Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

PAPR Reduction in OFDM Using Wavelet Packet

Pre-processing
Mohan Baro and Jacek Ilow
Dalhousie University
Department of Electrical and Computer Engineering
Halifax, NS, Canada
Email: {mbaro, j.ilow}@dal.ca

Abstract— This paper introduces a novel peak-to-average reduction schemes, Discrete Parameter Optimization methods
power ratio (PAPR) reduction method in orthogonal frequency such as Selective Mapping (SLM) or variants of Multiple
division multiplexing (OFDM) systems by deploying wavelet Signal Representations including Tone Injection are closest to
packet pre-processing of the quadrature amplitude modulation
(QAM) symbols. Specifically, joint inverse discrete wavelet packet the PAPR reduction methods developed in this paper [1], [2].
and Fourier transformations (IDWPT and IDFT) of QAM They are all based on adding time dependent blocks to
symbols are calculated at the transmitter with the purpose of the original OFDM frame to reduce its peak by changing
minimizing the PAPR of the OFDM frame to be transmitted. some or all subcarriers (tones) in the frequency domain. The
The wavelet packet tree representing a reversible IDWPT chosen discrete parameter optimization PAPR reduction methods are
at the transmitter is communicated to the receiver as side
information, where the output of the DFT block is passed through formulated as a solution to:
the DWPT to recover the QAM symbols. As the DWPT is a   
orthonormal transformation, the proposed method preserves the smin = min P AP R IDFT T s (X) , s = 1, · · · , M (1)
s
average transmitted energy while maintaining the integrity of the
transmitted information. With a small level of redundancy for where T s (·) represents a reversible transformation of the
side information, the proposed scheme achieves 5.5 dB reduction
in PAPR over the traditional OFDM system as measured using
multicarrier frequency-domain vector of QAM symbols X,
the complementary cumulative distribution function (CCDF) of and M is the number of possible transformations of X. The
the OFDM signals at the clipping probability of 10−4 . problem setup in (1) stresses that the vector output of T s (·)
is further processed through the IDFT of the conventional
I. I NTRODUCTION OFDM transmitter. The index smin represents the T s (X) and
Orthogonal frequency division multiplexing (OFDM) has the modified QAM symbols, resulting in the minimum PAPR
emerged as an efficient multicarrier modulation scheme for and thus chosen for transmission. To recover the QAM data,
wireless, frequency selective, communication channels. Ease the transmitter sends as side information smin , introducing
of implementation, high spectral efficiency, resilience to im- log2 (M ) bits of overhead per one OFDM frame. Without af-
pulse noise and multipath are a few advantages of OFDM fecting the bit error rate performance, a number of researchers
systems. However, a major drawback in the OFDM signal is have independently proposed transformations, referred to as
its large envelope fluctuation, which limits the efficiency of SLM, based on phase shifts of the data symbols. In this kind
the non-linear power amplifiers specific to wireless commu- of method, it is of great interest not only to find a computa-
nications by forcing them to operate at lower average power. tionally efficient transformation T s (·), but also to have a small
This phenomenon is quantified by the peak-to-average power number of these transformations whose application achieves a
ratio (PAPR) and results from the superposition of a large significant PAPR reduction for any given OFDM frame.
number of usually statistically independent subchannels that The basic idea in the PAPR reduction method proposed
can constructively sum up to high peaks. Also, from central in this paper is to apply in (1) an IDWPT as a transfor-
limit theorem, this causes the OFDM signal to have complex mation T s (·) . With some modifications this technique does
Gaussian process behavior and the envelope to be Rayleigh not destroy the multicarrier character of the OFDM signal.
distributed. In [3], a specialized PAPR reduction scheme for Wavelet
Various schemes have been developed to reduce high PAPR Packet Modulation (WPM) was introduced for a problem setup
in OFDM signals [1], [2]. For the purposes of this paper, we similar to that in (1) except for a lack of IDFT block. In
classify them here as distortion and distortion-less methods. this paper, a PAPR reduction using the wavelet packet tree
The first class of methods, in general, introduces an acceptable pruning (WPTP) technique is developed for PAPR problems in
level of in-channel distortions resulting in an increased, irre- OFDM transmissions. The motivation for this approach is that
ducible, symbol error rate at the receiver. The latter methods, the calculation of the time domain output of the IDWPT and
pursued in this paper, allow perfect recovery of the original IDFT cascade can performed by simple element-by-element
data at the expense of transmitting side information resulting multiplications of QAM symbols and frequency responses of
in a reduced throughput. Among the distortion-less PAPR cascaded filter banks.

195
1-4244-1457-1/08/$25.00 © IEEE
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE CCNC 2008 proceedings.

II. BACKGROUND H −1 {x}(n) = x(k)h(n − 2k) (6)
kεZ
A. OFDM Signal 
In the baseband-equivalent OFDM transmitter with N sub- G−1 {x}(n) = x(k)g(n − 2k) (7)
carriers, N modulation symbols, X[k] k = 0, 1, · · · , N − 1, kεZ

are mapped over the time interval [0, T ] on to the continuous Symbol sequences processed by the IDWPT can now be
time OFDM signal, x(t), with its discrete time version x[n] represented by discrete-time signals that are built by successive
referred to as the OFDM frame: iterations of H −1 and G−1 operators. At the receiver, a DWPT
N −1 is used to recover the original transmitted symbols deploying
1 
x[n] = X[k]ej2πkn/N (2) inverse operations represented by H and G. In the digital
N communication terminology, the usage of IDWPT and DWPT
k=0
is equivalent to transmit and matched filtering, respectively [6].
where n=0, ..., N−1, and x(t) is sampled at the Nyquist rate T1 . The IDWPT and DWPT operations are usually represented
In (2), the OFDM symbols x[n] are related to X[k]’s through by a wavelet packet tree as shown in Fig. 1. The nodes of
an IDFT and when N is power of two, they are calculated us-
ing the inverse fast Fourier transform (IFFT). The modulation
symbols X[k] are usually QAM symbols or, as it is the case
in this paper, pre-processed QAM symbols with the objective
of minimizing the PAPR of the transmitted OFDM frame.
The continuous PAPR, PAPRc , of x(t) is the ratio of the
maximum instantaneous power to the average power. The
discrete PAPR of x[m], upsampled version of x(t), denoted
by PAPR is calculated as:
2
|x[m]|
PAPR = max   (3)
0≤m<up·N −1 2
E |x[m]|
Fig. 1. Wavelet packet trees : (a) IDWPT (synthesis/modulation) Tree and
(b) DPWT (analysis/demodulation) Tree.
where E(·) is the expectation operator and up is the upsam-
pling factor over the Nyquist rate. The PAPR of an over wavelet packet tree T (D) are represented as a pair (−l, p)
sampled version of x(t) calculated as in (3) provides an where l, (l =< D) is current depth or level of node in the tree
accurate approximation of the PAPRc provided that the over of depth D, and p is position of the node at the level. Thus
sampling factor up is at least four. In this paper, up = 4 is an unpruned tree of depth D will have the following nodes:
used when deciding on the transformation used to construct
the signal going into the channel. ∇(T D ) = {(0, 0),
(−1, 0), (−1, 1),
B. Wavelet Packet Transform
(−2, 0), .., (−2, 3),
Wavelet packets are a class of generalized Fourier trans-
forms with basis functions localizing well in both time and fre- (−(D − 1), 0), .., (−(D − 1), 2(D−1) − 1),
quency domains. They are constructed using quadrature mirror (−D, 0), .., (−D, 2D − 1)} (8)
filter (QMF) pairs h(n) and g(n), satisfying some conditions where {(−D, 0), .., (−D, 2D −1)} = ∇l (T D ) are the leaf (ter-
leading usually to h(n) and g(n) being low- and high-pass minal) nodes and {(0, 0), .., (−D, 2(D−1) − 1)} = ∇nl (T D )
filters, respectively [4], [5]. The QMFs h(n) and g(n) of length are the non-leaf (non-terminal) nodes of the binary tree.
L are recursively used to define the sequence of basis functions Pruning involves the reconfiguration of a wavelet packet tree
ϕn (t), called wavelet packets, which satisfy various orthogo- by selective joining or splitting of appropriate nodes of a tree.
nality properties useful in calculating the discrete WPT. The 
A node join operation, (T, n) reconfigures a binary tree, T ,
latter takes at its input N numbers and produces N numbers at by converting a non-leaf node n into a leaf node. The child
the output, similarly to an DFT operation. The IDWPT can be nodes (and sub-trees) of n are reconstituted into node n. A
viewed as a synthesis operation representing the data symbols node split operation, (T, n) reconfigures a binary tree T by
as a sum of shifted ϕj [k] waveforms. An alternate way to converting a leaf node n into two child nodes which are also
visualize wavelet packet transforms is by defining a set of leaf nodes. In this paper, pruned trees are created by joining
operators (H −1 , G−1 , H and G) based on h(n) and g(n) as of one or more non-leaf nodes in ∇nl (T D ) of a tree T (D) .
follows [4]:
 III. S ELECTED M APPING USING WAVELET PACKET T REE
H{x}(2n) = x(k)h(k − 2n) (4) P RUNING
kεZ
In this section a OFDM PAPR reduction technique utilizing

G{x}(2n) = x(k)g(k − 2n) (5) wavelet packet tree pruning (WPTP) is presented. This tech-
kεZ nique was originally proposed for PAPR reduction in wavelet

196

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE CCNC 2008 proceedings.

packet modulation [3]. This technique is similar to selective (T (6) ) using successive node join, (T (D) , n), operations as
mapping (SLM) [1], [2] for OFDM where a large set of described earlier.
statistically independent signals are generated from the same
information and the signal with the lowest PAPR is selected
for transmission.
Fig. 2 shows a OFDM communication system with a
wavelet packet tree pruning (WPTP) block for PAPR reduc-
tion. The transmitter has a WPTP block, which is the prunable
version of the IDWPT (synthesis) tree shown in Fig.1a.
and is represented by Fig. 3a. Its corresponding prunable
DWPT(analysis) block is present at the receiver and is repre-
sented by Fig. 3b. Encoded or modulated data, X m is input to
a wavelet packet tree, the tree is iteratively pruned to generate
alternative sequences, Am , which are then passed through an
IDFT block. A peak acceptor then determines which output
Fig. 4. Full wavelet packet tree of depth=6.
sequence, sm , to transmit. If the sequence, sm , achieves a low
PAPR (global minimum) or if the PAPR is below a certain The multi-node join procedure for PAPR reduction starts
threshold, P AP Rth , it is selected for transmission, otherwise with an unpruned wavelet packet tree T (6) , which is iteratively
new sequences are generated until a sequence with low PAPR pruned using an allowable set of nodes P = ∇nl (T (6) ) E (6)
is found or all pruned trees are searched. for a search of a tree which achieves the lowest PAPR. The
exclusion set E (6) contains the set of nodes for which are not
allowed to be joined. The E set can be used restrict certain
sub-trees from being used in the PAPR reduction procedure.
For each iteration i <= N , starting with a unpruned tree,
Ti = T (6) , the PAPR, ρn is calculated for the output xm n of the
IFFT for each pruned tree, Tin formed by joining each node,
n ∈ Pi . The pruned tree with the minimum PAPR ρnmin
is stored as Ti+1 = Tinmin for the next iteration. In each
iteration Ti is successively pruned. The allowable set Pi+1 for
the next iteration is recalculated based on starting tree Ti+1 for
the next iteration as Pi+1 = ∇nl (Ti+1 )  E (6) . Finally, after
Fig. 2. OFDM system with wavelet packet tree pruning. all N iterations have been complete, the output of the pruned
tree which achieves the minimum PAPR of all pruned trees
is selected for transmission. Information about which pruned
tree achieved the minimum PAPR is also transmitted as side
information.
In the tree T (6) all nodes in ∇nl (T 6 ) may be used for
pruning, however, the pruning set nodes are limited to the
following:

PA = {(−5, 0), (−5, 1), ..., (−5, 31)} (9)


Fig. 3. One sided WPTP blocks. (a) Transmitter Pruning (b) Receiver
Pruning.
PB = {(−3, 0), (−3, 1), ..., (−3, 7),
The transformation performed by the pruning operation is (−4, 0), (−4, 1), ..., (−4, 15),
reversible. At the receiver using information sent by the trans- (−5, 0), (−5, 1), ..., (−5, 31)} (10)
mitter about the pruned tree, the corresponding WPTP receiver
block (Fig. 3b.) is used to recover the original information Note that the exclusion set E, dictates where on the original
transmitted. The DWPT in the receiver WPTP block is pruned tree pruning should occur. The E set can be modified to obtain
identical to the DWPT in the transmitter WPTP block, together allowable sets P A and P B .

forming a perfect reconstruction structure. A node join operation (Ti , n) with n ∈ Pi reconfigures
In the PAPR reduction algorithm of [3], a single node in the tree Ti . The node join operation will delete the child nodes
∇nl (T D ) was joined to provide a PAPR reduction. In this and sub-trees of node n. The set of child nodes and nodes of
paper, a multi node join procedure is utilized. Fig. 4 shows a sub-trees of a node n = (−l, p), defined by C (−l,p) are all
unpruned tree, T (6) of depth D = 6. A number of pruned reconstituted to form a new leaf node at node n. Thus in the
trees (or sub-trees) can generated from the unpruned tree multi-join algorithm Pi must be recalculated after each join

197

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE CCNC 2008 proceedings.
8
At the transmitter, the output of the two sided WPTP block,
OFDM (Unoptimized) X can visualized as a scatter pattern as shown in Figs. 7. It was
WPTP (Optimized)
7 be observed that pruning of the wavelet packet tree generates
alternative signalling points for the same data sequence input
6
similar the well known TI scheme [2], [7]. In the iterative tree
5 pruning process, data points are assigned to these signalling
Peak Level

points such that the PAPR of the output signal is minimized.


4
Fig. 7 shows the optimized output after termination of the
3

6
2 Unpruned WP Tree
Pruned WP Tree
1 4

0
0 10 20 30 40 50 60 2
Time Index

Quadrature
0
Fig. 5. Peak reduction using wavelet packet tree pruning.
−2
as many nodes and sub-trees are affected by a join operation.
Pi is recalculated after each join as follows:
−4
(−l,p)
Pi+1 = Pi  C (11)
−6
−6 −4 −2 0 2 4 6
where  is the set difference operation. For example, for

In−Phase

(T (6) , n) with the allowable set P B , and pruning node


n = (−4, 0), Fig. 7. Optimum reconfiguration data onto alternate signalling points.
WPTP algorithm involving node joins at different depths of
C (−4,0)={(−5, 0),(−5, 1),(−6, 0),(−6, 1),(−6, 2),(−6, 3)} (12) the tree. Each node join operation, modifies the original QAM-
16 scatter pattern by introducing new signalling points. Some
This joining operation prunes the tree and reduces the allow-
of the original data signalling points occupy the new signalling
able set P B to the following:
points. In some cases the original QAM signalling points

PB = P B  C (−4,0) (13) are not used for the representation of the input sequence.
 The new locations of the alternative signalling points are
PB = {(−3, 0), (−3, 1), ..., (−3, 7),
observed to be scaled versions of the original signalling points
(−4, 0), (−4, 1), ..., (−4, 15), {(−3, −3), (−3, −1), ..., (1, 1), (1, 3), (3, 3)} of the QAM-16
(−5, 2), ..., (−5, 31)} (14) constellation with the scaling factor depending on the depth
and position of the joined node, n = (−l, p). The scaling
Fig. 5 shows the time domain envelope of the outputs, factor is also dependent on the filters h(n) and g(n) which
xm , of the IFFT block using the full (unpruned) tree and are used in the construct of the wavelet packet tree, for the√
the optimum pruned tree after the application of the WPTP Haar (Daub-1) wavelet the scaling factor are multiples of ±2 2
algorithm with the latter exhibiting reduced peaks. for both the quadrature and imaginary axes. At the transmitter,
Another variation of PAPR reduction algorithm using WPTP only the IDWPT tree of the two sided WPTP block is pruned.
is developed by observing that the tree pruning operation The DWPT tree is an unpruned tree of the same depth (D).
modifies the signalling points in the frequency domain. This is However, the depth of the DWPT can theoretically be any
similar to tone injection or constellation extension methods of value less than D. At the receiver the corresponding two sided
PAPR reduction for OFDM [2]. This is clearly evident when WPTP block (Fig. 6a.) is present, however the only the DWPT
the original one sided WPTP pruning block of Fig. 3a. is is pruned to recover the transmitted sequence. The tree pruning
replaced with a two sided WPTP block shown in Fig. 6a. procedure is the same as int the one sided WPTP algorithm
with the only difference being that the pruned tree is now a
part of a two sided WPTP block.
IV. O PTIMIZATION FOR L OWER C OMPLEXITY AND
A MOUNT OF S IDE I NFORMATION
The number of iterations in the algorithm does have not be
large. From the analysis of PAPR reduction versus the number
of iteration it was observed that at most times minimum tree
Fig. 6. Two sided WPTP blocks. (a) Transmitter Pruning (b) Receiver Pruning was found within 5 or less iterations. Note that the number

198

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE CCNC 2008 proceedings.
of iterations is dependent on which nodes in allowable set 0
10
P are joined during previous iterations. With each node join OFDM64−Ref
removing the joined node and its subtrees from the set P . WPTP (5)
WPTP (2,3,4,5)
Therefore, the size of P reduces rapidly after each iteration. WPTP D(5)
Thus number of iterations or passes can be set to N = 5 to
−1 WPTP D(2,3,4,5)
10
TI − Ref
reduce the processing delay. The loop controlling the number

Pr(PAPR>PAPR )
0
of iterations can also be optimized by setting a desired PAPR
threshold value, P AP Rth . In this scheme the search procedure −2
10
stops as soon as one of iterations finds a pruned tree with a
PAPR below P AP Rth . This reduces the number of IDWTs
which have to be calculated; however, it offers slightly lower −3
10
PAPR reduction performance.
Another area of optimization is in the DWPT/IDWPT
operations themselves. Through the recursive application of
the Noble identities it is possible to reduce the number of −4
10
4 5 6 7 8 9 10 11 12
filters and samplers required in the WPM system. Noble PAPR0 (dB)
identities provide the means in (4-7) to interchange the order
of operation of the filters and sampler and thus allow for Fig. 8. CCDF Results.
the grouping and combining of multiple filters and samplers
of a wavelet packet tree. Noble identities can be used to
reduce multiple filters and samplers to one of each per data carried out for OFDM with 64 subcarriers using a QAM-16
carrier [5]. The information about which pruned tree was modulation. The Haar (Daub-1) wavelet packet bases were
used to obtain the reduced PAPR needs to be transmitted used to construct the wavelet packet trees. A raised cosine
to the receiver as side information, I. The receiver uses this filter of α = 0.1 with a upsampling factor of 5 was used at the
information to reconstruct the pruned tree for demodulation. transmitter. The CCDF results are are shown in Fig. 8. In this
The number of possible pruned trees, M , and thus maximum figure, the performance curves WPTP(5) and WPTP(2, 3, 4, 5)
number of transformation, T s (·) in (1), affect the amount of refer to the application of the allowable sets P A and P B ,
side information [3]. The lowering of the size of the allowable respectively. The figure also includes the performance curve
set P leads to the reduction in the number of pruned trees and for alternative signalling point scheme, WPTP D(5) and WPTP
thus the amount of the side information to be transmitted. D(2, 3, 4, 5) also utilizing the allowable sets P A and P B ,
Furthermore, the exclusion set E can further reduce the size respectively. Fig. 8 demonstrates the performance gains are
of P by disallowing certain subtrees from being pruned in the in the range of 3.5dB and 5.5dB over the original OFDM
algorithm. signal. CCDF plot of PAPR reduction using the Tone Injection
The size of the index id used to represent the pruned trees (TI) method is also included as reference. The simulations
is given by: using the larger allowable set P B performed better in both
schemes. The larger allowable set P B provides more degrees
I = log2 (Size(P ) × N ) bits (15)
of freedom in the creation of pruned trees, however, its larger
where P is the allowable set and N is the maximum number size also increases the amount of redundant side information,
of iterations used in the search for the pruned wavelet packet I, transmitted to the receiver.
tree that achieves the minimum PAPR.
R EFERENCES
V. S IMULATION R ESULTS [1] M. Breiling, S. H. Müller, and J. B. Huber, “SLM peak power reduction
without explicit side information,” IEEE Commun. Lett., vol. 5, no. 6, pp.
Simulations were conducted to verify the performance of the 239–241, 2001.
proposed PAPR reduction schemes using WPTP. In this paper, [2] J. Tellado, Multicarrier Modulation with Low PAR. Applications to DSL
the performance of the proposed PAPR reduction schemes are and Wireless. New York: Kluwer Academic Publishers, 2002.
[3] M. Baro and J. Ilow, “PAPR reduction in wavelet packet modulation
demonstrated through the complementary cumulative distribu- using tree pruning,” in 2007 IEEE 65th Vehicular Technology Conference
tion function (CCDF) of PAPR, which is a performance metric VTC2007-Spring, Dublin, Ireland, Apr. 2007.
independent of the transmitter amplifier. Given the reference [4] A. R. Lindsey, “Wavelet packet modulation: a generalized method for
orthogonally multiplexed communications,” in IEEE 27th Southeastern
level P AP R0 > 0, the probability of a PAPR being higher Symposium on System Theory, 1995, pp. 392–396.
than the reference value is the CCDF and is expressed as [5] D. Daly, C. Heneghan, A. Fagan, and M. Vetterli, “Optimal wavelet packet
follows [2]: modulation under finite complexity constraint,” in Proc. ICASP, vol. 3,
2002, pp. 2789–2792.
CCDF (P AP R0 ) = P r{P AP R > P AP R0 } (16) [6] A. Jamin and P. Mähönen, “Wavelet packet based modulation for wire-
less communications,” Wiley Wireless Communications and Networking
For practical reasons, the CCDF of PAPR is calculated based Journal, vol. 5, no. 2, pp. 123–137, Mar. 2005.
[7] A. Jayalath and C. Athaudage, “On the PAR reduction of OFDM signals
on the percentage of the OFDM/WPM frames for which using multiple signal representation,” IEEE Commun. Lett., vol. 8, no. 7,
PAPR exceeds the threshold P AP R0 . The simulations were pp. 425–427, 2004.

199

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE CCNC 2008 proceedings.

You might also like