Stability of optimal transport on metric measure spaces

This paper establishes the quantitative stability of Kantorovich potentials and optimal transport maps on metric measure spaces with lower Ricci curvature bounds, confirming a recent conjecture by Kitagawa, Letrouit, and Mérigot through a novel proof using heat kernel-regularized $c$-transforms that avoids reliance on linear structures or sectional curvature bounds.

The Gist

Here is an explanation of the paper "Stability of optimal transport on metric measure spaces" using simple language and creative analogies.

The Big Picture: Moving Mountains of Sand

Imagine you have two piles of sand. One pile is your starting point (Source), and the other is your destination (Target). Your goal is to move every grain of sand from the source to the destination in the most efficient way possible, minimizing the total distance traveled. This is the problem of Optimal Transport.

In the real world, this isn't just about sand. It's about:

• Moving data in AI.
• Redistributing wealth in economics.
• Matching supply to demand in logistics.

Mathematicians have known for decades how to solve this on "nice" surfaces (like a flat table or a smooth sphere). But what happens if the ground is bumpy, craggy, or even made of fractal shapes (like a snowflake)? This paper tackles that exact problem.

The Core Question: Is the Solution Stable?

The authors ask a crucial question: If I slightly change the destination pile of sand, does the plan to move the sand change drastically, or just a little bit?

• Unstable: A tiny shift in the destination causes the entire transport plan to flip upside down. (Imagine a house of cards collapsing with a gentle breeze).
• Stable: A tiny shift in the destination results in a tiny shift in the transport plan. (Imagine a heavy boulder rolling slightly when nudged).

The paper proves that even on very rough, non-smooth, and abstract mathematical landscapes, the transport plan is stable. If you nudge the target, the plan only wobbles a little.

The Problem: The "Rough Terrain"

In smooth worlds (like a smooth hill), mathematicians can use calculus (slopes and derivatives) to find the best path. But on "rough" spaces (like a craggy mountain range or a space with sharp corners), the ground is too jagged for standard calculus. The "slopes" don't exist everywhere.

Previous attempts to prove stability on these rough terrains relied on specific geometric rules (like "sectional curvature") that don't apply to all rough spaces. The authors wanted a proof that works for any space that has a "synthetic" lower bound on curvature (a fancy way of saying "it doesn't curve inward too sharply").

The Solution: The "Heat Kernel" Flashlight

To solve this, the authors invented a new tool: the Heat Kernel-Regulated c-transform.

Here is the analogy:
Imagine you are trying to find the shortest path through a dense, dark fog (the rough space). You can't see the ground clearly, so you can't calculate the slope.

1. The Old Way: Try to walk blindly and hope you don't fall off a cliff. (This is what previous methods tried to do on rough spaces).
2. The New Way (The Authors' Method): You turn on a heat lamp (the Heat Kernel).
   • Instead of looking at the sharp, jagged ground directly, the heat lamp "smears" the ground out. It turns the jagged rocks into a smooth, warm hill.
   • You solve the problem on this smooth, warm hill.
   • Then, you slowly turn off the heat lamp (let the temperature $t$ go to zero). As the fog clears, the smooth hill settles back into the jagged rocks, but because you solved the problem on the smooth version first, you know the solution won't jump wildly.

This "smearing" technique allows them to bypass the jagged edges of the math and prove that the solution remains stable even when the ground is rough.

The "John Domain" Rule

The paper also mentions a specific type of shape called a John Domain.

• Analogy: Imagine a cave system. A "John Domain" is a cave where, no matter where you are standing, you can always find a path to the exit that doesn't get too narrow or twisty.
• Why it matters: The authors prove that as long as your starting area (the source of the sand) is a "nice" cave (a John Domain) and your destination is compact, the stability holds. If the starting area is a weird, disconnected shape with dead ends, the math gets messy and stability might break.

The Results: What Did They Find?

1. For Potentials (The Map): They proved that the "map" (Kantorovich potential) used to guide the sand changes predictably. If you move the target by a tiny amount, the map changes by a tiny amount.
2. For Transport Maps (The Drivers): They extended this to show that the actual drivers (the optimal transport maps) are also stable. If you move the target, the drivers just take a slightly different route; they don't panic and drive in the opposite direction.

Why This Matters

This paper is a "universal translator" for math.

• It confirms a conjecture made by Kitagawa, Letrouit, and M´erigot.
• It works on Alexandrov spaces (spaces with curvature bounds but no smooth structure) and RCD spaces (spaces that behave like Riemannian manifolds but might be singular).
• It doesn't rely on the space being smooth or having a specific type of curvature.

In summary: The authors built a mathematical "heat lamp" that smooths out the jagged edges of rough, abstract worlds. By doing so, they proved that even in these chaotic, non-smooth environments, the most efficient way to move things remains stable and predictable. This gives scientists and engineers confidence that their optimization algorithms will work even when the data or the physical space they are modeling is imperfect or "rough."

Technical Summary

Here is a detailed technical summary of the paper "Stability of optimal transport on metric measure spaces" by Bang-Xian Han and Zhuo-Nan Zhu.

1. Problem Statement and Motivation

The Core Problem:
The paper addresses the quantitative stability of optimal transport (OT) in non-smooth geometric settings. Specifically, it investigates how the Kantorovich potentials (and consequently, optimal transport maps) change when the target probability measure is perturbed.

Context:

• In Euclidean spaces and smooth Riemannian manifolds, quantitative stability estimates (bounding the difference between potentials or maps by the Wasserstein distance between target measures) have been established recently (e.g., by Kitagawa, Letrouit, and Mérigot).
• The Conjecture: Kitagawa, Letrouit, and Mérigot conjectured that these stability results extend to general metric measure spaces with synthetic lower Ricci curvature bounds (specifically $\text{RCD}(K, N)$ spaces) and Alexandrov spaces (spaces with lower sectional curvature bounds).
• The Challenge: In non-smooth spaces, Kantorovich potentials lack the regularity (differentiability) required for standard linearization techniques used in smooth settings. Furthermore, the lack of a linear structure prevents the direct application of tools like the Hessian of the cost function.

Specific Constraints:
The authors assume the source measure $\rho$ is supported on a John domain $S$ with a density bounded away from zero and infinity ($a_1 m|_S \leq \rho \leq a_2 m|_S$). This condition is necessary to avoid instability caused by unbounded densities or non-open domains, as shown in previous counterexamples.

---

2. Methodology: Heat Kernel-Regularized $c$-Transform

The central innovation of this paper is the introduction of a heat kernel-regularized $c$-transform to bypass the low regularity of Kantorovich potentials in non-smooth settings.

Key Technical Steps:

1. Regularization via Heat Flow:
   Instead of working directly with the standard $c$-transform $\psi^c(x) = \inf_y \{c(x,y) - \psi(y)\}$, the authors define a regularized version using the heat kernel $p_t(x,y)$:
   $$\Phi_t[\psi](x) = -t \log \int_X e^{\frac{\psi(y)}{t}} p_{t/2}(x, y) \, dm(y)$$
   As $t \to 0$, $\Phi_t[\psi]$ converges to the standard $c$-transform (Varadhan's formula).

2. The Regularized Kantorovich Functional:
   They define a functional $K_t[\psi] = \int_S \Phi_t[\psi] \, d\rho$. The core of the proof relies on establishing the strong concavity of this functional.

3. Gradient and Hessian Estimates:

   • Gradient: The derivative of $K_t$ relates to the expectation of a test function under a probability measure $\mu_t$ derived from the heat kernel.
   • Hessian (Strong Concavity): The second derivative involves the variance of the test function. To prove strong concavity, the authors must bound this variance from below.
   • Local-to-Global Argument:
     • They first derive a local gradient estimate using heat kernel bounds and Li-Yau type estimates on $\text{RCD}(K, N)$ spaces.
     • They then globalize this estimate over the John domain $S$ using a Boman chain argument. This technique stitches together local Poincaré inequalities to control the global behavior of the potential, overcoming the lack of global convexity in the domain.

4. Handling Boundaries:
   A critical difficulty is that the target set $Y$ may have a boundary, which could disrupt the heat kernel regularization. The authors resolve this by utilizing the measure concentration property of the heat kernel and carefully constructing Lipschitz extensions of the potentials.

---

3. Key Contributions and Results

The paper provides two main theorems confirming the conjecture in different geometric settings.

Result 1: Stability of Kantorovich Potentials on $\text{RCD}(K, N)$ Spaces

Theorem 1.1:
Let $(X, d, m)$ be an $\text{RCD}(K, N)$ space. Let $S$ be a John domain and $\rho$ be a source measure with bounded density. For any two target measures $\mu, \nu$ supported on a compact set $Y$:
$$\|\phi_\mu - \phi_\nu\|_{L^1(\rho)} \leq C W_1^{1/2}(\mu, \nu)$$

• Significance: This establishes the first quantitative stability estimate for potentials in the general non-smooth $\text{RCD}$ setting. The exponent $1/2$ is consistent with known results in Euclidean spaces.
• Corollary 1.2: Applies to compact $\text{RCD}(K, N)$ spaces where the whole space is the domain.

Result 2: Stability of Optimal Transport Maps on Alexandrov Spaces

Theorem 1.3:
Let $(X, d)$ be an $n$-dimensional Alexandrov space with curvature bounded from below by $k$. If the source domain $S$ has finite perimeter, then for optimal transport maps $T_\mu, T_\nu$:
$$\int_S d^2(T_\mu(x), T_\nu(x)) \, d\rho(x) \leq C W_1^{1/6}(\mu, \nu)$$

• Significance: This extends stability results to Alexandrov spaces, which are the metric analogues of manifolds with lower sectional curvature bounds.
• Methodology: The proof lifts the problem to the unit tangent bundle and utilizes the geodesic flow. It exploits the one-dimensional convexity of the potentials along geodesics (semiconcavity) and uses a BV (Bounded Variation) estimate to control the number of times geodesics cross the boundary of the domain $S$.

---

4. Significance and Impact

1. Confirmation of a Major Conjecture: The paper definitively confirms the conjecture by Kitagawa, Letrouit, and Mérigot, proving that quantitative stability is a robust property of optimal transport that persists even in spaces lacking smooth structure.
2. New Tools for Non-Smooth Geometry: The heat kernel-regularized $c$-transform is a novel tool. Unlike previous methods that relied on linear structures or sectional curvature bounds, this approach works purely with synthetic curvature conditions ($\text{RCD}$) and heat flow properties.
3. Independence from Smoothness: The proof is new even in the smooth Riemannian setting, suggesting that the heat kernel regularization approach might offer a more robust or alternative perspective to classical linearization methods.
4. Applications: These results are foundational for:
   • Numerical Analysis: Justifying the convergence of numerical schemes for OT in non-smooth domains.
   • Machine Learning: Providing theoretical guarantees for Wasserstein-based metrics in data spaces that are not manifolds (e.g., graphs, fractals, or data with singularities).
   • Geometric Analysis: Deepening the understanding of the interplay between optimal transport, curvature, and heat flow in metric measure spaces.

Summary of Technical Flow

1. Regularize: Replace the non-smooth $c$-transform with a smooth heat-kernel version.
2. Analyze: Prove the regularized functional is strongly concave using local heat kernel estimates and Li-Yau inequalities.
3. Globalize: Use the Boman chain condition (specific to John domains) to extend local estimates to the global domain.
4. Limit: Pass to the limit $t \to 0$ to recover the stability of the original Kantorovich potentials.
5. Map Stability (Alexandrov): Lift to the tangent bundle, use geodesic flow and 1D convexity to relate potential stability to map stability, controlling boundary crossings via perimeter estimates.