Adaptive AR Channel Model Identication of

Time-varying Communication Systems

Zarko B.Krusevac
Australian National University
Canberra, ACT 2052 Australia
zarko.krusevac@rsise.anu.edu.au
Predrag B. Rapajic
University of Greenwich at Med. Pemb.
KENT ME4 4TB United Kingdom
P.Rapajic@greenwich.ac.uk
AbstractThis paper implements an adaptive identication
of autoregressive AR model coefcients for model-based lter-
ing over time-varying communication channels. The presented
approach does not require a-priori knowledge of the dynamics
of the system which overcomes the issue of determining model
coefcients that capture the dynamics of unknown time-varying
channels. Simulation MSE performance analysis in a multiuser
environment shows superior experimental performance of the
AR(2) model-based adaptive algorithm with adaptive model
identication, comparing to the AR(1) model-based adaptive
algorithm with adaptive model identication, the same algorithm
with xed model coefcients and standard observation-only-
based LMS and RLS adaptive algorithms.
I. INTRODUCTION
Background and Motivation - Model-based adaptive al-
gorithms exploit additional information of the propagation
conditions, in the form of a predetermined model, to improve
tracking performance over time-varying channels [1]. The dy-
namics of the channel time variations are generally described
in state equation form, in addition to the observation equation
[2], [3].
If the model captures the dynamics of the propagation
conditions, the model-based adaptive algorithms can provide
very good performance in non-stationary environments [4].
A stochastic modeling of time-varying channels helps to
overcome the complexity of the real propagation environment,
provides a focus on the critical broad attributes and tends to
average out nuisance parameters. Finite AR models [5], [6]
are often considered as proper stochastic description of the
channel dynamic in the form of linear difference (differential)
state equations.
The model-based approach to adaptive ltering usually
lends itself to a Kalman lter formulation or the problem
of jointly solving the process and measurement equations
[3]. However, the direct Kalman problem formulation for
the adaptive equalization over an unknown channel with ISI
requires information of the channel [7]. This can be achieved
by using observation-only based adaptive algorithms such as
LMS or RLS to provide the required channel estimate [8],
[9]. In [8] this technique is extended to adaptive equalization
over non-stationary communication channels. Even though the
model-based Kalman ltering theory is used and the system
is adaptive, the whole approach has the crucial drawback
of not recognizing the dynamics of time-varying channels,
neither by Kalman state (process) equation nor by observation-
only adaptive algorithms used for the channel estimation [4].
However, capturing the dynamics of channel time variations
by the Kalman process (state) equation is essential for the
overall performance of the system [4].
Since the channel output observation equation is linear in
the state process, assuming a known input sequence (training
sequence during the training period or sequence of tentative
decisions in decision directed mode), the Kalman ltering
problem can be directly formulated for channel process estima-
tion of an unknown time-varying channel [6], [10], [11]. Given
that the Kalman process equation is the channel state equation,
the dynamics of the channel time variations are explicitly
incorporated into the structure of the algorithm. However
this approach assumes separate explicit channel estimation
followed by equalization and data detection, which may not be
optimal in terms of information rate in the presence of channel
noise [12].
A specic approach to the Kalman equalization, originated
from [7] and named the Godard-Kalman approach [9], uses
the state vector to be the true (optimal) tap weights of
the equalizer. Optimal tap weights are calculated to minimize
the expected mean squared distortion of the lter output [7].
Since time-invariant channels are considered in [7], the optimal
tap weights are constant in time and the state transition
matrix is the identity matrix. However, if the channel is
time-varying then the optimal tap weights are time-varying
as well. Therefore, the Godard-Kalman approach can be
extended to time-varying channels by dening properly the
optimal tap weights transition matrix to implicitly capture
the dynamic of the time-varying channel process [4]. This
approach enables adaptive model-based (Kalman) equalization
without prior separate channel estimation. However, the issue
of determining model coefcients to capture the dynamics
of the optimal weights time-variations still remains. In this
paper, an adaptive identication of AR model coefcients is
performed to overcome this issue.
This paper extends and generalizes the Godard-Kalman
model based adaptive approach [7] to adaptive ltering over
time-varying communication channels. The extended Godard-
Kalman approach presented here utilizes RLS-based adaptive
identication of AR model coefcients. The adaptive model
identication, for a given model order, enables adaptive model
978-1-4244-2204-3/08/$25.00 2008 IEEE 618
based equalization without any a-priori information of fading
conditions. Simulation analysis in a multiuser environment
shows the superior experimental MSE performance of Godard-
Kalman algorithm with adaptive model identication com-
pared to one with xed model coefcients and to standard
observation-only-based LMS and RLS algorithms.
II. ADAPTIVE MODEL-BASED (KALMAN) FILTERING
THEORY
A. Model-based Design
The main idea behind the model-based adaptive algorithm
design is to use the information of the propagation conditions
to stochastically describe the time-varying channel process
using the channel state equation [2][4]. The observation equa-
tion and the state equation dene a state-space formulation
of a communication system over time-varying communication
channels. A discrete time, state-space model can be expressed
as follows
y
n
= S
n
x
n
+n
n
(1a)
S
n+1
= H
n
S
n
+Q
n
(1b)
where y
n
is the observable output of the channel, S
n
is the
channel process matrix, x
n
is the input of the channel, n
n
is
observation noise and Q
n
is channel process noise matrix, all
at the time instant n. Furthermore, H
n
is the channel transition
(system) matrix.
The observation (measurement) equation (1a) describes the
observation vector and relates the observable output of the
channel y
n
to the channel process matrix S
n
. The channel
process matrix S
n
is dened as the minimal set of data
that is sufcient to uniquely describe the unforced dynamical
behavior of the channel by the state (process) equation (1b).
B. Statement of the Kalman Filtering Problem
The Kalman ltering problem is the problem of jointly
solving the state (process) and observation (measurement)
equations for the unknown state matrix in an optimal manner
[2], [3].
The Kalman ltering problem may be formally stated as
follows [3]: Use the entire observed data y
1
,..., y
n
to nd, for
each n 1, the minimum mean-square estimate of the state at
the time i. The problem is called ltering if i = n, prediction
if i > n and smoothing if 1 i < n.
The concept of state is fundamental for the Kalman ltering
theory formulation [3] and it is not necessarily related to the
channel process S
n
in (1). For instance, the Godard-Kalman
formulation [7] denes the state as the vector of optimal lter
coefcients which are calculated to minimize the expected
mean squared distortion of the lter output. Furthermore, the
use of Kalman ltering presupposes that the system under
consideration is described by a set of linear difference (differ-
ential) equations [7]. However, the observation equation (1a)
is not necessarily linear in channel input x
n
. Thus, the state-
space formulation (1) is more general than the Kalman ltering
formulation.
III. GODARD-KALMAN ADAPTIVE ALGORITHMS FOR
TIME-VARYING CHANNELS
A. Godard-Kalman approach
A direct Kalman formulation for the adaptive ltering
for the communication system (1) (solution in x
n
) over an
unknown time-varying channel with ISI is not possible due to
non linearity of the observation equation (1a) in channel input
x
n
. The straightforward solutions requires information of the
channel (channel estimation).
Alternatively, the Godard-Kalman approach [7] to the
Kalman equalization [9] uses the state vector to be the true
(optimal) tap weights of the equalizer. Optimal tap weights are
calculated to minimize the expected mean squared distortion
E[|e(n)|
2
] of the lter output x
n
[7], where
e
n
= x
n
x
n
, (2)
and
x
n
=
H
n
y
n
. (3)
Vector
n
in (3) is the vector of the adaptive lter coefcients,
y
n
is the observable output of the communication system/the
lter input sequence and x
n
in (2) is the training sequence
(or sequence of tentative decisions), a priori known to the
receiver.
We use
opt
n
to denote the optimal (true) lter coef-
cients, which minimize the E[|e
n
|
2
] of the error (2). While in
time-invariant environments the optimal coefcients
opt
n
=

opt
are constant in time [7], if the channel is time-varying,
then
opt
n
becomes time-varying as well.
The main reason for adopting the Godard-Kalman algorithm
for adaptive model-based equalization over time-varying chan-
nels is the possibility to use the information of the propagation
conditions to describe how true lter coefcients
opt
n
vary
with time. Here we assume vector AR stochastic model for
the time-varying true coefcients
opt
n

opt
n+1
= H
n

opt
n
+q
n
(4)
where H
n
is a system (transition) matrix, q
n
is a white noise
sequence with var(q) =
2
q
I.
The lter output observation equation is given by
x
n
=
H
n
y
n
+
n
(5)
where
n
= x
n
(
opt
n
)
H
y
n
is the true noise with the
variance
2

.
The observation equation (5) and the state equation (4) form
the state-space model for the Kalman ltering problem state-
ment. The iterative solution can be derived in the following
form [4]

n+1|n
= H
n

n+1|n
+K
n

n
= x
n
y
n

n|n1
K
n
=
H
n
P
n|n1
y
T
n

+y
n
P
n|n1
y
n+1
(6)
P
n+1|n
=
_
H
n
P
n|n1
H
T
n

K
n
(
2

+y
n
P
n|n1
y
T
n
)K
T
n

+
2
q
I
619
where

n+1|n
= E(
n+1
|y
n
, y
n1
, ..., y
1
) (7)
and
P
n+1|n
= E(
n+1

n+1|n
)(
n+1

n+1|n
)
T
(8)
with initial conditions given by P(0) = P
0
.
The main advantage of the Godard-Kalman adaptive ap-
proach is adaptive ltering (equalization) without prior sepa-
rate channel estimation.
B. Adaptive Model Identication for the Kalman-Godard Fil-
ter
While the Godard-Kalman adaptive approach implicitly
captures the dynamics of time-varying channels, the issue of
determining model coefcients still remains.
AR(p)
model-based
Adaptive
receiver
Training
sequence
Decision
device
Adaptive AR(p)
model coefficients
identification
yk(n) xk(n)
xk(n)
adaptive
algorithm
Fig. 1. AR(p) model-based ltering with adaptive AR(p) coefcients
identication
In Fig. 1 we propose the Godard-Kalman ltering with
the adaptive model coefcient H
n
identication. Due to the
physical foundation of the channel process time-variations and
mass inertia effects, time variations of the system matrix H
n
are usually slow enough (at the time-scale of the higher rate
data). As shown in [4], the RLS algorithm is well suited
to adaptively estimate only very slow time-invariant system
matrix H
n
, Fig. 1, as follows

H
n+1
=

H
n
+K
n

n
= x
n
(

H
n

n1
)
H
y
n
(9)
K
n
=

1
P
n1
y
n
1 +
1
y
H
n
P
n1
y
n
P
n
=
1
P
n1

1
K
n
y
H
n
P
n1
We emphasise that the RLS algorithm is ill-suited to adapting
(time-varying)
opt
n
, but through the state-space formulation
well-suited to adapting the very slow time varying system
matrix H
n
.
IV. ADAPTIVE ALGORITHMS PERFORMANCE ANALYSIS
A. RLS Performance Analysis
1) Convergence: The rst order RLS convergence criterion
is given by [2], [4]
0 <
0
< 2 (10)
where
0
= 1 is the forgetting factor. The time constant
for convergence is given by [2], [4]
=
1
log |1
0
|
. (11)
The convergence criterion (10) and the time constant for
convergence (11) do not depend on lter input y(n) statistics.
Consequently the RLS is near-far resistant. As a Newton
method algorithm, the RLS provides a fast initial convergence.
However, in non-stationary environments, this fast initial con-
vergence usually has to be traded-off to improve stability and
tracking, as it is shown in simulations, Section V.
2) Second Order Stability and Tracking Analysis: The
relative steady state excess MSE or misadjustment for the RLS
algorithm with time-varying true weights is [2], [4]:
M
tot
=
E

=
p
0
2
+

2

2
y

p
2
+O(
2
0
)
= M
noise
+M
lag
+O(
2
0
) (12)
where E

= lim
n
E[|e(n)|
2
]; p = tr(R
y
)/
2
y
;
2
y
is
the variance of y(n); tr(R
y
) = trace(R
y
) is the sum of
the eigenvalues of R
y
= E[y(n)y
H
(n)];
2

=
2
q
/
2
; and
= I H is true speed of change of the time varying
true weights. The expression M
noise
in (12) is the excess
MSE due to noise adaptation and M
lag
is the excess lag
terms due to the time-varying optimal weights.
The second order stability requires [2], [4]
M
tot
= M
noise
+M
lag
<
1
3
(13)
Furthermore,
0,opt
and M
opt
can be found to be [2], [4]
M
opt
= p

for
0,opt
=

(14)
B. Model-based Algorithm Performance Analysis
1) Convergence: Following [2], [4], the rst order conver-
gence criterion is given by
0 <
1
2
max
< 2 (15)
where =
2

/
2

and
max
=
max
(R
y
) is the maximum
eigenvalue of R
y
.
The time constant for convergence for convergence is given
by
=
1
log
_

1
2 1

1
2 +1
_
(16)
where =
max
/
min
is condition number of R
y
and
min
is
the minimum eigenvalue of R
y
. Time constant for convergence
(16) depends on eigenvalue spread but much more weakly
than for LMS. For example, for for conditional number =

max
/
min
= 16, the LMS time constant is equal to 8 [4] and
the model-based algorithm time constant (16) is 3. Thus, the
model-based adaptive algorithm is much less near-far sensitive
than LMS.
620
2) Second Order Stability and Tracking Analysis: Again
following [2], [4], the relative steady state excess MSE or
misadjustment for the model-based adaptive algorithm with
time-varying true weights is
M
tot
=
E

=
p c
y
2
+
p c
y
2

+O(
2
)
= M
noise
+M
lag
+O(
2
) (17)
where c
y
= tr(R
1
2
y
)/p and =
2

/
2

Furthermore,
k,opt
and M
opt
can be found
M
opt
= p c
y

for
k,opt
=

. (18)
The second order stability requires M
tot
= M
noise
+M
lag
<
1
3
[2]. In addition
c
y
=
1
p
tr(R
1
2
y
)=
1
p
p

1
2
u

1
p

_
p

1
2
u
p=
1
p
_
p
2

2
x
=
x
(19)
leads to M
opt
(Kalman) M
opt
(RLS), (and
M
opt
(Kalman) M
opt
(LMS) [4]) and which means
that model-based adaptive algorithm offers superior tracking
to RLS [4].
V. SIMULATION STUDY
A. System Model for Simulation Study Adaptive MMSE Mul-
tiuser Receiver
The adaptive MMSE multiuser receiver structure Fig. 2,
presented in [13], [14], consists of a bank of adaptive MMSE
FIR lters along with the centralized (ML) detector part of the
receiver for data detection. The number of users in the cell of
interest K is assumed to be known. However, the number
K
I
of intercell interferers (I
1
, I
2
, ...I
KI
in Fig. 2) is unknown
[14].
While the non-linear centralized ML detection improves
signicantly the performance of the adaptive MMSE multiuser
receiver [14], it does not inuence the tracking performance
of the adaptive lters during the training period (the training
sequence is a priori known at the receiver). Since we focus on
the tracking performance of adaptive algorithms assuming an
a priori known input sequence, without loss of generality, we
consider only the bank of adaptive MMSE lters (the linear
front-end of the receiver).
The discrete-time received sample vector or the observation
of adaptive lter inputs at time n, y
n
can be expressed as
follows
y
n
= F
n
x
n
+n
n
(20)
where F
n
is the matrix of sampled system signatures which
includes spreading sequences and channel signatures, n
n
is
AWGN and x
n
contains the transmitted symbols.
The output of the MMSE lter for the kth user, k = 1, ..., K
at the time n is
x
k,n
=
H
k,n
y
k,n
(21)
Adaptive MMSE receivers bank
Decision
device
AR(p) model-based algorithm
with
adaptive AR(p) coefficients estimation
x
x
Training
sequence
Decision
directed
sequence
Adaptive
receiver K
Adaptive
receiver 2
Adaptive
receiver 1
y
Fig. 2. AR(p) model-based adaptive multiuser receiver with adaptive AR(p)
model coefcients estimation
where y
k,n
is discrete time received samples of kth user at
the time n,
k,n
are the adaptive lter coefcients of kth user
at the time n.
The coefcients are obtained adaptively during the train-
ing period, by minimizing the MSE, E[|e
k,n
|
2
],
e
k,n
= x
k,n
x
k,n
, (22)
where x
k,n
is the kth user training sequence at the time n, a
priori known to the receiver.
B. Simulation results
10
6
10
5
10
4
10
2
10
1
E
x
p
e
r
i
m
e
n
t
a
l

M
S
E
Normalized fading rate f
D
T
s
h=[1 0.25 0.5 0.125]

2
n
=0.1;
Number of users:6
RLS
LMS
Adaptive ident.
AR(2) GK
Fig. 3. Experimental MSE performance for AR(2) model-based algorithm
with adaptive model coefcients identication and observation-only based
LMS and RLS algorithms
We use the MATLAB multiple path fading channel
rayleighchan.m, 6-users, spreading gain of 6. The multi-
path channel consists of four-spaced channel taps with initial
621
values h = [1 0.25 0.5 0.125]. The channel noise variance
is given by
2
n
= 0.25. The feedforward lter for each user
consists of 24 taps. User k = 1 is the user of interest.
Simulations are performed for a range of normalized fading
rates f
D
T
S
.
Fig. 3 based on the average square error performance (ex-
perimental MSE) corroborates the theoretical performance
analysis provided in Section IV. The AR(2) model-based
adaptive algorithm with adaptive model identication provides
superior tracking performance to LMS and RLS, especially for
the medium speed fading conditions and the LMS algorithm
outperforms the RLS which provides very limited tracking
performance at medium and high-medium speed fading con-
ditions. On the other hand, if the channel is very slow (quasi
time-invariant) with f
D
T
S
0, then the superior initial
convergence speed of RLS provides a performance advantage
compared to LMS, given the reduced tracking requirements.
Even more, overall performance of the RLS approaches that of
the model-based algorithm, since the model-based algorithm
structure converges to the RLS algorithm.
10
6
10
5
10
4
10
2
10
1
Normalized fading rate f
D
T
S
E
x
p
e
r
i
m
e
n
t
a
l

M
S
E
data1nonadaptive iden. AR(1) GK
data2adaptive iden. AR(1) GK
data3adaptive iden. AR(2) GK
h=[1 0.25 0.5 0.125]

2
n
=0.1;
Number of users:6
1
3
11
2
Fig. 4. Experimental MSE performance for AR(2) and AR(1) model-based
algorithms with adaptive model coefcients identication and AR(1) model-
based algorithm with xed model coefcients H= I, = 0.85
Fig. 4 based on the experimental MSE shows that the
AR(2) model-based approach with adaptive model identi-
cation achieves notable performance improvement over the
AR(1) model-based approach with adaptive model identi-
cation, especially at the fast fading rate. Furthermore, Fig. 4
conrms the superiority of the model-based approach with
adaptive model identication over the same approach with
xed model parameters for a range of f
D
T
S
. At medium
fading rate f
D
T
S
= 5 10
4
, the adaptive approach with
a xed model parameters H
1
= I; = 0.85 provides
some performance improvement comparing to slow and high-
medium fading rate conditions, since the model is better suited
for particular fading conditions. However, the model-based
approach with adaptive model identication does not use any
a-priori information of the fading conditions.
REFERENCES
[1] P. Newson and B.Mulgrew, Model based adaptive channel identica-
tion algorithms for digital mobile radio, in Proc. IEEE International
Conference on Communications (ICC), vol. 3, May 1994, pp. 1531
1535.
[2] V.Solo and X. Kong, Adaptive Signal Processing Algorithms - Stability
and Performance, 2nd ed. Englewood Cliffs, NJ: Prentice Hall, 1995.
[3] S. Haykin, Adaptive Filter Theory, 4th ed. Upper Saddle River: Prentice
Hall, 2002.
[4] Z. B. Krusevac, P. B. Rapajic, and R. A. Kennedy, Model-based
adaptive algorithms for time-varying communication channels with
application to adaptive multiuser detection, in Proc. IEEE Vehicular
Technology Conference 2006-spring (VTC-spring06), Melbourne, VIC,
Australia, May 2006.
[5] K. E. Baddour and N. C. Beaulieu, Autoregressive modeling for fading
channel simulation, IEEE Transactions on Wireless Communications,
vol. 4, no. 4, pp. 16501662, July 2005.
[6] M.McGuire and M.Sima, Low-order Kalman lters for channel esti-
mation, in Proc. IEEE Pacic Rim Conference on Communications,
Computers and Signal Processing,(PACRIM), Aug. 2005, pp. 352355.
[7] D. Godard, Channel equalization using a Kalman lter for fast data
transimision, IBM J.Res.Dev., pp. 267273, 1974.
[8] S. McLaughlin, B. Mulgrew, and C. F. N. Cowan, A novel adaptive
equaliser for nonstationary communication channels, in Proc. IEEE
International Conference on Acoustics, Speech and Signal Processing,
(ICASSP), vol. 2, May 1989, pp. 944 947.
[9] S. McLaughlin, Adaptive equalization via Kalman ltering techniques,
IEE Proc.-F, vol. 138, no. 4, pp. 388396, Aug. 1991.
[10] C. Komninakis, C. Fragouli, A. Sayed, and R. Wesel, Channel estima-
tion and equalization in fading, in Thirty-Third Asilomar Conference on
Signals, Systems, and Computers, Pacic Grove, CA, 1999, pp. 1159
1163.
[11] S. Haykin, A.H.Sayed, J. R.Zeidler, P.Yee, and P.C.Wei, Adaptive
tracking of linear time-variant systems by extended RLS algorithms,
IEEE Transactions on Signal Processing, vol. 45, no. 5, May 1997.
[12] Z. B. Krusevac, P. B. Rapajic, and R. A. Kennedy, Optimal implicit
channel state estimation for nite-state Markov communication chan-
nels, in Proc. IEEE Int. Symp. on Inform. Theory (ISIT06), Seattle,
Washington, USA, July 2006, pp. 26572661.
[13] P. B. Rapajic and B. Vucetic, Adaptive receiver structures for asyn-
chronous cdma systems, IEEE Journal on Selected Areas in Commu-
nications, vol. 12, no. 4, pp. 685697, May 1994.
[14] P. B. Rapajic and D. K. Borah, Adaptive MMSE maximum likelihood
CDMA multiuser detection, IEEE Journal on Selected Areas in Com-
munications, vol. 17, no. 12, Dec. 1999.
622