Weak transport curvature, convex geodesic domains and continuous optimal maps

← All digests

Family 360 · Differential geometry
Importance 5/5 · Provisional editorial rating
Sources consulted 2026-10-26

1. The question

Imagine redistributing material across a curved closed surface or manifold. The initial and final densities are prescribed, and moving a unit of material from $x$ to $y$ costs half the squared geodesic distance. An optimal transport map chooses destinations to minimize total cost.

Even when both densities stay positive and bounded, existence of an almost-everywhere defined optimal map does not guarantee continuity. Nearby starting points might be sent to widely separated destinations. The geometry of the squared-distance function determines whether such tearing can occur.

The Ma–Trudinger–Wang condition, or MTW condition, is a curvature condition on that cost. Family 360 claims that its weak nonnegative form has two major consequences on every fixed smooth compact boundaryless manifold. First, every tangent injectivity domain is convex. Second, optimal maps between measurable densities with fixed positive upper and lower bounds have uniformly Hölder-continuous representatives and inverses.

The challenge is the cut locus, where squared distance ceases to be smooth. The papers allow conjugate cut endpoints rather than assuming that all relevant geodesics avoid them.

2. Background and essential definitions

The cost and its geodesic coordinates

Let $(M,g)$ be a smooth Riemannian manifold, and let $d$ denote geodesic distance. The cost is

$$c(x,y)=\frac12d(x,y)^2.$$

The exponential map $\exp_x$ takes an initial tangent vector $p\in T_xM$ to the point reached at time one by its geodesic. The open tangent injectivity domain is

$$I(x)=\{p:\ d(x,\exp_x(ap))=a|p|\text{ for some }a>1\}.$$

Thus the geodesic with initial vector $p$ remains minimizing slightly beyond its endpoint. In this domain, the minimizing geodesic is unique and the exponential map is nonsingular.

The closed minimizing domain is

$$G_x=\{p:\ d(x,\exp_xp)=|p|\}.$$

Its boundary can contain cut vectors. At a cut point, another shortest geodesic may compete, or the exponential map may become singular. The latter is conjugacy: infinitesimally different geodesics focus to the same endpoint.

Convexity of $I(x)$ means ordinary linear convexity inside the vector space $T_xM$. If two initial velocities lie in it, so does every straight interpolation between them. This is not a claim that the manifold itself is a convex subset of Euclidean space.

Weak MTW only tests transverse directions

In the papers’ convention, the MTW quantity is

$$\mathfrak S_{x,p}(\xi,\eta) =-\frac32\left.\frac{\partial^4}{\partial s^2\partial t^2} c\bigl(\exp_x(t\xi),\exp_x(p+s\eta)\bigr)\right|_{s=t=0}.$$

It measures how a second derivative of the cost changes as the destination velocity changes. Weak MTW requires

$$\langle\xi,\eta\rangle_x=0 \quad\Longrightarrow\quad \mathfrak S_{x,p}(\xi,\eta)\ge0 \qquad(p\in I(x)).$$

Only orthogonal direction pairs are tested, and the lower bound may be zero. No condition on nonorthogonal pairs or strictly positive lower bound is assumed. The derivatives are taken where the cost is smooth; the condition does not pretend that the fourth derivative exists on the cut locus.

Potentials and supporting functions

A cost potential has the form

$$u(x)=\max_{y\in M}\{-c(x,y)-v(y)\},$$

where $v$ is a continuous function. A term that attains the maximum is an active supporting function, sometimes called a supporting mountain. It lies below $u$ everywhere and touches it at the chosen point.

These potentials are semiconvex: locally, adding a sufficiently large quadratic makes them convex. Their ordinary subdifferential $\partial u(x)$ records slopes supporting them to first order. On a manifold vectors and covectors are identified using the metric. At a smooth contact, $\nabla u(x)$ is the initial minimizing velocity to the active target.

3. Development history and previous partial results

McCann’s polar factorization theorem supplied existence and almost-everywhere uniqueness of optimal maps for squared distance under suitable absolute-continuity hypotheses. Regularity is a separate issue.

Ma, Trudinger and Wang introduced a strict cost-curvature condition in their regularity theory; Trudinger and Wang developed its weak form, often called A3w. Loeper established the connection between this geometry, supporting functions and continuity. Kim and McCann supplied a geometric formulation of the supporting-mountain principle.

Villani proposed that weak MTW should force convexity of tangent injectivity domains. Figalli, Gallouët and Rifford proved the implication for nonfocal manifolds, where minimizing geodesics reach their cut points strictly before the first conjugate point. Loeper and Villani obtained related conclusions under stronger curvature and geometric assumptions. The possible coincidence of cut and conjugacy remained a significant obstruction.

For regularity, Caffarelli’s section and contact-volume methods were foundational. Figalli, Kim and McCann obtained Hölder continuity and injectivity for rough densities under weak MTW and appropriate smooth-cost, convexity and nondegeneracy assumptions. Rough densities themselves are therefore not the new feature here. The issue is uniform control on a compact manifold as transport graphs approach singular cost pairs.

Uniform estimates were already known in important settings, including strict MTW with additional assumptions and products of round spheres. Figalli, Rifford and Villani also established necessary conditions for transport continuity and a two-dimensional characterization involving weak MTW together with convex injectivity domains.

M1 claims to derive the missing convexity from weak MTW alone in the compact setting. M2 uses that foundation to obtain a common Hölder estimate across the entire class of densities with fixed positive bounds, without requiring uniform separation from conjugate pairs.

4. What the new manuscripts claim

Both papers assume $(M,g)$ is smooth, connected, compact, without boundary, and of dimension $n\ge2$, satisfying weak MTW as defined above.

M1: global geometric support without density assumptions

M1 claims that both $I(x)$ and $G_x$ are convex for every $x$. This is ordinary convexity, with no assertion that their boundaries are smooth or strictly convex.

More generally, every ordinary subgradient $p\in\partial u(x)$ of any cost potential is a minimizing velocity and supplies a global support:

$$u(x')\ge u(x)+c(x,\exp_xp)-c(x',\exp_xp) \qquad(x'\in M).$$

The distinction is that an ordinary first-order supporting slope is promoted to a genuine global squared-distance support, even at the minimizing boundary.

For each $0<t<1$, the map

$$F_t(x,p)=\exp_x(tp),\qquad (x,p)\in\Gamma_u:=\{(x,p):p\in\partial u(x)\},$$

is claimed to be a homeomorphism from the whole subgradient graph onto $M$. Intermediate infimum-regularized potentials are $C^{1,1}$, meaning differentiable with Lipschitz derivative, and the inverse pole and momentum maps have geometry-and-time-controlled Lipschitz bounds. Lifted sections of the dual potential gap are compact and convex.

These statements use no densities. The homeomorphism and intermediate regularity are asserted only for $0<t<1$. At time one, the support and minimizing-velocity conclusions remain, but an arbitrary potential can send many starting points to one target.

M2: uniform transport and inverse estimates

Fix constants $0<\lambda\le\Lambda<\infty$ for which the density class is nonempty. Let $\rho_0,\rho_1$ be measurable probability densities with

$$\int_M\rho_i\,d\mathrm{vol}=1, \qquad\lambda\le\rho_i\le\Lambda \quad\text{almost everywhere}.$$

The optimal map pushes $\rho_0\,d\mathrm{vol}$ to $\rho_1\,d\mathrm{vol}$: the mass arriving in each measurable target set equals its prescribed target mass.

M2 claims a homeomorphic representative $\widetilde T:M\to M$ of that almost-everywhere unique map, with

$$d(\widetilde T(x),\widetilde T(x'))\le C d(x,x')^\alpha,$$

$$d(\widetilde T^{-1}(y),\widetilde T^{-1}(y'))\le C d(y,y')^\alpha.$$

Here $\alpha>0$ and $C<\infty$ depend only on the fixed manifold and $\lambda,\Lambda$. They are common to all such density pairs, independently of density derivatives or oscillations. No continuity of the densities is assumed.

The exponent is obtained by contradiction and is not explicit. The theorem does not give uniform constants while the manifold or its metric varies. Combined with the prior necessity theorem, the source states an equivalence between weak MTW and transport continuity for all positive bounded density pairs in this setting.

5. Why the advance matters

Importance: 5/5, conditional on the claims. This would resolve Villani’s convexity conjecture in the stated compact setting and connect one weak local cost-curvature condition to global geodesic support and quantitative transport regularity.

The common estimate in M2 is stronger than separately proving continuity for each density pair. A sequence of increasingly irregular but uniformly bounded positive densities would retain the same Hölder control on the fixed manifold. The inverse estimate prevents uncontrolled compression as well as uncontrolled stretching.

The treatment of conjugate cut endpoints is central. At those points the usual smooth-coordinate formulas degenerate, and one cannot simply apply a smooth-cost theorem with constants silently depending on distance from the singular set.

The pair of papers also separates two sources of structure. M1 obtains support geometry without any measure bounds. M2 uses actual mass constraints to deduce uniqueness of final contact endpoints and regularity of the transport map.

6. The main idea

Continue the whole subgradient graph

Every ordinary subgradient can be expressed as a convex combination of finitely many active minimizing velocities. The potential may have infinitely many possible targets, but each local representation needs only finitely many. All minimizing velocities to each active cut target are retained.

M1 begins with short geodesic segments, where the geometry is controlled, and continues the projections $F_t$ toward time one. At a hypothetical first failing scale, it must exclude two mechanisms: a conjugacy obstruction and collisions of different points of the subgradient graph.

Radial shortening supplies smooth upper supports for distance and a strictly decreasing divided-action Hessian. Weak MTW controls only directions perpendicular to a velocity segment. The manuscript combines that transverse control with a radial Hessian identity and a Jacobi-field variation to exclude a conjugacy kernel at the first obstruction.

For collisions, it compares the finite active representations before passing to limits. Their first-order relations place relevant velocities in a hyperplane perpendicular to the collision direction. Linear terms are canceled at the finite stage before division by the squared separation. Weak MTW and strict radial comparison then provide the proposed contradiction.

Local injectivity, invariance of domain and a homotopy from the short-time map make the global projection one-to-one and onto. Closure as $t\uparrow1$ gives global supporting mountains. Importantly, closure does not turn the time-one projection of every arbitrary potential into a homeomorphism.

Measure how large a near-contact section can be

For a dual pair $(u,v)$, define the lifted gap section

$$S_x(r)=\left\{p\in G_x: u(x)+\frac{|p|^2}{2}+v(\exp_xp)\le r\right\}.$$

The gap is nonnegative and vanishes at exact transport contacts. Thus a small $r$ selects almost-optimal destinations from the starting point $x$.

M1 gives convexity of these lifted sets. M2 studies the oscillation of $v$ on their endpoint images, and takes the supremum over all density pairs in the fixed class. A power bound for this one scalar quantity yields a power bound for endpoint diameter. At zero height, the endpoint must then be unique.

Taking the supremum from the outset is how the argument aims to obtain class-uniform constants, rather than a separate small scale for each pair.

Control degeneration without assuming cut avoidance

Near a conjugate endpoint, the differential of $\exp_x$ can have small singular values. The manuscript factors the relevant Jacobi matrix through an intermediate joining point and a broken-action Hessian. Radial shortening adds a positive term, and transverse MTW supplies comparisons along affine velocity segments.

These estimates compare the degeneracies at related endpoints rather than demanding that all exponential Jacobians have a fixed positive lower bound. This distinction allows the analysis to approach conjugate configurations.

If every power bound for section oscillation failed, the manuscript finds two separated gap scales where that oscillation barely changes. Modified supporting functions then create a controlled interior contact configuration. Lower quadratic tests at selected endpoints transfer information between its center and outer contacts.

The order of this transfer matters. Contact-volume estimates using the original densities first exclude conjugacy at a selected outer endpoint and control its determinant. Only after obtaining the needed nonsingularity does the argument use the almost-everywhere density equation at sampled center endpoints. It does not assign a smooth density equation to an arbitrary singular point.

A determinant gain contradicts the geometric Jacobian comparisons and the fixed mass bounds. This yields a positive power estimate for the sections. Reversing the density pair gives the corresponding result for the inverse map, and compactness extends the local estimates globally.

7. Scope, limitations, and open questions

The claims concern squared geodesic distance on a fixed smooth compact connected manifold without boundary. They are not assertions for arbitrary costs, noncompact manifolds or domains with boundary.

Weak MTW is not replaced by ordinary nonnegative sectional curvature. It is the stated fourth-order cost condition on transverse pairs. No strict MTW inequality, nonfocality or prior convexity of injectivity domains is assumed.

M2 requires both densities to be bounded above and bounded away from zero over the whole manifold. Removing the lower bound, allowing singular measures, or asking for Lipschitz or smooth transport changes the problem. No explicit optimal exponent or uniformity across varying metrics is claimed.

The density-free intermediate homeomorphism in M1 and the final transport homeomorphism in M2 should also remain distinct. The latter uses mass bounds that an arbitrary potential does not possess. The manuscript arguments and claimed consequences are reported without independent proof verification.

8. Sources and further reading

M1 provides the convexity, support and intermediate geometry used as an explicit input to M2. M2 develops the class-uniform section estimate and explains the order of its limiting argument near conjugacy.

Primary predecessors include Figalli–Gallouët–Rifford on nonfocal injectivity domains, Figalli–Kim–McCann on Hölder optimal maps, Figalli–Rifford–Villani on continuity criteria, and Villani’s formulation of the geometric conjecture.

Source manuscripts

• M1. OpenAI, Global Support and Convex Injectivity Domains under Weak MTW (September 25, 2026).
• M2. OpenAI, Uniform Bi-Holder Transport from Weak MTW (September 25, 2026).

Sources consulted on 26 October 2026 at repository revision 1581d4c. Download the manuscript citations.