BANDWIDTH AND POWER EFFICIENCIES OF

TRELLIS CODED MODULATION SCHEMES

Stefan Ramseier, Christian Schlegel
Stefan Ramseier, Ascom Tech Ltd., Gewerbepark, CH-5506Migenwil, Switzerland
Christian Schlegel, Univ. of South Australia, The Levels, South Australia, 5095, Australia

ABSTRACT In this paper we discuss and compare ad- width/power tradeoff. In Section 2, bandwidth- and power-
vances in bandwidth efficient coding and signalling tech- efficiency will be defined. In Section 3, the unified system
niques, especially trellis-coded modulation (TCM). We con- model is presented. In Section 4, we present uncoded, con-
sider power efficient TCM schemes, such as Ungerbock volutionally coded, trellis-coded (TCM), correlative trellis
TCM, correlative trellis-coded modulation (C-TCM) [l], coded (C-TCM) and coded overlapped quadrature modula-
which is optimized with respect to bandwidth efficiency, tion schemes, and their respective power- and bandwidth-
and coded overlapped quadrature modulation [2], which uses efficiencies.
overlapping pulses. The bandwidth/power tradeoff of these
schemes is presented, thereby illuminating the different de- 2. Bandwidth and Power Efficiency
sign philosophies. The parameter which characterizes how efficiently a sys-
tem uses its allotted bandwidth is the bandwidth efficiency
1. Introduction q. It is defined as
The rapid increase in the volume of radio communications
Bit Rate
has quickly made bandwidth and power two of the most
treasured resources. While bandwidth is important because '= Channel Bandwidth W
[bits/s/Hz]. (1)
it essentially determines the number of independent chan- While it is obvious how to define the channel bandwidth
nels which can be used simultaneously, power determines W for Nyquist signalling, as we will see in Section 4, this be-
the quality of transmission and is of primary concern wher- comes more arbitrary for modulation schemes with infinite
ever it is expensive to generate, like in satellite systems, or spectral occupancy. We chose the 99% bandwidth definition
even more though, in deep space probes and battery pow- in this paper, i.e., we define W such that
ered terminals. While these resources can be traded for each
other for certain applications, one is generally interested in
using them jointly as efficiently as possible.
In designing digital transmission systems, one has natu-
1,W
IA(f)12df = 0.99J_,
03

IA(f)I2df, (2)

rally focussed on the one resource more at premium. In or- where A ( f ) is the spectrum of the transmitted signal. This
der to conserve bandwidth, elaborate pulse shaping schemes 99% bandwidth corresponds to an out-of-band power of -
were proposed, from Nyquist pulses to continuous phase 20dB.
modulation, CPM [3]. Shannon's information theory [4], on In order to compare different transmission systems, an-
the other hand, by predicting the existence of power-efficient other parameter, which expresses the power efficiency, has
codes started the highly successful branch of coding theory, to be considered also. This parameter is often chosen to be
used successfully at overcoming power limitations. For quite the bit error probability Pa.
some time, these two disciplines were pursued rather inde- Shannon [4] showed that on an additive white Gaussian
pendently and it was conjectured that coding had little to noise (AWGN) channel, the maximum bandwidth efficiency,
offer on bandwidth-limited channels [5], thus consigning it the Shannon limit, is given by
to power limited channels.
Then, however, with the introduction of trellis coded mod-
ulation (TCM) [6,7,8], which combines modulation and cod-
17" = log, (1 + 5) [bits/s/Hz], (3)
ing, power savings of up to 6dB were realized without ex- where SIN is the signal-tenoise power ratio. The signal-
panding the bandwidth [9]. In the wake of this discovery to-noise power ratio can be expressed as SIN = REb/NoW,
many researchers have presented results in this new area of where Eb is the average bit energy, No is the spectral noise
coded modulation (see references in [7] and [SI). It is now power density, and R is the bit rate. The Shannon limit
believed that the gain of 6dB is very close to the maximum then becomes
one can expect [6,10].
Because most results on bandwidth efficient coding are
(4)
presented and derived in a discrete signal space, the influ-
ence of pulse shaping, and how the dimensions of the signal Since R/W = qmax is the limiting spectral bit rate, we ob-
space fit into the available bandwidth, are not considered
tain a bound on the minimum bit energy required for reliable
explicitly in most papers on the subject. transmission, given by the Shannon bound
In this paper we will present uncoded and several types
of coded modulation schemes in the light of both power 29m.x -1
and bandwidth performance, and we will discuss the band- Eb/No 2 (5)
Bmax
Q 1993 IEEE
0-7803-0917-0/93$03.00
1634
 3. System Model trellis-coded modulation (C-TCM) [l], where bandwidth-
In this section we present the complex baseband model of efficiency was the optimization criterion. We also discuss
our communications system. The transmitter consists of a the performance of coded overlapped quadrsture modula-
shift-register followed by a mapper and a pulse shaping filter, tion [2], a TCM scheme which uses overlapping pulses, i.e.,
and the receiver includes a matched filter and a maximum intentional intersymbol interference.
likelihood sequence estimator (MLSE). 4.1 Nyquist Signalling
One of the most straightforward implementations of a dig-
ital transmission system is a linear, uncoded system. It can
be represented by our system model of Fig. 1 by removing
the shift register, leaving only the mapper and the pulse
shaping filter. It produces the output signal [ll]
M

s(t) = u(n)q(t - nT),

Figure 1: Schematic diagram of the L-stage shift-register where the sequence u(n) is the output sequence of the map-
encoder/modulator per with input symbols z(n), and q ( t ) denotes the signal
pulse with spectrum Q(f ) .
The transmitter can generate generalized coded q u a d m In the sequel, we restrict ourselves to Nyquist pulses q(t),
ture amplitude modulation signals. Its schematic diagram is i.e. pulses that satisfy [ll]
shown in Fig. 1. The transmitted complex baseband signal
s ( t ) is generated by feeding an independent and identically
distributed (iid) M-ary data sequence at a symbol rate 1/T
n=--oo
into an M-ary shift register with L stages. The mapper then
produces a U-ary complex-valued Dirac pulse depending on Such Nyquist pulses are all orthogonal and produce no in-
the current and the L previous input symbols at each time tersymbol interference if sampled at times k/2W. They are
interval. This pulse is filtered by the pulse shaping filter strictly band-limited, but of infinite duration.
Q ( f ) , whose output is the desired baseband signal. A Nyquist pulse which is very popular in applications is
the spectral raised cwine pulse whose spectrum is

If1 5 9
(e
Figure 2: Schematic diagram of the receiver
Q ( f )=

{b
TCOS’ (2fT+ - 1))
*
9 < If1
IIf1
< 9,

(8)
where a E [0,1] is the roll-off factor of the pulse. For a = 0
The block diagram of the receiver is shown in Fig. 2. The we obtain the pulse with the smallest spectrum, the sine(.)
received signal ~ ( t )which
, is assumed to be corrupted by pulse, which, however, presents insurmountable problems
additive white Gaussian noise, is filtered by the matched fil- for implementation. In practical systems values of cy as low
ter Q ” ( f ) . Its output is sampled at the symbol rate 1/T. a 0.3 can be approximated by realizable filters, producing
The input signal is then reconstructed by a maximum like- about 30% excess bandwidth as compared to the “ideal”
lihood sequence estimator (MLSE), whose output Z(n - n,) sine(.) pulse. With the 99% bandwidth definition of (2) the
is the estimate of the input symbol at time n - n,, where excess bandwidth is only about 13.5%.
n, denotes the decoding delay. The MLSE is a finite state Systems using Nyquist signalling have s ectra which are
machine with M L states. P
strictly band-limited (100% bandwidth = p ( l + c y ) for spec-
With this model, uncoded as well as most types of coded trally raised cosine pulses). The average signal power P for
modulation schemes can be represented, as will be shown in Nyquist pulses can be calculated as
the next section.
1
E [u(n)u(m)*] q(t - (TI - m)T)dt = REa,
4. Bandwidth- and Power-Efficient P =T m
Modulation Schemes (9)
In this section, we describe uncoded and four types of where Eb is the energy per bit.
coded modulation schemes. The first coded modulation With the above analysis, it is possible to compute the
scheme that is considered is a combination of QPSK with a power spectral density and, hence, the bandwidth efficiency
convolutional encoder. Next, we investigate ‘classical’ trel- q of several uncoded modulation schemes. This was per-
lis coded modulation (TCM), introduced by Ungerboeck [6]. formed using the 99% bandwidth in (1). The power effi-
It is optimized for power-efficiency; i.e., for the same er- ciency was measured by a computer simulation on an addi-
ror performance, the signal power can be reduced with re- tive white Gaussian noise (AWGN) channel.
spect to uncoded modulation without affecting bandwidth- For the comparisons, the 99% bandwidth and a bit error
efficiency [9]. Another approach was used for correlative rate of lov4 in an AWGN channel are chosen. For the pulse

1635
 4.2 Convolutionally Coded Modulation
5.0 . For a long time, modulation and error-correction coding
were two separate entities of a digital transmission system.

8 4*0
The linear modulator and demodulator produced a discrete
channel with a certain error rate as discussed in Section 4.1.
A code of rate k/n < 1adds n- k parity or check symbols to
every k information symbols z(n), lowering the effective rate
m of information and the bandwidth efficiency to qc = q k / n .
1 3.0 I

3 On the other hand, power efficiency is improved. Convc-

2 lutionally coded modulation is shown here for comparison
2.0. only, as it cannot be represented by the transmitter model
of Fig. 1.

0 2 4 6 8 10 12 14 16 18

Figure 3: Practical bandwidth efficiencies achieved by QAM

with spectrally raised cosine pulses and a = 0.3. The plots
are drawn for Pa = and the 99% bandwidth definition
is used.

BPSK
OC-QPSK
shaping filter, a spectral raised cosine filter with a rolloff fac- 0.0
tor of 0.3. ia used. For all modulation schemes, we compute 0 2 4 6 8 10 12 14 16 18
the bandwidth-efficiency in bits/s/Hz, and we determine
the signal-to-noise ratio required for the desired bit error
rate by computer simulation. As an absolute reference, we
choose the Shannon bound [4], as defined in Eq. (5). This is
the solid line in Fig. 3. The closer a modulation scheme ap- Figure 4: Bandwidth efficiencies achieved by convolutionally
proaches this bound, the better it is with respect to power- coded QPSK and TCM with spectrally raised cosine pulses
and bandwidth-efficiency. and Q = 0.3. Pb = 99% bandwidth.
Fig. 3 also shows the performance of some uncoded mod-
ulation schemes. For the parameters mentioned above, Two examples of convolutionally coded modulation
uncoded BPSK leads to a bandwidth-efficiency of 0.88 schemes are shown in Fig. 4 for the same parameters as
bits/s/Hz at an Eb/No of 8.4 dB. The bandwidth-efficiency above, i.e., 99% bandwidth and Pb = For the rate
of QPSK is twice the one of BPSK without change of the k/n = 1/2 convolutional code with constraint length 7, de-
E*/No, because QPSK uses one complex dimension in the noted (2,1,7), the reduction in bandwidth efficiency by a
signal space. For 8-ary QAM, three different mappings are factor of 2 is balanced by a 5dB gain in the required signal-
used. For 8-PSK, all 8 signal points lie on a circle. The nota- to-noise ratio. Similarly for the k/n = 1/4 code, denoted
tion 8-AMPM normal denotes a signal set where eight points (4,1,14), which gains almost 7dB at the cost of reducing the
form two squares of length one and 2cos(7r/l2) = 1.93, bandwidth efficiency by a factor of 4.
respectively, where the bigger square is rotated by 45 de- Here again, there is direct trading of bandwidth for power.
grees [lo]. Finally, &AMPM hez denotes a hexagonal signal Note that the two coded modulation schemes lie closer to
set with eight points [lo]. The bandwidth-efficiency of all the Shannon bound than the uncoded ones.
three signal sets is equal, and 8-AMPM hez leads to the best
power efficiency because of its compactness. 4.3 Trellis Coded Modulation
For 16-QAM and 64-QAM, the standard square constel- An alternative to conventional coding is to absorb the
lations are used. In the bandwidth-efficiency versus signal- parity symbols in 'an expanded signal set (e.g., going from
to-noise ratio plot of Fig. 3, QPSK, 8-AMPM, 16-QAM and QPSK to 8-PSK). This is the main idea behind trellis-coded
64-QAM roughly lie on a line parallel to the Shannon bound, modulation (TCM) [6,7,8]. TCM has evolved as a com-
hence they all have a similar performance. However, it can bination of coding and modulation techniques: by intrc-
be seen how bandwidth-efficiency can be traded for power- ducing redundancy in the signal alphabet and keeping the
efficiency and vice versa by modifying the signal alphabet transmission symbol rate fixed, no bandwidth expansion is
size. suffered [9]. Substantial gains in power efficiency can be

1636
achieved with this concept, lowering the required signal-t- The pulses used are time raised cosine pulses, analogous
noise ratio by up to 6dB. to (8) in the time domain. They are transmitted at a rate
Although trellis codes are usually modeled by a convolu- 2/T, i.e., only neighbors overlap. This gives the same trans-
tional encoder followed by a mapper [6], they can also be mission rate as coded QPSK using standard Nyquist pulses.
represented by the system model of Fig. 1. The original n o m Fig. 4 we see that the performance of the two schemes
mapper of an M-ary trellis code consists of 2M+1 symbols is also similar, with OC-QPSKhaving the benefit of a simple
whereas in the representation of Fig. 1, there are M L + l and manageable quadrature pulse. A more in-depth discus-
symbols. With the additional symbols, the effect of the sion of OC-QPSK can be found in [2].
convolutional code can be modeled.
Take the &state 8-PSK Ungerboeck code as an example. 4.5 Correlative Trellis Coded Modulation
In its original representation, the encoder consists of 3 bi- The classical trellis codes described above are designed
nary (A4= 2) delay cella and a mapper containing the eight to increase the Euclidean distance of the transmitted signal
(2') 8-PSK complex signal points. In our representation, and, hence, they increase the power efficiency without af-
the encoder consists of L = 2 quaternary (M = 4) delay fecting the bandwidth. In this section we present a class of
cells and a mapper containing 4(2+') = 64 elements, each of trellis codes that are optimized with respect to bandwidth
which also being a complex 8-PSK signal point. efficiency (see [l]). Unlike TCM, the trellis introduces con-
Since the encoder introduces memory in the form of a trolled correlation between close data symbols, hence the
finite state machine, a complex sequence estimator has to name correlative trellis coded modulation, C-TCM. This
be used to decode the sequence of received symbols. This correlation is controlled such that the spectrum of the re-
results in the error performance being dominated by the sulting signal is narrowed.
Euclidean distances between all passible coded sequences A CTCM transmitter can be represented by the system
rather than by the distance between individual signals as model of Fig. 1. For an M-ary input alphabet size and
in QAM. This is astonishing considering that the distance L shift-register stages, the mapper consists of ML+' com-
between the constituent 8-PSK signals is smaller than the plex symbols. The mapper was optimized to achieve the
distance between the signals in a QPSK signal set; how- maximum possible inband power of the transmitted signal;
ever, the minimum distance between all coded sequences results are documented in [l], [12] and [13]. Take again the
is considerably greater, explaining the increased power effi- example used in the previous section with L = 2 quater-
ciency of TCM. Combined with Nyquist shaping, this is an nary (M = 4) delay cells. The 64 elements of the optimized
attractive technique for bandwidth limited channels, which mapper for C-TCM assume 36 different complex values, as
has quickly found its way into international standards for opposed to 8 values for the Ungerbock code described above.
voice-band data modems and satellite systems (see refer-
ences in [7,8]).

t
The bandwidth versus signal-t-noise ratio plot of Fig. 4 5.0

/
shows the performance of several TCM schemes. The small
number below each point denotes the number of states of the
decoder in the receiver, which is an indication of its com- C - T 16C M Y
4.0
plexity. Note that power-efficiency can be improved by in- n
N
creasing complexity, without affecting bandwidth-efficiency.
Note further how close these TCM schemes get to the Shan- G
non bound.
-$ 3.0
Y

4.4 Coded Overlapped Quadrature Modulation

Another method to increase power efficiency is called F 2.0
JQPSK
coded overlapped quadrature modulation [2]. This method
is similar to TCM, but the redundancy necessary for coding
is introduced by pulse overlapping instead of signal alphabet 1 .o
BPSK
expansion.
The encoder and decoder can also be described by Figs. 1
and 2 in the same way as explained in Sec. 4.3. Since
the overlapping quadrature pulses q ( t ) cause intersymbol 0 2 4 6 8 10 12 14 16 18
interference, the MLSE has to be modified to account for Eb/% (dB)
this, which will result in an increase of complexity.
It can further be shown that overlapping Nyquist pulses
entails no advantage, and the pulses q ( t ) are chosen accord- Figure 5: Bandwidth efficiencies achieved by C-TCM with
ing to other criteria, such as maximum permissible ampli- spectrally raised cosine pulses and a = 0.3. Pb = 99%
tude fluctuation for use on non-linear channels or for ease of bandwidth.
synchronization. The mast simple pulses to use are duration
limited. Also, the original signal constellation can be kept In general, it could be shown that with C-TCM, a con-
simpler for a given transmission rate. siderable increase of bandwidth-efficiency can be achieved
Fig. 4 includes the performance of coded overlapped with respect to uncoded QAM, at the cost of a decreased
quadrature modulation using a rate 1/2 convolutional en- performance, however. Fig. 5 shows bandwidth- and power-
coder with a mapping into a QPSK signal set (OC-QPSIC). efficiencies of C-TCM schemes with a quaternary input al-

1637
phabet and one to five delay cells, corresponding to four to tegrate coding and modulation, i.e., coding takes place in
64 states of the trellis. Here again, the 99% bandwidth and signal space. They all are essentially trellis codes in the
a bit error rate Pb = were chosen. The small numbers sense that the optimal decoder is a maximum-likelihood se-
beside each point denote the number of states of the Viterbi quence estimator, which makes the schemes comparable on
decoder, as already noted in Section 4.3. the basis of complexity, where we may use the number of
As with M-ary QAM, the increased bandwidth-efficiency trellis states as the principal (linear) measure of complexity.
is achieved at the cost of an increased Ea/No. The points Note that TCM schemes, which are optimized with re-
roughly lie on a line parallel to the Shannon bound. Com- spect to power-efficiency, increase power-efficiency without
pared to M-ary QAM, C-TCM allows t o increase the affecting bandwidth-efficiency. On the other hand, C-TCM,
bandwidth-efficiency by up to 0.5 bit/s/Hz, or to decrease which is bandwidth-optimum, behaves like uncoded modula-
the signal power by about 1 dB, where the increased receiver tion schemes, as an increase in bandwidth-efficiency results
complexity is the price that has to be paid. in decreased performance. An interesting question for fu-
ture work is how to combine both TCM and C-TCM to get
5. Summary and Outlook a modulation scheme which is both power- and bandwidth-
We have discussed several coding and modulation schemes efficient, i.e., which combines the advantages of both ap-
which all aim at improving bandwidth and power efficiency. proaches.
Fig. 6 compares them in a bandwidth efficiency versus
&/No plot. The Shannon bound shows that increasing the Acknowledgment
bit rate per unit bandwidth increases the required energy The authors would like to thank Dr. Dacfey Dzung, A s
per bit. This behavior is reflected by the different modula- com Radiocom Ltd., Switzerland, for the fruitful discussions
tion schemes. lladitional rate-expansion coding and multi- and his work on the original report [14].
level signalling clearly trade bandwidth for power. Related
schemes, i.e., schemes with roughly comparable decoding References
complexities all lie on diagonal lines. In this view the dif-
ferent design philosophies become apparent. In TCM, 9 is [l] S. Ramseier, “Bandwidth Efficient Trellis Coded Modulation
unchanged and a complexity increase achieves a horizontal Schemes,” Proc. of the Intern. Conf. on Communications 1990,
shift of the performance points, an increase in power effi- ICC’90, pp. 340.3.1-5, April 1990
ciency. The performance points for correlated trellis coding [2] C. B. Schlegel, “Coded Overlapped Quadrature Modulation,”
show an improvement in bandwidth efficiency and in power Proc. of the Global C o d . on Communications 1991, GLOBE
COM’91, pp. 1192-1199, December 1991
efficiency with respect t o standard quadrature modulation
with increasing complexity. [3] J. B. Anderson, T. Aulin and C. E. Sundberg, Digital Phase
Modulation, Plenum Press, New York, 1986.
[4] C. E. Shannon, “A Mathematical Theory of Communication”,
Bell Syst. Tech. J., Vol. 27, pp. 379-423, July 1948.
5.0 - / [5] G. D. Forney, “Coding and its Application in Space Communi-
cations”, IEEE Spectrum, Vol. 7 , pp. 47-58, 1970.
[SI G. Ungerboedr, “Channel Coding with Multilevel/Phase Sig-
h
4.0 - nals”,IEEE Trans. Inform. Theory, Vol. IT-28, No.1, pp. 55-67,
January 1982.
N
Grn [A G. Ungerboedc, “TrellikCoded Modulation with Redundant Sig-
- nal Sets Part I: Introduction”, IEEE Communications Magazine,
‘3j

g
c)

F
3.0

2.0 -
/-/ TCM
Vol. 25, No.2, pp. 5-11, February 1987.
[8] G. Ungerboeck, “Trellis-Coded Modulation with Redundant Sig-
nal Sets Part 11: State of the Art”, IEEE Communications Mag-
azine, Vol. 25, No.2, pp. 12-21, February 1987.
~QPSK
[9] E. Biglieri, “Ungerboeck Codes Do Not Shape the Signal Power
Spectrum”, IEEE Trans. Inform. Theory, Vol. IT-32, No.4, pp.
BPSK 595-596, July 1986.
[lo] G.D. Forney et. al., “Efficient Modulation for Band-Limited
Channels,” IEEE Journal on Selected Areas in Communications,
vol. SAC-2, pp. 632-647, September 1984.
[ll] J. G. Proakis, Digital Communications, McGraw-Hill, London,
1983.
[12] S. Ramseier, “Bandwidth-Efficient Correlative Trellis Coded
Modulation Schemes,” Ph. D.Thesis No. 9502, ETH Zurich, 1991
Figure 6: Power/Bandwidth-’hdeoff of several modulation
[13] S. Rameeier, “Bandwidth-Efficient Correlative Trellis Coded
schemes with spectrally raised cosine pulses and (Y = 0.3. Modulation Schemes,” IEEE Transactions on Communications,
Pb = loq4, 99% bandwidth. accepted for publication
[14] C. Schlegel, D. Dzung, S. b e i e r , “Bandwidth and Power Ef-
All schemes discussed in this paper share the objective of ficient Coding Schemes: An Overview,” Asea Brown Boveri Re-
using bandwidth and power most efficiently. Even though search Report, CRB-W006C, October 1990
the different angles of approaches are different, they all in-

1638