betaFeedback

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 aa and upward rate bb satisfy a>b>0a>b>0. Its quasi-stationary distributions form a one-parameter family. We show that if the initial distribution (mn)n≥1(m_n)_{n\geq 1} is strictly positive and satisfies mn+1/mn→rm_{n+1}/m_n\to r, where (b/a)1/2<r<1(b/a)^{1/2}<r<1, 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 aa and upward rate bb, where a>b>0a>b>0. 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 ρ=(b/a)1/2\rho=(b/a)^{1/2}. If the initial distribution m=(mn)n≥1m=(m_n)_{n\ge1} is strictly positive and

mn+1mn→r∈(ρ,1),\frac{m_{n+1}}{m_n}\to r\in(\rho,1),

then the conditional distribution of the walk given survival converges in total variation to a corresponding nonminimal QSD q(r)q^{(r)}. 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 rr, multiplied by a periodic factor. Although its successive-term ratios do not converge, its conditional distribution converges to the nonminimal QSD q(r)q^{(r)}. 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 (Xt)t≥0(X_t)_{t\geq0} be the continuous-time random walk on the positive integers, killed when it reaches zero. From each state i≥1i\geq1, the process jumps to i+1i+1 at rate bb and to i−1i-1 at rate aa, where a>b>0a>b>0. In queueing terminology, this is the M/M/1 queue with arrival rate bb and service rate aa, stopped when the queue first empties. Write C={1,2,…}C=\{1,2,\ldots\} for the irreducible class of transient states. Let T=inf⁡{t≥0:Xt=0}T=\inf\{t\geq0:X_t=0\} be the absorption time. Write Pt=(pij(t))i,j≥1P_t=(p_{ij}(t))_{i,j\geq1} for the substochastic transition semigroup of the killed process, where

pij(t)=Pi(Xt=j,T>t),p_{ij}(t)=\mathbb{P}_i(X_t=j,T>t),

and write QQ for its generator.

Seneta [8] proved that, for each fixed initial state i≥1i\geq1,

Pi(Xt=j∣T>t)→qj∗:=(1−ρ)2jρj−1,j≥1,\mathbb{P}_i(X_t=j\mid T>t)\to q_j^*:=(1-\rho)^2j\rho^{j-1},\qquad j\geq1,

as t→∞t\to\infty, where ρ=(b/a)1/2\rho=(b/a)^{1/2}. 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 q=(qj)j≥1q=(q_j)_{j\geq1} is called a quasi-stationary distribution (QSD) if

Pq(Xt=j∣T>t)=qj,j≥1,t≥0.\mathbb{P}_q(X_t=j\mid T>t)=q_j,\qquad j\geq1,\quad t\geq0.

Standard theory (see, for example, [11]) shows that this is equivalent to qq being an xx-invariant probability distribution, that is, qPt=e−xtqqP_t=e^{-xt}q, t≥0t\geq0, or, componentwise,

∑i≥1qipij(t)=e−xtqj,j≥1,t≥0,\sum_{i\geq1}q_i p_{ij}(t)=e^{-xt}q_j,\qquad j\geq1,\quad t\geq0,

for some x>0x>0. Summing over jj gives Pq(T>t)=e−xt\mathbb{P}_q(T>t)=e^{-xt}, t≥0t\geq0, which says that xx is the exponential decay rate of the survival probability when qq 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 qQ=−xqqQ=-xq becomes

bqj−1+aqj+1−(a+b)qj=−xqj,j≥1,bq_{j-1}+aq_{j+1}-(a+b)q_j=-xq_j,\qquad j\geq1,

with q0=0q_0=0. It is easy to show that the normalized solutions are

qj(r)=(1−r)(1−s)r−s(rj−sj),j≥1,(1)q_j^{(r)}=\frac{(1-r)(1-s)}{r-s}(r^j-s^j),\qquad j\geq1, \tag*{(1)}

where r∈(ρ,1)r\in(\rho,1), s=ρ2/rs=\rho^2/r, and

x(r)=a+b−ar−br.x(r)=a+b-ar-\frac{b}{r}.

In particular, 0<s<ρ<r<10<s<\rho<r<1.

As r↓ρr\downarrow\rho, we have s↑ρs\uparrow\rho, so the two parameters coalesce at ρ\rho. At the same time, x(r)x(r) increases to the decay parameter of the class C, given by

λC=x(ρ)=a+b−2ab=(a−b)2,\lambda_C=x(\rho)=a+b-2\sqrt{ab}=(\sqrt{a}-\sqrt{b})^2,

while (1) converges componentwise to q∗q^*. Thus q∗q^* is the family member with the largest decay rate, λC\lambda_C. 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 m=(mn)n≥1m=(m_n)_{n\geq1}; we say that mm lies in the domain of attraction of the QSD qq if

Lm(Xt∣T>t)→q\mathcal{L}_m(X_t\mid T>t)\to q

as t→∞t\to\infty; sometimes we say that mm selects qq. 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 q∗q^*. Every QSD has a nonempty domain of attraction, since a QSD is invariant under conditioning on survival: in particular, q(r)q^{(r)} itself selects q(r)q^{(r)}. 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 mj(r)=(1−r)rj−1m_j^{(r)}=(1-r)r^{j-1}, j≥1j\geq1, where r∈(ρ,1)r\in(\rho,1). Since

rj=rmj(r)/(1−r)andsj=smj(s)/(1−s),r^j=rm_j^{(r)}/(1-r)\qquad\text{and}\qquad s^j=sm_j^{(s)}/(1-s),

a direct calculation gives

m(r)=A(r)q(r)+B(r)m(s),A(r)=r−sr(1−s),B(r)=s(1−r)r(1−s).(2)m^{(r)}=A(r)q^{(r)}+B(r)m^{(s)},\qquad A(r)=\frac{r-s}{r(1-s)},\qquad B(r)=\frac{s(1-r)}{r(1-s)}. \tag*{(2)}

Both coefficients are positive and their sum is one, so this is a convex decomposition.

The positive sequence hi=iρ−ih_i=i\rho^{-i}, i≥1i\ge1, with h0=0h_0=0, satisfies

bhi+1+ahi−1−(a+b)hi=−λChi,i≥1.(3)b h_{i+1}+a h_{i-1}-(a+b)h_i=-\lambda_C h_i,\qquad i\ge1. \tag*{(3)}

The associated Doob hh-transformed process is a birth-death process with birth and death rates

ab i+1iandab i−1i,\sqrt{ab}\,\frac{i+1}{i}\qquad\text{and}\qquad\sqrt{ab}\,\frac{i-1}{i},

respectively, and is honest. Hence [5 Theorem 3.1(ii)] Pth=e−λCth\mathbb{P}_t h=e^{-\lambda_C t}h. Thus, whenever ∑iμihi<∞\sum_i\mu_i h_i<\infty,

Pμ(T>t)≤Eμ[h(Xt);T>t]=∑i≥1μi(Pth)i=e−λCt∑i≥1μihi.\mathbb{P}_{\mu}(T>t)\le\mathbb{E}_{\mu}[h(X_t);T>t]=\sum_{i\ge1}\mu_i(\mathbb{P}_t h)_i=e^{-\lambda_C t}\sum_{i\ge1}\mu_i h_i.

For μ=m(s)\mu=m^{(s)}, the sum on the right is finite because s<ρs<\rho. Consequently,

Pm(s)(T>t)=O(e−λCt).\mathbb{P}_{m^{(s)}}(T>t)=O(e^{-\lambda_C t}).

On the other hand, a process started with distribution q(r)q^{(r)} survives to time tt with probability e−x(r)te^{-x(r)t}, so the q(r)q^{(r)} component in (2) contributes A(r)e−x(r)tA(r)e^{-x(r)t} to the survival probability. Moreover, λC−x(r)=a(r−ρ)2/r>0\lambda_C-x(r)=a(r-\rho)^2/r>0. Thus the contribution of the m(s)m^{(s)} component to the survival probability is asymptotically negligible relative to that of the q(r)q^{(r)} component. It follows that

Lm(r)(Xt∣T>t)→q(r)\mathcal{L}_{m^{(r)}}(X_t\mid T>t)\to q^{(r)}

in total variation and that

Pm(r)(T>t)∼A(r)e−x(r)t.\mathbb{P}_{m^{(r)}}(T>t)\sim A(r)e^{-x(r)t}.

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:

pij(t)=e−(a+b)tρj−i{Ii−j(ct)−Ii+j(ct)},i,j≥1,p_{ij}(t)=e^{-(a+b)t}\rho^{j-i}\{I_{i-j}(ct)-I_{i+j}(ct)\},\qquad i,j\ge1,

where c=2abc=2\sqrt{ab} and InI_n 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 m=(mn)n≥1m=(m_n)_{n\ge1} be a probability distribution such that mn>0m_n>0 and

mn+1mn→r∈(ρ,1).\frac{m_{n+1}}{m_n}\to r\in(\rho,1).

Then

∥Lm(Xt∣T>t)−q(r)∥TV→0.\left\|\mathcal{L}_m(X_t\mid T>t)-q^{(r)}\right\|_{\mathrm{TV}}\to0.

More precisely, if

W(t)=e−(a+b)t∑n≥1mnρ−nIn(2abt),W(t)=e^{-(a+b)t}\sum_{n\ge1}m_n\rho^{-n}I_n(2\sqrt{ab}t),

then, for every fixed j≥1j \ge1, and t→∞t \to\infty,

Pm(Xt=j,T>t)∼(rj−sj)W(t),\mathbb{P}_{m}(X_t=j,T>t)\sim(r^j-s^j)W(t),

and

Pm(T>t)∼r−s(1−r)(1−s)W(t).\mathbb{P}_{m}(T>t)\sim\frac{r-s}{(1-r)(1-s)}W(t).

In particular,

lim⁡t→∞−1tlog⁡Pm(T>t)=x(r).\lim_{t\to\infty}-\frac{1}{t}\log\mathbb{P}_{m}(T>t)=x(r).

Proof. For convenience, put κn(t)=e−(a+b)tρ−nIn(ct)\kappa_n(t)=e^{-(a+b)t}\rho^{-n}I_n(ct), n∈Zn\in\mathbb{Z}. Since I−n=InI_{-n}=I_n, we have κ−n(t)=ρ2nκn(t)\kappa_{-n}(t)=\rho^{2n}\kappa_n(t), n≥0n\ge0. The transition probabilities may therefore be written as

pij(t)=κi−j(t)−ρ2jκi+j(t).p_{ij}(t)=\kappa_{i-j}(t)-\rho^{2j}\kappa_{i+j}(t).

For j∈Zj\in\mathbb{Z}, set wj(t)=∑i≥1miκi−j(t)w_j(t)=\sum_{i\ge1}m_i\kappa_{i-j}(t). Then, for j≥1j\ge1,

uj(t):=Pm(Xt=j,T>t)=wj(t)−ρ2jw−j(t),u_j(t):=\mathbb{P}_{m}(X_t=j,T>t)=w_j(t)-\rho^{2j}w_{-j}(t),

and w0(t)=∑i≥1miκi(t)=W(t)w_0(t)=\sum_{i\ge1}m_i\kappa_i(t)=W(t). We first determine the exponential order of WW. The ratio hypothesis implies log⁡mn=nlog⁡r+o(n)\log m_n=n\log r+o(n). Choose ε>0\varepsilon>0 sufficiently small that ρ<r−ε<r+ε<1\rho<r-\varepsilon<r+\varepsilon<1. Then the tail of mnm_n is bounded above and below by constant multiples of (r+ε)n(r+\varepsilon)^n and (r−ε)n(r-\varepsilon)^n, respectively. The generating function of the sequence (In)(I_n) [1 9.6.33] is

∑n∈ZIn(ct)yn=exp⁡(ct2(y+y−1)),y>0.\sum_{n\in\mathbb{Z}}I_n(ct)y^n=\exp\left(\frac{ct}{2}(y+y^{-1})\right),\qquad y>0.

Applying this identity with y=(r±ε)/ρy=(r\pm\varepsilon)/\rho shows that the bilateral sums have exponential rates −x(r±ε)-x(r\pm\varepsilon). Extending the original sums from n≥Nn\ge N to n∈Zn\in\mathbb{Z} adds the finitely many terms with 1≤n<N1\le n<N and the terms with n≤0n\le0. We now show that these additional terms have exponential rate at most −λC-\lambda_C. For each fixed index, the standard fixed-order asymptotic formula In(ct)∼ect/2πctI_n(ct)\sim e^{ct}/\sqrt{2\pi ct} [1 equation (9.7.1)] shows that the associated term has exponential rate −(a+b)+c=−λC-(a+b)+c=-\lambda_C. Moreover, if y=(r±ε)/ρ>1y=(r\pm\varepsilon)/\rho>1, then

e−(a+b)t∑n≤0In(ct)yn=e−(a+b)t∑k≥0Ik(ct)y−k≤e−(a+b)t∑k≥0Ik(ct)≤e−λCt.e^{-(a+b)t}\sum_{n\le0}I_n(ct)y^n=e^{-(a+b)t}\sum_{k\ge0}I_k(ct)y^{-k}\le e^{-(a+b)t}\sum_{k\ge0}I_k(ct)\le e^{-\lambda_Ct}.

Finally,

λC−x(r)=ar+br−2ab=(ar−br)2>0.\lambda_C-x(r)=ar+\frac{b}{r}-2\sqrt{ab}=\left(\sqrt{ar}-\sqrt{\frac{b}{r}}\right)^2>0.

By continuity, λC>x(r±ε)\lambda_C>x(r\pm\varepsilon) when ε\varepsilon is sufficiently small. Thus the additional terms are exponentially negligible and do not affect the rates −x(r±ε)-x(r\pm\varepsilon). The geometric lower and upper bounds therefore give

−x(r−ε)≤lim inf⁡t→∞1tlog⁡W(t)≤lim sup⁡t→∞1tlog⁡W(t)≤−x(r+ε).-x(r-\varepsilon)\le\liminf_{t\to\infty}\frac{1}{t}\log W(t)\le\limsup_{t\to\infty}\frac{1}{t}\log W(t)\le-x(r+\varepsilon).

Letting ε↓0\varepsilon\downarrow0 and using the continuity of xx gives

1tlog⁡W(t)→−x(r),\frac{1}{t}\log W(t)\to-x(r),

or equivalently, W(t)=e−x(r)t+o(t)W(t)=e^{-x(r)t+o(t)}. Define a probability measure μt\mu_t on N\mathbb{N} by

μt(n)=mnκn(t)W(t).\mu_t(n)=\frac{m_n\kappa_n(t)}{W(t)}.

For each fixed NN, the contribution of 1≤n≤N1\le n\le N is O(t−1/2e−λCt)O(t^{-1/2}e^{-\lambda_C t}), whereas W(t)=e−x(r)t+o(t)W(t)=e^{-x(r)t+o(t)}. Hence μt({1,…,N})→0\mu_t(\{1,\ldots,N\})\to0 for every fixed NN, so the mass of μt\mu_t escapes to infinity.

For each fixed jj, the sequences mn+j/mnm_{n+j}/m_n and mn−j/mnm_{n-j}/m_n, with the latter restricted to n≥j+1n\ge j+1, converge to rjr^j and r−jr^{-j}, respectively, and are therefore bounded. Thus, since the mass of μt\mu_t escapes to infinity, their expectations under μt\mu_t converge to these limits. For j≥1j\ge1, splitting the sum defining wj(t)w_j(t) gives wj(t)=Dj(t)+Bj(t)w_j(t)=D_j(t)+B_j(t), where

Dj(t)=∑n≥1mn+jκn(t),Bj(t)=∑k=0j−1mj−kκ−k(t).D_j(t)=\sum_{n\ge1}m_{n+j}\kappa_n(t),\qquad B_j(t)=\sum_{k=0}^{j-1}m_{j-k}\kappa_{-k}(t).

while

w−j(t)=∑n≥j+1mn−jκn(t).w_{-j}(t)=\sum_{n\ge j+1}m_{n-j}\kappa_n(t).

For each fixed j≥1j\ge1,

Dj(t)W(t)=∑n≥1mn+jmnμt(n)→rj,\frac{D_j(t)}{W(t)}=\sum_{n\ge1}\frac{m_{n+j}}{m_n}\mu_t(n)\to r^j,

because mn+j/mn→rjm_{n+j}/m_n\to r^j as n→∞n\to\infty and the mass of μt\mu_t escapes to infinity. Similarly,

ρ2jw−j(t)W(t)=ρ2j∑n≥j+1mn−jmnμt(n)→ρ2jr−j=sj.\frac{\rho^{2j}w_{-j}(t)}{W(t)}=\rho^{2j}\sum_{n\ge j+1}\frac{m_{n-j}}{m_n}\mu_t(n)\to\rho^{2j}r^{-j}=s^j.

The finite remainder terms Bj(t)B_j(t) are negligible even after summation over jj. Indeed, by Tonelli’s theorem, and using ∑j≥k+1mj−k=1\sum_{j\ge k+1}m_{j-k}=1 for each k≥0k\ge0,

∑j≥1Bj(t)=∑k≥0κ−k(t).\sum_{j\ge1}B_j(t)=\sum_{k\ge0}\kappa_{-k}(t).

Since 0<ρ<10<\rho<1 and Ik(ct)≥0I_k(ct)\ge0,

∑k≥0κ−k(t)=e−(a+b)t∑k≥0ρkIk(ct)≤e−(a+b)t∑k≥0Ik(ct).\sum_{k\ge0}\kappa_{-k}(t)=e^{-(a+b)t}\sum_{k\ge0}\rho^k I_k(ct)\le e^{-(a+b)t}\sum_{k\ge0}I_k(ct).

The Bessel generating-function identity at y=1y=1 gives ∑n∈ZIn(ct)=ect\sum_{n\in\mathbb{Z}}I_n(ct)=e^{ct}. Consequently,

∑j≥1Bj(t)=∑k≥0κ−k(t)≤e−(a+b)tect=e−λCt.\sum_{j\ge1}B_j(t)=\sum_{k\ge0}\kappa_{-k}(t)\le e^{-(a+b)t}e^{ct}=e^{-\lambda_Ct}.

Since W(t)=e−x(r)t+o(t)W(t)=e^{-x(r)t+o(t)} and λC>x(r)\lambda_C>x(r), it follows that Bj(t)/W(t)→0B_j(t)/W(t)\to0 for every fixed jj. Consequently,

uj(t)W(t)→rj−sj,j≥1.\frac{u_j(t)}{W(t)}\to r^j-s^j,\qquad j\ge1.

It remains to obtain summable bounds for Dj(t)D_j(t) and ρ2jw−j(t)\rho^{2j}w_{-j}(t). Choose ρ<η<r<R<1\rho<\eta<r<R<1. There is an integer NN such that

η≤mn+1mn≤R,n≥N.(4)\eta\le\frac{m_{n+1}}{m_n}\le R,\qquad n\ge N. \tag*{(4)}

Since mn>0m_n>0 for every n≥1n\ge1, the finitely many ratios involving indices below NN can be absorbed into constants: there are constants A1,A2<∞A_1,A_2<\infty such that

mn+j≤A1Rjmn,n,j≥1,m_{n+j}\le A_1R^j m_n,\qquad n,j\ge1,

and

mn−j≤A2η−jmn,j≥1,n≥j+1.m_{n-j}\le A_2\eta^{-j}m_n,\qquad j\ge1,\quad n\ge j+1.

It follows that

Dj(t)W(t)=∑n≥1mn+jmnμt(n)≤A1Rj\frac{D_j(t)}{W(t)}=\sum_{n\ge1}\frac{m_{n+j}}{m_n}\mu_t(n)\le A_1R^j

and

ρ2jw−j(t)W(t)=ρ2j∑n≥j+1mn−jmnμt(n)≤A2(ρ2η)j.\frac{\rho^{2j}w_{-j}(t)}{W(t)}=\rho^{2j}\sum_{n\ge j+1}\frac{m_{n-j}}{m_n}\mu_t(n)\le A_2\left(\frac{\rho^2}{\eta}\right)^j.

Both geometric upper bounds are summable over jj. Dominated convergence, applied separately to Dj(t)D_j(t) and ρ2jw−j(t)\rho^{2j}w_{-j}(t), together with the aggregate estimate for the finite remainder terms Bj(t)B_j(t), therefore gives

Pm(T>t)W(t)=∑j≥1Dj(t)W(t)+∑j≥1Bj(t)W(t)−∑j≥1ρ2jw−j(t)W(t)⟶∑j≥1rj−∑j≥1sj=r−s(1−r)(1−s).\frac{\mathbb{P}_m(T>t)}{W(t)} =\sum_{j\ge1}\frac{D_j(t)}{W(t)} +\sum_{j\ge1}\frac{B_j(t)}{W(t)} -\sum_{j\ge1}\frac{\rho^{2j}w_{-j}(t)}{W(t)} \longrightarrow\sum_{j\ge1}r^j-\sum_{j\ge1}s^j =\frac{r-s}{(1-r)(1-s)}.

Finally, since 0≤uj(t)≤Dj(t)+Bj(t)0\le u_j(t)\le D_j(t)+B_j(t), the geometric bound on Dj(t)D_j(t) gives

∑j>Juj(t)W(t)≤A1∑j>JRj+1W(t)∑j≥1Bj(t).\sum_{j>J}\frac{u_j(t)}{W(t)} \le A_1\sum_{j>J}R^j+\frac{1}{W(t)}\sum_{j\ge1}B_j(t).

The second term tends to zero by the aggregate estimate above. Taking the superior limit and then letting J→∞J\to\infty shows that the conditioned laws are tight. After division by the survival probability, the conditioned probabilities therefore converge pointwise to qj(r)q_j^{(r)}. Since both the conditioned probabilities and (qj(r))j≥1(q_j^{(r)})_{j\ge1} 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 rjr^j, whereas the second term, which enforces absorption at zero, contributes ρ2jr−j=sj\rho^{2j}r^{-j}=s^j. Their difference is therefore rj−sjr^j-s^j, as in the formula for qj(r)q_j^{(r)}.

Remark 3.3. The theorem applies well beyond geometric initial distributions. For example, suppose that mn=crnLnm_n=cr^nL_n, n≥1n\ge1, where cc is the normalizing constant and Ln+1/Ln→1L_{n+1}/L_n\to1. Then mn+1/mn→rm_{n+1}/m_n\to r, so, if ρ<r<1\rho<r<1, the initial distribution mm selects q(r)q^{(r)}. This includes polynomial and logarithmic corrections, such as

Ln=nβandLn=(log⁡(n+1))γ,L_n=n^\beta\qquad\text{and}\qquad L_n=(\log(n+1))^\gamma,

as well as subexponential corrections of the form

Ln=exp⁡(c0nδ),0<δ<1.L_n=\exp(c_0n^\delta),\qquad0<\delta<1.

A finite mixture of geometric distributions provides another example: if its largest geometric ratio is r∗∈(ρ,1)r_\ast\in(\rho,1), then the mixture selects q(r∗)q^{(r_\ast)}.

Lighter tails and the critical boundary

The theorem deals with r>ρr > \rho, where the initial tail selects a nonminimal QSD. For initial distributions with sufficiently light tails, the conditional distribution instead converges to the minimal QSD q∗q^*.

Proposition 4.1. Let m=(mn)n≥1m = (m_n)_{n\ge1} be a probability distribution satisfying

Hm:=∑n≥1mnnρ−n<∞.H_m := \sum_{n\ge1} m_n n \rho^{-n} < \infty.

Then, for every fixed j≥1j \ge1,

Pm(Xt=j,T>t)∼γHmjρjt−3/2e−λCt,γ=22π c3/2,\mathbb{P}_m(X_t = j,T > t) \sim\gamma H_m j \rho^j t^{-3/2}e^{-\lambda_C t}, \qquad\gamma= \frac{2}{\sqrt{2\pi}\,c^{3/2}},

and

Pm(T>t)∼γHmρ(1−ρ)2t−3/2e−λCt.\mathbb{P}_m(T > t) \sim\gamma H_m \frac{\rho}{(1-\rho)^2}t^{-3/2}e^{-\lambda_C t}.

Moreover,

∥Lm(Xt∣T>t)−q∗∥TV→0.\lVert\mathcal{L}_m(X_t \mid T > t) - q^* \rVert_{\mathrm{TV}} \to0.

In particular, these conclusions hold if mn>0m_n > 0 and mn+1/mn→r<ρm_{n+1}/m_n \to r < \rho.

Proof. From the Bessel representation of the transition probabilities and the fixed-order large-argument expansion,

Ii−j(z)−Ii+j(z)∼2ijez2π z3/2.I_{i-j}(z) - I_{i+j}(z) \sim\frac{2ij e^z}{\sqrt{2\pi}\,z^{3/2}}.

for each fixed i,j≥1i,j \ge1. Consequently, t3/2eλCtpij(t)→γiρ−ijρjt^{3/2}e^{\lambda_C t}p_{ij}(t) \to\gamma i\rho^{-i}j\rho^j. We also require a bound uniform in ii and jj. The integral representation for the modified Bessel functions [1 equation (9.6.19)] gives

Ii−j(z)−Ii+j(z)=2π∫0πezcos⁡θsin⁡(iθ)sin⁡(jθ) dθ.I_{i-j}(z) - I_{i+j}(z) = \frac{2}{\pi}\int_0^\pi e^{z\cos\theta}\sin(i\theta)\sin(j\theta)\,d\theta.

Since ∣sin⁡(nθ)∣≤nsin⁡θ|\sin(n\theta)| \le n\sin\theta, for 0≤θ≤π0 \le\theta\le\pi, we have

0≤Ii−j(z)−Ii+j(z)≤2ijπ∫0πezcos⁡θsin⁡2θ dθ=2ijI1(z)z,0 \le I_{i-j}(z) - I_{i+j}(z) \le\frac{2ij}{\pi}\int_0^\pi e^{z\cos\theta}\sin^2\theta\,d\theta= \frac{2ijI_1(z)}{z},

where the last equality follows from [1 equations (9.6.19) and (9.6.26)]. Since z1/2e−zI1(z)z^{1/2}e^{-z}I_1(z) is continuous on [c,∞)[c,\infty) and has a finite limit as z→∞z \to\infty, it is bounded there. Hence, for t≥1t \ge1,

t3/2eλCtpij(t)≤Aiρ−ijρjt^{3/2}e^{\lambda_C t}p_{ij}(t) \le A i\rho^{-i}j\rho^j

for a constant AA independent of ii, jj and tt. Let

uj(t)=Pm(Xt=j,T>t)=∑i≥1mipij(t).u_j(t) = \mathbb{P}_m(X_t = j,T > t) = \sum_{i\ge1}m_i p_{ij}(t).

Dominated convergence in the sum over ii gives t3/2eλCtuj(t)→γHmjρjt^{3/2}e^{\lambda_C t}u_j(t) \to\gamma H_mj\rho^j for each j≥1j \ge1, while t3/2eλCtuj(t)≤AHmjρjt^{3/2}e^{\lambda_C t}u_j(t) \le AH_mj\rho^j. Since ∑j≥1jρj=ρ/(1−ρ)2\sum_{j\ge1}j\rho^j = \rho/(1-\rho)^2, a second application of dominated convergence gives

Pm(T>t)∼γHmρ(1−ρ)2t−3/2e−λCt.\mathbb{P}_m(T > t) \sim\gamma H_m \frac{\rho}{(1-\rho)^2}t^{-3/2}e^{-\lambda_C t}.

The same domination gives convergence in ℓ1\ell^1 of the scaled unnormalized distributions. After normalization, their limit is

jρj∑k≥1kρk=(1−ρ)2jρj−1=qj∗,\frac{j\rho^j}{\sum_{k\ge1} k\rho^k}=(1-\rho)^2j\rho^{j-1}=q_j^*,

and hence

∥Lm(Xt∣T>t)−q∗∥TV→0.\left\|\mathcal{L}_m(X_t\mid T>t)-q^*\right\|_{\mathrm{TV}}\to0.

Finally, suppose that mn>0m_n>0 and mn+1/mn→r<ρm_{n+1}/m_n\to r<\rho. Choose r<r0<ρr<r_0<\rho. Then, for some A0<∞A_0<\infty, mn≤A0r0nm_n\le A_0r_0^n eventually, and therefore ∑n≥1mnnρ−n<∞\sum_{n\ge1}m_n n\rho^{-n}<\infty. The preceding conclusions now apply.

Remark 4.2. For the critical case mn+1/mn→ρm_{n+1}/m_n\to\rho the ratio condition alone does not determine whether ∑n≥1mnnρ−n\sum_{n\ge1}m_n n\rho^{-n} is finite. For example, if mn=cρnn−βm_n=c\rho^n n^{-\beta}, n≥1n\ge1, where cc is the normalizing constant, then

∑n≥1mnnρ−n=c∑n≥1n1−β.\sum_{n\ge1}m_n n\rho^{-n}=c\sum_{n\ge1}n^{1-\beta}.

Thus Proposition 4.1 applies when β>2\beta>2, and these critical-ratio initial distributions select q∗q^*. More generally, polynomial or slowly varying factors multiplying ρn\rho^n 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 r∈(ρ,1)r\in(\rho,1) and 0<ε<10<\varepsilon<1, and let

mj=Arj−1(1+ε(−1)j),j≥1,m_j=Ar^{j-1}\left(1+\varepsilon(-1)^j\right),\qquad j\ge1,

where

A−1=∑j≥1rj−1(1+ε(−1)j)=11−r−ε1+r.A^{-1}=\sum_{j\ge1}r^{j-1}\left(1+\varepsilon(-1)^j\right)=\frac{1}{1-r}-\frac{\varepsilon}{1+r}.

Then m=(mj)j≥1m=(m_j)_{j\ge1} is a strictly positive probability distribution,

mj+1mj=r1−ε(−1)j1+ε(−1)j\frac{m_{j+1}}{m_j}=r\frac{1-\varepsilon(-1)^j}{1+\varepsilon(-1)^j}

has two distinct subsequential limits, and

∥Lm(Xt∣T>t)−q(r)∥TV→0.\left\|\mathcal{L}_m(X_t\mid T>t)-q^{(r)}\right\|_{\mathrm{TV}}\to0.

Proof. Put s=ρ2/rs=\rho^2/r. For z≠0z\ne0, define

z†=ρ2z,Vj(z)=zj−(z†)j,x(z)=a+b−az−bz.z^\dagger=\frac{\rho^2}{z},\qquad V_j^{(z)}=z^j-(z^\dagger)^j,\qquad x(z)=a+b-az-\frac{b}{z}.

If ∣z∣<1|z|<1 and ∣z†∣<1|z^\dagger|<1, then V(z)V^{(z)} belongs to ℓ1\ell^1. A direct substitution gives

bVj−1(z)+aVj+1(z)−(a+b)Vj(z)=−x(z)Vj(z),j≥1,bV_{j-1}^{(z)}+aV_{j+1}^{(z)}-(a+b)V_j^{(z)}=-x(z)V_j^{(z)},\qquad j\ge1,

with V0(z)=0V^{(z)}_0=0. Thus V(z)V^{(z)} is an ℓ1\ell^1 left eigenvector of QQ, with eigenvalue −x(z)-x(z). Since the killed generator is a bounded operator on ℓ1\ell^1, with

∥vQ∥1≤2(a+b)∥v∥1,v∈ℓ1,\lVert vQ\rVert_1 \le2(a+b)\lVert v\rVert_1,\qquad v\in\ell^1,

the semigroup is given by the operator-norm convergent series

Pt=etQ=∑n=0∞tnQnn!.P_t=e^{tQ}=\sum_{n=0}^{\infty}\frac{t^nQ^n}{n!}.

Since this series converges in operator norm and V(z)∈ℓ1V^{(z)}\in\ell^1, left multiplication by V(z)V^{(z)} may be passed through the sum. The eigenvector relation implies, by induction, that

V(z)Qn=(−x(z))nV(z),n≥0.V^{(z)}Q^n=(-x(z))^nV^{(z)},\qquad n\ge0.

Consequently,

V(z)Pt=∑n=0∞tnn!V(z)Qn=∑n=0∞(−x(z)t)nn!V(z)=e−x(z)tV(z).V^{(z)}P_t=\sum_{n=0}^{\infty}\frac{t^n}{n!}V^{(z)}Q^n=\sum_{n=0}^{\infty}\frac{(-x(z)t)^n}{n!}V^{(z)}=e^{-x(z)t}V^{(z)}.

We shall use this identity for z=rz=r and z=−rz=-r. In these two cases r†=sr^\dagger=s, (−r)†=−s(-r)^\dagger=-s, and all four geometric sequences are summable because 0<s<ρ<r<10<s<\rho<r<1.

Since we may write

mj=Ar(rj+ε(−r)j),m_j=\frac{A}{r}\left(r^j+\varepsilon(-r)^j\right),

using rj=Vj(r)+sjr^j=V_j^{(r)}+s^j and (−r)j=Vj(−r)+(−s)j(-r)^j=V_j^{(-r)}+(-s)^j, gives the exact decomposition

m=ArV(r)+AεrV(−r)+D,m=\frac{A}{r}V^{(r)}+\frac{A\varepsilon}{r}V^{(-r)}+D,

where the signed remainder D=(dj)j≥1D=(d_j)_{j\ge1} is given by

dj=Ar(sj+ε(−s)j).d_j=\frac{A}{r}\left(s^j+\varepsilon(-s)^j\right).

In particular,

∣dj∣≤A(1+ε)rsj.\lvert d_j\rvert\le\frac{A(1+\varepsilon)}{r}s^j.

For the two eigenvector components, we have

mPt=Are−x(r)tV(r)+Aεre−x(−r)tV(−r)+DPt.mP_t=\frac{A}{r}e^{-x(r)t}V^{(r)}+\frac{A\varepsilon}{r}e^{-x(-r)t}V^{(-r)}+DP_t.

The alternating V(−r)V^{(-r)} component is exponentially negligible relative to the positive V(r)V^{(r)} component. Indeed,

x(−r)=a+b+ar+br>λC,x(-r)=a+b+ar+\frac{b}{r}>\lambda_C,

whereas

λC−x(r)=ar+br−2ab>0.\lambda_C-x(r)=ar+\frac{b}{r}-2\sqrt{ab}>0.

Consequently, x(−r)>λC>x(r)x(-r)>\lambda_C>x(r).

Finally, we need to control the remainder. Recall that hj=jρ−jh_j=j\rho^{-j} satisfies Pth=e−λCthP_t h=e^{-\lambda_C t}h. Since hj≥1h_j\ge1 for j≥1j\ge1 (because 0<ρ<10<\rho<1), positivity of the killed semigroup gives Pt1≤Pth=e−λCthP_t1\le P_th=e^{-\lambda_C t}h. Therefore

∥DPt∥1≤∑j≥1∣dj∣Pj(T>t)≤e−λCt∑j≥1∣dj∣hj≤A(1+ε)re−λCt∑j≥1j(sρ)j.\lVert DP_t\rVert_1\le\sum_{j\ge1}|d_j|P_j(T>t)\le e^{-\lambda_C t}\sum_{j\ge1}|d_j|h_j\le\frac{A(1+\varepsilon)}{r}e^{-\lambda_C t}\sum_{j\ge1}j\left(\frac{s}{\rho}\right)^j.

Because s<ρs<\rho, the series converges, and hence

∥DPt∥1=O(e−λCt)=o(e−x(r)t).\lVert DP_t\rVert_1=O(e^{-\lambda_C t})=o(e^{-x(r)t}).

Also,

∥Aεre−x(−r)tV(−r)∥1=o(e−x(r)t).\left\lVert\frac{A\varepsilon}{r}e^{-x(-r)t}V^{(-r)}\right\rVert_1=o(e^{-x(r)t}).

It follows that, in ℓ1\ell^1,

mPt=Are−x(r)tV(r)+o(e−x(r)t).mP_t=\frac{A}{r}e^{-x(r)t}V^{(r)}+o(e^{-x(r)t}).

The vector V(r)V^{(r)} is positive, and

∥V(r)∥1=∑j≥1(rj−sj)=r−s(1−r)(1−s).\lVert V^{(r)}\rVert_1=\sum_{j\ge1}(r^j-s^j)=\frac{r-s}{(1-r)(1-s)}.

Summing the preceding ℓ1\ell^1 equivalent over j≥1j\ge1 gives

Pm(T>t)=Are−x(r)tr−s(1−r)(1−s)+o(e−x(r)t).\mathbb{P}_m(T>t)=\frac{A}{r}e^{-x(r)t}\frac{r-s}{(1-r)(1-s)}+o(e^{-x(r)t}).

Moreover,

Vj(r)∥V(r)∥1=(1−r)(1−s)r−s(rj−sj)=qj(r).\frac{V_j^{(r)}}{\lVert V^{(r)}\rVert_1}=\frac{(1-r)(1-s)}{r-s}(r^j-s^j)=q_j^{(r)}.

Dividing the ℓ1\ell^1 equivalent for mPtmP_t by its total mass therefore yields

∥Lm(Xt∣T>t)−q(r)∥TV→0.\left\lVert\mathcal{L}_m(X_t\mid T>t)-q^{(r)}\right\rVert_{\mathrm{TV}}\to0.

Finally,

mj+1mj=r1−ε(−1)j1+ε(−1)j.\frac{m_{j+1}}{m_j}=r\frac{1-\varepsilon(-1)^j}{1+\varepsilon(-1)^j}.

Along even and odd values of jj, this ratio takes the two values

r1−ε1+εandr1+ε1−ε,r\frac{1-\varepsilon}{1+\varepsilon}\quad\text{and}\quad r\frac{1+\varepsilon}{1-\varepsilon},

respectively. Hence the successive-term ratio does not converge, even though the conditioned distribution converges to q(r)q^{(r)}. □

The example shows that an exact domain criterion cannot be formulated solely through convergence of the raw ratios mn+1/mnm_{n+1}/m_n. 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 rr-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 mm lies in the domain of attraction of q(r)q^{(r)}, then the Markov property gives, for every fixed u≥0u \ge0,

Pm(T>t+u)Pm(T>t)→e−x(r)u.\frac{\mathbb{P}_m(T > t+u)}{\mathbb{P}_m(T > t)} \to e^{-x^{(r)}u}.

Indeed, by the Markov property, the ratio is the conditional expectation of PXt(T>u)\mathbb{P}_{X_t}(T > u) given survival to time tt. 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 uu is e−x(r)ue^{-x^{(r)}u}. Hence

Pm(T>t)=e−x(r)tℓ(t),ℓ(t+u)ℓ(t)→1.\mathbb{P}_m(T > t)=e^{-x^{(r)}t}\ell(t), \qquad\frac{\ell(t+u)}{\ell(t)} \to1.

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 r∈(ρ,1)r \in(\rho,1) a strictly positive initial distribution with tail ratio rr selects the nonminimal QSD q(r)q^{(r)}. 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 ∑n≥1mnnρ−n<∞\sum_{n \ge1}m_n n\rho^{-n}<\infty is sufficient for convergence to the minimal QSD q∗q^*. This includes every initial distribution with tail ratio strictly less than ρ\rho, as well as part of the critical regime in which the tail ratio is ρ\rho. For example, the proposition applies to mn=cρnn−βm_n=c\rho^n n^{-\beta}, when β>2\beta>2. Thus, for strictly positive initial distributions having a limiting tail ratio r<1r<1, our results give the following partial classification:

{r<ρ:selection of the minimal QSD q∗,r=ρ:a critical regime not determined by the tail ratio alone,ρ<r<1:selection of the nonminimal QSD q(r).\begin{cases} r<\rho& : \text{selection of the minimal QSD }q^*,\\ r=\rho& : \text{a critical regime not determined by the tail ratio alone},\\ \rho<r<1 & : \text{selection of the nonminimal QSD }q^{(r)}. \end{cases}

At the level of generality of the ratio hypothesis, the factor W(t)W(t) in Theorem 3.1 naturally retains subexponential information in the sequence mn/rnm_n/r^n; 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

Pm(T>t)∼γHmρ(1−ρ)2t−3/2e−λCt,Hm=∑n≥1mnnρ−n.\mathbb{P}_m(T > t)\sim\gamma H_m\frac{\rho}{(1-\rho)^2}t^{-3/2}e^{-\lambda_Ct}, \qquad H_m=\sum_{n\ge1}m_n n\rho^{-n}.

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 mn+1/mn→ρm_{n+1}/m_n\to\rho but ∑n≥1mnnρ−n=∞\sum_{n\ge1}m_n n\rho^{-n}=\infty, 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 mn+1/mnm_{n+1}/m_n to rr is not necessary for selection of q(r)q^{(r)}. 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. [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. [2]W. J. Anderson, Continuous-Time Markov Chains: An Applications-Oriented Approach, Springer-Verlag, New York, 1991.
  3. [3]S. Asmussen and H. Hering, Branching Processes, Birkhäuser, Boston, 1983.
  4. [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. [5]P. K. Pollett, Reversibility, invariance and μ-invariance, Advances in Applied Probability 20 (1988), 600–621.DOI
  6. [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. [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. [8]E. Seneta, Quasi-stationary behaviour in the random walk with continuous time, Australian Journal of Statistics 8 (1966), 92–98.DOI
  9. [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. [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. [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. [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.

Paper details

vlnc:W_jhcr5ucg
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents