{"id":"d9a2550a-2cbb-45c0-99de-60effa9a96ab","arxiv_id":"2412.01489","paper_version":2,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"In the small-diffusivity limit, the spectral gap of the symmetric inclusion process with any number of particles equals that of the two-particle process, while adding reservoirs restores the one-particle random walk identity.","lead":"The paper proves new spectral gap identities for the symmetric inclusion process, a model where particles diffuse on a graph while attracting each other. In the limit of strong stickiness the whole system relaxes like two particles, while adding particle reservoirs makes the system relax like a single random walk.","discovery_kind":"extension","skeptic_critique":null,"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the spectral gap of the symmetric inclusion process (SIP) on finite weighted graphs, in both conservative and open (reservoir) settings, and transfers some results to the Brownian energy process. The main results are: (i) Theorem 2.2, a non-asymptotic lower bound for gap_{SIP}(G,α) that is linear in α_min and thus sharp up to constants depending only on the geometry; (ii) Theorem 2.3, an asymptotic two-particle spectral gap identity lim_{ε→0} gap_k(G,ε α̂)/gap_2(G,ε α̂) = 1 for every k ≥ 3; (iii) Theorem 2.4, giving graphs for which the one-particle identity fails, with lim gap_2/gap_1 < 1; and (iv) Theorem 6.1 and Corollary 6.4, a one-particle spectral gap identity for the non-conservative SIP with absorbing or reversible reservoirs, valid for all α_min > 0. The proofs combine the rigid eigenstructure and annihilation/creation operators of SIP, Dirichlet-form comparison inequalities via constructed paths and flows, slow-fast limit theorems with absorption estimates, and orthogonal-polynomial duality for the reservoir system.","tokens_in":51321,"tokens_out":9970,"duration_ms":97486,"significance":"If the proofs are correct, these are substantial contributions to the spectral theory of interacting particle systems. The two-particle asymptotic identity is a new structural result that gives a precise sense in which the two-particle dynamics dominates the relaxation of the many-particle conservative SIP in the low-diffusivity limit; the sharp α_min-linear lower bound resolves a natural scaling question left open by the earlier quadratic bound (1.5). The non-conservative one-particle identity extends the short list of models satisfying Aldous-type spectral gap identities to a strongly interacting, genuinely reversible open system. The paper is largely self-contained: the main arguments are explicit, the comparison constants are explicit, and the slow-fast analysis is supported by two technical appendices. The paper does not rely on fitted parameters or circular assumptions. The main weaknesses are localized technical points in the proof of Proposition A.1 and in the overlap-counting step of Section 4.5, both of which appear repairable.","major_comments":[{"comment":"The proof asserts that τ_{∆_k}, the first hitting time of the transient set, stochastically dominates an exponential random variable with rate c_max k α̂_max. This is not correct as stated: from a configuration ξ ∈ Ω_k, the total rate of jumps that enter ∆_k is generally larger than c_max k α̂_max by a factor comparable to the maximum degree of G, because several directed edges can contribute to entering ∆_k simultaneously. Consequently the displayed stochastic domination fails. The intended conclusion, however, follows from the same argument after replacing the single-exponential domination by a union bound using a uniform upper bound on the total jump rate into ∆_k, so Proposition A.1 should remain valid once the lemma is corrected. Since Proposition A.1 is used in Proposition 5.1 and hence in Corollary 5.2, this correction is load-bearing.","section":"Appendix A, Lemma A.4"},{"comment":"The proof of (4.3) relies on the claim that each Dirichlet summand ∇²f(η, η − δ_z + δ_w) is produced by at most six triples (ℓ, m, σ), and Case 3, the reverse orientation (x_{s−1}, x_s) = (w, z), is dismissed as 'almost identical' without details. Because the constant 6 enters the final comparison constant of Theorem 3.2, a full verification of Case 3, or a more systematic counting argument covering both orientations, should be supplied so that the reader can check the bound.","section":"Section 4.5, Case 3"}],"minor_comments":[{"comment":"There are typographical errors in the derivation after (7.6): 'ω_x ξ_x (θ_x − σ)' should read 'ω_x ξ_x (θ_x − ̺)', and 'D_σ(ξ − δ_x, η)' should read 'D_{α,̺}(ξ − δ_x, η)'; later in the same display, 'F^{b̺_{α,ω,θ,k−1}}_{α,̺}' should be 'F^{b^θ_{α,ω,̺,k−1}}_{α,̺}'.","section":"Section 7.2"},{"comment":"The proof of Theorem 6.1 establishes gap_k(G,α,ω) ≥ gap_1(G,α,ω) via the survival-probability bound (7.2), but the reverse inequality gap_k(G,α,ω) ≤ gap_1(G,α,ω), which is needed for equality, is not explicitly proved; it follows from the annihilation/lifting intertwining for the killed dynamics, but should be stated.","section":"Section 6.2"},{"comment":"In the proof of Theorem 2.4, the inequalities (5.29) and the use of [Her23, Eq. (1.7)] are quoted from external sources without displaying the exact statement used; please state the precise inequalities so that the reader can verify the direction of the bound λ_{2,2} ≤ 1/E_{κ(·|W)}[τ_{W^c}].","section":"Section 5.5"},{"comment":"The paragraph after (Case 2.III) says that forward jumps in type [III] are of three types, but the correspondence between these three types and the arrows in Figure 4.3 is not fully spelled out; a short table or explicit map from the figure's arrows to the three cases would improve reproducibility of the count.","section":"Section 4.5"}],"recommendation":"major_revision","confidential_remarks":"The paper's central claims appear sound and the overall argument is coherent. The main issue is the incorrect stochastic-domination assertion in Lemma A.4, which is load-bearing for Proposition A.1 but easily repaired by a Poisson-thinning or union-bound argument. The overlap-counting omission in Section 4.5 should also be addressed. If the authors fix these two points, I would be happy to see the paper accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a substantial and largely rigorous paper. The genuinely new content is the asymptotic two-particle spectral gap identity in the vanishing-diffusivity limit (Theorem 2.3), the failure of the one-particle identity outside the log-concave regime on explicit geometries (Theorem 2.4), and the non-conservative one-particle identity with reservoirs (Theorem 6.1 and Corollary 6.4). The proofs are built honestly: the Dirichlet-form comparison in Section 4 is explicit, the slow-fast asymptotics in Section 5 rest on a cited limit theorem plus absorption estimates proven in Appendix A, and the paper openly says the exact two-particle identity remains unresolved. There are no fitted parameters and no circularity. The self-citation to [KS24] is appropriate; the new claims do not assume the target results.\n\nSoft spots, in proportion: the slow-fast step is load-bearing for Theorems 2.3 and 2.4. Proposition 5.1 passes from semigroup convergence to exact convergence of the spectral gap, which requires the fast inclusion dynamics to thermalize on the absorbing set uniformly in the starting configuration. The appendix supplies explicit estimates, but they are delicate and deserve independent checking, especially the log k dependence in Lemma A.2. The overlap count in Section 4.5 is another place where the text says \"tedious computations\" and waves at Case 3; it is probably correct, but a referee should verify it. The non-reversible extension in Remark 6.5 is flagged by the authors as plausible but not a theorem, which is acceptable but limits the scope of Corollary 6.4. None of these issues undermines the main claims as far as I can tell, but Section 4 and Appendix A are where a referee should spend time.\n\nWho this is for: researchers working on spectral gaps, metastability, and interacting particle systems, particularly those interested in Caputo's conjecture and its variants. It adds SIP to the short list of models where a spectral gap identity can be proved or sharply disproved on general geometries. I would send it to a serious referee and, conditional on a careful check of Sections 4 and 5, accept it. My own verdict is close to accept, with the main caveat being independent verification of the technical estimates.","headline":"Sharp and mostly rigorous asymptotic two-particle reduction for SIP; the slow-fast eigenvalue step is the thing to check.","tokens_in":51875,"tokens_out":1456,"would_cite":true,"duration_ms":16171,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60K35","60J27","05C50"],"pacs":[],"model":"deepseek-v4-flash","headline":"As diffusion vanishes, the k-particle inclusion process has the two-particle spectral gap; with reservoirs, the one-particle identity holds for all α_min>0.","keywords":["spectral gap","symmetric inclusion process","two-particle spectral gap identity","slow-fast systems","Dirichlet form comparison","Brownian energy process","open systems","random walk"],"falsifier":"Exact diagonalization of the three- and two-particle SIP generators on a small graph, for instance the three-site path with α=(ε,ε,ε), for a sequence ε→0 should show gap_3/gap_2→1; any graph for which the numerical liminf exceeds 1 falsifies Theorem 2.3. For the complementary failure claim, on the two-dimensional torus T²_N the same kind of computation should show gap_2/gap_1→0 like 1/log N.","tokens_in":51175,"feed_emoji":"🎲","tokens_out":10482,"duration_ms":94171,"temperature":0.7,"pith_summary":"Many-particle systems with mutual attraction are hard to analyze because the particles' tendency to clump couples all degrees of freedom. This paper studies the symmetric inclusion process, a graph model in which particles jump onto occupied sites faster than onto empty ones, and asks whether the spectral gap—the rate of convergence to equilibrium—can be read off from a small subsystem. It establishes two complementary reductions: in the closed, conservative system, when the independent-diffusion parameter tends to zero, the gap of any number k≥3 of particles is asymptotically equal to the gap of two particles, while the older one-particle identity is genuinely false outside the log-concave regime α_min≥1. It also proves, by comparison with the complete graph, a sharp lower bound showing the gap decays at most linearly in the minimal site weight, rather than quadratically as the naive two-sided bound suggests. In the open system in contact with reservoirs, the gap is exactly the gap of a single random walk with killing, for every positive diffusivity, and the identities carry over to the Brownian energy process.","feed_headline":"Two particles set the many-particle mixing rate","feed_subtitle":"Many-particle spectral gaps collapse to the two-particle gap; open systems keep a one-particle law.","key_machinery":"The argument runs on two tracks. In the closed system, the engine is the rigid eigenstructure encoded by the annihilation operator a_k and its creation adjoint a†_{α,k-1}: consistency L_{G,α,k} a_k = a_k L_{G,α,k-1} lifts every eigenvalue of the (k-1)-particle system into the k-particle spectrum, and the orthogonal decomposition L²(µ_{α,k}) = Im a_k ⊕ Ker a†_{α,k-1} reduces the spectral gap to a Dirichlet-form quotient on Ker a†. The comparison Theorem 3.2 bounds the Dirichlet form of any graph by that of the complete graph using distinguishing paths and, for large stacks, two-dimensional flows; the complete-graph form satisfies the explicit identity E_{K,α,k}(f)=k(|α|+k-1)‖f‖²_{α,k}. In the vanishing-diffusivity limit the paper switches to a slow-fast decomposition $ε^{{-1}}$L_{εα̂,k}=A_{α̂,k}+$ε^{{-1}}$B_k, where B_k is the fast, purely attractive process absorbed on the set Ω_k of configurations with no two particles at distance one; projective semigroup convergence yields exact convergence of gaps to eigenvalues of the limiting generator Π_k A_{α̂,k}, whose block-triangular structure on the levels Ω_{k,m} of m separated stacks is then analyzed. A consistency relation for the limiting generators gives λ_{k,m}(α̂)=λ_{m,m}(α̂), and a variational comparison gives λ_{k,k}(α̂)≥λ_{2,2}(α̂), so the lowest transient eigenvalue is always the two-particle one. The open system instead uses orthogonal-polynomial dualities to lift eigenfunctions of the absorbing k-particle system to eigenfunctions of the reservoir process, producing a complete orthogonal basis and exact spectral gap transfer.","core_discovery":"The paper's central claim is that the low-lying spectrum of the conservative symmetric inclusion process is dictated by the two-particle dynamics exactly in the limit where the independent diffusion parameter vanishes. Concretely, for any finite connected graph G and fixed normalized site weights α̂, Theorem 2.3 gives lim_{ε→0} gap_k(G, ε α̂)/gap_2(G, ε α̂)=1 for every k≥3, so the many-particle spectral gap is asymptotically the two-particle gap; this is the ε-asymptotic form of the conjectured identity gap_SIP=gap_2. The companion Theorem 2.4 shows the older one-particle identity gap_SIP=gap_RW is genuinely false outside the log-concave regime α_min≥1: on large tori, the two-particle gap is strictly smaller than the random-walk gap in the vanishing-diffusivity limit. For the non-conservative process in contact with reservoirs and constant reservoir density, Corollary 6.4 restores the one-particle identity for every α_min>0: gap_SIP(G,α,ω,ρ)=gap_RW(G,α,ω), where the random walk is killed at rate ω_x at x. The same statements transfer to the Brownian energy process by the isospectrality of the two generators.","pith_inferences":["Beyond the stated theorems, the exact identity gap_SIP=gap_2 for all α rather than only in the ε→0 limit remains open; upgrading Proposition 5.1 to a quantitative comparison with constants uniform in ε would be a natural route to prove the full conjecture for all graphs.","The dichotomy suggests a general principle for Dirichlet-reversible interacting systems: closed systems are governed by the smallest subsystem capable of forming a condensate, while reservoirs erase the interaction slowdown. This predicts the same two-particle reduction for Beta-Binomial splitting and KMP-type models on arbitrary graphs, where the paper only records partial results.","A direct testable extension: on a graph where the closed SIP gap is strictly below the random-walk gap, adding a single reservoir should replace the gap by the killed random-walk gap, discontinuously at reservoir strength zero; this could be checked numerically for moderate particle numbers.","The explicit comparison constants in Theorem 2.2 depend on |V|² diam(G) and an exponential in α_max/α_min, and the paper notes they are not optimal. On boxes and tori one expects the sharp prefactor to be governed by the ratio of relaxation time to meeting time, as in the torus calculation behind Theorem 2.4."],"forward_implications":["On every finite connected graph, the many-particle spectral gap of the closed symmetric inclusion process is, in the small-diffusivity limit, the two-particle gap; the earlier (k−1)-particle eigenvalue inclusion cannot reverse the ordering among k≥3.","The one-particle identity fails generically in the attractive regime: on large two- or higher-dimensional tori with uniform weights, gap_2/gap_1 tends to a constant below 1, with the two-particle eigenvalue of order 1/(N² log N) in two dimensions.","Even without the asymptotics, gap_SIP is controlled from below by a constant times α_min, with an explicit constant in terms of |V|, diameter, and conductivity, so the decay in diffusivity is at most linear.","Opening the system to reservoirs with a constant reservoir density restores the one-particle law: the gap equals that of a single random walk killed at the reservoir rates, for every α_min>0 and every graph.","The Brownian energy process, being isospectral with the symmetric inclusion process, inherits both the sharp lower bound and the failure or success of the one-particle identities."],"supporting_citations":[{"why":"Establishes the α_min≥1 one-particle identity and supplies the consistency, annihilation/creation eigenstructure, and Brownian-energy-process isospectrality on which Theorems 2.2–2.4 and Section 8 rely.","marker":"[KS24]"},{"why":"Provides the limit theorem for perturbed operator semigroups used in Proposition 5.1 to pass from semigroup convergence to exact spectral-gap asymptotics.","marker":"[Kur73]"},{"why":"Supplies the Markov-process formulation (Theorem 7.6) of the same slow-fast convergence used to obtain the limiting generator Π_k A_{α̂,k}.","marker":"[EK86]"},{"why":"Gives the meeting-time asymptotics for coalescing random walks on the torus used in the proof of Theorem 2.4 to show λ_{2,2} is much smaller than the random-walk gap.","marker":"[Cox89]"},{"why":"Supplies the inequality bounding the smallest eigenvalue of a killed process by an expected stationary hitting time, used in inequality (5.30) for λ_{2,2}.","marker":"[Her23]"},{"why":"Provides the orthogonal polynomial dualities and the L² basis for boundary-driven SIP used in Theorem 6.3 to lift absorbing-system eigenfunctions to the reservoir process.","marker":"[FRS22]"},{"why":"Introduces the comparison of Dirichlet forms via distinguishing paths that Section 4 adapts for the graph-to-complete-graph comparison in Theorem 3.2.","marker":"[DSC93]"}],"fun_headline_variants":["Many-particle gap collapses to two-particle gap","One-particle identity holds for open SIP","Two-particle gap dictates many-particle mixing","Symmetric inclusion process: gap follows two particles"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The asymptotic two-particle identity rests on the assumption that the fast attraction mechanism always drives the system into fully piled-up configurations quickly enough, uniformly over the starting configuration, before the slow free diffusion has time to act; if a geometry allowed a long-lived intermediate configuration of separated particles, the limiting spectrum could differ.","fun_headline_variants_meta":{"raw":{"variants":["Many-particle gap collapses to two-particle gap","One-particle identity holds for open SIP","Two-particle gap dictates many-particle mixing","Symmetric inclusion process: gap follows two particles"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000176,"raw_usage":{"total_tokens":1306,"prompt_tokens":980,"completion_tokens":326,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":272}},"tokens_in":596,"tokens_out":326,"duration_ms":3561,"temperature":1.0,"reasoning_tokens":272,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:17:41.998878+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Exact diagonalization of the three- and two-particle SIP generators on a small graph, for instance the three-site path with α=(ε,ε,ε), for a sequence ε→0 should show gap_3/gap_2→1; any graph for which the numerical liminf exceeds 1 falsifies Theorem 2.3. For the complementary failure claim, on the two-dimensional torus T²_N the same kind of computation should show gap_2/gap_1→0 like 1/log N.","supporting_citations":[],"review_version":1}