124
1, JANUARY 2009
CFO Estimation Schemes for Differential OFDM Systems
Xiang Nian Zeng, Student Member, IEEE, and Ali Ghrayeb, Senior Member, IEEE
Abstract—This paper proposes two blind carrier frequency
offset (CFO) estimation schemes for differentially modulated
orthogonal frequency division multiplexing (OFDM) systems.
The proposed schemes estimate the fractional part of the CFO
with only two consecutive OFDM blocks, and they exploit
two implicit properties associated with differentially modulated
OFDM (DOFDM) systems, i.e., the channel keeps constant over
two consecutive OFDM blocks, and the DOFDM systems employ
an M -ary phase-shift keying constellation. One of the schemes
is based on the finite alphabet (FA) constraint and the other
one is based on the constant modulus (CM) constraint. They
provide a trade-off between the performance and computational
complexity. The constrained Cramer-Rao lower bound is also
derived. Several numerical examples are presented to validate
the efficacy of the proposed schemes.
Index Terms—Carrier frequency offset (CFO), constant mod-
ulus (CM), differential modulation, finite alphabet (FA), orthog-
onal frequency division multiplexing (OFDM).
I. I NTRODUCTION
O RTHOGONAL frequency division multiplexing
(OFDM) systems employing differential modulation
have been proposed for situations where channel estimation
becomes impossible. By using differential modulation,
symbols can be detected without channel knowledge and thus
the receiver complexity can be reduced. Like OFDM system,
time domain differentially modulated OFDM (DOFDM)
system, which is the focus of this paper, is also sensitive
to carrier frequency offset (CFO). This DOFDM system is
adopted by the European digital audio broadcasting (DAB)
standard [1]. In [2], the authors propose an estimation scheme
for the fractional part of CFO (FFO) in particular for DOFDM
systems by exploiting the finite alphabet (FA) property of
the transmitted signal. We notice that in their data model
the accumulated phase rotation introduced by CFO is not
considered. However, as their proposed cost function relies on
comparing the phase difference between the symbols in two
consecutive OFDM blocks, the accumulated phase rotation
should not be ignored. With this phase rotation included into
the data model, their cost function actually does not give
a minimum at the true FFO and in turn could not give an
accurate estimation.
Manuscript received April 18, 2008; revised July 14, 2008; accepted August
27, 2008. The associate editor coordinating the review of this letter and
approving it for publication was O. Simeone.
X. Zeng was with the Department of Electrical and Computer Engineering,
Concordia University, Montreal, Quebec, Canada. She is now a postdoctoral
fellow in the Department of Electrical and Computer Engineering, McGill
University, Montreal, Quebec, Canada (e-mail: xiang.zeng@mail.mcgill.ca).
A. Ghrayeb is with the Department of Electrical and Computer
Engineering, Concordia University, Montreal, Quebec, Canada (e-mail:
aghrayeb@ece.concordia.ca).
This work was supported in part by NSERC Grant N00858 and an NSERC
Canada Graduate Scholarship.
Digital Object Identifier 10.1109/T-WC.2009.080546
1536-1276/09$25.00 
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:53 from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO.
In this paper, we propose two FFO estimation schemes by
exploiting two implicit properties associated with the DOFDM
systems, i.e., the channel keeps constant over two consecutive
OFDM blocks, and the DOFDM systems employ an M -ary
phase-shift keying (M -PSK) constellation. The first scheme
is based on the FA constraint. Different from that in [2],
we use differentiation in the frequency domain to bypass the
accumulated phase rotation introduced by CFO in designing
the cost function. The second scheme is based on the constant
modulus (CM) constraint. The two schemes rely on only two
consecutive OFDM blocks. Their cost functions are periodic
even in systems having virtual subcarriers (VSCs) [3], imply-
ing that they can be applied to systems with or without VSCs,
whereas some of the existing FFO estimators, e.g., the CM
based subspace FFO estimator in [4] can not be applied to the
former scenario. As the FA based cost function is somewhat
complicated, we use the modified Newton method to minimize
it efficiently. We also derive the constrained Cramer-Rao lower
bound (CRLB) [6] and use it as a bench mark in assessing
the performance of the proposed schemes.
We demonstrate through several numerical examples that
the two proposed estimators provide a trade-off between
performance and computational complexity. That is, the FA
based scheme is more complicated but it can achieve better
performance at high signal-to-noise ratios (SNRs) as compared
to the CM based scheme. Furthermore, the FA based scheme
also performs a little better than the CM based subspace
scheme in [4] at high SNRs, where the latter has been shown
to have a superior performance over several other schemes.
The rest of the paper is organized as follows. In Section
II, we describe the system model. In Section III, we propose
the FA based cost function and the CM based cost function
respectively for FFO estimation. In Section IV, the constrained
CRLB for CFO estimation is derived. Section V shows the
simulation results. Finally, conclusion is made in Section VI.
II. S YSTEM M ODEL
We consider a DOFDM system with N subcarriers. The
information symbols are differentially modulated along the
time direction on each subcarrier. The channel is assumed to
keep constant over two consecutive OFDM blocks such that
differential detection can be performed at the receiver with
the channel information bypassed. Let us denote the vector of
differentially modulated symbols transmitted in the nth OFDM
block as s(n) = [s0 (n), s1 (n), . . . , sN −1 (n)]T , where sk (n)
denotes the differentially modulated symbol transmitted on
the kth subcarrier. Let d(n) = [d0 (n), d1 (n), . . . , dN −1 (n)]T
denote the vector of information symbols carried in the nth
OFDM block, where dk (n) denotes the information symbol on
the kth subcarrier. The information symbol dk (n) belongs to
c 2009 IEEE
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 1, JANUARY 2009 125

a M -PSK constellation, i.e., A := {β0 , β1 , . . . , βM−1 } with its DFT output can be represented as
i
βi = ej2π M . Then, the differential modulation can be repre-
y(n, ε̃)
sented as s(n) = D(n)s(n− 1), where D(n) =diag(d(n)). In
general, OFDM systems usually have a number of VSCs [3] = FC(ε̃)H r(n)
ε(n−1)Q
to help with spectrum shaping. So, let us denote the number of = ejθ ej2π N FC(ε̆f )C(εi )FH Hs(n) + FC(ε̃)H v(n),
VSCs as P and the number of information bearing subcarriers
(2)
as M = N − P . The set of indices of the VSCs is denoted
by ϕ. For any k in ϕ, we have sk (n) = 0 and dk (n) = 0 for where ε̆f = εf − ε̃. Let yk (n, ε̃) denote the kth element of
all n. y(n, ε̃). From (2), for ε̃ = εf , we have
At the transmitter, the differentially modulated signal is ε(n−1)Q

first passed through the inverse discrete Fourier transform yk (n, εf ) = ejθ ej2π H((k−εi ))N s((k−εi ))N (n) + wk (n),
N

(IDFT), and then appended with a cyclic prefix (CP) before (3)
it is propagated through the multipath fading channel. We where wk (n) denotes the kth element of FC(εf )H v(n), and
denote the equivalent baseband discrete time channel impulse ((x))N represents x modulo N . Eqn. (3) reveals that if εf is
L accurately compensated then the DFT outputs do not have any
response over the nth OFDM block as h(n, t) = l=0
hl (n)δ(t − l), where L is the channel order and hl (n) with inter-carrier-interference.
l = 0, . . . , L, are channel coefficients modeled as independent We observe from (3) that, in the absence of noise,
complex Gaussian random variables (r.v.s) with zero mean and yk1 (n, εf )yk∗2 (n, εf ) = H((k1 −εi ))N s((k1 −εi ))N (n)
any normalized power delay profile. The channel frequency ∗
×H((k s∗
2 −εi ))N ((k2 −εi ))N
(n).
response onthe kth subcarrier can then be represented as
L −j2π lk
Hk (n) = l=0 hl (n)e
N . The length of CP which is Consequently, for k1 = k2 , we have
denoted by LCP , is always assumed to be larger than L such
that there is no inter-block-interference. Therefore, the total yk1 (n, εf )yk∗2 (n, εf ) = yk1 (n − 1, εf )yk∗2 (n − 1, εf )
number of samples in an OFDM block is Q = LCP + N . ×d((k1 −εi ))N (n)d∗((k2 −εi ))N (n).
The received signal vector in the time domain with the CP As both d((k1 −εi ))N (n) and d∗((k2 −εi ))N (n) belong to the M -
removed can be represented as PSK constellation, so does their multiplication denoted by
r(n) = ejθ ej2π
ε(n−1)Q
C(ε)FH H(n)s(n)+v(n), n = 1, 2, . . .
N
αk1 ,k2 . Therefore, heuristically, we can have a cost function
(1) for the joint estimation of εf and detection of αk1 ,k2 , that is
where θ denotes the initial phase offset of the first OFDM Jk1 ,k2 (ε̃, α̃k1 ,k2 ) = |yk1 (n, ε̃)yk∗2 (n, ε̃) − α̃k1 ,k2
block, and ε is the normalized CFO with respect to the
×yk1 (n − 1, ε̃)yk∗2 (n − 1, ε̃)|2 . (4)
subcarrier interval. The CFO has two parts, i.e., ε = εi + εf .
The integer part is denoted by εi , and the fractional part εf This is because in the absence of noise, we have
is within the range [−0.5, 0.5). The matrices C(ε) and H(n) Jk1 ,k2 (ε̃, α̃k1 ,k2 ) = 0 at ε̃ = εf and α̃k1 ,k2 = αk1 ,k2 .
ε ε(N −1)
are defined as C(ε) =diag([1, ej2π N , . . . , ej2π N ]), and Or, alternatively, we can have a FFO estimator given by
H(n) =diag([H0 (n), H1 (n), . . . , HN −1 (n)]). The N ×N dis- minimizing Jk1 ,k2 (ε̃, α̂k1 ,k2 (ε̃)) over ε̃ ∈ [−0.5, 0.5), with
crete Fourier transform (DFT) matrix F has its element on the α̂k1 ,k2 (ε̃) = arg min Jk1 ,k2 (ε̃, α̃).
−j2πmn α̃∈A
mth row and nth column defined as √1N e N . The noise By exploiting the frequency diversity of the OFDM system,
vector is defined as v(n) = [v0 (n), v1 (n), . . . , vN −1 (n)]T , finally we propose a cost function as
where vk (n) with k = 0, . . . , N − 1, are modeled as inde- −1 −1

N 
N
pendent complex Gaussian r.v.s with zero mean and variance JF A (ε̃) = Jk1 ,k2 (ε̃, α̂k1 ,k2 (ε̃)). (5)
σv2 . k1 =0 k2 =k1 +1

The FFO estimation is obtained by finding ε̃ that minimizes

III. P ROPOSED FFO E STIMATORS
A. FA Based Scheme
We first derive the FA based cost function for DOFDM
systems without VSCs, and then show that this cost function
can also be applied directly to systems having VSCs. As
the proposed schemes involve only two consecutive OFDM
blocks, i.e., (n − 1)th and nth, in the rest of the paper we
denote H(n − 1) and H(n) by H for simplicity of notation.1
Let us assume that the received time domain signal r(n) is
first compensated with a FFO candidate ε̃ ∈ [−0.5, 0.5). Then,
1 We note that requiring the channel to remain constant at least over
two consecutive blocks is a general requirement for all systems employing
differential detection.
JF A (ε̃), that is, ε̂f = arg min JF A (ε̃).
ε̃∈[−0.5,0.5)
The cost function derived in (5) is based on a system
without VSCs. However, this cost function can be applied
directly to DOFDM systems having VSCs. This is supported
by two facts. Firstly, in the absence of noise, all of the terms
in the summation of (5) have a minimum at ε̃ = εf . For
example, by assuming ε̃ = εf , let us consider those terms
that have either ((k1 − εi ))N or ((k2 − εi ))N as a VSC,
or have both of them as VSCs. For such terms, we have
Jk1 ,k2 (εf , α̂k1 ,k2 (εf )) = 0 where α̂k1 ,k2 (εf ) can take any
value in A. Secondly, JF A (ε̃) is periodic with a period of one
such that J(ε̃ + m) = J(ε̃) holds for any integer m ∈ [0, N )
and ε̃ ∈ [−0.5, 0.5). This property also holds for DOFDM
systems having VSCs and its proof is shown in Appendix
A. The reason we emphasize the application of the proposed

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:53 from IEEE Xplore. Restrictions apply.
126 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 1, JANUARY 2009

scheme as an FFO estimator to DOFDM systems having VSCs
is that some of the existing blind FFO estimators are developed
for systems without VSCs and can not be applied to systems
having VSCs, e.g., the CM based subspace scheme proposed
in [4].
In minimizing the cost function in (5), we use the mod-
ified Newton method as an alternative to the grid search
to reduce the computational complexity. The gradient and
Hessian of the cost function can be obtained from the terms
d2
dε̃ Jk1 ,k2 (ε̃, α̂k1 ,k2 (ε̃)) and dε̃2 Jk1 ,k2 (ε̃, α̂k1 ,k2 (ε̃)) which are
d
where x̆(n − 1), d̆(n), and F̆ are matrices obtained by
removing those rows with the indices in the set ϕ from
x(n − 1), d(n), and F, respectively, and D̆(n) =diag(d̆(n)).
The unknown deterministic parameters are ε, x̆(n − 1) and
d̆(n). As each element of d̆(n) takes discrete value in a
M -PSK constellation, the regular CRLB, which can only be
applied to the estimation problem with continuous unknown
parameters, does not make sense. So, we use the constrained
CRLB presented in [6]. The constrained CRLB with a FA
constraint is obtained by treating the symbols as if they were

given in Appendix B. As the modified Newton method is known prior [7]. Therefore, in the next, we derive the CRLB
efficient when the initial point is near the actual minimum, of ε, by assuming that d̆(n) is known already.
we select two points that are separated apart by 0.5, e.g., Let us define the observation vector as r = [r(n −
ε(0) = −0.25 and 0.25, and evaluate the cost function at these 1)T , r(n)T ]T and the vector of unknown deterministic param-
two points respectively and choose the point with a smaller eters as Θ = [ε, x̄T , x̃T ]T , where x̄ and x̃ denote the real and
cost value as the initial value. Then, the Newton iteration imaginary part of x̆(n − 1), respectively. Then, the likelihood
as specified in [5, pp. 114] is repeated. Let us denote the function of r conditioned on Θ is given by
CFO estimation obtained after the ith iteration as ε(i+1) . If
f (r|Θ)
|ε(i+1) − ε(i) | < 0.001 is satisfied, then the iteration process
1 2
is terminated and the fractional part of ε(i+1) is accepted as = (πσv2 )−2N exp[− (r(n − 1) − C(ε)F̆H x̆(n − 1)
an FFO estimation; otherwise, continue to the next iteration. σv2
εQ
+ r(n) − ej2π N C(ε)F̆H D̆(n) x̆(n − 1)2 )].
B. CM Based Scheme Then the Fisher information matrix (FIM) can be obtained as
  T 
We can also propose a CM based FFO estimation ∂ ln f (r|Θ) ∂ ln f (r|Θ)
scheme. We observe from (3) that, in the absence of noise, FIM = E
∂Θ ∂Θ
|yk (n, εf )|2 = |H((k−εi ))N |2 |s((k−εi ))N (n)|2 . Consequently, ⎡ 2 ⎤
4π 2π 2π
we have |yk (n, εf )|2 = |yk (n − 1, εf )|2 . This is true because 2 ⎣ 2π N 2 a ∗ N Im{b }
H
N Re{b }
H

= ⎦,
the square amplitude operation essentially removes the data
σv2 N Im{b } 2IM 0M×M
2π ∗
symbols when CM signaling is used. Therefore, the CM based N Re{b } 0M×M 2IM
cost function for FFO estimation is given by 2 2
where a = E1 F̆H x̆(n − 1) + E2 F̆H D̆(n) x̆(n − 1) ,

N −1 H
b = F̆E1 F̆H x̆(n − 1) + D̆(n) F̆E2 F̆H D̆(n)x̆(n − 1),
JCM (ε̃) = (|yk (n, ε̃)|2 − |yk (n − 1, ε̃)|2 )2 . (6) E1 =diag([0, 1, . . . , N − 1] ), and E2 = QIN + E1 . By
T
k=0
using the block matrix inversion lemma, we have
Comparing these two schemes, we notice that the FA based
N 2 σv2 1
cost function in (5) exploits more side information of the CRLB(ε) = (FIM−1 )1,1 = [a − Re(bH b)]−1 .
8π 2 2
M -PSK constellation. Accordingly, the computation of the
FA based cost function is more demanding than that of the V. S IMULATION R ESULTS
CM based cost function due to two factors. Firstly, the FA
We consider a DOFDM system with N = 64 and LCP =
based cost function consists of O(N 2 ) terms while the CM
16. There are P = 12 VSCs. One VSC is on the direct current
based cost function only has N terms. Secondly, the symbol
(DC) subcarrier. The others are located at the two edges of the
detection process is actually carried out implicitly within the
spectrum with each side having 6 and 5 VSCs, respectively.
FA based FFO estimation which indicates that the complexity
The multipath fading channel has an exponential power delay
of the FA based scheme increases with the constellation size.
profile with L = 7, i.e., E{|hl (n)|2 } =  7 1 e−i e−l , l =
However, we will show via simulations that, compared with i=0

the CM based scheme, the FA based scheme can achieve 0, 1, ..., 7. The FFO is estimated by using two consecutive
an additional performance improvement at high SNRs in DOFDM blocks during which the fading channel is keeping
exchange for the increased complexity. constant. The mean squared error (MSE) and the CRLB are
obtained by averaging over 10, 000 runs of simulation. The
SNR is defined as Es /σv2 where Es is the symbol energy on
IV. C ONSTRAINED CRLB D ERIVATION each information bearing subcarrier. For the CM based scheme
In our underlying estimation problem, by defining x(n − (6) and the CM based subspace scheme [4], the grid search
ε(n−2)Q
1) =ejθ ej2π N Hs(n − 1), the observation vectors can be method is used in minimizing their cost functions. In the range
represented as [−0.5, 0.5), 1, 000 equally spaced FFO candidates are used to
achieve a precision of 10−3 .
r(n − 1) = C(ε)F̆H x̆(n − 1) + v(n − 1), Unless stated otherwise, the CFO is uniformly distributed
in the range [−5, 5]. Due to the periodicity of the cost
and
εQ
functions for FFO estimation, we use the modified MSE as
r(n) = ej2π N C(ε)F̆H D̆(n)x̆(n − 1) + v(n), a performance measurement, in which the modified squared

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:53 from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 1, JANUARY 2009 127

−1 −1
10 10
FA based scheme, 2−PSK FA based scheme
FA based scheme, 4−PSK CM based subspace scheme
FA based scheme, 8−PSK CM based scheme
CM based scheme, 2−PSK
−2 CM based scheme, 4−PSK −2
10 CM based scheme, 8−PSK 10

−3 −3
10 10
modified MSE

modified MSE
−4 −4
10 10

−5 −5
10 10

−6 −6
10 10
0 5 10 15 20 25 30 35 0 5 10 15 20 25 30 35
SNR (dB) SNR (dB)

Fig. 1. Comparison of the modified MSE of the FA based FFO estimator and Fig. 2. Comparison of the modified MSE of the FA based FFO estimator,
the CM based FFO estimator for a DOFDM system with VSCs and ε ∈[-5,5]. the CM based FFO estimator and the CM based subspace FFO estimator for
a 4-PSK DOFDM system without VSCs and ε ∈[-5,5].
TABLE I
E FFICIENCY OF THE N EWTON METHOD FOR THE FA BASED FFO TABLE II
ESTIMATOR OVER A DOFDM SYSTEM WITH VSC S AND ε ∈[-5,5]. COMPARISON OF COMPUTATIONAL COMPLEXITY OF THE FA BASED FFO
ESTIMATOR , CM BASED FFO ESTIMATOR , AND THE CM BASED
# of cost func. SUBSPACE FFO ESTIMATOR FOR A SYSTEM WITHOUT VSC S .
complexity # of iterations
evaluations
SNR 10dB 25dB 10dB 25dB Schemes 10dB 25dB
2-PSK 3.7 3.0 5.9 5.5 FA based scheme 1.05s 0.58s
4-PSK 6.8 3.1 8.8 5.1 CM based scheme 0.21s
8-PSK 8.8 4.0 10.8 6.0 CM based subsapce scheme 0.27s

error is defined as SEmodified = min[(ε̂f − εf )2 , (ε̂f − ε̄f )2 ],
where ε̄f is the complement of εf given by ε̄f = εf − 1
if εf > 0; ε̄f = εf if εf = 0; ε̄f = εf + 1 if εf < 0.
The reason for not using regular MSE is that when CFO is
in the neighborhood of m + 0.5, with integer m, the regular
MSE may not be able to provide a correct measurement on
the estimation performance. For example, let us assume that in
one realization, the CFO is 0.48. The FFO estimation could be
−0.49 because of noise. This is an acceptable FFO estimation
as after the FFO compensation by using this estimation, not
much inter-carrier-interference is left at the DFT outputs.
2 with the CM based subspace scheme.
However, the regular squared error is given by (0.97) , which
indicates that it is a rather poor estimation on the contrary.
In Fig. 1, we compare the modified MSE of the FA based
scheme (given by (5)) and the CM based scheme (given by
(6)) for 2-PSK, 4-PSK and 8-PSK, respectively. We observe
that the FA based scheme performs better than the CM based
scheme at high SNRs. This is more prominent for 4-PSK and
8-PSK, in which the SNR gain is about 4dB at high SNR.
We also examine the efficiency of the Newton method for
the FA based scheme. In Table I, we show the required number
of iterations and the number of cost function evaluations at
several SNRs. It shows that at high SNR, the modified Newton
method can converge at around 3 or 4 iterations on average.
In Fig. 2, we compare the modified MSE of the FA based
scheme, the CM based scheme and the CM based subspace
scheme [4] for a 4-PSK DOFDM system without VSCs. The
CM based subspace scheme requires the knowledge of L,
which is usually unknown and is assumed to be equal to 10 in
the estimation process. In [4], the CM based subspace scheme
is shown to be superior to several blind schemes, i.e., VSC
method [3], CM method [8], and VSC+CM method [8]. In this
figure, we show that the FA based scheme performs a little
better than the CM based subspace scheme at high SNRs.
However, at low SNRs, the proposed estimators perform
a little worse than the CM based subspace scheme. This
is because the proposed cost functions are in the form of
comparing two noisy signals. We might say that there is twice
as much noise associated with the proposed schemes compared
In Table II, we consider the same system setup as that of
Fig. 2 and compare the computational complexity of the three
schemes. The complexity is measured by the average time
in seconds used for each estimation process performed by
computer simulation with Matlab. The complexity of the FA
based scheme decreases with SNR. We observe that, although
the Newton method is very efficient, the computational burden
of the FA based scheme is still higher than that of the other
two schemes.
In Fig. 3, we examine the regular MSE of the FA based
scheme by setting ε = 4.32, and compare it with the con-
strained CRLB. The modulation scheme is 4-PSK. To show the
robustness of the modified Newton method, in each run of the
simulation, we randomly select two points that are separated
apart by 0.5 in the range [−0.5, 0.5) as the candidates for the
initial estimate. As expected, the constrained CRLB is much

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:53 from IEEE Xplore. Restrictions apply.
128 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 1, JANUARY 2009

0 0
10 10
FA based scheme, No VSCs FA based scheme, fdT=0.003
constrained CRLB, No VSCs FA based scheme, f T=0.015
FA based scheme, With VSCs d
−1
10 constrained CRLB, With VSCs FA based scheme, ideal case
−1
10 CM based scheme, fdT=0.003
CM based scheme, f T=0.015
d
−2 CM based scheme, ideal case
10
−2
10

−3
10
regular MSE

regular MSE
−3
10
−4
10

−4
10
−5
10

−5
−6 10
10

−7 −6
10 10
0 5 10 15 20 25 30 35 0 5 10 15 20 25 30 35
SNR (dB) SNR (dB)

Fig. 3. Regular MSE performance of the FA based FFO estimator for a Fig. 4. Regular MSE performance of the FA based FFO estimator and the
4-PSK DOFDM system with or without VSCs, and ε = 4.32. CM based FFO estimator in a time variant channel for a 4-PSK DOFDM
system having VSCs and ε = 4.32.

better than the performance of the FA based scheme as it

is actually derived by assuming that the information symbols A PPENDIX A
are known. The performance of the FA based scheme is also P ROOF OF THE P ERIODICITY
compared for systems with or without VSCs. It shows that the
presence of the VSCs only degrades the performance slightly.
In Fig. 4, we examine the regular MSE of the FA based
From (2), we have yk (n, ε̃ + m) = y((k+m))N (n, ε̃). So, we
scheme and the CM based scheme for a 4-PSK DOFDM
have
system in a time variant channel with ε = 4.32. The fading

N −1 
N −1
channel is assumed to keep constant within each OFDM block
but change slowly from one block to another according to the J(ε̃ + m) = Jk1 ,k2 (ε̃, α̂k1 ,k2 (ε̃)),
k1 =0 k2 =k1 +1
Jakes’ fading model [9]. The underlying OFDM system oper-
ates at a central carrier frequency of 5GHz with a 1.25MHz where k1 = ((k1 + m))N and k2 = ((k2 + m))N . Then, we
total bandwidth. The normalized Doppler frequency fd T is set have
to 0.003, and 0.015, where fd denotes the Doppler frequency 
N −1 
N −1
and T denotes the duration of an OFDM block including CP. J(ε̃ + m) = Jk1 ,k2 (ε̃, α̂k1 ,k2 (ε̃))
As shown in the figure, the FA based scheme is slightly more k1 =m k2 =k1 +1
sensitive to channel variations than the CM based scheme. We 
m−1 
m−1
remark here that the CM based subspace scheme is insensitive + Jk1 ,k2 (ε̃, α̂k1 ,k2 (ε̃))
to channel variations. This is because when multiple OFDM k1 =0 k2 =k1 +1
blocks are available for CFO estimation, the CM based sub- −1
 N
m−1
space scheme does not require channel to keep constant over + Jk2 ,k1 (ε̃, α̂k2 ,k1 (ε̃)). (7)
them. k1 =0 k2 =m

From (4), we also have that

VI. C ONCLUSIONS
Jk2 ,k1 (ε̃, α̂k2 ,k1 (ε̃)) = |[yk1 (n, ε̃)][yk2 (n, ε̃)]∗ − α̂∗k2 ,k1 (ε̃)
In this paper, we proposed two FFO estimators which can
×yk1 (n − 1, ε̃)[yk2 (n − 1, ε̃)]∗ |2 .
be applied for DOFDM systems employing M -PSK only. The
proposed schemes rely on two consecutive OFDM blocks over (8)
which channel is required to keep constant. The constrained It can be easily shown that α̂∗k ,k (ε̃) = α̂k1 ,k2 (ε̃). Therefore,
CRLB was derived. Simulation results showed that the FA we have from (8) that
2 1

based scheme has a higher computational complexity than

that of the CM based scheme and the existing CM based Jk2 ,k1 (ε̃, α̂k2 ,k1 (ε̃)) = Jk1 ,k2 (ε̃, α̂k1 ,k2 (ε̃)).
subspace scheme, but it can achieve a little better performance
Finally, from (7), we obtain
at high SNRs. The proposed schemes also provide a wider
 −1  −1
application to systems having VSCs as compared to the CM N N

based subspace scheme. However, they are sensitive to channel J(ε̃ + m) = Jk1 ,k2 (ε̃, α̂k1 ,k2 (ε̃)) = J(ε̃).
variations, which is more significant for the FA based scheme. k1 =0 k2 =k1 +1

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:53 from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 1, JANUARY 2009 129

A PPENDIX B where ak1 ,k2 , āk1 ,k2 , ăk1 ,k2 , bk1 ,k2 , b̄k1 ,k2 , b̆k1 ,k2 are el-
G RADIENT AND H ESSIAN ements on the (k1 )th row and (k2 )th column of matrices
Let us define A(ε(i) , n), Ā(ε(i) , n), Ă(ε(i) , n), A(ε(i) , n − 1), Ā(ε(i) , n −
1), Ă(ε(i) , n − 1), respectively.
A(ε̃, n) = [FC(ε̃)H r(n)][FC(ε̃)H r(n)]H ,
4π R EFERENCES
Ā(ε̃, n) = − Im{[FE1 C(ε̃)H r(n)][FC(ε̃)H × r(n)]H },
N [1] ETSI EN 300 401 v1.3.3 (2001-05), “Radio broadcasting systems; Digital
and audio broadcasting (DAB) to mobile, portable and fixed receivers," May
2001.
8π 2 [2] Z. Liu and B. Weng, “Finite-alphabet based blind carrier frequency offset
Ă(ε̃, n) = − (Re{[FE21 C(ε̃)H r(n)][FC(ε̃)H r(n)]H } estimation for differentially coded OFDM," in Proc. IEEE Globecom, vol.
N2 1, pp. 327-331, Dec. 2003.
−[FE1 C(ε̃)H r(n)][FE1 C(ε̃)H r(n)]H ). [3] H. Liu and U. Tureli, “A high-efficiency carrier estimator for OFDM
communications," IEEE Commun. Lett., pp. 104-106, Apr. 1998.
Then, the gradient and Hessian of Jk1 ,k2 (ε̃, α̂k1 ,k2 (ε̃)) at ε̃ = [4] T. Roman and V. Koivuen, “Subspace method for blind CFO estimation
ε(i) , are given by for OFDM systems with constant modulus constellations," in Proc. IEEE
 Veh. Technol. Conf., May 2005, vol. 2, pp. 1253-1257.
d  [5] G. P. McCormick, Nonlinear Programing, Theory, Algorithm and Appli-
Jk1 ,k2 (ε̃, ᾱk1 ,k2 (ε̃)) cations. John Wiley & Sons, Inc., 1983.
dε̃ (i)
ε̃=ε [6] P. Stoica and B. C. Ng, “On the Cramer-Rao bound under parametric
= 2Re{āk1 ,k2 a∗k1 ,k2 + b̄k1 ,k2 b∗k1 ,k2 constraints," IEEE Signal Processing Lett., no. 7, pp. 177-179, July 1998.
[7] B. M. Sadler, R. J. Kozick, and T. Moore, “Bounds on bearing and symbol
−α̂∗k1 ,k2 (ε(i) )āk1 ,k2 b∗k1 ,k2 − α̂∗k1 ,k2 (ε(i) )ak1 ,k2 b̄∗k1 ,k2 }, estimation with side information," IEEE Trans. Signal Processing, vol.
49, no. 4, pp. 822-834, Apr. 2001.
and [8] M. Ghogho and A. Swami, “Blind frequency-offset estimator for OFDM
 systems transmitting constant-modulus symbols," IEEE Commun. Lett.,
d2  vol. 6, pp. 343-345, Aug. 2002.
J (ε̃, ᾱ (ε̃))
dε̃2
k1 ,k2 k1 ,k2  (i) [9] P. Dent, G. E. Bottomley, and T. Croft, “Jakes fading model revisited,"
ε̃=ε Electron. Lett., vol. 29, no. 13, pp. 1162-1163, June 1993.
= 2Re{ăk1 ,k2 a∗k1 ,k2 + āk1 ,k2 ā∗k1 ,k2 + b̆k1 ,k2 b∗k1 ,k2
+b̄k1 ,k2 b̄∗k1 ,k2 − α̂∗k1 ,k2 (ε(i) )ăk1 ,k2 b∗k1 ,k2
−α̂∗k1 ,k2 (ε(i) )āk1 ,k2 b̄∗k1 ,k2 − α̂∗k1 ,k2 (ε(i) )āk1 ,k2 b̄∗k1 ,k2
−α̂∗k1 ,k2 (ε(i) )ak1 ,k2 b̆∗k1 ,k2 },

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:53 from IEEE Xplore. Restrictions apply.