ScalingStacks

4.1. Linear algebra of Legendre transform and Monge-Ampére equations [05CW]

Original official author HTML, exact retained edition. Historical TeX conversion verdicts remain unchanged. Cited-edition alignment and mathematical self-containment are not assessed.

Complete original source context · Original author HTML

4.1. Linear algebra of Legendre transform and Monge-Ampére equations

The classical fact, implicitly used in [GW00], is that solving the two-dimensional Laplace equation is equivalent to solving the real two-dimensional Monge-Ampère equation. This can be easily generalized to higher dimensions. Namely, the chain rule and elementary linear algebra for a particular coordinate transformation yields the following.

Lemma 4.1.

If K⁡(s,t)K(s,t) is a local solution to the split Monge-Ampère equation (17), then Ψ⁡(y)\Psi(y) is a (local) solution of the classical (real) Monge-Ampère equation

(19) det∂2Ψ∂yi​∂yp=1,1≤i,j≤n,\det\frac{\partial^{2}\Psi}{\partial y_{i}\partial y_{p}}=1,\quad 1\leq i,j\leq n,

where

(20) yi=∂K∂si, 1≤i≤k,yp=tp,k+1≤p≤n,y_{i}=\frac{\partial K}{\partial s_{i}},\ 1\leq i\leq k,\qquad y_{p}=t_{p},\ k+1\leq p\leq n,

and Ψ⁡(y)\Psi(y) is the partial Legendre transform of K⁡(s,t)K(s,t), defined by

(21) ∂Ψ∂yi=si, 1≤i≤k,∂Ψ∂yp=−∂K∂tp,k+1≤p≤n.\frac{\partial\Psi}{\partial y_{i}}=s_{i},\ 1\leq i\leq k,\qquad\frac{\partial\Psi}{\partial y_{p}}=-\frac{\partial K}{\partial t_{p}},\ k+1\leq p\leq n.
Proof.

First, we check the very existence of the partial Legendre transform. Consider the following Jacobian and Hessian matrices:

(22) ∂(yi,yp)∂(s,t)=(∂2K∂sj​∂si∂2K∂tq​∂si0𝟙)=(VB0𝟙),Hess⁡K⁡(s,t)=(VBBt−W).\frac{\partial(y_{i},y_{p})}{\partial(s,t)}=\left(\begin{array}[]{cc}\frac{\partial^{2}K}{\partial s_{j}\partial s_{i}}&\frac{\partial^{2}K}{\partial t_{q}\partial s_{i}}\\ 0&\mathbbm{1}\end{array}\right)=\left(\begin{array}[]{cc}V&B\\ 0&\mathbbm{1}\end{array}\right),\quad\operatorname{Hess}K(s,t)=\left(\begin{array}[]{cc}V&B\\ {}^{t}B&-W\end{array}\right).

Then

(23) ∂(s,t)∂(yi,yp)=(V−1−V−1​B0𝟙),Hess⁡Ψ⁡(y)=(V−1−V−1​Bt(−V−1B)W+Bt​V−1​B).\frac{\partial(s,t)}{\partial(y_{i},y_{p})}=\left(\begin{array}[]{cc}V^{-1}&-V^{-1}B\\ 0&\mathbbm{1}\end{array}\right),\quad\operatorname{Hess}\Psi(y)=\left(\begin{array}[]{cc}V^{-1}&-V^{-1}B\\ {}^{t}(-V^{-1}B)&W+{{}^{t}B}V^{-1}B\end{array}\right).

This shows that if both VV and WW are symmetric and positive definite, then Hess⁡Ψ\operatorname{Hess}\Psi is also a symmetric positive definite matrix, so there exists (locally) a convex function Ψ\Psi – the partial Legendre transform.

On the other hand the inverse of a non-degenerate 2×22\times 2 block matrix with detA≠0\det A\neq 0 and detD≠0\det D\neq 0 is:

(24) (ABCD)−1=((A−B​D−1​C)−1(−A+B​D−1​C)−1​B​D−1(−D+C​A−1​B)−1​C​A−1(D−C​A−1​B)−1),\left(\begin{array}[]{cc}A&B\\ C&D\end{array}\right)^{-1}=\left(\begin{array}[]{cc}(A-BD^{-1}C)^{-1}&(-A+BD^{-1}C)^{-1}BD^{-1}\\ (-D+CA^{-1}B)^{-1}CA^{-1}&(D-CA^{-1}B)^{-1}\end{array}\right),

which implies that

(25) det(ABCD)=1⟺detA−1=det(D−CA−1B).\det\left(\begin{array}[]{cc}A&B\\ C&D\end{array}\right)=1\quad\Longleftrightarrow\quad\det A^{-1}=\det(D-CA^{-1}B).

Applied to Hess⁡Ψ\operatorname{Hess}\Psi the last observation shows that Ψ\Psi is a local Monge-Ampère solution if and only if detV=detW\det V=\det W. ∎

Original mathematics by the credited authors. Source-backed reader collection; mathematical self-containment is not assessed.