Abstract
In this note, we prove that the speed of convergence of the workload of a Lévy-driven queue to the quasi-stationary distribution is of order 1/t. We identify also the Laplace transform of the measure giving this speed and provide some examples.
Similar content being viewed by others
1 Introduction
In this paper, we consider a storage system with Lévy netput. In other words, the workload process \(\{Q(t), t\ge 0\}\) is a spectrally one-sided Lévy process X(t) that is reflected at 0:
where \({A^{-}}=\min \{A,0\}\). We assume that the drift of the process X(t) is negative; that is, we have \({\mathbb E}X(1)<0\). This stability condition guarantees the existence of a stationary distribution \(\pi \) of Q, which by virtue of “Reich’s identity” can be expressed in terms of the all-time supremum:
In the sequel, we consider the initial distribution Q(0) sampled from this steady-state distribution, which is indicated by adding the subscript \(\pi \) to the probability measure \({\mathbb P}\) and to the associated expectation \({\mathbb E}\).
Now, let T denote the busy period; that is,
We will further consider the Yaglom limit
where the convergence is to be understood in the weak sense. Yaglom [30] limits form a probability measure and a particular case of the quasi-stationary (QS) distribution, which is an invariant distribution for the process conditioned on non-extinction; that is, we condition on the event that the process survives some killing event (for example, related to exiting from some subset of possible values).
Yaglom [30] was the first to explicitly identify QS distributions for the subcritical Bienaymé–Galton–Watson branching process. This result has been generalized in the context of the continuous-time branching process and the Fleming–Viot process; see [1, 9, 18]. Similar results were also derived for Markov chains on positive integers with an absorbing state at the origin; see Seneta and Vere-Jones [27], Tweedie [29], Jacka and Roberts [13] and the bibliographic database of Pollet [25]. Recently, Foley and McDonald showed that Yaglom limits may depend on the starting state [10].
The research on QS distributions has been very extensive. Martinez and San Martin [22] analyze the Brownian motion with drift exiting from the positive half-line, complementing the result for random walks obtained by Iglehart [12]. In addition, QS laws have been studied for various Lévy processes. Kyprianou [15] found the Laplace transform of the QS distribution for the workload process of the stable M/G/1 queue with service times that have a rational moment generating function. Kyprianou and Palmowski [17] identified the QS distribution associated with a general light-tailed Lévy process. Haas and Rivero [26] found (after appropriate scaling) the QS distribution when the Lévy process under study has a jump measure with a regularly varying tail. The speed of convergence (in total variation) to the QS distribution for population processes has been studied in [5]. Finally, Mandjes et al. [21] derived the QS distribution of the workload process Q(t). This paper builds upon [21].
A contribution of this paper lies in proving that the speed of convergence to the quasi-stationary distribution is surprisingly slow (of order 1/t). We also identify a measure \(\xi (dx,dy)\) (which we call the second-order quasi-stationary measure) such that
where the above convergence is in the weak sense. We hence prove the conjecture posed in Polak and Rolski [24], which proved the above statement for a birth–death process by using an asymptotic expansion of a transition function and certain properties of Bessel functions. In this paper, we suggest a new method, which relies on a refined Tauberian-type expansion of the Laplace transform. We also analyze in detail the \(M/E(2,\nu )/1\) queue and a Brownian-driven queue.
If we want to simulate the quasi-stationary distribution directly from its definition given in (3), then the result stated in (3) shows that the speed of such a simulation is very slow. Still, in the literature there are papers giving other efficient algorithms of simulation of quasi-stationary measures; see, for example, Blanchet et al. [4] and references therein.
The main result in (3) contrasts the typical results derived for the regular stationary distribution of Markov processes where, in most of the cases, the rate is exponential. More precisely, for many models the distance between the distribution of the stochastic process at time t and its stationary distribution decays exponentially fast in t. The typical distances used are the total variation distance, the separation distance, and the \(L^2\) distance. The classical results concern mainly Markov chains and use Perron and Frobenius theory, renewal equations or the coupling method; see, for example, [6, 8, 14, 19, 20] and references therein. Another method concerns Harris recurrent Markov processes, and it is based on the construction of a special Lyapunov function and then the application of Foster-Lyapunov criteria; see, for example, [2, 23, 28]. All the above-mentioned methods, though, are different from the one used in this paper, which is based on expansions of Laplace transforms.
The paper is organized as follows: In next section, we introduce the notation and basic facts that are used later. In Sect. 3, we present the main results. The central step for the proof of the main results is given in Sect. 4. Finally, the last section provides some examples.
2 Preliminaries
We follow [16] for definitions, notation and basic facts on Lévy processes. Let \(X\equiv (X(t))_t\) be a Lévy process, which is defined on the filtered space \((\Omega ,\mathcal { F},\{\mathcal {F}_t\}_{t\ge 0}, {\mathbb P})\) with the natural filtration that satisfies the usual assumptions of right continuity and completion. We define \({\mathbb P}_x\) as \({\mathbb P}_x(X(0)=x)=1\) and \({\mathbb P}_0={\mathbb P}\); similarly, \({\mathbb E}_x\) is the expectation with respect to \({\mathbb P}_x\). We denote by \(\Pi (\cdot )\) the jump measure of X. If X is spectrally negative (resp. spectrally positive), then \(\Pi \) is supported in the non-positive half-line (resp. in the non-negative half-line); in other words, jumps are non-positive (resp. non-negative). We define the Laplace exponent \(\psi ({\eta })\) by
for \({\eta }\in \mathbb {R}\) such that the left-hand side of (4) is well defined (which holds at least for \(\eta \ge 0\)). We denote by \(\Phi (s):=\inf \{{\eta }\ge 0: \psi ({\eta })>s\}\) the right inverse of \(\psi \); see [16] for details.
Dual process We also consider the dual process \(\hat{X}_t=-X_t\) with jump measure \(\hat{\Pi }\left( 0,y\right) =\Pi \left( -y,0\right) \). The characteristics of \(\hat{X}\) are indicated by using the same symbols as for X, but with a ‘\(\hat{}\)’ added. We will write
We skip the symbol ‘\(\hat{}\)’ for Q(t) and hence for T and all quantities related to Q, as it will be clear from the context if a statement concerns the spectrally negative or the spectrally positive case.
Asymptotic expansions
Consider a function \(f:\mathbb {R}\rightarrow \mathbb {R}\) such that \(f(x)=0\) for \(x<0\). Let \(\tilde{f}(z):=\int _0^\infty e^{-z x}f(x)\,d x\) for \(z\in \mathbb {C}\) be its Laplace transform. Consider singularities of \(\tilde{f}(z)\); among these, let \(\mathfrak {R}(a_0)<0\) be the one with the largest real part. Notice that this yields the integrability of \(\int _0^\infty |f(x)|\, d x\). The inversion formula reads
for some (and then any) \(\mathfrak {R}(a)>\mathfrak {R}(a_0)\). In this paper, we need the following Tauberian theorem, found in Doetsch [7, Theorem 37.1], where the behaviour of the Laplace transform around the singularity \(a_0\) plays a crucial role.
First, recall the concept of the \(\mathfrak W\)-contour, centered at \(a_0\), with a half-angle of opening \(\pi /2<\psi \le \pi \), as depicted in [7, Fig. 30, p. 240]; also, for the purposes of our problem, \({\mathcal G}_{a_0}(\psi )\) is the region between the contour \(\mathfrak W\) and the line \(\mathfrak {R}(z)=0\). More precisely,
where \(\arg z\) is the principal part of the argument of the complex number z. In the following theorem, conditions are identified that provide an asymptotic expansion of the Laplace transform.
Theorem 1
([7, Theorem 37.1]) Suppose that for \(\tilde{f}:\mathbb {C}\rightarrow \mathbb {C}\) and \(\mathfrak {R}(a_0) < 0\) the following three conditions hold:
- (A1):
-
\(\tilde{f}(\cdot )\) is analytic in a region \({\mathcal G}_{a_0}(\psi )\) for some \(\pi /2<\psi \le \pi \);
- (A2):
-
\(\tilde{f}(z) \rightarrow 0\) as \(|z| \rightarrow \infty \) for \(z \in {\mathcal G}_{a_0}(\psi )\);
- (A3):
-
for some constants \(c_\nu \), \(\tilde{f}(s)\) has in \(|\mathrm {arg} (s - a_0)| < \psi \) the asymptotic expansion
$$\begin{aligned} \tilde{f}(s)\approx \sum _{\nu =0}^\infty c_\nu (s-a_0)^{\lambda _\nu }, \qquad (\mathfrak {R}(\lambda _0)<\mathfrak {R}(\lambda _1)<\ldots )\text {\ as\ }s\rightarrow a_0. \end{aligned}$$(6)
Then we conclude that, as \(t\rightarrow \infty \), f(t) has the asymptotic expansion
Assumptions For a spectrally negative Lévy process X(t), we impose the following assumptions:
(SN)
-
(SN1)
\(\psi ({\vartheta })\) attains its strictly negative minimum at \({\vartheta }{^*}>0\) (and hence \(\psi '({\vartheta }{^*})=0\));
-
(SN1)
\(\Phi \) is analytic in \(\mathcal {G}_{\zeta {^*}}(\phi )\) for \( \pi /2< \phi \le \pi \), where
$$\begin{aligned} \zeta {^*}:=\psi ({\vartheta }{^*})<0. \end{aligned}$$(7)
Similar conditions are assumed for a spectrally positive Lévy process X(t), for which \(\hat{X}_t\) is spectrally negative with Laplace exponent \(\hat{\psi }\) :
(SP)
-
(SP1)
\(\hat{\psi }({\vartheta })\) attains its strictly negative minimum at \({\vartheta }{^*}<0\) (and hence \(\hat{\psi }'({\vartheta }{^*})=0\));
-
(SP1)
\(\hat{\Phi }\) is analytic in \(\mathcal {G}_{\zeta {^*}}(\phi )\) for \( \pi /2< \phi \le \pi \), where
$$\begin{aligned} \zeta {^*}:=\hat{\psi }({\vartheta }{^*})<0. \end{aligned}$$(8)
To check the above assumptions, we can use the concept of semiexponentiality of a function f (see [11, p. 314]).
Definition 2
(Semiexponentiality) A function f is said to be semiexponential if, for some \(0< \phi \le \pi /2\), there exists a finite and strictly negative function \(\gamma ({\vartheta })\), called the indicator function, defined as the infimum of all \(a\in \mathbb {R}\) such that
for all sufficiently large r; here \(-\phi \le {\vartheta }\le \phi \) and \(\sup \gamma ({\vartheta })<0\).
It was proved in [21] using [11, Thm. 10.9f] that if there exists a density of \(\Pi \) (resp. \(\hat{\Pi }\)) which is of semiexponential type, then \(\Phi \) (resp. \(\hat{\Phi }\)) is analytic in \(\mathcal {G}_{\zeta {^*}}(\phi )\) for \( \pi /2< \phi \le \pi \). In particular, this assumption holds, for example, for a Brownian motion \(X(t)=\sigma B(t)-c t\) with linear drift, where \(c>0\) and B is the standard Brownian motion.
Quasi-stationary distribution Denote by
the bivariate Laplace transform of the quasi-stationary measure. The quasi-stationary distribution of the workload process Q(t) in the stationary regime was identified in [21].
Theorem 3
(Mandjes et al. [21])
(i) Under (SN),
(ii) Under (SP),
Note that conditions (SN) and (SP) are satisfied since we assumed that \({\mathbb E}X_1<0\). The two key components of the proof of this result are based first on Wiener–Hopf factorization, from which master formulas can be derived (given below), and second on either some expansion theorems (see [3] and [17]) or on some Tauberian-type theorems.
Master formulas Recall that Q(t) given in (1) is a workload process with stationary distribution (2) and busy period T. We define now the double Laplace–Stieltjes transform:
In [21], the following representations of L were derived.
Proposition 4
[Mandjes et al. [21]]
-
(i)
Under (SN),
$$\begin{aligned} L({\vartheta };\alpha ,\beta )= \frac{\Phi ({\vartheta })-\alpha -\Phi (0)}{\Phi ({\vartheta })+\beta }\frac{\Phi (0)}{\alpha +\beta +\Phi (0)}\frac{1}{{\vartheta }-\psi (\alpha +\Phi (0))}. \end{aligned}$$ -
(ii)
Under (SP),
$$\begin{aligned} L({\vartheta };\alpha ,\beta )=\frac{\hat{\psi }'(0+)}{{\vartheta }-\hat{\psi }(\beta )}\left( \frac{\alpha +\beta }{\hat{\psi }(\alpha +\beta )} - \frac{\alpha +\hat{\Phi }({\vartheta })}{\hat{\psi }(\alpha +\hat{\Phi }({\vartheta }))}\right) . \end{aligned}$$
From the proposition above, it follows that under assumptions (SN) or (SP) one can extend analytically \(L({\vartheta };\alpha ,\beta )\) into \({\mathcal G}_{\zeta {^*}}(\psi )\) for some \(\pi /2<\psi \le \pi \).
To prove the main result we will expand \(L(\vartheta , \alpha ,\beta )\) around \(\vartheta =\zeta ^*\) using Theorem 3. Then we will apply Theorem 1 to identify the expansion of \({\mathbb E}_\pi [e^{-\alpha Q(0)-\beta Q(t)}|T>t]\) for large values of t. In the last step we will use the definition of the quasi-stationary distribution.
3 Main results
We state now the main results of this paper. Define the constants
We start with the following expansion for the double Laplace–Stieltjes transform.
Proposition 5
If (SN) or (SP) hold, then
for \(\zeta ^*<0\) defined in (7) and (8).
- (i):
-
Under (SN),
$$\begin{aligned} C_0(\alpha ,\beta )&= -\frac{\Phi (0) }{\alpha +\beta +\Phi (0)} \frac{\alpha -\vartheta ^*+\Phi (0)}{\left( \beta +\vartheta ^*\right) \left( \zeta ^*-\psi (\alpha +\Phi (0))\right) },\\ C_1(\alpha ,\beta )&= \frac{A_1 \Phi (0)}{\left( \beta +\vartheta ^*\right) ^2 \left( \zeta ^*-\psi (\alpha +\Phi (0))\right) },\\ C_2(\alpha ,\beta )&= \frac{\Phi (0)\left( \frac{\left( \beta +\vartheta ^*\right) ^2 \left( \alpha -\vartheta ^*+\Phi (0)\right) }{\alpha +\beta +\Phi (0)}-\left( A_2 \left( \beta +\vartheta ^*\right) -A_1^2\right) \left( \psi (\alpha +\Phi (0))-\zeta ^*\right) \right) }{\left( \beta +\vartheta ^*\right) ^3\left( \zeta ^*-\psi (\alpha +\Phi (0))\right) ^2},\\ C_3(\alpha ,\beta )&=\frac{\Phi (0) A_3}{\left( \beta +\vartheta ^*\right) ^2\left( \zeta ^*-\psi (\alpha +\Phi (0))\right) } -\frac{\Phi (0) }{\left( \beta +\vartheta ^*\right) ^4\left( \zeta ^*-\psi (\alpha +\Phi (0))\right) ^2}\\&\quad \cdot \left( A_1 \left( \beta +\vartheta ^*\right) \left( 2 A_2 \left( \zeta ^*-\psi (\alpha +\Phi (0))\right) +\beta +\vartheta ^*\right) \right. \\&\qquad \qquad \left. +A_1^3\left( \psi (\alpha +\Phi (0))-\zeta ^*\right) \right) . \end{aligned}$$ - (ii):
-
Under (SP),
$$\begin{aligned} C_0(\alpha ,\beta )&= \left( \frac{\alpha +\beta }{\hat{\psi }(\alpha +\beta )}-\frac{\alpha +\vartheta ^*}{\hat{\psi }\left( \alpha +\vartheta ^*\right) }\right) \frac{\hat{\psi }'(0)}{\zeta ^*-\hat{\psi }\left( \beta \right) }, \\ C_1(\alpha ,\beta )&= -\hat{\psi }'(0)B_1\frac{ \hat{\psi }\left( \alpha +\vartheta ^*\right) -(\alpha +\vartheta ^*) \hat{\psi }'\left( \alpha +\vartheta ^*\right) }{\hat{\psi }\left( \alpha +\vartheta ^*\right) ^2 \left( \zeta ^*-\hat{\psi }(\beta )\right) }, \\ C_2(\alpha ,\beta )&=\frac{ \hat{\psi }'(0)}{\zeta ^*-\hat{\psi }(\beta )} \Bigg [ \frac{B_1^2 \hat{\psi }'\left( \alpha +\vartheta ^*\right) }{\hat{\psi }\left( \alpha +\vartheta ^*\right) ^2} -\frac{B_2}{\hat{\psi }\left( \alpha +\vartheta ^*\right) } -\frac{\frac{\alpha +\beta }{\hat{\psi }(\alpha +\beta )}-\frac{\alpha +\vartheta ^*}{\hat{\psi }\left( \alpha +\vartheta ^*\right) }}{\zeta ^*-\hat{\psi }(\beta )} \\&\quad +\frac{\left( \alpha +\vartheta ^*\right) \left( B_1^2 \left( \hat{\psi }\left( \alpha +\vartheta ^*\right) \hat{\psi }''\left( \alpha +\vartheta ^*\right) -2 \hat{\psi }'\left( \alpha +\vartheta ^*\right) ^2\right) +2 B_2 \hat{\psi }\left( \alpha +\vartheta ^*\right) \hat{\psi }'\left( \alpha +\vartheta ^*\right) \right) }{2 \hat{\psi }\left( \alpha +\vartheta ^*\right) ^3} \Bigg ], \end{aligned}$$
Theorem 6
If assumptions (SN) hold for a spectrally negative Lévy process X or (SP) hold for a spectrally positive Lévy process \(\hat{X}\), then the measure \(\xi (dx,dy)\) defined formally in (3) exists. That is, the speed of convergence to the quasi-stationary distribution is of order 1/t. Moreover, the Laplace transform \(\tilde{\xi } (\alpha , \beta ):=\int _0^\infty \int _0^\infty e^{-\alpha x}e^{-\beta y}\xi (dx,dy)\) equals
Proof
Recall that under the imposed assumptions (SN) or (SP), the Laplace exponent \(L({\vartheta };\alpha ,\beta )\) as a function of \(\vartheta \) satisfies the assumptions of Theorem 1. Hence, from Proposition 5 and Theorem 1 we have that, as \(t\rightarrow \infty \),
Further, note that by Theorem 3
Now, straightforward calculations give
which completes the proof due to the definition of the measure \(\xi \).
Remark 7
Formally starting from Proposition 4, then using a generalization of Proposition 5 in the next step, combined with Theorem 1, will give more terms of the expansion of \({\mathbb E}_\pi [e^{-\alpha Q(0)-\beta Q(t)}|T>t]\), and hence a longer expansion of the measure \({\mathbb P}_\pi (Q(0)\in dx, Q(t)\in dy|T>t)\) as \(t\rightarrow \infty \).
4 Proof of Proposition 5
Presume now that assumptions (SP) hold. We start from deriving the expansion of \(\hat{\Phi }(\vartheta )\):
as \(s\downarrow \zeta ^*\). Indeed, from a Taylor series expansion of \(\hat{\psi }\) around \(\vartheta ^*\) and the condition that \(\hat{\psi }'({\vartheta ^*})=0\), we have
From the above equation we can derive
The main idea of getting (14) is to include this expansion (14) into (13) and to match respective powers of \(\hat{\psi }({\vartheta })-\hat{\psi }({\vartheta ^*})\). Indeed, let us start from the first-order expansion
Then
Note that in the last equation \(o\left( \sqrt{\hat{\psi }({\vartheta })-\hat{\psi }({\vartheta ^*})}\right) = o(\vartheta -\vartheta ^*)\) by (15). Now, we continue with the second-order computation:
Then putting expression (16) in all powers of \({\vartheta }-{\vartheta ^*}\) in the increments on the right-hand side of the above equation, we can conclude that the last increment in (16) equals
where the last equality holds by observing that \(o\left( ({\vartheta }-{\vartheta ^*})^3\right) = o(\hat{\psi }({\vartheta })-\hat{\psi }({\vartheta ^*}))\), which follows from (15). Hence, plugging this back into (16) we derive
In the last step, we do the third-order computations required to identify \(o(\hat{\psi }({\vartheta })-\hat{\psi }({\vartheta ^*}))\) in (18) by again plugging (18) to (13). This produces (14). Substituting \({\vartheta }=\hat{\Phi }(s)\) and using \(\hat{\psi }(\hat{\Phi }(s))=s\) completes the proof of (12).
Having proven (12), we plug it into Proposition 4 and order the outcome according to powers of \(s-\zeta ^{*}\). This will complete the proof of the spectrally positive case (SP). Similarly, when (SN) holds then
as \(s\downarrow \zeta ^*\). Then, using Proposition 4 in the same way as before gives the required assertion after some simple manipulations. \(\square \)
5 Examples
In this section, we illustrate our theory through two examples.
Example 8
In this case
where \(\sigma _i\) (where \(i=1, 2, \ldots \)) are i.i.d. service times that have an Erlang\((2,\nu )\) distribution, i.e., with rate \(\nu \) and two phases. The arrival process is a homogeneous Poisson process N(t) with rate \(\lambda \). We assume that \(\varrho :=2\lambda /\nu <1\). For the Laplace exponent, we have that
which attains its minimum at
and it is equal to
One can easily check that all assumptions (SP) are satisfied. In particular, \(\hat{\Phi }(z)\) is analytic in \(\mathbb {C}\setminus (-\infty ,\zeta {^*}]\). Then, Proposition 5 gives for \(C_1(\alpha ,\beta )\) that
which is sufficient to compute the bivariate Laplace transform of the quasi-stationary measure given in (11). As a numerical illustration, if \(\lambda =1\) and \(\nu =4\), then
To identify the bivariate Laplace transform of the second-order quasi-stationary measure \(\xi \) given in Theorem 6, we need to find \(C_3(\alpha , \beta )\). Its expression is rather complex:
The expression is easy to evaluate numerically. Taking \(\lambda =1\) and \(\nu =4\) gives
where
and
Observe that the Laplace transforms \(\tilde{\mu }(\alpha ,\beta )\) and \(\tilde{\xi }(\alpha ,\beta )\) can be inverted as they can be written as linear combinations of powers of \(1/(\mathrm{const}_\alpha +\alpha )\) and \(1/(\mathrm{const}_\beta +\beta )\) for various constants \(\mathrm{const}_\alpha \) and \(\mathrm{const}_\beta \).
Example 9
In this case, \(X(t)= B(t)-t,\) where B(t) is a standard Brownian motion. Note that this process is spectrally positive and spectrally negative. We apply the spectrally positive results. It is not hard to check that
so that \(\vartheta ^*= -1\) and \(\zeta ^*=-1/2.\) Furthermore, assumptions (SP) are satisfied and
It is a matter of straightforward computations now to obtain that
and thus
In conclusion, the quasi-stationary distributions of Q(0) and Q(t) (conditioned that the busy period lasts longer than t, for large t) are both Erlang(2,1) with mean 2, whereas the stationary workload itself has an exponential distribution with mean 1/2; see [12, 21, 22]. Moreover, simple calculations lead to
Hence,
One can observe that the second-order quasi-stationary measure \(\xi \) is given by
It is worth noting that \(\xi \) may be negative (as is the case here), since it pertains to the speed of convergence to the Yaglom limit.
6 Conclusions
In this paper we proved that the speed of convergence of Yaglom limits is very slow, namely of order 1/t. We also identified the prefactor in front of this speed, called the second-order quasi-stationary law. Additionally, we analysed in detail two examples.
Future research directions lie in Markov additive processes or affine processes. It is also interesting to analyse the Yaglom limit for multivariate Lévy processes where instead of the busy period one considers the first exit time from a positive quadrant in \(\mathbb {R}^d\) for \(d>1\).
References
Asselah, A., Ferrari, P.A., Groisman, P., Jonckheere, M.: Fleming–Viot selects the minimal quasi-stationary distribution: the Galton–Watson case. Ann. Inst. Henri Poincaré Probab. Stat. 52(2), 647–668 (2016)
Asmussen, S.: Applied Probability and Queues, 2nd edn. Springer, Berlin (2003)
Bertoin, J., Doney, R.A.: Some asymptotic results for transient random walks. Adv. Appl. Probab. 28, 207–226 (1996)
Blanchet, J., Glynn, P., and Zheng, S.: Empirical analysis of a stochastic approximation approach for computing quasi-stationary distributions. In: EVOLVE—A Bridge Between Probability, Set Oriented Numerics, and Evolutionary Computation II, pp. 19–37, Springer (2013)
Champagnat, N., and Villemonais, D.: Population processes with unbounded extinction rate conditioned to non-extinction. arXiv:1611.03010v1 (2016)
Diaconis, P., Stroock, D.: Geometric bounds for eigenvalues of Markov chains. Ann. Probab. 1, 36–61 (1991)
Doetsch, G.: Introduction to the Theory and Application of the Laplace Transformation. Springer, Berlin (1974)
Down, D., Meyn, S.P., Tweedie, R.L.: Exponential and uniform ergodicity of Markov processes. Ann. Probab. 23(4), 1671–1691 (1995)
Ferrari, P.A., Marić, N.: Quasi stationary distributions and Fleming–Viot processes in countable spaces. Electron. J. Probab. 12(24), 684–702 (2007)
Foley, R.D., and McDonald, D.R.: Yaglom limits can depend on the starting state. arXiv:1709.07578 (2017)
Henrici, P.: Applied and Computational Complex Analysis, vol. 2. Wiley, New York (1977)
Iglehart, D.: Random walks with negative drift conditioned to stay positive. J. Appl. Probab. 11, 742–751 (1974)
Jacka, S., Roberts, G.: Weak convergence of conditioned processes on a countable state space. J. Appl. Probab. 32, 902–916 (1995)
Kingman, J.F.C.: Regenerative Phenomena. Wiley, New York (1972)
Kyprianou, E.: On the quasi-stationary distribution of the virtual waiting time in queues with Poisson arrivals. J. Appl. Probab. 8, 494–507 (1971)
Kyprianou, A.: Introductory Lectures on Fluctuations of Lévy Processes with Applications. Springer, Berlin (2006)
Kyprianou, A., Palmowski, Z.: Quasi-stationary distributions for Lévy processes. Bernoulli 12, 571–581 (2006)
Lambert, A.: Quasi-stationary distributions and the continuous-state branching process conditioned to be never extinct. Electron. J. Probab. 12(14), 420–446 (2007)
Lindvall, T.: Lectures on the Coupling Method. Wiley, New York (1992)
Lund, R.B., Meyn, S.P., Tweedie, R.L.: Computable exponential convergence rates for stochastically ordered Markov processes. Ann. Appl. Probab. 6(1), 218–237 (1996)
Mandjes, M., Palmowski, Z., Rolski, T.: Quasi-stationary workload in a Lévy-driven storage system. Stoch. Models 28(3), 413–32 (2012)
Martinez, S., San Martin, J.: Quasi-stationary distributions for a Brownian motion with drift and associated limit laws. J. Appl. Probab. 31, 911–920 (1994)
Meyn, S.P., Tweedie, R.L.: Markov Chains and Stochastic Stability. Communications and Control Engineering Series. Springer, London (1993)
Polak, M., Rolski, T.: A note on speed of convergence to the quasi-stationary distribution. Demonstratio Mathematica XLV(2), 384–397 (2012)
Pollett, P.: Quasi-stationary distributions: a bibliography. www.maths.uq.edu.au/~pkp/papers/qsds/qsds.pdf
Haas, B., Rivero, V.: Quasi-stationary distributions and Yaglom limits of self-similar Markov processes. Stoch. Process. Appl. 122(12), 4054–4095 (2012)
Seneta, E., Vere-Jones, D.: On quasi-stationary distributions in discrete-time Markov chains with a denumerable infinity of states. J. Appl. Probab. 3, 403–434 (1966)
Thorisson, H.: Coupling, Stationarity, and Regeneration. Springer, New York (2000)
Tweedie, R.: Quasi-stationary distributions for Markov chains on a general state space. J. Appl. Probab. 11, 726–741 (1974)
Yaglom, A.M.: Certain limit theorems of the theory of branching random processes. Doklady. Akad. Nauk. SSSR (N. S.) 56, 795–798 (1947)
Author information
Authors and Affiliations
Corresponding author
Additional information
Dedicated to Sergey Foss on the occasion of his 65th birthday.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
This work is partially supported by an NWO individual Grant through project 632.003.002 and by the National Science Centre under the grant 2018/B/ST1/00756 (2019-2022).
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Palmowski, Z., Vlasiou, M. Speed of convergence to the quasi-stationary distribution for Lévy input fluid queues. Queueing Syst 96, 153–167 (2020). https://doi.org/10.1007/s11134-020-09664-w
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11134-020-09664-w
Keywords
- Lévy processes
- Storage systems
- Quasi-stationary distribution
- Laplace transforms
- Fluctuation theory
- Speed of convergence