Adaptive Sparse Channel ation Re- Weighted Zero … · dable, accurat e receiver for e approaches)....
Transcript of Adaptive Sparse Channel ation Re- Weighted Zero … · dable, accurat e receiver for e approaches)....
AkeyStaappis obe by zeralgspathecoemealgmecon
KestNL
masysfadis nOnestFig(LMalgandalgon andsta[3]ver
Ad
Weig
Abstract—Accuy technical isandard normaplied to adaptioften describeexploited and adaptive spa
ro-attracting gorithm. Howarsity efficiente reason why efficients unifoethod using regorithm. Simuethod achievenventional one
Keywords—nortimation (ASCLMF).
Broadband ainstream techstems. Due tding is unavoinecessary at thne of effectivtimation (ACEg. 1. It is weMF) algorithmgorithm in achd steady-statgorithm is unsn the followingd weight initia
able normalize]. Recently, mrified that bro
daptive
ghted Z
urate channel ssues in broaalized least meive channel estd by sparse chthen estimatio
arse channel enormalized l
wever, this algtly. By virtual
-norm spaormly. In this e-weighted zeroulation results es better este.
rmalized LMF CE), re-weighte
I. IN
signal transmhniques in theto the fact thidable, accurathe receiver forve approachesE). A typical ell known tham outperformhieving a goote performanstable due to tg three factorsalization [2]. Ted LMF (NLMmany channel oadband chann
e Spar
Zero-A
Guan D
{gui, mehbo
estimation pradband wirelean fourth (NLtimation (ACEhannel model, on performancestimation (ASleast mean fgorithm cannof geometrical
arse constrainpaper, we proo-attracting N
show that thimation perfo
F (NLMF), adapted zero-attrac
NTRODUCTION mission is bece next generathat frequencyte channel statr adaptive cohs is adoptingframework of
at ACE usingms the least mod balance bences. Howevethe fact that i: input signal pTo improve thMF) algorithmmeasurement
nels often exhi
rse Ch
Attrac
Gui, AbolfazlDepartment of
GraduateTd}@mobile.e
roblem is one ess communicaLMF) algorithmE). Since the ch
such sparsity e could be imp
SCE) methods fourth (ZA-Nnot exploit chl figures, we et penalizes ch
opose a novel NLMF (RZA-N
he proposedormance than
aptive sparse chcting NLMF
coming one otion communicy-selective chte informationherent detectiog adaptive chf ACE is sho
g least mean fmean square (
tween converer, standard ts stability depower, noise p
he stability ofm was propost experimentsibit sparse stru
hannel
cting N
Fourth
l Mehbodniyaf Communicate School of En
Tohoku UniverSendai, Japa
cei.tohoku.ac.
of the ations. m was hannel
could proved
using NLMF) hannel xplain hannel ASCE
NLMF) ASCE n the
hannel (RZA-
of the cation hannel n (CSI) on [1]. hannel wn in fourth LMS)
rgence LMF
epends power LMF, sed in s have ucture
as sof aUnfthe the
TestialgoAS(beprothe dire[3],NLprebas
Fi
Estim
Norma
h
and Fumiyuktion Engineeringineering, rsity an .jp, adachi@ec
shown in Fig. a very few chafortunately, Ainherent sparestimation pe
To estimate imation (ASCorithm (ASCECE-LMF metlow 5dB). Ba
oposed a normstability, but
ectly in sparse, we propose
LMF) algorithsented a deta
sed sparse c
ig. 1. ASCE for b
mation
alized
ki Adachi ing
cei.tohoku.ac.j
2. In other woannel coefficie
ACE using NLrse structure ierformance.
such a chanCE) methods E-LMF) werethod is not sta
ased on the denmalized LMF t stable NLMe channel estimed ASCE usinm [5]. From ailed explanatchannel estim
broadband commu
Using
Least
jp
ords, sparse chents and mostLMF algorithminformation an
nnel, adaptivusing sparse
e proposed in able except innse channel m(NLMF) algo
MF algorithm mation. Basedng zero-attrac
a geometrication of ℓ -no
mation. Since
unication systems
g Re-
Mean
hannel is const of them are zm always negnd it may deg
ve sparse chaleast mean f[4]. However
n low SNR remodel, author iorithm to impcannot be ap
d on the methcting NLMF al perspective
orm zero-attrae ℓ -norm
.
n
sisted zeros. glects grade
annel
fourth r, the egime in [3] prove pplied hod in
(ZA-e, we acting zero-
attracting introduces a uniform sparse constraint on different magnitudes of channel coefficients, the performance of ZA-NLMF degrades.
Inspired by re-weighted ℓ -norm minimization algorithm in [6], in this paper, we propose an improved ASCE method using re-weighted zero-attracting NLMF (RZA-NLMF) algorithm. Unlike the ZA-NLMF algorithm [5], RZA-NLMF algorithm can penalize different magnitudes of channel coefficients with different sparse constraint strength. The effectiveness of our proposed method is confirmed by computer simulation.
The remainder of this paper is organized as follows. A system model is described and standard LMF and NLMF algorithms are introduced in Section II. In section III, sparse ASCE using ZA-NLMF algorithm is introduced and improved ACSE using RZA-NLMF algorithm is highlighted. Computer simulations are presented in Section IV in order to evaluate and compare performances of the proposed ASCE methods. Finally, we conclude the paper in Section V.
II. SYSTEM MODEL AND STANDARD LMF ALGORITHM
Consider a baseband frequency-selective fading wireless communication system where FIR sparse channel vector = [ℎ , ℎ , … , ℎ ] is -length and it is supported only by
nonzero channel taps. Assume that an input training signal ( ) is input to probe the unknown sparse channel. At the receiver side, observed signal ( ) is given by ( ) = ( ) + ( ), (1)
where ( ) = [ ( ), ( − 1), … , ( − + 1)] denotes the vector of training signal ( ) , and ( ) is the additive white Gaussian noise (AWGN) assumed to be independent with ( ). The objective of ASCE is to adaptively estimate the unknown sparse channel estimator using the training signal ( ) and the observed signal ( ). According to [2], we can apply standard LMF algorithm to adaptive channel estimation, with the cost function ( ) = 14 ( ), (2)
where ( ) = ( ) − ( ) ( ) is -th adaptive updated error. The update equation of the filter can be written as ( + 1) = ( ) + ∂ ( )∂ (n) = ( ) + ( ) ( ), (3)
where denotes the step-size of gradient descend. Unfortunately, channel estimation using standard LMF algorithm is not stable in adaptive updating process and hence it cannot be employed directly [2]. To improve the reliability of LMF, an stable LMF algorithms was proposed by virtual of normalization in [3], which was termed as normalized LMF (NLMF) algorithm. The update equation is given by ( + 1) = ( ) + ( ) ( )‖ ( )‖ ‖ ( )‖ + ( )
= ( ) + ( ) ( )‖ ( )‖ , (4)
where = ( )‖ ( )‖ + ( ), (5)
denotes variable step-size of gradient descent and ‖∙‖ is the Euclidean norm operator and ‖ ‖ = ∑ | | .
Unfortunately, interpretation of the relation between
and ( ) is not correct in [7]. Here, we observe that when ( ) ≫ ‖ ( )‖ , then approaches to ; when ( ) ≈‖ ( )‖ , then approaches to 2⁄ ; when ( ) ≪ ‖ ( )‖ , then approaches to 0. Fig. 3 illustrates the intuitive relation between and ( ) . For the standard NLMF algorithm, consider three step-sizes : 0.2, 0.6 and 1.2 . As ( ) increases, according to the depicted curves in Fig. 3, it is easy to find that the stability of NLMF approaches NLMS which stability was proven in [8]. Please also note that when ( )
Fig. 2. A typical example of sparse multipath channel.
Fig. 3. Relations between and ( ).
dec
A.
cha
can
whrowdismavecCSiso
hoLAgiv
whthegeoestspabet
congeoest
B.
stamecauutiLAconfun
whtheup
creases, the st
III. AD
Geometrical estimation
Consider a bannel estimati= [ , , …n be written as
here ∈w ; = [stribution atrix; = [ctor which is
S, the traininometry propert∈ (0,1), i.e., (1 − )lds for all
ASSO algorithven by = arhere denotee mean-squareometrical intimation is dearse channel tween ℓ -norm= 0, howevenstraint and ometrical figtimation using
ASCE using Recall that th
andard NLMFethod does notused by its oilize the sparsASSO algoritnstraint to thenction as
here denotee mean-fourthdate equation
tep size als
DAPTIVE SPAR
l interpretation
baseband systion. Assume ] , and s = [=
is a Toeplitz, … , , … ,(0, ) and, , … ,same as in
ng signal maty (RIP) [9] of∈ RIP( ,‖ ‖ ‖having no m
hm [10] base
rg lim 12 ‖ −es a regularize error (MSE) nterpretation epicted. Whenestimator canm sparse conser, there is no
solution plagure to expg ZA-NLMF a
ZA-NLMF alghe adaptive c
F algorithm int take advantaoriginal cost e constraint o
thm in (7), e cost functio
( ) = 14 (es a regularizh error (MFEof ZA-NLMF
so reduces to e
RSE CHANNEL
n of CS-based
tem model foa -length trthen the rece
, … , , … ,+ , z training sign] ∈ id denotes ] is an 1Eq. (1). Fromatrix satisf order with) if ‖ (1 +
more than nd sparse chan
− ‖ + ‖zation parameterm and sparof CS-basedn > 0 , as
n be obtained straint and solo convex poinane. One canplain adaptivalgorithm in [5
gorithm channel estimn Eq. (4), howage of the chanfunction in (2
or penalty funcwe introduce
on in (4) and
) + 2‖ ( )zation parameE) term and sF algorithm [5
ensure stability
ESTIMATION
d sparse chann
or CS-based sraining sequeneived signal m
] nal matrix wits an AWGNan id1 unknown ch
m the perspectsfies the resth positive para
)‖ ‖ , nonzero taps.nnel estimator
‖ ‖ , ter which balrsity of . In Fd sparse ch the figure sfrom convex
lution plane;nt between ℓ -n also extende sparse ch
5].
mation methodwever, the pronnel sparsity. I2) which doection. Accorde ℓ -norm s
obtain a new
)‖1, ter which balsparsity of ] is given by
y.
nenl
sparse nce as model
(6)
th first N with dentity hannel tive of tricted ameter
(7)
Then, r is
(8)
lances Fig. 4, hannel shows,
point When -norm d the hannel
d uses oposed It was es not
ding to sparse
w cost
(9)
lances . The
whegen
whe{0,1conesti
(coespa
Fi
Fico
( + 1) =ere =
nerated from
sgn ( )ere ( ) =1, … , }. It is
nstraint to aimation [11].
As shown in= 0) on solefficient. In oarsity efficient
ig. 4. Sparse chan
ig. 5. Sparse conoefficients by virt
( ) + (‖ and sgn(∙)
= ‖ ( )‖( )= [ℎ ( ), ℎ (s well knownapproximate
n Fig. 5, howlution plane other words, tly. If we int
nnel estimation wi
nstraint strength ctual of re-weighte
( ) ( )( )‖ + sgdenotes sign
= 1, ℎ (0, ℎ (−1, ℎ () … , ℎ (n that ZA-NL
the optimal
wever, ℓ -normis uniform ZA-NLMF c
troduce a re-
ith ℓ -norm spars
comparison betweed ℓ -norm zero-a
gn ( ) , function whi
( ) > 0( ) = 0( ) 0, )] and
LMF uses ℓ -l sparse cha
m sparse consfor each cha
cannot exploi-weighted ℓ -
se constraint.
een different chaattracting.
(10)
ich is
(11)
∈norm annel
traint annel it the norm
annel
sparse constraint function to the cost function in (4), adopt different re-weighted factors, e.g., = 1,5,10 and 20, then we are able to penalize different sparse constraint strength on different channel coefficients. In other words, RZA can penalize stronger constraint strength than smaller channel coefficients, and vice versa. Based on this idea, an improved ASCE using RZA-NLMF algorithm is proposed in the following.
C. ASCE using RZA-NLMFalgorithm The ZA-NLMF cannot distinguish between zero taps and
non-zero taps since all the taps are forced to zero uniformly as show in Fig. 5. Unfortunately, ZA-NLMF based approach will degrade amount of estimation performance. Motivated by reweighted ℓ -minimization sparse recovery algorithm [6] in CS [12,13], we proposed an improved ASCE method using RZA-NLMF algorithm. The cost function of RZA-NLMF is given by ( ) = 14 ( ) + log(1 + |ℎ |/ ), (10)
where > 0 is a regularization parameter which trades off the estimation error and channel sparsity. The corresponding update equation is ( + 1) = ( ) + ( ) ( )‖ ( )‖ + sgn ( )1 + | ( )|, (11)
where = / is a parameter which depends on step-size , regularization parameter and threshold , respectively. In
the second term of (11), if magnitudes of ℎ ( ), =0,1, … , − 1 are smaller than 1⁄ , then these small coefficients will be replaced by zeros in high probability.
D. Equivalence between CS-based sparse channel estimation and ASCE
According to LASSO algorithm, we explained the connections between CS-based sparse channel estimation method and ASCE. Here, we further interpretation their equivalence between them. It will be useful to enrich sparse signal processing theory. Let take LASSO and ZA-NLMF for an example. If each row of training matrix in system model (6) is , then by virtual of matrix vectorization, then it can be written as [ , … , , . . , ] . The -th receive signal is obtained as = + ζ . Their corresponding functions of different vectors are listed in Tab. I.
TAB. I. EQUIVALENCE BETWEEN LASSO AND ZA-LNMF.
LASSO ZA-NLMF
Training signal ( )
Received signal ( )
Channel vector = [ , , … , ] = [ℎ , ℎ , … , ℎ ]Sparse constraint ‖ ‖ sign( )
Iterative times = mod( − 1, ) + 1
IV. COMPUTER SIMULATIONS In this section, the proposed ASCE method using RZA-
NLMF algorithm is evaluated. For achieving average
performance, 1000 independent Monte-Carlo runs are adopted. The length of channel vector is set as = 16 and its number of dominant taps is set to = 1 and 4, respectively. Each dominant channel tap follows random Gaussian distribution as (0, σ ) and their positions are randomly allocated within the length of which is subject to E{|| || =1} . The received signal-to-noise ratio (SNR) is defined as 10log ( ⁄ ), where = 1 is the unit transmission power. Here, we set the SNR as 3dB and 5dB in computer simulation. All of the step sizes and regularization parameters are listed in Tab. II. The estimation performance is evaluated by average mean square error (MSE) which is defined by Avergae MSE{ ( )} = E{‖ − ( )‖ }, (12)
where E{∙} denotes the expectation operator, and ( ) are the actual channel vector and its -th iterative adaptive channel estimator, respectively.
TAB. II. SIMULATION PARAMETERS.
Parameters Values
Channel length = 16
No. of nonzero coefficients = 1 and 4
Step-size: 1.0 and 1.5
Regularization parameter: and 1e − 5
Re-weighted factor: 5, 10, 20
Two experiments are considered in this section. In the first
experiment, ASCE methods are evaluated in SNR = 5dB . Regularization parameters, i.e., λ and λ , are set as λ =λ = 1 − 6. Fig. 6-8 shows that proposed method can achieve better estimation performance than ZA-NLMF in different step-sizes, i.e., = 0.5 , 1.0 and 1.5 . When the number of nonzero coefficients is = 1, bigger re-weighted factor can obtain better estimation and vice versa. Please note that both ZA-NLMF and RZA-NLMF algorithms apply ASCE
Fig. 6. Average MSE performance comparisons at SNR = 5dB and step-size = 0.5.
which can achieve better estimation performance for sparser channel. As shown in three figures, when = 1, the depicted curves’ gap between NLMF and either ZA-NLMF or RZA-NLMF larger than the case with = 4 . It is worth mentioning that bigger reweighted factor obtains better estimation performance, e.g., RZA-NLMF using ε = 20 achieves better estimation performance than ε = 10 . It is mainly due to the bigger reweighted factor which can obtain variable sparse constraint for different magnitude of channel coefficients.
To verify the flexibility of the proposed RZA-NLMF
method at different SNR regimes, ASCE methods are evaluated in SNR = 10dB . Also, regularization parameters, i.e., λ and λ , are set to λ = λ = 5e − 8 to improve the
estimation performance. According to Fig. 9, RZA-NLMF works well and achieves better estimation performance than ZA-NLMF and NLMF.
V. CONCLSION In this paper, the disadvantage of conventional ASCE
method using ZA-NLMF algorithm was discussed from a geometrical perspective point of view. We found that ℓ -norm sparse constraint penalizes uniformly on different channel coefficients. Inspired by re-weighted ℓ -norm algorithm in CS, we proposed an improved ASCE using RZA-NLMF which penalizes smaller channel coefficients with stronger sparse constraint and vice versa. It was confirmed by computer simulation that the proposed ASCE using RZA-NLMF algorithm achieves better performance than either ZA-NLMF or NLMF approaches.
ACKNOWLEDGMENT This work was supported by the Japan Society for the
Promotion of Science (JSPS) postdoctoral fellowship and the National Natural Science Foundation of China under Grant 61261048.
REFERENCES
[1] F. Adachi, H. Tomeba, K. Takeda, and S. Members, “Introduction of frequency-domain signal processing to broadband single-carrier transmissions in a wireless channel,” IEICE Transactions on Communicationns, vol. E92-B, no. 9, pp. 2789–2808, Sept. 2009.
[2] E. Walach and B. Widrow, “The least mean fourth (LMF) adaptive algorithm and its family,” IEEE Transactions on Information Theory, vol. 30, no. 2, pp. 275–283, Mar. 1984.
[3] E. Eweda, “Global stabilization of the least mean fourth algorithm,” IEEE Transactions on Signal Processing, vol. 60, no. 3, pp. 1473–1477, Mar. 2012.
[4] G. Gui and F. Adachi, “Sparse LMF algorithm for adaptive channel estimation in low SNR region,” International Journal of Communication Systems, DOl:10.1002/dac.2531, pp. 1-11, Apr. 2013.
[5] G. Gui and F. Adachi, “Sparse least mean forth filter with zero-attracting L1-norm constraint,” Submitted for International Conference
Fig. 7. Average MSE performance comparisons at SNR = 5dB and step-size = 1.0.
Fig. 8. Average MSE performance comparisons at SNR = 5dB and step-size = 1.5.
Fig. 9. Average MSE performance comparisons at SNR = 10dB and step-size = 1.5.
on Information, Communications and Signal Processing (ICICS), Tainan, Taiwan, Dec. 10-13, 2013.
[6] E. J. Candès, M. B. Wakin, and S. P. Boyd, “Enhancing sparsity by reweighted L1 minimization,” Journal of Fourier Analysis and Applications, vol. 14, no. 5–6, pp. 877–905, Oct. 2008.
[7] E. Eweda and N. J. Bershad, “Stochastic analysis of a stable normalized least mean fourth algorithm for adaptive noise canceling with a white Gaussian reference,” IEEE Transactions on Signal Processing, vol. 60, no. 12, pp. 6235–6244, Dec. 2012.
[8] B. Widrow and D. Stearns, Adaptive signal processing, New Jersey: Prentice Hall, 1985.
[9] E. J. Candes, “The restricted isometry property and its implications for compressed sensing,” Comptes Rendus Mathematique, vol. 1, no. 346, pp. 589–592, May 2008.
[10] R. Tibshirani, “Regression shrinkage and selection via the Lasso,” Journal of the Royal Statistical Society (B), vol. 58, no. 1, pp. 267–288, Jan. 1996.
[11] D. L. Donoho and Y. Tsaig, “Fast solution of L1-norm minimization problems when the solution may be sparse,” IEEE Transactions on Information Theory, vol. 54, no. 11, pp. 4789–4812, Nov. 2008.
[12] D. L. Donoho, “Compressed sensing,” IEEE Transactions on Information Theory, vol. 52, no. 4, pp. 1289–1306, Apr. 2006.
[13] E. J. Candes, J. Romberg, and T. Tao, “Robust uncertainty principles : exact signal reconstruction from highly incomplete frequency information,” IEEE Transctions on Information Theory, vol. 52, no. 2, pp. 489–509, Feb. 2006.