{"id":"d5d6004d-5894-42a6-b0f6-1b78c42a5566","arxiv_id":"1908.03406","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Connecting two networks by the highest eigenvector-centrality product links maximizes SIS spreading prevalence near the critical point and minimizes the outbreak threshold.","lead":"This paper studies how to add multiple edges between two separate networks to make a disease spread as much as possible. It shows that near the epidemic threshold, the best simple rule is to link nodes that already rank high in eigenvector centrality, and this also lowers the outbreak threshold.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The optimality claim rests entirely on neglecting δu in Eqs. (16)–(17); for finite real networks this reduction of ⟨ρ⟩ to a function of E(C) is unverified, and LEC is only shown to beat two heuristics.","rationale":"The reader and I converge on the same load-bearing point: the derivation's validity for finite real-world networks. I do not see an internal inconsistency in the algebra leading to Eqs. (23)–(26); within the two-eigenvector ansatz the reduction to E(C) and hence LEC is coherent. The paper also provides non-trivial support for the weaker claim: Fig. 1 shows the QMF iteration matches Monte Carlo except for the Facebook pair, Fig. 2 shows small eigenvalue error and high cosine similarity for LEC-added edges, and Figs. 3–4 show LEC beats LDC and RC near criticality for all five pairs. These are real empirical supports. What remains unprotected is the word 'optimal' as applied to the exact prevalence: the test against LDC and RC is not a test against the true optimum, and the δu correction is dismissed by an asymptotic scaling argument rather than measured on the actual networks. The paper's own Discussion acknowledges the approximation regime but does not quantify its error. A small-network exhaustive search plus a swap search on the real pairs would settle whether LEC is actually optimal or merely a good heuristic. Hence the reader's CONDITIONAL verdict should stand unchanged.","tokens_in":98,"tokens_out":7655,"duration_ms":145232,"concrete_test":"On small two-layer networks (e.g., two random or deterministic graphs with 12–16 nodes, including a low-degree, high-modularity case resembling Facebook), generate all inter-layer edge sets for δM = 1, 2, 3 and compute the exact stationary prevalence from the discrete-time Markov chain equations iterated to convergence at λ just above the original critical point. Rank all candidates and compare the true maximizer with the LEC set. Repeat with a local-search swap algorithm seeded at LEC on the real Advogato–Facebook and HepPh–HepTh pairs: if any swap increases the exact QMF prevalence, LEC is not optimal. Also report the residual norm ‖u_true − P u_true‖, where P is the projector onto span{v̂_a1, v̂_b1}, for both LEC and alternative edge sets; if the residual is not small, the E(C)-only reduction is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central step is the approximation u ≈ u0 = βa v̂a1 + βb v̂b1 in Sec. III. Eq. (13) replaces the power-iteration limit by one application of G, and Eqs. (16)–(17) then drop δu because ω_a1 diverges for scale-free networks. This is an asymptotic statement. For the finite test networks, particularly Facebook (⟨k⟩ ≈ 2.06, modularity 0.809), the spectral gap and localization properties are far from that regime. If δu is not negligible, ⟨ρ⟩ depends on C not only through E(C) but through overlaps of δG v̂ with non-principal eigenvectors, so the reduction that ⟨ρ⟩ depends on C only in the form of E(C) fails. Then selecting top products v_a1(i)v_b1(j) is the maximizer of an approximate objective, not of the true prevalence. The paper's simulations compare LEC with LDC and RC only; those are weak baselines, and Fig. 2's cosine-similarity check is performed for LEC-added edges, so it does not test whether a different edge set would outperform LEC. The paper's own Discussion concedes the method is limited to small p and near-threshold λ but gives no quantitative estimate of the δu error for the real networks. Thus the stronger wording 'optimal' is not established; only 'near-optimal within the two-eigenvector ansatz and better than LDC/RC near criticality' is supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies how to add a fixed small number of inter-layer edges between two isolated networks so as to maximize the stationary prevalence of the susceptible-infected-susceptible (SIS) model near the epidemic threshold. Using the quenched mean-field formula of Eq. (8), the authors approximate the leading eigenvalue and eigenvector of the interconnected adjacency matrix by a two-dimensional projection onto the two principal eigenvectors, obtaining an expression in which the prevalence depends on the inter-layer matrix C only through E(C)=<v_a1, C v_b1>. Maximizing E(C) leads to the 'large eigenvector connecting' (LEC) rule: connect pairs with the largest products of eigenvector centralities. The rule is tested against large-degree connecting (LDC) and random connecting (RC) on three synthetic and two real-world network pairs, with LEC reported to give the largest prevalence increase near the critical point. The paper concludes that LEC is the optimal inter-layer structure and that it simultaneously minimizes the outbreak threshold.","tokens_in":12464,"tokens_out":4894,"duration_ms":54906,"significance":"If the optimality claim were established, the paper would provide a simple and useful criterion for designing interconnected networks to promote spreading, extending earlier single-edge results to multiple edges. The derivation in Sec. III is transparent, and the numerical comparison in Figs. 3 and 4 is clearly presented and consistent with LEC being a good heuristic near criticality for the tested networks. However, the central claim is stronger than what is demonstrated: the approximation neglects a potentially non-negligible eigenvector correction, the monotonicity of prevalence in E(C) is asserted empirically rather than proven, and optimality is tested only against two heuristics rather than over the full space of edge sets. These gaps do not appear fatal to the heuristic value of LEC, but they do mean that the paper currently establishes 'better than LDC and RC near criticality within a two-eigenvector approximation', not unconditional optimality.","major_comments":[{"comment":"The reduction of the prevalence to a function of E(C) rests on neglecting the eigenvector correction δu. The justification given in the text is that the leading eigenvalue of scale-free networks diverges in the thermodynamic limit, so that δu is small. This is an asymptotic statement and is not verified for the finite networks used here, particularly Facebook, which has mean degree 2.06 and modularity 0.809 and is far from the divergent-eigenvalue regime. Please provide a quantitative estimate of ||δu||/||u0|| and of the neglected terms in Eq. (18) for all five network pairs as a function of p, and state explicitly the range of p and λ over which Eq. (25) is accurate. Without such a check, the central reduction of the optimization problem to E(C) is not established for the real-world cases.","section":"Sec. III, Eqs. (12)-(17)"},{"comment":"The step from '⟨ρ⟩ depends on C only through E(C)' to 'therefore maximize E(C)' uses the assertion that ⟨ρ⟩ is an increasing function of E(C), which is described as empirical ('Empirically, for all the networks considered...'). This monotonicity is load-bearing: if it fails, a smaller E(C) could give a larger prevalence and the LEC rule would not follow. Please show the actual curve of ⟨ρ⟩ versus E(C) over the candidate range for each network pair, either from Eq. (26) or from iterating Eq. (4), and either prove the monotonicity or specify the exact interval of E(C) where it holds. The Discussion already concedes non-monotonicity for large E(C), so the domain restriction must be quantified.","section":"Sec. III, after Eq. (26)"},{"comment":"The word 'optimal' is used in the abstract, Sec. I, and Sec. V, but the numerical evidence compares LEC with only two heuristics, LDC and RC. The cosine-similarity check in Fig. 2 is performed on LEC-chosen edge sets, so it does not test whether a different edge set would yield a larger prevalence. Please provide an exhaustive comparison for at least small δM on small synthetic instances (for example, δM = 1 and 2, or a greedy marginal-gain search) or, alternatively, revise the title, abstract, and discussion to claim only 'near-optimal within the stated approximation and better than LDC/RC near criticality'.","section":"Secs. III-IV, Figs. 3-4"},{"comment":"The power-method limit u = lim_{n→∞}(ω_a1)^{-n} G^n u0 is approximated by a single application of G (n=1). This is a separate approximation whose error is not bounded. The contributions of δG acting on non-principal eigenvectors are discarded along with δu, and their size depends on the spectral gap of G0, which is not characterized for the real networks. Please either bound this term analytically in terms of ||δG|| and the spectral gap or test its magnitude numerically on the finite networks used in the paper.","section":"Sec. III, Eqs. (13)-(14)"}],"minor_comments":[{"comment":"In the text and caption, 'expect Advogato-Facebook' should be 'except Advogato-Facebook'.","section":"Sec. IV, Fig. 1"},{"comment":"The variability measure is written as the square root of ⟨ρ⟩^2 divided by ⟨ρ⟩^2 minus one; the notation is ambiguous. Please define ⟨ρ⟩^2 explicitly as the mean over independent runs of the squared prevalence and clarify the denominator.","section":"Sec. IV, Eq. (28)"},{"comment":"The caption says 'The while lines correspond to the contour'; this should be 'white lines'.","section":"Fig. 4 caption"},{"comment":"No error bars are reported for δ⟨ρ⟩. Since the differences between LEC and LDC are small in some panels, please add standard errors or state explicitly that the reported differences are above Monte Carlo noise.","section":"Sec. IV, Fig. 3"},{"comment":"The introduction says previous studies mostly considered adding a single edge, but Ref. [42] is later described as generalizing the optimal strategy to two edges. Please clarify precisely which cases are new in the present paper.","section":"Sec. I"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a plausible heuristic with a partially derived motivation, but the optimality language is not supported by the current analysis. I would ask the editor to require the authors to either add the quantitative checks described in the major comments or substantially temper the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is that the single-edge result of Aguirre et al. is extended to a fixed number of inter-layer edges. The derivation is transparent: linearize the SIS Markov chain near the threshold, use the Goltsev prevalence formula, approximate the leading eigenvector as a combination of the two layer principal eigenvectors, and reduce the objective to E(C), the sum of eigenvector-centrality products. Maximizing E(C) for a fixed edge count is a simple top-k selection. That is a real step forward, and the resulting rule is natural.\n\nThe paper also earns credit for testing its approximations directly. Figure 2 reports the eigenvalue error and the cosine similarity between the true and approximated leading eigenvector for p up to 0.01. The cosine similarities are high for all five network pairs, including the two real-world ones, so the two-eigenvector ansatz is not just a convenient fiction for these cases. The performance comparison in Figures 3 and 4 shows LEC beating LDC and RC in the near-critical band, consistent with the theory.\n\nThe soft spots are real but not fatal. The word 'optimal' overstates what is shown: the optimization is exact only inside the ansatz, and outside it the simulations only establish that LEC beats two heuristics, not that no other edge set does better. The neglect of δu is justified asymptotically for scale-free networks, but the Facebook layer has average degree about 2 and modularity 0.81, where the asymptotic argument is not obviously valid. Figure 2 mitigates this by showing the ansatz holds empirically for LEC-added edges, but it does not test the ansatz for other edge selections, so the optimality proof has a gap. The monotonicity of prevalence in E(C) is asserted from numerical examples rather than proven. To the paper's credit, the Discussion acknowledges the near-critical, small-p regime and the limits of linear approximation, but it does not quantify the δu error.\n\nThis is a paper for network scientists working on multilayer spreading or structural optimization. The derivation is reproducible from the text and the simulations are consistent. It deserves a serious referee, not a desk reject. The authors should be asked to soften the optimality claim, to state clearly that the approximation error is not controlled for all candidate edge sets, and to tighten the monotonicity discussion. With those revisions, it is a solid contribution.","headline":"A clean approximate answer to the multi-edge interconnection problem; 'optimal' is only within an uncontrolled approximation, but the paper is worth refereeing.","tokens_in":12942,"tokens_out":4184,"would_cite":false,"duration_ms":45740,"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":"Adding edges between the most eigenvector-central nodes maximizes epidemic spread.","keywords":["interconnected networks","SIS model","spreading dynamics","eigenvector centrality","multilayer networks","epidemic threshold","network optimization","Markov chain approximation"],"falsifier":"On a network pair with comparable leading eigenvalues and high modularity, add LEC edges and compare the true leading eigenvector with $u_0$; if the cosine similarity drops well below 1 or the predicted prevalence ordering reverses relative to LDC, the spectral approximation and with it the rule is falsified in that regime.","tokens_in":11923,"feed_emoji":"🦠","tokens_out":5544,"duration_ms":55060,"temperature":0.7,"pith_summary":"The paper studies how to add a fixed small number of edges between two isolated networks to maximize the stationary prevalence of an SIS epidemic. Its central claim is that near the outbreak threshold, the edges should join node pairs ranked by the product of their eigenvector centralities in the two layers (the LEC strategy), rather than by degree or at random. The same choice also lowers the epidemic threshold, so one placement rule optimizes both spreading measures. Numerical iterations of the Markov-chain equations on three synthetic and two real-world network pairs support the strategy near the critical point.","feed_headline":"Link top eigenvector-central nodes to maximize epidemic spread","feed_subtitle":"Near the outbreak threshold, this rule beats large-degree and random linking in five network pairs.","key_machinery":"The LEC strategy, whose load-bearing object is the scalar $E(C) = \\langle v_1^a, C v_1^b \\rangle$, the sum of products of eigenvector centralities across the chosen inter-layer edges. With the approximation $u \\approx u_0$, the leading eigenvalue shift is $\\delta\\omega_1 = \\tfrac{1}{2}(\\sqrt{4E(C)^2 + g^2} - g)$, where $g = \\omega_1^a - \\omega_1^b$, and the prevalence depends on $C$ only through $E(C)$. This turns a combinatorial search over candidate edge sets into a simple ranking problem; the paper then checks monotonicity in $E(C)$ and selects the top-ranked edges.","core_discovery":"For two layers with adjacency matrices $G_a$ and $G_b$ and principal eigenvectors $v_1^a$, $v_1^b$, the paper approximates the leading eigenpair of the interconnected network by taking the leading eigenvector as $u_0 = \\beta_a \\hat{v}_1^a + \\beta_b \\hat{v}_1^b$. All dependence of the stationary prevalence on the inter-layer connection matrix $C$ then enters through the single scalar $E(C) = \\langle v_1^a, C v_1^b \\rangle$. Since the near-critical prevalence and the threshold shift $\\delta\\omega_1$ are increasing in $E(C)$ for the cases considered, the optimal strategy is to choose the $\\delta M$ edges with the largest values $v_1^a(i) v_1^b(j)$. The paper calls this LEC and verifies that it outperforms large-degree and random connecting near the critical point.","pith_inferences":["Because $E(C)$ is bilinear in the two layer eigenvectors, the top-product rule is a rank-1 approximation to a constrained quadratic assignment; on small instances one could test whether exact enumeration ever beats LEC, and if so how large the gap is.","The same spectral reduction may carry over to other dynamics controlled by the leading eigenpair, such as synchronization or network reliability, replacing prevalence with the relevant spectral objective.","A refined strategy for high-modularity, low-average-degree layers could keep the $\\delta u$ correction term instead of dropping it; that would yield a testable prediction for when LEC should fail."],"forward_implications":["Near the epidemic threshold, LEC outperforms large-degree and random connection strategies for small added-edge budgets across all five tested network pairs.","The LEC structure that maximizes prevalence also minimizes the outbreak threshold, so a single design rule serves both promotion and containment objectives.","On synthetic scale-free networks LEC and LDC nearly tie, but on the real-world pairs with complex structure LEC has a clear advantage near threshold.","Away from the critical point and for very large added-edge counts, the linear approximation degrades and LDC or RC can take over."],"supporting_citations":[{"why":"Supplies the formula expressing near-critical prevalence through the leading eigenvalue and eigenvector.","marker":"[45]"},{"why":"Gives the optimal single-edge strategy based on eigenvector centrality that this paper extends to multiple edges.","marker":"[40]"},{"why":"Provides the discrete-time Markov chain equations used as the dynamical model.","marker":"[46]"},{"why":"Justifies neglecting $\\delta u$ by the divergence of the leading eigenvalue in scale-free networks.","marker":"[47]"},{"why":"Shows that a small fraction of inter-layer edges can trigger SIS outbreaks in multilayer networks, framing the optimization problem.","marker":"[31]"}],"fun_headline_variants":["Optimal inter-layer links: pick top eigenvector pairs","Maximize spread by linking eigenvector hubs","Eigenvector pairing beats degree for inter-layer links","Near criticality, connect top eigenvector nodes","LEC rule: add edges where eigenvector products peak"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The derivation assumes the interconnected network's leading eigenvector is well approximated by the fixed linear combination $\\beta_a \\hat{v}_1^a + \\beta_b \\hat{v}_1^b$, with the correction $\\delta u$ negligible; if that fails, the reduction of the optimization to $E(C)$ and the LEC rule does not follow.","fun_headline_variants_meta":{"raw":{"variants":["Optimal inter-layer links: pick top eigenvector pairs","Maximize spread by linking eigenvector hubs","Eigenvector pairing beats degree for inter-layer links","Near criticality, connect top eigenvector nodes","LEC rule: add edges where eigenvector products peak"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000218,"raw_usage":{"total_tokens":1406,"prompt_tokens":876,"completion_tokens":530,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":492,"completion_tokens_details":{"reasoning_tokens":455}},"tokens_in":492,"tokens_out":530,"duration_ms":5640,"temperature":1.0,"reasoning_tokens":455,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:13:27.814324+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a network pair with comparable leading eigenvalues and high modularity, add LEC edges and compare the true leading eigenvector with $u_0$; if the cosine similarity drops well below 1 or the predicted prevalence ordering reverses relative to LDC, the spectral approximation and with it the rule is falsified in that regime.","supporting_citations":[{"cited_title":"Dickison, S","cited_arxiv_id":null,"evidence_quote":"Supplies the formula expressing near-critical prevalence through the leading eigenvalue and eigenvector."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the optimal single-edge strategy based on eigenvector centrality that this paper extends to multiple edges."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the discrete-time Markov chain equations used as the dynamical model."},{"cited_title":"Wang, Q.-H","cited_arxiv_id":null,"evidence_quote":"Justifies neglecting $\\delta u$ by the divergence of the leading eigenvalue in scale-free networks."}],"review_version":1}