Sunteți pe pagina 1din 5

IEEE SIGNAL PROCESSING LETTERS, VOL. 24, NO.

2, FEBRUARY 2017 161

Second-Order Statistics Driven LMS Blind


Fractionally Spaced Channel Equalization
Gaetano Scarano, Andrea Petroni, Mauro Biagi, and Roberto Cusani

AbstractA novel cooperative least mean square adaptation rule


is proposed with application to fractionally spaced blind channel
equalization schemes used in quadrature amplitude modulation
data links. A novel cost function based on second-order statistics
only of the received signal is defined and jointly used with an-
other chosen among classical blind cost functions such as decision Fig. 1. IIR-MMSE-FSE.
directed and other Bussgang-type ones. Numerical results show
that severe misconvergence observed in classical blind equaliza- is sampled at rate P Ts1
before entering the equalizer, P being
tion algorithms is drastically mitigated using the here described the fractional sampling factor. Naturally, the filter output rate is
technique. equal to the symbol rate Ts1 and filter coefficients are updated
at the same rate.
Index TermsBussgang equalization, cross relations (CR),
Referring to the common case of P = 2, a novel blind adap-
decision directed equalization, fractionally spaced equalization,
least mean square (LMS). tive cooperative scheme is here described, in which the FSE
coefficients are updated at rate 2Ts1 , i.e., twice the rate of
I. INTRODUCTION classical implementations. More in detail, two different cost
functions (CFs) are employed and jointly minimized. The first
NTERSYMBOL interference (ISI) is the main impairment
I in quadrature amplitude modulation (QAM) data links and
its reduction is usually performed by digital filtering aimed at
CF can be anyone of those employed in present blind receivers
and works in correspondence of even time instants. The second
CF works instead in correspondence of odd instants and here
obtaining the channel equalization condition, in which the appears for the first time; it is derived from a here rediscov-
symbol detection error probability is kept below the desired ered property of the IIR minimum mean square error (MMSE)
level. The filter is adaptively implemented and trained by us- FSE [7]. Joint minimization of these two CFs is feasible since
ing a finite sequence of transmitted symbols known at receiver; both are minimized by the same filter even though they refer
moreover, to mitigate noise amplification effects as well as to to different instants. We show that the proposed adaptive fil-
reduce timing sensitivity, filtering is carried out after sampling tering scheme drastically mitigates the severe misconvergence
the received signal at rates greater than the nominal one, typ- phenomenon that affects certain blind equalization algorithms,
ically twice the transmitted symbol rate. This is referred to as especially for large symbol constellation order and frequency
fractionally spaced channel equalization [1], and the filter is selective channels.
so-called fractionally spaced equalizer (FSE).
During frame synchronization acquisition training symbols
cannot be used and, hence, self-trained or blind channel II. FRACTIONALLY SPACED EQUALIZATION
equalization techniques are often employed, among which the For P = 2, the discrete-time models of fractionally sampled
most popular is the constant modulus algorithm (CMA) [2][4], received signals in a QAM data link is
and its variations like that recently described in [5]. This class of
blind channel equalization techniques exploits the information 
+

contained in the higher order statistics of the non-Gaussian re- x[n] = a[l]g[n 2l] + w[n] ,
ceived signal. On the other hand, a certain number of other blind l=
channel estimation techniques exploits the remarkable property
of ciclostationarity of fractionally sampled data signals, which where g[n] = (gT hc )(nTs /2) takes into account the channel
allows, in certain conditions, for blind channel estimation car- hc (t) and the transmitting shaping filter gT (t), w[n] are samples
ried out using second-order statistics (SOS) only (see for in- of white additive noise, and Ts1 is the rate at which the QAM
stance [6] and its references). symbols a[l] are transmitted. The FSE operation is shown in
Very often, the FSE is adaptively implemented using a simple Fig. 1, where G(ej ) is the Fourier transform of g[n].
least mean square (LMS) rule [1][3], where the received signal Due to the excess bandwidth of the shaping filter gT (t),
e.g., a root-raised cosine (RRC) filter with roll-off > 0, the
Manuscript received October 7, 2016; revised November 11, 2016; accepted fractionally sampled data signal s[n] results in a cyclostationary
November 27, 2016. Date of publication December 1, 2016; date of current process, an useful property exploited in channel estimation as
version January 12, 2017. The associate editor coordinating the review of this well as in determining optimal FSEs [6][10]. In particular, the
manuscript and approving it for publication was Dr. Arash Amini. IIR-MMSE-FSE frequency response is [9], [10]:
The authors are with the DIET, Universita La Sapienza di Roma, Roma
00184, Italy (e-mail: gaetano.scarano@uniroma1.it; andrea.petroni@uniroma1.
2G(ej )
it; mauro.biagi@uniroma1.it; roberto.cusani@uniroma1.it). FMMSE
(IIR)
(ej ) = (1)
Digital Object Identifier 10.1109/LSP.2016.2635034 |G(ej )| + |G(ej +j )|2 + 2Pw Pa1
2

1070-9908 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
162 IEEE SIGNAL PROCESSING LETTERS, VOL. 24, NO. 2, FEBRUARY 2017

IV. FIR-FSE-CR
Let us now consider a FIR-FSE f (FIR) [k], and its -modulated
version f(FIR) [k] = f (FIR) [k](1)k , both defined on a finite support
Sf = {k1 , . . . , kL 1 } with L large enough to still consider valid
the IIR-MMSE-FSE-CR property (3)

Fig. 2. IIR-MMSE-FSE-CR. f(FIR) [k] Rx [m k; k mod 2]  Pw f(FIR) [m]. (4)
k S f
where Pw is the power of the stationary white additive noise
w[n], statistically independent of the stationary white symbols
Since, the cardinality of Sf is finite, we can write (4) only for
a[l] of power Pa , and () denotes complex conjugation. m Sf , so that, after collecting the FSE coefficients f(FIR) [k] in
the vector f(FIR) , we have
III. IIR-MMSE-FSE CROSS RELATIONS (CR)
k Rx f(FIR)  Pw f(FIR) = (f(FIR) )H Rx f(FIR)  Pw Ef (FIR) (5)
Let us denote by fMMSE( IIR)
, [k] = fMMSE [k](1) the sequence
(IIR)

j +j
having Fourier transform FMMSE (e (IIR)
), which we will refer to where (Rx )m ,k = Rx [m
as the coefficients of the -modulated IIR-MMSE-FSE. k; k mod 2], for m Sf , k Sf , and
Interestingly enough, fMMSE ( IIR) Ef (FIR) = (f(FIR) )H f(FIR) = |f (FIR) [k]|2 is the energy of the
, [k] satisfies the following prop- k S f
erty, presented for the ideal noise-free case in [7] sequence f (FIR) [k]. Stemming from (5), we restate Property 1
as follows, thus definitively characterizing any FIR-FSE suffi-

+
ciently close to the IIR-MMSE-FSE.
fMMSE
( IIR)
, [k] (x[2n + 1 + k] w[2n + 1 + k]) = 0. (2)
Property 2: (FIR-FSE-CR) The coefficients of any FIR-FSE
k =
defined on a sufficiently large support Sf form a quasi-
As shown in Fig. 2, the proof of (2) considers that expanded-by-2 eigenvector of the autocorrelation matrix Rx , precisely that
sequences passing through filters having a periodic modulo corresponding to the minimum quasi-eigenvalue, this latter
frequency response remain expanded-by-2 sequences. given by the noise power Pw .
It is worth noting that (2) constitutes the IIR-MMSE-FSE It is worth noting that (5) obtains the FIR-FSE coefficients
counterpart of the well-known CR property, which instead re- exclusively using SOS of the received signal x[n]; for a long
gards the channel coefficients g[n] [11]. For this reason, we will enough support Sf , a unique solution would ideally be obtained
refer to (2) as IIR-MMSE-FSE-CR. It is also worth noting that only when: i) the channel is a purely FIR one, of length Lg +
(2) is entirely due to the excess bandwidth of the transmitting 1; ii) the support is connected with L = Lg : Sf = {0, . . . ,
filter, measured by the roll-off factor in the case of RRC filter. Lg 1}
Since the noise is white, expectations of both sides of (2) In practice, a sufficiently accurate FIR-FSE cannot be ob-
multiplied by x[2n + 1 + m] yields tained from the eigenvector associated to the minimum eigen-
value of Rx , especially when this matrix is estimated from

+
observed data and the excess bandwidth of the transmit filter
fMMSE
( IIR)
, [k] Rx [m k; k mod 2] = Pw fMMSE, [m] (3)
( IIR)
is relatively small. For supports Sf longer than the channel,
k =
one shall resort to techniques that aim at identifying the sub-
where Rx [mk; k mod 2] = E{x[2n+1 + m] x[2n+1+k]} space associated to the smallest eigenvalues of Rx , where, in
denotes the biargumental autocorrelation function of the principle, the vector f(FIR) lies; unfortunately, since eigenvectors
cyclostationary random series x[n]. The following property, associated with nearby eigenvalues are wobbly [12][13], the
which definitively characterizes the IIR-MMSE-FSE, stems subspace spanned by the subset of eigenvectors associated to the
from (3). smallest eigenvalues merges with its orthogonal complementary
Property 1: (IIR-MMSE-FSE-CR) The coefficients subspace and, hence, even those techniques do not obtain a suf-
fMMSE
( IIR)
, [k] of the -modulated IIR-MMSE-FSE form an
ficiently accurate estimation [14].
eigensequence of the kernel Rx [mk; k mod 2], precisely that
corresponding to the minimum eigenvalue, this latter given by V. FIR-FSE-CR DRIVEN LMS ADAPTATION
the noise power Pw .
Proof: The noise power Pw is the minimum eigenvalue Here, we show how the FIR-FSE-CR property can be usefully
since. employed to drive the convergence of blind FIR-FSE synthesis
1) Due to noise w[n] and symbols a[l] uncorrelation, after techniques adaptively implemented using LMS. In particular,
denoting by [] the unit sample sequence, it results bearing in mind Figs. 1 and 2, we present a novel, coopera-
Rx [mk; k mod 2] = Rs [mk; k mod 2] + Pw [m tive adaptation rule that jointly minimizes two different CFs,
k], which, after substitution in (3), shows that the coef- one concerning with the error samples e[n/2] available on even
ficients fMMSE
( IIR)
, [k] form the eigensequence of the kernel
time indexes n, the other concerning with the error samples
Rs [mk; k mod 2] corresponding to the null eigenvalue. o[(n + 1)/2] available on odd time indexes. Specifically, col-
2) The kernel Rs [mk; k mod 2] is non-negative def- lecting the FIR-FSE coefficients in the vector f (FIR) , the even and
inite, i.e., for every sequence a[k] it results in the odd CFs are, respectively, defined as
+ +  
a[m]a[k] Rs [mk; k mod 2] 0, so Je (f (FIR) ) = E |e[n/2]|2 (even n) (6)
m = k =
that all its eigenvalues are non-negative.  
 Jo (f (FIR) ) = E |o[(n + 1)/2]|2 (odd n) (7)
SCARANO et al.: SECOND-ORDER STATISTICS DRIVEN LMS BLIND FRACTIONALLY SPACED CHANNEL EQUALIZATION 163

TABLE I Table I is obtained from the MMSE one in the limit 2 +,


DECISION DIRECTED AND MMSE ZMNL DEFINED ON THE CONSTELLATION
CM = {c0 , . . . , cM 1 }
and thus it can be correctly considered as a proper MMSE
ZMNL characterizing a Bussgang blind channel equalization
scheme. In this respect, it is also worth noting that the IDD
ZMNL is completely different of ZMNLs that characterize
the so-called Multiple Modulus based blind equalization
algorithms such as those in [18][20].

B. Cooperative Adaptation
Before proceeding, let us first collect the received signal sam-
ples in the following pair of vectors:
xe [l] = [ x[2lk] ]T , k Sf (11)
xo, [l] = [ (1)k x[2l + 1 + k] ]T k Sf . (12)
Hence, following the indication of [15], the complex gradients
where the even and the odd errors are (see also Figs. 1 and 2) of the two CFs in (6) and (7) are readily evaluated

e [n/2] = f (FIR) [k] x[n k] a[l] (even n) (8) f (FIR) Je (f (FIR) ) = E{e[n/2] xe [n/2]} (even n)
k S f (13)
 f (FIR) Jo (f (FIR) ) = E{o[(n + 1)/2] xo, [(n + 1)/2]} (odd n).
o[(n + 1)/2] = f (FIR) [k] (1)k x[n + k] (odd n). (9)
k S f Considering sample estimates of the complex gradients (13), we
end up with the following novel incremental LMS adaptation
The CF Jo () descends from Property 2 and is minimized by rule, whose adaptive implementation is shown in Fig. 3:
FIR-FSEs sufficiently close to the IIR-MMSE-FSE; on the other        
hand, the CF Je () is minimized by the unique FIR-MMSE-FSE fo(FIR) n2 +1 = fe(FIR) n2 + e e n2 xe n2 (even n) (14)
so that we can argue that jointly minimization of (6) and (7) will
 n+1   n+1   n+1   n+1 
lead to a FIR-FSE close enough to both the IIR-MMSE and fe(FIR) = fo(FIR) +o o
2 2 2 xo, 2 (odd n). (15)
FIR-MMSE ones. It is also worth noting that the CFs (6) and (7)
consider errors computed at different times; precisely, samples In (14) and (15), e and o are the learning factors and the LMS
e[n/2] in Je () are available for even values of n while samples adaptation starts after suitably choosing fe(FIR) [0].
o[(n + 1)/2] in Jo () are available for odd values of n. Remark: Attention must be paid to not confuse the incremen-
tal adaption rules (14) and (15) with the adaptive filtering scheme
A. Blind Equalization illustrated in [25] or the incremental cooperative-strategy pre-
sented in [26]; both these schemes rely on a single input that
The error computation (8) requires the knowledge of the trans- drives two different filters, which cooperate to minimize a sin-
mitted symbols a[l], available in trained equalization. Blind er- gle, common, CF. What makes unique our incremental adaptive
ror computation is carried out using a suitably defined symbol filtering scheme is that it exploits the fractionally spaced nature
estimate a[n] in lieu of a[n], usually obtained through a zero of a single filter, whose input is cyclically alternated between
memory nonlinearity (ZMNL) () of the FSE output two different sets of samples, namely xe [ n2 ] and xo, [ n +1
2 ], so
 that two CFs, each capturing a different aspect of blind equaliza-
(filtering) a [n/2] = f (FIR) [k] x[n k](even n) (10) tion, can be simultaneously minimized. We remark again that at
k S f
odd time indexes it is carried out the minimization of the novel
(estimation) a [n/2] = (a [n/2]) (even n). CF Jo (), here derived from Property 2 and exclusively based
on SOS of the received signal.
1
The class known as Bussgang equalizers collect blind equaliz-
ers obtained by different optimization criteria, each correspond- VI. NUMERICAL RESULTS
ing to a different ZMNL () [16]. Table I reports two remark-
able incoherent ZMNLs.2 In particular, the incoherent MMSE We performed a numerical study of the convergence success
ZMNL has been presented in [17], and its parameter 2 mea- of the here described novel FIR-FSE-CR cooperative LMS rule
sures the amount of residual ISI plus additive noise. It is worth applied to blind equalization of the 128QAM microwave urban
noting that the incoherent decision directed (IDD) ZMNL of data link channel with impulse response h[n] found in [20];
in our numerical simulations the channel impulse response was
1 The name is due to the circumstance that convergence is reached when
set to hC [n] = ej h[n], being a uniformly distributed random
a[n/2] becomes a Bussgang process satisfying the cross-correlation invari-
phase. The equalizer support Sf is a connected one with L = 33
ance E{a[n/2] x e } E{a[n/2] x e }; generalization of Bussgang invariance and the LMS adaptation has been initialized with a square-root
are found in [21] and [22]. Bussgang blind equalizers have been successfully raised-cosine filter with roll-off = 0.35; the signal-to-noise
employed also in image deblurring [23][24]. (SNR) is measured at the channel output, i.e., SNR = Ps /Pw ,
2 Coherent ZMNLs, as that derived in [16], are able to recover also a constant
being Ps the power of s[n].
channel phase factor and are not employed if phase recovery is concurrently
carried out by other dedicated tasks. On the other hand, incoherent ZMNLs Fig. 4 reports the percentage of good MonteCarlo runs
yield equalized symbols within a residual phase factor, without destructively (MCR) for which the symbol error rate (SER) achieved by
interfering with any concurrent phase recovery tasks. the IDD blind equalizers, operated with learning factors e =
164 IEEE SIGNAL PROCESSING LETTERS, VOL. 24, NO. 2, FEBRUARY 2017

Fig. 3. Incremental adaption rule: x[n] = [x[n K ], . . . , x[n + K ]]T , being K = max{|k 1 |, . . . |k L |} the maximum absolute index in the support S f .

Fig. 4. Good runs percentage observed in 1000 MCR (SNR = 24 dB/bit). Fig. 6. fMSE learning curves averaged over 1000 MCR (SNR = 24 dB/bit).

reports learning curves of the following quadratic distance:


  1/2
 2
 2
fMSE = |f [k] fMMSE [k]|
(FIR) (IIR)
|fMMSE [k]|
(IIR)

k S f k S f

Fig. 6 clearly shows the improvement in speed of convergence


achieved by CR driven LMS blind channel equalization. It also
Fig. 5. MSE learning curves averaged over 1000 MCR (SNR = 24 dB/bit). gives evidence that the CF Jo () helps the other CF Je () to
adaptively drive the equalizer toward the IIR-MMSE-FSE; we
can argue that this is the main reason of the drastic misconver-
o = 102 , is less than 102 ; ideal phase compensation was gence reduction achieved by the here described CR driven LMS
applied before evaluating SER on frames of 512 symbols as blind channel equalization technique.
long as the adaptation proceeds, the phase was redrawn, and
the adaptation restarted in each and every of 1000 MCR. VII. CONCLUSION
Convergence failure was declared if SER still remains above A novel cooperative LMS adaptation scheme suitable for
102 after 100 frames, and this never occurred in our novel IDD fractionally spaced blind channel equalization has been pro-
FIR-FSE-CR, while 115 convergence failures occurred using posed and computer simulation results have been presented to
the classical IDD FIR-FSE, this latter operated using the usual assess its merits. Compared to the classical LMS adaptation, a
fractionally spaced LMS rule with e = 102 supplementary step has been added, in which it is exploited a
 n +1  n  n  n  remarkable property of the FIR-FSE equalizer. This property,
fe(FIR) 2 = fe(FIR) 2 + e e 2 xe 2 (even n). here called FIR-FSE-CR, constitutes the equalizer counterpart
of the well-known CR, a second-order statistical property used
For comparison purposes, Fig. 4 reports also the results obtained for blind channel estimation [6].
using a CMA-FSE [3], which shall be operated with smaller The significant reduction of convergence failures obtained
learning factors, here set to e = 3.16 104 and o = 3 by driving LMS adaptation using FIR-FSE-CR makes the pro-
103 , to guarantee convergence. It is worth noting that CMA posed solution a serious candidate for cold start equalization,
FIR-FSE-CR is not able to outperform our novel IDD FIR-FSE- especially when phase recovery is concurrently carried out and
CR, since this latter can be operated with a learning factor much the equalization task is limited to recover transmitted symbols
higher, namely 102 versus 3.16104 . Moreover, we observe within a phase rotation. This incoherent scenario is the most
that IDD FIR-FSE-CR saves a time equivalent to about 8200 interesting one, even because coherent blind equalization tech-
samples since it is able to obtain 97.5% of good runs in about niques, which attempt to jointly estimate both channel phase and
74 512 = 38 400 samples, while CMA FIR-FSE-CR obtains transmitted symbols, in general presents a convergence failure
the same result in 90 512 = 46 080 samples. rate too high for practical applications.
To discuss the speed of convergence we present learning Finally, we outline that the novel CR based cooperative LMS
curves of the mean square error (MSE) in Fig. 5, while Fig. 6 rule can be jointly used with any blind equalization CF.
SCARANO et al.: SECOND-ORDER STATISTICS DRIVEN LMS BLIND FRACTIONALLY SPACED CHANNEL EQUALIZATION 165

REFERENCES [14] T. J. Endres, S. D. Halford, C. R. Johnson, Jr, and G. B. Giannakis,


Simulated comparison of blind equalization algorithms for cold start-
[1] G. Ungerboeck, Fractional tap-spacing equalizer and consequences for up applications, Int. J. Adaptive Control Signal Process., vol. 12,
clock recovery in data modems, IEEE Trans. Commun., vol. 24, no. 8, pp. 283301, 1998.
pp. 856864, Aug. 1976. [15] B. Widrow, J. McCool, and M. Ball, The complex LMS algorithm, Proc.
[2] D. N. Godard, Self-recovering equalization and carrier tracking in IEEE, vol. 63, no. 4, pp. 719720, Apr. 1975.
two-dimensional data communication systems, IEEE Trans. Commun., [16] S. Bellini, Bussgang techniques for blind deconvolution and equaliza-
vol. 28, no. 11, pp. 18671875, Nov. 1980. tion, in Blind Deconvolution, S. Haykin, Ed. Englewood Cliffs, NJ, USA:
[3] C. R. Johnson et al., The core of FSE-CMA behavior theory, in Unsu- Prentice Hall, 1994, pp. 859.
pervised Adaptive Filtering, vol. 2, S. Haykin, Ed. New York, NY, USA: [17] G. Panci, S. Colonnese, P. Campisi, and G. Scarano, Fractionally spaced
Wiley, 2000, pp. 13112. Bussgang equalization for correlated input symbols: A Bussgang ap-
[4] R. Cusani and A. Laurenti, Convergence analysis of the CMA blind proach, IEEE Trans. Signal Process., vol. 52, no. 5, pp. 18601869,
equalizer, IEEE Trans. Commun., vol. 43, no. 24, pp. 13041307, Apr. May 2005.
1995. [18] W. Sethares, G. A. Rey, and C. R. Johnson, Approaches to blind equal-
[5] S. Abrar and A. K. Nandi, Blind equalization of square-QAM sig- ization of signals with multiple modulus, in Proc. Int. Conf. Acoustics,
nals: A multimodulus approach, IEEE Trans. Commun., vol. 158, no. 6, Speech, Signal Process., Glasgow, U.K., May 1989, pp. 972975.
pp. 16741685, Jun. 2010. [19] M. J. Ready and R. P. Gooch, Blind equalization based on radius di-
[6] Special issue on blind systems identification and estimation, Proc. IEEE, rected adaptation, in Proc. Int. Conf. Acoust., Speech, Signal Process.,
vol. 86, Oct. 1998. Albuquerque, NM, USA, Apr. 1990, pp. 16881702.
[7] S. Colonnese, G. Panci, and G. Scarano, Using SOS in blind Buss- [20] J. J. Shynk, R. P. Gooch, G. Krishnamurthy, and C. K. Chan, A compar-
gang equalization for QAM systems, in Proc. IEEE 5th Workshop Signal ative performance study of several blind equalization algorithms, Proc.
Process. Adv. Wireless Commun., Lisbon, Portugal, Jul. 2004, pp. 415 SPIE Adaptive Signal Process., vol. 102, pp. 102117, Dec. 1991.
418. [21] G. Scarano, Cumulant series expansion of hybrid non-linear moment of
[8] G. Jacovitti, G. Panci, and G. Scarano, Bussgang-zero crossing equal- complex random variables, IEEE Trans. Signal Process., vol. 39, no. 4,
ization: An integrated HOS-SOS approach, IEEE Trans. Signal Process., pp. 291297, Apr. 1991.
vol. 48, no. 12, pp. 27982812, Dec. 2001. [22] G. Scarano, D. Caggiati, and G. Jacovitti, Cumulant series expansion of
[9] G. B. Giannakis and S. D. Haldford, Blind fractionally spaced hybrid nonlinear moments of n variates, IEEE Trans. Signal Process.,
equalization of noisy FIR channels: Direct and adaptive solutions, vol. 41, no. 1, pp. 486489, Jan. 1993.
IEEE Trans. Signal Process., vol. 45, no. 9, pp. 22772292, Sep. [23] G. Panci, S. Colonnese, P. Campisi, and G. Scarano, Multichannel blind
1997. image deconvolution using the Bussgang algorithm: Spatial and mul-
[10] C. B. Papadias and D. T. M. Slock, Fractionally spaced equalization of tiresolution approaches, IEEE Trans. Image Process., vol. 12, no. 11,
linear polyphase channels and related blind techniques based on multi- pp. 13241337, Nov. 2003.
channel linear prediction, IEEE Trans. Signal Process., vol. 47, no. 3, [24] S. Colonnese, P. Campisi, G. Panci, and G. Scarano, Blind image de-
pp. 641654, Mar. 1999. blurring driver by nonlinear processing in the edge domain, EURASIP J.
[11] L. A. Baccala and S. Roy, A new blind time-domain channel identification Appl. Signal Process., vol. 2004, no. 1, pp. 24622475, Jan. 2004.
method based on cyclostationarity, IEEE Signal Process. Lett., vol. 1, [25] J. Arenas-Garca, L. A. Azpicueta-Ruiz, M. T. M. Silva, V. H. Nascimento,
no. 6, pp. 8991, Jun. 1994. and A. H. Sayed, Combinations of adaptive filters, IEEE Signal Process.
[12] G. H. Golub and C. F. Van Loan, Matrix Computation, 3rd ed. Baltimore, Mag., vol. 33, no. 1, pp. 120140, Jan. 2016.
MD, USA: The Johns Hopkins Univ. Press, 1996. [26] W. B. Lopes and C. G. Lopes, Incremental-cooperative strategies in
[13] G. H. Wilkinson, The Algebraic Eigenvalue Problem. London, U.K.: Ox- combination of adaptive filters, in Proc. IEEE Int. Conf. Acoust., Speech
ford Univ. Press, 1978. Signal Process., Prague, Czech Republic, May 2011, pp. 41324135.

S-ar putea să vă placă și