A PDE approach to regularity of solutions to finite horizon optimal switching problems (Q1044477)

Let \((X_{t})\) be a Itô diffusion with infinitesimal generator \({\mathcal A}=\sum_{i,j=1}^{n}c_{ij}\partial^{2}/\partial x_{i}\partial x_{j} +\sum_{i=1}^{n}b_{i}\partial /\partial x_{i} \). A production activity can take place in \(q\) modes. The time is finite \([0,T]\). There are stopping times \(0<\tau_{1}<\tau_{2}<\dots <\gamma \leq T\), at \(\gamma\) the activity stops, the mode is constant \(u_{t}\) between two successive \(\tau\)'s and the profit is \(\int_{_{0}}^{^{\gamma }}\Phi (s,X_{s},u_{s})ds\) minus the sum of costs \(r_{ij}(\tau_{k})\) of swiching from the mode \(i\) to \(j\) at time \(\tau_{k}\) and minus the cost of stopping \(F(\gamma ,X_{\gamma },u_{\gamma })\). The problem of chosing the \(\tau\)'s and \(\gamma\) in order to maximize the mean value of the profit is solved by \(\tau\)'s and \(\gamma\) defined as the first moments when some processes reach \(0\), these processes being expressed in terms of \(r_{ij}\), \(F\) and the solution \((v_{i})_{i=1,\dots,q}\) of a system of the form: \(\min(v_{i}(t,x)-\max(\max_{j\neq i}(v_{j}(t,x)-a_{ij}(t)),F_{i}(t,x)) , -{\mathcal H}v_{i}(t,x)-\psi_{i}(t,x))=0\), \(i=1,\dots,q\), \(v_{i}(T,\cdot )=0\), where \({\mathcal H}=\partial /\partial t+{\mathcal A}\) and \(a_{ij}> 0\). By a viscosity subsolution of the system is meant a \((v_{i})\) such that, for every \(\varphi_{1},\varphi_{2}\in C^{2}([0,T]\times {\mathbb R}^{n})\) and every \((t,x)\) which is a local maximum of \(\varphi_{i}-v_{i}\) and in which \(\varphi_{i}-v_{i}=0\), the left members are \(\geq 0\). A subsolution is the same, with ``local minimum'' and \(\leq 0\), and a viscosity solution is one which is a super and subsolution. The following notations are introduced. For \(x,y\in {\mathbb R}^{n}\), \(r>0\), \(B_{r}(x)=(\| \cdot -x\| <r)\), for \(X=(t,x)\), \(Y=(s,y)\), \(s,t\in {\mathbb R}\), \(Q_{t}^{-}(X)=(t-r^{2},t]\times B_{r}(x)\), \(d(X,Y)=(\| x-y\|^{2}+|t-s|)^{1/2}\), \(\| f\|_{C(\Omega )}= \sup_{\Omega }|f|\), \(\| f\|_{_{C}k_{(\Omega )}}= \sum_{|i|+2j\leq k}\| D_{x}^{i}D_{t}^{j}f\|_{C(\Omega )}\) (\(D_{t}^{j}=\partial^{j}/\partial t^{j}\) etc), \(\| f\|_{_{C}0,1_{(\Omega )}}= \| f\|_{C (\Omega )}+\sup_{X,Y\in \Omega }(|f(X)-f(Y)|d(X,Y^{-1}))\), \([f]_{1,1}(X) = \inf_{_{a\in {\mathbb R},b\in {\mathbb R}}n}\sup_{r>0}r^{-1}\| f-(a+b\cdot ) \|_{_{C(Q_{t}}-_{(X)\cap \Omega )}}\), \(\| f\|_{_{C}1,1_{(\Omega )}}= \| f\|_{_{C}1_{(\Omega )}}+\sup_{X\in \Omega }[f]_{1,1}(X)\). One result is: if \(\| F_{i}\|_{_{C}1,1_{(\Omega )}}\), \(\| a_{ij}\|_{_{C}1,1_{(\Omega )}}<\infty\), \((v_{i})\) is a viscosity solution and \(A_{ij}=(v_{i}=\max_{k\neq i}(v_{k}-a_{ik})=v_{j}-a_{ij})\), then \(v_{i}\in C^{0,1}([0,T]\times {\mathbb R}^{n})\cap C^{1,1} (([0,T]\times {\mathbb R}^{n})\backslash (\cup_{j\neq i}\partial A_{ij}))\). When \(q=2\), the authors prove the unicity of the solution, in three particular cases: 1. \(F_{1}=F_{2}=-\infty\), 2. \(F_{1}=F_{2}\), \(a_{1}, a_{2}\) constant and \(\psi_{1}\neq \psi_{2}\) everywhere, 3. \(\psi_{1}=\psi_{2}\), \(F_{1}=F_{2}\) and \(a_{1}, a_{2}\) constant. They give also an example when the regularity result does not extend to \(\partial A_{12}\), \(A_{12}=(v_{1}=v_{2}-a_{12})\) in the case \(q=2\). From the last paragraph (numerical results). ``We apply the iterative method of projected successive over relaxation'', ``the Crank Nicolson scheme''. In order to reduce the number of calculations and get faster convergence, the problem is reduced to one in which \({\mathcal A}=(1/2)\partial^{2}/\partial x^{2}\). An algorithm is presented, at the end.