26.12.2013 Views

Adaptivity with moving grids

Adaptivity with moving grids

Adaptivity with moving grids

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

<strong>Adaptivity</strong> <strong>with</strong> <strong>moving</strong> <strong>grids</strong> 65<br />

norm is the least-squares norm given by<br />

∫<br />

I = |F(ξ,t) − ξ| 2 dξ. (3.39)<br />

Ω C<br />

Minimizing I subject to the equidistribution principle is in fact the celebrated<br />

Monge–Kantorovich problem from differential geometry. This principle<br />

is often called optimum transport, as it leads to a transformation, the<br />

creation of which takes a minimum amount of work as a deviation from the<br />

identity. Intuitively, this is likely to deliver a regular mesh, as this mesh<br />

will be as close (in an averaged sense) to the most regular possible mesh,<br />

i.e., a completely uniform one. Remarkably, this minimization problem has<br />

a unique solution <strong>with</strong> a very elegant expression for the transformation.<br />

Theorem 3.3. There exists a unique optimal mapping F(ξ,t) satisfying<br />

the equidistribution equation. This map has the same regularity as M.<br />

Furthermore, F(ξ,t) is the unique mapping from this class which can be<br />

written as the gradient (<strong>with</strong> respect to ξ) ofaconvex (mesh) potential<br />

P (ξ,t), so that<br />

F(ξ,t)=∇ ξ P (ξ,t), ∆ ξ P (ξ,t) > 0. (3.40)<br />

Proof. See Brenier (1991) or Caffarelli (1992, 1996) for an abstract proof<br />

and Delzanno et al. (2008) for a proof in the context of adaptive mesh<br />

generation.<br />

The following is then immediate.<br />

Lemma 3.4. The map F is irrotational so that ∇ ξ × F = 0, and the<br />

Jacobian of F is symmetric.<br />

Significantly, the transformation above is an example of a Legendre transformation<br />

(Sewell 2002). Such transformations include translations and<br />

linear maps by positive definite symmetric matrices. We now show how to<br />

calculate such a transformation.<br />

3.3.2. Properties of optimally transported meshes<br />

It is immediate that if x = ∇ ξ P then<br />

∂X<br />

= H(P ),<br />

∂ξ<br />

where H(P ) is the Hessian of P . Additionally, if the measure M ∈ W 2 (Ω P )<br />

is strictly positive on its supports (assumed to be convex), then the potential<br />

P ∈ Wloc 2 (Ω C), and satisfies, in the classical sense, the Monge–Ampère<br />

equation<br />

M(∇ ξ P, t)H(P )=θ(t). (3.41)

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!