This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts
for publication in the WCNC 2007 proceedings.
A New Carrier Frequency Offset Estimation
Using CP-ICA Scheme in OFDM Systems
Jong-Deuk Kim
Dept. of Electronics Engineering, University of InCheon
177 Dohwa-Dong, Nam-Gu,
InCheon 402-749, Korea
jdkim@incheon.ac.kr
Abstract— a well-known problem of OFDM is its sensitivity to
carrier frequency offset which leads to inter-carrier interference
(ICI) in the OFDM symbol. This ICI causes severe degradation of
the BER performance of the OFDM receiver. In this paper, we
propose a new ICI cancellation algorithm which estimates
frequency offset at the time-domain by using Cyclic Prefix -
Independent Component Analysis (CP-ICA) method to the
received subcarriers phase rotation. This algorithm is based on a
statistical blind estimation method, which mainly utilizes the
eigenvalue decomposition (EVD), rotating phasor and the 4th
cumulants.
Keywords-OFDM;Frequency Offset; ICI; CP-ICA
I. INTRODUCTION
O RTHOGNAL Frequency Division Multiplexing(OFDM)
is emerging as the preferred modulation scheme in
modern high data rate wireless communication systems. A
well known problem of OFDM, however, is its sensitivity to
frequency offset between the transmitted and received signal,
which may be caused by Doppler shift in the channel, or by
the difference between the transmitter and receiver local
oscillator frequencies. The carrier frequency offset causes loss
of orthogonality between subcarriers, and the signals
transmitted on each carrier are not independent of each other,
thus leading to the ICI [1, 2]. Therefore, a synchronization
problem of carrier frequency is very crucial for possible
application of OFDM in high data rate transmission wireless
communications. Many researchers have proposed various
methods to combat the ICI in OFDM systems. The existing
approaches that have been developed to reduce ICI can be
categorized as using pilot method, blind scheme using Cyclic
Prefix and the ICI self-cancellation (SC) scheme [3]. Most
existing frequency offset estimation techniques in OFDM
systems require pilot symbols [4, 5, 6, 7]. However, the use of
pilot symbols reduces the bandwidth efficiency because the
pilots occupy some part of the useful bandwidth. Therefore, a
blind approach scheme for frequency offset estimation was
introduced [8]. However, the algorithm relies on virtual
subcarriers, which also lowers the bandwidth efficiency. For
more realistic synchronization, some methods have been
proposed for blind estimation of timing offset as well as
1525-3511/07/$25.00 ©2007 IEEE
Youn-Shik Byun
Dept. of Electronics Engineering, University of InCheon
177 Dohwa-Dong, Nam-Gu,
InCheon 402-749, Korea
byun@incheon.ac.kr
frequency offset [9, 10, 11, 12]. These algorithms mainly
utilize the cyclic prefix, and some methods [11] require the
OFDM signal to be cyclostationary. In addition, statistical
separation approaches have also been explored to do a blind
source separation [13, 14, 15].
In this paper, we propose a blind method to estimate
frequency offset in OFDM systems. This method is based on
Cyclic Prefix– Independent Component Analysis (CP-ICA)
approach that statistically estimates the frequency offset and
corrects the offset using the estimated value at the receiver. The
paper is organized as follows: Section Ⅱ describes the
standard system model and the problem of ICI in OFDM.
Section Ⅲ describes the independent component analysis
scheme and the proposed CP-ICA methods in this paper.
Section Ⅳ provides a simulation result that tests the BER
performance and also compares it with the SC, ML and CP-
ML[3, 5, 10]. Finally, concluding remarks are presented in
section Ⅴ.
II. SYSTEM MODEL AND INTER-CARRIER INTERFERENCE
PROBLEM IN OFDM SYSTEMS
In OFDM systems, the input bit streams are multiplexed
into N symbol streams, each with symbol period Ts, and each
symbol stream is used to modulate parallel, synchronous
subcarriers [16]. The subcarriers are spaced by 1/NTs in
frequency, and thus they are orthogonal over the interval (0,
Ts). A typical discrete-time baseband OFDM transceiver
system is shown in Figure 1. First, a serial-to-parallel (S/P)
converter groups the stream of input bits from the source
encoder into groups of log2Ð bits, where Ð is the alphabet of
size of the digital modulation scheme employed on each
subcarrier. A total of N such symbols, Xl (m), are created.
Then, the N symbols are mapped to bins of an inverse fast
Fourier transform (IFFT). These IFFT bins correspond to the
orthogonal subcarriers in the OFDM symbols.
1 N −1 2π nm
(1)
∑
i
xl (n ) = X (m )e N
N m =0
l
where Xl (m) is the baseband l-symbols on m-th subcarrier.
1031
 This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the WCNC 2007 proceedings.

N −1
Yl ( k ) = ∑ y l ( n ) e − i 2 π kn / N , k = 0,1, ...., N − 1
n=0

1 N −1 N −1 2π n ( m +ε − k ) (4)
∑X (m )∑ e
i
= l
N
+ Wl ( k )
N m =0 n=0
N −1

Figure 1. Baseband OFDM system = X l ( k ) I l (0) + ∑

m = 0, m ≠ k
X l ( m )I l ( m − k ) + W l ( k )

At the receiver, the signal is converted back to a discrete N The ICI components are the interfering signals transmitted
point sequence yl(n), corresponding to each subcarrier. This on subcarriers other than the k-th subcarrier. The complex
discrete signal is demodulated using the N point fast Fourier coefficients are given by
transform (FFT) operation at the receiver. The demodulated
symbol stream is given by 1 s in (π ( m + ε − k )) 1
iπ (1 − )( m + ε − k ) (5)
Il (m − k ) = e N
N s in (π ( m + ε − k ) / N )
N −1 2 π kn
−i
(2)
Yl ( k ) = ∑
n=0
yl (n )e N
+ W l (k )
The first term in (4) denotes the useful component. The
transmitted symbol Xl(k) is rotated and attenuated. The second
where Wl (k) corresponds to the FFT of the samples of wl term is the complex coefficients for the ICI components in the
(n), which is the Additive White Gaussian Noise (AWGN) received signal. For sufficiently large N, ICI can be modeled as
introduced in the channel. Multipath distortion can also cause a Gaussian random process with zero mean variance σ by the I
2

Inter-symbol Interference (ISI) when adjacent symbols central limit theorem. Therefore, we can estimate the frequency
overlap with each other. This is prevented in OFDM system offsets from the estimate of phase rotation.
by the insertion of a cyclic prefix between successive OFDM
symbols. This CP is discarded at the receiver to cancel out ISI. III. FREQUENCY OFFSET ESTIMATION USING CP-ICA
It is due to the robustness of OFDM to ISI and multipath To rigorously define Independent Component Analysis
distortion that it has been considered for various wireless (ICA), we can use a statistical “latent variables” model [13-
application and standards. The main disadvantage of OFDM, 15]. We have an observable random vector y ( n ) .
however, is its susceptibility to small differences in frequency
at the transmitter and receiver, normally referred to as I −1 I −1
frequency offset. The frequency offset can be caused by y(n) = ∑∑aij xjn (n) + wjn (n) , n = 0,1,...., N −1 (6)
Doppler shift due to relative motion between the transmitter i=0 j =0

and receiver, or by the difference between the transmitter and

receiver local oscillator frequencies. In this paper, the carrier
frequency offset is modeled as a multiplicative factor
introduced in the channel, as shown in Figure 2.
y(n) is assumed to be linear mixtures of some mutually
statistically independent variables aij. Mathematically, Y is
generated according to Y=AX. The matrix A is mixing matrix.
In order to estimate unmixing matrix ζ = A-1., whitening
processing is needed. One popular method for whitening is to
use the eigenvalue decomposition of covariance matrix.

[ E v , D v ] = eig ( Ε {YY T }) = E v D v E v T (7)

Figure 2. Carrier Frequency Offset Model where Ev is the orthogonal matrix of eigenvectors of
eig ( E{YY T }) and Dv is the diagonal matrix of its eigenvalues.
The received signal is given by
The whiteness or uncorrelatedness is a necessary condition for
i
2 π nε the stronger statistically independent condition [13 - 14].
y l ( n ) = xl ( n ) e N
+ wl ( n ) (3)
O = Ev Dv−(1/2) EvT , Y = OY = O( AX +W ) = RX
 +W (8)
where ε(= foffset / ∆f) is the normalized frequency offset,
foffset is the frequency difference between the transmitted and where O is whitening matrix and after whitening, new
receiver frequencies and ∆f is the subcarrier spacing. wl (n) is  is white. i.e., its components are uncorrelated and
vector Y
the AWGN introduced in the channel. The effect of this their variance equal unity. The whitening transforms the
frequency offset on the received symbol stream can be mixing matrix A into a new rotating matrix R (= O A ) . We
understood by considering the received symbol Yl(k) on the k-
just need to find an orthogonal transformation U to make
th subcarrier.
components of X = U Y .

1032
 This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the WCNC 2007 proceedings.

UY = U ( R X + W ) = U (OAX + OW ) (9) where i is 1st and j is 2nd index of max eigenvalues.
= UO ( AX ) + UO (W ) = UO ( AX + W ) = X + Wˆ
Step 5. Buffering as respects important eigenvector
Therefore, ζ can be constructed as ζ = UO = A . -1
 m1 = m1[ M × K ] = E1 (i ) (15)
In order to mitigate the ICI in OFDM systems, we will have  , B = [ m1 m2 ], M = K = 2
to estimate the carrier frequency offset (CFO). The proposed  m2 = m2[ M × K ] = E1 ( j )
CFO estimation using CP-ICA performs as follows: ψ i (i=1,2) where B is buffer with respect to max eigenvectors matrix.
are linear mixtures of statistically independent each other by
the frequency offset and AWGN. Step 6. Rotating to the Step 5

 ψ = −1 y (n)   ψ : cyclic prefix parts  m ax ( E1 ( i )) − m in ( E 1 ( i )) m ax ( E1 ( j )) − m in ( E 1 ( j )) 

 1 n∑
1

g =  
l  ψ : useful data parts 
m 1 (1, K ) m 2 (1, K )

(16)
(10)  
ψ[2×( N / 4)] =  ,  m 1 ( M ,1) m 2 ( M ,1)
=− L 2

ψ = yl (n)  
L : CP length(= N / 4)
N −1
 co s θ sin θ   c s 


2 ∑   yl (n) : received l th symbols
R= = 
 − sin θ co s θ   − s c 

n=N −L
 1 0  T T
[ E 2 , D 2 ] = eig  0  , G = E2 ( g ⋅ g )E2
  R  
Step 1. Polar Decomposition
where g is real and imaginary axis separation to the max
 ψψ T  eigenvectors(=B), R is basis rotating matrix, G is rotated
[ E , D ] = e ig ( ) (11)
 l e n g t h (ψ )  matrix of independent components.
1
T −1
P = ED 2
E , O = P
Step 7. Extraction max eigenvector from the G matrix and
obtains basis rotating matrix (=R).
where E is eigenvectors, D is eigenvalues, P is whitener
matrix, and O is pseudo-inverse of P.
[ E3 , D3 ] = eig (G ) (17)
Step 2. Centering and Whitening k = arg max( diag { D 3}) , A = {1, 2,..., n}
k ∈A

ψ r = ψ − Ε{ψ }, ϒ = ( O ⋅ψ r ) (12)
θ = E3 ( k ), c =
1 θ (1)
+ , s=
(
θ (2) − − 1 θ (3) ( ) )
2 2 2⋅c
where ψ r is centering, ϒ is statistically independent where k is index of max eigenvalues and Ө is max
⋅ denotes the expectation operator.
matrix and Ε{} eigenvector.

Step 3. 4th cumulants of independent components [15] Step 8. Update basis rotating matrix.

c − con j ( s )  1 0 (18)
ℜ yy = Ε{( ϒϒ T )}, ℑ yy = Ε{( ϒϒ H )} Ru =   , Ub = 
1 
, U = U b ⋅ Ru
s c  0
C = ℜ yy − ℜ yy [τ 0 , τ 1 ]ℜ yy [τ 2 , τ 3 ] (13)
where Ru is updated basis rotating matrix, Ub is basis
− ℜ yy [τ 0 , τ 2 ]ℜ yy [τ 1 , τ 3 ]
Unitary, U is updated new rotating unitary matrix.
− ℑ yy [τ 0 , τ 3 ]conj ( ℑ yy [τ 1 , τ 2 ])
Step 9. Comparison for the optimal rotated matrix
where T is transpose, H is Hermitian, C is the 4th
cumulants and τ i is delay factor. If | s | > Th [→ Go Step 5. updating buffer]

Step 4. Extraction important eigenvector from step 3 B = Ru T ⋅ B (19)

[ E1 , D1 ] = e ig (C ) Else if | s | < Th [→ Obtain the optimal rotated matrix]

(14)
i = a r g m a x ( d i a g { D 1 } ) , A = {1, 2 , . . . , n }
i∈ A ζ [ M × K ] = (U ⋅ O ) (20)
j = a rg m a x ( d i a g { D 1 } )
j∈ A , j ≠ i
End

1033
 This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the WCNC 2007 proceedings.

where Th is decision threshold (=1 / No. of subcarriers) and

ζ is the optimal rotated square matrix.

Step 10. Frequency Offset Estimation

ζ M :(M )th rowvector

ε = tan −1 ( ∑ζ M ⋅ conj(ζ M −1 )) / 2π ,  (21)
ζ M −1 : ( M − 1) rowvector
th

if (ε < −10−6 ), ε = 1+ ε, end

IV. SIMULATION RESULTS Figure 4. BER Performance with ICI Cancellation, ε = 0.032
In order to compare the three different ICI cancellation
schemes, we will assume that timing synchronization is
perfect for all methods at the receiver. BER curves were used
to evaluate the performance of each scheme. The OFDM
transceiver system was implemented as specified by Figure 1.
Frequency offset was introduced as the phase rotation as given
by using (3). In this simulation, we choose 16-QAM and 64-
QAM of modulation schemes and the AWGN channel is
assumed. The OFDM system parameters chosen are as
follows: the number of subcarriers N = 256, the number of CP
length L= N/4, and normalized frequency offsets ε = {0.032,
0.49, 0.65, 0.98}. The simulation is carried out over 1000 Figure 5. BER Performance with ICI Cancellation, ε = 0.49
OFDM symbols for each SNR value. The BER curve for
Figure 6 and Figure 7 show comparisons of the performance
standard OFDM without ICI cancellation shows in Figure 3.
of the SC, ML, CP-ML and CP-ICA schemes for different
alphabet sizes and different frequency offsets in 0.5< ε <1.0.
We can observe the BER error floor. The SC, ML and CP-ML
methods are very hardly depredated BER. Since these
techniques do not completely cancel the ICI from adjacent
subcarriers and the effect of this residual ICI increases for
larger alphabet sizes and frequency offset values. However, by
using the proposed CP-ICA method for high normalized
frequency offset, we can compensate the frequency drift
caused by the difference between oscillators in the transmitter
and the receiver. This is attributed to the fact that the CP-ICA
method estimate the frequency offset very accurately and
Figure 3. BER performance without ICI cancellation of a standard OFDM cancel the frequency offset. The proposed CP-ICA method
system have an extremely performance overall in 0.0< ε <1.0. It gives
a significant boost to performance.
In the case of larger alphabet sizes and larger frequency
offset, OFDM systems do deteriorate the performance too
greatly. Therefore, we can conclude that larger alphabet sizes
and larger frequency offsets are more sensitive to ICI.
Figure 4 and Figure 5 provide comparisons of the
performance of the SC, ML, CP-ML and CP-ICA schemes for
different alphabet sizes and different values of the frequency
offset in 0.0< ε <0.5. In the presence of very small frequency
offset and small alphabet size, self-cancellation [3] gives the
best results. However, for larger alphabet sizes and larger
frequency offset such as 64-QAM and frequency offset of
0.49, SC and ML [5] method does offer a serious decrease in
performance. But the CP-ML [10] and CP-ICA method gives Figure 6. BER Performance with ICI Cancellation, ε = 0.65
the best overall results.

1034
 This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the WCNC 2007 proceedings.
Figure 7. BER Performance with ICI Cancellation, ε = 0.98
Frequency offset estimation for different values ε ={0.032,
0.49, 0.65, 0.98} and 64-QAM subcarrier modulation is shown
in Figure 8. This figure shows that using CP-ICA method can
estimate the frequency offsets very accurately overall.
Figure 8. Frequency Offset Estimation for different values ε = {0.032, 0.49,
0.65, 0.98}, 64-QAM
V. CONCLUSION
Inter-carrier interference which results from the frequency
offset degrades the performance of the OFDM systems. The
self-cancellation (SC), the maximum likelihood (ML) and CP-
ML estimation techniques were proposed in previous
publications [3, 5, 10] for mitigation of the ICI.
The SC [3] does not require very complex hardware or
software for implementation. However, it is not bandwidth
efficient as there is a redundancy of 2 for each carrier. The ML
[5] method also introduces the same level of redundancy but
provides better BER performance, since it accurately estimates
the very small frequency offset. Its implementation is more
complex than the SC method. On the other hand, the CP-ML
[10] method does not reduce bandwidth efficiency as the
frequency offset can be estimated from the Cyclic Prefix in
each OFDM frame. Also, it performed the best of the three
methods in 0.0< ε <0.5. However, this method performs at
frequency domain for estimation of the frequency offset.
In this paper, we propose using CP-ICA scheme for
estimation of frequency offsets in OFDM systems. This
method is a blind scheme and performs more efficiently at
time-domain than at frequency domain. Since our scheme does
not need any pilot symbol in estimation, we can expect
enhanced bandwidth efficiency in OFDM systems. We derive
the frequency offset estimation technique by using ICA
analyzing the phase rotation of the OFDM symbols. The carrier
frequency offset is very accurately estimated using our
technique based on Cyclic Prefix-Independent Component
Analysis. While the proposed CP-ICA method has little
weakness that complexity is a little heavy and the number of
symbols used is enough needed, simulation results show that
the proposed frequency offset estimator is more accurate than
the other estimators in 0.0< ε <1.0.
REFERENCES
[1] T. Pollet, M. van Bladel and M. Moeneclaey, “BER sensitivity of
OFDM systems to carrier frequency offset and Wiener phase noise,”
IEEE Trans. on Communications, vol. 43, no. 2, pp. 887-895, February
1995.
[2] L. J. Cimini Jr., “Analysis and Simulation of a Digital Mobile Channel
Using Orthogonal Frequency Division Multiplexing, IEEE Trans. on
Communications, vol. 33, no. 7, pp. 665 – 675, July 1985.
[3] Y. Zhao and S. Häggman, “Intercarrier interference self-cancellation
scheme for OFDM mobile communication systems,” IEEE Trans. on
Communications, vol. 49, no. 7, pp. 1185 – 1191, July 2001.
[4] C. Muschallik, “Improving an OFDM reception using an adaptive
Nyquist windowing,” IEEE Trans. on Consumer Electronics, vol. 42,
no. 3, pp. 259 – 269, August 1996.
[5] P. H. Moose, “A Technique for Orthogonal Frequency Division
Multiplexing Frequency Offset Correction,” IEEE Trans. on
Communications, vol. 42, no. 10, pp. 2908-2914, October 1994.
[6] U. Lambrette, M. Speth and H. Meyr, “OFDM burst frequency
synchronization by single carrier training data,” IEEE Communications
Letters, vol. 1, pp. 46-48, March 1997.
[7] F. Classen and H. Meyr, “Frequency synchronization algorithms for
OFDM systems suitable for communication over frequency-selective
fading channels,” Proc. of the IEEE Vehicular Technology Conference,
pp. 1655-1659, Stockholm, Sweden, June 1994.
[8] U. Tureli and H. Liu, “Blind carrier synchronization and channel
identification for OFDM communications,” Proc. of IEEE Int. Conf. on
Acoustics, Speech and Signal Processing, pp. 3509-3512, Seattle, WA,
May 1998.
[9] M. Luise and R. Reggiannini, “Carrier acquisition and tracking for
OFDM systems,” IEEE Trans. on Communications, vol. 44, no. 11, pp.
1590-1598, November 1996.
[10] J. Beek, M. Sandell, P. 0. Borjesson, “ML estimation of time and
frequency offset in OFDM systems,” IEEE Trans. on Signal
Processing, vol. 45, no. 7, pp. 1800-1805, July 1997.
[11] H. Bolcskei, “Blind estimation of symbol timing and carrier frequency
offset in pulse shaping OFDM systems,” Proc. of IEEE Int. Conf. on
Acoustics, Speech and Signal Processing, pp. 2749-2752, Phoenix, AZ,
March 1999.
[12] J. H. Gunther, H. Liu and A. L. Swindlehurst, “A new approach for
symbol frame synchronization and carrier frequency estimation in
OFDM communication,” Proc. of The IEEE Int. Conf. on Acoustics,
Speech and Signal Processing, pp.2725-2728, Phoenix, AZ, March
1999.
[13] J. F. Cardoso, "Source Separation using higher order moments", In Proc.
ICASSP'89, pp. 2109-2112, May 1989.
[14] P. Comon, P. Chevalier and V. Capdeville, “Performance of contrast-
based blind source separation,” In proc. SPAWC'97, vol. 1, no. 1, pp.
345-348, April 1997.
[15] A. Hyvarinen, J. karhune, E. Oja, “Independent Component Analysis,”
Wiley Interscience, 2001.
[16] R.V. Nee and R. Prasad, “OFDM for Wireless Multimedia
Communications,” Artech House, 2000.
1035