{"id":"12555509-f341-4fb3-803a-4be99d05fdee","arxiv_id":"2411.12825","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"TopoCode encodes the persistent homology of images as a small side packet and uses it to guide noise-robust reconstruction, outperforming LDPC and convolutional codes at low SNR in simulations.","lead":"A new coding scheme, TopoCode, transmits a compressed topological signature of an image alongside the raw data to detect and correct message-level errors. It is aimed at immersive media like XR and holographic communications, where preserving the overall shape of content matters more than perfect bits.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"TopoCode's correction target is the received—not original—persistence diagram, and the paper never models bit errors in the TopoCode itself, so the central reconstruction claim rests on an unstated assumption.","rationale":"The reader's weakest assumption—that the transmitted persistence diagram is received error-free—is exactly the load-bearing weakness I find. The paper's own transmission description in Section IV makes clear that the TopoCode bits are modulated and sent over the same noisy BPSK channel, yet the correction algorithm in Eq. (3) and the experiments in Section IV treat the received diagram as the true target. Because a corrupted target can pull the reconstructed image away from the original while still yielding a small optimization loss, the reported topological-fidelity and image-reconstruction results cannot be trusted until this gap is addressed. The issue is concrete, internal to the paper's setup, and readily testable, so it is the appropriate basis for a conditional verdict. I agree with the reader's identification and do not recommend changing the verdict: the paper should remain conditional pending a demonstration that the scheme works with realistic TopoCode bit errors or with appropriate protection of the side information.","tokens_in":7288,"tokens_out":3552,"duration_ms":38830,"concrete_test":"Rerun the Section IV MNIST experiment with two receiver branches: (i) ideal TopoCode, as implicitly assumed, and (ii) the actual demodulated TopoCode bits, including bit errors. For branch (ii), plot recovered-image PSNR/SSIM and the Wasserstein distance between the recovered PD and the original source PD. If branch (ii) degrades substantially at SNR=3 dB or differs from branch (i), the central claim is unsupported until the TopoCode is protected or the decoder is made robust to PD corruption. A complementary analytical check: compute the expected number of corrupted birth-death coordinates in a 256-bit diagram at the stated SNR and compare that with the diagram's size to confirm the assumption is violated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central reconstruction step, Eq. (3), minimizes Wasserstein distance to the persistence diagram \"sent by the transmitter.\" But Section IV describes transmitting the TopoCode bits over the same BPSK channel at the same SNR as the image, with no error protection or correction for those bits. At SNR=3 dB, standard BPSK gives a BER of roughly 2.3%, so a 256-bit TopoCode (the paper's own example) has an expected ~5.8 bit errors and near-certain corruption: the probability of an error-free diagram is below 1%. Since each birth-death pair is encoded as 8-bit coordinates, a single bit flip changes birth/death values and can change which pairs survive the alpha-thresholding in error detection. The optimization then drives the received image to match a corrupted diagram, so the reported small Wasserstein distances and high PSNR/SSIM do not establish fidelity to the original message. No experiment in Figs. 7–9 includes the case where the received TopoCode differs from the transmitted one. This is an internal gap, not merely a disagreement with coding-theory practice: the paper's own setup violates the assumption that the target PD is available at the receiver. The comparison with LDPC and convolutional codes is also unfair for the same reason, since those codes protect all information bits while TopoCode leaves its side information unprotected. Stability inequality (2) does not rescue the argument: it bounds PD distance in terms of pixel differences, not in terms of bit errors in the PD itself, and a corrupted target can be close to a reconstructed image that is far from the true source.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TopoCode, a message-level error detection and correction scheme for image communication. At the transmitter, it computes a persistence diagram (PD) of a grayscale image via cubical filtration, encodes the PD as side information, and concatenates this TopoCode with the binarized uncoded image before BPSK modulation. At the receiver, it compares the PD of the received image with the decoded TopoCode, thresholds birth-death pairs by total persistence to detect significant topological errors, and solves a p-Wasserstein minimization (Eq. 3) to reconstruct the image. Experiments on MNIST and Omniglot compare PSNR, SSIM, Wasserstein distance, and packet length with LDPC and convolutional codes, reporting advantages in low-SNR regimes.","tokens_in":7469,"tokens_out":7324,"duration_ms":68332,"significance":"TopoCode addresses a real gap: bit-level channel codes do not directly capture semantic or topological importance, and the proposed use of persistent homology is training-free and potentially applicable to multiple data modalities. The paper also provides explicit complexity statements and uses a standard TDA library. However, the central validation currently assumes the receiver has an accurate copy of the transmitter's PD even though the TopoCode bits are sent over the same noisy channel without protection, and the reported Wasserstein metric may be tautological because it is the same quantity minimized by the correction algorithm. If the side-information errors are modeled or the TopoCode is protected, the idea could be a useful message-level supplement; as presented, the claimed reconstruction gains are not established.","major_comments":[{"comment":"The correction target is assumed to be the transmitter's persistence diagram, but the TopoCode bits are transmitted unprotected over the same BPSK channel. At SNR=3 dB, BPSK has a bit error rate near 2.3e-2; a 256-bit TopoCode, the paper's own example, has an expected 5.8 bit errors and an essentially zero probability of arriving intact. Because each birth-death pair is encoded as 8-bit coordinates, bit flips corrupt the target diagram before the optimization in Eq. (3) is applied. The experiments in Figs. 7-9 never include errors in the TopoCode bits, so the reported PSNR, SSIM, and Wasserstein values do not demonstrate reconstruction of the original message. This is an internal gap: the receiver does not actually have the quantity Dh(fl(X)) that Eq. (3) requires unless the TopoCode is separately protected.","section":"§IV, Fig. 7; §III, Eq. (3)"},{"comment":"The loss minimized in Eq. (4) is the p-Wasserstein distance between the received image's PD and the target PD. If the 'Wasserstein distance' reported in Figs. 8c and 9c is computed with respect to the same received TopoCode used as the optimization target, the small values are tautological and do not measure fidelity to the transmitted image. The text does not specify the reference diagram. The authors should state which PD is used as the reference in those figures and, ideally, report the Wasserstein distance to the original PD of the clean image for the full system including TopoCode bit errors.","section":"§IV, Figs. 8c/9c; §III, Eq. (4)"},{"comment":"The comparison against LDPC and convolutional codes is not apples-to-apples. LDPC and convolutional coding protect all information bits, while TopoCode transmits its side information with no protection at the same SNR. The redundancy budgets also differ: TopoCode adds only about 4% overhead in the example, whereas the LDPC and convolutional codewords have substantially larger overhead. The shorter packet length reported in Figs. 8d and 9d is therefore not a fair system-level comparison. A fair comparison would protect the TopoCode bits with the same code, include TopoCode bit-error effects in the end-to-end reconstruction, and equalize the total overhead or data rate.","section":"§IV, Figs. 8 and 9"},{"comment":"The sentence 'Equation (2) provides a lower bound of the distance between X and X̂' states the opposite of the displayed inequality; Eq. (2) gives an upper bound on the PD distance in terms of the image distance. This verbal error should be corrected. More importantly, the stability inequality concerns perturbations of the underlying image and does not control bit errors in the encoded PD representation, so it does not rescue the correction algorithm from the corrupted-TopoCode problem.","section":"§III, Eq. (2)"}],"minor_comments":[{"comment":"There is a typo: 'BD paris' should be 'BD pairs'.","section":"§II"},{"comment":"The name is written inconsistently as 'Topocode' and 'TopoCode'; one spelling should be used throughout.","section":"Throughout"},{"comment":"The code rate of the regular LDPC code should be stated explicitly; the degree pair (5,20) alone does not determine the rate, and the number of parity bits is needed to interpret the packet-length comparison.","section":"§IV"},{"comment":"The 'TopoCode without Boundary' variant is not defined precisely; the authors should state how boundary pixels are removed and explain why this does not change the comparison conditions, since LDPC and convolutional decoding still reconstruct the full image.","section":"§IV, Fig. 7"},{"comment":"The paper's own limitation statement that TopoCode can only recover topological information is important and should be reflected in the abstract and introduction, because the reported PSNR and SSIM gains could be driven by non-topological regularization rather than by topological correction.","section":"§IV, last paragraph"}],"recommendation":"major_revision","confidential_remarks":"The paper is a concise letter and the requested corrections are substantive: the end-to-end simulation must model TopoCode bit errors or explicitly protect the side information, the Wasserstein reference must be clarified, and the baseline comparison should be made fair. These additions may exceed the letter format, so the editor may wish to consider whether a full paper venue is more appropriate. The idea itself is original and worth pursuing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: this is the first paper I know of that uses persistent homology for message-level error detection and correction in communication systems. That is a legitimate new application, and the framing — semantic communication without training — is relevant to XR and holographic-type traffic. The paper is clearly written and the authors openly acknowledge several limitations: limited gain at high SNR, inability to capture non-topological information, and variable-length PD codes. Those are honest disclosures.\n\nWhat it does well: the error detection scheme using total persistence thresholding is sensible, and the Fig. 5 example — two images with identical PSNR/SSIM but different topological error significance — makes the point effectively. The simplicity of the approach is a real strength: it is training-free, uses established TDA tools (Gudhi), and the redundancy in the example is genuinely small (256 bits vs 6,272 data bits).\n\nWhere it falls down. First, circularity: Eq. (3) minimizes the p-Wasserstein distance to the transmitted PD, and Figs. 8c/9c report that same distance as the success metric. The optimized image is forced to be close in that metric, so the numbers do not establish fidelity to the original message. The comparison with LDPC and convolutional codes is not apples-to-apples because those codes protect all information bits, whereas TopoCode acts as a data-dependent prior with unprotected side information; a denoising baseline is missing. Second, and more serious: the TopoCode bits are transmitted over the same BPSK channel at the same SNR, without any protection. At 3 dB, a 256-bit side packet has an expected ~5.8 bit errors and the probability of receiving it intact is below 1%. The correction algorithm assumes the receiver's copy of the PD equals the transmitter's. The paper never models bit errors in the side information. The stability bound (2) does not save it: that inequality relates PD distance to pixel differences, not to corruption of the PD itself.\n\nNone of this is fatal to the idea. Protecting the PD with a strong code, adding a standard denoising baseline, running more than 200 Monte Carlo trials, and testing sensitivity to PD bit errors would directly address the concerns. As it stands, the evidence is not sufficient to support the central claim, but the concept is worth engaging with.\n\nThis paper is for researchers in semantic communication and topological data analysis. I would send it to peer review, not desk-reject it, because the idea is new and the flaws are fixable. I would not cite it in its current form.","headline":"Original idea with a real methodological gap: the correction target is assumed error-free while being sent over the same noisy channel, and the evaluation metric is the same function being minimized.","tokens_in":8110,"tokens_out":2519,"would_cite":false,"duration_ms":24918,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Persistence diagrams can serve as message-level error-detection and correction codes for images.","keywords":["persistent homology","topological data analysis","error detection","error correction","image communication","message-level QoS","Wasserstein distance","semantic communication"],"falsifier":"Transmit the TopoCode bits over the same BPSK channel with bit errors and run the proposed correction: if the received diagram is corrupted, the optimizer may chase a wrong target; a clear test is to compare recovered-image PSNR and SSIM with and without error protection on the TopoCode bits at 3 dB SNR. If performance degrades sharply when the diagram has even a few bit errors, the scheme's low-overhead premise is undermined.","tokens_in":6954,"feed_emoji":"📡","tokens_out":4982,"duration_ms":43738,"temperature":0.7,"pith_summary":"This paper argues that a compact topological summary of an image—its persistence diagram, the birth-death pairs of connected components and holes—can be transmitted alongside the image as a 'TopoCode' that lets the receiver detect which errors matter and then reconstruct the image's shape. The point is to move error control from bit fidelity to message-level fidelity: in applications like extended reality and holographic communication, a few bit errors that smear a boundary are irrelevant, while errors that destroy the overall structure are critical. The author shows that comparing persistence diagrams via Wasserstein distance flags significant topological damage, and that gradient descent on that distance with a regularization term can remove noise and restore the image's topology. Simulations on MNIST and Omniglot indicate that at low SNR TopoCode recovers images more faithfully than LDPC or convolutional codes while adding only a few percent of overhead. The claim is that meaningful communication should care about the shape of the data, not just the bits.","feed_headline":"Topological shape codes fix image errors bit-level codes miss","feed_subtitle":"A persistence-diagram TopoCode adds ~4% overhead and restores image structure at low SNR where LDPC and convolutional codes struggle.","key_machinery":"The persistence diagram (PD) is the central object: a multiset of birth-death pairs tracking the appearance and disappearance of homology classes (connected components in H0, loops in H1, voids in H2) under a cubical-complex filtration. The p-Wasserstein distance between two PDs—the minimal cost of matching their points, with unmatched points sent to the diagonal—is the comparison engine, backed by a stability bound that ties PD distance to the input image difference. Total persistence (the sum of distances of BD pairs to the diagonal) sets a noise floor for removing insignificant pairs, and the correction step minimizes the Wasserstein loss against the transmitted PD with a regularizer that pushes pixel values toward 0 or 1. These pieces make the pipeline fully differentiable, so no training is required.","core_discovery":"The central claim is that persistent homology gives a training-free, data-modality-agnostic way to detect and correct message-level errors. Specifically, the persistence diagram of a grayscale image computed with a cubical complex is transmitted as side information; the receiver computes the diagram of the received image and uses the p-Wasserstein distance (with p=2) to measure topological mismatch. By thresholding the total persistence, insignificant noise pairs near the diagonal are filtered, so the distance reflects only meaningful structural changes. For correction, the receiver minimizes the Wasserstein distance between the received and transmitted diagrams plus a binarizing regularizer, using gradient descent that is valid because the filtration and distance are differentiable. The author demonstrates that this restores image topology at low SNR where bit-level codes produce visibly noisy reconstructions, and that the resulting code occupies roughly 4% of the source length.","pith_inferences":["A testable extension is to protect the TopoCode bits themselves (e.g., with a lightweight CRC or by embedding them in a pilot symbol) and measure how sensitive the correction is to PD corruption; the current scheme assumes the transmitted diagram arrives intact, which is not guaranteed over the same channel.","The method's reliance on the Wasserstein distance as an optimization target may be sensitive to local minima; the paper does not analyze convergence guarantees, so a natural follow-up is to compare gradient descent against alternative PD-based losses or regularizers.","Since the paper uses simple MNIST and Omniglot images with few BD pairs, a stress test on natural images could reveal whether the overhead stays small when PDs are large or when the topology is dominated by texture-like noise.","If message-level fidelity becomes the QoS metric, TopoCode could be paired with adaptive modulation so that only structurally important packets receive protection, changing how link-layer retransmission is triggered."],"forward_implications":["If TopoCode's claims hold, low-SNR image links can be made semantically robust by adding a few percent of topological side information rather than heavy bit-level coding.","The same machinery extends, in principle, to any data modality with a well-defined filtration, such as point clouds (using Vietoris-Rips or Alpha complexes), time series, and text.","Error significance becomes a tunable quantity: raising the persistence threshold lets the receiver accept transmissions whose noise is confined to small-scale features, reducing retransmission requests.","The approach offers a training-free alternative to deep-learning semantic communication, with different trade-offs in generalization and overhead.","Because correction targets topology only, it can be combined with bit-level codes: TopoCode handles structural damage, while conventional codes clean residual bit errors at higher SNR."],"supporting_citations":[{"why":"Provides the p-Wasserstein distance computation used to compare persistence diagrams in error detection.","marker":"[7]"},{"why":"Supplies the Wasserstein stability bound that justifies using PD distance as a proxy for image difference.","marker":"[9]"},{"why":"Establishes differentiability of persistent-homology optimization, enabling the gradient-descent correction step.","marker":"[10]"},{"why":"Gives efficient cubical-complex persistence computation, the filtration used for images.","marker":"[11]"},{"why":"The software library used in the simulations to compute persistence diagrams and Wasserstein distances.","marker":"[8]"}],"fun_headline_variants":["TopoCode adds 4% overhead to fix image errors bit codes miss","Persistent homology repairs images at low SNR with minimal cost","Shape-based error correction restores images where bit codes struggle","TopoCode: topological error correction for low-SNR images"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire method assumes that the receiver's copy of the transmitted persistence diagram is accurate enough to be the optimization target, even though the diagram's own bits travel over the same noisy channel with no extra protection.","fun_headline_variants_meta":{"raw":{"variants":["TopoCode adds 4% overhead to fix image errors bit codes miss","Persistent homology repairs images at low SNR with minimal cost","Shape-based error correction restores images where bit codes struggle","TopoCode: topological error correction for low-SNR images"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000729,"raw_usage":{"total_tokens":3208,"prompt_tokens":834,"completion_tokens":2374,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":450,"completion_tokens_details":{"reasoning_tokens":2302}},"tokens_in":450,"tokens_out":2374,"duration_ms":17841,"temperature":1.0,"reasoning_tokens":2302,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:08:57.795886+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Transmit the TopoCode bits over the same BPSK channel with bit errors and run the proposed correction: if the received diagram is corrupted, the optimizer may chase a wrong target; a clear test is to compare recovered-image PSNR and SSIM with and without error protection on the TopoCode bits at 3 dB SNR. If performance degrades sharply when the diagram has even a few bit errors, the scheme's low-overhead premise is undermined.","supporting_citations":[{"cited_title":"Geometry helps to compare persistence diagrams,","cited_arxiv_id":null,"evidence_quote":"Provides the p-Wasserstein distance computation used to compare persistence diagrams in error detection."},{"cited_title":"Optimizing persistent homology based functions,","cited_arxiv_id":null,"evidence_quote":"Establishes differentiability of persistent-homology optimization, enabling the gradient-descent correction step."},{"cited_title":"Efficient computation of persistent homology for cubical data,","cited_arxiv_id":null,"evidence_quote":"Gives efficient cubical-complex persistence computation, the filtration used for images."},{"cited_title":"[Online]","cited_arxiv_id":null,"evidence_quote":"The software library used in the simulations to compute persistence diagrams and Wasserstein distances."}],"review_version":1}