1608 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 57, NO.

6, JUNE 2009

Intercarrier Interference Reduction in

OFDM Systems Using Low Complexity Selective Mapping
Abolfazl Ghassemi, Student Member, IEEE, and T. Aaron Gulliver, Senior Member, IEEE

Abstract—Selective mapping (SLM) has been employed to
reduce the peak interference-to-carrier ratio (PICR) in OFDM
systems. Here, we significantly reduce the computational com-
plexity of SLM using subsets of intermediate signals within the
inverse fast Fourier and fast Fourier transforms, while achieving
performance close to that previously obtained.
Index Terms—Orthogonal frequency-division multiplexing,
peak interference-to-carrier ratio (PICR), selective mapping
(SLM).
I. I NTRODUCTION
computational complexity of the FFTs used to compute the
PICR. We examine our proposed SLM technique and compare
it with the technique in [1]. Numerical results are given which
show that there is only a slight degradation in PICR reduction
for the same number of SLM sequences, while achieving a
significant complexity reduction.
II. PICR AND SLM
We consider Moose’s frequency offset estimation method
N −1
[4]. Let {X(k)}k=0 denote the frequency-domain OFDM

O RTHOGONAL frequency division multiplexing signal at the transmitter, where N is the number of IFFT
(OFDM) is an effective multicarrier transmission points (subcarriers) and k is the frequency index. The time-
technique for wireless communications over frequency- domain OFDM signal is obtained by taking an N -point inverse
selective channels. However, OFDM is sensitive to frequency discrete Fourier transform (IDFT) of X(k)
offset due to oscillator errors and/or doppler spread. N −1
Frequency offset introduces inter-carrier interference (ICI) 1 
x(n) = X(k)TN−nk 0≤n≤N −1 (1)
with characteristics analogous to Gaussian noise [2], N
k=0
degrading the bit error rate (BER) performance.
Selective mapping (SLM) was employed in [1] to reduce where n is the discrete-time index, TN = e−j2π/N (known
the peak interference-to-carrier ratio (PICR) [1]. This tech- as the twiddle factor), and j 2 = −1. The corresponding
nique was adapted from peak-to-average power ratio (PAPR) analog signal is then up converted to the carrier frequency,
reduction techniques, exploiting the similarity between the and transmitted over the channel. We assume the channel is
PAPR and PICR. With SLM, several sequences are gener- AWGN. At the receiver, the reverse operations are employed.
ated from the data using a set of phase sequences, and the With a mismatch between carrier frequencies in the transmitter
one with the lowest PICR is chosen for transmission. This and receiver, the received signal has a frequency offset. Using
mitigates the effect of frequency offset. As a PAPR reduction (1) and taking the FFT of the received signal in order to
technique, each SLM sequence is generated using one inverse recover the original signal, we obtain
fast Fourier transform (IFFT) operation. Therefore, to obtain 
N −1
several sequences SLM has a high computational complexity. R(k) = X(b)S(b, k) + n(k)
This complexity is compounded when SLM is employed for b=0
PICR reduction. This is because ICI occurs at the receiver, 
N −1

so the PICR computation for each SLM sequence requires an = X(k)S(k, k) + X(b́)S(b́, k) + n(k) (2)
IFFT and an FFT [1]. b́=0
To reduce the computational complexity, we exploit the
where b́ = k,
IFFT/FFT structure which has a number of identical discrete
Fourier transforms (IDFTs) or discrete Fourier transforms sin(π(b́ − k + ))
S(b́, k) = ejπ(b́−k+) , (3)
(DFTs) at each stage. We first randomly divide the identical π(b́ − k + )
IDFTs into a number of subsets. Then, the inputs to each
subset are phase rotated before generating the time-domain and  is the normalized frequency offset. The first term in (2)
outputs. These time-domain signals are combined to generate is the original signal shifted by S(k, k). Since S(k, k) is only
different SLM sequences. This approach can significantly a function of , the frequency offset has the same effect on
reduce the computational complexity. Since the FFT and all subcarriers. The second term in (2) represents the ICI on
IFFT algorithms have reciprocal stages, it also reduces the the kth subcarrier and is a function of  and the transmitted
N −1
data X = {X(k)}k=0 . Let I(k) denote the second term in
Paper approved by A. Anastasopoulos, the Editor for Iterative Detection, (2). The PICR of X is defined as [1]
Estimation and Coding of the IEEE Communications Society. Manuscript
received September 15, 2007; revised February 28, 2008. max |I(k)|2
The authors are with the Department of Electrical and Computer Engineer- 0≤k≤N −1
PICR(X) = (4)
ing, University of Victoria, P.O. Box 3055, STN CSC, Victoria, BC, V8W |S(k, k)X(k)|2
3P6 Canada (e-mail: {aghassem, agullive}@ece.uvic.ca).
Digital Object Identifier 10.1109/TCOMM.2009.06.070469 Minimizing (4) will reduce the ICI.
0090-6778/09$25.00 
c 2009 IEEE

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 2, 2009 at 04:43 from IEEE Xplore. Restrictions apply.
 GHASSEMI and GULLIVER: INTERCARRIER INTERFERENCE REDUCTION IN OFDM SYSTEMS USING LOW COMPLEXITY SELECTIVE MAPPING 1609

Consider a phase rotated version of X(k) given by P u (k) = and indices for an intermediate signal are denoted by y and
 for an input y and time index n, respectively, and 
u
X(k)ejφ (k) where the φu (k) are randomly chosen from a n k for a
set of phases, typically {0, π} [6], [7]. In SLM as a PAPR frequency index k. For each identical N/rv−1 -point DFT at
reduction technique, the time-domain OFDM signal pu (n) stage v, (9) can be expressed as
is obtained using the IDFT of P u (k) according to (1). The N
rv −1
 r−1  
   N
lowest PAPR signal pu (n) is chosen for transmission from Y (r 
α
k+k0 ) = y (
α
n + v i)T ik0
T  k0
n 
TN/r
k
n
v
r
r N/r v−1
among U candidate OFDM signals, including the original  =0
n i=0
x(n). To determine the selected sequence at the receiver, (10)
log2 U bits of side information must be sent along with the where  k = 0, . . . , N/rv − 1, n
 = 0, . . . , N/rv − 1 and α, α =
data. Constructing the sequences pu (n) requires U IDFTs, 1, . . . , r v−1
, denotes a particular N/rv−1 -point DFT at stage
which results in high computational complexity for typical v, v = 1, . . . , m. Since there are rv−1 identical N/rv−1 -point
values of U . DFTs at stage v and each is individually reduced to N/rv -
Considering SLM as a PICR reduction technique, the re- point DFTs in the remaining m − v stages, we can formulate
sulting ICI is given by (9) corresponding to the inputs at stage v using (10) as
N
v−1
r v−1
r −1

N −1 
rv

ISLM (k) = e jφu (b́)

X(b́)S(b́, k) (5) Y (rk + k0 ) = Y (r
α
k + k0 ) =
α=1  =0
α=1 n
b́=0,b́=k r−1  
 N 
and the lowest PICR is obtained from y (
α
n + v i)T ik0
T  k0
n
T kn
(11)
 2
 r r N/r v−1 N/r v
max |ISLM (k)| i=0
0≤k≤N −1
PICRlowest (X) = min |S(k,k)X(k)|2 (6)
φ1 (k),...,φU (k) A phase rotated version of yα (
n+ r v i)
N
is given by
Similar to PAPR reduction with SLM, at the transmitter we N jφu (n)
n) = yα (
Pαu ( n+
i)e (12)
require U IDFTs to obtain the time-domain sequences pU (n). rv
However, from (6), we need the PICR in the frequency-domain and the phase rotated SLM sequences within the IFFT are then
which requires U DFTs at the transmitter. Hence, this results in
n), P2u (
P u (n) = [P1u ( n), . . . , Pαu (
n), . . . , Pruv−1 (
n)] . (13)
extremely high computational complexity and motivates us to
propose a low complexity SLM technique for PICR reduction. We propose a low complexity SLM technique where a subset
of the Pαu (
n) are phase rotated so the time-domain pu (n) are
III. P ROPOSED SLM T ECHNIQUE FOR PICR R EDUCTION obtained using only some of the rv−1 × N/rv−1 -point DFTs.
We use intermediate signals within an N -point IFFT and Let λ denote the number of subsets of N/rv−1 -point DFTs
FFT using decimation in frequency (DIF) to obtain U par- where λ ≤ U , 2 ≤ λ < rv−1 , and λ is a power of 2 (λ = 20
tially phase rotated SLM sequences and compute U PICRs, when there are no partial subsets). The location of the DFTs
respectively. Hence, we first consider time-domain SLM se- is randomly selected using the following steps.
quences which are obtained by taking the IFFT of X. The 1) Construct
 a pseudo-random sequence S1 =
v−1
corresponding FFT is considered since the IFFT operation is s1 s2 . . . , sα , . . . , sr over Zλ , the integers
[8] N −1 ∗ modulo λ, where sα corresponds to the phase rotated
1  ∗ sequence Pαu ( n) in (13).
 T
x(n) = X (k)T N nk
(7)
N 2) Obtain the matrix C = S1 S2 . . . S U of dimension
k=0 λ

∗
Let y(.) represent X (.). The expression inside the brackets
U
λ × rv−1 , where
in (7) is the discrete Fourier transform (DFT) of X ∗ (.), i.e. U
Sl+1 = Sl + λI mod U, 0≤l≤ −1 (14)

N −1 λ
Y (k) = y(n)T nk
N (8) and I is the identity vector of dimension rv−1 .
n=0 Hence, the elements of C can be expressed as
It is well known that an FFT algorithm recursively converts ⎡ 1 v−1 ⎤
s1 ( n) s21 ( n ) . . . sα
1 ( n) . . . sr1 ( n)
the DFT computation in (8) to r identical N/r-point DFTs ⎢ s1 ( 2 r v−1 ⎥
⎢ 2 n) s2 ( n) . . . s2 ( n) . . . s2 ( n)
α
through m = logr N stages where each N/r-point DFT is ⎥
⎢ .. .. .. .. .. .. ⎥
individually reduced for the remaining stages. This results in ⎢ . . . . . . ⎥
⎢ ⎥
rv−1 identical N/rv−1 -point DFTs at a particular stage v. The C=⎢ 1 2 r v−1 ⎥
⎢ sl ( n) sl ( n) . . . sl (
α
n) . . . sl ( n) ⎥
value of r corresponds to a radix-r FFT algorithm. The DIF ⎢ .. .. .. .. .. .. ⎥
⎢ ⎥
radix-r algorithm can be derived from (8) as ⎣ . . . . . . ⎦
1 2 r v−1
N r−1   s U ( n) s U ( n) . . . s U (
α
n) . . . s U ( n)
r −1
  N λ λ λ
(15)
λ

Y (rk + k0 ) = y(n + i)Tr ik0

TN
nk0
TN/r
kn

n=0 i=0
r where each row of C corresponds to λ SLM sequences. Each
(9) element sα n) denotes the location of the αth DFT which is
l (
where k = 0, . . . , N/r − 1, n = 0, . . . , N/r − 1, and k0 , 0 ≤ phase rotated based on the randomly chosen φu ( n) ∈ {0, π}
k0 ≤ r−1, denotes the index of the butterfly outputs. Symbols in (12) [6], [7].

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 2, 2009 at 04:43 from IEEE Xplore. Restrictions apply.
 1610 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 57, NO. 6, JUNE 2009

The proposed technique has v stages with one N -point DFT,

and (m−v) stages with ( Uλ +1)rv−1 × rv−1 N
-point DFTs. Note
that the first sequence is computed over all rv−1 -point DFTs.
This is illustrated in Fig. 1 and shows a significant complexity
reduction compared to the SLM in [1], which requires m
stages of U × N -point DFTs.
Example: consider N = 32, v = 4, r = 2, U = 4, and λ = 2.
We generate S = [01001110] over Z2 and obtain
 
0 1 0 0 1 1 1 0 Fig. 1. PICR computation for the proposed SLM sequences including the
C= (16) original sequence: (a) IFFT, (b) FFT.
2 3 2 2 3 3 3 2
Assume zero corresponds to the original data which is com- 10
0

Original OFDM
puted over all rv−1 DFTs i.e. this SLM sequence is not rotated. SLM [1]
From (16), the 4 phase rotated sequences are Proposed SLM m−v=6
Proposed SLM m−v=5
 
−1
10 Proposed SLM m−v=4
0
P (n) = P10 (
n)P20 (
n)P30 (
n)P40 (
n)P50 (
n)P60 (
n)P70 (
n)P80 (
n) , Proposed SLM m−v=3
 Proposed SLM m−v=2
1
P (n) = P11 ( n)P31 (
n)P21 ( n)P41 (
n)P51 (
n)P61 ( n)P81 (
n)P71 ( n) ,
 10
−2
2
P (n) = n)P22 (
P12 ( n)P32 ( n)P52 (
n)P42 ( n)P62 (
n)P72 (
n)P82 (
n) ,

CCDF

P 3 (n) = P13 ( n)P33 (
n)P23 ( n)P43 (
n)P53 (
n)P63 ( n)P83 (
n)P73 ( n) . −3
10

where a bold superscript represents a rotated DFT within an

SLM sequence. −4
In order to calculate the frequency-domain SLM sequences, 10

as indicated in [1], we require a value for  at the transmitter

to obtain I(k). However, the offset frequency  is not known −5
10
at the transmitter. Thus, to obtain the PICR, a worst case  is −13 −12 −11 −10 −9 −8 −7 −6 −5 −4 −3 −2 −1
considered, wc , where || < wc [1]. We assume the inputs PICRo(dB)

and outputs of the IFFT are in reverse and normal order,

Fig. 2. PICR CCDF for the proposed SLM compared to SLM in [1] for
respectively. Hence, the FFT computation is symmetric to the various m − v, with  = 0.1, r = 2, U = 8, N = 1024, and λ = 20 .
IFFT computation. As a consequence, to obtain the PICR of U
SLM sequences, we require v stages of ( Uλ + 1)rv−1 × rv−1 N
-
point DFTs and (m−v) stages of U ×N -point DFTs. A block additive complexity of the IFFT computation for our SLM (see
diagram illustrating this computation for our SLM is depicted Fig. 1) is
in Fig. 1. For our technique, we can compute the ICI based
on (5), and the lowest PICR is obtained using (6). 
v−1
U  m
Mtotal
IF F T
= Mβ + ( + 1) Mβ (17)
In the above example, consider P 1 (n) and P 2 (n). The β=1
λ
β=v
corresponding
 time-domain SLM sequences are 
p1 (n) = p11 ( n)p13 (
n)p12 ( n)p14 (
n)p15 (
n)p16 ( n)p18 (
n)p17 ( n), and
U
p2 (n) = p21 (n)p22 (
n)p23 ( n)p25 (
n)p24 ( n)p26 (
n)p27 (
n)p28 (
n) . total = Av [v + (
AIF FT
+ 1)(m − v)] (18)
p1 (n) and p2 (n) can be combined to generate another time- λ
domain SLM p c (n), i.e. respectively. Similarity, we can obtain the corresponding com-
 sequence 
p c (n) = p21 (
n)p12 (
n)p23 (
n)p24 (
n)p15 (
n)p16 (
n)p17 (
n)p28 (
n) . plexities for the FFT.
It is obvious that this sequence does not require the computa- To compare the multiplicative and additive complexity
tion of N/rv−1 -point DFTs thus further reducing complexity. between two PICR techniques, we define the reduction ratios
1 2
The locations of the phasing and de-phasing sequences Rmul = 1 − (Mtotal /Mtotal ) and Radd = 1 − (A1total /A2total ),
n) are designed once (offline) for both the FFT and IFFT.
P u ( respectively. Table I presents the multiplicative and additive
As a consequence, our SLM approach does not introduce complexity reductions for different λ and m − v values with
any complexity to the system and requires the same side N = 1024 and r = 2. We consider BPSK modulated OFDM
information as the technique in [1]. signals with N = 1024, U = 8, and  = 0.1, as in [1].
The effect of m−v and λ on the PICR reduction is examined The complementary cumulative distribution function (CCDF)
in the next section. of the PICR (4) is used to evaluate the PICR performance. We
randomly choose the phase sequences φu (ñ) from {0, π}.
Fig. 2 shows the CCDF of the proposed SLM with λ = 1
IV. N UMERICAL R ESULTS
and various m − v. For m − v = 5, the performance of the
To obtain the computational complexity, let Mv and Av proposed SLM is similar to that with the SLM in [1]. The
 k0
denote the number of twiddle factors TN/rn
v−1 (representing CCDF of the PICR for different values of λ is shown in Fig.
complex multiplications) and additions at stage v for the 3 for m − v = 5. This shows that randomly selected α has
DIF algorithm, respectively [3]. The overall multiplicative and better PICR performance than equally spaced or adjacent α.

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 2, 2009 at 04:43 from IEEE Xplore. Restrictions apply.
 GHASSEMI and GULLIVER: INTERCARRIER INTERFERENCE REDUCTION IN OFDM SYSTEMS USING LOW COMPLEXITY SELECTIVE MAPPING 1611

TABLE I
M ULTIPLICATIVE AND A DDITIVE C OMPLEXITY R EDUCTION WITH N = 1024, U = 8, λ = 2, 4, AND 2 ≤ m − v ≤ 6

m−v IFFT FFT

Proposed SLM compared to Proposed SLM compared to Proposed SLM compared to Proposed SLM compared to
SLM in [1], λ = 2 SLM in [1], λ = 4 SLM in [1], λ = 2 SLM in [1], λ = 4
Rmul (%) Radd (%) Rmul (%) Radd (%) Rmul (%) Radd (%) Rmul (%) Radd (%)
2 86 83 87 86 59 50 74 63
3 84 80 86 85 53 44 66 55
4 80 78 85 84 46 38 58 47
5 78 75 84 83 40 33 48 39
6 75 73 82 81 31 25 39 31

0
10
Original OFDM V. C ONCLUSIONS
SLM [1]
Proposed SLM λ=2
0 The main drawback in using the SLM technique in [1]
10
−1 Proposed SLM λ=21, random to reduce the peak interference-to-carrier ratio (PICR) is the
1
Proposed SLM λ=2 , equally spaced computational complexity of the multiple inverse fast Fourier
1
Proposed SLM λ=2 , adjacent transforms (IFFTs) and fast Fourier transforms (FFTs). To
2
Proposed SLM λ=2 , random
−2 3 reduce this complexity, we generated the time-domain SLM
10 Proposed SLM λ=2 , random
sequences using partial inputs to the middle stages of the
CCDF

IFFT. As a consequence, the computational complexity of

10
−3 the multiple FFTs to obtain the PICR was also reduced. A
comparison between this new technique and the approach in
[1] was presented in terms of PICR reduction and compu-
10
−4
tational complexity. Our technique significantly reduces the
computational complexity while providing approximately the
same PICR performance.
−5
10
−13−12−11−10 −9 −8 −7 −6 −5 −4 −3 −2 −1 0
PICRo(dB) ACKNOWLEDGMENT
The authors would like to thank the anonymous reviewers
Fig. 3. PICR CCDF for the proposed SLM compared to SLM in [1] for
various λ, with  = 0.1, r = 2, U = 8, N = 1024, and m − v = 5. for their helpful comments.

−4
With CCDF=10 and randomly selected α, there is a slight
performance degradation with λ = 2 compared to λ = 1.
In this case, the proposed SLM achieves multiplicative and
additive complexity reductions of Rmul = 78% and Radd =
75%, respectively, for the IFFT compared with the SLM in
[1]. The corresponding complexity reductions for the FFT are
Rmul = 40% and Radd = 33%.
The performance of the proposed technique is slightly worse
than the SLM in [1] when an equal number of sequences is
employed. However, SLM in [1] with U = 6 performs very
close but slightly worse than our proposed technique with U =
8, while we achieve multiplicative and additive complexity
reductions of Rmul = 53% and Radd = 50%, respectively,
for the IFFT. The corresponding complexity reductions for
the FFT are Rmul = 15% and Radd = 8%.
R EFERENCES
[1] K. Sathananthan and C. Tellambura, “Partial transmit sequence and
selected mapping schemes to reduce ICI in OFDM systems,” IEEE
Commun. Lett., vol. 6, no. 8, pp. 313-315 Aug. 2002.
[2] J. Armstrong, “Analysis of new and existing methods of reducing
intercarrier interference due to carrier frequency offset in OFDM,” IEEE
Trans. Commun., vol. 47, no. 3, pp. 365–369, Mar. 1999.
[3] A. Ghassemi and T. A. Gulliver, “PTS peak power reduction of OFDM
signals with low complexity IFFTs,” in Proc. IEEE Commun. Networks
and Services Research Conf., pp. 85–92, May 2007.
[4] P. H. Moose, “A technique for orthogonal frequency division multiplexing
frequency offset correction,” IEEE Trans. Commun., vol. 42, no. 10, pp.
2908-2914, Oct. 1994.
[5] J. Tellado, “Peak to average power reduction for multicarrier modulation,”
Ph.D. dissertation, Stanford University, Sept. 1999.
[6] G. T. Zhou and L. Peng, “Optimality condition for selected mapping in
OFDM,” IEEE Trans. Signal Processing, vol. 54, no. 8, pp. 3159-3165,
Aug. 2006.
[7] N. Ohkubo and T. Ohtsuki, “Design criteria for phase sequences in
selected mapping,” in Proc. IEEE Vehic. Tech. Conf., pp. 373-377, Apr.
2003.
[8] L. R. Rabiner and B. Gold, Theory and Application of Digital Signal
Processing. Englewood Cliffs, NJ: Prentice-Hall, 1975.

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 2, 2009 at 04:43 from IEEE Xplore. Restrictions apply.