You are on page 1of 8

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO.

10, OCTOBER 2008 3791

Theoretical Studies and Efficient Algorithm of


Semi-Blind ICI Equalization for OFDM
Hsiao-Chun Wu, Senior Member, IEEE, Xiaozhou Huang, Yiyan Wu, Fellow, IEEE,
and Xianbin Wang, Senior Member, IEEE

Abstract—The intercarrier interference (ICI) due to the advantage to combat the fading channels, time-varying
Doppler frequency shift, sampling clock offset, time-varying channel characteristics, sampling clock offset, Doppler
multipath fading and local oscillator frequency offset becomes frequency shift, and carrier frequency offset would destroy
the major difficulty for the data transmission via the wireless
orthogonal frequency division multiplexing (OFDM) systems. The the orthogonality of OFDM subcarrier waveforms and
existing ICI mitigation schemes involve the frequency-domain therefore intercarrier interference (ICI) occurs [1]. Different
channel estimation/equalization or the additional coding and time- and frequency-factors have been isolated and the
therefore require the pilot symbols which reduce the throughput. associated problems have been separately tackled one at a
The frequency-domain channel estimation/equalization relies on time in OFDM [2], [3]. However, this sequential procedure
the huge matrix inversion with high computational complexity
especially for the OFDM technologies possessing many subcar- operating on the discrete-time received sequences in [2], [3]
riers such as digital video broadcasting (DVB) systems and may not be ideal, since practically those factors cannot be
wireless metropolitan-area networks (WMAN). In our previous addressed separately and independently in the time domain.
work, we proposed a semi-blind ICI equalization scheme using According to [4], it is very difficult to accurately estimate
the joint multiple matrix diagonalization (JMMD) algorithm individual parameters such as sampling clock offset, channel
and empirically showed that the proposed method significantly
improved the symbol error rates for QPSK- and 16QAM-OFDM delays, frequency offset, phase noise, altogether in the time
systems. In this paper, we discuss the sufficient condition for domain and all of them can be combined to form an ICI
the theoretical ICI equalizability and also propose an alter- matrix. Such an ICI matrix involves either huge computational
native semi-blind ICI equalization method based on the joint complexity or significant throughput reduction according
approximate diagonalization of eigen-matrices (JADE) algorithm, to [5]–[7]. Therefore, the semi-blind ICI equalization (ICI
which is much more computationally efficient than our previous
method. matrix inverter) scheme, which relies on very few pilots, has
been proposed to improve the spectral efficiency [8].
Index Terms—Intercarrier interference (ICI), orthogonal fre-
quency division multiplexing (OFDM), joint multiple matrix Because the throughput or the spectral efficiency is a major
diagonalization (JMMD), joint approximate diagonalization of concern in the OFDM systems, we restrict our discussion
Eigen-matrices (JADE), equalizability. solely on the semi-blind and blind ICI equalization in this
paper. Our previous work in [8] provided a semi-blind ICI
I. I NTRODUCTION equalization method using the joint multiple matrix diagonal-
ization (JMMD) algorithm and a sparse ICI equalizer matrix
O RTHOGONAL Frequency Division Multiplexing
(OFDM) has been applied for contemporary wireless
local-area networks (WLAN), wireless metropolitan-area
was assumed therein. To the best of our knowledge, there
exists no theoretical studies on the ICI equalizability (ICI
matrix invertibility) and the validity of such a sparse ICI
networks (WMAN), wireless personal area network (WPAN)
equalizer matrix in [8] has not been justified either. In this
and digital video broadcasting (DVB) [1]. Despite of the
paper, we will show the sufficient conditions for the ICI
Manuscript received March 30, 2007; revised November 29, 2007 and equalizability and the approximability of the ICI equalizer
August 5, 2008; accepted August 15, 2008. The associate editor coordinating matrix as a sparse ridge matrix. Then, we will propose a
the review of this paper and approving it for publication was J. Olivier.
This work was supported by Information Technology Research Award for
new efficient semi-blind ICI equalization algorithm using the
National Priorities (NSF-ECCS 0426644) from National Science Foundation, joint approximate diagonalization of eigen-matrices (JADE)
Faculty Research Initiation Award from the Southeastern Center for Electrical recursion [9], which will reach the similar performance to the
Engineering Education, Research Enhancement Award from the Louisiana-
NASA Space Consortium and Faculty Research Award from Louisiana State
method in [8] but involve much less computational complexity.
University. The rest of this paper is organized as follows. The ICI
H.-C. Wu is with the Department of Electrical and Computer Engineering,
Louisiana State University, Baton Rouge, Louisiana 70803, USA (e-mail: matrix formulation for the OFDM systems is introduced in
wu@ece.lsu.edu). Section II. The ICI equalizer matrix or the inverse of the
X. Huang is with Texas Instruments, Germantown, Maryland 20784, USA ICI matrix is studied in Section III. The ICI equalizability
(e-mail: xiaozhouh@hotmail.com).
Y. Wu is with Communications Research Centre, Ottawa, Ontario, Canada or the invertibility of the ICI matrix, the approximability of
K2H 8S2 (e-mail: yiyan.wu@crc.ca). the ICI equalizer matrix both are addressed in Section III as
X. Wang is with the Department of Electrical and Computer Engineering, well. The new efficient semi-blind ICI equalization using the
University of Western Ontario, London, Ontario, Canada N6A 5B9 (e-mail:
wang@eng.uwo.ca). JADE algorithm is introduced in Section IV. The Monte Carlo
Digital Object Identifier 10.1109/T-WC.2008.070342 simulation results to compare different schemes for the QPSK-
1536-1276/08$25.00 
c 2008 IEEE

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 08:54 from IEEE Xplore. Restrictions apply.
3792 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 10, OCTOBER 2008

and 16QAM-OFDM systems are presented in Section V. where the transmitted data matrix Si is
Concluding statements will be drawn in Section VI.
Notations: C, R, Z denote the fields of complex, real and Si ≡ [si si+1 · · · si+q−1 ] ∈ CN ×q , (5)
 denotes
integer numbers respectively; a denotes a vector and A
 T  H  i is
and the noise matrix Γ
a matrix; A and A are the transposition and the Hermitian
 respectively; diag(A)
adjoint of A,  is the diagonalized version
 i ≡ [ni ni+1 · · · ni+q−1 ] ∈ CN ×q .
Γ (6)

of A by setting all of the off-diagonal elements in the square
matrix A  to be zero; if a = [an ]1≤n≤N is a vector, then the
As discussed in [6], the rank of Si is at most N , and each
circular-shift operation is defined as
ICI vector cα , 0 ≤ α ≤ N −1, has N entries. Therefore even if
shif tk (a) ≡ [a(n−k)N ]1≤n≤N , k ∈ Z, Si is fully-loaded with pilot symbols, only one ICI vector can
be estimated. Nevertheless, we have to maintain a reasonable
where the subscript (n − k)N of a(n−k)N denotes the integer throughput for the information data transmission together
(n − k) modulo N . with the ICI matrix estimation. To combat the difficulty of
estimating such an ICI matrix in Eq. (2), we will study the
II. I NTERCARRIER I NTERFERENCE IN OFDM characteristics of the ICI matrix and show how to approximate
the corresponding matrix inverse based on some usual channel
In the OFDM systems, the cyclic prefix is assumed to be conditions for OFDM systems in the subsequent section.
longer than the maximum channel delay spread [1], [4], [10].
The interblock interference from the previous OFDM symbols
is restricted within the cyclic prefix (guard interval) thereupon III. S TUDIES AND A PPROXIMATION OF ICI E QUALIZER
and the samples in the guard interval are discarded at the M ATRIX
receiver. Therefore we need to be concerned with the ICI  involves N 2
According to Eq. (2), the ICI matrix C
only [1], [4], [10]. Provided N subcarriers in each OFDM
parameters and N is usually fairly large. Even though the
block (symbol), according to [4], [11]–[13], the ith block of
training pilot sequences are carefully designed and the mean-
demodulated signals [ri,0 ri,1 · · · ri,N −1 ]T at the OFDM
square error ICI estimation schemes in [5]–[7] are applied, the
receiver, after the discrete Fourier transform, can be written
corresponding inverse or the ICI equalizer can still require
as
 si + ni , additional huge computational burden. Therefore, the direct
ri = C (1)
approach to design a semi-blind ICI equalizer has been pro-
where posed in our previous work [8], which can avoid the ICI matrix
⎡ ⎤ estimation and greatly reduce the computational complexity.
c0 The goal of an ICI equalizer can be described as follows:
⎢ c1 ⎥ i in Eq. (4),
⎢ ⎥ given q blocks of demodulated OFDM signals R
⎢ .. ⎥  N ×N
⎢ . ⎥ we carry out an ICI equalizer matrix D ∈ C such that

C ≡ [cα,β ] = ⎢ ⎥
⎢ cα ⎥
⎢ ⎥ ˆ
⎢ .. ⎥ Si = D
Ri , (7)
⎣ . ⎦
ˆ

cN −1
⎤ where Si corresponds to the matrix consisting of symbol
c0,0 c0,1 ··· c0,N −1 estimates. According to Eqs. (4) and (7), the perfect ICI
⎢ c1,0 c1,1 ··· c1,N −1 ⎥ equalization is achieved when
⎢ ⎥
= ⎢ .. .. .. .. ⎥, (2)
⎣ . . . . ⎦ ˆ
Si = Si + D
Γi , (8)
cN −1,0 cN −1,1 ··· cN −1,N −1
0 ≤ α, β ≤ N − 1. Note that cα,β stands for the ICI or
weighting coefficient associated with the interference from C
D  = IN , (9)
the βth subcarrier to the αth subcarrier; C  ∈ CN ×N
is the ICI matrix assumed to be stationary within a few where IN is an N × N identity matrix. According to
OFDM blocks in case of carrier frequency offset and time Eqs. (8), (9), the perfect ICI equalizer matrix should be
synchronization error and ri ≡ [ri,0 ri,1 · · · ri,N −1 ]T D =C −1 . However, such a matrix is not feasible in practice
∈ CN ×1 ; si ≡ [si,0 si,1 · · · si,N −1 ]T ∈ CN ×1 represents since (i) the precise estimate of C requires no throughput or
the ith block of transmitted information symbols; ni ≡ fully-loaded pilots in every OFDM symbol block as discussed
[ni,0 ni,1 · · · ni,N −1 ]T ∈ CN ×1 denotes the ith block of −1 cannot be real-time
in Section II; (ii) the computation of C
additive white noise. If we collect q blocks of demodulated on the subscriber’s platform for large subcarrier number N .
OFDM symbols to form a demodulated signal matrix, which In addition, even if the precise information of C is available,
is the existence of its inverse C  −1 is not always guaranteed.
Ri ≡ [ri ri+1 · · · ri+q−1 ] ∈ CN ×q . (3) Therefore, here we would like to provide the in-depth studies
on these crucial problems for the ICI matrix C  and the
Thus, we obtain corresponding inverse matrix (ICI equalizer matrix) in OFDM
i = C
R  Si + Γ
i, (4) systems.

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 08:54 from IEEE Xplore. Restrictions apply.
WU et al.: THEORETICAL STUDIES AND EFFICIENT ALGORITHM OF SEMI-BLIND ICI EQUALIZATION FOR OFDM 3793

A. ICI Equalizability rule given by Eq. (12) has to be investigated to check if the
 is invertible
First of all, we have to study if the ICI matrix C good approximability exists such that

or C is equalizable. Then, we have to make the approximation  2  2
 −1     −1  
on the ICI equalizer matrix D  to simplify the problem. Since C − D(Mt ) ≥ C − D(Mt ) , (13)
F F

C = [cα,β ], 0 ≤ α, β ≤ N − 1, according to Eq. (2),
 −1 depends on the characteristics of the for 1 ≤ Mt ≤ Mt ≤ N , and
the existence of C  2
ICI weighting coefficients cα,β . Usually, in the absence of  −1  
C − D(N ) = 0. (14)
subcarrier permutation ambiguity, according to [4], [14], [15], F
the ICI matrix is a strictly diagonally dominant matrix such Obviously, it is not always true that the approximation rule
that for C−1 in Eq. (12) will lead to the good approximability
N
−1 described by Eqs. (13) and (14). To specify when good ap-
|cα,α | > |cα,β |, ∀α = 0, 1, . . . , N − 1. (10) proximability occurs for C −1 , we need to define two particular
α=0 matrices here first.
Definition 1: A ridge matrix A  = [aα,β ]0≤α,β≤N −1 satisfies
α = β
the following conditions:
Remark 1: If C  is a strictly diagonally dominant matrix as Given any α = 0, 1, . . . , N − 1,
defined in Eq. (10), C −1 exists.
|aα,β1 | < |aα,β2 | ,
Proof: It arises from Theorem 6.1.10 (a) in [16]. 
 N −1 
According to Remark 1, in the normal channel condi- for ≥ (β1 − α)N > (β2 − α)N > 0 and β1 , β2 > α,
2
tions [4], [14], [15], the sum of the square-root ICI powers and
N
−1
|cα,β | is less than the subject frequency fading factor |aα,β1 | < |aα,β2 | ,
α=0  N −1 
for N − 2 ≤ (β1 − α)N < (β2 − α)N ≤ N − 1 and
α = β β1 , β2 < α, where
and are integer rounding down and
|cα,α |, α = 0, 1, . . . , N − 1. Thus the ICI matrix C  is up operators. 
equalizable or C −1 exists thereby. Definition 2: A sparse ridge matrix A  = [aα,β ]0≤α,β≤N −1
satisfies the following conditions: (a) A  is a ridge matrix
B. Approximability of the ICI Equalizer Matrix according to Definition 1. (b) For 0 ≤ α ≤ N − 1,

Even though the perfect ICI equalizer matrix D  = C−1 ⎪
⎨= 0, (β− α)N ≤ M2 , β ≥ α
exists in the normal conditions according to the analysis in aα,β (β − α)N ≥ N − M1 , β < α ,
Section III-A, it is very computationally complicated when N ⎪

≈ 0, otherwise
is large. Besides, in this paper, we are focused on the studies
of the blind ICI equalizer and do not need to calculate C −1 where M1 + M2 + 1 ≤ N is the matrix bandwidth. 
in this line of approach. Hence, we may relax the constraint
to approximate such an inverse, i. e., Now we can establish the approximability of the ICI
⎡ ⎤ equalizer matrix or the ICI matrix inverse C −1 based on the
d0
⎢ d ⎥ following remark.
⎢ 1 ⎥ Remark 2: If the ICI matrix C  is a strictly diagonally
⎢ .. ⎥
⎢ ⎥ dominant matrix as described in Eq. (10) and also a ridge
 = [dα,β ]0≤α,β≤N −1 = ⎢ ⎢ . ⎥  −1 ,
D ⎥≈C (11) −1 is a
matrix as described by Definition 1, then its inverse C
⎢ dα ⎥
⎢ ⎥ ridge matrix as well and the approximation rule for D =C  −1
⎢ .. ⎥
⎣ . ⎦ given by Eq. (12) leads to the good approximability as stated
dN −1 in Eqs. (13) and (14).
 Note: The first part of this remark is rather complicated to
where dα , 0 ≤ α ≤ N − 1, is the αth row vector of D. prove. However, we have tested enormously many random
According to [8], an intuitive approximation was made to matrices C  of different sizes via computer, which are both
 to be estimated, which
reduce the number of parameters in D strictly diagonally dominant and ridge matrices, and always
is achieved the corresponding inverse matrices C  −1 as ridge
⎧ 
⎪ matrices. Since C  is always a ridge matrix, it can easily be
−1
⎨= 0, for (β − α)N ≤ M2 β ≥ α
dα,β or (β − α)N ≥ N − M1 β < α , (12) justified that the good approximability described by Eqs. (13)

⎩ and (14) can result from the approximation rule given by
≈ 0, otherwise
Eq. (12).
where Mt ≡ M1 + M2 + 1 is the bandwidth of the approxi- According to Remark 2, if the ICI matrix C  in Eq. (2) is a
mated circulant-band matrix and M1 + M2 + 1 ≤ N . Express strictly diagonally dominant and ridge matrix, then C  −1 can
the approximate of the ICI inverse matrix D  = [dα,β ] in be well approximated as the sparse ridge matrix according to
terms of the matrix bandwidth Mt such that D  = D(M
 t ). Definition 2 using the rule of Eq. (12) and the total number
The asymptotical approximation error based on the Frobenius of nontrivial entries in the ICI equalizer matrix D is greatly
norm [16] between C  t ) using the approximation
 −1 and D(M reduced from N 2 to N (M1 +M2 +1). We may extend the same

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 08:54 from IEEE Xplore. Restrictions apply.
3794 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 10, OCTOBER 2008

strategy in [8] to further simply the ICI equalization problem. Since Si consists of independently identically distributed non-
The circular-shifted ICI equalizer matrix D   is defined as Gaussian information symbols, Eq. (21) can be formulated
⎡ ⎤ ⎡  ⎤ as a blind separation problem [8]. Our previous work using
shif tM1 (d0 ) d0 the JMMD involves a large computational complexity [8]. In
⎢ shif tM (d1 ) ⎥ ⎢ d ⎥
 ≡ ⎢
D ⎢ ..
1 ⎥ ⎢
⎥=⎢ ..
1 ⎥
⎥. (15)
this paper, we introduce a more efficient approach using the
⎣ . ⎦ ⎣ . ⎦ JADE algorithm instead. First, we segment the demodulated
signal matrix into L submatrices σ   and the corresponding
l R
shif tM (dN −1 ) d i
1 N −1 
partial correlation matrices ϑl ∈ C(M1 +M2 +1)×(M1 +M2 +1) ,
Thus 1 ≤ l ≤ L, are defined as
⎡ ⎤
H T
shif t−α (dα ) = ⎣dα 0
 ·
· · 0 ⎦, (16) ϑl ≡ σ " i R
l R " i σ
l , (22)
(N −M1 −M2 −1) zeros
l ∈ R(M1 +M2 +1)×N is defined
where σ as
where dα ∈ C1×(M1 +M2 +1) consists of the (M1 + M2 + 1) ⎡ ⎤
nontrivial values in dα . Only L representative ICI equal- 1 (l−1)N
ization vectors dα , for α = (l−1)N , 1 ≤ l ≤ L, (α = ⎢ L ⎥
L ⎢ 1 (l−1)N ⎥
(L−1)N ⎢ +1 ⎥
0, N , 2N
, · · · , ), need to be determined to reduce the σl ≡ ⎢ L
. ⎥, (23)
L L L ⎢ .. ⎥
computational complexity. L represents the number of vectors ⎣ ⎦
in the ICI equalization matrix, which are actually estimated 1 (l−1)N
+M1 +M2
L
directly from the received signals. All the other N − L vectors
in the ICI equalization matrix will be interpolated through such and (M1 + M2 + 1) is the individual truncated blind ICI
L vectors. Consequently, only (M1 + M2 + 1)L parameters equalization
 vector length in each segment. For example,
are to be estimated. Usually (M1 + M2 + 1)L N 2 . The rest 
0 = IM1 +M2 +1 0
σ  where IM1 +M2 +1 is the (M1 +M2 +1)×
of (N − L) ICI equalization vectors can be carried out using
(M1 + M2 + 1) identity matrix and  0 is the (M1 + M2 + 1) ×
the linear interpolation in [7] such that
(N −M1 −M2 −1) matrix containing all zeros. The whitening
α − α1 matrix is obtained by invoking the eigen-decomposition of
shif t−α (dα ) = shif t−α2 (dα2 )
(N/L) ϑl , which is the partial correlation matrix of the demodu-
(α2 − α) lated OFDM signal R i . Let λ1 , . . . , λM1 +M2 +1 denote the
+ shif t−α1 (dα1 ), (17)
(N/L) (M1 + M2 + 1) eigenvalues of ϑl , and v1 , . . . , vM1 +M2 +1
denote the corresponding eigenvectors. Thus, the whitening
 0 ≤α ≤ N − 1 and α modulo (N/L) = 0, where
for  l1 =  matrix W " is given by
α α
(N/L) is the closest rounding-down integer, l2 = (N/L)
l1 N l2 N # %H
is the closest rounding-up integer, α1 ≡ L , and α2 ≡ L . 1 1
"
W = √ v1 , · · · , $ vM1 +M2 +1
λ1 λM1 +M2 +1
IV. E FFICIENT S EMI -B LIND ICI E QUALIZATION U SING
JADE A LGORITHM ∈ C(M1 +M2 +1)×(M1 +M2 +1) , (24)
1×N
Define 1α ∈ R
 as a row vector (indicator vector) which and the whitened signal submatrices zl ∈ C(M1 +M2 +1)×q are
contains a ”1” as its (α + 1)th element and zero values as obtained as
the other elements, for α = 0, 1, . . . , N − 1. For example,
11 = [0 1 0 · · · 0]. Thus, an M1 -step permutation matrix is "σ  . (25)
zl = [zl (m, n)]1≤m≤M1 +M2 +1, i≤n≤i+q−1 =W l Ri
defined as

PM1 ≡ 1TN −M1 1TN −M1 +1 The fourth-order cumulants of the whitened signals zl (m, n)
 are given by [9]:
· · · 1TN −1 1T0 1T1 · · · 1TN −M1 −1 . (18)
cum[zl (i, n), zl∗ (j, n), zl (p, n), zl∗ (m, n)]
According to Eqs. (3)-(9), (15), (18), the ICI equalization is #i+q−1 %
achieved when 1 ∗ ∗
≡ zl (i, n)zl (j, n)zl (p, n)zl (m, n)
ˆ q n=i
Si = D
R  PM1 PT R
i = D    T 
M1 i = D PM1 Ri , (19) #i+q−1 % #i+q−1 %
1
∗ ∗
where D  = D PM1 and PM1 PT = IN . Thus, we can define − 2 zl (i, n)zl (j, n) zl (p, n)zl (m, n)
M1 q
the circular-shifted demodulated OFDM signal matrix as n=i
#i+q−1
n=i
% #i+q−1 %
1
R T 
i ≡ PM Ri . (20) − 2 zl (i, n)zl (p, n) ∗ ∗
zl (j, n)zl (m, n)
1
q n=i n=i
According to Eq. (19), we need to enforce the condition for #i+q−1 % #i+q−1 %
1
the representative ICI equalization vectors dα as − 2 zl (i, n)zl∗ (m, n) zl∗ (j, n)zl (p, n) .
! q
 = 1α , α = (l − 1)N , 1 ≤ l ≤ L. (21)
n=i n=i
shif t−M1 dα C (26)
L

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 08:54 from IEEE Xplore. Restrictions apply.
WU et al.: THEORETICAL STUDIES AND EFFICIENT ALGORITHM OF SEMI-BLIND ICI EQUALIZATION FOR OFDM 3795

Define the cumulant submatrix ζij ∈


Information Information
C(M1 +M2 +1)×(M1 +M2 +1) as Symbol
Pilot
Symbol
Pilot
Symbols Symbols
M1 +M Segment Segment
2 +1 M1 +M
2 +1
ζij ≡ cum[zl (i, n), zl∗ (j, n),
p=1 m=1

zl (p, n), zl∗ (m, n)] 1Tp−11m−1 , (27) OFDM blocks OFDM blocks
i to i+q-1 i+q+N to
where 1p−1 and 1m−1 are the indicator vectors of size (M1 + i+2q+N-1
M2 + 1) now. According to Eq. (27), we can form a cumulant
2 2
matrix υ  ∈ C(M1 +M2 +1) ×(M1 +M2 +1) as [9]:
OFDM blocks OFDM blocks
⎡ ⎤
ζ11 ··· ζ1,(M1 +M2 +1) i+q to i+q+N-1 i+2q+N to
i+2q+2N-1
⎢ ζ21 ··· ζ2,(M1 +M2 +1) ⎥
⎢ ⎥
≡⎢
υ . . ⎥.
⎣ .. ··· .
. ⎦
Fig. 1. The proposed OFDM transmission scheme.
ζ(M1 +M2 +1),1 · · · ζ(M1 +M2 +1),(M1 +M2 +1)
(28)
Then the (M1 + M2 + 1) most significant eigenpairs Step 1) Arrange the symbols at the OFDM transmitter accord-
{λr , εr , 1 ≤ r ≤ M1 + M2 + 1} of υ  are carried out, ing to Figure 1.
where λr and  εr are the M1 + M2 + 1 largest eigenval- Step 2) At the OFDM receiver, choose the segment number
ues and the corresponding eigenvectors. The set of matrices L and the blind equalizer length M1 + M2 + 1. Then circular
 r , 1 ≤ r ≤ M1 + M2 + 1} result from the vectors εr using
{Σ shift q blocks of demodulated OFDM symbols as R   = P T R 
i M1 i
a classic ”stacking-unstacking” device [9]. A unitary matrix according to Eq. (20).
 is determined to jointly diagonalize the set of matrices
U Step 3) Segment the circular-shifted demodulated signal ma-
 r , 1 ≤ r ≤ M1 + M2 + 1}. The ultimate blind separation
{Σ trix R into L submatrices and form the partial correlation
i
using the JADE algorithm is formulated as matrices ϑl , 1 ≤ l ≤ L, according to Eq. (22).
JADE = U ".
HW Step 4) Set l=1. Calculate the whitening matrix W " according
B (29)
to Eq. (24).
In the JADE algorithm, the joint diagonalization is imple- Step 5) Construct the set of cumulant eigen-matrices {Σ  r, 1 ≤
mented by extending the single-matrix Givens rotations for r ≤ M1 + M2 + 1} according to Eqs. (25)-(28).
several matrices. It achieves the optimization of the joint di- Step 6) Determine the blind separation matrix B opt according
agonalization criterion via successive Givens rotations, which to Eqs. (30) and (31).
is given by Step 7) Follow Steps 7-13 of [8] to determine dα , for
! M1 +M
2 +1 && &2 0 ≤ α ≤ N − 1.
ΨJADE−DIAG U  ≡ HΣ
& diag(U  &&
 r U)
' r=1 !( , (30)
opt = arg max  ΨJADE−DIAG U
U  V. S IMULATION
U
Through the simulations, we compare the system perfor-
where Uopt is the unitary matrix for the optimal JADE blind mances of the new OFDM scheme using our proposed semi-
separator as given by Eq. (29) such that blind ICI equalizer and the plain OFDM system. The OFDM
opt = U
H W ". systems under test comply with the IEEE 802.11 standard as
B opt (31)
specified in [17]. The information symbols are assumed to
However, as formulated in Eq. (21), the equalization vector be modulated as QPSK or 16QAM. The subcarrier number
for each subcarrier α is unique while the blind separation is chosen to be N =64 and N =256. The COST 207 channel
matrix Bopt ∈ C(M1 +M2 +1)×(M1 +M2 +1) in Eq. (31) consists model for rural areas in [18] are chosen to benchmark all of
of (M1 + M2 + 1) row vectors. Besides, the phase and the tested OFDM systems with normalized local frequency
permutation ambiguities are two major unsolved problems offset Δfo T = 0.2 [4]. The number of segments in our semi-
remaining in Bopt . As addressed in [8], a pilot-based iterative blind equalization scheme is L=4. For our proposed algorithm,
phase/permutation identification technique can be employed to there is a trade-off between the computational complexity and
solve the aforementioned problems. the system performance. If L is large, the system perfor-
Similar to Figure 1 of [8], we can pad N blocks of mance would improve at the expense of higher computational
indicator vectors in addition to every q blocks of OFDM complexity. As for the M1 , M2 , M values, they depend on
information symbols such that si+q = 1T0 , si+q+1 = 1T1 , . . ., the channel conditions. For example, in the fast time-varying
si+q+N −1 = 1TN −1 . Such pilot symbols would not sacrifice multiple fading channels, as Doppler frequency increases,
much throughput but can be used for determining the phase more and more adjacent subcarriers are involved to generate
and permutation ambiguities incurred by the blind separation ICI. In this case, we should pick large M1 , M2 , M values
q
scheme [8]. As a result, the ultimate throughput is q+N . The accordingly. To be consistent with the symbolic notation in [8],
new efficient semi-blind ICI equalization method using the we set M1 = M2 = M − 1 in the simulations of this paper.
JADE algorithm can be summarized as follows: We re-depict the proposed transmission scheme from [8] in
Figure 1. Two system performance measures, namely symbol

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 08:54 from IEEE Xplore. Restrictions apply.
3796 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 10, OCTOBER 2008

0 0
10 10

−1
10

−2
Symbol Error Rate

Symbol Error Rate


10

−3 −1
10 10

−4
10

−5
10 JADE, M=2, q=250 JADE, M=2, q=250
JADE, M=2, q=500 JADE, M=2, q=500
Raw Data RawData
−6 −2
10 10
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Signal−to−noise ratio η in dB Signal−to−noise ratio η in dB

Fig. 2. Symbol error rate versus signal-to-noise ratio resulting from the Fig. 4. Symbol error rate versus signal-to-noise ratio resulting from the
QPSK-OFDM systems (Raw Data: plain OFDM, JADE: our proposed semi- 16QAM-OFDM systems (Raw Data: plain OFDM, JMMD: our proposed
blind ICI equalizer) with 64 subcarriers. semi-blind ICI equalizer) with 64 subcarriers.

0
10 0
10

−1
10
Symbol Error Rate

Symbol Error Rate


−2
10 10
−1

−3
10
JADE, M=2, q=250
JADE, M=2, q=500 JADE, M=2
RawData Raw Data
−4
10 10
−2
0 5 10 15 20 25 30 35 40 200 400 600 800 1000 1200 1400 1600 1800
Signal−to−noise ratio η in dB Number of blocks q

Fig. 3. Symbol error rate versus signal-to-noise ratio resulting from the Fig. 5. Symbol error rate versus the number of blocks q resulting from
QPSK-OFDM systems (Raw Data: plain OFDM, JADE: our proposed semi- the 16QAM-OFDM systems (Raw Data: plain OFDM, JMMD: our proposed
blind ICI equalizer) with 256 subcarriers. semi-blind ICI equalizer) with 64 subcarriers when the signal-to-noise ratio
is 20 dB.

error rate and computational complexity, averaged over one


hundred Monte Carlo trials are provided in the following. Figures 5 and 6, it is noted that the symbol error rates will
decrease when the number of OFDM blocks collected for cu-
mulant estimates increases. Similar performance to Figures 2-6
A. Symbol Error Rate
can be achieved using our previous semi-blind ICI equalization
The symbol estimates can be obtained when the ICI equal- scheme in [8].
ization vectors dα are determined according to Step 7 in
Section IV. The truncation length is defaulted as M =2. Then
the corresponding probability of symbol error or symbol error B. Computational Complexity Comparison
rate can be calculated. We depict the symbol error rates
versus signal-to-noise ratios η for the aforementioned OFDM As stated in Section V-A, the symbol error performances
systems in Figures 2-6. According to Figures 2-4, the symbol of the semi-blind ICI equalization using the JADE algorithm
error rates are greatly improved by our proposed semi-blind in this paper and that using the JMMD algorithm in [8] are
ICI equalizer for QPSK-OFDM; nevertheless the improve- very similar. The computational complexity comparison is
ments for the 16QAM-OFDM are not so significant. This crucial thereby. Based on the same set of system parameters as
phenomenon results from the cumulant estimation variances Section V-A, we depict the average numbers of exact complex
for the 16QAM symbols are much larger than those for the multiplications per Monte Carlo trial for the JADE and JMMD
QPSK symbols. If we collect more data blocks (q is larger) algorithms involved in the semi-blind ICI equalization in
for ICI equalization, we will reduce such estimation variances. Figures 7-10. For QPSK- and 16QAM-OFDM systems, the
We provide Figures 5 and 6 to illustrate the relationship JADE algorithm is generally much more efficient than the
between the symbol error rates and the q values for 16-QAM JMMD algorithm especially for low truncation length M or
OFDM systems when the signal-to-noise ratio is 20 dB. From high signal-to-noise ratios according to Figures 7-10.

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 08:54 from IEEE Xplore. Restrictions apply.
WU et al.: THEORETICAL STUDIES AND EFFICIENT ALGORITHM OF SEMI-BLIND ICI EQUALIZATION FOR OFDM 3797

0 6
10 10

Number of Complex−valued Multiplication


5
10
Symbol Error Rate

4
10

3
10 JMMD, M=2
JADE, M=2
JADE, M=2 JMMD, M=4
Raw Data JADE, M=4
−1
10 0 5 10 15 20 25 30 35 40
0 500 1000 1500 2000
Number of blocks q Signal−to−noise ratio η in dB

Fig. 6. Symbol error rate versus the number of blocks q resulting from Fig. 8. Computational complexities in terms of average complex multiplica-
the 16QAM-OFDM systems (Raw Data: plain OFDM, JMMD: our proposed tions per Monte Carlo trial versus signal-to-noise ratio for JMMD and JADE
semi-blind ICI equalizer) with 256 subcarriers when the signal-to-noise ratio algorithms in semi-blind ICI equalization (QPSK-OFDM systems with 256
is 20 dB. subcarriers).

6
6 10

Number of Complex−valued Multiplication


10
Number of Complex−valued Multiplication

5
10
5 10

4
10
4 10

3
10
3 10 JMMD, M=2
JMMD, M=2
JADE, M=2 JADE, M=2
JMMD, M=4 JMMD, M=4
JADE, M=4 JADE, M=4

0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Signal−to−noise ratio η in dB Signal−to−noise ratio η in dB

Fig. 7. Computational complexities in terms of average complex multiplica- Fig. 9. Computational complexities in terms of average complex multiplica-
tions per Monte Carlo trial versus signal-to-noise ratio for JMMD and JADE tions per Monte Carlo trial versus signal-to-noise ratio for JMMD and JADE
algorithms in semi-blind ICI equalization (QPSK-OFDM systems with 64 algorithms in semi-blind ICI equalization (16QAM-OFDM systems with 64
subcarriers). subcarriers).

VI. C ONCLUSION R EFERENCES


[1] R. V. Nee and R. Prasad, Eds., OFDM for Wireless Multimedia Com-
In this paper, we provide the theoretical studies on the munications. Artech House Publisher, 2000.
existence condition and the approximability of the ICI equal- [2] M. Speth, S. Fechtel, G. Fock, and H. Meyr, “Optimum receiver design
izer matrix for the OFDM systems. We conclude that the ICI for wireless broad-band systems using OFDM: part I,” IEEE Trans.
Communications, vol. 47, no. 11, pp. 1668–1677, Nov. 1999.
matrix inverse or the ICI equalizer matrix exists when the [3] ——, “Optimum receiver design for OFDM-based broadband transmis-
ICI matrix is diagonally dominant. The ICI equalizer matrix sion: part II-a case study,” IEEE Trans. Commun., vol. 49, no. 4, pp.
can be well approximated as a sparse ridge matrix if the ICI 571–578, Apr. 2001.
[4] H.-C. Wu, “Analysis and characterization of intercarrier and interblock
matrix is a diagonally dominant and ridge matrix. Based on interferences for wireless mobile OFDM systems,” IEEE Trans. Broad-
the sparse ridge matrix approximation, we may design an casting, vol. 52, no. 2, pp. 203–210, June 2006.
efficient semi-blind ICI equalization method using the JADE [5] H.-C. Wu and Y. Wu, “A new ICI matrix estimation technique using
padded m-sequences for wireless OFDM systems,” in Proc. IEEE Global
algorithm. Through numerous Monte Carlo simulations, we Telecommunications Conference, vol. 5, Nov.-Dec. 2005.
discover that our new efficient semi-blind ICI equalization [6] ——, “A new ICI matrices estimation scheme using hadamard sequence
technique can reduce the symbol error rates for the QPSK- and for OFDM systems,” IEEE Trans. Broadcasting, vol. 51, no. 3, pp. 305–
314, Sept. 2005.
the QAM-OFDM systems. The symbol error rate performance
[7] ——, “Distributive pilot arrangement based on modified m-sequences
margin of this semi-blind ICI equalization scheme for QPSK- for OFDM intercarrier interference estimation,” IEEE Trans. Wireless
OFDM systems can even reach several orders-of-magnitude. Commun., vol. 6, no. 5, pp. 1605–1609, May 2007.
Moreover, the JADE algorithm leads to much less computa- [8] H.-C. Wu, X. Huang, and D. Xu, “Novel semi-blind ICI equalization
algorithm for wireless OFDM systems,” IEEE Trans. Broadcasting,
tional complexity than the JMMD algorithm especially for low vol. 52, no. 2, pp. 211–218, June 2006.
truncation length M or high signal-to-noise ratios. [9] J. F. Cardoso and A. Souloumiac, “Blind beamforming for non-Gaussian

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 08:54 from IEEE Xplore. Restrictions apply.
3798 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 10, OCTOBER 2008

6
10
Number of Complex−valued Multiplication
Xiaozhou Huang received the B.S. and M.S. de-
grees in wireless communication from Northern
Jiaotong University, Beijing, China, in 1993 and
5 1996, respectively, and the M.S. degree in physics
10 science from the Physical Department and the M.S.
degree in digital signal processing from the Depart-
ment of Electrical and Computer Engineering, both
4
from the University of Massachusetts Dartmouth,
10 North Dartmouth, in 1999 and 2002, respectively.
She received the Ph.D. degree in Electrical and
Computer Engineering, Louisiana State University
3
in 2006.
10 JMMD, M=2 Dr. Huang worked for Beijing University of Post and Telecommunication-
JADE, M=2 NORTEL Research Laboratory from 1996 to 1998. Currently, she works
JMMD, M=4 as a software application engineer for Texas Instruments, Germantown,
JADE, M=4 Maryland. Her research interests are in the areas of digital communication
0 5 10 15 20 25 30 35 40 and communication theory.
Signal−to−noise ratio η in dB
Yiyan Wu (Fellow, IEEE) received the B. Eng
Fig. 10. Computational complexities in terms of average complex multiplica- degree from the Beijing University of Posts and
tions per Monte Carlo trial versus signal-to-noise ratio for JMMD and JADE Telecommunications, and M.Eng. and Ph.D. degrees
algorithms in semi-blind ICI equalization (16QAM-OFDM systems with 256 in electrical engineering from Carleton University,
subcarriers). Ottawa, Canada, in 1986 and 1990, respectively.
After graduation, he worked at Telesat Canada as
a senior satellite communication systems Engineer.
In 1992, He joined Communications Research Cen-
signals,” IEE Proceedings F, Radar and Signal Processing, vol. 140, tre Canada (CRC) and is now a Principle Research
no. 6, pp. 362–370, Dec. 1993. Scientist. Dr. Wu’s research interests include broad-
[10] L. Hanzo, M. Munster, B. J. Choi, and T. Keller, OFDM and MC-CDMA band multimedia communications, digital broadcast-
for Broadband Multi-user Communications, WLANs and Broadcasting. ing, and communication systems engineering. He is an IEEE Fellow, an
John Wiley and Sons, 2003. adjunct professor of Carleton University, Ottawa, Canada; Shanghai Jiaotung
[11] J. Armstrong, “Analysis of new and existing methods of reducing University; and Beijing University of Posts and Telecommunications. Dr. Wu
intercarrier interference due to carrier frequency offset in OFDM,” IEEE is a member of the IEEE Broadcast Technology Society Administrative Com-
Trans. Commun., vol. 47, no. 3, pp. 365–369, Mar. 1999. mittee, and a member of the ATSC Board of Directors, representing IEEE.
[12] H.-C. Wu and X. Huang, “Joint phase/amplitude estimation and symbol He is the Editor-in-Chief of the IEEE T RANSACTIONS ON B ROADCASTING.
detection for wireless ICI self-cancellation coded OFDM systems,” IEEE Dr. Wu has more than 200 publications and received many technical awards
Trans. Broadcasting, vol. 50, no. 1, pp. 49–55, Mar. 2004. for his contribution to the research and development of digital broadcasting
[13] S. G. Kang, Y. M. Ha, and E. K. Joo, “A comparative investigation on and broadband multimedia communications.
channel estimation algorithms for OFDM in mobile communications,”
IEEE Trans. Broadcasting, vol. 49, no. 2, pp. 142–149, June 2003. Xianbin Wang (S’98-M’99-SM’06) received his
[14] X. Huang and H.-C. Wu, “Robust and efficient intercarrier interference Ph.D. degree in electrical and computer engineering
mitigation for OFDM systems in time-varying fading channels,” IEEE from National University of Singapore in 2001.
Trans. Vehicular Technology, vol. 56, no. 5, pp. 2517–2528, Sept. 2007. He was with Institute for Infocomm Research,
[15] ——, “Intercarrier interference analysis for wireless OFDM in mobile Singapore (formerly known as Centre for Wireless
channels,” in Proceedings IEEE Wireless Communications and Network- Communications), as a Senior R and D engineer
ing Conference (WCNC), vol. 4, Apr. 2006, pp. 1848–1853. in 2000. From Dec. 2000 to July 2002, he was a
[16] R. A. Horn and C. R. Johnson, Matrix Analysis. Cambridge University system designer at STMicroelectronics, where he
Press, 1990. was responsible for system design for DSL and
[17] Part 11: Wireless LAN Medium Access Control (MAC) and Physical Gigabit Ethernet chipsets. From July 2002 to Dec.
Layer (PHY) Specifications: Higher-Speed Physical Layer Extension in 2007, he was with Communications Research Centre
the 2.4 GHz Band, IEEE Std. IEEE 802.11b, 1999. Canada as Research Scientist. He joined The University of Western Ontario
[18] A. F. Molisch, Wideband Wireless Communications. New York: in Jan. 2008 as an Associate Professor and Canada Research Chair in
Prentice-Hall, 2000. Wireless Communications. His current research interests include broadband
wireless system, communication theory, and digital signal processing. Dr.
Hsiao-Chun Wu (M’00 SM’05) received a B. S. Wang has over 90 publications on various communication system design
E. E. degree from National Cheng Kung University, issues, including around 60 IEEE Journal and Conference papers, 10 granted
Taiwan, in 1990, and the M. S. and Ph. D. degrees in and pending patents, and several standard contributions.
electrical and computer engineering from University Dr. Wang is a Senior Member of IEEE. He is the recipient of the IEEE Scott
of Florida, Gainesville, in 1993 and 1999 respec- Helt Memorial Award for the Best Paper published in IEEE T RANSACTIONS
tively. ON B ROADCASTING in 2004. He currently serves as an Associate Editor for
From March 1999 to January 2001, he had worked IEEE T RANSACTIONS ON W IRELESS C OMMUNICATIONS , IEEE T RANS -
for Motorola Personal Communications Sector Re- ACTIONS ON V EHICULAR T ECHNOLOGY and IEEE T RANSACTIONS ON
search Labs as a Senior Electrical Engineer. Since B ROADCASTING. Dr. Wang was involved in a number of IEEE conferences
January 2001, he has joined the faculty in De- including GLOBECOM, WCNC, VTC, and ICME, on different roles such
partment of Electrical and Computer Engineering, as TPC and session chair. He was also a track co-chair for IEEE VTC 2008
Louisiana State University, Baton Rouge. Dr. Wu has published more than Spring in Singapore.
100 peer-refereed technical journal and conference articles in electrical and
computer engineering. His research interests include the areas of wireless
communications and signal processing. Dr. Wu is an IEEE Senior Member
and currently serves as an Associate Editor for IEEE T RANSACTIONS ON
B ROADCASTING, IEEE T RANSACTIONS ON V EHICULAR T ECHNOLOGY,
I NTERNATIONAL J OURNAL OF C OMPUTERS AND E LECTRICAL E NGINEER -
ING , and J OURNAL OF THE F RANKLIN I NSTITUTE. He also serves for
numerous textbooks, IEEE conferences and journals as the technical com-
mittee, symposium and track chair or the reviewer in signal processing,
communications, circuits and computers.

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 08:54 from IEEE Xplore. Restrictions apply.

You might also like