You are on page 1of 8


5, MAY 2010 1485

Joint Estimation of I/Q Imbalance, CFO and

Channel Response for MIMO OFDM Systems
Yuan-Hwui Chung and See-May Phoong, Senior Member, IEEE

Abstract—In this paper, we study the joint estimation of in- only one OFDM block for training and can achieve a good
phase and quadrature-phase (I/Q) imbalance, carrier frequency performance. For CFO estimation, a low complexity maximum
offset (CFO), and channel response for multiple-input multiple- likelihood (ML) technique was proposed in [5] and [6].The
output (MIMO) orthogonal frequency division multiplexing
(OFDM) systems using training sequences. A new concept called joint estimation of CFO and I/Q imbalance was investigated
channel residual energy (CRE) is introduced. We show that by in [7]. The authors in [7] combined the techniques in [3]
minimizing the CRE, we can jointly estimate the I/Q imbalance and [5] to jointly estimate the I/Q and CFO parameters using
and CFO without knowing the channel response. The proposed two repeated training blocks. The Cramer-Rao bound of CFO
method needs only one OFDM block for training and the training estimation was derived in [8].
symbols can be arbitrary. Moreover when the training block
consists of two repeated sequences, a low complexity two-step Recently, there have been a lot of interests in combining
approach is proposed to solve the joint estimation problem. the OFDM systems with the multiple-input multiple-output
Simulation results show that the mean-squared error (MSE) of (MIMO) technique. These systems are known as MIMO
the proposed method is close to the Cramer-Rao bound (CRB).
OFDM systems. Many methods have been proposed to deal
Index Terms—MIMO OFDM, CFO, I/Q imbalance, channel with the channel estimation in MIMO OFDM systems. One
of these methods is to send the training sequences (known
to the receiver) from the transmitter. The design of training
I. I NTRODUCTION sequences for MIMO OFDM systems was investigated in

I N recent years, direct conversion receiver has drawn a

lot of attention due to its low power consumption and
low implementation cost. However some mismatches in direct
[9]-[12]. In [9], the author derived a special class of the
optimal training sequences in one OFDM block. The general
solution for the optimal sequences and the conditions for
conversion receiver can seriously degrade the system perfor- their existence were derived in [10] and [11] respectively.
mance, such as in-phase and quadrature-phase (I/Q) imbalance However these methods assume that there is no mismatch of
and carrier frequency offset (CFO). The I/Q imbalance is the local oscillators. Several methods have been developed for
due to the amplitude and phase mismatches between the I the estimation of the I/Q and CFO mismatches [13]-[21]. The
and Q-branch of the local oscillator whereas the CFO is due compensation of I/Q imbalance for MIMO OFDM systems
to the mismatch of carrier frequency at the transmitter and was investigated in [13][14][15]. In [13], the authors applied
receiver. It is known [1] that the I/Q imbalance and CFO can the I/Q compensation method in [1] to MIMO OFDM systems.
cause a serious inter-carrier interference (ICI) in orthogonal The authors in [14] combined the MMSE and MLD decoders
frequency division multiplexing (OFDM) systems. As a result, for the I/Q compensation. A combination of the data-aided
the bit error rate (BER) has an error-flooring. There have been and decision directed approaches was proposed in [15]. In
many reports in the literature on the compensation of the I/Q [16], the authors derived the Cramer-Rao bound (CRB) for
imbalance and CFO [1]-[7]. Several compensation methods the joint estimation of CFO and channel response for MIMO
for I/Q imbalance in OFDM systems have been proposed systems. The design of training sequence for MIMO OFDM
in [1]. In [3] assuming that the channel frequency response systems in the presence of CFO was investigated in [17][18].
is smooth, a frequency-domain estimation method has been In [19], the authors used orthogonal training sequences for
proposed to jointly estimate the I/Q imbalance and channel CFO estimation. In [20], the authors studied the positions of
frequency response. Recently, exploiting the fact that the size pilot symbols in different OFDM blocks. Both the data-aided
of the DFT matrix is usually larger than the channel length and the non-data-aided methods were investigated in [21].
in OFDM systems, a time-domain method was proposed in Most of these methods need more than one training block
[4] for the joint estimation of I/Q and channel response. to achieve a good performance.
Both the frequency-domain and time-domain methods need In this paper, we extend the method in [4] for the joint
Paper approved by J. Wang, the Editor for Wireless Spread Spectrum of estimation of I/Q imbalance, CFO and channel response in
the IEEE Communications Society. Manuscript received December 3, 2008; MIMO OFDM systems. By introducing a new concept called
revised May 5, 2009 and September 30, 2009. channel residual energy (CRE), we propose a new estimation
The authors are with the Dept. of Electrical Engineering and Graduate
Institute of Communication Engineering, National Taiwan University, Taipei, method that needs only one OFDM block for training. By
Taiwan 106, R.O.C. (e-mail:, minimizing the CRE, we are able to estimate the I/Q imbal-
This work is supported by the National Science Council, Taiwan, R.O.C., ance and CFO without knowing the channel responses. Based
under Grant NSC 97-2628-E-002-044-MY3 and by the Aim for Top Univer-
sity Project of National Taiwan University under Grant 96R8044. on the estimated I/Q imbalance and CFO parameters, we can
Digital Object Identifier 10.1109/TCOMM.2010.05.080627 easily calculate the estimate of channel responses. Moreover
0090-6778/10$25.00 ⃝
c 2010 IEEE

when the training data consists of two repeated sequences, Suppose in addition to the CFO, there is also I/Q mismatch at
we derive a low complexity two-step approach for solving the receiver. The received vector due to I/Q mismatch becomes
the joint estimation problem. Simulation results show that the [13]
MSEs of the two proposed methods are very close to the
Cramer-Rao bound (CRB) [8][16]. z𝑘 = 𝜇𝑘 y𝑘 + 𝜈𝑘 y𝑘∗ , (6)
The rest of this paper is organized as follows. The system where 𝜇𝑘 and 𝜈𝑘 are the I/Q parameters at the receiver. They
model is described in Section II. In Section III, we briefly are related to the amplitude mismatch 𝜖𝑘 and phase mismatch
review the channel estimation problem for MIMO OFDM 𝜙𝑘 as
systems [10]. The joint estimation of the I/Q imbalance, CFO
and channel response is studied in Section IV. The two-step 1 + 𝜖𝑘 𝑒−𝑗𝜙𝑘 1 − 𝜖𝑘 𝑒𝑗𝜙𝑘
𝜇𝑘 ≜ and 𝜈𝑘 ≜ . (7)
approach is discussed in Section V. Simulation results are 2 2
given in Section VI. Conclusions are drawn in Section VII. Substituting (4) into (6), we get
Parts of the results in this paper were presented in [22].
z𝑘 = 𝜇𝑘 E𝑘 r𝑘 + 𝜈𝑘 E∗𝑘 r∗𝑘 . (8)
† 𝑇 ∗
Notation: The symbols A , A and A denote respectively The received vector z𝑘 consists of not only the desired base-
the transpose-conjugate, the transpose, and the complex con- band vector r𝑘 but also its complex conjugate r∗𝑘 . Moreover,
jugate of A. the presence of E𝑘 due to CFO will also destroy the subcarrier
orthogonality [5]. In later sections, we will show how to jointly
II. S YSTEM D ESCRIPTION estimate the I/Q imbalance, CFO and MIMO channel response
using training sequences.
Fig. 1 shows an MIMO OFDM system where the numbers
Suppose that we have estimates of the I/Q imbalance and
of the transmit and receive antenna are 𝑁𝑡 and 𝑁𝑟 respectively.
CFO at the receiver. We will show how to recover the desired
The input vector s𝑗 (see Fig. 1) is an 𝑀 × 1 vector containing
baseband vector r𝑘 from z𝑘 . Define a parameter 𝛼𝑘 that is
the modulation symbols. After taking the 𝑀 -point IDFT of
related to the I/Q imbalance parameters as
s𝑗 , we obtain the 𝑀 × 1 vector x𝑗 . After the insertion of a
CP of length 𝐿 − 1, the signal is transmitted from the 𝑗th 𝜈𝑘
𝛼𝑘 ≜ ∗ . (9)
transmit antenna. Let the channel impulse response from the 𝜇𝑘
𝑗th transmit antenna to the 𝑘th receive antenna be ℎ𝑘,𝑗 (𝑛). If 𝛼𝑘 is known at the receiver, from (6) we can get [3]
We assume that the lengths of all the channels are ≤ 𝐿 and
the length of the cyclic prefix (CP) is 𝐿 − 1. So there is no z𝑘 − 𝛼𝑘 z∗𝑘
𝜇𝑘 y𝑘 = . (10)
interblock interference between adjacent OFDM blocks after 1 − ∣𝛼𝑘 ∣2
CP removal. The received vector at the 𝑘th receive antenna If 𝜃𝑘 is also known at the receiver, from (4) we can recover
can be written as a scaled version of the desired baseband vector by
⎡ x0 ⎤
z𝑘 − 𝛼𝑘 z∗𝑘
[ ]⎢ x1 ⎥ 𝜇𝑘 r𝑘 = E∗𝑘 𝜇𝑘 y𝑘 = E∗𝑘 . (11)
r𝑘 = H𝑘,0 H𝑘,1 ⋅⋅⋅ H𝑘,𝑁𝑡 −1 ⎢ .. ⎥ + q𝑘 (1) 1 − ∣𝛼𝑘 ∣2
⎣ ⎦
where H𝑘,𝑗 is an 𝑀 ×𝑀 circulant matrix with the first column In the following, we briefly review the problem of MIMO
[ ]𝑇 channel estimation using training sequences described in [10].
h𝑘,𝑗 ≜ ℎ𝑘,𝑗 (0) ⋅ ⋅ ⋅ ℎ𝑘,𝑗 (𝐿 − 1) 0 ⋅ ⋅ ⋅ 0 , (2) Given s𝑖 , or equivalently x𝑖 (see Fig. 1), we want to estimate
the channel response from the received vector. Since it has
and q𝑘 is the 𝑀 × 1 blocked version of channel noise. been shown [16] that the channel responses can be estimated
After passing r𝑘 through the 𝑀 -point DFT, we can employ separately for each receive antenna, we drop the receive
a frequency domain equalizer (FEQ) to recover the transmit antenna index 𝑘 in the following discussion for simplicity.
signal s𝑗 . Suppose that there are no CFO and I/Q imbalance. For the
Suppose now that the system suffers from carrier frequency purpose of channel estimation, we rewrite the received signal
offset (CFO) Δ𝑓𝑘 . Define the normalized CFO as in (1) as
Δ𝑓𝑘 ⎡ ⎤
𝜃𝑘 ≜ = Δ𝑓𝑘 𝑀 𝑇, (3) h0
𝑀𝑇 [ ]⎢⎢ h1 ⎥

r = X0 X1 ⋅ ⋅ ⋅ X𝑁𝑡 −1 ⎢ .. ⎥ + q, (12)
where 𝑀 is the size of the DFT matrix and 𝑇 is the sample ⎣ . ⎦
spacing. The vector due to CFO is [18] h𝑁𝑡 −1
y𝑘 = E𝑘 r𝑘 , (4) where h𝑗 is the 𝑀 × 1 vector defined in (2) and X𝑗 is an
𝑀 × 𝑀 circulant matrix with the first column x𝑗 . Define the
where r𝑘 is the desired baseband vector in (1) and E𝑘 is an 𝑀 × 𝐿 submatrix A𝑗 which consists of the first 𝐿 columns
𝑀 × 𝑀 diagonal matrix of X𝑗 and let c𝑗 be
[ ]
ℎ𝑗 (𝐿 − 1) ]𝑇 .
2𝜋 2𝜋
E𝑘 ≜ 𝑑𝑖𝑎𝑔 1 𝑒𝑗 𝑀 𝜃𝑘 ⋅ ⋅ ⋅ 𝑒𝑗 𝑀 (𝑀−1)𝜃𝑘 . (5) c𝑗 ≜ [ ℎ𝑗 (0) ℎ𝑗 (1) ⋅ ⋅ ⋅ (13)

Ant 0 h0,0 (n) Ant 0

s0 x0 r0 s^ 0
Add Remove
. .
. .
. h0, Nt 1 (n) hNr 1,0 (n) FEQ
sNtà1 xNtà1 Ant Nt à 1
AntNr à 1 rNrà1 s^ Ntà1
Add Remove
CP hNr 1, Nt 1 (n) CP
Fig. 1. An MIMO OFDM system.
[ ]𝑇
[ c ≜ c𝑇0 c𝑇1] ⋅ ⋅ ⋅ c𝑇𝑁𝑡 −1 and A ≜ Collecting all the vectors d𝑖 for 0 ≤ 𝑖 ≤ 𝑁𝑡 − 1, we form the
A0 A1 ⋅ ⋅ ⋅ A𝑁𝑡 −1 . Then (12) can be rewritten as 𝑀 × 1 vector
[ ]𝑇
r = Ac + q. (14) d ≜ d𝑇0 d𝑇1 ⋅ ⋅ ⋅ d𝑇𝑁𝑡 −1 . (21)

From (14), we see that the 𝐿𝑁𝑡 × 1 channel vector c is Let A𝑘 be any 𝑀 × (𝜌 − 𝐿) matrix such that the following
identifiable if and only if the 𝑀 × 𝐿𝑁𝑡 matrix A has 𝑀 × 𝑀 matrix B is invertible
full column rank. Thus a necessary condition for channel [ ′ ′ ′ ]
identifiability is 𝑀 ≥ 𝐿𝑁𝑡 . Assume q is a complex Gaussian B ≜ A0 A0 A1 A1 ⋅ ⋅ ⋅ A𝑁𝑡 −1 A𝑁𝑡 −1 . (22)
random vector with covariance matrix R𝑞 = 𝒩0 I. Then a Using (21) and (22), the vector r in (14) can be rewritten as
least-squares estimate of c is given by r = Bd + q. Since B is 𝑀 × 𝑀 , the estimate of MIMO
ĉ = (A† A)−1 A† r. (15) channel response is given by
[ ]𝑇
Define the error vector as e = ĉ − c = (A† A)−1{A† q. }The d̂ = d̂𝑇0 d̂𝑇1 ⋅ ⋅ ⋅ d̂𝑇𝑁𝑡 −1 = B−1 r. (23)
design that minimizes the mean squared error E ∥e∥2 is
given in [10]. It was shown that the optimal training sequences The estimated channel response ĉ𝑗 is given by the first 𝐿
from different transmit antennas must satisfy entries of d̂𝑗 . When the training sequences are orthogonal as
in (16) with ℰ𝑠 = 1, the columns in A𝑘 are orthonormal. In
A†𝑘 A𝑖 = ℰ𝑠 𝛿(𝑘 − 𝑖)I, (16) ′
this case, one can choose the matrices A𝑘 in (22) such that B
where ℰ𝑠 is the signal power. That means the training se- is unitary. The channel estimate is thus given by d̂ = B† r. If in
quences from different transmit antennas must be orthogonal. addition, the orthogonal sequences are the optimal sequences
The general closed-form solution for x𝑖 that satisfies the proposed in [9], B can be chosen to be unitary and circulant.
orthogonality condition in (16) is derived in [10]. When the The unitary and circulant properties of B is shown below.
training sequences are orthogonal, the least-squares estimate In [9], the sequences for the 𝑖th transmit antenna are given
of channel response becomes by 𝑠𝑖 (𝑘) = 𝑠0 (𝑘)𝑒−𝑗 𝑀 𝜌𝑖𝑘 , where 𝑠0 (𝑘) are constant modu-
lus modulation symbols transmitted through the 0th transmit
ĉ = A† r. (17) antenna. Define two 𝑀 × 𝑀 diagonal matrices
In the following, we will describe the above estimation [ ]
S0 ≜ diag 𝑠0 (0) 𝑠0 (1) ⋅ ⋅ ⋅ 𝑠0 (𝑀 − 1) , (24)
method using a different formulation. This formulation is [ 2𝜋 2𝜋 ]
needed later for the joint estimation of I/Q imbalance, CFO Λ𝑖 ≜ diag 1 𝑒−𝑗 𝑀 𝜌𝑘 ⋅ ⋅ ⋅ 𝑒−𝑗 𝑀 (𝑀−1)𝜌𝑘 .(25)
and channel response. Define The 𝑀 × 𝑀 circulant matrix X𝑖 in (12) is given by
𝜌= . (18) X𝑖 = W† S0 Λ𝑖 W. (26)
In this paper, we assume that 𝑀 is a multiple of 𝑁𝑡 and the Denote the 𝑘th column of the DFT matrix W as w𝑘 . From
proposed method can be extended to the case where 𝑀 is (25), one can verify by direct multiplication that
not a multiple of 𝑁𝑡 by simple modification. Furthermore, we Λ𝑖 w𝑘 = w((𝑘+𝑖𝜌))𝑀 , (27)
assume that
[ ((∙))𝑀 means modulo 𝑀 . (27) ] implies that Λ𝑖 W =
𝜌 ≥ 𝐿, (19)
w((𝑖𝜌))𝑀 w((𝑖𝜌+1))𝑀 ⋅ ⋅ ⋅ w((𝑖𝜌−1))𝑀 ; the columns of Λ𝑖 W
so that the channel identifiability condition 𝑀 ≥ 𝑁𝑡 𝐿 is are simply those of W circularly shifted by 𝑖𝜌 columns.
guaranteed1. Suppose we append (𝜌−𝐿) zeros to the length-𝐿 From (26), we see that the 𝑘th column of X0 is equal to
vectors c𝑖 to obtain the 𝜌 × 1 vectors W† S0 w𝑘 (where Λ0 = I) and the 𝑘th column of X𝑖 is equal
[ ] to W† S0 (Λ𝑖 w𝑘 ). Using (27), we conclude that the columns
d𝑖 ≜ , for 𝑖 = 0, 1, . . . , 𝑁𝑡 − 1. (20) of X𝑖 are those of X0 circularly shifted by 𝑖𝜌 columns. That
1 Supposing 𝑁 OFDM blocks are available for training, the channel [ ]
x0,𝑖 x1,𝑖 ⋅ ⋅ ⋅ x𝑀−1,𝑖 =
identifiability condition becomes 𝑁 𝑀 ≥ 𝑁𝑡 𝐿. That means the symbol 𝜌 [ ]
satisfies 𝜌 ≥ 𝑁 𝐿
so that 𝑁 𝑀 ≥ 𝑁𝑡 𝐿 is achieved. x((𝜌𝑖))𝑀 ,0 x((𝜌𝑖+1))𝑀 ,0 ⋅ ⋅ ⋅ x((𝜌𝑖−1))𝑀 ,0 (28)

where x𝑘,𝑖 is the 𝑘th column of X𝑖 and x((𝑘))𝑀 ,0 is the Our goal is to find 𝛼 that minimizes the CRE. Since for most
((𝑘))𝑀 th column of X0 . Recall that A𝑖 are formed by the applications, 𝛼 is small, (33) can be approximated as

first 𝐿 columns of X𝑖 . If the 𝑀 × (𝜌 − 𝐿) matrix A𝑖 is

defined such that [ A𝑖 A𝑖 ] contains the first 𝜌 columns of CRE ≈ ∥PB−1 (z − 𝛼z∗ )∥2 . (34)
X𝑖 , using (28) we see that the 𝑀 × 𝑀 matrix B defined in From linear algebra, it is known that the optimal 𝛼 that
(22) is in fact equal to X0 , which is unitary and circulant. minimizes the CRE is
𝛼𝑜𝑝𝑡 = . (35)
In this section, we propose a new method to estimate the ∥PB−1 z∗ ∥2
channel response when there are CFO and I/Q imbalances. We By substituting 𝛼𝑜𝑝𝑡 into (30), we get the estimated MIMO
will first consider the simpler problem of the joint estimation channel response 𝜇d̂. For the compensation of I/Q imbalance,
of channel response and I/Q imbalance under the assumption one can employ (29) to obtain 𝜇r. Notice that there is no need
that there is no CFO. In this special case, the optimal solution to compensate the factor 𝜇 because it will be canceled when
is given in closed form. Then the joint estimation of the we use 𝜇d̂ to implement the FEQ. From (35), we see that to
channel response, CFO and I/Q imbalance will be studied. get 𝛼𝑜𝑝𝑡 , we only need to compute B−1 z (as B is fixed, B−1
Below we will show how to estimate 𝛼𝑘 and 𝜃𝑘 from one can be precomputed) and perform vector inner products at the
received vector z𝑘 at the 𝑘th receive antenna. For notational numerator and denominator2. When the training sequence in
simplicity, we will drop the receive antenna index 𝑘 as the [9] is used, B becomes unitary and circulant. As B−1 is also
problem can be solved separately for each receive antenna. circulant and unitary, B−1 z can be efficiently realized using
circular convolution.
A. Joint Estimation of Channel Response and I/Q Imbalance
In this subsection, we assume that there is no CFO. Hence B. Joint Estimation of Channel Response, I/Q Imbalance and
we have 𝜃 = 0 and E = I. From (11), 𝜇r is related to the CFO
received vector z as
z − 𝛼z∗ When the receiver suffers from both CFO and I/Q mismatch,
𝜇r = . (29) the received vector z is given by (8). From Sec. 2, we know
1 − ∣𝛼∣2
that if 𝜃 and 𝛼 are known, we can recover the desired baseband
From (23) and (29), if 𝛼 is given, an estimate of the MIMO vector from z using (11) and it is given by
channel response can be obtained as
[ ]𝑇 z − 𝛼z∗
𝜇r = E∗ 𝜇y = E∗ , (36)
𝜇d̂ = 𝜇 d̂𝑇0 d̂𝑇1 ⋅ ⋅ ⋅ d̂𝑇𝑁𝑡 −1 1 − ∣𝛼∣2
z − 𝛼z∗ where the diagonal matrix E is given in (5). We can obtain
= B−1 𝜇r = B−1 , (30)
1 − ∣𝛼∣2 an estimate of the MIMO channel response as
where B is defined in (22). When 𝛼 is estimated perfectly, the z − 𝛼z∗
first 𝐿 entries of each ĥ𝑗 in the above expression will give us 𝜇d̂ = B−1 𝜇r = B−1 E∗ . (37)
1 − ∣𝛼∣2
an estimate of the channel response and the last (𝜌−𝐿) entries
of d̂𝑗 are solely due to the channel noise. For moderately high From the above equation, when 𝛼 and 𝜃 are perfectly esti-
SNR, the energy of these entries should be small. Let us define mated, the last 𝜌 − 𝐿 entries of d̂𝑗 are again solely due to the
a quantity called the channel residual energy (CRE) as channel noise. By summing up the energy of these entries, we
have the CRE
∑ 𝜌−1
𝑡 −1 ∑

CRE ≜ ∣[𝜇d̂𝑙 ]𝑖 ∣2 , (31) CRE(𝛼, 𝜃) = ∥P𝜇d̂∥2 , (38)

𝑙=0 𝑖=𝐿
where P is defined in (32). Notice that the CRE is a function
where [𝜇d̂𝑙 ]𝑖 denotes the 𝑖th entry of 𝜇d̂𝑙 . Any error in the of both 𝛼 and 𝜃. Substituting (37) into the above equation, we
estimation of 𝛼 will increase the CRE (see the analysis at the can rewrite the CRE as
end of this section). Based on this observation, by minimizing  
the CRE we are able to estimate the I/Q parameter 𝛼 without  z − 𝛼z∗ 2
CRE(𝛼, 𝜃) = F  , (39)
knowing the channel response. To do this, we first define the 1 − ∣𝛼∣2 
(𝑀 − 𝑁𝑡 𝐿) × 𝑀 matrix
⎡ ⎤ where F = PB−1 E∗ . Following the argument in the previous
0 I𝜌−𝐿 0 ⋅⋅⋅ ⋅⋅⋅ 0 subsection, we get an estimate of 𝜃 and 𝛼 by minimizing the
⎢ 0 0 0 I𝜌−𝐿 0 ⋅⋅⋅ 0 ⎥ CRE. The joint optimization problem is solved in 2 steps: For
⎢ ⎥
P=⎢ . .. .. ⎥ . (32) a given 𝜃, we derive the optimal 𝛼, based on that we optimize
⎣ .. . ⋅⋅⋅ . ⎦
0 0 ⋅⋅⋅ ⋅⋅⋅ 0 I𝜌−𝐿 𝜃. Since ∣𝛼∣2 ≪ 1, the above expression can be approximated
Suppose that 𝜌 > 𝐿 so that P is not a zero matrix. Multiplying
𝜇d̂ by P, we can rewrite the CRE as CRE(𝛼, 𝜃) ≈ ∥Fz − 𝛼Fz∗ ∥2 . (40)
∗ 2
 −1 z − 𝛼z 
CRE = ∥P𝜇d̂∥ = PB  . (33)
2 If the real and imaginary parts of B−1 z are computed separately, one can

1 − ∣𝛼∣2  obtain B−1 z∗ by simple additions.


Given 𝜃, the optimal 𝛼 is given by V. J OINT E STIMATION U SING T WO R EPEATED T RAINING

∗ †
𝛼𝑜𝑝𝑡 (𝜃) = . (41) Suppose that two repeated training sequences are available.
∥Fz∗ ∥2
That means the training block is a (𝑀 +𝐿−1)×1 vector in the
Note that 𝛼𝑜𝑝𝑡 (𝜃) is a function of 𝜃 because F depends on 𝜃. [ ]𝑇
form of CP x𝑇𝑖 x𝑇𝑖 , where the training sequence x𝑖
Substituting 𝛼𝑜𝑝𝑡 (𝜃) into (40), the CRE can be written as
is an 𝑀2 × 1 vector, and the CP length is 𝐿 − 1. This repeated
∣(Fz)† (Fz∗ )∣2 structure has been proposed in [5] to solve the problem of
CRE(𝜃) = ∥Fz∥2 − . (42)
∥Fz∗ ∥2 CFO estimation. In what follows, we will exploit the repeated
Then the optimal estimate of CFO is given by structure to solve the joint estimation of CFO, I/Q and channel
response. Suppose that there are CFO and I/Q mismatch. From
𝜃𝑜𝑝𝑡 = arg min CRE(𝜃). (43) (8), the two received 𝑀 2 × 1 vectors are in the form of

Once the optimal 𝜃𝑜𝑝𝑡 is obtained from the above optimization, z𝑎 = 𝜇y + 𝜈y∗ + q𝑎 , (48)
the optimal 𝛼𝑜𝑝𝑡 can be obtained by substituting 𝜃𝑜𝑝𝑡 into (41) z𝑏 = 𝜇𝑒𝑗𝜋𝜃 y + 𝜈(𝑒𝑗𝜋𝜃 y)∗ + q𝑏 , (49)
and the estimated channel response is found by substituting
𝛼𝑜𝑝𝑡 and 𝜃𝑜𝑝𝑡 into (37). Then we can use (36) for symbol re- where 𝑎 and 𝑏 are the OFDM block indexes and y is an ×1 2
covery. Note that no iteration is needed in the above optimiza- vector in (4). Our goal is to jointly estimate CFO, I/Q and
tion process. However, a one-dimensional search is needed to channel response from z𝑎 and z𝑏 . Below we will first show
obtain the CFO estimate. In many practical applications, the how to solve the two subproblems: (A) given 𝛼, estimate 𝜃 and
training data often consist of repeated sequences. In this case, (B) given 𝜃, estimate 𝛼 and ℎ(𝑛). Then the joint estimation
the one-dimensional search problem in (42) can be avoided of 𝛼, 𝜃 and ℎ(𝑛) will be solved by a two-step approach.
and the joint optimization problem can be solved efficiently (A) Given 𝛼, Estimate CFO 𝜃: Rearranging (48) and (49),
using a two-step approach as demonstrated later. we can obtain
An Analysis of CRE: In the following, we assume that 𝛼 z𝑎 − 𝛼z∗𝑎
is small so that the second order term can be ignored. Also we 𝜇y = , (50)
1 − ∣𝛼∣2
assume that SNR is moderately high so that 𝛼q can be ignored z𝑏 − 𝛼z∗𝑏
in the analysis below. Let the estimate of 𝛼 and 𝜃 be 𝛼 ˆ and 𝑒𝑗𝜋𝜃 𝜇y = . (51)
1 − ∣𝛼∣2
𝜃ˆ respectively. Replacing 𝛼 in (40) and 𝜃 in F with 𝛼 ˆ and 𝜃ˆ
respectively and substituting the relation z = E𝜇r + 𝛼(E𝜇r)∗ So given 𝛼, one can estimate CFO by
into (40), we approximate the CRE as 1 { }
 𝜃ˆ = angle (z𝑎 − 𝛼z∗𝑎 )† (z𝑏 − 𝛼z∗𝑏 ) , (52)
CRE ≈ PB−1 Ê∗ EB(𝜇d)
2 where we have used the fact that (𝜇y)† (𝜇y) and (1 − ∣𝛼∣2 )2

+(𝛼 − 𝛼 ˆ)PB−1 Ê∗ E∗ B∗ (𝜇d)∗ + PB−1 Ê∗ E(𝜇q)(44) . are both positive. Notice that the estimate CFO 𝜃ˆ in the above
equation is in the range of −1 ≤ 𝜃ˆ < 1. The case where the
Define the estimation errors Δ𝛼 ≜ 𝛼 − 𝛼 ˆ and Δ𝜃 ≜ 𝜃 − CFO is outside this range (i.e. ∣𝜃∣ > 1) will be discussed later.
ˆ Assume Δ𝜃 is small so we can make the approximation
Ê∗ E ≈ I + Δ𝜃Υ, where Υ is 𝑀 × 𝑀 diagonal with the 𝑘th (B) Given 𝜃, Estimate 𝛼 and ℎ(𝑛): Using (37), we obtain
entry 𝑗 2𝜋𝑘 two independent estimates of 𝜇d from (50) and (51) as
𝑀 . Then we can rewrite the CRE as
 z𝑎 − 𝛼z∗𝑎
CRE ≈ P(𝜇d) + Δ𝜃PB−1 ΥB(𝜇d) 𝜇d̂𝑎 = B−1 E∗ (53)
2 1 − ∣𝛼∣2

+ Δ𝛼PB−1 (E∗ )2 B∗ (𝜇d)∗ + PB−1 Ê∗ (𝜇q) .(45) z𝑏 − 𝛼z∗𝑏
𝜇d̂𝑏 = 𝑒−𝑗𝜋𝜃 B−1 E∗ , (54)
1 − ∣𝛼∣2
Using (32), (20) and (21), it is found that P(𝜇d) = 0. Suppose
that the noise samples are i.i.d. complex random variables with where B and E are defined in (22) and (5) respectively (except
zero-mean and variance 𝒩0 . The mean value of CRE is given that their dimensions are 𝑀 𝑀
2 × 2 now). From the above two
by equations, we define the CRE as
[ ]
[ ] Δ𝜃 CRE𝑎 = ∥P𝜇d̂𝑎 ∥2 and CRE𝑏 = ∥P𝜇d̂𝑏 ∥2 , (55)
𝐸 {CRE} = Δ𝜃 Δ𝛼∗ V† V
respectively. Our goal is to find 𝛼 by minimizing the CRE𝑎
+ ∣𝜇∣2 (𝑀 − 𝑁𝑡 𝐿)𝒩0 , (46)
[ ] and CRE𝑏 . Substituting (53) and (54) into the CRE𝑎 and CRE𝑏
where V ≜ PB−1 ΥB(𝜇d) PB−1 (E∗ )2 B∗ (𝜇d)∗ . and using (41), we obtain two independent estimates of 𝛼 as
From the right hand side of (46), we see that the CRE contains (Fz∗𝑎 )† (Fz𝑎 ) (Fz∗𝑏 )† (Fz𝑏 )
two parts: the first term is channel dependent and the second ˆ𝑎 =
𝛼 , and 𝛼
ˆ 𝑏 = , (56)
∥Fz∗𝑎 ∥2 ∥Fz∗𝑏 ∥2
term is the noise term, which is independent of 𝛼 and 𝜃. Since
V† V is in general positive definite, we have where F = PB−1 E∗ . By taking the average, one can estimate
𝛼 as
𝐸 {CRE} ≥ ∣𝜇∣2 (𝑀 − 𝑁𝑡 𝐿)𝒩0 , (47)
where the equality is achieved if and only if Δ𝛼 = Δ𝜃 = 0. ˆ = (ˆ
𝛼 𝛼𝑎 + 𝛼ˆ𝑏 ). (57)

Substituting 𝛼
ˆ into (53) and (54), we obtain two estimates of −2
channel response 𝜇d̂𝑎 and 𝜇d̂𝑏 . By taking the average, we
have the estimated channel vector
1 10
𝜇d̂ = (𝜇d̂𝑎 + 𝜇d̂𝑏 ). (58)
(C) CFO with ∣𝜃∣ > 1: When the CFO is outside the range 10

of [−1, 1), we can write it as 𝜃 = 𝜃ˆ+ 2𝜏 where 𝜏 is an integer
and 𝜃ˆ is in the range of [−1, 1). Then 𝜃ˆ can be estimated by −5
using (52). To get an estimate of 𝜏 , we can substitute 𝜃ˆ +
2𝜏 into (55). Using (42), we obtain different CRE values for −6
different 𝜏 . Using the observation that when 𝜏 is estimated 10 Proposed (Sec. 4)
perfectly, the CRE should be a minimum, the estimate of 𝜏 is Two−Step (Sec. 5)
therefore given by −7
0 5 10 15 20 25 30 35
𝜏ˆ = arg min (CRE𝑎 + CRE𝑏 ) . (59) SNR(dB)
𝜏 ∈integer
Notice that in practice, 𝜏 is usually a small integer. We only
need to search for 𝜏 within a narrow range. 10

(D) Two-Step Approach for the Joint Estimation: Using

the results in (A), (B) and (C), we propose a two-step approach 10
to solve the joint estimation problem.
Step One: Obtain the initial estimates. −4
MSE(I/Q) 10
1. In practice, the value of 𝛼 is usually small. So we first
assume that 𝛼 ≈ 0. Substituting 𝛼 = 0{into (52), } we obtain −5
an initial estimate of 𝜃 as 𝜃ˆ0 = 𝜋1 angle z†𝑎 z𝑏 . 10

2. Estimate 𝜏 of CFO using (59). Replace 𝜃ˆ0 with 𝜃ˆ0 + 2ˆ𝜏.

3. Obtain an initial estimate of I/Q parameter 𝛼ˆ0 by substi- 10 Proposed (Sec. 4)
ˆ 0
tuting 𝜃 into (56) and taking the average as in (57). Two−Step (Sec. 5)
Step Two: Refine the estimates. −7
0 5 10 15 20 25 30 35
4. Obtain 𝜃ˆ by substituting 𝛼 ˆ0 into (52). Replace 𝜃ˆ with SNR(dB)
𝜃 + 2ˆ𝜏.
ˆ by substituting 𝜃ˆ in (56) and taking the average
5. Obtain 𝛼
as in (57). Fig. 2. MSE of I/Q in (a) case A; (b) case B.
6. Obtain 𝜇d̂ by substituting 𝜃ˆ and 𝛼
ˆ into (53) and (54) and
taking the average as in (58). ∑𝐿−1 { }
is normalized by 𝑙=0 E ∣ℎ𝑘,𝑗 (𝑙)∣2 = 1 for all 𝑘, 𝑗. The
(E) Complexity: From (52)-(59) and (42), the main com-
channel noise is AWGN. The training data are QPSK symbols.
putation is the calculations of the vectors B−1 E∗ z𝑖 and
The training sequences in the experiments are the optimal
B−1 E∗ z∗𝑖 . All the other computations are scalar multipli-
sequences in [9]. The size of the DFT matrix is 𝑀 = 1024.
cations and vector additions. The numerical search in (43)
The CP length is 𝐿 − 1 = 64.
is avoided. As mentioned at the end of Sec. 3, when the
Two cases of different parameters are considered:
training sequence is the orthogonal sequence proposed in [9],
the matrix B becomes unitary and circulant. The matrix B−1 (A) 𝑁𝑡 = 2, 𝑁𝑟 = 1, 𝜖 ∼ 𝒰 [1, 1.1], 𝜙 ∼ 𝒰 [0, 10∘ ], 𝜃 ∼ 𝒰 [−1, 1],
can be implemented efficiently using circular convolution with (B) 𝑁𝑡 = 4, 𝑁𝑟 = 1, 𝜖 ∼ 𝒰 [1, 1.1], 𝜙 ∼ 𝒰 [0, 10∘ ], 𝜃 ∼ 𝒰 [−3, 3],
a complexity in the order of 𝑀 log2 𝑀 . Compared with [7], where 𝒰 [𝑥, 𝑦] denotes the uniform distribution between 𝑥 and
our proposed method needs one more DFT operation than the 𝑦. Note that Case B has a wider range for the CFO parameters.
IQ-CFO-FD method [7].
The MSE performance is shown in Fig. 2-4. For comparison,
we also plot the MSE of the IQ-CFO-FD method proposed in
VI. S IMULATION R ESULTS [7]. It needs 𝑁𝑡 OFDM blocks for training.
In this section, we carry out Monte-Carlo experiments to The two proposed methods and the IQ-CFO-FD method
verify the performance of the proposed methods. We denote use different training sequences. The 3 different sequences
the proposed method in Sec. 4.2 as “Proposed (Sec. 4)” and are listed in Table 1. The first proposed method in Sec. 4 uses
the proposed method in Sec. 5 as “Two-Step (Sec. 5)”. A Type I sequence and the second proposed method in Sec. 5
total of 5000 random channels for each transmit and receive uses Type II sequence. The IQ-CFO-FD method uses Type III
antenna pair are generated in the experiments. The channel sequence since it is actually based on SISO technique.
taps are i.i.d. complex Gaussian random variables and the Fig. 2(a) and (b) show the MSEs of the I/Q parameters
channel length is 𝐿 = 65. The variance of the channel taps estimation for Case A and Case B. From these figures, we


Ant0 𝑎0 ⋅⋅⋅ 𝑎 𝑀 −1 𝑎𝑀 ⋅⋅⋅ 𝑎𝑀 −1

2 2
Type I Ant1 𝑏0 ⋅⋅⋅ 𝑏 𝑀 −1 𝑏𝑀 ⋅⋅⋅ 𝑏𝑀 −1
2 2
Ant0 𝑎0 ⋅⋅⋅ 𝑎 𝑀 −1 𝑎0 ⋅⋅⋅ 𝑎 𝑀 −1
2 2
Type II Ant1 𝑏0 ⋅⋅⋅ 𝑏 𝑀 −1 𝑏0 ⋅⋅⋅ 𝑏 𝑀 −1
2 2
′ ′ ′ ′
Ant0 𝑎0 ⋅⋅⋅ 𝑎𝑀 𝑎0 ⋅⋅⋅ 𝑎𝑀 0 ⋅⋅⋅ 0 0 ⋅⋅⋅ 0
−1 2
′ ′ ′ ′
Type III Ant1 0 ⋅⋅⋅ 0 0 ⋅⋅⋅ 0 𝑏0 ⋅⋅⋅ 𝑏𝑀 𝑏0 ⋅ ⋅ ⋅ 𝑏𝑀
−1 −1
′ √ 2
′ √ 2
𝑎𝑘 = 2𝑎𝑘 𝑏𝑘 = 2𝑏𝑘

−3 −2
10 10
Proposed (Sec. 4) Proposed (Sec. 4)
Two−Step (Sec. 5) Two−Step (Sec. 5)
−4 IQ−CFO−FD [7] 10 IQ−CFO−FD [7]
CRB [8][16], Type I CRB [8][16], Type I
CRB [8][16], Type II CRB [8][16], Type II

−5 CRB [8][16], Type III 10 CRB [8][16], Type III

−6 10

−8 0 5 10 15 20 25 30 35
10 SNR(dB)
0 5 10 15 20 25 30 35

0 −2
10 10

Proposed (Sec. 4)

10 Two−Step (Sec. 5) −4
CRB [8][16], Type I
CRB [8][16], Type II Proposed (Sec. 4)
−6 Two−Step (Sec. 5)
CRB [8][16], Type III 10 IQ−CFO−FD [7]
CRB [8][16], Type I
10 CRB [8][16], Type II
−8 CRB [8][16], Type III
0 5 10 15 20 25 30 35
10 SNR(dB)
0 5 10 15 20 25 30 35
SNR(dB) (b)
(b) Fig. 4. MSE of channel responses in (a) case A; (b) case B.
Fig. 3. MSEs of CFO in (a) case A; (b) case B.

flat. Figures 3 and 4 show the MSEs of the estimation of

channel response and CFO. In these figures we also plot the
found that our proposed methods provide a good performance, Cramer-Rao bound (CRB) on the CFO and channel response
and they are robust for different mismatched parameters. It is estimation in [8] and [16]. From Fig. 3(a) and (b), we see
also found that the proposed methods outperform the IQ-CFO- that our proposed methods provide a good performance for
FD method in both cases. Note that the IQ-CFO-FD method the estimation of CFO and their performance is very close
suffers an error-flooring in high SNR. This is because among the CRB in both cases. Note that the IQ-CFO-FD method can
5000 random channels, some do not have smooth frequency not estimate the CFO outside the range of [−1, 1) in Case B.
response and for these channels, the smoothness assumption Fig. 4(a) and (b) show the MSE of channel estimation. For
in [7] is not satisfied. As the results, the MSEs of [7] become Case A and Case B, both our proposed methods provide a

performance close to the CRB. Notice that the MSE of Case [13] A. Tarighat and A. H. Sayed, “MIMO OFDM receivers for systems
B is larger than the MSE of Case A since there are more with I/Q imbalance," IEEE Trans. Signal Process., vol. 53, no. 9, pp.
3583-3596, Sep. 2005.
transmit antennas and thus there are more channel taps in the [14] Y. Tanabe, Y. Egashira, T. Aoki, and K. Sato, “Suitable MIMO-OFDM
estimation. Next we compare our methods with the IQ-CFO- decoders to compensate IQ imbalance," in Proc. IEEE WCNC, Mar.
FD method. In Case A, the IQ-CFO-FD method performs very 2007.
well when the SNR is low. In Case B, the proposed methods [15] T. C. W. Schenk, P. F. M. Smulders, and E. R. Fledderus, “Estimation
and compensation of Tx and Rx I/Q imbalance in OFDM-based MIMO
are significantly better than the IQ-CFO-FD method. systems," in Proc. IEEE ICC, June 2006.
[16] O. Besson and P. Stoica, “On parameter estimation of MIMO flat-fading
VII. C ONCLUSION channels with frequency offsets," IEEE Trans. Signal Process., Mar.
In this paper, we propose new methods for the joint esti- [17] M. Ghogho and A. Swami, “Training design for multipath channel
mation of the I/Q imbalance, CFO and channel response for and frequency-offset estimation in MIMO systems," IEEE Trans. Signal
Process., Oct. 2006.
MIMO OFDM systems by using training sequences. When [18] H. Minn, N. Al-Dhahir, and Y. Li, “Optimal training signals for MIMO
only one OFDM block is available for training, the first OFDM channel estimation in the presence of frequency offset and phase
method is able to give an accurate estimate of the CFO, noise," IEEE Trans. Commun., vol. 54, no. 10, pp. 1754-1759, Oct. 2006.
[19] K. Lee and J. Chun, “Frequency-offset estimation for MIMO and
I/Q parameter and channel responses. The CFO is obtained OFDM systems using orthogonal training sequences," IEEE Trans. Veh.
through one-dimensional search algorithm. When two repeated Technol., Jan. 2007.
OFDM blocks are available for training, a low complexity two- [20] X. Ma, M.-K. Oh, G. B. Giannakis, and D.-J. Park, “Hopping pilots for
estimation of frequency-offset and multiantenna channels in MIMO-
step approach is proposed to solve the joint estimation prob- OFDM," IEEE Trans. Commun., vol. 53, no. 1, pp. 162-172, Jan. 2005.
lem. Simulation results show that the MSEs of the proposed [21] F. Simoens and M. Moeneclaey, “Reduced complexity data-aided and
methods are very close to the CRB. code-aided frequency offset estimation for flat-fading MIMO channels,"
IEEE Trans. Wireless. Commun., vol. 5, no. 6, June 2006.
[22] Y.-H. Chung and S.-M. Phoong, “Joint estimation of I/Q imbalance and
ACKNOWLEDGEMENT channel response for MIMO OFDM system," in Proc. EUSIPCO, Sept.
The authors would like to thank all the reviewers for their 2007.
helpful comments and suggestions.
Yuan-Hwui Chung (S’06) was born in Tainan,
R EFERENCES Taiwan, in 1979. He received the B. S. degree in
electrical engineering from National Cheng Kung
[1] A. Tarighat, A. H. Sayed, “Compensation schemes and performance University (NCKU), Tainan, Taiwan, in 2002 and
analysis of I/Q imbalances in OFDM receivers," IEEE Trans. Signal M.S. degree and Ph.D. degree in electrical engineer-
Process., Aug. 2005. ing and communication engineering from National
[2] M. Windisch and G. Fettweis, “Preample design for an efficient I/Q Taiwan University (NTU), Taipei, Taiwan, in 2004
imbalance compensation in OFDM direction-conversion receivers," in and 2009 respectively. He is now a Postdoctoral Fel-
Proc. 10th Intl. OFDM Workshop, Aug./Sep. 2005. low at NTU. His current research interests include
[3] J. Tubbax, B. Come, L. V. der Perre, S. Donnay, M. Engels, H. D. Man, signal processing for MIMO communications and
and M. Moonen, “Compensation of IQ imbalance and phase noise in OFDM systems.
OFDM systems," IEEE Trans. Wireless Commun., May 2005.
[4] W.-J. Cho, T.-K. Chang, Y.-H. Chung, S.-M. Phoong, and Y.-P. Lin,
“Frame synchronization and joint estimation of IQ imbalance and See-May Phoong (M’96-SM’04) was born in Johor,
channel response for OFDM systems," in Proc. IEEE ICASSP, Mar. Malaysia, in 1968. He received the B.S. degree
2008. in electrical engineering from the National Taiwan
[5] P. H. Moose, “A technique for orthogonal frequency division multplex- University (NTU), Taipei, Taiwan, R.O.C., in 1991
ing frequency offset correction," IEEE Trans. Commun., vol. 42, no. 10, and M.S. and Ph.D. degrees in electrical engineer-
pp. 2908-2914, Oct. 1994. ing from the California Institute of Technology
[6] T. M. Schmidl and D. C. Cox, “Robust frequency and timing synchro- (Caltech), Pasadena, California, in 1992 and 1996,
nization for OFDM," IEEE Trans. Commun., vol. 45, no. 12, pp. 1613- respectively.
1621, Dec. 1997. He was with the Faculty of the Department
[7] J. Tubbax, A. Fort, L. Van der Perre, S. Donnay, M. Engels, M. Moonen, of Electronic and Electrical Engineering, Nanyang
and H. De Man, “Joint compensation of I/Q imbalance and frequency Technological University, Singapore, from Septem-
offset in OFDM systems," in Proc. IEEE Globecom, 2003. ber 1996 to September 1997. In September 1997, he joined the Graduate
[8] M. Morelli and U. Mengali, “Carrier-frequency estimation for transmis- Institute of Communication Engineering and the Department of Electrical
sions over selective channels," IEEE Trans. Commun., Sep. 2000. Engineering, NTU, as an Assistant Professor, and since August 2006, he has
[9] Y. Li, “Simplified channel estimation for OFDM systems with multiple been a Professor.
transmit antennas," IEEE Trans. Wireless Commun., vol. 1, no. 1, pp. Dr. Phoong has previously served as an Associate Editor for IEEE
[10] I. Barhumi, G. Leus, and M. Moonen, “Optimal training design for S IGNAL P ROCESSING (Jan. 2002 – Dec. 2003), IEEE S IGNAL P ROCESSING
MIMO OFDM systems in mobile wireless channels," IEEE Trans. L ETTERS (March 2002 - Feb. 2005) and IEEE T RANSACTIONS ON C IRCUITS
Signal Process., June 2003. AND S YSTEMS I: R EGULAR PAPERS (Jan. 2006 - Dec. 2007). His interests
[11] H. Minn and N. Al-Dhahir, “Optimal training signals for MIMO OFDM include multirate signal processing, filter banks and their application to com-
channel estimation," in IEEE Trans. Wireless Commun., May 2006. munications. He received the Charles H. Wilts Prize (1997) for outstanding
[12] S. Sun, I. Wiemer, C. K. Ho, and T. T. Tjhung, “Training sequence independent research in electrical engineering at Caltech. He was also a
assisted channel estimation for MIMO OFDM," in Proc. IEEE WCNC, recipient of the Chinese Institute of Electrical Engineering’s Outstanding
Mar. 2003. Youth Electrical Engineer Award (2005).