The quasi-stationary distribution of the subcritical contact process

Franco Arrejor\'ia; Leonardo T. Rolla; Pablo Groisman

REVIEW 3 major objections 3 minor 15 references

The quasi-stationary distribution of the subcritical contact process

T0 review · 3 major / 3 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The subcritical contact process modulo translations has exactly one quasi-stationary distribution, the Yaglom limit.

desk verdict A strong and likely correct result whose proof currently hides a load-bearing d≥2 extension of a d=1 lemma. read the letter →

arxiv 1908.04175 v1 pith:ACDHZHLH submitted 2019-08-12 math.PR

Franco Arrejoría , Pablo Groisman , Leonardo T. Rolla This is my paper · ORCID

classification math.PR MSC 60K3560J27

keywords quasi-stationarydistributionsubcriticalcontactprocessYaglomlimitmodulotranslationscutbreakpointgoodestimatecomingdownfrominfinityuniqueness

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

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.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

The 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.

What 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.

The 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.

Is 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.

Who 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.

Referee Report

3 major / 3 minor

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.

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 (3)

[Section 3, Lemma 3.5] 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 3, definition of (Y,S)] 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 3, the 'other approximately-equal' step] 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.

minor comments (3)

[Section 2.2] There is a typo: 'there exits a QSD' should be 'there exists a QSD'.
[Section 2.3] 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 3, proof of (star)=] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: uniqueness is deduced from a uniform Yaglom limit; self-citations are independent intermediate estimates.

full rationale

The derivation is not circular. The paper imports from [AEGR15, Prop. 3.2] the existence of a specific QSD ν* via the Yaglom limit (2.2) for each fixed initial configuration, then proves the stronger uniform version in Theorem 3.2, eq. (3.4): the conditional law of ζ_t given survival is asymptotically ν*(ζ) times the probability that the diameter is below R_t, uniformly in ζ0. For any QSD ν, the left side of (3.4) is exactly ν(ζ) by the definition of a quasi-stationary distribution, and the diameter factor is ν({diam < R_t}) → 1 because ν is a probability on Λ, the space of finite configurations modulo translations. Therefore ν(ζ) = ν*(ζ). The input (2.2) is convergence from arbitrary starting points to a known object, not uniqueness of QSDs; indeed the paper explicitly notes that partial uniqueness results do not imply full uniqueness, citing the Galton-Watson example with many QSDs sharing the same Yaglom limit. The good-point and cut-point estimates (2.3), (3.6), and (3.7) are imported from [AEGR15] and [DR17], which have overlapping authors with the present paper, but those results are intermediate probabilistic estimates proved independently and do not contain the target uniqueness theorem. The statement that the proof of (3.7) is 'the same as that of Lemma 2.8 in [AEGR15] which was proved for d=1' could indicate a proof gap in the d≥2 extension, but a gap is a correctness concern, not a circular reduction. No parameter is fitted and no prediction is renamed as an input, so the central claim has independent mathematical content.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The proof is a pure mathematical derivation. It postulates no free parameters and no invented entities. It relies on the standard graphical construction of the contact process, on the existence and convergence of a QSD established in [AEGR15], and on the good-point estimate from [AEGR15, DR17].

assumptions (4)

domain assumption The contact process has the Markov property and the quotient process modulo translations is absorbed at the empty configuration (Section 2.2).
Standard from the graphical representation; used throughout to define QSDs and absorption times.
domain assumption There exists a QSD ν* satisfying the Yaglom limit (2.2) for every fixed initial state and state ζ.
Quoted from [AEGR15, Proposition 3.2]. The proof assumes this ν* exists and then shows all QSDs equal it.
domain assumption Good-point estimate (2.3): P(G^0_0) ≥ 1 - e^{-ρt} for β and t large.
Quoted from [AEGR15, Lemma 2.10]. Used to ensure paths do not travel more than βt, a key step in the cut break point argument.
standard math Standard probability, Markov chain, and stochastic domination tools.
Used implicitly in the proof of Lemma 3.5 and the factorization arguments.

how reviews work

0 comments

Cite this review

Pith. "Pith review of The quasi-stationary distribution of the subcritical contact process." pith.science (2026). https://pith.science/paper/ACDHZHLH

@misc{pith2026190804175,
  author       = {Pith},
  title        = {Pith review of: The quasi-stationary distribution of the subcritical contact process},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ACDHZHLH}},
  note         = {Machine review of arXiv:1908.04175}
}

abstract

We show that the quasi-stationary distribution of the subcritical contact process on $\mathbb{Z}^d$ is unique. This is in contrast with other processes which also do not come down from infinity, like stable queues and Galton-Watson, and it seems to be the first such example.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 10 canonical work pages

[1]

Andjel , F

E. Andjel , F. Ezanno , P. Groisman , L. T. Rolla . Subcritical contact process seen from the edge: convergence to quasi-equilibrium. Electron J Probab 20:32, 2015. doi:10.1214/EJP.v20-3881

work page doi:10.1214/ejp.v20-3881 2015
[2]

Bansaye , S

V. Bansaye , S. M\' e l\' e ard , M. Richard . Speed of coming down from infinity for birth-and-death processes. Adv in Appl Probab 48:1183--1210, 2016. doi:10.1017/apr.2016.70

work page doi:10.1017/apr.2016.70 2016
[3]

Cattiaux , P

P. Cattiaux , P. Collet , A. Lambert , S. Mart\' nez , S. M\' e l\' e ard , J. San Mart\' n . Quasi-stationary distributions and diffusion models in population dynamics. Ann Probab 37:1926--1969, 2009. doi:10.1214/09-AOP451

work page doi:10.1214/09-aop451 1926
[4]

Collet , S

P. Collet , S. Mart \'i nez , J. S. Mart \'i n . Quasi-stationary Distributions: M arkov Chains, Diffusions and Dynamical Systems . Probability and its Applications (New York). Springer, Heidelberg, 2013. doi:10.1007/978-3-642-33131-2

doi:10.1007/978-3-642-33131-2 2013
[5]

Champagnat , D

N. Champagnat , D. Villemonais . Exponential convergence to quasi-stationary distribution and Q -process . Probab Theory Related Fields 164:243--283, 2016. doi:10.1007/s00440-014-0611-7

work page doi:10.1007/s00440-014-0611-7 2016
[6]

General criteria for the study of quasi-stationarity

--- --- ---. General criteria for the study of quasi-stationarity, 2017. Preprint. arXiv:1712.08092 http://arxiv.org/abs/1712.08092

work page Pith review arXiv 2017
[7]

Deshayes , L

A. Deshayes , L. T. Rolla . Scaling limit of subcritical contact process. Stochastic Process Appl 127:2630--2649, 2017. doi:10.1016/j.spa.2016.12.005

work page doi:10.1016/j.spa.2016.12.005 2017
[8]

F. Ezanno . Syst \`e mes de particules en interaction et mod \`e les de d \'e position al \'e atoire . Ph.D. thesis, Universit \'e d'Aix Marseille, 2012

work page 2012

Show all 15 references

[9]

P. A. Ferrari , H. Kesten , S. Mart \'i nez . R -positivity, quasi-stationary distributions and ratio limit theorems for a class of probabilistic automata . Ann Appl Probab 6:577--616, 1996. doi:10.1214/aoap/1034968146

1996
[10]

T. Kuczek . The central limit theorem for the right edge of supercritical oriented percolation. Ann Probab 17:1322--1332, 1989. doi:10.1214/aop/1176991157

1989
[11]

T. M. Liggett . Stochastic interacting systems: contact, voter and exclusion processes, vol. 324 of Grundlehren der Mathematischen Wissenschaften. Springer-Verlag, Berlin, 1999. doi:10.1007/978-3-662-03990-8

1999 doi
[12]

M \'e l \'e ard , D

S. M \'e l \'e ard , D. Villemonais . Quasi-stationary distributions and population processes. Probab Surveys 9:340--410, 2012. doi:10.1214/11-PS191

2012 doi
[13]

Sturm , J

A. Sturm , J. M. Swart . Subcritical contact processes seen from a typical infected site. Electron J Probab 19:53, 2014. doi:10.1214/EJP.v19-2904

2014 doi
[14]

Seneta , D

E. Seneta , D. Vere-Jones . On quasi-stationary distributions in discrete-time M arkov chains with a denumerable infinity of states . J Appl Probability 3:403--434, 1966. doi:10.2307/3212128

1966 doi
[15]

A. M. Yaglom . Certain limit theorems of the theory of branching random processes. Doklady Akad Nauk SSSR (NS) 56:795--798, 1947

1947

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Andjel , F

E. Andjel , F. Ezanno , P. Groisman , L. T. Rolla . Subcritical contact process seen from the edge: convergence to quasi-equilibrium. Electron J Probab 20:32, 2015. doi:10.1214/EJP.v20-3881

work page doi:10.1214/ejp.v20-3881 2015

[2] [2]

Bansaye , S

V. Bansaye , S. M\' e l\' e ard , M. Richard . Speed of coming down from infinity for birth-and-death processes. Adv in Appl Probab 48:1183--1210, 2016. doi:10.1017/apr.2016.70

work page doi:10.1017/apr.2016.70 2016

[3] [3]

Cattiaux , P

P. Cattiaux , P. Collet , A. Lambert , S. Mart\' nez , S. M\' e l\' e ard , J. San Mart\' n . Quasi-stationary distributions and diffusion models in population dynamics. Ann Probab 37:1926--1969, 2009. doi:10.1214/09-AOP451

work page doi:10.1214/09-aop451 1926

[4] [4]

Collet , S

P. Collet , S. Mart \'i nez , J. S. Mart \'i n . Quasi-stationary Distributions: M arkov Chains, Diffusions and Dynamical Systems . Probability and its Applications (New York). Springer, Heidelberg, 2013. doi:10.1007/978-3-642-33131-2

doi:10.1007/978-3-642-33131-2 2013

[5] [5]

Champagnat , D

N. Champagnat , D. Villemonais . Exponential convergence to quasi-stationary distribution and Q -process . Probab Theory Related Fields 164:243--283, 2016. doi:10.1007/s00440-014-0611-7

work page doi:10.1007/s00440-014-0611-7 2016

[6] [6]

General criteria for the study of quasi-stationarity

--- --- ---. General criteria for the study of quasi-stationarity, 2017. Preprint. arXiv:1712.08092 http://arxiv.org/abs/1712.08092

work page Pith review arXiv 2017

[7] [7]

Deshayes , L

A. Deshayes , L. T. Rolla . Scaling limit of subcritical contact process. Stochastic Process Appl 127:2630--2649, 2017. doi:10.1016/j.spa.2016.12.005

work page doi:10.1016/j.spa.2016.12.005 2017

[8] [8]

F. Ezanno . Syst \`e mes de particules en interaction et mod \`e les de d \'e position al \'e atoire . Ph.D. thesis, Universit \'e d'Aix Marseille, 2012

work page 2012