Pith. sign in

REVIEW 3 major objections 5 minor 46 references

A Family of Robust Generalized Adaptive Filters and Application for Time-series Prediction

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read One tunable cost function is claimed to let a single adaptive filter stay competitive as Gaussian, impulsive, heavy-tailed, or asymmetric noise changes.

desk verdict A useful parametric extension of Barron's loss to arbitrary error order, but the kernel-positive-definiteness proof has a gap that the simulations actually lean on. read the letter →

arxiv 2506.00397 v1 pith:PAZ3EQYH submitted 2025-05-31 eess.SP

Yi Peng , Haiquan Zhao , Jinhui Hu This is my paper
classification eess.SP
keywords robustadaptivefiltergeneralizedGaussiannoisecostfunctiondesignkernelfilteringasymmetricChua'scircuittime-seriespredictionrecursivealgorithmmaximumcorrentropycriterion
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Adaptive filters are usually built around one cost function—mean-square error, fourth-order moment, log-square, or correntropy—and each works well only for certain noise. This paper proposes a single parametric cost function, the RGA cost, whose two shape parameters ($\alpha$ and $\beta$) and scale ($\lambda$) interpolate among these classic costs, so one filter can be retuned instead of redesigned when the noise changes. The authors show that in the limits the RGA cost recovers LMS, LMF, LMP, LMLS, RLMLS, MCC, and GMCC, and their simulations report competitive or better steady-state error across six symmetric noise environments, three asymmetric environments, and two chaotic time-series prediction tasks. Two extensions are derived from the same family: an asymmetric version (NARGA) using different scales for positive and negative errors, and a kernel recursive version (KRNRGA) for nonlinear prediction. If the central claim holds, robust filtering no longer requires switching algorithms as noise statistics drift; it only requires adjusting parameters of one cost.

What carries the argument

The load-bearing object is the RGA cost function of Eq. (22), a two-shape-parameter, one-scale-parameter loss built from $(1+|e|^\beta/\lambda^\beta)^{\alpha/\beta}$ with an affine offset. Its role is to make the algorithm's error nonlinearity tunable: $\beta$ sets the effective error moment, $\alpha$ sets the robustness and sensitivity region, and $\lambda$ sets the scale at which errors count as large. The update in Eq. (25) is a stochastic-gradient step on this cost. For the kernel variants, the companion NRGA kernel of Eq. (10) is the same function applied to the error $x-y$; the positive-definiteness claim of Theorem 3 ($0<\beta\le2$) is what licenses the Hilbert-space interpretation, the metric of Theorem 5, and the KRNRGA recursion. The asymmetric NARGA kernel replaces the single scale $\lambda$ with $\lambda_+$ and $\lambda_-$ depending on the sign of the error.

What would settle it

If, in a mixed skewed-plus-impulsive noise environment not listed in the paper, exhaustive tuning of the three RGA parameters at matched initial convergence never brings the filter within a few decibels of the best specialized algorithm's steady-state error, the central adaptability claim would be refuted.

Watch

Extended reading notes

Core claim

The central claim is that the RGA cost function of Eq. (22), constructed by raising $1+|e|^\beta/\lambda^\beta$ to the power $\alpha/\beta$ and adding an affine offset, is flexible enough to reproduce several established adaptive-filter cost functions at special parameter values and to interpolate smoothly among them. In the limits $\alpha\to-\infty$, $\alpha\to 0$, and $\alpha\to\beta$, it becomes the generalized maximum correntropy, robust log-square, and least mean $p$-power costs, respectively; with $\beta=2$, it contains LMS, LMLS, and MCC as special cases. The paper therefore claims that "the robustness of the RGA algorithm can be controlled by tuning the parameters to handle various noise environments." On the kernel side, the negative-RGA (NRGA) kernel is claimed positive definite if and only if $0<\beta\le2$, which makes the NRGA entropy a second-order statistic in a reproducing kernel Hilbert space and induces a metric; the NARGA variant uses separate scales $\lambda_+$ and $\lambda_-$ for positive and negative errors to fit asymmetric distributions; and the KRNRGA algorithm applies the NRGA kernel in a recursive kernel filter for nonlinear time-series prediction. Simulations on linear system identification and on Chua's-circuit prediction are reported to show that the proposed filters "generally outperform other competing algorithms in various noise environments."

Load-bearing premise

The mathematical claim that the NRGA kernel is a genuine kernel for all $0<\beta\le2$ is only demonstrated for the lower half of that range, while the simulations use $\beta=1.56$; the kernel-recursive variant stands or falls on that unproved extension.

Editorial extensions

If this is right

  • A single RGA filter can be deployed where the noise regime changes: retuning $\alpha$, $\beta$, and $\lambda$ replaces swapping the whole algorithm.
  • Because the limiting parameter values recover LMS, LMF, LMP, LMLS, RLMLS, MCC, and GMCC, RGA gives a common testing ground in which all these algorithms appear as one family.
  • If Theorem 3's range $0<\beta\le2$ is correct, the NRGA kernel offers a tunable, correntropy-like metric for robust filtering across that whole range.
  • The NARGA filter's sign-dependent scales make its error nonlinearity asymmetric, and the paper reports that this fits skewed error densities better than symmetric kernel filters.
  • KRNRGA extends the same family into online kernel-recursive learning, giving a nonlinear predictor for chaotic time series; the reported Chua's-circuit experiments support this use.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: the claimed adaptability would become automatic if $\alpha$, $\beta$, and $\lambda$ were tuned online from an estimate of the noise distribution or from a risk criterion; the paper leaves parameter selection manual.
  • Beyond the paper: the asymmetric two-scale construction could be transferred to other correntropy-type costs, for instance by replacing the Gaussian kernel in existing asymmetric correntropy filters, and tested on the same skewed-noise benchmarks.
  • Beyond the paper: a numerical eigenvalue test of the NRGA kernel matrix at $\beta=1.56$ would either close the proof gap in Theorem 3 or bound the usable parameter range, a check the paper does not report.
  • Beyond the paper: the steady-state MSD formula suggests a near-linear dependence on noise variance and step size, which could be inverted into an adaptive step-size rule for the RGA algorithm; the paper does not explore that.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper introduces a parametric cost function, the RGA loss, which contains LMS, LMP, LMLS, MCC, and GMCC as special or limiting cases, and derives three algorithms from it: the stochastic-gradient RGA filter, the asymmetric NARGA filter, and the kernel recursive KRNRGA filter. The authors also provide a convergence analysis and a steady-state MSD formula for RGA, a computational complexity table, and extensive simulations for linear system identification and Chua-circuit time-series prediction under Gaussian, impulsive, Laplace, binary, uniform, generalized-Gaussian, asymmetric, and F-distributed noises. The central claim is that this one family can replace multiple cost functions because its parameters can be tuned to handle different noise environments.

Significance. The RGA family is a natural unification of several robust cost functions, and the stochastic-gradient derivations and steady-state MSD analysis follow standard, internally consistent patterns. The Monte Carlo study with 1000 runs is a strength, and the existence proofs of the limiting cases (LMS, LMP, LMLS, MCC, GMCC) are clearly useful for practitioners. The significance is currently limited by three issues: the key positive-definiteness theorem for the NRGA kernel is proved only on part of the claimed interval, the 'adaptivity' demonstration consists of offline parameter tuning rather than an online adaptation mechanism, and the simulation comparison protocol is under-specified. If the proof gap is closed and the claims are calibrated to what is actually demonstrated, the paper would be a solid contribution to robust adaptive filtering.

major comments (3)
  1. [III.C, Theorem 3] The proof of Theorem 3 is incomplete in exactly the range needed by the later algorithms. The argument invokes [42, p.3 (1.6)] to show that a composition of the form f(lambda u^beta + a) is completely monotone for 0 < beta <= 1, and then concludes positive definiteness for 0 < beta <= 2. No argument is supplied for 1 < beta <= 2, and the 'only if' direction for beta > 2 is not addressed at all. The sentence 'And, F(u) is also c.m. by D.1 and Leibniz formula' does not bridge the gap. This is load-bearing because Theorems 4 and 5 and the KRNRGA derivation in Section IV.C assume a Mercer kernel, while Tables VI and VII use beta = 1.56 for NARGA and KRNRGA. Please provide a complete proof or a precise citation covering the full interval 0 < beta <= 2 and the failure for beta > 2, or restrict the kernel-based claims and the corresponding simulations to the range actually proved.
  2. [Abstract, Section III.A Remark 2, Section VII.B] The central claim that the RGA family can 'smoothly adapt' to varying noise environments is not supported by the experiments as stated. In Tables V-VII the RGA/NARGA/KRNRGA parameters alpha, beta, lambda, lambda_+ and lambda_- are chosen separately for each noise condition, i.e., by offline hand tuning, and no online adaptation rule for these parameters is proposed. There is also no experiment in which the noise distribution changes during the adaptation process; the tracking experiment in Fig. 8 changes only the unknown system, not the noise. The paper should either add an online parameter-selection or adaptation mechanism and test it under a changing noise environment, or explicitly restate the contribution as a flexible parametric family that requires per-environment retuning. As written, 'adaptability' overstates what is demonstrated.
  3. [Section VII.B and Tables V-VI] The comparison protocol is not described sufficiently to support the conclusion that the proposed algorithms 'generally outperform' existing techniques. The text states that parameters are chosen to ensure the same initial convergence rate (Section VII.B), but the matching criterion is not given, nor is any sensitivity analysis with respect to the chosen step sizes and kernel parameters. Tables V and VI list only nominal parameter values, with no indication of how sensitive the conclusions are to those choices. Without this information, the reader cannot distinguish an intrinsic advantage of the proposed cost functions from a favorable parameter selection. Please describe the matching procedure, report results over a range of step sizes, and provide error bars or statistical tests over the 1000 Monte Carlo runs.
minor comments (5)
  1. [III.C, Eqs. (15)-(16)] The statement that 'm is c.m.' is confusing: the function m(u) = lambda u^beta + a is a Bernstein function for 0 < beta <= 1, not a completely monotone function. What is meant is that the composition 1/m(u) is completely monotone. Please rewrite this passage to distinguish the two notions.
  2. [IV.C, Eqs. (30)-(42)] The index notation is inconsistent: L denotes the dictionary size in some places and the matrix dimension in others, while n is used for the current time index. This makes the recursive update, especially Eqs. (38)-(42), harder to follow. Please harmonize the notation.
  3. [Table III and Section VII.C] The KRNRGA experiments do not report the ALD threshold used for dictionary growth, and the role of the regularization parameter gamma in Eq. (41) is not fully specified. These quantities are needed for reproducibility.
  4. [Section V.B] The steady-state MSD derivation in Eqs. (59)-(62) assumes Gaussian output noise and small steady-state errors, but this restriction is stated only implicitly. Please state explicitly that the closed-form MSD formula holds in this regime and is not claimed for general non-Gaussian noise.
  5. [Throughout] Several display equations and tables are visually garbled in the manuscript version reviewed (e.g., Eq. (61), Table IV, Table V). The final version should ensure all symbols and table entries are typeset unambiguously.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the RGA family is defined as a generalization of externally cited losses, recovery of known algorithms is consistency rather than circularity, and the main analysis is not forced by its inputs.

full rationale

The paper's central construction is the RGA cost function (Eq. 22), a parametric generalization of the GA loss of Barron ([38]) and Chebrolu et al. ([39]), which are external citations. The limit cases in Section III.A (Eqs. 5-8) show that the RGA cost reduces to LMP/LMF/LMS, RLMLS/LMLS, GMCC/MCC, and an exponential form; this is consistency with known algorithms, not circularity, because the RGA cost is defined independently and then shown to contain these cases. The steady-state MSD in Section V.B is derived from the proposed update and the stated stochastic assumptions, then verified numerically in Fig. 14; it is not a fitted parameter renamed as a prediction. The self-citations ([25], [44]) are used only for contextual statements about generalized correntropy and standard convergence assumptions; they do not carry the load-bearing claim. The proof of Theorem 3 is incomplete as a mathematical argument (it establishes complete monotonicity only for 0<β≤1 while the theorem asserts 0<β≤2 and simulations use β=1.56), but an incomplete proof is a correctness risk, not a circularity: the theorem is not assumed as an input and the claim does not reduce to the paper's own definitions by construction. No step in the derivation chain is equivalent to its own inputs, and no fitted quantity is presented as an independent prediction. Therefore the appropriate finding is no significant circularity.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The central algorithm depends on hand-tuned shape and scale parameters (α, β, λ, and step sizes), and the theoretical claims rely on standard adaptive-filter independence assumptions plus an incompletely proven kernel positive-definiteness assertion. No new physical entities are introduced.

free parameters (7)
  • α (RGA/NRGA shape parameter) = e.g., -100, 2, 1000; per noise in Tables V-VII
    Controls the tail shape of the RGA loss; hand-selected for each noise environment, so the claimed adaptivity is achieved by manual parameter choice.
  • β (RGA/NRGA shape parameter) = e.g., 1.5, 2.1, 6, 8; per noise in Tables V-VII
    Sets the error moment order; hand-selected per experiment.
  • λ (RGA scale parameter) = e.g., 0.01, 0.1, 1.1 in Tables V-VII
    Tradeoff between convergence speed and steady-state error (Fig. 6); tuned per noise.
  • λ_+ and λ_- (NARGA asymmetric scale parameters) = e.g., λ_+=0.061, λ_-=0.6 in Table VI
    Separate scales for positive and negative errors, chosen per asymmetric noise distribution.
  • Step size μ (or η) = 0.0027 to 0.5 depending on algorithm and noise (Tables V-VII)
    Determines convergence and stability; tuned per algorithm, with the paper stating parameters are chosen for equal initial convergence rates.
  • γ (KRNRGA regularization) = 1e-10 (Table VII)
    Regularization scale in the kernel recursive least-squares cost (Eq. 30).
  • ALD threshold for dictionary growth = not specified
    Controls when new kernel dictionary elements are added; the paper mentions the ALD criterion [34] but gives no value, making exact reproduction impossible.
assumptions (4)
  • domain assumption The weight error vector is uncorrelated with the desired signal and input (Assumption 1, Section V.A).
    Standard independence assumption used to derive mean convergence (Eqs. 44-48) and steady-state MSD; approximate in practice.
  • domain assumption Input and desired signal are zero-mean jointly Gaussian (Assumption 2, Section V.A).
    Used with Lemma 1 of [45] to simplify E[f(e_i) x_i]; restricts the theoretical analysis to Gaussian inputs.
  • ad hoc to paper The NRGA kernel is positive definite for 0<β≤2 (Theorem 3).
    The provided proof only demonstrates complete monotonicity for 0<β≤1; the extension to β≤2 is asserted, not shown. KRNRGA and the metric theorem rely on this.
  • ad hoc to paper For every noise environment there exist RGA parameter values that achieve optimal or superior performance.
    The paper's central claim presumes such parameters exist and can be found; no selection rule or sensitivity analysis is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Family of Robust Generalized Adaptive Filters and Application for Time-series Prediction." pith.science (2026). https://pith.science/paper/PAZ3EQYH

@misc{pith2026250600397,
  author       = {Pith},
  title        = {Pith review of: A Family of Robust Generalized Adaptive Filters and Application for Time-series Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PAZ3EQYH}},
  note         = {Machine review of arXiv:2506.00397}
}
read the original abstract

The continuous development of new adaptive filters (AFs) based on novel cost functions (CFs) is driven by the demands of various application scenarios and noise environments. However, these algorithms typically demonstrate optimal performance only in specific conditions. In the event of the noise change, the performance of these AFs often declines, rendering simple parameter adjustments ineffective. Instead, a modification of the CF is necessary. To address this issue, the robust generalized adaptive AF (RGA-AF) with strong adaptability and flexibility is proposed in this paper. The flexibility of the RGA-AF's CF allows for smooth adaptation to varying noise environments through parameter adjustments, ensuring optimal filtering performance in diverse scenarios. Moreover, we introduce several fundamental properties of negative RGA (NRGA) entropy and present the negative asymmetric RGA-AF (NAR-GA-AF) and kernel recursive NRGA-AF (KRNRGA-AF). These AFs address asymmetric noise distribution and nonlinear filtering issues, respectively. Simulations of linear system identification and time-series prediction for Chua's circuit under different noise environments demonstrate the superiority of the proposed algorithms in comparison to existing techniques.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 46 canonical work pages

  1. [1]

    Haykin, Adaptive Filter Theory

    S.S. Haykin, Adaptive Filter Theory. Pearson Education India, 2008

    work page 2008

  2. [2]

    A variable step size LMS algo- rithm,

    R. H. Kwong and E. W. Johnston, “A variable step size LMS algo- rithm,” IEEE Transactions on Signal Processing , vol. 40, no. 7, pp. 1633-1642, Jul. 1992

    work page 1992

  3. [3]

    On the convergence behavior of the LMS and the normalized LMS algorithms,

    D. T. M. Slock, “On the convergence behavior of the LMS and the normalized LMS algorithms,” IEEE Transactions on Signal Processing, vol. 41, no. 9, pp. 2811-2825, Sept. 1993

    work page 1993

  4. [4]

    Fundamentals of adaptive filtering,

    Sayed and H. Ali, “Fundamentals of adaptive filtering,” IEEE Control Syst., vol. 25, no. 4, pp. 77–79, 2003

    work page 2003

  5. [5]

    Robust Widely Linear Affine Projection M-Estimate Adaptive Algorithm: Performance Analysis and Applica- tion,

    S. Lv, H. Zhao and W. Xu, “Robust Widely Linear Affine Projection M-Estimate Adaptive Algorithm: Performance Analysis and Applica- tion,” IEEE Transactions on Signal Processing, vol. 71, pp. 3623-3636, 2023

    work page 2023

  6. [6]

    Lawson -Norm-Based Adaptive Filter for Channel Estimation and In -Car Echo Cancellation,

    X. Huang, Y. Li, X. Han, and H. Tu, “Lawson -Norm-Based Adaptive Filter for Channel Estimation and In -Car Echo Cancellation,” IEEE Transactions on Circuits and Systems II: Express Briefs , vol. 69, no. 4, pp. 2376–2380, Apr. 2022

    work page 2022

  7. [7]

    The least mean fourth (LMF) adaptive algorithm and its family,

    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

    work page 1984

  8. [8]

    Least mean p -power error criterion for adaptive FIR filter,

    S. C. Pei and C. C. Tseng, “Least mean p -power error criterion for adaptive FIR filter,” IEEE Journal on Selected Areas in Communica- tions, vol. 12, no. 9, pp. 1540–1547, Dec. 1994

    work page 1994

Show all 46 references
  1. [9]

    Improved convergence analysis of sto- chastic gradient adaptive filters using the sign algorithm,

    V. Mathews and Sung Cho, “Improved convergence analysis of sto- chastic gradient adaptive filters using the sign algorithm, ” IEEE Trans- actions on Acoustics, Speech, and Signal Processing, vol. 35, no. 4, pp. 450-454, Apr. 1987

    1987

  2. [10]

    Steady -State Mean-Square Performance Analy- sis of the Affine Projection Sign Algorithm,

    Z. Zheng and Z. Liu, “Steady -State Mean-Square Performance Analy- sis of the Affine Projection Sign Algorithm,” IEEE Transactions on Circuits and Systems II: Express Briefs, vol. 67, no. 10, pp. 2244–2248, Oct. 2020

    2020

  3. [11]

    A Novel Family of Adap- tive Filtering Algorithms Based on the Logarithmic Cost,

    M. O. Sayin, N. D. Vanli and S. S. Kozat, “A Novel Family of Adap- tive Filtering Algorithms Based on the Logarithmic Cost, ” IEEE Transactions on Signal Processing , vol. 62, no. 17, pp. 4411 -4424, Sept., 2014

    2014

  4. [12]

    Robust least mean logarithmic square adaptive filtering algorithms,

    K. Xiong and S. Wang, “Robust least mean logarithmic square adaptive filtering algorithms,” Journal of the Franklin Institute , vol. 356, no. 1, pp. 654–674, Jan. 2019

    2019

  5. [13]

    Least mean M -estimate algorithms for robust adaptive filtering in impulse noise,

    Y. Zou, S. C. Chan, and T. S. Ng, “Least mean M -estimate algorithms for robust adaptive filtering in impulse noise,” IEEE Transactions on Circuits and Systems II: Analog and Digital Signal Processing, vol. 47, no. 12, pp. 1564–1569, Dec. 2000

    2000

  6. [14]

    Bias -Compensated Sparsity -Aware NLMM Algorithms for Robust Adaptive Echo Cancellation,

    Z. Zheng, Z. Liu, and J. Zhao, “Bias -Compensated Sparsity -Aware NLMM Algorithms for Robust Adaptive Echo Cancellation,” IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 67, no. 7, pp. 2383–2396, Jul. 2020

    2020

  7. [15]

    A robust mixed -norm adaptive filter algorithm,

    J. Chambers and A. Avlonitis, “A robust mixed -norm adaptive filter algorithm,” IEEE Signal Processing Letters , vol. 4, no. 2, pp. 46 -48, Feb. 1997

    1997

  8. [16]

    Continuous Mixed p-Norm Adaptive Algorithm for Sys- tem Identification,

    H. Zayyani, “Continuous Mixed p-Norm Adaptive Algorithm for Sys- tem Identification,” IEEE Signal Processing Letters, vol. 21, no. 9, pp. 1108-1110, Sept. 2014

    2014

  9. [17]

    A Robust Stable Laplace Con- tinuous Mixed Norm Adaptive Filter Algorithm,

    H. Zayyani, M. Korki and A. Taghavi, “A Robust Stable Laplace Con- tinuous Mixed Norm Adaptive Filter Algorithm,” IEEE Sensors Letters, vol. 8, no. 3, pp. 1-4, March 2024

    2024

  10. [18]

    Correntropy: Properties and Applications in Non -Gaussian Signal Processing,

    W. Liu, P. P. Pokharel and J. C. Principe, “Correntropy: Properties and Applications in Non -Gaussian Signal Processing, ” IEEE Transactions on Signal Processing, vol. 55, no. 11, pp. 5286-5298, Nov. 2007

    2007

  11. [19]

    Kernel recursive max- imum correntropy,

    Z. Wu, J. Shi, X. Zhang, W. Ma, and B. Chen, “Kernel recursive max- imum correntropy,” Signal Processing, vol. 117, pp. 11-16, Dec. 2015

    2015

  12. [20]

    Ran- dom Fourier Filters Under Maximum Correntropy Criterion,

    S. Wang, L. Dang, B. Chen, S. Duan, L. Wang, and C. K. Tse, “Ran- dom Fourier Filters Under Maximum Correntropy Criterion,” IEEE Transactions on Circuits and Systems I: Regular Papers , vol. 65, no. 10, pp. 3390–3403, Oct. 2018

    2018

  13. [21]

    Correntropy - Based Data Selective Adaptive Filtering,

    Y.-R. Chien, S.-T. Wu, H.-W. Tsao, and P. S. R. Diniz, “Correntropy - Based Data Selective Adaptive Filtering,” IEEE Transactions on Cir- cuits and Systems I: Regular Papers , vol. 71, no. 2, pp. 754 –766, Feb. 2024

    2024

  14. [22]

    A Robust Markovian Block Sparse Adaptive Algorithm with Its Convergence Analysis,

    Z. Habibi, H. Zayyani and M. Korki, “A Robust Markovian Block Sparse Adaptive Algorithm with Its Convergence Analysis, ” IEEE Transactions on Circuits and Systems II: Express Briefs , vol. 71, no. 3, pp. 1546-1550, March 2024

    2024

  15. [23]

    Generalized Correntropy for Robust Adaptive Filtering,

    B. Chen, L. Xing, H. Zhao, N. Zheng and J. C. Prı´ncipe, “Generalized Correntropy for Robust Adaptive Filtering,” IEEE Transactions on Signal Processing, vol. 64, no. 13, pp. 3376-3387, Jul., 2016

    2016

  16. [24]

    Parametric generalized Gaussian density estimation,

    M. K. Varanasi and B. Aazhang, “Parametric generalized Gaussian density estimation,” The Journal of the Acoustical Society of America , vol. 86, no. 4, pp. 1404-1415, Oct. 1989

    1989

  17. [25]

    Robust Generalized Maxi- mum Correntropy Criterion Algorithms for Active Noise Control,

    Y. Zhu, H. Zhao, X. Zeng and B. Chen, “Robust Generalized Maxi- mum Correntropy Criterion Algorithms for Active Noise Control, ” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 28, pp. 1282-1292, 2020

    2020

  18. [26]

    A Novel Family of Sparsity-Aware Robust Adaptive Filters Based on a Logistic Distance Metric,

    K. Kumar, M. L. N. S. Karthik and N. V. George, “A Novel Family of Sparsity-Aware Robust Adaptive Filters Based on a Logistic Distance Metric,” IEEE Transactions on Signal Processing , vol. 70, pp. 6128 - 6141, 2022

    2022

  19. [27]

    Generalized Modi- fied Blake–Zisserman Robust Sparse Adaptive Filters,

    K. Kumar, M. L. N. S. Karthik, and N. V. George, “Generalized Modi- fied Blake–Zisserman Robust Sparse Adaptive Filters,” IEEE Transac- tions on Systems, Man, and Cybernetics: Systems , vol. 53, no. 1, pp. 647–652, Jan. 2023

    2023

  20. [28]

    Asymmetric Gaussian and its appli- cation to pattern recognition

    T. Kato, S. Omachi, and H. Aso, “Asymmetric Gaussian and its appli- cation to pattern recognition” Proc. Joint IAPR Int. Workshops Stat. Techn. Pattern Recognit. (SPR) Struct. Syntactic Pattern Recognit. (SSPR), 2002, pp. 405-413

    2002

  21. [29]

    Robust Localization Using IMM Filter Based on Skew Gaussian -Gamma Mixture Distribution in Mixed LOS/NLOS Condition,

    W. Youn, Y. Huang and H. Myung, “Robust Localization Using IMM Filter Based on Skew Gaussian -Gamma Mixture Distribution in Mixed LOS/NLOS Condition, ” IEEE Transactions on Instrumentation and Measurement, vol. 69, no. 7, pp. 5166-5182, Jul. 2020

    2020

  22. [30]

    Asymmetric Correntropy for Robust Adaptive Filtering,

    B. Chen, Y. Xie, Z. Li, Y. Li and P. Ren, “Asymmetric Correntropy for Robust Adaptive Filtering, ” IEEE Transactions on Circuits and Sys- tems II: Express Briefs, vol. 69, no. 3, pp. 1922-1926, Mar. 2022

    1922

  23. [31]

    Generalized Asymmetric Correntropy for Robust Adaptive Filtering: A Theoretical and Simulation Study,

    H. Qu, M. Wang, J. Zhao, S. Zhao, T. Li, and P. Yue, “Generalized Asymmetric Correntropy for Robust Adaptive Filtering: A Theoretical and Simulation Study,” Remote Sensing, vol. 14, no. 15, Art. no. 15, Jan. 2022

    2022

  24. [32]

    A Survey on Time -Series Pre -Trained Models,

    Q. Ma et al., “A Survey on Time -Series Pre -Trained Models,” IEEE Transactions on Knowledge and Data Engineering, vol. 36, no. 12, pp. 7536-7555, Dec. 2024

    2024

  25. [33]

    Support Vector Echo -State Machine for Chaotic Time-Series Prediction,

    Z. Shi and M. Han, “Support Vector Echo -State Machine for Chaotic Time-Series Prediction,” IEEE Transactions on Neural Networks , vol. 18, no. 2, pp. 359-372, March 2007

    2007

  26. [34]

    The kernel recursive least -squares algorithm,

    Y. Engel, S. Mannor and R. Meir, “The kernel recursive least -squares algorithm,” IEEE Transactions on Signal Processing, vol. 52, no. 8, pp. 2275-2285, Aug. 2004

    2004

  27. [35]

    Interval Type-2 Fuzzy Neural Networks for Chaotic Time Series Prediction: A Concise Overview,

    M. Han, K. Zhong, T. Qiu and B. Han, “Interval Type-2 Fuzzy Neural Networks for Chaotic Time Series Prediction: A Concise Overview, ” IEEE Transactions on Cybernetics , vol. 49, no. 7, pp. 2720 -2731, July 2019

    2019

  28. [36]

    Kernel Recursive Generalized Maximum Cor- rentropy,

    J. Zhao and H. Zhang, “Kernel Recursive Generalized Maximum Cor- rentropy,” IEEE Signal Processing Letters , vol. 24, no. 12, pp. 1832 - 1836, Dec. 2017

    2017

  29. [37]

    Gauss Hermite Fourier Features Based on Maximum Correntropy Criterion for Adaptive Filtering,

    T. Zhou, S. Wang, J. Qian, Y. Zheng, Q. Zhang and Y. Xiao, “Gauss Hermite Fourier Features Based on Maximum Correntropy Criterion for Adaptive Filtering,” IEEE Transactions on Circuits and Systems I: Regular Papers, doi: 10.1109/TCSI.2024.3476150

  30. [38]

    A General and Adaptive Robust Loss Function,

    J. T. Barron, “A General and Adaptive Robust Loss Function,” 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Long Beach, CA, USA, 2019

    2019

  31. [39]

    Adap- tive Robust Kernels for Non -Linear Least Squares Problems,

    N. Chebrolu, T. Lä be, O. Vysotska, J. Behley and C. Stachniss, “Adap- tive Robust Kernels for Non -Linear Least Squares Problems, ” IEEE Robotics and Automation Letters , vol. 6, no. 2, pp. 2240 -2247, Apr . 2021

    2021

  32. [40]

    Generalized correla- tion function: definition, properties, and application to blind equaliza- tion,

    I. Santamaria, P. P. Pokharel and J. C. Principe, “Generalized correla- tion function: definition, properties, and application to blind equaliza- tion,” IEEE Transactions on Signal Processing , vol. 54, no. 6, pp. 2187-2197, Jun. 2006

    2006

  33. [41]

    V. N. Vapnik and V. Vapnik , Statistical Learning Theory . New York, NY, USA: Wiley, 1998, vol. 1

    1998

  34. [42]

    Completely monotonic functions,

    K. S. Miller and S. G. Samko, “Completely monotonic functions,” Integral Transforms and Special Functions, Dec. 2001

    2001

  35. [43]

    A family of robust adaptive filter- ing algorithms based on sigmoid cost,

    F. Huang, J. Zhang, and S. Zhang, “A family of robust adaptive filter- ing algorithms based on sigmoid cost,” Signal Process., vol. 149, pp. 179–192, Aug. 2018

    2018

  36. [44]

    Robust Variable Step -Size Re- weighted Zero -Attracting Least Mean M -Estimate Algorithm for Sparse System Identification,

    G. Wang, H. Zhao, and P. Song, “Robust Variable Step -Size Re- weighted Zero -Attracting Least Mean M -Estimate Algorithm for Sparse System Identification,” IEEE Transactions on Circuits and Sys- tems II: Express Briefs, vol. 67, no. 6, pp. 1149–1153, Jun. 2020

    2020

  37. [45]

    Robust adaptive filter with lncosh cost,

    C. Liu and M. Jiang, “Robust adaptive filter with lncosh cost,” Signal Process., vol. 168, Mar. 2020, Art. no. 107348

    2020

  38. [46]

    Adaptive filters with error nonlin- earities: Mean-square analysis and optimum design,

    T. Y. Al-Naffouri and A. H. Sayed, “Adaptive filters with error nonlin- earities: Mean-square analysis and optimum design,” EURASIP J. Adv. Signal Process., vol. 2001, no. 1, pp. 192–205, 2001

    2001

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.