IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO.
6, JUNE 2008
MIMO-OFDM Channel Estimation in the
Presence of Frequency Offsets
Zhongshan Zhang, Member, IEEE, Wei Zhang, Student Member, IEEE,
and Chintha Tellambura, Senior Member, IEEE
Abstract—Optimal pilot design and placement for channel
estimation in Multiple-input Multiple-output (MIMO) Orthog-
onal Frequency-Division Multiplexing (OFDM) systems with
frequency offsets are considered. Both the single-frequency-
offset case and the multiple-frequency-offset case are treated.
We show that the Constant-Envelope (CE) condition is sufficient
but not necessary for pilot design, and that pilots with multiple
envelopes can also achieve the optimal performance in terms
of the Mean Square Error (MSE) minimization, provided that
an additional constraint on the pilot placement is satisfied
simultaneously. New pilot designs, which take into account the
multiple-frequency-offset case, are proposed to eliminate Inter-
Pilot-Interference (IPI) and to optimize the MSE performance.
The Least-Squares (LS) and Linear Minimum Mean Square
Error (LMMSE) channel estimators for the multiple-frequency-
offset case are designed for uncorrelated and correlated MIMO-
OFDM channels, respectively. The LMMSE estimator requires
the channel covariance matrix. Both optimal adaptive pilot power
allocation and suboptimal uniform pilot power allocation are
developed for the proposed LMMSE estimator. The adaptive
allocation performs 4 dB better than the uniform allocation in the
high noise region, but they both perform identically in the low
noise region. Performance comparisons are made against several
previous pilot designs due to [1], [2]. The proposed LMMSE
estimator significantly outperforms the LS estimator.
Index Terms—Frequency offset, channel estimation, MIMO,
OFDM.
I. I NTRODUCTION
C ONSIDERABLE research has focused on Multiple-input
Multiple-output (MIMO) technology for increasing the
wireless system capacity. Compared to a single-input single-
output (SISO) system, a MIMO system can improve the
capacity by a factor of the minimum number of transmit
and receive antennas. Space-time coding, including space-time
block codes (STBC) and space-time trellis codes (STTC) [3],
can extract transmit diversity in MIMO systems.
The frequency-selective fading MIMO channel can be trans-
formed into a set of flat-fading MIMO channels by using
Orthogonal Frequency-Division Multiplexing (OFDM). This
transformation achieves a high capacity at a low cost of
equalization and demodulation [4], [5]. However, just as with
the SISO-OFDM, MIMO-OFDM systems too are sensitive
to frequency offset. Many SISO-OFDM frequency offset
estimators have been proposed [6]–[8]. A synchronization
Manuscript received January 14, 2007; revised July 17, 2007; accepted
September 16, 2007. The associate editor coordinating the review of this
letter and approving it for publication was C. Xiao.
The authors are with the Department of Electrical and Computer Engi-
neering, University of Alberta, Edmonton, AB T6G 2V4, Canada (e-mail:
{zszhang, wzhang, chintha}@ece.ualberta.ca). The corresponding author is
Dr. Chintha Tellambura.
Digital Object Identifier 10.1109/TWC.2008.070044.
1536-1276/08$25.00 
c 2008 IEEE
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 09:15 from IEEE Xplore. Restrictions apply.
2329
algorithm for MIMO-OFDM systems is proposed in [9], where
identical timing offset and frequency offset with respect to
each transmit-receive antenna pair are assumed. Parameters
estimation of MIMO flat-fading channels is discussed in [10],
where frequency offsets for different transmit-receive antennas
are assumed to be different, and the Cramer-Rao Lower Bound
(CRLB) for either the frequency offsets or channel estimation
variance error is derived.
Another prevalent impairment is the channel estimation
error, which degrades the bit error rate of MIMO-OFDM
systems. Robust channel estimation for OFDM systems is
discussed in [11]. Optimal pilot design and placement for
channel estimation is developed in terms of minimization of
the CRLB [12]. Optimal training signal design for frequency-
selective block fading channel estimation in MIMO-OFDM
systems is discussed in [13]. Some other constraints for train-
ing signal designs, such as low peak-to-average energy ratio
(PAR) and robustness to frequency offsets, are also considered
in [13]. Barhumi, Leus and Moonen (BLM) propose a high-
quality channel estimator for a MIMO-OFDM channel based
on frequency-domain uniformly placed pilots (the frequency
offset and channel correlation are not considered) [1]. Hu
improves the BLM pilots with a nonuniform placement de-
sign, which mitigates the performance loss due to the non-
modulated pilots in virtual subcarriers [14]. MIMO-OFDM
channel estimation for a correlated channel is discussed in
[15], [16]. However, neither [15] nor [16] considers the effect
of frequency offset. Training sequence design for MIMO chan-
nel estimation in the presence of a single frequency offset is
discussed in [17], [18]. In [17], optimal training signal design
for MIMO-OFDM channel estimation while considering a
single frequency offset and phase noise is proposed, and its
performance is much more robust than the algorithms that
consider only the effect of frequency offset. In [18], the
estimation of frequency-selective channel and frequency offset
in either SISO or MIMO systems is discussed. By using the
exact CRLB as a metric, a power efficient training preamble is
designed in [18] to reduce the complexity of estimation. Joint
frequency offset and channel estimation with either single
or multiple frequency offsets for MIMO frequency selective
fading channels is discussed in [2], where pilots for different
transmit antennas are orthogonal in the time-domain, and pilot
optimization in terms of MSE reduction is also studied.
This paper considers optimal pilot design and placement for
channel estimation in Multiple-input Multiple-output (MIMO)
Orthogonal Frequency-Division Multiplexing (OFDM) sys-
tems in the presence of frequency offsets. Both the single-
frequency-offset case and the multiple-frequency-offset case
 2330 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 6, JUNE 2008

are treated1 . Each frequency offset is modeled as an Indepen- Nt transmit antennas and Nr receive antennas, a N × 1
dent and Identically Distributed (i.i.d.) random variable (RV). vector xi (z) is used to represent the z-th block of frequency-
In our proposed scheme, several subcarriers are allocated to domain symbols sent by the i-th transmit antenna, where
each transmit antenna to transmit either pilots or data symbols. i ∈ {1, 2, · · · , Nt }. In the following sections, the temporal
The pilots of different transmit antennas may be either orthog- index z will be omitted for the sake of simplicity. Without
onal (in the frequency-domain or time-domain) or not. We loss of generality, each entry of xi is assumed to be an
show that the Constant-Envelope (CE) condition is sufficient Es
i.i.d. RV with mean zero and variance σx2 = , where
but not necessary for pilot design, and that pilots with multiple N Nt
envelopes can also achieve the optimal performance in terms Nt  
2
of the Mean Square Error (MSE) minimization, provided that Es = E xi 2 is the total transmit power.
an additional constraint on the pilot placement is satisfied i=1
By using hk,i (z) to represent the discrete-time impulse
simultaneously. New pilot designs, which take into account response of the z-th tap channel between the i-th transmit
the multiple-frequency-offset case, are proposed to eliminate and the k-th receive antenna, the related channel response
Inter-Pilot-Interference (IPI) and to optimize the MSE perfor- vector can be represented as hk,i = [h̃Tk,i , 0TN −Lmax ]T =
mance. The Least-Squares (LS) and Linear Minimum Mean
[hk,i (0), hk,i (1), · · · , hk,i (Lmax − 1), 0TN −Lmax ]T with Lmax
Square Error (LMMSE) channel estimators for the multiple-
representing the maximum length of all channels. The cor-
frequency-offset case are designed for uncorrelated and cor-
responding frequency-domain  channel attenuation matrix
 is
related MIMO-OFDM channels, respectively. The LMMSE (0) (1) (N −1)
estimator requires the channel covariance matrix. Both optimal given by Hk,i = diag Hk,i , Hk,i , · · · , Hk,i with
adaptive pilot power allocation and suboptimal uniform pilot (n) 
Lmax −1
hk,i (d)e− N representing the channel at-
j2πnd
Hk,i =
power allocation are developed for the proposed LMMSE d=0
estimator. Performance comparisons are made against several tenuation at the n-th subcarrier.
previous pilot designs due to [1], [2]. The estimator of [1] is In the following sections, ψk,i and εk,i are used to rep-
found to exhibit high error floors for the multiple-frequency- resent the initial phase and the normalized frequency offset
offset case. The proposed LMMSE estimator significantly (frequency offset normalized to a subcarrier spacing of OFDM
outperforms the LS estimator. symbols) between the i-th transmit and the k-th receive anten-
The remainder of this paper is organized as follows. A nas. Since the initial phase is independent of the frequency off-
MIMO-OFDM system model with multiple frequency offsets sets and channel attenuation and also because the estimation of
is discussed in Section II, and the LS frequency offset and the initial phase is beyond the discussion of this paper, without
channel estimation are analyzed in Section III. The optimal loss of generality, we assume that ψk,i for each (k, i) has been
pilot design and placement for LS channel estimation in the estimated and compensated for and that, therefore, ψk,i = 0.
presence of multiple frequency offsets are discussed in Section By considering the channel attenuations and frequency offsets,
IV, and an LMMSE estimator for a correlated MIMO-OFDM the q-th received vector can be represented as
channel is proposed in Section V. Numerical results are given y(q) = [y1T (q), y2T (q), · · · , yN
T
(q)]T , (1)
in Section VI, followed by conclusions in Section VII2 . r


Nt
where yk (q) = Ek,i FHk,i xi (q) + wk (q), Ek,i =
II. MIMO-OFDM S IGNAL M ODEL  i=1 
j2πεk,i j2πεk,i (N −1)
Input data bits of MIMO-OFDM are mapped to com- diag 1, e N , · · · , e N , and wk (q) is a vector
plex symbols drawn from a typical signal constellation, e.g., of additive white Gaussian noise (AWGN) with wk (q)[i] ∼
phase-shift keying (PSK) or quadrature amplitude modulation CN (0, σw2
).
(QAM). An OFDM symbol is generated by taking the Inverse At the k-th receive antenna, by taking the DFT operation
Discrete Fourier Transform (IDFT) of N input sub-symbols, to the received vector, we obtain
where N is the IDFT size, and the IDFT matrix F is defined
1 j2πnk 
Nt
√
as [F]nk = √ e N for 0 ≤ n, k ≤ N − 1. We also rk (q) = FH yk (q) = N Ecir p
k,i Xi (q)F(Lmax ) h̃k,i
H
N
assume that each OFDM symbol has a useful part of duration i=1  
Pk,i (q) (N ×Lmax )
Ts seconds and a cyclic prefix of length Tg seconds to mitigate
Inter-Symbol-Interference (ISI), where Tg is longer than the 
Nt
√
+ N Ecir
k,i Xi (q)F(Lmax ) h̃k,i + F wk (q) ,
d H H
channel-response duration. For a MIMO-OFDM system with    
i=1
Dk,i (q) (N ×Lmax ) η k (q) (N ×1)
1 This case can occur when, for example, MIMO OFDM is used in a
multiple user scenario. This case can also occur in cooperative relaying, which
(2)
can be seen as virtual MIMO systems. The single frequency offset case is a
special case of this case.
where Ecir k,i = F Ek,i F is a circulant matrix, F(Lmax ) is
H

2 Notation: (·)−1 , (·)T and (·)H are the inverse, transpose and complex
√
the first Lmax rows of F, and Xi (q) = Xdi (q) + Xpi (q) =
conjugate transpose of a matrix. The imaginary unit is j = −1. {x}, diag{xdi (q)} + diag{xpi (q)} with xdi (q) and xpi (q) being some
and {x} are the real and imaginary part of x, respectively. A circularly N × 1 data and pilot vectors, respectively. Note that the
symmetric complex Gaussian RV w with mean m and variance σ2 is denoted
by w ∼ CN (m, σ2 ). A real Gaussian RV x with mean a and variance σ2 averaged power of the data and pilot may be different. With
is denoted by x ∼ N (a, σ2 ). IN is the N × N identity matrix, and ON the pilots modulated over consecutive M symbols (M ≥ 1),
is the N × N all-zero matrix. 0N  is the N × 1 all-zero vector. a[i] is the the received vector becomes
i-th entry of vector a, and a22 = |a[i]|2. [B]mn is the mn-th entry of
T
i
matrix B. (x)n represents the remainder after division of x by n. E{x} and rk = rTk (0), · · · , rTk (M − 1) = Pk hk + Dk hk + η k ,
Var{x} are the mean and variance of x. (3)

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 09:15 from IEEE Xplore. Restrictions apply.
 ZHANG et al.: MIMO-OFDM CHANNEL ESTIMATION IN THE PRESENCE OF FREQUENCY OFFSETS 2331

⎡ ⎤
Pk,1 (0) ... Pk,Nt (0) inversely proportional to the Pilot-to-Noise Ratio (PNR), as
⎢ .. .. .. ⎥ 2
where Pk = ⎣ . . . ⎦, defined by PNR = Ep /σw . Note that we will mainly focus
P (M − 1) . . . P (M − 1) on the channel estimation in the following sections, and the
⎡ k,1 k,Nt ⎤
design of a frequency offset estimator that satisfies (5) is
Dk,1 (0) ... Dk,Nt (0)
⎢ .. .. .. ⎥ beyond the scope of this paper. Many existing algorithms,
Dk = ⎣ . . . ⎦, hk = e.g., that proposed in [20]–[22], can be used to perform the
Dk,1 (M − 1) . . . Dk,Nt (M − 1) frequency offset estimation (although these algorithms were
 T
T
h̃k,1 , · · · , h̃Tk,Nt
T
and η k = ηTk (0), · · · , η Tk (M − 1) . originally designed for OFDMA systems, they can be easily
tailored to MIMO-OFDM with each user in OFDMA being
virtually seen as a transmit antenna in MIMO-OFDM). In
III. LS F REQUENCY O FFSET AND C HANNEL E STIMATION Section IV, when the variance error of the frequency offset
IN MIMO-OFDM S YSTEMS
estimator is smaller than 10−2 (this accuracy is easy to achieve
This section and the next assume that the channel coeffi- in [20]–[22]), a robust channel estimation can be performed
cients for the different transmit and receive antennas indepen- by using the proposed pilots.
dent. Section V considers correlated MIMO-OFDM channels. When performing LS channel estimation, P̂k , the estimate
Using rk , the frequency offsets and channel coefficients are of Pk , should be full column rank, so that M N ≥ Lmax Nt .
jointly estimated as Since the frequency offsets have been estimated with negligi-
   2 ble errors, Pk instead of P̂k is used in channel estimation (the
 
ε̂k,1 , · · · , ε̂k,Nt ; ĥk = arg min rk − P̂k ĥk  , effect of a large frequency offset estimation error on the per-
ε̂k,1 ,··· ,ε̂k,Nt ;ĥk 2
(4) formance of the proposed channel estimation will be analyzed
⎡ ⎤ in the following sections, where P̂k = Pk isconsidered).
−1 By
Γk,1 (0)
... Γk,Nt (0) defining the pseudo-inverse of Pk as P†k = PH P PHk ,
√ ⎢ .. .. ⎥
k k
where P̂k = N⎣ . . ⎦, the LS estimation of hk is given by
. . .
Γk,1 (M − 1) . . . Γk,Nt (M − 1)
ĥk|εk,1 ,··· ,εk,Nt = P†k rk = hk + P†k Dk hk + P†k η k , (7)
Γk,i (m) = Êcir p
k,i Xi (m)F(Lmax ) ,
H
and [Êcirk,i ]ls =
j(N −1)((s−l)N +ε̂k,i )
sin (π ((s − l)N + ε̂k,i )) where the subscript of ĥk|εk,1 ,··· ,εk,Nt means the LS estimator
π  · e N , and
N sin N ((s − l)N + ε̂k,i ) is a function of frequency offsets εk,1 , · · · , εk,Nt . The MSE
ε̂k,i represents the estimate of εk,i . of LS estimator ĥk|εk,1 ,··· ,εk,Nt is given by
We perform an estimation in two steps: first, we keep each
channel attenuation unchanged and design a robust frequency    2 
1  
offset estimator whose performance is insensitive to the wire- MSE ĥk|LS = E ĥk|εk,1 ,··· ,εk,Nt − hk 
Lmax Nt 2
less channel; second, based on the frequency offset estimation    
−2
result, Pk can be achieved and used for channel estimation. trace VkH PH k Pk Vk Φk
The frequency offset can be estimated as = (8)
L max Nt
 H −1 
{ε̂k,1 , · · · , ε̂k,Nt } 2
σw trace Pk Pk
 2 +

M−1 Nt
p

 Lmax Nt
,
= arg min yk (n) − Êk,i FHk,i xi (n) ,
ε̂k,1 ,··· ,ε̂k,Nt    
n=0 i=1 2 where Φk = E hk hH k , and Vk = PH k Dk (Vk
(5) represents the power spread of Dk to the signal space
 
j2π ε̂k,i j2π ε̂k,i (N −1)
of Pk ). When in the presence
 of frequency offsets,
where Êk,i = diag 1, e N , · · · , e N . By trans-  −2
trace VkH PH k P k V Φ
k k = 0 is achieved if and only if

Nt
mitting the pilot vector mk (n) = Ek,i FHk,i xpi (n), from the following two conditions are satisfied simultaneously:
i=1 Preposition 1:
[19, page 926], for an unbiased estimator, the CRLB for a 1. xdi (m)xpH
k (q) = 0 for each 1 ≤ i, k ≤ Nt , 0 ≤
variance error of ε̂k,i is given by m, q ≤ M −1; i.e., the subcarrier spaces allocated to data
2
σw symbols are orthogonal to each pilot subcarrier space of
Var {ε̂k,i } ≥   each transmit

M−1
k (n) ∂mk (n)
∂mH  antenna.
−1 H
 ∂εk,i ∂εk,i 2. P†k Dk = PH k Pk Pk Dk = OLmax Nt .
n=0
2
σw The second condition in Preposition 1 can be derived as
= (6) follows:

M−1
2
ΛFHk,i xpi (n)2   −2 
n=0 trace VkH PH k Pk V k Φk = 0
2
3N Nt σw    H
= ,  −1 −1
2π 2 (N
− 1)(2N − 1)Ep ⇔ trace PH
k Pk
1
Vk Φk2 PH
k Pk
1
Vk Φk2 =0
 
2π 2π(N − 1)  −1
where Λ = diag 0, ,··· , , and Ep is the 1
N N ⇔ PH
k Pk Vk Φk2 = OLmax Nt .
total transmit power of the pilots. Evidently, Var {ε̂k,i } is (9)

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 09:15 from IEEE Xplore. Restrictions apply.
 2332 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 6, JUNE 2008

Since both PH k Pk and Φk are non-zero Hermitian matrices, 1 ≤ m, n ≤ Nt , with the ls-th element of Ecir
k,m,n being given
the only solution of (9) is Vk = OLmax Nt , which proves the by
second condition in Preposition 1.
sin (π ((s − l)N + (εk,n − εk,m )))
[Ecir
k,m,n ]ls = π 
N sin N ((s − l)N + (εk,n − εk,m )) (14)
IV. O PTIMAL P ILOT D ESIGN AND P LACEMENT FOR LS j(N −1)((s−l)N +(εk,n −εk,m ))
C HANNEL E STIMATION IN T HE P RESENCE OF M ULTIPLE ×e N .
F REQUENCY O FFSETS
Note that Gk,m,n = GH k,n,m represents the Inter-Pilot-
In the multiple frequency offsets case, define λ2p,k,j , λ2v,k,j Interference (IPI). To satisfy the requirement of (11), the
and λ2h,k,j as the j-th eigenvalue of PH k Pk , Vk Vk and Φk ,
H
following conditions should be satisfied simultaneously:
2 2
respectively, where both λp,k,j and λv,k,j are functions of Preposition 2:
εk,1 , · · · , εk,Nt , and, therefore, (8) can be rewritten as Ep
  1. Gk,n,n = IL , 1 ≤ n ≤ Nt .
Nt max
MSE ĥk|LS 2. Gk,m,n=m = OLmax for each (m, n = m).
⎛ ⎞
Nt −1 2 Nt −1 The first condition of Preposition 2 is easy to explain: since
1 
Lmax
λv,k,j λ2h,k,j Lmax  σ 2
= ⎝ + w ⎠
, k Pk to be identical,
(11) requires that all the eigenvalues of PH
4 2
Lmax Nt j=0
λp,k,j j=0
λp,k,j each Gk,n,n should be a diagonal matrix with an identical
Nt −1
diagonal element. Since the first item in (8) is noise-like,

Lmax
which cannot be minimized by optimizing the pilots, so that
s.t. λ2p,k,j = Lmax Ep .
   −1 
j=0
arg min MSE ĥk|LS ⇔ arg min trace PH k Pk (15)
(10) Pk Pk
   H 
The optimal λ2p,k,j that minimizes MSE ĥk|LS can be de- For a given power  constraint
 trace Pk Pk = Lmax Ep ,
−1
rived based on (10). Unfortunately, for a frequency selective minimizing trace PH k Pk requires the minimization
fading MIMO-OFDM channel where λ2h,k,0 = λ2h,k,1 = of the off-diagonal power of Pk Pk , and when the second
H

· · · = λ2h,k,Lmax Nt −1 , obtaining a closed-form resolution of condition of Preposition 2 is satisfied, (15) is met.

λ2p,k,j appears intractable, if not impossible. A closed-form In order to find the pilots and their placements that satisfy
resolution of λ2p,k,j is achievable when λ2h,k,0 = λ2h,k,1 = Preposition 2, let us first assume that a total of Np pilots are
· · · = λ2h,k,Lmax Nt −1 = λ2h,k , i.e., Φk = λ2h,k ILmax Nt . allocated to each transmit antenna, and that the frequency-
From the definition of Vk , we can readily represent its j- domain indexes of the pilots for the n-th transmit antenna is
th eigenvalue as λ2v,k,j = α2d,k,j λ2p,k,j , which means that (θn,1 , · · · , θn,Np ), where 0 ≤ θn,1 < · · · < θn,Np ≤ N −1. To
λ2v,k,j is proportional to λ2p,k,j (Note that α2d,k,j here is a make PH k Pk full-rank, Np ≥ Lmax should be satisfied. Note
function of the frequency offset with α2d,k,j ≤ λ2d,k,j , and that θn,z for each 1 ≤ n ≤ Nt may be either identical or not.
that λ2d,k,j is the n-th eigenvalue of Dk DH When M ≥ Nt , the pilots and data transmitted by different
k . Equality holds
only if the subcarriers allocated to the data and pilots for each transmit antennas can also be orthogonal in the time-domain.
transmit antenna are totally overlapped.) For MIMO-OFDM For each transmit antenna, to satisfy condition 1 of Prepo-
systems with frequency offsets, the interference of the pilots, sition 2, the total transmit power of the consecutive M
contributed by the data subcarriers, is noise-like; therefore, it symbols in each pilot subcarrier should be identical (i.e.,
is reasonable to assume that the average interference power of 
M−1
Ep
XpH n (i)Xn (i) =
p
diag {0, · · · , 1, · · · , 0}), and the
each pilot subcarrier is identical, i.e., α2d,k,0 = α2d,k,1 = · · · = i=0
Np Nt
α2d,k,Lmax Nt −1 = α2d,k , and the optimal λ2p,k,j can be obtained pilots’ placement requires
as  
Ep (θn,2 · l − θn,1 · l)N = · · · θn,Np · l − θn,Np −1 · l N
λ2p,k,j = , 0 ≤ j ≤ Lmax Nt − 1. (11)   (16)
Nt = θn,1 · l − θn,Np · l N
for each 1 ≤ n ≤ Nt , 1 ≤ l ≤ Lmax − 1. Note that
A. The Optimal Pilot Design and Placement 
M−1
Ep
XpH
n (i)Xn (i) =
p
diag {0, · · · , 1, · · · , 0} should
Before designing pilots that satisfy (11), let us first analyze i=0
N t Np
k Pk :
PH not be understood as indicating that each non-zero element
⎡ ⎤ of Xpn (i) must be modulated as a constant-envelope (CE)
Gk,1,1 . . . Gk,1,Nt pilot, although a CE is a perfect resolution (e.g., CE pilots
⎢ .. .. .. ⎥
PHk Pk = ⎣ . . . ⎦, (12) are proposed in [1]). A CE is required only when M = 1. If
Gk,Nt ,1 . . . Gk,Nt ,Nt M > 1, we need neither to modulate each pilot subcarrier for
each symbol nor to modulate these non-zero pilots as a CE.
where For example, each pilot shown in Table I to Table II satisfies
this requirement.

M−1
Gk,m,n = N F(Lmax ) XpH cir Since the single frequency offset (SFO) case is a special
m (i)Ek,m,n Xn (i)F(Lmax ) ,
p H

i=0
condition of the multiple frequency offsets (MFO) case, the
(13) optimal pilot for a multiple frequency offsets case should also

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 09:15 from IEEE Xplore. Restrictions apply.
 ZHANG et al.: MIMO-OFDM CHANNEL ESTIMATION IN THE PRESENCE OF FREQUENCY OFFSETS 2333

TABLE I
be optimal in a single frequency offset case. Therefore, we O PTIMAL PILOTS ALLOCATION IN MIMO-OFDM WITH N = 32, Nt = 4,
can simplify the condition 2 of Preposition 2 as Lmax = 4, Np = 4 AND M = 2: C ONSTANT E NVELOPE .


M−1
Symbol 1
F(Lmax ) XpH
m (i)Xn (i)F(Lmax ) = OLmax , (SFO);
p H
θz Tx 1 Tx 2 Tx 3 Tx 4
i=0 ρ
0 √ √ρ ρ
√ √ρ
(17a) 2
jπ
2 2
j3π
2
ρ ρ
√ ejπ ρ ρ
√ ej2π
8 √ e2 √ e 2
F(Lmax ) Πmn FH
(Lmax ) = OLmax , (MFO), (17b) 2
ρ jπ
2
ρ j2π
2
ρ j3π
2
ρ j4π
16 √
2
e √
2
e √
2
e √
2
e

M−1
cir 24 ρ
√ e 2
j3π
ρ
√ ej3π ρ
√ e 2
j9π
ρ
√ ej6π
where Πmn = XpH
m (i)Ek,m,n Xn (i), and (17b) implies
p 2 2 2 2
i=0 Symbol 2
(17a). Also define a Np ×Np matrix Πmn , which is generated θz Tx 1 Tx 2 Tx 3 Tx 4
by deleting all the zero rows and columns of Πmn . In Ap- 0 √ ρ √ρ √ρ √ρ
2 2 2 2
pendix A, it is shown that when pilot subcarriers allocated to ρ jπ
ρ jπ ρ j3π
ρ j2π
8 √ e2 √ e √ e 2 √ e
the different transmit antennas are orthogonal in the frequency- 2
ρ jπ
2
ρ j2π
2
ρ j3π
2
ρ j4π
16 √ e √ e √ e √ e
domain, Gk,m,n=m = OLmax for each m = n, and that, 2
j3π
2 2
j9π
2
ρ
√ e 2 ρ
√ e j3π ρ
√ e 2 ρ j6π
therefore, all pilots and their placements that satisfy (18) are 24 √ e
2 2 2 2
optimal pilots in terms of the minimum MSE:
  2πθ K m TABLE II
m,z p O PTIMAL PILOTS ALLOCATION IN MIMO-OFDM WITH N = 32, Nt = 4,
arg [Xpm (i)]θm,z θm,z = , Lmax = 4, Np = 4 AND M = 3: M ULTIPLE E NVELOPES .
N
N
s.t. Lmax ≤ Np ≤ N, = integer; Symbol 1
Np
θz Tx 1 Tx 2 Tx 3 Tx 4
θm,z Kp (n − m)
Kp ≥ Lmax , = integer; 0 ρ ρ 0 0
N jπ j3π
% %2 % %2 8 √ρ
e 2 √ρ jπ
e ρ
√ e 2
ρ j2π
√ e
% p % % % 3 3 3 3
%[Xm (i)]θm,z θm,z % = %[Xpm (i)](θm,z + N ) (θm,z + N ) % ; 16 ρejπ ρej2π 0 0
2 N 2 N j3π j9π
ρ ρ j3π ρ ρ j6π
M % %2 24 √ e 2 √ e √ e 2 √ e
 % p % Ep 3 3 3 3
%[Xm (i)]θm,z θm,z % = ; Symbol 2
i=1
Np Nt θz Tx 1 Tx 2 Tx 3 Tx 4
ρ ρ
Np % %2 0 0 0 √ √
M 
% p % Ep 2 2
%[Xm (i)]θm,z θm,z % =
jπ j3π
ρ ρ jπ ρ ρ j2π
. 8 √
3
e 2 √
3
e √
3
e 2 √
3
e
Nt ρ j3π ρ j4π
i=1 z=1 16 0 0 √
2
e √
2
e
(18) ρ j3π
ρ j3π ρ j9π
ρ j6π
24 √ e 2 √ e √ e 2 √ e
3 3 3 3
where 1 ≤ z ≤ Np , 1 ≤ m ≤ Nt and 0 ≤ i ≤ M − 1. Symbol 3
When θm,z = θn=m,z = θz for each 1 ≤ m, n = m ≤ θz Tx 1 Tx 2 Tx 3 Tx 4
Nt and 1 ≤ z ≤ Np , the Nt transmit antennas share Np 0 0 0 ρ
√ √ρ
2 2
pilot subcarriers. From (34) in Appendix A, we know that 8 ρ
√ e
jπ
2
ρ jπ
√ e ρ
√ e
j3π
2
ρ j2π
√ e
in order to make Gk,m,n = OLmax , μTu Πmn μ∗t = 0 should 3 3 3
ρ j3π
3
ρ j4π
16 0 0 √ e √ e
be satisfied for each 1 ≤ u, t ≤ Lmax . One way to achieve ρ j3π
ρ j3π ρ
2
j9π
2
ρ j6π
this result is to make Πmn = ONp , which requires the pilots 24 √ e 2 √ e √ e 2 √ e
3 3 3 3
transmitted by different transmit antennas be orthogonal in
the time-domain. These time-domain orthogonal pilots achieve
optimal MSE performance, but at the cost of low spectral suboptimal pilots only in terms of the MSE. These suboptimal
efficiency. pilotscan be designed  by minimizing the expectation of
MSE ĥk|εk,1 ,··· ,εk,Nt ; i.e.,
B. The Suboptimal Pilot Design and Placement to Minimize   
the Expectation of MSE Psubopt
k = arg min E MSE ĥk|εk,1 ,··· ,εk,Nt ,
Note that the MSE given by (8) is a function
Pk
 H  (19)
of εk,1 , · · · , εk,Nt . In the multiple frequency offsets s.t. trace Pk Pk = Lmax Ep ,
case, designing
&M−1 the optimal ' pilots that satisfy where the expectation is with respect to εk,1 , · · · , εk,Nt . When
 pH
F(Lmax ) Xm (i)Ecirk,m,n Xn (i) F(Lmax )
p H
= OLmax the eigenvalues of PH
k Pk are identical, the minimum MSE is
i=0 achieved, which requires that
is difficult, if the pilots transmitted by the different transmit
antennas are neither orthogonal in the time-domain nor 
M−1
orthogonal in the frequency-domain. PH
k,m (i)Pk,n=m (i) = OLmax , (SFO); (20a)
However, orthogonal pilot placement considerably reduces i=0
%  M−1 %2
the spectral efficiency, and in this subsection, we consider  %%  %
only the case of θm,z = θn=m,z = θz for each (1 ≤ %
arg min %E trace PH
k,m (i)Pk,n (i) % , (MFO).
m, n = m ≤ Nt , 1 ≤ z ≤ Np ) with M < Nt . Since Pk % %
n=m i=0
the IPI cannot be totally eliminated in this case, we can get (20b)

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 09:15 from IEEE Xplore. Restrictions apply.
 2334 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 6, JUNE 2008

(20a) can be achieved by all the pilots that satisfy (18). To V. L INEAR M INIMUM M EAN S QUARE E RROR (LMMSE)
resolve (20b), we have E STIMATOR BY E XPLOITING T HE C HANNEL C ORRELATION
 M−1
 Sections III to IV considered pilot design for an uncorre-
E trace PHk,m (i)Pk,n (i)
i=0
lated MIMO-OFDM channel. However, channel correlation is
M−1 usually exploited to improve the performance, as discussed in

= trace F(Lmax ) XpH [15], [16].
m (i)F Λp FXn (i)F(Lmax )
H p H
,
i=0
(21)
  A. LMMSE Estimator
4π 2 × 4 2 4π 2 (N − 1)2 2
where Λp ∼
= diag 0, − σ , · · · , − σ ε .
N2 ε
N2 Define the
received vector
To minimize (21), it requires T on the Nr receive anten-
nas as r = rT1 , · · · , rTNr = Ph + Dh + η, where
⎡ ⎤
P = diag {P1 , · · · , PNr }, D = diag {D1 , · · · , DNr },h =
 
Nt M−1
T T
T
Ep ⎢ ⎥ h1 , · · · , hTNr , η = η T1 , · · · , η TNr . Based on it, we
F(Lmax ) XpHm (i)F
H
= ⎣ BL , · · · , BL ⎦ ,
m=1 i=0
Nt   define
N/Np times BL
⎡ (22)
⎤
h̃ = P† r = h + P† Dh + P† η, (24)
⎢ ⎥
where BL = ⎣ILmax , · · · , ILmax , OLmax ×(Np −Lmax Nt ) ⎦,  H −1 H
  where P† = P P P . The
Lmax ×(Lmax Nt )  Hchannel

and Np ≥ Lmax Nt should be satisfied. By resolving (20) correlation
⎡ matrix is given by⎤ Φ = E hh =
to (22), the suboptimal pilots and their placements are Φ1,1 Φ1,2 . . . Φ1,Nr
⎢ Φ2,1 Φ2,2 . . . Φ2,Nr ⎥
  2πθz (m − 1)Lmax ⎢ ⎥
⎢ .. .. .. .. ⎥, where Φk,i =
arg [Xpm (i)]θz θz = , ⎣ . . . . ⎦
N
N ΦNr ,1 ΦNr ,2⎡ . . . ΦNr ,Nr ⎤
s.t. Lmax Nt ≤ Np ≤ N, = integer; Φk,i,1,1 Φk,i,1,2 . . . Φk,i,1,Nt
Np
    ⎢ Φk,i,2,1 Φk,i,2,2 . . . Φk,i,2,Nt ⎥
(θ2 · l − θ1 · l)N = · · · θNp · l − θNp −1 · l N ⎢ ⎥
  E hk hH = ⎢ .. .. .. .. ⎥
i
⎣ . . . . ⎦
= θ1 · l − θNp · l N , l = 1, 2, · · · , Lmax − 1;
Φk,i,Nt ,1 Φk,i,Nt
,2 . . . Φk,i,Nt ,Nt
θz (n − m)Lmax
= integer; and Φk,i,m,n = E h̃k,m h̃H = k,i,m,n ·
N  2 
i,n
 
[Xm (i)]θ1 θ1 = · · · = [Xpm (i)]θNp θNp ≥ 0;
p 2 2 2 2
diag σh,0 , σh,1 , · · · , σh,L max −1
with σh,l = E |h k,i (l)|

M
% p %2 Ep for each (k, i). In the following sections, we assume
%[Xm (i)] % = ; −1
Lmax

i=1
θz θz
Np Nt that 2
σh,l = 1. Note that k,k,m,m = 1 and that
Np
M 
% p %2
l=0
0 ≤ k,i,m,n ≤ 1 for each k = i or m = n. The correlation
%[Xm (i)] % = Ep .
θz θz
Nt matrix of h̃ is given by
i=1 z=1
(23)  H −1
2 (25)
Ch̃h̃ = ΩΦΩH + σw P P ,
When pilots for different transmit antennas are orthogonal
in the frequency-domain, pilots and their placements that ⎛ ⎞
satisfy (18) are the optimal; if pilots for different transmit  H −1 H
antennas are overlapped in the frequency-domain, pilots and where Ω = ⎝ILmax Nt Nr + P P ⎠
PD . We also have
their placements that satisfy (23) are sub-optimal in terms of   V
LS MSE minimization. For a given (N, Nt , Lmax , Np , M ), Chh̃ = E hh̃H = ΦΩH . From [23], an LMMSE estimator
the number of the optimal pilots is so large that numerically can be designed as
exemplifying all the optimal pilots as well as their placements
would be trivial. Table I to Table II present some examples of    −1 H  H 
2
the optimal pilots and their placements that satisfy both (18) ĥLMMSE = σw Φ−1 + ΩH PH P Ω Ω P P h̃
and (23) when N = 32, Nt = 4, Lmax = 4, Np = 4 and (26)
M = 2, 3.
Note that both the optimal pilots and the suboptimal pilots with its covariance matrix of estimation error being given by
are uniformly placed in the frequency-domain, and may suffer   H 
a performance degrade due to the pilots losses at the virtual
Ce = E ĥLMMSE − h ĥLMMSE − h . (27)
subcarriers. One way to mitigate this performance loss here
is to make the pilot distance larger than the virtual-subcarrier
bandwidth and to guarantee that Np ≥ Lmax Nt simultane-   trace {Ce }
ously. The MSE of ĥLMMSE is MSE ĥLMMSE = .
Lmax Nt Nr

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 09:15 from IEEE Xplore. Restrictions apply.
 ZHANG et al.: MIMO-OFDM CHANNEL ESTIMATION IN THE PRESENCE OF FREQUENCY OFFSETS 2335

N = 128; M = 1; Lmax = 4; Np = 4× Lmax; Nt = 4; Overlapped Pilots; Single Frequency Offset; LS Estimators

B. Adaptive Power Allocation to Minimize MSE 0
10
BLM: σε = 10−3
2

Using the eigenvalue decomposition of Ce , we have Proposed: σ2 = 10

ε
−3

  2
BLM: σε = 10−2

MSE ĥLMMSE 2
Proposed: σε = 10
−2

Nt Nr −1
Lmax 2 2
σw λφ,j λ2p,j
=   2 2  , −1

2 λ2 + λ2
Lmax Nt Nr σw λ + λ
10
j=0 p,j φ,j p,j v,j
(28)

MSE
where λ2φ,j represents the j-th eigenvalue of Φ, λ2p,j and λ2v,j
represent the j-th eigenvalue of PH P and VVH , respectively.
2
By resolving
 (36) in  Appendix B, the optimal λp,j to minimize −2
10

the MSE ĥLMMSE can be derived as

 Lmax 
Nt Nr −1

2
σw
Nr Lmax Ep + λ2φ,m σ2
λ2p,j = m=0
− 2w , (29)
Lmax Nt Nr − 1 λφ,j
−3
10

where 0 ≤ j ≤ Lmax Nt Nr − 1. By consid-

0 2 4 6 8 10 12 14 16 18 20
NPNR (dB)
Nt Nr −1
Lmax
ering λ2p,j = Nr Lmax Ep , the minimum Fig. 1. LS channel estimation in single frequency offset MIMO-OFDM
systems with M = 1, Nt = 4 and pilot subcarriers of different transmit
 j=0 antennas are overlapped.
MSE ĥLMMSE is given by 3
   N = 128; M = 1; L
max
= 4; N = 4× L
p max
; N = 2; Overlapped Pilots; Multiple Frequency Offsets; LS Estimators
t

min MSE ĥLMMSE

0
10
2 −3
BLM: σε = 10
2 −3
Proposed: σε = 10
Nt Nr −1
Lmax 2
σw (30) 2
BLM: σε = 10
−2

= Lmax 
Nt Nr −1
. Proposed: σ2 = 10
ε
−2

σw2
j=0 Nr Lmax Ep + λ2φ,m
m=0

As compared to (30), when uniform power allocation to the −1

10

Ep
pilots is applied; i.e., when PH P = IL N N , the
  Nt max t r
suboptimal MSE ĥLMMSE is given by
MSE

 % 
% Ep
MSE ĥLMMSE %PH P = ILmax Nt Nr −2
Nt 10

LmaxNt Nr −1
σ 2 λ2 (31)
=  w φ,j .
Nr Lmax Nt σw 2 + λ2 E
j=0 φ,j p

VI. N UMERICAL R ESULTS −3

10
0 2 4 6 8 10 12 14 16 18 20
A multipath-fading channel with path gains hk,i (l) = e−l NPNR (dB)

and the channel correlation coefficient k,i,m,n=n = 0.5
is considered, where l = 0, 1, 2, · · · , Lmax − 1, k, i =
1, 2, · · · , Nr and m, n = 1, 2, · · · , Nt . We also assume that
N = 128, M = 1, 2, 4, Lmax = 4, Nt = 2, 4, Nr = 2, 4 and
Np = Lmax Nt . Here, we use the Normalized-Pilot-to-Noise
Ep
Ratio (NPNR), i.e., NPNR = , instead of the
Nt · Np · σw 2
PNR to represent the normalized PNR of each pilot subcarrier
for each transmit antenna, and without loss of generality, the
average power of each pilot subcarrier is also assumed to
be identical to that of each data subcarrier, unless otherwise
stated. Multiple frequency offsets are considered for different
transmit and receive antennas with εk,i ∼ N (0, σε2 ).
In a MIMO-OFDM system, pilots transmitted by Nt trans-
mit antennas are modulated into consecutive M ≥ 1 symbols
(Sections III to IV). When M = 1, the pilot subcarriers
for the different transmit antennas can be either orthogonal
Fig. 2. LS channel estimation in multiple frequency offsets MIMO-OFDM
systems with M = 1, Nt = 2 and pilot subcarriers of different transmit
antennas are overlapped.
in the frequency-domain or not. Fig. 1 to Fig. 5 compares
the performance of the proposed LS estimator and the BLM
estimator with M = 1. Fig. 1 shows that with a single
frequency offset, the BLM estimator exhibits a performance
floor at high NPNR, and a larger σε2 implies a worse MSE
performance. However, the proposed estimator eliminates this
floor, and for different σε2 , it achieves an identical MSE.
Fig. 2 to Fig. 5 are for the multiple frequency offsets case.
Fig. 2 and Fig. 3 present the simulation results with overlapped
pilot subcarriers allocated for the different transmit antennas
when Nt = 2 and Nt = 4, respectively. In the multiple fre-
quency offsets case, both the BLM estimator and the proposed

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 09:15 from IEEE Xplore. Restrictions apply.
 2336 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 6, JUNE 2008

N = 128; M = 1; Lmax = 4; Np = 4× Lmax; Nt = 4; Overlapped Pilots; Multiple Frequency Offsets; LS Estimators N = 128; M = 1; Lmax = 4; Np = 4× Lmax; Nt = 4; Orthogonal Pilots; Multiple Frequemcy Offsets; LS Estimators
0 0
10 10
2 −3
BLM: σε = 10−3
2
BLM: σε = 10

Proposed: σ2 = 10−3 Proposed: σ2 = 10

−3
ε ε
BLM: σ2 = 10
−2
BLM: σε = 10−2
2
ε
2 −2 2 −2
Proposed: σε = 10 Proposed: σε = 10

−1 −1
10 10
MSE

MSE
−2 −2
10 10

−3 −3
10 10
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
NPNR (dB) NPNR (dB)

Fig. 3. LS channel estimation in multiple frequency offsets MIMO-OFDM Fig. 5. LS channel estimation in multiple frequency offsets MIMO-OFDM
systems with M = 1, Nt = 4 and pilot subcarriers of different transmit systems with M = 1, Nt = 4 and pilot subcarriers of different transmit
antennas are overlapped. antennas are orthogonal.

N = 128; M = 1; Lmax = 4; Np = 4× Lmax; Nt = 2; Orthogonal Pilots; Multiple Frequency Offsets; LS Estimators 2

N = 128; M = 4; Lmax = 4; Np = 4× Lmax; Nt = 4; σε = 10−2; Orthogonal Pilots; Multiple Frequency Offsets; LS Estimators
0
10
2 −3
BLM: σε = 10 BLM
Proposed: CE
Proposed: σε = 10−3
2
Proposed: Multiple−Envelop
BLM: σε = 10−2
2

2 −2
Proposed: σε = 10

−1
10
MSE

MSE

−1
10

−2
10

−3 −2
10 10
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
NPNR (dB) NPNR (dB)

Fig. 4. LS channel estimation in multiple frequency offsets MIMO-OFDM Fig. 6. LS channel estimation in multiple frequency offsets MIMO-OFDM
systems with M = 1, Nt = 2 and pilot subcarriers of different transmit systems with M = 4, Nt = 4 and pilot subcarriers of different transmit
antennas are orthogonal. antennas are orthogonal.

one exhibit a performance floor. However, the latter maintains BLM estimator. Since in frequency-domain orthogonal pilot
a performance advantage. For example, when Nt = 2 and placement, the IPI is reduced as compared to that in the
NPNR=20 dB, the MSE of the proposed estimator is about overlapped pilot placement, the MSE is smaller than that of the
1.6 × 10−3 (or 5 × 10−3) for σε2 = 10−3 (or σε2 = 10−2 ), and latter, but at the cost of lower spectral efficiency. For example,
that for the BLM estimator is about 5.1×10−3 (or 3.1×10−2) when Nt = 4, σε2 = 10−2 and NPNR=20 dB, the MSE for the
for σε2 = 10−3 (or σε2 = 10−2 ). proposed estimator with overlapped pilots placement is about
Fig. 4 and Fig. 5 compare the proposed estimator and 1.7 × 10−2 , whereas that with orthogonal pilots placement is
the BLM estimator with orthogonal pilot placement in the about 1.2 × 10−2 , as shown in Fig. 3 and Fig. 5.
frequency-domain. Note that even with orthogonal pilots When Nt > 1, pilots for each transmit antenna can be
placement, a performance floor will always appear at the modulated into consecutive M > 1 symbols, the envelope can
proposed estimator in the presence of multiple frequency either be CE or not. The simulation results with Nt = M = 4
offsets, although the proposed estimator still outperforms the are illustrated in Fig. 6, where frequency-domain orthogonal

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 09:15 from IEEE Xplore. Restrictions apply.
 ZHANG et al.: MIMO-OFDM CHANNEL ESTIMATION IN THE PRESENCE OF FREQUENCY OFFSETS 2337

N = 128; M = 4; L = 4; N = 4× L ; N = 4; σ2 = 10−2; Multiple Frequency Offsets; LS Estimators N = 128; L = 4; N = 4; N = L N ; Orthogonal Pilots; Multiple Frequency Offsets
max p max t ε max t p max t
0
0
10 10
2 −2
Proposed: Multiple−Symbol Modulation; CE Cui: σε = 10
Proposed: Multiple−Symbol Modulation; Multiple−Envelop
Proposed: σε = 10−2; Ed/Ep = 0
2
Proposed: Time−domain Perfectly−Orthogonal Modulation
Proposed: σ2 = 10−2; E /E = 0.1
ε d p
2 −1
Cui: σε = 10

Proposed: σ2 = 10 ; E /E = 0
−1
ε d p
Proposed: σε = 10−1; Ed/Ep = 0.1
2

−1
−1 10
10

MSE
MSE

−2
−2 10
10

−3
−3 10
10 0 2 4 6 8 10 12 14 16 18 20
0 2 4 6 8 10 12 14 16 18 20
NPNR (dB)
NPNR (dB)

Fig. 7. LS channel estimation in multiple frequency offsets MIMO-OFDM Fig. 9. Performance comparison between the proposed LS estimator and
systems with M = 4, Nt = 4 and pilots modulated with either consecutive Cui’s estimator with Lmax = 4 and Nt = 4.
multiple symbols or perfect orthogonal in time-domain.
N = 128; M = 1; Lmax = 4; Np = 4× Lmax; σε = 10−2; Multiple Frequency Offsets; Overpalled Pilots
2
0
10
N = 128; M = 4; Lmax = 4; Nt = 4; Np = LmaxNt; Orthogonal Pilots; Multiple Frequency Offsets
0
LS: Nt = 2, Nr = 2
10
LMMSE: Nt = 2, Nr = 2; CE
Flat−fading Channel; σε = 10−3
2
LMMSE: Nt = 2, Nr = 2; Adaptive Power Allocation
2 −3
Multipath−fading Channel; σε = 10 LS: Nt = 4, Nr = 4
2 −2
Flat−fading Channel; σε = 10 LMMSE: Nt = 4, Nr = 4; CE
Multipath−fading Channel; σ2 =
−2 LMMSE: Nt = 4, Nr = 4; Adaptive Power Allocation
10
ε

−1
10
−1
10
MSE
MSE

−2
10
−2
10

−3
10
−3 0 2 4 6 8 10 12 14 16 18 20
10 NPNR (dB)
0 2 4 6 8 10 12 14 16 18 20
NPNR (dB)

Fig. 10. Performance comparison between LS and LMMSE estimators.

Fig. 8. Performance degradation of LS channel estimation in the multipath-
fading channels with M = 4, Nt = 4 and pilot subcarriers of different
transmit antennas are orthogonal.
pilots are used, multiple symbol pilots have no performance
advantage over that of single-symbol pilots in terms of MSE,
pilot placement is assumed for the different transmit antennas and more seriously, multiple-symbol pilot modulation has
in each symbol (a similar result can be achieved in the lower spectrum efficiency. Multiple-symbol pilot modulation
case of overlapped pilots placement for different transmit is nevertheless applied for two main reasons: (1) it reduces
antennas). The proposed pilots, either a CE or Multiple- the peak power of each pilot subcarrier in each symbol; and
Envelope, can achieve the same performance advantage over (2) it has an advantage over single-symbol pilot modulation in
that of the BLM estimator, provided that the total pilots power tracking the time-variant channel. In the presence of multiple
Ep remains fixed. For a given Ep , we can also conclude that frequency offsets, non-zero Inter-Antenna-Interference will
an identical performance can be achieved in the proposed degrade the channel estimation performance if neither time-
estimator with either M = 1 or M > 1. For example, domain nor frequency-domain orthogonal pilot placement is
when Nt = 4, σε2 = 10−2 and NPNR=20 dB, an MSE utilized. To eliminate this Inter-Antenna-Interference, we can
of about 1.5 × 10−2 can be achieved with either M = 1 make M = Nt and ensure that the pilots and data transmit-
or M = 4, as shown in Fig. 5 and Fig. 6. No matter what ted by the different transmit antennas are orthogonal in the

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 09:15 from IEEE Xplore. Restrictions apply.
 2338 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 6, JUNE 2008

time-domain; i.e., when one transmit antenna is transmitting, the transmitter and the receiver, an LMMSE estimator with
the other transmit antennas remain silent. Fig. 7 presents CE pilots modulation was designed based on the proposed
the performance analysis of the proposed estimator in two optimal pilots. This estimator outperformed the proposed LS
pilot placement cases: the time-domain orthogonal and non- estimator in terms of the MSE. By applying an adaptive pilot
orthogonal pilot placements. The simulation results show that power allocation at the transmitter, a much lower MSE as
the performance floor at high NPNR can be mitigated in the compared to that achieved with a CE pilot modulation was
case of orthogonal pilot placement, but at the cost of lower achieved in the proposed LMMSE estimator.
spectral efficiency than that of non-orthogonal placement.
Note that (11) is optimal in flat fading, but the MSE A PPENDIX A
performance may degrade in frequency selective fading. Fig. 8 A NALYSIS OF Gk,m,n=m
shows that a performance degradation of about 0.4 dB (or Define Gk,m,n = N F(Lmax ) Πmn FH (Lmax ) where Πmn =
about 0.9 dB) is achieved in the multipath-fading channel at 
M−1
cir
a high NPNR when σ 2 = 10−3 (or when σ 2 = 10−2 ), as F(Lmax ) XpH
m (i)Ek,m,n Xn (i), where the θm,l θn,s -th ele-
p

compared to the MSE performance in the flat-fading channels. i=0

For a small σ 2 , this performance degradation is negligible. ment of Πmn is non-zero, as given by
We also compare the performance of the proposed LS [Πmn ]θm,l θn,s
estimator with Cui’s estimator [2] with multiple frequency 
M−1
∗

offsets, and the numerical results are shown in Fig. 9. Note = [Xpm (i)]θm,l θm,l Ecir
k,m,n θ [Xpn (i)]θn,s θn,s ,
m,l θn,s
that in Cui’s estimator, the total preamble is used for training i=0
only. For a fair comparison, we use 16 pilot subcarriers in (32)
the proposed estimator and use a length-16 training sequence where 1 ≤ l, s ≤ Np . Let us consider the following two cases:
in Cui’s estimator, with an identical total power Ep being 1) Pilot Subcarriers for the m-th and the n-th Transmit
allocated to the pilots/training in each estimator. If the data Antennas are Orthogonal in the frequency-domain: In this
subcarriers in the proposed symbol are not modulated, the case, θm,l = θn,s for each l and s, and it is easy to show
proposed estimator always outperforms Cui’s estimator at a that F(Lmax ) Πmn FH(Lmax ) = OLmax , so that Gk,m,n=m =
high NPNR, and a larger performance gain can be achieved as OLmax for each m = n.
σ 2 increases. For example, at a high NPNR, the performance 2) Pilot Subcarriers for the m-th and the n-th Transmit
improvement of the proposed estimator with Ed = 0 over Antennas are Overlapped in the frequency-domain (θm,z =
Cui’s estimator is about 0.7 dB (or 2.6 dB) with σ 2 = 10−2 θn,z = θz for z = 1, 2, · · · Np ): Define Πmn as
(or σ 2 = 10−1 ). A performance degradation will be achieved
in the proposed estimator if Ed = 0. In this simulation, we 
M−1


[Πmn ]θl θs = [Xpm (i)]∗θl θl Ecir
k,m,n θ θ [Xn (i)]θs θs ,
p
assume that Ed /Ep = 10−1 , and the MSE performance of l s
i=0
the proposed estimator is about 0.5 dB worse than that of (33)
Cui’s estimator. In other words, Cui’s estimator outperforms where 1 ≤ l, s ≤ Np . Also define a Np × Np matrix Πmn ,
the proposed estimator by sacrificing the spectral efficiency. which is generated by deleting all the zero rows and columns
When the covariance matrix is available at both the trans- of Πmn . Evidently, Πmn = ΠH mn . Gk,m,n can be rewritten
mitter and the receiver, an LMMSE estimator can be designed as
to improve the estimation accuracy. Fig. 10 compares the    H
channel estimation performance of the proposed LS and the Gk,m,n = N fθ1 , · · · , fθNp Πmn fθ1 , · · · , fθNp
proposed LMMSE estimators. Multiple frequency offsets are    
F(Lmax ) FH
considered with σε2 = 10−2 . The LMMSE estimator with (Lmax )

T

CE pilot modulation can considerably reduce the MSE by = N μ1 , · · · , μLmax Πmn μ∗1 , · · · , μ∗Lmax ,
applying adaptive pilot power allocation at the transmitter. For (34)
example, when Nt = Nr = 2, the performance advantage
of the proposed LMMSE estimator with adaptive pilot power where μTz is the z-th row of F(Lmax ) . Therefore, the ut-
allocation over that with CE pilot modulated at a low NPNR th element of Gk,m,n is [Gk,m,n ]ut = N μTu Πmn μ∗t , and
is about 3.1 dB, and its performance advantage over that of Gk,m,n = OLmax is achieved only when μTu Πmn μ∗t = 0,
the LS estimator is about 5.8 dB. 1 ≤ u, t ≤ Lmax .
A PPENDIX B
VII. C ONCLUSIONS O PTIMAL E IGENVALUES FOR A DAPTIVE A LLOCATION
The optimal pilot design and placement for channel es- Define a cost function,
timation in MIMO-OFDM with multiple frequency offsets  2 
C λp,0 , · · · , λ2p,Lmax Nt Nr −1
were discussed. The IPI was eliminated in the proposed
Nt Nr −1
Lmax 2 2
estimator, and, therefore, a performance advantage over that σw λφ,j λ2p,j
=   2 2 
of conventional estimators was achieved. Given a total pilot 2 λ2 + λ2
j=0Lmax Nt Nr σw λ + λv,j
power Ep , the pilots for the different transmit antennas was ⎛
p,j φ,j
⎞
p,j

modulated into one or consecutive multiple OFDM symbols Nt Nr −1

Lmax
with each pilot subcarrier in each symbol being modulated +β⎝ λ2p,j − Nr Lmax Ep ⎠ ,
as a CE or not (a CE is required in pilot design only when j=0
M = 1). With the channel covariance matrix known at both (35)

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 09:15 from IEEE Xplore. Restrictions apply.
 ZHANG et al.: MIMO-OFDM CHANNEL ESTIMATION IN THE PRESENCE OF FREQUENCY OFFSETS
where β is a positive real coefficient.
To minimize (35), let us take the partial derivative to
C λ2p,0 , · · · , λ2p,Lmax Nt Nr −1 with respect to each λ2p,j (0 ≤
j ≤ Lmax Nt Nr − 1) and set the result to zero. The result is
  2 2
2 2
Lmax Nt Nr σw λp,j + λ2φ,j λ2p,j + λv,j β [21] X. Fu, H. Minn, and C. Cantrell, “Two novel iterative joint frequency-
2 4
 4 2
 offset and channel estimation methods for OFDMA uplink,” in Proc.
− σw λφ,j λp,j − λv,j = 0,
(36)
LmaxNt Nr −1
2 [22] Z. Zhang, W. Zhang, and C. Tellambura, “Optimal pilots for OFDMA
st. λp,j = Nr Lmax Ep .
j=0
R EFERENCES
[1] I. Barhumi, G. Leus, and M. Moonen, “Optimal training design for
MIMO OFDM systems in mobile wireless channels,” IEEE Trans.
Signal Processing, vol. 51, no. 6, pp. 1615–1624, June 2003.
[2] T. Cui and C. Tellambura, “Channel and frequency offset estimation
and training sequence design for MIMO frequency selective channels,”
IEEE Trans. Wireless Commun., submitted, to appear.
[3] V. Tarokh, N. Seshadri, and A. R. Calderbank, “Space-time codes for
high data rate wireless communication: performance analysis and code
construction,” IEEE Trans. Inform. Theory, vol. 44, pp. 744–765, Mar.
1998.
[4] Y. Li, J. Winters, and N. Sollenberger, “MIMO-OFDM for wireless
communications: signal detection with enhanced channel estimation,”
IEEE Trans. Commun., vol. 50, no. 9, pp. 1471–1477, Sept. 2002.
[5] H. Bolcskei, “MIMO-OFDM wireless systems: basics, perspectives, and
challenges,” IEEE Wireless Commun., vol. 13, no. 4, pp. 31–37, Aug.
2006. (See also IEEE Personal Commun.)
[6] T. Schmidl and D. Cox, “Robust frequency and timing synchronization
for OFDM,” IEEE Trans. Commun., vol. 45, no. 12, pp. 1613–1621,
Dec. 1997.
[7] Z. Zhang, M. Zhao, H. Zhou, Y. Liu, and J. Gao, “Frequency offset
estimation with fast acquisition in OFDM system,” IEEE Commun. Lett.,
vol. 8, no. 3, pp. 171–173, Mar. 2004.
[8] Z. Zhang, W. Jiang, H. Zhou, Y. Liu, and J. Gao, “High accuracy
frequency offset correction with adjustable acquisition range in OFDM
systems,” IEEE Trans. Wireless Commun., vol. 4, no. 1, pp. 228–237,
Jan. 2005.
[9] A. Mody and G. Stuber, “Synchronization for MIMO OFDM systems,”
in Proc. IEEE Global Telecommn. Conf. (GLOBECOM), vol. 1, San
Antonio, TX, 2001, pp. 509–513.
[10] O. Besson and P. Stoica, “On parameter estimation of MIMO flat-
fading channels with frequency offsets,” IEEE Trans. Signal Processing,
vol. 51, no. 3, pp. 602–613, Mar. 2003.
[11] Y. Li, J. Cimini, and N. Sollenberger, “Robust channel estimation for
OFDM systems with rapid dispersive fading channels,” IEEE Trans.
Commun., vol. 46, no. 7, pp. 902–915, July 1998.
[12] M. Dong and L. Tong, “Optimal design and placement of pilot symbols
for channel estimation,” IEEE Trans. Signal Processing, vol. 50, no. 12,
pp. 3055–3069, Dec. 2002.
[13] H. Minn and N. Al-Dhahir, “Optimal training signals for MIMO OFDM
channel estimation,” IEEE Trans. Wireless Commun., vol. 5, no. 5, pp.
1158–1168, May 2006.
[14] D. Hu, L. Yang, Y. Shi, and L. He, “Optimal pilot sequence design for
channel estimation in MIMO OFDM systems,” IEEE Commun. Lett.,
vol. 10, no. 1, pp. 1–3, Jan. 2006.
[15] J. Kotecha and A. Sayeed, “Transmit signal design for optimal esti-
mation of correlated MIMO channels,” IEEE Trans. Signal Processing,
vol. 52, no. 2, pp. 546–557, Feb. 2004.
[16] H. Zhang, Y. Li, A. Reid, and J. Terry, “Channel estimation for
MIMO OFDM in correlated fading channels,” in Proc. IEEE Int. Conf.
Communications (ICC), vol. 4, May 2005, pp. 2626–2630.
[17] H. Minn, N. Al-Dhahir, and Y. Li, “Optimal training signals for MIMO
OFDM channel estimation in the presence of frequency offset and phase
noise,” IEEE Trans. Commun., vol. 54, no. 6, pp. 1081–1096, June 2006.
[18] M. Ghogho and A. Swami, “Training design for multipath channel
and frequency-offset estimation in MIMO systems,” IEEE Trans. Signal
Processing, vol. 54, no. 10, pp. 3957–3965, Oct. 2006.
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 3, 2009 at 09:15 from IEEE Xplore. Restrictions apply.
2339
[19] H. L. V. Trees, Detection, Estimation, and Modulation Theory, Part IV,
Optimum Array Processing. New York: John Wiley & Sons, 2002.
[20] Z. Zhang and H. Kayama, “Robust uplink carrier frequency offset
estimation with interference mitigation in OFDMA systems,” in Proc.
IEEE Wireless Commun. and Networking Conf., vol. 2, pp. 816–821,
Apr. 2006.
IEEE Global Telecommn. Conf. (GLOBECOM), San Francisco, CA,
2006.
uplink frequency offsets and channel estimation,” IEEE Trans. Veh.
Technol., to be submitted.
[23] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation
Theory. Englewood Cliffs, NJ: Prentice Hall, 1993.
Zhongshan Zhang received the M.S. degree in
computer science in 2001 and Ph.D. degree in
electrical engineering in 2004 from the Beijing Uni-
versity of Posts and Telecommunications (BUPT).
From Aug. 2004 he joined DoCoMo Beijing Labora-
tories as an associate researcher, and was promoted
to be a researcher in Dec. 2005. From Feb. 2006,
he joined University of Alberta, Edmonton, AB,
Canada, as a postdoctoral fellow. His main research
interests include statistical signal processing, syn-
chronization and channel estimation in Multiple-
input Multiple-output (MIMO) orthogonal frequency division multiplexing
(OFDM) systems, and cooperative communications.
Wei Zhang received her B.E. and M.E. degrees
in control engineering from Shandong University,
Jinan, P.R. China, in 2000 and 2004., respectively. In
Sept. 2005, she joined the Department of Electrical
and Computer Engineering at University of Alberta
where she is currently pursuing her Ph.D degree.
Her current research interests include wireless com-
munications theory, signal detection and estimation,
diversity and cooperative communications.
Chintha Tellambura (SM’02) received the B.Sc.
degree (with first-class honors) from the Univer-
sity of Moratuwa, Moratuwa, Sri Lanka, in 1986,
the M.Sc. degree in electronics from the Univer-
sity of London, London, U.K., in 1988, and the
Ph.D. degree in electrical engineering from the
University of Victoria, Victoria, BC, Canada, in
1993.
He was a Postdoctoral Research Fellow with the
University of Victoria (1993-1994) and the Univer-
sity of Bradford (1995-1996). He was with Monash
University, Melbourne, Australia, from 1997 to 2002. Presently, he is a Profes-
sor with the Department of Electrical and Computer Engineering, University
of Alberta. His research interests include Diversity and Fading Counter-
measures, Multiple-Input Multiple-Output (MIMO) Systems and Space-Time
Coding, and Orthogonal Frequency Division Multiplexing (OFDM).
Prof. Tellambura is an Associate Editor for the IEEE T RANSACTIONS
ON C OMMUNICATIONS and the Area Editor for Wireless Communications
Systems and Theory in the IEEE T RANSACTIONS ON W IRELESS C OMMU -
NICATIONS . He was Chair of the Communication Theory Symposium in
Globecom’05 held in St. Louis, MO.