TopoCode: Topologically Informed Error Detection and Correction in Communication Systems

Hongzhi Guo

REVIEW 4 major objections 5 minor 11 references

TopoCode: Topologically Informed Error Detection and Correction in Communication Systems

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Persistence diagrams can serve as message-level error-detection and correction codes for images.

desk verdict 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. read the letter →

arxiv 2411.12825 v3 pith:EZSOY2MR submitted 2024-11-19 cs.MM

Hongzhi Guo This is my paper · ORCID

classification cs.MM

keywords persistenthomologytopologicaldataanalysiserrordetectioncorrectionimagecommunicationmessage-levelQoSWassersteindistancesemantic

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

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.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the 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.

What 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).

Where 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.

None 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.

This 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.

Referee Report

4 major / 5 minor

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.

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 (4)

[§IV, Fig. 7; §III, Eq. (3)] 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.
[§IV, Figs. 8c/9c; §III, Eq. (4)] 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.
[§IV, Figs. 8 and 9] 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.
[§III, Eq. (2)] 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.

minor comments (5)

[§II] There is a typo: 'BD paris' should be 'BD pairs'.
[Throughout] The name is written inconsistently as 'Topocode' and 'TopoCode'; one spelling should be used throughout.
[§IV] 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.
[§IV, Fig. 7] 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.
[§IV, last paragraph] 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.

Circularity Check

1 steps flagged · score 6.0 of 10

The reported Wasserstein-distance metric is exactly the loss being minimized, so the topological-fidelity claim is partly forced by construction; the unmodeled corruption of the Topocode reference is a separate serious gap.

fitted input called prediction [Sec. III, Eq. (4); Sec. IV, Figs. 8(c)-9(c)]
"L =γ PHh=0 dw,p,h(Dh(fl(X)), Dh(fl( ˆX)))+ Pi min(| ˆX[i]|, |1 − ˆX[i]|). ... We can minimize L and obtain ˜X which is the recovered version of X. ... Topocode always shows small Wasserstein distance and short code length."

The corrected image is defined as the minimizer of L, whose first term is the p-Wasserstein distance between the candidate image's persistence diagram and the Topocode diagram. The paper then reports the p-Wasserstein distance of the recovered image as a headline performance metric (Figs. 8c and 9c) and concludes that Topocode always achieves small Wasserstein distance. Minimizing a loss and then reporting the minimized loss as evidence of success is not an independent test; it only confirms that the optimizer reduced its own objective. PSNR and SSIM remain independent evidence, but the Wasserstein-distance panels are forced small by construction, and the comparison with LDPC/convolutional codes on that metric is therefore not meaningful.

full rationale

The paper has no load-bearing self-citation or imported uniqueness theorem: refs. [2] and [3] are contextual, and the stability bound in Eq. (2) is an external mathematical result, so those patterns do not raise the score. The central circularity is narrower but real: Eq. (4) explicitly minimizes the p-Wasserstein distance to the Topocode diagram, and the paper reports that same distance as a performance metric, so the claim that Topocode always shows small Wasserstein distance is a tautological consequence of the optimization. The PSNR and SSIM results are independent, non-circular evidence, and the comparisons with LDPC and convolutional codes are external benchmarks, which prevents a higher score. However, a serious non-circular gap compounds this: the Topocode bits are transmitted over the same BPSK channel with the same SNR and no error protection, while the correction procedure in Eq. (3) treats the received Topocode as the reference. At low SNR, a 256-bit Topocode (the paper's own example) will almost certainly contain bit errors, so the reported small Wasserstein distances and reconstruction gains assume an uncorrupted reference that the simulations do not model. This is a correctness gap rather than a definitional circularity, but it further undermines the claim that small Wasserstein distance reflects fidelity to the original message. The authors' stated limitation that Topocode cannot recover non-topological information is acknowledged honestly and does not add circularity. Overall, because one central metric reduces by construction while independent metrics remain, the score is 6.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical or mathematical entities are introduced. The paper relies on existing TDA concepts (persistence diagrams, Wasserstein distance) and on several domain assumptions about channel behavior and the sufficiency of topological information for reconstruction, the most fragile being that the transmitted PD arrives uncorrupted.

free parameters (3)

α (total persistence threshold) = 0%, 15%, 30%, 45% in simulations
The threshold fraction for removing insignificant PD pairs is chosen by hand and varied across a range; there is no principled selection criterion.
γ (loss function weight) = not specified
The weight balancing Wasserstein distance and the regularization term is stated to impact results but its value and selection rule are not given.
Optimization hyperparameters (learning rate, iterations) = not specified
The gradient descent details are omitted, making the exact reconstruction results unreproducible.

assumptions (4)

standard math Stability of persistence diagrams (Eq. 2): the p-Wasserstein distance between PDs is bounded by a constant times the Lp distance between the images.
This is a known theorem from Skraba and Turner [9], invoked to justify detection, and it is used as an upper bound, not a bidirectional equivalence.
domain assumption The transmitted persistence diagram is received accurately enough to serve as a valid target for reconstruction.
The PD is sent over the same noisy BPSK channel without error protection, yet the algorithm assumes it is correct. This is not validated or discussed.
domain assumption The optimization in Eq. (3) convergences to an image whose topology matches the transmitted PD and which approximates the original content.
No convergence proof or uniqueness result is provided; many images can share the same PD, so the global minimum may not correspond to the original image.
domain assumption Cubical complex filtration is an appropriate and sufficient topological descriptor for grayscale images in this application.
The paper uses cubical complexes for images, but does not compare with other filtrations or discuss whether H0 and H1 capture the information needed for reconstruction.

how reviews work

0 comments

Cite this review

Pith. "Pith review of TopoCode: Topologically Informed Error Detection and Correction in Communication Systems." pith.science (2026). https://pith.science/paper/EZSOY2MR

@misc{pith2026241112825,
  author       = {Pith},
  title        = {Pith review of: TopoCode: Topologically Informed Error Detection and Correction in Communication Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EZSOY2MR}},
  note         = {Machine review of arXiv:2411.12825}
}

read the original abstract

Traditional error detection and correction codes focus on bit-level fidelity, which is insufficient for emerging technologies like eXtended Reality (XR) and holographic communications requiring high-data-rate, low-latency systems. Bit-level metrics cannot comprehensively evaluate Quality-of-Service (QoS) in these scenarios. This letter proposes TopoCode which leverages Topological Data Analysis (TDA) and persistent homology to encode topological information for message-level error detection and correction. It introduces minimal redundancy while enabling effective data reconstruction, especially in low Signal-to-Noise Ratio (SNR) conditions. TopoCode offers a promising approach to meet the demands of next-generation communication systems prioritizing semantic accuracy and message-level integrity.

Figures

Figures reproduced from arXiv: 2411.12825 by the authors.

**Figure 4.** Illustration of topological total persistence. [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗

**Figure 5.** Example of error detection using Topocode. The PSNR and SSIM of [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗

Figures from the paper (2 more)

**Figure 8.** Figure 8: Error correction performance comparison using MNIST dataset. [PITH_FULL_IMAGE:figures/full_fig_p005_8.png]

**Figure 9.** Figure 9: Error correction performance comparison using Omniglot dataset. [PITH_FULL_IMAGE:figures/full_fig_p005_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

11 extracted references · 8 canonical work pages

[1]

Proakis and M

J. Proakis and M. Salehi, Digital Communications, 5th edition . McGraw-Hill Higher Education, 2008

2008
[2]

Holographic-type communication: A new challenge for the next decade,

I. F. Akyildiz and H. Guo, “Holographic-type communication: A new challenge for the next decade,” ITU Journal on Future and Evolving Technologies, vol. 3, no. 2, pp. 421–442, 2022

work page 2022
[3]

Mulsemedia com- munication research challenges for metaverse in 6G wireless systems,

I. F. Akyildiz, H. Guo, R. Dai, and W. Gerstacker, “Mulsemedia com- munication research challenges for metaverse in 6G wireless systems,” ITU Journal on Future and Evolving Technologies , vol. 4, no. 4, pp. 562–579, 2023

work page 2023
[4]

Deep learning enabled semantic communication systems,

H. Xie, Z. Qin, G. Y . Li, and B.-H. Juang, “Deep learning enabled semantic communication systems,” IEEE Transactions on Signal Pro- cessing, vol. 69, pp. 2663–2675, 2021

2021
[5]

Topological multimodal sensor data analytics for target recognition and information exploitation in contested environments,

P. T. Schrader, “Topological multimodal sensor data analytics for target recognition and information exploitation in contested environments,” in Signal Processing, Sensor/Information Fusion, and Target Recognition XXXII, vol. 12547. SPIE, 2023, pp. 114–143

work page 2023
[6]

A survey of topological machine learning methods,

F. Hensel, M. Moor, and B. Rieck, “A survey of topological machine learning methods,” Frontiers in Artificial Intelligence, vol. 4, p. 681108, 2021

work page 2021
[7]

Geometry helps to compare persistence diagrams,

M. Kerber, D. Morozov, and A. Nigmetov, “Geometry helps to compare persistence diagrams,” ACM J. Exp. Algorithmics , vol. 22, Sep. 2017

work page 2017
[8]

[Online]

The GUDHI Project, GUDHI User and Reference Manual , 2024, GUDHI. [Online]. Available: https://gudhi.inria.fr/doc/latest/

work page 2024

Show all 11 references

[9]

Wasserstein stability for persistence diagrams,

P. Skraba and K. Turner, “Wasserstein stability for persistence diagrams,” arXiv preprint arXiv:2006.16824 , 2020

2006 arXiv
[10]

Optimizing persistent homology based functions,

M. Carriere, F. Chazal, M. Glisse, Y . Ike, H. Kannan, and Y . Umeda, “Optimizing persistent homology based functions,” in International conference on machine learning . PMLR, 2021, pp. 1294–1303

2021
[11]

Efficient computation of persistent homology for cubical data,

H. Wagner, C. Chen, and E. Vuc ¸ini, “Efficient computation of persistent homology for cubical data,” in Topological methods in data analysis and visualization II: theory, algorithms, and applications . Springer, 2011, pp. 91–106

2011

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [1]

Proakis and M

J. Proakis and M. Salehi, Digital Communications, 5th edition . McGraw-Hill Higher Education, 2008

2008

[2] [2]

Holographic-type communication: A new challenge for the next decade,

I. F. Akyildiz and H. Guo, “Holographic-type communication: A new challenge for the next decade,” ITU Journal on Future and Evolving Technologies, vol. 3, no. 2, pp. 421–442, 2022

work page 2022

[3] [3]

Mulsemedia com- munication research challenges for metaverse in 6G wireless systems,

I. F. Akyildiz, H. Guo, R. Dai, and W. Gerstacker, “Mulsemedia com- munication research challenges for metaverse in 6G wireless systems,” ITU Journal on Future and Evolving Technologies , vol. 4, no. 4, pp. 562–579, 2023

work page 2023

[4] [4]

Deep learning enabled semantic communication systems,

H. Xie, Z. Qin, G. Y . Li, and B.-H. Juang, “Deep learning enabled semantic communication systems,” IEEE Transactions on Signal Pro- cessing, vol. 69, pp. 2663–2675, 2021

2021

[5] [5]

Topological multimodal sensor data analytics for target recognition and information exploitation in contested environments,

P. T. Schrader, “Topological multimodal sensor data analytics for target recognition and information exploitation in contested environments,” in Signal Processing, Sensor/Information Fusion, and Target Recognition XXXII, vol. 12547. SPIE, 2023, pp. 114–143

work page 2023

[6] [6]

A survey of topological machine learning methods,

F. Hensel, M. Moor, and B. Rieck, “A survey of topological machine learning methods,” Frontiers in Artificial Intelligence, vol. 4, p. 681108, 2021

work page 2021

[7] [7]

Geometry helps to compare persistence diagrams,

M. Kerber, D. Morozov, and A. Nigmetov, “Geometry helps to compare persistence diagrams,” ACM J. Exp. Algorithmics , vol. 22, Sep. 2017

work page 2017

[8] [8]

[Online]

The GUDHI Project, GUDHI User and Reference Manual , 2024, GUDHI. [Online]. Available: https://gudhi.inria.fr/doc/latest/

work page 2024