{"id":"1b3a6e22-c625-421f-baeb-357404c790ac","arxiv_id":"2508.20706","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Scale-invariant graph embeddings whose parameters add under node merging give consistent multi-scale network reconstructions, with accuracy that depends on the metric and model family.","lead":"This paper tests whether graph-embedding models that keep the same mathematical form across resolutions can reconstruct hidden economic and trade networks from aggregate data. It finds that multi-scale (scale-invariant) models give consistent embeddings when nodes are merged, while single-scale maximum-entropy models do not, though accuracy gains are mixed.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Additive renormalization is a model-internal identity; the paper does not establish that real networks obey the MSM's independent-edge/OR form, so the claim of consistent embeddings rests on an imposed assumption.","rationale":"The reader's weakest-assumption analysis correctly identifies the same load-bearing issue: the renormalizability of the MSM is a theorem about the model class, not an empirical property of the observed networks. The paper's own evaluation is honest about mixed results and lacks uncertainty quantification, which supports a CONDITIONAL verdict. I do not see a reason to move to REJECT or ACCEPT: the theoretical derivation is sound as far as it goes, and the paper does not overclaim beyond its model framework except in a few interpretive sentences. The most useful next step is a synthetic-data test that isolates whether the summed-parameter rule remains useful when the true generative process is not the MSM; if it does, the practical claim is stronger; if not, the limitation should be made explicit. Since the reader already flagged this as the weakest assumption and the verdict is CONDITIONAL, no change in verdict is warranted.","tokens_in":32410,"tokens_out":2852,"duration_ms":36770,"concrete_test":"Generate synthetic networks from a non-MSM process (e.g., a stochastic block model or a configuration model with degree sequence and edge correlations matching the WTW) at the microscopic level. Coarse-grain these networks via logical OR. Fit the MSM at level 0, compute summed block-level probabilities, and compare their predictive log-likelihood / reconstruction accuracy against a model refitted at each coarse level. If summed MSM predictions degrade substantially relative to refitted models on these misspecified data, the cross-scale consistency claim is conditional on the MSM's generative form being correct, and the paper should state that limitation explicitly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mathematical result—that block-level parameters are sums of micro parameters and that summed probabilities equal coarse-grained probabilities—is derived in Appendix B under two strong assumptions: (1) the microscopic network is generated by the independent-edge MSM with p_ij = 1 - e^{-x_i x_j}, and (2) aggregation is logical OR (Eq. B1). Equations B5-B8 then show an algebraic identity: if the micro model is true, the coarse-grained block distribution is exactly the MSM with summed parameters. This is a property of the model, not a property of the data. The paper's stronger language—e.g., MSM 'aligning with the hidden generative process' in the Conclusions—goes beyond what is demonstrated. Figure 2 only verifies that the fitted micro MSM satisfies its own identity; it does not validate that empirical coarse-grained networks are MSM-distributed. The empirical comparisons (Figs. 5-7) show reasonable but mixed results: fitnCM beats fitnMSM on several metrics, AUC-ROC/PR are often indistinguishable between single-scale and multi-scale models, and no error bars or repeated-sampling uncertainty are provided. Thus the practical claim that a model fitted once at the finest scale predicts coarser-scale structure correctly is contingent on the MSM being the true data-generating process, which is imposed rather than tested.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops graph-embedding methods for network reconstruction in a multi-resolution setting. It contrasts maximum-entropy single-scale models (CM, fitnCM) with scale-invariant multi-scale models (degcMSM, maxlMSM, fitnMSM), where block-level parameters are defined as sums of microscopic parameters and link probabilities take the MSM form p_ij = 1 - exp(-x_i x_j). The central mathematical claim is that this form is renormalizable under arbitrary node aggregation: the coarse-grained probability is exactly the MSM probability with summed parameters. The paper tests the methods on an ING input-output network at several NAICS aggregation levels and on the WTW, reporting scale-invariance plots, parameter agreement, signed relative errors of link counts, reconstruction accuracies, AUC-ROC/PR curves, and triangle-density comparisons. The claimed payoffs are that a model fitted once at the finest resolution predicts coarser scales consistently, and that degcMSM improves over CM even at the fitted level.","tokens_in":32794,"tokens_out":7070,"duration_ms":82159,"significance":"The additive renormalization rule for the MSM is mathematically correct and potentially useful: it gives an interpretable, parameter-free transformation of node embeddings under aggregation. The paper is also transparent about the model assumptions and releases code. However, the main empirical support for the practical claim is currently weak: the headline scale-invariance check is an identity by construction, and the comparative results are point estimates without uncertainty quantification. If the model were tested against non-MSM generative processes and with proper significance analysis, the contribution would be solid. As it stands, the theoretical part is sound but the empirical validation of the central claim is incomplete.","major_comments":[{"comment":"The scale-invariance evidence in Fig. 2 is a model-internal identity, not an empirical test. Eq. (B8) is derived by assuming independent edges p=1-exp(-x_i x_j), OR aggregation, and summed parameters; it holds by construction for any fitted MSM parameters. The text itself says 'By construction' in Section V A, and Section V B 1 restricts 'true' scale-invariance to Eq. B8. Yet the Conclusion (Section VI) claims MSMs 'align with the hidden generative process.' This overstates what is demonstrated. Please add a falsifiable test, e.g. apply the same identity to synthetic networks generated by a non-MSM model, compare level-l likelihoods of summed MSM versus refitted alternatives, or show how deviations from the independent-edge/OR assumption would be detected. Without this, the central claim that a model fitted once predicts coarser scales rests on an imposed assumption.","section":"Section V A, Appendix B (Eqs. B5–B8), Fig. 2"},{"comment":"The quantitative comparisons that support the paper's empirical claims are reported without error bars or significance tests. For instance, Fig. 5 reports signed relative errors as single curves; Fig. 7a reports reconstruction accuracies without dispersion; Fig. 7b reports AUC curves without bootstrap intervals. Section V G explicitly acknowledges that 'it remains difficult to differentiate single-scale from multi-scale models, as their AUC performances are often coinciding,' and that fitnCM consistently outperforms fitnMSM. Given that each dataset is a single realization, please provide uncertainty quantification (e.g., bootstrap over nodes/edges or over repeated random partitions) and state which differences are statistically significant. The current wording (e.g., 'degcMSM delivers the best performance at the fitted level') is stronger than the evidence supports.","section":"Section V F–G, Figs. 5, 7a, 7b"},{"comment":"The paper motivates the problem as reconstructing an unobserved microscopic network from aggregate information (e.g., firm-level links hidden behind sector-level data). However, all experiments fit the models at the finest available level (level 0) and then aggregate upward; there is no experiment where the micro network is withheld and inferred from coarse data. Thus the privacy-oriented reconstruction direction promised in the abstract is not actually evaluated. A proof-of-concept in the reverse direction—or a revision of the framing—is needed, because the practical relevance of the additive sum rule to the stated problem depends on that direction.","section":"Abstract / Introduction vs. Section V"}],"minor_comments":[{"comment":"The main text states that at the optimum of maxlMSM, nodes with identical neighbors have the same parameters, but Appendix B 2 b says the stationarity condition does not guarantee this and that statistical equivalence is imposed by hand. Please reconcile these statements.","section":"Section III C 2 vs. Appendix B 2 b"},{"comment":"The claimed stationary solution for w_i is not a stationary point of the likelihood: for a_ii=0 the derivative is identically -1, and for a_ii=1 the derivative 1/p_ii-1 vanishes only in the limit w_i→∞. These are boundary/limit solutions, not solutions of ∂_{w_i} L=0. Clarify the formulation.","section":"Appendix B 2 a, Eqs. (B10)–(B12)"},{"comment":"The caption says 'for the fitted models (level 0) and summed ones (level 2)' but the axes show all levels 0–3; update the wording. Also fix the typo 'Novertheless' in Section V G.","section":"Section V G, Fig. 7 caption"},{"comment":"The statement that 'true deviation from scale-invariance should be only assessed using Eq. (B8)' is circular given that Eq. (B8) is an identity of the model. The sentence should be rewritten to acknowledge that Eq. (B8) tests internal consistency, not empirical validity.","section":"Section V B 1"}],"recommendation":"major_revision","confidential_remarks":"The theoretical derivation of the MSM renormalization rule is sound, but the empirical section needs substantial additional work before the paper can be accepted. The scale-invariance plot is presented as evidence even though the text acknowledges it is by construction, and the performance comparisons lack uncertainty quantification. These issues are fixable and the manuscript's scope is suitable for the journal, but the current version overclaims on both the empirical and practical fronts."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's new pieces are the degree-corrected and maximum-likelihood MSMs (degcMSM, maxlMSM), the renormalized single-scale variants, and the two empirical applications on the ING bank data and the World Trade Web. The core mathematics is in good shape: the additive renormalization rule follows directly from the model's independent-edge, logical-OR structure, and the proof that the Configuration Model is not scale-invariant (Appendix A1) is clean and correct. The authors also deserve credit for releasing code, using publicly available WTW data, and reporting mixed results without hiding them—e.g., fitnCM beats fitnMSM on AUC/PR, which is not what a partisan author would highlight.\n\nThe soft spots are real but not fatal. Figure 2 is presented as evidence for scale invariance, yet it is a tautology: summed probabilities equal coarse-grained probabilities because the model was constructed that way. The paper actually acknowledges this in Section V B 1, which is good, but the Conclusions still say the MSM is \"aligning with the hidden generative process.\" That is overreach. The framework assumes the true micro network is MSM-distributed and that aggregation is logical OR; this is an imposed modeling assumption, and the paper never tests it against the data. If real firm or trade networks violate that form, the nice additive rule is just an algebraic identity, not a discovered property of the system.\n\nThe empirical comparisons also need more caution. The claim that degcMSM outperforms CM at the fitted level rests on one confidential dataset with no error bars or significance tests. The RA metric has dispersion intervals, but the AUC and PR comparisons are point estimates with no uncertainty. On the WTW, the fitness ansatz is weak by the paper's own analysis, so the fitness-model results are less informative. These are fixable issues: report bootstrap or repeated-sampling uncertainty, and soften the language about the hidden generative process.\n\nWho is this for? People doing network reconstruction, especially for systemic-risk applications, and anyone working on multi-scale or renormalizable graph models. The theoretical contribution is worth their time, and the benchmark comparison, despite the caveats, is a useful data point. I would send this to peer review—it deserves a serious referee—but with clear requests: quantify uncertainty, test the model assumptions on data where possible, and separate the built-in consistency of the MSM from empirical validation.\n\nIn short: the math is sound, the paper is honest, and the new model variants are useful, but the rhetoric about alignment with the true generative process should be tempered.","headline":"The additive renormalization is mathematically sound and the comparison is honest, but the paper sells a model-internal identity as empirical support, so the strongest claims need to be dialed back.","tokens_in":33230,"tokens_out":2054,"would_cite":true,"duration_ms":25583,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["89.65.Gh","89.75.-k"],"model":"deepseek-v4-flash","headline":"Scale-invariant multi-scale models yield graph embeddings that stay consistent under node aggregation, so one fit at the finest scale predicts links at every coarser scale, while maximum-entropy models must be refitted each level.","keywords":["multi-scale network reconstruction","graph embeddings","scale invariance","network renormalization","configuration model","maximum entropy","input-output networks","world trade web"],"falsifier":"Fit the MSM at the finest level of a network with strong community structure or correlated edges, sum the parameters to a coarser level, and compare predicted block-edge probabilities with the observed coarsened network. If the empirical fraction of linked block pairs falls systematically outside the model's Bernoulli prediction intervals, the scale-invariance is an algebraic identity of the model rather than a property of the data. The paper's own finding that fitnCM beats fitnMSM where the fitness proxy is weak offers the complementary check: a dataset where the refitted single-scale model b","tokens_in":32313,"feed_emoji":"🕸️","tokens_out":12142,"duration_ms":108833,"temperature":0.7,"pith_summary":"This paper asks how to reconstruct a hidden network from partial or aggregate data when the resolution of the data changes because nodes are grouped into blocks, as happens when firms are aggregated into sectors or countries into trade regions. It argues that the standard maximum-entropy reconstruction models—the configuration model (CM) and its fitness-based variant (fitnCM)—are single-scale: their parameters are only meaningful at the level where they were fitted and must be re-estimated at every other level. The proposed alternative, the multi-scale model (MSM) family derived from the principle of scale invariance, keeps the same functional form at every resolution and renormalizes block parameters by simply summing the microscopic ones. If the paper is right, a model fitted once at the finest observable scale predicts connection probabilities at all coarser scales without refitting, and node embedding vectors transform additively under aggregation. On a confidential national input-output network and on the World Trade Web, the degree-corrected MSM matches or beats the configuration model, including at the level where both models are fitted, while the fitness-based MSM reproduces the total number of links at every scale.","feed_headline":"One fit, every scale: summed embeddings renormalize","feed_subtitle":"Coarse-level link probabilities come from adding up node vectors, so networks never need re-fitting.","key_machinery":"The load-bearing identity is the scale-invariance condition of Equation 19: the renormalized (summed) probability of a block-edge must equal the coarse-grained probability that at least one microscopic edge is present, with both expressed in the same functional form. The unique functional form compatible with this constraint under the independent-edge assumption is the MSM link probability p_ij = 1 − e^{−x_i x_j} (plus a self-loop term), which in turn yields the additive renormalization rule x_I = Σ_{i∈I} x_i, w_I = Σ_{i∈I} w_i (Equation 20). This sum rule is the mechanism that carries the argument: it makes node parameters additive under aggregation, and it guarantees that a single fit at l","core_discovery":"The central claim is that the multi-scale model (MSM), with connection probability p_ij = 1 − e^{−x_i x_j} and a matching self-loop term, is the functional form satisfying scale invariance under logical-OR node aggregation: the probability that two blocks are linked equals the probability that at least one edge exists between their members, in the same functional form at every level. Consequently block parameters renormalize by summation, x_I = Σ_{i∈I} x_i and w_I = Σ_{i∈I} w_i, so the embeddings are additive under coarse-graining and never re-learned. The configuration model p_ij = x_i x_j/(1 + x_i x_j) is not renormalizable: coarse-graining its probabilities yields a function outside the C","pith_inferences":["The consistency is algebraic once the independent-edge, OR-aggregation generating process is assumed; whether real supply and trade networks obey it is an empirical question the paper does not settle. A direct test would compare the distribution of edge counts between blocks in observed data with the Bernoulli prediction.","The additive sum rule suggests a screening principle: any latent-space model whose link function satisfies the coarse-graining identity inherits renormalizability, so other functional forms could be tested for this property rather than derived from first principles.","The paper's own result that fitnCM beats fitnMSM wherever the fitness proxy is weak implies a practical hybrid: use degree constraints where available and additively consistent fitness constraints where privacy forbids them, since the MSM's sum rule keeps mixed constraints consistent across scales.","Because the paper's interlude (Section V B 1) warns that metric-level agreement between summed and refitted models should not be read as scale invariance, the practical advantage of MSMs should be evaluated with the identity of Equation B8, not just with reconstruction scores."],"forward_implications":["A model fitted once at the finest observable scale applies to any coarser representation, so no refitting is needed when the data resolution changes.","Node embedding vectors become additive under aggregation, giving a principled interpretation to vector sums in latent space and connecting embeddings across resolutions.","The degree-corrected MSM outperforms the configuration model at the fitted level 0 and keeps high reconstruction accuracy for degrees, nearest-neighbour degree, and clustering at coarser levels.","The fitness-based MSM needs only additively aggregating exogenous variables (strengths or GDP) and preserves the total number of links at every scale, which matters when degrees are confidential.","Renormalizing by summation instead of coarse-graining microscopic probabilities cuts the computational cost by roughly (N_ℓ/N_0)² at level ℓ."],"supporting_citations":[{"why":"Supplies the MSM functional form and the scale-invariance derivation (summed equals coarse-grained probability) that the paper adapts to reconstruction.","marker":"[18]"},{"why":"Introduces multi-scale node embeddings for the fully observable case and the additive vector-sum result that this paper extends to hidden networks.","marker":"[13]"},{"why":"Defines the configuration model and its connection probability p = xy/(1+xy), the main single-scale benchmark.","marker":"[28]"},{"why":"Introduces the fitness ansatz and GDP as fitness for the World Trade Web.","marker":"[29]"},{"why":"Supplies strength as fitness for firm-level economic networks and the fitnCM reconstruction baseline.","marker":"[30]"},{"why":"Provides the reconstruction-accuracy metric (fraction of observed statistics inside dispersion intervals) used for the main comparisons.","marker":"[42]"},{"why":"Provides the World Trade Web dataset used for the geographic multi-scale application.","marker":"[22]"},{"why":"The single-scale logistic PCA embedding benchmark that the degree-corrected MSM is reported to beat at the fitted level.","marker":"[12]"}],"fun_headline_variants":["Sum nodes, keep links: scale-free embeddings","Add once, reuse at every scale: renormalizable embeddings","One embedding fits all resolutions: just sum","No refitting needed: embeddings add up across scales","Scale-invariant graphs: sum node vectors to rebuild"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The whole construction assumes the real network is generated by independent edges combined with logical-OR aggregation—a block is linked if at least one of its members is linked—so that the model's additive summation rule mirrors the true coarse-graining; if real edges are correlated or aggregation works differently, the 'consistent embeddings' are a property of the model, not of the data.","fun_headline_variants_meta":{"raw":{"variants":["Sum nodes, keep links: scale-free embeddings","Add once, reuse at every scale: renormalizable embeddings","One embedding fits all resolutions: just sum","No refitting needed: embeddings add up across scales","Scale-invariant graphs: sum node vectors to rebuild"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000546,"raw_usage":{"total_tokens":2461,"prompt_tokens":774,"completion_tokens":1687,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":1612}},"tokens_in":518,"tokens_out":1687,"duration_ms":12120,"temperature":1.0,"reasoning_tokens":1612,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T14:53:25.894514+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fit the MSM at the finest level of a network with strong community structure or correlated edges, sum the parameters to a coarser level, and compare predicted block-edge probabilities with the observed coarsened network. If the empirical fraction of linked block pairs falls systematically outside the model's Bernoulli prediction intervals, the scale-invariance is an algebraic identity of the model rather than a property of the data. The paper's own finding that fitnCM beats fitnMSM where the fitness proxy is weak offers the complementary check: a dataset where the refitted single-scale model b","supporting_citations":[{"cited_title":"Chanpuriya, C","cited_arxiv_id":null,"evidence_quote":"Supplies the MSM functional form and the scale-invariance derivation (summed equals coarse-grained probability) that the paper adapts to reconstruction."},{"cited_title":"Squartini, G","cited_arxiv_id":null,"evidence_quote":"Introduces multi-scale node embeddings for the fully observable case and the additive vector-sum result that this paper extends to hidden networks."},{"cited_title":"Multi-scale reconstruction of large supply networks","cited_arxiv_id":"2412.16122","evidence_quote":"Defines the configuration model and its connection probability p = xy/(1+xy), the main single-scale benchmark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the fitness ansatz and GDP as fitness for the World Trade Web."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies strength as fitness for firm-level economic networks and the fitnCM reconstruction baseline."},{"cited_title":"Baptista, R","cited_arxiv_id":null,"evidence_quote":"Provides the reconstruction-accuracy metric (fraction of observed statistics inside dispersion intervals) used for the main comparisons."},{"cited_title":"Villegas, T","cited_arxiv_id":null,"evidence_quote":"Provides the World Trade Web dataset used for the geographic multi-scale application."},{"cited_title":"Fagiolo, J","cited_arxiv_id":null,"evidence_quote":"The single-scale logistic PCA embedding benchmark that the degree-corrected MSM is reported to beat at the fitted level."}],"review_version":1}