Professional Documents
Culture Documents
From (3), we see that the ideal baseband signal r(n) is inter- 3. MIMO CHANNEL ESTIMATION
fered by its own complex conjugate due to the I/Q imbalance.
When the I/Q imbalance is vary small, i.e., ε ≈ 1 and φ ≈ 0, 3.1 Optimal Orthogonal Sequences
µ is close to 1 and ν is close to zero. Then the interference We consider MIMO channel estimation using training se-
from the complex conjugate r∗ (n) becomes negligible. quences. Given xi , we want to estimate the channel response
from the received vector rk . For the purpose of channel esti-
2.2 MIMO OFDM Systems mation, we can rewrite (5) as
Fig. 1 shows an MIMO OFDM system where the number of hk,0
transmit and receive antennas are Nt and Nr respectively. Let hk,1
the channel impulse response from the jth transmit antenna rk = [ X0 X1 · · · XNt −1 ] .. + qk ,
to the kth receive antenna be hk, j (n). We assume that the .
lengths of hk, j (n) are ≤ L and the length of cyclic prefix (CP) hk,Nt −1
is L − 1. Let M be the DFT size so that the vectors si , xi and (11)
ri in Fig. 1 are M × 1. In practice, M is usually much larger
than L for the bandwidth efficiency. Due to the assumptions where hk, j is the M × 1 vector defined in (6) and X j is an
on the CP length and channel lengths, the received vector at M × M circulant matrix with the first column x j . Define the
the kth receive antenna (after CP removal) can be written as M × L submatrix A j which consists of the first L columns
of X j and let ck, j = [ hk, j (0) hk, j (1) · · · hk, j (L − 1) ]T .
Then (11) can be rewritten as
x0
x1
rk = [ Hk,0 Hk,1 ··· Hk,Nt −1 ] .. + qk , (5) ck,0
. ck,1
xNt −1 rk = [ A0 A1 · · · ANt −1 ] .. + qk
.
ck,Nt −1
where Hk, j is an M ×M circulant matrix with the first column
= Ack + qk . (12)
hk, j = [ hk, j (0) hk, j (1) · · · From (12), we see that the channel vector ck is identifiable if
hk, j (L − 1) 0 ··· 0 ]T . (6) and only if the M × LNt matrix A has full column rank. Thus
a necessary condition for the channel identifiability is M ≥
The M × 1 vector qk is the AWGN vector. After the DFT LNt . Assume that qk is a complex Gaussian random vector
operations, one can employ a frequency domain equalizer with covariance matrix Rqk = N0 I. Then, a least squares
(FEQ) to recover the transmitted vector ŝk (see Fig. 1). When estimator of ĉk is given by
there is I/Q imbalance, the received vectors are no longer
given by rk in (5). In the presence of I/Q imbalance, the I/Q ĉk = (A† A)−1 A† rk . (13)
corrupted received vector at the kth receive antenna becomes Define the error vector as ek = ĉk − ck . The design of
[7] optimal sequences that minimize the mean squared error
Nt −1
∑k=0 Ekek k2 is given in [4]. It was shown that the optimal
zk = µk rk + νk r∗k . (7) training sequences from different antennas must satisfy
Using (5), we can write A†k Ai = δ (k − i)I. (14)
Nt −1 Nt −1 That means the training sequences from different transmit
zk = µk ∑ H j x j + νk ∑ H∗j x∗j + µk qk + νk q∗k . (8) antennas must be orthogonal. The general closed-form solu-
j=0 j=0 tions have been derived in [4].
From (8), we find that the interferences consist of not only 3.2 Circularly-Shift Orthogonal Sequences
the vector x j but also its complex conjugate x∗j . Define
In this paper, we propose a special type of orthogonal se-
quences for the estimation of MIMO channels. Due to the
νk space limit, we consider the case where M is a multiple of
αk = . (9)
µk∗ Nt . The general case is given in [10]. The training sequence
sent by the ith transmit antenna is given by
Then if αk is correctly estimated at the receiver, we can com- µµ ¶¶
pensate the I/Q imbalance either in the time domain [2] or in M
xi (n) = x0 n− i , (15)
the frequency domain [1]. It can be verified that Nt M
zk − αk z∗k where x0 (n) is the training sequence sent by the 0th trans-
µ k rk = . (10) mit antenna and its DFT coefficients satisfy |s0 (k)| = √1M for
1 − |αk |2
all k. In other words, the training sequences for all trans-
In Sec. 4, we will discuss how to jointly estimate the chan- mit antennas are obtained by circularly shifting the training
nel coefficients hk, j (n) and the I/Q imbalance factor αk using sequence of the 0th transmit antenna. These circularly-shift
training sequences. sequences satisfy the following property:
Lemma 1. When the training sequences are chosen as (15), Then we can rewrite the received vector rk at the kth receive
the corresponding matrices Ai satisfy the orthogonality con- antenna in (11) as
dition in (14)
dk,0
Proof. Let si (k) be the M-point DFT coefficients of xi (n). dk,1
We have rk = [ B0 B1 ··· BNt −1 ]
.. + qk ,
.
2π
si (k) = s0 (k)e− j Nt ik , for i = 1, 2, . . . , Nt − 1. (16) dk,Nt −1
(21)
Notice that for all i |si (k)| = √1M because |s0 (k)| = √1M for
where Bi is the M × NMt submatrix consisting of the first NMt
all k. Partition the DFT matrix as W = [ W0 W1 ], where
W0 consists of the first L columns of W. Since Ak is a sub- columns of the M ×M circulant matrix Xi . Define the M ×M
matrix of Xk and the circulant matrix Xk can be diagonalized matrix B = [ B0 · · · BNt −1 ] and the M × 1 vector dk =
by the DFT/IDFT matrices, Ak can be represented as [ dTk,0 · · · dTk,Nt −1 ]T . Then we have
where Sk is an M × M diagonal matrix with the diagonal en- If the training sequences are chosen as the circularly-shift
tries sk (n). From (16) and (17), the (l, m)th entry of the ma- orthogonal sequences in (15), then we can have the following
trix A†k Ai is given by lemma (a proof is given in [10])
4. JOINT ESTIMATION OF I/Q IMBALANCE AND is AWGN. The transmission and training data are both QPSK
CHANNEL RESPONSE symbols. The training symbols are randomly generated at
each estimation. The size of the DFT matrix is M = 64. To
When there is I/Q imbalance, the received vector at the kth avoid inter-block interference, the CP length is L − 1 = 3.
receive antenna is zk , not rk . The channel coefficients ob-
First we consider the following mean squared errors
tained by B† zk will not be an accurate estimate due to the
(MSE) for the estimation of the channel response and I/Q
I/Q imbalance. Below we will show how to estimate the I/Q
imbalance respectively:
imbalance factor αk and the MIMO channels hk, j (n) from
only one received vector zk . Recall that the vector µk rk can
be represented as (10). Let α̂k be an estimate of αk . Then 1 Nr −1 Nt −1 L−1
from (23) the channel estimator d̂k can be obtained as
MSE(channel) = ∑ ∑ ∑ E|hk, j (l) − ĥk, j (l)|2
Nt Nr L k=0 j=0 l=0
½ ¾
zk − α̂k z∗k νk 2
µk d̂k = B† (µk rk ) = B† . (25) MSE(IQ) = E |αopt − ∗ | . (30)
1 − |α̂k |2 µk
When α̂k ≈ αk , the I/Q imbalance can mostly be compen- Fig. 2 shows the MSEs of the estimation of the channel
sated and (25) will give an accurate estimate of the MIMO response and I/Q imbalance. We consider two cases: (i)
channels. In other words, when α̂k = αk , the last ( NMt − L) Nt = 2, Nr = 1 and (ii) Nt = 4, Nr = 1. Three different mis-
entries of d̂k, j defined in (24) are solely due to noise and any match parameter pairs of I/Q imbalance are considered for
error in the estimation of αk will increase the energy of these both cases: (1) ε = 1, φ = 0, (2) ε = 1.1, φ = 10◦ and (3)
M ε = 1.2, φ = 15◦ . When ε = 1 and φ = 0◦ , there is no I/Q
Nt − L entries. Using this observation, we are able to esti-
mate αk accurately as we will demonstrate below. In many imbalance in the system. From Fig. 2(a), we see that for
applications, |αk |2 is small. Then the channel estimator in all cases, the I/Q imbalance factor αk can be accurately esti-
(25) can be approximated as mated. For a SNR of 30dB, the MSE is 10−5 . Notice that the
accuracy of our method is independent of the values of ε and
µk d̂k ≈ B† (zk − α̂k z∗k ) . (26) φ . The 4 transmit antenna case is only slightly worse than the
2 transmit antenna case. Fig. 2(b) shows the corresponding
Define the (M − Nt L) × M matrix Θ as MSEs of the channel estimation. Again we observe that the
accuracy is independent of ε and φ . Comparing the cases of
0 I M −L 0 ··· ··· 0 2Tx and 4Tx, it is seen that the 4Tx case is worse than the
Nt
0 0 0 I M −L 0 ··· 0 2Tx case because we need to estimate 4 channels in the 4Tx
Θ= . Nt
. (27) case rather than 2 channels in the 2Tx case.
.. ..
.. . ··· . Fig. 3 shows the BER performance of OFDM systems of
0 0 ··· ··· 0 I M −L three cases: (i) Nt = Nr = 1, (ii) Nt = Nr = 2 and (iii) Nt =
Nt
Nr = 4. The mismatch parameters for the three cases are:
M
Multiplying d̂k by Θ will collect all the last ( Nt − L) entries (i) ε0 = 1.1, φ0 = 10◦ , (ii) ε0 = ε1 = 1.1, φ0 = 5◦ , φ1 = 10◦ ,
and (iii) ε0 = ε1 = ε2 = ε3 = 1.1, φ0 = 2.5◦ , φ1 = 5◦ , φ2 =
of d̂k,i . From (26), we have 7.5◦ , φ3 = 10◦ . Zero-forcing FEQs are used at the receiver.
For the purpose of symbol recovery, the number of antennas
Θd̂k ≈ ΘB† (zk − α̂k z∗k ). (28) should satisfy Nr ≥ Nt and in our simulation, we let Nr = Nt .
The BER performance of the systems with I/Q imbalance is
We want to find αk such that kΘd̂k k2 is minimized. From also given in the figure. It can be found that the performance
the theory of linear algebra, it is known that the optimal αk is seriously degraded and the BER suffers an error flooring.
is given by In the same figure, we also show the BER performance of
† ∗ †
(ΘB zk ) (ΘB zk ) † the ideal case where the receiver knows the exact value of
αk,opt = ∗ ∗ . (29) channel coefficients and αk . From the figure, we see that for
(ΘB† zk )† (ΘB† zk ) all cases, the BER performance of the proposed method is
Once αk,opt is obtained, one can get the estimate of MIMO very close to the ideal case. Note that the 4Tx-4Rx case is
worse than the 2Tx-2Rx case and 1Tx-1Rx case because of
channels, µk d̂k , using (25). For the compensation of I/Q im- higher data rate.
balance, one can employ (10) to obtain µk rk . Notice that
there is no need to compensate the factor µk because the fac-
6. CONCLUSION
tor µk in (µk rk ) will be canceled when we use (µk d̂k ) to im-
plement the FEQ. In this paper, we proposed a method for jointly estimating
the I/Q imbalance and channel response for MIMO OFDM
5. SIMULATION RESULTS systems. The proposed method can accurately estimate both
the I/Q imbalance and channel response. The implementa-
In this section, we carry out Monte-Carlo experiments to ver-
tion complexity of this method is very low. Moreover, the
ify the performance of the proposed method. A total of 5000
proposed method can be extended for jointly estimating the
random channels for each transmit and receive antenna pair
carrier frequency offset, I/Q imbalance and channel response
are generated in the simulation process. Each transmit and
[10].
receive channel pair are i.i.d. complex Gaussian random vari-
ables with 4 taps, and the variance of channel taps is normal-
ized by ∑L−1 2
l=0 E{|hk, j (l)| } = 1 for all k, j. The channel noise REFERENCES
−1 0
10 10
°
2Tx−1Rx ε=1 φ=0
−2 2Tx−1Rx ε=1.1 φ=10°
10
°
2Tx−1Rx ε=1.2 φ=15 −1
10
°
4Tx−1Rx ε=1 φ=0
−3
10 °
4Tx−1Rx ε=1.1 φ=10
°
4Tx−1Rx ε=1.2 φ=15 −2
−4 10
10
MSE(IQ)
BER
−5
10 −3
10 1Tx−1Rx proposed
1Tx−1Rx ideal
−6 1Tx−1Rx no compensation
10 2Tx−2Rx proposed
−4 2Tx−2Rx ideal
10 2Tx−2RX no compensation
−7
10 4Tx−4Rx proposed
4Tx−4Rx ideal
4Tx−4Rx no compensation
−8 −5
10 10
0 10 20 30 40 50 0 5 10 15 20 25 30 35
SNR(dB) SNR(dB)
(a)
Figure 3: BER performance of MIMO OFDM systems
−1
10
2Tx−1Rx ε=1 φ=0°
2Tx−1Rx ε=1.1 φ=10°
−2
°
10 2Tx−1Rx ε=1.2 φ=15 MIMO OFDM channel estimation,” in Proc. IEEE Global
°
4Tx−1Rx ε=1 φ=0 Telecommun. Conf., 2004.
°
4Tx−1Rx ε=1.1 φ=10
−3
10
4Tx−1Rx ε=1.2 φ=15°
[6] S. Sun, I. Wiemer, C. K. Ho, T. T. Tjhung, “Training
sequence assisted channel estimation for MIMO OFDM,”
MSE(Channel)