This work is motivated by numerical solutions to Hamilton-Jacobi-Bellman quasivariational inequalities (HJBQVIs) associated with combined stochastic and impulse control problems. In particular, we consider (i) direct control, (ii) penalized, and (iii) semi-Lagrangian discretization schemes applied to the HJBQVI problem. Scheme (i) takes the form of a Bellman problem involving an operator which is not necessarily contractive. We consider the well-posedness of the Bellman problem and give sufficient conditions for convergence of the corresponding policy iteration. To do so, we use weakly chained diagonally dominant matrices, which give a graph-theoretic characterization of weakly diagonally dominant M-matrices. We compare schemes (i)-(iii) under the following examples: (a) optimal control of the exchange rate, (b) optimal consumption with fixed and proportional transaction costs, and (c) pricing guaranteed minimum withdrawal benefits in variable annuities. We find that one should abstain from using scheme (i).
Definition 2.1 (Monotone matrix). A real square matrix A is monotone (in the sense of Collatz) if for all real vectorsWe use the following assumptions:Proposition 2.2 (Convergence of policy iteration). Suppose (H0), (H1), and that A(P ) is a monotone matrix for all P in P.(v ) ≥1 defined by Policy-Iteration is nondecreasing and converges to the unique solution v of (2.1). Moreover, if P is finite, convergence occurs in at most |P| iterations (i.e. v |P| = v |P|+1 = · · · ).The monotone convergence of (v ) ≥1 to the unique solution of (2.1) can be proven similarly to Theorem A.3 of Appendix A. See [7, Theorem 2.1] for a proof of the finite termination when P is finite. In practice, P is often finite, in which case (H0) and (H1) are trivial.Remark 2.3. Theorem A.3 establishes the existence and uniqueness of solutions to (2.1) independent of (H1.ii). Owing to this, results that rely on Proposition 2.2 can be relaxed to exclude (H1.ii), with the caveat that when P is infinite, Policy-Iteration be replaced by -Policy-Iteration (see Appendix A). In this case, the resulting sequence (v ) ≥1 is not necessarily nondecreasing.
Weakly chained diagonally dominant matricesWe say row i of a complex matrix A := (a ij ) is strictly diagonally dominant (SDD) if |a ii | > j =i |a ij |. We say A is SDD if all of its rows are SDD. Weakly diagonally dominant (WDD) is defined with weak inequality instead.
Definition 3.1. A complex square matrix A is said to be a weakly chained diagonally dominant (WCDD) if:(i) A is WDD; (ii) for each row r, there exists a path in the graph of A from r to an SDD row p.Recall that the directed graph of an M × M complex matrix A := (a ij ) is given by the vertices {1, . . . , M } and edges defined as follows: there exists an edge from i to j if and only if a ij = 0. Note that if r is itself an SDD row, the trivial path r → r satisfies the requirement of (ii) in the above.The nonsingularity of WCDD matrices is proven in [27]. We provide an elementary proof for the convenience of the reader: Lemma 3.2....