betaFeedback

WorksMathematics

A height transportation approach to the quantitative Prékopa–Leindler inequality

Abstract

We give a new, simpler proof of sharp stability for the Prékopa–Leindler inequality. In contrast to the main arguments from Figalli, van Hintum and Tiba~\cite{FigalliVanHintumTibaPL}, our proof hinges mostly on the sharp stability theory for the Brunn–Minkowski problem, developed by the same authors \cite{figalli2024BM}, and on the idea of using a height transportation approach in order to match level sets. As a by-product of our techniques, we are able to prove the sharp dependency on τ=min⁡{λ,1−λ}τ= \min\{λ,1-λ\} in the Prékopa–Leindler stability constant whenever one of the inputs is logarithmically concave, producing a new, simplified proof of two of the main results from~\cite{FigalliVanHintumTibaPL}. Finally, in the same fashion as Kristály and Pîrvuceanu \cite{kristaly2026quantitative}, we are able to obtain new stability results for the Bakry-Emery log-Sobolev inequality and for the Talagrand inequality. Due to the sharpness of the inputs, these results are optimal in terms of exponents. We moreover show that the L1L^1 nature of the distance in such results cannot in general be improved to any other LpL^p distance, where p>1.p>1.

2020 Mathematics Subject Classification. Primary 39B62, 52A40; Secondary 26D15, 49Q22, 52A20, 26D10, 60E15.

Introduction

In recent years there has been an increasing interest in the geometry of log-concave functions. These objects and their theory can be interpreted as a functional version of the Brunn–Minkowski theory developed for convex bodies [23, 31, 29], and play a pivotal role in theoretical guarantees of sampling algorithms [25]. One of the major structuring tools in this theory is the Prékopa-Leindler inequality, which gives a concavity principle to the natural generalization to the notion of log-concave volume in this geometry, as observed for instance in [15].

The Prékopa-Leindler inequality says that given three measurable functions f,g,h:Rn→R+f,g,h : \mathbb{R}^n \to\mathbb{R}_+ satisfying the Prékopa condition: for some λ∈(0,1)\lambda\in(0,1) it holds that

h(λx+(1−λ)y)≥f(x)λg(y)1−λ,for all x,y∈Rn,(1)h(\lambda x + (1-\lambda)y) \geq f(x)^\lambda g(y)^{1-\lambda}, \qquad\text{for all } x,y \in\mathbb{R}^n, \tag*{(1)}

then this inequality can be lifted to their integrals, namely

∫Rnh(x) dx≥(∫Rnf(x) dx)λ(∫Rng(x) dx)1−λ.(2)\int_{\mathbb{R}^n} h(x)\,\mathrm{d}x \ge\left(\int_{\mathbb{R}^n} f(x)\,\mathrm{d}x\right)^\lambda\left(\int_{\mathbb{R}^n} g(x)\,\mathrm{d}x\right)^{1-\lambda}. \tag*{(2)}

To emphasize the geometry of log-concave functions, it is often formulated in terms of two functions f,gf,g and their Asplund sum, defined as

hλ(z)≔sup⁡z=λx+(1−λ)yf(x)λg(y)1−λ.(3)h_\lambda(z) \coloneqq\sup_{z=\lambda x+(1-\lambda)y} f(x)^\lambda g(y)^{1-\lambda}. \tag*{(3)}

Then of course, any function hh such that f,g,hf,g,h satisfy (1.1) is bigger than hλh_\lambda and this one is the sharpest function for which (1.2) may hold.

Since the recent contribution of Figalli-van-Hintum-Tiba [22], this inequality has a sharp quantitative version, which can be interpreted as a strong concavity principle for the natural notion of volume in this geometry. This has found many applications to the stability of other functional and transport inequalities, as well as other variational problems that are central in log-concave geometry [MR25, BR26, KP26].

Theorem 1.1. For every n≥1n \ge1, there exists a dimensional constant Cn>0C_n > 0 such that the following holds. Let f,g,h:Rn→R+f,g,h : \mathbb{R}^n \to\mathbb{R}_+ be integrable functions satisfying the Prékopa condition (1.1) for some λ∈(0,1)\lambda\in(0,1), with ∥f∥L1,∥g∥L1>0\lVert f\rVert_{L^1}, \lVert g\rVert_{L^1} > 0. Assume that at least one of f,gf,g is log-concave such that almost equality holds

∫Rnh(x) dx≤(1+δPL)(∫Rnf(x) dx)λ(∫Rng(x) dx)1−λ,\int_{\mathbb{R}^n} h(x)\,\mathrm{d}x \le(1+\delta_{\mathrm{PL}})\left(\int_{\mathbb{R}^n} f(x)\,\mathrm{d}x\right)^\lambda\left(\int_{\mathbb{R}^n} g(x)\,\mathrm{d}x\right)^{1-\lambda},

where τ≔min⁡{λ,1−λ}\tau\coloneqq\min\{\lambda,1-\lambda\}. Then we have that

inf⁡z∈Rn∥f∥f∥L1−g(⋅+z)∥g∥L1∥L1(Rn)≤CnδPLτ.(4)\inf_{z\in\mathbb{R}^n}\left\lVert\frac{f}{\lVert f\rVert_{L^1}}-\frac{g(\cdot+z)}{\lVert g\rVert_{L^1}}\right\rVert_{L^1(\mathbb{R}^n)} \le C_n\sqrt{\frac{\delta_{\mathrm{PL}}}{\tau}}. \tag*{(4)}

The exponent 1/21/2 and the order of dependence on λ\lambda are optimal. Moreover, one may choose Cn≤eCnC_n \le e^{C n} for an absolute constant CC.

Remark 1.2. Keeping track of the dimension is useful here. The factor CenC e^n obtained when joining the translations is only one part of the stability constant: one must also include the polynomial factors from the height estimates and the dimensional constant from Brunn–Minkowski. Together, these give the bound eCne^{C_n} in Theorem 1.1, both for two log-concave inputs and for a single log-concave input.

For comparison, Theorem 1.9 and Remark 1.10 in the second arXiv version of [21] (6 October 2026), specialized to p=0p=0, give a constant of order nO(n)n^{O(n)} multiplying the same factor δPL/τ\sqrt{\delta_{\mathrm{PL}}/\tau}. Thus the dimensional upper bound obtained here improves the one stated there; we do not, however, claim that this specific dependence on nn is optimal. For general inputs, the constants and the smallness threshold in the Brunn–Minkowski input remain part of our estimate, as recorded in § 6.3.

With minor modifications to our proof, the same result holds for arbitrary integrable functions f,gf,g and hh, but with a non-optimal dependence on λ\lambda. We refer to § 6.3 for a discussion of this case. On the other hand, as discussed below in this introduction, we believe that our proof is particularly well suited to the case of log-concave functions, and we expect that the optimal dependence on λ\lambda is rightly simplified by the following argument due to [5].

Corollary 1.3. Under the assumptions of Theorem 1.1, assume in addition that both ff and gg are log-concave. Then the midpoint case of estimate (1.4) implies the full estimate for every λ∈(0,1)\lambda\in(0,1), up to a universal multiplicative change in CnC_n.

Proof. Let f,gf,g be log-concave functions, and assume without loss of generality that f,gf,g are normalized so that

∫Rnf(x)dx=∫Rng(x)dx=1.\int_{\mathbb{R}^{n}} f(x)\mathrm{d}x=\int_{\mathbb{R}^{n}} g(x)\mathrm{d}x=1.

Define their Asplund interpolation as

ht(z)≔sup⁡z=tx+(1−t)yf(x)tg(y)1−t,and setΦ(t)≔∫Rnht(z)dz.h_t(z)\coloneqq\sup_{z=tx+(1-t)y} f(x)^t g(y)^{1-t}, \qquad\text{and set}\quad \Phi(t)\coloneqq\int_{\mathbb{R}^{n}}h_t(z)\mathrm{d}z.

As a result, the Prékopa-Leindler inequality implies that

t↦log⁡Φ(t) is concave, and Φ(0)=Φ(1)=1.t\mapsto\log\Phi(t)\text{ is concave, and }\Phi(0)=\Phi(1)=1.

Up to swapping f,gf,g, we can assume that λ≤1/2\lambda\leq1/2, so that log concavity gives

log⁡Φ(λ)=log⁡Φ(2λ⋅12+(1−2λ)⋅0)≥2λlog⁡Φ(1/2).\log\Phi(\lambda)=\log\Phi\left(2\lambda\cdot\frac{1}{2}+(1-2\lambda)\cdot0\right)\geq2\lambda\log\Phi(1/2).

Thanks to the chosen normalization, the deficit of the Prékopa-Leindler inequality with convexity parameter tt is given by

t↦δPL(t)≔Φ(t)−1,t\mapsto\delta_{\mathrm{PL}}(t)\coloneqq\Phi(t)-1,

hence assuming that the stability result holds for t=1/2t=1/2 with a universal stability constant CnC_n, we obtain the log-concavity of Φ\Phi that

1Cninf⁡z∈Rn∥f−g(⋅+z)∥L1(Rn)2≤δPL(1/2)≤(1+δPL(λ))12λ−1.\frac{1}{C_n}\inf_{z\in\mathbb{R}^{n}}\left\|f-g(\cdot+z)\right\|_{L^1(\mathbb{R}^{n})}^{2}\leq\delta_{\mathrm{PL}}(1/2)\leq\left(1+\delta_{\mathrm{PL}}(\lambda)\right)^{\frac{1}{2\lambda}}-1.

Since ∥f−g(⋅+z)∥L1(Rn)≤2\left\|f-g(\cdot+z)\right\|_{L^1(\mathbb{R}^{n})}\leq2, up to increasing the constant CnC_n it suffices to treat δPL(λ)≤λ\delta_{\mathrm{PL}}(\lambda)\leq\lambda. In this case,

(1+δPL(λ))12λ−1≤exp⁡(δPL(λ)2λ)−1≤δPL(λ)λ,\left(1+\delta_{\mathrm{PL}}(\lambda)\right)^{\frac{1}{2\lambda}}-1\leq\exp\left(\frac{\delta_{\mathrm{PL}}(\lambda)}{2\lambda}\right)-1\leq\frac{\delta_{\mathrm{PL}}(\lambda)}{\lambda},

and the above estimates give

inf⁡z∈Rn∥f−g(⋅+z)∥L1(Rn)≤CnδPL(λ)λ.\inf_{z\in\mathbb{R}^{n}}\left\|f-g(\cdot+z)\right\|_{L^1(\mathbb{R}^{n})}\leq C_n\sqrt{\frac{\delta_{\mathrm{PL}}(\lambda)}{\lambda}}.

Since τ=λ\tau=\lambda and the original hh dominates hλh_\lambda, the result follows by mass normalization. □

Scheme of the proofs. Our first contribution is a new height transport proof of the classical Prékopa–Leindler inequality, which also yields a full characterization of the equality cases discovered by Dubuc [17]. The height transport method is based on the intuition that the Prékopa–Leindler inequality is the functional version of Brunn–Minkowski, and therefore it would be natural to expect a proof of the former by matching pairs of heights of the functions f,gf,g and applying the latter inequality to the corresponding matched level-sets, as illustrated in Figure 1.

References

  1. [1]Giovanni Alberti and Luigi Ambrosio. A geometrical approach to monotone functions in. Mathematische Zeitschrift, 230(2):259–316, 1999.DOI
  2. [2]Luigi Ambrosio, Elia Brué, and Daniele Semola. Lectures on optimal transport, volume 130. Springer, 1 edition, 2021.
  3. [3]João Henrique Andrade, José Francisco de Oliveira, João Marcos do Ó, Abiel Costa Macedo, and Jesse Ratzkin. Quantitative stability for the Trudinger–Moser inequality, 2026.arxiv.org/abs/2605.03668
  4. [4]Luigi Ambrosio, Nicola Fusco, and Diego Pallara. Functions of Bounded Variation and Free Discontinuity Problems. Oxford Mathematical Monographs. Oxford University Press, March 2000.
  5. [5]Károly J Böröczky and Apratim De. Stability of the prékopa-leindler inequality for log-concave functions. Advances in Mathematics, 386:107810, 2021.arxiv.org/abs/2007.15304
  6. [6]Gabriele Bianchi and Henrik Egnell. A note on the Sobolev inequality. Journal of Functional Analysis, 100(1):18–24, 1991.DOI
  7. [7]Marco Barchiesi and Vesa Julin. Robustness of the gaussian concentration inequality and the Brunn–Minkowski inequality. Calculus of Variations and Partial Differential Equations, 56(3):80, 2017.DOI
  8. [8]Michel Bonnefont, Aldéric Joulin, and Yutao Ma. Spectral gap for spherically symmetric log-concave probability measures, and beyond. Journal of Functional Analysis, 270(7):2456–2482, 2016.arxiv.org/abs/1406.4621
  9. [9]Haïm Brezis and Elliott H Lieb. Sobolev inequalities with remainder terms. Journal of Functional Analysis, 62(1):73–86, 1985.
  10. [10]Sergey G Bobkov and Michel Ledoux. From Brunn-Minkowski to Brascamp-Lieb and to logarithmic Sobolev inequalities. Geometric and Functional Analysis, 10(5):1028–1052, 2000.DOI
  11. [11]S. Bobkov. Extremal properties of half-spaces for log-concave distributions. Annals of Probability, 24(1):35 – 48, 1996.DOI
  12. [12]Károly J Böröczky, Yaozhong W Qiu, and Cyril Roberto. L²-stability for the variance Brascamp–Lieb inequality. arXiv preprint arXiv:2602.13679, 2026.arxiv.org/abs/2602.13679
  13. [14]Y. Brenier. Polar factorization and monotone rearrangement of vector-valued functions. Communications on pure and applied mathematics, 44(4):375–417, 1991.DOI
  14. [15]Dario Cordero-Erausquin and Bo’az Klartag. Moment measures. Journal of Functional Analysis, 268(12):3834–3866, 2015.
  15. [16]Jean Dolbeault, Maria J. Esteban, Alessio Figalli, Rupert L. Frank, and Michael Loss. Sharp stability for Sobolev and log-Sobolev inequalities, with optimal dimensional dependence. Cambridge Journal of Mathematics, 13(2):359–430, 2025.DOI
  16. [17]Serge Dubuc. Critères de convexité et inégalités intégrales. Annales de l’Institut Fourier, 27(1):135–165, 1977.DOI
  17. [18]Alessio Figalli, Francesco Maggi, and Aldo Pratelli. A refined brunn–minkowski inequality for convex sets. 26(6):2511–2519, 2009.
  18. [19]Alessio Figalli, Francesco Maggi, and Aldo Pratelli. A mass transportation approach to quantitative isoperimetric inequalities. Inventiones mathematicae, 182(1):167–211, 2010.
  19. [20]Alessio Figalli, Peter van Hintum, and Marius Tiba. Sharp quantitative stability of the Brunn–Minkowski inequality. arXiv preprint arXiv:2310.20643, 2023.arxiv.org/abs/2310.20643
  20. [21]Alessio Figalli, Peter van Hintum, and Marius Tiba. Sharp quantitative stability for the Prékopa–Leindler and Borell–Brascamp–Lieb inequalities. 2025.arxiv.org/abs/2501.04656
  21. [22]Alessio Figalli, Peter van Hintum, and Marius Tiba. Sharp quantitative stability for the Prékopa–Leindler and Borell–Brascamp–Lieb inequalities. arXiv preprint arXiv:2501.04656, 2025.arxiv.org/abs/2501.04656
  22. [23]Richard J. Gardner. The Brunn–Minkowski inequality. Bulletin of the American Mathematical Society, 39(3):355–405, 2002.DOI
  23. [24]Alexandru Kristály and Alexandru Pîrvuceanu. Quantitative stability for Bakry–Émery log-sobolev and talagrand inequalities, 2026.arxiv.org/abs/2608.07996
  24. [25]László Lovász and Santosh Vempala. The geometry of logconcave functions and sampling algorithms. Random Structures & Algorithms, 30(3):307–358, 2007.DOI
  25. [26]Bernard Maurey. Some deviation inequalities. Geometric and Functional Analysis, 1(2):188–197, 1991.
  26. [28]Vitali D. Milman and Gideon Schechtman. Asymptotic Theory of Finite Dimensional Normed Spaces, volume 1200 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 1986. With an appendix by M. Gromov.
  27. [29]Liran Rotem. Surface area measures of log-concave functions. Journal d’Analyse Mathématique, 147:373–400, 2022.arxiv.org/abs/2006.16933
  28. [30]Filippo Santambrogio. Optimal transport for applied mathematicians, volume 55. Springer, 2015.DOI
  29. [31]Rolf Schneider. Convex bodies: the Brunn–Minkowski theory, volume 151. Cambridge university press, 2013.
  30. [32]Adrien Saumard and Jon A. Wellner. Log-concavity and strong log-concavity: A review. Statistics Surveys, 8:45 – 114, 2014.arxiv.org/abs/1404.5886
  31. [33]Filippo Santambrogio and Xu-Jia Wang. Convexity of the support of the displacement interpolation: Counterexamples. Applied Mathematics Letters, 58:152–158, 2016.DOI

Paper details

vlnc:W_e39nm7fd
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY-SA 4.0
  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents