{"id":"f8f831f5-ff3e-480a-aa75-ef1e87559e54","arxiv_id":"2506.13279","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Using a rough point cloud of room boundaries as a Bayesian prior improves simulated sound field reconstruction from microphone measurements.","lead":"This paper introduces a Bayesian method that uses rough 3D room boundary information, such as a point cloud from a smartphone camera, to improve sound field reconstruction from microphone measurements. A generalist reader may care because it could make accurate spatial audio control possible with cheap, imperfect room scans.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Tikhonov and Lasso baseline tuning is not specified; the reported gains may stem from the empirical Bayes hyperparameter optimization rather than from the boundary-informed prior itself.","rationale":"The reader's weakest_assumption focused on the single-impedance β assumption, which I agree is a scope limitation. However, the paper explicitly acknowledges that the ISM simulation with reflection coefficient 0.95 does not exactly satisfy the impedance boundary condition, so the numerical setting already tests robustness to model mismatch of the prior. The more pressing concern is that the experimental comparison may be confounded: the authors never state how the Tikhonov σ² and σ²_α are set when µ=0, nor how the Lasso λ is chosen. Since the proposed method optimizes four hyperparameters jointly, its improvement over baselines could be due to the additional model-fitting flexibility rather than to the boundary-informed structure. This is a correctable but load-bearing issue because the central claim is explicitly empirical and comparative. A concrete re-run with properly tuned baselines would settle it. The reader's overall CONDITIONAL verdict remains appropriate; I do not propose a change, but I would require the baseline specification as a condition of acceptance.","tokens_in":8108,"tokens_out":4547,"duration_ms":49479,"concrete_test":"Re-run the Figure 2 experiment with the Tikhonov estimator's σ² and σ²_α estimated by the same marginal-likelihood maximization used for the proposed method but with µ fixed to 0, and with the Lasso λ selected by k-fold cross-validation. Report the resulting NMSE curves for all methods, ideally with error bars over the 10 Monte-Carlo runs. If the proposed method's NMSE advantage over the tuned Tikhonov baseline shrinks to within the Monte-Carlo uncertainty, the central claim that the boundary-informed prior itself enhances reconstruction is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is comparative: the boundary-informed prior 'significantly enhances the reconstruction' relative to Tikhonov and Lasso. In Section IV, the Tikhonov estimator is described only as 'constructed by setting µ = 0 in (16)', with no statement about how σ² and σ²_α are chosen, and the Lasso regularizer λ is never specified. If the Tikhonov baseline uses a poorly chosen or fixed regularization strength, the proposed method's advantage may be an artifact of the additionally optimized hyperparameters (a, b, d, η) rather than of the boundary information encoded in Σ_α. The paper also omits any cross-validation or grid-search procedure for Lasso, making the comparison incomplete. Because the abstract makes a quantitative claim of 'significant enhancement', the experiment must isolate the effect of the prior; as reported, it does not. The uniform-β assumption is a real scope limitation, but the simulation already deliberately violates the exact impedance model (reflection coefficient 0.95), so the prior is stress-tested in that respect; the baseline tuning issue directly threatens the validity of the numerical evidence supporting the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Bayesian estimator for reconstructing a frequency-domain sound field in an enclosed space from microphone measurements and partial knowledge of the room boundary given as a point cloud. The sound field is represented by a Herglotz/plane-wave expansion, and a Gaussian prior for the expansion coefficients is derived from impedance boundary conditions, with covariance depending on the matrix (beta Psi + Phi_tilde). The hyperparameters, including noise variance, signal variance, prior weight, and the complex specific impedance beta, are estimated by minimizing the marginal likelihood using gradient-based optimization. Numerical experiments with ISM-simulated shoebox rooms compare the proposed estimator with Nearest-neighbor, Tikhonov, and Lasso baselines, reporting NMSE versus the number of boundary points, boundary-position perturbation, microphone-position perturbation, and frequency. The central claim is that the boundary-informed prior significantly improves reconstruction, even with only a few hundred boundary points or with boundary-point errors up to 1 dm.","tokens_in":8465,"tokens_out":7685,"duration_ms":79065,"significance":"If the numerical evidence is reliable, this is a useful contribution: it provides a principled way to inject coarse and uncertain geometry into sound-field reconstruction, with all hyperparameters estimated from the measured data, and it gives a concrete target for calibration accuracy when using consumer-grade point-cloud sensors. The derivation in Section III is clear and internally consistent, and the evaluation is deliberately unfavorable to the proposed model because the ISM simulations use a reflection coefficient of 0.95 rather than exactly satisfying the impedance boundary condition. The robustness experiments in Figures 3 and 4 address a realistic source of model mismatch. The main weaknesses are that the baseline comparisons are incompletely specified and that the Monte Carlo results are reported without uncertainty, so the strength of the central comparative claim cannot yet be fully assessed.","major_comments":[{"comment":"The comparison that supports the central claim is incomplete. The Tikhonov estimator is described only as 'constructed by setting mu = 0 in (16)', with no statement of whether sigma^2 and sigma_alpha^2 for this baseline are estimated by the same marginal-likelihood procedure as the proposed method or are fixed. The Lasso estimate is introduced as adding lambda times the L1 norm of alpha to (7), but the value of lambda, or any grid search or cross-validation used to select it, is never reported. If the baselines use fixed or poorly chosen regularization while the proposed method optimizes the hyperparameters in (18), the reported gain could be caused by the empirical Bayes hyperparameter optimization rather than by the boundary information encoded in the prior. Please specify the baseline tuning procedures and, ideally, report results over a range of regularization strengths.","section":"Section IV, Tikhonov and Lasso baselines"},{"comment":"The NMSE in Eq. (20) is averaged over J validation points and N=10 Monte-Carlo simulations, but Figures 2-5 show only the mean curves, with no error bars, confidence intervals, or per-run spread. With N=10 and a single room configuration, the differences between the proposed method and Tikhonov/Lasso could lie within run-to-run variability, particularly in the regimes where the curves are close in Figures 3-5. Please add error bars or shaded intervals and state which random quantities vary across the N runs (noise, boundary perturbations, microphone perturbations).","section":"Figures 2-5 and Eq. (20)"},{"comment":"The marginal-likelihood objective in (18) is non-convex, but the paper does not report the initialization of the conjugate-gradient optimization, the use of restarts, or any sensitivity analysis. In addition, the reparametrization beta = e^eta with complex eta does not enforce a physical constraint such as a positive real part for the specific impedance, so the optimizer may select non-physical boundary parameters. Because the reported results depend on the quality of the local optimum reached, please report the initialization and add multiple restarts or a sensitivity study.","section":"Section III-B, Eq. (18)"}],"minor_comments":[{"comment":"The notation 'J(theta) = arg min ...' is an abuse of notation; please define J(theta) as the objective after reparameterization and state the minimization separately.","section":"Section III-B, Eq. (18)"},{"comment":"The numerical setup omits several quantities needed for reproducibility: room dimensions, source position and source signals, the exact sampling procedure for the microphones and validation points, and the generation of boundary and microphone perturbations. Please provide these details or make the code available.","section":"Section IV, experimental setup"},{"comment":"The paper does not state whether the plotted NMSE values are in dB; Eq. (20) defines a linear average. Please clarify the vertical-axis scaling.","section":"Figures 2-5"},{"comment":"At B=0, the covariance in (16) reduces to the Tikhonov covariance with mu=0; please confirm that the Proposed and Tikhonov curves coincide at that point, or explain any discrepancy.","section":"Figure 2"},{"comment":"There are minor typographical issues, including 'distribtuion' in Section III and 'significance value' in Section IV (should be 'significant value'), and 'Pola k-Ribi`ere' should be 'Polak-Ribiere'.","section":"Throughout"},{"comment":"The Nearest baseline is very weak; adding a non-boundary-informed kernel or Gaussian-process baseline without boundary information would make the comparison more informative.","section":"Section IV"}],"recommendation":"major_revision","confidential_remarks":"The core idea is sound and within the journal's scope. The main blockers are the missing specification of baseline regularization (especially Lasso lambda and the Tikhonov hyperparameters) and the absence of uncertainty information in the Monte Carlo comparisons. These are fixable in revision; I do not see grounds for rejection. No citation-pattern concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about this one. First, the central idea is genuinely neat: instead of needing full room geometry, the authors derive a Gaussian prior on plane-wave coefficients from impedance boundary conditions at a few hundred point-cloud points, with uniform complex impedance and boundary weight treated as hyperparameters. Second, the numerical evidence is under-specified enough that the abstract's claim of 'significantly enhances' is plausible but not established.\n\nWhat's actually new: previous work uses either full simulation or generic GP regularization; this specific construction—Σα = σ²α (I + µ(βΨ+Φ̃)ᴴ(βΨ+Φ̃))⁻¹—does not appear in the cited literature, and the joint empirical Bayes optimization of noise, signal, µ, and β is a natural and useful extension. The derivation in Section III is internally consistent, and the appendix gives explicit gradients. Credit where due: the simulation deliberately makes the evaluation unfavorable by using reflection coefficient 0.95, which does not exactly match the impedance model for oblique incidence, so the uniform-β assumption is partially stress-tested. The citation pattern looks reasonable, covering the relevant GP and impedance estimation work.\n\nSoft spots, in order of importance. The baseline comparison is not fully specified. Tikhonov is defined as µ=0 in (16), but the paper does not state how σ² and σ²α are chosen; if they are not also estimated from the marginal likelihood, the proposed method's advantage may come from the extra hyperparameter optimization rather than from the boundary information. The Lasso λ is never stated. There are no error bars on the N=10 Monte Carlo runs in Figures 2–5, so apparent differences of a few dB may not be significant. The evaluation is limited to one shoebox geometry with simulated data; no real point clouds, no code or data released. The uniform-β and plane-wave model are limitations, but the authors are upfront that this is preliminary work.\n\nOverall, the math looks solid, the idea is worth building on, and the paper deserves a serious referee. The missing tuning details and error bars need to be addressed before the comparative claim is taken at face value.\n\nRecommendation: send to peer review with a request to clarify baseline tuning, add error bars, and ideally release code/data.","headline":"Clean derivation of a boundary-informed Gaussian prior that deserves a serious referee, but the numerical comparison is under-specified: Tikhonov and Lasso baselines need explicit tuning before the headline claim is established.","tokens_in":8831,"tokens_out":4060,"would_cite":true,"duration_ms":38154,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes a Bayesian estimator that turns a coarse point cloud of room boundaries into a prior for plane-wave coefficients, and claims this boundary-informed prior improves sound-field reconstruction even with only a few hundred…","keywords":["sound field reconstruction","Bayesian estimation","impedance boundary conditions","point cloud prior","Herglotz wave functions","spatial audio","hyperparameter optimization","room acoustics"],"falsifier":"In a room with one heavily absorbing wall and three hard walls, run the same 100-microphone, 300 Hz experiment from Section IV with a point cloud of the true geometry; if the boundary-informed estimator does not beat Tikhonov because the spatially varying absorption is not represented in the prior, the single-impedance assumption is the breaking point.","tokens_in":7888,"feed_emoji":"🎧","tokens_out":6778,"duration_ms":59379,"temperature":0.7,"pith_summary":"The paper sets out to show that a room's sound field can be reconstructed more accurately by feeding a statistical estimator a rough point cloud of the room's boundaries, rather than ignoring the geometry as standard regularized methods do. The proposed linear Bayesian estimator builds a prior for plane-wave coefficients from the impedance boundary condition evaluated at the point-cloud locations, and it tunes the impedance and variance hyperparameters automatically. Numerical simulations with 100 microphones and a 1,000-plane-wave basis show the boundary-informed prior lowers normalized mean-squared error compared with nearest-neighbour, Tikhonov, and Lasso baselines. The claimed benefit appears with only a few hundred boundary points and survives boundary-position errors up to about 1 dm.","feed_headline":"Rough room scans cut sound-field reconstruction error below baselines","feed_subtitle":"A few hundred coarse wall points, off by 1 dm, beat standard Tikhonov and Lasso reconstruction.","key_machinery":"The load-bearing object is the boundary-constraint matrix pair $(\\Psi, \\tilde{\\Phi})$ sampled at the point-cloud locations, combined with a shared complex impedance $\\beta$. It enters through the prior covariance $\\Sigma_\\alpha = \\sigma_\\alpha^2 (I + \\mu(\\beta\\Psi+\\tilde{\\Phi})^H(\\beta\\Psi+\\tilde{\\Phi}))^{-1}$, which softly enforces the impedance boundary condition rather than imposing it as a hard constraint. Setting $\\mu=0$ recovers the Tikhonov estimator, so any improvement is attributable to the boundary term. The marginal-likelihood objective and its Wirtinger gradients, given in the appendix, let the solver jointly tune the noise variance, signal variance, boundary-prior strength, and impedance.","core_discovery":"The central claim is that a partial, imprecise boundary model is enough to regularize the ill-posed sound-field reconstruction problem. At a fixed frequency, the field is represented as a superposition of plane waves $\\alpha$, measured as $y = \\Phi\\alpha + \\varepsilon$, and the paper assigns $\\alpha$ a Gaussian prior with covariance $\\Sigma_\\alpha = \\sigma_\\alpha^2 \\left(I + \\mu(\\beta\\Psi + \\tilde{\\Phi})^H(\\beta\\Psi + \\tilde{\\Phi})\\right)^{-1}$, where $\\beta\\Psi + \\tilde{\\Phi}$ encodes the impedance boundary condition $\\beta\\, n(r)\\cdot\\nabla u(r) + i k u(r)=0$ at the boundary points. Using this prior in the maximum-a-posteriori estimate (or the equivalent Gaussian-process prediction) is what produces the reported gains; the hyperparameters $\\sigma^2$, $\\sigma_\\alpha^2$, $\\mu$, and $\\beta$ are jointly estimated by maximizing the marginal likelihood. The paper further claims the approach is robust to coarse, uncertain boundary data, specifically a few hundred point-cloud samples and up to 1 dm position error.","pith_inferences":["An extension the authors leave implicit: the same prior could be tested with per-patch impedance values instead of one global $\\beta$; if per-patch $\\beta$ gives further gains, real rooms with mixed surfaces become tractable.","The 1 dm robustness result suggests consumer photogrammetry point clouds are already precise enough for the prior, which would let spatial audio systems be calibrated from a quick phone scan rather than a full room survey.","The formulation is basis-agnostic: any basis satisfying the Helmholtz equation could carry the same boundary-informed covariance, so the method could be combined with sparse or dictionary bases when the field is directional."],"forward_implications":["With 100 microphones and 1,000 plane-wave basis functions in a simulated room, the proposed estimator lowers NMSE by several dB compared with Tikhonov and Lasso baselines across the tested frequencies.","The reconstruction error falls sharply as boundary points are added, saturating near 1000 points, so a few hundred wall samples already capture most of the benefit.","The gain over baselines persists when boundary positions are perturbed up to 1 dm and when microphone positions are also uncertain, so precise geometric calibration is not required.","Because the Tikhonov estimator is recovered by setting $\\mu=0$, the improvement is attributable entirely to the boundary information entering the prior covariance.","The joint marginal-likelihood optimization automatically selects the impedance and variance hyperparameters, so no manual tuning of the boundary term is required."],"supporting_citations":[{"why":"Supplies the impedance boundary condition $\\beta n\\cdot\\nabla u + iku = 0$ that the prior enforces at wall points.","marker":"[23]"},{"why":"Provides the Gaussian-process regression equations and marginal-likelihood framework used for prediction and hyperparameter estimation.","marker":"[27]"},{"why":"Defines Herglotz wave functions, the plane-wave superposition basis used to represent the field.","marker":"[25]"},{"why":"Establishes Gaussian processes for sound-field reconstruction, the class of estimators this prior extends.","marker":"[3]"},{"why":"Supplies the image-source-room simulator used to generate the numerical experiments and baselines.","marker":"[32]"},{"why":"Provides the Bayesian sound-source reconstruction formulation used to justify the joint hyperparameter optimization.","marker":"[29]"}],"fun_headline_variants":["Partial wall scans sharpen sound field maps","Coarse room scans beat dense mic grids","Boundary priors rescue sound field reconstruction","A few wall points fix sound field estimates","Imprecise room scans boost audio reconstruction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes all boundary points share a single complex impedance value and a locally reacting surface, so a real wall that absorbs differently at different positions or angles violates the prior.","fun_headline_variants_meta":{"raw":{"variants":["Partial wall scans sharpen sound field maps","Coarse room scans beat dense mic grids","Boundary priors rescue sound field reconstruction","A few wall points fix sound field estimates","Imprecise room scans boost audio reconstruction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000459,"raw_usage":{"total_tokens":2325,"prompt_tokens":995,"completion_tokens":1330,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":611,"completion_tokens_details":{"reasoning_tokens":1274}},"tokens_in":611,"tokens_out":1330,"duration_ms":10669,"temperature":1.0,"reasoning_tokens":1274,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:05:29.586416+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In a room with one heavily absorbing wall and three hard walls, run the same 100-microphone, 300 Hz experiment from Section IV with a point cloud of the true geometry; if the boundary-informed estimator does not beat Tikhonov because the spatially varying absorption is not represented in the prior, the single-impedance assumption is the breaking point.","supporting_citations":[{"cited_title":"Kuttruff, Room acoustics , CRC Press, 2016","cited_arxiv_id":null,"evidence_quote":"Supplies the impedance boundary condition $\\beta n\\cdot\\nabla u + iku = 0$ that the prior enforces at wall points."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines Herglotz wave functions, the plane-wave superposition basis used to represent the field."},{"cited_title":"Gaussian processes for so und field reconstruction,","cited_arxiv_id":null,"evidence_quote":"Establishes Gaussian processes for sound-field reconstruction, the class of estimators this prior extends."},{"cited_title":"Room impulse response generator,","cited_arxiv_id":null,"evidence_quote":"Supplies the image-source-room simulator used to generate the numerical experiments and baselines."},{"cited_title":"A bayesian approach to sound source reconstr uction: Optimal basis, regularization, and focusing,","cited_arxiv_id":null,"evidence_quote":"Provides the Bayesian sound-source reconstruction formulation used to justify the joint hyperparameter optimization."}],"review_version":1}