Hostname: page-component-cd9895bd7-dzt6s Total loading time: 0 Render date: 2024-12-22T22:02:37.539Z Has data issue: false hasContentIssue false

Moderate deviations of many-server queues via idempotent processes

Published online by Cambridge University Press:  20 December 2024

Anatolii Puhalskii*
Affiliation:
Institute for Problems in Information Transmission (IITP)
*
*Postal address: 19 B. Karetny, Moscow, Russia, 127051. Email address: puhalski@iitp.ru
Rights & Permissions [Opens in a new window]

Abstract

This paper obtains logarithmic asymptotics of moderate deviations of the stochastic process of the number of customers in a many-server queue with generally distributed inter-arrival and service times under a heavy-traffic scaling akin to the Halfin–Whitt regime. The deviation function is expressed in terms of the solution to a Fredholm equation of the second kind. A key element of the proof is the large-deviation principle in the scaling of moderate deviations for the sequential empirical process. The techniques of large-deviation convergence and idempotent processes are used extensively.

MSC classification

Type
Original Article
Copyright
© The Author(s), 2024. Published by Cambridge University Press on behalf of Applied Probability Trust

1. Introduction

Many-server queues are important in applications, but their analysis beyond Markovian assumptions is difficult; see, e.g., [Reference Asmussen2]. Various heavy-traffic asymptotics have been explored when the arrival and service rates tend to infinity. Of particular interest for applications is the set-up proposed in [Reference Halfin and Whitt8], where the service time distributions are held fixed, whereas the number of servers, n, and the arrival rate, $\lambda$ , grow without bound in such a way that $\sqrt{n}(1-\rho)\to\beta\in\mathbb R$ , with $\rho$ representing the traffic intensity: $\rho=\lambda/(n\mu)$ , where $\mu$ represents the reciprocal mean service time. With Q(t) denoting the number of customers present at time t, assuming the initial conditions are suitably chosen, in a fairly general situation the sequence of processes $(Q(t)-n)/\sqrt{n}$ , considered as random elements of the associated Skorokhod space, converges in law to a continuous-path process; see [Reference Aghajani and Ramanan1, Reference Halfin and Whitt8, Reference Kaspi and Ramanan11, Reference Puhalskii and Reed21, Reference Reed23]. Unless the service time distribution is exponential, the limit process is a process with memory, depends in an essential way on the service time cumulative distribution function (CDF), and is not well understood.

In order to gain additional insight, the paper [Reference Puhalskii20] proposed the study of moderate deviations of Q(t) and conjectured a large-deviation principle (LDP) for the process $(Q(t)-n)/(b_n\sqrt{n})$ under the heavy-traffic condition $ \sqrt{n}/b_n(1-\rho)\to\beta$ , where $b_n\to\infty$ and $b_n/\sqrt n\to0$ . (It has been observed that moderate-deviation asymptotics may capture exponents in the distributions of corresponding weak convergence limits; cf. [Reference Puhalskii18].) The deviation function (a.k.a. rate function) was purported to solve a convex variational problem with a quadratic objective function. In this paper we verify the conjecture and prove the LDP in question. Furthermore, we express the deviation function in terms of the solution to a Fredholm equation of the second kind, and we propose a framework for evaluating it numerically.

The proofs are arguably of methodological value, as they systematically use weak convergence methods and the machinery of idempotent processes. As in [Reference Puhalskii17, Reference Puhalskii19, Reference Pukhalskii22], the LDP is viewed as an analogue of weak convergence, the cornerstone of the approach being the following analogue of the celebrated tightness theorem of Prokhorov: a sequence of probability measures on a complete separable metric space is exponentially tight if and only if every subsequence of it admits a further subsequence that satisfies an LDP (Theorem (P) in [Reference Puhalskii, Shervashidze and De Gruyter16]). Consequently, once exponential tightness has been proved, the proof of the LDP is accomplished by proving the uniqueness of a subsequential large-deviation limit point. In order to take full advantage of weak convergence methods, it is convenient to recast the definition of the LDP for stochastic processes as large-deviation convergence (LD convergence) to idempotent processes; see [Reference Puhalskii19] and Appendix A for more detail. With tools for the study of weak convergence properties of many-server queues in heavy traffic being well developed, this paper derives the moderate-deviation asymptotics by using similar ideas. The main limit theorem asserts LD convergence of the process $(Q(t)-n)/(b_n\sqrt{n})$ to a certain idempotent process, which is analogous to the stochastic-process limit in [Reference Puhalskii and Reed21]. A key element of the proof is an LD limit for the sequential empirical process (see Lemma 1), a result that complements developments in [Reference Krichagina and Puhalskii12] and in [Reference Puhalskii and Reed21] and may be of interest in its own right. It identifies the limit idempotent process through finite-dimensional distributions. Whereas in weak convergence looking at second moments usually suffices to establish tightness, establishing the stronger property of exponential tightness calls for more intricate arguments and necessitates working with exponential martingales. In addition, a study of idempotent counterparts of the standard Wiener process, the Brownian bridge, and the Kiefer process is carried out. The properties of those idempotent processes are integral to the proofs.

The paper is organised as follows. Section 2 provides a precise specification of the model as well as the main result on the logarithmic asymptotics of moderate deviations of the number-in-the-system process. An added feature is the moderate-deviation asymptotics of the number of customers in an infinite-server queue in heavy traffic, which is also stated in the form of an LDP. The proofs of the LDPs in Section 2 are presented in Section 3. The techniques of LD convergence are employed. Section 4 is concerned with evaluating the deviation functions by reduction to solving Fredholm equations of the second kind. For the reader’s convenience, Appendix A gives a primer on idempotent processes and the use of weak convergence methods for proving LD convergence. Appendix B is concerned with the absolute continuity of the solution to a nonlinear renewal equation which is needed in Section 4.

2. Trajectorial moderate-deviation limit theorems

Assume as given a sequence of many-server queues with unlimited waiting room indexed by n, where n represents the number of servers. Service is performed on a first-come-first-served basis. If, upon a customer’s arrival, there are available servers, then the customer starts being served by one of the available servers, chosen arbitrarily. Otherwise, the customer joins the queue and awaits her turn to be served. When the service is complete, the customer leaves, relinquishing the server.

Let $Q_n(t)$ denote the number of customers present at time t. Of those customers, $Q_n(t)\wedge n $ customers are in service and $(Q_n(t)-n)^+$ customers are in the queue. The service times of the customers in the queue at time 0 and the service times of customers exogenously arriving after time 0 are denoted by $\eta_1,\eta_2,\ldots$ (in the order in which they enter service) and come from a sequence of independent and identically distributed (i.i.d.) positive unbounded random variables with continuous CDF F. It is thus assumed that

\[F(0)=0, F(x)<1, \text{ for all }x.\]

The mean service time $\mu^{-1}=\int_0^\infty x\,dF(x)$ is assumed to be finite. The residual service times of customers in service at time 0 are denoted by $\eta^{(0)}_1,\eta^{(0)}_2,\ldots$ and are assumed to be i.i.d. with CDF $F_0$ , which is the CDF of the delay in a stationary renewal process with inter-renewal CDF F. Thus,

(2.1) \begin{equation} F_0(x)=\mu\int_0^x(1-F(y))\,dy.\end{equation}

Let $A_n(t)$ denote the number of exogenous arrivals by time t, with $A_n(0)=0$ . It is assumed that the process $A_n(t)$ has unit jumps. The entities $Q_n(0)$ , $\{\eta^{(0)}_1,\eta^{(0)}_2,\ldots\}$ , $\{\eta_1,\eta_2,\ldots\}$ , and $A_n=(A_n(t),\,t\in\mathbb R_+)$ are assumed to be independent. All stochastic processes are assumed to have right-continuous paths with left-hand limits. Let ${{\hat A}_n}(t)$ denote the number of customers that enter service after time 0 and by time t, with ${{\hat A}_n}(0)=0$ . Since the random variables $\eta_i$ are continuous, the process $\hat A_n=(\hat A_n(t),t\in\mathbb R_+)$ has unit jumps almost surely. Balancing the arrivals and departures yields the equation

(2.2) \begin{equation} Q_n(t)=Q_n^{(0)}(t)+(Q_n(0)-n)^++A_n(t)-\int_0^t\int_0^t \mathbf{1}_{\{x+s\le t\}}\,d\,\sum_{i=1}^{{{\hat A}_n}(s)}\mathbf{1}_{\{\eta_i\le x\}},\end{equation}

where

(2.3) \begin{equation} Q_n^{(0)}(t)= \sum_{i=1}^{Q_n(0)\wedge n}\,\,\mathbf{1}_{\{\eta^{(0)}_i> t\}}\,,\end{equation}

which represents the number of customers present at time t out of those in service at time 0, and

\[ \int_0^t\int_0^t \mathbf{1}_{\{x+s\le t\}}\,d\,\sum_{i=1}^{{{\hat A}_n}(s)}\mathbf{1}_{\{\eta_i\le x\}}=\sum_{i=1}^{\hat A_n(t)}\,\mathbf{1}_{\{\eta_i+\hat\tau_{n,i}\le t\}}\,,\]

which represents the number of customers that enter service after time 0 and leave by time t, with $\hat\tau_{n,i}$ denoting the ith jump time of $\hat A_n$ , i.e., $ \hat\tau_{n,i}=\inf\{t:\,\hat A_n(t)\ge i\}.$ In addition, since each customer that is either in the queue at time 0 or has arrived exogenously by time t must either be in the queue at time t or have entered service by time t,

(2.4) \begin{equation} (Q_n(0)-n)^++A_n(t)=(Q_n(t)-n)^++{\hat A}_n(t).\end{equation}

For the existence and uniqueness of a solution to (2.2)–(2.4), the reader is referred to [Reference Puhalskii and Reed21].

Given $r_n\to\infty$ , as $n\to\infty$ , a sequence $\mathbb P_n$ of probability laws on the Borel $\sigma$ -algebra of a metric space M, and a $[0,\infty]$ -valued function I on M such that the sets $\{y\in M:\,I(y)\le \gamma\}$ are compact for all $\gamma\ge0$ , the sequence $\mathbb P_n$ is said to obey the LDP for rate $r_n$ with deviation function I, also referred to as a rate function, provided $ \lim_{n\to\infty}1/r_n\,\ln \mathbb P_n(W)=-\inf_{y\in W}I(y),$ for all Borel sets W such that the infima of I over the interior and the closure of W agree.

We now introduce the deviation function for the number-in-the-system process. For $T>0$ and $m\in\mathbb N$ , let $\mathbb D([0,T],\mathbb R^m)$ and $\mathbb D(\mathbb R_+,\mathbb R^m)$ represent the Skorokhod spaces of right-continuous $\mathbb R^m$ -valued functions with left-hand limits defined on [0, T] and $\mathbb R_+$ , respectively. These spaces are endowed with metrics rendering them complete separable metric spaces; see [Reference Ethier and Kurtz6, Reference Jacod and Shiryaev9] for more detail. Given $q=(q(t),t\in\mathbb R_+)\in\mathbb D(\mathbb R_+,\mathbb R)$ and $x_0\in\mathbb R$ , let

(2.5) \begin{equation} I_{x_0}^Q(q)=\frac{1}{2}\inf\bigg\{\int_0^1\dot w^0(x)^2\,dx+\int_0^\infty \dot w(t)^2\,dt+\int_0^\infty\int_0^1 \dot k(x,t)^2\,dx\,dt\bigg\}, \end{equation}

the infimum being taken over $w^0=(w^0(x),x\in[0,1])\in \mathbb D([0,1],\mathbb R)$ , $w=(w(t),t\in\mathbb R_+)\in\mathbb D(\mathbb R_+,\mathbb R)$ , and $k=((k(x,t),x\in[0,1]),t\in\mathbb R_+)\in\mathbb D(\mathbb R_+,\mathbb D([0,1],\mathbb R))$ such that $w^0(0)=w^0(1)=0$ , $w(0)=0$ , $k(x,0)=k(0,t)=k(1,t)=0$ ; $w^0$ , w, and k are absolutely continuous with respect to the Lebesgue measures on [0, 1], $\mathbb R_+$ , and $[0,1]\times \mathbb R_+$ , respectively; and, for all t,

(2.6) \begin{multline}q(t)=(1-F(t))x_0^+-(1-F_0(t))x_0^- -\beta F_0(t) +\int_0^tq(t-s)^+\,dF(s)+w^0(F_0(t))\\[5pt] +\int_0^t\bigl(1-F(t-s)\bigr)\sigma\,\dot w(s)\,ds+\int_{\mathbb R_+^2} \,\mathbf{1}_{\{x+s\le t\}}\,\,\dot k(F(x),\mu s)\,dF(x)\,\mu \,ds,\end{multline}

where $x_0^-=(\!-x_0)^+$ , and $\dot w^0(x)$ , $\dot w(t)$ and $\dot k(x,t)$ represent the respective Radon–Nikodym derivatives. If $w^0$ , w, and k as indicated do not exist, then $I_{x_0}^Q(q)=\infty$ . Note that $I_{x_0}^Q(q)=\infty$ unless $q(0)=x_0$ and q(t) is a continuous function, as the right-hand side of (2.6) is a continuous function of t. It is proved in Lemma 8 that if F is, in addition, absolutely continuous with respect to Lebesgue measure, then q(t) in (2.6) is absolutely continuous too. By Lemma B.1 in [Reference Puhalskii and Reed21], the equation (2.6) has a unique solution q(t) in the space of essentially locally bounded functions.

Let the process $X_n=(X_n(t),t\in\mathbb R_+)$ be defined by

(2.7) \begin{equation} X_n(t)=\frac{\sqrt{n}}{b_n}\bigg(\frac{Q_n(t)}{n}-1\bigg).\end{equation}

The next theorem verifies and refines Conjecture 1 in [Reference Puhalskii20]. Its proof is presented in Section 3.

Theorem 1. Suppose, in addition, that $A_n$ is a renewal process of rate $\lambda_n$ . Let $\rho_n=\lambda_n/(n\mu)$ , $\beta\in\mathbb R$ , $x_0\in\mathbb R$ , and $\sigma>0$ . Suppose that, as $n\to\infty$ ,

(2.8) \begin{equation} \frac{\sqrt{n}}{b_n}(1-\rho_n)\to\beta\end{equation}

and the sequence of random variables $\sqrt{n}/b_n\,(Q_n(0)/n-1)$ obeys the LDP in $\mathbb R$ for rate $b_n^2$ with deviation function $I_{x_0}(y)$ such that $I_{x_0}(x_0)=0$ and $I_{x_0}(y)=\infty$ , for $y\not=x_0$ . Suppose that the sequence of processes $\bigl((A_n(t)-\lambda_nt)/(b_n\sqrt{n}),t\in\mathbb R_+\bigr)$ obeys the LDP in $\mathbb D(\mathbb R_+,\mathbb R)$ for rate $b_n^2$ with deviation function $I^A(a)$ such that $I^A(a)=1/(2\sigma^2)\int_0^\infty\dot a(t)^2\,dt$ , provided $a=(a(t),t\in\mathbb R_+)$ is an absolutely continuous function with $a(0)=0$ , and $I^A(a)=\infty$ , otherwise. If, in addition,

(2.9) \begin{equation} b_n^6n^{1/b_n^2-1}\to0,\end{equation}

then the sequence $X_n$ obeys the LDP in $\mathbb D(\mathbb R_+,\mathbb R)$ for rate $b_n^2$ with deviation function $I_{x_0}^Q(q)$ .

Remark 1. In order that the LDP for $\bigl((A_n(t)-\lambda_nt)/(b_n\sqrt{n})\,t\ge0\bigr)$ in the statement hold, it suffices that $\mathbb E(n\xi_n)\to1/\mu$ , $\text{Var}(n\xi_n)\to \sigma^2/\mu^3$ , and that either $\text{sup}_n\mathbb E(n\xi_n)^{2+\epsilon}<\infty$ , for some $\epsilon>0$ , and $\sqrt{\ln n}/b_n\to\infty$ , or $\text{sup}_n\mathbb E\exp(\alpha (n\xi_n)^{\delta})<\infty$ and $n^{\delta/2}/b_n^{2-\delta}\to\infty$ , for some $\alpha>0$ and $\delta\in(0,1]$ , where $\xi_n$ represents a generic inter-arrival time for the nth queue; see [Reference Puhalskii18].

Remark 2. The condition (2.9) implies that $b_n^6/n\to 0$ , so that the condition that $b_n/\sqrt n\to 0$ necessarily holds. On the other hand, if $b_n^6/n^{1-\epsilon}\to0$ for some $\epsilon>0$ , then (2.9) holds.

As suggested by a referee, the next statement provides a version for the case of infinitely many servers. Consider a $GI/GI/\infty$ queue with renewal arrival process $A_n$ of rate $\lambda_n=n\lambda$ . All the assumptions and notation concerning the service times are the same as in Theorem 1. The arrival process, the initial number of customers, and the service times are independent. With $\overline Q_n(t)$ denoting the number of customers present at time t, the equations (2.2) and (2.3) are replaced with the respective equations

(2.10) \begin{equation} \overline{Q}_n(t)=\overline{Q}_n^{(0)}(t)+A_n(t)-\int_0^t\int_0^t \mathbf{1}_{\{x+s\le t\}}\,d\,\sum_{i=1}^{{A_n}(s)}\mathbf{1}_{\{\eta_i\le x\}}\end{equation}

and

(2.11) \begin{equation} \overline{Q}_n^{(0)}(t)= \sum_{i=1}^{\overline{Q}_n(0)}\,\,\mathbf{1}_{\{\eta^{(0)}_i> t\}}.\end{equation}

Given $q_0\in\mathbb R_+$ , let

(2.12) \begin{equation} \overline q(t)=q_0(1-F_0(t))+\lambda t-\lambda\int_0^t(t-s)\,dF(s)\end{equation}

and

(2.13) \begin{equation} \overline{X}_n(t)=\frac{\sqrt{n}}{b_n}\bigg(\frac{\overline Q_n(t)}{n}-\overline q(t)\bigg).\end{equation}

Theorem 2. Suppose that the sequence of processes $\bigl((A_n(t)-\lambda_nt)/(b_n\sqrt{n}),t\in\mathbb R_+\bigr)$ obeys the LDP in the hypotheses of Theorem 1. Given $x_0\in\mathbb R$ , suppose that the sequence $\overline{X}_n(0)$ obeys the LDP with deviation function $\overline I_{x_0}(y)$ such that $\overline I_{x_0}(x_0)=0$ and $\overline I_{x_0}(y)=\infty$ , for $y\not=x_0$ . If, in addition, (2.9) holds, then the sequence $\overline{X}_n$ obeys the LDP in $\mathbb D(\mathbb R_+,\mathbb R)$ for rate $b_n^2$ with deviation function $\overline I^Q_{q_0,x_0}(q)$ given by the right-hand side of (2.5), provided

\begin{multline*} q(t)=(1-F_0(t))x_0+\sqrt{q_0}\,w^0(F_0(t))+\int_0^t\bigl(1-F(t-s)\bigr)\sigma\,\dot w(s)\,ds\\[5pt] +\int_{\mathbb R_+^2} \,\mathbf{1}_{\{x+s\le t\}}\,\,\dot k(F(x),\lambda s)\,dF(x)\,\lambda \,ds,\end{multline*}

and equals $\infty$ otherwise.

Remark 3. The parameter $q_0$ arises as a law-of-large-numbers limit for the scaled initial number of customers. The corresponding parameter for the many-server queue in Theorem 1 equals 1.

3. Large-deviation convergence and proofs of Theorems 1 and 2

It is convenient to recast Theorem 1 as a statement on LD convergence. Introduce

(3.1) \begin{equation} Y_n(t)=\frac{\sqrt{n}}{b_n}\,\bigl(\frac{A_n(t)}{n}-\mu t\bigr)\end{equation}

and let $Y_n=(Y_n(t),\,t\in\mathbb R_+)$ . For the statement and proof of the next theorem, Appendix A is recommended reading.

Theorem 3. Suppose that, as $n\to\infty$ , the sequence $X_n(0)$ LD converges in distribution in $\mathbb R$ at rate $b_n^2$ to an idempotent variable X(0), the sequence $Y_n$ LD converges in distribution in $\mathbb D(\mathbb R_+,\mathbb R)$ at rate $b_n^2$ to an idempotent process Y with continuous paths, and (2.9) holds. Then the sequence $X_n$ LD converges in distribution in $\mathbb D(\mathbb R_+,\mathbb R)$ at rate $b_n^2$ to the idempotent process $X=(X(t),t\in\mathbb R_+)$ that is the unique solution to the equation

(3.2) \begin{multline} X(t)=(1-F(t))X(0)^+- (1-F_0(t))X(0)^-+\int_0^tX(t-s)^+\,dF(s)+W^0(F_0(t))\\[5pt] +Y(t)-\int_0^t Y(t-s)\,dF(s)+\int_0^t\int_{0}^t \,\mathbf{1}_{\{x+s\le t\}}\,\,\dot K( F(x),\mu s)\,dF(x)\,\mu\,ds, \end{multline}

where $W^0=(W^0(x),x\in[0,1])$ is a Brownian bridge idempotent process and $K=(K(x,t),(x,t)\in [0,1]\times\mathbb R_+)$ is a Kiefer idempotent process, X(0), Y, $W^0$ , and K being independent.

Theorem 1 is obtained as a special case. Suppose $A_n$ is a renewal process of rate $\lambda_n$ , the condition (2.8) holds, and the sequence $((A_n(t)-\lambda_nt)/(b_n\sqrt{n})\,,t\in\mathbb R_+)$ LD converges in distribution in $\mathbb D(\mathbb R_+,\mathbb R)$ at rate $b_n^2$ to $\sigma W$ , where $\sigma>0$ and $W=(W(t),t\in\mathbb R_+)$ is a standard Wiener idempotent process. Then, in the statement of Theorem 3, $Y(t)=\sigma W(t)-\beta\mu t$ , so that the limit idempotent process X solves the equation

\begin{multline*} X(t)=(1-F(t))X(0)^+-(1-F_0(t))X(0)^- -\beta F_0(t)+\int_0^tX(t-s)^+\,dF(s)\\[5pt] +W^0(F_0(t))+\int_0^t\bigl(1-F(t-s)\bigr)\sigma\,\dot W(s)\,ds+\int_0^t\int_{0}^t \,\mathbf{1}_{\{x+s\le t\}}\,\,\dot K(F(x),\mu s)\,dF(x)\,\mu \,ds, \end{multline*}

with X(0), W, $W^0$ , and K being independent. The assertion of Theorem 1 follows on observing that $\exp(\!-I^Q(y))$ , with $y\in\mathbb D(\mathbb R_+,\mathbb R)$ , is the deviability density of the idempotent distribution of X. To see the latter, note that the mapping $(w^0,w,k)\to q$ , as specified by (2.6), is continuous when restricted to the set $\{(w^0,w,k):\Pi^{W^0,W,K}(w^0,w,k)\ge a\}$ , where $\Pi^{W^0,W,K}(w^0,w,k)=\Pi^{W^0}(w^0)\Pi^{W}(w)\Pi^K(k)$ and $a\in(0,1]$ , so that X is strictly Luzin on $\bigl(\mathbb D([0,1],\mathbb R)\times\mathbb D(\mathbb R_+,\mathbb R)\times\mathbb D(\mathbb R_+,\mathbb D(\mathbb R_+,\mathbb R)),\Pi^{W^0,W,K}\bigr)$ ; see Appendix A for the definition and properties of being strictly Luzin. Therefore,

\[ \Pi^X(q)=\Pi^{W^0,W,K}(X=q)=\text{sup}_{(w^0,w,k):\,\eqref{eq:1}\text{ holds}}\Pi^{W^0}(w^0)\Pi^W(w)\Pi^K(k).\]

It is noteworthy that the limit idempotent process in (3.2) is analogous to the limit stochastic process on p. 139 in [Reference Puhalskii and Reed21].

The proof of Theorem 3 relies on an analogue of the weak convergence of the sequential empirical process to the Kiefer process; see, e.g., [Reference Krichagina and Puhalskii12]. Let random variables $\zeta_i$ be independent and uniform on [0, 1]. Define the centred and normalised sequential empirical process by

(3.3) \begin{align} K_n(x,t)&=\frac{1}{b_n\sqrt{n}}\,\sum_{i=1}^{\lfloor nt \rfloor}\bigl(\,\mathbf{1}_{\{\zeta_i\le x\}}\,-x\bigr),\end{align}

and let

(3.4) \begin{align}B_n(x,t)&=\frac{1}{b_n\sqrt n}\sum_{i=1}^{\lfloor nt\rfloor}\bigg(\,\mathbf{1}_{\{\zeta_i\le x\}}\,-\int_0^{x\wedge \zeta_i}\frac{dy}{1-y}\bigg),\end{align}

where $x\in[0,1]$ and $t\in\mathbb R_+$ . It is a simple matter to check that

(3.5) \begin{equation}K_n(x,t)=-\int_0^x\frac{K_n(y,t)}{1-y}\,dy+B_n(x,t).\end{equation}

Let $K_n=\bigl((K_n(x,t),\,x\in[0,1]),t\in\mathbb R_+\bigr)$ and $B_n=\bigl((B_n(x,t),\,x\in[0,1]),t\in\mathbb R_+\bigr)$ . Both processes are considered as random elements of $\mathbb D(\mathbb R_+,\mathbb D([0,1],\mathbb R))$ . Let $B=((B(x,t),x\in[0,1]),t\in\mathbb R_+)$ represent a Brownian sheet idempotent process, which is the canonical coordinate process on $\mathbb D(\mathbb R_+,\mathbb D([0,1],\mathbb R))$ , endowed with deviability $\Pi$ . Let $K=((K(x,t),x\in[0,1]),t\in\mathbb R_+)$ be defined as the solution of the equation

(3.6) \begin{equation} K(x,t)=-\int_0^x\frac{K(y,t)}{1-y}\,dy+B(x,t).\end{equation}

It is a Kiefer idempotent process by Lemma 7.

Lemma 1. Under (2.9), the sequence of stochastic processes $(K_n,B_n)$ LD converges at rate $b_n^2$ in $\mathbb D(\mathbb R_+,\mathbb D([0,1],\mathbb R)^2)$ to the idempotent process (K,B).

Proof. The proof draws on the proof of Lemma 3.1 in [Reference Krichagina and Puhalskii12]; see also [Reference Jacod and Shiryaev9, Chapter IX, §4c]. Given $t\in\mathbb R_+$ , let $K_n(t)=(K_n(x,t),\,x\in[0,1])\in\mathbb D([0,1],\mathbb R)$ , $B_n(t)=(B_n(x,t),\,x\in[0,1])\in\mathbb D([0,1],\mathbb R)$ , $K(t)=(K(x,t),\,x\in[0,1])\in\mathbb D([0,1],\mathbb R)$ , and $B(t)=(B(x,t),\,x\in[0,1])\in\mathbb D([0,1],\mathbb R).$ We prove first that, for $0\le t_1<t_2<\ldots<t_k$ , the sequence of $\mathbb D([0,1],\mathbb R)^{2k}$ -valued stochastic processes $\bigl((K_n(t_1),B_n(t_1)), ((K_n(t_2),B_n(t_2)),\ldots,(K_n(t_k),B_n(t_k))\bigr)$ LD converges to the $\mathbb D([0,1],\mathbb R)^{2k}$ -valued idempotent process $\bigl((K(t_1),B(t_1)),(K(t_2),B(t_2)),\ldots,(K(t_k),B(t_k))\bigr)$ in $\mathbb D([0,1],\mathbb R)^{2k}$ , as $n\to\infty$ . Since both the stochastic processes $((K_n(t),B_n(t)),t\in\mathbb R_+)$ and the idempotent process $((K(t),B(t)),t\in\mathbb R_+)$ have independent increments in t (see Lemma 7), it suffices to prove convergence of one-dimensional distributions, so we work with $((K_n(x,t),B_n(x,t)), x\in[0,1])$ and $((K(x,t),B(x,t)),x\in[0,1])$ , holding t fixed. By (3.4) and [Reference Jacod and Shiryaev9, Chapter II, §3c], the stochastic process $(B_n(x,t),x\in[0,1])$ is a martingale with respect to the natural filtration with the measure of jumps

\[ \mu^{n,B}([0,x],\Gamma)=\,\mathbf{1}_{\{1/(b_n\sqrt{n})\in\Gamma\}}\,\sum_{i=1}^{\lfloor nt\rfloor}\,\mathbf{1}_{\{\zeta_i\le x\}}\,,\]

the predictable measure of jumps

\[ \nu^{n,B}([0,x],\Gamma)=\,\mathbf{1}_{\{1/(b_n\sqrt{n})\in\Gamma\}}\,\sum_{i=1}^{\lfloor nt\rfloor}\int_0^{x\wedge \zeta_i}\frac{dy}{1-y},\]

and the predictable quadratic variation process

(3.7) \begin{multline} \langle B_n\rangle (x,t)=\int_0^x\int_\mathbb R u^2\nu^{n,B}(dy,du)=\frac{1}{b_n^2n}\,\nu^{n,B}([0,x],\{1/(b_n\sqrt{n})\})\\[5pt] =\frac{1}{b_n^2n}\,\sum_{i=1}^{\lfloor nt\rfloor}\int_0^{x\wedge \zeta_i}\frac{dy}{1-y}=\frac{\lfloor nt\rfloor}{b_n^2n}\,x+\frac{1}{b_n\sqrt{n}}\,K_n(x,t)-\frac{1}{b_n\sqrt{n}}\,B_n(x,t),\end{multline}

where $\Gamma\subset\mathbb R\setminus\{0\}$ .

We show next that

(3.8) \begin{equation} \lim_{r\to\infty}\limsup_{n\to\infty}\mathbb P(\text{sup}_{x\in[0,1]}\lvert B_n(x,t)\rvert>r)^{1/b_n^2}=0.\end{equation}

Since the process

\begin{multline*} \exp\bigl( b_n^2 B_n(x,t)-\int_0^x\int_\mathbb R(e^{ b_n^2 u}-1- b_n^2u)\nu^{n,B}(dy,du)\bigr)\\[5pt] = \exp\bigg( b_n^2 B_n(x,t)-(e^{ b_n/\sqrt{n}}-1- \frac{b_n}{\sqrt{n}})\sum_{i=1}^{\lfloor nt\rfloor}\int_0^{x\wedge \zeta_i}\frac{dy}{1-y}\bigg)\end{multline*}

is a local martingale with respect to x (see, e.g., [Reference Puhalskii19, Lemma 4.1.1, p. 294]), for any stopping time $\tau$ ,

\[ \mathbb E\exp\bigg( b_n^2 B_n(\tau,t)-(e^{ b_n/\sqrt{n}}-1- \frac{b_n}{\sqrt{n}})\sum_{i=1}^{\lfloor n t\rfloor}\int_0^{\tau\wedge \zeta_i}\frac{dy}{1-y}\bigg)\le 1.\]

Lemma 3.2.6 on p. 282 in [Reference Puhalskii19] implies that, for $r>0$ and $\gamma>0$ ,

\begin{multline*} \!\!\!\!\mathbb P(\text{sup}_{x\in[0,1]}e^{ b_n^2 B_n(x,t)}\ge e^{ b_n^2 r})\le e^{ b_n^2(\gamma- r)}+\mathbb P\bigg(\!\!\exp\!\bigg(\!\bigg(\!e^{ b_n/\sqrt{n}}-1- \frac{b_n}{\sqrt{n}}\!\bigg)\!\sum_{i=1}^{\lfloor n t\rfloor}\int_0^{ \zeta_i}\frac{dy}{1-y}\!\bigg)\ge e^{ b_n^2\gamma}\!\bigg)\\[5pt] \le e^{ b_n^2(\gamma- r)}+ e^{- b_n^2\gamma}\mathbb E\bigg(\!\!\exp\bigg(\bigg(e^{ b_n/\sqrt{n}}-1- \frac{b_n}{\sqrt{n}}\bigg)\sum_{i=1}^{\lfloor n t\rfloor}\int_0^{ \zeta_i}\frac{dy}{1-y}\bigg)\bigg)\\[5pt] =e^{ b_n^2(\gamma- r)}+ e^{- b_n^2\gamma}\big(1-(e^{ b_n/\sqrt{n}}-1- b_n/\sqrt{n})\big)^{-\lfloor nt\rfloor},\end{multline*}

with the latter equality holding for all n large enough because $e^{ b_n/\sqrt{n}}-1- b_n/\sqrt{n}\to0$ . Hence, assuming that $e^{ b_n/\sqrt{n}}-1- b_n/\sqrt{n}\le1/2$ , we have

\[ \mathbb P(\text{sup}_{x\in[0,1]}e^{ b_n^2 B_n(x,t)}\ge e^{ b_n^2 r})^{1/b_n^2}\le e^{ \gamma- r}+e^{- \gamma}2^{\lfloor nt\rfloor/b_n^2}.\]

Since $n/b_n^2\to\infty$ , it follows that

(3.9) \begin{equation} \lim_{ r\to\infty}\limsup_{n\to\infty} \mathbb P(\text{sup}_{x\in[0,1]} B_n(x,t)\ge r)^{1/b_n^2}=0.\end{equation}

A similar convergence holds with $-B_n(x,t)$ substituted for $B_n(x,t)$ . The limit (3.8) has been proved.

We next prove that, similarly,

(3.10) \begin{equation} \lim_{r\to\infty}\limsup_{n\to\infty}\mathbb P(\text{sup}_{x\in[0,1]}\lvert K_n(x,t)\rvert>r)^{1/b_n^2}=0.\end{equation}

Since, by (3.3), $(K_n(x,t),x\in[0,1])$ is distributed as $(\!-K_n(1-x,t),x\in[0,1])$ , it suffices to prove that

(3.11) \begin{align} \lim_{r\to\infty}\limsup_{n\to\infty}\mathbb P(\text{sup}_{x\in[0,1/2]}\lvert K_n(x,t)\rvert>r)^{1/b_n^2}=0.\end{align}

By (3.5) and the Gronwall–Bellman inequality,

(3.12) \begin{equation} \text{sup}_{x\in[0,1/2]}\lvert K_n(x,t)\rvert\le e^2\text{sup}_{x\in[0,1/2]}\lvert B_n(x,t)\rvert,\end{equation}

so that (3.11) follows from (3.9).

By (3.7), (3.8), and (3.10), for $x\in[0,1]$ ,

(3.13) \begin{equation} \lim_{n\to\infty}\mathbb P(\lvert b_n^2\langle B_n\rangle (x,t)-tx\rvert>\epsilon)^{1/b_n^2}=0.\end{equation}

If we extend it past $x=1$ by letting $B_n(x,t)=B_n(1,t)$ , the process $(B_n(x,t),x\in\mathbb R_+)$ is a square-integrable martingale with predictable quadratic variation process $(\langle B_n\rangle(x\wedge1,t), x\in\mathbb R_+)$ , so, by (3.13) and Theorem 5.4.4 on p. 423 in [Reference Puhalskii19], where one takes $\beta_\phi=b_n\sqrt{n}$ , $\alpha_\phi=n$ , and $r_\phi=b_n^2$ , the sequence of the extended processes $(B_n(x,t),x\in\mathbb R_+)$ LD converges in $\mathbb D(\mathbb R_+,\mathbb R)$ to the idempotent process $(B(x\wedge1,t),x\in\mathbb R_+)$ . By (3.5), (3.6), and the continuous mapping principle, for $0\le x_1\le\ldots\le x_l<1$ , the $((K_n(x_i,t),B_n(x_i,t)),i\in\{1,2,\ldots,l\})$ LD converge in $\mathbb R^{2l}$ to $((K(x_i,t),B(x_i,t)),i\in\{1,2,\ldots,l\})$ . Since $K_n(1,t)=0$ and $K(1,t)=0$ (see Appendix A), the latter convergence also holds if $x_l=1$ .

We now show that the sequence $(K_n(x,t),x\in[0,1])$ is $\mathbb C$ -exponentially tight of order $b_n^2$ in $\mathbb D([0,1],\mathbb R)$ . (The definition and basic properties of $\mathbb C$ -exponential tightness are reviewed in Appendix A.) By Theorem 8, (3.10) needs to be complemented with

(3.14) \begin{equation} \lim_{\delta\rightarrow0}\limsup_{n\to\infty}\text{sup}_{x\in[0,1]}\mathbb P(\text{sup}_{y\in[0, \delta]}|K_n(x+y,t)-K_n(x,t)|\ge \eta)^{1/b_n^2}=0,\end{equation}

for arbitrary $\eta>0$ , where $K_n(x,t)=0$ when $x\ge1$ . We use an argument similar to that used in the proof of (3.10). Defining $\overline{K}_n(x,t)=-K_n(1-x,t)$ for $x\in[0,1]$ and $\overline{K}_n(x,t)=0$ for $x\ge 1$ , we have by (3.3) that

(3.15) \begin{align} &\text{sup}_{x\in[0,1]}\mathbb P(\text{sup}_{y\in[0, \delta]}|K_n(x+y,t)-K_n(x,t)|\ge \eta)\nonumber\\[5pt] & \le \text{sup}_{x\in[0,1/2]}\mathbb P(\text{sup}_{y\in[0, \delta]}|K_n( x+y,t)-K_n(x,t)|\ge \eta)\nonumber\\[5pt] &+\text{sup}_{x\in[1/2,1-\delta]}\mathbb P(\text{sup}_{y\in[0, \delta]}|K_n(x+y,t)- K_n(x,t)|\ge \eta)\nonumber\\[5pt] &+\text{sup}_{x\in[1-\delta,1]}\mathbb P(\text{sup}_{y\in[0, \delta]}|K_n(x+y,t)- K_n(x,t)|\ge \eta)\nonumber\\[5pt] &\le\text{sup}_{x\in[0,1/2]}\mathbb P(\text{sup}_{y\in[0, \delta]}|K_n( x+y,t)-K_n(x,t)|\ge \eta)\nonumber\\[5pt] &+\text{sup}_{x\in[\delta,1/2]}\mathbb P(\text{sup}_{y\in[0, \delta]}|K_n(1-x+y,t)- K_n(1-x,t)|\ge \eta) \\[5pt] &+\mathbb P(\text{sup}_{u\in[1- \delta,1]}|K_n(u,t)|\ge \eta/2)\le\text{sup}_{x\in[0,1/2]}\mathbb P(\text{sup}_{y\in[0, \delta]}|K_n( x+y,t)-K_n(x,t)|\ge \eta)\nonumber\\[5pt] &+\text{sup}_{x\in[\delta,1/2]}\mathbb P(\text{sup}_{y\in[0, \delta]}|\overline K_n(x-y,t)-\overline K_n(x,t)|\ge \eta)\nonumber\\[5pt] &+\mathbb P(\text{sup}_{u\in[0, \delta]}|\overline K_n(u,t)|\ge \eta/2)\nonumber\\[5pt] &\le\text{sup}_{x\in[0,1/2]}\mathbb P(\text{sup}_{y\in[0, \delta]}|K_n( x+y,t)-K_n(x,t)|\ge \eta)\nonumber\\[5pt] &+\text{sup}_{x\in[0,1/2]}\mathbb P(\text{sup}_{y\in[0, \delta]}|\overline K_n(x+y,t)-\overline K_n(x,t)|\ge \eta)\nonumber\\[5pt] & +\mathbb P(\text{sup}_{u\in[0, \delta]}|\overline K_n(u,t)|\ge \eta/2).\end{align}

Since the random variables $\zeta_i$ are independent and uniformly distributed on [0,1], $\overline{K}_n$ has the same finite-dimensional distributions as $K_n$ , so that

(3.16) \begin{multline} \text{sup}_{x\in[0,1]}\mathbb P(\text{sup}_{y\in[0, \delta]}|K_n(x+y,t)-K_n(x,t)|\ge \eta)\\[5pt] \le2\text{sup}_{x\in[0,1/2]}\mathbb P(\text{sup}_{y\in[0, \delta]}|K_n( x+y,t)-K_n(x,t)|\ge \eta)+\mathbb P(\text{sup}_{u\in[0, \delta]}| K_n(u,t)|\ge \eta/2).\end{multline}

Since $x+y\le 2/3$ when $x\in[0,1/2]$ and $y\in[0,\delta]$ provided $\delta$ is small enough, by (3.5), for $x\in[0,1/2]$ and $\delta$ small enough,

(3.17) \begin{align} & \text{sup}_{y\in[0,\delta]}\lvert K_n(x+y,t)-K_n(x,t)\rvert\le\nonumber\\[5pt]& 3\delta\text{sup}_{u\in[0,1]}\lvert K_n(u,t)\rvert+\text{sup}_{y\in[0,\delta]}\lvert B_n(x+y,t)-B_n(x,t)\rvert.\end{align}

Similarly,

(3.18) \begin{equation} \text{sup}_{u\in[0, \delta]}| K_n(u,t)|\le\frac{\delta}{1-\delta}\,\text{sup}_{u\in[0,1]}\lvert K_n(u,t)\rvert+\text{sup}_{u\in[0, \delta]}| B_n(u,t)|.\end{equation}

By (3.16), (3.17), (3.18), and the fact that $B_n(t)$ LD converges to B(t),

(3.19) \begin{align} &\limsup_{n\to\infty}\text{sup}_{x\in[0,1]}\mathbb P(\text{sup}_{y\in[0, \delta]}|K_n(x+y,t)-K_n(x,t)|\ge \eta)^{1/b_n^2}\nonumber\\[5pt]&\le\limsup_{n\to\infty}\mathbb P(3\delta\text{sup}_{u\in[0,1]}\lvert K_n(u,t)\rvert\ge\eta/2)^{1/b_n^2}\nonumber\\[5pt] &+\limsup_{n\to\infty}\mathbb P(\frac{\delta}{1-\delta}\,\text{sup}_{u\in[0,1]}\lvert K_n(u,t)\rvert\ge\eta/4)^{1/b_n^2}\nonumber\\[5pt]&+\Pi(\text{sup}_{x\in[0,1/2]}\text{sup}_{y\in[0,\delta]}\lvert B(x+y,t)-B(x,t)\rvert\ge \eta/2)\nonumber\\[5pt] &+\Pi(\text{sup}_{u\in[0, \delta]}| B(u,t)|\ge\eta/4).\end{align}

By (3.10),

(3.20) \begin{align} &\lim_{\delta\to0}\limsup_{n\to\infty}\text{sup}_{x\in[0,1]}\mathbb P(\text{sup}_{y\in[0, \delta]}|K_n(x+y,t)-K_n(x,t)|\ge \eta)^{1/b_n^2} \nonumber\\[5pt] &\le\lim_{\delta\to0} \Pi(\text{sup}_{x\in[0,1/2]}\text{sup}_{y\in[0,\delta]}\lvert B(x+y,t)-B(x,t)\rvert\ge \eta/2)\nonumber\\[5pt]&+\lim_{\delta\to0}\Pi(\text{sup}_{u\in[0, \delta]}| B(u,t)|\ge\eta/4).\end{align}

The idempotent process $B=((B(x,t),x\in[0,1]),\,t\in\mathbb R_+)$ has trajectories from the space of continuous functions $\mathbb C(\mathbb R_+,\mathbb C([0,1],\mathbb R))$ ; see Appendix A. Since the collections of sets $\{b\in\mathbb C(\mathbb R_+,\mathbb C([0,1],\mathbb R)) :\,\text{sup}_{x\in[0,1/2]}\text{sup}_{y\in[0,\delta]}\lvert b(x+y,t)-b(x,t)\rvert\ge\eta/2\}$ and $\{b\in\mathbb C(\mathbb R_+,\mathbb C([0,1],\mathbb R)) :\,\text{sup}_{x\in[0,\delta]}\lvert b(x,t)\rvert\ge \eta/4\}$ are nested collections of closed sets as $\delta\downarrow0$ , the limit on the right of (3.20) is (see Appendix A)

\[ \Pi(\text{sup}_{x\in[0,1/2]}\text{sup}_{y\in[0,0]}\lvert B(x+y,t)-B(x,t)\rvert\ge \eta/2)+\Pi(\text{sup}_{u\in[0,0]}| B(u,t)|\ge\eta/4)=0,\]

which concludes the proof of (3.14).

Since the sequence of stochastic processes $((K_n(x,t),B_n(x,t)),x\in[0,1])$ LD converges to the idempotent process $((K(x,t),B(x,t)),x\in[0,1])$ in the sense of finite-dimensional distributions and is $\mathbb C$ -exponentially tight, the LD convergence holds in $\mathbb D([0,1],\mathbb R^2)$ ; see Theorem 7 in Appendix A. It has thus been proved that the sequence of stochastic processes $\bigl(((K_n(x,t_1),B_n(x,t_1)),\,x\in[0,1]),\ldots,((K_n(x,t_l),B_n(x,t_l)),\,x\in[0,1])\bigr)$ LD converges in $\mathbb D([0,1],\mathbb R^2)^l$ to the idempotent process $\bigl(((K(x,t_1),B(x,t_1)),\,x\in[0,1]),\ldots,((K(x,t_l),B(x,t_l)),\,x\in[0,1])\bigr)$ , for all $t_1\le t_2\le\ldots\le t_l$ . The proof of the lemma will be complete if the sequence $\bigl(((K_n(x,t),B_n(x,t)),x\in[0,1]),t\in\mathbb R_+\bigr)$ is shown to be $\mathbb C$ -exponentially tight of order $b_n^2$ in $\mathbb D(\mathbb R_+,\mathbb D([0,1],\mathbb R^2))$ . The definition of exponential tightness implies that it is sufficient to prove that each of the sequences $\{K_n,n\ge1\}$ and $\{B_n,n\ge1\}$ is $\mathbb C$ -exponentially tight of order $b_n^2$ in $\mathbb D(\mathbb R_+,\mathbb D([0,1],\mathbb R))$ . By (3.10) and Theorem 8, the $\mathbb C$ -exponential tightness of $\{K_n,n\ge1\}$ would follow if, for all $L>0$ and $\eta>0$ ,

(3.21) \begin{eqnarray} \lim_{\delta\rightarrow0}\limsup_{n\to\infty}\text{sup}_{s\in[0,L]}\mathbb P(\text{sup}_{t\in[0, \delta]}\text{sup}_{x\in[0,1]}|K_n(x,s+t)-K_n(x,s)|\ge \eta)^{1/b_n^2}=0. \qquad \end{eqnarray}

Since, in analogy with the reasoning in (3.15),

\begin{multline*} \mathbb P(\text{sup}_{t\in[0, \delta]}\text{sup}_{x\in[0,1]}|K_n(x,t+s)-K_n(x,s)|\ge \eta)\\[5pt] \le\mathbb P(\text{sup}_{t\in[0, \delta]}\text{sup}_{x\in[0,1/2]}|K_n(x,t+s)-K_n(x,s)|\ge \eta)\\[5pt] +\mathbb P(\text{sup}_{t\in[0, \delta]}\text{sup}_{x\in[1/2,1]}|K_n(x,t+s)-K_n(x,s)|\ge \eta)\\[5pt] =\mathbb P(\text{sup}_{t\in[0, \delta]}\text{sup}_{x\in[0,1/2]}|K_n(x,t+s)-K_n(x,s)|\ge \eta)\\[5pt] + \mathbb P(\text{sup}_{t\in[0, \delta]}\text{sup}_{x\in[0,1/2]}|\overline{K}_n(x,t+s)-\overline{K}_n(x,s)|\ge \eta)\\[5pt] =2\mathbb P(\text{sup}_{t\in[0, \delta]}\text{sup}_{x\in[0,1/2]}|K_n(x,t+s)-K_n(x,s)|\ge \eta),\end{multline*}

(3.21) is implied by

(3.22) \begin{eqnarray} &&\lim_{\delta\rightarrow0}\limsup_{n\to\infty}\text{sup}_{s\in[0,L]}\mathbb P(\text{sup}_{t\in[0, \delta]}\text{sup}_{x\in[0,1/2]}|K_n(x,t+s)-K_n(x,s)|\ge \eta)^{1/b_n^2}=0.\nonumber\\ && \end{eqnarray}

By (3.3), with x being held fixed, the process $(K_n(x,t+s)-K_n(x,s),t\in\mathbb R_+)$ is a locally square-integrable martingale, so $(\text{sup}_{x\in[0,1/2]}(K_n(x,t+s)-K_n(x,s)),t\in\mathbb R_+)$ , is a submartingale; hence by Doob’s inequality [Reference Liptser and Shiryaev14, Theorem 3.2, p. 60],

(3.23) \begin{align} &\mathbb P(\text{sup}_{t\in[0, \delta]}\text{sup}_{x\in[0,1/2]}|K_n(x,t+s)-K_n(x,s)|\ge \eta)\notag \\[5pt] & \qquad \le\frac{1}{\eta^{2b_n^2}}\,\mathbb E\text{sup}_{x\in[0,1/2]}(K_n(x,s+\delta)-K_n(x,s))^{2b_n^2}.\end{align}

As noted earlier, with t being held fixed, the process $(B_n(x,t),x\in[0,1])$ is a square-integrable martingale [Reference Jacod and Shiryaev9, Chapter II, §3c]. Equation (3.5) yields, by the Gronwall–Bellman inequality, as in (3.12),

(3.24) \begin{eqnarray} \text{sup}_{x\in[0,1/2]}|K_n(x,s+\delta)-K_n(x,s)|\le e^2\text{sup}_{x\in[0,1]}|B_n(x,s+\delta)-B_n(x,s)|.\end{eqnarray}

Since $(B_n(x,s+\delta)-B_n(x,s),x\in[0,1])$ is a square-integrable martingale, by another application of Doob’s inequality (see [Reference Jacod and Shiryaev9, Theorem I.1.43] and [Reference Liptser and Shiryayev13, Theorem I.9.2]), as well as by Jensen’s inequality,

\begin{eqnarray} \notag \mathbb E\text{sup}_{x\in[0,1]}(B_n(x,s+\delta)-B_n(x,s))^{2b_n^2}\le \Bigl(\frac{2b_n^2}{2b_n^2-1}\Bigr)^{2b_n^2}\mathbb E(B_n(1,s+\delta)-B_n(1,s))^{2b_n^2}.\end{eqnarray}

By (3.4), the fact that $1-\zeta_1$ and $\zeta_1$ have the same distribution, and the bound (5.6) in the proof of Theorem 19 in [Reference Petrov15, Chapter III, §5],

\begin{multline*} \mathbb E(B_n(1,s+\delta)-B_n(1,s))^{2b_n^2}\le(b_n\sqrt{n})^{-2b_n^2}\bigl((b_n^2+1)^{2b_n^2} (n\delta+1) \mathbb E(1+\ln\zeta_1)^{2b_n^2}\\[5pt] +2b_n^2(b_n^2+1)^{b_n^2}e^{b_n^2+1} (n\delta+1)^{b_n^2}\bigl(\mathbb E(1+\ln\zeta_1)^2\bigr)^{b_n^2}\bigr).\end{multline*}

(More specifically, the following bound is used. Suppose $X_1,\ldots,X_n$ are i.i.d. with $\mathbb EX_1=0$ . Then, provided $p\ge2$ and $r>p/2$ ,

\[ \mathbb E\lvert \sum_{i=1}^nX_i\rvert^p\le r^pn\mathbb E\lvert X_1\rvert^p+pr^{p/2}e^rn^{p/2}(\mathbb EX_1^2)^{p/2}.\]

See also [Reference Whittle25] for similar results.)

As $\zeta_1$ is uniform on [0,1], $\mathbb E(\ln\zeta_1)^{2b_n^2}=(2b_n^2)!$ , so that, with the use of Jensen’s inequality,

\begin{multline*}\limsup_{n\to\infty}(\mathbb E(B_n(1,s+\delta)-B_n(1,s))^{2b_n^2})^{1/b_n^2}\\[5pt] \le \limsup_{n\to\infty}(b_n\sqrt{n})^{-2}\Bigl(4n^{1/b_n^2}(b_n^2+1)^{2}\bigl((2b_n^2)!\bigr)^{1/b_n^2}+(b_n^2+1)e( n\delta+1)\mathbb E(1+\ln\zeta_1)^2\Bigr),\end{multline*}

which implies, via Stirling’s formula, on recalling that $b_n^6n^{1/b_n^2-1}\to0$ , that

\[\lim_{\delta\to0}\limsup_{n\to\infty}\text{sup}_{s\in[0,L]}\Bigl(\mathbb E(B_n(1,s+\delta)-B_n(1,s))^{2b_n^2}\Bigr)^{1/b_n^2}=0.\]

Recalling (3.23), (3.24), and (3.25) yields (3.22). The proof of the $\mathbb C$ -exponential tightness of $B_n$ is similar. (It is actually simpler.)

Going back to the set-up of Theorem 3, let

(3.25) \begin{align} H_n(t)&=Y_n(t)-\int_0^tY_n(t-s)\,dF(s), \end{align}
(3.26) \begin{align} X^{(0)}_n(t)&=\frac{\sqrt{n}}{b_n}\,\bigl(\frac{1}{n}\, Q_n^{(0)}(t)-(1-F_0(t))\bigr), \end{align}

(3.27) \begin{align} U_n(x,t)&=\frac{1}{b_n\sqrt{n}}\,\sum_{i=1}^{{{\hat A}_n}(t)}\bigl(\,\mathbf{1}_{\{\eta_i\le x\}}\,-F(x)\bigr),\end{align}

and

(3.28) \begin{align} \Theta_n(t)&=-\int_{\mathbb R_+^2} \mathbf{1}_{\{x+s\le t\}}\,dU_n(x,s).\end{align}

As, owing to (2.1),

(3.29) \begin{equation} 1-F_0(t)+\mu t-\mu\int_0^t(t-s)dF(s)=1,\end{equation}

(2.2), (2.4), and (2.7) imply that

(3.30) \begin{align} X_n(t)&=(1-F(t))X_n(0)^++ X_n^{(0)}(t)+\int_0^tX_n(t-s)^+\,dF(s)+H_n(t)+\Theta_n(t).\end{align}

The equation (3.2) is written in a similar way: introducing

(3.31) \begin{align} H(t)&=Y(t)-\int_0^t Y(t-s)\,dF(s),\end{align}
(3.32) \begin{align}X^{(0)}(t)&=W^0(F_0(t))-(1-F_0(t))X(0)^-,\end{align}
(3.33) \begin{align}U(x,t)&=K(F(x),\mu t),\end{align}

and

(3.34) \begin{align} \Theta(t)&=-\int_{\mathbb R_+^2} \mathbf{1}_{\{x+s\le t\}}\,dU(x,s)=-\int_{\mathbb R_+^2} \mathbf{1}_{\{x+s\le t\}}\,\dot K(F(x),\mu s)\,dF(x)\,\mu ds\end{align}

yields

(3.35) \begin{equation} X(t)=(1-F(t))X(0)^++X^{(0)}(t)+\int_0^tX(t-s)^+\,dF(s)+H(t)+\Theta(t).\end{equation}

Let $H=(H(t),\,t\in\mathbb R_+)$ , $H_n=(H_n(t),\,t\in\mathbb R_+)$ , $X^{(0)}=(X^{(0)}(t),\,t\in\mathbb R_+)$ , $ X_n^{(0)}=( X_n^{(0)}(t),\,t\in\mathbb R_+)$ , $U=((U(x,t),x\in\mathbb R_+),t\in\mathbb R_+)$ , $U_n=((U_n(x,t),x\in\mathbb R_+),t\in\mathbb R_+)$ , $\Theta=(\Theta(t),t\in\mathbb R_+)$ , and $\Theta_n=(\Theta_n(t),\,t\in\mathbb R_+)$ .

Theorem 4. As $n\to\infty$ , the sequence $(X_n(0), X_n^{(0)},H_n,\Theta_n)$ LD converges in distribution at rate $b_n^2$ in $\mathbb R\times\mathbb D(\mathbb R_+,\mathbb R)^3$ to $(X(0),X^{(0)},H,\Theta)$ .

The groundwork needs to be laid first. Let

(3.36) \begin{equation} L_n(x,t)=\frac{1}{b_n\sqrt{n}}\,\sum_{i=1}^{{{\hat A}_n}(t)}\bigg(\mathbf{1}_{\{\eta_i\le x\}}\,-\int_0^{\eta_i\wedge x}\frac{dF(u)}{1-F(u)}\bigg).\end{equation}

Since the random variables $F(\eta_i)$ are i.i.d. and uniform on [0, 1], in view of (3.3), (3.4), and (3.27), it may be assumed that

(3.37) \begin{equation} L_n(x,t)=B_n(F(x),\frac{\hat A_n(t)}{n})\end{equation}

and that

(3.38) \begin{equation} U_n(x,t)=K_n\bigg(F(x),\frac{\hat A_n(t)}{n}\bigg).\end{equation}

By (3.5),

(3.39) \begin{equation} U_n(x,t)=-\int_0^x\frac{U_n(y,t )}{1-F(y)}\,dF(y)+L_n(x,t).\end{equation}

By (3.28),

(3.40) \begin{equation} \Theta_n(t)=J_n(t)-M_n(t),\end{equation}

with

(3.41) \begin{align} J_n(t)&= \int_0^t \frac{U_n(x,t-x)}{1-F(x)}\,dF(x)\end{align}

and

(3.42) \begin{align} M_{n}(t)&=\int_{\mathbb R_+^2} \,\mathbf{1}_{\{x+s\le t\}}\,\,dL_n(x,s).\end{align}

Lemma 2. Under the hypotheses,

(3.43) \begin{align} \lim_{r\to\infty}\limsup_{n\to\infty}\mathbb P(\text{sup}_{s\in[0,t]}\lvert U_n(s,t-s)\rvert>r)^{1/b_n^2}=0 \end{align}

and

(3.44) \begin{align} \lim_{r\to\infty}\limsup_{n\to\infty}\mathbb P(\text{sup}_{s\in[0,t]}\lvert M_n(s)\rvert>r)^{1/b_n^2}=0. \end{align}

Proof. Note that

(3.45) \begin{equation} \lim_{r\to\infty} \limsup_{n\to\infty}\mathbb P\bigg(\frac{A_n(t)}{n}>r\bigg)^{1/b_n^2}=0, \end{equation}

which is a consequence of the LD convergence at rate $b_n^2$ of the $Y_n$ to Y (see (3.1)). Similarly, the LD convergence of $X_n(0)$ to X(0) implies that

(3.46) \begin{equation} \lim_{r\to\infty} \limsup_{n\to\infty}\mathbb P\bigg(\frac{Q_n(0)}{n}>r\bigg)^{1/b_n^2}=0.\end{equation}

By (2.4), ${{\hat A}_n}(t)\le(Q_n(0)-n)^++A_n(t)$ , so that (3.45) and (3.46) imply that

(3.47) \begin{equation} \lim_{L\to\infty} \limsup_{n\to\infty}\mathbb P\bigg(\frac{\hat A_n(t)}{n}>L\bigg)^{1/b_n^2}=0. \end{equation}

By (3.38),

(3.48) \begin{equation} \mathbb P(\text{sup}_{s\in[0,t]}\lvert U_n(s,t-s)\rvert>r)\le\mathbb P\bigg(\frac{\hat A_n(t)}{n}>L\bigg)+\mathbb P(\text{sup}_{\substack{s\in [0,L],\\ x\in[0,1]}}\lvert K_n(x,s)\rvert>r).\end{equation}

By the LD convergence of $K_n$ to K in Lemma 1, and since the trajectories of K are continuous,

\[ \lim_{r\to\infty} \limsup_{n\to\infty}\mathbb P\bigl(\text{sup}_{\substack{s\in [0,L],\\ x\in[0,1]}}\lvert K_n(x,s)\rvert>r)^{1/b_n^2}=0.\]

Combined with (3.47) and (3.48), this proves (3.43).

Lemma 3.1 in [Reference Puhalskii and Reed21] implies that the process $M_n=(M_n(t),\,t\in\mathbb R_+)$ is a local martingale with respect to the filtration $\mathbf{G}_n$ defined as follows. For $t\in\mathbb R_+$ , let $\hat{\mathcal{G}}_n(t)$ denote the complete $\sigma$ -algebra generated by the random variables $\,\mathbf{1}_{\{\hat\tau_{n,i}\le s\}}\,\,\mathbf{1}_{\{ \eta_i\le x\}}$ , where $x+ s\le t$ and $i\in\mathbb N$ , and by the ${{\hat A}_n}(s)$ (or, equivalently, by the $\,\mathbf{1}_{\{\hat\tau_{n,i}\le s\}}\,$ for $i\in\mathbb N$ ), where $s\leq t$ . Define $\mathcal{G}_n(t)=\cap_{\epsilon>0}\hat{\mathcal{G}}_n(t+\epsilon)$ and $\mathbf{G}_n=(\mathcal{G}_n(t), t\in\mathbb R_+)$ . By (3.36) and (3.42),

(3.49) \begin{equation} M_n(t)=\frac{1}{b_n\sqrt{n}}\,\sum_{i=1}^{{{\hat A}_n}(t)}\bigg(\,\mathbf{1}_{\{\eta_i+\hat\tau_{n,i}\le t\}}\,-\int_0^{\eta_i\wedge (t-\hat\tau_{n,i})}\frac{dF(u)}{1-F(u)}\bigg).\end{equation}

Thus, the measure of jumps of $M_n$ is

(3.50) \begin{equation} \mu_n([0,t],\Gamma)=\,\mathbf{1}_{\{1/(b_n\sqrt{n})\in\Gamma\}}\,\sum_{i=1}^{{{\hat A}_n}(t)}\,\mathbf{1}_{\{\eta_i+\hat\tau_{n,i}\le t\}}\,\end{equation}

and the associated $\mathbf{G}_n$ -predictable measure of jumps is

(3.51) \begin{equation} \nu_n([0,t],\Gamma)=\,\mathbf{1}_{\{1/(b_n\sqrt{n})\in\Gamma\}}\,\sum_{i=1}^{{{\hat A}_n}(t)}\int_0^{\eta_i\wedge (t-\hat\tau_{n,i})}\frac{dF(u)}{1-F(u)}.\end{equation}

Note that it is a continuous process. (For $\hat A_n$ being predictable, see Lemma C.1 in [Reference Puhalskii and Reed21].) The associated stochastic cumulant is (see, e.g., [Reference Puhalskii19, p. 293])

(3.52) \begin{equation} G_n(\alpha,t)=\bigg(e^{\alpha/(b_n\sqrt{n})}-1-\frac{\alpha}{b_n\sqrt{n}}\bigg)\sum_{i=1}^{{{\hat A}_n}(t)}\int_0^{\eta_i\wedge (t-\hat\tau_{n,i})}\frac{dF(u)}{1-F(u)}.\end{equation}

By Lemma 4.1.1 on p. 294 in [Reference Puhalskii19], the process $(e^{\alpha M_n(t)-G_n(\alpha,t)},t\in\mathbb R_+)$ is a local martingale, so that $\mathbb E e^{\alpha M_n(\tau)-G_n(\alpha,\tau)}\le1$ , for arbitrary stopping time $\tau$ . Lemma 3.2.6 on p. 282 in [Reference Puhalskii19] implies that, for $\gamma>0$ ,

\begin{multline*} \mathbb P(\text{sup}_{s\in[0,t]}e^{\alpha b_n^2 M_n(s)}\ge e^{\alpha b_n^2 r})\le e^{\alpha b_n^2(\gamma- r)}+\mathbb P(e^{G_n(\alpha b_n^2,t)}\ge e^{\alpha b_n^2\gamma})\\[5pt] \le e^{\alpha b_n^2(\gamma- r)}+\mathbb P(e^{\hat G_n(\alpha b_n^2,t)}\ge e^{\alpha b_n^2\gamma}),\end{multline*}

where

(3.53) \begin{equation} \hat G_n(\alpha,t)=\bigg(e^{\alpha/(b_n\sqrt{n})}-1-\frac{\alpha}{b_n\sqrt{n}}\bigg)\sum_{i=1}^{{{\hat A}_n}(t)}\int_0^{t-\hat\tau_{n,i}}\frac{dF(u)}{1-F(u)}.\end{equation}

Hence, for $\alpha>0$ ,

(3.54) \begin{equation} \mathbb P(\text{sup}_{s\in[0,t]} M_n(s)\ge r)^{1/b_n^2}\le e^{\alpha(\gamma- r)}+\mathbb P(\hat G_n(\alpha b_n^2,t)\ge\alpha b_n^2\gamma)^{1/b_n^2}.\end{equation}

On writing

(3.55) \begin{multline} \hat G_n(\alpha b_n^2,t)=\bigg(e^{\alpha b_n/\sqrt{n}}-1-\frac{\alpha b_n}{\sqrt{n}}\bigg)\int_0^t\int_0^{t-s}\frac{dF(u)}{1-F(u)}\,d\hat A_n(s)\\[5pt] =\bigg(e^{\alpha b_n/\sqrt{n}}-1-\frac{\alpha b_n}{\sqrt{n}}\bigg)\int_0^t\hat A_n(t-u)\,\frac{dF(u)}{1-F(u)}\end{multline}

and noting that $(n/b_n^2) (e^{\alpha b_n/\sqrt{n}}-1-\alpha b_n/\sqrt{n})\to\alpha^2/2$ , one can see, thanks to (3.47), that

\[ \lim_{n\to\infty}\mathbb P(\hat G_n(\alpha b_n^2,t)\ge\alpha b_n^2\gamma)^{1/b_n^2}=0, \]

provided $\alpha$ is small enough, which proves that

\[ \lim_{r\to\infty}\limsup_{n\to\infty}\mathbb P(\text{sup}_{s\in[0,t]}M_n(s)>r)^{1/b_n^2}=0.\]

The argument for $\text{sup}_{s\in[0,t]}(\!-M_n(s))$ is similar. The convergence (3.44) has been proved.

Lemma 3. For arbitrary $\epsilon>0$ and $t>0$ ,

\[\lim_{n\to\infty}\mathbb P\bigg(\text{sup}_{s\in[0,t]}\lvert \frac{\hat A_n(s)}{n}-\mu s\rvert>\epsilon\bigg)^{1/b_n^2}=0.\]

Proof. By (2.4), (2.2), (3.27), and (3.28),

(3.56) \begin{multline} \frac{1}{n}\,Q_n(t)=\bigg(\frac{1}{n}\,Q_n(0)-1\bigg)^+(1-F(t))+\frac{1}{n}\,Q_n^{(0)}(t)+\frac{1}{n}\,A_n(t)-\frac{1}{n}\,\int_{0}^tA_n(t-s)\,dF(s)\\[5pt] +\frac{1}{n}\,\int_0^t(Q_n(t-s)-n)^+ \,dF(s)+\frac{b_n}{\sqrt{n}}\,\Theta_n(t).\end{multline}

By (3.40), (3.41), (3.43), and (3.44), on recalling that $F(t)<1$ , we have

(3.57) \begin{equation} \lim_{r\to\infty}\limsup_{n\to\infty}\mathbb P(\text{sup}_{s\in[0,t]}\lvert \Theta_n(s)\rvert>r)^{1/b_n^2}=0.\end{equation}

The LD convergence at rate $b_n^2$ of $Y_n$ to Y implies that, for $\epsilon>0$ ,

(3.58) \begin{equation} \lim_{n\to\infty}\mathbb P\bigg(\text{sup}_{s\in[0,t]}\lvert \frac{A_n(s)}{n}-\mu s\rvert>\epsilon\bigg)^{1/b_n^2}=0.\end{equation}

By (2.3), (3.46), and Lemma 1,

\[ \lim_{n\to\infty}\mathbb P\bigg(\text{sup}_{s\in[0,t]}\lvert \frac{1}{n}\,Q_n^{(0)}(s)-(1-F_0(s))\rvert>\epsilon\bigg)^{1/b_n^2}=0.\]

Recalling (3.29) implies that

\[ \lim_{n\to\infty}\mathbb P\bigg(\text{sup}_{s\in[0,t]}\lvert \frac{1}{n}\,Q_n^{(0)}(s)+\frac{1}{n}\,A_n(s)-\frac{1}{n}\,\int_{0}^sA_n(s-x)\,dF(x)-1\rvert>\epsilon\bigg)^{1/b_n^2}=0.\]

In addition, the LD convergence of $X_n(0)$ to X(0) implies that (3.46) can be strengthened as follows:

(3.59) \begin{equation} \limsup_{n\to\infty}\mathbb P\bigg(\lvert \frac{Q_n(0)}{n}-1\rvert>\epsilon\bigg)^{1/b_n^2}=0.\end{equation}

Hence, by (3.56) and (3.57),

\[ \frac{1}{n}\,Q_n(t)-1=\,\int_0^t\bigg(\frac{1}{n}\,Q_n(t-s)-1\bigg)^+ \,dF(s)+\theta_n(t),\]

where

\begin{multline*}\!\!\!\!\!\theta_n(t)= \bigg(\frac{1}{n}\,Q_n(0)-1\bigg)^+(1-F(t))+\frac{1}{n}\,Q_n^{(0)}(t)+\frac{1}{n}\,A_n(t)-\frac{1}{n}\,\int_{0}^tA_n(t-s)\,dF(s)+\frac{b_n}{\sqrt{n}}\,\Theta_n(t)\end{multline*}

and

\[\!\!\!\!\!\lim_{n\to\infty} \mathbb P(\text{sup}_{s\in[0,t]}\lvert \theta_n(s)\rvert>\epsilon)^{1/b_n^2}=0.\]

Lemma B.1 in [Reference Puhalskii and Reed21] implies that there exists a function $\rho$ , which depends only on the function F, such that

\[ \text{sup}_{s\in[0,t]}\lvert \frac{1}{n}\,Q_n(s)-1\rvert\le \rho(t)\text{sup}_{s\in[0,t]}\lvert \theta_n(s)\rvert.\]

Therefore,

(3.60) \begin{equation} \lim_{n\to\infty} \mathbb P\bigg(\text{sup}_{s\in[0,t]}\lvert \frac{1}{n}\,Q_n(s)-1\rvert>\epsilon\bigg)^{1/b_n^2}=0.\end{equation}

When combined with (2.4) and (3.58), this yields the assertion of the lemma.

Let $L_n=\bigl((L_n(x,t),x\in\mathbb R_+),t\in\mathbb R_+\bigr)$ and $L=\bigl((L(x,t),x\in\mathbb R_+),t\in\mathbb R_+\bigr)$ , with $L(x,t)=B(F(x),\mu t)$ .

Lemma 4. As $n\to\infty$ , the sequence $(X_n(0), X_n^{(0)},U_n,L_n)$ LD converges in distribution in $\mathbb R\times\mathbb D(\mathbb R_+,\mathbb R)\times\mathbb D(\mathbb R_+,\mathbb D(\mathbb R_+,\mathbb R))^2$ to $(X(0),X^{(0)},U,L)$ .

Proof. Let

(3.61) \begin{equation} \tilde X^{(0)}_n(x,t)=\frac{1}{b_n\sqrt{n}}\,\sum_{i=1}^{\lfloor nt\rfloor}\,(\,\mathbf{1}_{\{\eta^{(0)}_i> x\}}\,-(1-F_0(x))),\end{equation}

$\tilde X^{(0)}_n(t)=(\tilde X^{(0)}_n(x,t),x\in\mathbb R_+)$ and $\tilde X^{(0)}_n=(\tilde X^{(0)}_n(t),t\in\mathbb R_+)$ . By the hypotheses of Theorem 3, $X_n(0)$ LD converges to X(0), by Lemma 1 and because $F_0$ is strictly increasing, and $\tilde X^{(0)}_n$ LD converges to $((\tilde K(F_0(x),t),x\in\mathbb R_+), t\in\mathbb R_+)$ , where $\tilde K$ represents a Kiefer idempotent process that is independent of (X(0),K, B). Also $(K_n,B_n)$ LD converges to (K, B). By independence assumptions, these convergences hold jointly; cf. Appendix A. Since $Q_n(0)/n\to 1$ and $\hat A_n(t)/n\to\mu t$ super-exponentially in probability by (3.59) and (3.58), respectively, ‘Slutsky’s theorem’ (Lemma 6) yields joint LD convergence of $(X_n(0),\tilde X^{(0)}_n,K_n,B_n,Q_n(0)/n,(\hat A_n(t)/n,t\in \mathbb R_+))$ to $(X(0),((\tilde K(F_0(x),t),x\in\mathbb R_+), t\in\mathbb R_+),K,B,1,(\mu t,t\in\mathbb R_+))$ . In addition, by (2.3), (3.26), and (3.61),

\[ X^{(0)}_n(t)=\tilde X^{(0)}_n\bigg(t,\frac{Q_n(0)}{n}\wedge 1\bigg)-(1-F_0(t))X_n(0)^-.\]

In order to deduce the LD convergence of $(X_n(0), X_n^{(0)},U_n,L_n)$ to $(X(0), X^{(0)},U,L)$ , it remains to recall (3.32), (3.33), (3.37), and (3.38); note that, by Lemma 7, $(\tilde K(t,1),t\in[0,1])$ is a Brownian bridge idempotent process; and apply the continuous mapping principle, the associated composition mappings being continuous at continuous limits. (See [Reference Whitt24] for more background on continuous functions in the Skorokhod space context.)

Lemma 5. The sequence $\{\Theta_n,n\in\mathbb N\}$ is $\mathbb C$ -exponentially tight of order $b_n^2$ in $\mathbb D(\mathbb R_+,\mathbb R)$ .

Proof. Let

(3.62) \begin{equation}J(t)=\int_0^t \frac{U(x,t-x)}{1-F(x)}\,dF(x).\end{equation}

By Lemma 4, (3.41), (3.62), and the continuous mapping principle, $J_n=(J_n(t), t\in\mathbb R_+)$ LD converges to $J=J(t),t\in\mathbb R_+)$ , so the sequence $J_n$ is $\mathbb C$ -exponentially tight. By (3.40), it remains to check that the sequence $M_n$ is $\mathbb C$ -exponentially tight, which, according to Theorem 8, is implied by the following convergences:

\begin{align} \lim_{r\to\infty}\limsup_{n\to\infty}\mathbb P(\text{sup}_{s\in[0,t]}\lvert M_n(s)\rvert>r)^{1/b_n^2}=0\notag\end{align}

and

(3.63) \begin{align} \lim_{\delta\to0}\limsup_{n\to\infty}\text{sup}_{s\in[0,t]}\mathbb P(\text{sup}_{s'\in[0,\delta]}\lvert M_n(s+s')-M_n(s)\rvert>\epsilon)^{1/b_n^2}=0,\end{align}

where $t>0$ and $\epsilon>0$ . The former convergence has already been proved; see (3.44). The proof of (3.63) proceeds along similar lines. Since, with $\alpha\in\mathbb R$ , the process $ (\exp\bigl(\alpha( M_n(s+s')-M_n(s))-(G_n(\alpha,s+s')-G_n(\alpha,s))\bigr),s'\in\mathbb R_+)$ is a local martingale, so that, for arbitrary stopping time $\tau$ ,

$$\mathbb E e^{\alpha( M_n(s+\tau)-M_n(s))-(G_n(\alpha,s+\tau)-G_n(\alpha,s))}\le1,$$

by Lemma 3.2.6 on p. 282 in [Reference Puhalskii19], for arbitrary $\gamma>0$ , in analogy with (3.54), for $b_n\ge1$ , we have

\begin{multline*} \mathbb P(\text{sup}_{s'\in[0,\delta]}( M_n(s+s')-M_n(s))\\[5pt]\ge \epsilon)^{1/b_n^2}\le e^{\alpha(\gamma- \epsilon)}+\mathbb P(\hat G_n(\alpha b_n^2,s+\delta)-\hat G_n(\alpha b_n^2,s)\ge\alpha b_n^2\gamma)^{1/b_n^2}.\end{multline*}

By (3.55) and Lemma 3,

\[\frac{1}{b_n^2}\, \hat G_n(\alpha b_n^2,t)\to\frac{\alpha^2}{2}\,\mu\int_0^t( t-u)\frac{dF(u)}{1-F(u)}\]

super-exponentially in probability at rate $b_n^2$ . The fact that the latter super-exponential convergence in probability is locally uniform in t, as the limit is a monotonic continuous function starting at 0, implies that, for $\delta$ small enough, depending on $\alpha$ ,

\[ \limsup_{n\to\infty}\text{sup}_{s\in[0,t]} \mathbb P(\text{sup}_{s'\in[0,\delta]}( M_n(s'+s)-M_n(s))\ge \epsilon)^{1/b_n^2}\le e^{\alpha(\gamma- \epsilon)}.\]

Now, one chooses $\gamma<\epsilon$ and sends $\alpha\to\infty$ . A similar convergence holds with $-M_n(s')$ substituted for $M_n(s')$ . The limit (3.63) has been proved.

Lemma 6. The sequence $(X_n(0), X_n^{(0)},H_n,J_n,L_n,U_n,\Theta_n)$ LD converges in distribution in $\mathbb R\times\mathbb D(\mathbb R_+,\mathbb R)^3\times\mathbb D(\mathbb R_+,\mathbb D(\mathbb R_+,\mathbb R))^2\times\mathbb D(\mathbb R_+,\mathbb R)$ to $(X(0),X^{(0)},H,J,L,U,\Theta)$ , as $n\to\infty$ .

Proof. Let

(3.64) \begin{equation} \Theta^{(l)}_n(t)=-\int_{\mathbb R_+^2}I_{t}^{(l)}(x,s)dU_n(x,s),\end{equation}

where

\[ I_{t}^{(l)}(x,s)=\sum_{i=1}^\infty\,\mathbf{1}_{\{s\in(s^{(l)}_{i-1},s^{(l)}_i]\}}\,\,\mathbf{1}_{\{x\in[0,t-s^{(l)}_{i-1}]\}}\,,\]

for some $0=s^{(l)}_0<s^{(l)}_1<\ldots$ such that $s^{(l)}_i\to\infty$ , as $i\to\infty$ , and $\text{sup}_{i\ge1}(s^{(l)}_i-s^{(l)}_{i-1})\to0$ , as $l\to\infty$ . Evidently,

\[\Theta^{(l)}_n(t)=-\sum_{i=1}^\infty(U_n(t-s^{(l)}_{i-1},s^{(l)}_i)-U_n(t-s^{(l)}_{i-1}, s^{(l)}_{i-1}))\,\mathbf{1}_{\{s^{(l)}_{i-1}\le t\}}.\]

Similarly, let

\begin{multline*} \Theta^{(l)}(t)=-\int_{\mathbb R_+^2}I_{t}^{(l)}(x,s)\,dU(x,s)=\sum_{i=1}^\infty(U(t-s^{(l)}_{i-1},s^{(l)}_i)-U(t-s^{(l)}_{i-1},s^{(l)}_{i-1}))\,\mathbf{1}_{\{s^{(l)}_{i-1}\le t\}}.\end{multline*}

By the LD convergence of $Y_n$ to Y in the hypotheses of Theorem 3, Lemma 4, (3.25), (3.31), (3.41), (3.62), and the continuous mapping principle, the sequence $(X_{n}(0), X_n^{(0)},H_{n},J_{n},L_{n},U_{n})$ LD converges to $(X(0), X^{(0)},H,J,L,U)$ . Hence, the sequence $(X_{n}(0), X_n^{(0)},H_{n},J_{n},L_{n},U_{n},\Theta_{n}^{(l)})$ LD converges to $(X(0), X^{(0)},H,J,L,U,\Theta^{(l)})$ . In addition, by Lemma 5, the sequence $\Theta_n$ is $\mathbb C$ -exponentially tight. Since the idempotent processes $X^{(0)},H,J,L,U$ are seen to have continuous trajectories, the sequence $( X_n^{(0)},H_n,J_n,L_n,U_n,\Theta_n)$ is $\mathbb C$ -exponentially tight. That a limit point of $(X_n(0),X_n^{(0)},H_n,J_n,L_n,U_n,\Theta_n)$ has the same idempotent distribution as $(X(0), X^{(0)},H,J,L,U,\Theta)$ would follow from the LD convergence of finite-dimensional distributions of $(X_n(0), X_n^{(0)},H_n,J_n,L_n,U_n,\Theta_{n})$ to finite-dimensional distributions of $(X(0), X^{(0)},H,J,L,U,\Theta)$ . Owing to (3.33),

\[ \Theta^{(l)}(t)=-\int_{\mathbb R_+^2}I_{t}^{(l)}(x,s)\dot K(F(x),\mu s)\,dF(x)\,\mu ds,\]

which implies, by (3.34) and the Cauchy–Schwarz inequality, that $\Theta^{(l)}\to \Theta$ locally uniformly, as $l\to\infty$ . Since the sequence $(X_n(0), X_n^{(0)},H_n,J_n,L_n,U_n)$ LD converges to $(X(0), X^{(0)},H,J,L,U)$ , to prove the finite-dimensional LD convergence it suffices to prove that

(3.65) \begin{equation} \lim_{l\to\infty}\limsup_{n\to\infty}\mathbb P(\lvert \Theta_n^{(l)}(t)-\Theta_n(t)\rvert>\epsilon)^{1/b_n^2}=0.\end{equation}

Let

\begin{multline*} \hat \Theta_n^{(l)}(s,t)=\frac{1}{b_n\sqrt{n}}\sum_{i=1}^{\hat A_n(s)}\sum_{j=1}^\infty\,\mathbf{1}_{\{s^{(l)}_{j-1}\le t\}}\,\,\mathbf{1}_{\{\hat\tau_{n,i}\in(s^{(l)}_{j-1},s^{(l)}_j]\}}\,\bigl(\,\mathbf{1}_{\{\eta_i\in( t-\hat\tau_{n,i},t-s^{(l)}_{j-1}]\}}\,\\[5pt] -(F(t-s^{(l)}_{j-1})-F(t-\hat\tau_{n,i}))\bigr).\end{multline*}

By (3.28) and (3.64),

(3.66) \begin{equation} \Theta_n(t)-\Theta_n^{(l)}(t)=\hat \Theta_n^{(l)}(t,t).\end{equation}

Let $\mathcal{F}_{n}(s)$ represent the complete $\sigma$ -algebra generated by the random variables $\hat\tau_{n,j}\wedge \hat\tau_{n,\hat A_n(s)+1}$ and $\eta_{j\wedge \hat A_n(s)}$ , where $j\in \mathbb N$ . By Part 4 of Lemma C.1 in [Reference Puhalskii and Reed21], with t held fixed, the process $(\hat\Theta_n^{(l)}(s,t),s\in\mathbb R_+)$ is an $\mathbb F_n$ -locally square-integrable martingale. Its measure of jumps is

\begin{multline*} \mu_n^{(l)}([0,s],\Gamma)=\sum_{i=1}^{\hat A_n(s)}\sum_{j=1}^\infty\,\mathbf{1}_{\{s^{(l)}_{j-1}\le t\}}\,\,\mathbf{1}_{\{\hat\tau_{n,i}\in(s^{(l)}_{j-1},s^{(l)}_j]\}}\,\bigl(\,\mathbf{1}_{\{(1-(F(t-s^{(l)}_{j-1})-F(t-\hat\tau_{n,i}))/(b_n\sqrt{n})\in\Gamma\}}\,\\[5pt] \,\mathbf{1}_{\{\eta_i\in( t-\hat\tau_{n,i},t-s^{(l)}_{j-1}]\}}\,+\,\mathbf{1}_{\{(-(F(t-s^{(l)}_{j-1})-F(t-\hat\tau_{n,i}))/(b_n\sqrt{n})\in\Gamma\}}\,\,\mathbf{1}_{\{\eta_i\not\in( t-\hat\tau_{n,i},t-s^{(l)}_{j-1}]\}}\,\bigr),\end{multline*}

where $\Gamma\subset \mathbb R\setminus\{0\}$ . Accordingly, the $\mathbb F_n$ -predictable measure of jumps is

\begin{multline*} \nu_n^{(l)}([0,s],\Gamma)=\sum_{i=1}^{\hat A_n(s)}\sum_{j=1}^\infty\,\mathbf{1}_{\{s^{(l)}_{j-1}\le t\}}\,\,\mathbf{1}_{\{\hat\tau_{n,i}\in(s^{(l)}_{j-1},s^{(l)}_j]\}}\,\bigl(\,\mathbf{1}_{\{(1-(F(t-s^{(l)}_{j-1})-F(t-\hat\tau_{n,i}))/(b_n\sqrt{n})\in\Gamma\}}\,\\[5pt] (F(t-s^{(l)}_{j-1})-F( t-\hat\tau_{n,i}))+\,\mathbf{1}_{\{-(F(t-s^{(l)}_{j-1})-F(t-\hat\tau_{n,i}))/(b_n\sqrt{n})\in\Gamma\}}\,(1-(F(t-s^{(l)}_{j-1})-F( t-\hat\tau_{n,i})))\bigr).\end{multline*}

For $\alpha\in\mathbb R$ , as on p. 214 in [Reference Puhalskii19], define the stochastic cumulant

(3.67) \begin{multline} G^{(l)}_n(\alpha,s)=\int_0^s\int_\mathbb R(e^{\alpha x}-1-\alpha x)\nu^{(l)}_n(ds',dx)\\[5pt] =\sum_{i=1}^{\hat A_n(s)}\sum_{j=1}^\infty\,\mathbf{1}_{\{s^{(l)}_{j-1}\le t\}}\,\,\mathbf{1}_{\{\hat\tau_{n,i}\in(s^{(l)}_{j-1},s^{(l)}_j]\}}\,\bigl((e^{\alpha (1-(F(t-s^{(l)}_{j-1})-F(t-\hat\tau_{n,i})))/(b_n\sqrt{n})}\\[5pt] -1-\frac{\alpha}{b_n\sqrt{n}}(1-(F(t-s^{(l)}_{j-1})-F(t-\hat\tau_{n,i}))))(F(t-s^{(l)}_{j-1})-F(t-\hat\tau_{n,i}))\\[5pt] +(e^{-\alpha (F(t-s^{(l)}_{j-1})-F(t-\hat\tau_{n,i}))/( b_n\sqrt{n})}-1\\[5pt] +\frac{\alpha}{b_n\sqrt{n}}(F(t-s^{(l)}_{j-1})-F(t-\hat\tau_{n,i})))(1-(F(t-s^{(l)}_{j-1})-F(t-\hat\tau_{n,i})))\bigr).\end{multline}

The associated stochastic exponential is defined by

(3.68) \begin{equation} \mathcal{E}^{(l)}_n(\alpha,s)=e^{G^{(l)}_n(\alpha,s)}\prod_{0<s'\le s}(1+\Delta G^{(l)}_n(\alpha,s'))e^{-\Delta G^{(l)}_n(\alpha,s')},\end{equation}

where $\Delta G^{(l)}_n(s')$ represents the jump of $G^{(l)}_n(s'')$ with respect to s” evaluated at $s''=s'$ and the product is taken over the jumps. By Lemma 4.1.1 on p. 294 in [Reference Puhalskii19], the process $ \bigl( e^{\alpha b_n^2 \hat \Theta_n^{(l)}(s,t))}\mathcal{E}_n(\alpha b_n^2,s)^{-1},s\in\mathbb R_+\bigr)$ is a well-defined local martingale, so that, for any stopping time $\tau$ ,

\[ \mathbb Ee^{\alpha b_n^2 \hat \Theta_n^{(l)}(\tau,t))}\mathcal{E}^{(l)}_n(\alpha b_n^2,\tau)^{-1}\le1. \]

Lemma 3.2.6 on p. 282 in [Reference Puhalskii19] and (3.68) imply that, for $\alpha>0$ and $\gamma>0$ ,

(3.69) \begin{align} \mathbb P(\text{sup}_{s\in[0,t]} \hat \Theta_n^{(l)}(s,t)\ge \epsilon)\le e^{\alpha b_n^2(\gamma-\epsilon)}+\mathbb P (\mathcal{E}^{(l)}_n(\alpha b_n^2,t)\ge e^{\alpha\notag b_n^2\gamma})\\[5pt] \le e^{\alpha b_n^2(\gamma-\epsilon)}+\mathbb P (G^{(l)}_n(\alpha b_n^2,t)\ge \alpha b_n^2\gamma).\end{align}

By (3.67),

\begin{multline*}G^{(l)}_n(\alpha b_n^2,t)\le\hat A_n(t)\bigg(\text{sup}_{\lvert y\rvert\le1}\bigg(e^{\alpha b_ny/\sqrt{n}}-1-\alpha\frac{b_ny}{\sqrt{n}}\bigg)\,\text{sup}_j\big(F(t-s^{(l)}_{j-1})-F(t-s^{(l)}_j)\big)\\[5pt] +\text{sup}_j\text{sup}_{y\in[F(t-s^{(l)}_{j}),F(t-s^{(l)}_{j-1})]}\bigg(e^{-\alpha b_n y/\sqrt{n}}-1+\alpha\frac{b_n}{\sqrt{n}}y\bigg)\bigg). \end{multline*}

As $n/b_n^2(e^{\alpha b_n y/\sqrt{n}}-1-\alpha b_ny/\sqrt{n})\to\alpha^2y^2/2$ uniformly on bounded intervals, $\hat A_n(t)/n\to \mu t$ super-exponentially as $n\to\infty$ , and $\text{sup}_j(s^l_j-s^l_{j-1})\to0$ as $l\to\infty$ , it follows that

\[ \lim_{l\to\infty}\limsup_{n \to\infty}\mathbb P (G^{(l)}_n(\alpha b_n^2,t)\ge \alpha b_n^2\gamma)^{1/b_n^2}=0,\]

which implies, thanks to (3.69), that

\[\limsup_{l\to\infty} \limsup_{n\to\infty} \mathbb P(\text{sup}_{s\in[0,t]}\hat \Theta_n^{(l)}(s,t)\ge \epsilon)^{1/b_n^2}\le e^{\alpha (\gamma-\epsilon)}.\]

Picking $\gamma<\epsilon$ and sending $\alpha$ to $\infty$ shows that the latter left-hand side equals zero. A similar argument proves the convergence

\[\lim_{l\to\infty} \limsup_{n\to\infty} \mathbb P(\text{sup}_{s\in[0,t]}(\!-\hat \Theta_n^{(l)}(s,t))\ge \epsilon)^{1/b_n^2}=0.\]

Recalling (3.66) yields the convergence (3.65).

Theorem 4 has thus been proved. In order to obtain the assertion of Theorem 3, note that $-K$ has the same idempotent distribution as K and invoke the continuous mapping principle in (3.30), which applies by Lemma B.2 in [Reference Puhalskii and Reed21].

The proof of Theorem 2 proceeds along similar lines. First we prove an analogue of Theorem 3 to the effect that if the random variables $\overline X_n(0)$ LD converge to an idempotent variable $\overline X(0)$ , then the processes $\overline X_n$ LD converge to the idempotent process $\overline X$ that solves the following analogue of (3.2):

(3.70) \begin{multline} \overline X(t)=(1-F_0(t))\overline X(0)+\sqrt{q_0}\,W^0(F_0(t))+\int_0^t\bigl(1-F(t-s)\bigr)\sigma\,\dot W(s)\,ds\\[5pt] +\int_{\mathbb R_+^2} \,\mathbf{1}_{\{x+s\le t\}}\,\,\dot K( F(x),\lambda s)\,dF(x)\,\lambda\,ds.\end{multline}

The proof is a simpler version of the proof of Theorem 3. Essentially, one replaces $\hat A_n$ with $A_n$ and $\mu$ with $\lambda$ . A key element of the proof of Theorem 3 is the property, asserted in the statement of Lemma 3, that $\hat A_n(t)/n\to \mu t$ super-exponentially in probability. This property takes some effort to establish. Its counterpart for the infinite-server queue is that $ A_n(t)/n\to \lambda t$ super-exponentially in probability; this is a direct consequence of the hypotheses.

In some more detail, since, in analogy with (2.2),

\[ \overline Q_n(t)=Q_n^{(0)}(t)+A_n(t)-\int_0^t\int_0^t \mathbf{1}_{\{x+s\le t\}}\,d\,\sum_{i=1}^{{{ A}_n}(s)}\mathbf{1}_{\{\eta_i\le x\}},\]

grouping terms appropriately and recalling (2.12) yields

\begin{multline*} \!\!\!\!\!\frac{1}{n}\,\overline Q_n(t)-\overline q(t)=\bigg(\!\frac{1}{n}\,\overline{Q}_n(0)-q_0\!\bigg)(1-F_0(t))+\frac{1}{n}\,A_n(t)-\lambda t-\!\int_0^t\!\!\bigg(\frac{1}{n}\,A_n(t-s)-\lambda(t-s)\!\!\bigg) dF(s)\\[5pt] +\frac{1}{n}\,\sum_{i=1}^{\overline{Q}_n(0)}\,(\,\mathbf{1}_{\{\eta^{(0)}_i> t\}}\,-(1-F_0(t)))-\int_{\mathbb R_+^2}\,\mathbf{1}_{\{x+s\le t\}}\,d\sum_{i=1}^{A_n(s)}\bigl(\,\mathbf{1}_{\{\eta_i\le x\}}\,-F(x)).\end{multline*}

On introducing

\begin{align*}\overline Y_n(t)&=\frac{\sqrt{n}}{b_n}\,\bigg(\frac{A_n(t)}{n}-\lambda t\bigg),\\[5pt] \overline X_n^{(0)}(t)&=\frac{1}{b_n\sqrt{n}}\,\sum_{i=1}^{\overline{Q}_n(0)}\,(\,\mathbf{1}_{\{\eta^{(0)}_i> t\}}\,-(1-F_0(t))),\\[5pt] \overline H_n(t)&=\overline Y_n(t)-\int_0^t\overline Y_n(t-s)\,dF(s),\\[5pt] \overline U_n(x,t)&=\frac{1}{b_n\sqrt{n}}\,\sum_{i=1}^{{{ A}_n}(t)}\bigl(\,\mathbf{1}_{\{\eta_i\le x\}}\,-F(x)\bigr),\end{align*}

and

\begin{align*} \overline\Theta_n(t)&=-\int_{\mathbb R_+^2} \mathbf{1}_{\{x+s\le t\}}\,d\overline U_n(x,s),\end{align*}

we obtain the following analogue of (3.30):

(3.71) \begin{equation} \overline X_n(t)=(1-F_0(t))\overline X_n(0)+\overline X_n^{(0)}(t)+\overline H_n(t)+\overline \Theta_n(t).\end{equation}

The hypotheses imply that $\overline Y_n$ LD converges to $\sigma W$ , with the notation of the proof of Theorem 3 being reused. Since, in analogy with (3.38), $ \overline U_n(x,t)=K_n(F(x), A_n(t)/n)$ and $A_n(t)/n$ converges to $\lambda t$ super-exponentially in probability, the process $\overline U_n$ LD converges to $\overline U$ , where $\overline U(x,t)=K(F(x),\lambda t).$ Furthermore, similarly to Lemma 4, it is proved that $(\overline X_n(0),\overline X_n^{(0)},\overline U_n,\overline L_n)$ LD converges to $(\overline X(0),\overline X^{(0)},\overline U,\overline L)$ , where $\overline X^{(0)}(t)=\sqrt{q_0}W^0(F_0(t))$ and $\overline L(x,t)=B(F(x),\lambda t)$ . Put together, these properties imply the analogue of Theorem 4: that $(\overline X_n(0),\overline X_n^{(0)},\overline H_n,\overline \Theta_n)$ LD converges to $(\overline X(0),\overline X^{(0)},\overline H,\overline \Theta)$ , where

\begin{align*} \overline H(t)=\sigma W(t)-\sigma\int_0^t W(t-s)\,dF(s)=\sigma W(t)-\sigma\int_0^t(1- F(t-s))\,\dot W(s)\,ds\end{align*}

and

\begin{align*}\overline \Theta(t)=-\int_{\mathbb R_+^2} \mathbf{1}_{\{x+s\le t\}}\,d\overline U(x,s)=-\int_{\mathbb R_+^2} \mathbf{1}_{\{x+s\le t\}}\,\dot K(F(x),\lambda s)\,dF(x)\,\lambda ds. \end{align*}

An application of the continuous mapping principle to (3.71) concludes the proof.

4. Evaluating the deviation functions

This section is concerned with solving for $ I^Q_{x_0}(q)$ and $\overline I^Q_{q_0,x_0}(q)$ .

Theorem 5. Suppose that the CDF F is an absolutely continuous function and $I_{x_0}^Q(q)<\infty$ . Then q is absolutely continuous, $(\dot q(t)-\int_0^t\dot q(s)\,\mathbf{1}_{\{q(s)>0\}}\,F'(t-s)\,ds,t\in\mathbb R_+)\in\mathbb L_2(\mathbb R_+)$ , the infimum in (2.5) is attained uniquely, and

\[ I_{x_0}^Q(q)=\frac{1}{2}\,\int_0^\infty\hat{p}(t)\bigl(\dot q(t)-\int_0^t\dot q(s)\,\mathbf{1}_{\{q(s)>0\}}\,F'(t-s)\,ds+(\beta-x_0^-) F_0'(t)\bigr)\,dt,\]

where $\hat{p}(t)$ represents the unique $ \mathbb L_2(\mathbb R_+)$ -solution p(t) of the Fredholm equation of the second kind

(4.1) \begin{multline} (\mu+\sigma^2)p(t)= \dot q(t)-\int_0^t\dot q(s)\,\mathbf{1}_{\{q(s)>0\}}\,F'(t-s)\,ds+(\beta-x_0^-) F_0'(t)\\[5pt] +\sigma^2\int_0^\infty F'(\lvert s-t\rvert)p(s)\,ds-\sigma^2\int_0^\infty\int_0^{s\wedge t}F'(s-r)\, F'(t-r)\,dr\,p(s)\,ds,\end{multline}

with $\dot q$ , $F_0'$ , and F’ denoting derivatives.

Proof. Writing

(4.2) \begin{equation} \int_{\mathbb R_+^2} \,\mathbf{1}_{\{x+s\le t\}}\,\,\dot k(F(x),\mu s)\,dF(x)\,\mu\,ds= \int_0^{ t}\int_0^{F(t-s)}\dot k(x,\mu s)\,dx\,\mu\,ds,\end{equation}

we see that the equation (2.6) is of the form

\[ q(t)=f(t)+\int_0^tq(t-s)^+\,dF(s), t\in\mathbb R_+,\]

with the functions f(t) and F(t) being absolutely continuous. The function q(t) is absolutely continuous by Lemma 8. In addition, (4.2) implies that, almost everywhere (a.e.),

(4.3) \begin{equation} \frac{d}{dt}\,\int_{\mathbb R_+^2} \,\mathbf{1}_{\{x+s\le t\}}\,\,{\dot k}(F(x),\mu s)\,dF(x)\,\mu\,ds=\int_0^t{\dot k}(F(s),\mu (t-s))F'(s)\,\mu ds.\end{equation}

The infimum in (2.5) is attained uniquely by coercitivity and strict convexity of the function being minimised; cf. [Reference Ekeland and Temam5, Proposition II.1.2]. Differentiation in (2.6) with the account of (4.3) implies that, a.e.,

\begin{multline*}\dot w^0(F_0(t))F_0'(t)+\sigma\,\dot w(t)-\int_0^tF'(t-s)\sigma\,\dot w(s)\,ds+\int_0^t\dot k(F(s),\mu (t-s))F'(s)\,\mu\,ds\\[5pt] -\bigg( \dot q(t)-\int_0^t\dot q(s)\,\mathbf{1}_{\{q(s)>0\}}\,F'(t-s)\,ds+(\beta-x_0^-) F_0'(t)\bigg)=0.\end{multline*}

Introduce the map

\begin{multline*} \Phi:\, (\dot w^0,\dot w,\dot k)\to\bigg(\dot w^0(F_0(t))F_0'(t)+\sigma\,\dot w(t)-\int_0^tF'(t-s)\sigma\,\dot w(s)\,ds\\[5pt] +\int_0^t\dot k(F(s),\mu (t-s))F'(s)\,\mu\,ds,t\in\mathbb R_+\bigg).\end{multline*}

Since $F_0'(t)$ is bounded by (2.1), $\Phi$ maps $V=\mathbb L_2([0,1])\times\mathbb L_2(\mathbb R_+)\times\mathbb L_2( [0,1]\times\mathbb R_+)$ to $\mathbb L_2(\mathbb R_+)$ . For instance, using the fact that $\int_0^\infty F'(s)\,ds=1$ , we have

\[ \int_0^\infty\bigl(\int_0^tF'(t-s)\dot w(s)\,ds\bigr)^2\,dt\le \int_0^\infty\int_0^tF'(t-s)\dot w(s)^2\,ds\,dt=\int_0^\infty\dot w(s)^2\,ds<\infty\]

and

\begin{multline*} \int_0^\infty\bigl(\int_0^t\dot k(F(s),\mu (t-s))F'(s)\,\mu ds\bigr)^2\,dt\le \int_0^\infty\int_0^t\dot k(F(s),\mu (t-s))^2F'(s)\,\mu^2ds\,dt\\[5pt] =\mu^2\int_0^\infty\int_0^1\dot k( x,t)^2dx\,dt<\infty.\end{multline*}

The method of Lagrange multipliers, more specifically, [Reference Ekeland and Temam5, Proposition III.5.2] with $Y=\mathbb L_2(\mathbb R_+)$ and the set of nonnegative functions as the cone $\mathcal{C}$ , yields

(4.4) \begin{multline} I_{x_0}^Q(q)=\text{sup}_{p\in \mathbb L_2(\mathbb R_+)}\;\inf_{\substack{(\dot w^0,\dot w,\dot k)\in\mathbb L_2([0,1])\\[5pt] \times\mathbb L_2( \mathbb R_+)\times\mathbb L_2( [0,1]\times\mathbb R_+)}}\Bigl(\frac{1}{2}\int_0^1\dot w^0(x)^2\,dx+\frac{1}{2}\int_0^\infty \dot w(t)^2\,dt\\[5pt] +\frac{1}{2}\int_0^\infty\int_0^1 \dot k(x,t)^2\,dx\,dt+\int_0^\infty p(t)\bigl(\dot q(t)+F'(t)x_0^++(\beta-x_0^-) F_0'(t)\\[5pt] -\int_0^t\dot q(s)\,\mathbf{1}_{\{q(s)>0\}}\,F'(t-s)\,ds -\dot w^0(F_0(t))F_0'(t)-\sigma\,\dot w(t)+\int_0^tF'(t-s)\sigma\,\dot w(s)\,ds\\[5pt] -\int_0^t\dot k(F(s),\mu (t-s))F'(s)\,\mu\,ds\bigr)\,dt\Bigr).\end{multline}

Minimising in (4.4) yields, with $(\dot{\hat{ w}}^{0}(t),\dot{\hat w}(t), \dot{\hat{k}}(x,t))$ , being optimal,

\begin{align*} \dot{\hat w}^0(x)-p(F_0^{-1}(x))=0,\\[5pt] \dot{\hat w}(t)-\sigma p(t)+\sigma\int_0^\infty p(t+s)F'(s)\,ds=0,\\[5pt] \dot{\hat k}(x,t)-p(\frac{t}{\mu}+F^{-1}(x))=0.\end{align*}

(For the latter, note that

\begin{multline*} \int_0^\infty p(t)\int_0^t\dot k(F(s),\mu (t-s))F'(s)\,\mu ds\,dt=\int_0^\infty\int_s^\infty p(t)\dot k(F(s),\mu (t-s))F'(s)\,\mu\,dt\, ds\\[5pt] =\int_0^\infty\int_0^\infty p(t+s)\dot k(F(s),\mu t)F'(s)\,\mu\,dt\,ds=\int_0^\infty\int_0^1 p(\frac{t}{\mu}+F^{-1}(x))\dot k( x,t)\,dx\,dt.)\end{multline*}

Hence,

\begin{align*} I_{x_0}^Q(q)&=\text{sup}_{p\in \mathbb L_2(\mathbb R_+)}\bigg(\int_0^\infty p(t)\bigl(\dot q(t)-\int_0^t\dot q(s)\,\mathbf{1}_{\{q(s)>0\}}\,F'(t-s)\,ds+(\beta-x_0^-) F_0'(t)\bigr)\,dt\\[5pt] &\quad-\frac{1}{2}\bigg(\int_0^\infty p(s)^2F_0'(s)\,ds+\int_0^\infty(\sigma p(t)-\sigma\int_0^\infty p(t+s)F'(s)\,ds)^2\,dt\\[5pt]&\quad+\mu\int_0^\infty\int_0^\infty p(t+s)^2\,F'(s)ds\,dt\bigg)\bigg).\end{align*}

Noting that

(4.5) \begin{equation} \int_0^\infty p(s)^2F_0'(s)\,ds+\mu\int_0^\infty\int_0^\infty p(t+s)^2\,F'(s)ds\,dt=\mu\int_0^\infty p(s)^2\,ds\end{equation}

yields

(4.6) \begin{multline} I_{x_0}^Q(q)=\text{sup}_{p\in\mathbb L_2(\mathbb R_+)}\bigg(\int_0^\infty p(t)\bigg(\dot q(t)-\int_0^t\dot q(s)\,\mathbf{1}_{\{q(s)>0\}}\,F'(t-s)\,ds+(\beta-x_0^-) F_0'(t)\bigg)\,dt\\[5pt] -\frac{1}{2}\bigg(\mu\int_0^\infty p(s)^2\,ds+\int_0^\infty(\sigma p(t)-\sigma\int_0^\infty p(t+s)F'(s)\,ds)^2\,dt\bigg)\bigg).\end{multline}

The existence and uniqueness of a maximiser in (4.6) follows from Proposition II.1.2 in [Reference Ekeland and Temam5] because the expression in the supremum tends to $-\infty$ as $\lVert p\rVert_{\mathbb L_2(\mathbb R_+)}\to\infty$ . Varying p in (4.6) implies (4.1). As the maximiser in (4.6) is unique, so is an $\mathbb L_2(\mathbb R_+)$ -solution of the Fredholm equation.

It is noteworthy that the integral operator on $\mathbb L_2(\mathbb R_+)$ with kernel

\[ \tilde K(t,s)=F'(\lvert s-t\rvert)-\int_0^{s\wedge t}F'(s-r)\,F'(t-r)\,dr\]

is not generally either Hilbert–Schmidt or compact, so the existence and uniqueness of $\hat p(t)$ is not a direct consequence of the general theory.

The numerical solution of Fredholm equations, such as (4.1), is discussed at quite some length in the literature; see, e.g., [Reference Atkinson3] and references therein. For instance, the collocation method with a basis of ‘hat’ functions may be tried: for $i\in\mathbb N$ and $n\in\mathbb N$ , let $t_i=i/n$ and $\ell_i(t)=(1-\lvert t-t_i\rvert)\,\mathbf{1}_{\{t_{i-1}\le t\le t_i\}}$ , with $t_0=0$ . Then an approximate solution is

\[ p_n(t)=\sum_{i=1}^{n^2}p_n(t_i)\ell_i(t),\]

where the $p_n(t_i)$ , $i=1,\ldots,n^2$ , satisfy the linear system

\begin{multline*} (\mu+\sigma^2)p_n(t_i)-\sigma^2\sum_{j=1}^{n^2}p_n(t_j)\int_0^{n^2} \tilde K(t_i,s)\ell_j(s)\,ds\\[5pt] = \dot q(t_i)-\int_0^{t_i}\dot q(s)\,\mathbf{1}_{\{q(s)>0\}}\,F'(t_i-s)\,ds+(\beta-x_0^-) F_0'(t_i).\end{multline*}

For more background, see [Reference Atkinson3].

The evaluation of $\overline I^Q_{q_0,x_0}$ is carried out similarly:

\begin{multline*} \overline I^Q_{q_0,x_0}(q)=\text{sup}_{p\in \mathbb L_2(\mathbb R_+)}\; \inf_{ \substack{(\dot w^0,\dot w,\dot k)\in \mathbb L_2([0,1])\\[5pt] \times\mathbb L_2( \mathbb R_+)\times \mathbb L_2([0,1]\times\mathbb R_+)}}\Bigl(\frac{1}{2}\int_0^1\dot w^0(x)^2\,dx+\frac{1}{2}\int_0^\infty \dot w(t)^2\,dt\\[5pt] +\frac{1}{2}\int_0^\infty\int_0^1 \dot k(x,t)^2\,dx\,dt+\int_0^\infty p(t)\bigl(\dot q(t) +x_0 F_0'(t) -\sqrt{q_0}\dot w^0(F_0(t))F_0'(t) -\sigma\,\dot w(t)\\[5pt] +\int_0^tF'(t-s)\sigma\,\dot w(s)\,ds -\int_0^t\dot k(F(s),\lambda (t-s))F'(s)\,\lambda\, ds\bigr)\,dt\Bigr).\end{multline*}

The infimum is attained at $(\dot{\overline{w}}^0(t),\dot{\overline{w}}(t),\dot{\overline k}(x,t))$ such that

\begin{align*} \dot{\overline{w}}^0(x)-\sqrt{q_0} p(F_0^{-1}(x))=0,\\[5pt] \dot{\overline{w}}(t)-\sigma p(t)+\sigma\int_0^\infty p(t+s) F'(s)\,ds=0,\\[5pt] \dot{\overline k}(x,t)- p(\frac{t}{\lambda}+F^{-1}(x))=0.\end{align*}

Consequently, taking into account (2.1) and (4.5), we have

\begin{multline*} \overline I^Q_{q_0,x_0}(q)=\text{sup}_{p\in \mathbb L_2(\mathbb R_+)}\bigl(\int_0^\infty p(t)(\dot q(t) +x_0 F_0'(t))\,dt-\frac{1}{2}\bigl(q_0\mu\int_0^\infty p(s)^2\,(1-F(s))\,ds\\[5pt] +\lambda\int_0^\infty p(s)^2F(s)\,ds+\int_0^\infty(\sigma p(t)-\sigma\int_0^\infty p(t+s)F'(s)\,ds)^2\,dt\bigr)\bigr)\\[5pt] =\frac{1}{2}\int_0^\infty\overline p(t)(\dot q(t) +x_0 F_0'(t))\,dt,\end{multline*}

with $\overline p(t)$ being the $\mathbb L_2(\mathbb R_+)$ -solution p(t) to the Fredholm equation of the second kind

\[ (q_0\mu(1-F(t))+\lambda F(t)+\sigma^2)p(t)=\dot q(t)+x_0F^{\prime}_0(t)+\sigma^2\int_0^\infty \tilde K(t,s)p(s)\,ds.\]

Appendix A. Large-deviation convergence and idempotent processes

This section reviews the basics of LD convergence and idempotent processes; see, e.g., [Reference Puhalskii19]. Let $\mathbf E$ represent a metric space. Let $\mathcal{P}({\mathbf{E}})$ denote the power set of $\mathbf{E}$ . The set function $\Pi:\, \mathcal{P}(\mathbf{E})\to[0,1]$ is said to be a deviability if $\Pi(E)=\text{sup}_{y\in E}\Pi(\{y\}),\,E\subset \mathbf{E}$ , where the function $\Pi(y)=\Pi(\{y\})$ is such that $\text{sup}_{y\in \mathbf{E}}\Pi(y)=1$ and the sets $\{y\in \mathbf{E}:\,\Pi(y)\ge \gamma\}$ are compact for all $\gamma \in(0,1]$ . (One can also refer to $\Pi$ as a maxi-measure or an idempotent probability.) A deviability is a tight set function in the sense that $\inf_{K\in\mathcal{K}(\mathbf E)}\Pi(\mathbf E\setminus K)=0$ , where $\mathcal{K}(\mathbf E)$ stands for the collection of compact subsets of $\mathbf E$ . If $\Xi$ is a directed set and $F_\xi, \xi\in \Xi,$ is a net of closed subsets of $\mathbf E$ that is nonincreasing with respect to the partial order on $\Xi$ by inclusion, then $\Pi(\cap_{\xi\in \Xi}F_\xi)=\lim_{\xi\in \Xi}\Pi(F_\xi)$ . A property pertaining to elements of $\mathbf E$ is said to hold $\Pi$ -a.e. if the value of $\Pi$ of the set of elements that do not have this property equals 0.

A function f from $\mathbf E$ to metric space $\mathbf E'$ is called an idempotent variable. The idempotent distribution of the idempotent variable f is defined as the set function $\Pi\circ f^{-1}(\Gamma)=\Pi(f\in\Gamma),\,\Gamma\subset \mathbf E'$ . If f is the canonical idempotent variable defined by $f(y)=y$ , then it has $\Pi$ as its idempotent distribution. The continuous images of deviabilities are deviabilities; i.e., if $f:\,\mathbf E\to\mathbf E'$ is continuous, then $\Pi\circ f^{-1}$ is a deviability on $\mathbb E'$ . Furthermore, this property extends to the situation where $f:\,\mathbf E\to\mathbf E'$ is strictly Luzin, i.e., continuous when restricted to the set $\{y\in \mathbf E:\,\Pi(y)\ge \gamma\}$ , for arbitrary $\gamma\in(0,1]$ . Thus, the idempotent distribution of a strictly Luzin idempotent variable is a deviability. More generally, f is said to be a Luzin idempotent variable if the idempotent distribution of f is a deviability. If $f=(f_1,f_2)$ , with $f_i$ assuming values in $\mathbf E^{\prime}_i$ , then the (marginal) idempotent distribution of $f_1$ is defined by $\Pi^{f_1}(y^{\prime}_1)=\Pi(f_1=y^{\prime}_1)=\text{sup}_{y:\,f_1(y)=y^{\prime}_1}\Pi(y)$ . The idempotent variables $f_1$ and $f_2$ are said to be independent if ${\Pi}(f_1=y^{\prime}_1,\,f_2=y^{\prime}_2)=\Pi(f_1=y^{\prime}_1)\Pi(f_2=y^{\prime}_2)$ for all $(y^{\prime}_1,y^{\prime}_2)\in\mathbf E^{\prime}_1\times\mathbf E^{\prime}_2$ , so the joint distribution is the product of the marginal ones. Independence of finite collections of idempotent variables is defined similarly.

A sequence $Q_n$ of probability measures on the Borel $\sigma$ -algebra of $\mathbf E$ is said to large-deviation (LD) converge at rate $r_n$ to deviability $\Pi$ if for every bounded continuous non-negative function f on $\mathbb{E}$

\[\lim_{n\to\infty}\left(\int_{ \mathbb{E}}f(x)^{n}\,Q_n(dx)\right)^{1/r_n}=\text{sup}_{x\in \mathbf{E}}f(x)\Pi(x).\]

Equivalently, one may require that $\lim_{n\to\infty}Q_n(\Gamma)^{1/r_n}=\Pi( \Gamma)$ for every Borel set $\Gamma$ such that $\Pi$ of the interior of $\Gamma$ and $\Pi$ of the closure of $\Gamma$ agree. If the sequence $Q_n$ LD converges to $\Pi$ , then $\Pi(y)=\lim_{\delta\to0}\liminf_{n\to\infty}(Q_n(B_\delta(y))^{1/r_n}=\lim_{\delta\to0}\limsup_{n\to\infty}(Q_n(B_\delta(y))^{1/r_n}$ , for all $y\in\mathbf E$ , where $B_\delta(y)$ represents the open ball of radius $\delta$ about y. (Closed balls may be used as well.) The sequence $Q_n$ is said to be exponentially tight of order $r_n$ if $\inf_{K\in\mathcal{K}(\mathbf E)}\limsup_{n\to\infty}Q_n(\mathbf E\setminus K)^{1/r_n}=0$ . If the sequence $Q_n$ is exponentially tight of order $r_n$ , then there exists a subsequence $Q_{n'}$ that LD converges at rate $r_{n'}$ to a deviability. Any such deviability will be referred to as a large-deviation (LD) limit point of $Q_n$ . Given $\tilde E\subset \mathbf E$ , the sequence $Q_n$ is said to be $\tilde E$ -exponentially tight if it is exponentially tight and $\tilde\Pi(\mathbf E\setminus\tilde E)=0$ , for any LD limit point $\tilde \Pi$ of $Q_n$ .

It is immediate that $\Pi$ is a deviability if and only if $I(x)=-\ln \Pi(x)$ is a tight deviation function, i.e., the sets $\{x\in\mathbf E:\,I(x)\le \gamma\}$ are compact for all $\gamma\ge0$ and $\inf_{x\in\mathbf E}I(x)=0$ , and that the sequence $Q_n$ LD converges to $\Pi$ at rate $r_n$ if and only if it obeys the LDP for rate $r_n$ with deviation function I, i.e., $\liminf_{n\to\infty}(1/r_n)\,\ln Q_n(G)\ge -\inf_{x\in G}I(x)$ for all open sets G, and $\limsup_{n\to\infty}(1/r_n)\,\ln Q_n(F)\le -\inf_{x\in F}I(x)$ for all closed sets F.

LD convergence of probability measures can be also expressed as LD convergence in distribution of the associated random variables to idempotent variables. A sequence $\{X_n,\,n\in\mathbb N\}$ of random variables with values in $\mathbf E' $ is said to LD converge in distribution at rate $r_n$ as $n\to\infty$ to an idempotent variable X with values in $\mathbf E'$ if the sequence of the probability laws of the $X_n$ LD converges at rate $r_n$ to the idempotent distribution of X. If the random variables $X^{\prime}_n$ and $X^{\prime\prime}_n$ are independent, $X^{\prime}_n$ LD converges to X’, and $X^{\prime\prime}_n$ LD converges to X”, then the sequence $(X^{\prime}_n,X^{\prime\prime}_n)$ LD converges to (X’,X”) and X’ and X” are independent. If a sequence $\{{Q}_n,\,n\in\mathbb N\}$ of probability measures LD converges to a deviability ${\Pi}$ , then one has LD convergence in distribution of the canonical idempotent variables. A continuous mapping principle holds: if the random variables $X_n$ LD converge at rate $r_n$ to an idempotent variable X and $f:\,\mathbf E'\to\mathbf E''$ is a continuous function, then the random variables $f(X_n)$ LD converge at rate $r_n$ to f(X), where $\mathbf E''$ is a metric space. The following version of Slutsky’s theorem holds; see, e.g., [19, Lemma 3.1.42, p. 275]. Certainly, a direct proof along the lines of the argument in [4, Chapter 1, Theorem 4.1] is possible.

Theorem 6. Suppose that $\mathbf E'$ and $\mathbf E''$ are separable metric spaces. Let $Y_n$ be random variables with values in $\mathbb E''$ , and let $a\in\mathbf E''$ . If the sequence $X_n$ LD converges to X and $Y_n\to a$ super-exponentially in probability for rate $r_n$ , i.e., $\mathbb P(d(Y_n,a)>\epsilon)^{1/r_n}\to0$ for arbitrary $\epsilon>0$ , then the sequence $(X_n,Y_n)$ LD converges at rate $r_n$ to (X,a) in $\mathbf E'\times\mathbf E''$ , where d denotes the metric on $\mathbf E''$ .

A collection $(X_t,\,t\in\mathbb R_+)$ of idempotent variables on $\mathbf E$ is called an idempotent process. The functions $(X_t(y),\,t\in\mathbb R_+)$ for various $y\in\mathbf E$ are called trajectories (or paths) of X. Idempotent processes are said to be independent if they are independent as idempotent variables with values in the associated function spaces. The concepts of idempotent processes with independent and/or stationary increments mimic those for stochastic processes. Since this paper deals with stochastic processes having right-continuous trajectories with left-hand limits, the underlying space $\mathbf E$ may be assumed to be a Skorokhod space $\mathbb D(\mathbb R_+,\mathbb R^m)$ , for suitable m.

Suppose $X_n=(X_n(t),t\in\mathbb R_+)$ is a sequence of stochastic processes that assume values in metric space $\mathbf E'$ with metric d’ and have right-continuous trajectories with left-hand limits. The sequence $X_n$ is said to be exponentially tight of order $r_n$ if the sequence of the distributions of $X_n$ as measures on the Skorokhod space $\mathbb D(\mathbb R_+,\mathbf E')$ is exponentially tight of order $r_n$ . It is said to be $\mathbb C$ -exponentially tight if any LD limit point is the law of an idempotent process with continuous trajectories, i.e., $\Pi(\mathbb D(\mathbb R_+,\mathbf E')\setminus\mathbb C(\mathbb R_+,\mathbf E'))=0$ , whenever $\Pi$ is an LD limit point of the laws of $X_n$ .

The method of finite-dimensional distributions for LD convergence of stochastic processes is summarised in the next theorem [Reference Puhalskii, Shervashidze and De Gruyter16]. The proof mimics the one used in weak convergence theory.

Theorem 7. If, for all tuples $t_1< t_2<\ldots < t_l$ with the $t_i$ coming from a dense subset of $\mathbb R_+$ , the sequence $(X_n(t_1),\ldots,X_n(t_l))$ LD converges in $\mathbb R^l$ at rate $r_n$ to $(X(t_1),\ldots,X(t_l))$ and the sequence $X_n$ is $\mathbb C$ -exponentially tight of order $r_n$ , then X is a continuous-path idempotent process and the sequence $X_n$ LD converges in $\mathbb D(\mathbb R_+,\mathbf E')$ at rate $r_n$ to X.

The form of the conditions in the next theorem, which is essentially due to [Reference Feng and Kurtz7], is at odds with what is common in weak convergence theory, so a proof is warranted, although the argument is standard.

Theorem 8. Suppose $\mathbf E'$ is, in addition, complete and separable. The sequence $X_n$ is $\mathbb C$ -exponentially tight of order $r_n$ if and only if the following hold:

  1. (i) the sequence $X_n(t)$ is exponentially tight of order $r_n$ for all t from a dense subset of $\mathbb R_+$ , and

  2. (ii) for all $\epsilon>0$ and $L>0$ ,

    (A.1) \begin{equation} \lim_{\delta\to0}\limsup_{n\to\infty}\text{sup}_{t\in[0,L]} \mathbb P(\text{sup}_{s\in[0,\delta]}d'(X_n(t+s),X_n(t))>\epsilon)^{1/r_n}=0. \end{equation}

Proof. The necessity of the conditions follows from the continuity of the projection mapping and the continuous mapping principle. The sufficiency is proved next. For $L>0$ , $\delta>0$ and X from the Skorokhod space $\mathbb D(\mathbb R_+,\mathbf E')$ , let

\[ w_L(X,\delta)=\text{sup}_{t,s\in[0,L]:\,\lvert t-s\rvert\le\delta} d'(X(t),X(s)).\]

Since

\begin{multline*} \mathbb P(w_L(X_n,\delta)>\epsilon)\le \mathbb P(\bigcup_{i=0}^{\lfloor L/\delta\rfloor }\{3\text{sup}_{t\in[i\delta,(i+1)\delta]\cap[0,L]}d'(X_n(t),X_n(i\delta))>\epsilon\})\\[5pt] \le \sum_{i=0}^{\lfloor L/\delta\rfloor }\mathbb P(3\text{sup}_{t\in[i\delta,(i+1)\delta]\cap[0,L]}d'(X_n(t),X_n(i\delta))>\epsilon)\\[5pt] \le(\lfloor\frac{ L}{\delta}\rfloor +1)\text{sup}_{t\in[0,L]}\mathbb P(\text{sup}_{s\in[t,t+\delta]}d'(X_n(s),X_n(t))>\frac{\epsilon}{3}),\end{multline*}

the hypotheses imply that

(A.2) \begin{equation} \lim_{\delta\to0}\limsup_{n\to\infty}\mathbb P(w_L(X_n,\delta)>\epsilon)^{1/r_n}=0.\end{equation}

Let

\[ w^{\prime}_L(X,\delta)=\inf_{\substack{0=t_0<t_1<\ldots<t_k=L:\\[5pt] t_j-t_{j-1}>\delta}}\,\,\max_{j=1,\ldots,k}\text{sup}_{u,v\in [t_{j-1},t_j)}d'(X(u),X(v)).\]

Since $w^{\prime}_L(X,\delta)\le w_L(X,2\delta)$ , provided $\delta< L/2$ , by (A.2),

(A.3) \begin{equation} \lim_{\delta\to0}\limsup_{n\to\infty}\mathbb P(w_L'(X_n,\delta)>\epsilon)^{1/r_n}=0.\end{equation}

As each $X_n$ is a member of the Skorokhod space $\mathbb D(\mathbb R_+,\mathbb E')$ ,

\[ \lim_{\delta\to0}\mathbb P(w_L'(X_n,\delta)>\epsilon)=0,\]

so, by (A.3),

\[ \lim_{\delta\to0}\text{sup}_{n}\mathbb P(w^{\prime}_L(X_n,\delta)>\epsilon)^{1/r_n}=0.\]

Let $\{t_1, t_2,\ldots\}$ represent a dense subset of $\mathbb R_+$ such that $X_n(t_1),X_n(t_2),\ldots$ are exponentially tight of order $r_n$ . Since $\mathbf E'$ is complete and separable, every probability measure on $\mathbf E'$ is tight, so it may be assumed that there exist compact subsets $K_1,K_2,\ldots$ such that, for all n,

\[ \mathbb P(X_n(t_i)\notin K_i)^{1/r_n}<\frac{\epsilon}{2^i}, \qquad i\in\mathbb N.\]

Choose positive $\delta_1,\delta_2,\ldots$ such that

\[ \mathbb P(w_{L}'(X_n,\delta_i)>\frac{1}{2^i})^{1/r_n}<\frac{\epsilon}{2^i}, \qquad i\in\mathbb N.\]

The set

\[ A=\bigcap_{i\in \mathbb N}\{X:\,w^{\prime}_{L}(X,\delta_i)\le\frac{1}{2^i},X(t_i)\in K_i\}\]

has compact closure (see, e.g., [10, Theorem A2.2, p. 563]), and

\[ \mathbb P(X_n\notin A)^{1/r_n}\le\sum_{i=1}^\infty \mathbb P(w^{\prime}_{L}(X_n,\delta_i)>\frac{1}{2^i})^{1/r_n}+\sum_{i=1}^\infty \mathbb P(X_n(t_i)\notin K_i)^{1/r_n}<2\epsilon.\]

Thus, the sequence $X_n$ is exponentially tight of order $r_n$ in $\mathbb D(\mathbb R_+,\mathbf E')$ . Let $\Pi$ represent a deviability on $\mathbb D(\mathbb R_+,\mathbf E')$ that is an LD limit point of the distributions of $X_n$ . It is proved next that $\Pi(X)=0$ if X is a discontinuous function. Suppose X has a jump at t, i.e., $d'(X(t),X(t-))>0$ . Let $\rho$ denote a metric in $\mathbb D(\mathbb R_+,\mathbf E')$ . It may be assumed that if $\rho(X',X)<\delta$ , then there exists a continuous nondecreasing function $\lambda(t)$ such that $\text{sup}_{s\le t+1}d'(X'(\lambda(s)),X(s))< \delta$ and $\text{sup}_{s\le t+1}\lvert \lambda (s)-s\rvert<\delta$ ; see, e.g., [Reference Ethier and Kurtz6, Reference Jacod and Shiryaev9]. Then, assuming that $2\delta<t$ and $\delta<1$ , we have $\inf_{s'\in[s-\delta,s+\delta]}d'(X(s'),X(s))< \delta$ , where $s\in[0,t+1]$ . Note that

\begin{multline*} d'(X(t),X(t-))\le d'(X(t),X'(s_1))+d(X'(s_1),X'(t))+d'(X'(t),X'(s_2))\\[5pt] +d'(X'(s_2),X(s_3))+d'(X(s_3),X(t-)),\end{multline*}

so that, with $s_1\in[t-\delta,t+\delta]$ such that $d'(X'(s_1),X(t))<\delta$ , $s_3\in[t-\delta,t]$ such that $d'(X(s_3),X(t-))<\delta$ , and $s_2\in[s_3-\delta,s_3+\delta]$ such that $d'(X'(s_2),X(s_3))<\delta$ , we have

\[ d'(X(t),X(t-))< 3\delta+2\text{sup}_{s\in[t-2\delta,t+\delta]}d(X'(s),X'(t)),\]

which implies that, for $\delta$ small enough,

\[ \text{sup}_{s\in[t-2\delta,t+\delta]}d'(X'(s),X'(t))>\frac{d'(X(t),X(t-))}{3}.\]

Since

\begin{multline*} \text{sup}_{s\in[t-2\delta,t+\delta]}d'(X'(s),X'(t))\le \text{sup}_{s\in[t,t+\delta]}d'(X'(s),X'(t))\\[5pt] + \text{sup}_{s\in[t-2\delta,t-\delta]}d'(X'(s),X'(t))+\text{sup}_{s\in[t-\delta,t]}d'(X'(s),X'(t))\\[5pt] \le \text{sup}_{s\in[t,t+\delta]}d'(X'(s),X'(t)) + \text{sup}_{s\in[t-2\delta,t-\delta]}d'(X'(s),X'(t-2\delta))\\[5pt] +\text{sup}_{s\in[t-\delta,t]}d'(X'(s),X'(t-\delta))+d'(X'(t-\delta),X'(t-2\delta))+2d'(X'(t),X'(t-\delta)),\end{multline*}

for $\epsilon=d'(X(t),X(t-))/18$ ,

\begin{align*} &\mathbb P(\rho(X_n,X)<\delta)\le\mathbb P(\text{sup}_{s\in[t,t+\delta]}d'(X_n(s),X_n(t))>\epsilon)\\[5pt]&+ \mathbb P(\text{sup}_{s\in[t-2\delta,t-\delta]}d'(X_n(s),X_n(t-2\delta))>\epsilon)\\[5pt] &+\mathbb P(\text{sup}_{s\in[t-\delta,t]}d'(X_n(s),X_n(t-\delta))>\epsilon).\end{align*}

Therefore, assuming $L\ge t$ and $r_n\ge1$ ,

\begin{align*} & \mathbb P(\rho(X_n,X)<\delta)^{1/r_n}\le \mathbb P(\text{sup}_{s\in[t,t+\delta]}d'(X_n(s),X_n(t))>\epsilon)^{1/r_n}\\[5pt]&+\mathbb P(\text{sup}_{s\in[t-\delta,t]}d'(X_n(s),X_n(t-\delta))>\epsilon)^{1/r_n}\\[5pt] &+\mathbb P(\text{sup}_{s\in[t-2\delta,t-\delta]}d'(X_n(s),X_n(t-2\delta))>\epsilon)^{1/r_n}\\[5pt]&\le3\text{sup}_{t'\in[0,L]} \mathbb P(\text{sup}_{s\in[0,\delta]}d'(X_n(t'+s),X_n(t'))>\epsilon)^{1/r_n}.\end{align*}

Since

\[ \Pi(X)=\lim_{\delta\to0}\liminf_{n\to\infty}\mathbb P(\rho(X_n,X)<\delta)^{1/r_n},\]

(A.1) implies that $\Pi(X)=0$ .

The discussion below concerns the properties of the idempotent processes that feature prominently in the paper. All the processes assume values in $\mathbb R$ . The standard Wiener idempotent process, denoted by $W=(W(t),t\in\mathbb R_+)$ , is defined as an idempotent process with idempotent distribution

\[ \Pi^{W}(w)=\exp\bigg(\!-\frac{1}{2}\int_0^\infty\dot w(t)^2\,dt\bigg),\]

provided $w=(w(t),t\in\mathbb R_+)\in\mathbb D(\mathbb R_+,\mathbb R)$ is absolutely continuous and $w(0)=0$ , and $\Pi^{W}(w)=0$ otherwise. It is straightforward to show that W has stationary independent increments. The restriction to [0, t] produces a standard Wiener idempotent process on [0, t] which is specified by the idempotent distribution $ \Pi_t^{W}(w)=\exp\bigl(\!-1/2\,\int_0^t\dot w(s)^2\,ds\bigr)$ . The Brownian bridge idempotent process on [0, 1], denoted by $W^0=(W^0(x),\,x\in[0,1])$ , is defined as an idempotent process with the idempotent distribution

\[ \Pi^{W^0}(w^0)=\exp\bigg(\!-\frac{1}{2}\int_0^1\dot w^0(x)^2\,dx\bigg),\]

provided $w^0=(w^0(x),x\in[0,1])\in\mathbb D([0,1],\mathbb R)$ is absolutely continuous and $w^0(0)=w^0(1)=0$ , and $\Pi^{W^0}(w^0)=0$ otherwise. The Brownian sheet idempotent process on $ [0,1]\times\mathbb R_+$ denoted by $(B(x,t),x\in[0,1],\,t\in\mathbb R_+)$ is defined as a two-parameter idempotent process with the distribution

\[ \Pi^B(b)=\exp\bigg(\!-\frac{1}{2}\int_{[0,1]\times\mathbb R_+}{\dot b}(x,t)^2\,dx\,dt\bigg),\]

provided $b=(b(x,t),x\in[0,1], t\in\mathbb R_+)$ is absolutely continuous with respect to the Lebesgue measure on $ [0,1]\times\mathbb R_+$ and $b(x,0)=b(0,t)=0$ , and $\Pi^{B}(b)=0$ otherwise. The Kiefer idempotent process on $[0,1]\times\mathbb R_+$ , denoted by $(K(x,t),x\in[0,1],t\in\mathbb R_+)$ , is defined as a two-parameter idempotent process with the idempotent distribution

\[ \Pi^K(k)=\exp\bigl(\!-\frac{1}{2}\int_{[0,1]\times\mathbb R_+}\dot k(x,t)^2\,dx\,dt\bigr),\]

provided $k=(k(x,t),x\in[0,1], t\in\mathbb R_+)$ is absolutely continuous with respect to the Lebesgue measure on $ [0,1]\times\mathbb R_+$ and $k(0,t)=k(1,t)=k(x,0)=0$ , and $\Pi^{K}(k)=0$ otherwise. It is considered as an element of $\mathbb D(\mathbb R_+,\mathbb D([0,1],\mathbb R_+))$ . Furthermore, as the deviabilities that the idempotent processes W, $W^0$ , B, and K have discontinuous paths are equal to zero, these idempotent processes can be considered as having paths from $\mathbb C(\mathbb R_+,\mathbb R)$ , $\mathbb C([0,1],\mathbb R)$ , $\mathbb C(\mathbb R_+,\mathbb C([0,1],\mathbb R))$ , and $\mathbb C(\mathbb R_+,\mathbb C([0,1],\mathbb R))$ , respectively. Being LD limits of their stochastic prototypes, the idempotent processes introduced here have similar properties, as summarised in the next lemma.

Lemma 7.

  1. (i) For $x>0$ , the idempotent process $(B(x,t)/\sqrt{x},t\in\mathbb R_+)$ is a standard Wiener idempotent process.

  2. (ii) For $t>0$ , $(K(x,t)/\sqrt{t},x\in[0,1])$ is a Brownian bridge idempotent process. For $x\in(0,1)$ , the idempotent process $(K(x,t)/\sqrt{x(1-x)},t\in\mathbb R_+)$ is a standard Wiener idempotent process.

  3. (iii) The Kiefer idempotent process can be written as

    (A.4) \begin{equation} K(x,t)=-\int_0^x\frac{K(y,t)}{1-y}\,dy+B(x,t),x\in[0,1], \end{equation}
    where B(x,t) is a Brownian sheet idempotent process. Conversely, if B(x,t) is a Brownian sheet idempotent process and (A.4) holds, then K(x,t) is a Kiefer idempotent process. Similarly,
    \[ W^0(x)=-\int_0^x\frac{W^0(y)}{1-y}\,dy+W'(x),x\in[0,1],\]
    where W’(x) is a standard Wiener idempotent process.

Proof. Parts 1 and 2 are elementary. For instance,

\begin{align*} &\Pi\bigg(\bigg(\frac{K(x,t)}{\sqrt{t}},x\in[0,1]\bigg)=(w^0(x),x\in[0,1])\bigg)\\[5pt]&=\text{sup}_{k:\,k(x,t)=\sqrt{t}w^0(x),x\in[0,1]}\exp\bigg(\!-\frac{1}{2}\int_{\mathbb R_+}\int_0^1\dot k(x,s)^2\,dx\,ds\bigg). \end{align*}

An application of the Cauchy–Schwarz inequality shows that the optimal k is

\[ k(x,s)=\frac{s\wedge t}{\sqrt{t}}\,w^0(x).\]

As for $(K(x,t)/\sqrt{x(1-x)},t\in\mathbb R_+)$ , the optimal trajectory $(k(y,t),y\in[0,1])$ to get to $w(t)\sqrt{x(1-x)}$ at x is

\[ k(y,t)= \begin{cases} w(t)\sqrt{x(1-x)}\,\displaystyle\frac{y}{x} &\text{ if }y\in[0,x],\\[9pt] w(t)\sqrt{x(1-x)}\,\displaystyle\frac{1-y}{1-x} &\text{ if }y\in[x,1]. \end{cases}\]

To prove Part 3, it suffices to show that if

(A.5) \begin{equation} k(x,t)=-\int_0^x\frac{k(y,t)}{1-y}\,dy+b(x,t),\end{equation}

with k and b being absolutely continuous and with $\Pi^B(b)>0$ , then $k(1,t)=0$ and

(A.6) \begin{equation} \int_0^\infty\int_0^1 \dot k(x,t)^2\,dx\,dt= \int_0^\infty\int_0^1 \dot b(x,t)^2\,dx\,dt .\end{equation}

Solving (A.5) yields, for $x<1$ ,

(A.7) \begin{equation} k(x,t)=(1-x)\int_0^x\frac{b_y(y,t)}{1-y}\,dy,\end{equation}

where $b_y(y,t)=\int_0^t \dot b(s,y)\,ds$ . By the Cauchy–Schwarz inequality,

\begin{multline*} \lvert \int_0^x\frac{b_y(y,t)}{1-y}\,dy\rvert\le\sqrt{\int_0^x\frac{1}{(1-y)^2}\,dy}\sqrt{\int_0^1b_y(y,t)^2dy}\le\frac{\sqrt{t}}{\sqrt{1-x}}\,\sqrt{\int_{ [0,1]\times \mathbb R_+}\dot b(y,s)^2\,dy\,ds}.\end{multline*}

It follows that $k(x,t)\to0$ as $x\to1$ , provided $\Pi^B(b)>0$ .

Let $k_t(x,t)=\int_0^x\dot k(y,t)\,dy$ . We show first that if $\Pi^B(b)>0$ , then, a.e.,

(A.8) \begin{equation} \frac{1}{1-x}\, k_t(x,t)^2 \to0\text{ as } x\to 1.\end{equation}

Given arbitrary $a>0$ , by (A.7),

\begin{multline*} \frac{k_t(x,t)^2}{1-x}=(1-x)\Bigl(\int_0^x\frac{\dot b(y,t)\,dy}{1-y}\Bigr)^2\le 2(1-x)a^2\Bigl(\int_0^x\frac{dy}{1-y}\Bigr)^2\\[5pt] +2(1-x)\int_0^x\frac{dy}{(1-y)^2}\int_0^x\dot b(y,t)^2\,\mathbf{1}_{\{\lvert \dot b(y,t)\rvert> a\}}\,\,dy\\[5pt] \le 2(1-x)\lvert \ln(1-x)\rvert^2a^2+2\int_0^x\dot b(y,t)^2\,\mathbf{1}_{\{\lvert \dot b(y,t)\rvert> a\}}\,\,dy.\end{multline*}

Since $\int_0^1\dot b(y,t)^2\,dy<\infty$ a.e., the latter right-hand side tends to 0 a.e., as $x\to1$ and $a\to\infty$ .

Next we prove (A.6). By (A.5),

(A.9) \begin{equation} \int_0^\infty \int_0^1\dot b(x,t)^2\,dx\,dt=\int_0^\infty \int_0^1\Bigg( \dot k(x,t)^2+2\dot k(x,t)\frac{k_t(x,t)}{1-x}+\bigg(\frac{k_t(x,t)}{1-x}\bigg)^2\Bigg)\,dx\,dt.\end{equation}

Integration by parts with the account of (A.8) yields, for almost all t,

\[ \int_0^1\dot k(x,t)\frac{k_t(x,t)}{1-x}\,dx=- \int_0^1 k_t(x,t)\bigg(\frac{\dot k(x,t)}{1-x}+\frac{ k_t(x,t)}{(1-x)^2}\bigg)\,dx,\]

so that

\[ \int_0^1\dot k(x,t)\frac{k_t(x,t)}{1-x}\,dx=-\frac{1}{2}\,\int_0^1\frac{ k_t(x,t)^2}{(1-x)^2}\,dx.\]

Recalling (A.9) implies (A.6).

Appendix B. A nonlinear renewal equation

This section is concerned with the properties of the equation

(B.1) \begin{equation} g(t)=f(t)+\int_0^tg(t-s)^+\,dF(s), t\in\mathbb R_+.\end{equation}

It is assumed that f(t) is a locally bounded measurable function and that F(t) is a continuous distribution function on $\mathbb R_+$ with $F(0)=0$ . The existence and uniqueness of an essentially locally bounded solution g(t) to (B.1) follows from Lemma B.2 in [Reference Puhalskii and Reed21].

Lemma 8. If the functions f and F are absolutely continuous with respect to Lebesgue measure, then the function g is absolutely continuous too.

Proof. Use Picard iterations. Let $g_0(t)=f(t)$ and

(B.2) \begin{equation} g_{k}(t)=f(t)+\int_0^tg_{k-1}(t-s)^+\,dF(s). \end{equation}

The functions $g_k$ are seen to be continuous. Let $\epsilon>0$ , $T>0$ , and $0\le t_0\le t_1\le\ldots\le t_l\le T$ . Since $g_k\to g$ locally uniformly (see [Reference Puhalskii and Reed21, Lemma B.1]), the function g is continuous and $\text{sup}_k\text{sup}_{t\in[0,T]}\lvert g_k(t)\rvert\le M$ for some $M>0$ . Note that

(B.3) \begin{multline} \lvert \int_0^{t_i}g_{k-1}(t_i-s)^+\,dF(s)-\int_0^{t_{i-1}}g_{k-1}(t_{i-1}-s)^+\,dF(s)\rvert\\[5pt] \le \int_{t_{i-1}}^{t_i}\lvert g_{k-1}(t_i-s)\rvert\,dF(s)+\int_0^{T}\,\mathbf{1}_{\{s\le t_{i-1}\}}\,\lvert g_{k-1}(t_i-s)-g_{k-1}(t_{i-1}-s)\rvert\,dF(s).\end{multline}

Let

\[ \psi(\delta)=\text{sup}_{0\le t_0\le t_1\le\ldots\le t_l\le T}\{\sum_{i=1}^l \lvert f(t_i)-f(t_{i-1})\rvert+M\sum_{i=1}^l \lvert F(t_i)-F(t_{i-1})\rvert:\,\sum_{l=1}^l(t_i-t_{i-1})\le\delta\}\]

and

\[ \phi_k(\delta)=\text{sup}_{0\le t_0\le t_1\le\ldots\le t_l\le T}\{\sum_{i=1}^l \lvert g_k(t_i)-g_k(t_{i-1})\rvert:\,\sum_{l=1}^l(t_i-t_{i-1})\le\delta\}.\]

By (B.2) and (B.3), for $k\ge1$ ,

(B.4) \begin{equation} \phi_k(\delta)\le \psi(\delta)+\phi_{k-1}(\delta)F(T).\end{equation}

Let

\[ \phi(\delta)=\text{sup}_{0\le t_0\le t_1\le\ldots\le t_l\le T}\{\lvert g(t_i)-g(t_{i-1})\rvert:\,\sum_{l=1}^l(t_i-t_{i-1})\le\delta\}.\]

Suppose that $F(T)<1$ . Since $g_k\to g$ locally uniformly, as $k\to\infty$ , $\phi_k(\delta)\to\phi(\delta)$ . Letting $k\to\infty$ in (B.4) implies that $\phi(\delta)\le\psi(\delta)/(1-F(T))$ , so that $\phi(\delta)\to0$ , as $\delta\to0$ . Hence g(t) is absolutely continuous on [0, T]. Next, as in [Reference Puhalskii and Reed21], write, for $t\in[0,T]$ ,

\[ g(t+T)=f(t+T)+\int_t^{t+T}g(t+T-s)\,dF(s)+\int_0^tg(t+T-s)\,d F(s).\]

By what has been proved, the sum of the first two terms on the right-hand side is an absolutely continuous function of t on [0, T]. The preceding argument implies that $g(t+T)$ is absolutely continuous in t on [0, T]. Iterating the argument proves the absolute continuity of g(t) on $\mathbb R_+$ .

Funding information

There are no funding bodies to thank in relation to the creation of this article.

Competing interests

There were no competing interests to declare which arose during the preparation or publication process of this article.

References

Aghajani, R. and Ramanan, K. (2020). The limit of stationary distributions of many-server queues in the Halfin-Whitt regime. Math. Operat. Res. 45, 10161055.CrossRefGoogle Scholar
Asmussen, S. (2003). Applied Probability and Queues, 2nd edition. Springer, New York.Google Scholar
Atkinson, K. E. (1997). The Numerical Solution of Integral Equations of the Second Kind. Cambridge University Press.CrossRefGoogle Scholar
Billingsley, P. (1968). Convergence of Probability Measures. John Wiley, New York.Google Scholar
Ekeland, I. and Temam, R. (1976). Convex Analysis and Variational Problems. North-Holland, Amsterdam.Google Scholar
Ethier, S. N. and Kurtz, T. G. (1986). Markov Processes: Characterization and Convergence. John Wiley, Hoboken, NJ.CrossRefGoogle Scholar
Feng, J. and Kurtz, T. G. (2006). Large Deviations for Stochastic Processes. American Mathematical Society, Providence, RI.CrossRefGoogle Scholar
Halfin, S. and Whitt, W. (1981). Heavy-traffic limits for queues with many exponential servers. Operat. Res. 29, 567588.CrossRefGoogle Scholar
Jacod, J. and Shiryaev, A. N. (1987). Limit Theorems for Stochastic Processes. Springer, Berlin.CrossRefGoogle Scholar
Kallenberg, O. (2002). Foundations of Modern Probability, 2nd edition. Springer, New York.CrossRefGoogle Scholar
Kaspi, H. and Ramanan, K. (2013). SPDE limits of many-server queues. Ann. Appl. Prob. 23, 145229.CrossRefGoogle Scholar
Krichagina, E. V. and Puhalskii, A. A. (1997). A heavy-traffic analysis of a closed queueing system with a $GI/\infty$ service center. Queueing Systems 25, 235280.CrossRefGoogle Scholar
Liptser, R. S. and Shiryayev, A. N. (1989). Theory of Martingales. Kluwer, Dordrecht.CrossRefGoogle Scholar
Liptser, R. S. and Shiryaev, A. N. (2001). Statistics of Random Processes: I. General Theory. Springer, Berlin.Google Scholar
Petrov, V. V. (1987). Limit Theorems for Sums of Independent Random Variables, 2nd edition. Nauka, Moscow (in Russian).Google Scholar
Puhalskii, A. (1991). On functional principle of large deviations. In Proceedings of the Bakuriani Colloquium in Honour of Yu. V. Prohorov, Bakuriani, Georgia, USSR, 24 February–4 March, 1990 (New Trends in Probability and Statistics 1), eds V. Sazonov and Shervashidze, T., De Gruyter, Amsterdam, pp. 198–218.Google Scholar
Puhalskii, A. (1992). Weak convergence theory approach to large deviations. In Large Deviations and Applications (Workshop Report 1992.51), Mathematisches Forschungsinstitut Oberwolfach, p. 16.Google Scholar
Puhalskii, A. (1999). Moderate deviations for queues in critical loading. Queueing Systems 31, 359392.CrossRefGoogle Scholar
Puhalskii, A. (2001). Large Deviations and Idempotent Probability. Chapman and Hall/CRC, Boca Raton.CrossRefGoogle Scholar
Puhalskii, A. A. (2022). Moderate deviation asymptotics of the $GI/G/n$ queue in the Halfin–Whitt regime. Queueing Systems 100, 341343.CrossRefGoogle Scholar
Puhalskii, A. A. and Reed, J. E. (2010). On many-server queues in heavy traffic. Ann. Appl. Prob. 20, 129195.CrossRefGoogle Scholar
Pukhalskii, A. A. (1994). On the theory of large deviations. Theory Prob. Appl. 38, 490497.CrossRefGoogle Scholar
Reed, J. (2009). The G/GI/N queue in the Halfin–Whitt regime. Ann. Appl. Prob. 19, 22112269.CrossRefGoogle Scholar
Whitt, W. (1980). Some useful functions for functional limit theorems. Math. Operat. Res. 5, 6785.CrossRefGoogle Scholar
Whittle, P. (1960). Bounds for the moments of linear and quadratic forms in independent variables. Theory Prob. Appl. 5, 302305.CrossRefGoogle Scholar