acceptodds
Under review as a conference paper at ICLR 2027

Wave Physics as Representation: Tensor-Structured Neural Reconstruction for Ocean Acoustic Fields

Abstract

Reconstructing three-dimensional (3D) complex ocean acoustic fields from sparse measurements is challenging due to highly oscillatory wave structures and restricted sensing patterns/routes. Existing learning-based approaches generally fall into two categories. First, purely data-driven models can learn complex field patterns from data, but do not explicitly exploit the underlying wave physics that can potentially enhance interpretability and reduce training costs. Second, physics-informed methods impose a specified physical equation as a reconstruction constraint, but usually require the environment-dependent parameters to be known, which is difficult to acquire in practice. As a result, practical 3D acoustic field reconstruction that is both physically grounded and robust to uncertain environmental information remains largely unexplored. This work takes an alternative route: instead of solving a fully specified wave equation, it extracts a structured representation from the mathematical form of wave physics and learns its components from data. Starting from the Helmholtz equation and normal mode propagation, we derive a structured spatial–angular tensor representation, termed the representation. Under a low-rank specialization, it takes the form of a rank- block-term tensor decomposition, which connects structured sampling patterns/routes often used in ocean sensing practice to sufficient conditions for field recovery. Furthermore, building on this tensor representation, we propose TIDAL, a neural reconstructor that uses the representation as an inductive bias for structured decoding and aligns information aggregation with the sampling geometry. The neural learner removes the need for known model parameters required by conventional physics-based recovery methods. Experiments across three structured sampling geometries show that TIDAL achieves the best reconstruction performance and fastest inference among baselines, with clear benefits from the physics-derived representation under limited data and model capacity.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Ocean acoustic field reconstruction, Complex-valued tensor, LL1 decomposition, Structured sensing, Physics-inspired neural reconstruction

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_ULkGOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a method called TIDAL to reconstruct 3D complex ocean acoustic fields from sparse H/V observations. The main idea is to use wave physics to derive a spatial-angular tensor representation (S/c) representation. The authors further connect this low-rank form to LL1 tensor decomposition and give recovery conditions based on existing results. Based on this representation, they design a neural network with attention and a structured decoder. Experiments on simulated ocean acoustic fields show better reconstruction accuracy than several baselines.

Critical strengths
  1. The problem is meaningful. Reconstructing complex acoustic fields from sparse measurements has practical value.

  2. The paper is generally well organized. The motivation and the overall framework are easy to follow.

  3. The experimental results show improvements over the selected baselines.

Critical weaknesses
  1. Novelty is limited. The model applies existing tensor decomposition and neural network representation to a standard tensor recovery application.

  2. The recovery theorem is based on an exact low-rank LL1 decomposition. However, the real TIDAL model uses dense spatial components. This means the theoretical guarantee does not apply to the proposed model. The theory is also based on existing results.

  3. It is unclear whether the sampling conditions in the theory are satisfied in the actual experiments. As a result, I do not think the theoretical part provides a strong justification for the actual neural network. The theory and the implementation seem to be two somewhat separate parts.

  4. The proposed neural network learns both spatial and angular components freely. The angular factors are not fixed Fourier basis. So the actual decoder looks more like a general learnable tensor factorization. I am not convinced that the proposed representation contains a physics-specific insight.

  5. All experiments are conducted on simulated acoustic fields generated from a limited set of configurations. There is no convincing experiment on unseen ocean real environments. The proposed model uses information from multiple angular directions, while several baseline reconstruct individual angular slices. This is an unfair setting. Also, some baselines are adapted from quite different tasks. For example, MxT is designed for image inpainting only.

Questions for authors

My questions are listed in Critical Weaknesses.

4 comments

We thank the reviewer for the comments and the opportunity to clarify our contributions and theoretical motivation.

Q1. Novelty of the proposed method

Our main contribution lies in deriving a structured representation from wave propagation physics and incorporating it into neural reconstruction under sparse H/V observations. The LL1 representation was not obvious a priori; it was revealed by our derivation from the governing physics equations.

Specifically, the physics-derived S/c representation guides the design of the structured decoder, while the H/V sampling geometry motivates the sampling-aligned axial attention. These two components jointly adapt the reconstruction architecture to the physical field structure and observation geometry.

The ablation studies in the original paper further demonstrate the effectiveness of this design. Compared with direct-output decoding, the S/c decoder consistently improves reconstruction accuracy, with particularly clear advantages under limited training data and reduced model capacity. These results support the value of the proposed representation-guided reconstruction framework.

Q2–Q3. Scope of the recovery theory and its connection to TIDAL

The theoretical analysis addresses a fundamental question in practical ocean sensing: whether structured H/V observations can uniquely determine the full 3D acoustic field. Starting from the physics-derived S/c representation, we consider its low-rank LL1 specialization and build on established tensor recovery results to characterize sufficient recoverability conditions for complex acoustic fields under H/V sampling. This establishes a principled connection between the physical field structure, practical sensing geometry, and theoretical recoverability.

The H/V observation geometry used in our experiments follows the coupled slab-sampling formulation considered in the theory. Importantly, this H/V sampling geometry is motivated by practical ocean sensing, where measurements are collected along horizontal transects and vertical arrays rather than at freely selected spatial locations. The low-rank assumption is also motivated by the structure of acoustic fields. We conducted an additional experiment to examine the low-rank properties of the angular Fourier coefficient fields computed directly from all 80 complete reference fields. The results reveal substantial spectral concentration: spatial rank 5 retains 32.6% of the energy, while ranks 52 and 69 retain 90% and 95%, respectively, for 200 × 200 matrices. These findings provide empirical evidence of approximate low-rank structure, while the additional energy captured at higher ranks reflects the rich spatial variations in realistic acoustic fields. This structural characterization supports the relevance of the theoretical analysis and motivates a natural extension toward more flexible spatial modeling.

Accordingly, TIDAL retains the physics-derived S/c representation and learns dense spatial components that naturally exhibit low-rank structure, as evidenced by Fig. 3(a), while also capturing richer spatial variations without imposing an explicit rank constraint. The resulting architecture preserves the structured spatial–angular coupling and accommodates complex interference patterns, as further supported by the rank-constrained ablations in the original paper. The theoretical guarantee characterizes the exact low-rank specialization under its sufficient conditions, while the practical architecture extends the same representation to more complex acoustic fields. The scope of the theoretical conditions and their connection to the practical architecture are now clarified in Secs. 4.2 and 5, with supporting analyses in Sec. 6.3 (Fig. 3).

Q4. Physics-specific insight of the S/c representation

The S/c structure originates from the spatial–angular separation in wave propagation physics, providing a physically motivated organization of complex acoustic fields. The physical insight is incorporated through the construction of the representation, while the angular factors remain learnable to capture complex, environment-dependent angular variations.

To further examine this design, we conducted an additional Fixed-Fourier experiment, replacing the learnable angular factors with a prescribed Fourier basis while retaining the S/c decoding structure. The unobserved-region RMSE results are summarized below.

ModelEquispacedShiftedJittered
Fixed-Fourier0.48580.48860.5028
TIDAL0.26920.27440.2950

TIDAL consistently achieves better reconstruction accuracy across all three sampling geometries. These results demonstrate the benefit of data-adaptive angular factors within the physics-derived S/c representation, allowing the model to capture complex spatial–angular variations more effectively than the fixed-basis alternative. This comparison is now presented in Sec. 6.5 (Table 2).

Q5. 3D baselines and generalization

To provide direct comparisons with methods that jointly exploit all 72 angular directions, we introduced four additional 3D reconstruction baselines. These include two training-free methods, PRGD-Tucker [1] and ShearLab 3D [2], which perform three-dimensional low-rank tensor completion and shearlet-based sparse reconstruction, respectively. We also included two training-based generative methods, conditional 3D EDM and conditional 3D GAN [3]. All four methods reconstruct the complete 200 × 200 × 72 acoustic field, jointly utilizing the spatial and angular dimensions. We evaluated these methods under the same H/V sampling geometries as TIDAL. The results for the primary H15/V15 setting are summarized below.

Table 1. Reconstruction performance under H15/V15 sampling (Unobserved relative RMSE, mean ± std).

MethodEquispacedShiftedJittered
PRGD-Tucker [1]1.034 ± 0.0581.009 ± 0.0561.028 ± 0.064
ShearLab 3D [2]0.972 ± 0.0380.975 ± 0.0130.978 ± 0.012
Conditional 3D EDM [3]0.356 ± 0.0710.358 ± 0.0710.372 ± 0.075
Conditional 3D GAN [3]0.544 ± 0.0870.543 ± 0.0820.559 ± 0.081
TIDAL0.269 ± 0.0850.274 ± 0.0850.295 ± 0.089

We additionally examined H10/V10 and H20/V20 sampling to assess performance across different observation densities.

Table 2. Average unobserved relative RMSE across the three sampling geometries.

MethodH10/V10H15/V15H20/V20
PRGD-Tucker [1]1.11801.02340.9627
ShearLab 3D [2]0.99160.97510.9557
Conditional 3D EDM [3]0.41010.36200.3376
Conditional 3D GAN [3]0.62000.54870.5216
TIDAL0.33700.27950.2620

These results further demonstrate the effectiveness of TIDAL in comparison with methods that jointly model the complete 72-angle acoustic field.

Regarding generalization, our dataset is generated using the KRAKEN solver with realistic oceanographic and bathymetric information, spanning multiple source locations, frequencies, source depths, and propagation conditions. The simulation settings and evaluation scope are described in Sec. 6.1, while generalization to measured acoustic fields and unseen ocean environments is discussed as future work in Sec. 7. Additional reconstruction results are now reported in Table 1 and Appendix C (Table 3).

[1] Zhang Y, Bian F, Zhang X, et al. Preconditioned Riemannian gradient descent algorithm for low-multilinear-rank tensor completion[C]//Forty-second International Conference on Machine Learning. 2025.

[2] Kutyniok G, Lim W Q, Reisenhofer R. Shearlab 3D: Faithful digital shearlet transforms based on compactly supported shearlets[J]. ACM Transactions on Mathematical Software (TOMS), 2016, 42(1): 1-42.

[3] Oommen V, Khodakarami S, Bora A, et al. Learning turbulent flows with generative models for super resolution and sparse flow reconstruction[J]. Nature Communications, 2026, 17(1): 3707.

This reviewer appreciates the new experiments and clarifications and would like to increase the score to 2. However, this reviewer believes that the overall novelty and new conceptualization remain limited in this manuscript. It is essentially an application of some existing tensor models to a new type of data, with considerable engineering effort. I think a score of 2 is fair.

We thank the reviewer for reconsidering our work. We would like to further clarify the central conceptual contribution of this work, which may not have been sufficiently emphasized in our previous response.

Our approach does not start by applying an existing tensor decomposition to acoustic data. Instead, we begin with the governing wave physics and derive a structured tensor representation that reveals the underlying spatial–angular structure of acoustic fields. This physics-derived representation then serves as a principled inductive bias for neural reconstruction, directly guiding how the neural network organizes and reconstructs the acoustic field. In this way, the architecture is designed based on the physical structure revealed by the governing equations, providing a way to incorporate wave physics into neural reconstruction without requiring known environment-dependent physical parameters.

Therefore, our main contribution lies in establishing the connection from wave physics to structured tensor representation and then to neural architecture design, rather than simply applying existing tensor models to a new type of data.

Reviewer_2STVOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes TIDAL, a physics-inspired method for reconstructing 3D ocean acoustic fields from sparse observations. It uses a spatial–angular tensor representation (S/c), connects its low-rank form to LL1 recoverability theory, and combines axial attention with a structured decoder. The results show good reconstruction accuracy and inference efficiency.

Critical strengths

1. The idea of using wave physics to guide the representation, without requiring known environmental parameters, is interesting.

2. TIDAL generally performs better than the reported baselines across different sampling settings.

Critical weaknesses

1. Limited novelty and disconnect between theory and implementation.

The novelty is somewhat limited, as the S/c representation follows a Fourier expansion and the recoverability analysis builds on existing LL1 results. More importantly, the theory assumes low-rank components, while TIDAL uses dense learned components. For example, taking Q = 40 under H15/V15, Theorem 2 requires L ≤\leq 5, whereas Fig. 3(a) reports a 90%-energy rank of 53. Although these ranks are not directly comparable, it is unclear how much the theory explains the actual model. A comparison with fixed Fourier factors or other tensor-based decoders would help.

2. Incomplete experimental validation and reproducibility concerns.

The experimental evaluation could be stronger. Classical tensor recovery and compressive-sensing baselines are missing, and only one training seed is used. Also, MxT, LaMO, and DiffusionPDE process individual angular slices, while TIDAL uses all 72 angles jointly, which may make the comparison unfair. Some improvements over MxT are quite small. The normalization procedure is also unclear, particularly whether it uses the complete field at test time.

Questions for authors

Q1. What makes the S/c representation specifically physics-based rather than a general Fourier decomposition? Could the authors compare it with fixed Fourier or other factorized decoders?

Q2. How do the LL1 recovery guarantees relate to TIDAL, given that it uses dense spatial components instead of low-rank ones?

Q3. Could the authors include classical reconstruction baselines, multiple training seeds, and comparisons where the learning-based baselines can also use all 72 angles?

Q4. Is the 99.5th-percentile normalization calculated using only observed measurements, or does it require the full ground-truth field?

Q5. How does TIDAL generalize to unseen source locations, frequencies, and environments? Will the authors release code, model configurations, and the dataset-generation pipeline for reproducibility?

7 comments

We thank the reviewer for the constructive comments and suggestions. We address the questions below.

Q1. Physics-derived representation and Fixed-Fourier comparison

The S/c representation is derived from the spatial–angular separation inherent in wave propagation physics. The physical knowledge is incorporated at the representation level, providing a structured way to model the coupling between spatial acoustic patterns and angular variations.

To further examine the role of the angular representation, we conducted an additional experiment using a Fixed-Fourier variant, where the angular factors are prescribed by a Fourier basis rather than learned adaptively. The unobserved-region RMSE results are summarized below.

ModelEquispacedShiftedJittered
TIDAL0.26920.27440.2950
Fixed-Fourier0.48580.48860.5028

TIDAL consistently achieves lower reconstruction errors across all three sampling geometries. These results highlight the benefit of data-adaptive angular modeling within the physics-derived S/c structure, which allows the representation to better capture complex spatial–angular variations in acoustic fields. This comparison is now presented in Sec. 6.5 (Table 2).

Q2. Connection between the theoretical guarantees and the practical architecture

The theoretical analysis addresses a fundamental question in practical ocean sensing: whether structured H/V observations can uniquely determine the full 3D acoustic field. By considering a low-rank specialization of the physics-derived S/c representation, we establish sufficient recoverability conditions under this sampling geometry through coupled LL1 theory. This characterizes when structured H/V measurements contain sufficient information for full-field recovery.

The low-rank assumption is motivated by the structure of acoustic fields rather than introduced solely for theoretical convenience. We directly analyzed the angular Fourier coefficient fields computed from all 80 complete reference fields. The results are summarized below.

MetricValue
Energy retained at spatial rank (L = 5)32.6%
Spatial rank required to retain 90% energy52
Spatial rank required to retain 95% energy69

These results reveal substantial spectral energy concentration, providing empirical support for the approximate low-rank structure of the reference fields. The additional energy captured at higher spatial ranks further reflects the rich spatial variations in realistic acoustic fields, motivating more flexible spatial modeling. Accordingly, TIDAL retains the physics-derived S/c representation and learns dense spatial components that naturally exhibit low-rank structure, as evidenced by Fig. 3(a), while also capturing richer spatial variations without imposing an explicit rank constraint.

Thus, the theory characterizes recoverability for a low-rank specialization of the S/c representation, while the practical architecture preserves the underlying spatial–angular structure with greater expressive flexibility. This connection and the scope of the theoretical guarantee are now clarified in Secs. 4.2 and 5, with supporting empirical evidence in Sec. 6.3.

Q3. Additional baselines and robustness

Given the limited rebuttal period, we focused on evaluating TIDAL and MxT, the most closely related baseline, using three independent training seeds to assess the robustness of their reconstruction performance. The unobserved-region RMSE results under H15/V15 sampling are summarized below.

Sampling geometryTIDALMxT
Equispaced0.2656 ± 0.00410.3002 ± 0.0116
Shifted0.2709 ± 0.00420.3050 ± 0.0111
Jittered0.2916 ± 0.00430.3185 ± 0.0116

TIDAL consistently outperforms MxT across all three sampling geometries, achieving lower reconstruction errors and substantially smaller variation across training seeds. Moreover, as reported in Table 1 of the original paper, TIDAL achieves approximately 5.1× faster inference than MxT (0.14 s versus 0.72 s per field). These results demonstrate simultaneous improvements in reconstruction accuracy, training stability, and inference efficiency.

To strengthen the comparison with established reconstruction methods, we included PRGD-Tucker [1], a classical low-rank tensor completion method, and ShearLab 3D [2], a compressive sensing-based method. Both methods perform training-free reconstruction directly in 3D from sparse H/V measurements, rather than independently reconstructing 2D slices. We further included two learning-based baselines, conditional 3D EDM and conditional 3D GAN [3], which jointly model all 72 angular directions. All four additional methods were evaluated on the same 200 × 200 × 72 complex acoustic fields under identical H/V observation protocols.

Table 1. Reconstruction performance under H15/V15 sampling (Unobserved relative RMSE, mean ± std).

MethodEquispacedShiftedJittered
PRGD-Tucker [1]1.034 ± 0.0581.009 ± 0.0561.028 ± 0.064
ShearLab 3D [2]0.972 ± 0.0380.975 ± 0.0130.978 ± 0.012
Conditional 3D EDM [3]0.356 ± 0.0710.358 ± 0.0710.372 ± 0.075
Conditional 3D GAN [3]0.544 ± 0.0870.543 ± 0.0820.559 ± 0.081
TIDAL0.269 ± 0.0850.274 ± 0.0850.295 ± 0.089

We further evaluated these methods under H10/V10 and H20/V20 sampling to assess their performance across different observation densities.

Table 2. Average unobserved relative RMSE across the three sampling geometries.

MethodH10/V10H15/V15H20/V20
PRGD-Tucker [1]1.11801.02340.9627
ShearLab 3D [2]0.99160.97510.9557
Conditional 3D EDM [3]0.41010.36200.3376
Conditional 3D GAN [3]0.62000.54870.5216
TIDAL0.33700.27950.2620

TIDAL consistently achieves the lowest reconstruction errors across all sampling densities and geometries. These results demonstrate its effectiveness against both classical tensor/compressive sensing methods and learning-based reconstruction models that jointly utilize all 72 angular directions. Detailed reconstruction results are now reported in Table 1 and Appendix C (Table 3), with the training-seed comparison provided in Appendix F (Table 8).

Q4. Normalization protocol

For each acoustic field, the normalization scale is computed solely from its observed H/V measurements. This scale is used to normalize the input and subsequently rescale the reconstructed field. The same procedure is applied to training, validation, and test samples. No unobserved values or complete reference fields are used to determine the normalization scale. This normalization protocol is now detailed in Appendix B.1.

Q5. Generalization and code availability

Our dataset includes diverse acoustic conditions, covering multiple source locations, frequencies, source depths, and environmental configurations. TIDAL is trained on these varying conditions and achieves consistent reconstruction performance on test fields, demonstrating its ability to learn acoustic field structures across the conditions represented in the dataset.

Generalization to entirely unseen source locations, frequencies, or ocean environments is an important direction for further investigation. We consider this beyond the scope of the present work and plan to explore it in future research.

To facilitate reproducibility, we will publicly release the implementation, model configurations, and scripts for generating the acoustic-field dataset upon publication.

[1] Zhang Y, Bian F, Zhang X, et al. Preconditioned Riemannian gradient descent algorithm for low-multilinear-rank tensor completion[C]//Forty-second International Conference on Machine Learning. 2025.

[2] Kutyniok G, Lim W Q, Reisenhofer R. Shearlab 3D: Faithful digital shearlet transforms based on compactly supported shearlets[J]. ACM Transactions on Mathematical Software (TOMS), 2016, 42(1): 1-42.

[3] Oommen V, Khodakarami S, Bora A, et al. Learning turbulent flows with generative models for super resolution and sparse flow reconstruction[J]. Nature Communications, 2026, 17(1): 3707.

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

This paper studies the reconstruction of three-dimensional complex ocean acoustic fields from sparse horizontal and vertical transect observations. Starting from normal-mode propagation and angular Fourier expansion, the authors introduce a spatial–angular representation. With an additional low-rank assumption on the spatial components, they connect this representation to existing LL1 tensor recovery theory. The proposed network, TIDAL, uses axial attention to encode the observations and a structured decoder to predict spatial components and angular factors, which are combined to reconstruct the full field. Experiments on simulated data evaluate reconstruction accuracy, inference efficiency, and performance under different training-set sizes, model capacities, and spatial-rank settings.

Critical strengths

The paper addresses three-dimensional complex acoustic-field reconstruction under structured transect sampling. Recovering both the amplitude and phase of acoustic pressure from sparse observations is a challenging and practically relevant task.

The S/c decoder is motivated by modal propagation and angular Fourier expansion, and uses shared spatial components to organize field predictions across angles. In controlled comparisons using the same encoder and similar parameter counts, it outperforms direct-output decoding across different training-set sizes and model capacities. These results provide empirical support for the structured output design.

Under the evaluated simulation conditions, TIDAL achieves better reconstruction accuracy than the selected baselines across all three sampling geometries at the default sampling budget. It also reports fast inference, showing a favorable combination of accuracy and efficiency.

Critical weaknesses

It is unclear to me what role the low-rank recovery analysis in Section 4.2 plays in the final network. That section assumes that each spatial component Sq has rank at most L, and discusses unique recovery of the full field when additional conditions hold. However, in Section 5.2, the network directly predicts dense spatial components without explicitly enforcing this rank constraint. Even accepting the low-rank assumption itself, I do not see how the recovery result applies to the implemented network. Section 6.3 and Appendix D examine the singular values of learned components, full-field fitting errors under different spatial ranks, and reconstruction performance with factorized versus dense decoders. These experiments indicate some compressibility in the learned components and support the choice of a dense decoder, but do not establish whether the actual network satisfies the recovery theorem’s conditions. The paper should clarify what the theory establishes for the proposed method. If it only characterizes an idealized special case, this scope should be stated more clearly.

The authors motivate the low rank of Sq by suggesting that modal contributions across angles may be strongly correlated, but I could not find a direct validation of this property for the angular Fourier coefficient fields. An individual directional slice can be represented by a finite number of modes, but a weighted combination of slices from different angles does not necessarily preserve the same low rank. It would be helpful to provide supporting theoretical or empirical evidence, or to compute the angular Fourier coefficient fields directly from complete reference fields and examine their singular-value spectra and approximation errors at different ranks. This would test whether the low-rank assumption holds for the fields under study more directly than analyzing the spatial components learned by the network.

Questions for authors

The recovery theory in Section 4.2 requires low-rank spatial components, whereas the network in Section 5.2 does not explicitly enforce this constraint. Which conclusions of the theory still apply to TIDAL? If the theory only establishes recoverability for an idealized special case, what is its specific connection to the final network design?

Beyond the singular-value analysis of learned neural components, is there theoretical or existing empirical evidence supporting the low rank of the angular Fourier coefficient fields defined in Section 4.1? Could the authors compute these coefficient fields from complete reference fields and report their singular-value spectra or low-rank approximation errors to assess this assumption?

3 comments

We thank the reviewer for the positive assessment and constructive suggestions. We address the two questions below.

Q1. Connection between the theoretical analysis and the TIDAL architecture

The theoretical analysis addresses a fundamental question in practical ocean sensing: whether structured H/V observations can uniquely determine the full 3D acoustic field. By considering a low-rank specialization of the physics-derived S/c representation, we establish sufficient recoverability conditions under this sampling geometry through coupled LL1 theory. This provides a theoretical characterization of when structured H/V measurements contain sufficient information for full-field recovery.

Importantly, the low-rank assumption is motivated by the structure of acoustic fields rather than introduced solely for theoretical convenience. As detailed in our response to Q2, our analysis of the angular Fourier coefficient fields computed directly from 80 complete reference fields reveals substantial spectral concentration, with 90% of the energy retained at spatial rank 52 for 200 × 200 matrices. However, these fields also contain complex spatial variations that cannot be adequately represented by very low ranks. This motivates TIDAL to retain the physics-derived S/c representation while learning dense spatial components. As demonstrated by the learned-component spectra in Fig. 3(a), TIDAL naturally captures low-rank structure from data without explicitly enforcing a prescribed spatial rank, while also accommodating richer spatial variations.

The original rank-constrained ablations further support this design choice, showing that dense spatial components substantially improve reconstruction accuracy over explicitly rank-constrained alternatives. Thus, the theory characterizes recoverability for a tractable structural specialization, while TIDAL preserves the underlying spatial–angular representation and extends its expressive capacity to realistic acoustic fields. The scope of the theoretical guarantee and its connection to the practical TIDAL architecture are now clarified in Secs. 4.2 and 5, with supporting empirical evidence in Sec. 6.3.

Q2. Low-rank properties of the angular Fourier coefficient fields

We directly analyzed the singular-value spectra of the angular Fourier coefficient fields computed from all 80 complete reference fields. The results are summarized below.

MetricValue
Energy retained at spatial rank (L=5)32.6%
Spatial rank required to retain 90% energy52
Spatial rank required to retain 95% energy69

These results reveal substantial spectral energy concentration in the reference Fourier coefficient fields, providing direct empirical evidence of their approximate low-rank structure. The additional energy captured at higher spatial ranks further reflects the rich spatial variations in realistic acoustic fields, motivating a more flexible representation to preserve these details.

This observation supports both the practical relevance of the low-rank assumption and the motivation for adopting dense spatial components in TIDAL. The corresponding singular-value spectra is now presented in Sec. 6.3 (Fig. 3).

Thank you very much for your detailed and constructive review. As the discussion period is approaching its end, we wanted to kindly follow up. We have updated our rebuttal and revised manuscript to address the concerns and questions you raised. We would greatly appreciate it if you could let us know whether our responses have addressed your main concerns or if any further questions remain. We would be happy to provide additional clarification. Thank you for your time and consideration.

Reviewer_NgHmOpenReview ↗

Comments by Reviewer NgHm

I thank the authors for the additional experiments and clarification. The analysis of Fourier coefficient fields computed from complete reference fields provides direct evidence of their approximate low-rank structure, addressing my concern about the lack of validation. The response also clarifies that the recovery guarantee applies to an idealized low-rank case rather than directly to TIDAL. Although the connection to the actual network remains limited, these additions help clarify the scope of the theory. I will maintain my score of 3 (weak acceptance).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.