Adaptive Channel SVD Estimation for

MIMO-OFDM Systems
H. Zamiri-Jafarian1, 2 and G. Gulak1
1
Department of Electrical and Computer Engineering, University of Toronto, Toronto, Ontario, Canada
2
Electrical Engineering Department, Ferdowsi University, Mashhad, Iran
[email protected]

Abstract—In this paper we propose an adaptive estimation x l (k ) = H l s l (k ) + nl (k ) for l = 1, L , L (1)

algorithm for channel matrix singular value decomposition
(SVD) in multiple-input multiple-output (MIMO) orthogonal where H l is an M × N channel matrix of the lth subcarrier
frequency division multiplexing (OFDM) systems. The SVD
method is an efficient approach to design space-time and n l (k ) is the M × 1 vector of the complex additive white
coding/decoding and detection algorithms in MIMO-OFDM Gaussian noise (AWGN) with zero-mean and autocorrelation
systems. However, the SVD estimation may involve complex
nonlinear optimization methods. The proposed algorithm in this matrix Rn l (k ) = σ n2 I M δ ( k ) for l = 1, L , L , while I M is
paper is developed based on a two-step recursive method by the M × M identity matrix. The SVD of the lth subcarrier
utilizing the linear constrained least mean square (LMS)
technique. channel matrix, H l , can be given as
H
Index Terms— MIMO-OFDM systems, MIMO channel H l = U l Σ lV l (2)
estimation, SVD estimation, constrained LMS.
where U l and V l are M × P and N × P unitary matrices,

I. INTRODUCTIION respectively. Note that P is the rank of H l where

Channel matrix singular value decomposition (SVD) P ≤ min( M , N ) . Σ l = diag (σ l1 , σ l 2 , L , σ lP ) is a
method is employed in MIMO-OFDM systems in order to over diagonal matrix containing the singular values of the lth
come subchannel interference, to allocate transmitted power subcarrier channel matrix and (.) H denotes transposed
through subchannels in an optimum manner and also to design
the space-time coding algorithm efficiently. Thus the SVD complex conjugate. Knowing the U l , V l and Σ l matrices
estimation is an important technique to exploit the full for all subcarriers l = 1, L , L are vital information to design
capability of MIMO-OFDM systems. The SVD is a nonlinear space-time coding, decoding and detection schemes
function and its estimation may become a more complicated
efficiently. Due to the unitary property of U l and V l
problem because it involves nonlinear optimization methods. In
this paper we propose an adaptive SVD estimation algorithm matrices, estimation of the subcarrier channel SVD at the
that uses a two-step recursive method. The estimation receiver side based on minimizing the mean square error
algorithm is developed based on the linear constrained LMS criterion is a complicated procedure because of nonlinear
technique and achieves good performance. The performance of H H
constraints U l U l = I P and V l V l = I P . A two-step
the proposed algorithm is evaluated by computer simulations
under different scenarios. adaptive method is proposed in the Section III based on the
constrained least mean square (LMS) algorithm that employs
linear constraints to estimate the SVD of subcarrier channel
II. SYSTEM MODELS matrices.
Consider a discrete MIMO-OFDM system with N transmitting
antennas, M receiving antennas and L subcarriers when III. ADAPTIVE SVD ESTIMATION

S (k) = [ s1 (k ),..., s L (k )] is a transmitted OFDM symbol We consider one subcarrier channel matrix, H l , and develop
matrix and s l (k) = [ s l1 (k ),..., s lN ( k )]T is a symbol vector the adaptive algorithm to estimate its SVD . Also, due to
T simplicity, we drop the subscript of H l in deriving the
transmitted from the lth subcarrier. Note that (.) represents
algorithm. Defining W 1 = UΣ and W 2 = VΣ matrices, it is
the transpose operation. Defining H = [ H 1 ,..., H L ]T as an
straightforward to show
L-point FFT of the M × N channel impulse response
matrix, H (k ) , after removing the cyclic prefix, the M × 1 w 1i = Hv i = σ i u i (3)
received vector of the lth sub-carrier, x l (k ) , becomes w 2iH = u iH H = σ i v iH (4)
where u i , v i , w 1i and w 2 i
E  Hˆ l - H l 
are the ith columns of U , 2

1 L
 
V , W 1 and W 2 , respectively, and σ i is the ith diagonal ∑
F
NMSE ( Hˆ ) =
element of Σ . L l =1 [
E Hl F
2
] (12)

Utilizing (3) and (4) and based on a training sequence s(k) ,

the adaptive estimation algorithm can be developed with two where Ĥ l is the estimation of H l . Note that the MIMO
steps. channel H l is estimated based on its SVD estimation from
In the first step, by receiving x (k ) and assuming vˆ i ( k − 1) is the following relation.
Hˆ l = Uˆ l Σˆ lVˆl
H
the estimated v i at time (k-1), w 1i at time k can be estimated (13)
by One hundred independent channel impulse response matrices
(
ˆ 1i ( k ) = I M − Wˆ 1i ( k )(Wˆ 1i ( k ) H Wˆ 1i ( k )) −1 Wˆ 1i ( k ) H
w ) are generated for evaluating the performance of the algorithm.
Fig. 1 shows the performance of the proposed adaptive method
× w (
ˆ 1i ( k − 1) + µ ( x ( k ) s ( k ) H vˆ i ( k − 1) − wˆ 1i ( k − 1)) ) for different values of N and M when SNR=30dB. The NMSE
for i = 1, L , P (5) of the estimated subcarrier channel matrices versus SNR for
N=M=2,4 and 8 has been shown in Fig. 2. Additional details of
ˆ
where W 1i (k ) = [w ˆ 1i −1 ( k )] and µ is a positive
ˆ 11 ( k ), L , w the adaptive estimation algorithm derivation and some
scalar step-size of the LMS algorithm. Note that the recursive modifications to the algorithm with more extensive simulation
equation of (5) has been derived by using the Lagrange results will be given in the full paper. Also, references will be
multiplier method in order to satisfy the orthogonality property included.
of W 1 based on the following constraint.
0
Wˆ 1i ( k ) H w
ˆ 1i ( k ) = 0 for i = 1, L , P (6)
From (5) the estimation of u i at time k is given by -5

1
−
uˆ i ( k ) = w
ˆ 1i ( k )( w
ˆ 1i ( k ) H w
ˆ 1i ( k )) 2 for i = 1, L , P (7) -10 N=8, M=8
N=8, M=2
NMSE (dB)

N=6, M=2
By having uˆ i ( k ) from (7), in the second step, w 2 i at time k -15 N=4, M=4
is estimated as
(
ˆ 2i ( k ) = I N − Wˆ 2i ( k )(Wˆ 2i ( k ) H Wˆ 2i ( k )) −1 Wˆ 2i ( k ) H
w ) -20

× w (
ˆ 2i ( k − 1) + µ ( s ( k ) x ( k ) H uˆ i ( k ) − w
ˆ 2i ( k − 1)) ) -25 N=4, M=2
N=2, M=6
for i = 1, L , P (8) N=2, M=4 N=2, M=2
where Wˆ 2i ( k ) = [w
-30
ˆ 21 ( k ), L , w
ˆ 2i−1 (k )] . To satisfy the 0 100 200 300 400 500 600 700 800 900 1000
Number of training symbols
orthogonality property of W 2 , Eq. (8) has been derived under
the following constraint. Fig. 1. The normalized MSE of the adaptive estimation algorithm versus the
number of training symbols.
Wˆ 2i ( k ) H w
ˆ 2i (k ) = 0 for i = 1, L , P (9)
ˆ 2i ( k ) from (8),
Using w v i and σ i for i = 1, L , P at time k -14
N=2, M=2
are estimated by -16
N=4, M=4
N=8, M=8
1
−
vˆ i ( k ) = w
ˆ 2i ( k )( w
ˆ 2i (k ) H w
ˆ 2i ( k )) 2 (10)
-18
1
σˆ i ( k ) = ( wˆ 2i (k ) H ˆ 2i
w ( k )) 2 (11)
NMSE (dB)

-20

IV. COMPUTER SIMULATIONS -22

A MIMO-OFDM system with L=64 for different N and M is

considered in the simulations when the channel impulse -24

response, H (k ) , has an exponential delay spread profile with

-26
duration of LC = 16 that is equal to the cyclic prefix interval.
The elements of the H (k ) are mutually independent -28
0 5 10 15 20 25 30
complex Gaussian random variables with zero-mean. To SNR (dB)
evaluated the performance of the estimation algorithm, the
normalized mean-square error (NMSE) criterion is employed. Fig. 2. The normalized MSE of the adaptive estimation algorithm versus
SNR.