acceptodds
Under review as a conference paper at ICLR 2027

HarmoCore: Functional Latent Diffusion for Sparse Reconstruction of Oscillatory Wave Fields

Abstract

Reconstructing oscillatory wave fields from scattered sensors is a severely underdetermined inverse problem. Beyond the challenges of general physical-field reconstruction, wave responses are complex-valued, frequency-sensitive, and highly oscillatory, while costly simulation and sensing often leave only extreme-sparse observations. Existing low-rank, operator, and diffusion approaches are largely designed for real-valued, smoother fields; dense pixel-space diffusion is particularly inefficient for oscillatory complex fields and difficult to scale to 3D. We propose , which places a generative prior in a compact, continuous, and structured wave-field latent. HarmoCore represents joint real–imaginary channels with Functional Tucker cores over shared continuous spatial bases, learns a frequency-conditioned core diffusion prior, and performs Diffusion Posterior Sampling directly in core space. At fixed sensor coordinates, the multilinear decoder induces an explicit likelihood guidance operator, avoiding dense pixel-space correction. Optional target-equation residual guidance further promotes physical consistency. Experiments on 2D Helmholtz, 2D synthetic wave fields, and 3D Helmholtz show substantial gains under – sensing while remaining practical in three dimensions.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Sparse Wave-Field Reconstruction, Functional Tucker Decomposition, Latent Diffusion Models, Physics-Informed Generative Modeling, Inverse Problems

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_m9onOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed HarmoCore to reconstruct oscillatory wave fields from scattered sensors using shared continuous spatial bases and sample-specific Functional Tucker cores. In particular, it proposed a frequency-aware, compact representation of complex oscillartory wave fields based on functional Tucker models. Then it trains a frequency-conditioned diffusion prior over normalized cores and reconstructs fields through observation guidance in core space, optionally supplemented by Helmholtz-equation residual guidance. Experiments show improvements under 1%–2% sensing.

Critical strengths

It reports significantly better results in 2D and 3D field reconstruction using the proposed method.

Critical weaknesses

Both the 2D and 3D fields are pretty simple and symmetric. It seems that they are generated from single source and single frequency. It would be good to consider a variety of fields with different sparsity levels.

Questions for authors
  1. What is the main benefits of using stacked real and imaginary channels over using stacked amplitude and phase channels?
    2.How about the performance of field reconstruction generated by multiple sources and multiple frequencies? Is it necessary to train a different model for a different source and frequency?
  2. Does diffusion outperform regularized inference in the same frozen basis when observations and physical constraints are held fixed?
  3. Figures 5–7 label rows as amplitude and phase, while their captions describe real and imaginary components.
4 comments

We sincerely thank you for the careful review and constructive feedback. We particularly appreciate the valuable suggestions, which have helped us improve our experiments and clarify the contributions and limitations of our work.

R1 — Multi-Source Fields, Multiple Frequencies, and Sensing Ratios

We thank you for highlighting the importance of field diversity. Our benchmarks include 1–4 sources in 2D Helmholtz, 1–3 in 3D Helmholtz, and 1–3 in synthetic fields, with randomized source positions and complex phases. Figures 7–9 provide multi-source examples.

The benchmarks cover 51, 41, and 17 frequencies for 2D Helmholtz, 3D Helmholtz, and synthetic fields, respectively. For 2D Helmholtz, 11 frequencies are used for training and 40 for interpolation. A single frequency-conditioned model is used across all frequencies; no separate model is trained per frequency.

We also added an independent-source evaluation (Appendix C.7). The 20 test configurations have unseen source-position/complex-amplitude combinations relative to the 500 training configurations. Without retraining, HarmoCore achieves 0.03656±0.033920.03656\pm0.03392 Rel. L2 at 5% sensing over 1,020 source-frequency cases, and 0.03482±0.030140.03482\pm0.03014 over the 800 interpolated-frequency cases.

The main results cover 1%, 2%, and 5% sensing, with additional 10% results in Appendix B.1. Our frequency-generalization claim is limited to interpolation within the evaluated range.

R2 — Why Real/Imaginary Channels Instead of Amplitude/Phase?

We use real and imaginary components as the modeling channels because they provide a natural Cartesian representation of the complex wave field.

For a complex field,

u=uR+iuI,u=u_R+iu_I,

the amplitude and phase are

\qquad \phi=\arg(u_R+iu_I).$$ There are two main reasons for choosing the real/imaginary representation. First, the complex Helmholtz equation naturally couples the real and imaginary components. Using these components directly allows the governing-equation residual to be formulated and differentiated without introducing additional transformations. Second, phase is a periodic variable with a branch discontinuity at the phase boundary and becomes poorly defined near zero amplitude. Modeling real and imaginary components avoids explicit phase unwrapping and provides a continuous Euclidean representation for diffusion training. We emphasize that the joint-versus-independent ablation in Table 7 evaluates the benefit of jointly modeling the two Cartesian components. It is not a direct experimental comparison between real/imaginary and amplitude/phase representations. We therefore present the above as the motivation for our representation choice rather than claiming that its superiority over amplitude/phase has been experimentally established. For visualization, the qualitative figures display amplitude and phase derived from the reconstructed real/imaginary channels. ## R3 — Does Diffusion Outperform Regularized Inference in the Same Frozen Basis? We added two complementary comparisons. **Learned versus Gaussian prior.** Table 11 replaces only the learned diffusion score with a conditional Gaussian score, keeping the frozen FTM, observations, source, initialization, normalization, guidance, and sampling fixed. At 1%/10% sensing, the learned score achieves Rel. L2 errors of 0.130/0.031, versus 0.284/0.149 for the Gaussian score. **HarmoCore versus fixed-basis LS.** Table 12 compares HarmoCore with observation-only least squares using the same frozen spatial factors. On 20 independent source configurations, their Rel. L2 errors at 1%/2% sensing are 0.128/0.067 and 0.750/0.500, respectively. At 5%, they are comparable (0.038 vs. 0.039), while at 10% LS performs better (0.011 vs. 0.032). These results indicate that the learned prior is particularly beneficial under extreme sparsity, but does not always outperform LS at higher sensing rates. We acknowledge that LS does not use PDE guidance; the Gaussian-score comparison provides the more controlled comparison under identical physical guidance. ## R4 — Amplitude/Phase and Real/Imaginary Figure Labels We thank you for identifying this inconsistency. The model is trained on joint Re/Im channels, while the figures visualize the derived amplitude and phase. We have corrected the row labels and captions to clarify this distinction.

Thank you very much for your detailed and constructive review. As the discussion is approaching its end, we wanted to kindly follow up. We have updated our response/paper to address the concerns and questions you raised (please see our latest reply/update). We would greatly appreciate it if you could let us know your current thoughts and any further questions or concerns you may have. We remain happy to clarify or provide additional information. Thank you for your time and consideration.

Thank you again for your review. As the discussion is closing soon, I wanted to kindly follow up. We would greatly appreciate any feedback you may have. Thank you for your time.

Thank you again for your review. As the discussion deadline is approaching, we would be grateful for any final comments or confirmation you may have. Thank you for your time.

Reviewer_g1NfOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HarmoCore for reconstructing complex-valued oscillatory wave fields from sparse observations. It combines a Functional Tucker representation with shared continuous spatial bases, a frequency-conditioned latent diffusion model, and observation-guided posterior sampling with optional PDE-residual guidance. Experiments on 2D Helmholtz, synthetic wave fields, and 3D Helmholtz report substantial improvements over several baselines under extremely sparse observations.

Critical strengths
  1. Strong empirical performance. The method achieves substantial improvements over several reconstruction baselines, particularly under 1% observations.

  2. Useful continuous representation. Functional Tucker provides a compact latent representation with continuous-coordinate decoding and an explicit observation operator, offering potential advantages for irregular sensing and posterior reconstruction.

Critical weaknesses
  1. Limited novelty relative to prior work. The core framework is highly similar to SDIFT, which already combines Functional Tucker representations, latent diffusion, and observation-guided posterior sampling. The main differences are frequency conditioning, joint real–imaginary modeling, and PDE-residual guidance. These modifications appear incremental, and the paper does not compare against SDIFT or adequately isolate their contributions through ablations. Consequently, the novelty beyond adapting an existing framework to oscillatory wave fields remains unclear.

  2. Insufficient experiments of the representation and its claimed advantages. A central motivation in the Introduction is that conventional representations and diffusion methods are less suitable for complex, highly oscillatory fields. However, the paper does not directly compare Functional Tucker against standard autoencoder-based latent diffusion under controlled settings, nor does it demonstrate that the proposed representation is particularly advantageous for oscillatory rather than smoother fields. Furthermore, despite emphasizing continuous-coordinate reconstruction and computational efficiency, all experiments use grid-sampled observations and provide no evaluation of off-grid reconstruction or runtime and memory advantages. Finally, several neural-operator baselines exhibit relative errors near or above 1, raising concerns about whether these methods are adequately adapted.

Questions for authors
  1. How does HarmoCore compare with SDIFT, particularly under the same representation capacity and observation protocol?

  2. Can the authors provide ablations isolating frequency conditioning, joint versus independent real–imaginary modeling, and core-space diffusion against direct optimization of the same Functional Tucker representation?

  3. Given that the Helmholtz operator and source are available at inference, what advantages does the proposed reconstruction offer over directly solving the known PDE? Can the method handle heterogeneous or partially unknown media?

  4. Can the authors report inference time and memory consumption, particularly for larger 3D grids, and clarify the relatively poor performance of several neural-operator baselines?

5 comments

We sincerely thank you for the thoughtful and constructive comments. We particularly appreciate the suggestions regarding the novelty of our method, the evaluation of the proposed representation, and the practical scope of the framework. These comments have helped us substantially improve the experimental evaluation and clarify the contributions and limitations of our work. We address each concern below.

R1 — Novelty Relative to SDIFT

We acknowledge that HarmoCore shares key components with SDIFT, including Functional Tucker representations, diffusion priors, and observation-guided posterior sampling. We do not claim these general ingredients as novel. Our contribution is their adaptation and validation for oscillatory, frequency-dependent complex wave fields, requiring joint complex-channel modeling and physics-constrained inference.

We added a matched SDIFT adaptation (Appendix C.4, Tables 9–10). Using the same frozen representation, training data, comparable model capacity, and PDE information, HarmoCore achieves Rel. L2 errors of 0.130/0.068 at 1%/2% sensing, versus 0.509/0.454 for SDIFT. Ablations further isolate the contributions of joint complex modeling and the learned diffusion prior.

We revised Related Work and the contribution statements to clarify this distinction. Our contribution is a validated framework for sparse complex wave-field reconstruction, rather than a new general Functional Tucker diffusion paradigm.

R2 — Representation Advantages, Off-Grid Reconstruction, and Computational Efficiency

Pixel-space diffusion.

We added a full-field pixel-space diffusion baseline (Appendix C.5), using a matched denoiser capacity and training budget. Both methods use identical observations and physical guidance at inference. At 1%/2% sensing, HarmoCore achieves Rel. L2 errors of 0.130/0.068, versus 0.843/0.757 for pixel-space diffusion. This supports the advantage of compact functional cores over direct pixel-space diffusion in our setting. We acknowledge that this is not a direct comparison with autoencoder-based latent diffusion.

Off-grid reconstruction.

Appendix C.8 evaluates displaced sensor coordinates using frozen continuous bases, without retraining. Independent fine-grid PDE solutions provide references.

SensingGridOff-grid
1%0.163690.16241
5%0.041270.04102

The comparable errors demonstrate the feasibility of off-grid reconstruction, although broader sensor geometries remain to be investigated.

Runtime and memory.

Appendix C.9 reports prior-training times on 5,500 fields/cores: HarmoCore 559.5 s, SDIFT 465.4 s, and pixel-space diffusion 8,693.2 s. These exclude FTM fitting and core extraction.

For a 32332^3 3D case, inference takes 4.455±0.0074.455\pm0.007 s on an RTX 5070 Ti (16 GB VRAM), with 2,158.7 MiB peak allocated and 2,656.0 MiB reserved memory. Basis and operator preparation are reported separately. Larger-resolution scaling remains future work.

Neural-operator baselines.

Appendix A.6 now details FNO/F-FNO architectures and training protocols. These models use sparse observed-value supervision. At 1% sensing, 2D FNO achieves an observed-point Rel. L2 of 0.042 but a full-field error of 2.427, illustrating the difficulty of recovering unobserved oscillatory structures. This result reflects our sparse-supervision setting, not an inherent limitation of neural operators.

R3 — Direct PDE Solvers and Unknown Media

We agree that directly solving the PDE is natural when the complete operator, source, medium, and boundary conditions are known. HarmoCore targets a different problem: posterior reconstruction from sparse measurements, using a learned prior over wave-field responses with available physical information as guidance.

Appendix C.3 evaluates incomplete-physics settings, including unknown source amplitudes, source width, constant wave speed, and PML information. For example, with unknown source amplitudes, residual guidance reduces the 1%-sensing Rel. L2 error from 0.431 to 0.145.

These results do not establish generalization to arbitrary heterogeneous media or unrestricted operator mismatch. Our benchmarks assume homogeneous media, and heterogeneous-media reconstruction remains future work.

Thank you very much for your detailed and constructive review. As the discussion is approaching its end, we wanted to kindly follow up. We have updated our response/paper to address the concerns and questions you raised (please see our latest reply/update). We would greatly appreciate it if you could let us know your current thoughts and any further questions or concerns you may have. We remain happy to clarify or provide additional information. Thank you for your time and consideration.

Thank you again for your review. As the discussion is closing soon, I wanted to kindly follow up. We would greatly appreciate any feedback you may have. Thank you for your time.

Reviewer_E5RBOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

HarmoCore is a functional latent diffusion framework for sparse wave field reconstruction. It represents the amplitude and phase of each field as joint channels of a compact Functional Tucker core over shared continuous spatial bases. The bases capture continuous coordinate dependence, while the core retains compact, sample- and frequency-specific coefficients. A frequency-conditioned diffusion model is trained on these cores, and Diffusion Posterior Sampling is performed directly in core space. At fixed sensor coordinates, the multilinear decoder induces an explicit likelihood guidance operator, avoiding dense pixel-space correction. The framework also includes optional governing-equation guidance. Experiments cover 2D Helmholtz, 2D synthetic wave fields, and 3D Helmholtz, and sensing covers 1%–2% field. The method relies on globally learned spatial parameterization and on first obtaining a sufficiently expressive FTM representation.

Critical strengths
  1. The paper shows a meaningful and challenging problem. It presents a clear framework that combines compact wave-field representations with diffusion-based reconstruction. One outstanding point is the shared continuous spatial bases induce an explicit linear observation operator on the core at fixed sensor coordinates, which is reused across reverse steps.
  2. The experiments cover 2D and 3D settings and provides several useful analyses. The authors also report negative results and acknowledge some limitations.
Critical weaknesses
  1. Mismatch between code and paper. The supplied code does not implement Eq. (14). In 2D Helmholtz, the physics-residual metric combines the normalized predicted field with the unnormalized source. In 3D Helmholtz, the metric uses a fixed Gaussian source, grid spacing, and discretized operator that differ from those used for data generation. This residual is also used for governing-equation guidance. These inconsistencies may affect the claim of “best physics consistency”.
  2. The benefits of the proposed extensions are insufficiently supported. In Section 4 SDIFT is regarded as the closest work, but I cannot find quantitative comparison between them. Table 2 does not separately evaluate frequency conditioning and joint modeling of two complex components, as well as the learned core-space prior. The supplied DiffusionPDE implementation uses observation guidance without PDE guidance, but HarmoCore can use the known operator and source. Moreover, the 2D Helmholtz LRTFR implementation jointly optimizes randomly initialized spatial bases and core coefficients for each test case, which is contrary to the linear least-squares fit against a fixed spatial basis. Thus the benefits of the proposed extensions are insufficiently supported.
Questions for authors
  1. Which metric and guidance code produced the main tables? Could you please provide the at least one test case, and explain the field normalization and the 3D operator differences. Please report Eq. (14) with the ground truth and predictions. This would strongly affect my decision if you could provide more experimental evidence to support your original claims.
  2. Another main concern is what baseline implementations and training data were used for the main tables? Do the reconstruction gains remain when compared with SDIFT adapted to time-harmonic complex wave fields, DiffusionPDE with PDE guidance, and linear least-squares reconstruction using the same fixed spatial basis on 2D Helmholtz? If possible, could you please make ablation to show how much does each of frequency conditioning and joint modeling of complex components contribute to the outcome?
4 comments

We sincerely thank the reviewer for the thorough evaluation and constructive feedback, especially the detailed comments on implementation consistency, physical-residual evaluation and experimental validation. These suggestions helped us identify weaknesses in our original submission and strengthen implementation, add controlled ablations and improve reproducibility. We address each concern below.

R1 — Physics-Residual Consistency and Reproducibility

Thank you for identifying the implementation inconsistencies. We revised residual evaluation and PDE guidance to adopt the original discrete Helmholtz operator, PML coefficients, grid spacing and case-specific sources.

The revised Eq. (14) defines

\sqrt{\frac{\mathbb E_{\mathcal I}\left|\mathcal A_\omega\hat u+f\right|^2} {\mathbb E_{\mathcal I}\left|f\right|^2}}$$ where $\mathcal I$ excludes only the outermost boundary layer while keeping PML cells. For normalized fields $u_n=u/s$, the source is scaled consistently: $$\mathrm{PhysRes}= \frac{\left\|\left[\mathcal A_\omega u_n+f/s\right]_{\mathcal I}\right\|_2} {2\left\|\left[f/s\right]_I\right\|_2}.$$ We use a global field scale in 2D and frequency-specific scaling in 3D. Appendix C.10 provides reproducible 1% sensing examples: - **2D Helmholtz:** four sources, $\omega=15$; ground-truth/prediction PhysRes: $1.9814\times10^{-5}/0.77153$. - **3D Helmholtz:** three sources, $\omega=7$; ground-truth/prediction PhysRes: $2.1229\times10^{-6}/0.58455$. The 3D guidance now uses the same discretization ($h=L/N$), stretched-coordinate PML and complex source as evaluation. The prior approximate operator and fixed Gaussian forcing are removed. Table 16 reports 3D comparisons with this consistent operator and source. **Runnable reproduction.** We release unchanged input/mask and pretrained weights for the 2D Helmholtz case ($\omega=15$, 1% sensing) in `code/examples/2d_helmholtz/`. Run ```bash python code/reproduce_helmholtz_case.py --device cuda --check ``` to reproduce full-field Rel. L2 = **0.05141156**, ground-truth Eq. (14) = **$2.135147\times10^{-5}$**, prediction Eq. (14) = **0.64855612**. This case contributes to main-table Ours statistics. For CPU verification: ```bash python code/verify_physics.py python -m unittest discover -s code/tests -v ``` The CPU check evaluates the 3D main-table prediction ($\omega=7$, 1% sensing), yielding ground-truth/prediction residuals $2.122899\times10^{-6}/0.58455477$. Corrected metric code is `code/common/2D_Helmholtz/physics_metric.py` and `code/common/3D_Helmholtz/physics_metric_3d.py`, sharing the operator backend `code/common/helmholtz_physics.py`. PDE guidance uses the squared norm-ratio residual. These updates are documented in Section 3.3, Appendix A.5, Algorithm 2 and Appendix C.10. ## R2 — Insufficient Evidence for the Proposed Components and Baselines We agree original experiments did not sufficiently ablate individual components. We have added controlled comparisons below. **SDIFT comparison.** We adapt $M=1$ SDIFT for time-harmonic fields, sharing the same frozen FTM, training data, scalar frequency conditioning, comparable denoiser capacity and 500 training epochs. Temporal convolutions and message passing are omitted (frequency-domain fields, no temporal sequence); EDM denoising and posterior sampling remain. With matched PDE guidance, Table 10 shows: | Method | 1% Rel. L2 | 2% Rel. L2 | |---|---:|---:| | HarmoCore | 0.130 | 0.068 | | SDIFT (adapted) | 0.509 | 0.454 | **Pixel-space diffusion with PDE guidance.** We add a DiffusionPDE-style baseline: a 3.48M-parameter frequency-conditioned UNet trained 500 epochs on full complex fields. Both methods use identical sparse observations and PDE guidance at inference. At 1%/2% sensing, its Rel. L2 is 0.843/0.757 vs. 0.130/0.068 for HarmoCore (Table 10). This is a controlled pixel-space adaptation, not an exact reproduction of original DiffusionPDE. **Joint versus independent Re/Im modeling.** Table 7 contrasts a joint denoiser (3.48M params) with two independent denoisers (3.54M total), under identical FTM, observations, initialization and coupled physical guidance. Joint modeling achieves 0.130/0.068 Rel. L2 (1%/2% sensing), versus 0.196/0.139 for independent modeling. **Learned prior versus Gaussian score.** Table 11 replaces only the learned diffusion score with a frequency-conditional Gaussian score, fixing all other settings. The learned prior yields lower errors at all sensing rates: 0.130 vs. 0.284 at 1% sensing. **Frequency conditioning.** Table 6 compares different frequency encodings; scalar normalized-frequency conditioning performs best. **LRTFR clarification.** We distinguish per-case LRTFR (jointly optimizes new spatial bases and core coefficients) from Fixed-basis LS (freezes HarmoCore bases and estimates only core coefficients). Additional implementation and experimental details are in Appendix A.6 and Appendices C.2–C.7.

Thank you very much for your detailed and constructive review. As the discussion is approaching its end, we wanted to kindly follow up. We have updated our response/paper to address the concerns and questions you raised (please see our latest reply/update). We would greatly appreciate it if you could let us know your current thoughts and any further questions or concerns you may have. We remain happy to clarify or provide additional information. Thank you for your time and consideration.

Thank you again for your review. As the discussion is closing soon, I wanted to kindly follow up. We would greatly appreciate any feedback you may have. Thank you for your time.

Thank you again for your review. As the discussion deadline is approaching, we would be grateful for any final comments or confirmation you may have. Thank you for your time.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response for all reviewers

We sincerely thank all reviewers for their careful evaluation and constructive feedback. We appreciate the positive assessments of the problem formulation, the continuous Functional Tucker representation, and the potential of performing posterior reconstruction directly in a compact latent space.

We agree that the original submission needed stronger evidence regarding physical consistency, the contributions of individual components, and comparisons with closely related methods. We have substantially revised the manuscript and expanded the experimental evaluation. The main changes are summarized below.

1. Corrected and validated physics-residual evaluation and guidance.

We corrected the implementation of Eq. (14) and clarified its evaluation convention to ensure consistency with the numerical operators used to generate the benchmark data. Both physics-residual evaluation and equation guidance now use the generating discrete Helmholtz operator, including its PML coefficients, grid spacing, and the actual case-specific complex source. We also clarified the normalization convention and established the equivalence between normalized-unit and physical-unit residual evaluation in Eq. (20).

To make the implementation independently verifiable, Appendix C.10 provides two reproducible test cases, reporting the field scales and physics residuals for both ground-truth and reconstructed fields.

2. Added controlled comparisons and component ablations.

We added a matched Functional Tucker adaptation of SDIFT, a full-field pixel-space diffusion baseline with both observation and PDE guidance, and fixed-basis least-squares reconstruction. We further added controlled comparisons between joint and independent real/imaginary priors, as well as between the learned diffusion score and a conditional Gaussian score. These experiments provide more direct evidence for the contributions of joint complex modeling and the learned core-space prior.

3. Expanded evaluation of field diversity and continuous reconstruction.

We added multi-source examples, independent-source evaluation without retraining, and off-grid sensing experiments using independently generated fine-grid PDE references. We also report prior-training costs and 3D inference time and memory consumption.

4. Clarified the scope and limitations.

We have clarified that HarmoCore targets sparse-observation posterior reconstruction rather than replacing conventional forward PDE solvers when the complete physical problem is known. We also explicitly acknowledge that arbitrary heterogeneous media and frequency extrapolation remain open directions.

The revised manuscript provides detailed implementation and experimental settings in the appendices. We also provide a red/green marked-up manuscript to facilitate inspection of the changes.

We hope these revisions address the reviewers' concerns and make the technical contributions and experimental evidence substantially clearer.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.