{"id":"46855058-fe7f-47d3-9c47-be2eccb3fe28","arxiv_id":"2509.05720","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"A time-domain kernel ridge regression framework for sound field estimation is derived, with closed-form solutions and experimental gains from time-domain data weighting.","lead":"This paper extends kernel ridge regression from single-frequency to discrete-time sound field estimation, giving closed-form estimates that satisfy the wave equation and allow temporal priors like room impulse response decay. It shows on simulated and real microphone data that time-domain weighting improves estimation accuracy and combines with directional weighting.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (29) has the cross term mis-ordered: it should be ⟨Γ_r Q h,a⟩, not ⟨Q Γ_r h,a⟩; the printed objective does not imply the claimed solution (30) unless Q and Γ_r commute.","rationale":"The reader's weakest assumption, the source-free Herglotz plane-wave model, is a legitimate scope limitation and is explicitly acknowledged in the paper, but it is not the most actionable issue. The more pressing, checkable problem is the algebraic inconsistency in Eq. (29): the cross term is ordered incorrectly, and the claimed optimum (30) does not follow from the printed finite-dimensional objective unless Q and Γ_r commute, which they generally do not. The correct expansion does lead to (30), so the central estimator is likely sound; this is a derivation error, not a fatal flaw. The source-free/Herglotz limitation could be tested separately, but it is not by itself a threat to the central claim. The published experiments and code, if reproducible, would also confirm whether the implemented estimator matches (30) rather than the printed (29).","tokens_in":20252,"tokens_out":28215,"duration_ms":324587,"concrete_test":"Re-derive (29) from (26) for M=1 without assuming QΓ_r=Γ_rQ. Then instantiate L=2 with κ0≠κ1, e.g. Γ_r=[[1.5,0.5],[0.5,1.5]], Q=diag(1,4), λ=1, h=(1,0)^T. Compute the minimizer of the printed objective (29) and compare with a_opt=(Γ_r+λQ^{-1})^{-1}h. They differ; with the corrected cross term ⟨Γ_r Q h,a⟩ they agree. This isolates whether the typo is the only obstruction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Expanding the weighted data term in (26) for u(r)=Σ_m Γ_r(r,r_m)a_m gives -2⟨Γ_r Q h,a⟩ for the linear term, not -2⟨Q Γ_r h,a⟩ as printed in (29). The two coincide only if Q commutes with Γ_r. In this construction Q is a time-domain diagonal envelope while Γ_r=F^{-1}Γ_r^freq F is a DFT-transformed frequency-diagonal kernel, so they do not commute (e.g. M=1, L=2 gives Γ_r∝[[κ0+κ1, κ0−κ1],[κ0−κ1, κ0+κ1]] and Q=diag(q0,q1)). Taking the gradient of the printed objective yields (Γ_r Q Γ_r+λΓ_r)a = Q Γ_r h, which does not factor to (Γ_r+λQ^{-1})^{-1}h; the stated optimum (30) follows only from the correctly ordered cross term. This is a concrete algebraic inconsistency in the derivation of the central data-weighted estimator. The experimental code may well implement (30), but the derivation as written cannot be followed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper generalizes single-frequency kernel ridge regression (KRR) sound field estimation to discrete-time sound fields. A time-domain reproducing kernel Hilbert space is constructed via the DFT of the frequency-domain plane-wave kernel, yielding a closed-form estimator (Eqs. (20), (24)). The framework is extended to include directional regularization and a time-domain data weighting (Eqs. (25)-(30)), with envelopes motivated by RIR delay and reverberation time. The method is evaluated on free-field, simulated room, and real MeshRIR data under four noise models, and the data weighting is shown to improve NMSE, especially when combined with directional weighting.","tokens_in":20590,"tokens_out":11206,"duration_ms":122995,"significance":"If the derivation is correct, this is a useful and nontrivial extension of the KRR sound field estimation framework: it moves from per-frequency processing to joint time-domain estimation, permits temporal priors, and provides closed-form solutions with physical wave-equation constraints. The paper is careful in its appendices: the kernel construction, the DFT convention, and the equivalence with the frequency-domain solution are all addressed explicitly. The availability of code and the use of real recorded RIRs are also strengths. The main caveats are the source-free/Herglotz model assumption and the heuristic choices of λ, β, τ_init, which are not the central claims of the paper. The data-weighting envelopes use physical priors (delay, RT60) rather than being fitted to the evaluation data, and the oracle envelope is explicitly an upper bound.","major_comments":[{"comment":"The cross term in the expansion of the weighted data term is mis-ordered. Expanding the first sum in (26) with u(r)=Σ Γ_r(r,r_m)a_m gives -2⟨Qh, Γ_r a⟩, which, because Γ_r is self-adjoint, equals -2⟨Γ_r Q h, a⟩, not -2⟨Q Γ_r h, a⟩ as printed. The matrices Γ_r and Q do not commute in general (already for M=1, L=2, Γ_r is a non-diagonal symmetric matrix while Q is diagonal). With the printed expression, the optimality condition becomes (Γ_r Q Γ_r + λΓ_r)a = QΓ_r h, which does not factor to (30). With the corrected cross term, the optimality condition is Γ_r Q Γ_r a + λΓ_r a = Γ_r Q h; for invertible Γ_r this is equivalent to (Γ_r + λQ^{-1})a = h and hence to (30). Thus the final estimator is correct for the intended objective, but the derivation as written is algebraically inconsistent and must be corrected.","section":"§IV-B, Eq. (29)"}],"minor_comments":[{"comment":"The inequality in the linear envelope definition is malformed: \"l0 ≥ l ≤ l0 + f_s τ_decay\" should presumably be \"l0 ≤ l ≤ l0 + f_s τ_decay\" (or the intended interval should be stated more carefully).","section":"§V-B, Eq. (37)"},{"comment":"The caption leaves it to the reader to infer which panel is simulated and which is real. Please state this explicitly in the caption.","section":"§VI-A, Fig. 2"},{"comment":"The statement that the equivalence with the frequency-domain method holds \"except for the Nyquist frequency\" is made before the assumption of odd L is introduced. It would be clearer to state the odd-L assumption at the start of the derivation and then discuss the Nyquist caveat.","section":"Appendix D"},{"comment":"The code link is given as a version-less GitHub URL. For reproducibility, a release or persistent identifier (e.g., Zenodo DOI) would be preferable.","section":"Code availability"}],"recommendation":"major_revision","confidential_remarks":"The only substantive technical issue is the algebraic error in Eq. (29), which is load-bearing for the derivation of the data-weighted estimator but appears to be a typo: the intended solution (30) is correct once the cross term is fixed. I would be willing to review a revised version that corrects this equation and adds a short derivation or counter-comment showing the corrected objective implies (30). The rest of the paper is careful and the experimental evaluation is convincing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something genuinely new: it constructs a time-domain RKHS for sound fields with a closed-form kernel, adds directional and temporal data weighting, and derives a closed-form estimator. The equivalence to frequency-domain KRR in the identity case (Appendix D) is a nice sanity check, and the experiments are broad—two simulated scenarios, real data, four noise models—with a clear, honest story that temporal weighting helps most in noise. The appendices are careful, and the code is public. The Nyquist-frequency caveat is acknowledged and is reasonable.\n\nThe main problem is in Section IV-B. Equation (29) has the cross term mis-ordered. Expanding the weighted data term gives ⟨Γ_r Q h, a⟩, not ⟨Q Γ_r h, a⟩. These are equal only if Q and Γ_r commute, which they do not in this construction (Q is a time-domain diagonal envelope; Γ_r is a DFT-transformed frequency-diagonal kernel). As printed, the gradient of (29) cannot be manipulated into the claimed solution (30). I checked: you get (ΓQΓ + λΓ)a = QΓh, which does not factor to (Γ + λQ^{-1})^{-1}h unless Q and Γ commute. So the derivation as written is not followable.\n\nThat said, the final estimator (30) is actually correct for the correctly ordered objective—the expansion with ⟨Γ Q h,a⟩ does give (30) when Γ is invertible. So this looks like a typographical error in an intermediate equation rather than a fatal flaw in the method. But because Eq. (29) is the central step for the weighted estimator, it has to be fixed before the paper is publishable.\n\nOther soft spots are minor: λ, β, and τ_init are heuristic, and the results show means without variance bars, so effect sizes are hard to judge. The oracle envelope is infeasible but used as an upper bound, which is fine.\n\nWho is this for? Researchers in sound field estimation, active noise control, and kernel methods in acoustics. The extension is useful and the empirical evidence supports it, despite the derivation slip.\n\nRecommendation: this deserves serious peer review, not a desk reject. The authors should be asked to correct the cross term, and perhaps add confidence intervals. The method itself is sound enough to engage with seriously.","headline":"Useful time-domain KRR extension with broad experiments, but the central derivation of the data-weighted estimator has a fixable ordering error in Eq. (29).","tokens_in":21026,"tokens_out":4423,"would_cite":true,"duration_ms":49400,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["43.60.+d"],"model":"deepseek-v4-flash","headline":"A time-domain reproducing kernel Hilbert space lets kernel ridge regression estimate whole room impulse response functions in one closed form, wave equation included.","keywords":["sound field estimation","kernel ridge regression","room impulse response","reproducing kernel Hilbert space","time-domain audio","directional weighting","data weighting","Helmholtz equation"],"falsifier":"Synthesize a noise-free sound field from a point source placed inside the estimation region and estimate it with many microphones; the plane-wave RKHS cannot represent the source term, so the error should plateau above zero and reveal the source-free assumption as the limiting factor.","tokens_in":20239,"feed_emoji":"🎧","tokens_out":6670,"duration_ms":71952,"temperature":0.7,"pith_summary":"The paper extends kernel ridge regression (KRR) sound field estimation from single-frequency fields to discrete-time sound fields. It constructs a reproducing kernel Hilbert space whose functions are discrete-time room impulse responses defined over a spatial region, with a closed-form kernel derived from the Herglotz plane-wave model (a superposition of plane waves) and the DFT. In this space, the optimal estimate is given in closed form and automatically satisfies the homogeneous wave equation, so no soft physics penalty is needed. The paper shows that weighting microphone data by a time envelope, built from propagation delay and reverberation time, improves estimation error in noisy conditions, and that this temporal weighting combines with directional weighting. If right, the method makes a broad class of time-domain sound field reconstruction problems solvable by KRR with exact physical constraints.","feed_headline":"Time-domain kernels estimate room impulse responses in closed form","feed_subtitle":"A time-aware weighting built from delay and reverberation time cuts estimation error in noisy rooms.","key_machinery":"The load-bearing object is the time-domain reproducing kernel Gamma-tilde(r,r') = F^{-1} Gamma(r,r') F = Re[BC Gamma(r,r') F], whose frequency-domain diagonal entries are spherical Bessel kernels j0(omega_l / c * ||r - r'||). It turns the wave equation from a constraint into a built-in property of the function space, so the infinite-dimensional optimization collapses to a finite linear system via the representer theorem. A secondary mechanism is the data-weighting envelope Q_tilde_m: a diagonal operator over time samples that encodes the expected signal-to-noise ratio of the RIR over time, such as the exponential envelope constructed from propagation delay and RT60.","core_discovery":"The paper's central claim is that time-domain sound field estimation can be solved exactly, in closed form, by kernel ridge regression. Starting from the usual per-frequency RKHS of solutions to the Helmholtz equation, the authors use the DFT to build a time-domain RKHS whose kernel is Gamma-tilde(r,r') = F^{-1} Gamma(r,r') F. The representer theorem then reduces the infinite-dimensional estimation problem to solving the linear system a_opt = (Gamma + lambda I)^{-1} h, or, with regularization and data weighting, a_opt = (Gamma_r + lambda Q^{-1})^{-1} h; the estimate at any position is a weighted sum of kernel blocks. The paper proves that with identity weighting this is equivalent to solving","pith_inferences":["Because the RKHS separates space and time through the DFT, a natural extension would be to build space-time kernels for moving microphones, letting one KRR problem use recordings taken at different positions and times jointly.","The envelope weighting could be learned from a first-pass estimate of the RIR instead of being fixed by RT60, giving an iterative estimator that adapts the temporal prior to the data.","The same construction likely extends to other wave models, such as spherical or higher-order basis functions, whenever a closed-form Herglotz-type kernel exists, though the paper does not derive those.","An immediate testable extension is to compare the exactly constrained time-domain estimator against physics-informed neural networks in low-data regimes; the closed form should win when the sound field is truly source-free and noise-dominated."],"forward_implications":["With identity weighting, the time-domain estimate equals the inverse DFT of per-frequency KRR estimates (except the Nyquist bin), so the method reduces to prior KRR when no time-domain prior is used.","Problems requiring joint treatment of frequencies, such as estimating full RIRs rather than single tones, become solvable by KRR with a closed-form solution.","A practical exponential envelope using only propagation delay and reverberation time nearly matches the oracle envelope, so the temporal weighting is implementable in real rooms.","Combining temporal data weighting with directional weighting improves NMSE in free-field, simulated reverberant, and real-room tests, especially at low SNR and at low frequencies.","The estimator remains linear in the data, so it can be embedded in spatial active noise control and sound zone control pipelines just like the single-frequency version."],"supporting_citations":[{"why":"single-frequency KRR constrained by the Helmholtz equation; the method this paper generalizes to time domain.","marker":"[27]"},{"why":"directionally weighted wave field estimation built on the Herglotz plane-wave integral; supplies the directional weighting and identity used for the closed-form regularized kernel.","marker":"[33]"},{"why":"environment-adapted physics-constrained kernel interpolation; supplies flexible directional weighting functions and numerical integration for kernels.","marker":"[34]"},{"why":"generalized representer theorem for Hilbert space-valued functions; justifies the closed form of the optimal estimator.","marker":"[49]"},{"why":"continuous-time KRR sound field estimation without a closed-form solution; positioned as the motivation for a discrete-time closed-form alternative.","marker":"[51]"},{"why":"RIR decay model based on propagation delay and RT60; provides the basis for the exponential data-weighting envelope.","marker":"[59]"},{"why":"image-source method used to generate the simulated room impulse responses in experiments.","marker":"[60]"},{"why":"MeshRIR dataset of real room impulse responses used for the real-data evaluation.","marker":"[62]"}],"fun_headline_variants":["Closed-form time-domain sound field estimation via kernel ridge regression","Time-domain kernels solve sound field estimation in closed form","Estimate room impulse responses with time-aware kernel weighting","Kernel method extends to time-domain for realizable sound field estimates","Time-domain weighting boosts kernel sound field estimation accuracy"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The sound field in the target region is source-free and can be represented as a superposition of plane waves through the Herglotz integral; if a source or evanescent field lies inside the region, the RKHS model is incomplete and the estimate can be biased.","fun_headline_variants_meta":{"raw":{"variants":["Closed-form time-domain sound field estimation via kernel ridge regression","Time-domain kernels solve sound field estimation in closed form","Estimate room impulse responses with time-aware kernel weighting","Kernel method extends to time-domain for realizable sound field estimates","Time-domain weighting boosts kernel sound field estimation accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000193,"raw_usage":{"total_tokens":1191,"prompt_tokens":750,"completion_tokens":441,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":494,"completion_tokens_details":{"reasoning_tokens":362}},"tokens_in":494,"tokens_out":441,"duration_ms":5078,"temperature":1.0,"reasoning_tokens":362,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T05:06:06.435695+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Synthesize a noise-free sound field from a point source placed inside the estimation region and estimate it with many microphones; the plane-wave RKHS cannot represent the source term, so the error should plateau above zero and reveal the source-free assumption as the limiting factor.","supporting_citations":[{"cited_title":"Kernel ridge regression with constraint of Helmholtz equation for sound field interpolation,","cited_arxiv_id":null,"evidence_quote":"single-frequency KRR constrained by the Helmholtz equation; the method this paper generalizes to time domain."},{"cited_title":"Directionally weighted wave field estimation exploiting prior information on source direction,","cited_arxiv_id":null,"evidence_quote":"directionally weighted wave field estimation built on the Herglotz plane-wave integral; supplies the directional weighting and identity used for the closed-form regularized kernel."},{"cited_title":"Sound field estimation based on physics-constrained kernel interpolation adapted to environment,","cited_arxiv_id":null,"evidence_quote":"environment-adapted physics-constrained kernel interpolation; supplies flexible directional weighting functions and numerical integration for kernels."},{"cited_title":"A Generalized Representer Theorem for Hilbert Space - Valued Functions","cited_arxiv_id":"1809.07347","evidence_quote":"generalized representer theorem for Hilbert space-valued functions; justifies the closed form of the optimal estimator."},{"cited_title":"Sound field estimation using deep kernel learning regularized by the wave equation,","cited_arxiv_id":null,"evidence_quote":"continuous-time KRR sound field estimation without a closed-form solution; positioned as the motivation for a discrete-time closed-form alternative."},{"cited_title":"Towards optimal regularization by incorporating prior knowledge in an acoustic echo canceller,","cited_arxiv_id":null,"evidence_quote":"RIR decay model based on propagation delay and RT60; provides the basis for the exponential data-weighting envelope."},{"cited_title":"MeshRIR: A dataset of room impulse responses on meshed grid points for evaluating sound field analysis and synthesis methods,","cited_arxiv_id":null,"evidence_quote":"MeshRIR dataset of real room impulse responses used for the real-data evaluation."}],"review_version":1}