acceptodds
Under review as a conference paper at ICLR 2027

Motion Closes the Rank Gap: Bayes-Optimal Mobile Sound-Field Estimation

Abstract

Recovering a room's sound field from microphone recordings is basic to room acoustics and spatial audio. Most methods improve the estimator, with better priors, kernels and learned physics models. Unfortunately, a static array of m microphones resolves at most m independent directions of a K-mode field, so for K>m, absent sparsity, no estimator, noise level or recording length recovers the rest. In contrast, we treat the loss as one of acquisition. Moving the same array makes all K directions identifiable at equal observation time. Under a Bayes-optimal model its risk then falls inversely with observation time, with a constant we derive in closed form as a trace. That constant is governed by a contraction factor, which approaches one in some rooms as K grows. Therefore we prove which rooms keep it bounded away from one for every K. Facets in all four axis directions, each at least one grid spacing longer than the aperture, suffice, and the length threshold is sharp. Simulations with no fitted parameter match the predicted constant, and at the longer observation times the moving array beats the best static position across the room under independent estimators as well as our own. On a real 16-microphone array in a furnished room, under a pre-registered protocol, 16 surveyed positions resolve more independent directions than any static position can, and the held-out error falls from one position to 16 and ends below the best-predicted static position.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
sound field reconstruction, room acoustics, moving microphone array, inverse problems, optimal experimental design

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Ri8TOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies how moving a fixed-size microphone array changes the identifiability and estimation error of a room sound field. Under a known-mode linear Gaussian model, it establishes a rank obstruction for a static array, derives the asymptotic Bayes risk of a moving array, and relates uniform conditioning to geometric coverage.

Critical strengths
  • The separation of the three issues is clear.
  • It provides the quantitative risk characterization as well as the interpretable geometric result.
  • Overall it provides a theoretical foundation to understand the mobile sound field estimations.
Critical weaknesses
  • Proposition 1 assumes that both the coefficient vector and the observation operator at a fixed position remain unchanged across snapshots. Repeated measurements then cannot increase rank. A model with known modal dynamics, varying excitation, or cross-frequency constraints could yield different equations over time even with a stationary array. The impossibility claim should be tied explicitly to the physical measurement setting for which this observation model is appropriate.

  • Proposition 2 concerns the worst-conditioned direction over all modal dimensions, whereas the risk constant in Theorem 1 depends on a spectral trace at a particular (K). A small uncovered region can destroy the uniform guarantee without necessarily causing a large error in the frequency range of interest. The paper should state more precisely what follows when the coverage condition fails, especially for finite (K).

  • The Neumann modes can represent wave behavior associated with boundaries included in the assumed room geometry. The real-room reconstruction, however, uses a rigid-box modal basis that omits the furnishings. Diffraction, scattering, or boundary-condition errors arising from unmodeled features may produce position-dependent residuals rather than the independent noise assumed in Theorem 1. The paper should discuss whether such residuals create a non-vanishing reconstruction bias and which part of the claimed (1/T) error decay survives model mismatch.

  • Can you elaborate on a practical procedure to estimate B and d_eff directly from data when the room modes are not known, e.g., via sample Gram matrices across positions or random probing?

Questions for authors

See above.

3 comments

Reviewer Ri8T

We thank the reviewer for the careful reading. Changes in the revised PDF are mint, or sky blue when another reviewer shares the concern.

W1 Scope of the impossibility claim. Agreed. The physical setting is a steady field from a fixed source, so the coefficients stay constant during recording. Proposition 1 needs those coefficients free in ℝ^K and the same spatial operator at every snapshot. With known dynamics, a known excitation model, or cross-frequency constraints, the bound need not apply. Sections 3 and 4 now say so. (also mentioned in Reviewer i9Us, W1.1)

W2 When coverage fails at finite K. At finite K, the limit in Theorem 1 still holds. Every room where the array fits has γ_B(K) < 1 (Lemma 2), so the constant is finite. What fails is a bound on γ_B that holds for all K at once. The horizon factor (1 - γ_B)⁻² then has no uniform bound. In Table 10 the rise in d_eff grows with the uncovered volume. Room 13 fails the coverage test. Its γ_B still sits inside the range of covered rooms at all three tested K (Table 7). A 45° rotation of the array removes coverage in every room of Table 10. There, d_eff rises by 4% in the faithful-scale box, with 0.05% uncovered. It rises by 78% in the cube, with 2.1% uncovered. A remark after Proposition 2 now says this. The abstract and Section 8 now say that γ_B bounds the constant.

W3.1 Bias from unmodeled features. Yes, it does not vanish. The estimator explains part of the unmodeled field with the modeled modes, and more data does not undo that. In a synthetic check the risk levels off at the predicted floor. Details are in W3.2. (also mentioned in Reviewer eD6R, W1, and Reviewer i9Us, Q5)

W3.2 What survives of the 1/T decay. The noise part decays as 1/T with the constant of Theorem 1. The unmodeled modes add a second term, which decays to a nonzero floor. The total error stops at that floor. A synthetic check, which is not in the paper, confirms this. With 50 modeled and 110 true modes, the moving risk levels off at 0.444, against a predicted floor of 0.445. It also shows a cost. With unmodeled energy at 12% of the modeled energy, the moving array has higher risk than the best static position at T = 500 snapshots. It has lower risk on every path from T = 2000 on. Section 8 now states the floor and this early-time cost.

Q1 Estimating B and d_eff from data. A procedure exists for a covered room, with sample Gram matrices as the reviewer suggests. The procedure has three steps.

  1. Record at least K fields with independent coefficients at the same positions, for example from different sources.
  2. Take the top K left singular vectors of the stacked data. They give a basis at the sampled points.
  3. Form two Gram matrices of that basis over the samples. The plain matrix M estimates what the array observes. The matrix R, weighted by 1/κ at each sample, estimates the room's Gram matrix. Geometry gives κ. Then d_eff = m tr(M⁻¹R) and 1 - γ_B = λ_min(R⁻¹M)/m.

With the modes hidden in a synthetic box room, 200 fields and noise, single runs recover d_eff to within 0.8% and γ_B to within 0.04 at 20000 snapshots. The procedure needs K as an input and sees nothing inside an uncovered set. The real-room data hold too few positions for it, so we leave it as future work. (also mentioned in Reviewer eD6R, W1)

Thanks for your prompt reply. After reading the responses, I am largely satisfied with the authors’ responses to my concerns and will maintain my positive rating.

Thank you for reading our response and for the constructive review. Your questions on finite K and on model mismatch made the paper clearer. We appreciate your time and effort!

Reviewer_eD6ROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies sound-field reconstruction using a moving microphone array, showing how measurements collected at different positions can overcome the rank limitations of a static array. The authors establish (i) the rank limitations of static arrays and derives asymptotic Bayesian risk guarantees for spatially distributed measurements and (ii) geometric conditions ensuring uniform conditioning as the number of modes increases. The theoretical results are supported by numerical simulations and real-room experiments.

Critical strengths

S1. The connection between geometric coverage and uniform conditioning of Proposition 2 and Theorem 2 is interesting.

S2. The paper includes extensive theoretical analysis, numerical experiments, and a real-world validation.

S3. The authors provide detailed experimental protocols for the real-room setup and acknowledge the limitations of their approach.

Critical weaknesses

W1. The practical applicability relies on a room-specific set of ϕ\phi, assumed known. But moving to a different room requires recomputing the eigenfunctions, while uncertain boundary conditions or furnishings may significantly alter the original conditions. The authors don't provide any comment or sensitivity analysis addressing this dependency.

W2. The experimental setup differs substantially from the theoretical one.

--W2.1 The real-room rank experiment combines responses at different frequencies, whereas the theory assumes a single sound field with fixed modal coefficients cc. The observed rank increase therefore does not directly validate the theoretical model.

--W2.2 The array is kept at a fixed height and orientation, with measurements collected at discrete stationary positions, whereas the theoretical model considers trajectories exploring a three-dimensional domain.

Despite these gaps, do the real-room results still validate the proposed geometric coverage and uniform-conditioning guarantees? The authors should clarify this point.

W3. It seems the authors overclaim their novelty: indeed, the static rank limitation follows directly from linear algebra, spatially distributed acquisition is well established, and the asymptotic (1/T) rate is expected in linear-Gaussian estimation. The novelty seems to be related to the theoretical geometric results, that are interesting but should be more clearly distinguished from these standard observations.

W4. What is the practical value of the theoretical guarantees? How do the geometric results improve acquisition design compared with existing approaches? Especially given the gaps between theoretical assumptions and approximations in the practical experiments. For instance, the theory relies on reflected Brownian motion, while the simulations show that realistic movement constraints can substantially affect performance.

Questions for authors

The paper contains potentially valuable mathematical results, particularly the connection between spatial coverage and uniform conditioning. However, the central acquisition insight is not particularly novel, and the practical implications of the geometric analysis is not discussed by the authors.

My main concern is the gap between the theoretical assumptions and the experimental validation. The method assumes a known room-specific modal basis, while the real-room experiments do not seem to establish identifiability under the fixed-coefficient model nor directly test the geometric conditioning guarantees.

Focus should be more clearly devoted on their geometric contributions, evaluate robustness to environmental variations, and strengthen the real-room validation with fixed-frequency experiments, and show more comprehensive spatial reconstruction measurements.

3 comments

Reviewer eD6R

We thank the reviewer. The main concern is the gap between theory and experiment. The real room shows a rank above the static limit of 16, at an empirical noise level, and lower held-out error at 16 positions than at one. Changes are yellow in the revised PDF, or sky blue when shared.

W1 Room-specific basis and sensitivity. The dependence is partial. The reconstruction and our d_eff values use the basis, so a new room needs a new one. The rank crossing reads only the data matrix, and the coverage test reads only the room geometry and the array. The reconstruction already runs with a wrong basis, a rigid box for a furnished room. We also scaled the model box by up to 5% (Appendix G.3). Under the second prior of Figure 6, full-set error stays at 0.34 to 0.43, against 0.87 to 1.22 for the static position. Under the declared prior it is 0.38 to 0.65, against 10 to 15. We changed the model box, not the room, so boundary conditions and furnishings remain untested. (also mentioned in Reviewer i9Us, Q5, and Reviewer Ri8T, W3.1 and Q1)

W2.1 and Q3 Rank across frequencies. It does not test the fixed-coefficient model directly. It counts the spatial directions the positions observe. One source at one frequency gives one column, so that rank is one by construction. If mode shapes do not change with frequency and the field repeats across takes, the crossing implies at least 17 directions at one frequency. The reconstruction is the fixed-frequency test, because it fits each line separately. The full set beats one position and the static position on at least 34 of 35 lines, in both sessions and under both priors. A rank test at one frequency needs several source positions and is a next measurement. (also mentioned in Reviewer i9Us, Q3)

W2.2 Height, orientation and discrete positions. Fixed orientation matches the model. Discrete positions fit the observation model, because risk depends only on the positions visited. They do not meet Assumption 1. Fixed height also departs, because the array stays in one plane. The array is at rest in each take, so noise during motion is unmeasured.

W2.3 Do the real-room results validate coverage? No, and Section 7.3 now says so. Sixteen positions at one height do not cover a room. Theorem 2 rests on its proof, on exact coverage computation over sampled rooms, and on the closed-form room family of Figure 10. The constant of Theorem 1 is likewise tested in synthetic rooms only.

W3 and Q1 Novelty and focus. Agreed. The rank bound, the 1/T rate, and the value of distributed acquisition are standard. What is new is the coverage identity and the four-facet certificate. Existing design methods optimize for one room and one model size (Section 2). The certificate bounds conditioning for every model size at once, from the room geometry alone. The contribution list and Section 8 now lead with it. (also mentioned in Reviewer i9Us, W1.3)

W4 Practical value. The coverage test tells before recording whether conditioning stays bounded as the model grows. Section 8 now says so. It flags room 13, which passes every finite-K diagnostic. The four-facet condition is a quick sufficient check. Past the horizon, the constant of Theorem 1 gives the recording time for a target error in the model. A real room adds a bias floor. They do not select an optimal path, and we ran no head-to-head comparison with design methods. Reflected Brownian motion is a reference, not a requirement. In simulation, a raster path at equal speed comes within a factor of 1.5 of the constant. A random walk at walking speed stays far above it (Table 13).

Q2 Environmental variation. We tested this only in part. One room limits what the hardware shows. It is enough for the rank result, because the limit of 16 holds for any static 16-microphone array in any room. Han & Kim (2026) show by an aperture-to-wavelength bound that a compact static array resolves only a few directions, in rooms from 1 to 61 cubic meters. One room is not enough to claim the same error reduction elsewhere, and the paper does not claim it. Room diversity enters on the theory side, with exact coverage verdicts on 91 sampled rooms (Table 9). Within this room, the two sessions agree.

Q4 Are the spatial measurements comprehensive? Not fully. The main score uses four held-out array positions, 64 microphones, each at least 1 m from every fitting microphone. In session 2 the full-set error there is 0.40 under the second prior and 0.63 under the declared prior. The revision adds 13 more positions, 208 microphones, 12 to 45 cm from the nearest fitting microphone (Appendix G.3). There the full-set error is 0.07 and 0.09. One position gives 0.72 and 2.1, and the static position gives 0.72 and 2.2. The full set is lower than both at all 13 positions and on all 35 lines. A denser grid far from the fit needs a new session.

Thanks for your reply. After reading the responses, my concerns have been addressed but the contribution still seems to be insufficient to increase my score. So, I'll keep my previous score.

Thank you for reading our responses and for the follow-up. We are glad the concerns are addressed, and we appreciate the time you gave to the paper.

Reviewer_i9UsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies sound-field estimation using a rigid microphone array translated through a room. Under a finite-dimensional modal model with time-invariant coefficients and independent Gaussian observation noise, a static array has an information matrix of rank at most the microphone count. For stationary reflected Brownian sampling, the authors establish an asymptotic Bayes risk proportional to (T^{-1}), with a constant expressed through the inverse average information matrix. They connect uniform conditioning across modal dimensions to a spatial coverage count and give a sufficient geometric condition for convex prismatic rooms. Validation includes finite-element simulations, comparisons with published estimators, public-data diagnostics, and measurements at multiple surveyed positions in one furnished room.

Critical strengths
  1. The connection between spatial coverage and uniform conditioning is useful and mathematically substantive. The identity
[
\sup_K\gamma_B(K)=1-\kappa_{\min}/m
]
provides a clear interpretation of the information operator. Together with the Minkowski-sum construction in Theorem 2, it explains when conditioning remains controlled as the model dimension increases. This is the strongest contribution: it goes beyond showing that additional measurement positions improve reconstruction and identifies a geometric mechanism governing that improvement.
  2. The evaluation includes meaningful checks and discloses several limitations. The paper distinguishes its idealized stochastic trajectory from a physically speed-matched path, reports independently tuned baseline results in the appendix, and acknowledges that the hardware experiment does not validate the asymptotic constant. The use of spatially separated held-out probes and a second recording session also strengthens the empirical evidence for an acquisition benefit.
Critical weaknesses
  1. The central identifiability claim needs a narrower scope, and Lemma 1 lacks a necessary hypothesis.

The static rank obstruction holds when a time-invariant coefficient vector is repeatedly observed through the same spatial operator, but is not a general impossibility result for stationary microphones. Known temporal dynamics or distinct modal frequencies can provide additional independent equations at fixed locations. This restriction should be explicit in the abstract and introduction, particularly given the frequency-dependent experiments.

Lemma 1 (p. 5) is also incomplete. Incident outward normals lying in one closed coordinate quadrant ensure compatibility with incident edges but not clearance from nonincident edges. Consider a 2×22\times2 array with s=1s=1 and

P=conv⁡{(0,0),(1/2,0),(5/2,3),(−2,3)}.P=\operatorname{conv}\{(0,0),(1/2,0),(5/2,3),(-2,3)\}.

The placement domain is nonempty, as an array anchored at (0,1)(0,1) fits. At (0,0)(0,0), the outward normals (0,−1)(0,-1) and (−3,−2)(-3,-2) satisfy the quadrant condition, yet covering this vertex would require including (1,0)∉P(1,0)\notin P. Thus the vertex and its neighborhood remain uncovered. Appendix B.2 recognizes the missing nonincident-facet clearance condition, which should be incorporated into the lemma.

Although this does not invalidate Theorem 2, the paper should better distinguish its geometric contribution from the elementary static rank bound and standard Gaussian inverse-information asymptotics.

  1. The hardware validation and moving-versus-static comparison require stronger support.

Appendix G.2 uses Weyl’s inequality to justify rank crossing. However, for Y=S+EY=S+E, proving s17(S)>0s_{17}(S)>0 requires s17(Y)>∥E∥2s_{17}(Y)>\|E\|_2. A separate silence recording provides an empirical noise reference, not a guaranteed bound on the signal recording’s noise. Statistical calibration is needed, especially when the column-normalized margin is only about 1.7. Moreover, rank above 16 in a stacked microphone-frequency matrix demonstrates greater spatial response diversity, not necessarily identifiability of every coefficient of a single target field.

The primary baseline comparison also relies on shared regularization that Appendix F.2 acknowledges disadvantages the static Brunnström estimator. Independently tuned static errors are approximately 0.77, 0.72, and 0.67 at 1, 3, and 10 seconds, versus moving errors of 0.821, 0.685, and 0.359 in Table 2. The short-time ranking therefore reverses, while the intermediate-time advantage narrows. These results belong in the main comparison. Additionally, the static location is optimized for the authors’ estimator rather than each competing method.

These concerns do not negate the benefits of multiple positions, but weaken the claimed strict hardware validation and the reported superiority over static baselines.

Questions for authors
  1. Do the authors agree that Lemma 1 requires an additional clearance condition? Please provide a corrected statement and explain which downstream claims rely on the stronger version.
  2. How is the noise norm from the silence recording converted into an upper confidence bound for the noise in the signal recording? Please report the calibration procedure, confidence level, and treatment of column normalization and signal-dependent disturbances. Would the rank-crossing conclusion remain valid under that bound?
  3. What exactly do the rows and columns of the hardware rank matrix represent? What assumptions connect its rank to the modal observation operator at a fixed frequency? Which identifiability claims can be supported without those assumptions?
  4. Can the authors provide a main-result table with independently selected regularization for both acquisition settings, together with clarification of whether the static position is optimized for each estimator? Please distinguish parameters selected using validation data from oracle diagnostic tuning.
  5. The real-room reconstruction changes substantially when the prior variance is multiplied by (10^{-3}). How stable is the acquisition advantage over a prespecified range of prior strengths, and how much of the poor static performance is attributable to prior/model mismatch rather than the rank obstruction?
1 comment

Reviewer i9Us

We thank the reviewer. The review is right on Lemma 1, the noise bound, and the static baseline. Changes in the revised PDF are lavender, or sky blue when another reviewer shares the concern.

W1.1 Scope of the rank claim. Agreed. Proposition 1 needs time-invariant coefficients, free in ℝ^K, and the same spatial operator at every snapshot. With known dynamics or cross-frequency constraints, the bound need not apply. The abstract, Section 1 and Section 4 now say so. (also mentioned in Reviewer Ri8T, W1)

W1.2 and Q1 Lemma 1. Agreed, it needs a clearance condition. A vertex neighborhood is coverable only if both outward edge normals lie in one closed coordinate quadrant. It is coverable if, in addition, the array placed with its matching corner microphone at the vertex lies in P and touches no non-incident facet. The reviewer's polygon fails that second condition. No other claim relies on the stronger version. Theorem 2 uses only the slicing part, and the covered rooms of Appendix D come from exact computation.

W1.3 Geometry versus the standard parts. Agreed. The rank bound is linear algebra, and the 1/T rate is the standard Gaussian limit. What is new is a condition on the room that bounds conditioning for every model size at once. The contribution list now leads with it. (also mentioned in Reviewer eD6R, W3)

W2.1 and Q2 Noise bound for the rank crossing. The silence norm was an uncalibrated proxy, and that was a gap. The revision estimates the noise from the signal recordings (Appendix G.2). Differences of disjoint period pairs cancel the periodic signal and leave noise. We scale them to the two-period mean and stack them like the data. Column normalization divides by the data's column norms. There is no formal confidence level. Of 4000 draws per session, none reaches s17, the 17th singular value. The margin of s17 over the largest draw is 2.5 to 3.1 on the raw matrix and about 1.3 on the normalized one. Against silence it was 6 to 7 and 1.7 to 4.0. So the crossing holds for measurement noise at this empirical level. The draws include non-periodic disturbances. Periodic distortion cancels and counts as part of the field. Slow change between takes needs a separate control. Repeats of one position give s17 at 0.0005 of the largest singular value, against 0.013 at 16 positions. That control covers one position only. It does not rule out a contribution at 16 positions.

Q3 What the rank matrix shows. Rows are microphone-position pairs, 256 at 16 positions. Columns are the 35 scored frequency lines. Two assumptions link its rank to the modal operator. Mode shapes do not change with frequency, and the field repeats across takes. Then the noiseless matrix is S = ΦC, so rank(Φ) ≥ 17. That supports at least 17 identifiable directions at one frequency, not identifiability of every coefficient. Without those assumptions, the data show only that the responses span more dimensions than one 16-microphone placement records. (also mentioned in Reviewer eD6R, W2.1)

W2.2 and Q4 Tuned static baseline. Agreed. The shared regularizer rule disadvantaged the static arm. For the estimator of Brunnström et al., each arm now selects its regularizer on three validation fields with new sources and noise. Validation also selects the static position, among 3559 candidates. Table 2 now reports these values, and Table 17 lists every setting. Static error is 0.764, 0.707 and 0.663 at 1, 3 and 10 s. Moving error is 0.821, 0.685 and 0.359. So static leads at 1 s. At 3 s moving is only 3% lower, a near tie on this one field. At 10 s moving leads by a factor of 1.8. The oracle row is chosen on the test field and is a diagnostic. At 3 s it is slightly below the moving value. Our own estimator keeps its declared prior, and its moving arm is lower at all three times.

Q5 Prior strength. The full position set has lower held-out error than the static position at every prior tested (Table 20). The range spans six decades. We set it after the reviews, before running the sweep. The ratio of static to full-set error is 1.25 to 74 across three passes. Most of the poor static result depends on the prior, not on the rank. A prior does not change what the array measured, so the fall in static error from 44 to 0.87 is not a rank effect. We attribute it to a wide prior fitting the mismatch between the rigid box and the furnished room. The sweep does not separate that from noise amplified in weakly observed directions. At each array's best prior, the errors are 0.87 static and 0.33 at 16 positions. (also mentioned in Reviewer eD6R, W1, and Reviewer Ri8T, W3.1)

One further correction, marked gray. The submission said the room was noisier than the count prediction assumed. It was quieter, so the measured counts equal the noise-free prediction and do not show extra room structure (Table 1, Appendix G.4).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.