Adaptive Blind Sparse-Channel Equalization

Shafayat Abrara
a Associate Professor

School of Science and Engineering

Habib University, Gulistan-e-Jauhar
Block 18, Karachi 75290, Pakistan
Email: shafayat.abrar@sse.habib.edu.pk

Abstract
In this article, a fractional-norm constrained blind adaptive algorithm is presented for sparse channel equalization. In essence,
the algorithm improves on the minimisation of the constant modulus (CM) criteria by adding a sparsity inducing p -norm penalty.
Simulation results demonstrate that the proposed regularised equalizer exploits the inherent channel sparsity effectively and exhibits
faster convergence compared to its counterparts.
Keywords: Blind equalization; constant modulus algorithm; sparse channel; adaptive filter; channel equalization

1. Introduction 2. Proposed Algorithm

Consider the following instantaneous constant modulus
(CM) cost function subjected to a constraint for sparsity:
2
min J(wk ) = 12 (R wkH xk xkH wk ) , s.t. wk pp c (1)
w

The constant modulus algorithm (CMA) is a widely studied
and, under certain conditions, an admissible solution for adap-
tive blind channel equalization problem [1]. The performance
of the traditional CMA, however, is not satisfactory if the under-
lying channel is sparse. By sparse channel, it is meant that the
number of significant channel coefficients is much less than its
total dimensionality. To make CMA suitable for such channels,
Matrin et al. [2] devised a number of sparse versions of CMA
where they incorporated sparsity under the approximate natu-
ral gradient (ANG) framework and developed proportionate-
type updates (i.e. the changes in the equalizer parameters were
proportional to their magnitudes). These variants happened to
perform better than CMA on sparse channels but exhibited sig-
nificant jitter when forced to converge faster. More recently,
regularised sparse solutions have attracted serious attention in
adaptive signal processing community. Most of these efforts
have been centered around the sparsity promoting minimisa-
tion of 0 and 1 norms of the filter parameters [3, 4, 5, 6]. The
use of fractional-norm regularisation has also evolved as an ad-
missible candidate and has been found sparser than 1 and more
tractable computationally than 0 [7, 8].
In this work, motivated by the idea of norm-constrained opti-
misation [9], we design a sparse CMA by projecting the gradi-
ent vector of the cost onto an p -ball and exploiting the smallest
geometrical angle between the gradient vectors associated with
the cost and the constraint. We discuss the stability of the pro-
posed update, and provide simulation results to demonstrate its
superiority over the CMA and sparse variants of CMA.
Uploaded to sharing knowledge with community.
where wk = [w1,k , w2,k , , wN,k ]T is an N 1 linear finite-
impulse response equalizer vector, xk = [xk , xk1 , , xkN+1 ]T
is an N 1 channel observation vector, R > 0 is a statisti-
cal constant [1], and wk p is a pseudo 0 -norm defined as
wk p = (i=1N
wi,k p )1/p . The objective is to mitigate the
sparse channel interference and recover the transmitted sinal
using solely the equalizer output wkH xk . The a priori informa-
tion about the sparsity of channel is assumed to be available,
and as a result, the equalizer coefficients are also assumed to be
parameterised with sparse representation. By sparse,we mean
that the number of significant parameters in wk , M, is much less
than its total dimensionality (that is M N).
Note that for the feasible set Q = wk pp c, the minimum
value in CM cost is assumed to be attainable as the objective is
continuous (and admissible for equalizable channels), and the
set Q is compact. Also the set Q is nonconvex, so there might be
multiple projection points in general on the geodesic of wk pp =
c. For the purpose of equalization, however, any such minimiser
is acceptable. The Lagrangian for (1) reads
L(wk , k ) = (R wkH xk xkH wk )2 k (wk pp c), (2)
where k is a real-valued Lagrangian multiplier. The gradient-
based update for the minimisation is obtained as
wk+1 = wk L(wk , k )/wk , (3)
where the superscript denotes complex conjugate. Denoting
gk = J(wk )/wk and bk = wk pp /wk as two gradient vec-
August 5, 2017
tors, we get wk+1 = wk (gk k bk ). We have to select k real-valued k , we have a lemma from the theory of holomor-
such that wk+1 p = c, k, i.e., the p -norm of vector wk is con- phic geometry of complex vectors [12, Lemma 2.2.2] (below
served for all values of k. This property yields a flow equation a, b = aH b, and R denotes the real part):
in continuous time-domain: Lemma 1: Let , be a positive Hermitian form on a com-
H plex vector space EC . The underlying real vector space EM
dw(t) pp w(t) pp
=( )
dw(t) dw(t)
= b(t)H =0 (4) inherits a positive definite inner product. Let v1 , v2 EC be
dt w(t) dt dt non-zero vectors. They span complex lines Cvi EM whose
where the superscript H denotes the complex conjugate trans- angle satisfies

(Cv1 , Cv2 ) (v1 , v2 ).

pose operation. The two vectors b(t) and dw(t)/dt are orthog-
onal to each other, and they are normal and tangential to the
surface w(t) pp = c at w(t), respectively. Moreover, for a suffi- The algebraic expression for the smallest angle, independent of
ciently small , we can approximate the time derivative as fol- spanning, in terms of the Hermitian structure is obtained from
lows:
cos ((Cv1 , Cv2 )) = v1 , v2 /(v1 v2 ).
wk+1 wk (g k bk )

dw(t)
lim = k (5)
dt t=k 0
Proof: Replace vi by vi /vi to assume that v1 , v2 are unit vec-
Combining (4) and (5), we obtain an optimal value of k , as
tors. Minimising the angle between them, in Cv1 and Cv2 , is
given by
equivalent to maximising its cosine. The real-valued cosine be-
bkH gk tween unit vectors in these planes equals
bkH (gk k bk ) = 0 k =
bk 2
(6)
cos ((ei1 v1 , ei1 v2 )) = R[ei(1 2 ) v1 , v2 ]/(v1 v2 ),
The vector k bk = bkH gk bk /bk 2 is the component of gk pro- for 1 , 2 M. The maximum value of this expression over all
jected onto bk . The weight update computes the projection of 1 , 2 M equals v1 , v2 as desired.
gk onto bk which is given by gk k bk = (Ibk bkH /bk 2 )gk . So
Owing to Lemma 1, the optimal value of k is obtained as
that the required update is not only against the gradient gk but
also follows the geodesic of wk pp . Refer to Fig. 1 for the geo- bkH gk
k,optimal =
bk 2
metrical illustration for a real-valued two-tap equalizer. More- (7)
over, the term k bk serves as zero-point attraction [10], because
it reduces the distance between wk and the origin when wk is Since k,optimal > 0, therefore the resulting algorithm maximises
small. the p -ball of the equalizer coefficients until it coincides with
the extremum of minimising CM cost.
Remark: Let Mg = span({ gk }) and Mb = span({bk }) be
0.5
kwk kp
two complex-valued N-dimensional vector (sub)spaces. The
orthogonal projection theorem suggests that the minimum an-
gle k (Mg , Mb ) (or the maximum cosine c(Mg , Mb )) between
0.4
b bH
I k k2 gk
kbk k Mg and Mb is defined by (see Fig. 2)
0.3 dw(t)
dt bk c(Mg , Mb ) = sup {gk , bk gk Mg (Mg Mb ) ,
w0,k

(8)
0.2
gk bk Mb (Mg Mb ) , gk 2 = bk 2 = 1}

wk+1
The complex-valued cosine of the angle between two com-
plex vectors v1 and v2 is given generally as cos(C ) =
0.1

cos ((v1 , v2 )) = v1 , v2 /(v1 v2 ) = eiK , where C C is

bk bH
I
called the complex angle, and = cos(C ) = cos(H ) 1.
wk
k
gk
kbk k2

0
0 0.1 0.2 0.3 0.4 0.5 The angles 0 H /2 and < K are known as
w1,k the Hermitian angle, and the Kasners pseudo-angle, respec-
tively, between the vectors v1 and v2 . So the proposed equalizer
Figure 1: Geometrical interpretation of constrained optimization. p -SCMA exploits the Hermitian angle in the update process
which is not only the smallest angle but is insensitive to the
Note that k is (a sort of) complex-valued cosine [11] of the multiplication of the vectors by any complex scalars, and it is
angle between bk and gk . From the problem definition, how- desirable in the context of CM equalization where the equalizer
ever, the value of k is required to be real-valued. To obtain a update is required to be insensitive to multiplication by complex
2
 coefficients taking large values, and to drive the unnecessary co-
efficients (which fall below certain threshold) to zero. Elegant
closed-form solutions for 1/2 or 2/3 regularization have been
developed by Zhang and Ye [15].
Consider the following lemma:

Lemma 2 [16]: Let f denote the objective function minh f =

(h w)2 + h p , 0 < p < 1, > 0. It has a unique minimum h
for w (p, ), where

(1 p) 2p 2p .
2 p p1 1
(p, ) = (12)
2

Figure 2: The geometry of the planes spanned by bk and gk , and the minimum
Next we discuss closed form solutions for p = 1/2 and p =
angle between them. 2/3.

4.1. Closed-form solution for 1/2 regularization

exponentials which represents phase/frequency offset errors.
So we find an adaptive solution to the sparse equalization Once we have wk = wRk + iwkI from the update (9), we need to
problem (1) which minimises the CM criterion, moves along regularize wRk and wkI separately as formulated below:
the geodesic of the constraint surface, and exploits the smallest
hRk = arg min {(h wRk ) + R h1/2 }
angle between the gradient vectors (possibly spanning complex 2 1/2
(13a)
lines), as given by: h

hkI = arg min {(h wkI ) + I h1/2 }

2 1/2
bkH gk
(13b)
h
wk+1 = wk gk
bk 2
bk , (9)
where h MN is an auxiliary variable. The closed form solu-
tion to the above optimization problems are given as (below L
where gk and bk are specified as: denotes either R or I):
gk = (wkH xk xkH wk R) xk xkH wk ,
wk,i (1 + cos ( 2 ))) , wk,i
(10a)

2

2 L
23 (wk,i
L L
> 454 L3
= 2 wL (1 + cos ( 2 2 (wL ))) , wL < 3 54 3
T 3 3
[ ] .
p w1,k wN,k L 2
hk,i
bk =
2 w1,k 2p
, ,
wN,k 2p
(10b)


k,i k,i k,i L

3 3 3 4
0, otherwise
(14a)
) = arccos ( 3 wk,i ) , i = 1, 2, , N

3. Steady-state Stability L 3L L 23
(wk,i 8
(14b)
The update (9) is stable. Denoting = bkH gk /bk 2 , we
obtain energy of the update (9) as given by The regularized equalizer second-stage output is thus obtained
as sk = hkH xk , where hk = hRk + ihkI .
wk+1 2 = wk 2 + 2 gk 2 + 2 2 bk 2 Proof: Consider a scalar 1/2 optimization problem as follows:
+ 2R[wkH (bk gk ) 2 bk 2 ]
= wk 2 + 2 gk 2 2 2 bk 2 + 2R[wkH (bk gk )]. min {(h w)2 + h1/2 } (15)
h
(11)
Taking derivative with respect to h, and substituting to zero, we
Owing to Bussgang theorem [13], we have EwkH (bk gk ) = 0. get
h w + sign(h) = 0
Further exploiting the independency between xk and wk (inde-
(16)
pendence theorem [14]), we obtain E(gk 2 2 bk 2 ) = 0, 4 h
yielding Ewk+1 2 = Ewk 2 which implies that there is no
Substituting h = z2 , we obtain
growth in the energy of wk and thus proves the stability.

z3 sign(h) wz + sign(h) = 0 (17)
4. Explicit Regularization 4

We may add a second stage to perform explicit regulariza- Note that we require h < 0 (> 0) for w < 0 (> 0). Let h < 0, it
tion. The aim of this stage is to (further) prune the equalizer gives sign(h) = 1, and w = w, and we obtain
coefficients as obtained in the first stage by introducing explicit
z3 wz +

1/2 or 2/3 regularization as a brute-force method to prevent the =0 (18)
4
3
Same is the result when w > 0; so we proceed with (18). In (23), we obtain the desired value of h as follows:

order to have three real-valued roots, we need to ensure that the


Cardans discriminant1 is positive, this gives


2
w(1 + cos( 32 23 C))



3
w3 2 h= w

3 > 0 w > 2/3 .
3

3 > 54 2/3 (24)

for


4
33 4 4



to Lemma 2, however, we can find solution only if w > 0, otherwise

Due
> 43 2/3 . Further substituting z = y w/3, we get

3
54 2/3
4

4.2. Closed-form solution for 2/3 regularization
3
y 3y 2q = 0 (19) Consider a scalar 2/3 optimization problem as follows:

where q = (/8) (3/w) . Eq. (19) may be solved by con-

min {(h w)2 + h2/3 }
3/2
(25)
sidering the triangle in Fig. 3, where y1 represents one of the h
three roots of y. We outline the proof as conceived by Mitchell
[17] in the interest of readers. The solution of above is rigourously presented in [18, 15].
Here, we sketch a similar proof but in simpler steps. Taking
derivative with respect to h, and equating to zero, we get

C sign(h)
=0
3 h1/3
hw+ (26)
1 y1



Substituting h = z3 , we get
B A
y21 1
z4 sign(h) wz + sign(h) = 0 (27)
Figure 3: Triangular interpretation for solving cubic polynomial. 3
We exploit Ferraris idea to introduce a parameter t in (27)

(z2 + t)2 = (2t)z2 + wz + (t2 )

Using the cosine law, we obtain (28)
3
1 + y21 (y21 1)2 y1 (3 y21 )
q = cos(C) = = (20) such that the right hand side becomes a monic quadratic poly-
2y1 2 nomial in z, i.e., it has a real root with multiplicity 2, or equiv-
which justifies the claim that y1 is one of the roots of (19). Sim- alently the discriminant is zero, which gives

w2
ilarly we obtain
w2 4(2t)(t2 ) = 0 (t2 ) =

(29)
y2 2 y1 3 3 8t
cos(B) = 1 , and cos(A) = . (21)
2 2
Substituting the value of t2 3
in (28), we obtain:
From (21), we obtain B = 2A. Since A + B + C = , therefore
A = 13 13 C. Now employing the sine law, we obtain w
2
(z2 + t)2 = ( 2tz + ) (30)
sin(A) sin(B) sin(C) 8t
= = 2 (22)
1 y1 y1 1 which gives
w
which implies y1 = sin(B)/ sin(A), and gives z2 + t = ( 2tz + ) (31)

8t
w sin(A) Above the two roots associated with the negative sign are of no
cos( )
4w 4w C
z1 = = cos(A) = use, as they lead to an undesirable result h < 0 (h > 0) for w > 0
3 sin(B) 3 3 3 3
(23) (w < 0). Solving, however, for positive sign, we obtain
The other two (z2 and z3 ) roots may be found by adding 2
t w

3
in the argument of cos(); however, by inspecting these roots,
z=
t
(32)
we find that the root specified in (23) is the desired root. From 2 8t 2

where the root of our interest is the one with plus sign as fol-
lows:
t w
+
1
For a cubic polynomial z3 + cz + d = 0, the Cardans discriminant is defined t
z= (33)
as = 4c3 27d2 . 2 8t 2
4
The last task is to find out the value of t from the cubic expres- 600
sion (29); we specify it again
500

t3 t w2 = 0
1 1
(34) 400
3 8

Frequency
Evaluating the Cardans discriminant, we obtain 300

w2
2
3
= 4 ( ) 27 ( )
200

3 8
100
= 3 w4
4 27

2
27 64 0
0 5 10 15 20 25 30
3 (w >
4 3 27 16 3 4 (35)
< 33 due to Lemma 2) EVS
27 64 81 3
11 3 Figure 4: EVS histogram obtained from ten thousand randomly generated
< sparse channels. The mean and standard deviation of the EVS are 6.57 and
108 3.88, respectively.
< 0 ( > 0)

which implies that there is only one real-valued root of (34). standard deviation 1.8. The histogram of the eigen-value spread
Since Mitchells triangle method requires > 0, therefore it is shown in Fig. 4.
cannot help us find that root. Using Holmes formula [19], how- We consider a two-modulus 8-ary amplitude phase shift key-
ever, we immediately obtain the required real-valued root of ing (APSK) signaling at 30 dB signal-to-noise-ratio (SNR),
(34) in a closed-form as follows: and use inter-symbol interference (ISI) metric for performance
2
comparison averaged over 1000 channels randomly obtained
cosh ( cosh1 ( w2 3/2 )) .
1 27 using sparse channel.m. Equalizers are initialised such that
t= (36)
3 3 16 the central tap is set to 1 + i, and the rest
of the taps are set to
(1 + i)/N where N = 120 taps, and i = 1. Results for ISI
Owing to the relation h = z3 , we obtain h as follows:
traces are summarised in Fig. 5, where the step-sizes appear in


w
the legend. Note that the proposed equalizer RSCMA (with ex-

3

sign(w) (
t
)
t


plicit 1/2 closed form regularization) outperforms CMA and

+
2 8t 2
h=
its sparse variants ANG-CMA and SCMA(1/2) in terms of
w (37)

2 4
> 3

for 3 steady-state performance.


3




0, otherwise. 6. Conclusions
where t is as specified in (36).
An p -regularised sparse CMA equalizer, RSCMA, is ob-
tained and demonstrated for blind channel equalization of com-
5. Simulation Results plex valued signals by incorporating the so-called zeroth-norm
constraint in the traditional CM cost function. Simulation re-
sults have shown that RSCMA exhibited faster convergence rate
We compare the proposed two-stage regularized sparse CMA
on sparse channels as compared to the traditional CMA equal-
(RSCMA) equalizer with the traditional CMA [1] and two
izer and its sparse variants. Finally, our equalizer proved to be
sparse variants of CMA like ANG-CMA [2] and SCMA(p)
a substitute for the traditionally used ones.
[20]. The baseband model of the sparse channels have 100 taps
with five non-zero taps obtained using the following program:
References
h=zeros(1,100); i0=randi([1,10],1,1);
[1] R. Johnson, P. Schniter, T.J. Endres, James D. Behm, D.R. Brown, and
i1=randi([20,30],1,1); i2=randi([40,50],1,1); R.A. Casas. Blind equalization using the constant modulus criterion: A
i3=randi([70,80],1,1); i4=randi([90,100],1,1); review. Proceedings of the IEEE, 86(10):19271950, 1998.
h(i0)=0.1*(2*rand-1)+0.1*(2*rand-1)*1i; [2] R.K. Martin, W.A. Sethares, R.C. Williamson, and R.C. Johnson Jr. Ex-
h(i1)=1+(2*rand-1)*1i; ploiting sparsity in adaptive filters. Signal Processing, IEEE Trans.,
50(8):18831894, 2002.
h(i2)=0.5*(2*rand-1)+0.2*(2*rand-1)*1i; [3] Y. Gu, J. Jin, and S. Mei. 0 -norm constraint LMS algorithm for sparse
h(i3)=0.2*(2*rand-1)+0.2*(2*rand-1)*1i; system identification. Signal Processing Letters, IEEE, 16(9):774777,
h(i4)=0.1*(2*rand-1)+0.1*(2*rand-1)*1i; 2009.
[4] K. Shi and P. Shi. Adaptive sparse volterra system identification with
h(tt,:)=h(tt,:)/norm(h(tt,:));
0 -norm penalty. Signal Processing, 91(10):24322436, 2011.
[5] D. Angelosante, J.A. Bazerque, and G.B. Giannakis. Online adaptive
The average eigenvalue spread of the channels obtained from estimation of sparse signals: Where RLS meets the 1 -norm. Signal Pro-
the above program (sparse channel.m) is nearly 4.8 with cessing, IEEE Trans., 58(7):34363447, 2010.

5
 [6] K. Shi and P. Shi. Convergence analysis of sparse lms algorithms EVS(2,4], Nr=100, N=200, SNR=30dB, 8APSK
10
with 1 -norm penalty based on white input signal. Signal Processing, CMA: = 0.0001
90(12):32893293, 2010. 12 ANGCMA: = 0.01
[7] Z. Xu, X. Chang, F. Xu, and H. Zhang. 1/2 -regularization: A threshold- SCMA: = 0.0001
14 RSCMA: =0.00005
ing representation theory and a fast solver. Neural Networks and Learning
Systems, IEEE Trans., 23(7):10131027, 2012. 16
[8] F.Y. Wu and F. Tong. Gradient optimization p-norm-like constraint LMS

ISI [dB]
algorithm for sparse system estimation. Signal Processing, 93(4):967 18
971, 2013.
20
[9] S.C. Douglas, S. Amari, and S.-Y. Kung. On gradient adaptation with
unit-norm constraints. IEEE Trans. Signal Processing, 48(6):18431847, 22
2000.
[10] J. Jin, Y. Gu, and S. Mei. A stochastic gradient approach on compressive 24
sensing signal reconstruction based on adaptive filtering framework. Se-
26
lected Topics in Signal Processing, IEEE Journal of, 4(2):409420, 2010. 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Index x 10
4
[11] K. Scharnhorst. Angles in complex vector spaces. Acta Applicandae
Mathematica, 69(1):95103, 2001. (a)
[12] W.M. Goldman. Complex hyperbolic geometry. Oxford University Press, EVS(4,6], Nr=100, N=200, SNR=30dB, 8APSK
1999. 10
[13] Sandro Bellini. Bussgang techniques for blind deconvolution and equal- CMA: = 0.0001
12 ANGCMA: = 0.01
ization. Blind deconvolution, pages 859, 1994. SCMA: = 0.0001
[14] JE Mazo. Analysis of decision-directed equalizer convergence. The Bell 14 RSCMA: =0.00005
System Technical Journal, 59(10):18571876, 1980.
[15] Y. Zhang and W. Ye. 2/3 regularization: Convergence of iterative thresh- 16

ISI [dB]
olding algorithm. Journal of Visual Communication and Image Represen-
18
tation, 33:350357, 2015.
[16] C. Miao and H. Yu. A general-thresholding solution for p (0 < 20
p < 1) regularized ct reconstruction. Image Processing, IEEE Trans.,
24(12):54555468, 2015. 22

[17] D.W. Mitchell. 91.60 solving cubics by solving triangles. The Mathemat- 24
ical Gazette, 91(522):514516, 2007.
[18] W. Cao, J. Sun, and Z. Xu. Fast image deconvolution using closed-form 26
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
thresholding formulas of regularization. Journal of Visual Communica- Index x 10
4

tion and Image Representation, 24(1):3141, 2013.

[19] G.C. Holmes. 86.70 the use of hyperbolic cosines in solving cubic poly- (b)
nomials. The Mathematical Gazette, 86(507):473477, 2002. EVS(6,8], Nr=100, N=200, SNR=30dB, 8APSK
10
[20] S.S. Khalid and S. Abrar. Blind adaptive algorithm for sparse chan-
nel equalisation using projections onto p -ball. Electronics Letters, 12
51(18):14221424, 2015.
14

16
ISI [dB]

18

20

22 CMA: = 0.0001
ANGCMA: = 0.01
24 SCMA: = 0.0001
RSCMA: =0.00005
26
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Index x 10
4

(c)
EVS(8,10], Nr=100, N=200, SNR=30dB, 8APSK
10

12

14

16
ISI [dB]

18

20

22 CMA: = 0.0001
ANGCMA: = 0.01
24 SCMA: = 0.0001
RSCMA: =0.00005
26
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Index x 10
4

(d)

Figure 5: Comparison of residual ISI plots.