acceptodds
Under review as a conference paper at ICLR 2027

GSCP-Net: Gram-Structured Coded-Probe Recovery of Complex Optical Fields Through Atmospheric Turbulence

Abstract

Recovering complex optical fields through atmospheric turbulence is a challenging inverse problem. Our key idea is to cast the complex optical field recovery as parameter estimation, rather than to learn an end-to-end mapping from received to source fields. To make this estimation tractable, we compress the high-dimensional complex field into a low-dimensional latent representation with a variational autoencoder. Inspired by unitary space–time modulation in communications, we exploit the unitary structure of lossless turbulent propagation and propose a latent-variable estimation method that combines multiple unitary coded probes with Gram geometry. Under lossless reception, the Gram geometry is independent of the instantaneous turbulence channel, whereas a finite receiver aperture truncates the field. This truncation breaks the ideal invariance, and the joint effect of propagation and truncation reduces to an effective metric . Since can be modeled explicitly, we keep the autoencoder fixed and train only . Based on this formulation, we propose the Gram-Structured Coded-Probe Network (GSCP-Net) and evaluate it on a 2-km distributed-turbulence channel spanning 42 turbulence–aperture conditions. Compared with end-to-end baselines, GSCP-Net is substantially more accurate under strong turbulence and small apertures, while using orders of magnitude fewer channel-dependent parameters and only a fraction of their training time. Even under with a 140-mm receiver aperture, GSCP-Net reaches a recovery fidelity of 0.660 using 15 probes. Without retraining or channel tracking, the learned probes and fixed receiver keep the fidelity nearly unchanged over 20 coherence times of evolving turbulence. These results establish Gram-structured coded probing as a lightweight, physics-guided approach to robust complex-field recovery across a broad range of turbulence strengths and receiver apertures. Code is available at https://anonymous.4open.science/r/GSCP-Net/.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Atmospheric turbulence, Complex field recovery, Gram geometry, Multiple coded probes, Finite aperture reception, Physics-guided formulation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_uLhbOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose GSCP-Net, a physics-guided framework for optical field reconstruction under atmospheric turbulence. It combines a pretrained VAE with unitary coded probes to construct Gram matrices, reducing reconstruction to latent variable estimation. Experiments on a 2 km simulated turbulence channel with 42 combinations of turbulence strengths and receiver apertures demonstrate superior reconstruction accuracy, adaptability to dynamic turbulence, and training efficiency compared with the selected baselines.

Critical strengths
  1. The paper effectively exploits Gram invariance in lossless optical propagation and models finite-aperture effects through a low-dimensional effective metric, translating physical properties into a practical measurement design while avoiding explicit estimation of the full instantaneous propagation operator.
  2. By freezing the VAE and optimizing only the unitary probe codebook, the method reduces channel-dependent training costs while ensuring unitarity through structural parameterization, offering a simple design with potential applicability to other structure-constrained measurement tasks.
Critical weaknesses
  1. GSCP-Net uses 12 or 15 coded probes to obtain multiple measurements, whereas the U-Net and NAFNet baselines appear to use a different measurement protocol. Can the authors provide comparisons under matched measurement budgets and acquisition costs to demonstrate that the performance gains come from the proposed Gram-structured probe design rather than additional observations?
  2. Under finite-aperture reception, the effective metric R(ω;u)R(\omega;u) varies across turbulence realizations, while the receiver uses a training-derived mean metric R(u)R(u). Why does recovery remain stable when the instantaneous metric deviates substantially from its mean? Can the authors quantify the impact of metric estimation error and analyze sensitivity to the number of turbulence realizations used for estimating the mean metric?
  3. The method relies on a pretrained VAE to represent source fields in a low-dimensional latent space, and the experiments primarily use amplitude and phase maps constructed from MNIST. To what extent does the recovery performance depend on the learned source prior? Does the method remain effective for out-of-distribution complex fields or fields that cannot be accurately represented by the VAE?
Questions for authors

Please refer to the Weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer uLhb (1/3)

We thank the reviewer for the constructive feedback and the suggestion of a comparison under matched measurement budgets. We are glad that the reviewer found that the paper "effectively exploits Gram invariance in lossless optical propagation" and offers "a simple design with potential applicability to other structure-constrained measurement tasks". We address the three weaknesses below (parts 1–3).

We first summarize how the results are evaluated. A fidelity of 1 means perfect recovery. Before transmission, the VAE compresses the source into 64 numbers. Even if the receiver estimated all 64 numbers exactly, the decoded field would reach a fidelity of only 0.86 on average (Sec. 4.1). This ceiling is set by the VAE, and a different VAE gives a different ceiling. Since we focus on recovering the field after turbulence, the loss caused by turbulence and the aperture should be read relative to 0.86, not 1. Each result is evaluated on 42 conditions, namely all combinations of 6 turbulence strengths (Cn2C_n^2 from 10−1510^{-15} to 3×10−133\times10^{-13} m$^{-2/3},weaktostrong)and7receiverapertures(140–480mm).Themeanandworstfidelityaretheaverageandtheminimumofthe42per−conditionfidelities.ForGSCP−Net,theminimumoccursunderthestrongestturbulence(, weak to strong) and 7 receiver apertures (140–480 mm). The mean and worst fidelity are the average and the minimum of the 42 per-condition fidelities. For GSCP-Net, the minimum occurs under the strongest turbulence (C_n^2 = 3\times10^{-13}$ m$^{-2/3}$) with the smallest aperture (140 mm). All results are averaged over 3 random seeds, and the same definitions are used in parts 2 and 3.

Weakness 1: comparison under matched measurement budgets and acquisition costs.

GSCP-Net uses 12 or 15 coded probes to obtain multiple measurements, whereas the U-Net and NAFNet baselines appear to use a different measurement protocol. Can the authors provide comparisons under matched measurement budgets and acquisition costs to demonstrate that the performance gains come from the proposed Gram-structured probe design rather than additional observations?

The gain comes mainly from the structure of the Gram matrix, not from additional observations. As requested, all receivers in Table 3.1 use the same 12 probes as GSCP-Net, at the same measurement budget and acquisition cost. With these 12 observations, U-Net and NAFNet reading the received fields reach a mean fidelity of only 0.44 and 0.48, respectively, versus 0.83 for GSCP-Net. An MLP reaches a mean fidelity of 0.82 when it is trained on the Gram matrix, but only 0.45 when it is trained on the received fields.

Table 3.1. Comparison under the same measurement budget and acquisition cost.

ReceiverGram modelingMean fidelityWorst fidelity
GSCP-Net (paper)✓0.830.54
Received fields instead (receiver replaced by an MLP)×0.450.16
Same MLP, trained on the Gram matrix✓0.820.59
U-Net reading 12 received fields×0.440.22
NAFNet reading 12 received fields×0.480.17

Results for GSCP-Net are taken from Table 2 of the paper. All other results are new and averaged over 3 seeds. All receivers in Table 3.1 use the same number of transmissions, the same total energy and the same acquisition time. The worst fidelity leads to the same conclusion (Table 3.1). U-Net and NAFNet reach a worst fidelity of only 0.22 and 0.17, respectively, versus 0.54 for GSCP-Net. The MLP reaches a worst fidelity of 0.59 when it is trained on the Gram matrix, but only 0.16 when it is trained on the received fields. Moreover, with 12 received fields instead of the single field used by the baselines in the paper, the mean fidelity of U-Net and NAFNet rises only from 0.32 to 0.44 and from 0.28 to 0.48, respectively. The additional observations therefore account for only a small part of the gain of GSCP-Net, and the rest comes from the structure of the Gram matrix.

In the revision, the U-Net and NAFNet baselines that read 12 received fields will be added to Table 1 and to the reconstruction examples in Figure 3 of the paper. Table 2 of the paper will also include the two MLP rows (trained on the received fields and on the Gram matrix).

AuthorsOpenReview ↗

Response to Reviewer uLhb (2/3)

Weakness 2: deviation of the metric from its mean, impact of the metric error, and number of realizations.

Under finite-aperture reception, the effective metric (R(ω;u)R(\omega;u)) varies across turbulence realizations, while the receiver uses a training-derived mean metric (R(u)R(u)). Why does recovery remain stable when the instantaneous metric deviates substantially from its mean? Can the authors quantify the impact of metric estimation error and analyze sensitivity to the number of turbulence realizations used for estimating the mean metric?

(1) In most conditions, the effective metric deviates little from the mean metric, so recovery is stable. The deviation is large only under strong turbulence with a small aperture, where recovery also degrades.

(2) The recovery error caused by the metric error is bounded, and every measured error lies within the bound, on average at about one third of it.

(3) The mean fidelity is insensitive to the number of realizations. With 8 turbulence realizations, it is already 0.83, the same as with the 64 used in the paper. The worst fidelity rises slowly with the number of realizations, from 0.50 with 8 to 0.54 with 64. A few dozen realizations are therefore enough for calibration in practice.

Table 3.2. Ablation of the mean metric, including the number of turbulence realizations used for calibration.

Metric used by the receiverMean fidelityWorst fidelity
No channel statistics (codebook unchanged)0.800.35
Mean metric, 1 realization0.810.31
Mean metric, 8 realizations0.830.50
Mean metric, 64 realizations (paper)0.830.54
True metric of each realization (ideal, not available in practice)0.860.86

Results for the mean metric with 64 realizations are taken from Table 2 of the paper. All other results are new and averaged over 3 seeds.

(1) Why recovery stays stable. Under weak and moderate turbulence, as long as the aperture is not too small, the effective metric deviates from the mean metric by at most 10%. This is because the part that is the same in every realization (e.g., how much light enters the aperture on average) is already contained in the mean metric, and only the fluctuation of each realization remains. For example, for Cn2≤10−14C_n^2 \le 10^{-14} m$^{-2/3}$ with apertures of 220 mm or more, and for Cn2=3×10−14C_n^2 = 3\times10^{-14} m$^{-2/3}$ with apertures of 260 mm or more, the fidelity is 0.86, the same as with the true metric of each realization. Under the strongest turbulence (Cn2=3×10−13C_n^2 = 3\times10^{-13} m$^{-2/3}$) with the smallest aperture (140 mm), the deviation is as large as the mean metric itself, and the fidelity drops to 0.54. Even in this condition, the learned codebook raises the fidelity from 0.30 with the fixed codebook to 0.54 (Table 2 of the paper), because it is trained with the true effective metric of each realization.

(2) Error bound. To first order, the recovery error δz\delta z satisfies ∥δz∥≤∥e∥/σmin⁡(J)\Vert\delta z\Vert \le \Vert e\Vert / \sigma_{\min}(J), which follows from Eq. (22) in the main text and Eq. (42) in Appendix A. Here, ee is the error in the Gram matrix caused by the metric deviation, and σmin⁡(J)\sigma_{\min}(J) is the smallest singular value of the Jacobian JJ of the equations solved by the receiver. We verified this bound in every condition using 32 turbulence realizations and 64 images per realization (new). The measured error of every sample lies within the bound, on average at about one third of it.

(3) Number of realizations. With only 1 realization, the estimated mean metric still contains the fluctuation of that realization, and the worst fidelity is only 0.31. With 96 realizations instead of the 64 used in the paper, the worst fidelity rises only slightly, from 0.54 to 0.55. Therefore, 64 realizations are already enough.

In the revision, the main text will include this error bound. The appendix will also report the ablation of the mean metric and of the number of turbulence realizations used for calibration, as well as the verification of the bound.

AuthorsOpenReview ↗

Response to Reviewer uLhb (3/3)

Weakness 3: dependence on the learned source prior, and out-of-distribution fields.

The method relies on a pretrained VAE to represent source fields in a low-dimensional latent space, and the experiments primarily use amplitude and phase maps constructed from MNIST. To what extent does the recovery performance depend on the learned source prior? Does the method remain effective for out-of-distribution complex fields or fields that cannot be accurately represented by the VAE?

(1) The recovery depends on the VAE only through the ceiling. A better VAE raises the ceiling but does not make the method more robust to turbulence.

(2) The method remains effective for out-of-distribution fields. Their fidelity is lower only because the VAE represents them less well, which lowers the ceiling. After turbulence, the mean fidelity still retains at least 96% of the respective ceiling.

Table 3.3. Ablation of the source representation and out-of-distribution fields.

Source representationTest fieldsCeilingMean fidelityFraction of ceiling retained
Linear PCAMNIST0.690.6696%
Deterministic autoencoderMNIST0.880.8496%
VAE (paper)MNIST0.860.8396%
VAE (paper, not retrained)EMNIST letters (unseen)0.790.7696%
VAE (paper, not retrained)Fashion-MNIST clothing (unseen)0.610.5997%

Results for the VAE on MNIST are taken from Section 4.1 and Table 2 of the paper. All other results are new and averaged over 3 seeds. For PCA and the autoencoder, only the source representation is changed, and the codebook is retrained with the same recipe. For EMNIST and Fashion-MNIST, the VAE and codebook of the paper are used directly, without retraining. The ceiling is the fidelity of the decoded field when the receiver estimates all 64 numbers exactly.

(1) Dependence on the source prior. The VAE compresses the source: it maps the 28×28 amplitude and phase maps (1568 numbers) to the 64 numbers that are transmitted. After the receiver has estimated these 64 numbers, the VAE decoder restores the source field. With a deterministic autoencoder instead of the VAE, the mean fidelity remains essentially unchanged (0.84 versus 0.83 for the VAE). Linear PCA gives a lower mean fidelity of 0.66, but only because it compresses the source less well: its ceiling is 0.69 (VAE 0.86, autoencoder 0.88). After turbulence, all three representations retain about 96% of their ceilings (Table 3.3), and the mean fidelity of each is within 0.03–0.04 of its ceiling. This indicates that a better representation raises the ceiling but does not make the method more robust to turbulence. The method therefore does not depend on the VAE specifically.

(2) Out-of-distribution fields. We applied the VAE and codebook of the paper directly to unseen EMNIST letters and Fashion-MNIST clothing images. The VAE represents these fields less well, so the ceiling drops to 0.79 and 0.61, respectively. After turbulence, however, the mean fidelity still retains 96% and 97% of the respective ceilings (Table 3.3). The recovery through turbulence therefore works as well as on MNIST (96%). With a VAE trained on such data, the mean fidelity reaches 0.84 in both cases (Table 10 of the paper). The mean metric does not need to be recalibrated, because it depends only on the turbulence and the aperture, not on the source. However, richer fields, such as natural images, cannot be represented accurately by 64 numbers. They require more numbers and more probes, and we leave this extension to future work.

In the revision, the appendix will include the source-representation ablation (PCA, deterministic autoencoder, VAE) and the results on out-of-distribution fields. A new paragraph on practical requirements and limitations in the main text will state that 64 numbers cannot represent richer fields accurately.

We hope these responses address your concerns, and we are happy to discuss any further questions.

Thank you for the detailed responses and additional experiments, which have substantially addressed my original concerns. Although the comparable performance of the Gram-based MLP leaves some questions about the distinct contribution of the physics-based receiver, I appreciate the additional evidence and clarification provided by the authors. Overall, the rebuttal has strengthened my confidence in the paper, and I have decided to raise my rating.

Reviewer_hRr7OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses complex optical field recovery through atmospheric turbulence by reframing it as low-dimensional parameter estimation rather than end-to-end mapping. A frozen autoencoder compresses the source into a latent representation, embedded into a fixed transmit basis and modulated by unitary coded probes. Exploiting the unitarity of lossless propagation, the authors show the received probes' Gram matrix is invariant to the turbulence realization under full-field reception; finite apertures reduce this to a low-dimensional effective metric. Experiments across 42 turbulence-aperture conditions show large gains over U-Net and NAFNet with far fewer parameters, and stable fidelity over 20 coherence times.

Critical strengths
  1. The method has a sound physical basis. It exploits Gram invariance under unitary propagation and introduces an effective metric to account for finite-aperture effects, providing a coherent encoding and recovery framework.
  2. Probe learning provides measurable benefits. With the same number of probes, learned probes outperform fixed probes, particularly under strong turbulence and small apertures. This controlled ablation provides more convincing evidence than the main baseline comparisons.
Critical weaknesses
  1. The main comparisons do not isolate the source of the performance gains. The proposed method and the baselines differ in transmitter-side encoding, measurement protocol, and reconstruction method. While these comparisons demonstrate the performance of the complete system, they do not establish how much improvement comes specifically from Gram-based recovery. A comparison with an alternative receiver using the same coded observations would clarify this contribution.
  2. The methodological novelty is insufficiently articulated. The paper combines a pretrained VAE, unitary probe coding, Gram measurements, and nonlinear recovery, but does not clearly explain the substantive advance over related unitary coding and joint encoding–reconstruction approaches. The ablations examine probe design and probe number, providing limited insight into the roles of the source representation and mean-metric approximation. Clearer positioning and targeted component analysis are needed to substantiate the claimed novelty.
  3. Essential solver details are missing. The paper does not adequately specify the initialization, iterative updates, stopping criteria, or differentiation through the solver during training. These omissions hinder reproducibility and assessment of the recovery procedure’s reliability.
Questions for authors
  1. Can the authors compare GSCP-Net with an alternative receiver using the same coded observations and frozen source representation, while keeping the transmission and acquisition settings unchanged? How much of the reported improvement can be attributed specifically to Gram-based recovery?
  2. What is the substantive methodological advance over existing unitary coding and joint encoding–reconstruction approaches? Beyond probe design and probe number, can the authors clarify the contributions of the VAE representation and mean-metric approximation through targeted ablations or analysis?
  3. Could the authors clarify the details about the solver?
5 comments
AuthorsOpenReview ↗

Response to Reviewer hRr7 (1/4)

We thank the reviewer for the constructive feedback and the suggestion to separate the sources of the gain. We are glad that the reviewer recognized the "sound physical basis" of the method and found that "probe learning provides measurable benefits". We address the three weaknesses and the three questions below (parts 1–4).

We first summarize how the results are evaluated. A fidelity of 1 means perfect recovery. Before transmission, the VAE compresses the source into 64 numbers. Even if the receiver estimated all 64 numbers exactly, the decoded field would reach a fidelity of only 0.86 on average (Sec. 4.1). This ceiling is set by the VAE, and a different VAE gives a different ceiling. Since we focus on recovering the field after turbulence, the loss caused by turbulence and the aperture should be read relative to 0.86, not 1. Each result is evaluated on 42 conditions, namely all combinations of 6 turbulence strengths (Cn2C_n^2 from 10−1510^{-15} to 3×10−133\times10^{-13} m$^{-2/3},weaktostrong)and7receiverapertures(140–480mm).Themeanandworstfidelityaretheaverageandtheminimumofthe42per−conditionfidelities.ForGSCP−Net,theminimumoccursunderthestrongestturbulence(, weak to strong) and 7 receiver apertures (140–480 mm). The mean and worst fidelity are the average and the minimum of the 42 per-condition fidelities. For GSCP-Net, the minimum occurs under the strongest turbulence (C_n^2 = 3\times10^{-13}$ m$^{-2/3}$) with the smallest aperture (140 mm). All results are averaged over 3 random seeds, and the same definitions are used in parts 2–4.

Weakness 1 and Question 1: an alternative receiver on the same coded observations.

The main comparisons do not isolate the source of the performance gains. … Can the authors compare GSCP-Net with an alternative receiver using the same coded observations and frozen source representation, while keeping the transmission and acquisition settings unchanged? How much of the reported improvement can be attributed specifically to Gram-based recovery?

About three quarters of the reported improvement comes from the Gram-based recovery. As requested, we replaced the receiver of GSCP-Net by an MLP and kept the transmission, the acquisition and the frozen source representation unchanged. The MLP reaches a mean fidelity of 0.82 when it is trained on the Gram matrix, but only 0.45 when it is trained on the received fields.

Table 2.1. Ablation of the receiver.

ConfigurationRecovery: Gram modelingMean fidelityWorst fidelity
GSCP-Net (full)✓0.830.54
Received fields instead (receiver replaced by an MLP)×0.450.16
Same MLP, trained on the Gram matrix✓0.820.59

Results for GSCP-Net are taken from Table 2 of the paper. All other results are new and averaged over 3 seeds.

With the Gram matrix as input, the worst fidelity of the MLP also improves, from 0.16 to 0.59 (Table 2.1). Since nothing else changes, the improvement in both mean and worst fidelity can be attributed to the Gram-based recovery. Compared with the single-field U-Net of the paper (0.32), the 12 coded observations raise the mean fidelity to 0.45, and the Gram-based recovery raises it further to 0.82. The reason is that the received fields are scrambled by turbulence, so a network can hardly read the source from them. In contrast, the Gram matrix records inner products between probes, which turbulence can affect only through the part truncated by the aperture (Sec. 3.2, Eqs. (14)–(16)).

The physics-based solver of GSCP-Net reads the same Gram matrix and performs about as well as the trained MLP (mean fidelity 0.83 versus 0.82, worst fidelity 0.54 versus 0.59, Table 2.1). The advantage of the solver is that it needs no training and uses only the calibrated mean metric. The solver is also differentiable, so the codebook is learned end-to-end through it (Table 2.4).

In the revision, Table 2 of the paper will include the two MLP rows (trained on the received fields and on the Gram matrix).

AuthorsOpenReview ↗

Response to Reviewer hRr7 (2/4)

Weakness 2 and Question 2: methodological advance, and the roles of the VAE and the mean metric.

The methodological novelty is insufficiently articulated. … What is the substantive methodological advance over existing unitary coding and joint encoding–reconstruction approaches? Beyond probe design and probe number, can the authors clarify the contributions of the VAE representation and mean-metric approximation through targeted ablations or analysis?

(1) The substantive advance is that the receiver reads the full Gram matrix of the probes, a measurement matched to the physics of turbulence. Turbulent propagation is unitary, so it preserves the lengths of the probes and the angles between them, which are exactly what the Gram matrix records. As a result, turbulence can change the Gram matrix only through the part truncated by the aperture. The receiver therefore needs only the calibrated mean metric for the truncated part, rather than an estimate of each turbulence realization. For the same reason, the codebook, trained end-to-end through the solver on the Gram matrix, only has to design the probes, not undo the turbulence (see ① below for the comparison with both lines of work).

(2) The learned VAE representation raises the ceiling but does not make the method more robust to turbulence. With the VAE instead of linear PCA, the ceiling rises from 0.69 to 0.86 and the mean fidelity from 0.66 to 0.83 (Table 2.2 in part 3).

(3) The mean metric compensates for aperture truncation. Compared with ignoring the truncation, it mainly raises the worst fidelity, from 0.35 to 0.54. With the true metric of each realization, the worst fidelity would reach the ceiling of 0.86. The remaining gap therefore comes from how much the metric of a single realization deviates from the mean metric.

Points (2) and (3) are analyzed in part 3 with targeted ablations.

① Advance over unitary coding and joint encoding–reconstruction. While the individual components are known (we cite the invariance of the Gram matrix under unitary propagation as a lemma), our advance is to connect them through the physical structure of turbulent propagation and to learn the probes end-to-end through this structure.

  • Unitary space–time modulation is designed for an arbitrary unknown channel, so the receiver can only exploit the subspace spanned by the codewords. Turbulent propagation, in contrast, is physically unitary. Each phase screen only changes the phase, and diffraction loses no energy, so the lengths of the probes and the angles between them are preserved. We exploit this structure of the channel by using the full Gram matrix, a measurement matched to the physics of turbulence. This is the novelty of the method. The calibrated mean metric compensates for the bias caused by aperture truncation (③).
  • Joint encoding–reconstruction (end-to-end learning) does not exploit the physics of the channel, so the network must learn from data how to undo random turbulence. GSCP-Net is also trained end-to-end, but its receiver reads the Gram matrix, a measurement matched to the physics of turbulence. The learned codebook therefore only has to design the probes, not undo the turbulence. For comparison with this line of work, we ran a new experiment in which the probes were trained jointly with a U-Net that reads the received fields. This baseline reaches a mean fidelity of 0.68 and a worst fidelity of 0.24, versus 0.83 and 0.54 for GSCP-Net. With the codebook of GSCP-Net kept fixed, the MLP in part 1 likewise reaches a mean fidelity of 0.82 when it is trained on the Gram matrix, but only 0.45 when it is trained on the received fields. Constrained by the physics of turbulence, GSCP-Net learns only about 12k channel-dependent parameters, versus 7.8M for U-Net. It trains in about 40 minutes, versus 2.9–4.0 hours for the baselines (Table 12 of the paper). The mean metric of each condition is calibrated from only 64 turbulence realizations. Moreover, the mean fidelity of GSCP-Net remains at least 0.81 under phase drift, a turbulence strength misestimated by a factor of about 3, or non-uniform turbulence along the path (new).

(Continued in part 3: ② the role of the VAE representation and ③ the role of the mean metric.)

AuthorsOpenReview ↗

Response to Reviewer hRr7 (3/4)

Weakness 2 and Question 2 (continued from part 2). We now analyze points (2) and (3) of our answer with targeted ablations.

② Role of the VAE representation. In a new ablation, we change only the source representation. The codebook is retrained for each representation with the same recipe, and all other settings are kept unchanged.

Table 2.2. Ablation of the source representation.

Source representationCeilingMean fidelityFraction of ceiling retained
Linear PCA0.690.6696%
Deterministic autoencoder0.880.8496%
VAE (paper)0.860.8396%

Results for the VAE are taken from Section 4.1 and Table 2 of the paper. All other results are new and averaged over 3 seeds.

The VAE compresses the source: it maps the 28×28 amplitude and phase maps (1568 numbers) to the 64 numbers that are transmitted. After the receiver has estimated these 64 numbers, the VAE decoder restores the source field. With a deterministic autoencoder instead of the VAE, the mean fidelity remains essentially unchanged (0.84 versus 0.83 for the VAE). Linear PCA gives a lower mean fidelity of 0.66, but only because it compresses the source less well: its ceiling is 0.69 (VAE 0.86, autoencoder 0.88). After turbulence, all three representations retain about 96% of their ceilings (Table 2.2), and the mean fidelity of each is within 0.03–0.04 of its ceiling. This indicates that a better representation raises the ceiling but does not make the method more robust to turbulence. The method therefore does not depend on the VAE specifically.

③ Role of the mean metric. In a new ablation, we change only the metric used by the receiver and keep the codebook and all other settings unchanged.

Table 2.3. Ablation of the mean metric.

Metric used by the receiverMean fidelityWorst fidelity
No channel statistics (codebook unchanged)0.800.35
Mean metric (paper)0.830.54
True metric of each realization (ideal, not available in practice)0.860.86

Results for the mean metric are taken from Table 2 of the paper. All other results are new and averaged over 3 seeds.

The mean metric compensates for aperture truncation. A finite aperture cuts off part of the light, so the Gram matrix is no longer strictly invariant but depends on the effective metric of the current turbulence realization. For each turbulence strength and aperture, we average the effective metric over 64 turbulence realizations and use this average at the receiver. Compared with ignoring the truncation (no channel statistics in Table 2.3), the mean metric raises the mean fidelity from 0.80 to 0.83 and, more importantly, the worst fidelity from 0.35 to 0.54. With the true metric of each realization, which is not available in practice, every condition would reach the ceiling of 0.86. The remaining gap is therefore the cost of averaging the metric, not a loss of information in the Gram matrix.

In the revision, Section 2 of the paper will explain how GSCP-Net differs from unitary coding and joint encoding–reconstruction. The new joint encoding–reconstruction baseline (part 2) will be added to Table 1 of the paper. The appendix will also include the ablations of the source representation (PCA, deterministic autoencoder, VAE) and of the mean metric.

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

The paper propose a method to recover a complex optical field sent through turbulence. Instead of training a network to map received fields to source fields, the authors:

(i) compress the source field (phase and amplitude) into a latent space z with a frozen VAE; (ii) map z to a coefficient vector a(z) in a fixed dimension transmit subspace ; (iii) transmit K unitary-coded versions of it.

Based on the observation that for lossless propagation H is a unitary channel, so that the Gram matrix G = Y^H Y of the received probes does not depend on the turbulence if the full field is received. The receiver uses a training-set estimation to solves a nonlinear least-squares problem for z. Experiment wise, GSCP-Net reaches higher mean fidelity versus about 0.1 for U-Net/NAFNet under strong turbulence.

Critical strengths

The author proposed a clean idea. Based on the unitarity of the turbulence channel, a channel-invariant Gram geometry is utilized for the turbulent-perturbed source image reconstruction. The finite-aperture truncation is handled as a bounded perturbation, and a VAE is used to compress the image for transportation.

Critical weaknesses

Robustness to the real-world noise and settings The proposed method assumes perfect parameters setting. It needs phase-stable coherent measurement of K probes, which is not necessarily true under real-world settings. Moreover, the method assumes the channel statistics is known, while in the real world settings, the channel statistics would be inhomogeneous or time/spatial dependent. The author is expected to explain how applicable is the proposed method to the real-world turbulence.

Questions for authors
  1. In table 2, regarding the claimed benefits of the table, would the author provide a controlled ablation study to show if the gain comes from: (a) transmitter coding, (b) the VAE prior, (c) multiple probes, or (d) the Gram modeling?
  2. In Figure 3, the other two baselines (UNet and NAFNet) are taking single input but the GSCP is taking K = 12 observations. Would the author add baselines with the same measurement budget and prior for a fair comparison? For example, a network that takes the K received probe fields (with the same VAE) and reconstruct the scene.
  3. How does fidelity degrade with real world noise, for example: detector/shot noise, phase noise, imperfect channel statistics etc.? How should K be chosen as SNR drops?
3 comments
AuthorsOpenReview ↗

Response to Reviewer kshc (1/3)

We thank the reviewer for the constructive feedback and the question of applicability to real-world turbulence. We are glad that the reviewer appreciated the "clean idea" and recognized the "channel-invariant Gram geometry" based on the unitarity of the turbulence channel. We address the weakness and the three questions below (parts 1–3).

We first summarize how the results are evaluated. A fidelity of 1 means perfect recovery. Before transmission, the VAE compresses the source into 64 numbers. Even if the receiver estimated all 64 numbers exactly, the decoded field would reach a fidelity of only 0.86 on average (Sec. 4.1). This ceiling is set by the VAE, and a different VAE gives a different ceiling. Since we focus on recovering the field after turbulence, the loss caused by turbulence and the aperture should be read relative to 0.86, not 1. Each result is evaluated on 42 conditions, namely all combinations of 6 turbulence strengths (Cn2C_n^2 from 10−1510^{-15} to 3×10−133\times10^{-13} m$^{-2/3},weaktostrong)and7receiverapertures(140–480mm).Themeanandworstfidelityaretheaverageandtheminimumofthe42per−conditionfidelities.ForGSCP−Net,theminimumoccursunderthestrongestturbulence(, weak to strong) and 7 receiver apertures (140–480 mm). The mean and worst fidelity are the average and the minimum of the 42 per-condition fidelities. For GSCP-Net, the minimum occurs under the strongest turbulence (C_n^2 = 3\times10^{-13}$ m$^{-2/3}$) with the smallest aperture (140 mm). All results are averaged over 3 random seeds, and the same definitions are used in parts 2 and 3.

Weakness: applicability to real-world turbulence.

It needs phase-stable coherent measurement of K probes, which is not necessarily true under real-world settings. Moreover, the method assumes the channel statistics is known, while in the real world settings, the channel statistics would be inhomogeneous or time/spatial dependent.

(1) Phase-stable coherent measurement of the KK probes is feasible with current lasers and detectors. The relative phase between the probes drifts over time. The narrower the laser linewidth, the slower the drift, so the probes only need to be measured fast enough. For a typical laser linewidth of 1 kHz, the relative phase drifts by about 0.5 rad within 40 µs. All KK probes must be measured within 40 µs. This requires a rate of a few hundred kHz (about 280 kHz for K=12K = 12 and 350 kHz for K=15K = 15), which commercial detectors sampling at GHz rates can provide. We simulated a phase drift of about 0.5 rad between the probes and let the receiver estimate it jointly with the source. As Table 1.1 shows, the mean fidelity is still 0.81, versus 0.83 without drift. Moreover, even for the strongest turbulence (Cn2=3×10−13C_n^2 = 3\times10^{-13} m$^{-2/3}$), the channel coherence time is about 0.7 ms (Table 11 of the paper), so the channel is essentially frozen within 40 µs.

(2) The channel statistics used by our method depend on just two parameters, the aperture size and the turbulence strength, and the latter only needs to be known to within a factor of about 3. In practice, the aperture size is known, and the turbulence strength can be measured with standard instruments such as scintillometers. Moreover, channel statistics that are inhomogeneous or vary in time and space have little effect on the mean fidelity of our method. To test this robustness, we considered three cases: a turbulence strength misestimated by a factor of about 3, turbulence that is non-uniform along the path, and a receiver that uses no channel statistics at all. In the last case, the receiver assumes that the Gram matrix is exactly invariant and ignores the aperture truncation. As Table 1.1 shows, the mean fidelity stays above 0.8 in all three cases. After the turbulence has evolved for 20 coherence times, the fidelity differs from the static case by less than 0.01, without any retraining or tracking (Table 3 of the paper).

Table 1.1. Fidelity under practical conditions.

Practical conditionMean fidelityWorst fidelity
Ideal (paper)0.830.54
Phase drift of about 0.5 rad between probes0.810.38
Turbulence strength misestimated by a factor of about 30.820.54
No channel statistics (codebook retrained)0.820.42
Non-uniform turbulence along the path0.81–0.840.47–0.56

Results for the ideal case are taken from Table 2 of the paper. All other results are new and averaged over 3 seeds.

In the revision, the main text will include a new paragraph on practical requirements and limitations. Table 1.1 will be added to the appendix.

AuthorsOpenReview ↗

Response to Reviewer kshc (2/3)

Question 1: source of the gain.

In table 2, regarding the claimed benefits of the table, would the author provide a controlled ablation study to show if the gain comes from: (a) transmitter coding, (b) the VAE prior, (c) multiple probes, or (d) the Gram modeling?

The largest gain comes from (d) the Gram modeling: the same network reaches a mean fidelity of 0.82 when it is trained on the Gram matrix, but only 0.45 when it is trained on the received fields. The ablation also reveals what learning adds: the learned codebook (a) raises the worst fidelity from 0.30 to 0.54, and a learned source representation (b), whether a VAE or an autoencoder, raises the mean fidelity from 0.66 (linear PCA) to 0.83–0.84. With 12 instead of 9 probes (c), the gain is mainly in the worst fidelity, which rises from 0.19 to 0.54. We ablated GSCP-Net itself, changing one item at a time and keeping everything else fixed. Mean and worst fidelity are the "mean" and "worst" columns in Table 2 of the paper.

Table 1.2. Ablation of GSCP-Net.

Configuration(a) learned codebook(b) VAE prior(c) probes(d) Gram modelingMean fidelityWorst fidelity
GSCP-Net (full)✓✓12✓0.830.54
Codebook not trained (fixed codebook)×✓12✓0.790.30
VAE → deterministic autoencoder (ceiling 0.88)✓×12✓0.840.53
VAE → linear PCA (ceiling 0.69)✓×12✓0.660.42
9 probes instead of 12✓✓9✓0.750.19
Received fields instead (receiver replaced by an MLP)✓✓12×0.450.16
Same MLP, trained on the Gram matrix✓✓12✓0.820.59

Results in rows 1, 2 and 5 are taken from Table 2 of the paper. All other results are new and averaged over 3 seeds.

(d) Gram modeling. From Table 1.2, we see that the same MLP reaches a mean fidelity of 0.82 when it is trained on the Gram matrix (row 7), but only 0.45 when it is trained on the received fields (row 6). The most important gain therefore comes from the structure of the Gram matrix. The reason is that the received fields are scrambled by turbulence, so a network can hardly read the source from them. In contrast, the Gram matrix records inner products between probes, which turbulence can affect only through the part truncated by the aperture (Sec. 3.2, Eqs. (14)–(16)). Our experiments confirm this: for Cn2≥3×10−14C_n^2 \ge 3\times10^{-14} m$^{-2/3}$, the MLP achieves a fidelity of only 0.16–0.19 with the received fields for all apertures, even at 480 mm, where more than 82% of the energy is received. Under the same conditions, it achieves 0.59–0.84 with the Gram matrix.

(a) and (c). With a fixed codebook instead of the learned one, the mean fidelity drops only from 0.83 to 0.79, but the worst fidelity drops from 0.54 to 0.30. With 9 instead of 12 probes, the mean and worst fidelity drop to 0.75 and 0.19, respectively. Nine is the smallest number of probes whose Gram matrix gives at least as many equations as unknowns (73 versus 64, Sec. 3.3). The learned codebook and the extra probes therefore mainly raise the worst fidelity.

(b) VAE prior. As noted in part 1, the VAE compresses the source: it maps the 28×28 amplitude and phase maps (1568 numbers) to 64 numbers. After the receiver has estimated these 64 numbers, the VAE decoder restores the source field. As a result, the receiver only needs to estimate 64 numbers instead of the full image, so the minimum number of probes is small (9, as noted above). With a deterministic autoencoder instead of the VAE, the mean and worst fidelity remain essentially unchanged. Linear PCA gives a lower mean fidelity of 0.66, but only because it compresses the source less well: its ceiling is 0.69 (VAE 0.86, autoencoder 0.88). After turbulence, all three representations retain about 96% of their ceilings (Table 1.2), and the mean fidelity of each is within 0.03–0.04 of its ceiling. This indicates that a better representation raises the ceiling but does not make the method more robust to turbulence. On the data used in the paper, a learned nonlinear representation therefore raises the mean fidelity from 0.66 (linear PCA) to 0.83, while the variational prior itself brings no extra gain (autoencoder 0.84).

In the revision, Table 2 of the paper will include the two MLP rows (trained on the received fields and on the Gram matrix). The appendix will include the source-representation ablation (PCA, deterministic autoencoder, VAE). The learned codebook with 12 probes will also be added to Table 6 in the appendix, so that the fixed and learned codebooks can be compared with the same number of probes.

AuthorsOpenReview ↗

Response to Reviewer kshc (3/3)

Question 2: baselines with the same measurement budget and prior.

In Figure 3, the other two baselines (UNet and NAFNet) are taking single input but the GSCP is taking K = 12 observations. Would the author add baselines with the same measurement budget and prior for a fair comparison? For example, a network that takes the K received probe fields (with the same VAE) and reconstruct the scene.

As requested, we added U-Net and NAFNet baselines that use the same 12 probes and the same VAE as GSCP-Net. Their mean fidelity is 0.31 (U-Net) and 0.34 (NAFNet), versus 0.83 for GSCP-Net. The gap comes mainly from the structure of the Gram matrix (part 2). "Same VAE" means that the baselines also output 64 numbers, which the VAE decoder of GSCP-Net maps to the source field. The results are shown in Table 1.3.

Table 1.3. Baselines with the same measurement budget and prior.

ReceiverKKPrior (same VAE)Gram modelingMean fidelityWorst fidelity
GSCP-Net (paper)12✓✓0.830.54
U-Net12✓×0.310.26
NAFNet12✓×0.340.27
U-Net12××0.440.22
NAFNet12××0.480.17
U-Net (paper baseline)1××0.320.09
NAFNet (paper baseline)1××0.280.08

Results for GSCP-Net are taken from Table 2 of the paper. All other results are new and averaged over 3 seeds. With the same VAE, U-Net and NAFNet are also far below GSCP-Net in worst fidelity (0.26 and 0.27 versus 0.54). Without the VAE, U-Net and NAFNet output the amplitude and phase maps directly and reach a mean fidelity of 0.44 and 0.48, respectively. This is 0.12 to 0.20 higher than the single-field baselines of the paper, but still far below GSCP-Net. Therefore, the 12 probes help the baselines only to a limited extent. With the same VAE, their mean fidelity even drops from 0.44 to 0.31 (U-Net) and from 0.48 to 0.34 (NAFNet). This shows that neither the 12 probes nor the VAE prior explains the advantage of GSCP-Net.

In the revision, the new U-Net and NAFNet baselines (12 probes, with and without the same VAE) will be added to Table 1 and to the reconstruction examples in Figure 3 of the paper. Their results for all 42 conditions will be given in Table 4 in the appendix. The single-field baselines will also be reported as averages over 3 seeds.

Question 3: real-world noise and the number of probes.

How does fidelity degrade with real world noise, for example: detector/shot noise, phase noise, imperfect channel statistics etc.? How should K be chosen as SNR drops?

(1) GSCP-Net is robust to all three types of noise. At an SNR of 10 dB, detector and shot noise lower the mean fidelity by only 0.01 (Table 1.4). At 0 dB, the mean fidelity drops from 0.83 to 0.54, but it is still higher than that of the U-Net and NAFNet baselines without noise (at most 0.48, Table 1.3). Phase noise and imperfect channel statistics lower the mean fidelity by at most 0.02 (Table 1.1).

(2) The number of probes KK should not be reduced as the SNR drops. It should be the largest value the acquisition time allows. With the same total energy, more probes give a higher mean fidelity at every SNR in Table 1.4 (9 to 15 probes). At 0 dB, for example, the mean fidelity is 0.33 with 9 probes and 0.57 with 15 probes.

We model detector thermal noise and shot noise as white Gaussian noise. In coherent detection, the shot noise comes mainly from the strong local oscillator, so it does not depend on the signal. In Table 1.4, the total energy is split equally among the KK probes. The SNR is computed as if all of this energy were sent in one probe.

Table 1.4. Mean fidelity under noise.

Total-energy SNRK=9K = 9K=12K = 12 (paper)K=15K = 15
Noiseless0.750.830.84
10 dB0.720.820.83
5 dB0.620.780.79
0 dB0.330.540.57

Results without noise are taken from Table 2 of the paper. All other results are new and averaged over 3 seeds. The upper limit on KK comes from the acquisition time. All probes must be measured within the time allowed by the laser phase noise (about 40 µs, part 1).

In the revision, we will add Table 1.4 to the appendix, together with the worst fidelity. A sentence on choosing KK will also be added to the new paragraph on practical requirements and limitations (part 1).

We hope these responses address your concerns, and we are happy to discuss any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.