AMApplied Mathematics2152-7385Scientific Research Publishing10.4236/am.2014.517252AM-50345ArticlesComputer Science&Communications Engineering Physics&Mathematics A New Scheme for Discrete HJB Equations hanyongZou1*School of Mathematics and Statistics, Guangdong University of Finance &amp; Economics, Guangzhou, China* E-mail:yong_china@126.com091020140517264326492 August 201428 August 2014 10 September 2014© Copyright 2014 by authors and Scientific Research Publishing Inc. 2014This work is licensed under the Creative Commons Attribution International License (CC BY). http://creativecommons.org/licenses/by/4.0/

In this paper we propose a relaxation scheme for solving discrete HJB equations based on scheme II  of Lions and Mercier. The convergence of the new scheme has been established. Numerical example shows that the scheme is efficient.

Iterative Algorithm Relaxation Scheme HJB Equation Convergence Existence
1. Introduction

Consider the following Hamilton-Jacobi-Bellman (HJB) equation:

where is a bounded domain in are elliptic operators of second order. Equation (1.1) is arising in stochastic control problems. See  and the references therein.

Equation (1.1) can be discretized by finite difference method or finite element method. See   and the references therein. Then we obtain the following discrete HJB equation:

where. Equation (1.2) is a system of nonsmooth nonlinear equations. Many numerical algorithms for solving (1.2) have been proposed. See  - and the references therein.

 has given two iterative algorithms for solving (1.2). At each iteration, a linear complementarity subproblem or a linear equation system subproblem is solved. See also  .

Scheme I.

Step 1: Given for some we find such that

Step 2: Let For we find such that

Step 3: If then the output is otherwise and it goes to Step 2.

Assume Let

That is: the lth row of matrix is the lth row of matrix; the lth component of vector is the lth component of vector. Now we formulate Scheme II of Lions and Mercier in the notation above.

Scheme II.

Step 1: for some we find such that

Step 2: For we find such that

Step 3: Compute as the solution of

Step 4: If then the output is, otherwise and it goes to Step 2.

In the last decade many numerical schemes have been given for solving (1.2). But the above schemes are still playing a very important role. See  - and the references therein.

In this paper we propose, based on Scheme II above, a relaxation scheme with a parameter, which for is just Scheme II. In our numerical example, the new scheme with is faster than Scheme II. The monotone convergence of the new scheme has been proved.

2. New Scheme and Convergence

We propose a new scheme which is an extension of Scheme II.

New Scheme II.

Step 1: Given for some find such that

Step 2: For find such that

Step 3: Compute as the solution of

Step 4: Compute

Step 5: If then output otherwise and go to Step 2.

In  we proposed the following conditions for (1.2).

Condition All the matrices are M-matrices.

In  we have proved the following theorem.

Theorem 2.1 If Condition holds then (1.2) has a unique solution.

We have the following convergence theorem.

Theorem 2.2 Assume that Condition holds, and that are produced by New Scheme II. Then is monotonely decreasing and convergent to the solution of (1.2).

Proof Since all are M-matrices, in New Scheme II are well defined.

First, we prove is decreasing monotonically, i.e.,

By (2.3) we have

which combining with (2.1) and (2.2) yields

Since are M-matrices, (2.7) means

By (2.4) we obtain

By, (2.8) and (2.9) we know

and

which and (2.10) implies

Similarly, by (2.3) we derive

which combining with (2.2) and (2.6) implies

Hence we have

By (2.4), we have

By (2.12), (2.13) and, we know

which combining with and (2.11) we derive

By (2.11), (2.12) and (2.13) ,we get

which combining with (2.15) implies

It is easy to derive by induction that

and

It follows that (2.5) holds.

It follows from (2.2) and (2.3) that

Since the set is a finite set there exist positive integers and with such that

Therefore, we have

Then by (2.2) we obtain

which and (2.17) results in

From (2.4), (2.16) and (2.19) we have

It follows from (2.18), (2.19) and (2.20) that

which means is a solution of (1.2). The existence of solution has been proved.

Finally, we prove the uniqueness of solution. Assume and are solutions of (1.2), i.e.,

It is easy to see from (2.21) and (2.22) that there exist and such that

(2.23) and (2.26) implie. But (2.24) and (2.25) implies. Hence. The proof is complete. ,

3. Numerical Example

We use example 2 in  , i.e.,

where

The discretization of the above second order derivatives are:

where denote the forward and backward difference respectively in and, ,. We use New Scheme II to solve the discrete problem. Take, and 1.1, 1.3, 1.5, 1.8, 1.9 respectively.

Table 1 and Table 2 show the ∞-norm of the residual when iteration terminates.

We see that for and is big for.

Table 3 shows the relation between iteration number and relaxation number. Table 4 and Table 5 show the value of at for and respectively.

We can see from Table 3 that the algorithm for is faster than that for. Table 4 and Table 5 display the monotonicity of the algorithm.

∞-norm of the residual R
0.10.50.80.91.0
3.419e–0042.099e–0119.464e–0126.861e–0126.651e–012
6.630e–0031.784e–0086.653e–0116.062e–0118.169e–006
∞-norm of the residual R
1.11.31.51.81.9
3.440e–0002.314e+0014.670e+0018.421e+0019.730e–000
1.667e–0034.323e+0011.754e+0024.323e+0012.089e+002
Iteration number m
0.10.50.80.91.0
20019810790124
600495282258400
0.10.50.80.91.0
1.0914098001.0860339621.0820020831.0806581231.079314164
1.0897513771.0800227281.0748911941.0735338441.076283661
1.0882562931.0754499581.0720501611.0710728141.073086733
1.0867583641.0730600861.0694513021.0685869241.072407806
Last 1.0659639941.0658871091.0658871091.0658871091.065887109