{"id":"6df651f9-aaf8-469c-8f85-514099cfd59c","arxiv_id":"1908.09470","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"The ERGM parameter region stabilizing a target graph is a convex cone whose faces are change-scores, and this cone can be used to predict vulnerable dyads.","lead":"The paper shows that the set of ERGM parameters that make a given graph more likely than a set of alternatives is the interior of a convex cone defined by change-scores. It gives an algorithm to compute this stable cone and uses it to identify which network ties are most likely to change.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Core cone characterization is sound, but Algorithm 1's closure test and initialization are unproven and can fail on degenerate M (empty cone, all-parallel rows, lineality); the computational claim is load-bearing and unsupported.","rationale":"The reader's conditional verdict is appropriate. The central mathematical claim—that local stability is equivalent to Mθ < 0 and that the set of stabilizing parameters is the interior of a convex cone—is sound and is proved directly from the ERGM likelihood ratio. The convex-cone proof is valid, and the change-score derivation is correct. The weaker point is the computational algorithm: Algorithm 1's termination condition and initialization are not proven, and degenerate inputs (single unique row, empty cone, lineality) are not addressed. This is exactly the same load-bearing concern identified by the reader. Because the core theorem does not depend on the algorithm, the paper should not be rejected; because the algorithm is presented as a contribution and is needed for the applications, a conditional verdict requiring a correctness proof or explicit edge-case handling is appropriate. The reader's verdict of CONDITIONAL is therefore left unchanged. No additional concerns (e.g., lack of error bars, the malformed citation) rise to the level of altering the verdict.","tokens_in":16335,"tokens_out":9303,"duration_ms":109168,"concrete_test":"Run Algorithm 1 exactly as specified on a minimal degenerate input: M = [[1,0]] for a one-dimensional parameter space. The correct stable region is the open halfspace θ1 < 0, whose double description consists of one non-redundant halfspace and one extreme ray. The algorithm's initialization loop (lines 3–6) searches for a second non-parallel row and will never terminate. If the authors intend a different initialization for this case, specify it and rerun; also test M = [[1,0],[-1,0]] (empty stable region) to verify that the algorithm terminates with a correct empty-cone description rather than looping until M is exhausted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The geometric characterization of the stable region in Section 2.1.1 (interior of a convex cone defined by Mθ < 0) is correct and follows directly from Equations 3–4. The load-bearing weakness is the computational method, specifically Algorithm 1 (Section 2.3). Its initialization loop (lines 3–6) requires two non-parallel rows of M; if M has only one unique row, or all rows are parallel, the loop never terminates even though a well-defined stable cone exists (e.g., M = [[1,0]] gives the open halfspace θ1 < 0). More generally, the closure test in lines 15–18—'if the convex hull is closed, terminate'—is asserted without proof or edge-case analysis. For degenerate cases such as an empty stable cone, a cone with a non-trivial lineality space, or redundant constraints, the test may either never succeed (so the algorithm exhausts M and returns a malformed H/V description) or may return prematurely before all non-redundant constraints are found. The paper's claimed 'practical algorithm for efficiently finding the stabilizing region' therefore rests on an unverified and possibly incorrect procedure. The central theorem does not depend on this algorithm, but the paper's applicability and the reported computations (e.g., the Lazega analysis) do.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper defines local stability of a target graph relative to a set of alternative graphs under an exponential family random graph model (ERGM). It shows that the stabilizing region of the parameter space, Φ = {θ : Mθ < 0}, where M is the matrix of change scores t(G')−t(G), is the interior of a convex polyhedral cone through the origin. The authors propose two algorithms (Algorithm 1 and Algorithm 2) based on the double description method to compute a non-redundant H- and V-representation of this stable cone, and they demonstrate the method on a star graph and on the Lazega law-firm network, including a dyad-vulnerability analysis.","tokens_in":16513,"tokens_out":2244,"duration_ms":22869,"significance":"If the computational claims are made rigorous, the paper offers a useful and elegant tool: the cone characterization is a direct, parameter-free consequence of the ERGM likelihood ratio, and the local-stability notion gives a fast proxy for dynamic stability under Metropolis dynamics. The paper's strengths include the explicit derivation in Section 2.1.1, the transparent one-step transition probability formulas in Section 3.2, and the reproducible simulation checks in Sections 3 and 4. However, the central computational contribution is not fully established, which limits the significance of the practical method and the reported applications.","major_comments":[{"comment":"The initialization loop requires sampling two non-parallel rows of M, but for degenerate M with only one distinct row direction (e.g., M = [1 0] in K=2) the loop never terminates even though the stable cone is well-defined as the open halfspace θ1 < 0. This is a concrete failure mode that should be handled explicitly or excluded by an assumption on M; as written, the algorithm is not a general procedure for the stated problem.","section":"Section 2.3, Algorithm 1, lines 3–6"},{"comment":"The closure test—'if the convex hull is closed, terminate'—is asserted without proof or edge-case analysis. For an empty stable cone, a cone with a non-trivial lineality space, or redundant constraints, the test based on counting vertices and halfspaces from Quickhull may either never succeed (so the loop exhausts M and returns a malformed representation) or may return prematurely before all non-redundant constraints are identified. Since Algorithm 2 explicitly assumes a closed superset from Algorithm 1, this gap undermines the claimed practical algorithm and the correctness of the reported computations in Section 4.","section":"Section 2.3, Algorithm 1, lines 15–18 (closure test)"},{"comment":"The empirical analysis of the Lazega network (stable/unstable edges and nulls, dyad distances di) depends on the unproven correctness of the cone-computation procedure. Please either provide a formal correctness proof for the double-description implementation (including termination and the closure test) or replace it with a certified convex hull / vertex enumeration routine; otherwise the load-bearing computational claims are unsupported.","section":"Sections 2.3 and 4"}],"minor_comments":[{"comment":"The notation R^{|S|}_− is nonstandard and should be defined; also the surrounding text has '1 < i < |S|' where '1 ≤ i ≤ |S|' is intended.","section":"Section 2.1.1, Eq. (4)"},{"comment":"In the formula for Pr(accept|d−,G), the exponent should be written as θ^T(t(G−)−t(G)) with the transpose explicit; the current expression omits the transpose and the dependence on the parameter vector.","section":"Section 3.2"},{"comment":"The table is visually garbled: the row for G lists statistic values but no clear separator from the change-score columns. Reformatting would improve readability.","section":"Table 1"},{"comment":"The quantity di ('distance from the model to stabilization of graph G'_i') is used prominently but never formally defined; please define the distance metric and its sign convention precisely.","section":"Section 4"}],"recommendation":"major_revision","confidential_remarks":"The core geometric characterization is sound and likely publishable after revision, but the computational section needs either a rigorous correctness proof or a clearly stated use of a certified library. The degenerate-case failures are concrete and should be addressed before the paper can be accepted. The manuscript's fit to a discrete-mathematics venue is reasonable given the double-description and convexity content."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. The main theoretical result — that the parameter region where a target graph is locally stable is the interior of a convex cone defined by change-score inequalities Mθ<0 — is correct, and it follows immediately from the ERGM likelihood ratio; there is nothing deep there. The paper's real contribution is what it builds on top: a practical framing of 'stable cones,' a dyad vulnerability analysis, and worked applications.\n\nThe star example is well done. They derive the two halfspaces, compute one-step transition probabilities for dyad toggles, and show that the locally stable cone aligns with persistence under Metropolis dynamics. The Lazega analysis is also informative: they locate unstable edges and nulls under a fitted model and connect them to structural interpretations. For practitioners who want a fast, simulation-free way to identify parameter regions that stabilize a graph and to identify which dyads are most likely to flip, this is a useful tool.\n\nWhere it gets soft is the algorithm. Algorithm 1 is presented as a practical method for computing the stable cone, but its correctness is not established. The initialization requires two non-parallel rows of M; if all rows are parallel or there is only one unique constraint, the loop never terminates, even though a perfectly well-defined stable cone exists (e.g., a halfspace or empty). The closure test in lines 15–18 is heuristic: checking that the number of facets from Quickhull equals the current number of halfspaces is not proven to be a valid termination criterion for all degenerate cases. The paper's own worst-case complexity is admitted to be equivalent to brute force, so the 'speedup' claim depends on typical-case behavior that is not rigorously justified. None of this undercuts the core characterization, but the phrase 'practical algorithm' is stronger than what is actually demonstrated.\n\nMinor stuff: there's a malformed citation in section 4 ('[ ?, see e.g.]]hunter.et.al:jcgs:2012') that should be fixed, and the simulation plots lack error bars.\n\nBottom line: the paper deserves a serious referee. The theory is sound, the applications are genuinely useful, and the algorithm issue is addressable — it needs a proper proof or at least a careful statement of non-degeneracy assumptions, and a discussion of what happens when the cone is empty or not pointed. It is not desk-reject material.","headline":"The core stable-cone theorem is correct but elementary; the paper's value is in the framework and applications, with the algorithm's correctness being the main unresolved issue.","tokens_in":17113,"tokens_out":3458,"would_cite":true,"duration_ms":32153,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["05C80","52B11"],"pacs":[],"model":"deepseek-v4-flash","headline":"For exponential random graph models, the set of parameters that makes a target graph locally stable is always the interior of a convex cone through the origin.","keywords":["exponential random graph models","local stability","convex cone","change scores","stable cone","double description method","dyad vulnerability","network dynamics"],"falsifier":"Build a change-score matrix M whose stable region is empty, for example two rows that are exact negatives of each other such as (1,1) and (-1,-1), and run Algorithm 1; if the algorithm does not terminate, or if it returns a nonempty closed cone, the closure test is unsound and the practical method fails on this input.","tokens_in":16070,"feed_emoji":"🕸️","tokens_out":9789,"duration_ms":87123,"temperature":0.7,"pith_summary":"The paper asks when a given network structure is the most probable outcome of an exponential-family random graph model, compared with a set of alternative graphs such as its single-edge perturbations. It proves that the set of model parameters making the target graph locally stable is always the interior of a convex polyhedral cone with apex at the origin, with the cone's faces read off directly from the change-score matrix whose rows are the statistic differences between the target graph and each alternative. The authors supply a double-description algorithm that computes this stable cone efficiently, and show that membership in the cone predicts persistence under random-walk Monte Carlo dynamics and identifies which edges or non-edges are most likely to change first. The characterization is demonstrated on a star ('cult') network and on a law-firm collaboration network, where the fitted model turns out to be locally unstable for 100 of the 630 possible single-edge toggles.","feed_headline":"Stable network parameters always form a convex cone","feed_subtitle":"A single change-score inequality decides whether a target graph beats every alternative — no simulation needed.","key_machinery":"The central object is the change-score matrix $M$, whose $i$-th row is the vector $t(G'_i)-t(G)$ for an alternative graph $G'_i$, together with the stable cone $\\Phi=\\{\\theta: M\\theta<0\\}$ that it defines. The paper proves that $\\Phi$ is the interior of a convex polyhedral cone with apex at the origin, and computes it with the double description method, alternating between an H-representation (the halfspaces from the rows of $M$) and a V-representation (the cone's vertices, normalized onto a sphere), testing the closure of the hull with a convex hull algorithm.","core_discovery":"Under an ERGM with probability mass function $\\Pr(G=g|\\theta)=\\exp(\\theta \\cdot t(g))/K(\\theta)$, the condition that a target graph $G$ be more probable than each alternative $G'\\in S$ reduces to the linear inequalities $\\theta^T(t(G')-t(G))<0$. Collecting the change-score vectors into the $|S|\\times K$ matrix $M$, the stabilizing subset is $\\Phi=\\{\\theta: M\\theta<0\\}$, which is the interior of a convex cone through the origin; the paper proves convexity by showing that any positive linear combination of two stabilizing vectors is again stabilizing. Each row of $M$ defines an open halfspace bounded by a hyperplane through the origin, and the faces of the cone are a subset of these hyperplanes. Local stability against the Hamming sphere of radius 1 makes the target graph a local mode of the model, and the paper shows that membership in the stable cone is necessary but not sufficient for dynamic stability under random-walk Monte Carlo dynamics, with escape times growing exponentially as one moves away from the cone's faces. The paper also shows how to use the stable cone to separate stabilization by social forces from entropic effects, and how to score individual dyads by their signed distance to the cone faces to predict which ties are most likely to change.","pith_inferences":["The solid angle of the stable cone gives a quantitative measure of a model family's 'capacity' to stabilize a given structure, allowing different families (e.g. with different sufficient statistics) to be compared by how much parameter space they devote to a structure.","The dyad-vulnerability scores could be adapted into a fast anomaly-detection tool for longitudinal networks: in a fitted ERGM, edges whose change-score row pushes the parameter vector outside the stable halfspace are precisely the ones a stochastic process would sever first, a prediction testable on panel data without simulation.","Because the ERGM form matches a Boltzmann distribution, the stable cone also characterizes parameter regions where the target graph is a local free-energy minimum; a natural extension is to compute, for a given energy function, which microstates can be stabilized at all by inspecting the union of feasible cones."],"forward_implications":["Checking whether a fitted model stabilizes a given graph becomes an exact matrix-inequality test, with no simulation required.","The cone's facets locate the parameter values at which some alternative graph becomes as probable as the target, giving a geometric picture of how changes in social forces flip the most likely network structure.","Because local stability is necessary but not sufficient for dynamic stability, the cone can be used to focus simulation effort on the thin band near the faces where dynamic persistence may fail.","Signed distances from the fitted parameter vector to the cone's faces produce closed-form dyad-vulnerability predictions, identifying which ties or null ties are most likely to be toggled first."],"supporting_citations":[{"why":"Supplies the random-walk Monte Carlo framework used to test whether local stability matches dynamic persistence.","marker":"[21]"},{"why":"Provides the double description method for convex hulls and vertex enumeration that Algorithm 2 adapts to compute the stable cone.","marker":"[22]"},{"why":"Supplies the convex hull algorithm used in Algorithm 1 to test whether the current hull description is closed.","marker":"[23]"},{"why":"Provides the simulation software used for the star-network validation chains.","marker":"[26]"},{"why":"Provides the Lazega law-firm collaboration dataset used in the real-world case study.","marker":"[28]"},{"why":"Supplies the previously published ERGM specification for the law-firm network that is re-fitted and analyzed.","marker":"[29]"}],"fun_headline_variants":["ERGM stability: a convex cone, proven","Local stability is a convex cone in parameter space","Change-score inequalities define stable cone","Stable cone: necessary but not sufficient","Find stable parameters without simulation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The practical algorithm assumes that its convex-hull closure test reliably detects when the current halfspace-and-vertex description is the fully reduced stable cone; the paper offers no proof or edge-case analysis for degenerate or empty stable cones, so the method could fail to terminate or return an incorrect region in those cases.","fun_headline_variants_meta":{"raw":{"variants":["ERGM stability: a convex cone, proven","Local stability is a convex cone in parameter space","Change-score inequalities define stable cone","Stable cone: necessary but not sufficient","Find stable parameters without simulation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000241,"raw_usage":{"total_tokens":1608,"prompt_tokens":1117,"completion_tokens":491,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":733,"completion_tokens_details":{"reasoning_tokens":428}},"tokens_in":733,"tokens_out":491,"duration_ms":4938,"temperature":1.0,"reasoning_tokens":428,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:11:06.367716+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build a change-score matrix M whose stable region is empty, for example two rows that are exact negatives of each other such as (1,1) and (-1,-1), and run Algorithm 1; if the algorithm does not terminate, or if it returns a nonempty closed cone, the closure test is unsound and the practical method fails on this input.","supporting_citations":[{"cited_title":"Computational Statistical Methods for Social Network Analysis","cited_arxiv_id":null,"evidence_quote":"Supplies the random-walk Monte Carlo framework used to test whether local stability matches dynamic persistence."},{"cited_title":"A Pivoting Algorithm for Convex Hulls and Vertex Enumeration of Arrangements and Polyhedra","cited_arxiv_id":null,"evidence_quote":"Provides the double description method for convex hulls and vertex enumeration that Algorithm 2 adapts to compute the stable cone."},{"cited_title":"The Quick-hull Algorithm for Convex Hulls","cited_arxiv_id":null,"evidence_quote":"Supplies the convex hull algorithm used in Algorithm 1 to test whether the current hull description is closed."},{"cited_title":"ergm: A Package to Fit, Simulate and Diagnose Exponential-Family Models for Networks","cited_arxiv_id":null,"evidence_quote":"Provides the simulation software used for the star-network validation chains."},{"cited_title":"The Collegial Phenomenon: The Social Mechanisms of Cooperation Among Peers in a Corporate Law Partnership","cited_arxiv_id":null,"evidence_quote":"Provides the Lazega law-firm collaboration dataset used in the real-world case study."},{"cited_title":"Curved Exponential Family Models for Social Networks","cited_arxiv_id":null,"evidence_quote":"Supplies the previously published ERGM specification for the law-firm network that is re-fitted and analyzed."}],"review_version":1}