/ OEWEV HD28 .M414 /^-\ 0 no. 3<^/D '' ALFRED P. WORKING PAPER SLOAN SCHOOL OF MANAGEMENT NfUMERICAL ANALYSIS OF A FREE-BOUNDARY SINGULAR CONTROL PROBLEM IN FINANCIAL ECONOMICS Ayman Hindy Huang Hang Zhu Chi-fu Massachusetts Institute of Technology WP#3615-93-EFA August 1993 MASSACHUSETTS INSTITUTE OF TECHNOLOGY 50 MEMORIAL DRIVE CAMBRIDGE, MASSACHUSETTS 02139 NUMERICAL ANALYSIS OF A FREE-BOUNDARY SINGULAR CONTROL PROBLEM IN FINANCIAL ECONOMICS Ayman Hindy Huang Hang Zhu Chi-fu Massachusetts Institute of Technology WP#3615-93-EFA August 1993 M.f.T. OCT LIBRARIES 1 5 1993 RECEIVED NUMERICAL ANALYSIS OF A FREE-BOUNDARY SINGULAR CONTROL PROBLEM IN FINANCIAL ECONOMICS Ayman Hindy, Chi-fu Huang, and Hang Zhu * August 1993 Abstract We analyze a numerical scheme for solving the consumption and investment allocation problem studied in a companion paper by Hindy, Huang, and Zhu (1993). For lem, Bellman's equation is The ential equation with gradient constraints. at which consumption occurs. ity solution to this prob- a differential inequality involving second order partial We Bellman's equation. solution involves finding a free-boundary prove that the value function We differ- is the unique viscos- describe a numerical analysis technique based on Markov chain approximation schemes. We approximate the original continuous time problem by a sequence of discrete parameter Markov chains control problems. We prove that the value functions and the optimal investment policies in the sequence of the ap- proximating control problems converge to the value function and the optimal investment policy, if it exists, of the original and abstinence regions in problem. the sequence of We show that the optimal consumption approximating problems converge to those in also the original problem. Graduate School of Business, Stanford University, Stanford, CA 94305. Huang Management, Massachusetts Institute of Technology, Cambridge, MA 02139. Zhu is with the Derivatives Research Division at CitiBank. We acknowledge financial support from the National Science Foundation under grants NSF SES-9022937 and NSF SES-9022939. "Hindy is is at the at the Sloan School of 1 1 1 INTRODUCTION 1 Introduction We analyze a numerical scheme for solving the problem of optimal life-time consumption and investment allocation studied HH&Z (1993), denoted henceforth. companion paper by Hindy, Huang, and Zhu in a HH&Z study a model that represents, things, habit forming preferences over the service flows among other from irreversible purchases of a We provide a brief statement of a special HH&Z contains the complete formulation, the durable good whose stock decays over time. case of the control problem in section 2. motivation of the study, the economic interpretations of the variables, and a detailed discussion of the numerical solution. framework that supports the We ainalysis in In this paper, we only provide the theoretical HH&Z. use dynamic programming to solve the optimal stochastic control for the contin- uous time problem formulated in HH&Z. The important feature of this problem is that the admissible controls allow for cumulative consumption processes with possible jumps and singular sample paths. Furthermore, the controls do not appear directly As a jective functional. sequel, takes the tial differential result, the associated form of a Bellman's equation, formula differential inequality involving a nonlinear equation with gradient constraint. The in the ob(7) in the second order par- differential inequality reveals the "free-boundary" nature of the solution. In particular, consumption occurs only at some boundary surface is in the domain of definition of the value function. This free-boundary the area in the state space at which both the second order differential equation and the gradient constraint are equal to zero. Furthermore, the optimal consumption amount required boundaxy. to keep the trajectory of the controlled process is the from crossing the free- This typically results in optimal consumption, and associated wealth, with singular sample paths; hence the term "singular control". In section 3.1, we state the differential inequality that the value function solves. HH&Z contains the details of the derivation which is is assumed to be twice continuously and Huang (1993) and HH&Z differentiable. show that if heuristic because the value function Under the same assumption, Hindy a function solves the differential inequality, together with boundary and regularity conditions, then such a function is indeed the INTRODUCTION J 2 value function. Furthermore, candidate policy In general, it is HH&Z provide a verification theorem to check whether a optimal. difficult to establish is one can either rely on analytical smoothness of the value function. To do so, results characterizing the regularity of the solution of Bellman's equation or produce a closed form expression regularity can be verified. Neither option is for the value function whose available in our case. In paxticulax, standard results in the theory of nonlinear partial difFerenticd equations; see, for example, Friedman (1964) and Ladyzenskaja, Solonnikov, and Ural'ceva (1968), do not provide evidence for the smoothness of the solution to our differential inequality with We will show, however, in section 3 that the diflferentiaJ so-caJled viscosity sense. Furthermore, that solution dynamic program. In section 4, We inequality has a solution in the indeed the value function of our is the details of the numerical scheme used for obtaining the HH&Z. The numerical scheme we use is based on the Markov chain approximation method introduced by Kushner (1977) and described and Dupuis (1992). We in details in Kushner approximate the original controlled continuous time processes by a sequence of controlled Markov Markov chains by the gradient constraints. will define the notion of viscosity solutions in section 3. we provide solutions reported in its chains. size of the possible We jump peirameterize the sequence of controlled in each chain which we denote by approximating Markov chains are chosen such that, as /i | 0, the mean and mean change per step, under any control, resemble more and more closely the local h. The square drift and the local variance, respectively, of the controlled continuous time process. This property is called "local consistency". We also replace the objective functional for the original problem by a sequence of appropriate anailogs for the Markov chains. Given the chosen approximating Meirkov chains and objective functional, the approach is to approximate the original continuous time problem by a sequence of discrete parameter Markov decision problems. For each decision problem corresponding value function satisfies in the sequence, the an iterative Bellman's equation. equations are easy to program <ind solve numerically. Such iterative Furthermore, the sequence of discrete Bellman's equations can be viewed as finite-difference approximations of the original Bellman equation for the continuous time case. 1 — lar, INTRODUCTION 3 Next, we turn our attention to the convergence of the numerical scheme. In particu- we focus on convergence of the value functions in the sequence of the approximating decision problems to the value function of the original continuous time problem. We use analyticaJ techniques, in particular the notion of viscosity solutions, to address this question. The argument runs, roughly, as follows. We demonstrate that the vaJue function the unique viscosity solution of Bellman's equation. This steps. We first show that the value function is is achieved a viscosity solution. comparison result between sub- and super-solutions, defined establish the uniqueness result. We in section 3 in We is two next develop a in section 3, and use it to \ then focus on the sequence of Bellman's equations for the approximating Markov chain problems and consider those as finite-difference approximations to the original monotone, We establish that this sequence of finite-difference approximations is consistent, aad stable. We will define these notions in section 5.1. Using these properties, we show Bellman's equation. that the sequence of value functions for the approximating Markov decision problems converges to a viscosity solution of the original Bellman's equation. uniqueness result established in section 3 implies that this limit the original program. we We is the value function for present the convergence analysis in section 5. also prove that the sequence of optimal investment policies in the In that section, approximating decision problems converges to the optimal investment policy in the original problem latter indeed exists. Furthermore, and abstinence in the we show The if the that the sequence of regions of consumption approximating problems converge to the corresponding ones in the original problem. It is perhaps worthwhile to point out that the analysis tion to the possibilities of jumps policies, in this and singular sample paths paper pays special atten- in the admissible consumption and the special structure of Bellman's equation imposed by the presence of the gradient constraints. In particular, the proofs of theorems ments to handle the possible jumps in 1 to 4 require involved argu- consumption and the gradient constraints. These arguments extend the existing literature on viscosity solutions as reported, for example, in Fleming and Soner (1993). in dynamic programming STATEMENT OF THE CONTROL PROBLEM 2 4 Statement of the Control Problem 2 In this section, we state the control problem which is the focus of analysis in this paper. This problem, with constant investment opportunity set and infinite horizon, case of the general problem studied by HH&Z. a special is For a statement of the general formulation, a complete discussion of the motivations for studying this problem, and for the economic interpretation of the variables, Hindy and Huang (1993) and HH&Z. consider a frictionless securities market with two long lived and continuously traded securities: at to Financial Market 2.1 We we refer the reader time where r The t. is a stock and a bond. The stock The bond is riskless, pays no dividends, and sells for B{t) = sells for e'"' at S{t) time t, the constant riskless interest rate. S{t) = S{0)+ is a geometric Brownian Motion given by:^ [ fiS{s)ds+ Jo where B is one-dimensional standard space {Q,^,P) and bond the market is fi > r cire strictly given by F= and cr > f' aS{s)dB{s) are constants. 5 a.s., (1) Brownian Motion defined on a complete probability {^t;t € [0,oo)} where ^t < Vi>0 Jo Note that the prices of both the positive with probability one. J^ with respect to which {5(5); will risky, does not pay dividend, and price process^ for the stock stock and is < i} is is The available information in the smallest sub-sigma-field of measurable. All processes to be discussed be adapted to F.^ ^ 'A process y is a mapping Y:Q x [0,oo) — 3J that is measurable with respect to <Si B{[0,oo)), the and the Borel sigma-field of [0, oo). For each u £Q, Y{u, .): [0, oo) — product sigma-field generated by ^ a sample path and for each t € [0, oo), Y{., <): Q — 3? is a random variable. ^The notation a.s. denotes statements which are true with probability one. We will use weak relations. Increasing means nondecreasing and positive means nonnegative. ^We assume that Tt is complete in the sense that it contains all the probability zero sets of J-. The process Y is said to be adapted to F if for each t 6 [0,oo), Y{t) is ^,-measurable. 3t is 2 STATEMENT OF THE CONTROL PROBLEM Admissible Controls 2.2 The 5 control variables are a consumption plan and an investment strategy. -A- consumption C plan is with C{u}, a process whose seimple paths are positive, increasing, and right- continuous, t) denoting the cumulative consumption from time investment strategy is C is t before consumption and trading. said to be financed by the investment strategy = W{0)+ W{t) / in state An u;. one-dimensional process A, where A{t) denotes the proportion of wealth invested in the stock at time plan to time I {rW{s) + A{s){ti-r)W{s)) ds-C{t-) A + ^0 if, A consumption a.s., f aA{s)W{s)dB{3), V< > 0, Jo (2) where W{t) is wecdth at time t left-continuous sample paths and that H^(f'") A admissible of t. C consumption plan if (2) is = W{i) — and the investment strategy well-defined^ and, for Denote by C and Note that the wealth process has before consumption. all t, E[C{t)] < AC(<). A that finances Q{t), with Q{t) it axe said to be some polynomial A the spjice of admissible consumption plans and trading strategies, respectively. 2.3 The Objective Functional objective functional is over the stock of durable, defined as an integraJ, over time, of a utility function defined z, and the standard of living, y, which are given, respectively, by z{t) y{t) = z{0-)e-'^' = y(0-)e-^' f\-'^^'-'UC{s) + 13 + A /' e-^('-')ffC(5) a.s., (3) a.s., (4) Jq- no- where 2(0") > and the integrjJs *For this we and y(0") > in (3) mean both and are constants, (4) are defined /? cind A, with path by path > X, are weighting factors, in the Lebesgue-Stieltjes sense. the Lebesgue integral and the ltd integral are well-defined. When A(t) feedback control depending on the state variables and C{t) depends on the history of {W,S), we there exists a solution is a mean W to the stochastic differential equation (2) for every pair of feedback controls. 3 ANALYSIS OF THE VALUE FUNCTION Wealth managed is dynaimicaJly to solve the following program: C s.t. is financed by and W{t) The Wq is initial function u • u is is wealth, «(y, z) a fixed The second A e with W{0) - AC{t) > a discount is satisfies < K{y + V< fcictor > = Wq, (5) 0, and u(-,-) is the felicity function. the linear growth condition: Vj/ z)-, > 0, z > for 0. some constant /T > 0. (6) •) is a strictly increasing, concave, and differentiable function. z, u(-,z) is a strictly increasing, concave, and differentiable function. • For a fixed y, u{y, • For > <5 A cissumed to satisfy the following conditions: nonnegative and < - •• E[l^e-''u{y{t),z{t))dt] supc€C where 6 constraint in (5) is the requirement that wealth after consumption at any time must be positive. Condition (6) guarantees that the supremum in (5) is finite. Since the interest rate and the parameters of the return on the stock are constant, the supremum call the supremum J(H^, y,z). The in (5) We is a function of the state variables in (5) the value function of the observe that, in general, little is W , y, and z only. As usual, dynamic program and denote known about it we by the differentiability of J. choice of a technique to analyze the dynamic program in (5) is influenced by prior knowledge of the smoothness of the value function. 3 3.1 Analysis of Bellman's Equation In this section, (5). The Value Function Although we report Bellman's equation associated with the little is known about the dynamic program regularity of the value function J, Bellman's equation heuristicaJly by assuming that J is we can in derive twice continuously differentiable. 3 ANALYSIS OF THE VALUE FUNCTION This assumption allows us to utilize Ito's 7 lemma, together with the dynamic programming principle, to derive Bellmaui's equation. Bellman's equation turns out to be a difFerentiaJ inequality involving a nonlinear second order partial differential equation with gradient constraints. In section 3, of Bellman's equation we prove rigorously that the value function we derive heuristically here. solution in the viscosity sense The — a notion that we define Assuming that the value function tions for optimality are given is smooth, by the following max|max^[u(y,2) + JwW[r + A{fi M+ - indeed a solution value function, however, HH&Z differential inequjJity in (7) a show that the necessary condi- differential inequality: + ^JwwW'^A'^a - Jw} = 0, /3zJ^ - can also be viewed, - XyJy SJ] , (7) plus the condition that nontrivial consumption occurs only at states where Jw- The is in section 3. r)] AJv is in the region of /3 J, -}- XJy = no consumption, as a nonlinear second-order partial differential equation with the gradient constraint that 13 J, + XJy -Jw<0. In addition to the above necessary conditions, it is clear that J must satisfy the following boundary condition: ^(y,z) The tion 2 is = \unJ{W,y,z) = j^ e-"u{ye-'\ze-^')ds differential inequzJity (7) suggests that the (8) . dynamic program formulated in sec- a free-boundciry singulcir control problem. Consumption should occur only when the state variables {W^y^z) reach a certain boundary surface characterized by the condition that Jw — l3Jz + XJy. Furthermore, the optimal consumption is the amount required to keep the trajectory of the controlled processes from crossing this critical boundary. Since the returns on the stock follow a Brownian Motion, such a consumption policy, and the associated wealth process, may have singular sample paths "singular control". Computing the location of the free-boundary — is hence the term the focus of this paper. Hindy and Huang (1993) and HH&Z show that if there is a smooth solution to (7) that satisfies (8) and additional mild regularity conditions, then that solution is the ANALYSIS OF THE VALUE FUNCTION 3 value function for the program 8 These papers also provide verification theorems (5). that specify conditions for a candidate feasible consumption and investment policy to attain the supremum in (5). We refer the reader to HH&Z statements of (§3) for the the necessary and sufficient conditions of optimality, the verification theorems, and the proofs. Viscosity Solutions of tion 3.2 The verification function is theorem presented in know whether useful is is when the ca^e Hindy and Huang it is known when there (1993). that the value a closed form is In general, it is difficult the value function has enough regularity for the verification theorem to To handle the apply. HH&Z twice continuously differentiable. Such solution of Bellman's equation as in to The Dynamic Programming Equa- the value function, which there situations in we utilize the is no information about the regularity of notion of weak, or viscosity, solutions. Viscosity solutions were introduced by Crandall and Lions (1982) for first-order equations, and by Lions (1983) theory, We we for second-order equations. refer the reader to the User's We Guide by Crandall, Ishii, and Lions (1991). are interested in viscosity solutions in the context of proving convergence of the Markov chain numerical scheme we For a general overview of the to be presented To in the following sections. that end, characterize the value function as the unique viscosity solution of Bellman's equation. prove uniqueness of the solution to Bellman's equation by utilizing a comparison principle between viscosity subsolutions and viscosity supersolutions. result has additional significance to our subsequent analysis because in establishing Let Q = This comparison it is a key element convergence of the numerical scheme. (0, oo) X [0, oo) x [0, oo) and consider a second-order partial differential equation of the form: ma.x[-6v where 7 + F{d^v, Dv) + = — l,A,/3), D = ( u{y, z), 7 • D^ v) = {dw,dy,dz), $ is 0, on Q with u(0, y, z) some function of y and z, = $(y, 2) , (9) and ^ denotes ANALYSIS OF THE VALUE FUNCTION 3 transpose. F{dwv, dwv, Next, we F The function given by: is = max{- VK^A^a^5^u + W[r + dyV, d^v) - A{fi r)]dwv] state the notion of viscosity solutions as in Crandall, Ishii, - \ydyV -^zd^v. and Lions (1991). An upper semicontinuous function v is a (viscosity) subsolution of (9) on the domain Q if v{0,y,z) < $(j/,z) and 4> G C^'^'^{Q) and any local maximum point (VKo,yo, 2o) ^ Q of v ~ Definition 1 (Viscosity Solution) (i) : Q —^ R for every (j), max|-(5v(iyo, I/o, ^o) + F{d1y<i>{Wo, yo, zo), 4>w{Woyyo-, 2o), +u(yo, zo), 7 D^<i>{Wo, • yo, zo)} <?!'y(VKo, j/o, > ^o), <i>z{Wo, yo, ^o)) . (10) (ii) A the domain lower semicontinuous function v Q if v{0,y, z) point (Wo,yo,zo) e max|-(^u(V^o, Q yo, zo) > : Q —* R ^{y,z) and for every is (f> a (viscosity) supersolution of (9) on 6 C^'^'^{Q) and any local minim.um ofv-4>, + F{d^(^{Wo, yo, zo), <f>w{WQ, yo, zq), <f>y{Wo, yo, zo), <f>z{Wo, yo, Zo)) +u(yo,Zo),7--0''<?i(Wo,yo,Zo)} <0. (11) A continuous function v and : Q —^ R is a viscosity solution of (9) if it is both a subsolution a supersolution of (9). The main result of this section Bellman's equation (9) is that J(W,y,z) is the unique viscosity solution of with boundary condition u(0,y,z) = ^(y,2) as in (8). First, we state without proof the following proposition Proposition programming 1 satisfies the dynamic E[e-'^J{WiT),y{T),z{T))+ re-''u{y{t),z{t))dt], (12) (Optimality Principle) The value function J principle JiW,y,z)= sup for any Ft-adapted stopping time r the stopped processes ofW,y, and > 0, z at r. where W{t), y{T), and z{t) are, respectively, ANALYSIS OF THE VALUE FUNCTION 3 The dynamic programming example, for in 10 principle in proposition Fleming and Soner (1993, §5.2). from the standard proofs because controls allow for jumps and singular sample paths. Zhu may utilize the principle of The proof of the possible Theorem 1 concave (0, in all of its J(H^,0,0). We to prove that the value function of this result J is is is a not standard because continuous on the domain WG all (0, oo), of continuous in the region need to show that J the doamin Q. Since u of its elements, J(Vl^, 0,0). = is increasing, is and hence limy_o,z-.o J{^i Yimy^Q,,^oJ{W,y,z) > is We then (1975, §14, page 82). 0, 0), fact, then follows that J limy—o,z-.o «^(^) J/,^) V-, 2) > J(W^,0,0). From the dynamic programming we have J{W, 0, 0) > where (M^«,yj,2j) are the controls. we show how one a viscosity solution of Bellman's equation (9). It increasing in prove that, in Suppose that principle, is Next, II). and using standard reasoning, we can show that J Holmes continuous on the boundary (IV, the value function J is arguments. oo) X (0, oo) X (0, oo); see however, proof of the dynamic programming establish that the vcJue function Utilizing the concavity of u, Q. is first is, singular sample paths in the admissible controls. The value function J We Proof. jumps and the 1 our formulation the admissible in A what can be found, of proposition (1991, Chapter dynamic programming viscosity solution of Bellman's equation. similar to The proof slightly different principle in such a case can be found in 1 is i; [e-''J{Wu yu ~zt) + e-''u{y„ z,) ds] f^ trajectories of a continuous state process for Taking the limit as < J, 0, and (13) , some admissible utilizing the continuity of the trajectory of the controlled process, the right-hand side of (13) converges to a value strictly greater than J{W,0,0)- a We (i) J is clear contradiction. next show that (^ is a viscosity solution of Bellman's equation. a viscosity subsolution. subsolution, then the Let J be a We argue by contradiction and show that dynamic programming principle stated C^'*'^((5) function such that J— <l> has a local if in proposition maximum at J fails 1 is {Wq, to be a violated. yo, zq) 6 Q. . . U ANALYSIS OF THE VALUE FUNCTION 3 That is, 3 a neighborhood A/q of {Wo,yo,zo) such that - HWo.yo^zo) > J{W,y,z) - J{Wo,yo,zo) Without that <j> loss of generality, we assume + '^{yi^) < at = <f){Wo,yo,Zo), By F{<f>ww,4>w,4>y,4>z) {Wo,yoiZo). that J{Wo,yo,ZQ) there exists 7 neighborhood A/ of (Wo, 6<f>{W, y, z) (14) = <^(^o? J'o, ^o)- Now suppose — <f>w < and —SJ + \<f>y + ^(f>z continuity, > and using the assumption such that, for {W,y,2) in some J/0,20), X<f>y{W,y,z) - Afo- that J(VFo,yo,2o) In other words, both violates inequality (10). y{W,y,z) e 4>{W,y,z), + l3<f>,{W,y,z)-4>w{W,y,z)<0 and (15) + F{dU{^, y. ^), 4>w{W, y, z), <f>,{W, y, z), M^, y, z)) + u{y, z) < -7 (16) By we the existence result in Zhu (1991, Theorems IV. 2. 2 and IV. 2. 3), optimal policy {A',C') at (Wo,yo, does not prescribe an initial jump We now ^o)-^ consider the establish that the optimal policy at {Wo,yo,Zo). If not, then C prescribes C some jump Using the dynamic programming principle, we can conclude that T]o. < -^(Vl^o-'?,yo the property of the local maximum in ^(Vl^o,yo,2o) By + A7?,Zo + VO /?r/), < r/ < 770 (14), the previous inequality also holds for ^, and hence < where by T] < and t; < 7?o is letting 77 J. C) - r/, yo + At?, zq sufficiently small so that 0, < ^From the (i>{Wo - (Wo — t?, 4>{^o, Vo, Zo) yo + A77, zo + 0t]) € A/q. Dividing we obtain X4>y{Wo, yo, Zo) + ^4>ziWo, yo, zo) - Zhu (1991, Chapter analysis in /9r?) + <i>w{Wo, yo, zq) IV), the optimal policy {A' ,C') is , known to exist. However, a space of processes larger than the space of admissible controls {A, C). The crucial feature of the analysis is that the expected life time utility associated with {A',C'), starting from any point [W, y, z), is equal to the viJue function J evaluated at that point. For more details, we {A' , may be realized refer the reader to Zhu in (1991). . 3 ANALYSIS OF THE VALUE FUNCTION which We a contradiction to (15). is have now established that {W*,y*,Zf) to 12 is (15) holds, then the optimally controlled process if continuous in a neighborhood of (WoiJ/Oi^o). We be a subsolution, then the dynamic programming principle We define the optional time f by the established policy. Then, it fact that = {W^,y',z^) follows from (15) > inf{i and : (VF,',^^*, 2j') next show that is K. Now take r to be the ^ Ao}. Clearly, f > 0, a.s., continuous up to r under the optimal is a.s. K ini{t minimum fails violated. 4>z) + u](W:, yl z;)dt] +E[j\-'\\<i>, + ^<t>,-<i>w]{W:,ylz;)dC^] = J (16) that E ll^ e-*'[-<5<^ + F{dU, 4>w, Define the optional time f if > of f : Jq and < > K}, \AlW^<7\^ds Applying f. (17) -7E[f]. for some constant Ito's rule to the process e-*V(M^;,yr,2;),wehave <i>{^o,yo,zo) = E[e-''<f>iw:,y;,z;)] -E\I^ e-''[-S4> + FidU. 4>w^ -E [J^^ Note that the expectation of the integral is the local e-^m, + 134>, - maximum combine the is <f>.)]iW;, y;, z:)dt] boimded property in (14), we can replace A'W'a dBt till (^ vanishes because that the stopping time r. Again by by J in first line of (18) result with (17) to conclude that J{Wo,yo,zo) > E[e-'W{W:,y;,z;)] -E[j\-''[\<i>y + P4>, > E[J\-''n{ylz:)dt + (18) cf>w]{w;,y:, z:)dc:] stocheistic integral Jq a martingale as the integrand <f>y, - 4>w]{w:,y;,z;)dc:] e-'^JiW:,y;,z:)] + T^H , and then ANALYSIS OF THE VALUE FUNCTION 3 which clearly violates the optimality ming of 13 (A',C') dynamic prograxn- ajid the principle of (12). J (a) smooth 4) is a viscosity supersolution. such that J(Wo,yo,2o) - By we must show, definition, < J{W,y,z) - <?^(W^o, J/o,2o) every C^'^'^- for 4>{W,y,z) for {W,y,z) G A/q, that the following holds < [-8J + F{4>wwAwAyAz)^-u\{WQ,yQ,zo) A<?^„(H^o,yo,2o) First, we choose interval [0,t), W°, y°, is = zero consumption C, where r = inf{< < /3<^z(V^o,yo,2o)-<^w(V^o,yo,2o) > and some investment policy /q |i4,W,a|^<i3 : > K} for the principle of dynamic progrctmming, (20) A on p and < t, Let 0. A on where p an optional time, Replacing J by we then apply <^, stochastic integral is eIJ' inequality (19). e-^'[-6<^ To derive J{Wo, Since 7; J — > <i> is Ito's rule, . together with the observation that the a maxtingaJe, to obtain from the above inequality that Dividing by p and letting p yo, ^o) has a local +F+ <0 u]{W^,ylz'',)ds\ I 0, using the continuity of (20), the < the integrand, r . we J, 0, r/, yo + Ary, zq + ^77), VO < r/ < first Wq. at (Wo,yo, 20)1 this inequality also holds for sufficiently small such that we obtain the second derive the dynamic programming principle implies that > J{Wo - minimum VO < p {Wq — T],yo + \t],zq + 0t]) € Ao- subtracting the right-hand side from the left-hand side and dividing by 77 > ii' = all J{Wo,yo,zo)>E[e-"'J{W^,y''^,z°)+re-"u{ylz',)ds if the stochastic some constant follows that, for it (19) 0. and 2° be the state processes corresponding to the controls C, By [0,r). + and inequality (20). This completes the proof. t] I <i> as well Therefore, and sending 3 ANALYSIS OF THE VALUE FUNCTION Comparison of Subsolutions and Supersolutions 3.3 We 14 conclude this section by providing a theorem that compares the subsolutions of Bell- maji's equation to its supersolutions. standard results example, as, for The proof of this result Fleming and Soner (1993, in The comparison result viscosity solution of the result is used later is §5.8). We utilize additional Jw — ^Jy — ^Jz > arguments to handle the gradient constraint that tion. not a simple extension of is in Bellman's equa- then used to argue that the value function dynamic programming equation. is the unique In addition, the comparison convergence of the Markov chain approximations to be in the proof of presented in a subsequent section. Theorem 2 (Comparison Principle) a supersolution of (9) on the domain Q Let V and V be, respectively, a suhsolution in the sense of Definition 1, then V <V and on the domain Q. = {W,y,z) and x = {W,y,z). We rewrite Bellman's equation as ma.x{—6v + F{^^^rV,Dxv) + u{y,z),j Djv} = 0, where 7 = — 1, A,/3) and Dj is the transpose of {dw,dy,dz). We construct a function U{W,y,z) = Proof. Let us denote, for simplicity, x • ( /o°° e~*'u(j/e~^', 2~^') dt — mo for some constant mo > -SU + F{d^U, DM) + «(l/, 2) > and 0. One can \Uy + l3U,-Uw>0, where the second inequality follows from the assumption that u both y easily verify that is strictly increasing in eind z. Next, we will prove the comparison between which the proof of the theorem is V and V^ = 6V + {1 — 0)U, for all 9, from immediate. For this purpose, we choose the auxiliary function rl;{x, where a > and x) < c = < ||x 1. x\' + Assume e{\x\' + [x^ V(x, x) e Q x Q , to the contrary that: V\x') > V(i'), for some x' G Q. (21) 3 ANALYSIS OF THE VALUE FUNCTION By the linear growth of attain of its a and (1991, maximum at e. We V and V, Appendix A: proof — V{x) — easy to check that V*(x) Q^Q- Note a point (io,io) € xp{x,x) will that (lo, xq) depends on the values dependence in the notation. One can prove, as will suppress this + \xo\) -* We will show it is 15 of comparison principle), that a|io — io|^ —> as a in Zhu | oo, and as c i 0. e(|xo| From the that supposition (21) produces the following contradiction. we definition of (xo,Xo), have, for small e, = V\xo)-V{xo)-xl;{xo,Xo) V\x) - V{x) - max 4>{x,x) (z,x)€QxQ > V\x')-V{x')-2e\x'\^>0. However, we will show next that the respectively, implies that V*(io) — V'(xo) — V V* and fact that are sub- and supersolutions, as e V'(io,Xo) i Jind J, a which t oo, is a clear contradiction. Applying Definition 1 to the subsolution V^ and the super-solution V, we conclude that rmix{-SV^ + F{d^rp,Dr7P) + u{-),^- Djrp} + F(-5^V,-^xV') + max{-<5V' Now, since Dxxk = 2ex + a\x — x\ «(•), = and D^^ -7- -010} 2ex — a\x — > 0, at xq < 0,, at xo x|, it follows (22) (23) . from inequality (23) that -7 Moreover, we claim that holds. Subtracting it contradiction to the Djxl^ = -7 in (22), 7- [2cxo from fcict • (24), • [2exo - + q(x — xq)] < we conclude that 2e7 that e(|xo| + - a(xo |xo|) i • xq)] 0. If (xq < 0. (24) not, the opposite inequality + xq) > as e | 0. Therefore, 0, Va > , which we must have and (23) that -6V\xo) + F(5^V, OrV') + u(l/o, zo) > 0, at xq -6V{x) + F{-dl,ik,-Dsi^) + u{yo,zo) < 0, at xq and is a in (22) , 3 ANALYSIS OF THE VALUE FUNCTION 16 Subtracting one from the other yields the inequality 6[V\xo) - V{xo)] < F(5^V, D^rP) - F(-a^V, -D,^) + Meanwhile, we calculate (Note: F{d^w^, = D.'iP) r = max{^V^o'^'<7' - yo) + ^oAti[a{Wo - Wo) + + where the maximizers can be explicitly j/o) + computed as T. _ , ccWo < ^cc{Wo -a{X\yo and the proof Corollary 1 zq)\ c X and a °" -q(Wo - Wo) + 2tWo \Ji aWo ^2 we can verify that zoP) + Wo)' - 2e'-^{W^ + W^) [u(yo, zo) - "(yo, ^o)] This completes the demonstration of a contra- | 00. is Zo)\ - 2ef^iWo - - yoT + /?ko - of the theorem. The value function J Wo)' - ^zo{zo Substituting into (25) and rearranging the terms, diction 2eWo]} , - - Wo) + ItWo ^°-~<72 which goes to zero as - ^zq{zq I -a[Ayo(yo S{V\xo) - V{xo)\ (25) "(yo, ^0). = max{-^V^o'^'^' - WoAfi[-a{Wo - Wo) + 2eWo]} A F{~dlxP, -D,rP) ol{Wo - for simplicity) -Q:['^yo(yo \i u{yo, 20) I the unique viscosity solution of Bellman's equation (9). Proof. V Suppose that there is another viscosity solution V. and J are sub- and supersolutions, because J and V respectively. On We V< then have the other hand, are also sub- and supersolutions, respectively. Hence, J because we have J < J = V. I V , 4 NUMERICAL APPROACH 17 Numerical Approach 4 In this section utility we describe the numerical approach utilized to solve the infinite- horizon We maximization problem. replace the continuous time processes {W,y,z) by a sequence of approximating discrete parameter Markov chains. The important feature of the approximating sequence mean square change that, as the step size converges to zero, the is mean and per step, under any control, converge to the local drift and the local variance, respectively, of the controlled continuous called "local consistency". The original optimization a sequence of discrete parameter control problems. each Markov chain control problem discrete Bellman's equation. satisfies time process. This property problem We is is then approximated by show that the value function of an iterative, and hence easily programmable, The Bellman's equation of the discrete control problem is a natural finite-difference approximation to the continuous time Bellman's "nonlinear" second order partial differential equation with gradient constraint. 4.1 A Constrained Problem For numerical implementation, variables to a finite region we need Qm = to restrict the domain [0,Mw] x [0,My] x of definition of the state [0,71/^]. For this purpose, introduce the reflection processes Lt^Rt and Dt at the boundaries of the region. we We hence consider the following modified state variables: where Lt, Rt {W,z,y) We dyt = —\ytdt dzt = -^Zidt dWt = Wt{r + XdCt — dLt , + ^dCt - dRt + Atifi - r))dt , and dCt + WtAtadBt - dDt and Dt are nondecreasing and increase only when the gtate processes hit the cutoff boundaries W = Mw, ^ = M^, and y = My, recall that the stock price process follows the riskless interest rate is constant. We will, resp-^ tively. a geometric Brownian I ^tion and that hence, apply the Markov.^ hain approach NUMERICAL APPROACH 4 18 to solve numerically, on the restricted domain, the following infinite horizon program: /•oo = j'^{W,y,z) . E sup A,CGAC e-''u{yt,zt)dt (26) Jo subject to the dynamics of z and y and the dynamic budget constraint, where {W, y, z) 6 Qm is the initial state. The Bellman's equation max{-<5J^ + max{£^ J^} + Mw] X [0, the operator £^ on [0, addition, we My] x is [0, u{y, 2), given by Ajf + £^ = ^W^A^a^-^ + W[r + — boundary conditions was suggested - J^] = (27) 0, A{fji -r)]^- \y§^ - ^zj-^. In in W = 0, = Mw, — and y = My, = and 0, z = Kushner and Dupuis (1992, M^. This type of §5.7) and employed numerical analysis of the one-dimensional financial model studied by Fitzpatrick and Fleming (1992). These boundary constraints reflect the the value function on the original unbounded domain. in the theoretical analysis of It is asymptotic behavior of worthwhile to remark that convergence of the numerical scheme, the value functions Qm increases to Q, at the boundaries. The boundary converge pointwise to J as the sequence of restricted domains regardless of the specification of the behavior of conditions we J^ chose, however, affect the quality of the actual numerical solution. The Mcirkov Chain Scheme 4.2 We (iJ^ specify the boundary constraints respectively at the cutoff boundaries J^ dynamic program takes the form: Mz], together with the appropriate boundary conditions, where __.0, in the for this introdu-e a sequence of grids Gh = {{k. ):W = kxh,y = ixh,z=j xh;0<k< Nw,0<i < = My/h, N^ = M^/h, and iV„,0 is integers. Each grid point {k,i,j) £ Gh corresponds to a state {W,y,z) with step size, and where A^y N,} Nw = Mw/h where h i.ie <j < , are W=kx h, NUMERICAL APPROACH 4 y = i X h, = and z For a fixed grid, j x h. For simplicity, we introduce ACn = A where an is artificial We 19 {{A, taJce the saime step size in all three directions. the spa<;e of discrete strategies Cy,A = the step size of control, bound we lxA,AC = OoTh;0<l<NA} A'^^ = A/ A to be relaxed in the limit eis is A total number of control steps with A and A^^ x J, A f oo. approximate the continuous time process {W,z,y) by a sequence of discrete pa- rameter Markov chains {(H^,J,y^, z^);n = 1,2, •••} with h denoting the granularity of the grid and n denoting the steps of the Markov chain. For every h, the chain has the property that, at each step n, there is a choice between investment and consumption. At any time, a chain can either make an instantaneous jump, (AC = walk", is of the domain. Fix a chain and 1. The reflected to the interior in a Markov chain its investment policy I — (AC = move from + {k A= a "random h. its dynamics The in the interior transition probabilities are specified as follows: possibly neighboring states: consistent with corresponding grid size case of no consumption The chain can manner h), or follow At the cutoff boundaries the the neighboring states on the grid. 0), to chain for this (AC = y. l,i,j), {k 0): the current state {k,i,j) only to one of the four — l,t,j), {k,i — l,j), A, the transition probabilities and {k,i,j — 1). For any in this case are defined as P/1MVV2 + kh^r + fiA) Q'{k,i,j) fc^/tMV/2 + fc/t^(M) Q'{k,i,j) P^[{Ki,Jlik,i-l,j)] Pk[{k,iJ),{k,iJ)] Q'{k,i,jy = l-Pi^[ik,iJ),{k+hiJ)]-P;^[{k,iJ),ik-hhJ)] -P;^[{k,i,j),ik,i-lJ)]-Pl^[{k,ijy{k,iJ-l)] NUMERICAL APPROACH 4 where the normalization Q''{k,iJ) The = ^max 20 fcLctor (5''(^,t,i) is {Ph''{l X A)V2 + recipe for these transition probabiHties given by kh^[r + I x A{(x + r)] {\i +. ^j)h''} . a slightly modified version of those sug- is Note that the normalization factor gested by Kushner (1977). + independent of the is investment policy A. This specification implies that, at each step, the Markov-chain has a strictly positive probability of remaining at the same node. Furthermore, we define the incremental difference Ay^ and AWl^ z^J, = W„\i - The quantity At^{k,i,j) W^^. = is j/^^^ — y^, Az^ = z^^j — equal to h?IQ^[k,i,i) and interpreted, following Kusnher and Dupuis (1992), as the "interpolation time" for the Markov varies chain. Note that the interpolation time At^[k,i,j) from state to state. a random variable that is At step n of the chain, with the previously defined time scale A<^, one can verify that = -XyAt'^^-ALl = -^zAt^^-ARl £„^[Az„^] K\^K] = ^[r- + Mf^ - r)Wu - A^n, E^lAWl: E^lAW!:]]'' = H^MVAt;; + O(Aii), ^n'[Ay^] where E^ denotes expectation conditional on the nth-time state the reflecting processes y^, z^, and Wj^ reach moments We 2. call this The The {k of the — AL^, A/2^, and AZ)^ equal and where to the positive value h only when and second of the continuous process {Wt,y,,Zg). property "local" consistency of the Markov chain. case of consumption — (AC = chain jumps along the direction \,i ^ (M^,J,y^, z^), their respective boundaries. This implies that the first Markov chain approximate those °^ and + X,j + ^). However, the except in the trivial case A the direction vector { = 1 — \,\,^) between three grid points = ( h): — 1, A, ^) from the current state later state, {k — l,i + \,j + 0), ^. For ease of is {k, i,j) to the state usually not on the grid programming, we take the intersection of with the corresponding surface boundary and randomize axijacent to the intersection point. way that the expected random increment will We randomize be along the direction ( in such a — l,A,/3) and length equal to the distance from the starting state to the point of intersection. of NUMERICAL APPROACH 4 - Without 21 we assume loss of generality, hereafter that A + and {k — 1), + 1,1 l,j + The 1). transition probabilities can P^[{k,i,j),{k,iJ + + l)] = + 1J + 1)] = Pk[{k,i,j),ik-'^,ij we can In this case, also, E^fAy^:] where AC = 4 + 1), be defined as: E^[Az':i] = 0AC, -^, and ^, and the increment in consumption. 4.3 A AC E„^[AW^„^] The = -AC (29) quantities in the right-hand time process units of consumption. The Markov— Chain Decision Problem {A,C) policy is admissible conditional on where — ' side of the equalities in (29) are the respective changes in the continuous corresponding to {k verify the property of "local" consistency: = AAC, is 1,1 ^. In this case, the 0-1 = l)] /? P^[{k,iJ),{k- < 1 by the three grid points {k,i,j intersection occurs inside a triangle spanned 1,1, j < W problem = kh and preserves the Mairkov property in that {)Zvy,T,{~t<^'' W= for the discrete if it k'h, y = ih and y' = parameter Markov chain J''(^,i,j) = rnax ^t. I I i'h, z is = A'^K*.', j), jh and z' (*'. = i', j'h. j')l, The if AC = control then to solve the program: £ ^""""(yn, ^n)Ai^ (30) 71=0 where tjj = Xro</<n ^'l*- Note, (30) is analogous to (26) in the sense that the former approximates the expected integral in the latter. sum in the k , 4 NUMERICAL APPROACH The dynamic progreimming equation now takes the discrete = J'{k,i,j) 22 iterative form m^x{Zk'„',:'P^[{k,iJ)Ak\i\r)]J'{k',i',j'), maxo<KN.{e-*^''•('^•'•^•)Ev..^;'^/:*[(t,t,;),(fc^^^;0]^'(fc^^^/^ (31) +U{iJ)At'{k,i,j)} G Gn, with for {k,i,j) that Ph[*, *] D-J{k^i,j) and P^[*, *] are the transition probabilities of the chain. = nk,r,j)-nk,i-l,j) h = . Using this notation, ~ = Jjk J*" C^ = l) Furthermore, one of these two if it is l,i,j) D^ J{k,t,j) = - + 2J{k,i,j) - l,ij) . U{i,j) l) < 0, < 0, (32) where (33) W{r + Ati)Dt - WArD; - XyD- - ^zDj. inequcilities must hold as an equality at each {k, i,j) € Gh- optimal not to consume at {k,i,j) then (32) holds as an equality. I 0. Clearly, [1 - e-*^'''(*''-')]/Af''(Jt,z', j) This leads to a pair of discrete differential which gives a finite-difference approximation of Bellman's equation (31) the constrained maximization problem on the domain 5, J{k + m&x{Ctrik,iJ)] + approximates the discount factor 6 as h section and , discrete Bellman's equation (31) in the form: Otherwise, (33) holds as an equality at {k,ij). inequalities us denote h + l3Dp\k,iJ)-D;j''{k,iJ + Ih/MVD^ + let = Ak,iJ + l)-J{k,z,j)^ -^ At^{k,i,j) L + we can express the XDtJ''{k-l,i,j + In particular, D-Jik,i,j) ^ ^ , n2 jf, L>k J\k,i,j) Now, recall Ak,t,j)-Hk,i,J-l)^ D-Jik^iJ) = ^ Ak,i+hj)-J{k,i,j) DfJiKrJ) = DjJ{k,i,j) where we iterative reflection® at the artificial boundaries, we [0, Mw] x [0,My] x [0,Mi]. for In outline a proof that the above described sequence of finite-difference ap- proximations, together with the boundary specifications, converges to the solution of the continuous time Bellman's equation. ^This specification will be detailed in the end of this section. 4 NUMERICAL APPROACH Boundary 4.4 23 Specification of - For numerical implementation, we attach The Value Function a "fictitious" surface to each cutoff boundary We denote by G^ the grid including this attached exterior surfaces. The three surfaces {W = 0,z = 0, andy = 0) are natural boundaries whereas the three surfaces {W = Mw + h,z = Mz + h, and y = My + h) are artificial fictitious boundaries. The value function on the surrounding the cubic grid Gh in which the Markov chain artificial conditions we turn boundary = W= 0, the boundajy condition surfax:e sumption because of y = infinite 0, it is clecir marginal value function at this boundary = ^J'{k,0,j + • Similarly, at the boundary 2 diate consumption J'{k,i,0) Next consider the • is is given by the natural constraint is 1) = utility for y on that boundary. Hence, the given by: + 0, that the optimal policy calls for con- ^A^ - 1,0,; + 1) + jJ'ik - 1,1,; because of infinite marginal utility for z, + 1). imme- optimal. Hence, the value function satisfies: = i^jf^{k,i,l) + artificial cutoff At the top-level boundary ^J'{k - l,i, 1) + ^J'{k -l,i + result, the iterative the natural constraint l, 1). boundaries. W = M\y + h, that the agent will optimally choose to As a These are the ^iih,jh). At the boundary J'{k,0,j) surfaces. to next. At the boundary J'{0,ij) • defined. boundaries will be specified to capture the optimal policy in a manner locally consistent with that of the continuous processes at the • is scheme XDf J'' + we take it that the wealth level consume regardless ~ DtJ*" = 0. so high of the levels of z or y. of the value function at that I^Dfj'' is boundary is given by 5 CONVERGENCE OF THE NUMERICAL SCHEME • = My + At the y-side boundaxy y - /3D/ J'' • At the z-side boundary z 5 we let DfJ^ = 0, if = 0, if DtJ'' = M^ + XDfJ'' h, /i, we 24 AC = AC = and h. let Dp'' = 0, if - DtJ^ = 0, if AC = AC = and h. Convergence of the Numerical Scheme we analyze the convergence In this section, of the value function approximating Markov chain problems to those trols in the We time problem. scheme described first and the optimal con- in the original continuous analyze the properties of the finite-difference approximation in section 4. In particular, we show that the discrete Markov chain approximation approach produces a monotone, stable, and consistent finite-difference approximation to the Bellman's equation that characterizes the value function. Armed with these results, we proceed domain Qm and the size of the cubic discrete Markov chain problems converges to prove that, as the mesh size goes to zero increases, the sequence of value functions in the to the value function of the original problem uniformly on any compact subset of the domain Q. Furthermore, convergence obtains gardless of the boundary conditions The proof domains Qmfunction J is case with a equation. is based on the artificial boundaries of the restricted we already established, that the value on the fact, the unique viscosity solution of Bellman's equation (27). smooth value function follows directly situation is specified (see, for instance, from consistency and different because We, hence, we do utilize the re- Strikwerda (1989)), convergence stability of the finite-difference not, in general, In the classical have a scheme. Here, the classical solution of Bellman's comparison principle for subsolutions and supersolutions of Bellman's nonlinear partial differential equation in order to gu£u-antee convergence of the approximations to the value function of the continuous-time problem. CONVERGENCE OF THE NUMERICAL SCHEME 5 25 Properties of the Finite— Difference Scheme 5.1 Let B{Gh) denote the spa<:e of real-valued functions on the extended lattice Gh- To study the discrete Bellman's equation (31), we introduce the mappings T^: B{Gh)-^B{Gh) = {T;^V)ik,i,j) and e-'^'"^"''^^ Sh : B{Gk) ^ B{Gk) where VA Y: P,[{k,iJ),{k\i'J')]Vik',t'J'), k',i',j' = {S,V){k,iJ) Yl PK[{k,i,j)Ak',i\j')]V{k',i',j'), k',i',j' for any V G B{Gh)- In the case Obviously A/, II lloo > 0. of no consumption, Define the operator norm • || || as we set Ak = HrH = sup^gg m\nk,i,j At^{k,i,j). ||Tu||oo/||u||oo5 with being the L°°-norm of the bounded function space B. The discrete Bellman's equation can be rewritten as follows: J\k,ij) = max{max[T;f J''(fc,t-,j)] + u{iJ)At\k,iJ),SHJHk,iJ)}, V(^,t, ;) € Gh. (34) It is clear that the solution of the discrete Bellman's equation operator equation (34). The is a fixed point of the properties of the operators T/^ and 5a, which are useful for proving convergence, are recorded in the following proposition. Proposition 2 7^* '5 are strict contraction operators with a contraction operator with \\Sh\\ < 1. \\Tf^\\ < e"*^*" < 1 and Sh is Thus, for every h, there exists a function J^ that solves (34)Proof. We estimate directly \T,^V{k,i,j)-T;^V'ik,i,j)\ < Y: PH[ik,tJ)Ak\i'J')]\V{k\i\j')-V'{k',t\j')\ e-^^'''('=-'"-^) k'.i'j' < max e-*^" \V{k,iJ)-V'{k,i,j)\ and (*,i,»eGh \ShV{k,iJ)-SHV'ik,iJ)\ < Y Ph[{k,iJ),{k',i\j')]\V{k',i'J')-V{k',i'J')\ k',i',j' < max (*,t,j)eGh \V{k,iJ)-V'ikJ,j)\ , 5 CONVERGENCE OF THE NUMERICAL SCHEME 26 which implies the property of contraction. Given that both Tf^ and Sh are both contractions, the second assertion follows. I Next, we discuss the properties of the finite-difference scheme, (32) and (33), associ- On ated with the Markov chain approximation. finite-difference operator at all ^hV ^ when v < equation differential We v. will that v v'. The scheme ChV = / Cv = f if, pointwise at each grid point, as for ciny /i J. > such that there exists compact domain T) and some positive constant The scheme we constructed Theorem will each grid point, result, at v' \i is denote a given v{W,y,z) < v'{W,y,z) monotone said to be said to be consistent with the partial we have <f>, ||t;''l|t» ^ K, scheme < for all K independent of h, — £/,</> /i where < Second, we ||p for > 0, is any denotes in section 4.2. stable, and scheme for Bellman's equation. monotonicity of the scheme < operators T/^V /iq, • |1 — = / C4> Ch.v^ 3 The discTtte Markov chain approximation yields a monotone, First, if following result establishes the properties of the finite-difference consistent finite-difference Proof. we h, Finally, a finite-difference 0. if /iq is smooth function said to be stable the L°°('D)-norm. < Recall that a finite difference scheme {W,y,z) € Q. jC/iu' by Ch and interpret ChV to be the scheme to the function of applying the a grid of size ShV < ShV T;^V' and Cein verify directly that, for is for implied by monotonicity of the discrete any V,V' G B{Gh) such that any smooth function <^, V < V. and with the obvious notation: Tl^4>\k,i,j)-<f>\Ki,j) Sh<t>'{k,i,j) - <l>'{k, i,j) = = -[l-e-'''''^''''^^)]4>'{k,i,j) [XDr<f>'{k - l,i, j + 1) + + CU'{k.ijW{k,i,j), - D;4>\k,i,j + ^Dl<f>\k,i,j) l)]h, (35) where <i>'*{k,i,j) Df(i> D; (i> = ^{kh,ih,jh). = 4>w± = <i>y -^ ^<f>ww -4>yy X h The difference terms can be estimated xh + C>(/i') + 0{h^) , , a.nd Dl<f> = Dj 4> = 4>ww (f), + + by 0(/i') -<f>,, xh + 0{h^). 5 CONVERGENCE OF THE NUMERICAL SCHEME Therefore, £'^<f> — = €*<!> 27 0{h), which implies the consistency of our finite-difference from Proposition 2 that, any investment policy A, approximation. Finally, it the set {(T;^, uniformly bounded in operator norm, which implies the stability 5/,); V/i} is follows of the finite-difference scheme. Note that consistency I theorem as used in 3 is actually equivalent to the "local" consistency property (28) and (29) of the corresponding 5.2 Markov chain. Convergence of The Value Function In this section, we report the result that the sequence of value functions in the approxi- mating decision problems converges program. The proof of in the for to the Vcdue function of the original continuous time this result requires special attention to the gradient constraints sequence of Bellman's equations of the approximating control problems. Theorem 4 (Convergence) The sequence of solutions J'^ik, i,j) of the discrete Markov chain problems converges to the value function J{W,y,z) of the continuous-time problem uniformly on any compact subset of Ny t oo, in Proof. Q = (0, oo)"' and as h [ such a way that hk —* W, hi —* y, and hj —* Nw T oo? ^z T °°> '^^^ z. For each (W^,y, 2) 6 Q, define r{W,y,z) = limsup J^{W\y\z'), and (36) (W,v',x')-(V*',»,x) fciO, J.(iy,y,z) = (W',v',i')gG'> ,^,liminf, j\W\y\z'). (37) '>10.(W",>i',j')€G'' every point J' is finite at J' is upper semicontinuous (VF, j/,2) since ajid J, is the finite-difference scheme is stable. lower semicontinuous, and J>{W,y,z)<J'{W,y,z), on Q. Note that for {y^z) J''(0, y, 2) = 'I'(y, 2), G (0,00) X (0,oo). We Clearly, V(y, 2) G G^. Therefore, J'(0, y, 2) will verify here that J* is (38) = ^{y, 2) = J.(0, y, z) a viscosity subsolution and J» CONVERGENCE OF THE NUMERICAL SCHEME 5 is 28 a viscosity supersolution of Bellman's equation max{-<5J+max{£^J}+u(j/,2),AJy+^Jj-JH'} =0, on Q with J(0,y,z) = *(y,-2). (39) It < then follows from the comparison result in theorem 2 that J* together with (38), implies that J' there exists limit /i J., and {Nw-, Ny, N^) I some ho > domain spanned by the t cxd in such that for h grid Gh- By result, and hence, by uniqueness, both are equal to J. Q be a strict local mjiximumof J' ~(f> for some Let {Wo, yo, zo) G we take the = on Q. This J, < € (j) C^'^'^{Q). Since such a way that h x {Nw, Ny, N^) ho, {Wo.,yo,2o) is ] oo, an interior point of the the property of "limsup", there exists a sequence indexed by h such that {Wh,yh',Zh) -^ {Wo,yo,zo), For h A/'o < ho, J^ — 4> hais aloccJ maximum at {Wh.,yh-i Zh), and hence, for some neighborhood {W, y, z) of (VVo,2o,yo), J\W,, y,, z,) Since J'' is 4>{Wf,, y,, zh) > J\W, y, z) - 4>{W, y, z), if 6 A^o (40) . a solution of the discrete operator equation j''iWH,yk,ZK) = m^x{m^x[T^j''{WH,yH,ZH)] using (40) and the monotonicity of Tf^ and Sh, max{max[T^^(^ On ash 10. J''{W'h,yh,Zh) -^ J'{Wo,yo,ZQ) the other hand, - <f>]{Wh, we recall y/., + u{yH,ZH)At\S'HJ\WH,yH,Zh)}, we can ^0 + ^Vh, ZK)At\ replace J^ by [5/.<A - <!> <^]{Wh, yn, such that ^O} > from (35) that T;^<i>-<f> = -[l-e-^^''']<P SH<f>-<f> = [XDt4> + + C^4>At\ and (42) 0Dl<f>-D:<f>]h. (43) Therefore, dividing (42) by Af'' and (43) by h, the inequality (41) at {Wh,yh, max{-( (41) 0. ^^r— )^ + max[£;f + u(y,, z,), AD.+ ^ + /3Dlcf> .?i] 2/1) D;4>} > becomes 0, 5 CONVERGENCE OF THE NUMERICAL SCHEME for h < ho- Finally, 29 by the consistency of our fmite-diiference scheme, as we let /i j 0, we obtain max\^-S<f> Thus, J' is + max[£'^(^] + u(yo, zq), X<f>y + ^(f>^ a viscosity subsolution by Definition combined with 2, it follows (38), proves that, for all Hm j\W',y',z') exists that 4>w] > 0, at (Wq, j/o, -^o)- 1. Using similar reasoning, we can show that J, comparison principle, theorem - a viscosity supersolution. is J'{W,y,z) < J,(\V,y,z) on Q By the which, {W,y,z) 6 Q, = = J{W,y,z) J'{W,y,z) = J,{W,y,z). I hlO, (VV",v',i')6G'' In the ifications argument of convergence, we do not require knowledge of the particular specon the cutoff boundaries for the Mao-kov chains. The same limit for J^ will be obtained regardless of the conditions chosen at the cutoff boundaries. This results from the mesh fact that, in the theoretical sequence of approximations, as the creases, the size of the corresponding bounded domain increases. numerical implementation, the size of the restricted domain is However, size de- in the actual kept fixed and hence the choice of the boundary conditions affects the quality of the numerical solution. boundary specifications discussed in the previous sections The were chosen to obtain a good approximation to the solution of the continuous-time problem. 5.3 Convergence of Approximating Policies Existence of the value function and the feasibility of computing its values with high accuracy do not guarantee existence of optimal policies that achieve the value function. Contrast this with classical feedback control case. In that case, optimal controls can be expressed as explicit functions of the value function and existence of a its derivatives. smooth value function immediately implies existence In our case, with a free-boundary problem, there value function and the optimal controls, which, the value function is not known to be smooth. if is no As a consequence, of optimal controls. direct relationship exist, are not feedback. between the Furthermore, CONVERGENCE OF THE NUMERICAL SCHEME 5 30 we prove convergence of the optimal investment In this section, policies in the approx- imating Markov chains to the optimal investment policy of the diffusion control problem if We the latter indeed exists. will parameterize the sequence of approximating problems by h and denote the optimal consumption and investment and A''*, respectively. We policies in problem h by C'^ Markov define an interpolation of the value function for the chain problem using the following linear aifine interpolation: = j\W,y,z) for W< < kh {k -h + DtJHk,i + lJ + l)x{W-kh) +D-J\k,ij + 1) X (y - ih) + D-J\k,i,j) X {z-jh) j'{k,iJ) \)h, ih < y < + {i l)h, and jh < extensions at the cutoff boundaries. Clearly, J'*{W, y, z) [0, Mw] X Lemma [0, 1 Mj,] As W Proof. The proof on Qm = \)h; with appropriate continuous on the cubic region that is nondecreasing in {W,y,z) and con- M]^. [0, J'^ is as the proof of property 827). is + {j X [0,Mj]. One can easily show the following a function of{W,y,z), J^{W,y,z) cave in W< 1 nondecreasing in its arguments follows the same reasoning of the value function in Fitzpatrick Furthermore, the proof of concavity of (page 829) in the same reference. J'' in and Fleming (1991, page W follows the proof of property 2 I Next, we utilize the special monotonicity and concavity features of the problem to prove convergence of the first and second order derivatives dJw{k, i,j) dJwwi^-ihJ) ~^ Jww{^,y,z) is as /i J. and {hk,hi,hj) used to show that the optimal investment A'^ for > > Jw{W^, and for the Markov chain control continuous-time model latter exists. Theorem 5 Suppose that there exists an optimal control policy {A',C') for the original continuos time problem.. Then, the sequence of policies A'^{k,i.,j) converges at {W.,y.,z) y, ^) {W,y,z). This information the discrete problem converges to the optimal investment policy A' when the — — where Jw and Jww exist, which is all points almost everywhere, to the optimal investment 5 CONVERGENCE OF THE NUMERICAL SCHEME A'{W,y,z) as h policy hk — W, hi > Proof. —y y, and hj —* Nw and I 0, oo,Nz 1 1 oo, and Ny ] oo, in such a way that z. Using standard arguments, we can show that the value function J and concave in its arguments. Bellman's equation, J the maximization in satisfies that increasing result, since J is first a viscosity solution of equation in almost everywhere sense. Carrying out we conclude that A' (7), is then follows; see Royden (1988, §5.5), that J has It and second derivatives almost everywhere. As a We 31 = j—-^'^ almost everywhere. rewrite the "finite-difference" form of the Bellman's equation for the interpolated value function J^ _ -SAt" [—^-^]J'{k,tJ) + 1 in the case of no consumption ment policy by the formula A' grid. Note that only the first 0{h) AC = 0. {k, i,j) = = ^ma^^{£if /"(/:, i,;)} + From this, f/(i,j), (44) we can compute the optimal — wa^D^J^>^ ' ^°^ (^' ^'?) and second order derivatives ^ ^~ invest- *^^ interior of J, with respect to W, are used in that formula. Substituting the previous formula back into the difference equation (44), and after a long calculation, we can express DlJ'^ as H{Dfj'^, some continuous function H. Now, using arguments Fleming (1991, theorem and A'^ D~ J^ — A', > of the —^ J^. As a and lemma result, 1 , we can prove that DfJ'^ —^ Qm sumption and abstinence in the as /i | and that as the size approximating sequence converge to the corresponding Specifically, if at a given point (iy,y,z), the optimal jump in solutions, except for, at most, a finite grid points that converge to calls for , Jy, I solution of the original problem prescribes a time problem W » following result that shows that the optimal "regions" of con- regions in the original problem. mating optimal D~ J^ — D\J^ —+ Jww and consequently we must have increases to Q. we provide the Jw, for and similar to those in Fitzpatrick at the points of differentiability with respect to domain Finally, 4), D~ J^, Dj J^i J^) {W , y, z). Similarly, if no consumption at consumption, then, number, all also prescribe the approxi- jumps at the the optimal solution in the continuous {W,y,z), then almost all the approximating CONCLUDING REMARKS 6 optimal solutions Theorem 32 no consumption will also caJl for 6 Suppose that there exists converge to at the grid points that an optimal control policy (A'^C) for the original continues time problem. Consider the sequence of approximating problems indexed by h and h I let 0, and value function J Nw is T °°i ^z T °°> <*"<^ ^y Let x T oo- = Q {W, y,z) € twice continuously differentiate in a neighborhood of x. Let be the size of the optimal consumption jump = Let {x^} at x. {(fc, = O;, then AC*''(x'') (AC'\x^) = > 0), respectively, for AC(x) i,j)} be a sequence of grid points that converges to x in such a way that J'^{x^) —^ J{i)- (AC'{x) be such that the If all h, AC*{x) > except for, at most, a finite number. Consider the case when AC*(i) Proof. = 0. Since J is smooth in a neighborhood of follows from arguments similar to those used for proving the verification result 2— in HH&Z It it — theorem that pj, -\-\Jy- Now x, suppose that AC*''(x'') = for only Jw <Q a at X. finite (45) number of points in the sequence. then follows from Bellman's equation for the Markov chain decision problem that /3D/J'' + XDfJ'' - DtJ'' = a.t x^, for but a all sequence. However, from the proof of theorem lim(/3£>t J'' which contradicts ments and 6 The (45). + XDtJ'' The proof utilizes the result of - 5, DtJ''){x^) of the case finite number we conclude = {^Jz + XJy that - ^)(x) = when AC*(x) > theorem 5 that of elements in the follows the {A'''} converge to A". same argu- I Concluding Remarks analysis of this paper achieves its objective of providing technical support for the numerical solution reported in HH&Z. In particular, we provide proofs of convergence of the sequences of both the vaJue functions, the optimal investment policies, and the CONCLUDING REMARKS 6 33 optimal consumption regions in the approximating problems to the corresponding entities in the original problem. Such results justify the description There are results as e-optimal solutions. We still some We hope that further research We remark that there is silent on the in this area HH&Z of the numerical interesting theoretical have not proved existence of an optimal solution Furthermore, we are also quite in in the class of open questions. admissible controls. reguljirity properties of the free-boundary. sheds some light on these issues. another method of proving convergence of the sequence of value functions in the approximating Markov decision problems. That method is based on probabilistic techniques from the theory of weak convergence of probability measures. For a detailed exposition of the probabilistic approach, and Dupuis (1992, §9 — §10) and the references therein. we The refer the reader to idea there is, Kushner very roughly, as follows. One constructs continuous time interpolations of the optimally controlled Markov chains. Then one proves that the sequence of interpolated processes has a subsequence that converges weakly to an optimally controlled process of the original problem. Since the approximating sequence has the same objective functional as the original problem, one then concludes, using the theory of weak convergence, that the sequence of approximating value functions converges to the value function of the original problem. Furthermore, by extending the space of admissible controls to the so-called "relaxed controls"; see Kushner and Dupuis (1992, relaxed control for the original problem. be §4.6), one can show that there exists an optimal The optimal in the class of original admissible controls. In any relaxed controls, however, need not case, the sequence of approximating optimal controls converges weakly to the optimal relaxed control of the original problem. We chose not to present the probabilistic approach here because the notion of weak convergence does not guarantee that the optimeil relaxed control are adapted to the information structure in the economy. As a result, the optimal relaxed control to be feasible We and admissible control may fail policies. remark, however, that using the analytical approach we can prove that the ap- proximating optimal investment policies and the consumption regions converge pointwise to their corresponding counterparts in the original of admissible controls. This is problem if the latter exists in the class a stronger result than weak convergence to a relaxed con- 7 REFERENCES trol, which may 34 not be in the class of the original admissible controls, that obtains using probabilistic techniques.- Finally, optimaJ policies we note that neither approach establishes existence of in the originad admissible class of controls described in section 2.2. Es- tablishing such a result requires information about the regularity of the free-boundary. For more on this issue, we refer the reader to Soner and Shreve (1989) and Hindy and Huang (1993). References 7 1. M. Crandall and P. L. Lions, Viscosity Solutions of Hamilton- Jacobi Equations, Transactions of the AmericeLD Mathematical Society 277, 1982, pp. 1-42. 2. M. G. Crandall, H. Ishii, ond Order PartiaJ Differential Equations, Bulletin of the and P. L. Lions, User's Guide to Viscosity Solutions of Sec- American Mathematical Society, 1992. 3. B. Fitzpatrick and W. Fleming, Numerical Methods for Optimal Investment-Consumption, Mathematics of Operations Research 16, 1991, pp. 823-841. 4. 5. W. Fleming and H. Soner, Controlled Markov Processes and Viscosity Solutions, Springer- Verlag, New W. Fleming and Player, Zero-Sum York, 1993. P. Souganidis, On The Stochastic Differential Existence of Value Functions of Two- Games, Indiana University Mathematics Journal 38, 1989, pp. 293-314. 6. A. Hindy, C. Huang, and H. Zhu, Optimal Consumption and Portfolio Rules With Durability and Habit Formation, Working Paper, Stanford University, 1993. 7. A. Friedman, Partial Differential Equations of Parabolic Type, Prentice-Hall, New York, 1964. Inc., 7 REFERENCES 8. 35 A. Hindy and C. Huang, Optimal Consumption and Portfolio Rules With Durability and Local Substitution, Econometrica 61, 1993, pp. 85-122. 9. R. Holmes, Geometric Functional Analysis and Its Applications, Springer-Verlag, New 10. H. York, 1975. J. Kushner, Probability Methods for Approximations for Elliptic Equations, Academic Press, 11. H. J. Kushner and P. G. Dupuis, New York, 1977. New York, 1992. O. A. Ladyzenskaja, V. A. Solonnikov, and N. N. Ural'ceva, Linear and Quasilinear Equations of Parabolic Type, Academic Press, 13. Control and Numerical Methods for Stochastic Control Prob- lems in Continuous Time, Springer-Verlag, 12. in Stochastic P. L. New York, 1968. Lions and P. E. Souganidis, Differential Games, Optimal Control, and Direc- tional Derivatives of Viscosity Solutions of Bellman's and Isaac's Equations, SIAM Journal on Control 23, 1985, pp. 566-583. 14. P. L. Lions, Optimal Control of Diffusion Processes and Hamilton-Jacobi-Bellman Equations, Part II: Viscosity Solutions and Uniqueness, CommuDications in Pure and Applied Matbematic's 37, 1983, pp. 511-553. 15. H. L. Royden, Real Analysis, Third Edition, Macmillan Publishing Co., New York, 1988. 16. H. M. Soner and S. E. Shreve, Regularity of the Value Function of a Singular Stochastic Control Problem, SIAM Two-Dimensional Journal on Control and Optimization 27, 1989, pp. 876-907. 17. J. C. Strikwerda, Finite Difference Wadsworth & Schemes and Partial Brooks, California, 1989. Differential Equations, 7 REFERENCES 18. T. Zariphopoulou, Investment-Consumption Models with Transaction Fees and Markov Chain - 36 Paramieters, SIAM Journai on Control and Optimization 30, 1992, pp. 613-636. 19. H. Zhu, Variational Inequalities and Dynamic Programming for Singular Stochastic Control, Doctoral Dissertation, Division of Applied Mathematics, sity, 1991. Brown Univer- 2935 j6i. MIT LIBRARIES 3 lOaO D0flSbSE4 1 Date Due