Reconstruction of a convex body from its Gaussian curvature measure

in collaboration with Q. Mérigot



We describe below a method to reconstruct a 3D convex body from its Gaussian curvature measure based on a variational characterization based on optimal transport due to [V. Oliker (2007) and J. Bertrand (2010)]. All details can be found in our paper Discrete optimal transport: complexity, geometry and applications.

Let $K\subseteq R^d$ be a convex body in $R^d$, containing the origin in its interior. Any convex set admits an exterior unit normal vector field $ n_K: \partial K\to S^{d-1}$, which is uniquely defined almost everywhere. Let $\sigma$ be the probability measure on the unit sphere, obtained by rescaling the $(d-1)$-dimensional Hausdorff measure. The Gaussian measure $\mathcal{G}_K$ of $K$ is by definition the pullback of $\sigma$ by the Gauss map $n_K$. More explicitly,

\[ \forall B \subseteq \partial K,~ \mathcal{G}_K(B) := \sigma(n_K(B))\]

Since $K$ contains the origin in its interior, its boundary can be parameterized by a radial map $\rho_K: S^{d-1} \to \partial K$. For every direction $u$ in $S^{d-1}$, $\rho_K(u)$ lies in the intersection of $\partial K$ with the ray $\{tu; t \geq 0\}$. We can again pull-back the measure $\mathcal{G}_K$ by the map $\rho_K$, thus defining a measure on the unit sphere $\mathcal{G}_K^0$, which we will call Alexandrov measure.

\[ \forall B \subseteq S^{d-1},~ \mathcal{G}^0_K(B) := \sigma(n_K\circ \rho_K(B)).\]

Move your mouse on the pictures to animate

Reconstruction of a polytop (on the right) which Gaussian measure (on the left) is supported on three arc of circles

Alexandrov addressed the question of the existence and uniqueness (up to homotethy) of a convex body with prescribed Alexandrov measure $\mu$, under some conditions on $\mu$. The relationship between this reconstruction problem and a problem of optimal transport on the unit sphere for the cost $c(u,v) = -\log(\max(u . v,0))$ has been first remarked by Oliker, and then used by Bertrand to give a direct variational proof of Alexandrov theorem. Bertrand's version of Alexandrov's theorem says the following:

$ \texttt{\bf Theorem : }$ Given a probability measure $\mu$ on the unit sphere, there exists a convex body $K$ such that $\mathcal{G}^0_K = \mu$ if and only the following optimal transport problem between $\sigma$ and $\mu$ for the cost function $c(u,v) := -\log(\max(u.v,0))$ admits a solution with finite cost:

\[ \sup_{\phi,\psi} \int_{S^{d-1}} \phi(u)\, d\sigma - \int_{S^{d-1}}\psi(v)\, d\mu = \inf_{\pi} \int_{S^{d-1} \times S^{d-1}} c(u,v) \\, d \pi(u,v) \]

where the maximum is taken over functions $\phi,\psi$ satisfying the relation $\phi(u) - \psi(v) \leq c(u,v)$, and the infimum is taken over transport plans between $u$ and $v$.
Based on this observation we solve previous non-standard optimal transportation problem to reconstruct convex bodies from their Gauss measures.

Move your mouse on the pictures to animate

Reconstruction of OTO's polytop (on the right) from its Gaussian measure (on the left) which is uniform plus a singular linear part