Abstract
We consider a long-run impulse control problem for a generic Markov process with a multiplicative reward functional. We construct a solution to the associated Bellman equation and provide a verification result. The argument is based on the probabilistic properties of the underlying process combined with the Krein-Rutman theorem applied to the specific non-linear operator. Also, it utilises the approximation of the problem in the bounded domain and with the help of the dyadic time-grid.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
Impulse control constitutes a versatile framework for controlling real-life stochastic systems. In this type of control, a decision-maker determines intervention times and instantaneous after-intervention states of the controlled process. By doing so, one can affect a continuous time phenomenon in a discrete time manner. Consequently, impulse control attracted considerable attention in the mathematical literature; see e.g. [7, 13, 31] for classic contributions and [6, 14, 24, 26] for more recent results. In addition to generic mathematical properties, impulse control problems were studied with reference to specific applications including i.a. controlling exchange rates, epidemics, and portfolios with transaction costs; see e.g. [23, 30, 32] and references therein.
When looking for an optimal impulse control strategy, one must decide on the optimality criterion. Recently, considerable attention was paid to the so-called risk-sensitive functional given, for any \(\gamma \in \mathbb {R}\), by
where Z is a (random) payoff corresponding to a chosen control strategy; see [19] for a seminal contribution. This functional with \(\gamma =0\) corresponds to the usual linear criterion and the case \(\gamma <0\) is associated with risk-averse preferences; see [8] for a comprehensive overview. Also, the functional with \(\gamma >0\) could be linked to the asymptotics of the power utility function; see [36] for details. Recent comprehensive discussion on the long-run version with \(\mu ^\gamma \) could be found in [10]. We refer also to [28] and references therein for a discussion on the connection between (1.1) and the duality of the large deviations-based criteria.
In this paper we focus on the use of the functional \(\mu ^\gamma \) with \(\gamma >0\). More specifically, we consider the impulse control problem for some continuous time Markov process and construct a solution to the associated Bellman equation which characterises an optimal impulse control strategy. To do this, we study the family of impulse control problems in bounded domains and then extend the analysis to the generic locally compact state space. This idea was used in [2], where PDEs techniques were applied to obtain the characterisation of the controlled diffusions in the risks-sensitive setting. A similar approximation for the the average cost per unit time problem was considered in [37].
The main contribution of this paper is a construction of a solution to the Bellman equation associated with the problem, see Theorem 5.1 for details. It should be noted that we get a bounded solution even though the state space could be unbounded and we assume virtually no ergodicity conditions for the uncontrolled process. Also, note that present results for \(\gamma >0\) complement our recent findings on the impulse control with the risk-averse preferences; see [29] for the dyadic case and [20] for the continuous time framework. Nevertheless, it should be noted that the techniques for \(\gamma <0\) and \(\gamma >0\) are substantially different and it is not possible to directly transform the results in one framework to the other; see e.g. [21, 25] for further discussion.
The structure of this paper is as follows. In Sect. 2 we formally introduce the problem, discuss the assumptions and, in Theorem 2.3, provide a verification argument. Next, in Sect. 3 we consider an auxiliary dyadic problem in a bounded domain and in Theorem 3.1 we construct a solution to the corresponding Bellman equation. This is used in Sect. 4 where we extend our analysis to the unbounded domain with the dyadic time-grid; see Theorem 4.2 for the main result. Next, in Sect. 5 we finally construct a solution to the Bellman equation for the original problem; see Theorem 5.1. Finally, in Appendix A we discuss some properties of the optimal stopping problems that are used in this paper.
2 Preliminaries
Let \(X=(X_t)_{t\ge 0}\) be a continuous time standard Feller–Markov process on a filtered probability space \((\Omega , \mathcal {F}, (\mathcal {F}_t), \mathbb {P})\). The process X takes values in a locally compact separable metric space E endowed with a metric \(\rho \) and the Borel \(\sigma \)-field \(\mathcal {E}\). With any \(x\in E\) we associate a probability measure \(\mathbb {P}_x\) describing the evolution of the process X starting in x; see Section 1.4 in [33] for details. Also, we use \(\mathbb {E}_x\), \(x\in E\), and \(P_t(x,A):=\mathbb {P}_x[X_t\in A]\), \(t\ge 0\), \(x\in E\), \(A\in \mathcal {E}\), for the corresponding expectation operator and the transition probability, respectively. By \(\mathcal {C}_b(E)\) we denote the family of continuous bounded real-valued functions on E. Also, to ease the notation, by \(\mathcal {T}\), \(\mathcal {T}_x\), and \(\mathcal {T}_{x,b}\) we denote the families of stopping times, \(\mathbb {P}_x\) a.s. finite stopping times, and \(\mathbb {P}_x\) a.s. bounded stopping times, respectively. Also, for any \(\delta >0\), by \(\mathcal {T}^\delta \subset \mathcal {T}\), \(\mathcal {T}_x^\delta \subset \mathcal {T}_x\), and \(\mathcal {T}_{x,b}^\delta \subset \mathcal {T}_{x,b}\), we denote the respective subfamilies of dyadic stopping times, i.e., those taking values in the set \(\{0,\delta , 2\delta , \ldots \}\cup \{\infty \}\). Finally, note that in this paper we follow the conventions \(\mathbb {N}:=\{0,1,2, \ldots \}\) and \(\mathbb {R}_{-}:=(-\infty ,0]\).
Throughout this paper we fix some compact set \(U\subseteq E\) and we assume that a decision-maker is allowed to shift the controlled process to U. This is done with the help of an impulse control strategy, i.e. a sequence \(V:=(\tau _i,\xi _i)_{i=1}^\infty \), where \((\tau _i)\) is an increasing sequence of stopping times and \((\xi _i)\) is a sequence of \(\mathcal {F}_{\tau _i}\)-measurable after-impulse states with values in U. With any starting point \(x\in E\) and a strategy V we associate a probability measure \(\mathbb {P}_{(x,V)}\) for the controlled process Y. Under this measure, the process starts at x and follows its usual (uncontrolled) dynamics up to the time \(\tau _1\). Then, it is immediately shifted to \(\xi _1\) and starts its evolution again, etc. More formally, we consider a countable product of filtered spaces \((\Omega , \mathcal {F}, (\mathcal {F}_t))\) and a coordinate process \((X_t^1, X_t^2, \ldots )\). Then, we define the controlled process Y as \(Y_t:=X_t^i\), \(t\in [\tau _{i-1},\tau _i)\) with the convention \(\tau _0\equiv 0\). Under the measure \(\mathbb {P}_{(x,V)}\) we get \(Y_{\tau _i}=\xi _i\); we refer to Chapter V in [31] for the construction details; see also Appendix in [12] and Section 2 in [34]. A strategy \(V=(\tau _i,\xi _i)_{i=1}^\infty \) is called admissible if for any \(x\in E\) we get \(\mathbb {P}_{(x,V)}[\lim _{n\rightarrow \infty }\tau _n=\infty ]=1\). The family of admissible impulse control strategies is denoted by \(\mathbb {V}\). Also, note that, to simplify the notation, by \(Y_{\tau _i^-}:=X_{\tau _i}^i\), \(i\in \mathbb {N}_{*}\), we denote the state of the process right before the ith impulse (yet, possibly, after the jump).
In this paper we study the asymptotics of the impulse control problem given by
where, for any \(x\in E\) and \(V\in \mathbb {V}\), we set
with f denoting the running reward function and c being the shift-cost function, respectively. Note that this could be seen as a long-run standardised version of the functional (1.1) with \(\gamma >0\) applied to the impulse control framework. Here, the standardisation refers to the fact that we do not use directly the parameter \(\gamma \) (apart from its sign). Also, the problem is of the long-run type, i.e. the utility is averaged over time which improves the stability of the results.
The analysis in this paper is based on the approximation of the problem in a bounded domain. Thus, we fix a sequence \((B_m)_{m\in \mathbb {N}}\) of compact sets satisfying \(B_m\subset B_{m+1}\) and \(E=\bigcup _{m=0}^\infty B_m\). Also, we assume that \(U\subset B_0\). Next, we assume the following conditions.
- (\(\mathcal {A}1\)):
-
(Reward/cost functions). The map \(f:E\mapsto \mathbb {R}_{-}\) is a continuous and bounded. Also, the map \(c:E\times U \mapsto \mathbb {R}_{-}\) is continuous, bounded, and strictly non-positive, and satisfies the triangle inequality, i.e. for some \(c_0<0\), we have
$$\begin{aligned} 0>c_0\ge c(x,\xi )\ge c(x,\eta )+c(\eta ,\xi ), \quad x\in E,\, \xi ,\eta \in U. \end{aligned}$$(2.3)Also, we assume that c satisfies the uniform limit at infinity condition
$$\begin{aligned} \lim _{\Vert x\Vert , \Vert y\Vert \rightarrow \infty }\sup _{\xi \in U} |c(x,\xi )-c(y,\xi )|=0. \end{aligned}$$(2.4) - (\(\mathcal {A}2\)):
-
(Transition probability continuity). For any \(t>0\), the transition probability \(P_t\) is continuous with respect to the total variation norm, i.e. for any sequence \((x_n)\subset E\) converging to \(x\in E\), we have
$$\begin{aligned} \lim _{n\rightarrow \infty }\sup _{A\in \mathcal {E}}|P_t(x_n,A)-P_t(x,A)|=0. \end{aligned}$$ - (\(\mathcal {A}3\)):
-
(Distance control). For any compact set \(\Gamma \subset E\), \(t_0>0\), and \(r_0>0\), we have
$$\begin{aligned} \lim _{r\rightarrow \infty }M_{\Gamma }(t_0,r)=0,\qquad \lim _{t\rightarrow 0} M_{\Gamma }(t,r_0) =0, \end{aligned}$$(2.5)where \(M_{\Gamma }(t,r):= \sup _{x\in \Gamma } \mathbb {P}_x[\sup _{s\in [0,t]} \rho (X_s,X_0)\ge r]\), \(t,r>0\).
- (\(\mathcal {A}4\)):
-
(Recurrence of open sets). For any \(m\in \mathbb {N}\), \(x\in B_m\), \(\delta >0\), and any open set \(\mathcal {O}\subset B_m\), we have
$$\begin{aligned} \mathbb {P}_x\left[ \cup _{i=1}^\infty \{X_{i\delta }\in \mathcal {O}\}\right] =1. \end{aligned}$$Also, we assume that for any \(x\in E\), \(\delta >0\), and \(m\in \mathbb {N}\), we have
$$\begin{aligned} \mathbb {P}_x[\tau _{B_m}<\infty ]=1 \end{aligned}$$(2.6)where \(\tau _{B_m}:=\delta \inf \{k\in \mathbb {N}:X_{k\delta }\notin B_m\}\).
Before we proceed, let us comment on these assumptions.
Assumption (A1) states typical reward/cost functions conditions. In particular, the non-positivity assumption for f is merely a technical normalisation. Indeed, for a generic \(\tilde{f}\in \mathcal {C}_b(E)\) we may set \(f(\cdot ):=\tilde{f}(\cdot )-\Vert \tilde{f}\Vert \le 0\) to get
where \(J^{f}\) denotes the version of the functional J from (2.2) corresponding to the running reward function f. Next, the conditions for c are standard requirements for the shift-cost functions in the impulse control setting. In particular, inequality (2.3) implies that a decision maker considering an impulse from x to \(\eta \) followed by an immediate impulse from \(\eta \) to \(\xi \) should directly shift the process from x to \(\xi \). This condition is used in Theorem 3.1. Also, (2.4) states that, at infinity, the cost function is almost constant. This is used to extract a (globally) uniformly convergent subsequence of a specific function sequence; see the proofs of Theorem 4.2 and Theorem 5.1. Finally, note that all the assumptions regarding the shift-cost functions are satisfied e.g. for c of the form \(c(x,\xi )=h(\rho (x,\xi ))+c_0\), \(x\in E\), \(\xi \in U\), where \(c_0<0\), the map \(h:\mathbb {R}\rightarrow \mathbb {R}_{-}\) is continuous, bounded, non-increasing and superadditive (i.e. satisfying \(h(x+y)\ge h(x)+h(y)\), \(x,y\in \mathbb {R}\)), and \(\rho \) denotes the underlying metric on E. For example, we may set \(h(x):=-\min (x,K)\), \(x\in \mathbb {R}\), for some constant \(K>0\).
Assumption (A2) states that the transition probabilities \(\mathbb {P}_t(x,\cdot )\) are continuous with respect to the total variation norm. Note that this directly implies that the transition semi-group associated to X is strong Feller, i.e. for any \(t>0\) and a bounded measurable map \(h:E\mapsto \mathbb {R}\), the map \(x\mapsto \mathbb {E}_x[h(X_t)]\) is continuous and bounded.
Assumption (A3) quantifies distance control properties of the underlying process. It states that, for a fixed time horizon, the process with a high probability stays close to its starting point and, with a fixed radius, with a high probability it does not leave the corresponding ball with a sufficiently short time horizon. Note that these properties are automatically satisfied if the transition semi-group is \(\mathcal {C}_0\)-Feller; see Proposition 2.1 in [26] and Proposition 6.4 in [5] for details.
Assumption (A4) states a form of the recurrence property of the process X. It requires that the process visits a sufficiently rich family of sets with unit probability.
It should be noted that the process-related Assumptions (A2)–(A4) are satisfied e.g for non-degenerate ergodic diffusions. Here, the non-degeneracy refers to the existence of a continuous and bounded density \(p_t\) with respect to some measure \(\nu _t\) such that the transition probability satisfies
This directly implies (A2). Next, using Theorem 6.7.2 from [1] we get that diffusions (and, more generally, solutions to stochastic differential equations driven by Lévy processes) are \(\mathcal {C}_0\)-Feller, which combined with Proposition 2.1 in [26] and Proposition 6.4 in [5] shows that (A3) is satisfied. Finally, the ergodicity guarantees (A4).
To solve (2.1), we show the existence of a solution to the impulse control Bellman equation, i.e. a function \(w\in \mathcal {C}_b(E)\) and a constant \(\lambda \in \mathbb {R}\) satisfying
where the operator M is given by
note that in (2.7), the uncontrolled Markov process is considered.
We start with a simple observation giving a lower bound for the constant \(\lambda \) from (2.7). To do this, we define the semi-group type by
We refer to e.g. Proposition 1 in [35] and the discussion following Formula (10.2.2) in [18] for further properties of r(f).
Lemma 2.1
Let \((w,\lambda )\) be a solution to (2.7). Then, we get \(\lambda \ge r(f)\).
Proof
From (2.7), for any \(T\ge 0\), we get
Thus, using the boundedness of w and Mw, we get
Consequently, dividing both hand-sides by T and letting \(T\rightarrow \infty \), we get \(0\ge r(f-\lambda )\), which concludes the proof. \(\square \)
Let us now link a solution to (2.7) with the optimal value and an optimal strategy for (2.1). To ease the notation, we recursively define the strategy \(\hat{V}:=(\hat{\tau }_i,\hat{\xi }_i)_{i=1}^\infty \) for \(i\in \mathbb {N}{\setminus }\{0\}\) by
where \(\hat{\tau }_0:=0\) and \(\xi _0\in U\) is some fixed point. First, we show that \(\hat{V}\) is a proper strategy.
Proposition 2.2
The strategy \(\hat{V}\) given by (2.9) is admissible.
Proof
To ease the notation, we define \(N(0,T):=\sum _{i=1}^\infty 1_{\{\hat{\tau }_i\le T\}}\), \(T\ge 0\). We fix some \(T> 0\) and \(x\in E\), and show that we get
Recalling (2.9), on the event \(A:=\{\lim _{i\rightarrow \infty }\hat{\tau }_i<+\infty \}\), for any \(n\in \mathbb {N}\), \(n\ge 1\), we get \(w(X_{\hat{\tau }_n}^n)=Mw(X_{\hat{\tau }_n}^n)=c(X_{\hat{\tau }_n}^n,X_{\hat{\tau }_n}^{n+1})+w(X_{\hat{\tau }_n}^{n+1})\). Also, recalling that \(c(x,\xi )\le c_0<0\), \(x\in E\), \(\xi \in U\), for any \(n\in \mathbb {N}\), \(n\ge 1\), we have \(w(X_{\hat{\tau }_n}^{n+1})-w(X_{\hat{\tau }_n}^{n})=-c(X_{\hat{\tau }_n}^{n},X_{\hat{\tau }_n}^{n+1})\ge -c_0>0\). Using this observation and Assumption (A3), we estimate the distance between consecutive impulses which will be used to prove (2.10). More specifically, for any \(k,m\in \mathbb {N}\), \(k,m\ge 1\), we get
it should be noted that the specific values for k and m will be determined later. Using the continuity of w we may find \(K>0\) such that \(\sup _{x,y\in U}(w(x)-w(y))\le K\). Let \(m\in \mathbb {N}\) be big enough to get \(-(m-1)\frac{c_0}{2}>K\). Thus, noting that \(X_{\hat{\tau }_{k+m-1}}^{k+m}, X_{\hat{\tau }_k}^{k+1}\in U\), we have \((w(X_{\hat{\tau }_{k+m-1}}^{k+m})-w(X_{\hat{\tau }_k}^{k+1}))\le K<-(m-1)\frac{c_0}{2}\). Consequently, recalling (2.11), on A, we get
Recalling the compactness of U and the continuity of w we may find \(r>0\) such that for any \(x\in U\) and \(y\in E\) satisfying \(\rho (x,y)<r\) we get \(|w(x)-w(y)|<-\frac{c_0}{2}\). Let us now consider the family of events
and note that, for any \(k\in \mathbb {N}\), \(k\ge 1\), on \(B_k\cap A\) we have \(\sum _{n=k}^{k+m-2}(w(X_{\hat{\tau }_n}^{n+1})-w(X_{\hat{\tau }_{n+1}}^{n+1}))< -(m-1)\frac{c_0}{2}\). Thus, recalling (2.12), for any \(k\in \mathbb {N}\), \(k\ge 1\), we get \(\mathbb {P}_{(x_0,\hat{V})}[ B_k\cap A]=0\) and, in particular, we have
Let us now show that \(\limsup _{k\rightarrow \infty }\mathbb {P}_{(x_0,\hat{V})}[B_k^c\cap \{N(0,T)=\infty \}]=0\). Noting that \(\{N(0,T)=\infty \}=\{\lim _{i\rightarrow \infty }\hat{\tau }_i\le T\}\), for any \(t_0>0\) and \(k\in \mathbb {N}\), \(k\ge 1\), we get
Using Assumption (A3), for any \(\varepsilon >0\), we may find \(t_0>0\), such that
Thus, using the strong Markov property and noting that \(X_{\hat{\tau }_n}^{n+1}\in U\), for any \(k\in \mathbb {N}\), \(k\ge 1\), we get
Recalling that \(\varepsilon >0\) was arbitrary, for any \(k\in \mathbb {N}\), \(k\ge 1\), we get
Now, to ease the notation, let \(C_k:=\bigcup _{n=k}^{\infty }\{\hat{\tau }_{n+1}-\hat{\tau }_n> t_0\}\cap \{\lim _{i\rightarrow \infty }\hat{\tau }_i\le T\}\), \(k\in \mathbb {N}\), \(k\ge 1\), and note that \(C_{k+1}\subset C_k\), \(k\in \mathbb {N}\), \(k\ge 1\). We show that
For the contradiction, assume that \(\lim _{k\rightarrow \infty }\mathbb {P}_{(x_0,\hat{V})}\left[ C_k\right] >0\). Consequently, we get \(\mathbb {P}_{(x_0,\hat{V})}\left[ \bigcap _{k=1}^\infty C_k\right] >0\). Note that for any \(\omega \in \bigcap _{k=1}^\infty C_k\) we have \(\lim _{i\rightarrow \infty }\hat{\tau }_i(\omega )\le T\). In particular, we may find \(i_0\in \mathbb {N}\) such that for any \(n\ge i_0\) we get \(\hat{\tau }_{n+1}(\omega )-\hat{\tau }_n(\omega )\le \frac{t_0}{2}\). This leads to the contradiction as from the fact that \(\omega \in \bigcap _{k=1}^\infty C_k\) we also get
Consequently, we get \(\lim _{k\rightarrow \infty }\mathbb {P}_{(x_0,\hat{V})}\left[ C_k\right] =0\) and, in particular, we get
Hence, recalling (2.15) and (2.18), we get
Thus, recalling (2.14), for any \(k\in \mathbb {N}\), \(k\ge 1\), we obtain
and letting \(k\rightarrow \infty \), we conclude the proof of (2.10). \(\square \)
Now, we show the verification result linking (2.7) with the optimal value and an optimal strategy for (2.1).
Theorem 2.3
Let \((w,\lambda )\) be a solution to (2.7) with \(\lambda >r(f)\). Then, we get
where the strategy \(\hat{V}\) is given by (2.9).
Proof
The proof is based on the argument from Theorem 4.4 in [20] thus we show only an outline. First, we show that \(\lambda =J(x,\hat{V})\), \(x\in E\), where the strategy \(\hat{V}\) is given by (2.9). Let us fix \(x\in E\). Then, combining the argument used in Lemma 7.1 in [5] and Proposition A.3, we get that the process
is a \(\mathbb {P}_{(x,\hat{V})}\)-martingale. Noting that on the event \(\{\hat{\tau }_{k+1}<T\}\) we get \(w(X^{k+1}_{\hat{\tau }_{k+1}})=Mw(X^{k+1}_{\hat{\tau }_{k+1}})=c(X^{k+1}_{\hat{\tau }_{k+1}},\hat{\xi }_{k+1})+w(\hat{\xi }_{k+1})\), \(k\in \mathbb {N}\), for any \(n\in \mathbb {N}\) we recursively get
Recalling Proposition 2.2 we get \(\hat{\tau }_n\rightarrow \infty \) as \(n\rightarrow \infty \). Thus, letting \(n\rightarrow \infty \) in (2.19) and using Lebesgue’s dominated convergence theorem we get
Thus, recalling the boundedness of w, taking the logarithm of both sides, dividing by T, and letting \(T\rightarrow \infty \) we obtain
Second, let us fix some \(x\in E\) and an admissible strategy \(V=(\xi _i,\tau _i)_{i=1}^\infty \in \mathbb {V}\). We show that \(\lambda \ge J(x,V)\). Using the argument from Lemma 7.1 in [5] and Proposition A.3, we get that the process
is a \(\mathbb {P}_{(x,V)}\)-supermartingale. Noting that on the event \(\{\tau _{k+1}<T\}\) we have
for any \(n\in \mathbb {N}\) we recursively get
Recalling the admissibility of V, we get \(\tau _n\rightarrow \infty \) as \(n\rightarrow \infty \). Thus, letting \(n\rightarrow \infty \) in (2.20) and using Fatou’s lemma, we get
Thus, taking the logarithm of both sides, dividing by T, and letting \(T\rightarrow \infty \), we get
which concludes the proof. \(\square \)
In the following sections we construct a solution to (2.7). In the construction we approximate the underlying problem using the dyadic time-grid. Also, we consider a version of the problem in the bounded domain.
3 Dyadic Impulse Control in a Bounded Set
In this section we consider a version of (2.1) with a dyadic-time-grid and obligatory impulses when the process leaves some compact set. In this way, we construct a solution to the bounded-domain dyadic counterpart of (2.7). More specifically, let us fix some \(\delta >0\) and \(m\in \mathbb {N}\). We show the existence of a map \(w_{\delta }^m\in \mathcal {C}_b(B_m)\) and a constant \(\lambda _\delta ^m\in \mathbb {R}\) satisfying
In fact, we start with the analysis of an associated one-step equation. More specifically, we show the existence of a constant \(\lambda _{\delta }^m\in \mathbb {R}\) and a map \(w_{\delta }^m\in \mathcal {C}_b(B_m)\) satisfying
see Theorem 3.1 for details. Then, we link (3.2) with (3.1) in Theorem 3.4.
In the proof of Theorem 3.1 we use the Krein–Rutman theorem to get the existence of a positive eigenvalue with a non-negative eigenfunction to the specific non-linear operator associated with (3.2). This technique was primarily used in the context of diffusions; see e.g. [3, 4, 9] and references therein. See also [38] for the use with discrete time risk-sensitive Markov decision processes. It should be noted that, due to the difficulty of the verification of the theorem assumptions (including the complete continuity of a suitable operator), this approach is applied primarily in the compact state space setting and the extension to a non-compact space requires some additional arguments.
Theorem 3.1
There exists a constant \(\lambda _{\delta }^m>0\) and a map \(w_{\delta }^m\in \mathcal {C}_b(B_m)\) such that (3.2) is satisfied and we get \(\sup _{\xi \in U} w^m_\delta (\xi )=0\).
Proof
The idea of the proof is to use the Krein-Rutman theorem to get an eigenvalue and an eigenvector of a suitable operator. More specifically, we consider a cone of non-negative continuous and bounded functions \(\mathcal {C}^+_b(B_m)\subset \mathcal {C}_b(B_m)\) and, for any \(h\in \mathcal {C}^+_b(B_m)\), we define the operators
Now, we use the Krein-Rutman theorem to show that \(\tilde{T}_{\delta }^m\) admits a positive eigenvalue and a non-negative eigenfunction; see Theorem 4.3 in [11] for details. We start with verifying the assumptions. First, note that \(\tilde{T}_{\delta }^m\) is positively homogeneous, monotonic increasing, and we have
where \(\mathbbm {1}\) denotes the function identically equal to 1 on \(B_m\). Also, using Assumption (A2), we get that \(\tilde{T}_{\delta }^m\) transforms \(\mathcal {C}^+_b(B_m)\) into itself and it is continuous with respect to the supremum norm. Let us now show that \(\tilde{T}_{\delta }^m\) is in fact completely continuous. To see this, let \((h_n)_{n\in \mathbb {N}}\subset \mathcal {C}^+_b(B_m)\) be a bounded (by some constant \(K>0\)) sequence; using the Arzelà-Ascoli theorem we show that it is possible to find a convergent subsequence of \((\tilde{T}_{\delta }^m h_n)_{n\in \mathbb {N}}\). Note that, for any \(n\in \mathbb {N}\), we get
hence \((\tilde{T}_{\delta }^m h_n)\) is uniformly bounded. Next, let us fix some \(\varepsilon >0\), \(x\in B_m\), and \((x_k)\subset B_m\) such that \(x_k\rightarrow x\) as \(k\rightarrow \infty \). Also, to ease the notation, for any \(n\in \mathbb {N}\), we set \(H_n(x):=1_{\{x\in B_m\}}h_n(x)+1_{\{x\notin B_m\}}\tilde{M}h_n(x)\), \(x\in E\), and note that \(H_n\) are measurable functions bounded by 2K uniformly in \(n\in \mathbb {N}\). Then, for any \(n,k\in \mathbb {N}\), we get
Also, using Assumption (A1), we may find \(k\in \mathbb {N}\) big enough such that, for any \(n\in \mathbb {N}\), we obtain
Next, note that for any \(u\in (0,\delta )\) and \(n,k\in \mathbb {N}\), we get
Also, using the inequality \(|e^{y}-e^{z}|\le e^{\max (y,z)}|y-z|\), \(y,z\in \mathbb {R}\), we may find \(u>0\) small enough such that, for any \(n,k\in \mathbb {N}\), we get
Next, setting \(F_n^u(x):=\mathbb {E}_x\left[ e^{\int _0^{\delta -u}f(X_s)ds}H_n(X_{\delta -u})\right] \), \(n\in \mathbb {N}\), \(x\in E\), and using the Markov property combined with Assumption (A2), we may find \(k\in \mathbb {N}\) big enough such that for any \(n\in \mathbb {N}\), we get
Thus, recalling (3.5)–(3.6), we get that for \(k\in \mathbb {N}\) big enough and any \(n\in \mathbb {N}\), we get \(\left| \mathbb {E}_x\left[ e^{\int _0^\delta f(X_s)ds}H_n(X_\delta )\right] - \mathbb {E}_{x_k}\left[ e^{\int _0^\delta f(X_s)ds}H_n(X_\delta )\right] \right| \le \frac{\varepsilon }{2}\). This combined with (3.3)–(3.4) shows \( | \tilde{T}_{\delta }^m h_n(x)-\tilde{T}_{\delta }^m h_n(x_k)|\le \varepsilon \) for \(k\in \mathbb {N}\) big enough and any \(n\in \mathbb {N}\), which proves the equicontinuity of the family \((\tilde{T}_{\delta }^m h_n)_{n\in \mathbb {N}}\). Consequently, using the Arzelà-Ascoli theorem, we may find a uniformly (in \(x\in B_m\)) convergent subsequence of \((\tilde{T}_{\delta }^m h_n)_{n\in \mathbb {N}}\) and the operator \(\tilde{T}_{\delta }^m\) is completely continuous. Thus, using the Krein-Rutman theorem we conclude that there exists a constant \(\tilde{\lambda }_{\delta }^m>0\) and a non-zero map \(h_{\delta }^m\in \mathcal {C}^+_b(B_m)\) such that
After a possible normalisation, we assume that \(\sup _{\xi \in U}h_{\delta }^m(\xi )=1\).
Let us now show that \(h_{\delta }^m(x)>0\), \(x\in B_m\). To see this, let us define \(D:=e^{-\delta \Vert f\Vert } \frac{1}{\tilde{\lambda }_{\delta }^m}\) and let \(\mathcal {O}_h\subset B_m\) be an open set such that
note that this set exists thanks to the continuity of \(h_{\delta }^m\) and the fact that \(h_{\delta }^m\) is non-zero. Next, using (3.7), we have
Then, for any \(n\in \mathbb {N}\), we inductively get
Thus, letting \(n\rightarrow \infty \) and using Assumption (A4) combined with (3.8), we show \(h_{\delta }^m(x)>0\) for any \(x\in B_m\).
Next, we define \(w^m_\delta (x):=\ln h^m_\delta (x)\), \(x\in B_m\), and \(\lambda ^m_\delta :=\frac{1}{\delta }\ln \tilde{\lambda }^m_\delta \). Thus, from (3.7), we get that the pair \((w^m_\delta , \lambda ^m_\delta )\) satisfies
In fact, using (2.3) from Assumption (A1) and the argument from Theorem 3.1 in [20], we have
Finally, we extend the definition of \(w^m_{\delta }\) to the full space E by setting
note that the definition is correct since, at the right-hand side, we need to evaluate \(w^m_{\delta }\) only at the points from \(U\subset B_0\subset B_m\) and this map is already defined there. \(\square \)
As we show now, Eq. (3.2) may be linked to a specific martingale characterisation.
Proposition 3.2
Let \((w_{\delta }^m, \lambda ^m_\delta )\) be a solution to (3.2). Then, for any \(x\in B_m\), we get that the process
is a \(\mathbb {P}_x\)-supermartingale. Also, the process
is a \(\mathbb {P}_x\)-martingale, where \(\hat{\tau }^m_\delta :=\delta \inf \{k\in \mathbb {N}:w_\delta ^m(X_{k\delta })=M w_\delta ^m(X_{k\delta })\}\).
Proof
To ease the notation, we show the proof only for \(\delta =1\); the general case follows the same logic. Let us fix \(m,n\in \mathbb {N}\) and \(x\in B_m\). Then, using the fact \(w_{1}^m(y)=Mw_{1}^m(y)\), \(x\notin B_m\), and the inequality
we have
which shows the supermartingale property of \((z_{1}^m(n))\). Next, note that on the set \(\{\tau _{B_m}\wedge \hat{\tau }^m_1>n\}\) we get
Thus, we have
which concludes the proof. \(\square \)
Let us denote by \(\mathbb {V}_{\delta ,m}\) the family of impulse control strategies with impulse times in the time-grid \(\{0,\delta , 2\delta , \ldots \}\) and obligatory impulses when the controlled process exits the set \(B_m\) at some multiple of \(\delta \). Using a martingale characterisation of (3.2), we get that \(\lambda ^m_\delta \) is the optimal value of the impulse control problem with impulse strategies from \(\mathbb {V}_{\delta ,m}\); see Theorem 3.3 To show this result, we introduce a strategy \(\hat{V}:=(\hat{\tau }_i,\hat{\xi }_i)_{i=1}^\infty \in \mathbb {V}_{\delta ,m}\) defined recursively, for \(i=1, 2,\ldots \), by
where \(\hat{\tau }_0:=0\) and \(\xi _0\in U\) is some fixed point.
Theorem 3.3
Let \((w_{\delta }^m, \lambda ^m_\delta )\) be a solution to (3.2). Then, for any \(x\in B_m\), we get
Also, the strategy \(\hat{V}\) defined in (3.9) is optimal.
Proof
The proof follows the lines of the proof of Theorem 2.3 and is omitted for brevity. \(\square \)
Next, we link (3.2) with an infinite horizon optimal stopping problem under the non-degeneracy assumption.
Theorem 3.4
Let \((w_{\delta }^m, \lambda ^m_\delta )\) be a solution to (3.2) with \(\lambda ^m_\delta >r(f)\). Then, we get that \((w_{\delta }^m, \lambda ^m_\delta )\) satisfies (3.1).
Proof
As in the proof of Proposition 3.2, we consider only \(\delta =1\); the general case follows the same logic.
First, note that for any \(x\in B_m\), \(n\in \mathbb {N}\), and \(\tau \in \mathcal {T}^\delta _x\), using Proposition 3.2 and Doob’s optional stopping theorem, we have
Also, recalling the boundedness of \(w^m_1\), using Proposition A.2, and letting \(n\rightarrow \infty \), we get
Next, noting that \(w_1^m(X_{\tau \wedge \tau _{B_m}})\ge Mw_1^m(X_{\tau \wedge \tau _{B_m}})\), and taking the supremum over \(\tau \in \mathcal {T}^\delta _x\), we get
Second, using again Proposition 3.2, for any \(x\in B_m\) and \(n\in \mathbb {N}\), we get
Using again the boundedness of \(w^m_1\) and Proposition A.2, and letting \(n\rightarrow \infty \), we get
In fact, noting that \(w_1^m(X_{ \hat{\tau }^m_\delta \wedge \tau _{B_m}})=Mw_1^m(X_{ \hat{\tau }^m_\delta \wedge \tau _{B_m}})\), we obtain
thus we get
Finally, using Proposition A.4, we have
which concludes the proof. \(\square \)
Remark 3.5
In Theorem 3.4 we showed that, if \(\lambda ^m_\delta >r(f)\), a solution to the one-step equation (3.2) is uniquely characterised by the optimal stopping value function (3.1). If \(\lambda ^m_\delta \le r(f)\), the problem is degenerate and, in particular, we cannot use the uniform integrability result from Proposition A.2. In fact, in this case it is even possible that the one-step Bellman equation admits multiple solutions and the optimal stopping characterisation does not hold; see e.g. Theorem 1.13 in [27] for details.
4 Dyadic Impulse Control
In this section we consider a dyadic full-domain version of (2.1). We construct a solution to the associated Bellman equation which will be later used to find a solution to (2.7). The argument uses a bounded domain approximation from Sect. 3. More specifically, throughout this section we fix some \(\delta >0\) and show the existence of a function \(w_\delta \in \mathcal {C}_b(E)\) and a constant \(\lambda _\delta \in \mathbb {R}\), which are a solution to the dyadic Bellman equation of the form
In fact, we set
note that this constant is well-defined as, from Theorem 3.3, recalling that \(B_m\subset B_{m+1}\), we get \(\lambda ^m_\delta \le \lambda ^{m+1}_\delta \), \(m\in \mathbb {N}\).
First, we state the lower bound for \(\lambda _\delta \).
Lemma 4.1
Let \((w_\delta ,\lambda _\delta )\) be a solution to (4.1). Then, we get \(\lambda _\delta \ge r(f)\).
Proof
The proof follows the lines of the proof of Lemma 2.1 and is omitted for brevity. \(\square \)
Next, we show the existence of a solution to (4.1) under the non-degeneracy assumption \(\lambda _\delta >r(f)\).
Theorem 4.2
Let \(\lambda _\delta \) be given by (4.2) and assume that \(\lambda _\delta >r(f)\). Then, there exists \(w_\delta \in \mathcal {C}_b(E)\) such that (4.1) is satisfied and we get \(\sup _{\xi \in U} w_{\delta }(\xi )=0\).
Proof
We start with some general comments and an outline of the argument. First, note that from Theorem 3.1, for any \(m\in \mathbb {N}\), we get a solution \((w_\delta ^m,\lambda _\delta ^m)\) to (3.2) satisfying \(\sup _{\xi \in U}w^m_\delta (\xi )=0\). Also, from the assumption \(\lambda _\delta >r(f)\) we get \(\lambda ^m_\delta >r(f)\) for \(m\in \mathbb {N}\) sufficiently big (for simplicity, we assume that \(\lambda ^0_\delta >r(f)\)). Thus, using Theorem 3.4, we get that, for any \(m\in \mathbb {N}\), the pair \((w_\delta ^m,\lambda _\delta ^m)\) satisfies (3.1).
Second, to construct a function \(w_\delta \), we use the Arzelà-Ascoli theorem. More specifically, recalling that \(\sup _{\xi \in U}w^m_\delta (\xi )=0\) and using the fact that \(-\Vert c\Vert \le c(x,\xi )\le 0 \), \(x\in E\), \(\xi \in U\), for any \(m\in \mathbb {N}\) and \(x\in E\), we get
Also, note that, for any \(m\in \mathbb {N}\) and \(x,y\in E\), we have
Consequently, the sequence \((Mw^m_\delta )_{m\in \mathbb {N}}\) is uniformly bounded and equicontinuous. Thus, using the Arzelà-Ascoli theorem combined with a diagonal argument, we may find a subsequence (for brevity still denoted by \((Mw^m_\delta )_{m\in \mathbb {N}}\)) and a map \(\phi _\delta \in \mathcal {C}_b(E)\) such that \(Mw^m_\delta (x)\) converges to \(\phi _\delta (x)\) as \(m\rightarrow \infty \) uniformly in x from any compact set. In fact, using (2.4) from Assumption (A1) and the argument from the first step of the proof of Theorem 4.1 in [20], we get that the convergence is uniform in \(x\in E\). Then, we define
To complete the construction, we show that \(w^m_\delta \) converges to \(w_\delta \) uniformly on compact sets. Indeed, in this case we have
thus \(\phi _\delta \equiv Mw_\delta \) and from (4.3) we get that (4.1) is satisfied. Also, recalling that from Theorem 3.1 we get \(\sup _{\xi \in U} w^m_\delta (\xi )=0\), \(m\in \mathbb {N}\), we also get \(\sup _{\xi \in U} w_\delta (\xi )=0\).
Finally, to show the convergence, we define the auxiliary functions
We split the rest of the proof into three steps: (1) proof that \(|w_{\delta }^{m}(x)-w_{\delta }^{m,1}(x)|\rightarrow 0\) as \(m\rightarrow \infty \) uniformly in \(x\in E\); (2) proof that \(|w_{\delta }^{m,1}(x)-w_{\delta }^{m,2}(x)|\rightarrow 0\) as \(m\rightarrow \infty \) uniformly in \(x\in E\); (3) proof that \(|w_{\delta }^{m,2}(x)-w_{\delta }(x)|\rightarrow 0\) as \(m\rightarrow \infty \) uniformly in x from compact sets.
Step 1. We show \(|w_{\delta }^{m}(x)-w_{\delta }^{m,1}(x)|\rightarrow 0\) as \(m\rightarrow \infty \) uniformly in \(x\in E\). Note that, for any \(x\in E\) and \(m\in \mathbb {N}\), we have
Similarly, we get \(w_\delta ^m(x)\le w_\delta ^{m,1}(x)+\Vert \phi _\delta -Mw_\delta ^m\Vert \), thus
Recalling the fact that \(\phi _\delta \) is a uniform limit of \(Mw_\delta ^m\) as \(m\rightarrow \infty \), we conclude the proof of this step.
Step 2. We show that \(|w_{\delta }^{m,1}(x)-w_{\delta }^{m,2}(x)|\rightarrow 0\) as \(m\rightarrow \infty \) uniformly in \(x\in E\). Recalling that \(\lambda _\delta ^m\uparrow \lambda _\delta \), we get \(w_\delta ^{m,1}(x)\ge w_\delta ^{m,2}(x)\ge -\Vert \phi _\delta \Vert \), \(x\in E\). Thus, using the inequality \(|\ln y-\ln z|\le \frac{1}{\min (y,z)}|y-z|\), \(y,z>0\), we get
Then, noting that \(\phi _\delta (\cdot )\le 0\), for any \(m\in \mathbb {N}\) and \(x\in E\), we obtain
Also, recalling that \(\lambda _\delta ^0\le \lambda _\delta ^m\le \lambda _\delta \), \(m\in \mathbb {N}\), for any \(x\in E\) and \(T\ge 0\), we get
Recalling \(\lambda _\delta ^0>r(f)\) and using Lemma A.1, for any \(\varepsilon >0\), we may find \(T\ge 0\), such that
Also, using the inequality \(|e^x-e^y|\le e^{\max (x,y)}|x-y|\), \(x,y\ge 0\), we obtain
Thus, for fixed \(T\ge 0\), we find \(m\ge 0\), such that \(e^{|\lambda _\delta ^m | T} T(\lambda _\delta -\lambda _\delta ^m)\le \varepsilon \). Hence, recalling (4.6)–(4.8), for any \(x\in E\) and T, m big enough, we get
Recalling that \(\varepsilon >0\) was arbitrary, we conclude the proof of this step.
Step 3. We show that \(|w_{\delta }^{m,2}(x)-w_{\delta }(x)|\rightarrow 0\) as \(m\rightarrow \infty \) uniformly in x from compact sets. First, we show that \(w_\delta ^{m,2}(x)\le w_\delta (x)\) for any \(m\in \mathbb {N}\) and \(x\in E\). Let \(\varepsilon >0\) and \(\tau _m^\varepsilon \in \mathcal {T}_{x,b}^\delta \) be an \(\varepsilon \)-optimal stopping time for \(w_\delta ^{m,2}(x)\). Then, we get
As \(\varepsilon >0\) was arbitrary, we get \(w_\delta ^{m,2}(x)\le w_\delta (x)\), \(m\in \mathbb {N}\), \(x\in E\). In fact, using a similar argument, for any \(x\in E\), we may show that the map \(m\mapsto w_{\delta }^{m,2}(x)\) is non-decreasing.
Second, let \(\varepsilon >0\) and \(\tau _\varepsilon \in \mathcal {T}_{x,b}^\delta \) be an \(\varepsilon \)-optimal stopping time for \(w_\delta (x)\). Then, we obtain
Noting that \(\tau _{B_m}\uparrow +\infty \) as \(m\rightarrow \infty \) and using the quasi left-continuity of X combined with Lemma A.2 and the boundedness of \(\phi _\delta \), we get
Thus, using (4.10) and recalling that \(\varepsilon >0\) was arbitrary, we get \(\lim _{m\rightarrow \infty } w_\delta ^{m,2}(x) = w_\delta (x)\). Also, noting that by Propositions A.3 and A.4, the maps \(x\mapsto w_\delta (x)\) and \(x\mapsto w_\delta ^{m,2}(x)\) are continuous, and using the monotonicity of \(m\mapsto w_\delta ^{m,2}(x)\), from Dini’s Theorem we get that \(w_\delta ^{m,2}(x)\) converges to \(w_\delta (x)\) uniformly in x from compact sets, which concludes the proof. \(\square \)
We conclude this section with a verification result related to (4.1).
Theorem 4.3
Let \((w_\delta ,\lambda _\delta )\) be a solution to (4.1) with \(\lambda _\delta >r(f)\). Then, we get
where \(\mathbb {V}^\delta \) is a family of impulse control strategies with impulse times on the dyadic time-grid \(\{0,\delta , 2\delta , \ldots \}\).
Proof
The proof follows the lines of the proof of Theorem 2.3 and is omitted for brevity. \(\square \)
5 Existence of a Solution to the Bellman Equation
In this section we construct a solution \((w,\lambda )\) to (2.7), which together with Theorem 2.3 provides a solution to (2.1). The argument uses a dyadic approximation and the results from Sect. 4. More specifically, we fix \(\delta >0\) and consider a family of dyadic time steps \(\delta _k:=\frac{\delta }{2^k}\), \(k\in \mathbb {N}\). First, we specify the value of \(\lambda \). In fact, we define
where \(\lambda _{\delta _k}\) is a constant given by (4.2), corresponding to \(\delta _k\). In Theorem 5.1, we show that if \(\lambda (\delta )>r(f)\), then there exists a solution to (2.7) with the constant \(\lambda \) given by \(\lambda (\delta )\). Also, in this case \(\lambda \) does not depend on \(\delta \) and the limit inferior could be replaced by the usual limit.
Theorem 5.1
Let \(\delta >0\) and let \(\lambda (\delta )\) be given by (5.1). Assume that \(\lambda (\delta )>r(f)\). Then, there exists \(w\in \mathcal {C}_b(E)\) such that (2.7) is satisfied with \(\lambda =\lambda (\delta )\). Also, \(\lambda (\delta )=\lim _{k\rightarrow \infty } \lambda _{\delta _k}\) and \(\lambda (\delta )\) does not depend on \(\delta >0\), i.e. for any \(\delta _1>0\) and \(\delta _2>0\) such that \(\lambda (\delta _1)>r(f)\) and \(\lambda (\delta _2)>r(f)\), we get \(\lambda (\delta _1)=\lambda (\delta _2)\).
Proof
The argument is partially based on the one used in Theorem 4.2 and thus we discuss only the main points. First, from the assumption \(\lambda (\delta )>r(f)\) we get \(\lambda _{\delta _k}>r(f)\) for sufficiently big \(k\in \mathbb {N}\); to simplify the notation, we assume \(\lambda _{\delta _0}>r(f)\). Hence, using Theorems 4.2, 4.3, and the fact \(\mathbb {V}^{\delta _k}\subset \mathbb {V}^{\delta _{k+1}}\), we inductively show
Thus, the sequence \((\lambda _{\delta _k})_{k=k_0}^\infty \) is non-decreasing and, consequently, convergent. Hence, \(\lambda (\delta )=\lim _{k\rightarrow \infty }\lambda _{\delta _k}\). Second, using again Theorem 4.2, for any \(k\in \mathbb {N}\), we find a map \(w_{\delta _k}\in \mathcal {C}_b(E)\) satisfying
and such that \(\sup _{\xi \in U} w_{\delta _k}(\xi )=0\). Thus, we obtain
and the family \((Mw_{\delta _k})_{k\in \mathbb {N}}\) is uniformly bounded. Also, it is equicontinuous as we have
Thus, using the Arzelà-Ascoli theorem, we may choose a subsequence (for brevity still denoted by \((Mw_{\delta _k})\)), such that \((Mw_{\delta _k})\) converges uniformly on compact sets to some map \(\phi \). In fact, using (2.4) from Assumption (A1) and the argument from the first step of the proof of Theorem 4.1 from [20], we get that \(Mw_{\delta _k}(x)\) converges to \(\phi (x)\) as \(k\rightarrow \infty \) uniformly in \(x\in E\). Next, let us define
In the following, we show that \(w_{\delta _k}\) converges to w uniformly in compact sets as \(k\rightarrow \infty \). Then, we get that \(Mw_{\delta _k}\) converges to Mw, hence \(Mw\equiv \phi \) and (2.7) is satisfied.
To show the convergence, we define
In the following, we show that \(|w(x)-w_{\delta _k}^1(x)|\rightarrow 0\) and \(|w_{\delta _k}^1(x)-w_{\delta _k}(x)|\rightarrow 0\) as \(k\rightarrow \infty \) uniformly in x from compact sets. In fact, to show the first convergence, we note that
where
Thus, to prove \(|w(x)-w_{\delta _k}^1(x)|\rightarrow 0\) it is enough to show \(|w(x)-w_{\delta _k}^0(x)|\rightarrow 0\) and \(|w(x)-w_{\delta _k}^2(x)|\rightarrow 0\) as \(k\rightarrow \infty \).
For transparency, we split the rest of the proof into three parts: (1) proof that \(|w(x)-w_{\delta _k}^0(x)|\rightarrow 0\) as \(k\rightarrow \infty \) uniformly in x from compact sets; (2) proof that \(|w(x)-w_{\delta _k}^2(x)|\rightarrow 0\) as \(k\rightarrow \infty \) uniformly in \(x\in E\); (3) proof that \(|w_{\delta _k}^1(x)-w_{\delta _k}(x)|\rightarrow 0\) as \(k\rightarrow \infty \) uniformly in \(x\in E\); (4) proof that \(\lambda (\delta )\) does not depend on \(\delta \).
Step 1. We show that \(|w(x)-w_{\delta _k}^0(x)|\rightarrow 0\) as \(k\rightarrow \infty \) as \(k\rightarrow \infty \) uniformly in x from compact sets. First, note that we have \(w_{\delta _k}^0(x)\le w(x)\), \(k\in \mathbb {N}\), \(x\in E\). Next, for any \(x\in E\) and \(\varepsilon >0\), let \(\tau _\varepsilon \in \mathcal {T}_{x,b}\) be an \(\varepsilon \)-optimal stopping time for w(x) and let \(\tau _\varepsilon ^k\) be its \(\mathcal {T}^{\delta _k}_{x,b}\) approximation given by
Then, we get
Also, using Proposition A.2 and letting \(k\rightarrow \infty \), we have
Consequently, recalling that \(\varepsilon >0\) was arbitrary, we obtain \(\lim _{k\rightarrow \infty }w_{\delta _k}^0(x)=w(x)\) for any \(x\in E\). Next, noting that \(\mathcal {T}_{x,b}^{\delta _k}\subset \mathcal {T}_{x,b}^{\delta _{k+1}}\), \(k\in \mathbb {N}\), we get \(w_{\delta _k}^0(x)\le w_{\delta _{k+1}}^0(x)\), \(k\in \mathbb {N}\), \(x\in E\). This combined with Propositions A.3, A.4, and Dini’s theorem, we get that the convergence of \(w_{\delta _k}^0\) to w is uniform on compact sets, which concludes the proof of this step.
Step 2. We show that \(|w(x)-w_{\delta _k}^2(x)|\rightarrow 0\) as \(k\rightarrow \infty \) uniformly in \(x\in E\). First, note that \(-\Vert \phi \Vert \le w(x)\le w_{\delta _k}^2(x)\), \(k\in \mathbb {N}\), \(x\in E\). Thus, using the inequality \(|\ln y-\ln z|\le \frac{1}{\min (y,z)}|y-z|\), \(y,z>0\), we get
Also, recalling that \(\phi (\cdot )\le 0\), for any \(k\in \mathbb {N}\) and \(x\in E\), we obtain
Thus, repeating the argument from the second step of the proof of Theorem 4.2, we get \(w_{\delta _k}^2(x)\rightarrow w(x)\) as \(k\rightarrow \infty \) uniformly in \(x\in E\), which concludes the proof of this step.
Step 3. We show that \(|w_{\delta _k}^1(x)-w_{\delta _k}(x)|\rightarrow 0\) as \(k\rightarrow \infty \) uniformly in \(x\in E\). In fact, recalling that \(\Vert Mw_{\delta _k}-\phi \Vert \rightarrow 0\) as \(k\rightarrow \infty \), the argument follows the lines of the one used in the first step of the proof of Theorem 4.2. This concludes the proof of this step.
Step 4. We show that \(\lambda (\delta )\) does not depend on \(\delta \) as long as \(\lambda (\delta )>r(f)\). More specifically, let \(\delta _1>0\) and \(\delta _2>0\) be such that \(\lambda (\delta _1)>r(f)\) and \(\lambda (\delta _2)>r(f)\). Then, using Steps 1–3, we may construct \(w^{\delta _1}\in \mathcal {C}_b(E)\) and \(w^{\delta _2}\in \mathcal {C}_b(E)\) such that the pairs \((w^{\delta _1},\lambda (\delta _1))\) and \((w^{\delta _2},\lambda (\delta _2))\) satisfy (2.7). Then, using Theorem 2.3, for any \(x\in E\), we get
which concludes the proof. \(\square \)
Remark 5.2
By the inspection of the proof we get that the statement of Theorem 5.1 holds true if we replace the dyadic sequence of time steps \(\delta _k=\frac{\delta }{2^k}\), \(k\in \mathbb {N}\), by any sequence \((\delta _k)\) converging to zero, as long as we have \(\mathcal {T}_{x,b}^{\delta _k}\subset \mathcal {T}_{x,b}^{\delta _{k+1}}\), \(x\in E\), \(k\in \mathbb {N}\). Note that this condition guarantees the monotonic convergence of \(\lambda _{\delta _k}\) and \(w^0_{\delta _k}\).
References
Applebaum, D.: Lévy Processes and Stochastic Calculus. Cambridge Studies in Advanced Mathematics, 2nd edn. Cambridge University Press, Cambridge (2009)
Arapostathis, A., Biswas, A.: Infinite horizon risk-sensitive control of diffusions without any blanket stability assumptions. Stoch. Process. Appl. 128, 1485–1524 (2018)
Arapostathis, A., Biswas, A., Kumar, K.S.: Risk-sensitive control and an abstract Collatz–Wielandt formula. J. Theor. Probab. 29(4), 1458–1484 (2016)
Arapostathis, A., Biswas, A., Saha, S.: Strict monotonicity of principal eigenvalues of elliptic operators in \(\mathbb{R} ^d\) and risk-sensitive control. J. Math. Pures Appl. 124, 169–219 (2019)
Basu, A., Stettner, Ł: Zero-sum Markov games with impulse controls. SIAM J. Control Optim. 58(1), 580–604 (2020)
Bayraktar, E., Emmerling, T., Menaldi, J.: On the impulse control of jump diffusions. SIAM J. Control Optim. 51(3), 2612–2637 (2013)
Bensoussan, A., Lions, J.-L.: Impulse Control and Quasi-Variational Inequalities. Gauthier-Villars, Montrouge (1984)
Bielecki, T.R., Pliska, S.R.: Economic properties of the risk sensitive criterion for portfolio management. Rev. Account. Financ. 2, 3–17 (2003)
Biswas, A.: An eigenvalue approach to the risk sensitive control problem in near monotone case. Syst. Control Lett. 60, 181–184 (2011)
Biswas, A., Borkar, V.: Ergodic risk-sensitive control: a survey. arXiv:2301.00224 (2023)
Bonsall, F.: Lectures on Some Fixed Point Theorems of Functional Analysis. Tata Institute of Fundamental Research, Mumbai (1962)
Christensen, S.: On the solution of general impulse control problems using superharmonic functions. Stoch. Process. Appl. 124(1), 709–729 (2014)
Davis, M.H.A.: Markov Models and Optimization. Chapman & Hall/CRC, New York (1993)
Dufour, F., Piunovskiy, A.B.: Impulsive control for continuous-time Markov decision processes: a linear programming approach. Appl. Math. Optim. 74, 129–161 (2016)
Fakeev, A.: Optimal stopping rules for stochastic processes with continuous parameter. Theory Probab. Appl. 15(2), 324–331 (1970)
Fakeev, A.: Optimal stopping of a Markov process. Theory Probab. Appl. 16(4), 694–696 (1971)
Gikhman, I.I., Skorokhod, A.V.: The Theory of Stochastic Processes II. Springer, Berlin (1975)
Hille, E., Phillips, R.S.: Functional Analysis and Semi-groups. American Mathematical Society, Providence (1957)
Howard, R.A., Matheson, J.E.: Risk-sensitive Markov decision processes. Manag. Sci. 18(7), 356–369 (1972)
Jelito, D., Pitera, M., Stettner, Ł: Long-run risk sensitive impulse control. SIAM J. Control Optim. 58(4), 2446–2468 (2020)
Jelito, D., Pitera, M., Stettner, Ł: Risk sensitive optimal stopping. Stoch. Process. Appl. 136, 125–144 (2021)
Jelito, D., Stettner, Ł: Risk-sensitive optimal stopping with unbounded terminal cost function. Electron. J. Probab. 27, 1–30 (2022)
Korn, R.: Some applications of impulse control in mathematical finance. Math. Methods Oper. Res. 50, 493–518 (1999)
Menaldi, J., Robin, M.: On some ergodic impulse control problems with constraint. SIAM J. Control Optim. 56(4), 2690–2711 (2018)
Nagai, H.: Stopping problems of certain multiplicative functionals and optimal investment with transaction costs. Appl. Math. Optim. 55(3), 359–384 (2007)
Palczewski, J., Stettner, Ł: Finite horizon optimal stopping of time-discontinuous functionals with applications to impulse control with delay. SIAM J. Control Optim. 48(8), 4874–4909 (2010)
Peskir, G., Shiryaev, A.: Optimal Stopping and Free-Boundary Problems. Springer, Berlin (2006)
Pham, H.: Long time asymptotics for optimal investment. In: Friz, P.K., Gatheral, J., Gulisashvili, A., Jacquier, A., Teichmann, J. (eds.) Large Deviations and Asymptotic Methods in Finance, pp. 507–528. Springer International Publishing, Cham (2015)
Pitera, M., Stettner, Ł: Long-run risk sensitive dyadic impulse control. Appl. Math. Optim. 84(1), 19–47 (2021)
Piunovskiy, A., Plakhov, A., Tumanov, M.: Optimal impulse control of a SIR epidemic. Optim. Control Appl. Methods 41(2), 448–468 (2020)
Robin,M.: Contrôle impulsionnel des processus de Markov. thèse d’état. Université Paris Dauphine (1978). https://tel.archives-ouvertes.fr/tel-00735779
Runggaldier, W.J., Yasuda, K.: Classical and restricted impulse control for the exchange rate under a stochastic trend model. J. Econ. Dyn. Control 91, 369–390 (2018)
Shiryaev, A.: Optimal Stopping Rules. Springer, Berlin (1978)
Stettner, Ł.: On impulsive control with long run average cost criterion. In: Stochastic Differential Systems. Springer, Berlin (1982)
Stettner, Ł: On some stopping and implusive control problems with a general discount rate criteria. Probab. Math. Stat. 10, 223–245 (1989)
Stettner, Ł: Asymptotics of HARA utility from terminal wealth under proportional transaction costs with decision lag or execution delay and obligatory diversification. In: Di Nunno, G., Øksendal, B. (eds.) Advanced Mathematical Methods for Finance, pp. 509–536. Springer, Berlin Heidelberg (2011)
Stettner, Ł: On an approximation of average cost per unit time impulse control of Markov processes. SIAM J. Control Optim. 60(4), 2115–2131 (2022)
Stettner, Ł.: Discrete time risk sensitive control problem. arXiv:2303.17913 (2023)
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Statements and Declarations
Damian Jelito and Łukasz Stettner acknowledge research support by Polish National Science Centre grant no. 2020/37/B/ST1/00463. The authors have no relevant financial or non-financial interests to disclose. The authors contributed equally to this work.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendix A: Properties of Optimal Stopping Problems
Appendix A: Properties of Optimal Stopping Problems
In this section we discuss some properties of the optimal stopping problems that are used in this paper. Throughout this section we consider \(g,G \in \mathcal {C}_b(E)\) and assume \(G(\cdot )\le 0\) and \(r(g)<0\), where r(g) is the semi-group type given by (2.8) corresponding to the map g. We start with a useful result related to the asymptotic behaviour of the running cost function g.
Lemma A.1
Let a be such that \(r(g)<a<0\). Then,
-
(1)
The map \(x\mapsto U_0^{g-a} 1(x):=\mathbb {E}_x\left[ \int _0^\infty e^{\int _0^t (g(X_s) -a)ds}dt\right] \) is continuous and bounded.
-
(2)
We get
$$\begin{aligned} \lim _{T\rightarrow \infty }\sup _{x\in E}\sup _{\begin{array}{c} \tau \ge T\\ \tau \in \mathcal {T}_x \end{array}}\mathbb {E}_x\left[ e^{\int _0^\tau g(X_s)ds}\right] = 0. \end{aligned}$$
Proof
For transparency, we prove the claims point by point.
Proof of (1). First, we show the boundedness of \(x\mapsto U_0^{g-a} 1(x)\). Let \(\varepsilon <a-r(g)\). Using the definition of \(r(g-a)\) we may find \(t_0\ge 0\), such that for any \(t\ge t_0\) we get \(\sup _{x\in E} \mathbb {E}_x\left[ e^{\int _0^t (g(X_s)-a)ds}\right] \le e^{t(r(g)-a+\varepsilon )}\). Then, using Fubini’s theorem and noting that \(r(g)-a+\varepsilon <0\), for any \(x_0\in E\), we get
which concludes the proof of the boundedness of \(x\mapsto U_0^{g-a} 1(x)\).
For the continuity, note that using Assumption (A2) and repeating the argument used in Lemma 4 in Section II.5 of [17], we get that \(x\mapsto \mathbb {E}_x\left[ e^{\int _0^t (g(X_s) -a)ds}dt\right] \) is continuous for any \(t\ge 0\). Also, as in the proof of the boundedness, we may show
and the upper bound is integrable (with respect to t). Thus, using Lebesgue’s dominated convergence theorem, we get the continuity of the map \(x\mapsto U_0^{g-a} 1(x)=\int _0^\infty \mathbb {E}_x\left[ e^{\int _0^t (g(X_s) -a)ds}\right] dt\), which concludes the proof of this step.
Proof of (2). Noting that \(U_0^{g-a} 1(x)\ge \int _0^1 e^{-t(\Vert g\Vert -a)}dt\), \(x\in E\), we may find \(d>0\), such that \(U_0^{g-a}1(x)\ge d>0\), \(x\in E\). Thus, recalling that \(a<0\), we get
which concludes the proof. \(\square \)
Using Lemma A.1 we get the uniform integrability of a suitable family of random variables. This result is extensively used throughout the paper as it simplifies numerous limiting arguments.
Proposition A.2
For any \(x\in E\), the family \(\{e^{\int _0^\tau g(X_s)ds}\}_{\tau \in \mathcal {T}_x}\) is \(\mathbb {P}_x\)-uniformly integrable.
Proof
Let us fix some \(x\in E\) and, for any \(\tau \in \mathcal {T}_x\) and \(n\in \mathbb {N}\), define the event \(A_n^\tau :=\{\int _0^\tau g(X_s)ds\ge n\}\). Note that for any \(T\ge 0\), we get
Next, for any \(\varepsilon >0\), using Lemma A.1, we may find \(T>0\) big enough to get
Also, noting that for \(\tau \le T\), we get \(A_n^\tau \subset \{T\Vert g\Vert \ge n\}\), for any \(n>T\Vert g\Vert \), we also get
Consequently, recalling that \(\varepsilon >0\) was arbitrary, we obtain
which concludes the proof. \(\square \)
Next, we consider an optimal stopping problem of the form
note that here the non-positivity assumption for G is only a normalisation as for a generic \(\tilde{G}\) we may set \(G(\cdot )=\tilde{G}(\cdot )-\Vert \tilde{G}\Vert \) to get \(G(\cdot )\le 0\).
The properties of the map (A.1) are summarised in the following proposition.
Proposition A.3
Let the map u be given by (A.1). Then, \(x\mapsto u(x)\) is continuous and bounded. Also, we get
Moreover, the process
is a supermartingale and the process \(z(t\wedge \hat{\tau })\), \(t\ge 0\), is a martingale, where
Proof
For transparency, we split the proof into two steps: (1) proof of the continuity of \(x\mapsto u(x)\) and identity (A.2); (2) proof of the martingale properties of the process z.
Step 1. We show that the map \(x\mapsto u(x)\) is continuous and the identity (A.2) holds. For any \(T\ge 0\) and \(x\in E\), let us define
Using Assumption (A3) and following the proof of Proposition 10 and Proposition 11 in [21], we get that the map \((T,x)\mapsto u_T(x)\) is jointly continuous and bounded; see also Remark 12 therein. We show that \(u_T(x)\rightarrow \hat{u}(x)\) as \(T\rightarrow \infty \) uniformly in \(x\in E\). Noting that
and using the inequality \(|\ln y-\ln z|\le \frac{1}{\min (y,z)}|y-z|\), \(y,z>0\), to show \(u_T(x)\rightarrow \hat{u}(x)\) as \(T\rightarrow \infty \) uniformly in \(x\in E\) it is enough to show \(e^{u_T(x)}\rightarrow e^{\hat{u}(x)}\) as \(T\rightarrow \infty \) uniformly in \(x\in E\). Then, using Lemma A.1, for any \(\varepsilon >0\), we may find \(T\ge 0\) such that for any \(x\in E\), we obtain
Thus, letting \(\varepsilon \rightarrow 0\), we get \(e^{u_T(x)}\rightarrow e^{\hat{u}(x)}\) as \(T\rightarrow \infty \) uniformly in \(x\in E\) and consequently \(u_T(x)\rightarrow \hat{u}(x)\) as \(T\rightarrow \infty \) uniformly in \(x\in E\). Thus, from the continuity of \(x\mapsto u_T(x)\), \(T\ge 0\), we get that the map \(x\mapsto \hat{u}(x)\) is continuous.
Now, we show that \(u\equiv \hat{u}\). First, we show that \(\lim _{T\rightarrow \infty } u_T(x) = \tilde{u}(x)\), where \( \tilde{u}(x):=\sup _{\tau \in \mathcal {T}_x} \liminf _{T\rightarrow \infty } \ln \mathbb {E}_x\left[ e^{\int _0^{\tau \wedge T} g(X_s)ds +G(X_{\tau \wedge T})}\right] \), \(x\in E\). For any \(T\ge 0\) and \(x\in E\), we get
thus we get \(\lim _{T\rightarrow \infty } u_T(x) \le \tilde{u}(x)\). Also, for any \(x\in E\), \(\tilde{\tau }\in \mathcal {T}_x\), and \(T\ge 0\), we get
Thus, letting \(T\rightarrow \infty \) and taking supremum over \(\tilde{\tau }\in \mathcal {T}_x\) we get \(\lim _{T\rightarrow \infty } u_T(x) = \tilde{u}(x)\), \(x\in E\). Also, using the argument from Lemma 2.2 from [22] we get \( \tilde{u} \equiv u\). Thus, we get \(u(x) = \lim _{T\rightarrow \infty } u_T(x) = \hat{u}(x)\), \(x\in E\), hence the map \(x\mapsto u(x)\) is continuous. Also, we get (A.2).
Step 2. We show the martingale properties of z. First, we focus on the stopping time \(\hat{\tau }\). Let us define
Using the argument from Proposition 11 in [21] we get that \(\tau _T\) is an optimal stopping time for \(u_{T}\). Also, noting that the map \(T\mapsto u_T(x)\), \(x\in E\), is increasing, we get that \(T\mapsto \tau _T \) is also increasing, thus we may define \(\tilde{\tau }:=\lim _{T\rightarrow \infty } \tau _T\). We show that \(\tilde{\tau }\equiv \hat{\tau }\).
Let \(A:=\{\tilde{\tau }<\infty \}\). First, we show that \(\tilde{\tau }\equiv \hat{\tau }\) on A. On the event A, we get \(u_{T-\tau _T}(X_{\tau _T})=G(X_{\tau _T})\). Thus, letting \(T\rightarrow \infty \), we get \(u(X_{\tilde{\tau }})=G(X_{\tilde{\tau }})\), hence we get \(\hat{\tau }\le \tilde{\tau }\). Also, noting that \(u_S(x)\le u(x)\), \(x\in E\), \(S\ge 0\), on the set \(\{\hat{\tau }\le T\}\) we get \(u_{T-\hat{\tau }}(X_{\hat{\tau }})\le u(X_{\hat{\tau }})\le G(X_{\hat{\tau }})\), hence
Thus, recalling that \(\hat{\tau }\le \tilde{\tau }< \infty \) and letting \(T\rightarrow \infty \) in (A.6), we get \(\tilde{\tau }\le \hat{\tau }\), which shows \(\tilde{\tau }\equiv \hat{\tau }\) on A.
Now, we show that \(\tilde{\tau }\equiv \hat{\tau }\) on \(A^c\). Let \(\omega \in A^c\) and suppose that \(\hat{\tau }(\omega )<\infty \). Then, we may find \(T\ge 0\) such that \(\hat{\tau }(\omega )<T\). Also, for any \(S\ge T\) we get
Thus, we get \(\tau _S(\omega )\le \hat{\tau }(\omega )\) for any \(S\ge T\). Consequently, letting \(S\rightarrow \infty \) we get \( \tilde{\tau }(\omega )<\infty \), which contradicts the choice of \(\omega \in A^c\). Consequently, on \(A^c\) we have \(\tilde{\tau }=\infty = \hat{\tau }\).
Finally, we show the martingale properties. Let us define the processes
Using standard argument we get that for any \(T\ge 0\), the process \(z_T(t)\), \(t\ge 0\), is a supermartingale and \(z_T(t\wedge \tau _T)\), \(t\ge 0\), is a martingale; see e.g. [15, 16] for details. Also, recalling that from the first step we get \(u_T(x)\rightarrow u(x)\) as \(T\rightarrow \infty \) uniformly in \(x\in E\), for any \(t\ge 0\), we get that \(z_T(t)\rightarrow z(t)\) and \(z_T(t\wedge \tau _T)\rightarrow z(t\wedge \hat{\tau })\) as \(T\rightarrow \infty \). Consequently, using Lebesgue’s dominated convergence theorem, we get that the process z(t) is a supermartingale and \(z(t\wedge \hat{\tau })\), \(t\ge 0\), is a martingale, which concludes the proof. \(\square \)
Next, we consider an optimal stopping problem in a compact set and dyadic time-grid. More specifically, let \(\delta >0\), let \(B\subset E\) be compact and assume that \(\mathbb {P}_x[\tau _B<\infty ]=1\), \(x\in B\), where \(\tau _B:=\delta \inf \{n \in \mathbb {N}:X_{n\delta }\notin B\}\). Within this framework, we consider an optimal stopping problem of the form
The properties of (A.7) are summarised in the following proposition.
Proposition A.4
Let \(u_B\) be given by (A.7). Then, we get
Also, the map \(x\mapsto u_B(x)\) is continuous and bounded. Moreover, the process
is a supermartingale and the process \(z(n\wedge \hat{\tau }/\delta )\), \(n\in \mathbb {N}\), is a martingale, where
Proof
To ease the notation, let us define
and note that we get \(u_B^n(x)\le \hat{u}_B(x)\le u_B(x)\), \(x\in E\). Next, note that using the boundedness of G and Proposition A.2, by Lebesgue’s dominated convergence theorem, we obtain
Also, for any \(n\in \mathbb {N}\), \(x\in E\), and \(\tau \in \mathcal {T}^\delta \), we get
Thus, letting \(n\rightarrow \infty \) and taking the supremum with respect to \(\tau \in \mathcal {T}^\delta \), we get \(\lim _{n\rightarrow \infty }u_B^n(x)\ge u_B(x)\), \(x\in E\). Consequently, we have
which concludes the proof of (A.8).
Let us now show the continuity of the map \( x\mapsto u_B(x)\) and the martingale characterisation. To see this, note that using a standard argument one may show that, for any \(n\in \mathbb {N}\) and \( x\in B\), we get
and, for any \(n\in \mathbb {N}\) and \( x\notin B\), we get \(u^n_B(x)=G(x)\); see e.g. Section 2.2 in [33] for details. Thus, letting \(n\rightarrow \infty \), for \(x\in B\), we have
while for \(x\notin B\), we get \(u_B(x)=G(x)\). Also, using Assumption (A2), we get that the process X is strong Feller. Thus, repeating the argument used in Lemma 4 from Chapter II.5 in [17], we get that, for any bounded and measurable function \(h:E\mapsto \mathbb {R}\), the map
is continuous and bounded. Applying this observation to \(h(x):=1_{\{x\in B\}}e^{u_B(x)}\) and \(h(x):=1_{\{x\notin B\}}e^{G(x)}\), \(x\in E\), we get the continuity of \(x\mapsto u_B(x)\). Also, using the argument from Proposition 3.2 we get that \(z_\delta (n)\), \(n\in \mathbb {N}\) is a supermartingale and \(z(n\wedge \hat{\tau }/\delta )\), \(n\in \mathbb {N}\), is a martingale, which concludes the proof. \(\square \)
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Jelito, D., Stettner, Ł. Asymptotics of Impulse Control Problem with Multiplicative Reward. Appl Math Optim 88, 24 (2023). https://doi.org/10.1007/s00245-023-10005-5
Accepted:
Published:
DOI: https://doi.org/10.1007/s00245-023-10005-5