MIMO-OFDM Channel Estimation based on Subspace Tracking∗

Jianxuan Du and Ye (Geoffrey) Li

School of Electrical and Computer Engineering
Georgia Institute of Technology, Atlanta, GA 30332-0250
Email: jxdu@ece.gatech.edu

Abstract— In this paper, we propose a channel estimation algo- vary much faster, are updated every OFDM block based on
rithm for multiple-input and multiple-output orthogonal frequency decision-directed LS channel estimation and further refined
division multiplexing (MIMO-OFDM) systems, which has con- by a robust filter [10] to exploit time domain correlation.
siderably less leakage than DFT-based channel estimators. This
algorithm uses the optimum low-rank channel approximation Simulation results show that the proposed algorithm is able to
obtained by tracking the frequency autocorrelation matrix of significantly reduce the error floor compared with DFT-based
the channel response. The coefficients corresponding to dominant algorithms.
eigenvectors of the autocorrelation matrix are estimated every
OFDM block while the eigenvectors are only updated using
the training block that is transmitted every M blocks due II. S YSTEM M ODEL AND C HANNEL E STIMATION
to the slowly-varying feature of the frequency autocorrelation. In this section, we briefly introduce our MIMO-OFDM
Simulation results show that the proposed algorithm can effec-
tively reduce channel estimation error and thus improve system
system model and LS channel estimation.
performance.
A. MIMO-OFDM System
I. I NTRODUCTION The structure of the MIMO-OFDM system with proposed
The combination of orthogonal frequency division multi- channel estimator is shown in Figure 1.
plexing (OFDM) with space-time coding has received much at-
tention recently to combat multipath delay spread and increase
t1[n, k ] r1[n, k ]
system capacity [1][2][3]. Channel parameters are needed IFFT
FFT
in order to coherently decode the transmitted signal. Least b[n, k ] space-
space-
encoder t 2 [n, k ] r2 [n, k ]
square (LS) channel estimation for MIMO-OFDM systems time
time
decoder
has been addressed in [4][5]. But if the multipaths are not IFFT FFT
sample-spaced, the well known leakage problem for DFT-
based channel estimation [6] induces an irreducible error floor
for estimation error. To reduce this error floor, more taps have channel
estimator
to be used, which not only increases computational complexity
but also makes estimation problem more ill-conditioned and M subspace
tracker
thus enhances noise. As an alternative, channel estimation M
algorithm based on parametric model has been proposed in
[7] and extended to MIMO-OFDM in [8]. By using estimation Fig. 1. Subspace-base channel estimation for MIMO-OFDM system with 2
of signal parameters via rotational invariance techniques transmit and 2 receive antennas
(ESPRIT), channel estimation is achieved by only estimating
a few coefficients corresponding to different multipath delays
which can yield smaller error. However, when the channel The complex baseband equivalent of a fading channel
varies slowly, the number of training symbols needed for impulse response can be expressed as,
initial multipath delay acquisition may become too large to L
X
be practical. Another parametric model-based algorithm needs h (t, τ ) = αl (t) ∆ (τ − τl ) , (1)
exhaustive search to find the near-maximum likelihood time l=1
of arrival estimation[9], which suffers high computational where αl (t)’s are wide-sense stationary narrowband complex
complexity. Gaussian processes and are assumed to be independent among
In this paper, we estimate the dominant subspace spanned different paths. L is the number of multipaths. Thus for each
by the frequency responses of different multipaths. Due to the receive antenna, the channel frequency response for the kth
fact that in a mobile environment the multipath delays vary subcarrier at time n is
slowly the subspace can be tracked by using training OFDM X
blocks sent at regular interval. The coefficients corresponding H [n, k] = αl (t) e−j2πk∆f τl , (2)
to different eigenvectors of the resultant subspace, which l
B
* The work was jointly supported by the NSF under Grant CCR-0121565 where ∆f = is the subcarrier spacing, B is the total band-
K
and Nortel networks. width, and K is the total number of subcarriers. In addition, we
assume the same normalized time-domain correlation function which approximates the channel response as
for all paths, i.e., χ−1
X
 2 kl
∗ σl rt (∆t) , l = m, Hi [n, k] = hi [n, l] WK , (8)
E {αl (t + ∆t) αm (t)} = (3)
0, l 6= m. l=0
2π
Without loss of generality, we assume that the total average where WK = e−j K . Thus, with correct detection, equation
power of the channel impulse response to be unity, i.e., (6) can be written in matrix notation as
X
σl2 = 1. (4) Nt
X
l x [n] = D̂i [n] Uhi [n] + w [n] . (9)
i=1
For a MIMO-OFDM system with Nt transmit and Nr
receive antennas, the received signal at the k-th subcarrier of Let  T T
4 T T
the n-th block from the j-th receive antenna can be expressed ĥ [n] = ĥ1 [n] , ĥ2 [n] , · · · , ĥNt [n] ,
as:
Nt
X from [4], the LS estimate of the channel is
xj [n, k] = Hij [n, k] bi [n, k] + wj [n, k] , (5)
−1 H
i=1 ĥ [n] = TH [n] T [n] T [n] x [n] (10)
for j = 1, . . . , Nr and k = 0, . . . , K − 1, where bi [n, k] is where
the symbol transmitted from the i-th transmit antenna at the  
k-th subcarrier of the n-th block, Hij [n, k] is the channel’s T [n] = D̂1 [n] U, D̂2 [n] U, · · · , D̂Nt [n] U . (11)
frequency response at the k-th subcarrier of the n-th block
corresponding to the i-th transmit and the j-th receive antenna, In [4], it has been shown that the estimate is unbiased and the
and, wj [n, k] is additive (complex) Gaussian noise that is mean-square error of the channel estimator is
assumed to be independent and identically distributed (i.i.d.)
 2 
4 1
M SE [n] = E ĥ [n] − h [n]

with zero-mean and variance ρ. Nt χ
1 n
−1 o
= Tr ρ TH [n] T [n] . (12)
B. Decision-Directed Least-Square Channel Estimation Nt χ
After detection of each OFDM block, we use the detected It can be proved in the appendix using the bordering the-
symbols for channel estimation. Since channel responses are orem for Hermitian matrices [11] that condition number of
independent among different transmit-receive antenna pairs, TH [n] T [n] increases with χ. It implies the channel estimation
we consider the channel estimation for one particular receive becomes more ill-conditioned as the number of parameters to
antenna and omit the receive antenna subscript j in equation be estimated increases. Thus we should choose the number
(5) to get: of parameters as small as possible while preserving energy
Nt
X of the channel response, which is the reason for tracking the
x [n, k] = Hi [n, k] bi [n, k] + w [n, k] . (6) optimum basis to form the transform matrix U in the next
i=1 section .
Define
∆ T III. SUBSPACE TRACKING
x [n] = (x [n, 0] , x [n, 1] , · · · , x [n, K − 1]) ,
∆
n o The major problem of decision-directed channel estimation
D̂i [n] = diag b̂i [n, 0] , · · · , b̂i [n, K − 1] , is the randomness of the symbol sequences during data trans-
mission mode. For example, when the symbol sequences from
∆ T
w [n] = (w [n, 0] , w [n, 1] , · · · , w [n, K − 1]) , any two of the transmit antennas are same or very close, it
is impossible or very hard to distinguish channel responses
where b̂i [n, k]’s are detected symbols. Let rank-χ approxima-
corresponding to different transmit antennas. The more the
tion of the channel response be:
number of transmit antennas, the more likely the channel is
T
(Hi [n, 0] , · · · , Hi [n, K − 1]) = Uhi [n] , (7) unidentifiable, or the more ill-conditioned channel identifica-
tion is. Furthermore, to reduce the leakage of decision-directed
where DFT-based channel estimation in MIMO-OFDM systems, the
4 T number of taps representing the channel frequency response
hi [n] = (hi [n, 0] , · · · , hi [n, χ − 1]) ,
has to be increased, which will make channel identification
and U is the rank-χ transform matrix. For DFT-based channel more ill-conditioned at the same time, as shown in the previous
estimator as in [4], it is just section. Moreover, increasing the number of taps makes the in-

1 1 ··· 1
 verse operation of matrices in Equation (10) more complicated.
1 WK ··· WKχ−1  Hence, it is essential for low-complexity and high-performance
U =  ..

.. .. ,
 channel estimator to reduce the number of parameters to be
. ··· . .  estimated while preserving most energy of channel frequency
(K−1) (χ−1)(K−1)
1 WK · · · WK responses during the data transmission mode. Therefore, we
 TABLE I TABLE II
CHANNEL SEPARATION USING THE OPTIMUM TRAINING SUBSPACE TRACKING FOR CHANNEL ESTIMATION
SEQUENCES
Initialization: √
(Ui [0])k,l = WKkl / K, 0 ≤ k ≤ K − 1, 0 ≤ l ≤ χ − 1;
a) During each training block, η [n, k] = x [n, k] · b∗1 [n, k];
b) Perform IFFT on (η [n, 0] , η [n, 1] , · · · , η [n, K − 1]) Φ [0] = I, 0 ≤ α ≤ 1;
to get (ζ [n, 0] , ζ [n, 1] , · · · , ζ [n, K − 1]);
c) For the channel response of transmit antenna i, During each training block:
 T
circularly left shift (ζ [n, 0] , ζ [n, 1] , · · · , ζ [n, K − 1]) input vi [n] = Ĥi [n, 0] , Ĥi [n, 1] , · · · , Ĥi [n, K − 1]
by (i − 1) K0 toget (ζ 0 [n, 0] , ζ 0 [n, 1] , · · · , ζ 0 [n, K − 1]).
∆ ζ 0 [n, k] k ∈ [0, K0 − 1] , ci [n] = UH
i [n − 1] vi [n]
Let ζ 00 [n, k] = Ai [n] = αAi [n − 1] Φi [n − 1] + (1 − α)vi [n] cH
0 else; i [n]
Ai [n] = Ui [n] Ri [n]
d) A channel estimate Ĥi [n, k] is obtained by performing FFT on H
(ζ 00 [n, 0] , ζ 00 [n, 1] , · · · , ζ 00 [n, K − 1]). Φi [n] = Ui [n − 1] Ui [n]
Low rank channel approximation:
vi [n] = Ui [n] hi [n]

will develop subspace tracking approaches to estimate channel

parameters. rather than the particular eigenvectors. Let
Let the K×K channel autocorrelation matrix (Rf )k1 ,k2 =
Ûi [n] = Uopt Qi [n] , (15)
E {H [k1 ] H ∗ [k2 ]} have singular value decomposition as:
where Qi [n] is a χ × χ unitary matrix, which accounts
Rf = Uf ΛUH
f (13) for the change of dominant eigenvectors without changing
where Uf is a K × K unitary matrix. Λ = the subspace. Substituting (15) into (12), it can be easily
diag {λ1 , λ2 , · · · , λK }, λ1 ≥ λ2 ≥ · · · > 0. From verified that the MSE of the channel estimator is invariant
[12], optimum rank-χ estimator is to select eigenvectors to rotation of the dominant eigenvectors, which can also be
u1 , u2 , · · · , uχ corresponding to the χ biggest eigenvalues. seen in [10]. Therefore, it is the dominant subspace spanned
Then the optimum rank-χ transform matrix is: by channel frequency responses that affects the performance
of the subspace tracking based channel estimator.
Uopt = (u1 , u2 , · · · , uχ ) . (14)
IV. S IMULATION R ESULTS
Therefore, channel autocorrelation matrix is needed here for
We use typical-urban (TU) delay profile with Doppler
the optimum low-rank channel estimation.
frequency fd of 40 and 100 Hz in our simulation and assume
Since channel autocorrelation is determined by channel de-
channel parameters corresponding to different transmit and
lay profile, which varies very slowly, periodically sent training
receive antenna pairs are independent but with the same
blocks can be used for tracking channel autocorrelation as well
statistical properties. The system has 2 transmit and 2 receive
as for channel estimation. In [5] and [13] optimal training se-
antennas. A total bandwidth of 1.25 MHz is divided into 256
quences have been proposed to maximally separate frequency
subcarriers, 2 subcarriers on each end are used as guard tones,
responses of different transmit antennas while preserving most
leaving the rest for transmitting data. The symbol duration is
energy of each channel response. The training sequences are
204.8 µs and another 20.2 µs is added as cyclic prefix (CP),
−K0 (i−1)k resulting in a total block duration of 225 µs. A 16-state 2
bi [n, k] = b1 [n, k] WK ,
b/s/Hz 4-PSK space-time code proposed in [17] is used in
for i=1, 2, ..., Nt , where K0 = bK/Nt c ≥ dtd /ts e, td is the our simulation. Each data block of 500 bits is coded into two
maximum delay spread, and ts is the sampling interval which symbol blocks to be sent via the two transmit antennas at
1
is equal to K∆f . It is known that the leakage introduced by the different subcarriers. 4 bits are used for trellis termination.
DFT-based approximation is decreased as K0 increases. Since In each independent simulation, 2000 OFDM blocks of
we are using well designed training sequences rather than data are transmitted with 1 training block sent every M=10
arbitrary data sequences, we can set K0 to be big enough such blocks. The performance averaged over independent simula-
that the leakage is negligible while introducing little aliasing tions is evaluated. For channel estimator with 7-tap subspace-
between different channel responses [5]. The procedure to tracking, the first 50 blocks use 9-tap DFT-based estimator.
separate channel responses can be described in Table I. The estimated channel parameters are used for initial subspace
The dimension χ of the subspace can be either determined acquisition so that initial training overhead can be saved at
by minimum description length (MDL) criterion [14] that is the expense of negligible performance loss for continuous
not accurate for low SNR or slow channel variation, or by the data transmission. Then channel estimation is switched to the
approach in [12] and [15]. We just choose the latter approach estimator with subspace tracking and the subspace is updated
for its simplicity and effectiveness; therefore, χ = dtd /ts e. at each training block. The forgetting factor α is chosen to be
Subspace tracking approach can be summarized Table II, 0.99.
which is modified LORAF 1 in [16]. Figures 2 and 3 show the average word-error-rate (WER)
It should be noted that the robust channel estimator depends and mean square error (MSE) of the proposed estimator for
only on the subspace spanned by the dominant eigenvectors system with Doppler frequency of 40 and 100 Hz, respectively.
 0
10
The performance of DFT-based robust estimator and that with Ideal CSI
9−tap DFT−based estimator
ideal channel state information (CSI) are provided for compar- 7−tap subspace−based estimator
ison. It can be seen that for WER of 1%, the performance of
the proposed channel estimator has about 2 dB improvement
over the DFT-based channel estimator for fd = 40 Hz. This −1
10
performance improvement is 1.8 dB for fd = 100 Hz. The
performance improvement can be clearly explained by the

WER
considerable MSE gap between two estimators at high SNR’s
given by Figures 2.b and 3.b, when MSE becomes dominated −2
10
by leakage. Note that the MSE of proposed estimator is worse
than that of the DFT-bassed estimator at low SNR’s, which can
be explained by the subspace tracking error caused by strong
noise.
−3
10
0 0 2 4 6 8 10 12 14 16 18 20
10 SNR (dB)
Ideal CSI
9−tap DFT−based estimator
7−tap subspace−based estimator
(a)
−6
9−tap DFT−based estimator
7−tap subspace−based estimator
−8
−1
10
−10
WER

−12

−14
−2 MSE
10
−16

−18

−20
−3
10
0 2 4 6 8 10 12 14 16 18 20 −22
SNR (dB)

(a) −24
0 2 4 6 8 10 12 14 16 18 20
−6 SNR (dB)
9−tap DFT−based estimator

−8
7−tap subspace−based estimator (b)

−10 Fig. 3. (a) WER and (b) MSE of MIMO-OFDM systems for channels with
TU delay profile and fd =100 Hz
−12

−14
subspace spanned by the the frequency responses of different
MSE

−16 multipaths, we obtain the optimum basis for LS channel

estimation. The proposed algorithm considerably reduces the
−18
leakage of low-rank channel estimation, which is dominant at
−20 high SNR’s and the corresponding performance improvement
−22
is shown by simulation.

−24
0 2 4 6 8 10 12 14 16 18 20 A PPENDIX
SNR (dB)

(b) Let U = (U, uχ+1 ), and from U0 , we define T0 [n] as in

0

equation (11) by substituting U0 for U. We will show that

Fig. 2. (a) WER and (b) MSE of MIMO-OFDM systems for channels with
cond T0H [n] T0 [n] ≥ cond TH [n] T [n] ,



TU delay profile and fd =40 Hz

where cond(.) means condition number of a matrix.

Proof : Let the eigenvalues of TH[n] T [n] be γ1 ≥ γ2 ≥
· · · ≥ γχNt > 0, From equation (11) and by exchanging
V. C ONCLUSION columns which does not change the eigenvalues, we have
MIMO-OFDM is an attractive technique for high data
T0 [n] = (T [n] , Y [n]) ,
rate transmission and channel estimation for MIMO-OFDM
systems is essential for coherent detection. In this paper, where
we propose a subspace-based channel estimation algorithm  
for MIMO-OFDM systems. By tracking the slow varying Y [n] = D̂1 [n] uχ+1 , · · · , D̂Nt [n] uχ+1 .
 [5] Y. Li, “Simplified channel estimation for OFDM systems with multiple
transmit antennas,” IEEE Trans. Wireless Commun., vol. 1, pp. 67–75,
TH [n] T [n] TH [n] Y [n]
 
Jan. 2002.
T0H [n] T0 [n] = [6] J.-J. van de Beek, O. S. Edfors, M. Sandell, S. K. Wilson, and P. O.
YH [n] T [n] YH [n] Y [n] Borjesson, “On channel estimation in OFDM systems,” in Proc. 45th
IEEE Vehicular Technology Conf., Chicago, IL, July 1995, pp. 815–819.
Let the eigenvalues of T0H [n] T0 [n] be γ10 ≥ γ20 ≥ · · · ≥ [7] B. Yang, K. B. Letaief, R. S. Cheng, and Z. Cao, “Channel estimation for
0 OFDM transmission in multipath fading channels based on parametric
γ(χ+1)N t
> 0. By the bordering theorem for Hermitian channel modeling,” IEEE Trans. Commun., vol. 49, pp. 467–479, Mar.
matrices [11], we have 2001.
[8] R. Chen and K. B. Letaief, “Channel estimation for space time coded
γ10 ≥ γ1 ≥ γχNt ≥ γ(χ+1)N
0
>0 OFDM systems in non-sample-spaced multipath channels,” in Proc.
t WCNC’02, Orlando, FL, Mar. 2002, pp. 61–66.
[9] T. A. Thomas and F. W. Vook, “Broadband MIMO-OFDM channel
thus estimation via near-maximum likelihood time of arrival estimation,” in
Proc. ICASSP’02, vol. 3, Orlando, Fl, May 2002, pp. 2569–2572.
γ10 [10] Y. Li, L. J. Cimini, and N. R. Sollegberger, “Robust channels estimation
cond T0H [n] T0 [n]


= 0 for OFDM systems with rapid dispersive fading channels,” IEEE Trans.
γ(χ+1)Nt
Commun., vol. 46, pp. 902–915, July 1998.
γ1 [11] R. A. Horn and C. R. Johnson, Matrix Analysis. Cambridge, 1985.
≥ [12] O. Edfors, M. Sandell, J.-J. van de Beek, et al., “OFDM channel
γχNt
estimation by singular value decomposition,” IEEE Trans. Commun.,
cond TH [n] T [n] .


= vol. 46, pp. 931–939, July 1998.
[13] I. Barhumi, G. Leus, and M. Moonen, “Optimal training sequences
for channel estimation in MIMO OFDM systems in mobile wireless
channels,” in Proc. 2002 International Zurich Seminar on Broadband
R EFERENCES Communications, Zurich, Switzerland, Feb. 2002, pp. 44–1–44–6.
[14] M. Wax and T. Kailath, “Detection of signals by information theoretic
[1] D. Agrawal, V. Tarokh, A. Naguib, and N. Seshadri, “Space-time criteria,” IEEE Trans. Acoust., Speech, Signal Processing, vol. 33, pp.
coded OFDM for high data-rate wireless communication over wideband 387–392, Apr. 1985.
channels,” in Proc. IEEE VTC’98, May 1998, pp. 2232–2246. [15] H. J. Landau and H. O. Pollak, “Prolate spheroidal wave functions,
[2] H. Bolcskei and A. Paulraj, “Space-frequency coded broadband OFDM fourier analysis and uncertainty-III: The dimension of the space of
systems,” in Proc. WCNC’02, vol. 1, May 2000, pp. 1–6. essentially time- and band-limted signals,” Bell Syst. Tech. J., vol. 41,
[3] G. Yi and K. Letaief, “Space-frequency-time coded OFDM for broad- pp. 1295–1336, July 1962.
band wireless communications,” in Proc. GLOBECOM’01, vol. 1, 2001, [16] P. Strobach, “Low-rank adaptive filters,” IEEE Trans. Signal Processing,
pp. 519–523. vol. 44, pp. 2932–2947, Dec. 1996.
[4] Y. Li, N. Seshadri, and S. Ariyavisitakul, “Channel estimation for OFDM [17] Q. Yan and R. S. Blum, “Optimum space-time convolutional codes,” in
systems with transmitter diversity in mobile wireless channels,” IEEE J. Wireless Communications and Networking Conf., vol. 3, Chicago, IL,
Select. Areas Commun., vol. 17, pp. 461–471, Mar. 1999. July 2000, pp. 1351–1355.