A dimension-free interpolation of Caffarelli’s contraction theorem

Bader Ammari, Alessio Figalli · 2026

Read A dimension-free interpolation of Caffarelli’s contraction theorem by Bader Ammari, Alessio Figalli (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-24443.txt.

Introduction

Ammari and Figalli prove global Lipschitz bounds for Brenier maps between chance measures with weights from a knob family ranging from the size to endless. Source and target may use mixed knobs, with the source no larger than the target. Their global bound stays even in size and both knobs, and for matching finite knobs it removes the fast size link of prior bounds. Local bounds inside fixed balls stay stable as the knobs tend to endless, and they regain Caffarelli's shrink rule with sharp constants.

Catalog record

CREATORBader Ammari, Alessio Figalli
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE34112
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.24443

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

A dimension-free interpolation of Caffarelli’s contraction theorem Bader Ammari, Alessio Figalli Abstract . We prove global Lipschitz estimates for Brenier maps between probability measures on ℝ n \mathbb{R}^{n} whose densities belong to the family ρ U , p = Z U , p − 1 ​ exp ⁡ ( − Θ p ​ ( U ) ) , Θ p ​ ( t ) = p ​ log ⁡ ( 1 + t p ) , p ∈ [ n , + ∞ ] , \rho_{U,p}=Z_{U,p}^{-1}\exp(-\Theta_{p}(U)),\qquad\Theta_{p}(t)=p\log\Bigl(1+\frac{t}{p}\Bigr),\qquad p\in[n,+\infty], with finite normalization constant Z U , p Z_{U,p} , and with the convention Θ ∞ ​ ( t ) = t \Theta_{\infty}(t)=t . We allow different parameters for source and target, d , D ∈ [ n , + ∞ ] d,D\in[n,+\infty] , with d ≤ D d\leq D . Our global estimate is uniform in n , d , D n,d,D , and in the case d = D < + ∞ d=D<+\infty , it improves the bounds of [ 5 ] by removing their exponential dependence on the dimension. We also prove localized estimates inside fixed balls B R B_{R} whose constants are stable under the limits d , D → + ∞ d,D\to+\infty and they allow us to recover Caffarelli’s celebrated contraction theorem with sharp constants. 2020 Mathematics Subject Classification: 35J96 (primary), 35B65, 35A23 (secondary). Keywords: Monge-Ampère equation, Optimal transport, Caffarelli’s contraction theorem. 1. Introduction The L^{2} optimal transport (or Brenier-McCann) map between two absolutely continuous probability measures \mu=f(x)\,dx and \nu=g(y)\,dy is the unique solution of the Monge problem for quadratic cost: \min\left\{\int_{\mathbb{R}^{n}}|x-T(x)|^{2}f(x)\,dx:T_{\#}\mu=\nu\right\}. Here T_{\#}\mu=\nu means that \mu(T^{-1}(A))=\nu(A) for every Borel set A\subset\mathbb{R}^{n} . By the Brenier-McCann theorem (see [ 3 , 13 ] ), the solution of this quadratic optimal transport problem is the gradient of a convex function, T=\nabla\varphi . This mapping is a fundamental object in optimal transportation theory and has found applications in numerous fields such as functional and geometric inequalities, probability, and machine learning (see [ 10 , 1 , 2 ] ). In [ 4 ] , Caffarelli proved that if f=e^{-V} and g=e^{-W} for some V,W\in C_{\mathrm{loc}}^{1,1}(\mathbb{R}^{n}) with D^{2}V\leq\mathbf{C}_{V}^{(2)}\,\mathrm{Id},\qquad D^{2}W\geq\mathbf{c}_{W}^{(2)}\,\mathrm{Id} almost everywhere, where 0<\mathbf{c}_{W}^{(2)},\mathbf{C}_{V}^{(2)}<+\infty , then the Brenier-McCann map T pushing f\,dx forward to g\,dy is globally Lipschitz with (1.1) \|DT\|_{L^{\infty}(\mathbb{R}^{n})}\leq\sqrt{\mathbf{C}_{V}^{(2)}/\mathbf{c}_{W}^{(2)}}. Caffarelli’s contraction theorem is notable for its dimension-free Lipschitz bound Equation 1.1 . This dimension independence allows one to transfer functional inequalities, such as log-Sobolev and Poincaré inequalities, via the Lipschitz transport map between the two measures. For further references, extensions, alternative proofs, and applications, we refer to [ 12 , 11 , 7 , 14 , 6 , 8 ] . The purpose of this paper is to obtain analogous estimates for a family of probability measures interpolating between polynomial-type densities and log-concave densities. Our global estimate is uniform not only in the dimension n , but also in the interpolation parameters. For p\in[n,+\infty) and a potential U>-p , set \Theta_{p}(U):=p\log\Bigl(1+\frac{U}{p}\Bigr),\qquad Z_{U,p}:=\int_{\mathbb{R}^{n}}e^{-\Theta_{p}(U)}\,dx, whenever the integral is finite, and define \rho_{U,p}:=Z_{U,p}^{-1}e^{-\Theta_{p}(U)}=Z_{U,p}^{-1}\Bigl(1+\frac{U}{p}\Bigr)^{-p}. We also set \Theta_{\infty}(U)=U , Z_{U,\infty}:=\int_{\mathbb{R}^{n}}e^{-U}\,dx , and \rho_{U,\infty}=Z_{U,\infty}^{-1}e^{-U} . Since \Theta_{p}(t)\to t as p\to+\infty , the family \rho_{U,p} interpolates between the densities considered in [ 5 , 9 ] and the log-concave densities appearing in Caffarelli’s theorem. For p\in[n,+\infty) , we use the global shorthand \displaystyle\mathbf{C}_{V,p}^{(1)} \displaystyle:=\Biggl\|\frac{|\nabla V|}{\sqrt{p}+|\cdot|}\Biggr\|_{L^{\infty}(\mathbb{R}^{n})}^{2},\qquad \displaystyle\mathbf{C}_{W,p}^{(1)} \displaystyle:=\Biggl\|\frac{|\nabla W|}{\sqrt{p}+|\cdot|}\Biggr\|_{L^{\infty}(\mathbb{R}^{n})}^{2}, \displaystyle\mathbf{c}_{V,p}^{(0)} \displaystyle:=\Biggl\|\frac{p+|\cdot|^{2}}{p+V}\Biggr\|_{L^{\infty}(\mathbb{R}^{n})}^{-1},\qquad \displaystyle\mathbf{c}_{W,p}^{(0)} \displaystyle:=\Biggl\|\frac{p+|\cdot|^{2}}{p+W}\Biggr\|_{L^{\infty}(\mathbb{R}^{n})}^{-1}, \displaystyle\mathbf{C}_{V,p}^{(0)} \displaystyle:=\Biggl\|\frac{p+V}{p+|\cdot|^{2}}\Biggr\|_{L^{\infty}(\mathbb{R}^{n})},\qquad \displaystyle\mathbf{C}_{W,p}^{(0)} \displaystyle:=\Biggl\|\frac{p+W}{p+|\cdot|^{2}}\Biggr\|_{L^{\infty}(\mathbb{R}^{n})}. When a local bound is needed, we write \displaystyle\mathbf{C}_{W,p}^{(1)}(R) \displaystyle:=\Biggl\|\frac{|\nabla W|}{\sqrt{p}+|\cdot|}\Biggr\|_{L^{\infty}(B_{R}(0))}^{2},\qquad \displaystyle\mathbf{c}_{V,p}^{(0)}(R) \displaystyle:=\Biggl\|\frac{p+|\cdot|^{2}}{p+V}\Biggr\|_{L^{\infty}(B_{R}(0))}^{-1}, \displaystyle\mathbf{C}_{V,p}^{(0)}(R) \displaystyle:=\Biggl\|\frac{p+V}{p+|\cdot|^{2}}\Biggr\|_{L^{\infty}(B_{R}(0))},\qquad \displaystyle\mathbf{C}_{W,p}^{(0)}(R) \displaystyle:=\Biggl\|\frac{p+W}{p+|\cdot|^{2}}\Biggr\|_{L^{\infty}(B_{R}(0))}. Here and below, we use the convention B_{+\infty}(0)=\mathbb{R}^{n} , and omit the argument when R=+\infty . At the endpoint p=+\infty , for R<+\infty , we use the conventions \begin{gathered}\mathbf{c}_{V,\infty}^{(0)}(R)=\mathbf{C}_{V,\infty}^{(0)}(R)=\mathbf{c}_{W,\infty}^{(0)}(R)=\mathbf{C}_{W,\infty}^{(0)}(R)=1,\\ \mathbf{C}_{V,\infty}^{(1)}=\mathbf{C}_{W,\infty}^{(1)}=0.\end{gathered} Inequalities such as U>-p are understood as void when p=+\infty . We shall allow different source and target parameters, denoted by d and D . More precisely, we will study the monotone transport map from \rho_{V,d} to \rho_{W,D} , where d,D\in[n,+\infty],\qquad d\leq D. With this notation, the polynomial-to-polynomial case studied in [ 5 , 9 ] corresponds to D=d<+\infty , while the case D=d=+\infty corresponds to Caffarelli’s setup. Remark 1.1 . The lower bound n\leq d is natural for the maximum-principle argument: after taking the Monge-Ampère equation to the power 1/d , the concavity of \det^{1/d} is available. The restriction d\leq D is also natural from the point of view of tails. Indeed, if n\leq D<d<+\infty and V(x)=W(x)=|x|^{2} , then the radial Brenier map T(x)=t(|x|)x/|x| from \rho_{V,d} to \rho_{W,D} satisfies Z_{W,D}^{-1}\int_{t(r)}^{+\infty}\frac{s^{n-1}}{(1+s^{2}/D)^{D}}\,ds=Z_{V,d}^{-1}\int_{r}^{+\infty}\frac{s^{n-1}}{(1+s^{2}/d)^{d}}\,ds. Thus, t(r)\simeq r^{(2d-n)/(2D-n)} as r\to+\infty , which is superlinear when d>D . In particular, in this range, no global Lipschitz bound can hold in general for T . We first state a global Hessian estimate for the polynomial-to-polynomial case considered in [ 5 ] . Its constant is uniform in n,d,D . To keep the statement readable, we collect the following global structural quantities, understood in the extended sense; if they make C_{\mathrm{glob}}=+\infty , then the estimate is void: \displaystyle\mathfrak{q}_{V} \displaystyle:=\frac{\max\{1,\mathbf{C}_{V,n}^{(0)}\}}{\min\{1,\mathbf{c}_{V,n}^{(0)}\}},\qquad \displaystyle\mathfrak{q}_{W} \displaystyle:=\frac{\max\{1,\mathbf{C}_{W,n}^{(0)}\}}{\min\{1,\mathbf{c}_{W,n}^{(0)}\}}, \displaystyle\mathfrak{c}_{V} \displaystyle:=\min\{1,\mathbf{c}_{V,n}^{(0)}\},\qquad \displaystyle\mathfrak{C}_{W} \displaystyle:=\max\{1,\mathbf{C}_{W,n}^{(0)}\}, \displaystyle\mathfrak{L}_{V} \displaystyle:=\mathbf{C}_{V,n}^{(1)},\qquad \displaystyle\mathfrak{L}_{W} \displaystyle:=\mathbf{C}_{W,n}^{(1)}. These quantities control all larger parameters. Indeed, for fixed x , the ratio \frac{p+U(x)}{p+|x|^{2}} moves monotonically towards 1 as p increases, and (\sqrt{p}+|x|)^{-1} is decreasing in p . Thus, the zeroth-order upper and lower constants are controlled by their values at p=n and by the endpoint value 1 , while the first-order constants are largest at p=n . The ratios \mathbf{C}_{V,p}^{(0)}/\mathbf{c}_{V,p}^{(0)} and \mathbf{C}_{W,p}^{(0)}/\mathbf{c}_{W,p}^{(0)} need not be monotone in p , but they are bounded by \mathfrak{q}_{V} and \mathfrak{q}_{W} . Set \mathbf{M}_{\mathrm{glob}}:=10^{6}\,\mathfrak{q}_{V}^{2}\mathfrak{q}_{W}^{2},\qquad A_{\mathrm{glob}}:=\frac{\mathbf{C}_{V}^{(2)}\mathfrak{C}_{W}\mathbf{M}_{\mathrm{glob}}}{\mathbf{c}_{W}^{(2)}\mathfrak{c}_{V}},\qquad B_{\mathrm{glob}}:=\frac{8\mathfrak{L}_{V}\mathfrak{L}_{W}\mathbf{M}_{\mathrm{glob}}}{(\mathbf{c}_{W}^{(2)})^{2}\mathfrak{c}_{V}^{2}}, and C_{\mathrm{glob}}:=\sqrt{A_{\mathrm{glob}}+B_{\mathrm{glob}}}+\sqrt{B_{\mathrm{glob}}}. Theorem 1.2 (Global dimension-free Hessian estimate) . Let d,D\in[n,+\infty] with d\leq D . Let V>-d and W>-D , assume that the normalizing constants are finite, and let \nabla\varphi be the Brenier map pushing \rho_{V,d} forward to \rho_{W,D} . Assume that D^{2}V\leq\mathbf{C}_{V}^{(2)}\,\mathrm{Id},\qquad D^{2}W\geq\mathbf{c}_{W}^{(2)}\,\mathrm{Id}, where 0<\mathbf{c}_{W}^{(2)},\mathbf{C}_{V}^{(2)}<+\infty . Then \|D^{2}\varphi\|_{L^{\infty}(\mathbb{R}^{n})}\leq C_{\mathrm{glob}}. Taking D=d<+\infty in Theorem 1.2 gives the polynomial-to-polynomial estimate of [ 5 ] in a form that is uniform in the parameters n,d,D , contrary to the result in [ 5 ] which had an exponential dependence in d . While this theorem is extremely robust, it does not recover the sharp Caffarelli bound ( 1.1 ) in the case d=D=+\infty . To recover such a sharp result, we need first to prove quantitative bounds on the Hessian of \varphi in a fixed ball B_{R}(0) , then we will let D\to+\infty and d\to+\infty at first and finally let R\to+\infty . To prove this refined result, we need to introduce the quantities \mathcal{R}_{d,D}(R) and \mathcal{G}_{d,D}(R) . These are finite-parameter local growth constants coming from Fathi’s argument in [ 9 ] . For d,D\in[n,+\infty] with d\leq D , set (1.2) \Gamma_{d,D}:=\begin{cases}\displaystyle+\infty,&D=d<+\infty,\\[8.00003pt] \displaystyle\left(\frac{2d-D}{D-d}\right)_{+},&d<D<+\infty,\\[8.00003pt] 0,&D=+\infty.\end{cases} For d\leq D<+\infty and 0<R<+\infty , set s_{d,R}:=\max\{R,\sqrt{d}\} and \mathfrak{m}_{d,R}:=\frac{\omega_{n}(3s_{d,R})^{n}}{Z_{V,d}\bigl(\mathbf{C}_{V,d}^{(0)}(10s_{d,R})\bigr)^{d}}\left(1+\frac{100s_{d,R}^{2}}{d}\right)^{-d}. Here, \omega_{n} denotes the volume of the unit ball in \mathbb{R}^{n} . Also, define \Psi_{W,D}(r):=Z_{W,D}^{-1}\int_{B_{r}(0)^{c}}e^{-\Theta_{D}(W(y))}\,dy, and (1.3) \mathcal{R}_{d,D}(R):=3\inf\{r\geq 0:\Psi_{W,D}(r)\leq\mathfrak{m}_{d,R}\},\qquad\mathcal{G}_{d,D}(R):=1+\frac{\mathcal{R}_{d,D}(R)^{2}}{D}. Since \mathfrak{m}_{d,R}>0 and \Psi_{W,D}(r)\to 0 as r\to+\infty , the radius \mathcal{R}_{d,D}(R) is finite. For the localized estimate, with \Gamma_{d,D} as in Equation 1.2 , define (1.4) \displaystyle\Lambda_{d,D}(R) \displaystyle:=\begin{cases}\displaystyle\mathbf{C}_{W,D}^{(0)}(\mathcal{R}_{d,D}(R))\mathcal{G}_{d,D}(R),&D<+\infty,\\[8.00003pt] 1,&D=+\infty,\end{cases} \displaystyle\Xi_{d,D}(R) \displaystyle:=\begin{cases}\displaystyle\min\left\{\frac{D}{d}\,\frac{\mathbf{C}_{W,D}^{(1)}(\mathcal{R}_{d,D}(R))}{\mathbf{C}_{W,D}^{(0)}(\mathcal{R}_{d,D}(R))},\mathbf{c}_{W}^{(2)}\Gamma_{d,D}\right\},&D<+\infty,\\[8.00003pt] 0,&D=+\infty.\end{cases} Finally, set \displaystyle A_{V,W}^{d,D}(R) \displaystyle:=\frac{\mathbf{C}_{V}^{(2)}\Lambda_{d,D}(R)}{\mathbf{c}_{W}^{(2)}\mathbf{c}_{V,d}^{(0)}(R)},\qquad \displaystyle B_{V,W}^{d,D}(R) \displaystyle:=\frac{4\mathbf{C}_{V,d}^{(1)}\Lambda_{d,D}(R)\Xi_{d,D}(R)}{(\mathbf{c}_{W}^{(2)})^{2}(\mathbf{c}_{V,d}^{(0)}(R))^{2}}. Theorem 1.3 (Localized Hessian estimate) . Let d,D\in[n,+\infty] with d\leq D , and let 0<R<+\infty . Let V>-d and W>-D , assume that the normalizing constants are finite,

Read the complete text