{"id":"f0f94a9a-e328-4cc8-991a-9e24190945e6","arxiv_id":"2506.02686","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A maximum-entropy generalization of the S^1 geometric graph model that enforces arbitrary inter-community edge counts while retaining the standard popularity-similarity form.","lead":"The paper introduces the Random Hyperbolic Block Model, a generative network model that lets users set in advance how often different groups of nodes connect while keeping the familiar geometric structure of random hyperbolic graphs. It then shows that a standard geometric embedding tool, D-Mercator, recovers node degrees and clustering from these networks but fails to recover the imposed group mixing patterns.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The finite-size identification ϕ_i=f_i and Φ_IJ=F_IJ is asserted but never verified in the experiments; if the generated graphs do not realize F_in, D-Mercator's reconstruction errors conflate generator bias with geometric failure.","rationale":"The reader's weakest assumption identifies exactly the same gap: the thermodynamic-limit identification is asserted but not checked in the finite-N experiments. The paper itself acknowledges the need for a numerical finite-size correction, and Section 3.1 omits any description of it. Since the experimental conclusion about D-Mercator's failure is the paper's main empirical evidence, this omission is the single most load-bearing concern. The maximum-entropy derivation in Section 2 is internally consistent given the stated limit, and the model is plausibly correct and useful; the issue is not the mathematical construction but whether the reported experiments realize the intended F_in. A direct measurement of the input-side mixing error would settle whether the negative conclusion is an artifact of the generator or a genuine property of geometric embeddings. The reader's CONDITIONAL verdict remains appropriate: the concern does not invalidate the model but requires additional evidence, so no verdict change is needed.","tokens_in":9764,"tokens_out":10999,"duration_ms":111598,"concrete_test":"For the default parameter set (N=3000, n=10, k=10, β=2, ρ=0.5, q=1), generate the RHBM exactly as described in Section 3.1 (ϕ_i=f_i, Φ_IJ=F_IJ), and compute the realized block mixing matrix F_real, either as the expected block counts under the model or as the empirical average over many latent-coordinate draws. Calculate ||F_real−F_in||_1 / ||F_in||_1. Repeat for N=1000 and for n=100. If the input-side relative error is comparable to the 10–50% errors reported in Figure 1, then the D-Mercator experiment does not support the claim; if it is below ~5%, the concern is resolved and the original interpretation stands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2 explicitly limits the equivalence between the maximum-entropy solution and Eq. (2) to the thermodynamic limit, stating that for finite N the parameters ϕ_i and Φ_IJ satisfying constraints (5) and (6) 'must be found numerically'. Section 3.1, however, describes the generation protocol without any mention or implementation of this correction, using N=1000–5000 and n up to 100. If the graphs are generated with ϕ_i=f_i and Φ_IJ=F_IJ directly, the realized block mixing matrix may deviate from F_in by a finite-size bias. The experimental metric in Figure 1 compares D-Mercator's reconstructed F_out to the nominal F_in, not to the actually realized mixing matrix. Unless the input-side deviation is shown to be negligible, the reported 10–55% reconstruction errors cannot be attributed to the geometric embedding's inability to capture mesoscale structure; they may simply reflect that the input graphs do not have the intended block counts. This is load-bearing because the paper's central negative conclusion—that purely geometric models fail on arbitrary mesoscale mixing—depends on the input graphs genuinely realizing the intended F_in.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the Random Hyperbolic Block Model (RHBM), a generalization of the S1/H2 geometric network model in which the probability of a link between nodes i in block I and j in block J is modulated by a block-level parameter Phi_IJ in addition to individual fitness phi_i. The central theoretical claim is that, in the thermodynamic limit, the maximum-entropy ensemble constrained by total energy, per-block link counts, and the degree sequence reproduces exactly the RHBM edge probability (Eq. (11)), with phi_i = f_i and Phi_IJ = F_IJ. The authors also present an alternative formulation in which the RHBM is a union of n-choose-2 S1 subgraphs sharing angular coordinates, and they report experiments using D-Mercator to embed RHBM graphs in S^D for D = 1,...,5. The experiments show that D-Mercator reconstructs degrees and clustering well but has 10-55% relative error in recovering the input block mixing matrix, which the authors interpret as evidence that purely geometric models cannot realize arbitrary mesoscale mixing patterns.","tokens_in":9976,"tokens_out":14357,"duration_ms":145533,"significance":"If the technical gaps in the experimental section are resolved, the RHBM is a useful and principled contribution: it provides a maximum-entropy null model that couples the popular S1 geometry with explicit control of the block mixing matrix, and the asymptotic derivation from constraints (4)-(6) to Eq. (11) is internally consistent, with Eqs. (8)-(10) satisfying the block-degree consistency check. The alternative subgraph representation of the model is also a helpful conceptual tool. The negative experimental result about D-Mercator is interesting, but as it stands it is not yet load-bearing evidence for the paper's broader claim that geometry cannot represent such mixing patterns, because the input graphs' realized mixing matrices and the inference heuristic's optimality are not validated.","major_comments":[{"comment":"The paper states that the identification phi_i = f_i and Phi_IJ = F_IJ holds exactly only in the thermodynamic limit and that finite systems require a numerical correction to satisfy (5) and (6), but Section 3.1 describes no such correction and the experiments use N = 1000-5000. The error metric in Figure 1 compares D-Mercator's F_out to the nominal F_in, not to the mixing matrix actually realized by the input graph. If the graphs are generated with phi_i = f_i and Phi_IJ = F_IJ directly, the realized block counts may deviate from F_in, in which case the reported reconstruction errors conflate input-side generation bias with the geometric embedding's inability. Please (i) report the realized block mixing matrix of the generated graphs, (ii) evaluate F_out against the realized matrix rather than, or in addition to, the nominal F_in, and (iii) either implement the finite-size numerical correction or quantify that the input-side deviation is negligible across the tested N, k, and n ranges.","section":"Section 2 (after Eq. (11)) and Section 3.1"},{"comment":"The formula for F as written is not a valid expected-link-count matrix for the parameters used. Interpreting the first term 'rho + 1/n I' as an all-ones matrix multiplied by rho plus (1/n) times the identity gives diagonal entries rho + 1/n and off-diagonal entries rho + (1-rho) q^{...}/(2 sum); for the reported rho = -0.5 and n = 10 all entries are negative, which is incompatible with F_IJ being expected block edge counts in Eq. (11). In addition, no normalization is described that ties F to the specified average degree k, so the total number of expected edges in the generated network is not determined by the written formula. Please provide an unambiguous, reproducible definition of F, including the intended all-ones matrix if that is the notation, the normalization to total edges, and the valid range of rho for each n.","section":"Section 3.1, parametric mixing matrix F"},{"comment":"The conclusion that 'purely geometric representations may not suffice' is stronger than what the experiments establish. The experiments show that D-Mercator, a heuristic approximate inference algorithm, does not recover F_in; they do not show that no S^D model with suitably chosen hidden degrees and angular coordinates can realize the input mixing pattern. The triangle-inequality argument in the Discussion is informal and is not quantified. Please either restrict the claim to the observed behavior of D-Mercator, or add evidence that the limitation is intrinsic to the S^D family, for example by solving the exact maximum-likelihood or moment-matching problem on small graphs and showing a lower bound on the achievable mixing error.","section":"Section 3.3 and Section 4 (conclusion)"}],"minor_comments":[{"comment":"The parenthetical 'or twice that number if I = J' is ambiguous: please define explicitly whether F_II denotes the expected number of unordered intra-block edges or the expected sum of degrees inside the block, since this affects the normalization in Eq. (6).","section":"Section 2, Eq. (5)"},{"comment":"The sentence explaining that constraints (5)-(6) imply (13) 'under the maximum entropy principle' because of the 'least restrictive assumption' is obscure; the implication is actually a derived property of the explicit solution (11) in the N -> infinity limit, and the phrase about the least restrictive assumption should be removed or reformulated.","section":"Section 2, after Eq. (13)"},{"comment":"The experimental section does not state how many independent graph samples were generated for each parameter configuration or how error bars in Figures 1, 3, and 4 were computed; please specify the number of trials and the dispersion of the reported errors.","section":"Section 3.1"},{"comment":"Please define precisely how F_out is computed from the D-Mercator embedding, including whether the edge probabilities are integrated over the inferred angular coordinates and how the total number of edges is normalized before computing the relative error; otherwise the reported error may depend on an arbitrary rescaling between F_in and F_out.","section":"Section 3.2"},{"comment":"The phrase 'integrating p_ij over theta_j and lambda_j' mixes the uniform angular variable with the Lagrange multiplier lambda_j; please clarify the grand-canonical or hidden-variable interpretation so that the reader understands over which random variables the averages are taken.","section":"Section 2, Eq. (8)"},{"comment":"There are occasional grammatical and typographical issues, for example 'the dissimilarity between groups do not obey' in the abstract and the inconsistent spacing in 'S D model'; a careful copy edit would improve readability.","section":"Abstract and throughout"}],"recommendation":"major_revision","confidential_remarks":"The central maximum-entropy derivation is sound and the model is a worthwhile contribution to geometric network modeling. My main concern is that the experimental validation, which carries the paper's negative claim about purely geometric models, is not yet reproducible or internally calibrated: the finite-size identification is unchecked, the parametric F matrix appears to be mis-specified for the range of rho used, and the heuristic nature of D-Mercator is not accounted for in the conclusions. These are fixable with additional experiments and rewording, so I do not view the paper as beyond repair. The self-citation [15] is relevant to the alternative formulation and does not appear problematic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is worth your time if you care about generative models that combine latent geometry with explicit block mixing. The key new result is a maximum-entropy derivation: constraints on total energy, block link counts, and degree sequence produce exactly the S^1-style edge probability with block-dependent normalization. That derivation is clean, and the union-of-S^1 representation is a useful reformulation. The empirical case that D-Mercator preserves degrees and clustering but reconstructs mixing matrices poorly is also interesting and non-obvious, even if the error magnitudes are not precise.\n\nThe soft spots are the ones the reader flagged, and I agree with the stress-test. The paper explicitly states that phi_i=f_i and Phi_IJ=F_IJ hold only in the thermodynamic limit and that finite systems require numerical correction, but Section 3.1 does not describe or implement that correction. Experiments use N=1000 to 5000. If the actual realized mixing matrix deviates from F_in, the reported 10-55% error conflates generator bias with geometric failure. This is not fatal, but it needs to be addressed with either a verified correction or a check that finite-size deviations are negligible at these sizes.\n\nSecond, the relationship to the authors' own Hidden-Degree Geometric Block Model [15] is understated. The edge probability in Eq. (2) looks like the same model, and the paper should be explicit about what is new: the maximum-entropy derivation and the union-of-S^1 representation, not the model itself.\n\nThird, the experiments have no error bars or repetitions, and the conclusion that purely geometric models fail on arbitrary mixing is stronger than the evidence. The qualitative pattern is probably right, but the paper should temper the claim or add repetitions and a direct comparison of realized vs. nominal F.\n\nThe citation pattern is fine; self-citation to [15] is legitimate. The math in the thermodynamic limit is consistent. The discussion is honest about the Euclidean formulation.\n\nVerdict: conditional. The model is a genuine contribution and the experimental question is relevant. Deserves peer review, with major revisions to verify finite-size behavior, clarify novelty relative to [15], and bolster the experiments.\n\nRecommendation: send to review.","headline":"A clean max-entropy block extension of S^1 with a real but fixable finite-size gap in its experimental validation.","tokens_in":10525,"tokens_out":4758,"would_cite":true,"duration_ms":45848,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["05C80","68R10"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper introduces the Random Hyperbolic Block Model, an $S^1$-based generator with explicit constraints on group-to-group mixing, and shows it is the maximum-entropy ensemble realizing arbitrary block mixing in the thermodynamic limit.","keywords":["random hyperbolic graphs","block model","maximum entropy","community structure","mixing matrix","network embedding","D-Mercator","S1 model"],"falsifier":"Generate RHBM graphs at $N=1000,3000,5000$, measure the empirical block mixing matrix $F_{\\mathrm{emp}}$ directly from the adjacency matrix, and compare it with the target $F_{\\mathrm{in}}$ before any embedding. If the relative error is already in the 10–55 percent range, the reported embedding errors conflate generator inaccuracy with geometric inability; if generation errors are near zero and embedding errors persist, the paper's conclusion is supported.","tokens_in":9489,"feed_emoji":"🌐","tokens_out":13005,"duration_ms":106158,"temperature":0.7,"pith_summary":"The paper sets out to show that the popularity-similarity geometry behind random hyperbolic graphs cannot realize arbitrary community mixing patterns, and that a small extension of the model can. Its Random Hyperbolic Block Model (RHBM) keeps the $S^1$ edge-probability form — the $S^1$ model being a random geometric graph in which connection probability falls off with angular distance and grows with hidden degrees — but replaces hidden degrees with block-normalized fitnesses and adds a symmetric matrix of latent forces between blocks, giving each community pair its own connection propensity. The RHBM is derived as the maximum-entropy ensemble constrained by total energy, expected link counts per block pair, and expected degree sequence; in the limit of infinitely many nodes that ensemble reproduces the model exactly, making RHBM the least-biased geometric generator for prescribed block mixing. Experiments embed RHBM graphs back into hyperbolic spaces and find that degrees and clustering come back almost perfectly while the block mixing matrix is off by 10–55 percent, supporting the claim that attribute-driven group mixing is essentially non-geometric.","feed_headline":"Hyperbolic graphs get explicit control over community mixing","feed_subtitle":"Even the best hyperbolic embedding reproduces degrees and clustering but misses the block mixing patterns.","key_machinery":"The load-bearing object is the block-level hidden force $\\Phi_{IJ}$ acting together with the block-normalized fitness $\\phi_i$: they replace the single hidden degree of $S^1$ and make the edge probability block-dependent through the product $\\phi_i\\phi_j\\Phi_{IJ}$. The maximum-entropy derivation is the mechanism that certifies the model: constraining total energy, per-block-pair edge counts, and the degree sequence yields, in the thermodynamic limit, exactly the model's edge probabilities, making the RHBM the least-biased geometric ensemble for a target mixing matrix. The equivalent decomposition into $\\binom{n}{2}$ independent mono/bipartite $S^1$ subgraphs is what makes the model constructive: each block pair is its own $S^1$ graph, while drawing each node's angular coordinate once preserves transitivity and clustering across communities.","core_discovery":"On the paper's own terms, the discovery is that a random geometric graph can be given arbitrary mesoscale control without leaving the $S^1$ framework. The Random Hyperbolic Block Model (RHBM) assigns each node a hidden fitness $\\phi_i$ and each block pair $(I,J)$ a hidden force $\\Phi_{IJ}$, so the edge probability takes the same functional form as in $S^1$, with block-wise hidden degrees replacing the global hidden degree and normalization. The paper derives this model as a maximum-entropy ensemble constrained by expected total energy, expected edge counts $F_{IJ}$ between every block pair, and expected degree sequence; in the limit $N\\to\\infty$ that ensemble has exactly the RHBM edge probabilities, with $\\phi_i=f_i$ and $\\Phi_{IJ}=F_{IJ}$. It also shows the RHBM is the union of $\\binom{n}{2}$ independent mono/bipartite $S^1$ graphs, one per block pair, provided each node draws its fitness and angular coordinate once and reuses them across subgraphs. The experiments embed RHBM graphs back into $S^D$; degrees and clustering are recovered well, while the block mixing matrix is not, with relative errors from about 10 percent to 55 percent across $N$, $k$, $n$, $\\rho$, $q$, $\\beta$, and $D=1,\\ldots,5$.","pith_inferences":["Inference: Re-running the embedding experiments after applying the finite-size correction to $\\phi_i$ and $\\Phi_{IJ}$ would separate generator error from geometry error; if the 10–55 percent errors survive, the geometric-impossibility conclusion is robust.","Inference: Because block pairs are generated as independent subgraphs that share angular coordinates, the RHBM design maps directly onto scalable community-aware null models for empirical attributed networks, an application the authors list only as future work.","Inference: The trend of lower mixing-matrix error at higher embedding dimensions suggests that raising $D$ postpones but does not eliminate the geometric obstruction; testing $D$ well beyond 5 could locate the dimension at which the effect saturates.","Inference: RHBM graphs can serve as a benchmark for community-aware embedding algorithms: if some algorithm recovers the mixing matrix, the failure is specific to geometry-only embedding rather than to latent-space inference as a whole."],"forward_implications":["The RHBM can generate sparse, small-world, clustered networks with heterogeneous degrees and arbitrary symmetric block mixing matrices, including patterns that violate the triangle inequality.","For fixed expected degree sequence and block mixing, the RHBM is the least-biased ensemble within the popularity-similarity framework in the thermodynamic limit.","Because the model decomposes into independent mono/bipartite $S^1$ subgraphs with shared latent coordinates, it can be generated by reusing existing $S^1$ routines block by block.","Hyperbolic embeddings of RHBM graphs recover degrees and clustering but miss the mixing matrix, indicating that purely geometric representations fail to capture this mesoscale structure.","The RHBM provides a way to separate attribute-driven mixing from popularity-similarity geometry, which makes it a suitable null model for testing whether observed community structure has a geometric origin."],"supporting_citations":[{"why":"supplies the maximum-entropy derivation of the $S^1$ model that the RHBM extends to block-level constraints","marker":"[5]"},{"why":"introduces the stronger per-block degree constraints that yield the $(S^1)^{\\binom{n}{2}}$ equivalent formulation","marker":"[15]"},{"why":"provides the D-Mercator embedding method used to test whether geometry recovers mixing patterns","marker":"[16]"},{"why":"establishes the $S^1/H^2$ isomorphism and the thermodynamic-limit mapping of hidden degrees that justifies calling the model hyperbolic","marker":"[20]"},{"why":"defines the popularity-similarity optimization framework whose limitations motivate explicit mixing control","marker":"[23]"},{"why":"shows the angular-aggregation approach to communities, which the paper argues cannot realize arbitrary mixing because of the triangle inequality","marker":"[4]"}],"fun_headline_variants":["RHBM: hyperbolic graphs with arbitrary community control","Random Hyperbolic Block Model adds mesoscale control","Hyperbolic random graphs now support arbitrary block mixing","RHBM gives hyperbolic graphs explicit control over community structure","Arbitrary mesoscale control in hyperbolic graph models via RHBM"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The exact match between the RHBM's target mixing matrix and the maximum-entropy model holds only as the number of nodes goes to infinity; the experiments use 1,000 to 5,000 nodes and the finite-size correction the derivation calls for is never described, so the generated graphs may not have precisely the intended block mixing.","fun_headline_variants_meta":{"raw":{"variants":["RHBM: hyperbolic graphs with arbitrary community control","Random Hyperbolic Block Model adds mesoscale control","Hyperbolic random graphs now support arbitrary block mixing","RHBM gives hyperbolic graphs explicit control over community structure","Arbitrary mesoscale control in hyperbolic graph models via RHBM"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000678,"raw_usage":{"total_tokens":3107,"prompt_tokens":994,"completion_tokens":2113,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":2037}},"tokens_in":610,"tokens_out":2113,"duration_ms":14593,"temperature":1.0,"reasoning_tokens":2037,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:19:53.695242+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate RHBM graphs at $N=1000,3000,5000$, measure the empirical block mixing matrix $F_{\\mathrm{emp}}$ directly from the adjacency matrix, and compare it with the target $F_{\\mathrm{in}}$ before any embedding. If the relative error is already in the 10–55 percent range, the reported embedding errors conflate generator inaccuracy with geometric inability; if generation errors are near zero and embedding errors persist, the paper's conclusion is supported.","supporting_citations":[{"cited_title":"The Hidden- Degree Geometric Block Model","cited_arxiv_id":null,"evidence_quote":"introduces the stronger per-block degree constraints that yield the $(S^1)^{\\binom{n}{2}}$ equivalent formulation"},{"cited_title":"Popularity versus similarity in growing networks","cited_arxiv_id":null,"evidence_quote":"defines the popularity-similarity optimization framework whose limitations motivate explicit mixing control"},{"cited_title":"Sus- taining the Internet with hyperbolic mapping","cited_arxiv_id":null,"evidence_quote":"shows the angular-aggregation approach to communities, which the paper argues cannot realize arbitrary mixing because of the triangle inequality"}],"review_version":1}