{"id":"e580449c-694d-45aa-ba0a-0c872cf7dba6","arxiv_id":"2412.19825","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"Using persistence diagrams as transmitted semantic summaries can drastically cut bit rate and improve error robustness for topology-based classification, but the reported rate advantage rests on a self-defined metric.","lead":"A group of engineers proposes that instead of sending raw sensor data over a network, a transmitter should send a persistence diagram, a compact topological summary of the data's shape. In simulations on a loop-counting task, this 'PD semantics' appears to need far fewer bits and to survive channel errors better than raw data or autoencoder representations, though the rate measure used is non-standard.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PD rate in Eq. (13) multiplies per-symbol entropy H(q_m) by the codebook support size M_S instead of the random number of PD points N_S, so the reported bits/object are not actual message lengths.","rationale":"The paper's central claim is that PD semantics achieve large rate reductions and robustness gains; the specific numbers 15x and 23x are the headline results. The reader's identified weakest assumption is exactly the rate definition in Eq. (13), and my reading confirms this is the most load-bearing problem. In Section IV, each PD point is quantized to one of m^2 cells, so an object with N_S points produces N_S symbols. Eq. (13) instead uses M_S = m(m+1)/2 as the number of symbols per object, which is the number of cells with nonzero support under the triangular constraint, not a data-dependent count. Since N_S is random (Figure 8a), the formula produces a number that is not the source-coded message length. The error propagates to Section VI-C4, where r_PD = 30.58 bits/object is used to budget BCH codewords and claim 15x and 23x reductions. A corrected rate would need to include the entropy of N_S and the per-symbol entropy summed over actual features. I also note the task construction (class labels are loop counts, exactly the H1 invariant captured by PDs) makes the inference comparison favorable to PD, but the rate definition is the more decisive flaw because it directly invalidates the quantitative claims. The authors do provide a dataset link and detailed experimental setup, which is helpful for reproducibility, but without a valid rate metric the empirical evidence does not support the central conclusions.","tokens_in":19281,"tokens_out":8130,"duration_ms":71358,"concrete_test":"Recompute the PD and AE rates using the actual per-object symbol counts: for each object i, compute R_PD_corrected = (1/|N|) * sum_i N_{S_i} * H(q_m) (and, more accurately, add the entropy of the sequence including N_S, e.g., H(N_S) + E[N_S] * H(q_m) under conditional independence). Then regenerate Figures 6 and 7 and the reported 15x and 23x reductions. If R_PD_corrected(m=10) is substantially larger than 30.58 bits/object—for instance, above 100 bits/object given the distribution of N_S in Figure 8a—the headline rate advantages shrink and may no longer hold.","verdict_should_be":"REJECT","load_bearing_attack":"The rate definition in Section V-B, Eq. (13), sets R_PD = M_S * H(q_m) with M_S = m(m+1)/2. This confounds the number of quantization cells with nonzero support with the number of transmitted symbols per object. Per Section IV-A, each PD S = {s_1,...,s_N_S} is quantized element-wise via q_m (Eqs. 6-7), yielding one alphabet symbol per topological feature. H(q_m) is a per-symbol entropy (bits per feature). The actual number of symbols per object is N_S, which is random (Figure 8a shows a wide distribution, e.g., values up to about 150). The per-object rate should be at least E[N_S] * H(q_m), plus the cost of conveying N_S or a termination mechanism, not M_S * H(q_m). M_S is a geometric constant of the triangular grid and is unrelated to the number of features in a given PD. Consequently, r_PD = 30.58 bits/object at m=10 is not the number of bits needed to transmit a PD; it is the per-symbol entropy multiplied by the number of nonzero-probability cells. The same template is applied to AE-latent representations (Section V-C, MA = m^2) and raw data (Section V-D, MG = m^2), with the same defect. Because the headline claims (15x rate reduction over AE, 23x resource reduction over raw data) in Section VI-C4 are computed from these non-message-length quantities, the central quantitative evidence is invalidated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes transmitting persistence diagrams (PDs) as 'structural semantics' from a sensor to a decision-maker, defines semantic distortion and rate for PDs through uniform 2D vector quantization, and empirically compares rate-distortion-accuracy trade-offs against raw point clouds and autoencoder (AE) latent representations. In a binary symmetric channel with BCH coding, it claims PDs achieve over 80% inference accuracy at a roughly 15 times lower rate than AE latents and a 23-fold resource reduction versus raw data. The central quantitative claim rests on the rate formula R_PD = M_S * H(q_m) in Eq. (13), where M_S is the number of nonzero-probability quantization regions rather than the number of transmitted PD points per object. The paper also provides an appendix treating the number of PD points as random, which exposes an internal inconsistency in the rate definition.","tokens_in":19635,"tokens_out":8975,"duration_ms":80004,"significance":"The idea of using topological summaries as communication semantics is novel and timely, and the paper contributes a complete experimental pipeline: VR-filtration PDs, PersLay-based classifiers, cross-validation, and publicly available code/data. The empirical observation that PD point distributions are highly sparse is interesting and could motivate efficient coding schemes. However, the headline rate-reduction numbers are built on a rate definition that does not count the actual message length per object, so the quantitative contributions are not currently supported. The conceptual framework is potentially salvageable after a corrected rate formulation and re-evaluated experiments, but the order-of-magnitude efficiency claims must be substantially revised.","major_comments":[{"comment":"The semantic rate R_PD = M_S * H(q_m) does not correspond to the number of bits required to transmit a PD. A PD is a multiset of N_S points, each mapped to one alphabet symbol by the quantizer, so the per-object message length is the sum of N_S symbols plus the cost of conveying N_S (or a termination mechanism). The factor M_S = m(m+1)/2 is the number of quantization cells with nonzero probability, not the number of transmitted points. The paper itself models N_S as a random variable in Appendix A, and Fig. 8a shows a broad distribution with values up to about 150, so the rate definition is internally inconsistent. The same defect appears in §V-C and §V-D, where M_A = m^2 and M_G = m^2 are used instead of the actual numbers of transmitted points (d for AE, N_G for raw data). Consequently, the values r_PD = 30.58, r_AE = 466.19, and r_Raw = 6035.20 bits/object reported in §VI-C4 are not actual message lengths.","section":"§V-B, Eq. (13); Appendix A and Fig. 8a"},{"comment":"Because the 15-fold and 23-fold rate reductions are computed by applying BCH overhead to the incorrect r_PD and r_AE values, the paper's central efficiency claims are not established. With a correct rate computation, the AE rate for d=27 and m=10 would be d * H(q_a) = 27 * H(q_a) rather than 100 * H(q_a), while the PD rate would be E[N_S] * H(q_s) rather than 55 * H(q_s). Depending on E[N_S], the reported 15x advantage over AE could shrink by roughly an order of magnitude. The authors should re-derive the rates and rerun the comparisons before making any efficiency claim.","section":"§VI-C4, Eq. (16) and Fig. 7"},{"comment":"The three classes are defined by the number of loops in the handwritten digit (C1 = one loop, C2 = two loops, C3 = no loops), which is exactly the topological feature captured by H1 of the VR filtration. This makes the classification task intrinsically aligned with the PD representation and creates a favorable comparison against raw data and AE latent representations, whose features are not constructed from this prior knowledge. To support a general claim that PD semantics improve communication efficiency, the evaluation should include tasks whose labels do not coincide so directly with the transmitted topology, or the paper should at least clearly acknowledge this bias.","section":"§VI-A, class label definition"}],"minor_comments":[{"comment":"There is a typo in the caption: 'PD semnatics' should be 'PD semantics'.","section":"Fig. 7 caption"},{"comment":"In Eq. (14), A_PD(m,t) is described as 'average error' but is then used as inference accuracy; please fix the terminology.","section":"§VI-B, Eq. (14)"},{"comment":"The paper repeatedly describes the distortion and rate definitions as 'qualitative', but Eqs. (8) and (13) are quantitative definitions; this wording should be adjusted.","section":"§V-A, §V-B"},{"comment":"No confidence intervals or variability bands are reported for the estimated rates and distortions in Figs. 3 and 6; given the small dataset (600 images), bootstrap or repeated subsampling would strengthen the evidence.","section":"Fig. 3 and Fig. 6"}],"recommendation":"major_revision","confidential_remarks":"The rate-definition error is serious and affects the headline claims, so the paper cannot be accepted in its current form. I recommend major revision rather than reject because the conceptual framework is novel and could be repaired with a corrected rate formulation and re-evaluated experiments; however, the quantitative conclusions may change substantially and the authors must also address the task-selection bias in the dataset. The paper's code/data availability and reproducible pipeline are strengths that should be preserved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know this paper is the first to apply persistence diagrams (PDs) as structural semantics in goal-oriented communication, and it defines rate/distortion trade-offs for them. That part is genuinely new and worth reading. The empirical observation that PD point distributions are sparse, leading to low entropy per symbol, is also real and potentially useful.\n\nBut the central quantitative claims don't survive contact with the actual definitions. The rate in Eq. (13) is R_PD = M_S * H(q_m), where M_S is the number of quantization cells with nonzero probability, not the number of points in a PD. Each object's PD has a random number N_S of points (Figure 8a shows a wide distribution), and the quantizer produces one symbol per point. So the actual message length per object is around E[N_S] * H(q_m), not M_S * H(q_m). The same mistake applies to the AE and raw-data rates, which use m^2 instead of the actual number of latent vectors or point-cloud points. Because the inflation factors differ across representations (M_S vs. E[N_S], m^2 vs. d, m^2 vs. E[N_G]), the reported 15x and 23x gains are not trustworthy. The true ratios could be substantially smaller. This is a load-bearing flaw, not a minor bookkeeping issue.\n\nThe inference benchmark is also circular. The three classes are defined by loop counts (0, 1, or 2 loops), and the PD's H1 homology is exactly the invariant that counts loops. So the representation is essentially given the answer key. That makes the high accuracy at low rates unsurprising and undercuts the comparison with AE or raw data, which have to learn the relevant structure from scratch.\n\nWhat the paper does well: the exposition is clear, the dataset and code are linked, and the idea of using topological summaries as structural semantics is worth exploring. The sparsity observation could matter for practical systems if the rate were defined correctly.\n\nNet: this needs major revision before it can support its claims. The authors should redefine rate as a function of the actual number of transmitted points (or provide a rigorous joint source-channel treatment), and they should use a task where the label is not directly encoded in the PD. As is, I'd recommend rejecting, but sending it to peer review with a clear request to fix these issues. It's not a desk-reject; the idea is too fresh to bury without comment.\n\nRecommendation: engage with it as a borderline paper; give the authors a chance to correct the rate definition and benchmark.","headline":"Novel idea—PDs as semantics—but the rate metric is misdefined and the benchmark is circular, so the headline gains don't hold.","tokens_in":20123,"tokens_out":5153,"would_cite":false,"duration_ms":45376,"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":"Sending a persistence diagram instead of the raw point cloud lets a receiver classify the object at roughly a 15-fold lower coded rate than an autoencoder-latent benchmark, and about 23-fold lower than raw data, even on a noisy binary…","keywords":["persistence diagrams","semantic communication","topological data analysis","rate-distortion trade-off","inference accuracy","quantization","BCH error-correcting codes","structural semantics"],"falsifier":"Use the empirical distribution of $N_S$ from the paper's appendix to entropy-code each object's actual PD points and count the true per-object bits; if the coded PD rate is no longer roughly 15 times below the coded autoencoder-latent rate, or the 23-fold raw-data saving vanishes, the rate-efficiency claim is refuted.","tokens_in":19109,"feed_emoji":"📡","tokens_out":13757,"duration_ms":115160,"temperature":0.7,"pith_summary":"The paper argues that a persistence diagram—a multiset of birth–death pairs $(b,d)$ with $b<d$ that records when connected components, loops, and voids appear and disappear in a filtration of a point cloud—is a better object to transmit than the raw point cloud or an autoencoder's latent code when the receiver's job is classification. It defines a semantic distortion as the mean squared error between a PD point and its quantized representative, and a semantic rate as the quantizer entropy times the number $M_S=m(m+1)/2$ of occupied quantization regions. On a handwritten-digit point-cloud benchmark, the reported trade-offs show PD semantics holding over 80% inference accuracy at about 30.58 bits per object uncoded, and BCH-coded PD semantics beating coded autoencoder latents by roughly 15 times in rate while giving more than a 23-fold saving over raw data under a binary symmetric channel. If correct, this makes compact topological summaries a practical alternative in goal-oriented communication, leaving room in the bit budget for error-correction redundancy.","feed_headline":"Persistence diagrams need ~15x fewer coded bits than autoencoders","feed_subtitle":"Shape summaries keep classification accuracy high on noisy channels, beating raw data and autoencoder latents on rate.","key_machinery":"The load-bearing object is the persistence diagram, a multiset of birth–death pairs $(b,d)$ in the open half-plane $b<d$, computed through a Vietoris–Rips filtration of the point cloud. The rate mechanism is a uniform 2D vector quantizer: each PD point is mapped to the center of one of $m^2$ cells in a bounded box, and the semantic rate is $R_{\\mathrm{PD}} = M_S H(q_m)$ bits per object, where $M_S = m(m+1)/2$ counts the cells with nonzero probability and $H(q_m)$ is the entropy of the quantized symbol. The quantizer supplies both the distortion measure $D_{\\mathrm{PD}}^{\\mathrm{MSE}} = \\mathbb{E}\\|s - q_m(s)\\|^2$ and the rate law used in all comparisons. On the inference side, PDs are converted to fixed-size vectors by a permutation-invariant neural layer so a standard classifier can be trained at the receiver; the same quantization pipeline is applied to autoencoder latents and raw points as baselines, and in noisy-channel experiments the quantized symbol stream is entropy-coded and protected by BCH error-correcting codes (cyclic codes that correct up to a prescribed number of bit errors).","core_discovery":"The central claim is that persistence diagrams, quantized and transmitted as symbols, preserve enough shape information for a downstream neural classifier to maintain high accuracy at rates where raw data and autoencoder latents fail. The paper shows this by computing PDs from Vietoris–Rips filtrations of the point clouds, quantizing each PD point with a uniform $m\\times m$ grid on a bounded support box, and feeding the quantized coordinates through a permutation-invariant neural layer into a classifier. In a binary symmetric channel with crossover probabilities $\\alpha=0.1$ and $\\alpha=0.12$, BCH error-correcting codes wrapped around the quantized PD stream reach the error-free accuracy level at coded rates as low as 254.35 bits/object, about 15 times below the coded autoencoder-latent requirement; the same raw-data resource budget that infers one object supports about 23 PD-based inferences at comparable accuracy. The authors attribute this to sparsity of the PD distribution: probability mass concentrates in only $M_S$ of the $m^2$ cells, so semantic entropy is far lower than raw or latent coordinate entropy.","pith_inferences":["A testable extension is to recompute rates from each object's actual number of topological features: the rate formula counts $M_S=m(m+1)/2$ symbols per object, while $N_S$ per object is random (shown in the paper's own appendix), and real per-object message lengths may change the reported 15-fold and 23-fold comparisons.","The same quantization-plus-error-correction recipe should transfer to other structural summaries, such as witness complexes, cubical persistence, or persistent landscapes, whenever the summary distribution concentrates on few cells.","A natural next step is joint design of the quantization grid and the error-correcting code for a target accuracy, rather than fixing coarse quantization first and choosing a code afterwards.","The evidence opens a broader design question: when the transmitter's goal is a downstream task, the right semantic object may be the one with the sparsest distribution over the channel alphabet, not the one with the lowest raw-space distortion."],"forward_implications":["PD-based links can meet a target inference accuracy with far fewer bits than raw or learned-latent links, so the unused bit budget can be spent on forward error correction without exceeding the raw-data resource limit.","Because PD inference accuracy stays nearly flat as distortion and rate vary, an engineer can choose coarse quantization and still keep accuracy, simplifying the source code and improving tolerance to channel noise.","Under one raw-data transmission budget, a PD-based system can carry roughly 23 classifications at comparable accuracy, which translates directly into lower latency or higher throughput per unit bandwidth.","The rate-distortion gap between PD semantics and the alternatives is approximately a constant factor at the tested operating points, so the advantage is a structural property rather than a threshold artifact."],"supporting_citations":[{"why":"supplies the persistence homology and persistence diagram background defining the topological signatures used as semantics.","marker":"[18]"},{"why":"provides the permutation-invariant neural layer used to turn PDs into vector representations for the receiver's classifier.","marker":"[19]"},{"why":"is the computational topology reference for simplices, filtrations, and persistent homology behind PD generation.","marker":"[55]"},{"why":"supplies the high-rate entropy-coded quantization framework used to define semantic distortion and rate.","marker":"[57]"},{"why":"gives the entropy coding and binary symmetric channel model used in the noisy-channel experiments.","marker":"[61]"},{"why":"provides the BCH code tables and error-correction framework used to protect quantized PD symbols.","marker":"[62]"},{"why":"provides the handwritten-digit image source from which the point-cloud dataset is extracted.","marker":"[58]"}],"fun_headline_variants":["Persistence diagrams need ~15x fewer coded bits than autoencoders","Shape summaries slash channel rate 15x while keeping accuracy","PDs: 15x less rate than AE latents for high-accuracy inference","Tops-down: 23x more inferences per bit compared to raw data","Send geometry not points: PDs cut coded rate 15x"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The rate comparisons assume every object's persistence diagram can be described by exactly $M_S=m(m+1)/2$ quantized symbols, but the number of topological features per object is random, so the reported bits per object are not the actual source-coded message length.","fun_headline_variants_meta":{"raw":{"variants":["Persistence diagrams need ~15x fewer coded bits than autoencoders","Shape summaries slash channel rate 15x while keeping accuracy","PDs: 15x less rate than AE latents for high-accuracy inference","Tops-down: 23x more inferences per bit compared to raw data","Send geometry not points: PDs cut coded rate 15x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000242,"raw_usage":{"total_tokens":1547,"prompt_tokens":992,"completion_tokens":555,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":608,"completion_tokens_details":{"reasoning_tokens":459}},"tokens_in":608,"tokens_out":555,"duration_ms":6152,"temperature":1.0,"reasoning_tokens":459,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:55:12.355171+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use the empirical distribution of $N_S$ from the paper's appendix to entropy-code each object's actual PD points and count the true per-object bits; if the coded PD rate is no longer roughly 15 times below the coded autoencoder-latent rate, or the 23-fold raw-data saving vanishes, the rate-efficiency claim is refuted.","supporting_citations":[{"cited_title":"PersLay: A neural network layer for persistence diagrams and new graph topological signatures,","cited_arxiv_id":null,"evidence_quote":"provides the permutation-invariant neural layer used to turn PDs into vector representations for the receiver's classifier."},{"cited_title":"Edelsbrunner and J","cited_arxiv_id":null,"evidence_quote":"is the computational topology reference for simplices, filtrations, and persistent homology behind PD generation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the high-rate entropy-coded quantization framework used to define semantic distortion and rate."},{"cited_title":"Lin and D","cited_arxiv_id":null,"evidence_quote":"provides the BCH code tables and error-correction framework used to protect quantized PD symbols."}],"review_version":1}