{"id":"297a050b-52ff-4444-aff1-6ec84945b20a","arxiv_id":"2505.20747","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A kernel-based Volterra estimator that uses Wiener-Hammerstein structure through off-diagonal kernel blocks and runs in O(Nγ²) time for separable inputs.","lead":"The paper introduces a new way to build mathematical kernels for identifying Wiener-Hammerstein nonlinear systems from input-output data, using the known block structure of these systems. This leads to more accurate estimates and, for a common class of signals, substantial computational savings.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PSD guarantee in Theorem 3.1 does not cover the negative lags produced by the convolution in (10); for the proposed DC-based ζ, κ1−ζζ is not PSD on those lags.","rationale":"The reader's weakest assumption concerns applying the FIR/polynomial kernel to IIR and saturation systems. That is a legitimate scoping concern, but the paper explicitly acknowledges it, and the central contribution is for polynomial FIR WH systems. I found a more internal problem that affects even the scoped setting: the convolution in (10) evaluates κ1 and ζ at negative lags, while the PSD proof only covers nonnegative lags, and for the proposed DC-based ζ the condition κ1−ζζ is PSD fails on negative lags. If the concrete test shows indefiniteness, the regularization and empirical-Bayes interpretation in (4)–(5) lose their foundation; if the authors intended zero extension, then the efficient Q construction no longer matches the kernel. The paper has real strengths — the structured off-diagonal kernel, the O(N^3) construction, and the separability argument are nontrivial and the simulations are supportive — so this is not a rejection of the contribution. It is a request to state the domain of κ1, ζ, and the convolution support, and to verify that Theorem 3.1 covers the matrix actually used. That is exactly the kind of fixable but load-bearing gap that justifies a conditional verdict.","tokens_in":25621,"tokens_out":25936,"duration_ms":273646,"concrete_test":"Take n=2, M=2, a1=a2=1, κ1 and κ2 as DC kernels with β1,β2>0, α1=α2=1, and ζ(t)=c1 e^{−(α1+β1)t}. Form the 9×9 matrix P by literal evaluation of (10) over t_i,s_j∈{0,1}, using the DC formulas extended to negative lags, and compute its minimum eigenvalue. If it is negative for some hyperparameter values, Theorem 3.1's sufficient condition is not sufficient for the kernel actually used. If the authors instead zero-extend κ1/ζ for negative lags, compare the Q built from conv2(K2,Q^w) in (11) with Q=ΦPΦ^T from that P to check whether the efficient construction corresponds to the PSD kernel.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Kernel (10a) sums ξ1,ξ2 over 0..n−1 while t_i,s_j only range over 0..n−1. When t_i<ξ1 or s_j<ξ2, K^w_pq is evaluated at negative lags. Section 2.1 defines κ1, ζ only on R_+, and the proof of Theorem 3.1 in Appendix A.2 establishes PSD of P^w only for nonnegative lags. If κ1 and ζ are extended by the DC formulas to negative lags, the sufficient condition of Theorem 3.1 fails: at t=s=−r<0, κ1(−r,−r)−ζ(−r)^2 = c1^2 e^{2αr}(1−e^{2βr}) < 0 for the ζ of Corollary 3.1(1). Thus the actual coefficient-space matrix P formed by (10) with nonzero off-diagonal κ2 — i.e. nontrivial G2 — is either undefined or outside the PSD guarantee. The O(N^3) construction of Q in Theorem 4.1 shifts the input rather than the kernel, so it can be correct even if P is not PSD, but the central claim that kernel (10) is PSD under Theorem 3.1 is not established for the implemented kernel.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses kernel-based regularized identification of Volterra series for Wiener-Hammerstein (WH) systems with polynomial nonlinearity. Starting from the optimal kernel θ_p θ_q^T, the authors design structural kernels K_pq of the form (10a)-(10b), which embed the WH convolution structure through κ2 and products of κ1 and ζ. Theorem 3.1 states a positive semidefiniteness (PSD) condition for the resulting kernel matrix; Theorem 4.1 gives an O(N^3) construction of the output kernel matrix Q via a 2-D convolution; Theorem 4.2 shows that, for a delta-like κ2 and inputs satisfying a separability condition (15), Q has separability rank γ and the cost drops to O(Nγ^2). The paper includes Monte Carlo experiments on four databanks, comparing the proposed kernels against SED-MPK and two existing Wiener-system methods.","tokens_in":25935,"tokens_out":18371,"duration_ms":193782,"significance":"If the proof gaps identified below are fixed, this is a valuable contribution: it provides a direct, structural kernel for Volterra maps of WH systems with nonzero off-diagonal blocks, and it matches the state-of-the-art O(N^3) complexity while encoding more prior knowledge than previous kernels. The separability-rank analysis in Theorem 4.2 is nontrivial and the simulations support the practical usefulness of the approach. The authors provide detailed appendix proofs and a transparent experimental setup, which are strengths. The main weaknesses are the incomplete PSD guarantee for the implemented kernel at negative lags and the underspecified convolution in Theorem 4.1; both appear fixable, which is why I recommend major revision rather than rejection.","major_comments":[{"comment":"The proposed kernel evaluates κ1(t_i−ξ1, s_i−ξ2) and ζ at arguments that can be negative whenever ξ1>t_i or ξ2>s_i, since ξ1, ξ2 range over 0,...,n−1 while t_i, s_i start at 0. However, κ1 and ζ are introduced in Section 2.1 only on R_+, and the proof of Theorem 3.1 in Appendix A.2 establishes PSD of P^w only for nonnegative indices. If the DC formulas are naively extended to negative arguments, the sufficient condition of Theorem 3.1 fails: for ζ as in Corollary 3.1(1), κ1(−r,−r)−ζ(−r)^2 = c1^2 e^{2αr}(1−e^{2βr}) < 0 for r>0. Thus the PSD guarantee does not cover the implemented kernel whenever κ2 has nonzero off-diagonal entries, which is exactly the D1/D2 setting. The authors should specify a causal zero extension (or another explicit extension) for κ1, ζ, and Kw at negative lags and then prove that the convolution in (10a) preserves PSD under that extension; as written, the claim that P defined through (10) is PSD is not established.","section":"Section 3, Eq. (10a), and Theorem 3.1"},{"comment":"With K2 ∈ R^{n×n} and Qw ∈ R^{N×N}, the 2-dimensional convolution conv2(K2, Qw) as defined in the footnote has size (N+n−1)×(N+n−1), not N×N as stated for Q. Moreover, Eqs. (30)-(32) require Qw_pq(t−ξ1, s−ξ2) at negative indices when t or s is near 1, but Qw in (12) is only defined for indices 1,...,N. The theorem should specify the zero-padding convention (for example, u(t)=0 for t≤0, corresponding to the pre-windowing option in footnote 2) and state which submatrix of the convolution is taken. Without this, the claimed O(N^3) construction of Q is not well defined.","section":"Section 4.1, Eq. (11), and Theorem 4.1"},{"comment":"The structural prior (8) assumes FIR linear blocks G1 and G2 and a polynomial nonlinearity, but Databanks D1 and D2 use IIR linear blocks and Databank D3 uses a saturation nonlinearity. The only justification is the qualitative truncation and polynomial-approximation argument in Remarks 2.1 and 2.2. Since these experiments are used as evidence for the kernel's effectiveness, the paper should either provide a quantitative statement about the resulting approximation error or explicitly present these cases as robustness checks of a misspecified model rather than as instances where the structural prior holds.","section":"Section 5, databanks D1-D3"}],"minor_comments":[{"comment":"The abstract states that the complexity is lowered to O(Nγ^2), but Theorem 4.2 gives O(Nγ^2+n^2r) and requires the additional assumption that κ1 is extended-p semiseparable in order to remove the n^2r term; the abstract should state this extra condition.","section":"Abstract and Theorem 4.2"},{"comment":"The eigenfunction notation is inconsistent: the corollary statement uses ψ_i while the proof uses ϕ_i for the same objects, and the claim l≪N is not satisfied in the simulations of Databank D2, where l=100 and N=400.","section":"Corollary 3.1(2) and Appendix A.3"},{"comment":"The label 'R VS' is used without definition; it should be spelled out on first use, for example 'regularized Volterra series (R VS)'.","section":"Section 5.1"},{"comment":"It would be useful to state explicitly which test inputs satisfy the separability assumption; white Gaussian noise (D2) does not satisfy it exactly, while the damped cosine of D4 does with small r.","section":"Section 4.2, Eq. (15)"}],"recommendation":"major_revision","confidential_remarks":"The PSD and convolution issues are substantive but appear fixable within the manuscript's scope; I do not see concerns about novelty or fit with the journal. Please ask the authors to clarify the extension to negative lags and to correct the convolution dimensions before the next round."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one. It is the first, within this literature, to build off-diagonal blocks into Volterra-map kernels from Wiener-Hammerstein structure, and the complexity reductions are genuinely worked out. That part deserves attention. The catch: the main PSD theorem, as stated, has a gap on negative lags. Convince yourself by plugging t=s=-r into κ1(t,s)-ζ(t)ζ(s) with the Corollary 3.1(1) ζ: you get c1^2 e^{2αr}(1-e^{2βr})<0, while (10a) inevitably evaluates κ1 and ζ at such negative arguments. The proof of Theorem 3.1 establishes PSD only on nonnegative lags, then invokes a convolution theorem. With any straightforward extension to negative lags, the extension is not PSD, so the convolution argument does not hold. This is not a cosmetic quibble: the PSD guarantee is what makes the kernel matrix legitimate in the regularization. It is repairable—by restricting the convolution to causal terms, by designing ζ so that κ1-ζζ stays PSD on all integer lags, or by a direct PSD proof for the actual submatrix—but it needs to be fixed.\n\nWhat is genuinely good: the derivation from the optimal kernel (Prop 3.1), the explicit rank formula in Theorem 4.2, and the simulation work. D2/D3 comparisons against SED-MPK and semi-parametric baselines show real prediction improvements. The self-citations are legitimate building blocks, not a circularity. I also think the idea of using system structure instead of generic smoothness for Volterra kernels is the right direction.\n\nSoft spots beyond the PSD gap: Theorem 4.1 says Q=conv2(K2,Qw) with K2 n×n and Qw N×N; that output is (N+n-1)×(N+n-1), so some cropping or 'same' convolution convention is missing. Theorem 4.2 skips the proof that ϕPΦ^T is separable with the same rank, which is load-bearing for O(Nγ^2) of prediction. And the kernel is derived under FIR+polynomial assumptions, but applied to IIR and saturation in D1-D3 with only a qualitative truncation remark; the misspecification is not quantified. None of these look fatal; they look addressable.\n\nBottom line: this is a paper for kernel-based system identification people. I would send it to a serious referee, not desk reject. The PSD statement must be corrected before publication; with that fixed, the contribution stands.","headline":"A genuinely new structural kernel design for Volterra identification, with real complexity gains, but the central PSD theorem has a fixable gap on negative lags.","tokens_in":26500,"tokens_out":7493,"would_cite":true,"duration_ms":78429,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93B30","93C10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Structured kernels built from the Wiener-Hammerstein block prior identify Volterra series in cubic time and, for separable inputs, near-linear time.","keywords":["Wiener-Hammerstein systems","Volterra series identification","kernel-based regularization","kernel design","output kernel matrix","separability rank"],"falsifier":"For the Wiener special case $\\kappa_2(t,s)=1$ when $t=s=0$ and $0$ otherwise, take input $u(t)=e^{\\lambda t}$ (so $r=1$) and $M=3$; Theorem 4.2 says the output kernel matrix $Q_w$ has exact separability rank $\\gamma=5$. Compute the singular values of the numerically formed $Q_w$ for $N$ large: if the sixth singular value is not at machine precision relative to the largest, the rank formula fails. The same check applies for $u(t)=e^{-0.0003t}\\cos(0.1t+\\pi/3)$ with $r=2$ and $M=3$, where the predicted rank is $\\gamma=14$.","tokens_in":25385,"feed_emoji":"⚙️","tokens_out":14649,"duration_ms":142360,"temperature":0.7,"pith_summary":"This paper tries to establish that the bottleneck of regularized Volterra-series identification—kernel design and computation—can be resolved by building the kernel from the actual Wiener-Hammerstein (WH) block structure instead of from generic smoothness assumptions. For a WH system with finite-memory linear blocks and polynomial nonlinearity, the true Volterra maps factor as $h_m(t_1,\\ldots,t_m)=a_m\\sum_{\\tau=0}^{n-1}g_2(\\tau)\\prod_{i=1}^m g_1(t_i-\\tau)$, and the paper turns this factorization into a positive-semidefinite kernel with nonzero off-diagonal blocks coupling different Volterra orders. The convolution structure of the kernel makes the output covariance matrix computable in $O(N^3)$ time, and in a Wiener-system special case with separable input signals the cost drops to $O(N\\gamma^2)$, where $\\gamma$ is an explicit rank. Monte Carlo simulations show better average prediction accuracy than generic smoothness kernels and than Bayesian semiparametric and prediction-error baselines on the tested systems.","feed_headline":"Structure-aware kernels cut Volterra identification to O(N^3)","feed_subtitle":"Kernels built from block-structure priors beat generic smoothness and scale linearly for separable inputs.","key_machinery":"The load-bearing object is the factored Volterra-map identity $h_m(t_1,\\ldots,t_m)=a_m\\sum_{\\tau=0}^{n-1}g_2(\\tau)\\prod_{i=1}^m g_1(t_i-\\tau)$, which expresses every order of a Wiener-Hammerstein system as a shared convolution with $g_2$ of products of $g_1$. The designed kernel (10a)-(10b) replaces $g_1,g_2$ by kernels $\\kappa_1,\\kappa_2$ and $g_1$ by $\\zeta$; Theorem 3.1 guarantees the full kernel matrix is positive semidefinite exactly when $\\kappa_1(t,s)-\\zeta(t)\\zeta(s)$ is positive semidefinite. Computation is carried by two structural identities: $Q=\\mathrm{conv2}(K_2,Q_w)$ for the output kernel, and, in the Wiener special case, the separability decomposition $Q=\\bar{U}\\bar{V}^T$ whose rank $\\gamma$ is given by the combinatorial formula in Theorem 4.2.","core_discovery":"The central claim is that the optimal-kernel identity $P_{\\mathrm{opt}}=\\theta_0\\theta_0^T$, applied to the exact Volterra-map form of a WH system, produces an explicit cross-kernel $$K_{pq}(t_p,s_q;\\eta_{pq})=\\sum_{\\xi_1=0}^{n-1}\\sum_{\\xi_2=0}^{n-1}\\kappa_2(\\xi_1,\\xi_2;\\eta_2)K^w_{pq}(t_p-\\xi_1,s_q-\\xi_2;a_p,a_q,\\eta_1),$$ where $\\kappa_1,\\kappa_2$ are user-designed kernels for the two linear blocks and $\\zeta$ is a designed stand-in for $g_1$. This kernel has nonzero off-diagonal blocks, so it encodes correlations between different Volterra orders, and it is positive semidefinite whenever $\\kappa_1(t,s)-\\zeta(t)\\zeta(s)$ is positive semidefinite. The output kernel matrix has the convolution form $Q=\\mathrm{conv2}(K_2,Q_w)$, computable in $O(N^3)$; in the special case $\\kappa_2(t,s)=1$ if $t=s=0$ and $0$ otherwise, with inputs satisfying $u(t-b)=\\sum_{i=1}^r\\pi_i(t)\\rho_i(b)$, $Q$ is exactly separable with rank $\\gamma=\\frac{(r+M-1)!}{(r-1)!M!}+2\\sum_{m=1}^{M-1}\\frac{(r+m-1)!}{(r-1)!m!}$, reducing the empirical-Bayes and prediction computations to $O(N\\gamma^2)$.","pith_inferences":["The same kernel construction should extend to Hammerstein systems and to parallel interconnections of Wiener-Hammerstein blocks, because the Volterra map of a parallel structure is a sum of terms of type (8) and the kernel would be the corresponding sum, keeping the convolution form.","Because $\\gamma$ depends only on $M$ and $r$, the $O(N\\gamma^2)$ regime is best for weakly nonlinear systems or low-complexity inputs; a testable extension is to update the separable generators incrementally for online or adaptive identification.","When the FIR/polynomial prior is wrong, a natural safeguard is to add a generic smoothness kernel term to the designed kernel, giving an estimator that can fall back to smoothness when the block-structure prior is misspecified.","The comparisons emphasize prediction and block fits; a further useful test is whether the off-diagonal blocks recover the true cross-correlations between Volterra orders, something the block-diagonal baselines cannot do."],"forward_implications":["For Wiener-Hammerstein systems with polynomial nonlinearity, the regularization kernel can encode block-structure and cross-order correlations rather than only smoothness and decay.","The empirical-Bayes estimate and hyperparameter criterion can be evaluated in $O(N^3)$ time independent of the number of Volterra coefficients $n_\\theta$.","For Wiener systems with separable input signals, the output kernel matrix has explicit rank $\\gamma$ and the computation scales as $O(N\\gamma^2)$, i.e., linearly in sample size for fixed $\\gamma$.","On the tested Wiener-Hammerstein and Wiener databanks, the proposed kernels give better average prediction fit than the SED-MPK kernel, and better robustness than Bayesian semiparametric and prediction-error baselines in the D3 comparison.","Hyperparameters are still tuned by empirical Bayes, preserving automatic model-complexity control in short or low-SNR data."],"supporting_citations":[{"why":"Supplies the optimal-kernel identity $P_{\\mathrm{opt}}=\\theta_0\\theta_0^T$ and the design guideline from which the proposed kernel follows.","marker":"[17]"},{"why":"The state-of-the-art $O(N^3)$ Volterra-series kernel SED-MPK that the proposed kernels match in complexity and are compared against.","marker":"[9]"},{"why":"Establishes the Volterra representation (8) of Wiener-Hammerstein systems with FIR linear blocks and polynomial nonlinearity.","marker":"[11]"},{"why":"Earlier regularized Volterra kernel using smoothness priors and the D1 test system used as a comparison baseline.","marker":"[7]"},{"why":"Bayesian semiparametric Wiener identification method and the saturation test system used for the D3 comparison.","marker":"[22]"},{"why":"Maximum-likelihood prediction-error Wiener model used as the PEM baseline for D3.","marker":"[14]"},{"why":"Gives the DC-kernel eigenfunctions and eigenvalues used to construct $\\zeta$ for the DC-ob kernel.","marker":"[33]"},{"why":"Provides the semiseparable-kernel machinery and the input separability assumption used to derive the $O(N\\gamma^2)$ result.","marker":"[34]"}],"fun_headline_variants":["Structure-aware kernels cut Volterra ID to O(N^3)","Off-diagonal Volterra kernels: direct design, O(N^3) cost","Separable outputs shrink Volterra identification to O(Nγ^2)","Wiener-Hammerstein kernels with block priors speed up Volterra fits","Volterra kernel design: from O(N^3) to O(Nγ^2) for separable inputs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the true system is exactly a Wiener-Hammerstein system with finite-memory linear blocks of length $n$ and a static nonlinearity that is a polynomial of degree $M$; if the real dynamics have infinite memory or a non-polynomial nonlinearity, the kernel encodes the wrong structure and can introduce bias.","fun_headline_variants_meta":{"raw":{"variants":["Structure-aware kernels cut Volterra ID to O(N^3)","Off-diagonal Volterra kernels: direct design, O(N^3) cost","Separable outputs shrink Volterra identification to O(Nγ^2)","Wiener-Hammerstein kernels with block priors speed up Volterra fits","Volterra kernel design: from O(N^3) to O(Nγ^2) for separable inputs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000655,"raw_usage":{"total_tokens":3085,"prompt_tokens":1116,"completion_tokens":1969,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":732,"completion_tokens_details":{"reasoning_tokens":1866}},"tokens_in":732,"tokens_out":1969,"duration_ms":15895,"temperature":1.0,"reasoning_tokens":1866,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:48:42.929995+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For the Wiener special case $\\kappa_2(t,s)=1$ when $t=s=0$ and $0$ otherwise, take input $u(t)=e^{\\lambda t}$ (so $r=1$) and $M=3$; Theorem 4.2 says the output kernel matrix $Q_w$ has exact separability rank $\\gamma=5$. Compute the singular values of the numerically formed $Q_w$ for $N$ large: if the sixth singular value is not at machine precision relative to the largest, the rank formula fails. The same check applies for $u(t)=e^{-0.0003t}\\cos(0.1t+\\pi/3)$ with $r=2$ and $M=3$, where the predicted rank is $\\gamma=14$.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the optimal-kernel identity $P_{\\mathrm{opt}}=\\theta_0\\theta_0^T$ and the design guideline from which the proposed kernel follows."},{"cited_title":"Dalla Libera, R","cited_arxiv_id":null,"evidence_quote":"The state-of-the-art $O(N^3)$ Volterra-series kernel SED-MPK that the proposed kernels match in complexity and are compared against."},{"cited_title":"Bai and F","cited_arxiv_id":null,"evidence_quote":"Establishes the Volterra representation (8) of Wiener-Hammerstein systems with FIR linear blocks and polynomial nonlinearity."},{"cited_title":"Birpoutsoukis, A","cited_arxiv_id":null,"evidence_quote":"Earlier regularized Volterra kernel using smoothness priors and the D1 test system used as a comparison baseline."},{"cited_title":"Lindsten, T","cited_arxiv_id":null,"evidence_quote":"Bayesian semiparametric Wiener identification method and the saturation test system used for the D3 comparison."},{"cited_title":"Hagenblad, L","cited_arxiv_id":null,"evidence_quote":"Maximum-likelihood prediction-error Wiener model used as the PEM baseline for D3."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the DC-kernel eigenfunctions and eigenvalues used to construct $\\zeta$ for the DC-ob kernel."},{"cited_title":"Chen and M","cited_arxiv_id":null,"evidence_quote":"Provides the semiseparable-kernel machinery and the input separability assumption used to derive the $O(N\\gamma^2)$ result."}],"review_version":1}