{"id":"4a856770-dc6e-4690-a3de-b94766a2521d","arxiv_id":"1908.04175","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"The subcritical contact process modulo translations has a unique quasi-stationary distribution.","lead":"The subcritical contact process, a model of infection that always eventually dies out, has exactly one quasi-stationary distribution when viewed up to translations. This resolves a long-open question and provides the first example of a process with a unique quasi-stationary distribution that does not come down from infinity in finite time.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 3.5's uniform cut-point bound is imported from a d=1 result; the d≥2 extension needed for Theorem 3.2 is not proved in the paper.","rationale":"The paper's central claim is the uniqueness of the QSD for the subcritical contact process modulo translations, proved by reducing every QSD to the Yaglom limit via Theorem 3.2. The proof of Theorem 3.2 is structurally sound: the cut break point construction, the spatial factorization (∗)= and (⋆)=, and the final approximation by ν* are all coherent, and the use of positive association does not by itself invalidate the argument. The most load-bearing point is not the raw good-point estimate (2.3), on which the reader focused, but the step that converts it into a uniform conditional statement for the random cut point: Lemma 3.5(3.7). The manuscript's justification of (3.7) explicitly cites a lemma from [AEGR15] that was proved for d=1, and it does not reproduce or cite a d≥2 version. The asserted inequality P(X∉⌊η0⌋, τ>t) ≤ e^{−e^t+1} is a decreasing-event estimate where FKG gives only a lower bound, so the upper bound is not a formal consequence of the cited good-point estimate alone. This is a missing-support concern rather than a demonstrated falsehood: the result may well be true, and [DR17] may contain the needed generalization, but the paper as written leaves the gap. I therefore recommend CONDITIONAL acceptance pending a verified d≥2 proof of the uniform estimate in Lemma 3.5, rather than outright acceptance or rejection. No other load-bearing objection surfaced: the factorization step is justified by the large separation R_t ≫ βt, the use of lexicographic order is translation invariant, and the final QSD argument is clean.","tokens_in":7265,"tokens_out":37179,"duration_ms":373144,"concrete_test":"Check [DR17, §3] and [AEGR15, Lemma 2.8] for a d≥2 proof of the bound Pη0(X∉⌊η0⌋, τ>t) ≤ e^{−e^t+1} for arbitrary finite η0⊂Z^d. If the bound is neither proved there nor in the present manuscript, request that the authors supply the d≥2 proof of Lemma 3.5(3.7), or exhibit a counterexample; until then, Theorem 3.2 rests on an unverified imported estimate.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central proof of Theorem 3.2 requires Lemma 3.5(3.7), the uniform estimate Pη0(\\tilde{G}^Y_S | τ>t) ≈ 1. The proof of (3.7) is deferred with: 'the proof is the same as that of Lemma 2.8 in [AEGR15] which was proved for d=1.' No d≥2 proof is supplied. The key asserted inequality in that cited argument is Pη0(X∉⌊η0⌋, τ>t) ≤ (1−e^{−t})^{e^{2t}−1} ≤ e^{−e^t+1}, where ⌊η0⌋ is the first ⌊e^{2t}⌋ points of η0 in lexicographic order. This is a decreasing-event estimate: the first e^{2t} initial sites all fail to have descendants by time t. For the subcritical contact process, survival events are positively associated, so FKG gives P(all die) ≥ ∏(1−P(survive)), not an upper bound; the product-type exponential upper bound is therefore non-obvious and requires a real proof. In d=1 the lexicographic order is spatial and the interval structure is available; in d≥2 the first e^{2t} lexicographic points form a corner region and the argument does not automatically transfer. Since (3.7) is exactly what makes the conditioning on \\tilde{G}^S_Y, S<t/2 replace τ>t in the first and last ≈ of the Theorem 3.2 proof, failure of this estimate would break the factorization and the uniqueness conclusion.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proves that the subcritical contact process modulo translations on Z^d has a unique quasi-stationary distribution. The authors first recall the known existence of a Yaglom limit nu* for fixed initial configurations, then prove a uniform version of that limit (Theorem 3.2): for every fixed configuration zeta, the conditional probability P_{zeta0}(zeta_t = zeta | tau > t) is asymptotically equal to nu*(zeta) times the conditional probability that the diameter is smaller than R_t, uniformly in the initial configuration. Theorem 3.1 follows by taking zeta0 random with law equal to any QSD and letting t go to infinity. The novelty is that uniqueness holds even though the process does not come down from infinity, in contrast with stable queues and Galton-Watson processes.","tokens_in":7550,"tokens_out":17998,"duration_ms":196744,"significance":"If the proof is correct, this is the first example of a process with a unique quasi-stationary distribution that does not come down from infinity, and it establishes that spatial structure, not just the rate of coming down from infinity, controls uniqueness. The strategy is conceptually attractive: it avoids classification theorems and gives a direct factorization argument based on a cut break point. The paper also contains no parameter fitting or numerical work, and the main conclusion is a sharp, falsifiable mathematical statement. However, the proof is extremely terse at several load-bearing points, in particular the uniform cut-point estimate in Lemma 3.5, and the current text does not supply enough detail for the d >= 2 claim.","major_comments":[{"comment":"The proof of (3.7) is not supplied for d >= 2. The sentence \"the proof is the same as that of Lemma 2.8 in [AEGR15] which was proved for d = 1\" is not a proof, and the displayed estimate P_{eta0}(X not in floor(eta0), tau > t) <= (1 - e^{-t})^{e^{2t} - 1} is not an immediate consequence of the individual survival probabilities. For the contact process, the events that a given initial site has a descendant at time t are positively associated with each other, so the event that all of the first e^{2t} sites fail to have descendants is typically at least as likely as the product of the marginal failure probabilities, not at most. The one-dimensional proof can use the linear order of the first lexicographic sites; in d >= 2 those sites form a corner region and the same argument does not automatically transfer. Since (3.7) is used in the first and last approximations in the proof of Theorem 3.2, this gap is load-bearing and must be addressed with a complete proof or a precise reference covering d >= 2.","section":"Section 3, Lemma 3.5"},{"comment":"The cut break point is introduced as a random variable (Y,S) taking values in Z^d times the natural numbers, but the contact process evolves in continuous time and the first time that the infection from the minimal surviving site is localized at a single break point need not be an integer. The chain of equalities and approximations in the proof of Theorem 3.2 integrates over integer times s = 1, 2, ..., floor(t/2) and uses the identity H_{x,y,s} = {X=x, Y=y, S=s} intersected with the good event. No discretization argument or approximation by lattice times is provided. Without such an argument, the decomposition into disjoint regions and the factorization (star) are not justified for cut times that occur between integer times.","section":"Section 3, definition of (Y,S)"},{"comment":"In the chain following Lemma 3.5, the replacement of P_{eta0}(<eta_t> = zeta | (y,s) leads to L_t, G^s_y) by nu*(zeta) is stated to follow from (2.2) and Lemma 3.5, but the displayed uniform statement sup_{r >= t/2} |P_{0}(<eta_r> = zeta | (0,0) leads to L_r, G^0_0) - nu*(zeta)| approximately 0 is simply asserted. The conditioning event G^0_0 involves the whole time interval [0,r+t] and depends on the number of jumps before time t, whereas the Yaglom limit (2.2) concerns the unconstrained process conditioned on tau > r. A uniform version under the additional good-point event, with a time horizon longer than r, requires a separate argument. This step is responsible for producing the factor nu*(zeta), so it is load-bearing.","section":"Section 3, the 'other approximately-equal' step"}],"minor_comments":[{"comment":"There is a typo: 'there exits a QSD' should be 'there exists a QSD'.","section":"Section 2.2"},{"comment":"The sentence 'The product beta t means floor(beta t)' is confusing because beta t is later used as a radius in expressions such as B^y_{2R_t - beta t - 1}; please clarify whether the radius is an integer or a real number.","section":"Section 2.3"},{"comment":"The statement 'on the event H_{x,y,s}, the only sites that can be infected at time t are those in B^y_{beta t} and those in (B^y_{2R_t - beta t - 1})^c' is asserted without proof and is closely related to the conditioning issues in the major comments; it would help to spell out the argument in full.","section":"Section 3, proof of (star)="}],"recommendation":"major_revision","confidential_remarks":"The paper has a strong and attractive main idea, but the current version is too terse for the central d >= 2 claim. The proof of Lemma 3.5 is the main obstacle: it cites a d = 1 result and gives an inequality that is not transparent from positive association alone. I do not see this as a definite counterexample, but the burden is on the authors to supply a complete d >= 2 proof or to restrict the statement. The continuous-time discretization issue for the cut point should also be fixed before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this paper proves uniqueness of the QSD for the subcritical contact process modulo translations on Z^d, which is new and significant. The main theorem (Theorem 3.1) follows cleanly from the uniform limit in Theorem 3.2, and the proof structure is coherent. The novelty is real: prior work only got uniqueness among minimal QSDs or QSDs with finite mean, and here they get all QSDs. The remark that this is the first example of a process with a unique QSD that does not come down from infinity is a nice observation.\n\nWhat the paper does well: the cut-break-point method is adapted from their own previous work, and the factorization argument in Theorem 3.2 is explicit and mostly self-contained. Use of the good-point estimate is transparent. No fitting, no circularity; the QSD ν* is imported from existing Yaglom limit results and shown to be the only candidate.\n\nThe soft spot: Lemma 3.5, specifically (3.7), is load-bearing, and its proof is not really in the paper. The authors defer to Lemma 2.8 in [AEGR15], which was proved for d=1, and say 'the same' argument works in d≥2 using the lexicographic order. The key inequality is an upper bound on the probability that the first ⌊e^{2t}⌋ initial sites all die before time t. That is a decreasing event, and for a positively associated process like the contact process FKG gives a lower bound on this probability, not an upper bound. The product-type bound in the paper is non-obvious; it needs an actual proof that the lexicographic block in d≥2 behaves like an interval in d=1. The authors may have such a proof in mind, but it is not supplied. Since (3.7) is exactly what lets the conditioning on τ>t be replaced by the cut-break-point conditions, failure here would break the factorization and the uniqueness conclusion.\n\nIs the paper likely to be right? I think yes. The result is plausible, and the missing piece is a technical lemma that can probably be fixed. But the current preprint has a genuine gap in a central step, and I would not call it fully rigorous as written.\n\nWho this is for: probability folks working on QSDs and particle systems. It deserves a serious refereeing process; if the gap is filled, it is a solid publication. Recommendation: send it to review, and tell the referee to focus on Lemma 3.5.","headline":"A strong and likely correct result whose proof currently hides a load-bearing d≥2 extension of a d=1 lemma.","tokens_in":8109,"tokens_out":13654,"would_cite":true,"duration_ms":143203,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60K35","60J27"],"pacs":[],"model":"deepseek-v4-flash","headline":"The subcritical contact process modulo translations has exactly one quasi-stationary distribution, the Yaglom limit.","keywords":["quasi-stationary distribution","subcritical contact process","Yaglom limit","contact process modulo translations","cut break point","good point estimate","coming down from infinity","uniqueness"],"falsifier":"Simulate the subcritical contact process modulo translations in one dimension for a fixed subcritical λ, starting from a single infected site and from a large finite block; if the conditional distributions of the configuration given survival at large t do not converge to the same law, uniqueness fails. A more targeted check of the proof is to start from configurations whose diameter grows like $e^{\\sqrt t}$ and test whether the uniform approximation in Theorem 3.2 holds; any visible failure would falsify the claimed uniform Yaglom limit.","tokens_in":7040,"feed_emoji":"🦠","tokens_out":16491,"duration_ms":152077,"temperature":0.7,"pith_summary":"This paper proves that the subcritical contact process on $\\mathbb{Z}^d$, viewed modulo translations, has a unique quasi-stationary distribution (QSD). The unique law is the Yaglom limit $\\nu_*$ obtained by conditioning the process on survival from any fixed finite configuration. This settles a gap left by earlier partial results, which had only established uniqueness of a minimal QSD or uniqueness within the class of QSDs with finite expected number of infected sites. The result is notable because subcritical Galton-Watson processes, which have comparable negative drift, admit infinitely many QSDs, so uniqueness here is driven by spatial constraints rather than drift. The paper claims this is the first example of a process with a unique QSD that does not come down from infinity in finite time.","feed_headline":"Subcritical contact process has a unique survival law","feed_subtitle":"No matter how the infection starts, conditioning on survival leads to one fixed shape distribution.","key_machinery":"The argument is carried by cut break points combined with good points. A space-time point $(z,s)$ is good if every $\\lambda$-path starting there makes fewer than $\\beta t$ jumps over $[s, s+t]$, an event whose probability is at least $1 - e^{-\\rho t}$ for large $\\beta$; a cut break point $(Y,S)$ is the earliest time at which the infection emanating from the lexicographically first surviving initial site passes through a single site inside a large good boundary. The proof uses the cut break point to split the conditional law of the final configuration into two independent pieces: the inside region determines the rescaled cluster $C^y_{s,t}$, whose law approaches $\\nu_*$ by the Yaglom limit from a single site, and the disjoint outside region determines whether the diameter exceeds $R_t$. Factorizing over these independent space-time regions yields the uniform approximation in Theorem 3.2.","core_discovery":"The central claim is Theorem 3.1: the subcritical contact process modulo translations has exactly one quasi-stationary distribution. The proof establishes a stronger uniform Yaglom limit (Theorem 3.2): for $R_t = e^{\\sqrt t}$, every configuration $\\zeta$, and every initial configuration $\\zeta_0$, the conditional probability that $\\zeta_t = \\zeta$ given survival differs from $\\nu_*(\\zeta)$ times the conditional probability that $\\operatorname{diam}(\\zeta_t) < R_t$ given survival, and this difference vanishes uniformly as $t \\to \\infty$. Because a QSD $\\nu$ must satisfy $\\mathbb{P}_\\nu(\\zeta_t = \\zeta \\mid \\tau > t) = \\nu(\\zeta)$, substituting $\\nu$ into this uniform limit and letting $t$ tend to infinity forces $\\nu(\\zeta) = \\nu_*(\\zeta)$ for each $\\zeta$. Uniqueness follows.","pith_inferences":["A natural next test is whether the same uniqueness holds for other short-range spatial infection models and fails for long-range variants; the proof's reliance on good-point localization suggests the mechanism is robust to details but sensitive to long jumps.","Remark 3.3 indicates the theorem works for any diameter scale between linear and exponential in $t$ and fails outside that range, so the scale $e^{\\sqrt t}$ is not the mechanism itself; this points to a characteristic spatial scale of the conditioned process.","The contrast with Galton-Watson processes suggests that QSD uniqueness may hold exactly when spatial geometry destroys the extra infinite-mean QSDs; a contact process on a complete graph, which removes geometry, should therefore have multiple QSDs."],"forward_implications":["Every quasi-stationary distribution of the subcritical contact process modulo translations equals the Yaglom limit $\\nu_*$.","Since the QSD $\\nu_*$ satisfies $\\mathbb{P}_{\\nu_*}(\\tau > t) = e^{-\\alpha t}$ for a fixed $\\alpha > 0$, uniqueness implies every QSD has the same exponential absorption rate.","Uniqueness of the QSD is not determined by negative drift of the total number of infected sites: spatial structure alone can force uniqueness even when the process does not come down from infinity in finite time.","The uniform limit in Theorem 3.2 gives quantitative control: for large $t$, the conditional law of the configuration given survival is $\\nu_*$ times a diameter factor, uniformly over all starting configurations."],"supporting_citations":[{"why":"It supplies the good-point estimate (2.3), the Yaglom limit (2.2), and the cut-break-point construction that the proof extends to higher dimensions.","marker":"[AEGR15]"},{"why":"It constructs a QSD ν* with finite expected number of infected sites and identifies it with the Yaglom limit, the candidate whose uniqueness is proved.","marker":"[SS14]"},{"why":"It establishes a unique minimal QSD for a discrete-time version of the process, the partial uniqueness result that Theorem 3.1 completes.","marker":"[FKM96]"},{"why":"It generalizes the cut-break-point argument to higher dimensions, used in Lemma 3.5 and in the factorization step.","marker":"[DR17]"},{"why":"It shows that for birth-and-death processes QSD uniqueness is equivalent to coming down from infinity, the contrast that makes the contact process a first example.","marker":"[BMR16]"},{"why":"It gives the same uniqueness/coming-down-from-infinity equivalence for one-dimensional diffusions, framing the novelty of the present result.","marker":"[CCL+09]"}],"fun_headline_variants":["Subcritical contact process has exactly one survival law","Unique survival distribution for subcritical contact process","First unique quasi-stationary law among non-coming-down processes","Subcritical contact process: conditioning on survival yields unique law"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof rests on an estimate from prior work: for large parameters, with failure probability at most $e^{-\\rho t}$, a typical space-time point is 'good' in that no infection path from it makes more than $\\beta t$ jumps during a time interval of length $t$; if this estimate failed for the subcritical contact process, the localization and factorization steps would break.","fun_headline_variants_meta":{"raw":{"variants":["Subcritical contact process has exactly one survival law","Unique survival distribution for subcritical contact process","First unique quasi-stationary law among non-coming-down processes","Subcritical contact process: conditioning on survival yields unique law"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000353,"raw_usage":{"total_tokens":1829,"prompt_tokens":759,"completion_tokens":1070,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":375,"completion_tokens_details":{"reasoning_tokens":1006}},"tokens_in":375,"tokens_out":1070,"duration_ms":11173,"temperature":1.0,"reasoning_tokens":1006,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:48:47.283305+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the subcritical contact process modulo translations in one dimension for a fixed subcritical λ, starting from a single infected site and from a large finite block; if the conditional distributions of the configuration given survival at large t do not converge to the same law, uniqueness fails. A more targeted check of the proof is to start from configurations whose diameter grows like $e^{\\sqrt t}$ and test whether the uniform approximation in Theorem 3.2 holds; any visible failure would falsify the claimed uniform Yaglom limit.","supporting_citations":[],"review_version":1}