{"id":"bacabf50-19e6-482e-9782-83ac0dc092c1","arxiv_id":"2506.03943","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Hypergraph lower Ricci curvature (HLRC) is a new closed-form, bounded curvature score for hyperedges that separates intra-community from bridge-like hyperedges.","lead":"Researchers introduce HLRC, a fast closed-form curvature score for hyperedges that separates tightly connected community links from sparse bridge links. It gives network scientists a cheap geometric descriptor for community detection and clustering in higher-order data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper contradicts itself on which hyperedge class gets higher HLRC: Section 2.2 says intra-community hyperedges get lower curvature on Contact High School, while earlier text and the Fig.","rationale":"The reader's weakest_assumption focuses on the heuristic, non-derived nature of Eq. (1), and I agree that the formula lacks a foundational derivation. However, the most immediately falsifiable problem is internal: the paper contains two mutually incompatible empirical statements about the sign of the HLRC difference between intra- and inter-community hyperedges. This is not a disagreement with an external consensus; it is a contradiction within the reported results. If the true direction is the one stated in the Section 2.2 sentence, then the positive-curvature-means-cohesive interpretation, the Fig. 2a color scheme, and the narrative around the synthetic experiments are all wrong. If the true direction is the one in the earlier text and Fig. 2 caption, then the page 5 sentence is a sign error. Either way, the manuscript as written does not support a definite reading of the central claim. Because this can be settled by re-running the released code and reporting the intra/inter distributions, the reader's CONDITIONAL verdict remains appropriate: the contribution may be correct after correction, but it should not be accepted as-is. The HORC range inconsistency is a separate issue that should also be fixed in revision.","tokens_in":20591,"tokens_out":4463,"duration_ms":41605,"concrete_test":"Run the released code on the Contact High School dataset, recompute HLRC for every hyperedge, and reproduce the intra- versus inter-community distributions shown in Fig. 2b, reporting mean, median, Wilcoxon rank-sum p-value, and effect size. Determine which class has higher curvature. Then repeat the same computation on the synthetic SBM hypergraphs from Section 2.2. If intra-community hyperedges have higher HLRC in both settings, the Section 2.2 sentence is a sign error and the central claim stands; if the real-data direction is reversed, the geometric interpretation and community-detection claims require revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.2 contains an unresolved contradiction about the sign of the reported effect. On page 4, the authors state that in the high-school hypergraph 'hyperedges within the same classroom tend to exhibit higher (more positive) curvature than those spanning multiple classrooms' (Fig. 2 caption) and that in the synthetic SBM results 'intra-community hyperedges consistently exhibit significantly more positive curvature than inter-community ones.' Immediately afterward, on page 5, they write that 'both HLRC and HORC sharply distinguish these categories, assigning significantly lower curvature to intra-community hyperedges and higher curvature to inter-community hyperedges.' These two statements cannot both be true. The first is consistent with the paper's geometric story, in which positive curvature indicates cohesive, clique-like structure and negative curvature indicates bridge-like bottlenecks; the second reverses that story. Since every downstream interpretation, including community detection, venue-level semantic analysis, and hypergraph clustering, relies on this sign convention, the central empirical claim is ambiguous until the direction is verified. A related but secondary inconsistency appears in the stated HORC range: Supplementary Note S1.1 says HORC always lies in [-2,-1], while S1.2 and the main text use [-2,1], which further complicates the comparisons in Fig. 2 and Table 3.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Hypergraph Lower Ricci Curvature (HLRC), a closed-form scalar assigned to each hyperedge via Eq. (1), combining inverse node-degree sums, hyperedge neighborhood size, and max/min node-degree normalizations. The authors claim HLRC is provably bounded in (-1,1], computable in near-linear time, and able to separate intra-community from inter-community hyperedges, with applications to community detection, venue-level co-authorship analysis, temporal trend tracking, and clustering of hypergraph collections. The evaluation includes special uniform hypergraphs, synthetic HSBM hypergraphs, the Contact High School dataset, the MADStat and MAG-10 co-authorship hypergraphs, and the Stex and Mus collections, with comparisons against HORC and HFRC.","tokens_in":20930,"tokens_out":9006,"duration_ms":82287,"significance":"If the identified inconsistencies are resolved, HLRC would be a practically attractive descriptor: it is closed-form, provably bounded in (-1,1], shows a clear separation in the synthetic SBM experiments, runs orders of magnitude faster than HORC, and is accompanied by public code and data. The main value is empirical scalability rather than mathematical novelty. However, the paper's central claim to be a 'curvature' measure is weakened by the admitted heuristic nature of Eq. (1), and the current text contains a direct contradiction about the sign of the intra/inter-community effect that every downstream interpretation depends on. The empirical results are potentially useful, but the presentation must first be made internally consistent.","major_comments":[{"comment":"The direction of the reported effect is stated twice in opposite ways. The Fig. 2 caption and the paragraph preceding it say that intra-classroom hyperedges have higher (more positive) HLRC and boundary/inter-classroom hyperedges have strongly negative HLRC, while the paragraph after Fig. 2b-d says 'both HLRC and HORC sharply distinguish these categories, assigning significantly lower curvature to intra-community hyperedges and higher curvature to inter-community hyperedges.' These statements cannot both be true. Since Sections 2.3, 2.4, and the Discussion interpret positive curvature as cohesion and negative curvature as bridging, the entire empirical story depends on which direction is correct. Please correct the statement, verify the sign in Fig. 2b, and state the effect direction consistently throughout.","section":"Section 2.2, Fig. 2 caption and text after Fig. 2b-d"},{"comment":"The allowed range of HORC is given inconsistently. Supplementary Note S1.1 states that 'regardless of which reasonable choices one makes for AGG and mu, the resulting HORC value always lies in the interval [-2,-1]', while S1.2 and Section 4.4 use [-2,1] (with 60 histogram bins spanning [-2,1]). If HORC is confined to [-2,-1], the histogram binning in Section 4.4 and the comparisons in Fig. 2 and Table 3 are based on a miscalibrated feature representation; if the range is [-2,1], then S1.1's universal bound is wrong. Please reconcile the two statements and justify or re-run the clustering comparison under the correct range.","section":"Supplementary Note S1.1/S1.2; Section 4.4"},{"comment":"The statement of the hypercycle theorem is internally inconsistent. Section 4.2 introduces the closed-form expression under the condition 'when the hypercyle is large enough i.e. m>=2k', but Supplementary Theorem 3 specifies different regimes (m=2; m>=3; m=3; m>=4; m<2k; m>=2k), and the formula given for k>2s in the supplement is said to hold for m>=3, not only for m>=2k. For example, with k=5, s=2, m=3, the main-text condition m>=2k excludes a case that the supplement's m>=3 branch covers. Moreover, the 'intermediate regimes' are explicitly omitted from both the displayed formulas and the proof. Please state the exact conditions for each branch and either supply the omitted formulas or restrict the theorem to the cases actually proved.","section":"Section 4.2 vs Supplementary S2.3, Theorem 3"},{"comment":"The paper says HLRC is defined 'for a given hyperedge of degree greater than two', but the real datasets in Table 1 contain large fractions of hyperedges of size 2 (70.3% for Contact High School, 40.8% for MADStat, 29.9% for MAG-10), and Section 4.4 states that curvature values were computed for 'all hyperedges' and 'edges of all sizes'. It is therefore unclear whether Eq. (1) was applied to de=2 hyperedges, how n_e is defined for them, and why the definition excludes them. Please clarify the domain of Eq. (1), and if pairwise hyperedges are included, state this explicitly and adjust the wording 'greater than two' (and Theorem 1's 'de>1').","section":"Section 2.1 and Table 1"},{"comment":"The paper calls HLRC an extension of lower Ricci curvature for graphs (reference [25]) but provides no derivation connecting Eq. (1) to that graph notion; the coefficients 1/2 and the offset -1 are presented without derivation, and the Discussion concedes that the 'theoretical foundations remain largely heuristic'. Since the geometric interpretation underlies the interpretation of all experiments, either give a precise derivation or limiting argument from the graph lower Ricci curvature, or reframe HLRC as a closed-form heuristic descriptor rather than as a curvature extension. This is not a reason to reject the empirical results, but the current framing overstates the theoretical status.","section":"Section 3 Discussion and Eq. (1)"}],"minor_comments":[{"comment":"The notation for average hyperedge size is inconsistent: Table 4 uses \\bar d_e, while Supplementary S2.2 uses \\bar d_v for the same quantity, which is also confusing because d_v elsewhere denotes node degree.","section":"Table 4 vs Supplementary S2.2"},{"comment":"The real-data separation in Fig. 2b-d is not accompanied by effect sizes or test statistics, although significance is reported for the synthetic SBM results. Please add quantitative summaries (medians, p-values) for the high-school comparison.","section":"Figure 2b-d"},{"comment":"The abstract and Section 2.4 claim that HLRC outperforms both HORC and HFRC, but Table 3 and the clustering pipeline compare only HLRC with HORC; HFRC is absent from the clustering evaluation. Please either add HFRC results or adjust the claim.","section":"Table 3 and Sections 2.4, 3"},{"comment":"There are several typographical errors and grammatical slips, including '2-reguluar' in the Fig. 1 caption, 'a interpretable' in Section 2.1, 'These results demonstrates' in Section 2.2, and 'hypercyle' in Section 4.2. A careful proofreading pass is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The empirical contribution is the main strength, and the code/data availability is a plus. My main concern is that the paper's title and framing promise a mathematical curvature theory that is not delivered; the sign contradiction and the HORC range inconsistency must be resolved before the empirical claims can be trusted. The manuscript is likely to be acceptable after a major revision, but the authors should either substantially strengthen the theoretical connection to lower Ricci curvature or explicitly re-scope the paper as an empirical heuristic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know: this paper defines HLRC, a closed-form curvature for hyperedges, proves it lies in (-1,1], and shows it separates intra- from inter-community hyperedges at low computational cost. The formula is genuinely new, extending the authors' earlier graph lower Ricci curvature with max/min normalizations and a shared-neighbor term. It is not a re-derivation of HORC or HFRC.\n\nWhat it does well: the bound proof is coherent, the special-case derivations (complete hypergraphs reach 1, hypergrids 0, hypertrees negative) are real work, and the synthetic SBM experiments show clean separation. The runtime advantage over HORC is dramatic, and the clustering results on Stex are strong. Code is public.\n\nThe soft spots are real but mostly fixable. The biggest is a direct contradiction in Section 2.2. The text first says intra-classroom hyperedges in the high-school data have higher (more positive) curvature than inter-classroom ones, and the synthetic results say the same. A few paragraphs later it says HLRC and HORC assign significantly lower curvature to intra-community hyperedges and higher to inter-community. Both cannot be true. Every downstream interpretation depends on the sign convention, so this has to be fixed before the paper is publishable. It looks like a typo, but it is load-bearing. Also, S1.1 says HORC always lies in [-2,-1] while S1.2 and the main text use [-2,1]; minor but sloppy. The Mus clustering omits Palestrina post-hoc (67% of the data); the authors disclose it, but it weakens that comparison.\n\nThe coefficients in Eq. (1) are chosen so the toy examples behave nicely, and the authors admit the theory is largely heuristic. That is acceptable for a practical descriptor, but the paper should keep that caveat front and center.\n\nWho this is for: anyone working on hypergraph analytics, community detection, or geometric deep learning. It deserves a serious referee; the flaws are fixable and the central claim—a cheap, bounded, informative curvature—holds up. I would send it out with a request to fix the contradiction and clarify the HORC range.","headline":"New closed-form hypergraph curvature measure that deserves refereeing, but the paper contains a direct, load-bearing contradiction about the sign of its main effect.","tokens_in":21374,"tokens_out":1675,"would_cite":true,"duration_ms":16587,"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":"A closed-form curvature score for hypergraph edges separates community-internal from bridge edges and does so at a fraction of optimal-transport cost.","keywords":["hypergraph curvature","lower Ricci curvature","community detection","higher-order networks","closed-form curvature","Ollivier-Ricci curvature","Forman-Ricci curvature","hypergraph embedding"],"falsifier":"Take the three stochastic-block-model hypergraphs from Fig. 1 and apply hyperedge swaps that keep every node's degree and every hyperedge's size fixed while scrambling which nodes share hyperedges across communities. If the intra- versus inter-community HLRC separation persists with the same Wilcoxon significance, the signal is dominated by degree statistics rather than by the shared-neighbor geometry the formula claims to measure.","tokens_in":20424,"feed_emoji":"🕸️","tokens_out":8004,"duration_ms":79296,"temperature":0.7,"pith_summary":"The paper proposes a new numerical score for the curvature of a hyperedge—a group interaction involving any number of nodes—called hypergraph lower Ricci curvature (HLRC). The score is a closed-form expression combining inverse node degrees with shared-neighbor and degree terms normalized by max and min neighborhood sizes, so it costs about as little as combinatorial Forman-Ricci curvature yet stays bounded in $(-1,1]$. The authors claim this quantity captures higher-order geometry: hyperedges inside a community score positive, bridge hyperedges score negative, complete hypergraphs score 1, and grid-like structures score 0. Across stochastic block models, a high-school contact network, co-authorship networks, and collections of hypergraphs, HLRC separates intra- from inter-community hyperedges and yields clustering embeddings that beat the optimal-transport-based HORC and the degree-only HFRC. The pitch is that HLRC gives network analysts a scalable geometric descriptor for community detection, anomaly detection, and hypergraph neural networks.","feed_headline":"A cheap closed-form curvature tells hypergraph communities apart","feed_subtitle":"In (-1,1], HLRC splits community edges from bridge edges at a fraction of optimal-transport cost.","key_machinery":"The load-bearing object is Eq. (1) itself, the HLRC functional. Its three additive ingredients—the inverse-degree sum $\\sum_{v\\in e}1/n_v$ (local density), the shared-neighborhood count $n_e$ weighted by $d_e/2-1$ (higher-order overlap), and the same term normalized by the maximum and minimum node-neighborhood sizes (boundary sensitivity)—encode the paper's notion of curvature. The $-1$ offset and the $1/2$ coefficient are design constants that make the expression collapse to 1 when every node in $e$ has the minimal possible neighborhood $d_e-1+n_e$, and force the lower bound $-1$ by positivity of the three additive terms. The proof of boundedness (Theorem 1) and the closed-form evaluations on hypercycles, hypertrees, hypergrids, and complete hypergraphs are what carry the interpretation of HLRC as a geometric invariant.","core_discovery":"The central claim is that there exists a closed-form quantity, Eq. (1),\n$$\n\\mathrm{HLRC}(e) = \\sum_{v\\in e} \\frac{1}{n_v} + \\frac{n_e + d_e/2 - 1}{\\max_{v\\in e} n_v} + \\frac{n_e + d_e/2 - 1}{\\min_{v\\in e} n_v} - 1,\n$$\nwhich is a legitimate lower Ricci curvature for an unweighted, undirected hyperedge $e$: it is provably bounded in $(-1,1]$, attains 1 on complete hypergraphs, 0 on hypergrids, is non-negative on hypercycles, and becomes negative on bridging or bottleneck hyperedges. The authors show that in synthetic stochastic-block-model hypergraphs and in real-world contact and co-authorship hypergraphs, HLRC assigns significantly higher values to intra-community hyperedges than to inter-community bridges, separates venues or conferences by their theory-versus-application character, tracks temporal changes in collaboration cohesion, and produces hypergraph-cluster embeddings with higher adjusted Rand index and adjusted mutual information than HORC-based embeddings—all with runtimes on the order of HFRC, far below HORC's optimal-transport cost.","pith_inferences":["A direct test the paper does not run is to correlate HLRC with the Wasserstein-based HORC over a large random corpus of hypergraphs; the reported comparisons are task-based, so they show HLRC winning on selected separations, not that HLRC approximates HORC as a number.","If the shared-neighbor term is the main driver, HLRC should also separate communities on the graph projection of any hypergraph; computing it on the projected graph and comparing would isolate what the hypergraph formulation adds beyond clique expansion.","The formula suggests a null-model normalization: replacing $n_e$ by its expectation under a configuration model would yield a z-scored curvature, turning the observed bimodal separation into a statistical test for surprising bridges.","Because HLRC is cheap to evaluate, it could serve as a prefilter that selects hyperedges for expensive HORC computation, approximating full geometric curvature with a bounded budget."],"forward_implications":["Community detection on hypergraphs can be done by thresholding a single per-hyperedge number instead of solving an optimal transport problem for every edge.","Curvature histograms become a cheap global signature for a whole hypergraph, letting collections of hypergraphs be clustered by semantic or stylistic type.","The fixed $(-1,1]$ scale makes curvature values comparable across datasets, addressing HFRC's lack of an intrinsic reference range.","Bridge and bottleneck hyperedges—candidate anomalies, transmission paths, or interdisciplinary links—can be flagged as the most negative HLRC edges.","Hypergraph neural networks can use HLRC as an edge-level filter to counter over-smoothing and over-squashing, a direction the authors propose."],"supporting_citations":[{"why":"Defines the graph lower Ricci curvature that HLRC extends and from which the $(-1,1]$ boundedness is inherited.","marker":"[25]"},{"why":"Introduces Ollivier-Ricci curvature on directed hypergraphs, forming the theoretical lineage of the HORC baseline.","marker":"[44]"},{"why":"Supplies the unified HORC formulation used as the geometric baseline and provides the Stex and Mus hypergraph collections for clustering experiments.","marker":"[45]"},{"why":"Defines hypergraph Forman-Ricci curvature, the closed-form combinatorial baseline that HLRC is designed to surpass in expressiveness.","marker":"[46]"},{"why":"Generates the stochastic-block-model hypergraphs on which the intra- versus inter-community curvature separation is established.","marker":"[47]"},{"why":"Provides the high-school proximity-contact data used in the community-detection experiment.","marker":"[48]"},{"why":"Provides the MADStat co-authorship hypergraph with journal and year labels used for venue and temporal analysis.","marker":"[51]"},{"why":"Supplies the categorical-edge-label clustering setting used to build the MAG-10 conference-labeled hypergraph.","marker":"[52]"},{"why":"Provides the Microsoft Academic Graph source from which the MAG-10 co-authorship hypergraph is constructed.","marker":"[53]"}],"fun_headline_variants":["Closed-form curvature for hypergraphs separates communities fast","Hypergraph curvature in closed form: fast community splitter","HLRC: closed-form Ricci curvature for hypergraphs, beats cost","Hypergraph lower Ricci curvature: cheap and community-aware"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The formula's specific normalization—the 1/2 coefficient, the -1 offset, and the max/min denominators—is chosen by hand rather than derived from a theory, and the authors themselves describe the theoretical foundation as largely heuristic; if that particular combination is arbitrary, the observed separations could be artifacts of the chosen formula rather than evidence of a genuine geometric invariant.","fun_headline_variants_meta":{"raw":{"variants":["Closed-form curvature for hypergraphs separates communities fast","Hypergraph curvature in closed form: fast community splitter","HLRC: closed-form Ricci curvature for hypergraphs, beats cost","Hypergraph lower Ricci curvature: cheap and community-aware"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000427,"raw_usage":{"total_tokens":2208,"prompt_tokens":993,"completion_tokens":1215,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":609,"completion_tokens_details":{"reasoning_tokens":1149}},"tokens_in":609,"tokens_out":1215,"duration_ms":9953,"temperature":1.0,"reasoning_tokens":1149,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:53:32.633932+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the three stochastic-block-model hypergraphs from Fig. 1 and apply hyperedge swaps that keep every node's degree and every hyperedge's size fixed while scrambling which nodes share hyperedges across communities. If the intra- versus inter-community HLRC separation persists with the same Wilcoxon significance, the signal is dominated by degree statistics rather than by the shared-neighbor geometry the formula claims to measure.","supporting_citations":[{"cited_title":"Ollivier Ricci curvature of directed hypergraphs.Scientific Reports, 10(1):12466, 2020","cited_arxiv_id":null,"evidence_quote":"Introduces Ollivier-Ricci curvature on directed hypergraphs, forming the theoretical lineage of the HORC baseline."},{"cited_title":"Ollivier-Ricci curvature for hypergraphs: A unified framework","cited_arxiv_id":null,"evidence_quote":"Supplies the unified HORC formulation used as the geometric baseline and provides the Stex and Mus hypergraph collections for clustering experiments."},{"cited_title":"Forman–Ricci curvature for hypergraphs","cited_arxiv_id":null,"evidence_quote":"Defines hypergraph Forman-Ricci curvature, the closed-form combinatorial baseline that HLRC is designed to surpass in expressiveness."},{"cited_title":"Consistency of spectral partitioning of uniform hypergraphs under planted partition model.Advances in Neural Information Processing Systems, 27, 2014","cited_arxiv_id":null,"evidence_quote":"Generates the stochastic-block-model hypergraphs on which the intra- versus inter-community curvature separation is established."},{"cited_title":"Contact patterns in a high school: A comparison between data collected using wearable sensors, contact diaries and friendship surveys.PLOS ONE, 10(9):e0136497, 2015","cited_arxiv_id":null,"evidence_quote":"Provides the high-school proximity-contact data used in the community-detection experiment."},{"cited_title":"Co-citation and co-authorship networks of statisticians.Journal of Business & Economic Statistics, 40(2):469–485, 2022","cited_arxiv_id":null,"evidence_quote":"Provides the MADStat co-authorship hypergraph with journal and year labels used for venue and temporal analysis."},{"cited_title":"Clustering in graphs and hypergraphs with categorical edge labels","cited_arxiv_id":null,"evidence_quote":"Supplies the categorical-edge-label clustering setting used to build the MAG-10 conference-labeled hypergraph."},{"cited_title":"An overview of microsoft academic service (MAS) and applications","cited_arxiv_id":null,"evidence_quote":"Provides the Microsoft Academic Graph source from which the MAG-10 co-authorship hypergraph is constructed."}],"review_version":1}