IEEE Conference Template-1
IEEE Conference Template-1
Abstract—Recently, adaptive filter algorithms for jointly wide- rate, or any other periodic component [3]. Important examples
sense cyclostationary signals based on a minimal time-averaged are orthogonal frequency division multiplexing (OFDM) sig-
mean-square error (TA-MSE) framework have been developed. nals and narrowband (NB) power line communication (PLC)
These algorithms demonstrate their estimating efficacy in the
presence of Gaussian noise, particularly they are very effective systems [4]. Gardner and his co-workers in [5]–[7] have
when the total number of possible filter coefficients to implement shown that this modelling of signals as cyclostationary signals
the linear estimator for cyclostationary signals is limited. This allows a distinct improvement in the filtering performance over
brief presents an adaptive filtering algorithm for cyclostationary classical stationary signal filtering techniques in environments
signals in the presence of α-stable noise environments. The where the interferer power is very high. However, a family
proposed algorithm is based on the optimization of the time-
averaged fractional pth-order moments of the error cost function. of LMS adaptive filtering algorithms for the cyclostationary
The proposed algorithm is called the minimum time-averaged signal model was developed, and their performance analysis
least-mean fractional pth-order moments (TA-LMP) algorithm. was examined in the literature [8], [9].
We analyzed α-stable distribution model with the characteristic Adaptive filtering methods in cyclostationary environment,
exponents {1 < p < α ≤ 2} With an approximate solution using i.e., cyclic Wiener filter (CWF) are widely implemented by the
Taylor’s expansion, the transient analysis in terms of the mean
and mean-square convergent behavior is derived. The obtained frequency shift (FRESH) filter structure by using a set of linear
theoretical results are supported by experimental simulations. time-invariant (LTI) filters, whose inputs are frequency-shifted
Index Terms—Adaptive Filtering, Cyclostationary Signals, versions (cyclic frequencies) of the input signal refer figure 1.
Alpha-Stable Distribution, Impulsive Noise, Fractional Lower- The filte [6] would be more efficient if the signals exhibited
order Moments (FLOM). more different cyclic frequencies. However, the most difficult
challenge in the design of the FRESH filter is determining the
I. I NTRODUCTION
optimal subset of frequency shifts (cyclic spectrum) within
Recently, many efforts have been devoted on solving many some design constraints. Generally, the frequency shift error
engineering problems using adaptive algorithms. However, a corrector is added to the basic structure of the FRESH filter
well-known type of adaptation algorithm is the Least-mean- to eliminate the effects of the cyclic frequency errors. This
square (LMS) algorithm [1], [2] which has the advantages new structure is then called a linear periodically time-varying
of being easy to implement, low computation complexity (LPTV) system [7], [10]. Ultimately, the LPTV system re-
while providing a relatively good performance. In General, quires an infinite number of LTI filters in order to realize the
the second order statistics of the error signal, e.g., mean-square CWF, which may not be feasible. Nonetheless, the frequency
error (MSE), is used as the cost function to design LMS and shift errors can be compensated for by adopting an adaptive
its family of normalized LMS (NLMS) and variable step-size FRESH filtering technique as a practical alternative [9], [10].
LMS (VSS-LMS), which is considered as the key methods in Generally, the Wiener filter defines optimal (uses MSE
Gaussian noise environments [2]. criteria) LTI filter for estimation of the desired signal d(k)
Most digital communications signals are typically cyclo- based on the input signal u(k), when both d(k) and u(k) are
stationary; this property arises from the signals’ implied jointly stationary signals [6]. Therefore, the LTI characteristics
periodicity, which is related to the carrier frequency, baud (auto- and cross-correlation functions) are constant all the
Manuscript received May 15, 2023; revised XXXXXX, 2023. The associate time, and the estimation error results are constant too (they
editor coordinating the review of this manuscript and approving it for do not depend on the time). But in the scenario of the FRESH
publication was XXXXXXXX. This work was supported in part by the filter, its LTI’s filters characteristics are constant only over
National Natural Science Foundation of China under Grant 61971083, in part
by the fundamental research funds for the central university under Grants periodic time instants (in frequency shift cycles); consequently,
3132019341. (Corresponding author: Li Sen.). the estimation error results will not be constant either. How-
ever, one of the potential solutions to finding optimal LTI that we denote the operator mod(a, m) returns the remainder after
minimizes the MSE at all instants is applying random phase the division of a by m.
translation. In which the d(k) and u(k) are translated into
jointly asymptotically stationarized processes by a uniformly II. SYSTEM MODEL
distributed random time shift e.g., [−T0 /2, T0 /2], where T0 is A. Wide-Sense Cyclostationary Stochastic Processes.
the period of both signals [6]. In this scenario, the optimum Definition 1: The process d(k) is said to be second-order
cost function to be considered is a minimal time-averaged statistic (wide-sense) cyclostationary signal (WSCS) with pe-
mean-squared error (TA-MSE) [8]. In other words, we can riod T0 if and only if its mean and autocorrelation function
use the TA-MSE cost function to force the LTI characteristics (s) are constant with the period T0 , [6] i.e.:
to the optimal solution. In this sense, the LPTV filter can
be considered as a generalization form of the conventional E [d(k)] = E [d(k + mT0 )] ,
Wiener theory for optimal LTI filters when the involved signals
E [u(k)] = E [u(k + mT0 )] , ∀k ∈ (−∞, ∞) (1)
are cyclostationary signals. Additionally, the TA-LMSE cost
function is more convenient than the conventional MSE cost
Ru (k, l) = E u(k + l)uH (k)
function for the linear estimator for cyclostationary signals
when the number of LTI filters is smaller than the number = E u(k + mT0 + l)uH (k) ,
∀k ∈ (−∞, ∞) (2)
required for implementing the CWF [8], [11].
In the literature, it has been demonstrated that the perfor- Where m is an integer.
mance of filtering methods relay on the MSE seriously effected Definition 2: A pair of signals d1 (k) and d2 (k) said exhibit
in impulsive noise environments [12], [13]. To tackle this is- jointly WSCS (JWSCS) with period T0 if each signal is
sues, Shao in [14] has proposed estimation based on fractional wide-sense cyclostationary
with period T0 and their cross-
lower-order moments (FLOM’s) cost function, which mainly correlation function E d1 (k + l)dH 2 (k) is periodic with pe-
include least mean-pth (LMP) moments of error and mean riod T0 [6, 9, 18].
absolute error (MAE) [15]. Numerous studies have verified
B. Overview of α-Stable Distribution
performance and effectiveness of the algorithm based FLOM
framework [16], [17]. These works have shown that pth norm It is known that to model the environments that exhibit
algorithms achieves more robustness and superior performance heavy-tailed impulsive noise, the α-stable model is usually
compared to the algorithm based on the MSE cost function in used [16], [17]. The symmetric α-stable (SαS) model defined
the presence of heavy-tailed impulsive noise. in [17] is a special case of the α-stable distribution, which
Therefore, inspired by the advantages of minimum time- is controlled by the thickness of the tails of the distribution,
averaged criteria of optimality used in [8] and the fractional namely the characteristic exponent α, a scale factor, namely
lower-order statistic theory [17], in the work presented here, dispersion γ > 0 and a location parameter ϱ = 0. The
first we aim to model adaptive filtering for jointly proper characteristic function of (SαS) model is given by [17]
complex cyclostationary input and SOI signals that is ro- ψ(t) = exp (jϱ − γ |t| )
α
(3)
bust under impulsive noise environments. Secondly, with an
approximate solution using Taylor’s expansion [13], Bound However, the fractional pth-order moments of an symmetric
conditions in terms of the mean convergent behavior is derived. α-stable model SαS random variable X with 1 < p < α ⩽ 2,
We also denote that; this work adopts the same assumptions when the location parameter is zero, ϱ = 0, is described as
used in [8] except that the noise here follows the alpha-stable [17]
p p
distribution model as in [14], [17]. Moreover, to the best of F LOM (p, α, γ) = E [|X| ] = C (p, α) γ α (4)
the authors’ knowledge, to date, no studies have reported the
results of an adaptive filter for cyclostationary signals using where
2p+1 Γ p+1 Γ − αp
the optimization criteria we used in this work. 2
C (p, α) = (5)
απΓ − p2
A. Notations Where Γ (.) represents the gamma function, and γ > 0 is the
dispersion levels of the noise which behaves in a similar way to
Throughout this work, we denote to symbol for column
the variance of the Gaussian distribution for {1 < p < α ≤ 2}.
vectors with lower-case letters, e.g., u , matrices with upper-
case letters, e.g., U and denote to Schur-Hadamard (element-
wise) product as ⊙. We use E [.] , (.)H , (.)T , and (.)∗ to denote
the Stochastic expectation, Hermitian transpose, transpose and
complex conjugate, respectively. The vector u(k) is recovered
from matrix U via u(k) = vec(U (k)) similar to the procedure
in [1]. We denote the time-averaging operator as ⟨.⟩T0 , in
particular, the time-averaging forPthe sequence u(k) over Fig. 1. Linear estimation of scalar jointly cyclostationary signals [8].
T0 −1
period T0 > 0, ⟨u(k)⟩T0 = √1T i=0 u(k − 1) [8]. Also,
0
C. TA-LMS review III. PROPOSED ALGORITHM
The existing TA-LMS estimation algorithm has been devel- A. Proposed TA-LMP algorithm
oped for the discrete-time (DT) jointly proper-complex wide- Fortunately, as shown in [17], the dispersion of the α-stable
sense cyclostationary input and SOI signals under a Gaussian distribution process is similar to the variance of the Gaussian
environment based on a minimal TA-MSE framework. The distribution for the characteristic exponent 1 < α ≤ 2.
TA-MSE cost function was defined in [8], as As shown in (3), the dispersion of the α-stable model, is
D 2
E solely related to its pth-order moment by a constant. As a
J(w) = d(k) − wH (k)u(k) (6)
T0 result, minimizing dispersion is the same as minimizing the
fractional lower order of pth-order moments. Thus, based
where
K on this foundation and using the concept of minimum time-
d(k) = wM (k) + η(k) (7)
averaged framework, we formulate the cost function for the
With d(k) is a scalar SOI, wM (k) is an M×1 possibly optimal time-averaged signal estimation problem under consideration
time-varying coefficient vector (LPTV solution) and η(k) be based on the pth-order moments criterion as.
random estimation error, and background noise, which is
D p
E
J(w) = d(k) − wH (k)u(k) (13)
cyclostationary process with a period T0 . However, since d(k) T0
and u(k) are JWSCS with period T0 , it leads to that wM (k) where 1 ≤ p < α ≤ 2. The solution of the problem (10)
periodic sequence with period T0 . Also, we define results in the coefficient update equation of the proposed time-
˜ = wK (k)u(k)
d(k) (8) averaged least mean pth-order (TA-LMP) algorithm for k > 0
M
p−1
˜ w(k + 1) = w(k) + µU (k) |ϵ(k)| ⊙ sign (ϵ(k)) (14)
Which a scalar d(k) represents the output of the linear
estimator, based on an M×1 multivariate input signal u(k) where ϵ(k) = d(k) − wH (k)u(k) T0 .
, to estimate an M × 1 time-invariant coefficient vector w(k). We note that, when the p = 1, (14) turns to algorithm called
This linear estimator can be considered as scalar LTI or time-averaged least absolute deviation (TA-LAD) algorithm as
scalar FRESH estimators [8], [11], where the relationships following
between the expression (7) and the common scalar LTI and
scalar FRESH estimator structures can be easily achieved as w(k + 1) = w(k) + µU (k)sign (ϵ(k)) (15)
explained in [8]. However, as shown in Fig. 1, the multivariate IV. B OUND CONDITIONS ON µ
input signal u(k) is obtained by mapping the input scalar r(k).
To perform the bound conditions analysis to the (13), when
For the sake of simplification, we neglect the mapping stage
the characteristic exponent 1 < α ≤ 2 is considered, for that
and the filters applied directly to the cyclostationary input
aim the following assumptions are made
signal. As any DT LPTV system can be realized as a FRESH
Assumption 1: The signals d(k) and u(k) are proper
filter [11], [18], in this work, we consider the linear estimator
complex jointly zero-mean Gaussian WSCS with period T0 ,
in (7), is the scalar FRESH filter with L number of LTI filter
which leads wM (k) to be also periodic with period T0 , such
branches and each LTI filter has k th weight taps as described
that wM (k) = wM (k + T0 ).
in [18]. The total number of the coefficients of the system
Assumption 2: The corrupted noise η(k1) is cyclostationary
M = L × K. The received signal r(k) passes via number of
process with period T0 , and it’s independent of any other
n L corresponding
branches a set of the cyclic frequency shifts
o signal, ∀k1 ̸= k.
L−1
exp {2πθn /T0 }n=0 , θt ∈ {−(L − 1)/2 : (L − 1)/2} to
To derive transient analysis of the TA-LMP algorithm in
form an M × 1 input vector u(k) of the tth LTI filters as (14), first we assume that the time-average Wiener-Hopf vector
2πθt
wopt defined in ( [8], Eq. (5)) is existed. Then we define the
(u(k))n.K+v = r(k − v) exp j (k − v) (9) coefficient vector error by
T0
Where n ∈ {0, 1, ..., L − 1}, and v ∈ {1, 2, ..., K}. Thus, ŵ(k) = wopt − w(k) (16)
applying steepest decent method to (6), the TA-LMS adaptive By substituting the definition of w(k) in (13) and (15) into
filtering algorithm for k > 0 [8]. last expression above we obtain
w(k + 1) = w(k) + µU (k) d̃(k) − U H (K)w(k) (10) ϵ(k) = U H (k)wM (k) + η(k) − U H (k)(wopt − ŵ(k))
where ⇒ ϵ(k) = ϵa (k) + ϵg (k) + η(k) (17)
1 where g(k) = wM (k) − wopt , ϵa (k) = U H (k)ŵ(k), ϵg (k) =
U (k) = √ {u(k), u(k − 1), ..., u(k − T0 + 1)} , M × T0
T0 U H (k)g(k) and η(k) represents the time-averaged background
(11) noise and impulsive noise component.
1 T However, by employing (15) into (13), we obtain
d̃(k) = √ {d(k), d(k − 1), ..., d(k − T0 + 1)} , T0 × 1
T0 p−1
(12) ŵ(k + 1) = ŵ(k)) − µU (k) |ϵ(k)| ⊙ sign (ϵ(k)) (18)
To make the performance analysis easier and simpler, we A. Mean Convergent
modify the last expression as To evaluate the mean convergent, we use the approximation
of (22) that defined in [19], such that
T0
X p−1 ϵj (k)
ŵi (k + 1) = ŵi (k)) − µ uij (k) |ϵj (k)|
h i
(19) ′
E f (ε̃j (k) + η j (k)) ≈ E ε̃j (k)f (η j (k))
j=1
|ϵj (k)|
= E [ε̃j (k)] δη,j (k) (29)
An equivalently, (18) can be rewritten as
After substituting the last approximation in (28) into (21) and
T0
X p−2 take stochastic expectation, we obtained
ŵi (k + 1) = ŵi (k)) − µ uij (k) |ϵj (k)| ϵj (k) (20)
j=1 T0
X
E [ŵi (k + 1)] = E [ŵi (k)]−µ E [uij (k)ε̃j (k)] δη,j (k)
For i = 1, 2, ..., M , where uij is the ijth entry of the j=1
matrix U (k), and di (k) and d˜i (k) are the ith entries of the (30)
desired signal vector d(k) and the estimated signal vector
˜ Now after doing some mathematical manipulations to (29), we
d(k), respectively. However, for the sake of more simplicity,
can re-write the expression (29) in a vector form, Therefore,
we rewrite the last expression (19), as
the mean relation satisfies the following recursion for k > 0
T0
X
ŵi (k + 1) = ŵi (k)) − µ uij (k)f (ϵj (k)) (21) E [ŵi (k + 1)] = (IM − µδη Ru (k)) E [ŵi (k)]
j=1 − µδη E [Ru g(k)] (31)
p−2 Thus, from [18], we find that as k → ∞, E [Ru (k)g(k)] → 0,
where f (ϵ(k)) = |ϵj (k)| ϵj (k).
Thus, by using (15), we assumed that ε̃j (k) = thus, it follows that, for the mean convergent of the TA-LMP
⟨ϵa,j (k) + ϵg,j (k)⟩T0 , and η j (k) is time-averaged of the Cor- algorithm under α-stable noise distribution environments, the
rupt noise, then we define step-size should satisfy the condition
1
T0 0<µ≤ , ∀k ∈ T0 (32)
δη λmax (Ru (k))
X
ŵi (k + 1) = ŵi (k)) − µ uij (k)f (ε̃j (k) + η j (k)) (22)
j=1 Also, it can be written as
Now, by taking the Taylor series expansion of the nonlinearity 1
0<µ≤ p−2 (33)
function f (ϵj (k)) with respect to ε̃j (k) around the noise η j (k) (p − 1)C(p − 2, α)λmax (Ru (k)) γ α
as in [13]. It follows that where λmax (.) denotes to maximum eigenvalue of matrix (.).
′ ′′ It’s very easy to recognize that the expression derived in (32)
f (ϵj (k)) = f (η j (k)) + ε̃j (k)f (η j (k)) − ε̃2j (k)f (η j (k))
is a generalized expression form of the mean convergent which
(23)
′ ′′ is including TA-LMS, and the fractional pth order algorithms.
Where f (.) and f (.) are the first and second order deriva-
For example, if we considered the case of Gaussian noise
tives of the function f (.), such that
distribution where α = p = 2, then based on the (27), noise
f (η j (k)) = η j (k)
p−2
η j (k) (24) component δ = 1, and therefore (32) becomes identical to
the mean convergent relation of TA-LMS derived in ( [8], Eq.
′ p−2 (33)).
f (η j (k)) = (p − 1) η j (k) (25)
V. SIMULATION RESULT
′′ p−4
f (η j (k)) = (p − 1)(p − 2) η j (k) (26) In this part, we give some numerical simulation results in
alpha-stable noise to validate the theoretical analysis and sim-
If we consider the fractional pth-order moments noise distri- ulation performance of the TA-LMP algorithm. An unknown
bution, then by using definition (4) and (5), it holds that system wopt is randomly chosen and the input signal achieved
p−1 as an average over 100 independents Monte Carlo experiments.
ϑη,j (k) = E f (η j (k)) = C(p − 1, α)γ α (27) We consider the received signal at the end of the filter input
h ′ i has the structure as follows [8].
p−2
δη,j (k) = E f (η j (k)) = (p − 1)C(p − 2, α)γ α (28)
r(k) = (1 + ξsin(2π/T0 )) exp (jϕ(k)) , 0<ξ<1 (34)
where C(p − 1, α) and (p − 1)C(p − 2, α) can be computed where we used ξ = 0.5, T0 = 20, and ϕ(k) be independent
using (4). identical distribution (i.i.d.) process, uniformly distributed over
In the following TA-LMP analysis, we will investigate the ϕ(k). We assumed that the corrupt noise component η(k)
requirements that guarantee the convergence of the coefficients is mixed of SαS distributed with characteristic exponent
vector of the adaptive filter w(k) {1 < p < α ≤ 2} and Gaussian noise. For simplicity in the
implementation, we used a finite impulse response (FIR) filters
structure in the implementation of our designed FRESH filter
instead of LTI filters. The number of FIR filters branches
are L and the number of weigh taps at each FIR filter is
N = 1, which means that the total number of the coefficient
weights of the FRESH filter is M = L × N . Furthermore,
coefficient vector wM (k) is set the satisfy (wM (k))n =
the
1)
1 + mod(k,T
T0 e|n−1| , n ∈ {0, 1, ..., L − 1} with T1 ≥ T0 /2
as in [8], here we used T1 = 9. The generalized signal-to-
noise ratio (GSNR) used is 30 dB, where GSNR is defined in
( [20], Eq. (38)).
we consider the α-stable distribution noise model
Sα−stable (α, β, γ, ϱ) which is defined in [13] as the non-
Gaussian heavy-tailed impulsive noise model to test the robust-
ness of the proposed TA-LMP algorithm under the impulsive
noise.
R EFERENCES
[1] A. H. Sayed, Adaptive filters. John Wiley & Sons, 2011.
[2] E. Eweda, “Transient performance degradation of the LMS, RLS, sign,
Fig. 2. MSD performance of the TA-LMP algorithm versus the different signed regressor, and sign-sign algorithms with data correlation,” IEEE
choices of the exponent p, for Sα−stable (1.2, 0, 1, 0) and µ = 0.005. Transactions on Circuits and Systems II: Analog and Digital Signal
Processing, vol. 46, no. 8, pp. 1055–1062, 1999.
Fig.2 and Fig. 3 illustrate the MSD performance of the [3] W. A. Gardner, “Cyclostationarity in communications and signal pro-
TA-LMP algorithm over long iteration in the extremely SαS cessing,” tech. rep., Statistical Signal Processing Inc Yountville Ca,
1994.
distribution, which varies from highly strong impulsive noise [4] Y.-R. Chien, S. S.-D. Xu, and S.-H. Lu, “Cyclostationary impulsive
condition with Sα−stable (1.2, 0, 1, 0) to the environment that noise mitigation algorithm for narrowband powerline communications,”
exhibits less impulsive noise level with Sα−stable (1.5, 0, 1, 0), Journal of the Franklin Institute, vol. 357, no. 1, pp. 687–703, 2020.
[5] W. Gardner, “Stationarizable random processes,” IEEE Transactions on
respectively for different options of the exponent p. We used information theory, vol. 24, no. 1, pp. 8–22, 1978.
also, M = 5 i.e., (L = 5, N = 1) and step-size µ = 0.005 and [6] W. A. Gardner, “Cyclic wiener filtering: theory and method,” IEEE
0.0021 for Fig. 2 and Fig. 3 respectively. It can be observed Transactions on communications, vol. 41, no. 1, pp. 151–163, 1993.
[7] W. A. Gardner, A. Napolitano, and L. Paura, “Cyclostationarity: Half
from these two figures that the simulated results show that a century of research,” Signal processing, vol. 86, no. 4, pp. 639–697,
the proposed TA-LMP is robust against heavy-tailed noise 2006.
effectively, while the existing TA-LMS [8] shows very poor [8] N. Shlezinger, K. Todros, and R. Dabora, “Adaptive filtering based on
time-averaged mse for cyclostationary signals,” IEEE Transactions on
performance in such an environment as compare to the TA- Communications, vol. 65, no. 4, pp. 1746–1761, 2017.
LMP. [9] J. Zhang, K. Wong, and Z. Luo, “A generalized structure of blind
adaptive frequency-shift filter for signal extraction,” in 1997 IEEE
VI. CONCLUSION International Symposium on Circuits and Systems (ISCAS), vol. 2,
In the presented work here, we introduced the time-averaged pp. 1197–1200, IEEE, 1997.
[10] O. A. Y. Ojeda et al., “Adaptive-fresh filters for compensation of cycle-
fractionally pth-order moments TA-LMP algorithm for cy- frequency errors,” IEEE Transactions on Signal Processing, vol. 58,
clostationary signals in a non-Gaussian environment. The no. 1, pp. 1–10, 2009.
[11] N. Shlezinger and R. Dabora, “Frequency-shift filtering for ofdm signal
recovery in narrowband power line communications,” IEEE Transactions
on Communications, vol. 62, no. 4, pp. 1283–1295, 2014.
[12] O. M. Abdelrhman and L. Sen, “Robust adaptive filtering algorithms
based on the half-quadratic criterion,” Signal Processing, vol. 202,
p. 108775, 2023.
[13] O. M. Abdelrhman, Y. Dou, and S. Li, “A generalized robust logarith-
mic family-based adaptive filtering algorithms,” IEEE Transactions on
Circuits and Systems II: Express Briefs, 2023.
[14] P. G. Georgiou, P. Tsakalides, and C. Kyriakakis, “Alpha-stable modeling
of noise and robust time-delay estimation in the presence of impulsive
noise,” IEEE transactions on multimedia, vol. 1, no. 3, pp. 291–301,
1999.
[15] S. P. Talebi, S. Werner, and D. P. Mandic, “Distributed adaptive filtering
of α-stable signals,” IEEE Signal Processing Letters, vol. 25, no. 10,
pp. 1450–1454, 2018.
[16] M. Shao and C. L. Nikias, “Signal processing with fractional lower
order moments: stable processes and their applications,” Proceedings of
the IEEE, vol. 81, no. 7, pp. 986–1010, 1993.
[17] G. Aydin, O. Arikan, and A. E. Cetin, “Robust adaptive filtering
algorithms for/spl alpha/-stable random processes,” IEEE Transactions
on Circuits and Systems II: Analog and Digital Signal Processing,
vol. 46, no. 2, pp. 198–202, 1999.
[18] O. A. Y. Ojeda, J. Grajal, and L. Garcia, “Adaptive-fresh filtering,”
Adaptive Filtering, 2011.
[19] S. C. Douglas and T.-Y. Meng, “Stochastic gradient adaptation under
general error criteria,” IEEE transactions on signal processing, vol. 42,
no. 6, pp. 1335–1351, 1994.
[20] Q. Tian, T. Qiu, J. Ma, J. Li, and R. Li, “A simplified doa estimation
method based on correntropy in the presence of impulsive noise,” IEEE
Access, vol. 6, pp. 67010–67025, 2018.