betaFeedback

WorksMathematics

A centroaffine perspective on improved Brascamp–Lieb and dimensional Brunn–Minkowski inequalities

Abstract

We study dimensional concavity properties of even log-concave measures within the framework of weighted centroaffine differential geometry. We express the relevant local concavity inequalities in terms of a weighted Hilbert–Brunn–Minkowski operator and establish an anisotropic Poincaré inequality for odd functions and a strengthened Brascamp–Lieb inequality for even functions. As applications, we obtain quantitative dimensional concavity estimates for a class of geometric measures and measures with pinched Hessian, extending results of Cordero-Erausquin and Rotem for rotationally invariant measures and Eskenazis and Moschidis for the Gaussian measure.

Introduction

The classical Brunn–Minkowski inequality asserts that, for any two convex bodies K,L⊂RnK, L \subset\mathbb{R}^{n}, namely compact convex sets with nonempty interior, and every t∈[0,1]t \in[0,1],

∣(1−t)K+tL∣1n≥(1−t)∣K∣1n+t∣L∣1n.(1)|(1-t)K+tL|^{\frac{1}{n}} \ge(1-t)|K|^{\frac{1}{n}}+t|L|^{\frac{1}{n}}. \tag*{(1)}

Here, ∣⋅∣|\cdot| denotes the Lebesgue measure. Moreover, we denote by Kn\mathcal{K}^{n} the family of convex bodies in Rn\mathbb{R}^{n}. Considerable attention has been devoted to extending and strengthening (1), particularly under symmetry assumptions. One direction consists of replacing Minkowski addition by the LpL_{p}-addition of convex bodies. Of particular interest is the range 0≤p<10 \le p < 1, where the LpL_{p}-Brunn–Minkowski conjectures assert that

∣(1−t)K+ptL∣≥((1−t)∣K∣pn+t∣L∣pn)np,(2)|(1-t)K+_{p}tL| \ge\left((1-t)|K|^{\frac{p}{n}}+t|L|^{\frac{p}{n}}\right)^{\frac{n}{p}}, \tag*{(2)}

for all origin-symmetric K,L∈KnK,L \in\mathcal{K}^{n} and every t∈[0,1]t \in[0,1]. We defer the precise definition of LpL_{p}-addition to Section 2 and note that, in the chosen range 0≤p<10 \le p < 1, (2) strengthens (1). The limiting case p=0p=0 is the celebrated log-Brunn–Minkowski conjecture of Böröczky, Lutwak, Yang and Zhang [3], which asserts that

∣(1−t)K+0tL∣≥∣K∣1−t∣L∣t.|(1-t)K+_{0}tL| \ge|K|^{1-t}|L|^{t}.

An approach to these questions is provided by the spectral properties of the Hilbert–Brunn–Minkowski operator, which we briefly describe. Let Ke,+2,α\mathcal{K}^{2,\alpha}_{e,+} denote the class of origin-symmetric

MSC 2020 Classification: 52A40, 26D10, 53A15, 60E15. convex bodies with C2,αC^{2,\alpha} boundary and strictly positive principal curvatures. For K∈Ke,+2,αK \in\mathcal{K}^{2,\alpha}_{e,+}, let hKh_K be its support function and define the tensor

QK=∇Sn−12hK+hKId⁡Q_K = \nabla^2_{\mathbb{S}^{n-1}} h_K + h_K \operatorname{Id}

on the tangent bundle of Sn−1\mathbb{S}^{n-1}. Here, ∇Sn−1\nabla_{\mathbb{S}^{n-1}} denotes the spherical gradient. With the normalization used here, the Hilbert–Brunn–Minkowski operator acts on z∈C2(Sn−1)z \in C^2(\mathbb{S}^{n-1}) by

ΔKz=(QK−1)ij(hK∇ij2z+(hK)izj+(hK)jzi),(3)\Delta_K z = (Q_K^{-1})^{ij}\left(h_K\nabla^2_{ij}z + (h_K)_i z_j + (h_K)_j z_i\right), \tag*{(3)}

where the standard covariant derivative on Sn−1\mathbb{S}^{n-1} is considered and repeated indices are summed. Kolesnikov and E. Milman [23, 26] related local LpL_p-Brunn–Minkowski inequalities to lower bounds on the first nonzero even eigenvalue of −ΔK-\Delta_K. In particular, they established the local inequalities for every smooth origin-symmetric convex body in a dimension-dependent range of parameters p<1p < 1 sufficiently close to 1. E. Milman [26] subsequently identified ΔK\Delta_K with the Laplace–Beltrami operator on the boundary of KK (which we denote by ∂K\partial K), associated with a suitable Riemannian metric and a connection distinct from the Levi-Civita connection of that metric. Doing so, the treatment in [26] placed the problem in the context of so-called centroaffine differential geometry and led to results under curvature pinching conditions. The relation between this geometry and the Hessian metric of the Minkowski functional was further developed by Ivaki and E. Milman [18]. Similar results under pinching were later obtained in connection to improved Brascamp–Lieb inequalities by Kolesnikov, Livshyts and Rotem [19].

Another direction concerns replacing the Lebesgue measure by more general measures. For a measure μ\mu with density e−ue^{-u}, where uu is convex, the Prékopa–Leindler inequality asserts that

μ((1−t)K+tL)≥μ(K)1−tμ(L)t(4)\mu((1-t)K+tL) \geq\mu(K)^{1-t}\mu(L)^t \tag*{(4)}

for all convex bodies KK, LL and every t∈[0,1]t \in[0,1]. Equivalently (compare the work of Borell [2]), a Borel measure μ\mu satisfying (4) is said to be log-concave. For the Lebesgue measure, it is immediate via homogeneity that this multiplicative adimensional inequality is equivalent to (1). The same procedure evidently breaks for general log-concave measures. Improving the concavity of log-concave measures is a research topic that has produced a significant amount of work in the study of log-concave measures, the main example being the dimensional Brunn–Minkowski conjecture: It postulates that the volume in (1) can be replaced with an arbitrary even log-concave measure when the convex sets are assumed to be origin symmetric. As a natural question, the weighted and LpL_p extensions meet in the study of inequalities of the form

μ((1−t)K+ptL)≥((1−t)μ(K)pn+tμ(L)pn)np,(5)\mu((1-t)K +_p tL) \geq\left((1-t)\mu(K)^{\frac{p}{n}} + t\mu(L)^{\frac{p}{n}}\right)^{\frac{n}{p}}, \tag*{(5)}

for origin-symmetric K,L∈KnK,L \in\mathcal{K}^n, and even log-concave measures μ\mu, with p=0p=0 interpreted in the limiting sense. The choice p=1p=1 gives the dimensional Brunn–Minkowski conjecture, while p=0p=0 gives the weighted log-Brunn–Minkowski conjecture. These inequalities were studied within a common (p,q)(p,q)-concavity framework by Hosle, Kolesnikov and Livshyts [16].

For the choice of μ\mu as the Gaussian measure and p=1p=1, Gardner and Zvavitch [14] originally proposed the dimensional inequality (5) under the weaker assumption that the convex bodies contain the origin. Nayar and Tkocz [28] disproved this formulation, leaving the origin-symmetric case as the principal question. The corresponding dimensional problem for general even log-concave measures was investigated by Colesanti, Livshyts and Marsiglietti [6], who established local results near the Euclidean ball for rotationally invariant log-concave measures.

We briefly comment on the usual approach for this weighted setting, using a Bochner formula due to Kolesnikov and E. Milman [21, 22] and Kolesnikov and Livshyts [20], for these concavity inequalities it is sufficient to establish coercivity estimates for quantities of the form

∫K∥D2f∥HS+⟨(D2u)Df,Df⟩ dμ\int_K \lVert D^2 f\rVert_{HS} + \langle(D^2u)Df, Df\rangle\, \mathrm{d}\mu

for every origin symmetric convex set KK and even function ff. Using this approach, Eskenazis and Moschidis [12] subsequently proved the Gaussian inequality with the conjectured exponent 1/n1/n for all origin-symmetric convex bodies. Further extensions have since been obtained. Han and Zhu [15] replaced origin symmetry by the assumption that the Gaussian measures conditioned on the two convex sets have the same barycenter, and obtained an inequality with explicit barycenter terms in the general case. Recently Milman, Nakamura and Tsuji [27] established a heterovariant Gaussian Brunn–Minkowski inequality, involving Gaussian measures with different covariance matrices subject to a matrix compatibility condition. Their result contains the dimensional Gaussian inequality and connects it with strengthened Gaussian dilation inequalities, which relates to the (B) conjecture, which we now state. The Gaussian statement was proposed by Banaszczyk and popularized by Latała [24]; its extension to even log-concave measures asks whether, for every origin-symmetric convex body KK, the function

t⟼μ(etK)t \longmapsto\mu(e^t K)

is log-concave on R\mathbb{R}. Cordero-Erausquin, Fradelizi and Maurey [8] proved the Gaussian case. We also remark that the (B) property holds for a certain class of Gaussian mixtures [13]. Saroglou [30] showed that the log-Brunn–Minkowski conjecture for the Lebesgue measure implies its weighted version for even log-concave measures and, consequently, the (B) conjecture. Cordero-Erausquin and Rotem [10] established both the (B) property and the dimensional Brunn–Minkowski inequality for rotationally invariant log-concave measures, as well as for a broader class of measures with radial densities.

The general scheme is to prove an improved Poincaré inequality for odd functions, then apply this estimate to the coordinate functions of a gradient of an even function and conclude some coercivity estimate. In the general case that the measure is assumed to be merely log-concave this was combined with delicate gradient estimates to produce a weaker form of (5), in an approach put forth by Livshyts [25] and extended by Eskenazis, Giannopoulos and Tziotziou [11] and Yang [32].

Main results. The purpose of this work is to connect the spectral and centroaffine approach to the log-Brunn–Minkowski problem with the functional inequalities developed for dimensional concavity and the (B) conjecture. A similar path has been recently explored also by Hu and Ivaki [17].

We first express the weighted local concavity inequalities in terms of a spectral minimization problem for an appropriate operator defined on the boundary of convex sets. More precisely, given K∈Ke,+2,αK \in\mathcal{K}^{2,\alpha}_{e,+} and a log-concave measure μ\mu with density of class C1C^1, we define the weighted Hilbert–Brunn–Minkowski operator as

DK,μf:=−ΔK,μf+⟨∇u,x⟩f.(6)D_{K,\mu}f := -\Delta_{K,\mu}f+\langle\nabla u,x\rangle f. \tag*{(6)}

Here, ΔK,μ\Delta_{K,\mu} denotes the weighted centroaffine Laplace–Beltrami operator, which will be defined properly in § 2. Moreover, ⟨⋅,⋅⟩\langle\cdot,\cdot\rangle denotes the standard Euclidean scalar product in Rn\mathbb{R}^{n}. We use the same notation for its restriction to embedded submanifolds of Rn\mathbb{R}^{n}. The weighted cone-volume measure VK,μV_{K,\mu} on ∂K\partial K with respect to μ\mu is defined as the measure on ∂K\partial K with density

dVK,μ(x)=⟨x,νK(x)⟩nμ(K)dHn−1.\mathrm{d}V_{K,\mu}(x)=\frac{\langle x,\nu_{K}(x)\rangle}{n\mu(K)}\mathrm{d}\mathcal{H}^{n-1}.

We note that pushing (6) on the sphere through the Gauss map in the unweighted setting yields precisely the usual Hilbert–Brunn–Minkowski operator (3).

With these definitions at hand, the relevant functional for us is the following:

JK,μ(f):=∫∂K(DK,μf)f dVK,μ∫∂Kf2 dVK,μ−(∫∂Kf dVK,μ)2.J_{K,\mu}(f):=\frac{\int_{\partial K}(D_{K,\mu}f)f\,\mathrm{d}V_{K,\mu}}{\int_{\partial K}f^{2}\,\mathrm{d}V_{K,\mu}-\left(\int_{\partial K}f\,\mathrm{d}V_{K,\mu}\right)^{2}}.

We remark that whenever dμ(x)=e−u(x)dx\mathrm{d}\mu(x)=e^{-u(x)}\mathrm{d}x and uu is strictly convex and even, the denominator is never zero unless f≡0f\equiv0. Our first result records the common spectral formulation of the preceding concavity properties.

Theorem 1.1. Let μ\mu be an even log-concave measure on Rn\mathbb{R}^{n} with density e−ue^{-u}, where u∈C1(Rn)u\in C^{1}(\mathbb{R}^{n}) is strictly convex. Then the following characterizations hold:

  1. For p∈[0,1]p\in[0,1], inequality (5) holds for all origin-symmetric convex bodies K,LK,L and every t∈[0,1]t\in[0,1] if and only if

inf⁡f∈C2(∂K)Var⁡VK,μ(f)>0JK,μ(f)≥n−p(7)\inf_{\substack{f\in C^{2}(\partial K)\\ \operatorname{Var}_{V_{K,\mu}}(f)>0}}J_{K,\mu}(f)\ge n-p \tag*{(7)}

for every K∈Ke,+2,αK\in\mathcal{K}_{e,+}^{2,\alpha}. The case p=0p=0 is understood in the limiting sense.

  1. The measure μ\mu satisfies the (B) inequality for all origin-symmetric convex bodies if and only if

inf⁡f≡c≠0JK,μ(f)≥n\inf_{f\equiv c\ne0}J_{K,\mu}(f)\ge n

for every K∈Ke,+2,αK\in\mathcal{K}_{e,+}^{2,\alpha}.

We note that in the case of the Lebesgue measure, the weighted cone-volume measure VK,μV_{K,\mu} reduces to the usual normalized cone-volume measure and (7) reduces to the first even eigenvalue of the Hilbert–Brunn–Minkowski operator.

Fix now a smooth origin-symmetric convex body KK, and set

Φ(x)=∥x∥K22,\Phi(x)=\frac{\lVert x\rVert_{K}^{2}}{2},

where ∥⋅∥K\lVert\cdot\rVert_{K} denotes its Minkowski functional. Following the ideas of [10], we were able to extend their central Poincaré inequality for odd functions to the anisotropic setting. Let us remark that, during the preparation of this work, a preprint by Cordero-Erausquin, Livshyts and Rotem [9] appeared where the following result was obtained independently. The proof strategy is the same up to minor differences, one of them being that we use a weighted centroaffine formalism. This shall be of no surprise, as the blueprint is the one laid down in the substantial work by Cordero-Erausquin and Rotem [10]. Nonetheless, in this manuscript we manage to further leverage the ideas from [10] combining them with the approach of [12]. In particular, we obtain the conclusions expected in [9 p. 9] as a consequence of the following result. We present these implications in the remainder of this manuscript.

Theorem 1.2. Let KK be a smooth origin-symmetric convex body with strictly positive principal curvatures, let w∈C1((0,∞))w \in C^{1}((0,\infty)) be increasing, and let V:Rn→RV:\mathbb{R}^{n}\to\mathbb{R} be even and convex. Consider the finite measure

dη(x)=e−V(x)−w(∥x∥K) dx.\mathrm{d}\eta(x)=e^{-V(x)-w(\lVert x\rVert_{K})}\,\mathrm{d}x.

Then, for every odd φ∈C1(Rn)\varphi\in C^{1}(\mathbb{R}^{n}) for which the integrals are finite,

∫Rnw′(∥x∥K)∥x∥Kφ2 dη≤∫Rn⟨(D2Φ)−1∇φ,∇φ⟩ dη.\int_{\mathbb{R}^{n}} \frac{w'(\lVert x\rVert_{K})}{\lVert x\rVert_{K}}\varphi^{2}\,\mathrm{d}\eta\leq\int_{\mathbb{R}^{n}} \left\langle(D^{2}\Phi)^{-1}\nabla\varphi,\nabla\varphi\right\rangle\,\mathrm{d}\eta.

To pass from the above to an improved Brascamp–Lieb inequality for even functions, we impose the pinching condition

αId⁡≤D2Φ(x)≤βId⁡,x≠0,0<α≤β<∞.(8)\alpha\operatorname{Id} \leq D^{2}\Phi(x) \leq\beta\operatorname{Id}, \qquad x\ne0,\qquad0<\alpha\leq\beta<\infty. \tag*{(8)}

and write γ=α/β\gamma=\alpha/\beta. Adapting the approach from [10], we obtain the following consequence of Theorem 1.2.

Theorem 1.3. Let KK be a smooth origin-symmetric convex body with strictly positive principal curvatures satisfying (8), and let w∈C2((0,∞))w \in C^{2}((0,\infty)) be increasing and satisfy

w′′(s)+γw′(s)s≥0,s>0.(9)w''(s)+\gamma\frac{w'(s)}{s}\geq0,\qquad s>0. \tag*{(9)}

Let V∈C2(Rn)V\in C^{2}(\mathbb{R}^{n}) be even and convex, and suppose that

dη(x)=e−V(x)−W(x) dx,W(x)=w(∥x∥K),\mathrm{d}\eta(x)=e^{-V(x)-W(x)}\,\mathrm{d}x,\qquad W(x)=w(\lVert x\rVert_{K}),

is a probability measure. Then, for every even φ∈C1(Rn)\varphi\in C^{1}(\mathbb{R}^{n}),

Var⁡η(φ)≤∫Rn⟨(D2W+γw′(∥x∥K)∥x∥KD2Φ)−1∇φ,∇φ⟩ dη.\operatorname{Var}_{\eta}(\varphi)\leq\int_{\mathbb{R}^{n}}\left\langle\left(D^{2}W+\gamma\frac{w'(\lVert x\rVert_{K})}{\lVert x\rVert_{K}}D^{2}\Phi\right)^{-1}\nabla\varphi,\nabla\varphi\right\rangle\,\mathrm{d}\eta.

When the matrix is singular, the inverse quadratic form is understood as the associated dual quadratic form.

Here Var⁡η(φ)=∫φ2 dη−(∫φ dη)2\operatorname{Var}_{\eta}(\varphi)=\int\varphi^{2}\,\mathrm{d}\eta-\left(\int\varphi\,\mathrm{d}\eta\right)^{2}. When KK is the Euclidean unit ball, D2ΦD^{2}\Phi is just the identity matrix Id⁡\operatorname{Id} and one can take γ=1\gamma=1, recovering the corresponding inequalities of Cordero-Erausquin and Rotem [10].

Using the L2L_{2} method and the above inequalities we can conclude a dimensional Brunn–Minkowski inequality for this class of measures.

Theorem 1.4. Let KK be a smooth origin-symmetric convex body with strictly positive principal curvatures, and suppose that Φ=∥⋅∥K2/2\Phi= \lVert\cdot\rVert_K^2/2 satisfies (8). Let γ=α/β\gamma= \alpha/\beta, and let w∈C2((0,∞))w \in C^2((0,\infty)) be increasing and satisfy (9). Suppose that

dμ(x)=e−w(∥x∥K)dx\mathrm{d}\mu(x) = e^{-w(\lVert x\rVert_K)} \mathrm{d}x

defines a finite measure. Then, for every pair of origin-symmetric convex bodies A,B⊂RnA, B \subset\mathbb{R}^n and every t∈[0,1]t \in[0,1],

μ((1−t)A+tB)γ≥(1−t)μ(A)γ+tμ(B)γ.(10)\mu((1-t)A+tB)^{\gamma} \ge(1-t)\mu(A)^{\gamma}+t\mu(B)^{\gamma}. \tag*{(10)}

We point out that inequality (10) constitutes a quantitative version of the result of [10], since the parameter γ\gamma in the pinching above quantifies how far the set KK is from a ball, in centroaffine terms. We remark that some of the measures above need not necessarily be log-concave. Finally, we show that a similar argument, independent of the centroaffine machinery we develop, works for general measures with a suitable pinching on the potential, leading to the following result.

Theorem 1.5. Let V∈C2(Rn)V \in C^2(\mathbb{R}^n) be even and convex, and suppose that

k1Id⁡≤D2V,ΔV≤k2nk_1 \operatorname{Id} \le D^2V,\qquad\Delta V \le k_2n

for some 0<k1≤k2<∞0 < k_1 \le k_2 < \infty. Set r=k1/k2r = k_1/k_2 and let dμ=e−Vdx\mathrm{d}\mu= e^{-V}\mathrm{d}x. Then, for every pair of origin-symmetric convex bodies A,B⊂RnA, B \subset\mathbb{R}^n and every t∈[0,1]t \in[0,1],

μ((1−t)A+tB)r≥(1−t)μ(A)r+tμ(B)r.\mu((1-t)A+tB)^r \ge(1-t)\mu(A)^r+t\mu(B)^r.

Theorem 1.5 can be understood as an extension of the result of Eskenazis and Moschidis [12], since for the Gaussian k1=k2=1k_1=k_2=1. In particular, notice that Theorem 1.5 improves the previous dimensional inequalities from [20] and [1] under pinching conditions.

Structure of the manuscript. The paper is organized as follows. Section 2 collects the geometric notation and the properties of the weighted centroaffine Laplacian. Section 3 develops the weighted Hilbert–Brunn–Minkowski framework and discusses its relation to local concavity inequalities and spectral comparison questions. Section 4 establishes the anisotropic Poincaré and strengthened Brascamp–Lieb inequalities. Section 5 derives the dimensional concavity estimates for measures associated with a Minkowski functional and for general pinched potentials.

Further context for the results. As previously mentioned in this introduction, both the dimensional Brunn–Minkowski conjecture and the (B) conjecture are implied by the logarithmic Brunn–Minkowski conjecture. An AI-generated solution to this major problem was released by OpenAI at the beginning of October 2026. The elaboration of this manuscript and its content dates back to prior to the aforementioned release, and our proofs follow the path laid out by the various works mentioned in the introduction. We believe our results to be of independent interest for a deeper understanding of the aforementioned conjectures. Moreover, several of our statements involve measures beyond log-concave ones and, therefore, are not implied by the logarithmic Brunn–Minkowski conjecture.

Preliminaries

Throughout this work, let K⊂RnK \subset\mathbb{R}^{n} be a compact convex set whose boundary is a C2C^{2} manifold with bounded and strictly positive principal curvatures. We call a set of this type a smooth convex body and assume that the origin lies in its interior. We denote by νK:∂K→Sn−1\nu_K:\partial K\to S^{n-1} the corresponding Gauss map and by II\mathit{II} its second fundamental form. Identifying the latter with the associated shape operator, we have dνK=IId\nu_K=\mathit{II}. We refer to [31 Section 2.5] for further details on the differential geometry of smooth convex bodies.

Let hK:Sn−1→Rh_K:S^{n-1}\to\mathbb{R}, hK(ξ)=sup⁡x∈K⟨x,ξ⟩h_K(\xi)=\sup_{x\in K}\langle x,\xi\rangle, denote the support function of KK, and write hˉK=hK∘νK\bar h_K=h_K\circ\nu_K, so that hˉK\bar h_K is defined on ∂K\partial K. Following [23, 26], we equip ∂K\partial K with the centroaffine metric

gK=IIhˉK.g_K=\frac{\mathit{II}}{\bar h_K}.

For computational purposes, we work directly on ∂K\partial K. The corresponding spherical formulation is obtained through the Gauss map; see also [18, 17]. For convex bodies K,L∈KnK,L\in\mathcal{K}^{n} containing the origin in their interiors, p∈(0,1)p\in(0,1) and t∈[0,1]t\in[0,1], we define their LpL_p-sum by

(1−t)K+ptL:=⋂θ∈Sn−1{x∈Rn:⟨x,θ⟩≤((1−t)hK(θ)p+thL(θ)p)1/p}.(1-t)K+_{p}tL:=\bigcap_{\theta\in S^{n-1}}\left\{x\in\mathbb{R}^{n}:\langle x,\theta\rangle\leq\left((1-t)h_K(\theta)^p+th_L(\theta)^p\right)^{1/p}\right\}.

This is the Wulff shape associated with the function ((1−t)hKp+thLp)1/p\left((1-t)h_K^p+th_L^p\right)^{1/p}, which need not itself be a support function. The limiting case p=0p=0 is defined by replacing this function with hˉK1−thˉLt\bar h_K^{1-t}\bar h_L^t, while p=1p=1 recovers the usual Minkowski combination (1−t)K+tL(1-t)K+tL.

Let u∈C1(Rn)u\in C^{1}(\mathbb{R}^{n}) be convex, and let μ\mu be the log-concave measure defined by

dμ(x)=e−u(x) dx.d\mu(x)=e^{-u(x)}\,dx.

With a slight abuse of notation, we also use μ\mu to denote the surface measure e−u dHn−1e^{-u}\,d\mathcal{H}^{n-1} on ∂K\partial K. We define the μ\mu-cone-volume measure of KK by

dVK,μ(x)=hˉK(x)e−u(x)nμ(K) dHn−1(x).dV_{K,\mu}(x)=\frac{\bar h_K(x)e^{-u(x)}}{n\mu(K)}\,d\mathcal{H}^{n-1}(x).

For the Lebesgue measure, this agrees with the normalized cone-volume measure. In general, however, VK,μV_{K,\mu} need not be a probability measure, a distinction that will be relevant in what follows.

Building on [23, 26, 18], we define, for f∈C2(∂K)f\in C^{2}(\partial K),

ΔK,μf=hˉK∇∂K⋅(II−1∇∂Kf)+2⟨II−1∇∂Kf,∇∂KhˉK⟩−hˉK⟨II−1∇∂Kf,∇∂Ku⟩.\Delta_{K,\mu}f=\bar h_K\nabla_{\partial K}\cdot\left(\mathit{II}^{-1}\nabla_{\partial K}f\right)+2\left\langle\mathit{II}^{-1}\nabla_{\partial K}f,\nabla_{\partial K}\bar h_K\right\rangle-\bar h_K\left\langle\mathit{II}^{-1}\nabla_{\partial K}f,\nabla_{\partial K}u\right\rangle.

When u≡0u\equiv0, this reduces to the Hilbert–Brunn–Minkowski operator with the normalization of [26]. Equivalently, ΔK,μ\Delta_{K,\mu} is the weighted Laplace–Beltrami operator associated with (∂K,gK,VK,μ)(\partial K,g_K,V_{K,\mu}). Indeed, the corresponding metric gradient is

grad⁡gKf=gK−1∇∂Kf=hˉKII−1∇∂Kf,\operatorname{grad}_{g_K}f=g_K^{-1}\nabla_{\partial K}f=\bar h_K\mathit{II}^{-1}\nabla_{\partial K}f,

and integration by parts gives, for every f,g∈C2(∂K)f,g\in C^{2}(\partial K),

∫∂K(−ΔK,μf)g dVK,μ=∫∂KgK(grad⁡gKf,grad⁡gKg) dVK,μ=∫∂KhˉK⟨II−1∇∂Kf,∇∂Kg⟩ dVK,μ.(11)\begin{aligned} \int_{\partial K}(-\Delta_{K,\mu}f)g\,dV_{K,\mu} =\int_{\partial K}g_K\left(\operatorname{grad}_{g_K}f,\operatorname{grad}_{g_K}g\right)\,dV_{K,\mu} \\ &=\int_{\partial K}\bar h_K\left\langle\mathit{II}^{-1}\nabla_{\partial K}f,\nabla_{\partial K}g\right\rangle\,dV_{K,\mu}. \tag*{(11)} \end{aligned}

On the weighted Hilbert operator

In this section, we define the weighted analogue of the Hilbert operator, following [23], and study its connection to local formulations of global concavity inequalities and their spectral counterparts. Our starting point is the (p,q)(p,q)-concavity inequality

μ((1−t)K+ptL)qn≥(1−t)μ(K)qn+tμ(L)qn,(12)\mu\left((1-t)K +_{p} tL\right)^{\frac{q}{n}} \ge(1-t)\mu(K)^{\frac{q}{n}} + t\mu(L)^{\frac{q}{n}}, \tag*{(12)}

where 0≤q≤p≤10 \le q \le p \le1, K,LK,L are origin-symmetric convex bodies, and μ\mu is an even log-concave probability measure on Rn\mathbb{R}^{n} with density

dμ=e−u dx,\mathrm{d}\mu= e^{-u}\,\mathrm{d}x,

for an even, convex, lower semicontinuous function u:Rn→R∪{+∞}u:\mathbb{R}^{n} \to\mathbb{R}\cup\{+\infty\}. The case q=0q=0 is understood in the logarithmic sense. For the differential and spectral statements below, we additionally assume that uu is finite and belongs to C2(Rn)C^{2}(\mathbb{R}^{n}).

As established in [16 Lemma 2.1], the validity of (12) for all origin-symmetric convex bodies implies the following infinitesimal inequality: for every origin-symmetric K∈K2,+nK\in\mathcal{K}^{n}_{2,+} and every even f∈C1(∂K)f\in C^{1}(\partial K),

∫∂K(⟨I−1∇∂Kf,∇∂Kf⟩−Hμf2−(1−p)f2hK) dμ+n−qnμ(K)(∫∂Kf dμ)2≥0,(13)\int_{\partial K}\left(\left\langle\mathbb{I}^{-1}\nabla_{\partial K}f,\nabla_{\partial K}f\right\rangle-\mathrm{H}_{\mu}f^{2}-(1-p)\frac{f^{2}}{h_{K}}\right)\,\mathrm{d}\mu+\frac{n-q}{n\mu(K)}\left(\int_{\partial K}f\,\mathrm{d}\mu\right)^{2}\ge0, \tag*{(13)}

where Hμ=tr⁡(I)−⟨∇u,νK⟩\mathrm{H}_{\mu}=\operatorname{tr}(\mathbb{I})-\langle\nabla u,\nu_{K}\rangle. As in the preliminaries, dμ\mathrm{d}\mu on ∂K\partial K denotes e−udHn−1e^{-u}\mathrm{d}\mathcal{H}^{n-1}.

The converse implication is more delicate. In the unweighted setting, local-to-global arguments were developed using strongly isomorphic polytopes in [29] and a PDE approach in [4]. In the weighted setting, [16 Theorem 3.1] shows that, for 0≤q≤p<10\le q\le p<1, the validity of (13) for all such KK and ff implies (12) for all origin-symmetric convex bodies. For p=1p=1, this is [20 Lemma 3.1].

We begin with the observation that the local inequality takes a particularly convenient form when expressed in terms of the Riemannian structure discussed above. The following algebraic reformulation does not require symmetry.

Proposition 3.1. Let K∈K2,+nK\in\mathcal{K}^{n}_{2,+} contain the origin in its interior. For a given f∈C2(∂K)f\in C^{2}(\partial K), set ρ=f/hˉK\rho=f/\bar{h}_{K}. Then (13), evaluated at ff, can be further rewritten as

∫∂K(−ΔK,μρ)ρ dVK,μ+∫∂K[⟨∇u,x⟩−(n−p)]ρ2 dVK,μ+(n−q)(∫∂Kρ dVK,μ)2≥0.(14)\int_{\partial K}(-\Delta_{K,\mu}\rho)\rho\,\mathrm{d}V_{K,\mu}+\int_{\partial K}\left[\langle\nabla u,x\rangle-(n-p)\right]\rho^{2}\,\mathrm{d}V_{K,\mu}+(n-q)\left(\int_{\partial K}\rho\,\mathrm{d}V_{K,\mu}\right)^{2}\ge0. \tag*{(14)}

Equivalently,

∫∂K[−ΔK,μρ+⟨∇u,x⟩ρ]ρ dVK,μ≥(n−p)Var⁡VK,μ(ρ)−(p−q)(∫∂Kρ dVK,μ)2,\int_{\partial K}\left[-\Delta_{K,\mu}\rho+\langle\nabla u,x\rangle\rho\right]\rho\,\mathrm{d}V_{K,\mu}\ge(n-p)\operatorname{Var}_{V_{K,\mu}}(\rho)-(p-q)\left(\int_{\partial K}\rho\,\mathrm{d}V_{K,\mu}\right)^{2},

where, here and below with a slight abuse of notation,

Var⁡VK,μ(v):=∫∂Kv2 dVK,μ−(∫∂Kv dVK,μ)2.\operatorname{Var}_{V_{K,\mu}}(v):=\int_{\partial K}v^{2}\,\mathrm{d}V_{K,\mu}-\left(\int_{\partial K}v\,\mathrm{d}V_{K,\mu}\right)^{2}.

Proof. First, consider the substitution f=ρhˉf=\rho\bar{h}. Observe that, by the divergence theorem,

∫∂K⟨Π−1∇∂Kf,∇∂Kf⟩ dμ=−∫∂K∇∂K⋅(Π−1∇∂Kf)f dμ+∫∂K⟨Π−1∇∂Kf,∇∂Ku⟩f dμ.\int_{\partial K}\langle\boldsymbol{\Pi}^{-1}\nabla_{\partial K}f,\nabla_{\partial K}f\rangle\,d\mu=-\int_{\partial K}\nabla_{\partial K}\cdot(\boldsymbol{\Pi}^{-1}\nabla_{\partial K}f)f\,d\mu+\int_{\partial K}\langle\boldsymbol{\Pi}^{-1}\nabla_{\partial K}f,\nabla_{\partial K}u\rangle f\,d\mu.

Now, by direct computations,

∇∂K⋅(Π−1∇∂Kf)=∇∂K⋅(hˉKΠ−1∇∂Kρ)+∇∂K⋅(ρΠ−1∇∂KhˉK)=∇∂K⋅(Π−1∇Kρ)hˉK+2⟨Π−1∇∂Kρ,∇∂KhˉK⟩+ρ(n−1−tr⁡(Π)),\begin{aligned} \nabla_{\partial K}\cdot(\boldsymbol{\Pi}^{-1}\nabla_{\partial K}f)&=\nabla_{\partial K}\cdot(\bar{h}_K\boldsymbol{\Pi}^{-1}\nabla_{\partial K}\rho)+\nabla_{\partial K}\cdot(\rho\boldsymbol{\Pi}^{-1}\nabla_{\partial K}\bar{h}_K)\\ &=\nabla_{\partial K}\cdot(\boldsymbol{\Pi}^{-1}\nabla_K\rho)\bar{h}_K+2\langle\boldsymbol{\Pi}^{-1}\nabla_{\partial K}\rho,\nabla_{\partial K}\bar{h}_K\rangle+\rho(n-1-\operatorname{tr}(\boldsymbol{\Pi})), \end{aligned}

and

⟨Π−1∇∂Kf,∇∂Ku⟩=⟨Π−1∇∂KhˉK,∇∂Ku⟩ρ+⟨Π−1∇∂KhˉK,∇∂Ku⟩hˉK.\langle\boldsymbol{\Pi}^{-1}\nabla_{\partial K}f,\nabla_{\partial K}u\rangle=\langle\boldsymbol{\Pi}^{-1}\nabla_{\partial K}\bar{h}_K,\nabla_{\partial K}u\rangle\rho+\langle\boldsymbol{\Pi}^{-1}\nabla_{\partial K}\bar{h}_K,\nabla_{\partial K}u\rangle\bar{h}_K.

Recalling that

Hμ=tr⁡(Π)−⟨∇u,νK⟩H_\mu=\operatorname{tr}(\boldsymbol{\Pi})-\langle\nabla u,\nu_K\rangle

and

⟨∇u,x⟩=⟨∇u,νK⟩hˉK+⟨∇∂Ku,Π−1∇∂KhˉK⟩,\langle\nabla u,x\rangle=\langle\nabla u,\nu_K\rangle\bar{h}_K+\langle\nabla_{\partial K}u,\boldsymbol{\Pi}^{-1}\nabla_{\partial K}\bar{h}_K\rangle,

we see that the generalized mean curvature cancels and we are left with precisely with (14).

The log-concavity of μ\mu corresponds to p=1p=1 and q=0q=0, with the latter understood in the limiting sense. Since log-concavity holds without symmetry assumptions, this gives the following reformulation of the boundary Poincaré inequalities of Colesanti [5] and Kolesnikov–Milman [21, 22]. An equivalent result was independently obtained in [9 Lemma 8.2].

Theorem 3.2. Let μ\mu be a log-concave measure on Rn\mathbb{R}^{n} with density dμ=e−u dxd\mu=e^{-u}\,dx, where u∈C2(Rn)u\in C^{2}(\mathbb{R}^{n}) is convex. Then, for every K∈K2,+nK\in\mathcal{K}^{n}_{2,+} containing the origin in its interior and every ρ∈C1(∂K)\rho\in C^{1}(\partial K),

∫∂K(n−1−⟨∇u,x⟩)ρ2 dVK,μ−n(∫∂Kρ dVK,μ)2≤∫∂KgK(grad⁡gKρ,grad⁡gKρ) dVK,μ.(15)\int_{\partial K}\left(n-1-\langle\nabla u,x\rangle\right)\rho^{2}\,dV_{K,\mu}-n\left(\int_{\partial K}\rho\,dV_{K,\mu}\right)^{2}\leq\int_{\partial K}g_K(\operatorname{grad}_{g_K}\rho,\operatorname{grad}_{g_K}\rho)\,dV_{K,\mu}. \tag*{(15)}

Proof. It is immediate in virtue of Proposition 3.1 with p=1p=1 and q=0q=0, together with (11).

The preceding formulations admit a natural spectral interpretation. We now return to even measures and origin-symmetric bodies and introduce the corresponding operator.

Definition 3.3 (Weighted Hilbert operator). Let μ\mu be an even log-concave measure on Rn\mathbb{R}^{n} with density dμ=e−u dxd\mu=e^{-u}\,dx, where u∈C2(Rn)u\in C^{2}(\mathbb{R}^{n}), and let K∈K2,+nK\in\mathcal{K}^{n}_{2,+} be origin-symmetric. We define the weighted Hilbert–Brunn–Minkowski operator associated with KK and μ\mu by

DK,μf=−ΔK,μf+⟨∇u,x⟩f.D_{K,\mu}f=-\Delta_{K,\mu}f+\langle\nabla u,x\rangle f.

Notice that an integration by parts gives

∫∂K(DK,μf)f dVK,μ=∫∂K[gK(grad⁡gKf,grad⁡gKf)+⟨∇u,x⟩f2]dVK,μ.\int_{\partial K}(D_{K,\mu}f)f\,dV_{K,\mu}=\int_{\partial K}\left[g_K(\operatorname{grad}_{g_K}f,\operatorname{grad}_{g_K}f)+\langle\nabla u,x\rangle f^{2}\right]dV_{K,\mu}.

Therefore, DK,μ\mathcal{D}_{K,\mu} is self-adjoint in the Sobolev space H1(∂K,VK,μ)H^1(\partial K,V_{K,\mu}). In particular, it is reasonable to introduce the following spectral quantity:

λ1,e(DK,μ):=inf⁡f∈Ce2(∂K)Var⁡VK,μ(f)>0JK,μ(f)=inf⁡f∈Ce2(∂K)Var⁡VK,μ(f)>0∫∂K(DK,μf)f dVK,μVar⁡VK,μ(f).\lambda_{1,e}(\mathcal{D}_{K,\mu}) := \inf_{\substack{f\in C_e^2(\partial K)\\ \operatorname{Var}_{V_{K,\mu}}(f)>0}} \mathcal{J}_{K,\mu}(f) = \inf_{\substack{f\in C_e^2(\partial K)\\ \operatorname{Var}_{V_{K,\mu}}(f)>0}} \frac{\int_{\partial K}(\mathcal{D}_{K,\mu}f)f\,\mathrm{d}V_{K,\mu}}{\operatorname{Var}_{V_{K,\mu}}(f)}.

Here Ce2(∂K)C_e^2(\partial K) denotes the even C2C^2 functions on ∂K\partial K. Unlike in the unweighted setting [22, 26], constants do not, in general, belong to the kernel of DK,μ\mathcal{D}_{K,\mu}. Consequently, one cannot impose a zero mean constraint in this variational problem. Furthermore, notice that the divergence theorem gives

VK,μ(∂K)=1nμ(K)∫∂Kh‾K dμ=1−1nμ(K)∫K⟨∇u,x⟩ dμ≤1.V_{K,\mu}(\partial K)=\frac{1}{n\mu(K)}\int_{\partial K}\overline{h}_K\,\mathrm{d}\mu=1-\frac{1}{n\mu(K)}\int_K\langle\nabla u,x\rangle\,\mathrm{d}\mu\leq1.

where the last inequality follows from the convexity and evenness of uu. In the case that uu is strictly convex the above inequality is strict. Indeed, since ⟨∇u,x⟩≥0\langle\nabla u,x\rangle\geq0 if VK,μ(∂K)=1V_{K,\mu}(\partial K)=1 we would have

⟨∇u(x)−∇u(0),x−0⟩=0,∀x∈K.\langle\nabla u(x)-\nabla u(0),x-0\rangle=0,\qquad\forall x\in K.

But then strict convexity would imply that x=0x=0, which is a contradiction. Writing m=VK,μ(∂K)∈(0,1]m=V_{K,\mu}(\partial K)\in(0,1], the Cauchy–Schwarz inequality yields

(∫∂Kf dVK,μ)2≤m∫∂Kf2 dVK,μ.\left(\int_{\partial K}f\,\mathrm{d}V_{K,\mu}\right)^2\leq m\int_{\partial K}f^2\,\mathrm{d}V_{K,\mu}.

Thus,

Var⁡VK,μ(f)=∫∂Kf2 dVK,μ−(∫∂Kf dVK,μ)2≥(1−m)∫∂Kf2 dVK,μ≥1−mm(∫∂Kf dVK,μ)2≥0.\begin{aligned} \operatorname{Var}_{V_{K,\mu}}(f)=\int_{\partial K}f^2\,\mathrm{d}V_{K,\mu}-\left(\int_{\partial K}f\,\mathrm{d}V_{K,\mu}\right)^2\geq(1-m)\int_{\partial K}f^2\,\mathrm{d}V_{K,\mu} \\ &\geq\frac{1-m}{m}\left(\int_{\partial K}f\,\mathrm{d}V_{K,\mu}\right)^2\geq0. \end{aligned}

Thus for strictly convex uu, Var⁡VK,μ(f)>0\operatorname{Var}_{V_{K,\mu}}(f)>0 for every nonzero continuous ff.

Finally, when one discusses concavity principles for log-concave measures it is enough to assume that the function uu is strictly convex, via a usual approximation argument, so we focus our attention to this case in order to avoid ambiguity.

With these preliminaries in place, we use the above operator to set a spectral framework for studying dimensional concavity properties of even log-concave measures, as illustrated by the following characterization, which is Theorem 1.1 in the introduction.

Theorem 3.4. Let μ\mu be an even log-concave measure on Rn\mathbb{R}^n with density dμ=e−u dx\mathrm{d}\mu=e^{-u}\,\mathrm{d}x, where u∈C2(Rn)u\in C^2(\mathbb{R}^n). For p∈[0,1]p\in[0,1], the inequality

μ((1−t)K+ptL)pn≥(1−t)μ(K)pn+tμ(L)pn\mu\left((1-t)K+_p tL\right)^{\frac{p}{n}}\geq(1-t)\mu(K)^{\frac{p}{n}}+t\mu(L)^{\frac{p}{n}}

holds for all origin-symmetric convex bodies K,LK,L and every t∈[0,1]t\in[0,1] if and only if

λ1,e(DC,μ)≥n−p\lambda_{1,e}(\mathcal{D}_{C,\mu})\geq n-p

for every origin-symmetric C∈K2,+nC \in\mathcal{K}_{2,+}^n. The case p=0p = 0 is understood in the limiting sense as

log⁡μ((1−t)K+0tL)≥(1−t)log⁡μ(K)+tlog⁡μ(L).\log\mu((1-t)K +_0 tL) \ge(1-t)\log\mu(K) + t\log\mu(L).

Furthermore given an origin-symmetric convex set KK, we have that

t↦log⁡μ(etK),t \mapsto\log\mu(e^tK),

is concave at t=0t = 0, if and only if

JK,μ(1)≥n.\mathcal{J}_{K,\mu}(1) \ge n.

In particular the (B)(B) property holds for the measure μ\mu if and only if JK,μ(1)≥n\mathcal{J}_{K,\mu}(1) \ge n, for every origin symmetric convex set K∈K2,+nK \in\mathcal{K}_{2,+}^n.

Proof. For the first part, setting q=pq = p in Proposition 3.1, the local inequality becomes

∫∂C(DC,μf)f dVC,μ≥(n−p)Var⁡VC,μ(f).\int_{\partial C}(\mathcal{D}_{C,\mu}f)f\,dV_{C,\mu} \ge(n-p)\operatorname{Var}_{V_{C,\mu}}(f).

For functions with Var⁡VC,μ(f)>0\operatorname{Var}_{V_{C,\mu}}(f) > 0, this is exactly the bound on the Rayleigh quotient, which is the case for strictly convex potentials uu. For p<1p < 1, the equivalence with the global inequality follows from Lemma 3.1 in [16]. For p=1p = 1, one can refer to [20].

For the second part, fixing a convex set KK, consider the functional

F(t)=μ(etK),t∈R.F(t) = \mu(e^tK), t \in\mathbb{R}.

It is well known, see e.g. [7, 10], that

F′(0)=∫Kn−⟨Du,x⟩ dμ=nμ(K)∫∂KhK(νK) dVK,μF'(0) = \int_K n-\langle Du,x\rangle\,d\mu= n\mu(K)\int_{\partial K} h_K(\nu_K)\,dV_{K,\mu}

and

F′′(0)=μ(K)∫K⟨Du,x⟩2−⟨(D2u)x,x⟩−⟨Du,x⟩ dμ.F''(0) = \mu(K)\int_K \langle Du,x\rangle^2-\langle(D^2u)x,x\rangle-\langle Du,x\rangle\,d\mu.

Thus the inequality (log⁡(F(t)))′′(0)≤0(\log(F(t)))''(0) \le0, is equivalent to

∫K⟨(D2u)x,x⟩+⟨Du,x⟩ dμ≥∫K⟨Du,x⟩2 dμ−1μ(K)(∫K⟨Du,x⟩ dμ)2(16)\int_K \langle(D^2u)x,x\rangle+\langle Du,x\rangle\,d\mu\ge\int_K \langle Du,x\rangle^2\,d\mu-\frac{1}{\mu(K)}\left(\int_K \langle Du,x\rangle\,d\mu\right)^2 \tag*{(16)}

On the other hand, JK,μ(1)≥n\mathcal{J}_{K,\mu}(1) \ge n can be expanded as

1nμ(K)∫∂K⟨Du,x⟩⟨x,νK(x)⟩ dμ≥1μ(K)∫∂K⟨x,νK(x)⟩ dμ−1nμ(K)2(∫∂K⟨x,νK(x)⟩ dμ)2(17)\frac{1}{n\mu(K)}\int_{\partial K}\langle Du,x\rangle\langle x,\nu_K(x)\rangle\,d\mu\ge\frac{1}{\mu(K)}\int_{\partial K}\langle x,\nu_K(x)\rangle\,d\mu-\frac{1}{n\mu(K)^2}\left(\int_{\partial K}\langle x,\nu_K(x)\rangle\,d\mu\right)^2 \tag*{(17)}

For the term on the left hand side, an integration by parts yields

∫∂K⟨Du,x⟩⟨x,νK(x)⟩e−u dx=∫K⟨Du,x⟩Lμ(∣x∣22) dμ+∫K⟨D(⟨Du,x⟩),x⟩ dμ=n∫K⟨Du,x⟩ dμ−∫K⟨Du,x⟩2 dμ+∫K⟨(D2u)x,x⟩ dμ+∫K⟨Du,x⟩ dμ\begin{aligned} \int_{\partial K}\langle Du,x\rangle\langle x,\nu_K(x)\rangle e^{-u}\,dx &= \int_K \langle Du,x\rangle\mathcal{L}_{\mu}\left(\frac{|x|^2}{2}\right)\,d\mu+\int_K\langle D(\langle Du,x\rangle),x\rangle\,d\mu\\ &= n\int_K\langle Du,x\rangle\,d\mu-\int_K\langle Du,x\rangle^2\,d\mu\\ &\quad+\int_K\langle(D^2u)x,x\rangle\,d\mu+\int_K\langle Du,x\rangle\,d\mu \end{aligned}

For the right hand side of (17), we need only employ the following direct application of the divergence theorem

∫∂K⟨x,νK(x)⟩ dμ=nμ(K)−∫K⟨Du,x⟩ dμ.\int_{\partial K} \langle x,\nu_K(x)\rangle\,d\mu=n\mu(K)-\int_K\langle Du,x\rangle\,d\mu.

which yields that the right hand side of (17) is equal to

∫K⟨Du,x⟩ dμ−1nμ(K)(∫K⟨Du,x⟩2).\int_K\langle Du,x\rangle\,d\mu-\frac{1}{n\mu(K)}\left(\int_K\langle Du,x\rangle^2\right).

Putting everything together, 1n\frac{1}{n} cancels out and we are left precisely with (16), concluding the proof.

Anisotropic Poincaré and strong Brascamp–Lieb inequalities for gauge-symmetric measures

We start with a simple lemma to split the Hessian of inverse quadratic gauge potentials into radial and tangential components, with respect to the anisotropic Riemannian metric introduced in [18]. We remark that a proof can be given via a usage of Lemma 2.2 therein. See, for example, [9 Lemma 8.1]. We propose an independent proof.

Lemma 4.1. For KK smooth convex body, consider the function Φ(x)=∥x∥K22\Phi(x)=\frac{\lVert x\rVert_K^2}{2}. For every φ∈C1(Rn)\varphi\in C^1(\mathbb{R}^n) and y∈∂Ky\in\partial K, we have

⟨(D2Φ(y))−1Dφ(y),Dφ(y)⟩=⟨Dφ(y),y⟩2+gK(grad⁡gKφ(y),grad⁡gKφ(y)).\left\langle(D^2\Phi(y))^{-1}D\varphi(y),D\varphi(y)\right\rangle=\langle D\varphi(y),y\rangle^2+g_K(\operatorname{grad}_{g_K}\varphi(y),\operatorname{grad}_{g_K}\varphi(y)).

Proof. We first observe that, by Fenchel–Legendre duality,

(D2Φ(x))−1=D2(hK22)(DΦ(x))(D^2\Phi(x))^{-1}=D^2\left(\frac{h_K^2}{2}\right)(D\Phi(x))

for every x∈Rn∖{o}x\in\mathbb{R}^n\setminus\{o\}. The equality above can be found, for example, in [31 (1.49), Section 1.7]. Then, direct computations show

D2(hK22)(DΦ(x))=(hKD2hK+DhK⊗DhK)∣z=DΦ(x).D^2\left(\frac{h_K^2}{2}\right)(D\Phi(x))=\left(h_KD^2h_K+Dh_K\otimes Dh_K\right)\bigg|_{z=D\Phi(x)}.

If we now choose x=y∈∂Kx=y\in\partial K, by the properties of gauge functions DΦ(y)=νK(y)hˉK(y)D\Phi(y)=\frac{\nu_K(y)}{\bar h_K(y)}. Denoting by πz\pi_z, z∈Rn∖{o}z\in\mathbb{R}^n\setminus\{o\}, the orthogonal projection along zz, it is well-known that (compare, e.g., [31 Lemma 2.5.1])

D2hK(z)=∣z∣−1πztD2hK(z/∣z∣)πz,D^2h_K(z)=|z|^{-1}\pi_z^tD^2h_K(z/|z|)\pi_z,

where now we consider these objects as bilinear forms on Rn\mathbb{R}^n. In particular,

D2hK(νK(y)/hˉ(y))=hˉK(y)(πνK(y)tD2hK(νK(y))πνK(y)).D^2h_K(\nu_K(y)/\bar h(y))=\bar h_K(y)\left(\pi_{\nu_K(y)}^tD^2h_K(\nu_K(y))\pi_{\nu_K(y)}\right).

As hKh_K is 1-homogeneous and ∇hK\nabla h_K is 0-homogeneous, we infer

(D2Φ(y))−1=hˉK(y)(πνK(y)tD2hK(νK(y))πνK(y))+y⊗y,(D^2\Phi(y))^{-1}=\bar h_K(y)\left(\pi_{\nu_K(y)}^tD^2h_K(\nu_K(y))\pi_{\nu_K(y)}\right)+y\otimes y,

and the proof is concluded by direct computations, recalling that, for every y∈∂Ky\in\partial K,

D2hK(νK(y))=Π−1(y).D^2h_K(\nu_K(y))=\Pi^{-1}(y).

Finally, let us recall the following 1-dimensional Poincaré inequality, which is [10 Lemma 8].

Lemma 4.2. Consider v,w∈C([0,∞))∩C1((0,∞))v,w\in C([0,\infty))\cap C^{1}((0,\infty)), and f∈Cc1([0,∞))f\in C^{1}_{c}([0,\infty)) such that f(0)=0f(0)=0. Then, for every α≥0\alpha\geq0

∫0∞w′tf2tαe−w−v dt≤∫0∞((f′)2+α(ft)2−v′f2t)tαe−w−v dt.\int_{0}^{\infty}\frac{w'}{t}f^{2}t^{\alpha}e^{-w-v}\,\mathrm{d}t\leq\int_{0}^{\infty}\left((f')^{2}+\alpha\left(\frac{f}{t}\right)^{2}-v'\frac{f^{2}}{t}\right)t^{\alpha}e^{-w-v}\,\mathrm{d}t.

With these results at hand, we can now prove the main result of this section, which can be understood as the anisotropic analogue of [10 Theorem 4]. An equivalent formulation was recently proved in [9 Theorem 1.8].

Theorem 4.3. Consider an increasing function w∈C1(0,∞)w\in C^{1}(0,\infty) and KK an even and smooth convex body. Let u:Rn→Ru:\mathbb{R}^{n}\to\mathbb{R} be convex, then denoting by η\eta the measure on Rn\mathbb{R}^{n} satisfying dη(x)=e−u(x)−w(∥x∥K) dx\mathrm{d}\eta(x)=e^{-u(x)-w(\lVert x\rVert_{K})}\,\mathrm{d}x, for every odd function φ∈C1(Rn)\varphi\in C^{1}(\mathbb{R}^{n}) we have

∫Rnw′(∥x∥K)∥x∥Kφ2(x) dη(x)≤∫Rn⟨(D2Φ(x))−1Dφ(x),Dφ(x)⟩ dη(x),(18)\int_{\mathbb{R}^{n}}\frac{w'(\lVert x\rVert_{K})}{\lVert x\rVert_{K}}\varphi^{2}(x)\,\mathrm{d}\eta(x)\leq\int_{\mathbb{R}^{n}}\left\langle\left(D^{2}\Phi(x)\right)^{-1}D\varphi(x),D\varphi(x)\right\rangle\,\mathrm{d}\eta(x), \tag*{(18)}

where Φ(x)=∥x∥K2/2\Phi(x)=\lVert x\rVert_{K}^{2}/2.

Proof. We start by observing that, using polar coordinates, we can rewrite the left-hand side of (18) as

∫Rnw′(∥x∥K)∥x∥Kφ2(x) dη(x)=∫∂K∫0∞w′(r)rφ2(ry)rn−1e−u(ry)−w(r)⟨νK(y),y⟩ dr dHn−1(y).(19)\int_{\mathbb{R}^{n}}\frac{w'(\lVert x\rVert_{K})}{\lVert x\rVert_{K}}\varphi^{2}(x)\,\mathrm{d}\eta(x)=\int_{\partial K}\int_{0}^{\infty}\frac{w'(r)}{r}\varphi^{2}(ry)r^{n-1}e^{-u(ry)-w(r)}\langle\nu_{K}(y),y\rangle\,\mathrm{d}r\,\mathrm{d}\mathcal{H}^{n-1}(y). \tag*{(19)}

By Lemma 4.2 with the choice α=n−1\alpha=n-1, we infer

(19)≤∫∂K∫0∞(⟨Dφ(ry),y⟩2+(n−1)(φ(ry)r)2−1r⟨Du(ry),ry⟩(φ(ry)r)2)×rn−1e−w(r)−u(ry)⟨νK(y),y⟩ dr dHn−1(y).(20)\begin{aligned} (19)&\leq\int_{\partial K}\int_{0}^{\infty}\left(\langle D\varphi(ry),y\rangle^{2}+(n-1)\left(\frac{\varphi(ry)}{r}\right)^{2}-\frac{1}{r}\langle Du(ry),ry\rangle\left(\frac{\varphi(ry)}{r}\right)^{2}\right)\\ &\qquad\qquad\times r^{n-1}e^{-w(r)-u(ry)}\langle\nu_{K}(y),y\rangle\,\mathrm{d}r\,\mathrm{d}\mathcal{H}^{n-1}(y). \tag*{(20)} \end{aligned}

Next, observe that, by (15), since hh is odd and the measure as well as ∂K\partial K is symmetric, we have that

∫∂K((n−1)(φ(ry)r)2−⟨Du(ry),ry⟩(φ(ry)r)2)⟨νK(y),y⟩e−u(ry) dHn−1(y)≤∫∂KgK(grad⁡gK(φ(ry)r),grad⁡gK(φ(ry)r))⟨νK(y),y⟩e−u(ry) dHn−1(y).(21)\begin{aligned} \int_{\partial K}\Bigg((n-1)\left(\frac{\varphi(ry)}{r}\right)^{2} -\langle Du(ry),ry\rangle\left(\frac{\varphi(ry)}{r}\right)^{2}\Bigg) \langle\nu_{K}(y),y\rangle e^{-u(ry)}\,\mathrm{d}\mathcal{H}^{n-1}(y)\\ \leq\int_{\partial K}g_{K}\left(\operatorname{grad}_{g_{K}}\left(\frac{\varphi(ry)}{r}\right),\operatorname{grad}_{g_{K}}\left(\frac{\varphi(ry)}{r}\right)\right) \langle\nu_{K}(y),y\rangle e^{-u(ry)}\,\mathrm{d}\mathcal{H}^{n-1}(y). \tag*{(21)} \end{aligned}

If now we invert the order of integration in (20) and apply the inequality (21), we deduce

(20)≤∫∂K∫0∞(⟨Dφ(ry),y⟩2+gK(grad⁡gK(φ(ry)r),grad⁡gK(φ(ry)r)))×rn−1⟨νK(y),y⟩e−u(ry)−w(r) dr dHn−1(y).\begin{aligned} (20)&\leq\int_{\partial K}\int_{0}^{\infty}\Bigg(\langle D\varphi(ry),y\rangle^{2} +g_{K}\left(\operatorname{grad}_{g_{K}}\left(\frac{\varphi(ry)}{r}\right),\operatorname{grad}_{g_{K}}\left(\frac{\varphi(ry)}{r}\right)\right)\Bigg)\\ &\qquad\qquad\times r^{n-1}\langle\nu_{K}(y),y\rangle e^{-u(ry)-w(r)}\,\mathrm{d}r\,\mathrm{d}\mathcal{H}^{n-1}(y). \end{aligned}

Finally, by Lemma 4.1 we have

⟨Dφ(ry),y⟩2+gK(grad⁡gK(φ(ry)r),grad⁡gK(φ(ry)r))=⟨(D2Φ(ry))−1Dφ(ry),Dφ(ry)⟩,\langle D\varphi(ry),y\rangle^{2}+g_{K}\left(\operatorname{grad}_{g_{K}}\left(\frac{\varphi(ry)}{r}\right),\operatorname{grad}_{g_{K}}\left(\frac{\varphi(ry)}{r}\right)\right)=\left\langle\left(D^{2}\Phi(ry)\right)^{-1}D\varphi(ry),D\varphi(ry)\right\rangle,

and the proof is concluded by integrating in polar coordinates. ∊pp

Remark 4.4. The above inequality is sharp whenever V=0V=0 and h=⟨∇Φ,θ⟩h=\langle\nabla\Phi,\theta\rangle, with θ\theta fixed. We remark also that the above inequality is a Poincaré inequality on the Hessian weighted manifold

(Rn,D2Φ,e−V−w(∥x∥K) dx)(\mathbb{R}^{n},D^{2}\Phi,e^{-V-w(\|x\|_{K})}\,dx)

for odd functions. In the case that the underlying set KK is a ball, we have that D2ΦD^{2}\Phi is the identity and thus the above inequality reduces to a Poincaré inequality in the usual Euclidean space, which is the result of Cordero-Erausquin and Rotem [10].

The usage of the centroaffine metric (or the ambient anisotropic one) albeit convenient for the computations and the proof of the inequality still does not agree fully with the geometry of Rn\mathbb{R}^{n}. This problem is remedied in the works of Milman and Ivaki–Milman [26, 18], through a pinching condition on the set KK. Following this idea we establish an improved Brascamp–Lieb inequality. The proof is an adaptation of [10 Theorem 3].

Theorem 4.5. Consider an increasing function w∈C2((0,∞))w\in C^{2}((0,\infty)) and a probability measure η\eta on Rn\mathbb{R}^{n} satisfying dη(x)=e−V(x)−w(∥x∥K) dxd\eta(x)=e^{-V(x)-w(\|x\|_{K})}\,dx, where V∈C2(Rn)V\in C^{2}(\mathbb{R}^{n}) is convex and even and KK is an origin-symmetric smooth convex body. Suppose, additionally, that

αId⁡≤D2∥x∥K22≤βId⁡\alpha\operatorname{Id}\le D^{2}\frac{\|x\|_{K}^{2}}{2}\le\beta\operatorname{Id}

for every x∈Rnx\in\mathbb{R}^{n} some 0<α<β<∞0<\alpha<\beta<\infty, and that ww satisfies

w′′(s)+γw′(s)s≥0,s>0(22)w^{\prime\prime}(s)+\gamma\frac{w^{\prime}(s)}{s}\ge0,\qquad s>0 \tag*{(22)}

where γ=α/β≤1\gamma=\alpha/\beta\le1.

Then, writing W(x)=w(∥x∥K)W(x)=w(\|x\|_{K}), for every even function φ∈C1(Rn)\varphi\in C^{1}(\mathbb{R}^{n}) we have the inequality

Var⁡η(φ)≤∫Rn⟨(D2W(x)+γw′(∥x∥K)∥x∥KD2∥x∥K22)−1Dφ(x),Dφ(x)⟩ dη(x).(23)\operatorname{Var}_{\eta}(\varphi)\le\int_{\mathbb{R}^{n}}\left\langle\left(D^{2}W(x)+\gamma\frac{w^{\prime}(\|x\|_{K})}{\|x\|_{K}}D^{2}\frac{\|x\|_{K}^{2}}{2}\right)^{-1}D\varphi(x),D\varphi(x)\right\rangle\,d\eta(x). \tag*{(23)}

Proof. The arguments mimic those in [10] which, in turn, builds upon [8]. We streamline the presentation for the convenience of the reader.

Let us begin by showing that (22) implies that

the matrix A(x)=D2W(x)+γw′(∥x∥K)∥x∥KD2∥x∥K22A(x)=D^{2}W(x)+\gamma\frac{w^{\prime}(\|x\|_{K})}{\|x\|_{K}}D^{2}\frac{\|x\|_{K}^{2}}{2} is positive semidefinite.

Indeed, by direct computations, for every x∈Rn∖{o}x\in\mathbb{R}^{n}\setminus\{o\} and writing t=∥x∥Kt=\|x\|_{K}, we have the identity

A(x)=w′(t)D2∥x∥K+w′′(t)D∥x∥K⊗D∥x∥K+γw′(t)t(tD2∥x∥K+D∥x∥K⊗D∥x∥K)=(w′′(t)+γw′(t)t)D∥x∥K⊗D∥x∥K+w′(t)(1+γ)D2∥x∥K.\begin{aligned} A(x)&=w^{\prime}(t)D^{2}\|x\|_{K}+w^{\prime\prime}(t)D\|x\|_{K}\otimes D\|x\|_{K}+\gamma\frac{w^{\prime}(t)}{t}\left(tD^{2}\|x\|_{K}+D\|x\|_{K}\otimes D\|x\|_{K}\right)\\ &=\left(w^{\prime\prime}(t)+\gamma\frac{w^{\prime}(t)}{t}\right)D\|x\|_{K}\otimes D\|x\|_{K}+w^{\prime}(t)(1+\gamma)D^{2}\|x\|_{K}. \end{aligned}

The coefficient of the first matrix is non-negative thanks to (22) while, for the second one, this follows from the monotonicity of ww and, thus, AA is positive semidefinite. This is enough to follow through with the computations in the rest of the proof, as explained in [10 Remark 9]. We may (and do) assume that φ\varphi is in addition centered, i.e. ∫φ=0\int\varphi=0 subtracting a constant if that is not the case. With these preliminary remarks in place notice that (23) is equivalent to

∫Rn⟨A−1Dφ,Dφ⟩−φ2 dη≥0.(24)\int_{\mathbb{R}^{n}} \langle A^{-1}D\varphi,D\varphi\rangle-\varphi^{2}\,\mathrm{d}\eta\ge0. \tag*{(24)}

Denoting by LηL_{\eta} the standard Laplace–Beltrami operator on Rn\mathbb{R}^{n} with respect to η\eta, Lη(C2(Rn))L_{\eta}(C^{2}(\mathbb{R}^{n})) is dense in the subspace of L2(η)L^{2}(\eta) orthogonal to constant functions. See [8 Lemma 3] for a proof. In particular, we can approximate φ\varphi with these function and, therefore, (24) would follow if we are able to show that

∫Rn⟨A−1Dφ,Dφ⟩−2Lη(u)φ+(Lη(u))2 dη=∫Rn⟨A−1Dφ,Dφ⟩−φ2+(Lη(u)−φ)2 dη≥0∀u∈Cc2(Rn).(25)\begin{aligned} \int_{\mathbb{R}^{n}}\langle A^{-1}D\varphi,D\varphi\rangle-2L_{\eta}(u)\varphi+(L_{\eta}(u))^{2}\,\mathrm{d}\eta\\ =\int_{\mathbb{R}^{n}}\langle A^{-1}D\varphi,D\varphi\rangle-\varphi^{2}+(L_{\eta}(u)-\varphi)^{2}\,\mathrm{d}\eta\ge0\quad\forall u\in C_{c}^{2}(\mathbb{R}^{n}). \tag*{(25)} \end{aligned}

We now estimate the first integral in (25). We have −∫RnLη(u)φ dη=∫Rn⟨Du,Dφ⟩ dη-\int_{\mathbb{R}^{n}}L_{\eta}(u)\varphi\,\mathrm{d}\eta=\int_{\mathbb{R}^{n}}\langle Du,D\varphi\rangle\,\mathrm{d}\eta by integration by parts. The Cauchy–Schwarz inequality for positive quadratic forms implies that

⟨A−1Dφ,Dφ⟩+⟨ADu,Du⟩≥−2⟨Du,Dφ⟩,\langle A^{-1}D\varphi,D\varphi\rangle+\langle ADu,Du\rangle\ge-2\langle Du,D\varphi\rangle,

we obtain the bound

∫Rn⟨A−1Dφ,Dφ⟩−2Lη(u)φ dη≥∫Rn−⟨ADu,Du⟩ dη.(26)\int_{\mathbb{R}^{n}}\langle A^{-1}D\varphi,D\varphi\rangle-2L_{\eta}(u)\varphi\,\mathrm{d}\eta\ge\int_{\mathbb{R}^{n}}-\langle ADu,Du\rangle\,\mathrm{d}\eta. \tag*{(26)}

Concerning the remainder, integration by parts gives

∫Rn(Lη(u))2 dη=∫Rn∥D2u∥HS2+⟨(D2V+D2W)Du,Du⟩ dη,(27)\int_{\mathbb{R}^{n}}(L_{\eta}(u))^{2}\,\mathrm{d}\eta=\int_{\mathbb{R}^{n}}\|D^{2}u\|_{\mathrm{HS}}^{2}+\langle(D^{2}V+D^{2}W)Du,Du\rangle\,\mathrm{d}\eta, \tag*{(27)}

where ∥⋅∥HS\|\cdot\|_{\mathrm{HS}} denotes the Hilbert–Schmidt norm of a matrix. Then, (26) and (27) in combination with the fact that D2VD^{2}V is positive definite, give

∫Rn⟨A−1Dφ,Dφ⟩−2Lη(u)φ+(Lη(u))2 dη≥∫Rn∥D2u∥HS2−⟨γw′(∥x∥K)∥x∥KD2∥x∥K22Du,Du⟩ dη.\begin{aligned} \int_{\mathbb{R}^{n}}\langle A^{-1}D\varphi,D\varphi\rangle-2L_{\eta}(u)\varphi+(L_{\eta}(u))^{2}\,\mathrm{d}\eta\ge{}& \int_{\mathbb{R}^{n}}\|D^{2}u\|_{\mathrm{HS}}^{2}\\ &-\left\langle\gamma\frac{w'(\|x\|_{K})}{\|x\|_{K}}D^{2}\frac{\|x\|_{K}^{2}}{2}Du,Du\right\rangle\,\mathrm{d}\eta. \end{aligned}

Therefore to conclude the proof, it suffices to show that

∫Rn∥D2u∥HS2 dη≥∫Rnγw′(∥x∥K)∥x∥K⟨D2∥x∥K22Du,Du⟩ dη∀u∈Cc2(Rn),\int_{\mathbb{R}^{n}}\|D^{2}u\|_{\mathrm{HS}}^{2}\,\mathrm{d}\eta\ge\int_{\mathbb{R}^{n}}\gamma\frac{w'(\|x\|_{K})}{\|x\|_{K}}\left\langle D^{2}\frac{\|x\|_{K}^{2}}{2}Du,Du\right\rangle\,\mathrm{d}\eta\quad\forall u\in C_{c}^{2}(\mathbb{R}^{n}),

which is indeed true, as

∫Rn∥D2u∥HS2 dη≥α∫Rn∑i=1n⟨(D2∥x∥K22)−1Dui,Dui⟩ dη≥α∫Rnw′(∥x∥K)∥x∥K∣Du∣2 dη≥∫Rnγw′(∥x∥K)∥x∥K⟨D2∥x∥K22Du,Du⟩ dη,(28)\begin{aligned} \int_{\mathbb{R}^{n}}\|D^{2}u\|_{\mathrm{HS}}^{2}\,\mathrm{d}\eta&\ge\alpha\int_{\mathbb{R}^{n}}\sum_{i=1}^{n}\left\langle\left(D^{2}\frac{\|x\|_{K}^{2}}{2}\right)^{-1}Du_{i},Du_{i}\right\rangle\,\mathrm{d}\eta\\ &\ge\alpha\int_{\mathbb{R}^{n}}\frac{w'(\|x\|_{K})}{\|x\|_{K}}|Du|^{2}\,\mathrm{d}\eta\\ &\ge\int_{\mathbb{R}^{n}}\gamma\frac{w'(\|x\|_{K})}{\|x\|_{K}}\left\langle D^{2}\frac{\|x\|_{K}^{2}}{2}Du,Du\right\rangle\,\mathrm{d}\eta, \tag*{(28)} \end{aligned}

where uiu_i denotes the ii-th derivative of uu and, in the second step, we have used Theorem 4.3. □

We also have the following variant of the above inequality that can be deduced by a direct modification of the final step in the proof above (28), when the function ww is assumed to be in addition convex.

Theorem 4.6. Consider an increasing and convex function w∈C2((0,∞))w \in C^2((0,\infty)) and a probability measure η\eta on Rn\mathbb{R}^n satisfying dη(x)=e−V(x)−w(∥x∥K)dx\mathrm{d}\eta(x)=e^{-V(x)-w(\lVert x\rVert_K)}\mathrm{d}x, where V∈C2(Rn)V \in C^2(\mathbb{R}^n) is convex and even and KK is an origin-symmetric smooth convex body and set W(x)=w(∥x∥K)W(x)=w(\lVert x\rVert_K). Assuming that

D2∥x∥K22≥αId⁡,D^2\frac{\lVert x\rVert_K^2}{2} \ge\alpha\operatorname{Id},

then for every even function φ∈C1(Rn)\varphi\in C^1(\mathbb{R}^n),

Var⁡η(φ)≤∫Rn⟨(D2W(x)+αw′(∥x∥K)∥x∥KId⁡)−1Dφ(x),Dφ(x)⟩ dη(x).\operatorname{Var}_{\eta}(\varphi) \le\int_{\mathbb{R}^n} \left\langle\left(D^2W(x)+\alpha\frac{w'(\lVert x\rVert_K)}{\lVert x\rVert_K}\operatorname{Id}\right)^{-1}D\varphi(x),D\varphi(x)\right\rangle\,\mathrm{d}\eta(x).

Remark 4.7. Whenever KK is a C2C^2 smooth convex body that is origin symmetric, with strictly positive principal curvatures we may always find numbers α,β\alpha,\beta such that the above pinching or one sided control is satisfied.

Improved dimensional concavity for gauge-symmetric and pinched measures

The first aim of this section is to prove a dimensional Brunn–Minkowski inequality for the following type of measures. We consider W(x)=w(∥x∥K)W(x)=w(\lVert x\rVert_K), where ww is chosen as in Theorem 4.5, and the corresponding measure μ\mu such that dμ(x)=e−W(x)dx\mathrm{d}\mu(x)=e^{-W(x)}\mathrm{d}x. Again with the notation Φ(x)=∥x∥K2/2\Phi(x)=\lVert x\rVert_K^2/2, by homogeneity Φ\Phi we have DΦ(x)=(D2Φ(x))xD\Phi(x)=(D^2\Phi(x))x and the Laplace–Beltrami operator of μ\mu can be written as

Lμ(φ)=Δφ−w′(∥x∥K)∥x∥K⟨(D2Φ)x,Dφ⟩.L_{\mu}(\varphi)=\Delta\varphi-\frac{w'(\lVert x\rVert_K)}{\lVert x\rVert_K}\left\langle(D^2\Phi)x,D\varphi\right\rangle.

Consider now the measure η\eta in Theorem 4.5 and define, for φ,ψ∈Cc2(Rn)\varphi,\psi\in C_c^2(\mathbb{R}^n), the bilinear form

Qη(φ,ψ):=∫Rntr⁡(D2φ(D2Φ)−1D2ψ)−w′(∥x∥K)∥x∥K⟨Dφ,Dψ⟩ dη.Q_{\eta}(\varphi,\psi):=\int_{\mathbb{R}^n}\operatorname{tr}\left(D^2\varphi(D^2\Phi)^{-1}D^2\psi\right)-\frac{w'(\lVert x\rVert_K)}{\lVert x\rVert_K}\langle D\varphi,D\psi\rangle\,\mathrm{d}\eta.

Observe that, by Theorem 4.3, QηQ_{\eta} is positive semidefinite on the subspace of even functions. The operators QηQ_{\eta} and LμL_{\mu} are connected by the relation

Qη(φ,Φ)=∫RnLμ(φ) dη.(29)Q_{\eta}(\varphi,\Phi)=\int_{\mathbb{R}^n}L_{\mu}(\varphi)\,\mathrm{d}\eta. \tag*{(29)}

which is instrumental in the proof of the following result.

Theorem 5.1. Consider an origin-symmetric and smooth convex body K⊂RnK \subset\mathbb{R}^n such that the corresponding quadratic potential Φ\Phi satisfies

αId⁡≤D2Φ≤βId⁡(30)\alpha\operatorname{Id}\le D^2\Phi\le\beta\operatorname{Id} \tag*{(30)}

for some 0<α≤β<∞0<\alpha\le\beta<\infty. Consider, moreover, an increasing function w∈C2((0,∞))w\in C^2((0,\infty)) satisfying (22), and write W(x)=w(∥x∥K)W(x)=w(\lVert x\rVert_K). Then, for every convex function V∈C2(Rn)V\in C^2(\mathbb{R}^n) such that the measure η\eta with density dη(x)=e−V(x)−W(x)dx\mathrm{d}\eta(x)=e^{-V(x)-W(x)}\mathrm{d}x is a probability measure we have

∫Rn∥D2φ∥HS2+⟨(D2W)Dφ,Dφ⟩dη≥γn(∫RnLμ(φ)dη)2(31)\int_{\mathbb{R}^n}\lVert D^2\varphi\rVert_{\mathrm{HS}}^2+\left\langle(D^2W)D\varphi,D\varphi\right\rangle\mathrm{d}\eta\ge\frac{\gamma}{n}\left(\int_{\mathbb{R}^n}\mathcal{L}_\mu(\varphi)\mathrm{d}\eta\right)^2 \tag*{(31)}

for every even φ∈Cc2(Rn)\varphi\in C_c^2(\mathbb{R}^n).

Proof. Fix φ∈Cc2(Rn)\varphi\in C_c^2(\mathbb{R}^n) even. We start by observing that, by the pinching condition (30),

∫Rn∥D2φ∥HS2+⟨(D2W)Dφ,Dφ⟩dη≥∫Rnαtr⁡(D2φ(D2Φ)−1D2φ)+⟨(D2W)Dφ,Dφ⟩dη≥αQη(φ,φ)+∫Rn⟨(D2W+γw′(∥x∥K)∥x∥KD2Φ)Dφ,Dφ⟩dη≥αQη(φ,φ),(32)\begin{aligned} \int_{\mathbb{R}^n}\lVert D^2\varphi\rVert_{\mathrm{HS}}^2+\left\langle(D^2W)D\varphi,D\varphi\right\rangle\mathrm{d}\eta &\ge\int_{\mathbb{R}^n}\alpha\operatorname{tr}\left(D^2\varphi(D^2\Phi)^{-1}D^2\varphi\right)+\left\langle(D^2W)D\varphi,D\varphi\right\rangle\mathrm{d}\eta\\ &\ge\alpha Q_\eta(\varphi,\varphi)+\int_{\mathbb{R}^n}\left\langle\left(D^2W+\gamma\frac{w'(\lVert x\rVert_K)}{\lVert x\rVert_K}D^2\Phi\right)D\varphi,D\varphi\right\rangle\mathrm{d}\eta\\ &\ge\alpha Q_\eta(\varphi,\varphi), \tag*{(32)} \end{aligned}

where the last step follows from (22) as in Theorem 4.5.

Next, we note that, by the bilinearity of QηQ_\eta and (29), the function ψ=φ+λΦ\psi=\varphi+\lambda\Phi, λ∈R\lambda\in\mathbb{R}, satisfies

Qη(ψ,ψ)=Qη(φ,φ)+2λ∫RnLμ(φ)dη+λ2∫RnLμ(Φ)dη.Q_\eta(\psi,\psi)=Q_\eta(\varphi,\varphi)+2\lambda\int_{\mathbb{R}^n}\mathcal{L}_\mu(\varphi)\mathrm{d}\eta+\lambda^2\int_{\mathbb{R}^n}\mathcal{L}_\mu(\Phi)\mathrm{d}\eta.

Then, the assumed estimate Lμ(Φ)≤βn\mathcal{L}_\mu(\Phi)\le\beta n implies

Qη(φ,φ)≥Qη(ψ,ψ)−2λ∫RnLμ(φ)dη−λ2βnQ_\eta(\varphi,\varphi)\ge Q_\eta(\psi,\psi)-2\lambda\int_{\mathbb{R}^n}\mathcal{L}_\mu(\varphi)\mathrm{d}\eta-\lambda^2\beta n

for every λ∈R\lambda\in\mathbb{R}. Optimizing in λ\lambda yields

Qη(φ,φ)≥Qη(ψ,ψ)+1βn(∫RnLμ(φ)dη)2≥1βn(∫RnLμ(φ)dη)2,Q_\eta(\varphi,\varphi)\ge Q_\eta(\psi,\psi)+\frac{1}{\beta n}\left(\int_{\mathbb{R}^n}\mathcal{L}_\mu(\varphi)\mathrm{d}\eta\right)^2\ge\frac{1}{\beta n}\left(\int_{\mathbb{R}^n}\mathcal{L}_\mu(\varphi)\mathrm{d}\eta\right)^2,

where the last line follows since QηQ_\eta is positive semidefinite for the functions at hand. The proof of (31) is concluded comparing this last estimate with (32).

To show that a dimensional Brunn–Minkowski inequality holds for the measure μ\mu with concavity γ/n\gamma/n, we invoke the following result by Kolesnikov and Livshyts, which is part of [20 Lemma 2.3].

Theorem 5.2. Let μˉ\bar{\mu} be an even measure with density dμ(x)=e−U(x)dx\mathrm{d}\mu(x)=e^{-U(x)}\mathrm{d}x such that U∈C2(Rn)U\in C^2(\mathbb{R}^n), suppose that for every origin symmetric convex body A⊂RnA\subset\mathbb{R}^n and every φ∈C2(Rn)\varphi\in C^2(\mathbb{R}^n) with Lμˉ(φ)=1\mathcal{L}_{\bar{\mu}}(\varphi)=1 on AA we have that

∫A∥D2φ∥HS2+⟨(D2U)Dφ,Dφ⟩dμˉ≥p.(33)\int_A\lVert D^2\varphi\rVert_{\mathrm{HS}}^2+\left\langle(D^2U)D\varphi,D\varphi\right\rangle\mathrm{d}\bar{\mu}\ge p. \tag*{(33)}

Then,

μ((1−t)A+tB)p≥(1−t)μ(A)p+tμ(A)p,\mu((1-t)A+tB)^p\ge(1-t)\mu(A)^p+t\mu(A)^p,

for all origin-symmetric convex bodies A,BA,B.

With Theorem 5.2 at hand, the inequality

μ((1−t)A+tB)γ/n≥(1−t)μ(A)γ/n+tμ(B)γ/n\mu((1-t)A+tB)^{\gamma/n}\ge(1-t)\mu(A)^{\gamma/n}+t\mu(B)^{\gamma/n}

for all origin-symmetric convex bodies A,B⊂RnA,B\subset\mathbb{R}^{n} follows by (31) and standard approximation arguments, providing a proof of Theorem 1.4.

We finish this section by exporting the perturbative argument in the proof of Theorem 5.1 to more general pinched log-concave measures. Its proof is completely independent from most of the machinery developed above, but it came as a natural consequence of the procedures in the previous sections. The result improves the dependence on the pinching as established in [20] and [1].

Theorem 5.3. Consider an even and convex function V∈C2(Rn)V\in C^{2}(\mathbb{R}^{n}) such that

k1Id⁡≤D2VandΔV≤k2nk_{1}\operatorname{Id}\le D^{2}V\quad\text{and}\quad\Delta V\le k_{2}n

for some 0<k1≤k2<∞0<k_{1}\le k_{2}<\infty, and write r=k1/k2r=k_{1}/k_{2}. Denoting by μˉ\bar{\mu} the log-concave measure with density dμˉ(x)=e−V(x)dx\mathrm{d}\bar{\mu}(x)=e^{-V(x)}\mathrm{d}x we have, for every t∈[0,1]t\in[0,1],

μˉ((1−t)A+tB)r/n≥(1−t)μˉ(K)r/n+tμˉ(L)r/n(34)\bar{\mu}((1-t)A+tB)^{r/n}\ge(1-t)\bar{\mu}(K)^{r/n}+t\bar{\mu}(L)^{r/n} \tag*{(34)}

for every origin-symmetric convex bodies A,B⊂RnA,B\subset\mathbb{R}^{n}.

Proof. In order to prove (34), the strategy is again to prove that condition (33) holds for every origin-symmetric and smooth convex body A⊂RnA\subset\mathbb{R}^{n} and apply the conclusion of Theorem 5.2. The statement follows then by density.

Fix a smooth and origin symmetric convex body AA. Similarly as we did with QηQ_{\eta} at the beginning of this section, for φ,ψ∈Cc2(Rn)\varphi,\psi\in C_{c}^{2}(\mathbb{R}^{n}) we define the bilinear form

RV(φ,ψ)=∫Atr⁡(D2φ(D2V)−1D2ψ)−⟨Dφ,Dψ⟩ dμˉ.R_{V}(\varphi,\psi)=\int_{A}\operatorname{tr}\left(D^{2}\varphi\left(D^{2}V\right)^{-1}D^{2}\psi\right)-\langle D\varphi,D\psi\rangle\,\mathrm{d}\bar{\mu}.

Observe that RVR_{V} is positive definite on odd functions by the Brascamp–Lieb variance inequality. For φ\varphi fixed, consider the linear perturbation ψ=φ+λV\psi=\varphi+\lambda V, λ∈R\lambda\in\mathbb{R}. Then,

RV(ψ,ψ)=RV(φ,φ)+2λ∫ALμˉ(φ) dμˉ+λ2∫ALμˉ(V) dμˉR_{V}(\psi,\psi)=R_{V}(\varphi,\varphi)+2\lambda\int_{A}\mathcal{L}_{\bar{\mu}}(\varphi)\,\mathrm{d}\bar{\mu}+\lambda^{2}\int_{A}\mathcal{L}_{\bar{\mu}}(V)\,\mathrm{d}\bar{\mu}

and, since ΔV≤k2n\Delta V\le k_{2}n, we infer

RV(φ,φ)≥RV(ψ,ψ)−2λ∫ALμˉ(φ) dμˉ−λ2k2nR_{V}(\varphi,\varphi)\ge R_{V}(\psi,\psi)-2\lambda\int_{A}\mathcal{L}_{\bar{\mu}}(\varphi)\,\mathrm{d}\bar{\mu}-\lambda^{2}k_{2}n

for every λ∈R\lambda\in\mathbb{R}. Optimizing in λ\lambda,

RV(φ,φ)≥RV(ψ,ψ)+1k2n(∫ALμˉ(φ) dμˉ)2≥1nk2(∫ALμˉ(φ) dμˉ)2.(35)R_{V}(\varphi,\varphi)\ge R_{V}(\psi,\psi)+\frac{1}{k_{2}n}\left(\int_{A}\mathcal{L}_{\bar{\mu}}(\varphi)\,\mathrm{d}\bar{\mu}\right)^{2}\ge\frac{1}{nk_{2}}\left(\int_{A}\mathcal{L}_{\bar{\mu}}(\varphi)\,\mathrm{d}\bar{\mu}\right)^{2}. \tag*{(35)}

Finally, using k1Id⁡≤D2Vk_{1}\operatorname{Id}\le D^{2}V and (35), we estimate

∫A∥D2φ∥HS2+⟨(D2V)Dφ,Dφ⟩ dμˉ≥∫A∥D2φ∥HS2 dμˉ≥k1∫Atr⁡(D2φ(D2V)−1D2φ) dμˉ≥k1RV(φ,φ)≥rn(∫ALμˉ(φ) dμˉ)2,\begin{aligned} \int_{A}\|D^{2}\varphi\|_{\mathrm{HS}}^{2}+\langle(D^{2}V)D\varphi,D\varphi\rangle\,\mathrm{d}\bar{\mu}\ge\int_{A}\|D^{2}\varphi\|_{\mathrm{HS}}^{2}\,\mathrm{d}\bar{\mu} \\ &\ge k_{1}\int_{A}\operatorname{tr}\left(D^{2}\varphi\left(D^{2}V\right)^{-1}D^{2}\varphi\right)\,\mathrm{d}\bar{\mu}\ge k_{1}R_{V}(\varphi,\varphi)\ge\frac{r}{n}\left(\int_{A}\mathcal{L}_{\bar{\mu}}(\varphi)\,\mathrm{d}\bar{\mu}\right)^{2}, \end{aligned}

concluding the proof.

AI use disclosure. The authors used AI to streamline the organization of the paper and, in part, to improve the presentation. All the mathematical statements and proofs have been obtained and checked solely by the authors, without the aid of LLMs.

Acknowledgments. The first-named author was supported by the Austrian Science Fund (FWF): 10.55776/P36344N. The second-named author gratefully acknowledges the support of the Basic Research Program of HSE University. The third-named author was supported, in part, by the Austrian Science Fund (FWF): 10.55776/PAT3787224. The authors extend their gratitude to the Erdős Center in Budapest, to its director, Károly Böröczky, and to the administrative staff for hosting and supporting the “Focused Workshop on the log-Brunn-Minkowski conjecture: State of the art and ramifications”, during which the authors started the collaboration leading to this manuscript.

Sotiris Armeniakos

Institut für Stochastik und Wirtschaftsmathematik, TU Wien

e-mail: [email protected]

Alexander Kolesnikov

HSE University, Russian Federation

e-mail: [email protected]

Jacopo Ulivelli

Institut für Diskrete Mathematik und Geometrie, TU Wien

e-mail: [email protected]

References

  1. [1]S. Armeniakos and J. Ulivelli, A new infinitesimal form of the Prékopa–Leindler inequality with multiplicative structure and applications, arXiv:2602.10812, (2026).arxiv.org/abs/2602.10812
  2. [2]C. Borell, Convex set functions in d-space, Period. Math. Hungar., 6 (1975), pp. 111–136.
  3. [3]K. Böröczky, E. Lutwak, D. Yang, and G. Zhang, The logarithmic Minkowski problem, Journal of the American Mathematical Society, 26 (2013), pp. 831–852.DOI
  4. [4]S. Chen, Y. Huang, Q.-R. Li, and J. Liu, The L_p-Brunn-Minkowski inequality for p < 1, Adv. Math., 368 (2020), pp. 107166, 21.
  5. [5]A. Colesanti, From the Brunn–Minkowski inequality to a class of Poincaré-type inequalities, Commun. Contemp. Math., 10 (2008), pp. 765–772.DOI
  6. [6]A. Colesanti, G. V. Livshyts, and A. Marsiglietti, On the stability of Brunn-Minkowski type inequalities, J. Funct. Anal., 273 (2017), pp. 1120–1139.arxiv.org/abs/1606.06586
  7. [7]D. Cordero-Erausquin and A. Eskenazis, Concavity principles for weighted marginals, arXiv:2506.16941, (2025).
  8. [8]D. Cordero-Erausquin, M. Fradelizi, and B. Maurey, The (B) conjecture for the Gaussian measure of dilates of symmetric convex sets and related problems, J. Funct. Anal., 214 (2004), pp. 410–427.
  9. [9]D. Cordero-Erausquin, G. V. Livshyts, and L. Rotem, Remarks on the Brascamp–Lieb inequality, arXiv:2610.34289, (2026).
  10. [10]D. Cordero-Erausquin and L. Rotem, Improved log-concavity for rotationally invariant measures of symmetric convex sets, Ann. Probab., 51 (2023), pp. 987–1003.
  11. [11]A. Eskenazis, A. Giannopoulos, and N. Tziotziou, Functional perimeter and the dimensional Brunn–Minkowski inequality for log-concave measures, arXiv:2605.02747, (2026).arxiv.org/abs/2605.02747
  12. [12]A. Eskenazis and G. Moschidis, The dimensional Brunn-Minkowski inequality in Gauss space, J. Funct. Anal., 280 (2021), pp. Paper No. 108914, 19.arxiv.org/abs/2004.07146
  13. [13]A. Eskenazis, P. Nayar, and T. Tkocz, Gaussian mixtures: entropy and geometric inequalities, Ann. Probab., 46 (2018), pp. 2908–2945.DOI
  14. [14]R. Gardner and A. Zvavitch, Gaussian Brunn–Minkowski inequalities, Trans. Amer. Math. Soc., 362 (2010), pp. 5333–5353.DOI
  15. [15]B.-X. Han and Z.-N. Zhu, Sharp Poincaré Interpolation Along Wasserstein Geodesics, arXiv:2607.10769, (2026).arxiv.org/abs/2607.10769
  16. [16]J. Hosle, A. V. Kolesnikov, and G. V. Livshyts, On the L_p-Brunn-Minkowski and dimensional Brunn–Minkowski conjectures for log-concave measures, The Journal of Geometric Analysis, 31 (2021), pp. 5799–5836.arxiv.org/abs/2003.05282
  17. [17]Y. Hu and M. N. Ivaki, Weighted centro-affine Poincaré inequalities, arXiv:2606.04774, (2026).
  18. [18]M. N. Ivaki and E. Milman, Lᵖ-Minkowski problem under curvature pinching, Int. Math. Res. Not. IMRN, (2024), pp. 8638–8652.
  19. [19]A. V. Kolesnikov, G. Livshyts, and L. Rotem, On p-Brunn-Minkowski and Brascamp-Lieb inequalities, arXiv:2507.12099, (2026).arxiv.org/abs/2507.12099
  20. [20]A. V. Kolesnikov and G. V. Livshyts, On the Gardner-Zyavitch conjecture: symmetry in inequalities of Brunn-Minkowski type, Adv. Math., 384 (2021), pp. Paper No. 107689, 23.arxiv.org/abs/1807.06952
  21. [21]A. V. Kolesnikov and E. Milman, Brascamp-Lieb-type inequalities on weighted Riemannian manifolds with boundary, J. Geom. Anal., 27 (2017), pp. 1680–1702.arxiv.org/abs/1310.2526
  22. [22]———, Poincaré and Brunn–Minkowski inequalities on the boundary of weighted Riemannian manifolds, Amer. J. Math., 140 (2018), pp. 1147–1185.arxiv.org/abs/1711.08825
  23. [23]———, Local Lᵖ-Brunn-Minkowski inequalities for p < 1, Mem. Amer. Math. Soc., 277 (2022), pp. v+78.
  24. [24]R. Latała, On some inequalities for Gaussian measures, in Proceedings of the International Congress of Mathematicians, vol. II, Beijing, 2002, Higher Education Press, pp. 813–822.DOI
  25. [25]G. V. Livshyts, A universal bound in the dimensional Brunn–Minkowski inequality for log-concave measures, Trans. Amer. Math. Soc., 376 (2023), pp. 6663–6680.arxiv.org/abs/2107.00095
  26. [26]E. Milman, Centro-affine differential geometry and the log-Minkowski problem, J. Eur. Math. Soc. (JEMS), 27 (2025), pp. 709–772.
  27. [27]E. Milman, S. Nakamura, and H. Tsuji, Gaussian inequalities with different covariance matrices- the Heterovariant Gaussian Brunn-Minkowski inequality, arXiv:2610.13596, (2026).arxiv.org/abs/2610.13596
  28. [28]P. Nayar and T. Tkocz, A note on a Brunn–Minkowski inequality for the Gaussian measure, Proceedings of the American Mathematical Society, (2013), pp. 4027–4030.arxiv.org/abs/1202.0046
  29. [29]E. Putterman, Equivalence of the local and global versions of the Lᵖ-Brunn-Minkowski inequality, J. Funct. Anal., 280 (2021), pp. Paper No. 108956, 20.
  30. [30]C. Saroglou, More on logarithmic sums of convex bodies, Mathematika, 62 (2016), pp. 818–841.
  31. [31]R. Schneider, Convex bodies: the Brunn-Minkowski theory, vol. 151 of Encyclopedia of Mathematics and its Applications, Cambridge University Press, Cambridge, expanded ed., 2014.
  32. [32]K.-W. Yang, A uniform bound in the dimensional Brunn–Minkowski inequality for even log-concave measures, arXiv:2607.10104, (2026).arxiv.org/abs/2607.10104

Paper details

vlnc:W_ve43n9rz
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 1

downloads · 2

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents