You are on page 1of 4

A DETERMINISTIC APPROACH TO BLIND IDENTIFICATION OF

MULTI-CHANNEL FIR SYSTEMS*

Hui Liu, G u a n g h a n Xu Lang T o n g


Dept. of Electrical & C o m p u t e r Engineering Dept. of Electrical L S stems Engineering
The Universit of Texas at Austin university of Zonnecticut
Austin, $X 78712-1084 Storrs, CT 06269-3157

Abstract signals can also be modeled as a multichannel system [6].


linlike most of the probabilistic methods, our algorithm
Conventional blind channel identification algorithms is parametric and it only involves solving a set of linear
are based on channel outputs and the knowledge of the equations, which may lead to simple implementa.tion. Sim-
probabilistic model of a channel input. In some practical ulation results show that the identification can be realized
applications, however, the input statistical model may not based on only 50 output d a t a symbols. Let us begin our
be known, or there may not be sufficient data t o obtain ac- discussions with the problem statement.
curate enough estimates of certain statistics. In this paper,
we consider the system input to be an unknown deterniin- 2 Problem Statement
istic signal and study the problem of blind identification
of a channel which can be decomposed into multichannel Consider a baseband communication problem
FIR systems driven by a input signal. A new determinktic
blind identification algorithm based solely on the system
outputs are proposed. h'ecessary and sufficient identifiabil-
z ( t ) = C s ( k ) h ( t- kT) + n(t), (1)
k
ity conditions concerning the channel and the deterministic
input signal are also presented. where a typical channel impulse response h ( l ) is given in
Figure 1.
1 Introduction
Since Sato [l]first proposed the innovative idea of self-
recovering (blind) adaptive identification, blind channel
identification and equalization, as a research problem, has
been extensively studied by many researchers [2, 3, 4 , 5 , S]
and significant progress has been made in this area. How-
ever, the above existing blind identifiation methods are
blind only in the sense that they do not require the ex-
plicit knowledge of the system (channel) input; they all
require the same statistical assumption of the input, e.g.,
all the input symbols ace white. In many practical situa-
tions, such as a fast Raleigh fading environment, however,
the knowledge of the input may not be available, or there
are not enough d a t a samples to acquire a reasonably wcu-
rate statistical estimate. In this case, the input may only
be treated as an urbitrury deterministicsignal. This brings
up a new challenge in the u e a of channel identification: Is Figure 1: Special Channel with L =2 .
it possible to identify channel(s) without any knowledgv of
the input? Without loss of generality, we will establish our data
In this paper, we show that, under certain conditions, it model for the special case that h ( t ) lasts for L = 2 baud
is possible to achieve blind identification of FIR channels period T. If we oversample z ( t ) , Le., T, = T / M , and
based only on the channel outputs. A deterministic algo- collect the data samples of each baud period to form the
rithm of blind channel identification is presented. More new vectors x k = [ z ( M k ) , z ( M k + l ) , . . z ( M k + M - l ) I T ,
importantly, we also obtained the sufficient and necessary we have
conditions under which such blind identification is possi-
ble. The basic idea behind this deterministic approach is x(k) = h(a)s(k-2)+h(l)s(k- l)+h(O)s(k)+n(k).(2)
to exploit different exhibitions of the same input signal by
multiple FIR channels. Here, the concept of multichan- where nk donates the white noise. Let us look at the t t h el-
nel should not be limited to multiple physical receivers ement +,(k) of x ( k ) , and express it in terms of the symbols
or sensors. Temperaly ove [sampled digitd communication t.9(k)}:

"'Thiswork is supported in part by University Research In-


stitute at the Ilnivrraity d Texas at Austin
%(k) = h , ( l ) s ( k- 2) + h,(l)s(k - 1) + h,(O)s(k) + .,(e),
(3)

IV-581
0-7803-1775-0/94 $3.00 0 1994 IEEE
where n , ( k ) is the i t h element of n h and h , ( k ) denotes the them and write a larger set of linear equations in terms of
A
ittielement of h(n). n = 0 , 1 , 2 . For general cases, i . e . . h l , .. , , h, or simply h = [ h l , . . , hLITand solve all the
I. # 3, ( 3 ) becomes channel responses simultaneously. Denote X ' ( L ) as
L-1

z , ( k ) = x h , ( l ' ) s ( k- I ) = h , ( k )0 s ( k ) , (4) 0 '.. 0 Xi+,(L) -X,(L) 0 0


1=0
'. 0 (M-i)
where 0 denotes convolution. Spatial oversampling. i.e. 0 ... 0 XM(L) 0 "' -X,(L)
the same signal is sampled via multiple receivers, can hr,
treated the same way as temporal oversampling and (4)
still holds. where each block, e.g., 0 or {Xm(Z)},has the size, i.e.,
T h e blind channel identification problem can be stated (N - L +
+ 1) x (Z 1). In the noise free case, h is in the
w follows: given the observations of chunnel output { x ( k ) ] null space of the following large matrix.
or {z,(k),i= 1 , . . . , M . k = 0,. . . , N ) , estimate the chan-
nel {h,(.))E,. Blind channel equalization i s to further es-
timate the source sigrrals Is(.)}.
3 The Proposed Algorithm
By (4), for any pair of two noise-free outputs r , ( t ) and where the actual size ofX(L) is of (N-L+l)- X(Z+ '

s,(k), +
1)(M 1). When the channels are corrupted by noise, we
can estimate h by solving a least squares problem shown
r , ( k ) = h,(k) G ) . g ( k ) , z,(k) =: h l ( k ) C s(k1, (51
below:

-
where (1) indicates convolution. Then
h , ( k ) i . ) z * ( t )= h j ( k ) O ( h , ( k ) O s ( k ) l
= h , ( k ) O ) j h , ( t ) O s ( k ) )= h . ( k ) O ~ , ( k ) . ( 6 )
==,(k)
where h is subject t o certain nontrivial constraint, e.g.,
The above equation shows that the outputs of each chan- Ilhll = 1 or c"il = 1 for some constant vector e . Note
nel pair are linearly related by their channel responses. that the deterministic blind identification problem formu-
ITearly, if we have adequate data samples of the outputs, lated above is linear whereas most statistical approaches
by (6), we can write out an overdetermined set of linear are nonlinear. Alt.hough the size of the matrix X may be
equations involving h , ( . ) and h j ( . ) , Under certain condi- large for large N , M , and L,the computation cost may be
tions which we will elaborate later, h , ( . ) and h , ( . ) can be reduced by exploiting the sparsity and the block Hankel
det.ermined uniquely np to a scalar multiple. The use of
structure of this matrix.
surh a cross relation between each output pair is the basic
klea behind our deterministic blind channel identification. 3.1 Detection of Channel Order L
Note that this structure is not availa.ble in single channel In the above, we assumed that the maximum order of
system. multiple channels L is known. In practice, it is not nec-
More specifically. for k = 1 , . . . , N , where N is the niim- essarily true and the first step is to detect the maximum
her of synthols in t.he received data z , ( k ) and z , ( k ) , (ti) order L. One method of estimate L is constructing the
becomes N - L +
1 linear equations involving h , ( . ) and above X using a larger order L e ( > L ) and checking its
b,(.): rank. Let us first consider the noise-free case. If we use
a larger order L , and form X,(L,) as in (8). We note
+
that first L 1 columns of X , ( L , ) is a subset of the ma-
trix X,(L). By taking account of all the possible pairs of
{X,,Xj}, we can claim that any linear combination of the
following L, - L + 1 vectors-

L,-L Le-L-,
... n

X , " ( L )=
is a solution 1,o X(L,)h = 0, Z, > L. We can show
x,,,(,A-L) z",(N:Li-l) ...
that the vectors in (11) span the null space of X(Ze). The
(8)
+ +
rank of X will be Y~ = ( L , 1)(M 1) - ( L e - L + 1) =
For each pair of ( z , ~ ) , we can always write out. a set
+ +
hfL, ibf L . Therefore, L can be easily determined
by checking the rank of X ( L , ) . IC there is additive noise,
(if linear equations'. Here we propose to combine all of we can detect I,, - L + I n by checking how many smaller
'The equation f o r t w o channels i j similar to n special singular values of X are close to one another (on). More
case of the algorithm independently developed by Giirelli and objective criteria such as MDL, AIC can also be used to
Nikias [7]. detect L .

IV-582
3.2 Basic Algorithmic Procedure # of Symbols 1 # of Receivers I Oversamp. Rate
m - -m"""
-I
n I 3 I
I
3
I
I
1. Form X(L,) as in (9) based on an ouerestimatedorder SNR I Channel Length I # of Trials
L.. 10-40 I 4T I 500
2 . Perform an SVD or FSD to estimate how much
smaller singular values are the noise singular values
and detect the maximum channel order L. Table 1: Simulation P a r a m e t e r s
3 Use the estimated L to form X(L) as in (9) and find
h subject to some nontrivial constraint to minimize
By Theorems 1-2, if the coprime condition is satisfied,
IlX(L)hll. +
the channel is identifiable if p 2 2L 1 and it is not if
1 From h , we can have t.he channel responses for all the
channels, i.e., h j , . . , , h M .
p < L +
1. What happens when L +
1 2 p 5 2L I? +
The following two theorems will address this issue and give
necessary and sufficient identifiability conditions, in which
Now we have the algorithm for multichannel blind iden- the conditions on input modes and channels cannot be
tification. What is not clear is whether or not the true decoupled, unlike those given in Theorems 1-2.
channel responses are the unique solution of ( 1 0 ) .
Theorem 3 The suficient and necessary conditions for
4 Identifiability Conditions hlind multichannel identification can be stated as f o ~ ~ o w s ,
Identifiability conditions .have been extensively studied 1. the input s ( k ) has at least L 1 + +& modes
in system identification [8, 91. However, blind channel z ..' zp;
identification and conventional system identification are 2. t h e is no polynomial g ( z ) of order between 1 and
different since the latter assumes that the input signals P - 1 such that the following equation holds for e =
are known and controllable while the former does not. To 1 , . . .(M ,
study identifiability conditions of the proposed approach,
two issues need to be considered (1) the condition of the
identifiable channels; (2) the condition cf the input. These
two conditions may not be decoupled.
Observation:
The multichannel system c m be identified uniquely by solv-
ing linear equations Xti = 0 iff the data matrix X i.r of where u ( z ) i s the characteristic polynomial of the in-
rank M ( L + 1) - 1. put and { f k ( z ) ] f = , can be a n y polynomials.
l'he following three theorems give more explicit expres-
sioiis and provide more insight. Due tu space limitation, If there exists such a g ( z ) which satisties ( l a ) , {k,] become
we cannot provide proofs of the following theorems, which the ambiguous channel response which are not identifiable
will appear in the full paper [lo]. based on their outputs. An important observation can
be made regarding the number of channels M . First, if
Theorem 1 (Sufficient Condition) The bhnd ide ntifi-
M = 1, then the bound goes to infinity and the channels
calion problem f4ns a i ~ n i q u zsolution i f
are not identifiable which makes sense. As M increases, it
requires fewer modes to identify the channels. When A4
I . the polynomials { h , ( z ) ] E l are coprime or do not is reasonably large, the number of modes required on the
share a n y common roots, +
input is approximating L 1, which is the minimum num-
2. the input { s ( i ) , i = 1 , . ..,A')has 2L +
1 or more ber of modes required in standard system identification.
modes, or it can be ezpressed a8 a linear sum of 2 L + 1 In other words, there are sufficient number of ,hannels,
exponentials, blind identification (with unknown s(.)) and conventional
system identification (with known s(.)) are not SO different.
where h . ( z ) = h , ( O ) + h , ( l ) z - ' + - . . and L i s the mazimurn
order among { h ,( z ) ] z l 5 Simulation Results
Computer siniulat.ions were conducted to intuitively
I n fact, these conditions are the same ~ t sthose for identi- evaluate the performance of the proposed algorithm in
fying a rational function with the denominator and numer- comparison to that. of the existing stochastic algorithm
ator both of order I, 191. It is true that, our system is only [6]. In this simulation, two antennas were used and the re-
of order L , and can be sufficiently identified wit.h kiiown ceived data were sampled twice of the symbol rate. T h e in-
input of L + 1 modes (sometimes referecd to as persistcnt put signal type is QPSK. For simplicity of comparison, we
excitation of order L $- 1 [a]). For blind identification, t,he assumed that the channel order L is known. The key sim-
input is unknown and the order requirement on the input. ulation parameters are also summarized in Table 1. The
becomes 2 L + 1 . This relation can be shown more clearly FIR channels are shown in Figure 2.
by rdormulating the blind identification problem to be a In the first simulation study, we fixed the SNK to 20dB
standard SIMO system identification prcrblem [IO]. and varied the number of symbols from 50 300. Figure 3
sliows the Root-Mean-Square-Errors (RMSEs) of the chan-
Theorem 2 (Necessary Conditions) The channels c a n nel estimates from both the stochastic and deterministic
no/ be unrquely identrfii:d zf ( i ) there i:r o common zero methods. From this figure, we can see that the determin-
s h a w d by all charanek, OT (2) the input has less thun 1. + 1 istic method always performs better than the stochastic
moi1r.s. method, especially when the number of symbols is small.

IV-583
The main reason is that the stochastic method relies on the
statistical estimates of the received signal which are not so
accurate when we have a small number of d a t a symbols.
On the other hand, the deterministic method never as-
sumes any knowledge of the input statistics and it is not
so sensitive to the number of symbols.
In the second simulation study, we fixed the number of
symbols to be 50 and varied the SNR from 10
Figure 4 gives the RMSEs of the channel estimates from
40dB. -
these two methods. Despite the fact that the deterministic
method always surpass its counterpart, it is interesting to
iiote that the performance curve of the stochastic method
liats out after the SNR reaches 20dB, while that of the
proposed method still declines. This is due to the same
finite d a t a effect which cannot be cured by increasing SNR.
However, since the additive noise is the only cause of the
estimation error in the deterministic method, an increase
of SNR certainly leads t o its performance improvement
-
Figure 3: Comparison of t h e deterministic (solid) a n d
stochastic (dash) m e t h o d s for N = 50 300.

6 Conclusions
In this paper, we present a blind algorithm for identify-
ing multichannel FIR systems with an unknown determin-
istic input. The algorithm based solely on the system out-
puts and it only involves solving a set of linear equations
Necessary and sufficient identifiability conditions regard-
ing the input and channels are derived, which are useful in
further algorithm development of deterministic blind iden
tification. Computer simulations results demonstrate the
p t e n t i a l of the proposed algorithm.

1 ..... ......... $. . . . . . . . . . .
~

.............................. ....
., ....................... . ..
4 6
R s m i v e t i:real part
6
Receiver 1: lmginar/part -
Figure 4: Comparison of t h e deterministic a n d
stochastic m e t h o d s for S N R = 1 0 40 d B .

....... ........ ....... [5] Z. Ding, R.A. Kennedy, B.D.O. Anderson, and C.R.
Johnson, ‘%-convergence of Godard blind equalizers
........... .............. .... ....
in data communication systems”, IEEE Trans. Com-
. . munications, pages 1313-1337, September 1991.
.l2 4 6 6 [6] L. Tong, G. Xu, and T. Kailath, “Blind Identifica-
Rnwber 2:real pan Reoeiver 2: imginaIy pan
tion and Equalization Based on Second-Order Statis-
tics: A Time Domain Approach”, To appear in IEEE
Figure 2: Chitnnel responses ctf t,wo receivers Trans. on Information Theory.
[7] M.1. Giirelli and C.L. Nikias, “A New Eigenvector-
Based Algorithm for Multichannel Blind Deconvolu-
References tion of Input Colored Signals”, In Proc. ICASSP’93
[I] Y. Sato, “A Method of self-recovering equaliza- Conf., volume IV, pages 448-451, Minneapolis, MN,
tion for multilevel amplitude-modulation”, IEEE April 1993.
Trans. Commun., 23(6):679-682, June 1975. [8] E. W. Bai and S . S. Sastry, Persistency of excita-
[Z] A. Benveniste and M. Goursat, “Blind equalizers”, tion, suddicient richness and parameter convergence
IEEE Trans. Commun., 32(8):871-883, August 1.984. in discrete time adaptive control, System €4 Control
[3] J. K. Tugnait, Identificaiton of linear stochastic sys- Letters, G:153-163, 1985.
tem via second and fourth-order cumulant match- [9] L.Ljung, System Identification, Prentice-Hall, Engle-
ing, IEEE Tmns. Information Theory, IT-33:393- wood Cliffs, NJ, 1987.
407, May 1 9 8 i .
[IO] G. Xu, H. Liu, T. Lang, and T . Kailath, “A Determin-
[4] W . A . Gardner, “A New Method of Channel istic Approach to Blind Channel Identification”, sub-
Identification”, IEEE Trans. o n Communications, mitted IEEE Trans. on Signal Processing, Dec. 1993.
30(6):813-817, June 1991.

IV-584

You might also like