acceptodds
Under review as a conference paper at ICLR 2027

Receiver-Distribution-Based Alphabet Design for Structured-Light Communication

Abstract

Due to the deteriorating effect of atmospheric/oceanic turbulence on optical signals, the design of laser source profiles (viewed as alphabets) plays an important role in free-space optical communication. Alphabet design becomes difficult in such random media when separation between transmitted fields does not reliably translate into separable receiver distributions and no tractable receiver-side criterion is available. We introduce receiver-distribution alphabet design, which learns operational pairwise distinguishability from held-out cross-entropy induced by classifiers with calibration on receiver-intensity observations. The resulting empirical scores support finite-pool selection that jointly protects the least distinguishable selected pair and improves aggregate pairwise separability, without assuming an analytic channel metric. We evaluate this principle in physics-based simulations of structured-light propagation through random media, where the resulting alphabet improves mean classification accuracy by percentage points over an unadapted binary reference on propagation realizations disjoint from selection. These results support receiver-distribution-aware alphabet design when source-space geometry alone does not reliably predict receiver separability.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
structured light, free-space optical communication, constellation design, Jensen–Shannon divergence

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_GUJAOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper explores an alphabet design for structured light communications utilizing receiver distribution. It tackles the receiver-distribution alphabet design problem within a specific coefficient space and candidate pool size. The solution involves training an encoder, converting the pooled distinguishability matrix and employing a multistart greedy search to select the most suitable alphabet.

Critical strengths
  1. This paper considers a physics-related problem related to free-space optical communications, and provides a mathematically rigorous foundation to support their claims.

  2. The authors effectively bridge traditional optimization methods with modern learning-based techniques. The authors also provide their code, ensuring the proposed methods are transparent and easily reproducible.

Critical weaknesses
  1. The introduction lacks conviction and fails to adequately emphasize the significance of the problem. It raises questions about the practicality of the proposed methods and highlights the shortcomings of existing approaches. In addition, I do not see significant innovations in the first and second contribution points as what makes the work better in contrast with the state-of-the-art is not reflected clearly.

  2. This work uses strong assumptions, therefore, the generalizability might be a question. For example: 1) this work only considers a single fix beam, 2) this work fixes the discrete power with normalization, 3) this work uses only a single setting of Laguerre-Gaussian basis (how about others? Additionally, it appears that the authors repeat this information many times), 4) this work considers only limited selected alphabet size and candidate-pool size. Furthermore, this work has no comparison against state-of-the-art, hence, it is difficult to assess the performance gain.

Questions for authors
  1. When the authors mentioned random medium, could it be mathematically modelled using a statistical distribution? Additionally, what happens if we have multiple mediums?

  2. Is there any modification to the one-channel ResNet-18 classifier? If it is only the training procedure, how is the performance on the other networks? Why would the author choose to use this learning-based solution over conventional statistical approaches? This raises the importance of an ablation study.

  3. How is the computational complexity of the proposed framework? It seems that the Big-O notation in line 340 only applies for the multistart greedy search. How about the ResNet-18 inference?

  4. How is the performance of the proposed framework with different number of alphabets? This is important to assess the generalizability of the work.

Reviewer_3pP7OpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper selects transmitted beam patterns for structured-light communication through turbulence by scoring candidate pairs with a classifier-based pairwise cross-entropy on simulated receiver images. It compares two selection rules on this score, exact max–min and greedy min-sum. On held-out realizations, the min-sum alphabet gets 79.1% accuracy, compared with 74.7% for max–min and 65.3% for a binary reference alphabet.

Critical strengths
  1. The problem is well motivated. Separation at the transmitter does not translate into separation at the receiver after turbulence and intensity detection. The geometry and source-intensity controls show this convincingly: they decode worse than even the binary reference.

  2. The evaluation is careful. Decoders are retrained from scratch on realizations disjoint from those used for selection, all alphabets share 25 matched cells, and the comparisons are paired. The max–min solution has an exact optimality certificate over the candidate pool.

  3. The method is simple and black-box. It needs only samples from a simulator or measurements, with no differentiable channel model.

Critical weaknesses
  1. Limited ML novelty and questionable fit for ICLR. Each component is standard: a classifier-based divergence estimate, the JS/cross-entropy identity, p-dispersion with clique solvers, and greedy subset selection. The contribution is mainly applying these components to one optics problem.

  2. The experimental scope is narrow. a) Everything is in simulation. b) Only one turbulence regime is tested (σ=3, z=5). c) Only one basis (d=4, real coefficients), one pool, and one alphabet size (K=15). With a single setting, it is unclear whether min-sum beats max–min, or whether the gains generalize.

  3. Baselines are weak. The comparisons are a fixed binary alphabet and source-side heuristics. Missing comparisons: a) random 15-subsets of the pool; b) selecting directly from the 500-class encoder’s confusion matrix; c) an end-to-end learned constellation. Without these, it is hard to tell how much of the gain comes from the proposed pairwise-CE score rather than from any receiver-aware selection.

  4. No communication-level metrics. Results are 15-class classification accuracy at a single operating point. There is no symbol error rate versus turbulence strength, no information rate, and no cost comparison. A 79% symbol accuracy is also far from what a practical link would use, so the practical value is unclear.

Questions for authors

See above weaknesses

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

The paper picks 15 structured-light symbols from 500 superpositions of four Laguerre-Gaussian modes. It scores candidate pairs by the held-out pairwise cross-entropy of calibrated ResNet-18 classifiers on simulated receiver images and selects alphabets by exact max-min or multistart min-sum. Each alphabet is decoded by classifiers retrained on media disjoint from selection, in 25 cells (five partition seeds × five classifier-training seeds); in one simulated regime, min-sum decodes at 79.06% against 65.26% for a binary reference and 74.67% for max-min.

Critical strengths
  1. The evaluation is reproducible, and its main finding holds. The archive's records re-derive Table 1 and every paired difference, the multistart reproduces S_sum exactly, the Fig. 2 receiver images regenerate on CPU from their seed, and the gain over the binary reference survives a design-aware test (partition-level p = 0.002). This is why I trust the paper's numbers, including the ones that argue against its framing.

  2. In this regime the geometry-only control (56.72%, below the binary 65.26%; Table 1) shows that source-space max-min separation is a poor proxy for receiver separability. This is the most informative result in the paper and the main thing I learned from it.

Critical weaknesses
  1. No experiment shows that the classifier-induced score, the stated contribution (p.2 l.088-089), is needed. All decoded comparators are source-side (binary, S_geo, Table 11), so the results show only that receiver-aware selection beats source-side rules. They do not compare held-out cross-entropy with simpler receiver-side scores (mean-image distance, MMD) or with random subsets, and the archive notes "no uniform random alphabet has been decoded yet". Selection and decoding also use one regime (σ = 3, z⋆ = 5), one d = 4 basis and one decoder, while the abstract (p.1 l.026-027) extends support to settings where source geometry mispredicts receiver separability and the introduction claims applicability to measured channels (p.1 l.053-054). Since the score is the paper's stated contribution, this gap decides my rating.

  2. The comparison between objectives is overstated. The 25 cells re-partition one bank of 100 media (App. E.3; "not 25 independent experiments", per the archive), so "paired t = 6.38, p < 10^-5" (p.7 l.375) becomes p = 0.023 at the partition level. The max-min optimum is also not unique (single swaps from S_mm reach τ* with L_sum 1.699-1.756). The archive records min-sum as a later ablation, predicted to decode worse than a constrained one-swap alphabet (75.91%, +1.24 pp over max-min, partition-level p = 0.29) that only the README reports. The paper's recommendation of min-sum over max-min therefore rests on less evidence than stated, and the paper does not say that this comparison was added after decoding had begun.

Questions for authors
  1. Validation experiment: How do uniform random 15-subsets and a classifier-free receiver-side score (same solvers), decoded in the same cells, compare with S_sum?
  2. Validation experiment: Does min-sum still beat max-min when other tied max-min optima, such as the lowest-L_sum one, are decoded in the same cells?
  3. Clarification: Will you report every decoded variant, including the one-swap alphabet, and when each objective was added?
Reviewer_2bihOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary
  • The paper presents a framework to obtain suitable alphabet design for optical signalling based on receiver distributions.
  • The key contributions include the alphabet design with two objectives (namely, max-min and min-sum) along with some numerical results.
  • The presented results compare it with a binary reference and goemtry only control based alphabet design.
Critical strengths
  • The paper aims to utilize a receiver distribution based alphabet design for better performance.
Critical weaknesses
  • The practical impact of the work in its is current state is not clear. The run time considered is huge (177s). Is the system expected to be run once and then used in the setup forever. Or is the alphabet design regularly updated based on channel.
  • The discretization of RP^3 to 500 classes is not clear. Why not explore 100 classes for faster run time or 1000 classes for better performance. Based on the simulation setting considered is 500 a good number of classes.
Questions for authors
  • How is the final selected alphabet expected to be used in a practical communication system (in terms of life cycle of this alphabet-> selection, communication, discarding/reselection).
  • Primarily the results of the paper are based on simulated data, why not consider more variations in terms of classes and other such parameters.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.