv. ELABORATIONS
OF UNCODED MODULATION
the area of the hexagonal Voronoi region for a hexagonal
lattice with minimum distance 2 is 2 6 = 3.464, or 0.86.6 In this section we shall discuss further variants of un-
the size of the square region,whichaccordingto our coded modulation: constellations with nonuniform proba-
approximation principle-should translate to a 0.6 dB gain bilities, higher-dimensionaluncodedconstellations, and
for a hexagonal’constellation over a rectangular one with constellations for nonintegral numbers of bits/symbol.
the same boundary. (A hexagonal boundary, as suggested
in [12], has an energyefficiencywithin0.03dB of the A. Nonuniform Probabilities
circular boundary, or 0.03 dB better than the cross.)
Fig. 6 shows the besthexagonalpackings for n = 2 Attainment of the channel capacity bound requires that
through 6. For n > 4, the predicted 0.6 dB gain is effec- the signal points have a Gaussian probability distribution,
tively obtained over the best rectangular packings.(Histori-. whereas with all the constellationsof the previous sectionit
cal notes: suggestions that thehexagonal lattice would is implicit that points are to be used with equal probabil-
asymptotically be the best were made very early; see, e.g., ities. A uniform circular distribution of radius R has
[13]. The suboptimal n = 3 “double diamond” structure average energy S, = R 2 / 2 and entropy H , = log, vR2; a

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY TIRUCHIRAPALLI. Downloaded on November 14,2024 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 FORNEY et al. : MODULATION FOR BAND-LIMITED CHANNELS . 631

TABLE I11
ENERGY
SAVINGS FROM N-SPHERE MAPPING

N Gain dB
2 1.05 .m
4 . 1.11 .Ai
8 1.18 .73
18 1.25 .88
24 1.29 1.10
32 1.31 1.17
48 1.34 1.26
64 1.35 1.31

Fig. 7. Source-coded
constellation. Gain = n(N+2) [(N/2)l]-zN (for N even)
12

two-dimensional Gaussian distribution of variance u 2 in tion, ( n ! ) - l / n goes to e/n), or 1.53 dB; the improvement
each dimension has average energy Sg = 2 a 2 and entropy over N = 2 goes to e/2 or 1.33 dB, as computed above.
Hg = log, 277ea2. Thus, This is because for large N the probabilities of points in
any twodimensionsbecomenonuniform and ultimately
,
H , = log 2 sS, Gaussian. (It seemsremarkable that a purely geometric
Hg = log, res,. fact like the asymptotic ratio of the second moment of an
N-sphere to that of anN-cubecanbederivedfrom an
To yield the same entropy, the Gaussian distribution re- information-theoretic entropy calculation in 2-space, but so
quires a factor of e/2 =1.36 (or 1.33 dB)lessaverage it can.)
energy than the circular distribution. Implementation of such a schemealsoinvolvesadded
Implementation of a constellation withnonuniform complexity that mayoutweigh the performance gain. To
probabilities presents a number of practical problems. One send n bits/symbol in N dimensions (assuming N even),
possible way of achieving some of the potential gain is to incoming bits must be grouped in blocks of Nn/2. Some
divide the incoming data bits into words of nonuniform sort of mappingmust then bemade into the 2Nn/2 N-
length according to a prefix code, and then to map the dimensional vectors with odd-integer coordinates (assum-
prefix code words into signal points drawn as before from ing a rectangular lattice) which have least energy amongall
a regular two-dimensional lattice. The probability associ- such vectors. This can rapidly become a huge task; and a
ated with a prefix code word of length t bits is then 2-f. corresponding inverse mapping must be made at the re-
For example, Fig. 7 gives a set of prefix code words and a ceiver. Compromises can be madeto simplify the mapping,
mapping onto the hexagonal lattice that yields an average at the cost of some suboptimality in energy efficiency: e.g.,
energy of S = 7.02 while transmitting anaverage of 4 the cross is an effectivecompromisebetween the square
bits/symbol, an improvement of close to 1 dBover the and the circle in two dimensions.
best n = 4 uniform code known. Of course, the fact that
the number of data bits transmitted per unit time is a C. Nonintegral Number of Bits/Symbol
random variable leads to system problems (e.g., buffering,
delay) that mayoutweighanypossibleimprovement in It is sometimes desirable (as we shall see in Section VI)
signal-to-noise margin. to transmit a nonintegral number of bits/symbol. Since in
general an additional 1 bit/symbol costs aboutan ad-
B. Higher -DimensionalConstellations ditional 3 dB, it ought to be possible to send an additional
1/2 bit/symbol for about 1.5 dB. In this section we give a
.It is possible to achieve the same gain in another way by simple method that effectively achieves such performance.
coding blocks of datainto higher-dimensional constel- The method can begeneralized to other simple binary
lations without going to the true block coding to be de- fractions at the expectedcosts, but we shall omit the
scribed in later sections.(By “true coding,” we refer to generalization here.
schemes in which’the distance betweensequences in a To send n + 1/2 bits/symbol, we proceedasfollows.
higher number of dimensions is greater than that between Use a signal constellation comprising 2” “inner points”
points in two dimensions.) We have already seen in Section drawn from a regular grid, such as any of those of Section
111-A that a small (0.2 dB) gain is possible by going from 111, and an additional 2“-’ “outer points’’ drawn from the
one-dimensional PAM to two-dimensional QAM and samegrid and of as little averageenergy as possible,
choosing points on a two-dimensional rectangular lattice subject to whatever symmetry constraints may be imposed.
from within a circular rather than a square boundary. In Incoming bits are then grouped into blocks of 2n + 1 bits
the same way, by going to a higher number N of dimen- and sent in two successive symbolintervals as follows. One
sions and choosing points on an N-dimensional rectangu- bit in the block determines whether any outer point is to be
lar lattice from within an N-sphere rather than an N-cube, used. If not, the remaining 2n bits are used, n at a time, to
further .modest savings are possible.Table I11gives the select two inner points. If so, then one additional bit selects
energy savings possible in N dimensions,basedon the which of the two signals is to be an outer point, n - 1 bits
difference between average energy of.an N-sphere versus select which outer point, and the remaining n bits select
an N-cube of the samevolume. Note that as Ngoes to which inner point for the other signal. (That is, at most one
infinity, the gain goes to se/6 (by the Stirling approxima- outerpoint is sent.) Withrandom data,the average

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY TIRUCHIRAPALLI. Downloaded on November 14,2024 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 638 IEEE JOURNAL ON SELECTED
AREAS IN COMMUNICATIONS, VOL.
NO.SAC-2, 5, SEPTEMBER 1984

.. .).
. . ..

1; ;
additional energy is needed for signaling. (In practice,
. :I: . . . . ..... some of the “coding gain” may be lost dueto there being a
largenumber of sequences at distance d,, from the
correct sequence and therefore a large number of possibili-
ties for error, called the “error coefficient” effect. We shall
nS4 not be able to discuss the “error coefficient” much in this
S1=10.S0=26. g=14 (11.5dE) n-5
S,-ZO. So-54, 8-28.5 (14.5d6) paper, but offer some general remarksat the end of Section
.. .. .. .. VIII.)
Conventional coding techniques cannot be’directly ap-
.. .. .. .. . .,. . . . . .
plied in conjunction with band-limited modulation tech-

... ... ... ... ... ... ... ... .. .. .. .. .. .. .. ..

niques, at least withsignificantgain. (In 1970-1971, at
least four companies prototyped conventional coding
.. .. .. .. .. .. .. .. .. .. .. .. .. .. .. ..
schemes for use in high-speed modems; two of the compa-
nies failed, and two shortly withdrew their products from

n=6 ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
the market.) In recent years, however, a number of effec-
tive coding techniques have been developed for such appli-
S,-42. S0=102.~=57(17.ME)
.... ....
.... .... cations. The most important point to be made in this paper
....
... .. .I..
is that all of these coding schemes can be developed froma
common conceptual principle. This principle was set forth
n=7
dearly by Ungerboeck [15], who called it “mapping by set
S,=82. S.=218,~=116(20.6dE) partitioning,” although its roots mayperhapsbe found
Fig. 8. Constellations to send ( n + $) bits/symbol. elsewhere as well.We describe it in this section, and in
succeeding sections then use it to develop all known and
some new coding schemes, both block and trellis.
energy is 3/4 the average energy of inner points plus 1/4 We shall consider only 2-dimensionalconstellations with
the average energy of outer points. Fig. 8 shows constella- points drawn froma 2-dimensional rectangular grid. (From
tions of 24,48, 96, and 192 points that can be used in such research to date, we cannot find any advantage to starting
schemes for 4 < n < 7; the average energy in all cases for with hexagonal grids when higher orders of coding are to
+
n 1/2 bits/symbol is approximatelyhalfwaybetween be used.)
+
that needed for n and that for n 1 bits/symbol. Thus, Such a constellation can be divided into two subsets by
these constellations are intermediate between the Campo- assigning alternate points to each subset; i.e., according to
piano/Glazer constellations in the same way that the cross the pattern
constellations are intermediate between the squares. (In
fact, it can be shown that the 2-dimensional cross constel-
lations can be derived from 1-dimensional PAM constella- . . . . . . . . . . .
tions with “inner” and “outer” points in an analogous
I
. . . A B A B A . . .
way-) . . . B A B A B . . .
V. CODINGFUNDAMENTALS
. . . A B A B A . . .
. . . B A B A B . . .
Heretofore we havebeenconcernedwithmethods of
mapping input bits tosignal point constellations in two or . . . A B A B A .
more dimensions, where in higher -dimensions the bits . . . . . . . . . . .
simply lie on the lattice that is the Cartesian product of
two-dimensional rectangular lattices, so that the distance
between points in N-space is no different from that in two The resulting two subsets ( A and B ) have the following
dimensions. Now we shall begin to discussmethods of properties.
coding of sequences of signal points, where for the pur- a) The points in each subset lie on a rectangular grid
poses of thispaper we meanbycoding (or “channel (rotated 45” with respect to the original grid).
coding”) the introduction of interdependencies between b) The minimum squared distance between points within
sequences of signal points such that not all sequences are a subset is twice the minimum squared distance [ d : ] be-
possible; as a consequence, perhaps surprisingly, the tween points in the original constellation.
minimum distance d,, in N-space between two possible Furthermore, because of the first property, the partition-
sequences is greater than the minimum distance do in ingcanbe repeated to yield4, 8, 16, . . . subsets with
2-space between two signalpoints in the constellation from similar properties, and in particular within-subset squared
which signal points are drawn. Use of maximum likelihood distances of 4, 8, 16,. . times d:. Fig. 9 shows the
sequence detection at the receiver yields a “coding gain” of 64-point square constellation divided into two subsets of
a factor of d L / d ? in energyefficiency,lesswhatever 32 points, 4 subsets of 16 points, 8 subsets of 8 points, and

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY TIRUCHIRAPALLI. Downloaded on November 14,2024 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 FORNEY et a[.: MODULATION FOR BAND-LIMITED CHANNELS 639

A B A B A B A B
B A B A B A B A
A B A B A B A B
Selector
“Channel Coding”

pling “source coding” from “channel coding” and is im-

portant bothconceptually and in implementation.
We now show howthis general scheme can beapplied to
16 subsets of 4 points. The subset nomenclature is accord- both block and trellis codes, with performance approach-
ing to the following binary tree: ing the R , estimate.

ORIQINAL CONSTELLATION

This nomenclature is different from that of Ungerboeck VI.BLOCK CODES

[15], whouses a more natural subscript notation that
reflects the successive binary partitions; our nomenclature In Section IV-B we saw what could be achieved by using
will be useful in the next section, where we shall use the points froman N-dimensional rectangular lattice and using
fact (evident by inspection) that at both the 4-subset and an N-sphere, rather than an N-cube, as a boundary.
8-subset levels, the minimumsquared distance between The rectangular lattice comprising all N-dimensional
two A subsets, say, is 2d: times the Hammingdis- vectors with odd-integer coordinates is not the most densely
tance between their subscripts: e.g., d2(A,, A , ) = 2d:; packed for any N greater than 1; for example, for N = 2,
d2(A,, Aol) = 2d:; d2(A,, All) = 4d:; and so forth. the hexagonal lattice is 0.6 dB denser, as we haveseen.
These subsets may then be used to implement relatively Finding the densest lattice in N dimensions is an old and
simple but effective coding schemes, illustrated in general well-studied problem in the mathematical literature. Table
in Fig. 10. Certain incoming data bits are encoded in a IV gives the densest packings currently known for all N up
binary encoder, resulthg in a larger number of coded bits. to 24 and selectedlarger N , with the improvement in
The coded bits are then used to select which subsets are to packing density over the rectangular lattice given in abso-
be used for each symbol. The remaining incoming bits are lute and in dB terms [16]. The dimensionsN = 4, 8,16, and
not coded, but merelyselect points from the selected 24 are locally particularly goodand are known to be
subsets, with the signal constellation chosen large enough optimum in the sense of being the densest possible lattice
to accommodate all incoming bits. The coding scheme is packing in these dimensions.
thus more or less decoupled from the choice of constella- A body of recent work [ll], [12], [18]-[21] generalizes the
tion, as long asit is of the rectangular gridtype. The Campopiano-Glazer construction to N dimensions by tak-
coding gain is effectively determined by the distance prop- ing all points on the densest lattice in N-space that lie
erties of the subsets combinedwiththose of the binary within an N-sphere, where the radius of the sphere is
code, regardless of the size of the constellation. On the chosen just large enough to enclose 2mNpoints, to send rn
other hand, the constellation size, boundary, symmetries, bits per dimension. Thesecodes obtain a “coding gain”
and other “ uncoded” properties such as were investigated over PAM whichis a combination of both the lattice
in earlier sections are more or lessindependent of the packing density gain of Table IV (“channel coding”) and
coded bits and are determinedby the mapping of the the N-sphere/N-cube boundary gain of,Table I11 (“source
remaining bits. This may be regarded as effectively decou- coding”). The resulting coding gains achievable for N = 2,

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY TIRUCHIRAPALLI. Downloaded on November 14,2024 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 640 IEEE JOURNAL ON SELECTEDAREAS IN COMMUNICATIONS, VOL. SAC-2, NO. 5, SEPTEMBER 1984

TABLE IV TABLE V
SAVINGS
ENERGY FROM DENSE
LATTICES COMBINED ENERGY
SAVINGS

N Density Gain dB N Gain dB

1 .500 1.00 .w 2 1.21 .82

2 ' ' ,289 1.15 .&? 4 1.57 1.96
3 ,177 1.26 1.00 8 2.36 3.74
4 ,125 1.41 1.51 16 3.54 5.50
5 .oBB 1.52 1.81 24 5.16 7.12
6 ,072 1.67 2.21 32 5.24 7.19
7 1.81 2.58 48 8.04 9.04
s ,063

,063 2.00 3.01 8.89 64 9.40.

10
11 ,035 2.18 3.38
12 ,037 2.31 3.83
13 ,035 2.39 3.78
14 .036 2.49 3.96
15 044 2.64
~. 4.21
16 .& 2.83 4.52
points,. and further, in which the 16 subset subscripts (each
17 ,063 2.89 4.60
,072 2.99 4.75
18
19 ,088 3.10 4.91 subset nowhavingtwo subscripts) are codewords in the
20 ,125 3.25 5.11
21 ,177 3.39 5.30 (16,ll) &tended Hamming code, whose minimum Ham-
22 ,289 3.57 5.53
23 ,500 3.77 5.76 ming distance between codewords is4.
24 1.000 4.00 6.02
32 1." 4.00 6.02
6.19
The 24-dimensiond lattice consists of all sequencesof 12
36 2.000 4.16
40
48
16.000
16832.947
4.59
6.00
' 6.62
7.78
points in w h h all points are either A points or B points;
-64 4194304.000 8.44 8.09
the 24 ( i , j J subscripts are codewords in the (24,12) Golay
code, known to have minimum Hamming distance 8, and ,
further, in which the third subscripts k are constrained to
4, 8, 16, 24, 32, 48 and 64 are shown in TableGV. Because satisfy an overall parity check in the following way: if the
the numljer of near neighbors in thesedenselypacked sequence is of all A points, then overall k parity is even
lattices becomesverylarge, the total number of error (an even.number are equal to l), while if the sequenceis of
events ("error coefficient") becomes large, which reduces all B points, overall k parity is odd.
the coding gain realized in practice. Also, the mapping of If the minimum squared distance between points in the
all mN bit combinations to their corresponding signal 2-dimensional constellation .is d:, then the minimum
poihts can be an enormous task, even if all possible sym- squared distance between points (sequences) in these
metries and simplificationsare cleveriy exploited[20], [21]. higher-dimensional lattices can be shown to be 2d:, 4d2,
We will now show that certain of these dense N-dimen- 8d:, and 16d:, respectively, as follows.
sional lattices.can be constructed using 2-dimensional rect- a) A sequence of A points and a sequence of B points
angular lattices, the subset partitioning idea, and simple differ from each other by squared distance at least d: in
binary block codes. In particular, we shall give construc- every point and therefore by at least 2d:, 4d2, 8d:, and
tions for N = 4, 8, 16 and 24 that form a natural sequence 12d: in total. In fact, in the 24-dimensional casethere is a
both in complexity and iri nominal coding gain (respec- distance of at least 5d: in at. least one symbol, so the
tively 1.5, 3.0, 4.5, and 6.0 dB, using the simplest imple- minimum squared distance between A sequences and B
mentations). (Cusack[21a]hasrecentlyshownhow to sequences is at least 16d:. The proof depends on the
construct dense 2"-dimensional lattices from 2-dimensional properties of the Golaycode as well as the particular
lattices using Reed-Muller codes, for any n; for N = 4, 8, partitioliing shown in Fig. 9 and is in the Appendix.
and 16, the lattices obtained are the same as those we b) Two different sequenceswith points allfrom the
obtain here.) r: same sequence of subsets must differ in at least one point
To generate the optimumN-dimensional lattices for by the minimum within-subset squared distance, which is
N = 4, 8, 16, and 24, we shall use sequencesof 2, 4, 8, and 2d:,4d:, 8d2, or 16d:, respectively. This is all we need to
12 points from the 2-dimensional rectangularlattice, parti- establish 2d: as the minimumsquared distance between
tioned as shown in Fig. 9, into 2,4, 8, and 16 subsets, sequences in the 4-dimensional case.
respectively. c) For N = 8,16, and 24, the i or (i, j ) subset subscripts
The 4-dimension& lattice is generatedbytakingall are drawn from (4,3), (16,11), or (24,12)codeswith
sequences of two points in which both"points come from minimum Hamming distances 2, 4, and 8, respectively. By
the same subset, ice.,sequences of the form ( A , A ) or the relation between subscript Hamming distance d, and
(B, B). subset squared distance d; = 2d,d,2given in Section V,
The 8-dimensional lattice consists of allsequences of two sequences with points drawn from subsets of the same
four points jn which all points are either A ,points or B type ( A or B ) but different i or ( i , j ) subscripts must
points and further in which the 4 subset subscripts satisfy differ by squared distance at least 4d:, 8d?, or 16d:,
+ +
an overall parity check, i, + i, i , i, = 0; e g , sequences respectively. This is all we need for the 8- and 16-dimen-
of the form ( A o +I,,A,, Ao),( B o ,B,, Bo,B,), and so forth. sional cases. ' '

(In other words, the.subscripts must be codewords in the d) For N = 24, two sequences of points from subsets of
(4,3) single-parity-check block
code,
whose minimum the same type and withthesame (i; j ) subscripts but
Hamming distance between codewordsis 2.) different k subscripts must differ byat least 8d: in at least
The 16-dimensional lattice consists of all sequences of two symbols because of the overall k parity check, and the
eight points .in which all points are either A points or B fact that the minimum squared distance between points of

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY TIRUCHIRAPALLI. Downloaded on November 14,2024 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 FORNEY et al.: MODULATION FOR BAND-LIMITED CHANNELS 641

the same type and with the same (i, j ) subscripts is 8d:. the parity-check constraint. Repeat, assuming that B points
This concludes the 24-dimensional proof. were sent, to get the best B sequence. Compare the best A
To send m bits/symbol using these lattices, we need to and B sequences, and choose the better as the final deci-
encode a block of mN bits into one of 2mNlattice points. sion.
To maximize coding gain, the 2"N lattice points of least Detection for N=16' or 24 is harder, involving either
energy should be chosen; however, implementation of the generalization of soft-decision decoding of the (16,ll) or
mappings from bits to points and viceversabecomes (24,12) block codes to perform error correction on tenta-
complex.Simplermethods willnow begiven,using the tive decision subscripts (as above; the method of changing
binary codes used to construct the lattices, and the constel- the least reliable decision if an overall parity check fails
lations either of Fig. 3 (for N = 8 and 24) or of Fig. 8, was used in the 1950's as a soft-decision error-correction
along with the method of sending half-integral numbers of method for single-parity-check codes by Wagner), or ex-
bits/symbol given in Section IV-C (for N = 4 and 16). The haustive search of a neighborhood of the received sequence
cost in coding giin is relatively small, ranging from a few in N-space. Note that as its first step the decoder can
tenths of a decibel for N = 4 or 8, up to about 1 dB for alwayschoose the closest point in each subset to each
N = 24; it is upperbounded by the N-sphere/N-cube gain received point as representative of that subset, there being
given in Table 111. no reason to prefer anymore distant point, and then
Of the block ofmN bits, we alwaysuseone bit to proceed to determine the bestsequence of subsets using
specify whether A or B points will be used. For N = 8, 16, those points (with their distances from the received point)
and 24, a further set of bits is used as input to a binary as proxies for the corresponding subsets; the decoding task
block coder, which produces appropriate codewords to be may thus be partitioned in the sameway as coding is
used as subset subscript designators: 3 bits to produce 4 partitioned in Fig. 10.
for N = 8 , l l bits toproduce 16 for N = 16, and 12 11= 23 +
+
bits to produce 24 '12 = 36 for N = 24. Thus, a total of 1,
4, 12, or 24 incoming bits are used as in Fig. 10 to select VII. TRELLIS
CODES
the subsets, or i, 1, 15, or 2 bits/symbol, respectively.
The remaining bits are used to select points from the On power-limited channels (such as the satellite channel),
selected subsets. We use the rectangular constellations of convolutional coding techniques have more or less become
Figs. 3 and 8 as follows: for N = 4, constellations of the standard (although there are somewho continue to
1.5 X 2"' points as in Fig. 8, divided into two 1.5 X 2"-'- champion block codes [23]). Generally, anything that can
point subsets A and B; for N = 8, constellations of 2 m + 1 be achieved with a block code can be achieved with some-
points as in Fig. 3, divided into four 2"-'-point subsets; what greater simplicity with a convolutional code. We have
for N = 16, constellations of 1.5 X 2"'+' points as in Fig. 8, just seen that relatively simple ( N = 8)blockcodes can
divided into eight 1.5 X2"-2-point subsets; and for N = 24, achieve of the order of 3 dB coding gain on band-limited
constellations of 2m+2points as in Fig. 3, divided into 16 channels, and relatively complex ( N = 24) block codes can
2"-'-point subsets. (In all cases m must be large enough achieve of the order of 6 dB. We shall now see that trellis
so that the subsets resultingfrom the partitioning have codes can,do the same, perhaps a bit more simply.
equal size and other desired properties, e.g., symmetries.)
,Thesecanbeusedtosendm-f, m - 1 , m - l + , o r m - 2 A . Ungerboeck Codes
bits/symbol, where for N = 4and 16, the method of
sending half-integral bits/symbol of Section IV-C may be For band-limited channels, the trellis codes camefirst, in
used. the work of Ungerboeck [15]. Ih Ungerboeck's paper, to
Since the minimum squared distance in N-space is 2, 4, send n bits/symbolwithtwo-dimensional modulation, a
8, or 16 times the minimum squared distance for an constellation of 2"+' points is used, partitioned into 4 or 8
uncoded 2'"-point constellation, there is a distance gain of subsets. 1or 2 incoming bits/symbolenter a rate-+ or rate-#
3, 6, 9, or12 dB,respectively.However, the expanded binary convolutional encoder, and the resulting 2 or 3
constellations required with coded modulation cost 1.5, 3, coded bits/symbol specify which subset is to be used. The
4.5, and 6 dB, respectively, yielding a net coding gain of remainingincoming bits specifywhich point from the
1.5, 3, 4.5, and 6 dB for N = 4, 8, 16, and 24. The family selected subset is to be used.
relationship of this progression of codes is apparent. The coding gain obtainable increases withthe number M
Maximumlikelihoodsequence detection of the lattice of states in the convolutional encoder. Ungerboeck's sim-
point closest to a sequence of received points is easy for plest scheme usesa 4-state encoder and achieves a nominal
N = 4 and 8. General methods are given in [22]. For N= 8, 3. dB coding gain (a factor of 4, or 6 dB, in increased
given four received points, assume first that A points were sequence distance, less 3 dBdue to use of the larger
sent. Find the closest A point. to each received point, and 2"+l-point constellation). His most complex scheme usesa
check subscript parity of the four subsets tentatively de- 128-state encoder and gains 6 dB (the limit with 8 subsets
cided. If the panty checkfails,change the least reliable and a 2"+'-point constellation since the within-subset dis-
decision to the next closest A point (which must be in the tance is 8d2 for a 9 dB gain, less the 3 dB due to the larger
other A subset). This gives the best A sequence satisfying constellation). Table VI gives the coding gains obtained by

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY TIRUCHIRAPALLI. Downloaded on November 14,2024 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 642 IEEE JOURNAL ON SELECTEDAREAS IN COMMUNICATIONS, VOL. SAC-2, NO. 5, SEPTEMBER 1984

TABLE VI
CODING
GAINSFOR UNGERBOECK CODES
States Gain dB

4 2.0 3.0
8 ' 2.5 4.0
16 3.0 4.8
32 3.0 4.8
€4 3.5 5.4
128 4.0 6.0
N=4

4 -State Trellis

N=8
Fig. 12. Trellises corresponding to N =4 and N = 8 block codes.

2-state trellises. The N =16 block code can similarly be

associated with a 64-statetrellisdecomposable into two
parallel 32-state trellises, m d the N = 24 block codecan be
associated with a 2 X 4096-state trellis.)
8-state trellis codes with nominal4 dB cbding gainare in
&State Trellis the process of being adopted as international CCITT
Fig. 11. Ungerboeck 4-'and 8-state trellis codes. standards for 9600 bit/s transmissionovertheswitched
(dial) telephone network [25]and potentially for 14.4 kbit/s
transmission over private lines aswell [26]. A slight variant
[27j of the Ungerboeck scheme involving a nonlinear con-
Ungerboeck for these and intermediate numbers of states. volutional encoder is being used in these standards; with
Decoding is assumed to be by the Viterbi algorithm [24], thisvariant, whosetrellisisshown in Fig. 13, a 90"
a maximum likelihood sequence estimation procedure for rotation of a coded sequence is another coded sequence, so
any trellis code.Thecomplexity of such a decoderis that differential coding techniques may be used. The dis-
roughly~proportional tothe number of encoder states. With tance properties and therefore coding {gainof the variant
these codes,each b r q c h & the trellis corresponds to a are apparently identical to those of Ungerboeck's 8-state
subset rather than to an individual signal point; but if the scheme.
first step in decoding is to determine the best signal point
within each subset (the one closest to the received point), B. Other Trellis Codes
then that point and its met& (squared distance from the
received point) can be used thereafter for that branch, and The Ungerboeck codes seem to cover the range of possi-
Viterbi decoding canproceed in a conventional manner. ble coding gains with complexity of the order of what we
Fig. 11 gives trellises with branches labeled by subset for might expect, and may therefore be taken as benchmarks
Ungerboeck's 4-state and 8-state codes. ofhow muchcomplexityisneeded to achieve different
(Note: the, block codes of the previous section can be coding gains in the 3-6 dB range. Can they be improved
represented as trellises; Fig. 12 showsthetrellises corre- upon? From ourresearch, the answer seemsto be: yes, but
sponding to the N = 4 and N = 8 codes. Viterbi decoding not verymuch. In thissection we shalldescribetwo
could therefore be used for them as well. It is interesting schemes that exhibit modest improvements and some new
that the block code with 3 dB coding gain is also associated ideas: a 2-state code that has a nominalcoding gain of
with a 4-state trellis, albeit decomposable into two parallel almost 3 dB, and an 8-state trellis code with a coding gain

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY TIRUCHIRAPALLI. Downloaded on November 14,2024 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 FORNEY et al.: MODULATION FOR BAND-LIMITED CHANNELS

... m, A A A
.e.
643

Fig. 13. Nonlinear 8-statetrellis code with 90' symmetry.

of 4.5 dB. The new idea in the 2-state scheme is to use

subsets that arepartially overlapping and partially distinct. Fig.14.
2-state trellises. (a) Infinite
nonredundant2-state
trellis.
(b) 2-state trellis terminated every 4 symbols. (c) Time-invariant2-state
The new idea in the 8-state scheme is to use a 4-dimen- trelhs with A' # A", B' # B".
sional constellation rather than 2-dimensional as the basic
constellation. (4-dimensional trelliscodeshavealsobeen
studied by Wilson [28] and Fang et al. [28a].)
1) 2-state Trellis Code: If a 2"-point signal constellation
is partitioned into two subsets of A points and B points,
and channel codingisdoneusing the (infinite) 2-state
trellis shown at the top of Fig. 14,it would appear at first
glance that a 3 dB gain is obtained at no cost. T h i s scheme
6
B"'A E A B A E A B I A '

sends n bits/symbol with no signal constellation expan-

sion; further, any two sequences that start at one common
node and end atanother differ by a squared distance of at
least 2d:, because the paths differ by at least d: when Fig.15.80-point signalconstellation.32-pointsubsets A', A", B', B"
have 24 common points, 8 uniqugoints; e.g., A'= 24 A inner points
they diverge and another d: whentheymerge, so the plus 8 A' outer points. Sisi=31.3; So= 80; S = 43.5 (16.4 dB).
nominal codinggain is apparently a factor of 2 or 3 dB.
Of course, we cannot get something for nothing, and the
fallacy in this scheme(called "catastrophic error propa- for the bitlost at the fourth symbol, and allows a full 3 dB
gation" in the convolutional coding literature) is that there gain. This terminated trellis codemaybe regarded as a
are paths of infinite length starting from a common node generalization of a single-panty-checkblock code.
that neverremergeandhavesquared distance only d:, Another way of gaining almost 3 dBwhileusing a
namely any two paths of the form AXYZ . . * and BXYZ time-invariant trellis code is as follows. The signal constel-
- . (The "error coefficient"is infinite.) lation is modestly expanded to include (1 p)2" points, +
One way of curing this problem is to terminate the trellis arranged on a rectangular grid and divided as usual into A
+
every b symbols by forcing it to a single node, illustrated in and B subsets of (1 p)2"-l points each. The A and B
Fig. 14(b). In other words, at the bth symbol, the subset is subsets are further divided into (1 - p)2"-, "inner" points
constrained .to be A or B, as necessary to reach the and 2 p "outer" points. Finally, sets A', A", B', and B",
designated node. Only n - 1 bits can be used to determine each of 2"-' points, are created asfollows: A' and A"
the bth symbol, so there is a cost of 1 bit per b symbols of both include the (1-p)2"-' inner A points, but each
transmission capacity, but now a legitimate coding gain of includes a different half' of the 2p outer points; and
3 dB is obtained minus ( l / b ) x 3 dB for the rate loss. The similarly with B' and B". Thus A' and A'' are partially
8-space block code wouldoperate in justthis way if it used overlapping and partially disjoint, and so are B' and B".
only A points, partitioned into A , and A , (see the top half The probability that a random choice from A' will not bea
of its trellis in Fig. 12); happily it is possible to insert a member of A" is p. Such a construction is illustrated for
similar code made up of B points into the interstices of the n = 6 and p = 0.25 in Fig. 15.
A code lattice without compromising distance, and the Now the sets A', A", B', and B" may be used in a
additional bit involved in specifying A or B compensates 2-state trellis as illustrated in the last part of Fig. 14. The

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY TIRUCHIRAPALLI. Downloaded on November 14,2024 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 644 IEEE JOURNAL ON SELECTED
AREAS IN COMMUNICATIONS, VOL.
NO.SAC-2, 5, SEPTEMBER 1984

squared distance between two paths beginning and ending

at commonnodesremains 2d,’ since the basic distance
properties between A subsets and B subsets remain. But
now,although there are still pairs of infinite sequences
comprisingonly inner points that start from a common
node andnever accumulate distance of more than d:, their
probability is zero.The squared distance betweenany
branch that uses an outer point and its counterpart branch
must be at least 2df since the counterpart branch cannot
use the same outer point and the distance between differ- Fig. 16. Encoderfor8-statefour-dimensional code.
ent points in the same subset is at least 2d:. For practical
purposes, this means that a sequence contaihng sucha
branch cannot be confused with a sequence containing the subsets, then again the squared distance is at least 4df.
counterpart branch, their squared distance being at least Supposenow that the signal constellation contains 2”
3d:, so that in effect, whenever an outer point is sent, points. Then, over two symbol intervals, 2n - 1 bits enter
reconvergence to a common node is forced, as in the trellis the modem and one parity check is generated, giving 2n
termination method. Here, however, the reconvergence is bits to select the two signal points. Sigce n - 4 bits/symbol
probabilistic and happens on average every l / p symbols, enter the modem, there is a loss of 1.5 dB due to the larger
e.g., every 4 symbols if p = 0.25. I
signal constellation and a gain of 6 dB in distance for a net
There is a slight reduction in power due to the increased nominal coding gain of 4.5 dB.
constellation size; e.g., the average power usingthe constel- Section IV discussed encodingfor a half-integral number
lation of Fig. 15 is 43.5 or 16.4 dB, versus 42 or .16.2 dB for of bits/symbol, and that method can be applied here. An
the 8 X 8 constellation, or 41 (16.1 dB) for the Fig. 5 n = 6 alternative which is somewhat more attractive is to have an
cross constellation. Use of an integral approximation gives integer number n of bits/symbol enter the modem, with
an estimate of additional power required of a factor of three bits entering the convolutional encoder each pair of
l’+p 2 ,*or1.0625 (0.26 dB) for p = 0.25. This can be made symbols and 2n - 3 bits entering a prefix-code source
as small as desired by reducing p ; the cost, however, is a coder as described in SectionIV. This both yields an
greater average time to converge and a greater average integer number of bits/symbol and also gains some of the
number of near-neighbor sequences (“error coefficient”) possible 1.33 dB for nonuniform probabilities.
increasing inversely proportional to p ; for this code the
“error coefficient” is rather large and must be taken into VIII. CONCLUDING
REMARKS
account.Any p greater than zero in principle avoids
catastrophic error propagation: a p of about 0.25 seems a It has not been possible in this paper to cover a number
good choice in practice. of topics that are of importance in practice.
2) 8 -State Trellis Code: For the 8-state four-dimensional The only channel disturbance considered has been white
scheme,we shall use a two-dimensional rectangular grid Gaussian noise. Other disturbances are usually controlling
divided into four subsets as before. The binary convolu- on telephone channels. There is some accumulating experi-
tional encoder for this scheme, however, operates on pairs ence that the coded modulation schemes are often more
of symbols rather than singlesymbols. An appropriate robust relative ,to uncodedschemes than wouldbe pre-
encoder is shown in Fig. 16. During each pair of symbol dicted byGaussiannoise calculations against someim-
intervals, three bits enter the encoder and four coded bits portant disturbances, such as nonlinear distortion and
are produced. The first two coded bits select the subset for phase jitter, perhaps due to the memory inherent in coded
the first symbol and the second two’bits select the subset modulationand sequence estimation over multiple sym-
for the second symbol. bols.
If the Hamming distance betweentwoencodedse- Because of the symmetries of attractive constellations,
quences is K , then the squared distance between the map- e.g., the 90” symmetry of most of our rectangular constella-
pings onto grid points is at least Kdf. We now show that tions, there may be an ambiguity in phase at the receiver.
the minimum free Hamming distance of this convolutional In general, there are twoways to handle this ambiguity
code is 4. First note that the response of the encoder to a withcoded modulation. If the code is such that,on a
single 1 on any input line is a sequence with even weight, sequence basis, every 90” rotation of a code sequence is
from which it follows that all encoded sequences have even another legitimate code sequence, then it will be possible
weight and the minimum free distance is even. By inspec- by differential quadrantal coding to make end-to-end
tion, it is easy to verify that there is no encoded sequence transmission. transparent to 90” rotations. Alternatively, if
of weight 2, and a simultaneous 1 on all inputs yields an 90” rotations do not give valid code sequences in general,
encoded sequence of weight 4. Thus, the squared distance then it will be possibleeventually to detect this and to
between any twosequences corresponding to different force receiverphase to a valid setting. The former tech-
encoded outputs is at least 4d:. If the encoded outputs are nique is generally preferred. Of the codes we have dis-
the same, but different elements are chosen from the same cussed, the block codes generally are differentially codable

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY TIRUCHIRAPALLI. Downloaded on November 14,2024 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 FORNEY et al. : MODULATION FOR BAND-LIMITED CHANNELS 645

and the trellis codes are generally not, although they can b) For A points, i = 0 iff y = Omod4, and j = 0 iff
often be modified to be; e.g., the modification of Fig. 11 x = Omod4; for B points, i = 0 iff y = 1mod4, and j = 0
shown in Fig. 13. iff x = 1mod4.
Finally, we have mostly used nominal coding gain as a c)For A points, x + y = 2i + 2 j + 4 k m o d 8 ; for B
figure of merit of coding schemes. In fact, error probabili- points, x + y = 2 i + 2 j + 4 k + 2 m o d 8 .
ties for coded systems on Gaussian channels are typically Two sequences with points from different groups differ
[24] of the form P( E ) = Kexp( - E ) , where the exponent by at least 1 in every one of the 24 (x, y ) coordinates. They
E is governed by the nominal coding gain and the “error cannot all differ only by 1,however, because of the follow-
coefficient” K is of the order of the number of coded ing. The sum S of all coordinates satisfies
sequences at minimum distance from an average trans-
mitted sequence. In general, the error coefficient
S = 2wij + 4wk modulo8, for A sequences;
a) increases with the complexity of coding; 2wij + 4w, modulo8, for B sequences
b) can cost asignificant fraction of a dB for coding
schemes with moderate (3-4 dB) gain, for error probabili- where wij is the number of (i, j ) subscripts equal to 1,and
ties in the 10-s-10-6 range; wk the number of k subscripts equal to 1. Butsince the
c) can become very large for schemes withlarge (6 + dB) ( i , j ) subscripts formaGolaycodewordand all such
gain, such as the blockcodeswith N = 24, or the most words have weights equal to integer multiples of 4, and
complex trellis codes; and since wk is even for A sequences and odd for B sequences
d) is generally significantly larger for block codes than by construction,
for trellis codes with comparable nominal coding gain.
S = 0 modulo 8, for A sequences;
Thus the error coefficient cannot be ignored in a more
4modulo 8, for B sequences.
detailed assessment of coded systems.
Now suppose that there were a B sequence that differed
from an A sequence by + 1 in every coordinate, and let m
VIII. SUMMARY be the number of coordinates in which the difference was
+ 1. Since the sum SA of the A coordinates is Omod8 and
On the bad-limited channel, dense packing of 2-dimen- the sum S, of the B coordinates is 4mod8, and SA - S, =
sional constellations with optimal (circular) boundaries m -(24- m) = 2mmod8, it follows that 2m = 4mod8, or
yields less than 1dB improvement over simple pulseampli- m = 2mod4. Now, the construction of the array is such
’ tude modulation. Uncoded schemes in higher dimensions that if a B point has an x coordinate 1 larger than the x
or, alternatively, sourcecodingcangainsomewhatmore coordinate of an A point, then the j subscript is the same,
than 1 dB by using signal points with nonuniform proba- whereas if it is 1 smaller, then the j subscript is different;
bilities. These gains pale by comparison with what can be similarly a difference of + 1 in y gives the same i sub-
obtained with (channel) coding,whererelatively simple script, while a difference of - 1 gives the opposite one.
block or trellis codes easily yield coding gains of the order Thus, m = wij. But, wij= Omod4, so m = Omod4; con-
of 3 dB, or 1 bit/symbol. Relativelycomplexblock and tradiction. Hence, any B sequence must differ from every
trellis codes have been constructed that yield of the order
of 6 dB, or 2 bits/symbol. Because this is as much gain as
+
A sequenceby at least 3 in one coordinate, Q.E.D.
This lattice and its distance properties wereoriginally
would be predicted using the R , estimate and is only 3 dB discovered by Leech [29].
below the capacity limit, it seemsunlikely that further
majorimprovements are possible.However, within the
REFERENCES
spectrum of performance of already known schemes, there
will likely be some further embellishments that will reduce C. E. Shannon andW.Weaver, A Mathematical Theory of Com-
implementation complexity or have other desirable proper- munication. Urbya, IL: Univ. Illinois Press, 1949.
G. D. Forney, Jr., Coding and its applicationin space communi-
ties, such as the differentially coded variant of Un- cations,” IEEE Spectrum, vol. 7, pp. 47-58, 1970.
gerboeck’s 8-state trellis code that is likely to become an J. M. Wozencraft and R. S. Kennedy, “Modulation and demodula-
tion for probabilistic coding,” IEEE Trans. Inform. Theory, vol.
international standard. IT-12, pp. 29?;297, 1966.
J. L. Massey, Coding and modulation in digital communications,”
in Proc. 1974 Int. Zurich Seminar Digital Commun., Zurich,
Switzerland, pp. E2(1)-E2(4), 1974.
APPENDIX J. M. Wozencraft and I. M. Jacobs, Principles of Communication
Engineering. New York:Wiley, 1965.
PROOF THAT24-SPACE LATTICEHASd&, = 16d: R. W. Lucky, J. Salz, and E. J. Weldon, Jr., Principles of Data
Communications. New York:McGraw-Hill, 1968, ch. 3.
If the grid of triple-subscripted signal points illustrated C. R. Cahn, “Combined digital phase and amplitude modulation
communication systems,” IRE Trans.Commun.Syst., vol. CS-8,
in Fig. 9 is rotated 45’ with a point A , at the origin, and pp. 150-154, 1960.
scaled so that the coordinates (x,y ) of all points are I. C. Hancock and R. W. Lucky,“Performanceof combined
amplitude and phase-modulatedcommunication systems,” IRE
integers, then the following hold true. Trans. Commun. Syst., vol. CS-8, pp. 232-237, 1960.
a) For A points, both x and y are even; for B points, C. N. Campopiano and B. G. Glazer, “A coherent digital ampli-
tude and phase modulation scheme,” IRE Trans. Commun. Syst.,
both x and y are odd. vol. CS-10, pp. 90-95, 1962.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY TIRUCHIRAPALLI. Downloaded on November 14,2024 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 R. W. Lucky and J. C. Hancock, “On the optimum perfo.rmance tor of Technology and Planning of the Motorola Information Systems
of N-ary systems having twodegrees of freedom,” IRETrans. Group, Mansfield, MA. During the 1971-1972 academic year he was a
Commun. Syst., vol. CS-10, pp. 185-192, 1962. Visiting Scientist at Stanford University, Stanford, CA, and he has also
J. H. Conway and N. J. A. Sloane, “Voronoi regions of lattices, served as an Adjunct Professor at the Massachusetts Institute of Technol-
second moments of polytopes, and quantization,” ZEEE Trans.
Inform. Theory, vol. IT-28, pp. 211-226, 1982. ogy.
J. H. Conway and N. J. A. Sloane, “A fast encoding method for Dr. Forney was Editor of the IEEE TRANSACTTONS ON INFORMATION
lattice codes and quantizers,” ZEEE Trans. Inform. Theory, vol. THEORY from 1970 to 1973 and has held numerous other IEEE positions.
IT-29, pp. 820-824,1983. He was winner of the 1970 Information Theory Group Prize Paper Award
R. W. Lucky, “Digital phase and amplitude modulated communi- and the 1972 Browder J. Thompson Memorial Prize Paper Award. He was
cation systems,” Ph.D. dissertation, Purdue Univ., Lafayette, IN, elected to membership in the National Academy of Engineering in 1983.
1961 p. 89 (attributed to Cahn).
G. J. Foschini, R. D. Gitlin, and S . B. Weinstein, “Optimization of
two-dimensional signal constellations in the presence of Gaussian
noise,” ZEEE Trans. Commun., vol. COM-22, pp. 28-38, 1974.
G. Ungerboeck, “Ch,annel coding with myltilevel/phase signals,”
ZEEE Trans. Inform. Theory, vol. IT-28, pp. 55-67, 1982.
J. Leech and N. J. A. Sloane, “Sphere packings and error-cor-
recting codes,” Can. J . Math., vol. 23, pp. 718-745,1971.
G. R. Welti and J. S. Lee, “Digital transmission with coherent
four-dimensional modulation,” ZEEE Trans. Inform. Theory, vol.
IT-20, pp. 497-502, 1974.
Canada, “An aspect of future modem development,” CCITT Con-
trib. COM-Sp.A NO. 142, 1974. Robert G. Gallager (S’58-M61-F168)received
N. J. A. Sloane, “Tables of sphere packings and spherical codes,” the S.B. degree in electrical engineering from the
ZEEE Trans. Inform. Theory, vol. IT-27, pp. 327-338, 1981. University of Pennsylvania, Philadelphia, PA, in
P. deBuda, “Encoding and decoding algorithms for an optimal 1953, and the S.M. and Sc.D. degreesin electrical
lattice-based code,” in ZCC Conu. Rec., p. 65.3. 1-65.3.5, 1981. engineering from the Massachusetts Institute of
Canada, “Block coding for improve8 modemperformance,” Technology, Caybridge, MA, in 1957 and 1960,
CCITT Contrib. COM-XVII No. 112, 1983. respectively.
E. L. Cusack, “&or control codes for QAM signalling,” Electron. From 1953 to 1954 he was a Member of the
Lett., V O ~ .20, pp. 62-63, 1984.
J. H. Conway and N. J. A. Sloane, “Fast q;antizing and decoding Technical Staff at Bell Telephone Laboratories,
algorithms for lattice quantizers and codes, ZEEE Trans. Inform. and from 1954to 1956 wasin the Signal Corps of
Theory, vol. IT-28, p p 227-232, 1982. the U.S. Army. He has been with the Massachu-
E. R. Berlekamp, The technology of error-correcting codes,” setts Institute of Technology since 1956 and was Associate Chairman of
Proc. ZEEE, vol. 68, pp. 564-593,1980, the faculty from 1973 to 1975. He is currently a .Professor in the
G. D. Forney, Jr., “The Viterbi algorithm,” Proc. ZEEE, vol. 61, Department of Electrical Engineering and Computer Science, Associate
pp. 268-278,1973. Director of the Laboratory for Information and Decision Systems, and
Rapporteur on 9600 bit/s duplex family modem, “Draft recom- Co-chairman of the Department Area I (Control, Communication, and
mendation V.32 for a family of 2-wire, duplex modems operating Operations Research). He is a consultant to Codex Corporation, Western
at data signalling rates of up to 9600 bit/s for use on the general..
switched telephone network and on leased telephone-type circuits,” Union, and CNR. He is, the author of the textbook Znformafion Theory
CCITT Contrib. COM-XVII No. T30,Geneva, Switzerland, Mar. and Reliable Communication (New York: Wiley, 1968).His major research
1984: also see J. D. Brownlie and E. L. Cusack. “Dudex transmis- interests are data communication networks, information theory, and
sion ’at 4800 and. 9600 bit/s on the general ’switched telephone communication engineering.
network and the use of channel Coding with a partitioned signal Dr:Gallager was awarded the IEEE Baker Prize Paper Award in 1966
constellation,” in Proc. Zurich Znt. Sem.“Digital Commun., 1984. for the paper -A simple derivation of the codingtheorem and
[26] ~ ~ ~ ~lyic ~ $ ~ ~o d e v~ ~ ~ of a r ~ izp’i
$ ~ $c l
applications.” ~He has~been a member
~ of the Administrative Committee of
on poht-to-point 4-wire
leased telephone-type circuits,” C C I n the IEEE On Information from 1965 to 1970 and from
Study Group XVII,
1984. 1979 to the present, was Chairman of the Group in 1971, and was elected
[27] L. F. Wei, “ Rotationally invariant convolutional .channel coding a member of the National Academy of Engineering in 1979.
with expanded signal space-Part 11: Nonlinear codes,” ZEEE J .
Select. Areas Commun., vol. SAC-?: pp. xxx-xxx, Sept. 1984.
[28] S . G. Wilson and H. A. Sleeper, Four-dimensional modulation
and coding: an alternate to frequency reuse,” Univ. Virginia, Rep.
UVA/528200/EE83/107, Sept. 1983.
[28a] R. Fang,and.W. Lee, “Four-dimensionally coded PSK systems for
combattmg effects of severe IS1 and CCI,” in Proc. 1983 ZEEE
Globecom Conu. Rec., pp. 30.4. 1-30.4.7, 1983.
[29] J. Leech, “Notes on sphere packings,” Can. J . Math., vol. 19, pp.
251-267,1967.

G. David Forney, Jr. (S’58-M61-F’73) was born

in NewYork, NY, on March 6, 1940. He re-
ceived the B.S.E. degree in electrical engineering
from Princeton University, Princeton, NJ,in
1961, and the M.S. and Sc.D. degrees in electrical
engineering from the Massachusetts Institute of
Technology, Cambridge, MA, in 1963 and 1965,
respectively.
In 1965, he joined the Codex Corporation, and
became a Vice President and Director in 1970.
Since 1982he has been Vice President and Direc-

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY TIRUCHIRAPALLI. Downloaded on November 14,2024 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.
 IEEE JOURNAL ON SELECTED
AREAS IN COMMUNICATIONS, VOL. SAC-2, NO. 5, SEPTEMBER 1984 647

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY TIRUCHIRAPALLI. Downloaded on November 14,2024 at 18:06:34 UTC from IEEE Xplore. Restrictions apply.