🔢 mathematics

$\lambda$ -geodesic convexity of the squared sliced-Wasserstein distance

This paper establishes that the squared sliced-Wasserstein distance is $\lambda$ -geodesically convex on sets of measures with bounded densities under specific geometric conditions on the target measure, thereby extending convexity properties known in one dimension to higher dimensions and deriving consequences for the sliced-Wasserstein flow.

Original authors: Giacomo Cozzi, Filippo Santambrogio

Published 2026-10-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Giacomo Cozzi, Filippo Santambrogio

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to move a pile of sand from one shape to another. You want to do it with the least amount of effort, shifting every grain just enough to reach its new destination without wasting energy. This is the core idea behind a field of mathematics called optimal transport. For decades, mathematicians have studied how to measure the "distance" between two different distributions of mass, like two clouds of dust or two populations of data points. They have developed a precise way to calculate the cost of moving one shape into another, a tool that has become essential in everything from economics to machine learning. However, when the shapes exist in more than one dimension—when they are not just lines on a page but occupy a full space—this distance behaves in a tricky, unpredictable way. It does not always follow the smooth, straight-line logic that mathematicians prefer when they try to predict how a system will evolve over time.

To make these complex calculations manageable, researchers invented a shortcut called the sliced Wasserstein distance. Instead of trying to move the entire multi-dimensional shape at once, this method slices the shape into thin, one-dimensional strips, like cutting a loaf of bread. It calculates the cost of moving the sand along each of these individual slices and then averages the results. This approach is much faster and easier to compute than the traditional method, making it a favorite for modern computer algorithms. But there was a lingering question: does this shortcut preserve the smooth, predictable mathematical properties of the original, more difficult method? Specifically, does the "energy" of the system behave in a stable way as the shapes change?

In a recent study, Giacomo Cozzi and Filippo Santambrogio set out to answer this question. They focused on the mathematical behavior of the squared sliced Wasserstein distance, which acts as a measure of how far a current distribution is from a desired target. In the world of mathematics, a function is considered "convex" if it curves upward like a bowl, meaning that if you take a step toward a solution, you are guaranteed to get closer without hitting any unexpected bumps or valleys. This property is crucial because it ensures that algorithms designed to find the best solution will work reliably and not get stuck. The researchers discovered that while the traditional distance often fails to be convex in higher dimensions, the sliced version behaves much better, but only under specific conditions.

The team proved that the squared sliced Wasserstein distance is indeed "geodesically convex," a technical way of saying it curves upward along the most efficient paths between shapes, provided the target shape meets certain geometric criteria. They found that if the target distribution is not too concentrated in any single spot—meaning its density is bounded—and if it satisfies a particular condition regarding how its mass is distributed along different angles, the mathematical landscape becomes stable. This stability holds even when the starting shape has a high density of mass, as long as it does not exceed a certain limit. The researchers showed that the degree of this stability depends on the dimension of the space and the specific shape of the target, sometimes resulting in a very strong stability and other times in a weaker, though still useful, form of stability.

This finding has immediate consequences for how we understand the flow of these distributions over time. The sliced Wasserstein distance is often used to drive a process called the sliced Wasserstein flow, which is a continuous algorithm that gradually morphs a starting shape into a target shape. Because the researchers proved that the underlying distance function is convex under their conditions, they were able to demonstrate that this flow has a unique solution, provided the starting shape has a bounded density and the target shape satisfies specific geometric requirements. In simpler terms, if you start with a specific shape and let it evolve according to these rules, there is only one possible path it can take to reach the target. There are no branches or ambiguities; the path is determined and singular. This is a significant step forward, as it guarantees that the algorithms using this method will not produce conflicting results or behave erratically, as long as the initial and target distributions meet the necessary criteria.

The study also clarified the limits of this stability. The researchers noted that the stability is not infinite; it can weaken or even disappear if the starting shape becomes too dense or if the target shape does not meet the required geometric conditions. They provided examples of shapes that work well, such as those with uniform projections, and explained why certain high-dimensional shapes might fail to meet the criteria. By establishing these boundaries, the paper gives computer scientists and mathematicians a clear map of where this powerful tool can be safely applied. It confirms that the sliced Wasserstein distance is not just a computational trick, but a robust mathematical object with predictable behavior, opening the door for more reliable and efficient methods in data analysis and machine learning.

Technical Summary: $\lambda$ -Geodesic Convexity of the Squared Sliced Wasserstein Distance

Problem Statement
The paper addresses a fundamental limitation in optimal transport theory: the squared Wasserstein distance $W_2^2(\cdot, \nu)$ with respect to a fixed target measure $\nu$ is generally not geodesically convex on the space of probability measures $\mathcal{P}_2(\mathbb{R}^d)$ when the dimension $d \geq 2$ . While this function is 2-geodesically convex in dimension $d=1$ , the lack of convexity in higher dimensions complicates the analysis of gradient flows and the uniqueness of solutions.

To mitigate the computational cost of high-dimensional Wasserstein distances, the Sliced Wasserstein distance ( $SW_2$ ) has been introduced. It is defined by averaging one-dimensional Wasserstein distances over all projections of the measures onto lines. The central question of this work is whether the squared Sliced Wasserstein distance, $F_\nu(\rho) = \frac{1}{2}SW_2^2(\rho, \nu)$ , possesses $\lambda$ -geodesic convexity properties similar to the one-dimensional case, specifically on sets of measures with bounded density.

Methodology
The authors employ a combination of optimal transport theory, geometric analysis, and functional analysis to establish $\lambda$ -geodesic convexity. The methodology proceeds in three main stages:

Gradient Flow Identification: The paper first establishes that the Sliced Wasserstein Flow (SWF)—an isotropic, continuous-time evolution of a density $\rho$ towards a target $\nu$ —corresponds to the Wasserstein gradient descent of the functional $F_\nu$ . The velocity field $v[\rho]$ driving this flow is explicitly derived as the negative gradient of the first variation of $F_\nu$ .
Lipschitz Bounds on Velocity: The core technical strategy relies on Lemma 3.1, which states that if the velocity field $v[\rho] + \frac{1}{d}\text{id}$ is Lipschitz continuous with constant $L$ , then $F_\nu$ is $\lambda$ -geodesically convex with $\lambda = \frac{1}{d} - L$ . Consequently, the problem reduces to deriving uniform Lipschitz bounds for the velocity field $v[\rho]$ on sets of measures with bounded density.
Analysis of Projections: To bound the Lipschitz constant, the authors analyze the derivative of the optimal transport maps in the one-dimensional projections. They relate the derivative of the transport map $T_\vartheta$ $T_{ϑ}$ to the ratio of the source density projection $\rho_\vartheta$ $ρ_{ϑ}$ and the target density projection $\nu_\vartheta$ $ν_{ϑ}$ .
- Case 1 (Strong Assumptions): Initially, they consider the case where the projections of the target measure $\nu$ are uniformly bounded from below ( $\nu_\vartheta \geq c_0 > 0$ ). This allows for direct estimation but is shown to be a very restrictive condition, essentially limiting existence to dimensions $d \leq 3$ .
- Case 2 (General Assumptions): The authors then derive a more robust result (Theorem 4.4) requiring only that the target measure $\nu$ satisfies a specific geometric condition (Assumption 4.6). This condition imposes a lower bound on the density of $\nu$ near the boundary of its support, specifically $\nu_\vartheta(r) \geq \kappa (\min\{N_\vartheta(r), 1-N_\vartheta(r)\})^\alpha$ , where $N_\vartheta$ is the cumulative distribution of the projection.

Key Contributions and Results

Theorem 4.4 (Main Result): The paper proves that for a target measure $\nu$ satisfying Assumption (4.6) and a set of source measures $K = \{\rho \in \mathcal{P}_2(\mathbb{R}^d) : \|\rho\|_{L^\infty} \leq M\}$ , the functional $F_\nu$ is $\lambda$ -geodesically convex on $K$ . The convexity modulus is given by:
$\lambda = \frac{1}{d} - \frac{c(d)}{(1 - \frac{\alpha d}{d-1})\kappa} M^{1/d}$
where $c(d)$ is a constant depending only on the dimension $d$ , while $\kappa$ and $\alpha$ are parameters characterizing the target measure $\nu$ in Assumption (4.6). The authors note that $\lambda$ is generally non-positive and decreases as the density bound $M$ increases, indicating that convexity degenerates for high-density measures.
Geometric Conditions on Target Measures: The paper provides a detailed analysis of Assumption (4.6), demonstrating that it is satisfied by measures with densities bounded away from zero on smooth, uniformly convex domains (Example 4.10) and measures with specific power-law behaviors near the boundary of their support (Example 4.8).
Uniqueness of the Sliced Wasserstein Flow: As a direct consequence of the $\lambda$ -geodesic convexity result, the authors establish the well-posedness of the Sliced Wasserstein Flow. Specifically, Corollary 1.1 states that if $\nu$ satisfies (4.6) and the initial datum $\rho_0$ is in $L^\infty$ , there exists a unique solution $\{\rho_t\}_{t \geq 0}$ to the flow equation (1.3) such that the $L^\infty$ norm remains locally bounded.
Dimensional Constraints: The paper clarifies the geometric limitations of the "uniformly bounded projections" condition, showing via Remark 4.2 that such measures essentially only exist in dimensions $d \leq 3$ .

Significance and Claims
The paper claims that its primary significance lies in extending the theory of geodesic convexity to the Sliced Wasserstein distance, a metric widely used in machine learning and data science due to its computational efficiency. By proving $\lambda$ -geodesic convexity under realistic density bounds, the authors enable the application of the robust theory of Wasserstein gradient flows (developed in works like [AGS05] and [San17]) to the Sliced Wasserstein Flow.

Specifically, the authors highlight that this result:

Guarantees the uniqueness of the SWF solution, a property not previously established under these general conditions.
Implies the convergence rate of the discrete Jordan-Kinderlehrer-Otto (JKO) scheme to the continuous flow solution.
Allows for the use of the flow interchange technique in future studies involving $SW_2$ .

The authors remain modest regarding the scope of their convexity result, explicitly noting that $\lambda$ is often negative (implying weak convexity or even concavity in the standard sense) and that the convexity modulus deteriorates as the density bound $M$ grows. They do not propose new algorithms or experimental validations but rather provide the theoretical foundation necessary to rigorously analyze the behavior of existing sliced transport methods.

More like this