182

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 46, NO. 2, FEBRUARY 1998

Performance Analysis of Viterbi Decoding

for 64-DAPSK and 64-QAM Modulated
OFDM Signals
Thomas May, Hermann Rohling, and Volker Engels
Abstract The multilevel modulation techniques of 64quadrature amplitude modulation (QAM) and 64- differential
amplitude and phase-shift keying (DAPSK) have been
proposed in combination with the orthogonal frequency-division
multiplexing (OFDM) scheme for digital terrestrial video
broadcasting (DTVB) [1][3]. With this system a data rate of
34 Mb/s can be transmitted over an 8-MHz radio channel.
A comparison of these modulation methods in the uncoded
case has been presented in [1]. The channel coding scheme
proposed for DTVB in [5] consists of an inner convolutional code
concatenated with an outer ReedSolomon (RS) code. In this
paper the performance of the convolutional codes is analyzed
for the two modulation techniques. This analysis includes soft
decision Viterbi decoding of the convolutional code. For soft
decision decoding of DAPSK modulated signals a new metric
is developed.
Index TermsDAPSK, OFDM, soft decision decoding.

I. INTRODUCTION

OR DIGITAL terrestrial TV broadcasting, an orthogonal frequency-division multiplexing (OFDM) transmission system has been specified [4]. This system has to cope
with multipath propagation channels in a realistic radio environment. Intersymbol interferences (ISIs) can be avoided
due to the implemented guard interval, but there is still an
effect of frequency-selective fading, which means that some
of the subcarriers are strongly attenuated by the radio channel.
Therefore, an efficient channel coding scheme has to be used.
Since hard decision decoding effectively reduces the distance
of a code by a factor of two [15], soft decision decoding has
to be applied in order to achieve good performance.
64-quadrature amplitude modulation (QAM) and 64differential amplitude and phase-shift keying (DAPSK) are two
candidates that have been proposed as modulation schemes for
the digital terrestrial video broadcasting (DTVB) application.
64-QAM is a coherent modulation technique which makes
necessary the transmission of pilot signals for the channel
estimation. An equalizer must be implemented which leads

Paper approved by S. B. Wicker, the Editor for Coding Theory and Techniques of the IEEE Communications Society. Manuscript received February
21, 1996; revised January 28, 1997 and August 8, 1997.
T. May and H. Rohling are with the Institute of Telecommunications,
Technical University of Braunschweig, D-38092, Braunschweig, Germany (email: [email protected]; [email protected]).
V. Engels was with the Institute of Telecommunications, Technical University of Braunschweig, D-38092, Braunschweig, Germany. He is now with
Bosch Telecom, D-71522, Backnang, Germany (e-mail: volker.engels@bk.
bosch.de).
Publisher Item Identifier S 0090-6778(98)01930-8.

to an increased computation complexity in the receiver. As

an alternative, a multilevel purely differential modulation
technique, 64-DAPSK, has been developed, which does not
require any pilot symbols, channel estimation, or equalization.
An entire description of 64-DAPSK and a comparison with
64-QAM in the uncoded case can be found in [1]. For
a realistic comparison of these modulation techniques the
channel coding scheme should be taken into account because
the performance of soft decision decoding depends on the
modulation technique. Therefore, the objective of this paper
is to analyze the performance in the coded case.
For this analysis a soft decision metric for 64-DAPSK modulation is introduced. Soft decision Viterbi decoding of convolutional codes is analyzed both for 64-QAM and 64-DAPSK
using the OFDM transmission scheme. Although this investigation is made for an OFDM transmission system, the way of
processing soft decision decoding is in no point specialized for
OFDM. That means that the proposed metric for 64-DAPSK
can also be used for single-carrier transmission systems.
Recently, new metrics for differential phase-shift keying
(DPSK) modulation have been proposed for the additive white
Gaussian noise (AWGN) channel, which are discussed in [7].
In these schemes interleaving is not needed and, in this case,
much of the loss due to the noncoherent detection can be
regained by the sequence estimation. This is not possible in
the considered transmission system because interleaving is
necessary in a frequency-selective fading channel. As far as
we know, no proposal of a metric for DAPSK modulation has
been made before as the DAPSK modulation technique itself
is rather new. Uncoded 16-DAPSK has been analyzed in [6].
II. THE ANALYZED CODING AND TRANSMISSION SYSTEM
A. Channel Coding Scheme
The channel coding scheme proposed in [5] is depicted
in Fig. 1. A convolutional code is used as an inner code in
order to correct the error events which consist of few bit
errors. If a burst of bit errors occurs at the channel output, the
convolutional code can not be decoded correctly. Therefore,
bit interleaving is used in order to spread the bit errors over
the entire OFDM symbol. However, there will still be error
events which the decoder of the convolutional code is unable
to correct. In this case, a burst of errors always will occur at
the decoder output. The task of the outer ReedSolomon (RS)
code is to correct these burst errors. Because there is more than
a single RS codeword in each OFDM symbol, an additional

00906778/98$10.00 1998 IEEE

MAY et al.: ANALYSIS OF VITERBI DECODING FOR OFDM SIGNALS

183

Fig. 1. Analyzed coding scheme.

Fig. 2. Transmission system.

interleaving of the RS code symbols is implemented (see

Fig. 1). In the simulation, an OFDM system with
subcarriers has been considered. Each OFDM symbol transfers
6144 b in total. The convolutional code has the memory length
in accordance with [5]. The bit interleaver is a block
interleaver with 83 rows and 74 columns, which means that
only two bits of an OFDM symbol are not involved in the interleaving process. The performance of convolutional codes with
code rates
,
, and
from [17] has been analyzed.
A (204, 188) RS code with
b/symbol has been chosen
for the DTVB application. The objective of the concatenated
code is to fulfil the requirement of a residual bit-error rate
(BER) of 10
at the output of the RS decoder. For this,
the BER at the output of the Viterbi decoder is required to be
lower than 2 10
[5].
This concatenated coding scheme is very efficient for both
AWGN and frequency-selective radio channels. For fading
channels, the possibility of soft decision decoding of the
convolutional code leads to a high coding gain. Soft decision
decoding can easily be implemented if a coherent modulation
technique like 64-QAM is considered. For the incoherent 64DAPSK, a new soft decision metric has been developed,
which is described in this paper. Apart from the different
metric calculation algorithms, the data are processed in the
64-DAPSK receiver the same way as in the case of 64-QAM.

(a)

(b)
Fig. 3. Constellation diagrams of 64-DAPSK and 64-QAM.

B. Transmission System
Fig. 2 shows the considered OFDM transmission system.
The coded bit sequence
is mapped into 64-QAM or
64-DAPSK modulation states by the modulator for each subcarrier, i.e., for the th subcarrier in the current OFDM symbol,
the bits
are mapped onto the modulation state .
As the next step in the transmitter, an inverse fast Fourier
transform (IFFT) is processed in order to calculate the timediscrete OFDM signal. Furthermore, a guard interval is added
to the timediscrete signal as a periodic extension of the
IFFT output sequence. The transmitted signal is distorted
by the frequency-selective radio channel and, furthermore,
AWGN is added.

For a brief explanation, the constellation diagrams of the

incoherent 64-DAPSK and the coherent 64-QAM are shown
in Fig. 3. 64-QAM has a chessboard-like constellation diagram
whereas the constellation diagram of 64-DAPSK consists of
four different amplitude rings, each with 16 phase states. The
amplitude values of the rings are , , , and . This means
that the amplitudes differ by a constant factor which has
.
been optimized in [1] to
The six input bits which determine the modulation state of a
subcarrier are divided into two different groups for amplitude
and phase modulation. The phase modulation is independent
of the amplitude and identical with the well known 16-DPSK.

184

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 46, NO. 2, FEBRUARY 1998

TABLE I
DIFFERENTIAL AMPLITUDE MODULATION FOR 64-DAPSK:
CHOICE OF THE CURRENT AMPLITUDE STATE jxi j DEPENDING
~i j AND THE AMPLITUDE BITS
ON THE PREVIOUS STATE jx

The input bits

are used for this differential
phase modulation in the 64-DAPSK scheme.
The amplitude states
are chosen from the constellation
diagram depending on the previous amplitude state
and the
two information bits
and
according to Table I. This
means, e.g., for a subcarrier if the amplitude in the previous
OFDM symbol was
and the input information bits
and
are both zero, then the amplitude in the current OFDM
again. Table I shows the amplitude state
symbol is
for the subcarrier in the current OFDM symbol.
In each OFDM receiver after block synchronization, analogto-digital (A/D) conversion, and removing of the guard interval, a fast Fourier transform (FFT) produces the complex
output states
for each subcarrier .
If the coherent 64-QAM is used, after the FFT a channel
equalization must be performed, which is assumed to be ideal
throughout this paper. This means that the channel transfer
factor
is assumed to be known exactly for each subcarrier
. With this information the transmitted state
of each
subcarrier in the 64-QAM constellation diagram is evaluated
by a simple quotient
(1)
in order to get the coded bit sequence for hard decision
decoding or the metric increments for soft decision decoding.
For the incoherent 64-DAPSK demodulation, first the quotient
(2)
and the preceding state
of the currently received state
of the same subcarrier in the receiver (FFT output) are
calculated as described in [1]. The resulting complex quotient
is evaluated in order to get the phase and amplitude bits
. This quotient
is nearly independent of
the channel transfer factor
if the radio channel does not
change the transmission behavior too quickly. Therefore, pilot
symbols, channel estimation, and equalization are not needed
in the 64-DAPSK receiver, which reduces the computation
complexity.
The four bits
are determined depending
on the phase difference between
and
only. This phase
demodulation process is the same as the demodulation of 16DPSK. For the amplitude demodulation, Table II shows how
the amplitude bits
and
are obtained. For the evaluation
of
, simple amplitude thresholds are used.
C. Channel Model
Computer simulations have been performed with AWGN
and frequency-selective radio channels according to the statis-

Fig. 4. Pdf of the delay for the multipath channels (power delay profile
hilly terrain).

EVALUATION

TABLE II
AMPLITUDE INFORMATION BITS
THE 64-DAPSK DEMODULATION

OF THE

bi;5 AND bi;4 IN

tical wide sense stationary uncorrelated scattering (WSSUS)

multipath channel model described in [18]. For the frequencyselective channels, the profile hilly terrain from [19] has
been used, which is shown in Fig. 4.
According to this probability density function (pdf) of the
delay, the channel impulse responses have been generated
for the simulation of a Rayleigh and a Rice channel and,
in the case of the Rice channel, a line-of-sight (LOS) path
with a Rician factor of 10 dB has been added to the impulse
responses. In the simulations, the received signal has been
calculated in the time domain using these impulse responses.
The denotations Rayleigh and Rice refer to the fact that
the received subcarrier amplitudes are Rayleigh-distributed if
there is no LOS path, and Rician distributed if there is a
LOS path.
In the simulations the channel is frequency-selective but it is
not time-varying, i.e., the Doppler frequency is
. High
Doppler frequencies affect both the performance of 64-DAPSK
and the channel estimation for 64-QAM.
III. VITERBI DECODING
Viterbi decoding of convolutional codes is a maximumlikelihood sequence estimation (MLSE) of the received signal
[14], [16]. The simplest way of proceeding is hard decision
decoding, where the decision of the decoder is based only
on the bits which are produced by the demodulator. A better
performance can be achieved with soft decision decoding. For
soft decision decoding the decision is based on the exact states
of the subcarriers as they are received. Obviously, the way of
processing soft decision decoding depends on the modulation
technique. An optimal soft decision metric for the 64-QAM is
available, whereas a suitable metric for 64-DAPSK had to be
developed for this investigation.
Noncoherent detection always causes a loss compared with
coherent detection. In [8] and [10] it has been shown for
DPSK modulation that much of this loss can be regained by
a sequence estimator. This idea has been applied for trellis-

MAY et al.: ANALYSIS OF VITERBI DECODING FOR OFDM SIGNALS

185

pdf of

is a complex Gaussian distribution.

(4)

A maximum a posteriori sequence estimation (MAPSE)

symbol estimator for
performs the following maximization:
(5)
and of
Because of the independence of the noise terms
the modulation symbols , the symbol sequence estimation
can be written as

Fig. 5. Performance of ideal 64-QAM and 64-DAPSK (ring ratio a 1:4)

with a convolutional code m = 6; R = 3=4 hard decision decoding AWGN
channel.

coded DPSK signals in [11], resulting in a metric for the

Viterbi algorithm. In [7], this and other schemes are discussed.
An optimal metric for noncoherent detection of trellis-coded
DPSK is derived and it is shown that with this metric nearly
the performance of coherently detected trellis-coded DPSK
can be achieved. However, the optimal metric is not additive
so that it can not be used for the Viterbi algorithm without
simplifications. Furthermore, all of these schemes have been
designed for the AWGN channel and do not allow interleaving.
For OFDM transmission over a frequency-selective channel
we need interleaving in order to make the error probabilities
of successive bits to be decoded independently and avoid error
bursts. Thus, we have to accept the loss due to noncoherent
detection.
In this section the performance of 64-DAPSK and 64-QAM
in combination with convolutional codes is compared using
both hard and soft decision decoding. The comparison will be
made at a BER of 2 10 as explained in Section II-A.

(6)
(7)
(8)
Equation (6) implies that the error events on different subcarriers are independent. This is obviously true for coherent
modulation and demodulation. It is also valid in the case
of differential modulation, since the differential encoding
is organized between two symbols
and
of the same
subcarrier, not between adjacent symbols
and
. Considering a single carrier transmission system with differential
modulation, error events in different modulation symbols are
not statistically independent any more, but, if bit interleaving
is applied, this fact cannot be exploited like in [7].
It is assumed that the transmitted symbols
occur with
the same probability. Then the second sum in (8) is irrelevant
for maximization and can be left out. This results in an MLSE
optimization criterion where, for coherent modulation schemes
with (4), the following minimization is performed:
(9)

A. Hard Decision Decoding

Hard decision decoding only makes use of the bits which
are provided by the demodulator. The MLSE chooses the code
sequence with the least Hamming distance from the received
bit sequence. The performance of a hard decision decoded
convolutional code
is shown in Fig. 5
for 64-QAM and 64-DAPSK. The signal-to-noise ratio (SNR)
difference at
is 4.8 dB.
B. Soft Decision Decoding of 64-QAM Modulated Signals
For soft decision decoding the received modulation states
are taken into account. The (complex) modulation state of the
th subcarrier of the received OFDM signal is
(3)
where is the transmitted modulation state, is the complex
channel transfer factor of the considered subcarrier, and
is
. As the result of a channel
AWGN with the noise power
estimation, the receiver is assumed to know the transfer factors
. The modulation states
are mutually independent. The

This sum can directly be used as a metric of the Viterbi

algorithm in the case of binary phase-shift keying (BPSK)
modulation, the metric increment for
being
. This is possible because only a single bit per modulation state is transferred.
If more than a single bit per state is transmitted, the minimization in (9) is optimal, too, but then each contains more
than a single bit of an interleaved coded bit sequence. Thus, all
possible coded and interleaved bit sequences would have to be
tried and the corresponding symbol sequences
would have
to be calculated in order to evaluate (9). This can not be done
by an add/compare/select algorithm and so the computation
effort is too high for a reasonable sequence length.
If we wish to decode the convolutional code with the Viterbi
algorithm, we need metric increments
for each bit
which can be obtained by the following approximation:
(10)
(11)
is the point in the constellation diagram which is
where
nearest to the equalized received state
and by which the

186

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 46, NO. 2, FEBRUARY 1998

assumed bit
(0 or 1) is transferred. Using the Viterbi
algorithm with this metric results in a suboptimal MLSE not
only because of the approximation above but also because the
different bits contained in the same modulation symbol are
of the same subcarrier
evaluated independently, i.e., the
need not be equal for all in the suboptimal metric. The
decision criterion with this suboptimal metric is
(12)
This simplification allows the use of the Viterbi algorithm,
resulting in a computation effort which can be managed.
Alternatively, the system could be designed with a symbol
interleaver instead of a bit interleaver. In the case of 64QAM, coordinate interleaving could be used, which has been
proposed in [12]. Then (9) could be evaluated without any
approximations. In [13], a high gain has been obtained by
replacing a symbol interleaver with a coordinate interleaver
for trellis-coded QAM because the coordinate interleaver
increases the order of diversity of the system. Obviously,
the order of diversity is further increased by bit interleaving
instead of coordinate interleaving. For this reason, using a
symbol or coordinate interleaver does not seem to be a very
promising approach.

Fig. 6. Received modulation symbol and noise terms.

With this approximation, (17) becomes

(19)
(20)
Assuming that the channel transfer factor changes slowly
, the pdf for the MLSE can be approximated by a
Gaussian distribution conditioned on
and .
(21)
with expectation value

C. Soft Decision Decoding of 64-DAPSK Modulated Signals

1) Approximation of pdfs: For differential modulation
schemes the problem of soft decision decoding is difficult
to solve in an optimal way, but reasonable approximation
can be used (see also [23]). If the 64-DAPSK modulation
scheme [1] is used, the analytical description is quite different
compared with 64-QAM. The main question comes up about
the differentially encoded amplitude bits. The amplitude bits
are determined by calculating and evaluating the quotient of
two modulation states
(13)
and are the modulation states of the considered subcarrier
has to
in the preceding OFDM symbol. The pdf
be determined in order to use (8). The pdf analysis however
leads to complex expressions which furthermore depend on
the SNR. Instead of that, it is easier to use the parameter
and the new random variable

(22)
and variance
(23)
The approximation of the pdf of the random variable
by a Gaussian pdf in (21) is good for high SNR. This is
validated by simulations in Figs. 79. There are only small
differences for a high SNR and the approximation can still be
accepted at
dB
or at
dB
, respectively. For a code rate of
and a BER in the order of 10 , an SNR of 23 dB
is required for the AWGN channel or 30 dB for a Rayleighfading channel. In both cases the approximated curves show
small differences with simulation results.
For the phases and phase differences
(24)

(14)
(15)
(16)
(17)
At high SNR (i.e.,
,
) the
logarithm function in (17) can be approximated by the first
two summands of its Taylor series

(25)
the pdf
assumption of high
made.

has to be approximated. Again, the

is
(26)

(18)

(27)

is the component of the noise in the direction of

where
, a Gaussian-distributed real random variable (see Fig. 6).

(28)

MAY et al.: ANALYSIS OF VITERBI DECODING FOR OFDM SIGNALS

Fig. 7. Approximation of the pdf of the random variable

SNR = 10 dB.

Fig. 8. Approximation of the pdf of the random variable

SNR = 20 dB.

187

w i ; vi

wi ; vi

= 0,

0,

where
is a component of the noise according to Fig. 6.
With these simplifications, the pdf of the phase difference
between
and
can also be approximated by a Gaussian
distribution.

Fig. 9. Approximation of the pdf of the random variable

1, SNR = 20 dB.

ln a3

Fig. 10. Approximation of the pdf of the random variable

ence), SNR = 10 dB.

wi ; v i

'i (phase differ-

is not uniformly distributed but

is. Therefore,
is irrelevant for the decisions and can be removed. At this
point the approximations in (21) and (29) are used

(29)

(32)

where
(30)
This approximation of the pdf in (29) has been verified by
a simulation. The approximated function is compared with a
simulation result in Fig. 10.
2) Suboptimal MAPSE and MLSE for 64-DAPSK: Based on
the approximations above, the sequence estimation can easily
be performed according to (8). Amplitude bits and phase bits
are evaluated in two different steps. Again
is the assumed
bit sequence, and
and
are the best fitting logarithms
of amplitude quotients or the best fitting phase differences for
each bit.

If (32) is multiplied by , the minimization has the same

result. This equation can be used as a metric for the Viterbi
algorithm, written as
(33)
where
for amplitude bits
for phase bits

(35)
for amplitude bits
for phase bits.

(31)

(34)

(36)

Each summand in (33) is the metric increment for the

describes the reliability of the bits
corresponding bit.

188

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 46, NO. 2, FEBRUARY 1998

TABLE III
SNR [dB] REQUIRED FOR BER = 2 1 1004

according to the position of the received symbols in the constellation diagram.

is an information about the reliability
of the subcarrier (reliability information). In a practical
is used instead of
for calculating
application,
in (35). Thus, no a priori knowledge is needed at the receiver.
For the suboptimal MAPSE (sub-MAPSE),
takes into
consideration that the amplitude quotients of DAPSK do not
occur with the same probability. For the suboptimal MLSE
(sub-MLSE),
is left out.
For low noise power,
is very small. At
dB
with the sub-MLSE only 0.1 dB more signal power is needed
for the same BER compared to the sub-MAPSE. If
is left
out, the metric becomes independent of the SNR. This makes
the sub-MLSE simpler than the sub-MAPSE, where
has to
be determined additionally. Therefore, the sub-MLSE will be
used in the following.
The obtained metric for soft decision decoding of 64DAPSK modulated signals leads to a simple calculation algorithm without a high computation complexity.
In [9] a similar metric, termed Gaussian metric, for MDPSK is given where the decision variable
is used instead
of .
D. Performance of Soft Decision Decoding
Computer simulations have been carried out with the puncfrom
tured convolutional codes with memory length
subcarriers has been
[17]. An OFDM system with
analyzed for this paper, but the performance results will not
change if a
or
FFT is used. Interleaving is done by a
block interleaver with 83 rows and 74 columns.
Fig. 11 shows the performance of the convolutional code
with the code rate
for 64-QAM ( is assumed to be
known exactly) and 64-DAPSK for the AWGN channel.
It is a remarkable result that with soft decision decoding
64-DAPSK shows a higher additional gain compared to hard
decision decoding than 64-QAM does, which can be seen
from Figs. 5 and 11. For a convolutional code with code rate
, this effect reduces the SNR loss between ideal 64,
QAM and 64-DAPSK from 4.5 dB (at
hard decision decoding) to 4 dB (at
, soft
decision decoding). The SNR loss is caused both by the
differential coding and by the unequal distances in the 64DAPSK constellation diagram. The unequal distances in the
constellation diagram are considered by the new soft decision
metric and, thus, the SNR distance between QAM and DAPSK

Fig. 11. Performance of ideal 64-QAM and 64-DAPSK (ring ratio a = 1:4)
with a convolutional code m = 6; R = 3=4 soft decision decoding AWGN
channel.

can be reduced although the soft decision metric for DAPSK

is not optimum.
E. Comparison of 64-QAM and 64-DAPSK
The performance of QAM and DAPSK can be compared
by determining the required SNR for a fixed BER. For DTVB
a remaining BER below 2 10
after decoding of the
convolutional code is to be assured [5]. Therefore, the different
modulation schemes and code rates
for the convolutional
. The required SNR
code are compared at
values are given in Table III.
Table IV shows the differences of the SNR values between
64-QAM and 64-DAPSK for a fixed code rate . However,
this comparison does not take into account that different data
rates can be transmitted with the compared systems due to
the pilot signals which have to be transmitted for the channel
estimation if 64-QAM is used.
Pilot signals are not necessary for DAPSK so that a convolutional code with increased redundancy can be employed
instead, if the same user data rate is to be transmitted. Such
a comparison with a fixed data rate seems to be reasonable.
If 11% of the subcarriers are needed for the transmission of
pilot signals in the case of 64-QAM, the code rate
can be used for QAM and the code rate
can be used
for DAPSK modulation. The performances of these systems
are compared in Table V.
For Rice- and Rayleigh-fading channels, the BER curves of
64-QAM and 64-DAPSK are shown in Figs. 12 and 13. Again

MAY et al.: ANALYSIS OF VITERBI DECODING FOR OFDM SIGNALS

DIFFERENCE

OF

REQUIRED SNR [dB]

FOR

189

TABLE IV
64-QAM (IDEAL CHANNEL ESTIMATION) AND 64-DAPSK MODULATION

AT BER = 2

10 4 , FIXED CODE RATES

TABLE V
DIFFERENCE OF REQUIRED SNR [dB] FOR 64-QAM (IDEAL CHANNEL ESTIMATION) WITH R = 3=4
AND 64-DAPSK MODULATION WITH R = 2=3 AT BER = 2 1 1004 , FIXED USER DATA RATE

Fig. 12. Performance of ideal 64-QAM (convolutional code R = 3=4) and

64-DAPSK (R = 2=3) over the Rice-fading channel (fixed user data rate).

the code rates have been chosen as in Table V so that the

same data rate is transmitted.
In [20] Sandell analyzes several methods of channel estimation for OFDM based on pilot symbols. With a channel
estimator consisting of a two-dimensional separable filter,
i.e., a finite-impulse response (FIR) frequency filter with five
taps and an FIR time filter with two taps, he achieves the
best performance at the given complexity. The performance
degradation compared to an equalizer with exact knowledge
of the channel transfer function is 1.4 dB. The complexity of
this estimator is three complex multiplications per transmitted
modulation symbol, the storage of at least five complete
OFDM blocks before the estimation procedure can be started,
and the corresponding delay. The system parameters and the
properties of the channel Sandell considers require the use of
6% of the transmitted symbols as pilot symbols. However, he
shows that the amount of pilot symbols to be used depends
on the length of the OFDM blocks and the guard interval,
on the subcarrier spacing, on the maximum echo delay and
on the maximum Doppler frequency of the channel. In [22],
for example, an OFDM system with 12% pilot symbols is
proposed. In [21], an OFDM system with 20% pilot symbols
which also uses FIR filters both in frequency and time direction

Fig. 13. Performance of ideal 64-QAM (convolutional code R = 3=4) and

64-DAPSK (R = 2=3) over the Rayleigh-fading channel (fixed user data
rate).

shows a performance loss of 1.5 dB compared with an ideal

channel estimation. Finally, it must be mentioned that for the
design of the FIR filters in [20] and [21], knowledge of the
maximal Doppler frequency and of the maximal echo delay
has been used.
With regard to these results from literature, it seems that the
differences of the required SNR given in Tables IV and V for
an ideal channel estimation must be reduced by approximately
1.5 dB in the case of a real channel estimation. In this case, the
performance of 64-QAM and 64-DAPSK do not differ much.
IV. CONCLUSION
In this paper an OFDM transmission system with convolutional channel coding has been analyzed. The modulation
techniques of 64-QAM and 64-DAPSK have been used. In
both cases, soft decision Viterbi decoding has been applied
where, in the case of 64-DAPSK, a new metric has been
introduced. In the uncoded case, 64-DAPSK needs a 4.5dB higher SNR than the ideal 64-QAM (AWGN channel, at
). With a convolutional code (code rate
,
memory 6) and soft decision decoding, the new metric reduces
the SNR distance between 64-QAM and 64-DAPSK by 0.5 dB

190

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 46, NO. 2, FEBRUARY 1998

to 4 dB at
. The new metric can also be used
for single carrier transmission.
For frequency-selective fading channels the coherent 64QAM requires an exact channel estimation and equalization,
which leads to high computation complexity in real OFDM
receivers. In this paper a perfect channel estimation has been
assumed in the case of 64-QAM.
Furthermore, pilot signals have to be transmitted for 64QAM, which is not necessary for 64-DAPSK because 64DAPSK is a purely differential modulation scheme. Instead
of the pilot signals, channel coding with increased redundancy
can be used in the case of 64-DAPSK modulation. This reduces
the SNR distance between the two modulation techniques to
2 dB (Rice fading) or 1.6 dB (Rayleigh fading).
A real channel estimation and equalization for the coherent 64-QAM modulation technique is assumed to cause an
additional SNR loss in the order of 1.5 dB. This means, the
performance of the two considered modulation methods show
only small differences in SNR in the coded case. However,
the computation complexity is much lower in the case of the
incoherent 64-DAPSK receiver.
REFERENCES
[1] V. Engels and H. Rohling, Multilevel differential modulation techniques (64-DAPSK) for multicarrier transmission systems, European
Trans. Telecommun., vol. 6, pp. 633640, Nov./Dec. 1995.
[2] H. Rohling and V. Engels, Differential amplitude phase shift keying
(DAPSK)A new modulation method for DTVB, in Proc. Int. Broadcasting Convention, Amsterdam, The Netherlands, 1995, pp. 102108.
[3] R. Monnier, J. B. Rault, and T. de Couasnon, Digital television
broadcasting with high spectral efficiency, in Proc. Int. Broadcasting
Convention, Amsterdam, The Netherlands, 1992, pp. 380384.
[4] B. Marti, P. Bernard, N. Lodge, and R. Schafer, European activities on
digital television broadcastingFrom company to cooperative projects,
EBU Tech. Rev., pp. 2229, Summer 1993.
[5] R. Schafer, Terrestrial transmission of DTVB signalsThe European
specification, in Proc. Int. Broadcasting Convention, Amsterdam, The
Netherlands, 1995, pp. 7984.
[6] Y. C. Chow, A. R. Nix, and J. P. McGeehan, Analysis of 16-APSK
modulation in AWGN and Rayleigh fading channel, Electron. Lett.,
vol. 28, pp. 16081610, Aug. 1992.
[7] T. R. Giallorenzi and S. G. Wilson, Noncoherent demodulation techniques for trellis-coded M-DPSK signals, IEEE Trans. Commun., vol.
43, pp 23702380, Aug. 1995.
[8] S. G. Wilson, J. Freebersyser, and C. Marshall, Multi-symbol detection
of MPSK, in Proc. GLOBECOM89, Dallas, TX, Nov. 1989, pp.
16921697.
[9] M. K. Simon and D. Divsalar The performance of trellis coded
multilevel DPSK on a fading mobile satellite channel, IEEE Trans.
Veh. Technol., vol. 37, pp. 7891, May 1988.
[10] D. Divsalar and M. K. Simon, Multiple-symbol differential detection
of MPSK, IEEE Trans. Commun., vol. 38, pp. 300308, Mar. 1990.
[11] D. Divsalar, M. K. Simon, and M. Shahshahani, The performance
of trellis-coded MDPSK with multiple symbol detection, IEEE Trans.
Commun., vol. 38, pp. 13911403, Sept. 1990.
[12] K. Boulle and J. C. Belfiore, Modulation schemes designed for the
rayleigh channel, in Proc. CISS92, Princeton, NJ, 1992, pp. 288293.
[13] B. D. Jelicic and S. Roy, Design of trellis-coded QAM for flat fading
& AWGN channel, IEEE Trans. Veh. Technol., vol. 44, pp. 192201,
Feb. 1995.
[14] A. J. Viterbi, Convolutional codes and their performance in communication, IEEE Trans. Commun., vol. COM-19, pp. 751772, 1971.
[15] J. G. Proakis, Digital Communications, 3rd ed. New York: McGrawHill, 1995.
[16] G. C. Clark and J. B. Cain, Error Correction Coding for Digital
Communications. New York: Plenum, 1988.

[17] G. Begin, D. Haccoun, and C. Paquin, Further results on high-rate

punctured convolutional codes for Viterbi and sequential decoding,
IEEE Trans. Commun., vol. 38, pp. 19221928, 1990.
[18] P. Hoher, A statistical discrete-time model for the WSSUS multipath
channel, IEEE Trans. Veh. Technol., vol. 41, pp. 461468, Nov. 1992.
[19] Digital land mobile radio communications, Office for Publications
of the European Communities, Luxembourg, Final Rep. ISBN 92-8259946-9, 1989.
[20] M. Sandell, Design and analysis of estimators for multicarrier modulation and ultrasonic imaging, Ph.D. dissertation, Lulea Univ. Technol.,
Lulea, Sweden, Sept. 1996.
[21] P. Hoher, TCM on frequency-selective land-mobile fading channels,
in Proc. 5th Int. Workshop on Digital Communications, B. V., M. Luise
and E. Biglieri (Eds.), Pisa, Italy, Sept. 812, 1991.
[22] M. Bossert and A. Donder, Channel estimation and equalization in
orthogonal frequency division multiplexing systems, in Proc. VDE
Conf. 135, Neu-Ulm, 1995.
[23] T. Muller and H. Rohling, Channel coding for narrowband Rayleigh
fading with robustness against changes in doppler spread, IEEE Trans.
Commun., to be published.

Thomas May received the Dipl.-Ing. degree in electrical engineering from the Technische Universitat
Braunschweig, Braunschweig, Germany, in 1995.
Currently, he is with the Institute of Communications, Technische Universitat Braunschweig, Braunschweig, Germany, as a Research Engineer, working
in the areas of OFDM and channel coding.

Hermann Rohling received the Dipl.-Math. degree

from the University of Stuttgart, Stuttgart, Germany,
in 1977, and the Ph.D. degree from the Rheinisch
Westfalische Technische Hochschule, Aachen, Germany, in 1984.
From 1977 until 1988 he was with the Research
Institute of AEG, Ulm, Germany, as a Research
Mathematician, working in the area of digital radar
signal processing. Since 1988 he has been with
the Technische Universitat Braunschweig, Braunschweig, Germany, as a Professor of communications. His current research topics include mobile communications, especially
broadband OFDM transmission systems, self-organizing networks, air traffic
control, and digital radar signal processing.
Dr. Rohling is a member of the Deutsche Gesellschaft fur Ortung und
Navigation (DGON) and of the Informationstechnische Gesellschaft (ITG).

Volker Engels received the Dipl.-Ing. and the Ph.D.

degrees in electrical engineering from the Technische Universitat Braunschweig, Braunschweig, Germany, in 1993 and 1997, respectively.
From 1993 until 1997, he was with the Institute
of Communications, Technische Universitat Braunschweig, Braunschweig, Germany, as a Research
Engineer, working on digital transmission systems,
especially single and multicarrier techniques, coherent and incoherent modulation schemes, channel
coding, synchronization, and radio channel models.
Since 1997 he has been with the Development Center, Bosch Telecom,
Backnang, Germany, and he is doing research on future broadband radio
access systems.
Dr. Engels is a member of the Informationstechnische Gesellschaft (ITG).