WorksMathematics
Domains of attraction of quasi-stationary distributions for an absorbed random walk
Abstract
We study a continuous-time random walk on the positive integers, killed at zero, where the downward rate and upward rate satisfy . Its quasi-stationary distributions form a one-parameter family. We show that if the initial distribution is strictly positive and satisfies , where , then its conditional distribution given survival converges in total variation to the corresponding nonminimal quasi-stationary distribution. We obtain asymptotic equivalents for the joint probabilities of position and survival, and for the survival probability itself. We also give a moment condition under which the minimal quasi-stationary distribution is selected. Finally, we construct an initial distribution whose successive-term ratios do not converge but whose conditional distribution nevertheless converges to a nonminimal quasi-stationary distribution. Thus the ratio condition is sufficient but not necessary.
Introduction
We consider a continuous-time random walk on the positive integers, killed at zero, with downward rate and upward rate , where . For this process, the full family of quasi-stationary distributions (QSDs) is known, as is the limiting conditional distribution from each fixed initial state. Seneta established the fixed-state limit in [8], while Seneta and Vere-Jones developed a general framework for countably infinite Markov chains with absorption [9]. Van Doorn identified the full family of QSDs for birth–death processes [10 Theorem 3.2] and extended the fixed-state convergence result to initial distributions with finite support. He also observed that treating initial distributions with infinite support would require conditions comparing their tails with that of the minimal QSD [10 p. 695]. Related questions for discrete-time birth–death processes were addressed by van Doorn and Schrijner [12], who connected limiting conditional distributions with ratio limits for transition probabilities through their spectral representation.
The purpose of this paper is to identify tail conditions under which an initial distribution with infinite support selects a specified member of the QSD family. Sufficiently light tails lead to selection of the minimal QSD, whereas heavier, asymptotically geometric tails can lead to selection of a nonminimal QSD. Our main result makes this precise. Put . If the initial distribution is strictly positive and
then the conditional distribution of the walk given survival converges in total variation to a corresponding nonminimal QSD . We also obtain asymptotic equivalents for the joint probabilities of position and survival, and for the survival probability itself. The theorem applies not only to geometric distributions, but also to geometric tails modified by polynomial, logarithmic or more general subexponential factors.
Analogous dependence on the initial distribution occurs in other absorbing processes. For Brownian motion with negative drift, Martı́nez, Picco and San Martín [4] showed that the rate at which the initial density decreases determines the limiting QSD, and constructed an example with different limits along different sequences of times. For subcritical branching processes, tail conditions on the initial population distribution similarly determine the limiting QSD [7, 3]. We are not aware of a corresponding geometric-tail domain-of-attraction result for the absorbed random walk considered here.
In earlier work, we developed a general framework for polynomial corrections to exponential decay in absorbing continuous-time Markov chains, using the same killed random walk as an example [6]. The questions addressed here are different, but the two studies use related Bessel-function representations of the transition probabilities of the killed process.
The ratio condition in our main theorem is sufficient but not necessary. We construct an initial distribution whose probabilities form a geometric sequence with parameter , multiplied by a periodic factor. Although its successive-term ratios do not converge, its conditional distribution converges to the nonminimal QSD . Thus, a complete domain-of-attraction criterion cannot be formulated solely in terms of convergence of successive-term ratios.
The paper is organised as follows. Section 2 introduces the model and its QSDs, and gives an exact calculation for a geometric initial distribution. Section 3 establishes the geometric-tail domain-of-attraction theorem. Section 4 treats lighter tails and discusses the critical case. Section 5 gives an example with oscillating successive-term ratios, and Section 6 derives a necessary survival-ratio condition for convergence to a nonminimal QSD. Section 7 summarises the resulting partial classification and discusses the remaining open problems.
The model and its quasi-stationary distributions
Let be the continuous-time random walk on the positive integers, killed when it reaches zero. From each state , the process jumps to at rate and to at rate , where . In queueing terminology, this is the M/M/1 queue with arrival rate and service rate , stopped when the queue first empties. Write for the irreducible class of transient states. Let be the absorption time. Write for the substochastic transition semigroup of the killed process, where
and write for its generator.
Seneta [8] proved that, for each fixed initial state ,
as , where . Van Doorn’s extension of the fixed-state result shows that the same conclusion holds for every initial distribution with finite support [10 Section 4]. This limiting conditional distribution is one of a family of quasi-stationary distributions. A probability distribution is called a quasi-stationary distribution (QSD) if
Standard theory (see, for example, [11]) shows that this is equivalent to being an -invariant probability distribution, that is, , , or, componentwise,
for some . Summing over gives , , which says that is the exponential decay rate of the survival probability when is taken as the initial distribution.
The QSDs for an absorbing birth–death process are described in [11 Theorem 16]. For the present constant-rate process, the equation becomes
with . It is easy to show that the normalized solutions are
where , , and
In particular, .
As , we have , so the two parameters coalesce at . At the same time, increases to the decay parameter of the class C, given by
while (1) converges componentwise to . Thus is the family member with the largest decay rate, . It is called the minimal QSD because it is stochastically smaller than every other QSD in the family.
All members of the family can be limiting conditional distributions for an appropriately chosen initial distribution ; we say that lies in the domain of attraction of the QSD if
as ; sometimes we say that selects . Thus the results of Seneta and van Doorn show that every initial distribution with finite support lies in the domain of attraction of the minimal QSD . Every QSD has a nonempty domain of attraction, since a QSD is invariant under conditioning on survival: in particular, itself selects . The substantive question is whether one can identify natural classes of initial distributions that select a given member of the family. We next show that geometric initial distributions provide such a class.
Let , , where . Since
a direct calculation gives
Both coefficients are positive and their sum is one, so this is a convex decomposition.
The positive sequence , , with , satisfies
The associated Doob -transformed process is a birth-death process with birth and death rates
respectively, and is honest. Hence [5 Theorem 3.1(ii)] . Thus, whenever ,
For , the sum on the right is finite because . Consequently,
On the other hand, a process started with distribution survives to time with probability , so the component in (2) contributes to the survival probability. Moreover, . Thus the contribution of the component to the survival probability is asymptotically negligible relative to that of the component. It follows that
in total variation and that
This exact calculation anticipates the general tail-ratio result of the next section.
A geometric-tail domain of attraction
The exact calculation above suggests that an initial distribution with an asymptotically geometric tail should select the same QSD as the corresponding geometric initial law. The next theorem gives the resulting domain-of-attraction criterion. The proof uses Seneta’s explicit formula [8 equation (2.2)] for the transition probabilities:
where and is the usual modified Bessel function of the first kind. It allows the effect of the tail of the initial distribution to be isolated.
Theorem 3.1 (Ratio-tail sufficiency). Let be a probability distribution such that and
Then
More precisely, if
then, for every fixed , and ,
and
In particular,
Proof. For convenience, put , . Since , we have , . The transition probabilities may therefore be written as
For , set . Then, for ,
and . We first determine the exponential order of . The ratio hypothesis implies . Choose sufficiently small that . Then the tail of is bounded above and below by constant multiples of and , respectively. The generating function of the sequence [1 9.6.33] is
Applying this identity with shows that the bilateral sums have exponential rates . Extending the original sums from to adds the finitely many terms with and the terms with . We now show that these additional terms have exponential rate at most . For each fixed index, the standard fixed-order asymptotic formula [1 equation (9.7.1)] shows that the associated term has exponential rate . Moreover, if , then
Finally,
By continuity, when is sufficiently small. Thus the additional terms are exponentially negligible and do not affect the rates . The geometric lower and upper bounds therefore give
Letting and using the continuity of gives
or equivalently, . Define a probability measure on by
For each fixed , the contribution of is , whereas . Hence for every fixed , so the mass of escapes to infinity.
For each fixed , the sequences and , with the latter restricted to , converge to and , respectively, and are therefore bounded. Thus, since the mass of escapes to infinity, their expectations under converge to these limits. For , splitting the sum defining gives , where
while
For each fixed ,
because as and the mass of escapes to infinity. Similarly,
The finite remainder terms are negligible even after summation over . Indeed, by Tonelli’s theorem, and using for each ,
Since and ,
The Bessel generating-function identity at gives . Consequently,
Since and , it follows that for every fixed . Consequently,
It remains to obtain summable bounds for and . Choose . There is an integer such that
Since for every , the finitely many ratios involving indices below can be absorbed into constants: there are constants such that
and
It follows that
and
Both geometric upper bounds are summable over . Dominated convergence, applied separately to and , together with the aggregate estimate for the finite remainder terms , therefore gives
Finally, since , the geometric bound on gives
The second term tends to zero by the aggregate estimate above. Taking the superior limit and then letting shows that the conditioned laws are tight. After division by the survival probability, the conditioned probabilities therefore converge pointwise to . Since both the conditioned probabilities and have total mass one, pointwise convergence implies convergence in total variation.
Remark 3.2. The proof also explains why the limiting QSD involves the difference of two geometric terms. In the killed transition formula, the first Bessel term contributes , whereas the second term, which enforces absorption at zero, contributes . Their difference is therefore , as in the formula for .
Remark 3.3. The theorem applies well beyond geometric initial distributions. For example, suppose that , , where is the normalizing constant and . Then , so, if , the initial distribution selects . This includes polynomial and logarithmic corrections, such as
as well as subexponential corrections of the form
A finite mixture of geometric distributions provides another example: if its largest geometric ratio is , then the mixture selects .
Lighter tails and the critical boundary
The theorem deals with , where the initial tail selects a nonminimal QSD. For initial distributions with sufficiently light tails, the conditional distribution instead converges to the minimal QSD .
Proposition 4.1. Let be a probability distribution satisfying
Then, for every fixed ,
and
Moreover,
In particular, these conclusions hold if and .
Proof. From the Bessel representation of the transition probabilities and the fixed-order large-argument expansion,
for each fixed . Consequently, . We also require a bound uniform in and . The integral representation for the modified Bessel functions [1 equation (9.6.19)] gives
Since , for , we have
where the last equality follows from [1 equations (9.6.19) and (9.6.26)]. Since is continuous on and has a finite limit as , it is bounded there. Hence, for ,
for a constant independent of , and . Let
Dominated convergence in the sum over gives for each , while . Since , a second application of dominated convergence gives
The same domination gives convergence in of the scaled unnormalized distributions. After normalization, their limit is
and hence
Finally, suppose that and . Choose . Then, for some , eventually, and therefore . The preceding conclusions now apply.
Remark 4.2. For the critical case the ratio condition alone does not determine whether is finite. For example, if , , where is the normalizing constant, then
Thus Proposition 4.1 applies when , and these critical-ratio initial distributions select . More generally, polynomial or slowly varying factors multiplying may determine whether the moment condition is satisfied.
The ratio condition is not necessary
Theorem 3.1 gives a simple sufficient condition, but the raw ratios can oscillate without affecting the limiting QSD. The next example isolates the oscillatory mode and shows that it decays too quickly to contribute to the conditioned limit.
Proposition 5.1. Fix and , and let
where
Then is a strictly positive probability distribution,
has two distinct subsequential limits, and
Proof. Put . For , define
If and , then belongs to . A direct substitution gives
with . Thus is an left eigenvector of , with eigenvalue . Since the killed generator is a bounded operator on , with
the semigroup is given by the operator-norm convergent series
Since this series converges in operator norm and , left multiplication by may be passed through the sum. The eigenvector relation implies, by induction, that
Consequently,
We shall use this identity for and . In these two cases , , and all four geometric sequences are summable because .
Since we may write
using and , gives the exact decomposition
where the signed remainder is given by
In particular,
For the two eigenvector components, we have
The alternating component is exponentially negligible relative to the positive component. Indeed,
whereas
Consequently, .
Finally, we need to control the remainder. Recall that satisfies . Since for (because ), positivity of the killed semigroup gives . Therefore
Because , the series converges, and hence
Also,
It follows that, in ,
The vector is positive, and
Summing the preceding equivalent over gives
Moreover,
Dividing the equivalent for by its total mass therefore yields
Finally,
Along even and odd values of , this ratio takes the two values
respectively. Hence the successive-term ratio does not converge, even though the conditioned distribution converges to . □
The example shows that an exact domain criterion cannot be formulated solely through convergence of the raw ratios . Such a criterion would have to allow for oscillatory components, such as the alternating component above, that become asymptotically negligible under evolution by the killed semigroup because their temporal decay rates are strictly greater than that of the positive -component. A smoothed-tail condition, a generating-function condition, or a condition expressing asymptotic dominance of the positive component may provide a more intrinsic formulation.
A necessary survival-ratio condition
If an initial distribution lies in the domain of attraction of , then the Markov property gives, for every fixed ,
Indeed, by the Markov property, the ratio is the conditional expectation of given survival to time . This function is bounded by one, so total-variation convergence of the conditioned laws allows passage to the limiting QSD, under which the survival probability over an additional time is . Hence
This is a genuine necessary condition, not a sufficient one; moreover, it is a condition on the evolved survival law rather than a complete description of the initial tail.
Discussion
Our results identify substantial subdomains of attraction for every QSD of the absorbed constant-rate random walk. Theorem 3.1 shows that for each a strictly positive initial distribution with tail ratio selects the nonminimal QSD . This condition covers geometric tails with polynomial, logarithmic or other ratio-slowly-varying corrections. The oscillatory example of Section 5 shows, however, that convergence of the tail ratio is sufficient but not necessary.
For the minimal QSD, Proposition 4.1 shows that is sufficient for convergence to the minimal QSD . This includes every initial distribution with tail ratio strictly less than , as well as part of the critical regime in which the tail ratio is . For example, the proposition applies to , when . Thus, for strictly positive initial distributions having a limiting tail ratio , our results give the following partial classification:
At the level of generality of the ratio hypothesis, the factor in Theorem 3.1 naturally retains subexponential information in the sequence ; without additional regularity, one should not expect a universal polynomial prefactor. By contrast, under the moment condition of Proposition 4.1, the survival probability has the explicit asymptotic form
A complete description of the domains of attraction remains open. There are two principal gaps. The first is the remainder of the critical regime: when but , the limiting conditional behaviour is not determined by the present results. The second is the problem of finding an intrinsic necessary-and-sufficient condition for selecting a given nonminimal QSD. The counterexample in Section 5 shows that convergence of to is not necessary for selection of . The remaining problem is to find an intrinsic condition on the initial distribution that is both necessary and sufficient for this selection. Generating-function singularities and exponentially smoothed tails are natural candidates for formulating such a condition.
A natural extension of the results presented here would be to birth–death processes whose birth and death rates are eventually, or asymptotically, constant. One expects the limiting rates to determine the relation between the tail ratio of the initial distribution and the selected QSD, subject to the spectral effects of the modified rates near the absorbing boundary. For eventually constant rates, the birth–death recurrence has geometric solutions outside a finite set, suggesting an approach based on transfer matrices or finite-rank perturbations of the constant-rate generator. For asymptotically constant rates, analogous results would require sufficiently uniform asymptotics for the recurrence solutions or the associated orthogonal polynomials [2 Chapter 8]. The main difficulty is that the exact Bessel generating identity used here is unavailable, so both the exponential scale of the relevant convolution and the uniform tail bounds needed for total-variation convergence would have to be obtained by other means.
Use of AI-assisted technologies. The author used Microsoft Copilot, a GPT-based generative AI tool, to assist with phrasing, exposition, and preliminary checks of intermediate calculations. The author takes full responsibility for the content of the paper.
References
- [1]M. Abramowitz and I. A. Stegun (eds.), Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables, National Bureau of Standards Applied Mathematics Series 55, U.S. Government Printing Office, Washington, DC, 1964.
- [2]W. J. Anderson, Continuous-Time Markov Chains: An Applications-Oriented Approach, Springer-Verlag, New York, 1991.
- [3]S. Asmussen and H. Hering, Branching Processes, Birkhäuser, Boston, 1983.
- [4]S. Martínez, P. Picco and J. San Martín, Domain of attraction of quasi-stationary distributions for the Brownian motion with drift, Advances in Applied Probability 30 (1998), 385–408.DOI
- [5]P. K. Pollett, Reversibility, invariance and μ-invariance, Advances in Applied Probability 20 (1988), 600–621.DOI
- [6]P. K. Pollett, Polynomial decay in absorbing continuous-time Markov chains, Advances in Applied Probability Accepted: 2026-11-12.arxiv.org/abs/2609.27335
- [7]H. Rubin and D. Vere-Jones, Domains of attraction for the subcritical Galton–Watson branching process, Journal of Applied Probability 5 (1968), 216–219.DOI
- [8]E. Seneta, Quasi-stationary behaviour in the random walk with continuous time, Australian Journal of Statistics 8 (1966), 92–98.DOI
- [9]E. Seneta and D. Vere-Jones, On quasi-stationary distributions in discrete-time Markov chains with a denumerable infinity of states, Journal of Applied Probability 3 (1966), 403–434.DOI
- [10]E. A. van Doorn, Quasi-stationary distributions and convergence to quasi-stationarity of birth–death processes, Advances in Applied Probability 23 (1991), 683–700.
- [11]E. A. van Doorn and P. K. Pollett, Quasi-stationary distributions for discrete-state models, European Journal of Operational Research 230 (2013), 1–14.DOI
- [12]E. A. van Doorn and P. Schrijner, Ratio limits and limiting conditional distributions for discrete-time birth–death processes, Journal of Mathematical Analysis and Applications 190 (1995), 263–284.