Optimal impulse control on an unbounded domain with nonlinear cost functions Stefano Baccarin∗ and Simona Sanfelici† Keywords: impulse control, stochastic cash management, quasi-variational inequalities, finite element approximation. Mathematics Subject Classification (2000): 49J40, 60G40, 65N30. Abstract In this paper we consider the optimal impulse control of a system which evolves randomly in accordance with a homogeneous diffusion process in R1 . Whenever the system is controlled a cost is incurred which has a fixed component and a component which increases with the magnitude of the control applied. In addition to these controlling costs there are holding or carrying costs which are a positive function of the state of the system. Our objective is to minimize the expected discounted value of all costs over an infinite planning horizon. Under general assumptions on the cost functions we show that the value function is a weak solution of a quasi-variational inequality and we deduce from this solution the existence of an optimal impulse policy. The computation of the value function is performed by means of the Finite Element Method on suitable truncated domains, whose convergence is discussed. 1 Introduction We consider a fund x(t) whose dynamics in absence of control is described by a homogeneous diffusion process in <1 . Holding or carrying costs of the fund are continuously incurred at a rate f (x(t)) which is a positive function of x(t). To reduce these costs the controller can at any time increase or decrease the fund by a quantity ξ ∈ <, but each time the system is controlled a cost C(ξ) is incurred. We will assume for these controlling costs the form C(ξ) = k + c(ξ) where k > 0 is a fixed component and c(ξ) is a component which increases with the size of the intervention. Given a discount rate γ > 0 our objective is to minimize the expected discounted value of all these costs over an infinite planning horizon. ∗ Department of Applied Mathematics, University of Torino, Piazza Arbarello 8, I-10122 Torino, Italy. E-mail: stefano.baccarin@unito.it † Department of Economics, University of Parma, Via J.F. Kennedy 6, I-43100 Parma, Italy. E-mail: simona.sanfelici@unipr.it 1 For a concrete financial application of this model one may consider the socalled “stochastic cash management problem”. In this case x(t) is a cash fund which fluctuates randomly because it collects a great number of different receipts and payments. Whenever the cash level is positive an opportunity (holding) cost must be accounted for and whenever cash falls below zero penalty costs are incurred. Therefore it would be convenient for the treasurer to maintain at zero the cash fund by continuously buying (when x(t) > 0) or selling (when x(t) < 0) short-term securities, but each time he converts cash into securities he must bear transaction costs. The presence of a fixed component in the transaction costs makes a continuous control unprofitable and the financial manager intervenes only at isolated points of time and for discrete operations. From a theoretical point of view our problem is one of impulse control where a policy is made up of a sequence ti of stopping times and corresponding random jumps ξi enforced upon the system. It is well known that applying the optimality principle of Dynamic Programming we can associate an impulse control problem with a quasi-variational inequality1 . If one succeeds in finding a regular solution of this inequality it can be shown that this solution is the value function and from this solution it is possible to derive the optimal policy (see, for example, Richard (1977), Harrison, Sellke and Taylor (1983), Eastham and Hastings (1988), Korn(1997,1998)). In this paper using a variational approach and the functional analysis techniques developed in the monographs of Bensoussan e Lions (1982,1984) we go beyond “verification theorems”. Under general assumptions we show that the value function of our problem is always a weak solution of the associated quasi-variational inequality in a suitable Sobolev space. From this solution we can obtain the optimal control whose existence is therefore always ensured. The analytical solution in the form of a control band policy can be obtained only in some special cases involving, for instance, linear or quadratic holdingpenalty costs (see Harrison, Sellke and Taylor (1983), Baccarin (2002)). For more general specifications, a numerical approach to the problem is compulsory. We perform a Finite Element (FE) approximation by means of continuous, piecewise linear functions defined on a suitable truncated domain [−r, r] and vanishing on the boundary. The convergence of the solution ur of the truncated problem to the value function as r → +∞ can be shown. The FE approximation urh is obtained by means of the “Bensoussan-Lions discrete iterative process”, whose convergence analysis on bounded domains can be found in Loinger (1980), Cortey-Dumont (1980) and Boulbrachene (1998). The paper is organized as follows. In §2 we give a precise formulation of our impulse control problem and we recall some results of Bensoussan and Lions (1984) concerning a variational inequality in a weighted Sobolev space. In §3 we prove some properties of the non local operator associated with the transaction cost structure. Furthermore we consider a weak formulation of the quasi variational inequality in a weighted Sobolev space and we show the existence of 1 A formal argument which shows how to derive the quasi-variational inequality from Belmann’s optimality principle can be found, for the cash management problem, in Constantinides and Richard (1978) or in Baccarin (2002). 2 a minimum solution for this inequality. §4 is devoted to the characterization of the value function of the problem as the minimum solution of the quasi variational inequality. Moreover using the value function we prove the existence of an optimal impulse policy dividing < in two regions: a continuation set where the system evolves freely and a transaction region where the system is controlled. Finally, in §5 we show the FE approach for computing the value function and some numerical simulations, with different data specifications, are performed and analyzed in §6. 2 Problem formulation and preliminary results We consider a filtered probability space S = (Ω, F, P, Ft ) and a one-dimensional standard Wiener process Wt adapted to Ft . We are given the following functions ( σ(x) > λ > 0 , σ(x) ∈ W 1,∞ (R) (1) b(x) ∈ W 1,∞ (R) . The dynamics of the cash fund x(t), in absence of control, is described by the following Itô stochastic differential equation ( dx(t) = b(x)dt + σ(x)dWt x(0) = x . An impulse policy V = (t1 , ξ1 ; t2 , ξ2 ; ...; ti , ξi ; .....) is an increasing sequence (i ≥ 1) of stopping times 0 ≤ t1 ... ≤ ti ≤ ti+1 .....(with respect to Ft ) and corresponding random variables ξi , which represent the jumps enforced upon the system. An impulse control is said to be admissible if it satisfies the two feasibility conditions ti → ∞ a.s. when i → ∞ (2) ξi is Fti measurable ∀i ≥ 1 where Fti is the minimum σ-algebra of events prior to ti 2 . We will denote by A the set of admissible policies. When policy V is used, the controlled process y(t) is generated by the following set of stochastic differential equations with random initial conditions3 when ti ≤ t < ti+1 (i ≥ 0) dy(t) = b(y)dt + σ(y)dWt (3) y(ti ) = y(t− ) + ξ ∀i ≥1 i i y(t0 ) = x 2 t → ∞ implies that only a finite number of actions can be taken in any bounded interval; i ξi is Fti measurable means that the ξi decision depends only on the information available at ti . 3 Since σ(x) and b(x) are Lipschitz continuous this system has a strong solution: the probability space S = (Ω, F, P, Ft ) is fixed and it will not depend on the control V . 3 − 4 where t0 ≡ 0 and y(t− i ) = lim y(t) if ti−1 < ti , or y(ti ) = y(ti−1 ) if ti−1 = ti . t↑ti By yx (t) we will denote the controlled process y(t) starting in x = y(0). We make the following assumptions on the holding/penalty costs f (x) and on the variable part c(ξ) of the transaction costs C(ξ) : | s (4) c(ξ) → ∞ if | ξ |→ ∞, c(ξ) is sub-linear c(ξ1 + ξ2 ) ≤ c(ξ1 ) + c(ξ2 ) (5) f (x) is measurable; 0 ≤ f (x) ≤ f0 (1 + |x| ), s > 0 c(ξ) : < → <+ is continuous, c(0) = 0, ∀ξ1 , ξ2 ∈ < . Thus we are imposing polynomial growth conditions on f (x) and we suppose c(ξ) continuous, sublinear and unbounded from above when | ξ |→ ∞. The sublinearity condition on c(ξ) is largely verified in applications, because implementing the same transfer of ξ = ξ1 + ξ2 is usually cheaper if we make it with one transaction. To each feasible control V and initial condition x = y(0) it is associated the cost ) (∞ Z ∞ X −γs −γti e f (yx (s))ds , Jx (V ) = Ex χti <∞ + C(ξi ) e 0 i=1 where γ > 0 is the discount rate and 1 χti <∞ (ω) = 0 if ti (ω) < ∞ elsewhere . Our objective is to characterize the value function U (x) = inf Jx (V ) and V ∈A to show that, for every x ∈ <, there exists an optimal control V̂ such that inf Jx (V ) = Jx (V̂ ) . V ∈A We define the following functions a2 (x) = From (1) it results a2 (x) ∈ W 1,∞ (<), a2 (x) > a1 (x) ∈ L∞ . σ 2 (x) 2 , λ2 2 a1 (x) = −b(x) + a02 (x). ≡α>0 (6) Since the cost functions of our problem are unbounded in < it is convenient to characterize U (x) as a solution of a variational problem not in W 1,∞ (<) but in a weighted Sobolev space, in such a√way to allow for the limit condition 2 U (x) → +∞ as |x| → ∞. Let wµ (x) = e−µ 1+x be the weight function (µ > 0); we define Lp,µ (<) functions as follows Z p Lp,µ (<) = f | wµp (x) | f | dx < ∞ . 4 More than one control action is allowed at the same time instant. 4 We consider weighted Sobolev spaces that, in the sense of distributions, can be defined as W 1,p,µ (<) W 2,p,µ (<) = = {f ∈ Lp,µ | f 0 (x) ∈ Lp,µ } {f ∈ Lp,µ | f 0 (x), f 00 (x) ∈ Lp,µ } and in particular H 1,µ = W 1,2,µ (R), H 2,µ = W 2,2,µ (R) . We denote by Z Z Z 2 2 (u, v)µ = u v wµ dx, (u, v)H 1,µ = u v wµ dx + u0 v 0 wµ2 dx R R R the inner products respectively in L2,µ and H 1,µ , which are Hilbert spaces, and by |u|µ ||u||µ the corresponding norms. Let us consider the continuous bilinear form in H 1,µ Z Z 2a2 xµ 0 (7) aµ (u, v) = )u v wµ2 dx. a2 u0 v 0 wµ2 dx + (a1 − √ 1 + x2 R R For fixed µ > 0 it can be easily shown that there exist ρ, β > 0 such that it results 2 2 aµ (u, u) + ρ |u|µ ≥ β ||u||µ (8) that is to say the bilinear form Aρ,µ (u, v) = aµ (u, v)+ρ (u, v)µ becomes coercive on H 1,µ . We set Z ∞ u0 (x) = E e−γt f (x(t))dt (9) 0 which is the cost function corresponding to no action (i.e. t1 = +∞). From (4) it follows that u0 (x) is finite and continuous in <. In the sequel we will be interested in the following variational inequality aµ (u, v − u) + γ (u, v − u)µ ≥ (f, v − u)µ 1,µ (10) such that v ≤ ψ ∀v ∈ H1,µ 0 u ∈ H , u ≤ ψ, 0 ≤ u(x) ≤ u (x) where we are making the underlying assumptions on the obstacle ψ ψ continuous, ψ ∈ L2,µ , ψ ≥ 0. (11) In Bensoussan and Lions (1984) it is proved that there exists one and only one solution uψ (x) of (10). We consider now the process y(t) controlled by an admissible policy V. Bensoussan and Lions (1984) show that the only solution5 uψ (x) verifies the next essential theorem 5 In the sequel we will always take the continuous representative of uψ (x). 5 Theorem 1 Let ti ≤ ti+1 be two consecutive stopping times of V and y(t) the corresponding controlled process. The solution uψ (x) of (10) verifies Z ti+1 f (y(s)e−γs ds (12) e−γti uψ (y(ti ))χti <∞ ≤ E χti <∞ ti . +e−γti+1 uψ (y(t− i+1 ))χti+1 <∞ |Fti Furthermore if ti+1 = t = inf {t ≥ ti |uψ (y(t− )) = ψ(y(t− )) } then the equality holds Z t −γti uψ (y(ti ))χti <∞ = E χti <∞ e f (y(s)e−γs ds (13) ti +e−γt uψ (y(t))χt<∞ |Fti . This a fundamental result because the regularity properties of uψ ∈ H 1,µ are much weaker than those usually required to apply Itô’s Lemma. Moreover the solution of (10) ensures the existence and uniqueness of uψ . 3 The quasi-variational inequality associated with the impulse control problem In order to consider the controlling costs C(ξ) we introduce the non local operator M u(x) = inf [C(ξ) + u(x + ξ)] ξ∈< which from (5) is well defined for all u(x) : < → < bounded from below. The next theorem gives some properties of M u. Theorem 2 Given u(x) bounded from below and assumptions (5) the operator M satisfies the following properties. 1. If u(x) ≥ z(x) ≥ 0 then M u(x) ≥ M z(x) ≥ 0 . 2. If u(x) is continuous then there exists a Borel measurable function ξˆu (x) : < → < which verifies M u(x) = C(ξˆu (x)) + u(x + ξˆu (x)) ∀x ∈ < . (14) 3. If u(x) is continuous then M u(x) is continuous . 4. Assume 0 ≤ u0 (x) ≤ ... ≤ un (x) ≤ ... ≤ u(x) and kun − ukC 0 (I¯r ) → 0, when n → ∞, ∀r > 0 fixed. Then it follows kM un − M ukC 0 (I¯r ) → 0, ∀r > 0 fixed. Proof. Property 1 is evident from the definition of M u. To deduce property 2 we observe that, for fixed x, the function Gu,x (ξ) ≡ C(ξ) + u(x + ξ) is continuous and we have lim Gu,x (ξ) = +∞ because lim C(ξ) = +∞ and u(x) is bounded from be|ξ|→∞ |ξ|→∞ low. Therefore Gu,x (ξ) achieves its greatest lower bound and for every x ∈ < 6 we can choose a point of global minimum ξˆu (x) of Gu,x (ξ) in order to define a Borel measurable function ξˆu (x) : < → < which satisfies (14). We prove property 3 by showing that for any sequence xn → x, we have M u(xn ) → M u(x), ∀x ∈ <. For ξ ∈ < fixed it holds M u(xn ) ≤ C(ξ) + u(xn + ξ) and from xn → x and the continuity of u it follows lim sup M u(xn ) ≤ C(ξ) + u(x + ξ) . n→∞ Since we can choose ξ arbitrarily, we obtain for any sequence xn → x lim sup M u(xn ) ≤ M u(x) . n→∞ Let xn → x be a sequence such that the corresponding M u(xn ) converges. From (14) it holds M u(xn ) = C(ξˆu (xn )) + u(x + ξˆu (xn )) and from the continuity of u we obtain, for r > 0 and n large enough C(ξˆu (xn )) + u(x + ξˆu (xn )) ≤ C(0) + u(x) + r . Since lim C(ξ) = +∞ the sequence ξˆu (xn ) stays bounded and we can find |ξ|→∞ a subsequence (xn , ξˆu (xn )) converging to (x, ξ ∗ ). Since C(ξ) is continuous it follows lim M u(xn ) = C(ξ ∗ ) + u(x + ξ ∗ ) ≥ M u(x). n→∞ This result is true for any sequence xn → x such that M u(xn ) converges. Therefore we can deduce, for any sequence xn → x lim sup M u(xn ) ≤ M u(x) ≤ lim inf M u(xn ) n→∞ n→∞ and consequently M u is continuous. In order to prove property 4 we first show that M un (x) converges pointwise to M u(x), ∀x ∈ <. From un (x) ≤ u(x) it follows M un (x) = C(ξˆun (x)) + un (x + ξˆun (x)) ≤ C(0) + u(x) . Since C(ξˆun (x)) is bounded and ˆ ξun (x) ≤ L, ∀n ∈ N. Setting r = |x| + L we obtain M un (x) ≥ ≥ lim C(ξ) = +∞, for fixed x it must be |ξ|→∞ C(ξˆun (x)) + u(x + ξˆun (x)) − ku − un kC 0 (I¯r ) M u(x) − ku − un kC 0 (I¯r ) . 7 From property 1, M un (x) ≤ M u(x) and therefore it holds M u(x) − ku − un kC 0 (I¯r ) ≤ M un (x) ≤ M u(x) ∀n ∈ N. By assumption ku − un kC 0 (I¯r ) → 0, ∀r > 0 fixed, and this implies lim M un (x) = M u(x). We can repeat the preceding argument ∀x ∈ < obtain- n→∞ ing that M un (x) converges pointwise to M u(x) in <. Property 4 then follows from properties 1, 3 and Dini’s theorem. We consider now the set of functions K = z ∈ H 1,µ | 0 ≤ z ≤ u0 and we define an operator T : K → K, which relates z ∈ K to the continuous solution η ∈ K of (10) corresponding to the obstacle ψ = M z, where M z is continuous from property 3 of M. The operator T has the fundamental monotonicity property T z1 ≤ T z 2 if z1 ≤ z2 (15) which follows from property 1 of M and the fact that the solution of (10) increases when the obstacle ψ is increasing (see Bensoussan, Lions 1978). We will look for the value function U (x) of our problem among the solutions of the underlying quasi-variational inequality6 aµ (u, v − u) + γ (u, v − u)µ ≥ (f, v − u)µ (16) ∀v ∈ H 1,µ such that v ≤ M u u ∈ H 1,µ , u ≤ M u, 0 ≤ u ≤ u0 . It is immediate to observe that u is a solution of (16) if and only if u is a fixed point of the operator T. It should be noted also that if a regularity result of the type u ∈ H 2,µ holds true, than one can easily show the equivalence between (16) and the strong formulation Au ≤ f u ≤ Mu (17) (Au − f )(u − M u) = 0 0 ≤ u ≤ u0 where A is the second order differential operator Au = − du du d (a2 ) + a1 + γu . dx dx dx The strong formulation (17) of the quasi-variational inequality is the one commonly used to obtain verification theorems for stationary (infinite horizon) impulse control problems. Inequality (16) may have many solutions, as it is usual with unbounded domains, but the most relevant solution comes out to be the minimum solution. 6 The attribute quasi-variational points out that we have an implicit obstacle ψ = M u defined by means of the unknown function u. 8 Theorem 3 Under assumptions (4), (5), (6) the quasi variational inequality (16) has a minimum solution umin . Moreover, the function umin verifies umin (x) ≤ inf Jx (V ) V ∈A ∀x ∈ < . (18) Proof. We consider the sequence un = T un−1 starting from u0 = 0. By u1 = T u0 ≥ 0 = u0 and the monotonicity property (15) of T we obtain recursively un+1 = T un ≥ un = T un−1 and the sequence un ∈ C 0 (<) is increasing. Therefore un converges pointwise to a function umin , with 0 ≤ umin ≤ u0 . From un+1 ≤ M un it follows un+1 (x) ≤ C(ξ) + un (x + ξ) ∀ξ ∈ < and when n → ∞, as ξ is arbitrary, we deduce immediately umin (x) ≤ M umin (x), ∀x ∈ <. The sequence un converges strongly in L2,µ to umin because 0 ≤ un ≤ u0 and u0 ∈ L2,µ . Setting v = 0 into (10), which is always admissible, we obtain aµ (un , un ) + γ (un , un )µ ≤ (f, un )µ ∀n ∈ N. Using (8) we can deduce, for ρ large enough 2 β ||un ||µ 2 aµ (un , un ) + γ |un |µ + (ρ − γ) |un |µ ≤ (f, un )µ + (ρ − γ) |un |µ |un |µ (|f |µ + (ρ − γ) u0 µ ) ≤ and then it follows 2 ≤ 2 β ||un ||µ ≤ |f |µ + (ρ − γ) u0 µ ∀n ∈ N . The norms ||un ||µ stay bounded in H 1,µ , thus we can find a subsequence um which converges weakly to a function u∗ ∈ H 1,µ . Since the injection of H 1,µ in L2,µ is compact we obtain also that um converges strongly to u∗ in L2,µ and consequently u∗ = umin almost everywhere. If we consider any fixed interval Ir , the norms ||um ||H 1 (Ir ) of um in H 1 (Ir ), remain uniformly bounded by a constant Mr , which depends only on r, ∀m ∈ N . Then we can apply Morrey’s theorem on Sobolev spaces (see, for example, Brezis 1983) and we obtain 1 |um (x) − um (y)| ≤ Cr |x − y| 2 ∀m, ∀x, y ∈ I¯r that is the functions um (x) ∈ C 0 (I¯r ) are uniformly Hölder continuous on I¯r , ∀r > 0 fixed. This implies umin ∈ C 0 (I¯r ) and7 kum − umin kC 0 (I¯r ) → 0 7 The ∀r fixed, when m → ∞. (19) function umin , which we defined pointwise, is the continuous representative of u∗ . 9 Let v ∈ H 1,µ satisfy v ≤ M umin . If we choose a function g ∈ Cc∞ (<) such that 1 if x ∈ I1 g(x) = 0 if x ∈ </I2 and we define gr (x) = g( xr ), r > 0, one can easily show that, when r → ∞, vgr → v in H 1,µ . For ε > 0 we have (v − ε) gr ≤ = ≤ (M umin − ε) gr M um gr + (M umin − M um )gr − εgr M um gr + (kM umin − M um kC 0 (I¯2r ) − ε)gr . From Property 4 of M and (19) it follows kM umin − M um kC 0 (I¯2r ) → 0, when m → ∞, and for fixed r, ε we can find N = N (ε, r) such that (v − ε) gr ≤ M um gr ≤ M um ∀m > N (ε, r) . The function (v − ε) gr is therefore an admissible test function for every problem um+1 , with m > N, and substituting (v − ε) gr into (10) we obtain aµ (um+1 , (v − ε) gr − um+1 ) + γ (um+1 , (v − ε) gr − um+1 )µ ≥ ≥ (f, (v − ε) gr − um+1 )µ ∀m > N Since aµ (u, u) + γ(u, u)µ is weakly lower semi-continuous in H 1,µ , and um converges weakly to umin , we can deduce aµ (umin , (v − ε) gr − umin ) + γ (umin , (v − ε) gr − umin )µ ≥ ≥ (f, (v − ε) gr − umin )µ ∀v ∈ H 1,µ , v ≤ M umin , ∀ε, r > 0 . Making ε → 0 and afterwards r → ∞ it follows aµ (umin , v − umin ) + γ (umin , v − umin )µ ≥ (f, v − umin )µ ∀v ∈ H 1,µ such that v ≤ M umin and we have proved that umin is a solution of (16). It is immediate to show that umin is the minimum solution. Any solution u of (16) is a fixed point of the operator T and from u0 = 0 ≤ u we obtain by recurrence that um = T um−1 ≤ u = T m u. When m → ∞ this inequality implies umin ≤ u. In order to prove the second part of the theorem we observe that the function un (x) is the solution of (10), corresponding to the obstacle ψ = M un−1 (x). Thus we can apply (12) to tj , tj+1 , and un−j , with 0 ≤ j ≤ n − 1. It follows that e−γtj un−j (yx (tj ))χtj <∞ ≤ (20) i h R tj+1 ≤ E χtj <∞ tj f (yx (s)e−γs ds + e−γtj+1 un−j (yx (t− j+1 ))χtj+1 <∞ Ftj . 10 As un−j ≤ M un−j−1 we can deduce un−j (yx (t− j+1 )) ≤ = C(ξj+1 ) + un−j−1 (yx (t− j+1 ) + ξj+1 ) (21) C(ξj+1 ) + un−j−1 (yx (tj+1 )). Substituting (21) into (20) and taking the mathematical expectation we have j Ee−γt h un−jR(yx (tj ))χtj <∞ ≤ t ≤ E χtj <∞ tjj+1 f (yx (s)e−γs ds + e−γtj+1 (C(ξj+1 ) + un−j−1 (yx (tj+1 )))χtj+1 <∞ . If we recall that u0 ≡ 0, t0 ≡ 0, yx (t0 ) = x and we sum up all these inequalities for j varying from 0 to n − 1 we obtain Z tn n X un (x) ≤ E e−γtj C(ξj )χtj <∞ . f (yx (s)e−γs ds + 0 j=1 When n → ∞ this inequality implies umin (x) ≤ Jx (V ) 4 ∀x ∈ <, ∀V ∈ A. Value function and optimal policy We describe now the optimal policy V̂ . For this purpose we define in < two regions: a continuation region F where the system evolves freely, which is the open set F = {x ∈ < : umin (x) < M umin (x)} (22) and the complementary intervention region G = F C , where the system is controlled. The first optimal stopping time is defined to be the first exit time of the uncontrolled process from F t̂1 = inf {t ≥ 0 | yx (t) ∈ / F} . In t̂1 it is enforced the optimal jump ˆ x (t̂− )) ξ(y 1 ˆ ξ1 = arbitrary if t̂1 < ∞ if t̂1 = +∞ the chosenfunction which verifies (14) corresponding to where ξˆ = ξˆumin (x) is umin . The subsequent t̂i+1 , ξˆi+1 are defined recursively by (i ≥ 1) t̂i+1 = inf t ≥ t̂i | yx (t− ) ∈ /F ˆ x (t̂− )) ξ(y if t̂i+1 < ∞ i+1 ξˆi+1 = arbitrary if t̂i+1 = +∞ 11 . (23) It is interesting to note that, if the optimal stopping time t̂i is finite, ξˆi puts back the system inside F. This implies t̂i+1 > t̂i , that is to say the optimal stopping times are separated whenever they are finite. In fact we have umin (yx (t̂i )) = ˆ umin (yx (t̂− i ) + ξi ) = ˆ M umin (yx (t̂− i )) − C(ξi ) ˆ ˆ C(ξˆi + ξ) + umin (yx (t̂− i ) + ξi + ξ) − C(ξi ) ≤ ∀ξ ∈ < and from the sublinearity of c(ξ) we obtain umin (yx (t̂i )) ≤ = ˆ umin (yx (t̂− i ) + ξi + ξ) + C(ξ) − C(0) umin (yx (t̂i ) + ξ) + C(ξ) − C(0) ∀ξ ∈ < . As ξ is arbitrary we can deduce (t̂i < ∞) umin (yx (t̂i )) ≤ M umin (yx (t̂i )) − C(0) < M umin (yx (t̂i )) and therefore yx (t̂i ) ∈ F and consequently t̂i+1 > t̂i . For policy V̂ to be admissible it has still to be shown that t̂i → ∞ a.s. when i → ∞. Next theorem shows that V̂ is admissible and optimal. o n Theorem 4 Under assumptions (1), (4), (5), the policy V̂ = t̂i , ξˆi defined in (23) is admissible and optimal. Furthermore the minimum solution umin of (16) is the value function, that is to say umin (x) = inf Jx (V ) = Jx (V̂ ) . (24) V ∈A Proof. We look at umin (x) as the solution of (10) with ψ = M umin . Since it is always − umin (yx (t̂− i )) = M umin (yx (t̂i )) = ψ we can apply (13) between t̂i and t̂i+1 . It follows (25) e−γ t̂i umin (yx (t̂i ))χt̂i <∞ = i h R t̂i+1 F . f (yx (s)e−γs ds + e−γ t̂i+1 umin (yx (t̂− ))χ = E χt̂i <∞ t̂ t̂i+1 <∞ t̂i i+1 i But there holds ˆ umin (yx (t̂− i+1 ))χt̂i+1 <∞ = C(ξi+1 )χt̂i+1 <∞ + umin (yx (t̂i+1 ))χt̂i+1 <∞ and substituting (26) into (25) and taking expectation we obtain " Z t̂i+1 −γ t̂i f (yx (s)e−γs ds umin (yx (t̂i ))χt̂i <∞ = E χt̂i <∞ Ee t̂i + χt̂i+1 <∞ C(ξˆi+1 ) e−γ t̂i+1 i + E e−γ t̂i+1 umin (yx (t̂i+1 ))χt̂i+1 <∞ . 12 (26) (27) Summing up (26) for i varying from 0 to n − 1 it follows "Z t̂n umin (x) = E f (yx (s)e−γs ds (28) 0 + n X χt̂i <∞ C(ξˆi ) e −γ t̂i i=1 # +E e−γ t̂n umin (yx (t̂n ))χt̂n <∞ . As umin ≥ 0, f ≥ 0 this equality implies C(0) ∞ X i=1 χt̂i <∞ e−γ t̂i < ∞ a.s. whence we obtain immediately that V̂ is admissible t̂i → ∞ a.s. Furthermore from (28) it follows "Z t̂n umin ≥ E when i → ∞ . f (yx (s)e−γs ds + 0 n X # χt̂i <∞ C(ξˆi ) e−γ t̂i , i=1 ∀n and when n → ∞ we can deduce umin (x) ≥ Jx (V̂ ) and consequently from (18) we obtain (24). 5 Computation of the value function The numerical solution to problem (16) can be accomplished in two steps. First, we consider a sequence of approximated problems over suitable truncated domains Ir = [−r, r], r → +∞, with homogeneous Dirichlet boundary conditions. The convergence in L∞ loc (<) of the solutions of the truncated problems to u min as r → +∞ is shown in Theorem 5. Second, thanks to this convergence result, we perform a Finite Element (FE) approximation to (16) by means of continuous, piecewise linear functions defined on a convenient truncated interval I r and vanishing on the boundary. This procedure yields an accurate approximation to umin on any fixed compact domain, if we choose r sufficiently large with respect to the domain of interest. For r > 0 fixed, we consider the following quasi variational inequality in H01 (Ir ) ar (u, v − u) + γ (u, v − u)r ≥ (f, v − u)r ∀v ∈ H01 (Ir ) such that v ≤ Mr u (29) u ∈ H01 (Ir ), u ≤ Mr u, 0 ≤ u ≤ u0 . 13 where ar (u, v) = Z a2 u0 v 0 dx + Ir Z Mr u(x) = a1 u0 v dx, (u, v)r = Ir inf ξ∈< x+ξ∈I¯r Z u v dx, Ir C(ξ) + u(x + ξ) and u0 ∈ H01 (Ir ) is the solution of the equation ∀v ∈ H01 (Ir ) . ar (u, v) + γ (u, v)r = (f, v)r (30) Under assumptions (4), (5), (6), there is one and only one solution u0 ≥ 0 of (30) and one and only one solution ur of (29) (see Bensoussan and Lions (1984)). The solution ur (x) has the probabilistic interpretation ur (x) = umin (x) = M inJxr (V ) V ∈A where Jxr (V )=E (∞ X C(ξi ) e −γti χti <∞ + i=1 Z τr e −γs f (yx (s))ds 0 ) and τr = inf t ≥ 0 | yx (t− ) or yx (t) ∈ / Ir is the first exit time from the interior of Ir . We denote by ũr the function ur continued by 0 outside Ir . In the next theorem we show the above mentioned convergence result of ũr to umin when r → ∞. Theorem 5 Under assumptions (4), (5), (6), when r → ∞ we have kũr − umin kC 0 (K) kũr − umin kL2,µ → → 0 ∀K ⊂ <, compact 0, and there exists a subsequence um such that kũm − umin k → 0 weakly in H 1,µ . Proof. By restricting the integration to Ir , we define Z Z 2a2 xµ 0 )u v wµ2 dx (a1 − √ a2 u0 v 0 wµ2 dx + aµ,r (u, v) = 2 1 + x Ir I Zr 2 u v wµ dx . (u, v)µ,r = Ir Let w ∈ H01 (Ir ) satisfy w ≤ Mr ur . As (wµ2 ) = − √2µx w2 , setting u = ur and 1+x2 µ 2 v = ur + wµ (w − ur ) in the first inequality of (29), we obtain ar (ur , wµ2 (w − ur )) + γ ur , wµ2 (w − ur ) r = = aµ,r (ũr , w̃ − ũr ) + γ (ũr , w̃ − ũr )µ,r ≥ ≥ (f, wµ2 (w − ur ))r = (f, w̃ − ũr )µ,r . 0 14 Therefore for every w ∈ H01 (Ir ) such that w ≤ Mr ur we have aµ (ũr , w̃ − ũr ) + γ (ũr , w̃ − ũr )µ ≥ (f, w̃ − ũr )µ . (31) We consider now the sequence ũr , with r ∈ N. From the probabilistic interpretation of ur it follows ũ1 ≤ ũ1 ≤ ... ≤ ũr ≤ .. ≤ umin and ũr converges pointwise to a function u∗ . Setting w̃ = 0 into (31), ∀r ∈ N, and using (8) we obtain in the same way as in Theorem 3 β ||ũr ||µ ≤ |f |µ + (ρ − γ) |umin |µ ∀r ∈ N . (32) Therefore there exists a subsequence ũm which converges pointwise to u∗ , strongly in L2,µ , weakly in H 1,µ . From Morrey’s theorem the functions ũm are uniformly Hölder continuous on < and consequently u∗ is continuous and ũm converges uniformly to u∗ , on every compact subset of <. It has still to be shown that u∗ = umin . Since u∗ ≤ umin it is sufficient to show that u∗ is a solution of (16). From ur ≤ Mr ur it follows ũr ≤ Mr ũr and we obtain also that u∗ ≤ M u∗ . It remains to prove that for every w ∈ H 1,µ such that w ≤ M u∗ it holds aµ (u∗ , w − u∗ ) + γ (u∗ , w − u∗ )µ ≥ (f, w − u∗ )µ . This is shown in the same way as in the proof of Theorem 3. This convergence result allows us to compute umin by solving the truncated problem (29), for a conveniently chosen interval Ir including the domain of interest. The FE method performs a decomposition of Ir into some finite number Nh of subintervals −r = x0 < x1 < . . . < xNh = r. Let h > 0 be the maximum length of the intervals [xi , xi+1 ] of the decomposition and Vh (as h → 0) be the space of all continuous, piecewise linear functions defined on Ir and vanishing on the boundary. We remark that we shall only consider the piecewise linear element case due to the limited higher order regularity for the solution ur . The discrete problem reads ar (urh , vh − urh ) + γ(urh , vh − urh )r ≥ (f, vh − urh )r (33) urh ∈ Vh , urh ≤ πh M urh , ∀vh ∈ Vh , vh ≤ πh M urh , where πh is the interpolation operator from C 0 (Ir ) onto Vh . The convergence analysis of the FE approximation to models of the form (29) can be found in Loinger (1980), Cortey-Dumont (1980) for the coercive case and Boulbrachene (1998) for the noncoercive case. In particular, under the additional regularity assumptions ur ∈ W 2,p (Ir ), 2 ≤ p < ∞, Aur ∈ L∞ (Ir ) (34) the following optimal error estimate holds kur − urh kC 0 (Ir ) ≤ Ch2 | log h|3 . (35) Finally, combining (35) with the convergence result of Theorem 5, we get the following convergence result to the original problem (16) 15 Corollary 6 For any fixed compact set K, we get lim ( lim kumin − ũrh kC 0 (K) ) = 0. r→∞ h→0 Proof. It follows immediately from kumin − ũrh kC 0 (K) ≤ kumin − ũr kC 0 (K) + kũr − ũrh kC 0 (K) → 0, (36) as r → +∞ and h → 0. Analogously to the continuous case, the FE approximation urh is obtained by means of the following discrete iterative process starting from urh,0 = 0 ar (urh,n , vh − urh,n ) + γ(urh,n , vh − urh,n )r ≥ (f, vh − urh,n )r urh,n ∈ Vh , urh,n ≤ πh M urh,n−1 , ∀vh ∈ Vh , vh ≤ πh M urh,n−1 , (37) for n = 1, 2, . . . . The sequence {urh,n } defined by (37) is pointwise non dePNh −1 n ui ϕi (x), where creasing and converges to urh in Vh . If we set urh,n (x) = i=1 the ϕi ’s are the so called hat functions generating Vh , we get the following finite system of linear algebraic inequalities uTn AT (v − un ) ≥ bT (v − un ) un ≤ ψn−1 , v ≤ ψn−1 , (38) where A = (Ai,j ) is the stiffness matrix with Ai,j := ar (ϕj , ϕi ) + γ(ϕj , ϕi )r , un = (un1 , un2 , . . . , unNh −1 )T , b = (b1 , b2 , . . . , bNh −1 )T is the vector with bj := (f, ϕj )r and ψn−1 is the vector of the nodal values of M urh,n−1 . System (38) is equivalent to the following constrained system Aun ≤ b, un ≤ ψn−1 , (ψn−1 − un )T (Aun − b) = 0. (39) For the solution of (39), the projected SOR method has been used. 6 Numerical experiments In this section, we show some numerical results concerning different data specifications8 . The first set of simulations refers to a case with quadratic holding-penalty costs. Such kind of problems can be solved analytically as in Baccarin (2002) and the optimal control takes the form of a control band policy with barriers d < D < U < up. In particular, d and up are the free boundaries separating the continuation region {u < M u} from the intervention region {u = M u}. In Table 1, we list the norms of the approximation error against the analytical solution for decreasing h in the L2 , H 1 and L∞ -norms. The parameter values are a2 (x) = 0.5, a1 (x) = 0, k = 1, γ = 0.5, f (x) = x2 , c(x) = |x|. Each line i (i = 1, 2, 3, 4) of the table refers to a different approximation, where 8 All the simulations are run in Matlab on a Pentium IV processor. 16 i 1 2 3 4 ri 30 35 40 45 h 0.4 0.2 0.1 0.05 kumin − urhi kL2 0.060604 (0.00) 0.015399 (3.94) 0.002095 (7.35) 0.005108 (0.41) kumin − urhi kH 1 0.384473 (0.00) 0.185515 (2.07) 0.093810 (1.98) 0.048068 (1.95) kumin − urhi kL∞ 0.038770 (0.00) 0.008477 (4.57) 0.001784 (4.75) 0.002772 (0.64) Table 1: The columns list the norms in the spaces L2 (−3, 3), H 1 (−3, 3) and the L∞ norm of the approximation error umin − urhi for decreasing h and increasing ri . The ratios of error norms on successive grids are in parentheses. i 1 2 3 4 ri h 30 0.4 35 0.2 40 0.1 45 0.05 analytical d -2.2 -2.0 -2.0 -2.05 -2.0349 D -0.6 -0.4 -0.5 -0.5 -0.4943 U 0.6 0.4 0.5 0.5 0.4943 up 2.2 2.0 2.0 2.05 2.0349 Table 2: Numerical and analytical barriers of the control band policy. the computational domain Iri = [−30 − 5(i − 1), 30 + 5(i − 1)] is discretized by a nonuniform mesh having smaller elements of length h = 0.4/2i−1 in the core region [−3 − (i − 1), 3 + (i − 1)]. In parentheses, we record the ratios i k of errors of the numerical approximations sequence, kumin − urhi k/kumin − urh/2 obtained by halving the mesh size h on successive grids. The integrals in the discrete formulation (39) are computed by means of the fourth order GaussLegendre quadrature rule and the error estimates are confined to the inner region K = [−3, 3]. In order to reduce approximation errors in the computation i of the nonlinear obstacle, the nodal values of M urh,n−1 are computed using the ri interpolation values of uh,n−1 on five intermediate points for each finite element. As it can be seen from the table, the convergence rate is affected by the presence of the term kumin − ũr kC 0 (K) in (36). The evaluation of this truncation error and the question of how large the computational domain Ir should be conveniently chosen in terms of a given error tolerance will be the object of future research. The numerical barriers of the control band policy are compared with the exact ones in Table 2. The barriers are always mesh points and there is strong evidence that the computed barriers converge to the corresponding analytical values. In all our simulations a mesh size h smaller than a given tolerance ε gives an approximation error smaller than ε for the barriers. Figure 1 shows the i } and corresponding obstacles for i = 3. approximating sequence {urh,n In the following simulation, we try to mimic a more complex and verisimilar study case where we assume a variable diffusion coefficient and more realistic 17 8 7 6 5 4 3 2 1 0 −5 −4 −3 −2 −1 0 1 2 3 4 5 i } (solid line) and corresponding obstaFigure 1: Approximating sequence {urh,n cles (dotted line) for i = 3 in the quadratic case. cost functions are considered9 . Hence, we assume b(x) ≡ 0.05, 2 C(ξ) = 0.01 + c(ξ), σ(x) = 0.3 − 0.2 e−x , 2 0.05x − 0.01x e−x , x ≥ 0 f (x) = 2 −0.1x + 0.04x , x<0 γ = 0.03, where c(ξ) = 0.03|ξ| − 0.01e−|ξ| + 0.01. The results are displayed in Figure 2. The computed barrier values are d = −0.22, D = −0.04, U = 0.02, up = 0.28 for a discretization step h = 0.02. Since there is a small interval near the origin where the transaction costs are larger than the carrying costs, we observe that there exists a small continuation region around the cash level x = 0. However, due to the presence of a positive drift, we notice that this region is slightly asymmetric, although this effect is somehow compensated by the higher penalty cost rate. 9 For instance, in the cash management problem the carrying cost rate is certainly higher for negative than for positive cash levels. 18 0.34 0.335 0.33 0.325 u 0.32 0.315 0.31 0.305 0.3 0.295 d = −0.22; D = −0.04; U = 0.02; up = 0.28 0.29 −0.4 −0.3 −0.2 −0.1 0 x 0.1 0.2 0.3 Figure 2: Numerical solution uh (solid line) versus the obstacle M uh (dotted line) in the non quadratic case. References • Baccarin, S. (2002): Optimal impulse control for cash management with quadratic holding-penalty costs. Decisions in Economics and Finance 25 19-32. • Bensoussan, A. and Lions, J.L. (1982): Applications of variational inequalities in stohastic control. North Holland. • Bensoussan, A. and Lions, J.L. (1984): Impulse Control and QuasiVariational Inequalities. Gauthiers-Villars, Paris. • Boulbrachene (1998): The noncoercive quasi-variational inequalities related to impulse control problems, Comput. Math. Appl., 35 , no. 12, 101–108. • Brezis, H. (1983): Analyse fonctionnelle. Théorie et applications. Masson Editeur, Paris. • Cortey-Dumont, P. (1980): Approximation numerique d’une I.Q.V. liée a des problemesde gestion de stock. RAIRO, Num. Anal., 14 335-346. • Eastham, J. and Hastings K. (1988): Optimal impulse control of portfolios. Math. Oper. Res. 4 588-605. 19 • Harrison, J.M., Sellke, T. and Taylor, A. (1983): Impulse control of a Brownian motion. Math. Oper. Res. 8 454-466. • Korn, R. (1997): Optimal impulse control when control actions have random consequences. Math. Oper. Res. 22 639-667. • Korn, R. (1998): Portfolio optimisation with strictly positive transaction costs and impulse control. Finance and Stochastics 2 85-114. • Loinger, E. (1980): A finite element approach to a quasivariational inequality. Calcolo, 17 , no. 3, 197–209. • Richard, S.F. (1977): Optimal impulse control of a diffusion process with both fixed and proportional costs of control. SIAM J. Control Optimization 15 79-91. 20