Novel adaptive channel equalization algorithms

for increasing the efficiency

of wireless communication systems
Ph.D. Theses

Kovács, Lóránt

Scientific supervisor:
Prof. Dr. János Levendovszky

Budapest University of Technology and Economics

Department of Telecommunications
2007
1 Introduction
Wireless communication and networking tend to have a deep penetration into
present-day society. Services and trends such as e-business, e-administration ... etc.,
and the ever-present requirement for mobile access pose huge challenges to wireless
communication technologies, where available frequency bands have become scarce
and hard-to-come-by. The key concept to accommodate these needs is spectral
efficiency which summarizes the performance of given technology into a measure in-
dicating what is the nominal dataspeed achieved over 1 Hz of radiospectrum. Hence,
wireless development is driven by the motivation of increasing spectral efficiency in
order to pave the way towards low cost broadband services. This objective implies
that broadband services are to be implemented over narrowband radio channels
which makes them susceptible to selective fading due to multipath propagation re-
sulting in severe performance degradation [13]. This prompts the development of
novel algorithms being able to facilitate very high speed information transmission
and sophisticated services but adhering to the present limits of technology and spec-
tral resources at the same time. As a result, the thesis focusing on developing novel
channel equalization algorithms for communication networks which achieve low bit
error rate and therefore increase the spectral efficiency. These algorithms prove to
be instrumental to combat both InterSymbol Interference (ISI) and Multi Access
Interference (MUI) in order to avoid large scale drops in system performance and
increasing the real data speed.
New development took place in the following domains

1. Novel blind equalization algorithms have been introduced in order to increase

the real data speed by eliminating the training sequence. In this way, the
spectral efficiency can be increased by omitting the overhead related to the
training sequence in the GSM packet header.

2. New algorithms have been proposed which can directly minimize the Prob-
ability of Error (instead of minimizing the mean square error (MSE) or the
peak distortion (PD)). Hence, the new equalization supports better QoS com-
munication by significantly decreasing the bit error probability.

3. A new channel identifier algorithm have been developed where not only the
filter parameters but also the model degree is under adaptation, which helps
to avoid both over- and undermodelling. Hence, multiple simultaneous iden-
tification problems are carried out with optimal resource management (i.e.
multichannel identification can be performed by a single DSP). This problem
is receives a great interest in identifying several channels simultaneously on
the BS.

The convergence of the algorithms introduced by the thesis is established by

analytical tools relying on the theory and tools of stochastic approximation. The
speed of convergence and the performance are analyzed by extensive simulations on
several practical channel models. As a result, the novel algorithms can be directly
applied in high-data-rate wireless communication networks.

2
1.1 State-of-the-art
The first attempt to develop an adaptive channel equalizer appeared in the work
of Lucky, Weldon and Saltz in 1965 [14], where the optimization of the equalizer
coefficients is based on the Peak Distortion criterion. Widrow introduced the Mean
Square Error criterion in 1967 [15]. These classical methods apply finite impulse
response (FIR) filters as equalizer which is followed by a threshold detector in order
to carry out fast symbol-by-symbol decision. Furthermore, these criterions yield very
simple recursive algorithms (linear stochastic differentia equations) as well. The
mathematical background of stochastic approximation is worked out by Robbins,
Monroe, Kushner and Clark [16, 17].
The development of the classical methods is summarized in Figure 1. (Detailed
state-of-the-art can be found in the dissertation, in section 1.1.1.)

Evolution of channel equalization

? ? ?
Equalizer architec- Increasing the speed of Equalization methods
ture: Decision-feedback convergence: Recursive minimizing directly
equalization ([18]), Least Squares algorithm the Bit Error Rate
nonlinear equalization [21], averaged stochastic (minBER equalization)
(maximum likelihood approximation [22] [23, 24, 25]
sequence estimation
[19], nonparametric
detection [20])

? ?
Blind adaptive meth- Equalization of Mul-
ods for learning omit- tiuser Systems (the
ting the training se- equalizer should com-
quence [26, 27, 28, 29] pensate not only the
Intersymbol Interfer-
ence, but Multiuser
Interference as well)
[30, 31, 32, 33]

Figure 1: Development of channel equalization

Based on these preliminaries the dissertation addresses the following issues:

• construction of small-complexity, real time, blind equalizer for DS-CDMA;

• convergence analysis of the decision directed LMS algorithm;

3
 • new real time equalization based on the minimum Bit Error Rate strategy;

• Optimal resource management of multiple system identification tasks by adap-

tive model degree estimation;

2 The methodology
The aim of the research is the development of new algorithms where the mathe-
matical challenges are (i) proving the optimality of the equalizer in steady state
in the sense of a given criterion, (ii) proving the convergence in probability, (iii)
demonstration of the performance by extensive simulations.

3 Overview of the results

The structure of relations between technological motivations, new scientific results
and outcomes are depicted by Figure 2.

Figure 2: Motivations, results and outcomes

4 The model
The general model of the communication systems by which the new algorithms are
treated is depicted by Figure 3.
The denotions used in the text:

• yk is the transmitted information symbol in the kth time slot. yk is i.i.d.

sequence of Bernoulli distribution with p = 0.5 parameter;

• hk , k = 0, ..., M is the impulse response of the channel, where M + 1 denotes

the span of the distortion (which comes from ISI and/or MUI);

4
 νk noise
channel equalizer
? xk ȳk
yk - hk - +j - wk - sgn(.) - ŷk

Figure 3: Model of digital communication system with linear distortion and additive
Gaussian noise

• νk stands for the noise in the kth time slot, which is assumed to be Gaussian
with p.d.f.:
1 1 T −1
p(u) = p e− 2 u K u (1)
N
(2π) det (K)
where K is the covariance matrix defined as: Kij = E {νk−i νk−j }. (In the case
of white noise K = N0 I, where I is the unit matrix and N0 is the noise power;
• xk is the received signal:
M
X
xk = hj yk−j + νk ; (2)
j=0

• In the case of block-transmission a matrix-vector denotion can be introduced

as y = [yk , ..., yk−N ]T , x = [xk , ..., xk−N ]T , n = [νk , ..., νk−N ]T ; and the received
vector can be expressed as
x = Hy + n,
in the case of single user systems H is of Toeplitz-type, while in the case of
multiuser systems block-Toeplitz-type;
• The so called training sequence consists of known input-output symbol pairs:
τ (K) := {(yk , xk ) , k = 1, ..., K};
• Detection means a mapping ŷ = f (x) yielding the decision on the sequence y;
• the equalizer make a linear mapping on the received signal in order to com-
pensate the distortion of the channel:
J
X
ỹk = wj xk−j . (3)
j=0

where wk denotes the weights of the equalizer, and the detection is carried out
by threshold decision: ŷ = sgn {ỹ};
• the convolution of the channel and equalizer is denoted by
M
X
qk = hi wk−i (4)
i=0

5
5 New scientific results
In this section the new theses are introduced.

Thesis A. Novel blind equalization algorithms

In this section a new blind ZF equalizer algorithm, and the convergence analysis of
the decision directed LMS algorithm will be introduced.

Thesis A.1.
A new blind adaptive decorrelation method (ADM) is constructed, that can decor-
relate its stationary input process:

wl (k + 1) = wl (k) − ∆ {ỹk ỹk−l − δl,0 } (5)

where ∆ is the step-size parameter and δl,0 is the Kronecker-symbol. Note that
algorithm (5) does not require the transmitted sequence yk , only its own output ỹk
is used for updating the equalizer coefficients wk .

Thesis A.2.
It is shown that algorithm (5) can be applied as blind equalizer in DS-CDMA sys-
tems. Almost sure convergence to the inverse of the channel matrix (ZF equalization)
is proven by applying the Kushner-Clark theorem:
n o
P lim W(k) = H−1 = 1
k→∞

where H denotes the channel matrix, and W is the Toeplitz-matrix formed by the
equalizer coefficients.
The following extensions of ADM (5) have been introduced:

1. in the case of additive Gaussian noise unbiased estimation of the empirical

covariance matrix may be used;

2. applying the averaging technique for stochastic approximation algorithms in-

troduced by Polyak and Juditsky [22] can be applied to ADM in order to
increase the convergence speed.

Performance analysis of ADM

The BER-SNR performance of ADM and its extensions are depicted by Figure 4 for
a DS-CDMA uplink channel, considering 5 active users, bit-synchronism, AWGN
channels between users and the base, and Gold-codes of length of 31 chips. Note
that the results of ADM are very close to the BPSK bound and far outperform the
conventional RAKE receiver (Single User Detector).

6
 BER−SNR performance of ADM for channel H2

−1
10

−2
10

−3
10

BER
−4
10

−5
10

−6
10 ADM
UADM
−7 SUD
10
BPSK bound

2 4 6 8 10 12 14
SNR

Figure 4: BER vs. SNR for channel model H(2) (defined in the dissertation in section
3.1.4.)

Thesis A.3.
It has been proven that the optimal weight vector of the Decision Directed Least
Mean Square algorithm (DDLMS)
à à J ! J
!
X X
wi (k + 1) = wi (k) − ∆ sgn wj xk−j − wj xk−j xk−i (6)
j=1 j=1

in steady state differs from the optimal weight vector of the LMS only in a constant
DDLMS
factor: wopt = R−1 r(1 − 2Pb ) = wopt
LMS
(1 − 2Pb ), yielding the same bit error rate.
Almost sure convergence to this solution has also been proven.

Thesis B. Novel minBER equalization algorithms

The classical equalization algorithms are based on suboptimal criteria (peak dis-
tortion, mean square error) in order to achieve easy-to-compute optimization algo-
rithms. The current DSP technology makes possible to implement more complex
and more efficient algorithms. Hence, the dissertation applies the minimum BER
equalization strategy, i.e. the direct minimization of the BER which is the main
performance measure of the digital communication.
In order to establish the new cost function, the BER will be expressed as a
function of the equalizer coefficients:
 
X PN PM +n
wn hl−n yl 
Pb (w) = (1/2L ) Φ  n=0q Pl=n . (7)
y∈Y N0 N w
n=0 n
2

where No is the power spectral density of the additive white noise, L is the
length of the overall channel impulse response q (see (4)), and Φ(·) is the
standard normal c.d.f. Y denotes the set of binary vectors for which Y =
{y = (y0 , y1 , ..., yL )|y0 = −1 and yi ∈ {−1, 1}}.

7
 The new cost function Pb can be minimized by the steepest descent method:

∂Pb (w (k))
wi (k + 1) = wi (k) − ∆ (8)
∂wi
where w(k) denotes the equalizer coefficient vector at the kth iteration, and
 ³ ´2 
PJ PM +n
∂Pb (w) 1 X  − n=0 wn l=n hl−n yl 
= q P exp   P J (9)
·
∂wi 2N0 n=0 wn 2
2L 2πN0 ( Jn=0 wn2 )3 y∈Y
" J M +i J M +n
#
X X X X
· ( wn2 ) · ( hl−i yl ) − wi ( wn hl−n yl )
n=0 l=i n=0 l=n

Direct application of this method is not straightforward because of the exponentially

growing number of terms (with the length of q, i.e. the span of of channel-equalizer
cascade) in the summation in algorithm (8). To overcome this difficulty new analyt-
ical upperbounds on BER and statistical sampling techniques are applied in order
to achieve real time, low-complexity, high performance algorithms.

Thesis B.1.
New low-complexity equalization algorithms are introduced based on newly derived
analytical upperbounds on BER:
q P P
−q0 + L Ll=1 ( M j=0 wj hl−j )
2
d) Pb (w) ≤ Φ (G(w)) , ahol G(w) = q P
N0 Jn=0 wn2

∗
e) Pb (w) ≤ Q(w) where
  q P PM   q P PM 
L 2 L 2
1 −q 0 + L l=1 ( w h
j=0 j l−j ) −q 0 − L l=1 ( w h
j=0 j l−j )
Q(w) = Φ  q P  + Φ q P  .
2 J 2 J 2
N0 n=0 wn N0 n=0 wn
where equations provided with aP* stands in the case of eye-openness of the overall
channel impulse response: q0 > Lk=1 |qk |.
Based on these differentiable upperbounds new equalizers are introduced using
the steepest descent method:

∂G(w)
wi (k + 1) = wi (k) − ∆ (10)
∂wi
and
∂Q(w)
wi (k + 1) = wi (k) − ∆ (11)
∂wi
The gradients used in (10) and (11) can be analytically calculated and are defined
in the dissertation in section 4.4.2.

8
Thesis B.2.
The expression of Pb in (7) can be seen as an expectation in the case of uniformly
distributed information symbols (which has been assumed in the model) as follows:
 
X PJ PM +n
1 wn hl−n yl 
PE (w) = L Φ  n=0 q Pl=n =
2 y∈Y
N0 Jn=0 wn2
1 X
= L PE (w, y) = Ey {PE (w, y)} . (12)
2 y∈Y

This expectation can be approximated by statistical sampling techniques, such as

the Li-Silvester method, which was originally applied in reliability analysis. Using
the dominant samples in the expectation, a sharp lowerbound can be achieved. The
following algorithm provides the dominant samples.

Set K = 4;
y0 = [−1, 1, ..., 1]
y1 = s̃−1 ([−1, 1, ..., 1, −1])
y2 = s̃−1 ([−1, 1, ..., 1, −1, 1])
IF r1 + r2 <= r3
THEN y3 = s̃−1 ([−1, 1, ..., 1, −1, −1])
ELSE y3 = s̃−1 ([−1, 1, ..., 1, −1, 1, 1])
N = {y0 , y1 , y2 , y3 };
Let S(w) ≤ G(w) denote the Li-Silvester lowerbound.
The gradient of the lowerbound (s(w)) is the same as g(w) (9),
using Y against N ;
Calculate the recursion wi (k + 1) = wi (k) − ∆si (w).

The denotions used in this algorithm are explained in the dissertation in section
4.5.1.

Thesis B.3.
The complete summation in the original gradient (9) can be substituted by one
randomly chosen term leading to the instantaneous estimation of the gradient.

wi (k + 1) = wi (k) − ∆ĝi (w(k), Y = y) (13)

where P {Yi = −1} = P {Yi = 1} = 0.5, and

 ³P ´2 
J PM +n
1  − w n
n=0 l=n h y
l−n l

ĝi (w(k), Y = y) = q PJ exp  P J  ·
2N w 2
2πN0 ( n=0 wn )2 3 0 n=0 n
" J M +i J M +n
#
X X X X
· ( wn2 ) · ( hl−i yl ) − wi ( wn hl−n yl ) (14)
n=0 l=i n=0 l=n

9
The almost sure convergence of this stochastic approximation algorithm to the op-
timum of the true gradient algorithm (8) is proven. Note that this method does not
consists of exponential summation, hence real time equalization can be carried out.

Performance analysis of minBER algorithms in Thesis B.

Figure 5. depicts the BER-SNR performance of the new minBER algorithms for
a non minimum-phase channel with 4 coefficients. (TGS stands for True Gradient
Search (8), LISI4 stand for Li-Silvester method using 4 dominant terms, STG means
Stochastic Gradient method.) Note that the new algorithms far outperform the
classical ones, namely the BER can be decreased even to the one hundredth related
to the classical solution.

Channel h7 J=6
−2
10
TGS
STG
−4
10 LISI4
LMS

−6
10

−8
10
BER

−10
10

−12
10

−14
10

10 12 14 16 18 20 22 24 26 28
SNR

Figure 5: BER vs. SNR in the case of channel h(7) (defined in section 4.7.) and 6
equalizer coefficients

Thesis C.: Novel Adaptive Filter Degree Algorithm

A lot of channel equalization algorithms require channel state information which
implies the use of adaptive channel identifiers. One of the main issues of system
identification is the proper estimation of the model degree. Underestimation of the
model degree results in poor model that cannot satisfactory approximate the original
system. On the other hand overmodelling is equivalent to waste of computational
resources. Hence, finding the exact model degree yields optimal resource manage-
ment in the case of performing simultaneous identification tasks on the same DSP.
(This case is typical e.g. in a mobile base station, where all the mobiles in the
cell communicate over different radio channels with the base). The novelty here is
the adaptivity, which gives rise to automatic model degree identification against the
classical information theoretical criteria by Akaike and Rissanen [34, 35], that are
off-line and require a priori statistical information.

10
Thesis C.1.
The new Adaptive Filter Degree Algorithm (AFDA) is an extension of the classical
Minimum Mean Square Error (MMSE) identification algorithm. The AFDA method
adopts not only the coefficients of the identifier filter but the degree of this filter as
well. In order to establish a stop-criterion for the algorithm the analytical expression
for the residual error as a function of the model degree has been derived as follows:
Eres (J) ∼ O (k(M − J)) where J is the degree of the identifier, and M is the exact
model degree, while k is a constant that depends on the covariance matrix of the
desired signal.
The basic steps of AFDA are depicted by the flow-graph in Figure 6. The core of
AFDA consists of sequential execution of the MMSE recursion with increasing filter
degree. (The initial degree should be smaller than the exact one). After convergence
of MMSE with a given degree, the residual error will be measured. Stopping the filter
degree adaptation can be based on a predefined level of error, or on the recognition
that in the case of overmodelling there is no more decrease of the residual error.

Figure 6: Flow-graph of AFDA

Thesis C.2.
Upperbound on the convergence time of AFDA has been derived which demonstrate
that the increase of convergence time caused by the sequential execution of the

11
MMSE is under a tolerable level, but at the same time finding of the exact model
degree is satisfied. The initialization of the MMSE of increased degree by the last
optimum satisfies the relatively fast convergence, because the two optimums (at
degree j and j + 1) are close to each other. The expression of the convergence time:
M ε
X log |wn (0)−hn|
TcAF DA = (15)
n=1
log(1 − ∆)

where n stands for the number of degree adaptation steps, ∆ is the step-size, M
denotes the exact model degree.

Performance analysis of AFDA

Figure 7. depicts the convergence curves of AFDA and MMSE in the case of under-
modelling and in the case of hypothetical knowledge of the exact model degree. By
applying AFDA the dramatic decrease of the residual error can be achieved, at the
cost of increased convergence time.

−1
10

−2
10

−3
10

MSE of AFDA
−4
10 MSE of MMSE (with exact knowledge of M)
MSE of MMSE (undermodeling: J=M−1=3)

1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Number of iterations

Figure 7: The squared error vs. the number of iterations in the case of channel h(4)
(defined in section 4.7. in the dissertation) and of 30 dB SNR

6 Implementation issues
The new results were motivated by the challenges regarding modern services imple-
mented over high speed wireless communication systems. Hence, the implementation
possibilities of the new algorithms were calculated applying the current DSP tech-
nology. A 8000 MIPS, 32-bit fixed point DSP (Texas TMS320C6416 – designed for
communication applications) was considered.

6.1 Blind equalization: Decision directed LMS

Applying TMS320C6416, the equalization and parallel updating of all the equalizer
weights can be carried out at a rate of 20 MHz, in the case of 6 equalizer coefficients.

12
On the other hand, if the rate of updates is reduced to 10 MHz 40 coefficients can
be handled.

6.2 Blind equalization: ADM algorithm

The implementation of ADM is highly limited by the transmission block size,
while the dimension of the channel matrix depends linearly on the block size. If
smaller block size is chosen, ADM can be efficiently implemented using the Texas
TMS320C6416: equalization and updating the weights can be performed at 10 MHz
in the case of block size of 16 and 10 active users. If the number of users is 30, the
rate of updates decreases to 1 MHz. In the case of large block size, the convolutional
form of the algorithm can be efficiently implemented (for details see dissertation,
section 3.1.).

6.3 MinBER equalization

Figure 8. depicts the limits of implementation of the different minBER algorithms
applying Texas TMS320C6416 DSP. Parallel updating of all coefficients in each time
slot is assumed. The realizable number of channel+equalizer coefficients are shown
versus the possible rate of updating. The True Gradient Algorithm, due to its
exponential complexity, cannot be realized at high speed, namely even in the case
of 5 equalizer+channel coefficient the limit of the rate of updates is about 1MHz,
while with the same frequency the other algorithms can even be executed with over
100 coefficients.
100

True gradient
90 Stochastic gradient
Bound d)

80
Number of channel+equalizer coefficients

70

60

50

40

30

20

10

0
1 2 3 4 5 6 7 8
10 10 10 10 10 10 10 10
fs [Hz]

Figure 8: Different minBER algorithms implemented on Texas TMS320C6416

6.4 Adaptive Filter Degree Algorithm

One attractive application field of AFDA is the channel identification of mobiles in
the cell of a given base. All the mobiles have different position yielding different
propagation properties and channel impulse response functions, which all should

13
be identified separately. Using modern DSP technology, parallel identification of
several impulse response functions can be carried out by using a single DSP. In the
case of Texas TMS320C6416 DSP 1000 identifier coefficients can be realized at 1
MHz rate of updates. The typical span of mobile channels does not exceed 10 time
slots that implies the parallel identification of over 100 mobile channels satisfying
optimal DSP resource allocation.

These facts imply that the new algorithms can be directly implemented in high-
speed wireless communication system in a real time fashion.

References
[1] L. Kovács, J. Levendovszky, and E.C. van der Meulen. A novel blind chan-
nel equalization algorithm minimizing the peak distortion in ds-cdma systems.
WSEAS Trans. Comm., 6:289–294, February 2007.

[2] J. Levendovszky, L. Kovács, A. Oláh, and D. Varga. A new blind signal pro-
cessing algorithm for decorrelation and multiuser detection. Hiradástechnika,
57:15–20, December 2002.

[3] J. Levendovszky and L. Kovács. A new blind signal processing algorithm for
decorrelation and channel equalization. In Proc. of 3rd EURASIP-IEEE Re-
gion8 Symposium on Video Processing and Multimedia Communications, pages
135–138, 13-15 June 2001.

[4] G. Jeney, Levendovszky, and L. Kovács. Blind adaptive stochastic neural net-
work for multiuser detection. In IEEE Vehicular Technology Conference 2001
Spring, pages 1868–1872, 6-9 May 2001.

[5] J. Levendovszky, L. Kovács, G. Jeney, and E.C. van der Meulen. New blind
signal processing algorithm for multiuser detection. In Proc. of the 22th Sympo-
sium on Information Theory and Communication Theory in the Benelux, pages
17–24, 15-16 May 2001.

[6] J. Levendovszky, L. Kovács, and G. Jeney. Novel adaptive signal processing

algorithms for multiuser detection. In Proc. of 4th EURASIP-IEEE Region8
Symposium on Video Processing and Multimedia Communications, pages 361–
365, 16-19 June 2002.

[7] L. Kovács and J. Levendovszky. Nonsupervised learning algorithm for mul-

tiuser detection. In Proc. of Polish-Czech-Hungarian Workshop on Circuit
Theory, Signal Processing and Telecommunication Networks, pages 157–164,
14-17 September 2001.

[8] J. Levendovszky, L. Kovács, and E.C. van der Meulen. New blind signal pro-
cessing algorithm for channel equalization. In Proc of the 23rd Symposium on
Information Theory and Communication Theory in the Benelux, pages 277–284,
29-31 May 2002.

14
 [9] L. Kovács, J. Levendovszky, and E.C. van der Meulen. Minimum probability
of error based equalization algorithms for fading channels. EURASIP Journal
on Wireless Communications and Networking, 2007.

[10] L. Kovács and J. Levendovszky. Minimális bithibaarány-stratégián alapuló

közel optimális csatornakiegyenlítés statisztikai mintavételezéssel. Hiradástech-
nika, 62:3–7, May 2007.

[11] J. Levendovszky, L. Kovács, A. Olah, D. Varga, and E.C. van der Meulen.
Novel sampling methodes for increased spectral efficiency in wireless commu-
nication systems. In Proc of the 24th Symposium on Information Theory and
Communication Theory in the Benelux, 2003.

[12] J. Levendovszky, L. Kovács, I. Koller, and E.C. van der Meulen. Optimal
resource management algorithm for adaptive modelling. In Proc. of the 20th
Symposium on Information Theory in the Benelux, pages 197–204, 27-29 May
1999.

[13] R. Steele and L. Hanzo, editors. Mobile Radio Communications. Wiley, 1999.

[14] R.W. Lucky, J. Salz, and E.J. Weldon. Principles of Data Communication.
McGraw-Hill, 1968.

[15] B. Widrow and S.D. Stearns. Adaptive signal processing. Prentice Hall, 1985.

[16] H. Robbins and S. Monro. A stochastic approximation method. Annual Math.

Statis., 22:400–407, 1951.

[17] H.J. Kushner and D.S. Clark. Stochastic approximation methods for constrained
and unconstrained systems. Springer Verlag, 1978.

[18] M.E. Austin. Decision-feedback equalization for digital communication over

dispersive channel. Technical report, MIT Lincoln Laboratory Lexington, MA.,
1967.

[19] G.D. Forney, Jr. Maximum-likelihood sequence estimation of digital sequences

in the presence of intesymbol interference. IEEE Trans. Inf. Theory, 18:363–
378, 1972.

[20] J. Levendovszky, E.C. van der Meulen, and Zs. Elek. Nonparametric detection
by feedforward neural networks. Neural Network World, 6:929–957, 2000.

[21] D. Godard. Channel equalization using kalman filter for fast data transmission.
IBM J. Res. Develop., 18:267–273, 1974.

[22] B.T. Polyak and A.B. Juditsky. Acceleration of stochastic approximation by

averaging. SIAM J. Contr. Optim.,, 30:838–855, 1992.

[23] O. Shimbo and M. Celebiler. The probability of error due to intersymbol inter-
ference and gaussian noise in digital communication systems. IEEE Trans. on
Communication Technology, COM-19:113–119, 1971.

15
[24] E. Shamash and K. Yao. On the structure and performance of a linear decision
feedback equalizer based on the minimum error probability criterion. In Proc.
Int. Conf. on Communications, pages 25F1–25F5, 1974.

[25] C. Yeh and J.R. Barry. daptive minimum bit-error rate equalization for binary
signaling. IEEE Trans. Comm., 48:1226–1235, 2000.

[26] Y. Sato. A method of self-recovering equalization for multi-level amplitude

modulation. IEEE Trans. Comm., 23:679–682, 1975.

[27] D. Hatzinakos and C.L. Nikias. Blind equalization using a tricepstrum-based

algorithm. IEEE trans. on Commun., 39:669–682, 1991.

[28] Y. Sato. Blind equalization and blind sequence estimation. IEICE Trans.
Comm., E-77b:545–556, 1994.

[29] Z. Xu and P. Liu. Blind multiuser detection by kurtosis maximiza-

tion/minimization. IEEE Signal Proc. Letters, 11, 2004.

[30] S. Verdu. Multiuser detection. Cambridge University Pres, 1999.

[31] M.K. Varanasi and B. Aazhang. Multistage detection in asynchronous code

division multiple access communications. IEEE Trans. Comm., 38:508–519,
1990.

[32] G. Jeney and L. Levendovszky. Stochastic hopfield network for multiuser de-
tection. In European Conf. Wireless Techn., pages 147–150, 2000.

[33] X. Wang and V. Poor. Blind multiuser detection: A subspace approach. IEEE
Trans. Inform. Theory, 44:677–689, 1998.

[34] H. Akaike. A new look at the statistical model identification. IEEE Trans.
Automatic Control, AC-19:716–723, 1974.

[35] J. Rissanen. Modeling by shortest data description. Automatica, 14:465–471,

1978.

[36] K. Elek and L. Kovács. Hiradástechnika Példatár. Müegyetemi Kiadó, 2004.

[37] R. Prasad, W. Mohr, and W. Konhauser. Third Generation Mobile Communi-

cation Systems. Artech House, 2000.

[38] S. Haykin. Adaptive filter theory. Prentice Hall, 1985.

[39] M.L. Honig and D.G. Messerschmitt. Adaptive Filters, Structures, Algorithms
and Applications. Kluwer Academic Publishers, 1984.

[40] A.J. Viterbi. CDMA, Principles of spread spectrum communication. Addison-

Wesley, 1992.

[41] J.G. Proakis. Digital Communications. McGrawHill, 1995.

16
[42] P. Rózsa. Lineáris algebra és alkalmazásai. Tankönyvkiadó, Budapest, 1991.
[43] L. Pap and S. Imre. A mobil hirközlés alapjai. Müegyetemi Kiado, Budapest,
2001.
[44] H. Kwakernaak and R. Sivan. Linear Optimal Control Systems. Wiley-
Interscience, 1972.
[45] F. Balazs, S. Imre, and G. Jeney. Korezserű mobil vevőalgoritmusok. kézirat,
2003.
[46] E. Biglieri, J.G. Proakis, and Sh. Shamai. Fading channels: information-
theoretic and communications aspects. IEEE Trans. Information Theory,
44:2619–2692, 1998.
[47] B.T. Polyak. New stochastic approximation type procedures (in russian). Av-
tomat. i Telemekh., N7:98–107, 1990.
[48] D. Ruppert. Efficient estimators from slowly convergent robbins-monroe pro-
cess. Technical report, School of Operations Research and Industrial Engineer-
ing, Cornell University, Ithaca, NY, 1988.
[49] G. Ungerboeck. Adaptive maximum-likelihood receiver for carrier-modulated
data-transmission systems. IEEE Trans. Comm., 22:624–636, 1974.
[50] J. Hagenauer, E. Offer, C. Measson, and M. Mörz. Decoding and equaliza-
tion with analog non-linear networks. European Trans. Telecomm., 10:659–680,
1999.
[51] R. Gold. Optimum binary sequences for spread spectrum multiplexing. IEEE
Trans. Inf. Theory, 14:154–156, 1967.
[52] E.H. Dinan and B. Jabbari. Spreading codes for cdma and wideband cdma
cellular networks. IEEE Commun. Magazine, 1998.
[53] V. Krishnamurty. Averaged stochastic gradient algorithms for adaptive blind
multiuser detection in ds/cdma systems. IEEE Trans. Commun., 48:125–134,
2000.
[54] S. Ulukus and R.D. Yates. A blind adaptive decorrelating detector for cdma
systems. IEEE J. Select. Areas Commun., 16:1530–1541, 1998.
[55] D. Das and M.K. Varanasi. Blind adaptive multiuser detection for cellular
systems using stochastic approximation with averaging. IEEE J. on selected
areas in comm., 20:310–319, 2002.
[56] L. Tong, G. Xu, and T. Kailath. Blind identification and equalization based on
second order statistics: a time domain approach. IEEE Trans. Inform. Theory,
40:340–349, 1994.
[57] N. Seshadri. Joint data and channel estimation using fast blind trellis search
techniques. IEEE Trans. Comm., 42:1000–1011, 1994.

17
[58] D.N. Godard. Self-recovering equalization and carrier tracking in twodimen-
sional data communication systems. IEEE Trans. Comm., 28:1867–1875, 1980.

[59] G. Picchi and G. Prati. Blind equalization and carrier recovery using a stop-
and-go decision directed algorithm. IEEE Trans. Comm., 35:877–887, 1987.

[60] J. Tugnait and T. Li. Blind asynchronous multiuser cdma receivers for isi chan-
nels using code-aided cma. IEEE Journal on Selected Areas in Communications,
19:1520–1530, 2001.

[61] L. Li and H. Fan. Blind cdma detection and equalization using linearly con-
strained cma. In Proc. of IEEE Intl. Conf. on Acoustics, Speech, and Signal
Proc. (ICASSP’00), pages 2905–8, 2000.

[62] M. Hadef, S. Weiss, and M. Rupp. Adaptive blind multiuser ds-cdma downlink
equaliser. IEE Electronics Letters, page online no: 20052902, August 2005.

[63] J. Yang, J.J. Werner, and G. Dumont. The multimodulus blind equalization
and its generalized algorithms. IEEE J. Select. Areas Commun., 20:997–1015,
2002.

[64] E. Moulines et al. Subspace methods for the blind identification of multichannel
fir filters. IEEE Trans. Signal Processing, 43:516–525, 1995.

[65] J. Shen and Z. Ding. Zero-forcing blind equalization based on subspace esti-
mation for multiuser systems. IEEE Trans. on Communications, 49, 2001.

[66] D.J. Gesbert, J. Sorelius, and A.J. Paulraj. Blind multiuser mmse detection of
cdma signals. Proc. 1998 IEEE ICASSP, pages 3161–3164, 1998.

[67] M. Torlak and G. Xu. Blind multiuser channel estimation in asynchronous

cdma systems. IEEE Trans. Signal Processing, 45:137–147, 1997.

[68] G.B. Giannakis and J.M. Mendel. Identification of nonminimum phase sys-
tems using higer-order statistics. IEEE Trans. Acoust. Speech and Signal Proc.,
37:360–377, 1989.

[69] C.Y. Chi and Chen C.H. Cumulant-based inverse filter criteria for mimo blind
deconvolution: properties, algorithms, and application to ds/cdma systems in
multipath. IEEE Trans. Signal Processing, 49:1282–99, 2001.

[70] D.K. Borah and P.B. Rapajic. Optimal adaptive multiuser detection in un-
known multipath channels. IEEE J. on selected areas in comm., 19:1115–1127,
2001.

[71] A.J. Bell and T.J. Sejnowski. An information-maximization approach to blind

separation and blind deconvolution. Neural computations, 7:1129–1159, 1995.

[72] P. Comon. Independent component analysis - a new concept? Signal Processing,

36:287–314, 1994.

18
[73] A. Hyvarinen and E. Oja. A fast fixed-point algorithm for independent compo-
nent analysis. Neural Computations, 9(7):1483–1492, 1997.

[74] O. Shalvi and E. Weinstein. New criteria for blind equalization of nonminimum
phase system and channels. IEEE Trans. on Information Theory, 36:312–321,
1990.

[75] M.L. Honig, U. Madhow, and S. Verdu. Blind adaptive multiuser detection.
IEEE Trans. Inform. Theory, 41:994–960, 1995.

[76] J. Levendovszky, W. Mommaerts, and E.C. van der Meulen. Hysteretic neural
networks for global optimization of quadratic forms. Neural Network World,
2:475–496, 1992.

[77] G.I. Kechirotis and E.S. Manolakos. Hopfield neural network implementation
of optimal cdma multiuser detector. IEEE Trans. on Neural Networks, 7, 1996.

[78] J. Schoukens and R. Pintelon. Identification of linear systems. Pergamon press,

2001.

[79] G. Xu, H. Liu, L. Tong, and T. Kailath. A least-squares approach to blind

channel identification. IEEE Trans. Signal Processing, 43:1982–2993, 1995.

[80] F. Glave. An upper bound on the probability of error due to intersymbol

interference for correlated digital signals. IEEE Trans. Inf. Theory, 18:356–
363, 1972.

[81] P. Chandra Kumar, P. Saratchandran, and N. Sundararajan. Minimal radial

basis function neural networks for nonlinear channel equalisation. IEE Proc.
Vision, Image and Signal Processing, 147:428–435, 2000.

[82] S. Chen, B. Mulgrew, and L. Hanzo. Least bit error rate adaptive nonlinear
equalisers for binary signalling. IEE Proc. Comm., 150:29–36, 2003.

[83] S. Chen, A.K. Shamingan, B. Mulgrew, and L. Hanzo. Adaptive minimum-ber

linear multiuser detection. In ICASSP 2001, pages 2253–2256, 2001.

[84] A. Dua, U.B. Desai, and R.K. Mallik. Minimum probability of error-based
methods for adaptive multiuser detection in multipath ds-cdma channels. IEEE
Trans. Wireless Comm., 3, 2004.

[85] A. Hjorungnes and M. Debbah. Minimum ber fir receiver filters for ds-cdma
systems. In IEEE GLOBECOM, 2005.

19