Sie sind auf Seite 1von 6

Comparison of hybrid adaptive blind equalizers for QAM signals

A. Labed
1
, A. Belouchrani
3
, A. Aissa-El-Bey
2;4
and T. Chonavel
2;4
1
EMP, Computer Science Department, Algiers, Algeria
2
Institut TELECOM; TELECOM Bretagne; UMR CNRS 3192 Lab-STICC, Brest, France
3
ENP, Electrical Engineering Department, El Harrach, Algiers, Algeria
4
Universit Europenne de Bretagne, France
ABSTACT
This paper compares different hybrid blind equalization algorithms used for QAM signals. In hybrid
equalizers, a penalty term, with zeros at constellation points coordinates, called constellation matching error (CME) is
added to the criterion of one of the standard algorithms, such as the constant modulus algorithm (CMA), the
multimodulus algorithm (MMA) or the recently proposed extended constant modulus algorithm (ECMA). Among the
CMEs, to be considered, we have recently introduced a new one which is the product of l
1
-norm of the deviations of
equalizer output from the constellation points. The hybrid algorithms, obtained by combining different CMEs with the
CMA or the ECMA are compared through simulations on 16-QAM, 64-QAM and 256-QAM signals transmitted over
different channels.
Key words: QAM signals, blind equalization, CMA, modified CMA.
1. INTRODUCTION
Blind equalizers are designed to filter transmitted signals without need to training sequences. In the context of blind
equalization, the constant modulus algorithm (CMA) which is easy to implement and has good convergence properties
[1-3], is very famous. However, besides its insensitivity to phase rotations, the CMA does not lead to sufficiently low
residual errors, if applied to blindly equalize quadrature amplitude modulated (QAM) signals. That is why other
algorithms like the multimodulus (MMA)[4], the square contour algorithm (SCA)[5, 6] and the extended CMA
(ECMA)[8] have been proposed. Unfortunately, even these new algorithms do not ensure good equalization of high
order QAM constellations and other extensions like that based on a CMA and a decision directed (DD) equalizer
operating concurrently[7]or the modified CMA (MCMA) have been developed. In the latter, the CMA criterion is
augmented with a penalty term, with zeros at constellation points coordinates, called constellation matching error
Photonics Applications in Astronomy, Communications, Industry, and High-Energy Physics Experiments 2009,
edited by Ryszard S. Romaniuk, Krzysztof S. Kulpa, Proc. of SPIE Vol. 7502, 75021T 2009 SPIE CCC code:
0277-786X/09/$18 doi: 10.1117/12.838258
Proc. of SPIE Vol. 7502 75021T-1
(CME) [11-13]. This hybrid approach (CMA+ CME) has been demonstrated to drastically reduce the convergence time
together with a lower residual error.
Hybrid algorithms can also be obtained by combining a CME with the MMA or the ECMA criterion. The resulting
cost functions are minimized to adaptively update the equalizer coefficients, generally by using the stochastic gradient
descent algorithm.
In this paper, we compare performances of hybrid adaptive blind equalizers in terms of mean squared error (MSE)
and convergence speed. We consider two families of hybrid cost functions: i) CMA+CME; and ii) ECMA+CME. We
recall that, the most common CME functions encountered in the literature are polynomials of high order and powers of
cosine functions[12,13]. To our knowledge, we are the first to use the idea of combining the ECMA criterion with a
CME in our paper [15]. Moreover, we have introduced a new CME in [15], which is the product of l
1
-norm of the
deviations of equalizer output from the constellation points. The combination CMA+l
1
-norm CME is referred to as the
l
1
-modified constant modulus algorithm (l
1
-MCMA) and the ECMA+l
1
-norm CME as l
1
-ECMA. We also consider the
CME function (Gaussian) proposed by Barbarossa and Scaglione in [11], in the context of hybrid instead of dual mode
equalization[15]. The resulting criteria are referred to as Gauss-MCMA and Gauss-ECMA respectively. These criteria
are compared with the existing Cos-modified constant modulus algorithm (Cos-MCMA)[13]. The simulation results
show the performance advantages in terms of mean squared error and that hybrid ECMA is able to correct the phase
rotations independently of the choice of CME.
2. STANDARD BLIND EQUALIZATION ALGORITHMS
Let s
k
, k=1,,N be a sequence of QAM symbols to be transmitted over a channel h.
The blind equalization scheme is:
k
Lc
l
l k l k
n s h x + =

1
0
(1)
With s
k
the channel input at time k, h the impulse response of the channel (with Lc complex coefficients), x
k
the
equalizer input signal and n
k
the noise.
The equalizer output symbol y
k
is given by:
k
H
k
X w y = (2)
The equalizer weights (Le taps) vector w and the vector of symbols used to update it at time k are respectively given by:
w =[w
0
,w
1
,,w
Le-1
]
T
and X
k
= [x
k
, x
k-1
,, x
k-Le+1
]
T
; where (.)
T
denotes transpose and (.)
H
hermetian transpose.
To tune the equalizer coefficients, a cost function is minimized according to these coefficients.
In the sequel of this section we give a brief description of standard algorithms cost functions.
1) The standard CMA cost function is:
( )

=
2
2
) (
m k CMA
R y E w J (3)
With E the expectation symbol and R
m
a dispersion constant given by: { } { }
2 4
/
k k m
a E a E R = ; the a
k
are
the QAM constellation points.
The equalizer weights vector is updated by:
( )
k k
H
k m k k k
w X X R y w w
2
1
=
+
(4)
with the step size.
Proc. of SPIE Vol. 7502 75021T-2
2) The MMA cost function is: ( ) ( ) { }
2
2
2
2
) (
i r MMA
R y R y E w J
ki kr
+ = (5)
In this expression, y
kr
and y
ki
refer to the real and imaginary part of y
k
respectively and the constants R
r
and R
i
are: { } { }
2 4
/
kr kr
a E a E R R
i r
= = . The equalizer weights are updated according to:
( )
k i ki r kr k k
X R y jy R y y w w
ki kr
) ( ) (
2 2
1
=
+
(6)
3) The ECMA is a particular case of the generalized constant modulus algorithms proposed in [9]. The generalized
complex modulus (similarly to Minkowki distance) is defined as ( )
p
p
i
p
r
p
z z z
1
+ = for p 1, z
r
and z
i
the real and
imaginary part of z respectively.
The ECMA criterion is given by:
( )
)
`

=
2
2 , 4
2
4
) ( R y E w J
k ECMA
(7)
The constant { } { }
2
4
4
4
2 , 4
/
k k
a E a E R = and the updating formula
) ( 4
2
4
3 3
2 , 4
2
4
1 k
k
ki kr
k k k
X
y
jy y
R y w w

=
+
(8)
3. HYBRID ALGORITHMS
The general form of hybrid version of the CMA criterion is:
( ) ( ) ( ) ( )
ki kr m k CME CMA MCMA
y jg y g R y E J J w J + +
)
`

= + =
2
2
) ( (9)
While the hybrid version of the ECMA criterion is given by:
( ) ( ) ( ) ( )
ki kr k CME ECMA MCMA
y jg y g R y E J J w J + +
)
`

= + =
2
2 , 4
2
4
) ( (10)
Where g(.) is a positive function that vanishes at constellation coordinates and the tuning parameter is introduced to
make a tradeoff between the amplitude and the constellation-matched errors.
The updating procedures for hybrid CMA and hybrid ECMA are respectively:
) (
1 k k k
H
k k k k k
X w X X e w w + =
+
(11)
and
(
(

=
+ k k
k
ki kr
k k k
X X
y
jy y
R y w w
k 2
4
3 3
2 , 4
2
4
1
) ( 4 (12)
with
ki kr
y u y u
k
u g
du
d
j u g
du
d
= =
= ) ( ) ( (13)
The function g used in the CME term can be chosen among the following functions:
The cosine CMEs functions
|

\
|
=
d
x
x g
k
C
2
cos ) (
2
or
|

\
|
=
d
x
x g
k
S
2
sin 1 ) (
2
;
where k is an integer and 2d is the minimal distance between symbols. In our comparisons, we will take
|

\
|
=
d
x
x g
C
2
cos ) (
2
and call the resulting hybrid criteria Cos-MCMA and Cos-ECMA respectively.
Proc. of SPIE Vol. 7502 75021T-3
The Gaussian CME
2
2
2
1
1 ) (

l
s x
Nc
l
G
e x g

=

=
with s
l
, l=1,, Nc are the constellation points and a parameter that controls the width of the function around
these points. The resulting hybrid criteria will be noted Gauss-MCMA and Gauss-ECMA respectively.
Our l
1
-norm CME given by:
i
M
i
L
c x x g =

=1
) (
Where c
i
are the constellation points coordinates (1, 3,) and M refers to the cardinal of the alphabet of
symbols (M=4 for 16-QAM and M=8 for 64-QAM,).
All the previous CME functions have the property to vanish at each symbol of the QAM constellation and are
symmetric around these points. It is also required from a CME to be uniform so that it does not favor or penalize
information symbols over others. This is not the case for the l
1
-norm CME which tends to favor high modulus symbols
over the others, especially in the case of high order QAM constellations (256-QAM,). Hence, we expect l
1
-MCMA
and l
1
-ECMA to perform worse than the other hybrid algorithms.

4. COMPARISONS
To compare performances of hybrid versions of CMA and ECMA algorithms, independent and identically
distributed (iid) 16-QAM, 64-QAM and 256-QAM symbols were generated. A 20 taps equalizer (Le = 20) was used.
All convergence curves (MSE in dB, versus the iteration number), were averaged over 100 Monte Carlo independent
runs. For each QAM order, we have chosen the parameters of the updating formulas (, and ) to allow all hybrid
algorithms converge to the same residual error.
On Figure 1, we compare performances of CMA, MMA, ECMA, l
1
-MCMA, Cos-MCMA and Gauss-MCMA for
16-QAM signals transmitted over the frequency selective (with complex coefficients) channel used in[13], defined as
h
1
= [1 0.1294 + 0.483j]. The parameters of the stochastic gradient were set to (= 2 10
-5
), (= 2.5 10
-5
),
(= 7 10
-6
), (= 8 10
-6
, = 1.5), (= 8 10
-6
, =100/) and (= 7 10
-6
, =5, =0.5) respectively. From this
figure, it is obvious that hybrid algorithms perform much better than the standard ones. We also notice that l
1
-MCMA
has faster convergence than Gauss-MCMA and Cos-MCMA.
On figure 2, the convergence properties of l
1
-ECMA, Cos-ECMA and Gauss-ECMA are compared for 16-QAM
transmitted over channel h
1
. The parameters of the updating formulas are = 7 10
-6
for ECMA, (= 7 10
-6
, =2) for
l
1
-ECMA, (= 4 10
-6
, =100/) for Cos-ECMA and (= 1.8 10
-5
, =5, =0.5) for Gauss-ECMA. In this case the
Gauss-ECMA outperforms the l
1
-ECMA which by itself converges faster than Cos-ECMA.
For 64-QAM (Fig. 3 and 4), we have used the channel h
2
= [0.8421+0.3579j 0.0447+0.0494j 0.3602+ 0.1692j]
which was randomly generated and induces phase rotations. For the CMA, Cos-MCMA and Gauss-MCMA a phase
locked loop (PLL) was necessary to correct the phase. The parameters are = 810
-7
for MMA and CMA, (= 710
-7
,
=500/) for l
1
-MCMA, (=510
-7
, =200/) for Cos-MCMA and (= 610
-7
, =5, =0.3) for Gauss-MCMA.
While for the modified ECMA, they were set to (= 410
-7
, =100) for l
1
-ECMA, (= 610
-7
, =200/) for Cos-
ECMA and (= 5 10
-7
, =10, =0.3) for Gauss-ECMA. We can notice that the Gauss-MCMA converges slightly
faster than l
1
-MCMA and Cos-MCMA but Gauss-ECMA is better than l
1
-ECMA and Cos-ECMA respectively. We also
see that the modified ECMA algorithms perform much better than the modified CMA algorithms.
Proc. of SPIE Vol. 7502 75021T-4
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
-25
-20
-15
-10
-5
0
5
Iteration
M
S
E

(
d
B
)
MMA
CMA
ECMA
Cos-MCMA
l1-MCMA
Gauss-MCMA
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
-25
-20
-15
-10
-5
0
5
Iteration
M
S
E
(
d
B
)
Cos-ECMA
ECMA
l1-ECMA
Gauss-ECMA
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
x 10
4
-15
-10
-5
0
5
10
Iteration
M
S
E
(
d
B
)
Gauss-MCMA
MMA
CMA
Cos-MCMA
l1-MCMA
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
x 10
4
-15
-10
-5
0
5
10
Iteration
M
S
E
(
d
B
)
ECMA
Gauss-ECMA
Cos-ECMA
l1-ECMA

In the case of 256- QAM (Fig. 3 and 4), we have chosen the channel with five complex coefficients
h
3
= [-0.2+0.3j -0.5 +0.4j 0.7-0.6j 0.4+0.3j 0.2+0.1j], used by S. Chen in[7] to compare performances of
CMA equalizer, concurrent CMA and decision directed equalizer (CMA+DD) and concurrent CMA and soft decision
directed equalizer (CMA+SDD) respectively. This channel also provokes rotation effects.
In this simulations set, we have intentionally omitted the results of l
1
-MCMA and l
1
-ECMA since the combination of
l
1
-norm CME with CMA or ECMA did not lead to significant improvement as we could expect.
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
x 10
4
-20
-15
-10
-5
0
5
10
15
Iteration
M
S
E
(
d
B
)
Gass-MCMA
CMA
Cos-MCMA
MMA
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
x 10
4
-20
-15
-10
-5
0
5
10
15
Iteration
M
S
E
(
d
B
)
Cos-ECMA
Gauss-ECMA
ECMA
Figure 1 Comparison of convergence speed of
modified CMA for 16-QAM transmitted over h
1
under an SNR=30 dB.
Figure 2 Comparison of convergence speed of
modified ECMA for 16-QAM transmitted over h
1
under an SNR=30 dB.
Figure 3 Comparison of convergence speed of
modified CMA for 64-QAM transmitted over h
2
under an SNR=30 dB.
Figure 4 Comparison of convergence speed of
modified for ECMA 64-QAM transmitted over h
2
under an SNR=30 dB.
Figure 5 Comparison of convergence speed of
modified CMA for 256-QAM transmitted over h
3
under an SNR=50 dB.
Figure 6 Comparison of convergence speed of
modified ECMA for 256-QAM transmitted over h
3
under an SNR=50 dB.
Proc. of SPIE Vol. 7502 75021T-5
Effectively, as previously mensioned, since the l
1
-norm CME function does not respect the uniformity
requirement, it gives poor ameliorations while applied to 256-QAM symbols. The parameters are = 310
-8
for MMA,
= 210
-8
for CMA, (=210
-8
, =200) for Cos-MCMA and (= 310
-8
, =10, =0.2) for Gauss-MCMA. For the
ECMA based algorithms, we have taken (= 110
-8
, =800) for Cos-ECMA and (= 1.8 10
-8
, =200/, =0.2) for
Gauss-ECMA.
It is noticeable that for 256-QAM signals the Gauss-MCMA and Gauss-ECMA algorithms are faster than Cos-
MCMA and Cos-ECMA respectively.
Finally, by looking at the simulation results presented in[7], we can assert that performances obtained for
256-QAM using the hybrid algorithms are much higher than that obtained using CMA+SDD (SNR=90 dB). Indeed, the
latter converges to an MSE of about -7dB after 25000 iterations whereas; the Gauss-MCMA reaches an MSE of -15 dB
after 20000 iterations and the Gauss-ECMA an MSE of -15 dB after 10000 iterations under an SNR of only =50 dB.
5. CONCLUSIONS
Six hybrid adaptive blind equalizers (l
1
-MCMA, Cos-MCMA, Gauss-MCMA, l
1
-ECMA, Cos-ECMA, Gauss-
ECMA) were compared. Among these algorithms, two were recently introduced: l
1
-norm CME combined to CMA and
ECMA and Gaussian CME combined to CMA and ECMA. The superiority of the algorithms based on these new cost
functions, was demonstrated through simulations, in particular, for 16-QAM and 64-QAM symbols.
Finally, if we consider all these algorithms, we can conclude that combining the Gaussian CME with the ECMA
leads to the best performances.
REFERENCES
1. D. N. Godard, Self-recovering equalization and carrier tracking in two-dimensional data communication systems,
IEEE Transactions on Communications, vol. 28, no. 11, pp. 18671875, November 1980.
2. C. R. Johnson, P. Schniter, T. J. Endres, J. D. Behm, D. R. Brown, and R. A. Casas, Blind equalization using the
constant modulus criterion: A review, Proc. of The IEEE, vol. 86, no. 10, pp. 19271950, October 1998.
3. R. Cusani and A. Laurenti, Convergence analysis of the CMA blind equalizer, IEEE Transactions on
Communications, vol. 43, no. 2-3-4, pp. 13041307, February-March-April 1995.
4. J. Yang, J.J. Werner, and G.A. Dumont, The multimodulus blind equalization algorithm, in Proc. Int. Conf. DSP,
Santorini, Greece, July 1997, pp. 127130.
5. T. Thaiupathump and S.A. Kassam, Square contour algorithm: A new algorithm for blind equalization and carrier
phase recovery, in Proc. IEEE Asilomar Conf. on Sig., Syst. and Comp., 2003, pp. 647651.
6. S.A. Sheikh and P. Fan, Improved square contour algorithm for blind equalization of QAM signals, in Proc.
INMIC, December 2006, pp. 116119.
7. S. Chen, Low complexity concurrent constant modulus algorithm and soft decision scheme for blind
equalization, IEE Proc. Vis. Image Signal Process., Vol. 150, No. 5, October 2003.
8. J. A. Garcia and A. R. F. Vidal, Improved blind equalization via adaptive combination of constant modulus
algorithms, in Proc. ICASSP, Toulouse, France, May 2006, vol. 3, pp.756759.
9. X.L. Li and X.D. Zhang, A family of generalized constant modulus algorithms for blind equalization, IEEE
Transactions on Communications, vol. 54, no. 11, pp. 19131917, November 2006.
10. Z. Xiong, Z. Dongfeng, J. Zhigang, W. Anhong, and Z. Liyi, A modified blind equalization algorithm based on
kurtosis of output signal, in Proc. Asia-Pacific Radio Science Conf., August 2004, pp. 228231.
11. S. Barbarossa and A. Scaglione, Blind equalization using cost function matched to the signal constellation, in
Proc. Asilomar, Pacific Grove, CA, November 1997, vol. 1, pp. 550554.
12. L. He, M. G. Amin, and C. Reed, Adaptive equalization techniques for indoor dynamic wireless communication
channels, SPIE proceedings series, vol. 4395, no. 3, pp. 2838, April 2001.
13. L. He, M. G. Amin, C. Reed, and R. C. Malkemes, A hybrid adaptive blind equalization algorithm for QAM
signals in wireless communications, IEEE Trans. on Sig. Proc., vol. 52, no. 7, pp. 20582069, July 2004.
14. K. Banovic, E. Abdel-Raheem, and M.A.S. Khalid, Hybrid methods for blind adaptive equalization: New results
and comparisons, in Proc. 10
th
ISCC, June 2005, pp. 275280.
15. A. Labed, A. Assa-El-Bey, T. Chonavel and A. Belouchrani, New hybrid adaptive blind equalization algorithms
for QAM signals, to appear in Proc. ICASSP, Taipei, Taiwan, April 2009.
Proc. of SPIE Vol. 7502 75021T-6

Das könnte Ihnen auch gefallen