{"id":"34809a37-fd98-44c1-acf6-6c7252efd032","arxiv_id":"1908.08165","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"The paper derives an optimal step-size for the least-mean-fourth adaptive filter by minimizing mean-square deviation at each iteration, reporting large steady-state gains over NLMF and VSSLMFQ in low-SNR simulations.","lead":"The paper derives a step-size rule for the fourth-power least-mean (LMF) adaptive filter that minimizes the mean-square deviation at every iteration. It reports much lower steady-state error than two existing LMF variants in low-SNR simulations with Gaussian and non-Gaussian noises, though the rule depends on knowing the very error it is minimizing.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"OPLMF's step-size (Eq. 29) requires MSD(n), the unknown system's true deviation; Table 1 propagates it from W_O, so the claimed superiority is an oracle result, not an implementable algorithm.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the algorithm requires knowledge of MSD(n), which depends on the unknown system vector W_O. This is not a minor implementation detail; it is the input to the step-size formula at every iteration. The derivation of Eq. (29) is a legitimate one-step optimization of the approximate recursion (18), and under the oracle assumption the algebra is internally coherent. But the abstract and conclusions claim a practical algorithm that is superior to existing methods at low SNR. The reported simulations use W_O both to generate the MSD schedule and to measure performance, so the comparison is favorable by construction. A real system identifier would have to estimate MSD and the noise moments from observed signals, which the paper does not describe. For these reasons I agree with the reader's REJECT verdict; my stress-test does not change it. I credit the authors for deriving the recursion explicitly and for testing multiple noise distributions, but the missing implementability step is decisive for the central practical claim.","tokens_in":12631,"tokens_out":7285,"duration_ms":76116,"concrete_test":"Re-run Experiment 4 (time-varying system, SNR=1 dB, Gaussian noise, uncorrelated input) with an implementable MSD estimate: replace MSD(n) in Eq. (29) by MSD_hat(n) = (e_hat^2(n) - sigma_rho_hat^2)/sigma_x_hat^2, where e_hat^2(n) is a smoothed squared-error estimate and sigma_rho_hat^2 is estimated from the error signal. Keep all other settings identical and compare the final MSD with the reported -44.37 dB and with NLMF/VSSLMFQ. If the implementable version cannot reproduce the reported floor, the headline superiority depends on the oracle MSD.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on Eq. (29) and Table 1, where the step-size is computed from MSD(n)=E[||W_O(n)-W(n)||^2]. This is the very quantity an adaptive filter in system identification is supposed to learn; W_O is unknown. Table 1 initializes the recursion with MSD(0)=||W_O||^2 (since W(0)=0) and then propagates MSD via the approximate recursion Eq. (18), which additionally requires sigma_rho^2, E[rho^4], and E[rho^6]. No estimator for MSD or those noise moments is supplied. In every simulation in Section 4, the known W_O is used both to drive the MSD schedule and to evaluate the MSD curves, so the claimed 20-60 dB improvement over NLMF and VSSLMFQ is conditional on information a real implementation would not possess. Furthermore, the zero steady-state result MSD(infinity)=0 is derived from that same approximate recursion rather than from an implementable recursion with an estimated MSD. The mathematical construction may be coherent under oracle assumptions, but it does not establish a practical algorithm.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an optimized least mean absolute fourth (OPLMF) algorithm for adaptive system identification at low signal-to-noise ratio. The step-size is chosen at each iteration to minimize a mean-square deviation (MSD) recursion, Eq. (18), leading to the formula in Eq. (29). The authors claim that the resulting steady-state MSD is zero, assert this via Eq. (31), and support the claim with seven system-identification experiments comparing OPLMF with NLMF and VSSLMFQ under Gaussian, Uniform, Binary, Rayleigh, and Poisson noises. The paper also includes a computational-complexity comparison and a brief stability discussion.","tokens_in":12902,"tokens_out":4885,"duration_ms":47740,"significance":"A step-size rule that optimally trades convergence speed against steady-state misalignment for a high-order error-power filter could be a useful contribution, especially for low-SNR and non-Gaussian-noise regimes. The paper covers a relevant problem class and provides a broad simulation study, and the complexity comparison in Table 2 is a positive feature. However, the central construction depends on the true weight-error MSD, which is not available in practice, and the theoretical derivation relies on heuristic moment closures without error control. As it stands, the paper demonstrates an oracle-assisted design rather than an implementable algorithm, so the claimed practical superiority over NLMF and VSSLMFQ is not established.","major_comments":[{"comment":"The step-size formula uses MSD(n) = E[||W_O(n) - W(n)||^2], the true weight-error norm of the unknown system. Table 1 updates MSD(n) with the model recursion, but this recursion must be initialized; with W(0) = 0, the required initial value is MSD(0) = ||W_O||^2, which depends on the unknown system. No data-driven estimator for MSD or for the noise moments E[rho^4] and E[rho^6] is supplied. In every Section 4 experiment, the known W_O is used both to schedule mu(n) and to evaluate the MSD curves, so the reported 20-60 dB improvements over NLMF and VSSLMFQ are conditional on information a real system identifier does not possess.","section":"Section 2, Eq. (29) and Table 1"},{"comment":"The core recursion Eq. (18) is obtained through ad-hoc Gaussian input moment closures, for example E[(X^T V)^4] is replaced by 6 sigma_x^4 (E[V^T V])^2 and higher-order terms are factorized as E[(V^T V)^k] approximately equal to (E[V^T V])^k. The manuscript gives no error bound or supporting argument for these closures, and the input is assumed Gaussian while the algorithm is then optimized for non-Gaussian noises. The claimed agreement between simulation and theory in Figs. 1(B)-7(B) is not independent validation, because the theory curve is generated from the same approximate recursion that defines the step-size.","section":"Section 2, Eqs. (12)-(18)"},{"comment":"The conclusion MSD(infinity) = 0 is asserted after 'several computations' and 'Combine Eq. (23)', but the intermediate algebra is not shown. Eq. (31) has the structure MSD(infinity)^2 times a bracket equals zero, so it also admits a nonzero root; the manuscript does not prove that the bracket cannot vanish or that the zero root is the fixed point actually attained. Because the step-size depends on MSD(n), the fixed-point analysis must account for this coupling explicitly, which the paper does not do.","section":"Section 2, Eq. (31)"},{"comment":"The minimization leading to the 'optimal' step-size sets partial MSD(n)/partial mu = 0, treating the current MSD as independent of the step-size. In reality MSD(n) depends on all previous step-sizes, so this is a heuristic approximation. The manuscript does not state it as such or assess its effect on the optimality claim, and therefore the label 'optimal' is not formally justified.","section":"Section 2, Eqs. (26)-(29)"}],"minor_comments":[{"comment":"There are multiple typographical errors and inconsistencies: 'read' should be 'red' in figure captions, 'discription' should be 'description', 'Passion' should be 'Poisson', and 'OPLMAT' is used instead of 'OPLMF' in several places.","section":"Throughout"},{"comment":"Equation numbering is confusing: Eq. (19) is first defined as g(L, mu, sigma_x) and later reused for the reduced first-order recursion, and Eq. (22) repeats a formula already derived. Please renumber consistently.","section":"Section 2, Eq. (22) and Eq. (19)"},{"comment":"The text says gamma is a small positive number to keep the denominator finite, but then states (1 - 1/(2L)) <= gamma < 1, which is not derived and seems inconsistent with 'small'; the range should be justified or corrected.","section":"Section 2, Eq. (23)"},{"comment":"The correlated input is defined by y(n) = 0.5 y(n) + x(n), which is self-referential and presumably should be y(n) = 0.5 y(n-1) + x(n); please correct and define the filtering operation precisely.","section":"Section 4"},{"comment":"In Experiment 2 the VSSLMFQ result is described as divergent, but Table 3 does not show an entry; please clarify whether the algorithm diverged and how the table was compiled in that case.","section":"Table 3"},{"comment":"The 'MSD error' is described as the difference between simulation and Theory (Eq. (30)), but Eq. (30) is a recursion and not a closed-form curve; please specify exactly how the theoretical MSD trajectory is computed in the figures.","section":"Figures 1(B)-7(B)"},{"comment":"The notation mu_f,n is used in Eq. (29), while Eqs. (30) and (31) use mu_f,infinity without a definition; avoid switching subscripts without explanation.","section":"Eqs. (29)-(31)"}],"recommendation":"reject","confidential_remarks":"The paper appears under-polished, with many typos and notation inconsistencies, but the main obstacle is the oracle dependence of the proposed algorithm. If the authors could provide an implementable estimator for MSD and the noise moments and redo the simulations without using the true W_O, a revised manuscript might be worth reconsidering. In the current form, I do not see how the central practical claim can be salvaged without substantial new material, so I recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a direct extension of Guan and Li's optimized LMAT work to the LMF cost. What is genuinely new is the explicit step-size formula in Eq. (29), the MSD recursion Eq. (18), and the worked noise-moment values for five distributions. The computational complexity comparison is also useful. That is real content, and the authors deserve credit for spelling the recursion out rather than hand-waving. The math is internally coherent under the stated Gaussian-input moment closures.\n\nThe soft spot is load-bearing, not cosmetic. The step-size formula depends on MSD(n), which is the expected squared norm of the true weight-error vector. Table 1 initializes MSD(0) with ||W_O||^2, i.e. the unknown system's norm, and then propagates MSD through the approximate recursion. Every simulation in Section 4 uses the known W_O both to schedule the step-size and to evaluate the resulting MSD. So the claimed superiority over NLMF and VSSLMFQ is conditional on oracle information. No estimator for MSD or the noise moments is supplied or evaluated. This makes the central practical claim an open problem rather than a solved one.\n\nThe zero steady-state MSD result in Eq. (31) is asserted from the same approximate recursion without showing the intermediate algebra. That is a smaller issue in isolation, but it matters because the recursion itself uses unverified moment closures (Eqs. 12–16) with no error control. The paper also has many typos and OCR artifacts that make parts of the derivation hard to follow; I would not want to referee it without a clean TeX source.\n\nThat said, the oracle flaw is not fatal to the mathematical idea. If the authors can supply a practical MSD estimator, or clearly frame the contribution as a 'genie-aided' performance bound, the paper could be a reasonable contribution to the adaptive-filtering literature. As written, the evidence does not support the claimed practical gains.\n\nWho is this for? Specialists in LMF-family algorithms and variable step-size design. A serious editor should send it to peer review, because the derivation is substantive enough to warrant expert scrutiny and the oracle issue is a correctable framing problem rather than a trivial mistake. But I would not cite it in its current form.","headline":"The OPLMF step-size rule is an oracle method that requires the very MSD it is supposed to minimize, so the claimed 20–60 dB gains are not established for real implementations; the derivation is a legitimate but flawed extension of the authors' prior LMAT work.","tokens_in":13420,"tokens_out":1428,"would_cite":false,"duration_ms":16817,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A12"],"pacs":[],"model":"deepseek-v4-flash","headline":"A time-updated, MSD-minimizing step size makes the least mean fourth adaptive filter converge to zero steady-state misalignment, even at low SNR and under non-Gaussian noise.","keywords":["adaptive filtering","least mean fourth","optimal step-size","mean-square deviation","low signal-to-noise ratio","non-Gaussian noise","system identification","time-varying systems"],"falsifier":"Run OPLMF on the paper's low-SNR identification tasks while estimating $\\mathrm{MSD}(n)$ online from the filter's error signal instead of using the true coefficient error $\\|\\mathbf{W}_O(n)-\\mathbf{W}(n)\\|^2$; if the measured steady-state MSD then remains clearly above zero, or the filter diverges, the claimed zero steady-state MSD holds only when an oracle supplies the state.","tokens_in":12406,"feed_emoji":"📉","tokens_out":8319,"duration_ms":78198,"temperature":0.7,"pith_summary":"The paper proposes an optimized least mean absolute fourth (OPLMF) adaptive filter for system identification in low signal-to-noise-ratio conditions. At each iteration the step-size is chosen to minimize the next mean-square deviation, producing the update $\\mu(n) = 3\\,\\mathrm{MSD}(n)(\\sigma_\\rho^2 + \\sigma_x^2 \\mathrm{MSD}(n))/\\mu_{f,n}$. The paper derives the mean convergence and steady-state analysis from a recursive MSD model and shows that the steady-state MSD becomes zero, so the excess mean-square error reduces to the noise variance $\\sigma_\\rho^2$. If correct, this gives a low-complexity LMF variant that tracks time-varying systems under Gaussian, Uniform, Binary, Rayleigh, and Poisson noises better than the compared normalized and quotient-form variable step-size LMF algorithms.","feed_headline":"Optimal step size drives LMF error to zero at low SNR","feed_subtitle":"A per-iteration MSD-minimizing step size beats variable-step LMF baselines in low-SNR tests.","key_machinery":"The load-bearing object is the time-varying MSD recursion of Eq. (18), a cubic polynomial in $\\mathrm{MSD}(n)$ whose coefficients depend on filter length $L$, input variance $\\sigma_x^2$, step-size $\\mu$, and the noise moments $\\mathbb{E}[\\rho^4(n)]$ and $\\mathbb{E}[\\rho^6(n)]$. The optimal step-size is obtained by setting the partial derivative $\\partial \\mathrm{MSD}(n+1)/\\partial \\mu = 0$ under the assumption that $\\mathrm{MSD}(n)$ is known, which produces Eq. (29). The recursion is also used to derive the stability bound $0 < \\mu(n) < \\sigma_\\rho^2/(5(L+2)\\sigma_x^2 \\mathbb{E}[\\rho^4(n)])$ and to propagate the predicted MSD in Table 1. The paper uses the known moment formulas for Gaussian, Uniform, Binary, Rayleigh, and Poisson noises to instantiate the fourth and sixth noise moments.","core_discovery":"The central claim is that the step-size, rather than being a fixed tuning constant, can be treated as a state-dependent control that makes the LMF algorithm's misalignment vanish. Starting from the coefficient-error recursion $\\mathbf{V}(n+1)=\\mathbf{V}(n)+\\mu \\mathbf{X}(n)e(n)^3$, the paper derives a scalar recursion for $\\mathrm{MSD}(n)=\\mathbb{E}[\\mathbf{V}^T(n)\\mathbf{V}(n)]$ by approximating the fourth-, sixth-, and higher-order moments of the Gaussian input. Minimizing $\\mathrm{MSD}(n+1)$ with respect to $\\mu$ at every time $n$ yields Eq. (29), and inserting this step-size back into the recursion leaves an equation whose only steady-state solution is $\\mathrm{MSD}(\\infty)=0$. The companion statement is that the steady-state EMSE equals the additive noise variance, $\\mathrm{EMSE}(\\infty)=\\sigma_\\rho^2$, meaning the adaptive filter contributes no excess error. The paper validates this by seven Monte Carlo system-identification experiments at SNR levels from 0 to 3 dB, reporting steady-state MSD values far below NLMF and VSSLMFQ baselines.","pith_inferences":["A practical implementation still needs a running estimate of $\\mathrm{MSD}(n)$; because the paper's simulations supply the true coefficient error, a real identifier would have to estimate MSD from data, and the zero steady-state result would likely degrade to a small nonzero floor.","The same 'minimize next MSD' construction transfers naturally to other error-power cost functions, such as least mean absolute third or mixed-norm filters, where analogous closed-form step-sizes may exist.","A testable extension is to replace the known noise moments by online estimates (sample fourth and sixth moments of the error); if those estimates are accurate, OPLMF should track time-varying systems at even lower SNR without oracle knowledge."],"forward_implications":["Steady-state misalignment vanishes: in the model, $\\mathrm{MSD}(\\infty)=0$, so the adaptive filter's excess error disappears and only the ambient noise variance remains.","The same step-size rule applies across noise types; only the fourth and sixth noise moments in Eq. (29) change, so no redesign is needed for Gaussian, Uniform, Binary, Rayleigh, or Poisson noise.","The algorithm needs no divisions in its update and about $2L+16$ multiplications per iteration, making it cheaper than the VSSLMFQ baseline while giving lower steady-state MSD.","The automatically large initial step-size followed by a shrinking step-size provides fast initial convergence and fine tracking in slowly time-varying systems."],"supporting_citations":[{"why":"Establishes the least mean fourth algorithm and its stability trade-offs, which this paper extends with an optimized step-size.","marker":"[6]"},{"why":"Introduces the practice of choosing an approximately optimal step-size by minimizing MSD, the design principle adopted here.","marker":"[8]"},{"why":"Provides the optimized least mean absolute third algorithm whose MSD-minimization derivation is adapted to the LMF cost.","marker":"[9]"},{"why":"Supplies the normalized LMF baseline and its complexity count used for comparison in Table 2 and the simulations.","marker":"[13]"},{"why":"Gives the sufficient stability condition for the LMF recursion that the paper uses to bound the step-size in Eq. (25).","marker":"[19]"},{"why":"Presents the VSSLMFQ quotient-form variable step-size algorithm used as the main baseline in the simulations.","marker":"[24]"},{"why":"Provides the nonparametric variance estimator used to update the input variance estimate and the gamma range in the algorithm.","marker":"[25]"},{"why":"Gives the fourth and sixth moment formulas for the Gaussian and non-Gaussian noise distributions used to instantiate Eq. (29).","marker":"[26]"}],"fun_headline_variants":["Optimal step-size zeroes LMF steady-state MSD at low SNR","Per-iteration MSD-minimizing step-size reaches zero error","Low-SNR LMF step-size control eliminates misalignment","Adaptive step-size for LMF yields zero steady-state MSD","Step-size optimizer yields zero MSD for low-SNR LMF"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole scheme presumes that at iteration $n$ the filter knows its current mean-square deviation $\\mathrm{MSD}(n)$ and the noise moments $\\mathbb{E}[\\rho^4(n)]$ and $\\mathbb{E}[\\rho^6(n)]$, quantities that in real system identification depend on the very unknown system being identified.","fun_headline_variants_meta":{"raw":{"variants":["Optimal step-size zeroes LMF steady-state MSD at low SNR","Per-iteration MSD-minimizing step-size reaches zero error","Low-SNR LMF step-size control eliminates misalignment","Adaptive step-size for LMF yields zero steady-state MSD","Step-size optimizer yields zero MSD for low-SNR LMF"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000808,"raw_usage":{"total_tokens":3558,"prompt_tokens":965,"completion_tokens":2593,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":2502}},"tokens_in":581,"tokens_out":2593,"duration_ms":19609,"temperature":1.0,"reasoning_tokens":2502,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:49:38.796099+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run OPLMF on the paper's low-SNR identification tasks while estimating $\\mathrm{MSD}(n)$ online from the filter's error signal instead of using the true coefficient error $\\|\\mathbf{W}_O(n)-\\mathbf{W}(n)\\|^2$; if the measured steady-state MSD then remains clearly above zero, or the filter diverges, the claimed zero steady-state MSD holds only when an oracle supplies the state.","supporting_citations":[{"cited_title":"The least mean fouth (LMF) adaptive algorithm and its family [J]","cited_arxiv_id":null,"evidence_quote":"Establishes the least mean fourth algorithm and its stability trade-offs, which this paper extends with an optimized step-size."},{"cited_title":"An optimized NLMS algorithm for system identification [J]","cited_arxiv_id":null,"evidence_quote":"Introduces the practice of choosing an approximately optimal step-size by minimizing MSD, the design principle adopted here."},{"cited_title":"Optimal step size of least mean absolute third algorithm [J]","cited_arxiv_id":null,"evidence_quote":"Provides the optimized least mean absolute third algorithm whose MSD-minimization derivation is adapted to the LMF cost."},{"cited_title":"Stochastic anal ysis of a stable normalized least mean fourth algorithm for adaptive noise canceling with a white Gaussian reference [J]","cited_arxiv_id":null,"evidence_quote":"Supplies the normalized LMF baseline and its complexity count used for comparison in Table 2 and the simulations."},{"cited_title":"Global stabilization of the least mean fourth algorithm [J]","cited_arxiv_id":null,"evidence_quote":"Gives the sufficient stability condition for the LMF recursion that the paper uses to bound the step-size in Eq. (25)."},{"cited_title":"A robust and stable variable step-size design for the LMF using quotient form [J]","cited_arxiv_id":null,"evidence_quote":"Presents the VSSLMFQ quotient-form variable step-size algorithm used as the main baseline in the simulations."},{"cited_title":"A nonparametric VSSNLMS algorithm [J]","cited_arxiv_id":null,"evidence_quote":"Provides the nonparametric variance estimator used to update the input variance estimate and the gamma range in the algorithm."},{"cited_title":"Mathematical handbook of formulas and tables [M]","cited_arxiv_id":null,"evidence_quote":"Gives the fourth and sixth moment formulas for the Gaussian and non-Gaussian noise distributions used to instantiate Eq. (29)."}],"review_version":1}