ebook img

A Dual Method For Backward Stochastic Differential Equations with Application to Risk Valuation PDF

0.32 MB·
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview A Dual Method For Backward Stochastic Differential Equations with Application to Risk Valuation

ff A Dual Method For Backward Stochastic Di erential Equations with Application to Risk Valuation AndrzejRuszczyn´ski JianingYao ∗ † 7 1 January 24,2017 0 2 n a J Abstract 2 2 Weproposeanumericalrecipeforriskevaluationdefinedbyabackwardstochasticdiffer- entialequation. Usingdualrepresentationoftheriskmeasure,weconverttheriskvaluationto ] C astochasticcontrolproblemwherethecontrolisacertainRadon-Nikodymderivativeprocess. Byexploringthemaximumprinciple,weshowthatapiecewise-constantdualcontrolprovides O agoodapproximationonashortinterval. Adynamicprogrammingalgorithmextendstheap- . h proximationto a finite time horizon. Finally, we illustrate theapplicationofthe procedureto t a riskmanagementinconjunctionwithnestedsimulation. m Keywords:DynamicRiskMeasures,Forward–BackwardStochasticDifferentialEquations,Stochas- [ ticMaximumPrinciple,RiskManagement 1 v 1 Introduction 4 3 2 The main objective of this paper is to present a simple and efficient numerical method for solving 6 0 backward stochastic differential equations with convex and homogeneous drivers. Such equations . are fundamental modeling tools for continuous-time dynamic risk measures with Brownian filtra- 1 0 tion,butmayalsoariseinotherapplications. 7 Thekeyproperty ofdynamic riskmeasuresistime-consistency, whichallowsfordynamic pro- 1 grammingformulations. ThediscretetimecasewasextensivelyexploredbyDetlfsenandScandolo : v [11], Bion-Nadal [5], Cheridito et al. [7, 8], Föllmer and Penner [12], Fritelli and Rosazza Gianin i X [15],FrittelliandScandolo [16],Riedel[32],andRuszczyn´ski andShapiro[35]. r Forthecontinuous-time case,Coquet,Hu,MéminandPeng[9]discovered thattime-consistent a dynamic risk measures, with Brownian filtration, can be represented as solutions of Backward Stochastic Differential Equations (BSDE)[30];undermildgrowthconditions, thisistheonlyform Rutgers University, Department of Management Science and Information Systems, Piscataway, NJ 08854, USA, ∗ Email:[email protected] Rutgers University, Department of Management Science and Information Systems, Piscataway, NJ 08854, USA, † Email:[email protected] 1 possible. Specifically, the y-part solution of one-dimensional BSDE, defined below, measures the riskofavariableξ atthecurrent timet: T T T Y = ξ + g(s,Z ) ds Z dW , 0 t T, (1) t T s s s Z −Z ≤ ≤ t t withthedrivergbeinginterpretedasa“riskrate.” The -measurablerandomvariableξ isusually T T F afunction oftheterminalstateofacertainstochastic dynamical system. Inspired by that, Barrieu and ElKaroui provided a comprehensive study in [3, 4]; further con- tributionsbeingmadebyDelbean,Peng,andRosazzaGianin[10],andQuenezandSulem[31](for amoregeneral modelwithLevyprocesses). Inaddition, application tofinancewasconsidered, for example,in[19]. Usingtheconvergence resultsofBriand,DelyonandMémin[6],Stadje[39]finds thedriversofBSDEcorresponding todiscrete-time riskmeasures. Motivated by an earlier work on risk-averse control of discrete-time stochastic process [37], Ruszczyn´skiandYao[38]formulatearisk-aversestochasticcontrolproblemfordiffusionprocesses. Thecorresponding dynamicprogrammingequationleadstoadecoupledforward–backwardsystem ofstochastic differentialequations. While forward stochastic differential equations can be solved by several efficient methods, the main challenge is the numerical solution of (1), where ξ represents the future value function. In T particular, Zhang [41] and Touzi et al. [40] use backward Euler’s approximation and regression. Such an approach, however, is not well-suited for risk measurement, because it does not preserve the monotonicity of the risk measure. Alternatively, Øksendal [28] directly attacks continuous- timerisk-aversecontrolproblemwithjumpsbyderivingsufficientconditions. Algorithmsbasedon maximumprinciple wereinvestigated byLudwigetal. [23]. Ourideaistoderivearecursivemethodbasedonrisk-aversedynamicprogramming,sothatthe approximation becomesatime-consistent coherent riskmeasureindiscrete time. Thepaperisorganized asfollows. Insection 2,wequickly introduce theconcept ofadynamic risk measure and review its properties. In section 3 werecall the dual representation of a dynamic risk measure and formulate an equivalent stochastic control problem. The optimality condition for thedualcontrolproblem, aspecialformofamaximumprinciple, arederivedinsection4. Sections 5 and 6 estimate the errors introduced by using constant processes as dual controls. In section 7 we present the whole numerical method with piecewise constant dual controls and analyze its rate of convergence. Finally, in section 8, weillustrate the efficacy ofour approach on atwo-stage risk managementmodel. 2 The Risk Evaluation Problem Givenacompletefilteredprobabilityspace(Ω, ,P)withfiltration generatedbyd-dimen- t t [0,T] sionalBrownianmotion W ,weconsideFr thefollowingstoch{Fast}ic∈ differentialequation: t t [0,T] { }∈ dX = b(t,X) dt+σ(t,X) dW, X = x, t [0,T], (2) t t t t 0 ∈ withmeasurable b : [0,T] Rn Rn,andσ : [0,T] Rn Rn Rd. Weintroduce thefollowing × → × → × notation. 2 E[ ]:= E[ ]; t t • · ·|F 2(Ω, ,P;Rn): the set of Rn-valued -measurable random variables ξ such that ξ 2 := t t • L F F k k E[ ξ2] < ;forn = 1,wewriteit 2(Ω, ,P); | | ∞ L F 2,n[t,T]: thesetofRn-valuedadaptedprocessesYon[t,T],suchthat Y 2 := E T Y 2ds < • H k k 2,n[t,T] t t | s| ;forn= 1wewriteit 2[t,T];1 H hR i ∞ H k([t,T] Rn)thespaceoffunctions f : [t,T] Rn Rwhosehasderivativeuptok-th,in • Cb × × → themeanwhile, allthosederivatives arecontinuous andbounded; for f : Rn R,wedenote 7→ byCk(RN); b (B): thespaceofLipschitzcontinuous functions f : B R. L • C → Wemakefollowingassumptions aboutthedriftandvolatility terms. Assumption2.1. (i) b(,0) + σ(,0) 2[0,T]; | · | | · | ∈ H (ii) Thefunctions b,σ 1([0,T] Rn),theconstantC > 0denotes theLipschitzconstants ∈ Cb × b(t,x ) b(t,x ) + σ(t,x ) σ(t,x ) C x x a.s. 1 2 1 2 1 2 | − | | − |≤ | − | b(t,x ) + σ(t,x ) C x , a.s.. 1 1 1 | | | | ≤ | | (iii) The dimension of Brownian motion and the state process coincide, i.e., n = d, and we also assume 1 σ(t,x)σ (t,x) , (t,x) [0,T] Rd. ⊤ ≥ CI ∀ ∈ × Ourintention istoevaluateriskofaterminalcostgenerated bytheforwardprocess(2): ρ Φ(X ) , (3) 0,T T (cid:2) (cid:3) where Φ (Rn) is bounded, and ρ is a dynamic risk measure consistent with the ∈ CL s,t 0 s t T filtration t t [0,T]. Wereferthereader(cid:8)to[(cid:9)29≤]≤fo≤racomprehensivediscussiononriskmeasurement {F}∈ andfiltration-consistent evaluations. Special role in the dynamic risk theory is played by g-evaluations which defined by one- dimensional backwardstochastic differentialequations ofthefollowingform: dY = g(t,Y,Z)dt Z dW, Y = Φ(X ), t [0,T], (4) t t t t t T T − − ∈ g withρ Φ(X )]definedtobeequaltoY. ThedrivergisjointlyLipschitzin(y,z),andtheprocess t,T T t g(,0,0)(cid:2) 2[0,T]. · ∈ H As proved in [9] every F-consistent nonlinear evaluation that is dominated by a g-evaluation with g = µy + νz with some ν,µ > 0 is in fact a g-evaluation for some g; the dominance is | | | | understood asfollows: ρ [ξ+η] ρ [ξ] ρν,µ[η] 0,T − 0,T ≤ 0,T forallξ,η 2(Ω, ,P). T ∈ L F 1Whenthenormisclearfromthecontext,thesubscriptsareskipped. 3 Proposition 2.2. Forall0 t T andallξ,ξ (Ω, ,P),thefollowingproperties hold: ′ 2 T (i) Generalized constan≤tpr≤eservation: Ifξ ∈ L(Ω, F,P),thenρg [ξ] = ξ; (ii) Timeconsistency: ρg [ξ] = ρg [ρg [ξ]∈],Lfo2ralFl0t s t; t,t s,T s,t t,T ≤ ≤ (iii) Localproperty: ρg [ξ +ξ ]= ρg [ξ]+ ρg [ξ ],forallA . t,T 1A ′1Ac 1A t,T 1Ac t,T ′ ∈ Ft From now on, shall focus exclusively ong-evaluations asdynamic risk measures, and weshall skipthesuperscript ginρg. Theevaluation ofriskisequivalent tothesolution ofadecoupled forward–backward systemof stochastic differential equations (2)–(4). Animportant virtueofthissystemisitsMarkovproperty: ρ Φ(X ) = v(t,X), (5) t,T T t (cid:2) (cid:3) wherev: [0,T] Rn R. Wehave × → v(t,x) = ρx Φ(Xt,x) , (t,x) [0,T] Rn, (6) t,T T ∈ × (cid:2) (cid:3) where Xt,x isthesolutionofthesystem(2)restartedattimetfromstate x: s { } dXt,x = b(s,Xt,x) ds+σ(s,Xt,x)dW , s [t,T], Xt,x = x, (7) s s s s ∈ t and ρx Φ(Xt,x) is the (deterministic) value of Yt,x in the backward equation (4) with terminal t,T T t condition(cid:2) Φ(Xt,x(cid:3)). T Numericalmethodsforsolvingforwardequationsareverywellunderstood(see,e.g.,[18]). We focus, therefore, onthebackward equation (4). Sofar, alimitednumber ofresults areavailable for thispurpose. ThemostprominentistheEulermethodwithfunctional regression (see,e.g.,[24,25, 40, 41, 42]). Our intention is to show that for drivers satisfying additional coherence conditions, a muchmore effective method canbe developed, which exploits time-consistency, duality theory for riskmeasures, andthemaximumprinciple instochastic control. 3 The Dual Control Problem Wefurther restrict the risk measures under consideration tocoherent measures, bymaking the fol- lowingadditional assumption aboutthedriverg. Assumption3.1. Thedrivergsatisfiesforalmostallt [0,T]thefollowingconditions: ∈ (i) gisdeterministic andindependent ofy,thatis,g :[0,T] Rd R,andg(,0) 0; × → · ≡ (ii) g(t, )isconvex forallt [0,T]; · ∈ (iii) g(t, )ispositively homogeneous forallt [0,T]. · ∈ g Undertheseconditions, onecanderivefurtherpropertiesoftheevaluationsρ []fort [0,T], t,T · ∈ inadditiontothegeneralpropertiesofF-consistentnonlinearexpectationsstatedinProposition2.2. Theorem3.2. Supposegsatisfies Assumption3.1. Thenthedynamicriskmeasure ρ has t,r 0 t r T { } ≤≤ ≤ thefollowingproperties: 4 (i) TranslationProperty: forallξ (Ω, ,P)andη (Ω, ,P), 2 r 2 t ∈ L F ∈ L F ρ [ξ+η] = ρ [ξ]+η, a.s.; t,r t,r (ii) Convexity: forallξ,ξ (Ω, ,P)andallλ L (Ω, ,P)suchthat0 λ 1, ′ 2 r ∞ t ∈ L F ∈ F ≤ ≤ ρ [λξ+(1 λ)ξ ] λρ [ξ]+(1 λ)ρ [ξ ], a.s.. t,r ′ t,r t,r ′ − ≤ − (iii) PositiveHomogeneity: forall ξ (Ω, ,P)and allβ L (Ω, ,P)such thatβ 0, we 2 r ∞ t ∈ L F ∈ F ≥ have ρ [βξ] = βρ [ξ], a.s.. t,r t,r Itfollowsthat under Assumption 3.1,the operators ρ constitute afamilyofcoherent t,r 0 t r T { } ≤≤ ≤ conditional measuresofrisks. Particularly important forusisthedualrepresentation oftheriskmeasure ρ ,whichis t,T t [0,T] basedonthedualrepresentation ofthedriverg: (cid:8) (cid:9)∈ g(t,z) = maxνz, ν At ∈ whereA = ∂ g(t,0)isabounded,convex,andclosedsetinRn. Thefollowingstatementspecializes t z theresultsofBarrieuandElKaroui[3]toourcase. Theorem3.3. Suppose Assumption3.1issatisfied. Then ρg Φ(Xt,x) = sup E Γ Φ(Xt,x) , (8) t,T T t,T T (cid:2) (cid:3) µ∈At,T (cid:2) (cid:3) where isthespaceofA -valuedadaptedprocesseson[t,T],andtheprocess Γ satisfies At,T s t,s s [t,T] thestochastic differential equation: (cid:8) (cid:9) ∈ dΓ = µ Γ dW , s [t,T], Γ = 1. (9) t,s s t,s s t,t ∈ Moreover, asolutionµˆ oftheoptimalcontrolproblem (8)–(9)exists. Thefollowinglemmaprovides ausefulestimate. Lemma 3.4. A constant C exists, such that for all 0 t < s T and all Γ that satisfy (9), we t,s ≤ ≤ { } have Γ 1 2 C(s t). (10) t,s k − k ≤ − Proof. UsingItôisometry, weobtainthechainofrelations s s s Γ 1 2 = µ Γ 2dr µ 2 Γ 2dr µ 2 1+ Γ 1 2 dr. (11) t,s r t,r r t,r r t,r k − k Z k k ≤ Z k k k k ≤ Z k k k − k t t t (cid:0) (cid:1) IfK isauniformupperboundonthenormofthesubgradientsofg(r,0)wededucethat Γ 1 2 t,r ψ ,r [t,T],whereψsatisfiestheODE: dψr = K2(1+ψ ),withψ = 0. Consequently,k − k ≤ r ∈ dr r t Γ 1 2 ψ = eK2(r t) 1. (12) t,r r − k − k ≤ − Theconvexity oftheexponential function yieldsthepostulated bound. (cid:3) 5 Thedualrepresentationtheoremallowsustotransformtheriskevaluationproblemtoastochas- ticcontrolproblem. Ourobjectivenowistoapproximatetheevaluationofriskonashortinterval∆ sothat it reduces the functional optimization problem tovector optimization. Toproceed, wehave tofirstinvestigate thecorresponding maximumprinciple ofthecontrolproblem (8). 4 Stochastic Maximum Principle In this section, we decipher the optimality conditions of the stochastic control problem (8)–(9). Since only the process Γ is controlled, the analysis is rather standard. For completeness, t,s s [t,T] { } ∈ werepeatsomeimportantstepshere. Suppose µˆ isthe optimal control; then, forany µ and 0 α 1, wecanform aperturbed ∈ A ≤ ≤ controlfunction µα = µˆ +α(µ µˆ). − It is still an element of , due to the convexity of the sets A . The processes Γˆ, Γ and Γα are the s A stateprocesses underthecontrols µˆ,µ,andµ ,respectively. α Welinearize thestateequation(9)aboutΓˆ toget,for s [t,T], ∈ dηµ = µˆ ηµ+Γˆ (µ µˆ ) dW , ηµ = 0. (13) s s s t,s s− s s t (cid:2) (cid:3) Itisevidentthatthisequation hasauniquestrongsolution. Denote 1 hα = Γα Γˆ ηµ, s [0,T]. s α t,s− t,s − s ∈ (cid:2) (cid:3) Thefollowingresultjustifiestheusefulness ofthelinearized equation (13). Lemma4.1. lim sup hα 2 = 0. (14) α 00 s Tk sk → ≤ ≤ Proof. Wefirstprovethat lim sup Γα Γˆ 2 = 0. (15) α 0t s T k t,s− t,sk → ≤ ≤ Wehave d Γα Γˆ = µαΓα µˆ Γˆ dW = (µα µˆ )Γˆ +µα(Γα Γˆ ) dW . (16) t,s − t,s s t,s− s t,s s s − s t,s s t,s − t,s s (cid:0) (cid:1) (cid:0) (cid:1) (cid:0) (cid:1) ByItôisometry, r Γα Γˆ 2 = (µα µˆ )Γˆ +µα(Γα Γˆ ) 2 ds k t,r − t,rk Z k s − s t,s s t,s − t,s k t r r 2 (µα µˆ )Γˆ 2 ds+2 µα(Γα Γˆ ) 2 ds ≤ Z k s − s t,sk Z k s t,s− t,s k t t r r 2 (µα µˆ )Γˆ 2 ds+K Γα Γˆ 2 ds, ≤ Z k s − s t,sk Z k t,s− t,sk t t 6 where K is a constant. Since the first integral on the right hand side converges to 0, as α 0, the → Gronwallinequality yields(15). Wecannowprove(14). Combining(16)and(13),weobtainthestochasticdifferentialequation forhα: 1 dhα = (µˆ +α(µ µˆ ))Γα µˆ Γˆ µˆ ηµ Γˆ (µ µˆ ) dW s (cid:26)α s s− s t,s − s t,s − s s − t,s s − s (cid:27) s (cid:2) (cid:3) 1 = µˆ Γα Γˆ +(µ µˆ ) Γα Γˆ µˆ ηµ dW (cid:26)α s t,s− t,s s− s t,s− t,s − s s(cid:27) s (cid:2) (cid:3) (cid:2) (cid:3) = µˆ hα+(µ µˆ ) Γα Γˆ dW . (cid:26) s s s − s t,s− t,s (cid:27) s (cid:2) (cid:3) Sincetheprocesses µˆ and µ arebounded, Itôisometryyieldsagain s s { } { } r r hα 2 K hα 2 ds+K Γα Γˆ 2 ds, k rk ≤ Z k sk Z k t,s − t,sk t t where K isconstant. BytheGronwallinequality, using(15),wegetthedesiredresult. (cid:3) Theconvergence resultabovedirectly leadstothefollowingvariational inequality. Lemma4.2. Foranyµ wehave ∈ A E ξ ηµ 0. (17) T T ≤ (cid:2) (cid:3) Proof. Sinceµˆ istheoptimalcontrol, E ξ Γα Γˆ 0. T t,T − t,T ≤ (cid:2) (cid:0) (cid:1)(cid:3) Lemma4.1leadsto 1 limE ξ Γα Γˆ = E ξ ηµ 0, α 0 Tα t,T − t,T T T ≤ → h (cid:0) (cid:1)i (cid:2) (cid:3) asrequired. (cid:3) We now express the expected value in (17) as an integral, to obtain a pointwise variational inequality (the maximum principle). To this end, we introduce the following backward stochastic differential equation(theadjointequation): dp = k µˆ ds+k dW , p = ξ , s [t,T], (18) s s s s s T T − ∈ withξ = Φ Xt,x . Byconstruction, E ξ ηµ = E pˆ ηµ . ApplyingtheItôformulatotheproduct T T T T T T µ process psηs(cid:0),we(cid:1)obtain (cid:2) (cid:3) (cid:2) (cid:3) d(p ηµ) = k ηµ+ pˆ µˆ ηµ+Γˆ (µ µˆ ) dW +k Γˆ (µ µˆ ) ds. s s s s s s s t,s s s s s t,s s s − − (cid:16) (cid:2) (cid:3)(cid:17) Iffollowsthat T E ξ ηµ = E k Γˆ (µ µˆ )ds . (19) T T (cid:20) Zt s t,s s− s (cid:21) (cid:2) (cid:3) 7 We can summarize our derivations in the following version of the maximum principle. We define theHamiltonian H : R Rn Rn R: × × → H(γ,ν,κ) = kγν. Theorem4.3. Foralmostall s [t,T],withprobability 1, ∈ H(Γˆ ,µˆ ,k )= maxH(Γˆ ,ν,k ). t,s s s t,s s ν A ∈ Proof. Foranyµ ,wedefinetheset ∈ A = (ω,s) Ω [t,T] : k Γˆ (µ µˆ )> 0 . s t,s s s G { ∈ × − } Weconstruct anewcontrolµ : ∗ ∈ A µ , (ω,s) , µ = s ∈ G ∗s  µˆ , otherwise.  s Themeasurability andadaptedness ofµ canbeeasilyverified. Itfollowsfrom(17)and(19)that ∗ T E k Γˆ (µ µˆ )ds 0. (cid:20) Zt s t,s ∗s − s (cid:21) ≤ Bytheconstruction ofµ , ∗ k Γˆ (µ µˆ )ds P(dω) 0. " s t,s ∗s − s ≤ G Sincetheintegrandispositiveon ,theproductmeasureof mustbezero. (cid:3) G G 5 Regularity of the Integrand in the Adjoint Equation Wealsomakeastronger assumption aboutthedrift anddiffusion termsoftheforwardsystem, and abouttheterminalcostfunction. Assumption5.1. Thefunctions b,σ,Φ 2([0,T] Rn),and ∈ Cb × 1 σ(s,x) σ(t,x) C s t 2 | − | ≤ | − | forall s,t [0,T]andall x Rn. ∈ ∈ Consider the forward–backward system (7) and (18). The key to our further estimates is the followingregularity resultabouttheintegrand k intheadjointequation (18). t { } Lemma5.2. AconstantC exists, suchthatforall0 t < s T,andall x Rn, ≤ ≤ ∈ 1 ks kt C s t 2. (20) k − k ≤ | − | 8 Proof. Thequasilinearparabolicpartialdifferentialequationcorrespondingtotheforward–backward system(7)–(18)hasthefollowingform(see,e.g. [26,sec. 8.2]), 1 u(t,x)+u (t,x)b(t,x)+ tr u (t,x)σ(t,x)σT(t,x) +u (t,x)σ(t,x)µˆ = 0, (21) t x xx x t 2 (cid:0) (cid:1) with the boundary condition u(T,x) Φ(x). Due to the linearity of the driver of (18), the terms ≡ withu canbecollapsed. Thentheequation (21)istheFeynman-Kacequation for x u(s,X˜t,x) = E Φ X˜t,x , s [t,T], (22) s T Fs ∈ (cid:2) (cid:0) (cid:1)(cid:12)(cid:12) (cid:3) where (cid:12) dX˜t,x = b(s,X˜t,x)+σ(s,X˜t,x)µˆ ds+σ(s,X˜t,x)dW , s [t,T], X˜t,x = x. s s s s s s ∈ s (cid:2) (cid:3) Ma and Yong [26] consider it on page 195 in formula (1.12). Under Assumption 5.1, the equation (21)hasaclassical solution u(, ),andthentheprocess · · k = u (s,X˜t,x)σ(s,X˜t,x), s [t,T], (23) s x s s ∈ isthesolution oftheadjoint equation (18). By[26,Prop. 8.1.1], aprocess H 2,n n[t,T]exists, × suchthattheprocessG = u (s,X˜t,x)satisfiesthefollowingn-dimensional BS∈DEH: s x s T T G = Φ (X˜t,x)+ b (s,X˜t,x)+σ (s,X˜t,x)µˆ G +σ (s,X˜t,x)H ds H dW . (24) t x T Z x s x s s s x s s −Z s s t (cid:16)(cid:2) (cid:3) (cid:17) t Weobtainthefollowingestimate: k k = u (s,X˜t,x)σ(s,X˜t,x) u (t,x)σ(t,x) k s − tk k x s s − x k u (s,X˜t,x)σ(s,X˜t,x) u (s,X˜t,x)σ(t,x)+u (s,X˜t,x)σ(t,x) u (t,x)σ(t,x) (25) ≤ k x s s − x s x s − x k u (s,X˜t,x) σ(s,X˜t,x) σ(t,x) + σ(t,x) G G . ≤ k x s kk s − k k kk s − tk Thefirsttermontherighthandsideof (25)canbebounded withthehelpofAssumption5.1: σ(s,X˜t,x) σ(t,x) σ(s,X˜t,x) σ(s,x) + σ(s,x) σ(t,x) k s − k ≤ k s − k k − k C1 s t 21 +C2 X˜s x C3 s t 21, ≤ | − | k − k≤ | − | whereC ,C ,andC aresomeuniversalconstants. Itfollowsfrom(24)that 1 2 3 r r G G = b (s,X˜ )+σ (s,X˜ )µˆ G +σ (s,X˜ )H ds+ H dW . r t x s x s s s x s s s s − −Z Z t (cid:16)(cid:2) (cid:3) (cid:17) t Therefore, thesecondtermontherighthandsideof (25)canbeboundedas G G 2 C s t. s t 4 k − k ≤ | − | Integrating theseestimatesinto(25),weobtain(20)withauniversal constantC. (cid:3) 9 6 Error Estimates for Constant Controls on Small Intervals To reduce an infinite dimensional control problem to a finite dimensional vector optimization, we partition the interval [0,T] into N short pieces of length ∆ = T/N, and develop a scheme for evaluating the risk measure (3) by using constant dual controls on each piece. We denote t = i∆, i fori= 0,1,...,N. For simplicity, in addition to Assumption 3.1, we assume that the driver g does not depend on time,andthusallsets A = ∂g(0)arethesame. Wedenotethemwiththesymbol A;asweshallsee t lateronthisisnotamajorrestriction. If the system’s state at time t is x, then the value of the risk measure (6) is then the optimal i valueofproblem (8). Bydynamicprogramming, v(ti,x) = ρtxi,ti+1 v ti+1,Xttii+,x1 . (cid:2) (cid:0) (cid:1)(cid:3) The risk measure ρtxi,ti+1[·] is defined by problem (8), with terminal time ti+1 and the function Φ(·) replaced by v(ti+1, ·). Equivalently, it is equal to Yttii,x, in the corresponding forward–backward systemontheinterval[ti,ti+1]: dXti,x = b(s,Xti,x) ds+σ(s,Xti,x)dW , Xti,x = x, (26) s s s s ti −dYsti,x = g(Zsti,x)ds−Zsti,x dWs, Yttii+,x1 = v ti+1,Xttii+,x1 . (27) (cid:0) (cid:1) Under Assumption 5.1, the function v(, ) is the classical solution of the associated Hamilton– · · Jacobi–Bellman equation: 1 v(t,x)+v (t,x)b(t,x)+ tr v (t,x)σ(t,x)σT(t,x) +g v (t,x)σ(t,x) = 0, (28) t x xx x 2 (cid:0) (cid:1) (cid:0) (cid:1) withtheterminalcondition v(T,x) = Φ(x). Supposeweuseaconstantcontrol intheinterval[ti,ti+1]: µs := µˆti = argmax ktiν, ∀s ∈ [ti,ti+1], (29) ν A ∈ where(p,k)solvetheadjoint equationcorresponding to(18): dps = −ksµˆsds+ksdWs, s∈ [ti,ti+1], pti+1 = v ti+1,Xttii+,x1 . (30) (cid:0) (cid:1) Westill use Γˆ to denote the state evolution under the optimal control, while Γ is the process under control µ defined in (29). It is well-known that the value function v(, ) of the system (26)–(27) is · · in 2([0,T] Rn); see, for example, [42, Thm. 2.4.1]. Therefore, the bounds developed insection Cb × 5remainvalidfortheprocesses (p,k)in(30). Our objective is to show that a constant C exists, independent of x, N, and i, such that the approximation errorontheithintervalcanbebounded asfollows: 0 ≤ E v ti+1,Xttii+,x1 Γˆti,ti+1 −Γti,ti+1 ≤C∆23. (31) (cid:2) (cid:0) (cid:1)(cid:0) (cid:1)(cid:3) The fact that wedo not know k will not be essential; later, we shall generate even better constant ti controls bydiscrete-time dynamicprogramming. Wecannowderivesomeusefulestimates fortheconstant controlfunction (29). 10

See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.