Vous êtes sur la page 1sur 5

Peak-to-Average Power Ratio Reduction of an OFDM Signal

Using Partial Transmit Sequences With Embedded Side Information

Leonard J. Cimini, Jr., and Nelson R. Sollenberger
AT&T Labs - Research
100 Schulz Drive, Red Bank, New Jersey 07701
Tel/Fax: 732-345-3129/3039
Email: ljc@research.att.com
Abstract- OFDM is an attractive technique for achieving
high-bit-rate wireless data transmission. However, the
potentially large peak-to-average power ratio (PAP) of a
multicarrier signal has limited its application. Previously,
we described a technique, based on partial transmit
sequences, that provides improved PAP statistics with little
loss in efficiency, and which is more easily implemented
than the standard Partial Transmit Sequence approach.
One of the remaining challenges is how to transmit knowledge about the combining sequence to the receiver in a
reliable fashion. Here, we describe a technique, using no
overhead, for embedding this information within the
transmitted data and reliably recovering it at the receiver.

Orthogonal frequency division multiplexing (OFDM) [ I ]
is a very attractive technique for high-bit-rate transmission in a
radio environment. By dividing the total bandwidth into many
narrow subchannels which are transmitted in parallel, the
effects of multipath delay spread can be minimized. This
approach has been adopted or proposed for Digital Audio
Broadcasting [2], Digital Terrestrial Television Broadcasting
131, wireless LANs [ 4 ] ,and high-speed cellular data [5]. One
disadvantage to using OFDM for wireless applications is the
potentially large peak-to-average power ratio (PAP)
characteristic of a multicarrier signal with a large number of
subchannels. In particular, a baseband OFDM signal with N
subchannels has a PAP = N 2 / N = N ; for N =256, PAP-24 dB.
When passed through a nonlinear device, such as a transmit
power amplifier, the signal may suffer significant spectral
spreading and in-band distortion. With the increased interest
in OFDM for wireless applications, overcoming this problem
is a very active and important area of research.
The conventional solutions to the PAP problem are to use
a linear amplifier or to backoff the operating point of a
nonlinear amplifier; both approaches resulting in a significant
power efficiency penalty. Several alternative solutions have
been proposed. The simplest is to deliberately clip the OFDM
signal before amplification [6-71, which gives a good PAP at
the expense of some performance degradation. Another uses
nonlinear block coding [8], where the desired data sequence is
embedded in a larger sequence and only a subset of all the
possible sequences are used, specifically, those with low peak

powers. Using this approach, a 3-dB PAP can be achieved

with only a small bandwidth penalty. As originally described,
to implement this coding scheme, large look-up tables are
required at the transmitter and the receiver, limiting its
usefulness to applications with a small number of subchannels.
Progress has been made towards coding schemes that reduce
the PAP, can be implemented in a systematic form, and have
some error correcting capabilities [9]. Nevertheless, these
methods are difficult to extend to systems with more than a
few subchannels and the coding gains are small for reasonable
levels of redundancy.
Two promising techniques for improving the statistics of
the PAP of an OFDM signal have been proposed. These
techniques have been termed the Selective Mapping (SLM)
approach [lo-1 11 and the Partial Transmit Sequence (PTS)
approach [ 12-131. In SLM, M statistically independent
sequences are generated from the same information and that
sequence with the lowest PAP is chosen for transmission. In
PTS, the data block to be transmitted is partitioned into M
disjoint subblocks which are combined to minimize PAP.
Both introduce additional complexity but provide improved
PAP statistics for the OFDM signal with little cost in
efficiency. In [14], we described new algorithms for
combining partial transmit sequences. Based on simulations,
we showed that these suboptimal strategies, which are less
complex and more easily implemented, suffer little
performance degradation from optimum combining.
A requirement for both of the approaches above is that the
receiver must have knowledge about the generation process of
the transmitted OFDM signal in order to recover the
information. This can be sent as side information resulting in
some loss of efficiency. Alternatively, for the PTS approach,
differential encoding can be employed across subchannels
within a subblock; in this case, the overhead is a single
reference carrier per subblock. Here, we describe a technique
for embedding the side information within the transmitted data
with no overhead and a detection scheme for reliably
recovering this information.
In Section 2, we briefly summarize the PTS approach and
the lower complexity, suboptimal, alternative. Then, in
Section 3, we describe a method for transmitting the side
information without any loss in efficiency. Results are
presented in Section 4, followed by a summary in Section 5.

Q 2000 IEEE



approach is to form a weighted combination of the M clusters,


2.1 Peak-to-AveragePower Ratio

X = C b , , , X m

In OFDM, a block of N symbols, ( X , , n = O , l , ...,N - 1 1, is

formed with each symbol modulating one of a set of N
subcarriers, cf,, ,n =0, I ,... ,N - 1 ). The N subcarriers are
chosen to be orthogonal, that is, f , = n A f , where Af = I/NT
and T i s the original symbol period. The resulting signal after
D/A conversion can be expressed as

where { b n , , m =1,2, ...,M } are weighting factors and are

assumed to be pure rotations. After transforming to the time
domain, (3) becomes

xf = C b , x ,

c Xne2rrf,,r,o I t -< NT








The vector x,, called the partial transmit sequence, is the

IFFT of X m . The phase factors are then chosen to minimize
the PAP of x. A PTS transmitter is shown in Fig. 2.

n =o

The PAP of the transmitted signal in ( I ) can be defined as









Fig. 2. Partial transmit sequence approach

2.3 Low-Complexity Combining Algorithms


In the PTS approach described in the previous section, one

bottleneck is the optimization of the phase factors used for
combining the clusters. In [14], we first introduced a
suboptimal iterative algorithm for combining the P TSs. We
will briefly summarize the algorithm for the simplest case
where only binary (i.e., 1 and -1) weighting factors are
considered. After dividing the input data block into M
clusters, form the M N-point PTSs. As a first step, assume that
b,,=l for all m and compute the PAP of the combined signal
(4). Next, invert the first phase factor (bl =- 1) and
recompute the resulting PAP. If the new PAP is lower, retain
b I as part of the final phase sequence, otherwise, b reverts to
its previous value. The algorithm continues in this fashion
until all M possibilities for flipping the signs of the phase
factors have been explored. In what follows, we will refer to
(b,,,, m = I ,...M ) as the inversion sequence.

PAP. (dB)

Fig. 1. Comparison of PAP of unmodified and PAPreduced OFDM signal

In Fig. 1 , the complementary cumulative distribution
function (CCDF=Pr(PAP>PAP,,))ofthe
signal is shown for the particular case of 256 subcarriers. To
more accurately approximate the true PAP, the results in Fig. I
have been computed by oversampling (1) by a factor of four.
For 256 subcarriers, the absolute PAP should be 24 dB.
However, with a reasonably large number of subcarriers
(usually 16 is sufficient), the signal amplitude is
approximately Rayleigh distributed and the large peaks occur
with very low probability. In particular, the PAP exceeds 11
dB for only 0.1% of the possible transmitted OFDM blocks.

2.2 Partial Transmit Sequences

In the PTS approach, the input data block is partitioned
into disjoint subblocks or clusters which are combined to
{ x , , , ~ =,...,
o ,NI - I J , as a vector, x = [ xX ~I ... ~ N - 1 1
Then, partition X into M disjoint sets, represented by the
vectors { X N l , m =1,2,..., M I . The objective of the PTS

~ .

The results presented in [14] show that the suboptimal

iterative approach to combining PTSs provides significant
improvement with only a small degradation compared to the
optimum. Nevertheless, the iterative approach requires some
feedback for implementation. An alternative approach, which
avoids feedback, is to approximate the optimum by
multiplying the information sequence by a number of random
sequences and choosing the best to transmit. Based on this


observation, a practical solution is to use a known set of

sequences which are easily generated, such as Walsh
sequences. This was considered in [I41 and the results are
summarized in the next section.

similar fashion. Then, with the modulation removed, the data

symbols (in the frequency domain) are differentially detected
by computing, for each cluster, the test statistic

2.4 Summary of Results in [14]

In the results which follow, I00000 random OFDM blocks
were generated to obtain the CCDFs. To avoid missing peaks,
the transmitted signal was oversampled by a factor of four.
We assume 256 subcarriers throughout and QPSK data
symbols with the energy normalized to unity. In Fig. 1 , results
are shown for the case of a single OFDM block and 16 clusters
each composed of 16 subcarriers. The unmodified OFDM
signal has a PAP that exceeds 10.4 dB for less than 1 % of the
blocks. For the suboptimal algorithm, using the 16 Walsh
sequences of length 16 as the inversion sequence, a value of
about 8 dB is obtained. By using the PTS approach with the
optimum binary phase sequence for combining, the 1% PAP
reduces to 6.8 dB. While a degradation of 1 dB is encountered
using the suboptimal approach, the optimization process has
been reduced to 16 sets of 16 additions, a considerable savings
over trying to find the optimum set of phase factors.

m = 1,2,...,M


Given the decision statistic in ( 5 ) , several detection

schemes are possible:
The simplest approach is to quantize Re[Z,,] to +1 and
then make decisions independently for each cluster.
This approach, as we will see in the next section, does
not work well because there is no mechanism for
correcting errors.
Better performance can be achieved by quantizing the
individual Z,s and then decoding the entire sequence
to the nearest Walsh sequence. Specifically, we first
generate the sequence {Re[Z,] ,m = 1,2,...,M } and
quantize each component to +1 or -1. Then, we choose
the Walsh sequence of length M that is closest, in
Hamming distance, to the resulting sequence. This
allows for error correction since the received sequence is
mapped into one of only M possible Walsh sequences.
Clearly, the performance can be further improved by
retaining all of the information in the decision statistics.
Therefore, the best strategy should be to compute the
sequence, {Z,,m=1,2 ,...,M } , and, then choose the
Walsh sequence of length M that is closest, in Euclidean
distance, to the resulting sequence.

3.1 Marking Algorithm

To reliably detect the inversion sequence, the data

modulation must be removed. The approach used here is to
raise the frequency symbols to the fourth power. This is a
standard approach for removing QPSK modulation (e.g., see
[15]). Higher-order PSK modulations can be removed in a


where Y),, represents the jth tone in the mth cluster and *
denotes conjugation. It is easy to show that, in the absence of
noise, if cluster m was not altered by the inversion sequence,
then Z , will be + ( N / M - 1). However, if b , =- 1, then Z,
will be - ( N / M - 1). Therefore, a simple binary detection
scheme can recover the inversion sequence. The summation
over the tones in a cluster averages the noise and provides a
significant performance improvement.

To recover the data, the receiver must know which

inversion sequence was used to modify the transmitted data.
This can be sent as explicit side information at the expense of
some loss in efficiency. Here, we present an alternative
approach that suffers no spectral efficiency penalty. The basic
strategy is to embed a marker onto the transmitted data that
can be used to uniquely identify the inversion sequence at the
receiver. The detection of the inversion sequence must be
reliable enough so that it does not have a significant effect on
the overall system performance.

3.2 Detection Schemes

/= 1


The basic algorithm for generating the marker is as

follows: If the inversion sequence does not rotate the cluster
(i.e., 6 , = I), do nothing. Otherwise, if the inversion sequence
rotates the cluster (i.e., b, =-I), then rotate every other tone in
that cluster by 7c/4. This is equivalent to using two signal
constellations for the data symbols in a cluster: one for the
unrotated clusters and another, rotated by n/4, for the modified
clusters. This algorithm puts an embedded marker on those
clusters that have been rotated. The goal then is to design a
detection scheme for this marker that is reliable enough in the
presence of noise and multipath fading to only minimally
impact the system performance.



Zm =

3.3 Results
To quantify the performance of the marking and detection
algorithms, we will use the same parameters as in Section 2.4.
We assume 256 subcarriers (i.e., N = 2 5 6 ) which are divided
into 16 clusters (i.e., M=16), each composed of 16
subcarriers. QPSK is used to modulate the tones. The
measure of performance will be the word error rate (WER),
where word corresponds to one OFDM block or, equivalently,
the length of one inversion sequence. Initially, we will present
results on the detection performance of the schemes described
in Section 3.2. Specifically, performance will be presented for
the probability that the inversion sequence is received in error.
Then we will present some results estimating the effect of
inversion sequence errors on the system WER.
In Fig. 3 , results are shown for the probability of error in
detecting the inversion sequence as a function of the signal-


to-noise ratio (SNR) in an additive, white, Gaussian noise

(A WGN) environment. As expected, the simple cluster-by-

cluster detection scheme does not work well. By using

minimum distance decoding, whether based on Hamming or
Euclidean distances, significant improvements are obtained.
The benefit comes from the error correction that is possible
with minimum distance decoding. The 16 Walsh sequences of
length 16 have a minimum distance of 8 and, as such, can
correct up to 4 errors. Using the Hamming (Euclidean)
distance, a 1 % WER can be achieved with an SNR of about 3.2
dB (2.3 dB). In the remaining figures, we will concentrate on
the minimum Euclidean distance decoding.

performance in flat fading or with very low levels of delay

spread was discussed above. For very large delay spreads,
some penalty should occur. This is caused by the rapid
variations of the channel characteristic across the OFDM
frequency block. Because a differential-in-frequency
detection scheme is employed, the decorrelation between the
fading on adjacent tones will cause an error floor to appear.
This is analogous to the error caused by time variations in the
channel when a differential-in-time detection scheme is used.


r=5 psec

.---Hamming '%
---Euclidean ,





SNR (dB)

%' '


Fig. 4. Detection performance in a fading environment

Fig. 3. Detection performance in A WGN

For wireless transmission, the performance must be
evaluated in a fading environment. In Fig. 4, results are
shown in a flat, Rayleigh fading channel and in a dispersive
environment with a two-ray, equal-amplitude, power delay
profile. For the results shown here and assuming an OFDM
bandwidth of 800 kHz, the two rays are separated by 5 psec.
Results are shown assuming no diversity and for two-branch
maximal ratio combining. Clearly, diversity is an essential
component of the receiver. At the I % WER, diversity reduces
the required SNR by more than 10 dB in a flat fading
environment and by more than 6 dB in a dispersive
environment. Flat fading is the most difficult environment
because, when the channel fades, all of the tones will be in a
fade and error correction will not be effective. With twobranch receive diversity and 5-psec delay spread, a 1% WER
can be achieved with an SNR of only 4.2 dB.
Since the performance should depend on the amount of
frequency selectivity in the channel, the sensitivity of the
detection schemes to different levels of delay spread must be
evaluated. The results, shown in Fig. 5, are plotted assuming
two-branch receive diversity and SNR =3 dB. As can be seen,
the performance is relatively insensitive. The poorer

Two-Branch Diversity
SNR = 3 dB

Min. Euclidean Dist.






Fig. 5. Sensitivity of detection scheme to delay spread

As the final check on the efficacy of using an embedded
transmission approach, we have computed the effect of errors
in the detection of the inversion sequence on the system WER.
The results shown by the solid line in Fig. 6, taken from [19],
can be interpreted as the performance with perfect detection of
the inversion sequence. These results have been computed

with half-rate Reed-Solomon coding across the subchannels,

as well as convolutional coding with constraint lengths K = 6
and K =9. The effect of imperfect detection of the inversion
sequence was taken into account by assuming that the errors in
the data detection and the errors in the inversion sequence
detection are independent. Clearly, this is an upper bound,
since the occurrence of errors in detecting the data and the
inversion sequence are correlated. These results are
represented by the dashed line in Fig. 6. In all cases, the upper
bound on the penalty is less than 0.5 dB.





- N=256



SNR (dB)

Fig. 6. Effect of errors in the inversion sequence detection


OFDM is a very appealing technique for achieving highbit-rate wireless data transmission. However, the potentially
large PAP of an OFDM signal has limited its application. The
Partial Transmit Sequence approach can provide improved
PAP statistics with little loss in efficiency. The suboptimal
strategy for combining partial transmit sequences, presented in
[ 141 and summarized here, achieves similar performance but
with reduced complexity. In this paper, we described an
algorithm for transmitting knowledge of the combining or
inversion sequence without any overhead and strategies for
reliably detecting this sequence. Simulation results
demonstrated that the marking and detection algorithms cause
only minimal degradation in the system performance.
The performance of this approach can be improved by
increasing the number of tones per cluster and, thereby,
increasing the noise averaging benefit. So, this approach
should be even more effective for a wideband OFDM system
[16]. In addition, this technique can be easily extended to
higher-order PSK modulations. For square constellations,
such 16-QAM, the extension is not as readily apparent. The
differential detection and the modulation removal processes
cannot be easily adapted to square constellations. An

alternative is to employ two-level circular modulations, in

which case, the algorithm described here could be applied.

[I] R. van Nee and R. Prasad, OFDM for Wireless Multimedia
Communications, Artech, 2000.
[2] M. Alard and R. Lassalle, "Principles of Modulation and
Channel Coding for Digital Broadcasting for Mobile
Receivers," EBU Technical Review, 1987, pp. 168- 190.
[3] U. Reimers, "Digital Video Broadcasting," IEEE Commun.
Mug., Vol. 36, No. 6, June 1998, pp. 104-110.
[4] M. Aldinger, "Multicarrier COFDM Scheme in High BitRate Radio Local Area Networks," Proc. of PIMRC'94, pp.
[5] L. J. Cimini, Jr., J. C. Chuang, and N. R. Sollenberger,
"Advanced Cellular Internet Service," IEEE Commun. Mag.,
Vol. 36, No. IO, Oct. 1998, pp. 150-159.
[6] R. O'Neill and L. N. Lopes, "Envelope Variations and
Spectral Splatter in Clipped Multicarrier Signals," Proc. of
PIMRC'95, pp. 7 1-75.
[7] X. Li and L. J. Cimini, Jr., "Effects of Clipping and
Filtering on the Performance of OFDM," IEEE Commun.
Letts., Vol. 2, No. 5, May 1998, pp. 131-133.
[8] A. E. Jones, T. A. Willunson, and S. K. Barton, "Block
Coding Scheme for Reduction of Peak to Mean Envelope
Power Ratio of Multicarrier Transmission Scheme," Elec.
Letts., Vol. 30, No. 25, Dec. 1994, pp. 2098-2099.
[9] R. D. J. van Nee, "OFDM Codes for Peak-to-Average
Power Reduction and Error Correction," Proc. of
Globecom'96, pp. 740-744.
[lo] P. Van Eetvelt, G. Wade, and M. Tomlinson, "Peak to
Average Power Reduction for OFDM Schemes by Selective
Scrambling," Elec. Letts., Vol. 32, No. 21, Oct. 1996, pp.
[ I 11 R. W. Bauml, R. F. H. Fischer, and J. B. Huber,
"Reducing the Peak-to-Average Power Ratio of Multicarrier
Modulation by Selective Mapping," Elec. Letts., Vol. 32, No.
22, Oct. 1996, pp. 2056-2057.
[I21 S. H. Muller and J. B. Huber, "A Novel Peak Power
Reduction Scheme for OFDM," Proc. of PIMRC'97, pp.
[13] S. H. Muller and J. B. Huber, "OFDM with Reduced
Peak-to-Average Power Ratio by Optimum Combination of
Partial Transmit Sequences,' Elec. Letts., Vol. 33, No. 5, Feb.
1997, pp. 368-369.
[I41 L. J. Cimini, Jr., and N. R. Sollenberger, "Peak-toAverage Power Ratio Reduction of an OFDM Signal Using
Partial Transmit Sequences," IEEE Commun. Letts., Vol. 4,
No. 3, March 2000, pp. 86-88.
[ 151 J. J. Spilker, Digital Communications by Satellite,
Prentice-Hall, 1977, pp. 390-393.
[I61 J. C. Chuang, et al, "High-speed Wireless Data Access
Based on Combining EDGE with Wideband OFDM," ZEEE
Commun. Mag., Vol. 37, No. 11, Nov. 1999, pp. 92-98.