Quantitative stability for the Brascamp-Lieb inequality and moment measures

This paper establishes a sharp quantitative stability version of the Brascamp-Lieb inequality with a stability constant independent of the convex function, alongside uniform stability results for moment measures, by leveraging recent sharp stability versions of the Prékopa-Leindler inequality.

The Gist

Imagine you are trying to bake the perfect cake. You have a recipe (a mathematical formula) that tells you exactly how much flour, sugar, and eggs to use to get a specific texture. In the world of mathematics, this "recipe" is called a Moment Measure, and the "texture" is a specific shape or distribution of data.

For a long time, mathematicians knew that if you followed the recipe perfectly, you'd get the right cake. But they struggled with a tricky question: What happens if your recipe is slightly off? If you accidentally add a tiny bit too much sugar, does the cake collapse completely, or does it just taste a little sweeter? Can we predict exactly how much the cake will change based on how much sugar you added?

This paper by João Miguel Machado and João P. G. Ramos is like a new, ultra-precise ruler for bakers. It provides a way to measure exactly how much the final "cake" (the solution) will wobble if your "recipe" (the input data) is slightly imperfect.

Here is a breakdown of their discovery using everyday analogies:

1. The Core Problem: The "Wobbly Cake"

In mathematics, there is a famous rule called the Brascamp-Lieb Inequality. Think of this as a law of physics for shapes. It says that if you have a certain type of smooth, bowl-shaped hill (a convex function), there is a specific way to measure the "spread" of points on that hill.

The authors wanted to know: If we are slightly off from the perfect shape, how far off are we?

• Old way: Previous math tools could tell you that you were off, but they were like a blurry map. They didn't work well for all types of hills, and the "error bars" were huge.
• The new way: The authors created a sharp, uniform ruler. They proved that no matter what kind of smooth hill you are standing on, if you are slightly off, you can calculate exactly how far you are from the perfect spot.

2. The Secret Ingredient: The "Stability Principle"

To build this ruler, the authors used a clever trick involving a different mathematical rule called the Prékopa-Leindler inequality.

Imagine you are trying to balance a stack of plates.

• If the stack is perfect, it's stable.
• If you tilt it slightly, it wobbles.
• The authors found a way to measure that wobble so precisely that they could predict exactly how much the stack would fall over.

They realized that this "wobble measurement" (stability) works uniformly. It doesn't matter if your hill is steep, flat, wide, or narrow; the ruler works the same way. This is a huge deal because, in the past, mathematicians had to use different rulers for different shapes, and none of them worked perfectly for the complex problems they were trying to solve.

3. The Big Application: "The Moment Measure"

Now, let's connect this to the main goal: Moment Measures.

Think of a Moment Measure as a "shadow" cast by a 3D object.

• You have a cloud of data points (the object).
• You want to find a specific mathematical "lens" (a convex potential) that, when you shine light through it, casts a shadow that perfectly matches your data.

The paper asks: If my data points are slightly noisy (like a shaky camera taking a photo of the shadow), how much does my "lens" have to change to fix it?

The authors' new ruler allows them to say: "If your data is off by X amount, your lens only needs to shift by Y amount."

• Why this matters: In real life, data is always noisy. Sensors fail, measurements have errors, and computers have rounding issues. Knowing that the solution is "stable" means that if you have a slightly bad input, your computer won't spit out a completely garbage answer. It will give you an answer that is predictably close to the truth.

4. The "Magic" of Uniformity

The most exciting part of this paper is Uniformity.

Imagine you are a carpenter. You have a hammer that works great on soft wood, but if you hit a piece of oak, it breaks. You need a different hammer for every type of wood. That was the old math.

The authors built a universal hammer. It works on soft wood, oak, steel, and even glass. In math terms, their stability estimate works for any convex function, regardless of how weird or complex it is. This "one-size-fits-all" tool is what allows them to solve problems that were previously impossible.

5. Real-World Uses

The paper shows how this new tool helps in three specific scenarios:

• The Compact Box (Limited Space): Imagine trying to fit a puzzle into a small box. The authors show that if the box is small and tight, you can predict the puzzle's fit with extreme precision. This is great for computer simulations where we limit data to a specific area.
• The Regularization (The "Smoothing" Trick): Sometimes, math problems are too messy to solve directly. So, we add a little "smoothing" ingredient (regularization) to make them easier. The authors calculated exactly how much the solution changes as you add or remove this smoothing. It's like knowing exactly how much water to add to dough to get the perfect consistency.
• The Whole World (Infinite Space): Finally, they tackled the hardest case: data spread out over the entire universe (infinite space). They found that if the data is "spread out enough" (not squashed into a flat line), their ruler still works. This is crucial for things like sampling (generating random numbers for AI or physics simulations). If you know the solution is stable, you can trust your AI to generate realistic data even if the starting point isn't perfect.

Summary

In simple terms, this paper is about trust.

Mathematicians often worry that if you tweak a formula slightly, the answer might explode into chaos. Machado and Ramos have proven that for a very important class of problems (Moment Measures), the answer is robust. They built a precise, universal tool to measure exactly how much the answer changes, ensuring that even with imperfect data, the mathematical "cake" will still turn out deliciously close to perfection.

This is a major step forward for fields like Optimal Transport (moving things efficiently), Machine Learning (training AI models), and Physics (understanding how matter distributes itself).

Technical Summary

1. Problem Statement

The paper addresses the quantitative stability of the moment-measure correspondence and the Brascamp-Lieb (BL) variance inequality.

• Moment Measures: Given a Radon probability measure $\mu$ on $\mathbb{R}^d$, a moment measure representation seeks a convex potential $\psi$ such that $\mu = (\nabla \psi)_\# \varrho_\psi$, where $\varrho_\psi \propto e^{-\psi}$ is a Gibbs measure. This problem is central to optimal transport, convex geometry, and statistical sampling.
• The Stability Question: If two measures $\mu$ and $\nu$ are close (in terms of Wasserstein distance), are their associated optimal potentials $\psi_\mu, \psi_\nu$ and Gibbs measures $\varrho_\mu, \varrho_\nu$ also close?
• The Gap: While the existence and uniqueness (up to translation) of moment measures are well-established, quantitative stability estimates (bounds on the distance between solutions in terms of the distance between inputs) have been lacking, particularly those that are uniform over broad classes of convex functions. Previous results often depended on specific properties of the potential or lacked sharp exponents.

2. Methodology

The authors develop a novel framework linking the stability of the Prékopa-Leindler inequality to the Brascamp-Lieb variance inequality, and subsequently to the moment measure problem.

A. Linearization and Prékopa-Leindler Stability

The core strategy involves a linearization argument (inspired by Bobkov and Ledoux) applied to the Prékopa-Leindler inequality.

• They utilize recent sharp stability results for the Prékopa-Leindler inequality (specifically from Figalli, van Hintum, and Tiba) which provide $L^1$-stability bounds with a sharp exponent of $1/2$.
• By considering perturbations of the form $e^{-\phi + \delta f}$, they relate the deficit in the Prékopa-Leindler inequality to the deficit in the Brascamp-Lieb inequality.

B. The Brascamp-Lieb Variance Inequality

The paper establishes a sharp, uniform quantitative stability for the Brascamp-Lieb inequality:
$$\delta_{BL}(f) = \int_{\mathbb{R}^d} \langle (D^2\phi)^{-1}\nabla f, \nabla f \rangle d\varrho_\phi - \text{Var}_{\varrho_\phi}(f) \geq 0$$

• Key Innovation: The stability bound is uniform with respect to the convex potential $\phi$. The constant $C_d$ depends only on the dimension, not on the specific geometry of $\phi$.
• Technique: The proof handles the general case (where $\phi$ may not be strongly convex) by using an approximation argument involving quadratic regularization ($\phi_\alpha = \phi + \frac{\alpha}{2}|x|^2$) and leveraging the non-degeneracy of the support of moment measures (Lemma 2.3).

C. Transfer to Moment Measures

The authors use the uniform BL stability to control the variational functionals associated with moment measures:

• Primal Functional: $E_\mu(\varrho) = \int \log \varrho \, d\varrho + T(\varrho, \mu)$ (Entropy + Optimal Transport cost).
• Dual Functional: $J_\mu(\phi) = \log \int e^{-\phi^*} - \int \phi \, d\mu$.
• They show that the deficit in $J_\mu$ (or $E_\mu$) controls the distance of the solution to the manifold of optimizers via the BL stability.

3. Key Contributions and Results

I. Sharp Uniform Stability for Brascamp-Lieb (Theorem 1.1)

They prove that for any locally Lipschitz function $f$:
$$\text{dist}_{L^1(\varrho_\phi)}(f, \mathcal{O}_{BL}) \leq C_d \sqrt{\delta_{BL}(f)}$$
where $\mathcal{O}_{BL}$ is the manifold of affine functions $a \cdot \nabla \phi + b$.

• Significance: The exponent $1/2$ is sharp, and the constant $C_d$ is independent of the potential $\phi$. This uniformity is crucial for applying the result to nonlinear variational problems.

II. Quantitative Stability for Moment Measures (Theorem 1.2)

For a measure $\mu$ with barycenter at the origin and full-dimensional support, and any convex $\phi$:
$$\text{dist}_{L^1}(\varrho_{\phi^*}, \mathcal{M}_\mu) \leq C_d \sqrt{J_\mu(\bar{\phi}) - J_\mu(\phi)}$$
where $\bar{\phi}$ is the maximizer and $\mathcal{M}_\mu$ is the set of minimizers of the energy.

• This establishes a Polyak-Lojasiewicz type inequality for the moment measure functional, quantifying how close a sub-optimal potential is to the true solution.

III. Applications and Specific Stability Regimes

1. Compact Domains (Theorem 1.3):
For measures supported on a compact convex set $\Omega$, the stability is strong:
$$\text{dist}_{L^1}(\varrho_{\phi_\mu}, \varrho_{\phi_\nu}) \leq C_{d,\Omega} W_1(\mu, \nu)^{1/2}$$
This allows controlling strong norms ($L^1$) using weak topologies ($W_1$), which is vital for numerical applications.

2. Quadratic Regularization (Theorem 1.4):
They analyze the convergence of regularized moment measures (introduced in [DF25]) as the regularization parameter $\alpha \to 0$.
$$\text{dist}_{L^1}(\varrho_\alpha, \mathcal{M}_\mu) \leq C_{d,\mu} \alpha^{1/2}$$
This provides explicit convergence rates for regularized solutions to the classical moment measure problem.

3. Stability over $\mathbb{R}^d$ (Theorems 1.5, 5.6, 5.12):
To extend stability to the whole space, the authors identify two classes of measures where the second moments of the Gibbs measures are uniformly bounded:

• Class $K_\vartheta$: Measures with a lower bound on the directional first moment $\Theta(\mu) \geq \vartheta$ and supported on a compact set.
• Class $K_\Lambda$: Measures $\mu \propto e^{-V}$ where the curvature is bounded above ($D^2 V \leq \Lambda I$).
  • Novelty: They prove via a bootstrap argument using Caffarelli's contraction theorem that if $\mu \in K_\Lambda$, the optimal potential $\phi^*_\mu$ is $\Lambda^{-1/3}$-strongly convex.
• Result: For these classes, they derive stability in the $W_2$ distance:
  $$\text{dist}_{W_2}(\mathcal{M}_\mu, \mathcal{M}_\nu) \leq C W_2(\mu, \nu)^{\frac{p-2}{6p-4}}$$
  for any $p > 2$. The exponent approaches $1/6$ as $p \to \infty$.

4. Significance and Impact

1. Completing the Program: The paper completes the program of quantitative stability for the Brascamp-Lieb variance inequality by providing a result that is both sharp in exponent and uniform in the potential, a combination absent in prior literature.
2. Bridging Inequalities and Variational Problems: It successfully demonstrates how stability in functional inequalities (BL, Prékopa-Leindler) can be transferred to the stability of solutions in nonlinear variational problems (Moment Measures).
3. Numerical Implications: The results for compact domains and regularized measures provide rigorous error bounds for numerical algorithms used in sampling and generative modeling (e.g., Langevin dynamics, optimal transport solvers).
4. Sampling Applications: The discovery that measures with bounded curvature ($K_\Lambda$) induce strongly convex potentials ($\Lambda^{-1/3}$) implies that sampling algorithms (like Langevin dynamics) applied to the dual Gibbs measure will enjoy exponential convergence rates, even if the original measure $\mu$ only has polynomial convergence guarantees.
5. Geometric Insight: The work deepens the understanding of the geometry of moment measures, specifically the relationship between the curvature of the input measure and the strong convexity of the optimal transport potential.

In summary, this paper provides a robust, unified theory for the stability of moment measures, leveraging sharp functional inequalities to solve open problems in optimal transport and convex analysis.