6 JEAN-DAVID BENAMOU, BRITTANY D. FROESE, AND ADAM M. OBERMANlem:Op2dThe solution operator incorporates the Dirichlet boundary conditions (D).Remark. While it is possible to define the domain of the operator more carefully,for our purposes it is sufficient to observe the it is well-posed on H 2 (Ω).Lemma 1. Provided the solution, u, of (MA) is in H 2 (Ω), it is a fixed point of theoperator T .Proof. Let v be the solution of (MA, D, C). Since v is in H 2 (Ω), ∆v is in L 2 (Ω)and the Poisson equation∆u = |∆v| ,augmented by the Dirichlet boundary conditions (D), is well-posed. Inserting thesolution v into Definition 1 we obtain( √ )T [v] = ∆ −1 (∆v)2= ∆ −1 (|∆v|).Since v is convex, by (C), ∆v > 0, and consequently,Therefore, v is a fixed point of (M2).T [v] = ∆ −1 (∆v) = v.Method 2 consists of iterating u n+1 = T [u n ] by solving√M2 (M2) ∆u n+1 = (u n xx) 2 + (u n yy) 2 + 2(u n xy) 2 + 2f.along with the prescribed Dirichlet boundary conditions (D).2.3. Discussion of Method 1. In order to prove convergence of the finite differentmethod to viscosity solution of (MA), it is sufficient to prove that the method ismonotone. When written in the form (M1), it would be enough to show [19] thatthe value u i,j is a non-decreasing function of the neighboring values. However, thisis not the case for (M1). In addition, based on the discussion in [21] and [20] wedon’t expect that any narrow stencil method will be monotone and consistent forthe equation (MA).However, we can study the convexity requirement. Directional convexity alonggrid directions is the requirement that the solution satisfyu(x + v) + u(x − v)eq:convex (2.4) u(x) ≤2for all lattice (grid) directions v. For more details about enforcing convexity ofsolutions, using finite difference methods, we refer to [20].The directional convexity along the axes is satisfied by (M1).lem:conv1Lemma 2. The fixed point of (M1) satisfies the inequalities (2.4) for the grid directionsv ∈ {(1, 0), (0, 1)}.Proof. Using the notation of (2.2), assume without loss of generality that a 2 ≤ a 1 .Next, estimate u i,j from (M1), using the fact that f is non-negative,u i,j ≤ a 1 + a 2− a 1 − a 2= a 2 = u i,j+1 + u i,j−12 22to obtain convexity in the direction v = (1, 0). Using the first assumption in theproof, we also haveu i,j ≤ u i+1,j + u i−1,j,2□
TWO NUMERICAL METHODS FOR THE ELLIPTIC MONGE-AMPÈRE EQUATION 7which gives the convexity in the direction v = (0, 1).2.4. Discussion of Method 2. The method is convenient to implement because itsimply involves evaluating derivatives, and solving the Poisson equation, which canbe performed using the method of choice. However, the Poisson equation is solvedwith a source involving the norm of the Hessian of the previous solution. Whenthe solution is not in H 2 (Ω), this function can be unbounded. The regularityrequirement for solutions makes the the method is a model for the Finite Elementand Galerkin methods discussed in the introduction, because they require solutionsto be in H 2 (Ω) for convergence.□sec:overview3. Numerical Examples: OverviewIn this section we discuss the performance of the methods on a range of computationalexamples. These examples fall into three categories depending on theregularity of the solution.• Smooth or mildly singular solutions. These solutions are the mildest: weexpect any reasonable methods for <strong>Monge</strong>-Ampère will perform well. Inthese cases we find that (M2) is faster than (M1) by more than an order ofmagnitude.• Moderately singular solutions. These examples that are at the edge of thecapabilities of many of the existing methods. Both of our methods producesimilar results.• Strongly singular solutions. In this case many of the methods currentlyavailable either have not demonstrated convergence, or, as in the case of [9,15, 7, 8, 10], are known to fail. For these examples, at a given, finite,resolution, (M1) converges more quickly than (M2) by a couple of orders ofmagnitude. The effect is more substantial at the higher resolutions tested.The increasingly slow convergence rate for (M2) (both compared to (M1)and to (M2) on smoother solutions) is consistent with divergence.Regularity results can be used to determine when solutions are in H 2 (Ω), andthereby determine the method of choice. For strictly convex domains and smoothstrictly positive data, the solution is smooth [4]. However, as soon at f approacheszero, regularity breaks down, see [17] for W 2,p estimates on solutions.3.1. Comparison with previous works. There are four main examples computedby Dean and Glowinski. Feng and Neilan give computational examples forthe first three problems, but not the fourth. The first two, which are we callmild, are studied in §4.1,4.2. The third, for which no exact solution is available,but the solution is known to be singular, is studied in 5.3. The fourth solution isnot in H 2 (Ω). For this solution, the method of Dean and Glowinski diverges [8,p.1361], and Feng and Neilan did not present computational examples. While theconvegence rate is slower than for the smoother examples, both our methods (M1)and (M2) appear to converge for this example, §5.4.The solutions above we categorize as mild or moderate. In addition, we alsopresent computational results for more singular examples, §6.We compare the results of Dean and Glowinksi [9, 15, 7, 8, 10] to (M2), whichis the model for these methods. For the first three examples, these require roughlythe same number of iterations to converge. Although computation times were notgiven, presumably, each iteration of our method is more efficient, since it involves a