Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Adaptive Channel Modeling For MIMO Wireless Communications: Chengjin Zhang Robert R. Bitmead

Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

Adaptive Channel Modeling for MIMO Wireless

Communications
Chengjin Zhang

Robert R. Bitmead

Department of Electrical and Computer Engineering


University of California, San Diego
San Diego, CA 92093-0407
Email: zhangc@ucsd.edu

Department of Mechanical and Aerospace Engineering


University of California, San Diego
San Diego, CA 92093-0411
Email: rbitmead@ucsd.edu

Abstract The application of state-space-based subspace system identification methods to training-based estimation for timevarying MIMO frequency-selective channels is explored with the
motivation of possible parsimonious parametrization and direct
model complexity control. The comparison between the statespace-based channel estimation algorithm and the FIR-based
RLS algorithm shows the former is a more robust modeling
approach than the later.

I. I NTRODUCTION
Digital communication using multiple transmit and receive
antennas has been one of the most important technical developments in modern communications. In a rich scattering
environment, multi-input-multi-output (MIMO) systems offer
significant capacity gain at no cost of extra spectrum [1].
Furthermore, space-time channel codes [2] can be applied to
utilize multiple transmit antennas to build spatial redundancy
in the transmitted signal such that maximum spatial diversity
is achieved. Coherent space-time processing schemes assume
the availability of a channel model at the receiver. Therefore,
this model needs to be estimated at the receiver end.
For slowly-varying channels, training-based channel estimation is very common in practice. Most channel estimation
schemes assume an FIR channel model. FIR models for wireless channels have been widely used for their simplicity and
guaranteed stability. An FIR model represents the sub-channels
of a MIMO channel with separately parametrized finite-length
impulse responses. However, when the spatial subchannels
are correlated with each other due to insufficient separation
between adjacent antennas, an FIR model can be very nonparsimonious and contain excessive redundancy. Therefore it
may be beneficial to adopt a state-space model which treats
the whole channel as a single entity and hence captures the
structure in the channel while allowing a more parsimonious
description of the MIMO channel.
It is the subject of this paper to explore the application
of state-space models to represent MIMO frequency-selective
channels in the hope for possible parsimonious parameterization. Since any channel model is essentially an approximation
of the physical channel, the goal of channel modeling is to
find a model that is as close to the real channel as possible
and maintains manageable complexity as well. It is found that

state-space models are able to provide low-order models of


high-quality channel approximation.
Low-order, high-quality models are of interest because they
hold the prospect of requiring fewer parameters for their
description and consequently an improvement in adaptation
rate. Since we are considering learning with training data, the
adaptation rate is important, because training symbols contain
no new message information. The simulations of the paper
demonstrate comparable (but slower) adaptation performance
of state-space methods but improved approximation properties.
For the purpose of channel estimation, an algorithm for
estimating MIMO state-space models is needed. Subspace
system identification (SSI) algorithms are a group of methods
that identify a MIMO state-space system in a straightforward
way using numerically robust computational tools such as
singular value decomposition (SVD) and QR factorization
[3], [4]. SSI algorithms provide a direct way to control the
complexity of the estimated channel model. The order of the
channel model can be selected by the user by choosing the
number of the largest singular values of the estimated extended
observability matrix to include. Furthermore, the modeling
error is assessable as it is related to the sum of the neglected
Hankel singular values.
This paper is organized as follows. Section II describes the
redundancy in correlated MIMO channels and the motivation
for the adoption of state-space models. Section III reviews
the subspace system identification methods and discusses its
application to MIMO channel estimation. Section IV presents
numerical simulation results and compares the state-spacebased SSI channel estimation algorithm to the conventional
FIR-based RLS channel estimation algorithm. The paper is
concluded in Section V.
II. C HANNEL R EDUNDANCY AND S TATE -S PACE C HANNEL
M ODEL
In order to understand the redundancy in spatially correlated
MIMO channels and its effect on the order or the McMillan
degree of the channel, a computer experiment was carried
out to examine the singular values of the Hankel matrix of
correlated MIMO channels and compare them with those of
spatially white MIMO channels. It will be shown that channel

Normalized singular values of the Hankel matrix, 1000 2x2 channel realizations

Normalized singular values of the Hankel matrix, 1000 10x10 channel realizations

1
correlated channel
uncorrelated channel

correlated channel
uncorrelated channel

0.9

0.9

0.8

0.8

0.7

0.7

0.6

0.6

0.5

0.5

0.4

0.4

0.3

0.3

0.2

0.2

0.1

0.1

10

(a) 2 2 channels.

10

15

20

25

30

35

40

45

50

(b) 10 10 channels.

Fig. 1. Hankel singular values of MIMO channels (normalized with respect to the largest value). = 0.8 for the spatially correlated channels; = 0 for
spatially white channels.

correlation increases the number of negligible small Hankel and possess the structure

singular values and hence reduces the effective channel order.


1

2 . . . m1
In the simulation, the MIMO channel is assumed to be

1
. . . m2

composed of multiple finite impulse responses. Assuming all


RT X = RRX =
(5)
..

the subchannels have the same length L, an FIR MIMO


.
m1
m2
channel with m transmit antennas and p receive antennas can

...
1
be represented by assembling all the taps of the subchannel
impulse responses with the same delay into a matrix H where represents the correlation coefficient between the
fading of two adjacent transmit or receive antennas. The value
and represent the channel as a series of matrices
of is chosen to be 0.8 in the simulation.

h11 ( ) h12 ( ) . . . h1m ( )


As an indicator of the effective channel order, the singular
h21 ( ) h22 ( ) . . . h2m ( ) values of the Hankel matrix of the corresponding MIMO

.
[H0 , H1 , . . . , HL1 ], H =
channels
..
are computed and averaged over 1000 channel re

.
alizations. [The H -norm of the difference between a MIMO
hp1 ( ) hp2 ( ) . . . hpm ( )
(1) system and its low-order approximation is related to the sum of
the neglected Hankel singular values, see [6] (Theorem 7.11).]
The simulated channels are taken to be square symmetric,
Figure 1 shows the difference between the distributions of
and the length of the impulse responses is set to 6. The channel
Hankel singular values for spatially white channels and spataps are generated as zero-mean circularly-symmetric comtially correlated channels. It can be seen that spatial correlation
plexity Gaussian (ZMCSCG) random variables. The impulse
increases the number of negligible singular values and hence
response of each subchannel is normalized to have unit energy.
allows a low-order state-space model that is a high-quality
The MIMO channel is assumed to the uncorrelated in delay
approximation of the original full-order channel. Furthermore,
but correlated in spatial dimension, i.e.
the larger the size of the MIMO channel is, the lower the order

of a high-quality channel model can be achieved. Therefore,
0, k 6= l
E(vec(Hk )vec(Hl )H ) =
(2) a state-space model is capable of providing good low-order
R, k = l
approximations of large spatially correlated MIMO channels.
where vec() is an operator that stacks the columns of a
xk+1 = Axk + Buk
matrix on top of each other to form a vector. We adopt a
simple model for the spatial correlation structure of H [5]
yk = Cxk + Duk + nk
(6)
H
R

=
=

(RRX )1/2 Hw (RT X )T /2


RT X RRX

(3)
(4)

where Hw is a pm matrix with IID ZMCSCG elements.


denotes the Kronecker product. RT X and RRX are, respectively, the transmit covariance matrix and receive covariance
matrix. They are assumed to be the same in the simulation

where
uk m 1 input vector at time k;
yk p 1 output vector at time k;
xk n 1 state vector at time k;
nk p 1 white Gaussian noise vector at time k.
On the other hand, the order of an FIR channel can only be
controlled by the length of the impulse response. This is an

inferior approach compared to the model reduction in statespace models, as will be shown later in Section IV.
III. S UBSPACE S YSTEM I DENTIFICATION AND MIMO
C HANNEL E STIMATION
The MIMO FIR model can be fitted by forming a parameter
vector from the columns of H and a regressor vector from
a suitable rearrangement of the input signals at the transmit
antennas. Then Recursive Least Squares is applied to this
vector problem.
However, for state-space models, a special class of estimation algorithms is needed. Subspace system identification
(SSI) refers to a class of recent algorithms, such as MOESP
and N4SID, which apply input-output system identification
methods to determine directly a state-space realization of
a system. The key idea of SSI methods is to estimate the
extended observability matrix through the projection of future
input-output data onto past input-output data based on the
relationship between Hankel matrices of the input and output
given by
Y1,i,M = i X1,M + Hi U1,i,M + N1,i,M , i > n.

(7)

where U1,i,M , Y1,i,M and N1,i,M are Hankel matrices of the


input, output and noise, respectively, with the form of

y1
y2

yM
y2
y3

yM+1

Y1,i,M = .
(8)
.
..
..
..
..

.
.
.
yi

yi+1

yi+M1

X1,M is a matrix containing the state vectors




X1,M = x1 xM .

i and Hi are, respectively, the extended observability matrix


and the matrix of impulse response coefficients.

D
0
0
C
CB
CA
D
0
0

CAB
..
CB D 0 0
i = . , Hi =
.

..
..
..
..

CAi2
.
.
.
.
CAi1

CAi2 B

D
(9)
Then the system matrices A, B, C and D are computed
based on the estimated observability matrix and singular value
decomposition (SVD) algorithm, see [3], [4], [7] for in-depth
treatments.
Based on the channel model given in (6), SSI methods
require the input to satisfy the following requirements for the
channel to be identifiable.
1) The input uk is uncorrelated with the additive Gaussian
white noise nk .
2) The input uk is persistently exciting of order of at least
2 times the maximum order of the channel.
3) The symbols in the input sequence are contiguous and,
for consistency, the number of inputs goes to infinity.
The first assumption is usually satisfied for wireless communication systems. The second one requires the training

sequence to maintain a certain structure. Notice that the third


assumption places limitation on the application of SSI methods
to channel estimation in wireless communication systems
where the training sequences are usually not contiguous in
time.
In this paper, we consider a special type of SSI algorithms, the MOESP algorithm [8] and its recursive version
[9]. Detailed discussion of SSI for non-contiguous data and
the recursive MOESP for time-varying channels is given in
APPENDIX.
IV. S IMULATION

AND I NTERPRETATION

The purpose of the simulation study is to examine the


estimation performance of state-space-based recursive SSI
channel estimation algorithm and to compare it to the FIRbased recursive least squares (RLS) channel estimator.
The spatially correlated MIMO channels are generated in
the same way as discussed in Section II. The training sequences are random binary sequences that are uncorrelated in
both spatial dimension and time dimension. We choose the H2
norm of the difference between the estimated channel model
and the true channel as an indicator of the estimation error. For
a fixed MIMO FIR-6 channel with temporally and spatially
white input symbol sequence, both the SSI and FIR/RLS
channel estimators were run until they reached a stationary
value. Then, at time t = 0, the channel was changed and the
identification methods allowed to reconverge. This is depicted
in Figure 2(a) and 2(b) for 2 2 channels with SNR=20dB
and illustrates the transient performance or learning rate of the
identifiers.
For the upper figures, the horizontal axis shows the number
of training symbols sent after time t = 0. The vertical axis
shows the H2 error between the true channel and the model.
There are separate curves for each different model order, with
the dashed curves corresponding to the FIR models and the
solid or dotted curves corresponding to state-space models.
The items of interest are the steady-state error values, which
indicate the modeling accuracy, and the learning or adaptation
rate. The ordering of the curves shows that lower-order models
achieve poorer steady-state accuracy and faster learning. The
curves themselves are the result of averaged curves over 20
channel (initial and final) realizations and, for each channel
realization, 100 training symbol sequence realizations.
The lower figures are derived from the upper figures and plot
one point for each model order and identification method, with
triangles for the FIR/RLS method and circles for SSI/MOESP.
The vertical axis is the steadystate modeling error and the
horizontal axis shows the input symbol number from which
the channel model error remains below 110% of the steadystate value. This latter figure gives a measure of convergence
rate.
The state-space models provide low-order models of highquality approximation, especially in the presence of channel
redundancy, whereas the reduced-length FIR models fail to
provide comparable estimation error performance. This performance difference between reduced-order state-space models

Tracking performance MOESP and RLS


3

2.5

2.5
H error in dB

2
1.5

2
1.5

H error in dB

Tracking performance MOESP and RLS


3

0.5
0

0.5
0

50

100

150

200

250

50

100

symbol index
= 0.90, 100 data realizations, 20 channel realizations

200

250

= 0.97, 100 data realizations, 20 channel realizations

2
Recursive MOESP
RLS

steadystate H error

Recursive MOESP
RLS

1.5

1.5

steadystate H error

150
symbol index

0.5

10

15

20

25
30
35
convergence point

40

45

50

0.5

0
20

55

40

60

(a) = 0.90, fast convergence.

100
120
convergence point

140

160

180

(b) = 0.97, slow convergence.

Estimation error of MOESP and RLS for 2 2 spatially correlated channels with = 0.8.

Tracking performance MOESP and RLS


5

H error in dB

2
1
0

50

100

150

200

250

symbol index
= 0.93, 100 data realizations, 10 channel realizations
3

and reduced-length FIR models becomes even more evident


in large-dimension MIMO channels (Figure 3). The reason
for such a difference is that state-space models are a more
general model class than FIR models. The former includes
the latter as a special case and thus provides a larger model
set. Therefore, when the true order or length of the physical
channel is unknown, the state-space model is a more robust
channel modeling approach than the FIR model because the
former is less sensitive to model order selection error than the
latter.
The convergence rate of the recursive MOESP algorithm
is in general slower than that of the RLS algorithm, but the
difference stays in a comparable range. It is also observed that
the recursive MOESP algorithm exhibits a start-up delay which
is not present for the RLS algorithm. We suspect that this
delay can be remedied by careful modification of the MOESP
algorithm.
The faster convergence of RLS is to be expected because the
(spatially and temporally) white training sequence is optimal
for FIR model structure [7]. However, for the ultimate application of these models for the equalization of the MIMO channel,
where the models themselves might be used to capture FIR
or state-space approximations to the channel inverse, it is
not immediately apparent which method would be faster.
One suspects that, since MOESP includes an RLS section to
produce the B and D matrices of the state-variable realization,
that RLS would be always faster than SSI of the same order.

steadystate H error

Fig. 2.

80

Recursive MOESP
RLS

2.5
2
1.5
1
0.5
0
10

20

30

40

50
60
convergence point

70

80

90

100

Fig. 3. Estimation error of MOESP and RLS for 3x3 spatially correlated
channels with = 0.8, = 0.93.

identification (SSI) algorithm for non-contiguous training data


is presented for channel estimation in multiple-antenna communication systems. The comparison between the state-spacebased channel estimation and the FIR-based RLS algorithm
also shows the former is a more robust modeling approach
than the later. Future work may involve improvements of the
recursive SSI algorithms for faster convergence.

V. C ONCLUSION
In this paper, the state-space model is proposed for modeling
MIMO wireless channels with the motivation for a more
parsimonious parameterization. A recursive subspace system

ACKNOWLEDGMENT
This research was supported by CORE Grant No. 02-10109
sponsored by Ericsson.

A PPENDIX
A. SSI for non-contiguous data
Notice that the third assumption of conventional SSI algorithms
(Section III) places limitations on the application of SSI methods
to channel estimation in wireless communication systems where the
training sequences are usually not contiguous in time. Instead, they
lie in the overhead or mid-amble of a frame and are separated by data
symbols. In this circumstance, the state evolution of the received data
must be restarted at the frame boundaries. This is at variance with
the standard formulation of SSI.
A suitable modification of the SSI algorithms for non-contiguous
training data is proposed in [10]. It is shown that the non-contiguousdata approach is similar in estimation power to the contiguous-data
approach when the length of the training sequence is sufficiently
larger than the dimension of the extended observability matrix i
(9).

B. Recursive SSI for time-varying channels


When the channel is time-varying, channel estimation needs to be
carried out in a recursive manner so that any newly received data
can be used to produce an up-to-date estimate of the channel. In this
paper we mainly consider the application of a recursive version of
the ordinary MOESP SSI scheme proposed in [9], whose key idea is
to update the estimate of the extended observability matrix with the
newly received data using LQ factorization. An exponential forgetting
factor can also be added to remove the effect of data from the distant
past at a high rate.
To give a brief summary of the recursive ordinary MOESP
algorithm, consider that at time t we have an LQ decomposition

Without loss of generality, we only consider the case of (10). In order


to find out R22 (t + 1), we rewrite (10) as

U (t) u (t + 1)
U (t + 1)
=
Y (t) y (t + 1)
Y (t + 1)

R11 (t)
0
u (t + 1) Q 0
(11)
=
R21 (t) R22 (t) y (t + 1) 0 1

R11 (t + 1) 0mipi
0mi1
=
Q0
R21 (t + 1) R22 (t) (t + 1)
where the pi 1 vector (t + 1) is obtained through a series of
Givens rotation on the large lower triangular matrix. (0, 1] is an
exponential forgetting factor which could be used to adjust the effect
of the past input-output data on the estimate of the current channel
model. R22 (t + 1) can be found by taking an LQ factorization of
[R22 (t) (t + 1)] as follows.
[R22 (t) (t + 1)] = R0 Q0
Therefore, by substituting (12) into (11),


U (t + 1)
R11 (t + 1)
=
Y (t + 1)
R21 (t + 1)

R11 (t + 1)
=
R21 (t + 1)

R11 (t + 1)
=
R21 (t + 1)

(12)

we obtain

0 0
Q0
R0 Q0

0 0 Q1
R0 Q0 Q2

0
Q00
R0

where Q1 is a matrix that contains the first mi rows of Q0 and Q2


is a matrix that contains the last pi + 1 rows of Q0 . It is clear that

Q1
00
Q =
Q0 Q2
is a matrix with orthogonal rows, i.e. Q00 (Q00 )T = I. Thus

U (t)
R11 (t)mimi
=
Y (t)
R21 (t)pimi

0
Q
R22 (t)pipi

where U (t) and Y (t) are respectively the Hankel matrices of the
input and output data at time t with the form
2

6
6
Y (t) = 6
4

y0
y1
..
.
yi1

y1
y2
..
.
yi

...
...
...

yti+1
yti
..
.
yt

3
7
7
7
5

It was shown in [3] that R22 (t) is an estimate of the extended


observability matrix at time t, i (t), in the sense that they share
the same column space.
The task of a recursive MOESP scheme would be to compute
R22 (t + 1), an estimate of i (t + 1), based on the LQ factorization
at time t and the newly received data at time t + 1. Suppose at
time t + 1 a new set of input-output data, xt+1 and yt+1 , becomes
available. We can form the following data vectors
3
2u
3
2y
ti+2
ti+2
6 .. 7
6 . 7
. 7 , y (t + 1) = 6 .. 7
u (t + 1) = 6
5
4
5
4
ut
yt
ut+1
yt+1
Then the new input-output Hankel data matrix at time k + 1 can be
formed by appending the above vectors to the end of the Hankel data
matrix at time t.

U (t + 1)
U (t) u (t + 1)
=
(10)
Y (t + 1)
Y (t) y (t + 1)

R22 (t + 1) = R0
In summary, the update from R22 (t) to R22 (t + 1) is a two-step
procedure: computing (t+1) in (11) using Givens rotation followed
by computing R0 in (12).

R EFERENCES
[1] G. J. Foschini and M. J. Gans, On limits of wireless communications in
a fading environment when using multiple antennas, Wireless Personal
Communications, vol. 6, no. 3, pp. 311335, March 1998.
[2] V. Tarokh, H. Jafarkhani, and A. R. Calderbank, Space-time block
coding for wireless communications: Performance results, IEEE J.
Select. Areas Commun., pp. 451460, March 1999.
[3] M. Verhaegen and P. Dewilde, Subspace model identification Part 1.
the output-error state-space model identification class of algorithms,
International Journal of Control, vol. 56, no. 5, pp. 11871210, 1992.
[4] P. Van Overschee and B. De Moor, Subspace identification for linear
systems: theory, implementation, applications.
Dordrecht: Kluwer
Academic Publishers, 1996.
[5] K. Yu and B. Ottersten, Models for mimo propagation channels: a
review, Wireless Communications and Mobile Computing, vol. 2, pp.
273291, Nov. 2002.
[6] K. Zhou and J. C. Doyle, Essentials of robust control, 1st ed. New
Jersey, USA: Prentice Hall, 1997.
[7] L. Ljung, System identification, theory for the user, 2nd ed. New Jersey,
USA: Prentice Hall, 1999.
[8] M. Verhaegen, Identification of the deterministic part of MIMO state
space models given in innovations form from input-output data, Automatica, vol. 30, no. 1, pp. 6174, Jan. 1994.
[9] M. Lovera, T. Gustafsson, and M. Verhaegen, Recursive subspace
identification of linear and non-linear Wiener state-space models,
Automatica, vol. 36, no. 11, pp. 16391650, Nov. 2000.
[10] C. Zhang and R. R. Bitmead, Multiple antenna system equalization
using semi-blind subspace identification methods, Proceedings of 13th
IFAC Symposium on System Identification, Rotterdam, Aug. 2003.

You might also like