{"id":"c6218b57-9ba1-490c-aaf2-1b32d9f5e4c5","arxiv_id":"1908.02510","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A q-calculus variant of Volterra LMS is presented, but for equal q it is exactly Volterra LMS with a scaled step size, so the claimed improvement is largely a step-size effect.","lead":"This short paper proposes q-VLMS, a Volterra least mean square adaptive filter that replaces the usual derivative with a quantum-calculus (q) derivative and claims faster convergence for nonlinear channel estimation. The reported speedup appears to come mostly from an increased effective step size, and the mathematical support is not solid.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 2.31 dB gain claim is not established: Eq. (7) is exactly VLMS with per-tap step sizes µ(q_i+1)/2, and Protocol 2 compares only at a common µ, so any improvement may be a step-size artifact rather than a q-calculus effect.","rationale":"The central claim requires that Eq. (7) be a genuinely q-calculus-based update whose advantage over VLMS comes from the q-derivative rather than from a re-parameterization of the standard update. That condition is not met in the paper as written: Eq. (7) is the standard VLMS update with a constant diagonal gain, i.e., per-tap step-size control, and Protocol 2 does not separate this effect from the q-derivative. The reader's weakest-assumption analysis identifies this same confound, and an independent derivation of Eq. (6) from Eq. (5) also fails because the Jackson derivative of e^2 contains a (q_i-1)w_i u_i^2 term absent from Eq. (6). Both considerations point to the same conclusion: the 2.31 dB improvement is not shown to be a q-calculus effect. A matched-step-size control would settle the issue directly; absent such a control, the central claim is unsupported. The reader's REJECT verdict is therefore appropriate, and no change to that verdict is needed.","tokens_in":5757,"tokens_out":13793,"duration_ms":144620,"concrete_test":"Re-run Protocol 2 with a control arm that isolates the q-gain: use the same normalized regressor S^{-1}u described in Section III for both q-VLMS and VLMS; in the VLMS arm, set per-tap step sizes µ_i=µ(q_i+1)/2 (equivalently, run q=1 with a diagonal step matrix µG), keeping µ, q_i, channel draws, input draws, and SNR levels identical to the reported q-VLMS runs. Because Eq. (7) is a constant diagonal scaling of the VLMS update, the matched control should reproduce the q-VLMS curves exactly if Eq. (7) is implemented faithfully. If the reported 2.31 dB advantage over ordinary VLMS disappears under this matching, the claimed improvement is fully explained by the effective step-size increase and does not support a new q-calculus-based algorithm.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is Eq. (7). Because G is a constant diagonal matrix, w(r+1)=w(r)+µGu(r)e(r) is exactly the standard VLMS update with per-tap step sizes µ_i=µ(q_i+1)/2; for q_i>1 this is VLMS at a larger effective step. Protocol 2 compares q-VLMS and VLMS at the same scalar µ, so the reported average 2.31 dB NWD improvement is exactly the kind of transient acceleration expected from a larger step size, and it does not demonstrate a benefit of the q-derivative. The convergence bound in Eq. (16) makes the same point: q enters only by shrinking the allowed µ interval through (q+1), i.e., through the effective step size. There is also a derivation gap: applying the Jackson derivative in Eq. (5) to J(w)=e^2(r), with e=c-w_i u_i, gives D_{q,w_i}J = -2u_i e + (q_i-1)w_i u_i^2, not the -2[(q_i+1)/2]u_i e claimed in Eq. (6). The extra term is dropped without explanation, so the identification of Eq. (7) as a q-gradient algorithm is not established. The central claim therefore rests on an uncontrolled comparison and an unsupported gradient identity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a q-calculus (Jackson derivative) variant of the Volterra least mean square (LMS) algorithm for nonlinear channel estimation. The proposed q-VLMS update is w(r+1)=w(r)+mu G u(r)e(r) with G = diag((q_i+1)/2). The authors derive an optimal solution, analyze mean convergence, compute the step-size bound, and perform simulations at three SNRs. They report that q-VLMS converges faster than conventional VLMS, with an average 2.31 dB NWD improvement.","tokens_in":6095,"tokens_out":7596,"duration_ms":76555,"significance":"If the central claims were established, the paper would offer a simple modification of Volterra LMS with a tunable per-tap gain and would provide a useful convergence analysis. The authors do include an analytical convergence model and a 1000-trial Monte Carlo validation, which are strengths. However, the claimed improvement is not supported: the update reduces to standard VLMS with per-tap step sizes mu(q_i+1)/2, the q-gradient identity in Eq. (6) is not derived and is incorrect under the standard Jackson derivative, and the convergence bound in Eq. (16) is inconsistent with the update. The paper's main conclusion rests on an uncontrolled comparison and an unsupported gradient identity, so the contribution as stated is not established.","major_comments":[{"comment":"The q-gradient identity in Eq. (6) is asserted without derivation and is not the standard Jackson derivative defined in Eq. (5). For the instantaneous cost e^2 with e = o - u^T w, the component-wise Jackson derivative with respect to w_i equals -2u_i e + (q_i-1) w_i u_i^2, not -2[(q_i+1)/2]u_i e. The additional (q_i-1)w_i u_i^2 term is dropped without explanation. Because Eq. (7) is derived from Eq. (6), the algorithm is not established as a q-gradient descent method.","section":"Section III, Eqs. (5)-(6)"},{"comment":"The mean weight-error recursion (14) has a sign error: with Delta w = w_opt - w and e = u^T Delta w + eta, the update yields Delta w(r+1) = (I - mu G u u^T)Delta w(r) - mu G u eta (or the analogous expression with the opposite sign convention), not the plus sign shown before the u^T Delta w term. In addition, Eq. (16) is inconsistent with update (7): since G = diag((q_i+1)/2), the stability bound should be mu < 2/(max_i (q_i+1)lambda_i) (or 4/((q+1)lambda_max) under the assumed whitening), not 1/((q+1)lambda_max). The analytical convergence analysis therefore does not match the simulated algorithm.","section":"Section III-B, Eqs. (14)-(16)"},{"comment":"The comparison between q-VLMS and conventional VLMS at the same nominal mu is uncontrolled. Because G is a constant diagonal matrix, update (7) is exactly the standard VLMS update with per-tap step sizes mu_i = mu(q_i+1)/2. For q_i > 1, this is VLMS with a larger effective step size per tap. The reported average 2.31 dB NWD improvement is therefore fully consistent with a step-size scaling artifact and does not demonstrate any benefit of the q-derivative. To support the paper's claim, the authors would need to compare q-VLMS against VLMS with per-tap step sizes mu_i, or otherwise equalize the effective step sizes.","section":"Section IV, Protocol 2"},{"comment":"The sentence 'we repeat the same simulations with G = SR^-1S' is inconsistent with the definition G = diag((q_i+1)/2) used throughout the algorithm and analysis. If this phrase is intended to define the VLMS comparison, it must be clarified; if not, it introduces a different algorithm whose relationship to the preceding analysis is unclear. More fundamentally, since q enters only through the fixed diagonal gain G, the paper's central claim that q-calculus improves convergence is not supported once the equivalence to per-tap step-size VLMS is recognized.","section":"Section IV, Protocol 2"}],"minor_comments":[{"comment":"The notation in Eq. (5) is ambiguous: 'f(qn)' and '(q-1)i' should be written as f(qx) and (q-1)x, or the component-wise vector generalization should be stated explicitly.","section":"Section III, Eq. (5)"},{"comment":"The definition of diag(S) uses set braces; please use vector or diagonal-matrix notation for consistency with the rest of the paper.","section":"Section II, Eq. (8)"},{"comment":"Figures 1 and 2 are referenced but not described in the text; please ensure that the plots include axis labels, legends, SNR values, and the q values used in Fig. 2.","section":"Section IV"},{"comment":"The term 'Quantum Calculus' is used in the title and abstract, but the paper employs Jackson's q-calculus; consider using 'q-calculus' consistently to avoid confusion with quantum computing.","section":"Title and Abstract"}],"recommendation":"reject","confidential_remarks":"The reference list is heavily dominated by the authors' own prior work (e.g., [2], [3], [5], [6], [9]-[16]), and the paper does not appear to engage with the broader q-LMS literature beyond those papers. This is not by itself disqualifying, but the novelty relative to [15] and [16] should be assessed carefully. Given the load-bearing errors in the gradient derivation, the convergence bound, and the experimental comparison, the central claim is not currently supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this paper doesn't establish its central claim. The proposed q-VLMS update (Eq. 7) is exactly the standard Volterra LMS update with step size μ(q_i+1)/2 per tap. For equal q it's VLMS with step μ(q+1)/2. So the \"novel algorithm\" reduces to a known filter with a step-size change, and the experiments compare at the same nominal μ without controlling for that. The reported 2.31 dB gain is what you'd expect from simply running VLMS at a larger step.\n\nWhat's actually here: a short, readable paper that extends the q-LMS idea to Volterra filters, includes a mean convergence analysis, and tests on a nonlinear channel estimation problem with 1000 trials. The authors also cite the recent critiques of fractional LMS, which is good practice. The convergence bound in Eq. (16) is conservative but not obviously wrong at first glance (it has a factor-of-two offset relative to the standard 2/λ_max bound, likely from a sign error in Eq. (14)).\n\nThe soft spots are load-bearing. First, the q-gradient in Eq. (6) is asserted without derivation. Applying Jackson's derivative to e^2(r) gives an extra term (q_i-1)w_i u_i^2 that is dropped; without an explanation, Eq. (7) is not actually a q-gradient descent. Second, Eq. (14) has a sign error: the update gives Δw(n+1)=Δw − μ G S⁻¹u(...) (with the right definition), and the plus sign in (14) contradicts the subsequent (I−μA)^i form. Third, and most importantly, the equivalence to scaled VLMS is never acknowledged. Protocol 2 compares q-VLMS and VLMS at the same μ; if they had instead run VLMS at μ(q+1)/2, the curves would likely match. There is no evidence that the q-derivative itself helps.\n\nWho this is for: readers working on adaptive filtering might find the per-tap diagonal gain idea familiar—it is diagonal preconditioning—but the paper doesn't frame it honestly. A reviewer could salvage a minor note about variable-step VLMS, but as written it does not deserve publication. I would not cite it, and I would not send it to a serious referee; the core identification collapses on inspection.","headline":"The q-VLMS algorithm is just VLMS with a scaled step size, and the reported 2.31 dB gain is an artifact of that scaling rather than a benefit of q-calculus.","tokens_in":6633,"tokens_out":3936,"would_cite":false,"duration_ms":40044,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93E24","62L20"],"pacs":[],"model":"deepseek-v4-flash","headline":"q-VLMS, a q-calculus variant of Volterra LMS, converges faster in nonlinear channel estimation, with an average 2.31 dB gain over VLMS.","keywords":["q-calculus","Volterra LMS","nonlinear channel estimation","adaptive filtering","Jackson derivative","convergence analysis","step-size bound","least mean square"],"falsifier":"Run conventional VLMS on the same second-order Volterra channels with per-tap step sizes $\\mu(q_i+1)/2$ (that is, replace $\\mu$ by $\\mu G$) and compare NWD curves to q-VLMS at step $\\mu$; if the curves overlap, the reported 2.31 dB gain is an effective-step-size artifact, not a q-calculus effect.","tokens_in":5565,"feed_emoji":"📡","tokens_out":12837,"duration_ms":119907,"temperature":0.7,"pith_summary":"This paper claims that the Volterra least-mean-square (VLMS) adaptive filter can be made to converge faster for nonlinear channel estimation by replacing the ordinary gradient with the $q$-derivative, also called Jackson's derivative. The resulting $q$-VLMS update multiplies each tap's correction by $(q_i+1)/2$, which gives high-$q$ taps a larger effective step size. The paper derives a convergence bound $0<\\mu<1/\\max_i\\{(q_i+1)\\lambda_i\\}$ and validates it in simulations on a second-order Volterra channel at three noise levels. In those simulations, q-VLMS outperforms conventional VLMS by an average 2.31 dB in normalized weight deviation. If right, this gives a simple, low-cost way to speed up nonlinear channel estimation without increasing filter order.","feed_headline":"q-calculus speeds up nonlinear channel estimation by 2.31 dB","feed_subtitle":"A diagonal gain from the q-derivative makes the Volterra LMS update converge faster, cutting normalized weight error.","key_machinery":"The machinery is the $q$-derivative (Jackson's derivative), $D_q f(x)=(f(qx)-f(x))/((q-1)x)$, used in place of the ordinary gradient of the mean-square error. For a quadratic cost this replaces the VLMS update with $w(r+1)=w(r)+\\mu G u(r)e(r)$, where $G=\\mathrm{diag}((q_1+1)/2,\\ldots,(q_M+1)/2)$ rescales each coordinate's effective step. The paper whitens the quadratic regressor with $S^{-1}$ so that the input correlation is the identity, then derives the mean recursion $E[\\Delta w(r+1)]=(I-\\mu A)E[\\Delta w(r)]$ with $A=GE[S^{-1}uu^{\\top}S^{-1}]$, yielding the stability bound $0<\\mu<1/\\max_i\\{(q_i+1)\\lambda_i\\}$.","core_discovery":"The paper's central discovery is the update rule $w(r+1)=w(r)+\\mu G u(r)e(r)$, obtained by applying the $q$-derivative to the mean-square error cost of a second-order Volterra filter. The diagonal matrix $G=\\mathrm{diag}((q_1+1)/2,\\ldots,(q_M+1)/2)$ is the entire mathematical content left by the $q$-gradient; when $q_i=1$ the update reduces to VLMS. The authors show that under a Gaussian i.i.d. input model, the mean weight-error recursion becomes $E[\\Delta w(r+1)]=(I-\\mu A)E[\\Delta w(r)]$ with $A=GE[S^{-1}uu^{\\top}S^{-1}]$, giving the stability interval above. They report that the analytical mean absolute error tracks simulation with a correlation coefficient of 0.9995, and that q-VLMS outperforms VLMS by an average 2.31 dB in normalized weight deviation across SNRs of 10, 20, and 30 dB.","pith_inferences":["Because the update is exactly the VLMS update multiplied by $G$, q-VLMS at step $\\mu$ is algebraically VLMS at per-tap steps $\\mu(q_i+1)/2$; the paper never compares against VLMS run at those effective steps, so whether the gain is a q-derivative property or a step-size scaling remains open.","If the gain is step-size scaling, the practical lesson is to choose larger per-tap steps inside the stability bound, with the q-derivative functioning as a derivation device rather than a new mechanism.","The same diagonal-gain construction could be applied to normalized LMS, affine projection, or kernel adaptive filters, where eigenvalue spread slows convergence; this is an extension the paper does not test.","A time-varying $q_i$ driven by error-correlation energy, already used in the q-LMS literature the paper cites, would be a natural next test to see whether the gain compounds during adaptation."],"forward_implications":["Nonlinear channel estimators using q-VLMS can reach a target normalized weight deviation with fewer samples, which matters when the training sequence is short.","The step-size bound $0<\\mu<1/\\max_i\\{(q_i+1)\\lambda_i\\}$ gives a formula for setting $\\mu$ once the input eigenvalues are known, removing trial-and-error tuning.","The same convergence analysis used for linear LMS carries over to second-order Volterra filters when the quadratic input is whitened by $S^{-1}$, so the result extends the linear toolkit to nonlinear channels.","The added cost is only $K$ extra multiplications per update, so the reported gain does not require a heavier filter structure."],"supporting_citations":[{"why":"Supplies the Jackson-derivative-based q-LMS update that q-VLMS extends to the Volterra case.","marker":"[15]"},{"why":"Defines the second-order Volterra series expansion and the regressor/weight structure used throughout.","marker":"[17]"},{"why":"Introduces the diagonal scaling matrix that whitens the quadratic input, a prerequisite for the convergence analysis.","marker":"[18]"},{"why":"Provides the Gaussian i.i.d. assumptions and the mean/mean-square analysis framework used to derive the step-size bound.","marker":"[19]"}],"fun_headline_variants":["q-derivative gives Volterra LMS a 2.31 dB boost","q-VLMS: Volterra LMS with faster convergence, 2.31 dB better","q-calculus accelerates Volterra LMS, gains 2.31 dB in tests","Nonlinear channel estimation gets 2.31 dB boost from q-VLMS","q-calculus boosts Volterra LMS convergence by 2.31 dB"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that q-VLMS and VLMS are fairly compared at the same nominal step size; since q-VLMS is the VLMS update multiplied by $G=\\mathrm{diag}((q_i+1)/2)$, the measured 2.31 dB gain could come from the larger effective step size rather than from any property of the $q$-derivative.","fun_headline_variants_meta":{"raw":{"variants":["q-derivative gives Volterra LMS a 2.31 dB boost","q-VLMS: Volterra LMS with faster convergence, 2.31 dB better","q-calculus accelerates Volterra LMS, gains 2.31 dB in tests","Nonlinear channel estimation gets 2.31 dB boost from q-VLMS","q-calculus boosts Volterra LMS convergence by 2.31 dB"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000852,"raw_usage":{"total_tokens":3657,"prompt_tokens":854,"completion_tokens":2803,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":470,"completion_tokens_details":{"reasoning_tokens":2697}},"tokens_in":470,"tokens_out":2803,"duration_ms":18766,"temperature":1.0,"reasoning_tokens":2697,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:42:15.653144+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run conventional VLMS on the same second-order Volterra channels with per-tap step sizes $\\mu(q_i+1)/2$ (that is, replace $\\mu$ by $\\mu G$) and compare NWD curves to q-VLMS at step $\\mu$; if the curves overlap, the reported 2.31 dB gain is an effective-step-size artifact, not a q-calculus effect.","supporting_citations":[{"cited_title":"Second-order volterra filtering and its application to nonlinear system identification,","cited_arxiv_id":null,"evidence_quote":"Defines the second-order Volterra series expansion and the regressor/weight structure used throughout."},{"cited_title":"Optimizing the performance of polynomial adaptive filters: making quadratic filters converge like linear filters,","cited_arxiv_id":null,"evidence_quote":"Introduces the diagonal scaling matrix that whitens the quadratic input, a prerequisite for the convergence analysis."},{"cited_title":"Performance analysis of nonlinear adaptive filter based on lms algorithm,","cited_arxiv_id":null,"evidence_quote":"Provides the Gaussian i.i.d. assumptions and the mean/mean-square analysis framework used to derive the step-size bound."}],"review_version":1}