Frequency Offset Estimation and

1244 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO.
4, APRIL 2008
Frequency Offset Estimation and

Training Sequence Design for MIMO OFDM
Yanxiang Jiang, Member, IEEE, Hlaing Minn, Senior Member, IEEE, Xiqi Gao, Senior Member, IEEE,
Xiaohu You, Member, IEEE, and Yinghui Li, Student Member, IEEE
Abstract— This paper addresses carrier frequency offset compensation of CFO is very important [2]. A number of
(CFO) estimation and training sequence design for multiple- approaches have dealt with CFO estimation in a SISO-OFDM
input multiple-output (MIMO) orthogonal frequency division setup [3], [4], [2], [5], [6], [7]. According to whether the
multiplexing (OFDM) systems over frequency selective fad-
ing channels. By exploiting the orthogonality of the training CFO estimators use training sequences or not, they can be
sequences in the frequency domain, integer CFO (ICFO) is classified as blind ones [3] [4] and training-based ones [2],
estimated. With the uniformly spaced non-zero pilots in the [5], [6], [7]. Similar to SISO-OFDM, MIMO-OFDM is also
training sequences and the corresponding geometric mapping, very sensitive to CFO. Moreover, for MIMO-OFDM, there
fractional CFO (FCFO) is estimated through the roots of a exists multi-antenna interference (MAI) between the received
real polynomial. Furthermore, the condition for the training
sequences to guarantee estimation identifiability is developed. signals from different transmit antennas. The MAI makes CFO
Through the analysis of the correlation property of the training estimation more difficult, and a careful training sequence de-
sequences, two types of sub-optimal training sequences generated sign is required for training-based CFO estimation. However,
from the Chu sequence are constructed. Simulation results verify unlike SISO-OFDM, only a few works on CFO estimation
the good performance of the CFO estimator assisted by the for MIMO-OFDM have appeared in the literature. In [8], a
proposed training sequences.
blind kurtosis-based CFO estimator for MIMO-OFDM was
Index Terms— MIMO-OFDM, frequency selective fading chan- developed. For training-based CFO estimators, the overviews
nels, training sequences, frequency offset estimation. concerning the necessary changes to the training sequences
and the corresponding CFO estimators when extending SISO-
I. I NTRODUCTION OFDM to MIMO-OFDM were provided in [9], [10]. How-
ever, with the provided training sequences in [9], satisfactory
RTHOGONAL frequency division multiplexing
O (OFDM) transmission is receiving increasing attention
in recent years due to its robustness to frequency-selective
CFO estimation performance cannot be achieved. With the
training sequences in [10], the training period grows linearly
with the number of transmit antennas, which results in an
fading and its subcarrier-wise adaptability. On the other
increased overhead. In [11], a white sequence based maximum
hand, multiple-input multiple-output (MIMO) systems
likelihood (ML) CFO estimator was addressed for MIMO,
attract considerable interest due to the higher capacity and
while a hopping pilot based CFO estimator was proposed
spectral efficiency that they can provide in comparison with
for MIMO-OFDM in [12]. Numerical calculations of the
single-input single-output (SISO) systems. Accordingly,
CFO estimators in [11] [12] require a large point discrete
MIMO-OFDM has emerged as a strong candidate for beyond
Fourier transform (DFT) operation and a time consuming line
third generation (B3G) mobile wide-band communications
search over a large set of frequency grids, which make the
[1].
estimation computationally prohibitive. To reduce complexity,
It is well known that SISO-OFDM is highly sensitive to
computationally efficient CFO estimation was introduced in
carrier frequency offset (CFO), and accurate estimation and
[13] by exploiting proper approximations. However, the CFO
Manuscript received October 18, 2006; revised February 5, 2007; accepted estimator in [13] is only applied to flat-fading MIMO channels.
February 10, 2007. The associate editor coordinating the review of this
manuscript and approving it for publication was D. Huang. The work of When training sequence design for CFO estimation is
Yanxiang Jiang, Xiqi Gao, and Xiaohu You was supported in part by the concerned, it has received relatively little attention. It was
National Natural Science Foundation of China under Grants 60496311 and
60572072, the China High-Tech 863 Project under Grant 2003AA123310 investigated for single antenna systems in [14], where a white
and 2006AA01Z264, and the International Cooperation Project on Beyond sequence was found to minimize the worst-case asymptotic
3G Mobile of China under Grant 2005DFA10360. The work of Hlaing Minn Cramer-Rao bound (CRB). Recently, an improved training
and Yinghui Li was supported in part by the Erik Jonsson School Research
Excellence Initiative, the University of Texas at Dallas, USA. This paper was sequence and structure design was developed in [15] by
presented in part at the IEEE International Conference on Communications exploiting the CRB and received training signal statistics. In
(ICC), Istanbul, Turkey, June 2006. [16], training sequences were designed for CFO estimation
Y. Jiang, X. Gao, and X. You are with the National Mobile Communications
Research Laboratory, Southeast University, Nanjing 210096, China (e-mail: in MIMO systems using a channel-independent CRB. In
{yxjiang, xqgao, xhyu}@seu.edu.cn). [17], the effect of CFO was incorporated into the mean-
H. Minn and Y. Li are with the Department of Electrical Engineering, square error (MSE) optimal training sequence designs for
University of Texas at Dallas, TX 75083-0688, USA (e-mail: {hlaing.minn,
yinghui.li}@utdallas.edu). MIMO-OFDM channel estimation in [18]. Note that optimal
Digital Object Identifier 10.1109/TWC.2008.060846. training sequence design for MIMO-OFDM CFO estimation
1536-1276/08$25.00
c 2008 IEEE
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 09:22 from IEEE Xplore. Restrictions apply.
JIANG et al.: FREQUENCY OFFSET ESTIMATION AND TRAINING SEQUENCE DESIGN FOR MIMO OFDM 1245
in frequency selective fading channels is still an open problem. CFO. Define

In this paper, we propose a new training-based CFO es-
DN̄ (ε) = diag{[1, ej2πε/N , · · · , ej2πε(N̄ −1)/N ]T },
timator for MIMO-OFDM over frequency selective fading
channels. To avoid the large point DFT operation and the time where ε is the frequency offset normalized by the subcarrier
consuming line search and also to guarantee the estimate preci- spacing. Suppose the length-L channel impulse response from
sion, we propose to first estimate integer CFO (ICFO) and then the μ-th transmit antenna to the ν-th receive antenna is denoted
fractional CFO (FCFO). Next, we relate FCFO estimation to by the L × 1 vector h(ν,μ) . Then, after removing the CP at
direction of arrival (DOA) estimation in uniform linear arrays the ν-th receive antenna, the N × 1 received vector yν can be
(ULA). Then, we propose to exploit a geometric mapping to written as [12]
transform the complex polynomial related to FCFO estimation
N
t −1
into a real polynomial. Correspondingly, FCFO is estimated √
yν = N ej2πεNg /N DN (ε) FNH diag{h̃(ν,μ) }t̃μ
through the roots of the real polynomial. In the derivation
μ=0
of the CFO estimator, we also develop the identifiability
condition concerning the training sequences. Furthermore, + wν , (1)
exploiting the well-studied results for DOA estimation, we where
L−1 (ν,μ)
transform the problem of training sequence design for CFO h̃(ν,μ) = FN [e0N , e1N , · · · , eN ]h ,
estimation into the study of the correlation property of the
training sequences, and propose to construct them from the and wν is an N ×1 vector of additive white complex Gaussian
Chu sequence [19]. noise (AWGN) samples with zero-mean and equal variance of
2
The rest of this paper is organized as follows. In Section σw .
II, we briefly describe the MIMO-OFDM system model. The The goal of this paper is to design the training sequences
Nt −1 Nr −1
proposed CFO estimator including ICFO and FCFO estimation {t̃μ }μ=0 and estimate ε from the observation of {yν }ν=0 .
is presented in Section III. The aspect concerning training
sequence design is addressed in Section IV. Simulation results
are shown in Section V. Final conclusions are drawn in Section III. F REQUENCY O FFSET E STIMATION FOR MIMO OFDM
VI.
Let Q = N/P with (N )P = 0. Design
Notations : Upper (lower) bold-face letters are used for
matrices (column vectors). Superscripts ∗, T and H denote (C0) 0 ≤ i0 < i1 < · · · < iμ < · · · < iNt −1 < Q.
conjugate, transpose and Hermitian transpose, respectively.
(·)P denotes the residue of the number within the brackets Define
modulo P . (·) and (·) denote the real and imaginary parts q+(P −1)Q
Θq = [eqN , eN
q+Q
, · · · , eN ], 0 ≤ q < Q.
of the enclosed parameters, respectively. ·, || · ||2 , E[·] and
⊗ denote the floor, Euclidean norm-square, expectation and Let s̃μ denote a length-P sequence whose elements are all
Kronecker product operators, respectively. sign(·) denotes the non-zero. Then, we propose to construct the training sequence
signum function and sign(0) = 1 is assumed. [x]m denotes transmitted from the μ-th antenna as t̃μ = Θiμ s̃μ (C1).
the m-th entry of a column vector x. x(m) denotes the m- Without loss of generality, the total energy allocated to training
cyclic-down-shift version of x for m > 0 and |m|-cyclic- is supposed to be split equally between the transmit antennas,
up-shift version of x for m < 0. diag{x} denotes a diagonal i.e., ||s̃μ ||2 = N/Nt (C2). Note that the entire training symbol
matrix with the elements of x on its diagonal. [X]m,n denotes is utilized by our proposed CFO estimator.
the (m, n)-th entry of a matrix X. FN and IN denote the
N × N unitary DFT matrix and the N × N identity matrix,
respectively. ekN denotes the k-th column vector of IN . 1Q A. ICFO Estimation
(0Q ) and 0P ×Q denote the Q × 1 all-one (all-zero) vector and Let y = [y0T , y1T , · · · , yνT , · · · , yN
T
]T denote the Nr N ×
r −1
P × Q all-zero matrix, respectively. JQ denotes the Q × Q 1 cascaded vector from the Nr receive antennas. Then, y can
exchange matrix with ones on its anti-diagonal and zeros be written as
elsewhere. Unless otherwise stated, 0 ≤ μ ≤ Nt − 1 and √
0 ≤ ν ≤ Nr − 1 are assumed. y = N ej2πεNg /N {INr ⊗ [DN (ε)S]}h + w, (2)
where
II. S YSTEM M ODEL
h = [hT0 , hT1 , · · · , hTν , · · · , hTNr −1 ]T ,
Let us consider a MIMO-OFDM system with Nt transmit hν = [(h(ν,0) )T , (h(ν,1) )T , · · · , (h(ν,μ) )T , · · · , (h(ν,Nt −1) )T ]T ,
antennas, Nr receive antennas and N subcarriers. Suppose the
S = F̄ H diag{[s̃T0 , s̃T1 , · · · , s̃Tμ , · · · , s̃TNt −1 ]T }F̆ ,
training sequence transmitted from the μ-th antenna is denoted
F̄ = [Θi0 , Θi1 , · · · , Θiμ , · · · , ΘiNt −1 ]T F N ,
by the N × 1 vector t̃μ . Before transmission, this vector is
processed by an inverse discrete Fourier transform (IDFT), F̆ = [e0Nt ⊗ ΘTi0 , e1Nt ⊗ ΘTi1 , · · · , eμ T
Nt ⊗ Θiμ ,
and a cyclic prefix (CP) of length Ng is inserted. We assume Nt −1
· · · , eNt
⊗ ΘTiNt −1 ]{INt ⊗ [FN [IL , 0L×(N −L) ]T ]},
that Ng ≥ L − 1, where L is the maximum length of all the
w = [w0T , w1T , · · · , wνT , · · · , wN
T
r −1
]T .
frequency selective fading channels. We further assume that
all the transmit-receive antenna pairs are affected by the same
1246 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 4, APRIL 2008
From (2), the ML estimation of ε can be readily obtained and condition (C1), which implies the periodic property of the
[20] as follows training sequences, ȳ can be expressed in an equivalent form
H H −1 H
as shown in (7) at the bottom of the page. In (7),
ε̂ = arg max y {INr ⊗[DN (ε̃)S̃(S̃ S̃) S̃ DN (−ε̃)]}y ,
ε̃
(3) βμ = εf + iμ ,
where v = e−j2πεi Ng /N [INr ⊗ DN (−εi )]w.
S̃ = F̄ H diag{[s̃T0 , s̃T1 , · · · , s̃Tμ , · · · , s̃TNt −1 ]T }. It follows from (7) that the estimation of εf is equivalent to
Nt −1
the estimation of the Nt different equivalent CFOs {βμ }μ=0 .
With condition (C0), which implies the orthogonality of the
training sequences in the frequency domain, matrix inversion Furthermore, exploiting the periodic property of the training
involved in (3) can be avoided, and (3) can thus be simplified sequences again, we can stack ȳ into the Q × Nr P matrix
as follows Y = [Y0 , Y1 , · · · , Yν , · · · YNr −1 ], where
[Yν ]q,p = [((eνNr )T ⊗ IN )ȳ]qP +p , 0 ≤ q < Q, 0 ≤ p < P.
ε̂ = arg max ||{INr ⊗ [F̄ DN (−ε̃)]}y||2 . (4)
ε̃
Then, (7) can be expressed in the following equivalent form
Actually, with condition (C0), our proposed training sequences
can be treated as frequency-division multiplexing (FDM) pilot Y = BX + V , (8)
allocation type sequences [18].
where
Let l denote the pilot location vector which is given by
Nt −1
i B = [b0 , b1 , · · · , bμ , · · · , bNt −1 ],
l= eQμ . Then, we have as follows.
μ=0 bμ = [1, ej2πβμ /Q , · · · , ej2πβμ q/Q , · · · , ej2πβμ (Q−1)/Q ]T ,
Theorem 1: With ε̃, ε ∈ (−Q/2, Q − Q/2], ε̃ = ε
X = [X0 , X1 , · · · , Xν , · · · , XNr −1 ],
uniquely maximizes the cost function in (4) for any h(ν,μ) (=
0L ) with the following condition Xν = [x(ν,0) , x(ν,1) , · · · , x(ν,μ) , · · · , x(ν,Nt −1) ]T ,
√
x(ν,μ) = P ej2πεf Ng /N DP (βμ )
(C3) (N − Nt P ) ≥ Nt P & P ≥ L &
×FPH diag{s̃μ }ΘTiμ FN [IL , 0L×(N −L) ]T h(ν,μ) ,
(1Q − l)T l(q) > 0, ∀q ∈ {1, 2, · · · , Q − 1}.
Proof: See Appendix I. and V is the Q × Nr P matrix generated from v in the same
It follows immediately that the corresponding estimation is way as Y . From (8), we can see that there exists an inherent
identifiable for ε ∈ (−Q/2, Q − Q/2] if condition (C3) relationship between FCFO estimation and DOA estimation in
is satisfied. ULA [21] with P , Q and Nr satisfying Nr P > Q (C4).
From (4), the CFO can be estimated by exploiting fast The covariance matrix of Y can be estimated by R̂Y Y =
Fourier transform (FFT) interpolation and a time consuming Y Y H /(Nr P ). Let L denote the Q × Q unitary column
line search over a large set of frequency grids. But this conjugate symmetric matrix which is given by
approach is computationally complicated and the estimate
precision also depends on the FFT size used. To reduce [1 − (Q)2 ] IQ/2 jIQ/2
L= √
complexity, ε is divided into the ICFO εi and the FCFO εf . 2 JQ/2 −jJQ/2
⎡ ⎤
The ICFO εi can be estimated by invoking only the N point I(Q−1)/2 0(Q−1)/2
√ jI(Q−1)/2
(Q)2 ⎣ T
FFT as follows + √ 0(Q−1)/2 2 0T(Q−1)/2 ⎦ .
2 J(Q−1)/2 0(Q−1)/2 −jJ(Q−1)/2
ε̂i = arg max ||{INr ⊗ [F̄ DN (−ε̃i )]}y||2 . (5)
−Q/2<ε̃i ≤Q−Q/2
Then, with the aid of L, the complex matrix R̂Y Y can be
The discussion on the uniqueness of ε̂i = εi in maximizing transformed into a real matrix [22] as follows
the estimation metric for any h(ν,μ) (= 0L ) is provided in r H ∗
Appendix II. Other ICFO estimators may also be used. Once R̂Y Y = 1/2 · L (R̂Y Y + JQ R̂Y Y JQ )L
the ICFO is estimated, ICFO correction can be readily carried = (LH R̂Y Y L). (9)
out as follows r
The eigen-decomposition of the real matrix R̂Y Y can be
−j2π ε̂i Ng /N
ȳ = e [INr ⊗ DN (−ε̂i )]y. (6) obtained as
r H H
R̂Y Y = EX ΛX EX + EV ΛV EV , (10)
B. FCFO Estimation
where
The FCFO εf is estimated based on ȳ. Assume ε̂i = εi
and substitute (2) into (6). Then, by exploiting condition (C0) ΛX = diag{[λ0 , λ1 , · · · , λNt −1 ]T },
√
ȳ = N ej2πεf Ng /N INr ⊗ {[DN (β0 ), DN (β1 ), · · · , DN (βμ ), · · · , DN (βNt −1 )]

× (INt ⊗ 1Q ⊗ FPH )diag{[s̃T0 , s̃T1 , · · · , s̃Tμ , · · · , s̃TNt −1 ]T }F̆ } h + v, (7)
2
ΛV = σw IQ−Nt , which shows that Φ is a column conjugate symmetric matrix
2
λ0 ≥ λ1 ≥ · · · ≥ λNt −1 > σw , no matter Q is even or odd.
and EX and EV contain the unitary eigen-vectors that span By exploiting (13) and (14) in (11), the following equivalent
the signal space and noise space, respectively. Let polynomial can be obtained (ignoring the constant items)
z = ej2πβ/Q , f r (g) = aT (g)JQ Ar a(g), (17)

a(z) = [1, z, · · · , z Q−1 ]T . where
Ar = JQ ΦH L[IQ − EX EX
H
]LH Φ.
Then, by exploiting a(z), a polynomial of degree 2(Q − 1)
with complex coefficients can be obtained [23] as follows Due to the column conjugate symmetric property of L and
Φ, we have (18) as shown at the bottom of the page. Hence,
f (z) = aT (z)JQ Aa(z), (11)
ΦH L becomes a real matrix no matter Q is even or odd.
where Moreover, EX is also a real matrix. Accordingly, f r (g) is
H
A = L[IQ − EX EX ]LH . transformed into a polynomial with real coefficients. The roots
Nt −1
of f r (g) = 0 can be obtained by the fast root-calculating
Correspondingly, {βμ }μ=0 can be indirectly estimated by algorithms for real polynomials in [24], whose computational
calculating the pairwise roots of f (z) = 0 which are closest complexities are much less than those for complex ones.
to the unit circle.
Note that by its definition, z is always located on the After the roots of f r (g) = 0 are obtained, the FCFO can
unit circle. By analysis, we find that the following geometric be readily estimated according to the following steps.
mapping exists,
1) Find the Nt pairwise roots of f r (g) = 0 whose imagi-
g(z) = cot(πβ/Q) = j(z + 1)/(z − 1). (12) nary parts have the smallest absolute values, {(gμ ) ±
Nt −1
It follows immediately that g(z) is a monotonic reversible j(gμ )}μ=0 .
mapping with real values for β ∈ [−0.5, Q − 0.5). Hence, we 2) Calculate the equivalent CFOs corresponding to the Nt
have z(g) = (g + j)/(g − j). Then, a(z) can be expressed as roots,
a function of g as β̂μ = (Q/π × acot[(gμ )])Q . (19)
a(z) = (g − j)1−Q d(g), (13) 3) Calculate the FCFO ε̂fμ corresponding to each β̂μ ,
⎧
where ⎪
⎪ β̂μ − i0 , β̂μ ∈ ((i0 − th )Q , i0 + th )
⎪
⎪
⎨ β̂μ − i1 , β̂μ ∈ (i1 − th , i1 + th )
d(g) = [(g − j)Q−1 , (g + j)(g − j)Q−2 , · · · , (g + j)Q−1 ]T .
ε̂fμ = ··· ··· ,
⎪
⎪
Since the elements of d(g) are polynomials of degree (Q − 1) ⎪ β̂μ − iNt −1 ,
⎪ β̂μ ∈ (iNt −1 − th ,
⎩
with respect to g, d(g) in (13) can be further expressed as (iNt −1 + th )Q )
(20)
d(g) = Φa(g), (14) where th denotes a threshold which is predefined to
where Φ is the Q × Q coefficient matrix. By straight-forward avoid the ambiguous estimation and 0.5 < th < 1,
calculation, the elements of Φ are obtained as shown in [a, b) has the usual meaning for a < b, and [a, b) =

(15). In (15), Cqq = q!/[(q − q )!q !]. Then, from (15), [0, b) ∪ [a, Q) for a > b.
Nt −1
we immediately obtain 4) Obtain the final FCFO by averaging {ε̂fμ }μ=0 ,
Nt −1
[Φ]q,q = [Φ]∗Q−1−q,q & (Q)2 [Φ](Q−1)/2,q = 0, 1
ε̂f = ε̂f . (21)
∀q, q ∈ {0, 1, · · · , Q − 1}, (16) Nt μ=0 μ
min{q,q }

Q−1−q q −q
[Φ]q,q = j Cqq CQ−1−q (−1)Q−1−q−q +q , 0 ≤ q, q ≤ Q − 1, (15)
q =max{0,q+q −Q+1}
[1−(Q)2 ]Q/2+(Q)2 (Q−1)/2−1

∗
[ΦH L]q,q = [Φ]q ,q [L]q ,q + [Φ]q ,q [L]∗q ,q + (Q)2 [Φ](Q−1)/2,q [L](Q−1)/2,q
q =0
⎧ ⎫
⎨[1−(Q)2 ]Q/2+(Q)
2 (Q−1)/2−1 ⎬
= 2 [Φ]∗q ,q [L]q ,q + (Q)2 [Φ](Q−1)/2,q [L](Q−1)/2,q ,
⎩ ⎭
q =0
∀q, q ∈ {0, 1, · · · , Q − 1}. (18)
C. Computational Complexity sequences from a length-P Chu sequence s with its element
The computational load of our proposed CFO estimator given by
2
mainly involves the N point FFT, the eigen-decomposition [s]p = ejπvp /P , 0 ≤ p ≤ P − 1,
r r
of R̂Y Y and the root-calculation for f (g) = 0, which where v is coprime to P , and P is supposed to be even.
require 4N log2 N , 9Q3 and 64/3 · (Q − 1)3 real additions
Let s̃μ = Q/Nt FP s(μM ) with M = P/NI and NI ≥
or multiplications [24] [25], respectively. Compared with the Nt . Then, we refer to
direct CFO estimate from (4), which requires a large point the so-constructed training sequences
as TS 0. Let s̃μ = Q/Nt FP s. Then, we refer to the so-
DFT operation and an exhaustive line search, our proposed constructed training sequences as TS 1. Note that the TS 1
estimator has significantly lower complexity especially with training sequences are equivalent to the so-called repeated
a relatively small Q. Furthermore, by applying the approach phase-rotated Chu (RPC) sequences [28]. Define
in [26], the complexity of our proposed estimator can be
further decreased by calculating the roots from the first-order pμ,l = (1 − m)μM + l,
derivative of f r (g) = 0, but at the cost of a slight performance
where m = 0 for TS 0, and m = 1 for TS 1. Then, from the
degradation at low signal-to-noise ratio (SNR).
above constructions we have
1 2 2
IV. T RAINING S EQUENCE D ESIGN [T (μ,μ ) ]l,l = (−1)v(pμ,l −pμ ,l )+1 ejπv(pμ,l −pμ ,l )/P
With our design conditions in the previous section, the Nt
training sequences are determined completely by the base ×e−jπ(P −1)[v(pμ,l −pμ ,l )−μ,μ ]/P sin(π μ,μ )
Nt −1 Nt −1
sequences {s̃μ }μ=0 with fixed P , Q and {iμ }μ=0 . Let / sin{π[v(pμ,l − pμ ,l ) − μ,μ ]/P }. (24)
RXX denote the covariance matrix of X. It is pointed out in
It follows immediately from (24) that
[21] [22] that the optimal performance can be achieved with
(μ,μ )
uncorrelated signals, i.e., diagonal matrix RXX . Since RXX [T ]l,l p −p =0
can be estimated by R̂XX = XX H /(Nr P ), it is expected μ,l μ ,l
(μ,μ )
that good performance can be achieved with diagonal matrix
[T ]l,l p −p =0 < P/Nt . (25)
μ,l μ ,l
R̂XX . Making R̂XX to be a diagonal matrix is equivalent to (μ,μ )
making [R̂XX ]μ,μ μ=μ = 0. Define We see that [T ]l,l achieves its maximum for TS 0 when
(μ − μ )M = l − l and for TS 1 when l = l . Assume that

μ,μ = (iμ − iμ )/Q, the channel energy is mainly concentrated in the preceding
1 M channel taps and the first channel tap is the dominant one.
sμ = √ FPH s̃μ .
Q Then,
it can be inferred from (22) and (25) that the value
Assume that the channel taps remain constant during the of [R̂XX ]μ,μ μ=μ for TS 0 is very small, and it is much
training period. Then, [R̂XX ]μ,μ can be expressed as smaller than that for TS 1 in the same channel environment.
In this sense, we can say that TS 0 is superior to TS 1, which
[R̂XX ]μ,μ = will be verified through simulation results in the following
Q
Nr −1 L−1
L−1
section.
[h(ν,μ) ]l [T (μ,μ ) ]l,l [h(ν,μ ) ]∗l , (22) Note that our proposed training sequence structure is similar
Nr ν=0 to the one introduced recently in [16], and the identifiability
l=0 l =0
where of our CFO estimator, however, cannot be guaranteed with the

training sequences in [16].
(l )
[T (μ,μ ) ]l,l = (s(l) T ∗
μ ) DP ( μ,μ Q)(sμ ) .
Since the elements of h(ν,μ) are random variables, V. S IMULATION R ESULTS

[R̂XX ]μ,μ μ=μ = 0 can be achieved by making the training To evaluate our CFO estimator’s performance with the pro-
sequences satisfy posed training sequences for MIMO-OFDM, a number of sim-
ulations are carried out. Throughout the simulations, a MIMO
[T (μ,μ ) ]l,l = 0, if μ = μ & 0 ≤ l, l ≤ L − 1. (23)
OFDM system of bandwidth 20MHz operating at 5GHz with
It follows immediately that the optimal training sequences N = 1024 and Ng = 64 is used. Each channel has 4 inde-
should satisfy the condition in (23). However, with condition pendent Rayleigh fading taps, whose relative average-powers
(C0), μ,μ is definitely a decimal fraction, which greatly and propagation delays are {0, −9.7, −19.2, −22.8}dB and
complicates the satisfaction of the condition in (23) with {0, 0.1, 0.2, 0.4}μs, respectively. For the training sequences,
proper training sequences. To ease the above problem, we we set P and Q to 64 and 16, respectively. The normalized
construct two types of sub-optimal training sequences which CFO ε is generated within the range (−Q/2, Q − Q/2].
can make the off-diagonal elements of R̂XX as small as For description convenience, we henceforth refer to the pro-
possible. It has been proven in [15] that training sequence posed real polynomial based CFO estimator as RPBE.
with the zero auto-correlation (ZAC) property is optimal for In the following, we use the average CRB (avCRB), which
CFO estimation in SISO frequency selective fading channels. corresponds to the extended Miller and Chang bound (EMCB)
Besides, constant amplitude ZAC (CAZAC) sequence (e.g., [15] [29], to benchmark the performance of RPBE. The
[19], [27] and references therein) is often a preferred choice average CRB or EMCB is obtained by simply averaging the
for training. Therefore, we propose to construct the training snapshot CRB over independent channel realizations, which
Fig. 1. CFO estimation performance for different training Fig. 2. CFO estimation performance for different Q with
sequences with Nt = 3 and Nr = 2. Nt = 3 and Nr = 2.
can be calculated straight-forwardly [20] as follows
CRBε =
2
N σw
H
, (26)
8π 2 hH X B[INr N − X (X H X )−1 X H ]BX h
where
X = INr ⊗ S,
B = INr ⊗ diag{[Ng , Ng + 1, · · · , Ng + N − 1]T }.
In order to verify our analysis concerning training sequence

design with the Chu sequence, we construct the random
sequences (RS) by generating the Nt P pilots randomly, and
compare the MSE performances of RPBE with TS 0, TS 1
and RS. Also included for comparison are the performances
of the CFO estimators with their training sequences in [9] Fig. 3. CFO estimation performance for different numbers
and [10]. In Fig. 1, we present the corresponding simulation of transmit antennas with Nr = 2.
results with Nt = 3 and Nr = 2. It can be observed that
the performance of RPBE with TS 0 is slightly better than
that with TS 1, which coincides with the analytical results Studying the curves in Fig. 2, we can see that the performance
concerning training sequence design in the previous section. of RPBE with the same training sequences degrades with
It can also be observed that the performances of RPBE with smaller Q provided that Q > Nt and P ≥ L, which agrees
TS 0 and TS 1 are better than that with RS, which should well with the analysis in [21]. We can also see that a smaller
be attributed to the good correlation property of TS 0 and TS Q yields a larger average CRB, and the MSE performance
1. Actually, the performance improvements are more evident follows the same trend as the average CRB.
when the frequency selective fading channels with large delay In Fig. 3, we illustrate how the number of transmit antennas
spreads are considered. Furthermore, we observe that the Nt affects the performance of RPBE. It can be observed that
performances of RPBE with TS 0 and TS 1 are far better than the MSE performance for TS 1 deteriorates with increased Nt
that of the CFO estimator in [9], and almost the same as that due to the influence of MAI, whereas the MSE performance
of the CFO estimator in [10]. Noting that the overhead of the for TS 0 degrades slightly, which should be ascribed to the
training sequences presented in [10] grows linearly with the better correlation property of TS 0. Another observation is
number of transmit antennas, we can find certain advantages that the number of transmit antennas has little impact on the
of the proposed training sequences. We also observe that the average CRB.
average CRB for TS 0 is slightly smaller than that for TS 1. Fig. 4 shows how the number of receive antennas Nr affects
Since the average CRBs for TS 0 and TS 1 are very close, the performance of RPBE. We can observe that the MSE
only one curve is plotted subsequently. performance of RPBE is substantially improved for both TS 0
Depicted in Fig. 2 is the MSE performance of RPBE as a and TS 1 by increasing Nr . We have the similar observation
function of SNR for different Q with Nt = 3 and Nr = 2. for the average CRB. These observations imply that it is more
Then, D(Δ) can be expressed as

D(Δ) = dH (Δ)d(Δ), (28)
where
d(Δ) = [dT0 (Δ), dT1 (Δ), · · · , dTν (Δ), · · · , dTNr −1 (Δ)]T ,
dν (Δ) = C S̆ F̆ hν ,
⎡ (0,0) ⎤
C ··· C (0,Nt −1)
⎢.. . ⎥
C = ⎣. C (ω,μ) .. ⎦,
C (Q−Nt −1,0) · · · C (Q−Nt −1,Nt −1)
√
C (ω,μ) = N ΘTzω FN DN (Δ)FNH Θiμ .
Correspondingly, we have
D(Δ) ≥ 0 & D(0) = 0 &

Fig. 4. CFO estimation performance for different numbers D(Δ) = 0, iff. d(Δ) = 0Nr (N −Nt P ) . (29)
of receive antennas with Nt = 3. Therefore, for any h(ν,μ) (= 0L ), proving that ε̃ = ε is the
unique value to maximize P(Δ) with ε̃, ε ∈ (−Q/2, Q −
Q/2] is equivalent to proving that d(Δ) = 0Nr (N −Nt P ) for
efficient for performance improvement of RPBE to increase any Δ ∈ (−Q, 0) ∪ (0, Q).
the number of receive antennas than the number of transmit
antennas. Consider the following two cases:
1) When Δ is not an integer: We establish immediately
VI. C ONCLUSIONS from the definition of C (ω,μ) that it is a column-wise circulant
matrix. Denote the square matrix formed by the first Nt P rows
In this paper, we have presented a training sequence assisted
of C by C P (recall that (N − Nt P ) ≥ Nt P in our design
CFO estimator for MIMO OFDM systems by exploiting the
condition). Then, by exploiting the relation between circulant
properties of the training sequences and an efficient geometric
matrix and DFT matrix [25], C P can be decomposed as
mapping. We have developed the required conditions for √
the training sequences to yield estimation identifiability, and C P = P (INt ⊗ FPH )Λ(INt ⊗ FP ), (30)
proposed sub-optimal training sequences constructed from the
Chu sequence. Our proposed estimator and training sequences where
⎡ (0,0) ⎤
yield better or similar estimation performance with much Λ ··· Λ(0,Nt −1)
smaller training overhead than existing methods for MIMO- ⎢.. . ⎥
Λ = ⎣. Λ(μ,μ ) .. ⎦,
OFDM systems.
Λ(Nt −1,0) · · · Λ(Nt −1,Nt −1)

Λ(μ,μ ) = diag{c̃(μ,μ ) },
A PPENDIX I
c̃(μ,μ ) = FP C (μ,μ ) e0P .
This appendix presents the proof of Theorem 1. Substitute
(2) into the cost function in (4) and ignore the noise item. The element of c̃(μ,μ ) can be obtained from its definition as
Define Δ = ε − ε̃. Then, the cost function in (4) can be follows
expressed as 1
[c̃(μ,μ ) ]p = √ c(μ,μ ) (1 − ej2πΔ )ej2π(iμ −zμ +Δ)(−p)P /N
Q
P(Δ) = = 0, 0 ≤ p ≤ P − 1, (31)
N hH {INr ⊗ [S H DN (−Δ)F̄ H F̄ DN (Δ)S]}h. (27)
where
Let
c(μ,μ ) = aμ /(aμ − bμ ),
S̆ = diag{[s̃T0 , s̃T1 , · · · , s̃Tμ , · · · , s̃TNt −1 ]T }.
aμ = ej2πzμ /Q ,
Define
bμ = ej2π(iμ +Δ)/Q .
H H H
D(Δ) = N h [INr ⊗ (F̆ S̆ S̆ F̆ )]h − P(Δ).
It follows from (32) that rank{Λ(μ,μ ) } = P . Let
Then, it follows immediately that the maximum of P(Δ)
Λp = [INt ⊗ (epP )T ]Λ[INt ⊗ (epP )T ]T ,
corresponds to the minimum of D(Δ) with respect to Δ.
Suppose which denotes the Nt × Nt sub-matrix of Λ. Then, we can
decompose Λp as follows
0 ≤ ω < Q − Nt & 0 ≤ zω < Q & zω = iμ &
1
zω = zω , iff. ω = ω . Λp = √ (1 − ej2πΔ )Λap CNt Λbp , (32)
Q
where together with condition (C0), which implies iμ = iμ iff. μ =

−(−p)P /P −(−p) /P μ , we have
Λap = diag{a0 , · · · , aNt −1 P }, √
(−p) /P (−p) /P C (ω,μ) = N IP & C (ω,μ ) = 0P ×P , if μ = μ. (39)
Λbp = diag{b0 P , · · · , bNt −1P },
Hence,
and CNt is the Nt × Nt square matrix with its element given

by [CNt ]μ,μ = c(μ,μ ) . From the definitions of aμ and bμ , d(ν,ω) (Δ) =
we have √
N diag{s̃μ }ΘTiμ FN [IL , 0L×(N −L) ]T h(ν,μ) . (40)
aμ = 0 & aμ = bμ &
With the condition P ≥ L, we immediately obtain
aμ = aμ , bμ = bμ , aμ = bμ , ∀μ = μ . (33) d(ν,ω) (Δ) = 0P and then d(Δ) = 0Nr (N −Nt P ) for any
Moreover, we also have (34) as shown at the bottom of the h(ν,μ) (= 0L ).
page. Then, with the assumption that Nt and Q are not very Combining the above two cases, we draw the conclusion
large (for example Nt ≤ 8, Q = 16), the determinant of CNt that d(Δ) = 0Nr (N −Nt P ) for any h(ν,μ) (= 0L ) and any Δ ∈
can be obtained with its definition as follows (−Q, 0) ∪ (0, Q). This completes the proof.
t −1 μ−1
N Nt −1
(bμ − bμ )(aμ − aμ )aμ
det{CNt } =
μ=1 μ =0 μ =0
(aμ − bμ )(aμ − bμ )(aμ − bμ )
A PPENDIX II
= 0, (35)
which shows that rank{CNt } = Nt . From (32), we immedi- In this appendix, we will discuss the uniqueness of ε̂i = εi
ately obtain that rank{Λp } = Nt . With the special diagonal in maximizing the estimation metric for any h(ν,μ) (= 0L ). For
structure of Λ, we establish that rank{Λ} = Nt P and then the case that ε = εi , it follows immediately from Theorem 1
rank{C} = Nt P . Exploiting the following relationship that the uniqueness of ε̂i = εi in maximizing the estimation
metric for any h(ν,μ) (= 0L ) is guaranteed. Therefore, in the
rank{A0 } + rank{A1 } − n ≤ rank{A0 A1 } following, we only consider the case that ε = εi , i.e., εf = 0.
≤ min{rank{A0 }, rank{A1 }}, (36) Define Δi = ε − ε̃i . Then, the estimation metric P(Δi )
can be written into an equivalent form as shown in (41) at the
where A0 has n columns and A1 has n rows, we further
bottom of the next page. In (41),
establish that rank{C S̆ F̆ } = Nt L for any non-integer Δ

,μ )
(recall that P ≥ L in our design condition). Hence, for any M (μ,μ (Δi ) = SμH Gμ (Δi )Sμ ,
h(ν,μ) (= 0L ), we immediately obtain from its definition that Gμ (Δi ) = DN (−Δi )FNH Θiμ ΘTiμ FN DN (Δi ),
dν (Δ) = 0N −Nt P and then d(Δ) = 0Nr (N −Nt P ) . √
Sμ = N FNH Θiμ diag{s̃μ }ΘTiμ FN [IL , 0L×(N −L) ]T .
2) When Δ is an integer: By exploiting the structure of C,
From the definition of Gμ (Δi ), we can obtain
dν (Δ) can be transformed into the following equivalent form
Gμ (Δi ) =
dν (Δ) = [(d(ν,0) (Δ))T , (d(ν,1) (Δ))T , ⎡ ⎤
(0,0) (0,Q−1)
· · · , (d (ν,ω) T
(Δ)) , · · · , (d(ν,Q−Nt −1) T T
(Δ)) ] , Gμ (Δi ) ···
Gμ (Δi )
(37) ⎢ . ⎥
⎢ . (q,q ) . ⎥,
⎣ . Gμ (Δi ) .. ⎦
where (Q−1,0) (Q−1,Q−1)
Gμ (Δi ) · · · Gμ (Δi )
d(ν,ω) (Δ) = (42)
Nt −1
where
C (ω,μ) diag{s̃μ }ΘTiμ FN [IL , 0L×(N −L) ]T h(ν,μ) .
μ=0 G(q,q
μ
)
(Δi ) = Q−1 ej2π(q−q )(iμ −Δi )/Q · IP .
For any integer Δ ∈ (−Q, 0) ∪ (0, Q), with our design Define
1
condition we have sμ = √ FPH s̃μ .
Q
(1Q − l)T l(Δ) > 0. (38)
Then, we can also obtain that Sμ is an N × L column-wise
Without loss of generality, suppose (iμ + Δ)Q = zω . Then, circulant matrix with [DQ (iμ P )1Q ] ⊗ [DP (iμ )sμ ] being its
t −1 μ−1
N t −1 μ−1
N
(bμ − bμ )(aμ − aμ ) = (bμ − bμ )
μ=1 μ =0 μ=2 μ =1
N
t −1 N
t −1 μ−1
N
t −1 N
t −1

× (−1)μ · (aμ − aμ ) · (aμ − bμ ) · (aμ − b0 ) . (34)
μ =0 μ=1,μ=μ μ =0,μ =μ μ=1 μ=0,μ=μ

first column vector. Hence, we have Therefore, [M (μ,μ ,μ) (Δi )]l,l achieves its minimum 0 when

r(Δi ) ∈ {−Q/2, −Q/2+1, · · · , Q−Q/2−1}\{r(iμ −
[M (μ,μ ,μ ) (Δi )]l,l = α(μ, μ ; l, l )/Q and its maximum N/Nt when r(Δi ) = r(iμ −iμ ). While
iμ )} (μ,μ
Q−1
Q−q−1 Q−1
[M
,μ )
(Δi )]l,l achieves its minimum 0 when r(Δi ) ∈
jq ψ jqθ j(q+q )ψ −jqθ jqψ
× e e +e e − e , {−Q/2, −Q/2 + 1, · · · , Q − Q/2 − 1}.

q=0 q =0 q=0 Remark 2: Both [M (μ,μ ,μ) (Δi )]l,l and [M (μ,μ ,μ ) (Δi )]l,l
(43) are the continuous functions with respect to Δi . For any γ > 0
where (for example γ = 10−2 ), there exists δ > 0 (for example
(l )
δ = 0.1) which makes the relationships as shown in (49) and
α(μ, μ ; l, l ) = ej2π(liμ −l iμ )/N · (s(l) H
μ ) DP (iμ − iμ )sμ , (50) at the bottom of the next page hold.
θ = 2π(iμ − iμ − Δi )/Q, Remark 3: Assume 0 < |εf | < 0.5. Then, for r(εi − ε̃i ) ∈
ψ = 2π(iμ − iμ )/Q. {−Q/2, −Q/2 + 1, · · · , Q − Q/2 − 1}, we have (51)
as shown at the bottom of the next page.

Remark 4: ζ (μ,μ ,μ ) (Δi ) is the period-Q continuous func-
When μ = μ , imposing the condition that the elements tion with respect to Δi , which achieves its minimum 0 when
of s̃μ have constant amplitude, which translates to the time- r(Δi ) = r(iμ − iμ ) and its maximum +∞ when r(Δi ) =
domain zero auto-correlation (ZAC) property of sμ [15], from r(iμ − iμ ). Impose the condition that |α(μ, μ ; l, l )| < 1/Nt
(43) we immediately obtain (44) as shown at the bottom of for μ = μ . Then, there exists χ > 0 (for example χ = 5)
the page. When μ = μ , from (43) we obtain (45) as shown which makes the relationship as shown in (52) at the bottom
at the bottom of the page. Define of the next page hold.

We use an example as shown in Fig. 5 to illustrate the above
ζ (μ,μ ,μ ) (Δi ) = [M (μ,μ ,μ) (Δi )]l,l /[M (μ,μ ,μ ) (Δi )]l,l .
remarks.
Then, we immediately establish Assume

,μ )

ζ (μ,μ ,μ ) (Δi ) = P/[Nt · |α(μ, μ ; l, l )|] E{[h(ν,μ) ]∗l [h(ν ]l } = 0, ∀(ν, μ) = (ν , μ ).
× | sin(ψ/2)| · | cot(ψ/2) − cot(θ/2)|. (46) Then, it follows from Remark 2 and 4 that (53) and (54) as
shown at the bottom of the page can be obtained, respectively.
The following remarks are now in order.
According to its definition, [M (μ,μ ,μ) (Δi )]l,l must be one of
Remark 1: Both [M (μ,μ ,μ) (Δi )]l,l and [M (μ,μ ,μ ) (Δi )]l,l
the Q possible values whatever iμ or iμ is when we vary
are the period-Q functions with respect to Δi . Fix the true
ε̃i from (−Q/2 + 1) to (Q − Q/2). From Remark 3, we
CFO ε and vary the candidate ICFO ε̃i between (−Q/2, Q−
immediately obtain that
Q/2]. Then, we have

Δi ∈ [ε − Q + Q/2, ε + Q/2) ⊂ (−Q, Q). (47) [M (μ,μ ,μ) (εf )]l,l [M (μ,μ ,μ) (εf + (iμ − iμ )Q )]l,l 1,
For description convenience, we employ r(Δi ) to normalize if (iμ − iμ )Q > 1, (55)

Δi from [ε−Q+Q/2, ε+Q/2) to [−Q/2, Q−Q/2), and the items involved in the right hand side of (54) achieve
where their maximum when r(εi − ε̃i ) = r(iμ − iμ ) for any
h(ν,μ) (= 0L ). Then, we can establish that there are Nr Nt
r(Δi ) = Δi − [sign(Δi − Q + Q/2)
items involved in the right hand side of (54) which achieve
the
+ sign(Δi + Q/2)]Q/2. (48) maximum when ε̃i = εi (i.e., μ = μ ), and at most Nr Nt −
N
r −1 N
t −1 N
t −1

P(Δi ) = (h(ν,μ) )H M (μ,μ ,μ) (Δi )h(ν,μ)
ν=0 μ=0 μ =0
r −1 N
N t −1 N
t −1 N
t −1

+ (h(ν,μ) )H M (μ,μ ,μ ) (Δi )h(ν,μ ) , (41)
ν=0 μ=0 μ =0 μ =0,μ =μ

P 2P
M (μ,μ ,μ) (Δi ) = + [(Q − 1)ejθ − Qej2θ + ej(Q+1)θ ]/(1 − ejθ )2 ·IL
Nt Nt Q
P 1 − cos(Qθ)
= · · IL . (44)
Nt Q 1 − cos θ

[M (μ,μ ,μ ) (Δi )]l,l = α(μ, μ ; l, l )/Q · (1 − ejQθ )2 /[(1 − ejθ )(ejψ − ejθ )ej(Q−1)θ ]
1 − cos(Qθ)
= α(μ, μ ; l, l )/Q · . (45)
cos(ψ/2) − cos(θ − ψ/2)

min (1Q − l)T l(q) items that achieve the maximum
1≤q≤Q−1
when ε̃i = εi (i.e., μ = μ ). Due to the random snapshot
channel energies, a closed form of training design conditions
to yield the uniqueness of ε̂i = εi is intractable. However, it
follows from condition (C3) and the above analysis that the
reliability of the uniqueness can be maximized
by designing

the training sequences to maximize min (iμ − iμ )Q and
μ =μ
T (q)

min (1Q − l) l .
1≤q≤Q−1
ACKNOWLEDGMENT
The authors would like to thank the anonymous reviewers
for their valuable comments which helped to improve the
quality of the paper.
R EFERENCES
[1] G. L. Stuber, J. R. Barry, S. Mclaughlin, Y. Li, M. A. Ingram, and
[M (μ,μ ,μ) (Δi )]l,l , [M (μ,μ ,μ ) (Δi )]l,l and

Fig. 5. T. G. Pratt, “Broadband MIMO-OFDM wireless communications,” Proc.

ζ (μ,μ ,μ ) (Δi ) versus r(Δi ) with P = 64, Q = 16, Nt = 3, IEEE, vol. 92, no. 2, pp. 271–294, Feb. 2004.
[2] P. Moose, “A technique for orthogonal frequency division multiplexing
iμ = 3, iμ = 7, iμ = 4. frequency offset correction,” IEEE Trans. Commun., vol. 42, pp. 2908–
2914, Oct. 1994.
[3] J. van de Beek, M. Sandell, and P. O. Borjesson, “ML estimation of time
and frequency offset in OFDM systems,” IEEE Trans. Signal Processing,
vol. 45, pp. 1800–1805, July 1997.
Q−Q/2

[M (μ,μ ,μ) (Δi )]l,l < γ, if r(Δi ) ∈ (q − δ, q) ∪ (q, q + δ) \ (r(iμ − iμ ) − δ, r(iμ − iμ ) + δ)
q=−Q/2

\ (−Q/2 − δ, −Q/2) ∪ (Q − Q/2, Q − Q/2 + δ) , (49)
Q−Q/2

[M (μ,μ ,μ ) (Δi )]l,l < γ, if r(Δi ) ∈ (q − δ, q) ∪ (q, q + δ)
q=−Q/2

\ (−Q/2 − δ, −Q/2) ∪ (Q − Q/2, Q − Q/2 + δ) . (50)

[M (μ,μ ,μ) (Δi |r(εi − ε̃i ) = r(iμ − iμ ))]l,l [M (μ,μ ,μ) (Δi |r(εi − ε̃i ) = r(iμ − iμ ))]l,l

= sin2 {π[εf + r(εi − ε̃i ) − r(iμ − iμ )]/Q}/sin2 (πεf /Q)r(εi −ε̃i )=r(i > 1. (51)
μ −iμ )

ζ (μ,μ ,μ ) (Δi ) > P · | sin(ψ/2)| · | cot(ψ/2) − cot(ψ/2 + πδ/Q)| > χ,

if r(Δi ) ∈ (−Q/2, r(iμ − iμ ) − δ) ∪ (r(iμ − iμ ) + δ, Q − Q/2) . (52)
Q−Q/2
. N
t −1

P(Δi ) = 0, if r(Δi ) ∈ (q − δ, q) ∪ (q, q + δ) \ (r(iμ − iμ ) − δ, r(iμ − iμ ) + δ)
q=−Q/2 μ,μ =0

\ (−Q/2 − δ, −Q/2) ∪ (Q − Q/2, Q − Q/2 + δ) , (53)
Nr −1 Nt −1 Nt −1
.
P(Δi ) = (h(ν,μ) )H M (μ,μ ,μ) (Δi )h(ν,μ) ,
ν=0 μ=0 μ =0
Q−Q/2−1
N
t −1

if r(Δi ) ∈ [q + δ, q + 1 − δ] (r(iμ − iμ ) − δ, r(iμ − iμ ) + δ) . (54)
q=−Q/2 μ,μ =0
[4] U. Tureli, H. Liu, and M. D. Zoltowski, “OFDM blind carrier offset [28] J. Coon, M. Beach, and J. McGeehan, “Optimal training sequences for
estimation: ESPRIT,” IEEE Trans. Commun., vol. 48, pp. 1459–1461, channel estimation in cyclic-prefix-based single-carrier systems with
Sept. 2000. transmit diversity,” IEEE Signal Processing Lett., vol. 11, no. 9, pp.
[5] M. Morelli and U. Mengali, “An improved frequency offset estimator 729–732, Sept. 2004.
for OFDM applications,” IEEE Commun. Lett., vol. 3, pp. 75–77, Mar. [29] F. Gini and R. Reggiannini, “On the use of Cramer-Rao-like bounds in
1999. the presence of random nuisance parameters,” IEEE Trans. Commun.,
[6] D. Huang and K. B. Letaief, “Carrier frequency offset estimation for vol. 48, no. 12, pp. 2120–2126, Dec. 2000.
OFDM systems using null subcarriers,” IEEE Trans. Commun., vol. 54,
no. 5, pp. 813–823, May 2006. Yanxiang Jiang (S’03-M’08) received the B.S. degree in electrical engineer-
[7] ——, “Enhanced carrier frequency offset estimation for OFDM using ing from Nanjing University, Nanjing, China, in 1999, the M.E. degree and
channel side information,” IEEE Trans. Wireless Commun., vol. 5, the Ph.D. degree in electrical engineering from Southeast University, Nanjing,
no. 10, pp. 2784–2793, Oct. 2006. China, in 2003 and 2007, respectively.
[8] Y. Yao and G. B. Giannakis, “Blind carrier frequency offset estimation He received the NJU excellent graduate honor from Nanjing University in
in SISO, MIMO and multiuser OFDM systems,” IEEE Trans. Commun., 1999. His research interests include mobile communications, wireless signal
vol. 53, no. 1, pp. 173–183, Jan. 2005. processing, parameter estimation, synchronization, signal design, and digital
[9] A. N. Mody and G. L. Stuber, “Synchronization for MIMO OFDM implementation of communication systems.
systems,” in Proc. IEEE Globecom’01, vol. 1, Nov. 2001, pp. 509–513.
[10] A. van Zelst and T. C. W. Schenk, “Implementation of a MIMO OFDM Hlaing Minn (S’99-M’01-SM’07) received his B.E. degree in Electronics
based wireless LAN system,” IEEE Trans. Signal Processing, vol. 52, from Yangon Institute of Technology, Yangon, Myanmar, in 1995, M.Eng.
pp. 483–494, Feb. 2004. degree in Telecommunications from Asian Institute of Technology (AIT),
[11] O. Besson and P. Stoica, “On parameter estimation of MIMO flat- Pathumthani, Thailand, in 1997 and Ph.D. degree in Electrical Engineering
fading channels with frequency offsets,” IEEE Trans. Signal Processing, from the University of Victoria, Victoria, BC, Canada, in 2001.
vol. 51, no. 3, pp. 602–613, Mar. 2003. He was with the Telecommunications Program in AIT as a laboratory
[12] X. Ma, M. K. Oh, G. B. Giannakis, and D. J. Park, “Hopping pilots supervisor during 1998. He was a research assistant from 1999 to 2001 and a
for estimation of frequency-offset and multi-antenna channels in MIMO post-doctoral research fellow during 2002 in the Department of Electrical and
OFDM,” IEEE Trans. Commun., vol. 53, no. 1, pp. 162–172, Jan. 2005. Computer Engineering at the University of Victoria. Since September 2002, he
[13] F. Simoens and M. Moeneclaey, “Reduced complexity data-aided and has been with the Erik Jonsson School of Engineering and Computer Science,
code-aided frequency offset estimation for flat-fading MIMO channels,” the University of Texas at Dallas, USA, as an Assistant Professor. His research
IEEE Trans. Wireless Commun., vol. 5, no. 6, pp. 1558–1567, June 2006. interests include wireless communications, statistical signal processing, error
[14] P. Stoica and O. Besson, “Training sequence design for frequency offset control, detection, estimation, synchronization, signal design, and cross-layer
and frequency-selective channel estimation,” IEEE Trans. Commun., design. He is an Editor for the IEEE Transactions on Communications.
vol. 51, no. 11, pp. 1910–1917, Nov. 2003.
[15] H. Minn, X. Fu, and V. K. Bhargava, “Optimal periodic training signal Xiqi Gao (SM’07) received the Ph.D. degree in electrical engineering from
for frequency offset estimation in frequency-selective fading channels,” Southeast University, Nanjing, China, in 1997. He joined the Department
IEEE Trans. Commun., vol. 54, no. 6, pp. 1081–1096, June 2006. of Radio Engineering, Southeast University, in April 1992. Now he is a
[16] M. Ghogho and A. Swami, “Training design for multipath channel and professor of information systems and communications. From September 1999
frequency offset estimation in MIMO systems,” IEEE Trans. Signal to August 2000, he was a visiting scholar at Massachusetts Institute of
Processing, vol. 54, no. 10, pp. 3957–3965, Oct. 2006. Technology, Cambridge, and Boston University, Boston, MA. His current
[17] H. Minn, N. Al-Dhahir, and Y. Li, “Optimal training signals for MIMO research interests include broadband multi-carrier transmission for beyond
OFDM channel estimation in the presence of frequency offset and phase 3G mobile communications, space-time wireless communications, iterative
noise,” IEEE Trans. Commun., vol. 54, no. 10, pp. 1754–1759, Oct. detection/decoding, signal processing for wireless communications.
2006. Dr. Gao received the Science and Technology Progress Awards of the State
[18] H. Minn and N. Al-Dhahir, “Optimal training signals for MIMO OFDM Education Ministry of China in 1998 and 2006. He is currently serving as an
channel estimation,” IEEE Trans. Wireless Commun., vol. 5, no. 5, pp. editor for the IEEE Transactions on Wireless Communications.
1158–1168, May 2006.
[19] D. Chu, “Polyphase codes with good periodic correlation properties,”
Xiaohu You (M’91) received the M.S. and Ph.D. degrees in electrical
IEEE Trans. Inform. Theory, vol. 18, pp. 531–532, July 1972.
engineering from Southeast University, Nanjing, China, in 1985 and 1988,
[20] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation
respectively.
Theory. Prentical-Hall, 1993.
Since 1990 he has been working with National Mobile Communications
[21] P. Stoica and A. Nehorai, “MUSIC, maximum likelihood, and Cramer-
Research Laboratory at Southeast University, where he holds the ranks of
Rao bound,” IEEE Trans. Acoustics, Speech, Signal Processing, vol. 5,
professor and director. From 1993 to 1997 he was engaged, as a team leader,
pp. 720–741, May 1989.
in the development of China’s first GSM and CDMA trial systems. He was the
[22] M. Pesavento, A. B. Gershman, and M. Haardt, “Unitary root-MUSIC
Premier Foundation Investigator of the China National Science Foundation
with a real-valued eigendecomposition: a theroretical and expreimental
in 1998. From 1999 to 2001 he was on leave from Southeast University,
performance study,” IEEE Trans. Signal Processing, vol. 48, no. 5, pp.
working as the chief director of China’s 3G (C3G) Mobile Communications
1306–1314, May 2000.
R&D Project. He is currently responsible for organizing China’s B3G R&D
[23] Y. X. Jiang, X. Q. Gao, X. H. You, and W. Heng, “Training sequence
activities under the umbrella of the National 863 High-Tech Program, and
assisted frequency offset estimation for MIMO OFDM,” in Proc. IEEE
he is also the chairman of the China 863-FuTURE Expert Committee.
ICC’06, vol. 12, June 2006, pp. 5371–5376.
He has published two books and over 20 IEEE journal papers in related
[24] W. H. Press, Numerical Recipes in C++: the Art of Scientific Computing.
areas. His research interests include mobile communications, advanced signal
Cambridge University Press, 2002.
processing, and applications.
[25] G. H. Golub and C. F. Van Loan, Matrix Computations. The John
Hopkins University Press, 1996.
[26] F. Gao and A. Nallanathan, “Blind maximum likelihood CFO estimation Yinghui Li (S’05) received the B.E. and M.S. degree in Electrical Engineering
for OFDM systems via polynomial rooting,” IEEE Signal Processing from the Nanjing University of Aeronautics and Astronautics, Nanjing, China,
Lett., vol. 13, no. 2, pp. 73–76, Feb. 2006. in 2000 and 2003, respectively, and the Ph.D. degree in Electrical Engineering
[27] W. H. Mow, “A new unified construction of perfect root-of-unity se- from the University of Texas at Dallas in 2007. Her research interests are in
quences,” in Proc. IEEE Spread-Spectrum Techniques and Applications, the applications of statistical signal processing in synchronization, channel
vol. 3, Sept. 1996, pp. 955–959. estimation and detection problems in broadband wireless communications.

Frequency Offset Estimation and

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Frequency Offset Estimation and

Uploaded by

Copyright:

Available Formats

1244 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO.