betaFeedback

WorksMathematics

The sharp bound on the distance between the centers of mass of a convex body and its boundary

Abstract

We prove that the distance between the centers of mass c(K)c(K) and c(∂K)c(\partial K) of a convex body KK and its boundary ∂K\partial K, respectively, does not exceed Dn D_n times the width of KK in the direction of the vector joining c(∂K)c(\partial K) and c(K)c(K). The constant DnD_n is optimal and coincides with the sharp constant in the problem of bounding the distance between the orthogonal projection of the centroid of a convex body onto a hyperplane and the centroid of its orthogonal projection onto that hyperplane.

Introduction

Let K⊂RnK \subset\mathbb{R}^n be a convex body, i.e., a convex compact set with non-empty interior. Write

c(K)=1∣K∣∫Kx dx,c(∂K)=1∣∂K∣∫∂Kx dS(x),c(K) = \frac{1}{|K|}\int_K x\,dx,\qquad c(\partial K) = \frac{1}{|\partial K|}\int_{\partial K} x\,dS(x),

for the centroid of KK and its boundary, respectively, where ∣K∣|K| is the volume of KK, ∣∂K∣|\partial K| is its surface area, and dSdS denotes the surface area measure on ∂K\partial K.

Let Sn−1S^{n-1} denote the unit sphere in Rn\mathbb{R}^n. For u∈Sn−1u \in S^{n-1}, let

wK(u)=max⁡x∈K⟨x,u⟩−min⁡x∈K⟨x,u⟩w_K(u) = \max_{x\in K}\langle x,u\rangle-\min_{x\in K}\langle x,u\rangle

be the width of KK in the direction uu.

In their book Unsolved Problems in Geometry, Croft, Falconer, and Guy [1 p. 36, Problem A25] ask how far apart the centroid of a planar convex body and the centroid of its boundary can be, relative to such quantities as the diameter and the width. They also point out the analogous questions in higher dimensions.

The planar problem was solved by Nazarov, Ryabogin, and Yaskin [5]. They proved that for every planar convex body KK and every u∈S1u \in S^1,

∣⟨c(∂K)−c(K),u⟩∣≤16wK(u),\left|\langle c(\partial K)-c(K),u\rangle\right| \le\frac{1}{6}w_K(u),

and therefore

∣c(∂K)−c(K)∣≤16diam⁡K.|c(\partial K)-c(K)| \le\frac{1}{6}\operatorname{diam} K.

The constant 1/61/6 in both inequalities is best possible.

The purpose of the present paper is to solve the corresponding problem in all dimensions. In [5], it was conjectured that the supremum of

∣⟨c(∂K)−c(K),u⟩∣wK(u)\frac{|\langle c(\partial K)-c(K),u\rangle|}{w_K(u)}

2020 Mathematics Subject Classification. Primary 52A20, 52A40.

Key words and phrases. Convex body, center of mass, extremal problem.

The first, third and fourth authors are supported in part by NSERC. The second author is supported in part by U.S. National Science Foundation Grant DMS-2247771. over all convex bodies K⊂RnK \subset\mathbb{R}^n and all directions u∈Sn−1u \in S^{n-1} is the constant DnD_n defined below. For an affine function ℓ(t)=(1−t)α+tβ\ell(t)=(1-t)\alpha+t\beta on [0,1][0,1] such that α,β≥0\alpha,\beta\geq0 are not both zero, set

Ij(ℓ)=∫01tjℓ(t)n−2 dt,j=0,1,2,(1)I_j(\ell)=\int_0^1 t^j\ell(t)^{n-2}\,\mathrm{d}t,\qquad j=0,1,2, \tag*{(1)}

and let

Dn=max⁡ℓ≥0 affine on [0,1]ℓ≢0{I2(ℓ)I1(ℓ)−I1(ℓ)I0(ℓ)}.(2)D_n=\max_{\substack{\ell\geq0\ \text{affine on }[0,1]\\ \ell\not\equiv0}}\left\{\frac{I_2(\ell)}{I_1(\ell)}-\frac{I_1(\ell)}{I_0(\ell)}\right\}. \tag*{(2)}

Multiplying ℓ\ell by a positive constant does not change the expression in (2). Hence we may normalize ℓ\ell so that ℓ(0)+ℓ(1)=1\ell(0)+\ell(1)=1. The resulting parameter set is compact, so the maximum is attained. The constants DnD_n form an increasing sequence such that

D2=16,D3=1−23,lim⁡n→∞Dn=0.2016…,D_2=\frac{1}{6},\qquad D_3=1-\sqrt{\frac{2}{3}},\qquad\lim_{n\to\infty}D_n=0.2016\ldots,

see [4].

The main result of this paper is the following.

Theorem 1.1. For every convex body K⊂RnK\subset\mathbb{R}^n, n≥2n\geq2, and every u∈Sn−1u\in S^{n-1},

∣⟨c(∂K)−c(K),u⟩∣≤DnwK(u).(3)\left|\langle c(\partial K)-c(K),u\rangle\right|\leq D_nw_K(u). \tag*{(3)}

Consequently,

∣c(∂K)−c(K)∣≤Dndiam⁡K.(4)|c(\partial K)-c(K)|\leq D_n\operatorname{diam} K. \tag*{(4)}

The constant DnD_n is optimal in both inequalities.

The constants DnD_n first appeared in the following problem studied by Myroshnychenko, Tatarko, and Yaskin [4]. For a convex body K⊂RnK\subset\mathbb{R}^n and a hyperplane HH in Rn\mathbb{R}^n, consider PHc(K)P_Hc(K), the orthogonal projection of c(K)c(K) onto HH, and c(PHK)c(P_HK), the centroid of the orthogonal projection of KK onto HH. If these two points are different, let

u=PHc(K)−c(PHK)∣PHc(K)−c(PHK)∣.u=\frac{P_Hc(K)-c(P_HK)}{|P_Hc(K)-c(P_HK)|}.

It was shown in [4] that

∣PHc(K)−c(PHK)∣≤DnwK(u).|P_Hc(K)-c(P_HK)|\leq D_nw_K(u).

Here the constant DnD_n is best possible.

There is a geometric reason to expect the constants DnD_n from [4] to appear in Theorem 1.1. Fix a hyperplane HH and compress a convex body KK in the direction orthogonal to HH, preserving its projection onto HH. The centroid of the compressed body projects to PHc(K)P_Hc(K), while the projection of its boundary centroid approaches c(PHK)c(P_HK). Indeed, as the body becomes thinner, its surface area becomes concentrated on the two boundary graphs over PHKP_HK. Thus the inequality of [4] is recovered as a limiting case of the inequality in Theorem 1.1. In particular, the optimal constant in (3) cannot be smaller than DnD_n.

The paper is organized as follows. Section 2 outlines the proof and states its main ingredient, Proposition 2.1. In Section 3, we derive bounds (3) and (4) from this proposition, and in Section 4 we establish the sharpness of the constant DnD_n. Section 5 contains one-dimensional auxiliary inequalities. Finally, Section 6 uses localization to prove Proposition 2.1.

Outline of the proof

The proof of Theorem 1.1 is based on decomposing KK into parallel chords. The coordinate of c(K)c(K) in a fixed direction is expressed as an average of midpoint coordinates of chords weighted by chord length; see (6). To obtain the corresponding coordinate of c(∂K)c(\partial K), we first take the unweighted average of these midpoint coordinates for each chord direction and then average the resulting values over all directions, with weights given by the volumes of the corresponding projections; see (10). The following proposition allows us to compare these averages.

Proposition 2.1. Let Ω⊂Rn−1\Omega\subset\mathbb{R}^{n-1} be a bounded, open, and convex set. Let a,ba,b be bounded functions on Ω\Omega, with aa convex, bb concave, and 0≤a≤b≤10 \le a \le b \le1. Put

h=b−a,m=a+b2.h=b-a,\qquad m=\frac{a+b}{2}.

If h≢0h \not\equiv0, then

∣1∣Ω∣∫Ωm(y) dy−∫Ωh(y)m(y) dy∫Ωh(y) dy∣≤Dn.(5)\left|\frac{1}{|\Omega|}\int_{\Omega}m(y)\,\mathrm{d}y-\frac{\int_{\Omega}h(y)m(y)\,\mathrm{d}y}{\int_{\Omega}h(y)\,\mathrm{d}y}\right|\le D_n. \tag*{(5)}

The appearance of DnD_n in Proposition 2.1 can be understood through localization. Applied in Rn−1\mathbb{R}^{n-1}, the localization lemma of Kannan, Lovász, and Simonovits reduces the proof of (5) to one-dimensional integrals with weights proportional to ℓn−2\ell^{n-2}, where ℓ\ell is nonnegative and affine. These are precisely the weights occurring in the definition of DnD_n. For affine functions, the required one-dimensional bound follows from this definition. The one-dimensional auxiliary inequalities extend this bound to general pairs of convex and concave functions.

Proof of the bounds in Theorem 1.1

Proof. Assuming Proposition 2.1, we now prove inequalities (3) and (4). Fix u∈Sn−1u \in S^{n-1}. Since the inequalities are invariant under translations and dilations, we may assume that

min⁡x∈K⟨x,u⟩=0,max⁡x∈K⟨x,u⟩=1.\min_{x\in K}\langle x,u\rangle=0,\qquad\max_{x\in K}\langle x,u\rangle=1.

Thus it suffices to prove

∣⟨c(∂K)−c(K),u⟩∣≤Dn.\left|\langle c(\partial K)-c(K),u\rangle\right|\le D_n.

Fix v∈Sn−1v \in S^{n-1} with ⟨u,v⟩≠0\langle u,v\rangle\ne0. Replacing vv by −v-v if necessary, we may assume ⟨u,v⟩>0\langle u,v\rangle>0. Put

Ωv=int⁡(Pv⊥K),\Omega_v=\operatorname{int}(P_{v^\perp}K),

where Pv⊥P_{v^\perp} denotes the orthogonal projection onto the hyperplane v⊥={x∈Rn:⟨x,v⟩=0}v^\perp=\{x\in\mathbb{R}^n:\langle x,v\rangle=0\}, and the interior is taken relative to v⊥v^\perp.

For y∈Ωvy\in\Omega_v, define

av(y)=min⁡x∈K∩(y+Rv)⟨x,u⟩,bv(y)=max⁡x∈K∩(y+Rv)⟨x,u⟩.a_v(y)=\min_{x\in K\cap(y+\mathbb{R}v)}\langle x,u\rangle,\qquad b_v(y)=\max_{x\in K\cap(y+\mathbb{R}v)}\langle x,u\rangle.

By convexity of KK, the function ava_v is convex and bvb_v is concave on Ωv\Omega_v. Moreover, 0≤av<bv≤10\le a_v<b_v\le1. Put

hv=bv−av,mv=av+bv2.h_v=b_v-a_v,\qquad m_v=\frac{a_v+b_v}{2}.

Thus mv(y)m_v(y) is the uu-coordinate of the midpoint of the chord K∩(y+Rv)K\cap(y+\mathbb{R}v), while the length of this chord is hv(y)/⟨u,v⟩h_v(y)/\langle u,v\rangle.

We first express ⟨c(K),u⟩\langle c(K),u\rangle in terms of these functions. Writing x=y+tvx=y+tv, where y∈v⊥y\in v^\perp, and using Fubini’s theorem gives

∫K⟨x,u⟩ dx=∫Ωv(∫{t:y+tv∈K}⟨y+tv,u⟩ dt) dy=1⟨u,v⟩∫Ωv(∫av(y)bv(y)s ds) dy=1⟨u,v⟩∫Ωvbv(y)2−av(y)22 dy=1⟨u,v⟩∫Ωvhv(y)mv(y) dy.\begin{aligned} \int_K \langle x,u\rangle\,dx &=\int_{\Omega_v}\left(\int_{\{t:y+tv\in K\}}\langle y+tv,u\rangle\,dt\right)\,dy\\ &=\frac{1}{\langle u,v\rangle}\int_{\Omega_v}\left(\int_{a_v(y)}^{b_v(y)}s\,ds\right)\,dy\\ &=\frac{1}{\langle u,v\rangle}\int_{\Omega_v}\frac{b_v(y)^2-a_v(y)^2}{2}\,dy\\ &=\frac{1}{\langle u,v\rangle}\int_{\Omega_v}h_v(y)m_v(y)\,dy. \end{aligned}

Here, in the second equality, we used the change of variables s=⟨x,u⟩s=\langle x,u\rangle so that ds=⟨u,v⟩dtds=\langle u,v\rangle dt. Similarly,

∣K∣=1⟨u,v⟩∫Ωvhv(y) dy.|K|=\frac{1}{\langle u,v\rangle}\int_{\Omega_v}h_v(y)\,dy.

Consequently,

⟨c(K),u⟩=1∣K∣∫K⟨x,u⟩ dx=∫Ωvhv(y)mv(y) dy∫Ωvhv(y) dy.(6)\langle c(K),u\rangle=\frac{1}{|K|}\int_K\langle x,u\rangle\,dx =\frac{\int_{\Omega_v}h_v(y)m_v(y)\,dy}{\int_{\Omega_v}h_v(y)\,dy}. \tag*{(6)}

We next express ⟨c(∂K),u⟩\langle c(\partial K),u\rangle using the same chord midpoint coordinates mv(y)m_v(y). Let νK(x)\nu_K(x) denote the outer unit normal to KK, defined dSdS-almost everywhere on ∂K\partial K. For every nonnegative measurable function ff, the area formula (see [2 Theorem 3.2.3]) gives

∫∂Kf(x)∣⟨νK(x),v⟩∣ dS(x)=∫v⊥∑x∈∂K∩(y+Rv)f(x) dy.(7)\int_{\partial K}f(x)|\langle\nu_K(x),v\rangle|\,dS(x) =\int_{v^\perp}\sum_{x\in\partial K\cap(y+\mathbb{R}v)}f(x)\,dy. \tag*{(7)}

For y∈Ωvy\in\Omega_v, the fiber on the right consists of the two endpoints of the corresponding chord. The boundary of Pv⊥KP_{v^\perp}K has (n−1)(n-1)-dimensional Lebesgue measure zero and therefore does not contribute to the integral. We may apply (7) with f=1f=1 and f(x)=⟨x,u⟩f(x)=\langle x,u\rangle, obtaining

∫∂K∣⟨νK(x),v⟩∣ dS(x)=2∣Ωv∣(8)\int_{\partial K}|\langle\nu_K(x),v\rangle|\,dS(x)=2|\Omega_v| \tag*{(8)}

and

∫∂K⟨x,u⟩∣⟨νK(x),v⟩∣ dS(x)=2∫Ωvmv(y) dy.(9)\int_{\partial K}\langle x,u\rangle|\langle\nu_K(x),v\rangle|\,dS(x) =2\int_{\Omega_v}m_v(y)\,dy. \tag*{(9)}

Since vv and −v-v determine the same chords, the functions ava_v, bvb_v, hvh_v, mvm_v, as well as Ωv\Omega_v, are unchanged under v↦−vv\mapsto-v. Consequently, (6), (8), and (9) hold for every vv with ⟨u,v⟩≠0\langle u,v\rangle\ne0.

To apply Proposition 2.1, we now average (8) and (9) over all chord directions vv to obtain a suitable representation of the boundary centroid. Let dσd\sigma be the normalized rotation-invariant measure on Sn−1S^{n-1}. Since

∫Sn−1∣⟨u,v⟩∣ dσ(v)=γn>0\int_{S^{n-1}}|\langle u,v\rangle|\,d\sigma(v)=\gamma_n>0

is independent of the unit vector uu, Fubini’s theorem together with (9) gives

∫∂K⟨x,u⟩ dS(x)=1γn∫∂K⟨x,u⟩(∫Sn−1∣⟨vK(x),v⟩∣ dσ(v))dS(x)=1γn∫Sn−1(∫∂K⟨x,u⟩∣⟨vK(x),v⟩∣ dS(x))dσ(v)=2γn∫Sn−1(∫Ωvmv(y) dy)dσ(v).\begin{aligned} \int_{\partial K}\langle x,u\rangle\,dS(x) &=\frac{1}{\gamma_n}\int_{\partial K}\langle x,u\rangle\left(\int_{S^{n-1}}|\langle v_K(x),v\rangle|\,d\sigma(v)\right)dS(x)\\ &=\frac{1}{\gamma_n}\int_{S^{n-1}}\left(\int_{\partial K}\langle x,u\rangle|\langle v_K(x),v\rangle|\,dS(x)\right)d\sigma(v)\\ &=\frac{2}{\gamma_n}\int_{S^{n-1}}\left(\int_{\Omega_v}m_v(y)\,dy\right)d\sigma(v). \end{aligned}

By (8), we have

∫Sn−1∣Ωv∣ dσ(v)=12∫Sn−1(∫∂K∣⟨vK(x),v⟩∣ dS(x))dσ(v)=12∫∂K(∫Sn−1∣⟨vK(x),v⟩∣ dσ(v))dS(x)=γn2∣∂K∣.\begin{aligned} \int_{S^{n-1}}|\Omega_v|\,d\sigma(v) =\frac{1}{2}\int_{S^{n-1}}\left(\int_{\partial K}|\langle v_K(x),v\rangle|\,dS(x)\right)d\sigma(v) \\ &=\frac{1}{2}\int_{\partial K}\left(\int_{S^{n-1}}|\langle v_K(x),v\rangle|\,d\sigma(v)\right)dS(x) =\frac{\gamma_n}{2}|\partial K|. \end{aligned}

Therefore,

⟨c(∂K),u⟩=1∣∂K∣∫∂K⟨x,u⟩ dS(x)=∫Sn−1∫Ωvmv(y) dy dσ(v)∫Sn−1∣Ωv∣ dσ(v).(10)\langle c(\partial K),u\rangle =\frac{1}{|\partial K|}\int_{\partial K}\langle x,u\rangle\,dS(x) =\frac{\int_{S^{n-1}}\int_{\Omega_v}m_v(y)\,dy\,d\sigma(v)} {\int_{S^{n-1}}|\Omega_v|\,d\sigma(v)}. \tag*{(10)}

We now apply Proposition 2.1. Since 0≤av<bv≤10\le a_v<b_v\le1, it follows from (6) that

∣1∣Ωv∣∫Ωvmv(y) dy−⟨c(K),u⟩∣=∣1∣Ωv∣∫Ωvmv(y) dy−∫Ωvhv(y)mv(y) dy∫Ωvhv(y) dy∣≤Dn\left|\frac{1}{|\Omega_v|}\int_{\Omega_v}m_v(y)\,dy-\langle c(K),u\rangle\right| =\left|\frac{1}{|\Omega_v|}\int_{\Omega_v}m_v(y)\,dy- \frac{\int_{\Omega_v}h_v(y)m_v(y)\,dy} {\int_{\Omega_v}h_v(y)\,dy}\right| \le D_n

whenever ⟨u,v⟩≠0\langle u,v\rangle\ne0.

Using (10), we therefore obtain

∣⟨c(∂K)−c(K),u⟩∣=∣∫Sn−1∣Ωv∣(1∣Ωv∣∫Ωvmv(y) dy−⟨c(K),u⟩)dσ(v)∣∫Sn−1∣Ωv∣ dσ(v)≤Dn.|\langle c(\partial K)-c(K),u\rangle| =\frac{\left|\int_{S^{n-1}}|\Omega_v|\left(\frac{1}{|\Omega_v|}\int_{\Omega_v}m_v(y)\,dy-\langle c(K),u\rangle\right)d\sigma(v)\right|} {\int_{S^{n-1}}|\Omega_v|\,d\sigma(v)} \le D_n.

Here the directions satisfying ⟨u,v⟩=0\langle u,v\rangle=0 may be disregarded, since they form a set of spherical measure zero.

The last inequality is precisely (3) since wK(u)=1w_K(u)=1. If c(∂K)≠c(K)c(\partial K)\ne c(K), choosing uu in the direction of their difference and using wK(u)≤diam⁡Kw_K(u)\le\operatorname{diam}K gives (4).

Sharpness

To prove the sharpness of the constant DnD_n in (3), it is enough to exhibit a sequence of bodies for which the quotient in (3) tends to DnD_n in the direction e1e_1.

Assume first n≥3n\ge3. Choose an affine function ℓ≥0\ell\ge0 attaining the maximum in (2), and for ε>0\varepsilon>0 define

Kε={(t,y,z)∈R×Rn−2×R:0≤t≤1, ∥y∥∞≤εℓ(t), 0≤z≤ε2t}.(11)K_\varepsilon=\left\{(t,y,z)\in\mathbb{R}\times\mathbb{R}^{n-2}\times\mathbb{R}:0\le t\le1,\ \lVert y\rVert_\infty\le\varepsilon\ell(t),\ 0\le z\le\varepsilon^2t\right\}. \tag*{(11)}

The body KεK_\varepsilon is convex and wKε(e1)=1w_{K_\varepsilon}(e_1)=1. Fubini’s theorem gives

⟨c(Kε),e1⟩=I2(ℓ)I1(ℓ).(12)\langle c(K_\varepsilon),e_1\rangle=\frac{I_2(\ell)}{I_1(\ell)}. \tag*{(12)}

The bottom and top faces, z=0z=0 and z=ε2tz=\varepsilon^2t, have total area

(1+1+ε4)(2ε)n−2I0(ℓ),(1+\sqrt{1+\varepsilon^4})(2\varepsilon)^{n-2}I_0(\ell),

and the first coordinates of their centroids are both I1(ℓ)/I0(ℓ)I_1(\ell)/I_0(\ell). The 2(n−2)2(n-2) side faces have total area O(εn−1)O(\varepsilon^{n-1}). The part of the boundary at t=1t=1 has area O(εn)O(\varepsilon^n), and the part at t=0t=0 has zero (n−1)(n-1)-dimensional area. Hence

⟨c(∂Kε),e1⟩⟶I1(ℓ)I0(ℓ).(13)\langle c(\partial K_\varepsilon),e_1\rangle\longrightarrow\frac{I_1(\ell)}{I_0(\ell)}. \tag*{(13)}

Combining (12) and (13) gives

∣⟨c(∂Kε)−c(Kε),e1⟩∣wKε(e1)⟶I2(ℓ)I1(ℓ)−I1(ℓ)I0(ℓ)=Dn.\frac{\left|\langle c(\partial K_\varepsilon)-c(K_\varepsilon),e_1\rangle\right|}{w_{K_\varepsilon}(e_1)} \longrightarrow\frac{I_2(\ell)}{I_1(\ell)}-\frac{I_1(\ell)}{I_0(\ell)}=D_n.

Thus the constant in (3) is sharp for n≥3n\geq3.

Since KεK_\varepsilon contains (0,0,…,0)(0,0,\ldots,0) and (1,0,…,0)(1,0,\ldots,0) and lies within distance O(ε)O(\varepsilon) of the segment joining these points, we have

diam⁡Kε⟶1.\operatorname{diam} K_\varepsilon\longrightarrow1.

The yy-coordinates of both centroids vanish by symmetry and their zz-coordinates are O(ε2)O(\varepsilon^2). Therefore

∣c(∂Kε)−c(Kε)∣diam⁡Kε⟶Dn,\frac{\left|c(\partial K_\varepsilon)-c(K_\varepsilon)\right|}{\operatorname{diam} K_\varepsilon}\longrightarrow D_n,

which proves sharpness in (4) for n≥3n\geq3.

When n=2n=2, let KεK_\varepsilon be the triangle with vertices (0,0),(1,0),(1,ε2)(0,0),(1,0),(1,\varepsilon^2). Then ⟨c(Kε),e1⟩=2/3\langle c(K_\varepsilon),e_1\rangle=2/3, ⟨c(∂Kε),e1⟩→1/2\langle c(\partial K_\varepsilon),e_1\rangle\to1/2, and diam⁡Kε→1\operatorname{diam} K_\varepsilon\to1 as ε→0+\varepsilon\to0^+. Thus we obtain the same conclusions as above.

One-dimensional comparison inequalities

The remainder of the paper is devoted to proving Proposition 2.1. In this section we obtain a series of auxiliary one-dimensional integral inequalities that will be used in the proof of Proposition 5.5. The latter will be the main ingredient in the localization argument of Proposition 2.1.

Preliminary estimates

For any nonzero finite positive measure ν\nu and bounded measurable real-valued function ff, define

Aν(f)=∫f dν∫dνA_\nu(f)=\frac{\int f\,d\nu}{\int d\nu}

and

Vν(f)=Aν(f2)−Aν(f)2.V_\nu(f)=A_\nu(f^2)-A_\nu(f)^2.

Thus AνA_\nu denotes averaging, and VνV_\nu is the corresponding variance. Notice that

Vν(f)=Aν((f−Aν(f))2)≥0.V_\nu(f)=A_\nu\left((f-A_\nu(f))^2\right)\geq0.

For the remainder of the paper, put C=DnC = D_n. Fix a nonzero nonnegative affine function ℓ(t)\ell(t) on [0,1][0,1] and define the normalized weight

ω(t)=ℓ(t)n−2I0(ℓ),0≤t≤1.(14)\omega(t) = \frac{\ell(t)^{n-2}}{I_0(\ell)}, \qquad0 \le t \le1. \tag*{(14)}

Let dμ(t)=ω(t) dtd\mu(t) = \omega(t)\,dt. Throughout this section, we omit the subscript μ\mu and write AA and VV for AμA_\mu and VμV_\mu, respectively. Since μ([0,1])=1\mu([0,1]) = 1, we have

A(f)=∫01f(t)ω(t) dt.A(f) = \int_0^1 f(t)\omega(t)\,dt.

Let T(t)=tT(t) = t. Then

A(T)=∫01tω(t) dt=I1(ℓ)I0(ℓ),A(T) = \int_0^1 t\omega(t)\,dt = \frac{I_1(\ell)}{I_0(\ell)},

and

A(T2)=∫01t2ω(t) dt=I2(ℓ)I0(ℓ).A(T^2) = \int_0^1 t^2\omega(t)\,dt = \frac{I_2(\ell)}{I_0(\ell)}.

Therefore

V(T)=A(T2)−A(T)2=I2I0−I12I02.V(T) = A(T^2) - A(T)^2 = \frac{I_2}{I_0} - \frac{I_1^2}{I_0^2}.

From the definition of DnD_n,

V(T)A(T)=I2(ℓ)I1(ℓ)−I1(ℓ)I0(ℓ)≤C,\frac{V(T)}{A(T)} = \frac{I_2(\ell)}{I_1(\ell)} - \frac{I_1(\ell)}{I_0(\ell)} \le C,

and hence

V(T)≤CA(T).V(T) \le CA(T).

Since AA is linear and V(1−T)=V(T)V(1-T) = V(T), applying the same inequality after the reflection t↦1−tt \mapsto1-t gives

V(T)≤C(1−A(T)).V(T) \le C(1-A(T)).

Thus

V(T)≤CA(T),andV(T)≤C(1−A(T)).(15)V(T) \le CA(T), \qquad\text{and} \qquad V(T) \le C(1-A(T)). \tag*{(15)}

We shall also use

C≥16,C \ge\frac{1}{6},

which follows by taking ℓ\ell constant in (2).

In the lemmas below, we establish integral inequalities for bounded convex or concave functions on [0,1][0,1]. Since changing their endpoint values does not affect these integrals, we may assume that such functions are continuous on [0,1][0,1], replacing their endpoint values by their one-sided limits if necessary.

The following estimate will be used in Lemma 5.3.

Lemma 5.1. If 0≤h≤10 \le h \le1 is a concave function on [0,1][0,1], then

V(h)≤13A(h).(16)V(h) \le\frac{1}{3}A(h). \tag*{(16)}

Proof. Suppose first that n≥3n \ge3, and consider the convex body

Pℓ={(t,y)∈[0,1]×Rn−2:∥y∥∞≤ℓ(t)}.P_\ell= \{(t,y) \in[0,1] \times\mathbb{R}^{n-2} : \lVert y\rVert_\infty\le\ell(t)\}.

For fixed tt, the yy-section of PℓP_\ell has (n−2)(n-2)-dimensional volume (2ℓ(t))n−2(2\ell(t))^{n-2}. Note that the explicit construction of PℓP_\ell is not important. All we need is that the (n−2)(n-2)-dimensional volume of its cross-sections is proportional to ℓn−2(t)\ell^{n-2}(t). Hence, for every integrable function ff on [0,1][0,1],

1∣Pℓ∣∫Pℓf(t) dt dy=∫01f(t)ℓ(t)n−2 dt∫01ℓ(t)n−2 dt=A(f).\frac{1}{|P_\ell|}\int_{P_\ell} f(t)\,\mathrm{d}t\,\mathrm{d}y = \frac{\int_0^1 f(t)\ell(t)^{n-2}\,\mathrm{d}t}{\int_0^1 \ell(t)^{n-2}\,\mathrm{d}t} =A(f).

When n=2n=2, we take Pℓ=[0,1]P_\ell=[0,1], and the same identity holds.

If A(h)=0A(h)=0 or A(h)=1A(h)=1, then V(h)=0V(h)=0 and there is nothing to prove. Hence for the remainder of the proof we assume 0<A(h)<10<A(h)<1.

Let

F(s)=A(1{h>s})=∣{(t,y)∈Pℓ:h(t)>s}∣∣Pℓ∣,0≤s≤1.F(s)=A(\mathbf{1}_{\{h>s\}})=\frac{|\{(t,y)\in P_\ell:h(t)>s\}|}{|P_\ell|},\qquad0\leq s\leq1.

Since hh is nonnegative, concave, and not identically zero, it is positive on (0,1)(0,1), so F(0)=1F(0)=1.

Since (t,y)↦h(t)(t,y)\mapsto h(t) is concave on PℓP_\ell, the Brunn–Minkowski inequality implies that FF is log-concave. By the layer cake decomposition,

A(h)=∫01F(s) ds,andA(h2)=2∫01sF(s) ds.(17)A(h)=\int_0^1 F(s)\,\mathrm{d}s,\qquad\text{and}\qquad A(h^2)=2\int_0^1 sF(s)\,\mathrm{d}s. \tag*{(17)}

To bound V(h)V(h), we compare the function FF with an exponential function that has the same integral. In particular, this will allow us to bound A(h2)A(h^2) in terms of A(h)A(h). Choose λ>0\lambda>0 so that

A(h)=∫01e−λs ds=1−e−λλ,A(h)=\int_0^1 e^{-\lambda s}\,\mathrm{d}s=\frac{1-e^{-\lambda}}{\lambda},

which is always possible since 1−e−λλ\frac{1-e^{-\lambda}}{\lambda} decreases continuously from 11 to 00 as λ\lambda increases from 00 to ∞\infty.

Since log⁡F(s)+λs\log F(s)+\lambda s is concave and vanishes at s=0s=0, the difference F(s)−e−λsF(s)-e^{-\lambda s} changes sign at most once, from nonnegative to nonpositive. Thus there exists s0∈[0,1]s_0\in[0,1] such that

(s−s0)(F(s)−e−λs)≤0(s-s_0)\left(F(s)-e^{-\lambda s}\right)\leq0

for all s∈[0,1]s\in[0,1]. Since

∫01F(s) ds=∫01e−λs ds,\int_0^1 F(s)\,\mathrm{d}s=\int_0^1 e^{-\lambda s}\,\mathrm{d}s,

we obtain

∫01sF(s) ds−∫01se−λs ds=∫01(s−s0)(F(s)−e−λs) ds≤0.\int_0^1 sF(s)\,\mathrm{d}s-\int_0^1 se^{-\lambda s}\,\mathrm{d}s =\int_0^1(s-s_0)\left(F(s)-e^{-\lambda s}\right)\,\mathrm{d}s\leq0.

Together with (17), this implies

A(h2)=2∫01sF(s) ds≤2∫01se−λs ds=2λ2(1−(1+λ)e−λ).A(h^2)=2\int_0^1sF(s)\,\mathrm{d}s\leq2\int_0^1se^{-\lambda s}\,\mathrm{d}s=\frac{2}{\lambda^2}\left(1-(1+\lambda)e^{-\lambda}\right).

Therefore,

V(h)A(h)=A(h2)A(h)−A(h)≤2(1−(1+λ)e−λ)λ(1−e−λ)−1−e−λλ≤13.(18)\frac{V(h)}{A(h)}=\frac{A(h^2)}{A(h)}-A(h) \leq\frac{2(1-(1+\lambda)e^{-\lambda})}{\lambda(1-e^{-\lambda})}-\frac{1-e^{-\lambda}}{\lambda}\leq\frac{1}{3}. \tag*{(18)}

For the last inequality, multiplication by the positive denominator reduces it to

Φ(λ)=(λ−3)eλ+3e−λ+5λ≥0.\Phi(\lambda)=(\lambda-3)e^\lambda+3e^{-\lambda}+5\lambda\geq0.

Now Φ(0)=Φ′(0)=0\Phi(0)=\Phi'(0)=0 and

Φ′′(λ)=(λ−1)eλ+3e−λ>0.\Phi''(\lambda)=(\lambda-1)e^\lambda+3e^{-\lambda}>0.

For λ≥1\lambda\ge1 this is immediate. For 0≤λ≤10 \le\lambda\le1, it follows from (1−λ)e2λ≤e/2<3(1-\lambda)e^{2\lambda} \le e/2 < 3. Thus Φ\Phi is increasing and so (18) holds.

□\square

The goal of this section is to establish Proposition 5.5, i.e., we need to prove the inequality

−2CA(b−a)≤V(a)−V(b)≤2CA(b−a)-2CA(b-a) \le V(a)-V(b) \le2CA(b-a)

for continuous functions 0≤a≤b≤10 \le a \le b \le1 on [0,1][0,1], with aa convex and bb concave. Consider the functional

J(f)=V(f)+2CA(f).J(f)=V(f)+2CA(f).

Then the right-hand inequality above is equivalent to

J(a)≤J(b).J(a) \le J(b).

The advantage of this formulation is that JJ satisfies certain comparison inequalities for elementary piecewise affine functions, which can then be extended to general convex functions using the convexity of JJ.

To see that the functional JJ is convex, take 0≤λ≤10 \le\lambda\le1. Then

V(λf+(1−λ)g)=λV(f)+(1−λ)V(g)−λ(1−λ)V(f−g)≤λV(f)+(1−λ)V(g),V(\lambda f+(1-\lambda)g)=\lambda V(f)+(1-\lambda)V(g)-\lambda(1-\lambda)V(f-g)\le\lambda V(f)+(1-\lambda)V(g),

while AA is linear. Hence

J(λf+(1−λ)g)≤λJ(f)+(1−λ)J(g).J(\lambda f+(1-\lambda)g)\le\lambda J(f)+(1-\lambda)J(g).

One-sided elementary piecewise affine functions

For 0≤s<10 \le s<1, set

Rs(t)=(t−s)+1−s,R_s(t)=\frac{(t-s)_+}{1-s},

and let R1=0R_1=0.

Lemma 5.2. The function s↦J(Rs)s\mapsto J(R_s) is nonincreasing on [0,1][0,1]. In particular,

J(Rs)≤J(R0)=J(T).(19)J(R_s)\le J(R_0)=J(T). \tag*{(19)}

Proof. We first show that J(Rs)J(R_s) is nonincreasing on s∈[0,1)s\in[0,1) by proving that ddsJ(Rs)≤0\frac{d}{ds}J(R_s)\le0. For 0≤s<10\le s<1, let μs\mu_s be the probability measure on [s,1][s,1] defined by

dμs(t)=ω(t) dt∫s1ω(r) dr.d\mu_s(t)=\frac{\omega(t)\,dt}{\int_s^1\omega(r)\,dr}.

Also, for s≤t≤1s\le t\le1,

Rs(t)=t−s1−s.R_s(t)=\frac{t-s}{1-s}.

Under the change of variables

r=t−s1−s,r=\frac{t-s}{1-s},

Rs(t)R_s(t) becomes the coordinate function T(r)=rT(r)=r on [0,1][0,1], and the measure μs\mu_s is now defined on [0,1][0,1] and is given by

ℓ(s+(1−s)r)n−2∫01ℓ(s+(1−s)ρ)n−2 dρ dr.\frac{\ell(s+(1-s)r)^{n-2}}{\int_0^1\ell(s+(1-s)\rho)^{n-2}\,d\rho}\,dr.

Since r↦ℓ(s+(1−s)r)r \mapsto\ell(s+(1-s)r) is again a nonnegative affine function on [0,1][0,1], this is a weight of the form (14). Hence the second inequality in (15), applied to μs\mu_s, gives

Vμs(Rs)≤C(1−Aμs(Rs)).(20)V_{\mu_s}(R_s) \le C\left(1-A_{\mu_s}(R_s)\right). \tag*{(20)}

We now differentiate J(Rs)J(R_s). Recall that

A(Rs)=∫s1Rs(t)ω(t) dt,A(Rs2)=∫s1Rs(t)2ω(t) dt.A(R_s)=\int_s^1 R_s(t)\omega(t)\,dt,\qquad A(R_s^2)=\int_s^1 R_s(t)^2\omega(t)\,dt.

Observe that for s<t≤1s<t\le1,

∂Rs(t)∂s=−1−Rs(t)1−s.\frac{\partial R_s(t)}{\partial s}=-\frac{1-R_s(t)}{1-s}.

By Leibniz’s rule,

ddsA(Rs)=−Rs(s)ω(s)+∫s1∂Rs(t)∂sω(t) dt=−11−s∫s1(1−Rs(t))ω(t) dt=−μ([s,1])1−s(1−Aμs(Rs)).\begin{aligned} \frac{d}{ds}A(R_s)&=-R_s(s)\omega(s)+\int_s^1\frac{\partial R_s(t)}{\partial s}\omega(t)\,dt\\ &=-\frac{1}{1-s}\int_s^1(1-R_s(t))\omega(t)\,dt\\ &=-\frac{\mu([s,1])}{1-s}\left(1-A_{\mu_s}(R_s)\right). \end{aligned}

where we used Rs(s)=0R_s(s)=0 and dμ(t)=ω(t) dtd\mu(t)=\omega(t)\,dt. Similarly,

ddsA(Rs2)=∫s12Rs(t)∂Rs(t)∂sω(t) dt=−2μ([s,1])1−s(Aμs(Rs)−Aμs(Rs2)).\frac{d}{ds}A(R_s^2)=\int_s^1 2R_s(t)\frac{\partial R_s(t)}{\partial s}\omega(t)\,dt=-\frac{2\mu([s,1])}{1-s}\left(A_{\mu_s}(R_s)-A_{\mu_s}(R_s^2)\right).

Since

V(Rs)=A(Rs2)−A(Rs)2V(R_s)=A(R_s^2)-A(R_s)^2

and

A(Rs)=μ([s,1])Aμs(Rs),A(R_s)=\mu([s,1])A_{\mu_s}(R_s),

substitution of the preceding two derivatives and basic rearrangements give

ddsJ(Rs)=−2μ([s,1])1−s[C+(1−μ([s,1]))Aμs(Rs)](1−Aμs(Rs))−Vμs(Rs).\frac{d}{ds}J(R_s)=-\frac{2\mu([s,1])}{1-s}\left[C+(1-\mu([s,1]))A_{\mu_s}(R_s)\right]\left(1-A_{\mu_s}(R_s)\right)-V_{\mu_s}(R_s).

The expression in brackets can be written as

[C(1−Aμs(Rs))−Vμs(Rs)]+(1−μ([s,1]))Aμs(Rs)(1−Aμs(Rs)).\left[C(1-A_{\mu_s}(R_s))-V_{\mu_s}(R_s)\right]+(1-\mu([s,1]))A_{\mu_s}(R_s)(1-A_{\mu_s}(R_s)).

Both terms are nonnegative: the first by (20), and the second because

0≤μ([s,1])≤1,0≤Aμs(Rs)≤1.0\le\mu([s,1])\le1,\qquad0\le A_{\mu_s}(R_s)\le1.

Therefore

ddsJ(Rs)≤0.\frac{d}{ds}J(R_s)\le0.

Thus s↦J(Rs)s\mapsto J(R_s) is nonincreasing on [0,1)[0,1). Since J(Rs)≥0=J(R1)J(R_s)\ge0=J(R_1), (19) holds on [0,1][0,1]. This completes the proof.

Two-sided elementary piecewise affine functions

For 0<u≤10<u\leq1, put

Lu(t)=(u−t)+u,L_u(t)=\frac{(u-t)_+}{u},

and set L0=0L_0=0.

Lemma 5.3. If 0≤u≤s≤10\leq u\leq s\leq1 and 0≤z≤10\leq z\leq1, then

J(zLu+Rs)≤J(z(1−T)+T).(21)J(zL_u+R_s)\leq J(z(1-T)+T). \tag*{(21)}

Proof. Since u≤su\leq s, we have LuRs=0L_uR_s=0. Expanding the difference of the two sides of (21) as a quadratic polynomial in zz and regrouping gives

J(z(1−T)+T)−J(zLu+Rs)=(1−z)2[J(T)−J(Rs)]+2z(1−z)[C(1+A(T)−A(Lu)−2A(Rs))−V(Rs)+A(Lu)A(Rs)]+z2[J(1)−J(Lu+Rs)].\begin{aligned} J(z(1-T)+T)-J(zL_u+R_s) ={}&(1-z)^2[J(T)-J(R_s)]\\ &+2z(1-z)\left[C(1+A(T)-A(L_u)-2A(R_s))-V(R_s)+A(L_u)A(R_s)\right]\\ &+z^2[J(1)-J(L_u+R_s)]. \end{aligned}

To show (21), it is enough to show that every coefficient of (1−z)2(1-z)^2, 2z(1−z)2z(1-z), z2z^2 is nonnegative. The first coefficient is nonnegative by Lemma 5.2.

For the last coefficient, put

h=1−Lu−Rs.h=1-L_u-R_s.

Then 0≤h≤10\leq h\leq1 is concave,

V(Lu+Rs)=V(1−h)=V(h),A(h)=1−A(Lu)−A(Rs),V(L_u+R_s)=V(1-h)=V(h),\qquad A(h)=1-A(L_u)-A(R_s),

and hence

J(1)−J(Lu+Rs)=2CA(h)−V(h)≥0J(1)-J(L_u+R_s)=2CA(h)-V(h)\geq0

by Lemma 5.1 and 2C≥1/32C\geq1/3.

It remains to show that the mixed coefficient is nonnegative. Put q=A(Lu)q=A(L_u). Since 0≤Lu+Rs≤10\leq L_u+R_s\leq1, we have

0≤q≤1−A(Rs).0\leq q\leq1-A(R_s).

For fixed ss, the mixed coefficient

C(1+A(T)−q−2A(Rs))−V(Rs)+qA(Rs)C(1+A(T)-q-2A(R_s))-V(R_s)+qA(R_s)

is an affine function of qq on [0,1−A(Rs)][0,1-A(R_s)]. It therefore suffices to check its values at the endpoints of this interval.

At q=0q=0, it equals

C(1+A(T)−2A(Rs))−V(Rs)=[J(T)−J(Rs)]+[C(1−A(T))−V(T)]≥0C(1+A(T)-2A(R_s))-V(R_s)=[J(T)-J(R_s)]+[C(1-A(T))-V(T)]\geq0

by Lemma 5.2 and (15).

At q=1−A(Rs)q=1-A(R_s), it equals

C(A(T)−A(Rs))+A(Rs)(1−A(Rs))−V(Rs)≥0,C(A(T)-A(R_s))+A(R_s)(1-A(R_s))-V(R_s)\geq0,

since Rs≤TR_s\leq T and

V(Rs)=A(Rs2)−A(Rs)2≤A(Rs)−A(Rs)2=A(Rs)(1−A(Rs)).V(R_s)=A(R_s^2)-A(R_s)^2\leq A(R_s)-A(R_s)^2=A(R_s)(1-A(R_s)).

Thus all three coefficients are nonnegative, proving the claim.

Affine majorants

Lemma 5.4. Let bb be affine and let aa be convex on [0,1][0,1]. If 0≤a≤b≤10 \le a \le b \le1, then

J(a)≤J(b).(22)J(a) \le J(b). \tag*{(22)}

Proof. The idea of the proof is to approximate aa by convex piecewise affine functions, and then use convexity of JJ and Lemma 5.3.

We first reduce to the case where aa vanishes at some point. Subtracting min⁡[0,1]a\min_{[0,1]} a from both aa and bb preserves all the assumptions of the lemma and leaves J(b)−J(a)J(b)-J(a) unchanged, since VV is invariant under addition of constants and AA is linear. Thus we may assume that

min⁡[0,1]a=0.\min_{[0,1]} a = 0.

If b≡0b \equiv0, then a≡0a \equiv0 and there is nothing to prove. Reflecting both the functions and the weight if necessary, we may assume b(1)≥b(0)b(1) \ge b(0). We may further assume that b(1)=1b(1)=1. Indeed, if (22) holds for the functions a/b(1)a/b(1) and b/b(1)b/b(1), then

V(a)−V(b)≤2Cb(1)A(b−a)≤2CA(b−a),V(a)-V(b) \le2Cb(1)A(b-a) \le2CA(b-a),

which is equivalent to J(a)≤J(b)J(a) \le J(b). Here we used the linearity of AA, the fact that VV is homogeneous of degree two, and the inequalities 0<b(1)≤10 < b(1) \le1 and A(b−a)≥0A(b-a) \ge0.

These reductions preserve the condition min⁡[0,1]a=0\min_{[0,1]} a=0. We therefore have

b(t)=b(0)(1−t)+t,0≤b(0)≤1,b(t)=b(0)(1-t)+t,\qquad0 \le b(0) \le1,

and a(t0)=0a(t_0)=0 for some t0∈[0,1]t_0 \in[0,1].

For each NN, subdivide [0,t0][0,t_0] and [t0,1][t_0,1] into NN equal subintervals:

0=u0<⋯<uN=t0=s0<⋯<sN=10=u_0<\cdots<u_N=t_0=s_0<\cdots<s_N=1

(with the degenerate side omitted if t0=0t_0=0 or 11), and let pNp_N be the piecewise affine function equal to aa at all partition points. Then pNp_N is convex,

pN(t0)=0,0≤pN≤b,p_N(t_0)=0,\qquad0 \le p_N \le b,

and pN→ap_N \to a uniformly.

Write

pN=pN−+pN+.p_N=p_N^-+p_N^+.

where

pN−(t)={pN(t),0≤t≤t0,0,t0≤t≤1,pN+(t)={0,0≤t≤t0,pN(t),t0≤t≤1.p_N^-(t)= \begin{cases} p_N(t), & 0 \le t \le t_0,\\ 0, & t_0 \le t \le1, \end{cases} \qquad p_N^+(t)= \begin{cases} 0, & 0 \le t \le t_0,\\ p_N(t), & t_0 \le t \le1. \end{cases}

Since pN≥0p_N \ge0 and pN(t0)=0p_N(t_0)=0, both functions are continuous and convex. Moreover, pN−p_N^- vanishes on [t0,1][t_0,1], so it is a linear combination with nonnegative coefficients of the functions (uj−t)+(u_j-t)_+, where uj≤t0u_j \le t_0. The coefficients are the jumps in its slope. Similarly, pN+p_N^+ is a linear combination with nonnegative coefficients of the functions (t−si)+(t-s_i)_+, where si≥t0s_i \ge t_0. Thus we may write

pN−=∑jαjb(0)Luj,αj≥0,∑jαj≤1,uj≤t0,p_N^-=\sum_j \alpha_j b(0)L_{u_j},\qquad\alpha_j \ge0,\qquad\sum_j\alpha_j \le1,\qquad u_j \le t_0,

and

pN+=∑iβiRsi,βi≥0,∑iβi≤1,si≥t0.p_N^+=\sum_i \beta_i R_{s_i},\qquad\beta_i \ge0,\qquad\sum_i\beta_i \le1,\qquad s_i \ge t_0.

Here ∑iβi≤1\sum_i \beta_i \le1 follows from pN+(1)≤1p_N^+(1) \le1. If b(0)>0b(0)>0, then ∑jαj≤1\sum_j \alpha_j \le1 follows from pN−(0)≤b(0)p_N^-(0) \le b(0). If b(0)=0b(0)=0, then t0=0t_0=0 and pN−≡0p_N^- \equiv0; in this case we take a single term with α0=1\alpha_0=1 and u0=0u_0=0.

In order to apply Lemma 5.3 and use convexity of JJ, we now express pNp_N as a convex combination. Since L0=0L_0=0 and R1=0R_1=0, we may add them to the preceding sums so that

∑jαj=∑iβi=1.\sum_j \alpha_j=\sum_i \beta_i=1.

Hence

pN=∑i,jαjβi(b(0)Luj+Rsi).p_N=\sum_{i,j}\alpha_j\beta_i\bigl(b(0)L_{u_j}+R_{s_i}\bigr).

By Lemma 5.3 and the convexity of JJ,

J(pN)≤∑i,jαjβiJ(b(0)Luj+Rsi)≤∑i,jαjβiJ(b)=J(b).J(p_N)\le\sum_{i,j}\alpha_j\beta_iJ\bigl(b(0)L_{u_j}+R_{s_i}\bigr)\le\sum_{i,j}\alpha_j\beta_iJ(b)=J(b).

Letting N→∞N\to\infty gives J(a)≤J(b)J(a)\le J(b). □

Proposition 5.5. Let aa be convex and bb be concave on [0,1][0,1] with 0≤a≤b≤10\le a\le b\le1. Then

∣V(a)−V(b)∣≤2CA(b−a).(23)|V(a)-V(b)|\le2CA(b-a). \tag*{(23)}

Proof. We first prove that J(a)≤J(b)J(a)\le J(b). It suffices to do so when b>0b>0 on [0,1][0,1]: the general case follows by applying the result to (a+δ)/(1+δ)(a+\delta)/(1+\delta) and (b+δ)/(1+δ)(b+\delta)/(1+\delta) and letting δ→0+\delta\to0^+, using the continuity of JJ under uniform convergence.

Let L(t)=(1−t)b(0)+tb(1)L(t)=(1-t)b(0)+tb(1). Then

0≤a≤L≤b≤1,0\le a\le L\le b\le1,

and Lemma 5.4 gives J(a)≤J(L)J(a)\le J(L). Since L>0L>0 on [0,1][0,1], for sufficiently small ε>0\varepsilon>0 the function

g=(1+ε)L−εbg=(1+\varepsilon)L-\varepsilon b

is nonnegative. It is also convex and satisfies g≤Lg\le L, so Lemma 5.4 gives J(g)≤J(L)J(g)\le J(L). Since

L=g+εb1+ε,L=\frac{g+\varepsilon b}{1+\varepsilon},

the convexity of JJ yields

J(L)≤J(g)+εJ(b)1+ε≤J(L)+εJ(b)1+ε.J(L)\le\frac{J(g)+\varepsilon J(b)}{1+\varepsilon}\le\frac{J(L)+\varepsilon J(b)}{1+\varepsilon}.

Thus J(L)≤J(b)J(L)\le J(b), and hence J(a)≤J(b)J(a)\le J(b). Equivalently,

V(a)−V(b)≤2CA(b−a).(24)V(a)-V(b)\le2CA(b-a). \tag*{(24)}

Applying the same argument to the convex function 1−b1-b and the concave function 1−a1-a, satisfying 0≤1−b≤1−a≤10\le1-b\le1-a\le1, and using

V(1−f)=V(f),(1−a)−(1−b)=b−a,V(1-f)=V(f),\qquad(1-a)-(1-b)=b-a,

gives

V(b)−V(a)≤2CA(b−a).V(b)-V(a)\le2CA(b-a).

Together with (24), this proves (23). □

Proof of Proposition 2.1

Since h≥0h \ge0 is concave and not identically zero, h>0h>0 in Ω\Omega. Throughout the proof, AA and VV denote the functionals associated with Lebesgue measure on Ω\Omega. Thus,

A(f)=1∣Ω∣∫Ωf(y) dy,V(f)=A(f2)−A(f)2.A(f)=\frac{1}{|\Omega|}\int_{\Omega} f(y)\,dy,\qquad V(f)=A(f^2)-A(f)^2.

Observe that the quantity we need to bound is

1∣Ω∣∫Ωm dy−∫Ωhm dy∫Ωh dy=A(m)−A(hm)A(h).\frac{1}{|\Omega|}\int_{\Omega}m\,dy-\frac{\int_{\Omega}hm\,dy}{\int_{\Omega}h\,dy}=A(m)-\frac{A(hm)}{A(h)}.

Since AA is linear and

h=b−a,m=a+b2,h=b-a,\qquad m=\frac{a+b}{2},

we obtain

A(m)=A(a)+A(b)2,hm=b2−a22,A(h)=A(b)−A(a).A(m)=\frac{A(a)+A(b)}{2},\qquad hm=\frac{b^2-a^2}{2},\qquad A(h)=A(b)-A(a).

Consequently,

A(m)−A(hm)A(h)=12(A(a)+A(b)−A(b2)−A(a2)A(b)−A(a))=V(a)−V(b)2A(h).(25)A(m)-\frac{A(hm)}{A(h)}=\frac{1}{2}\left(A(a)+A(b)-\frac{A(b^2)-A(a^2)}{A(b)-A(a)}\right)=\frac{V(a)-V(b)}{2A(h)}. \tag*{(25)}

We consider the planar case and the higher-dimensional case separately. If n=2n=2, then Ω⊂R\Omega\subset\mathbb{R} is an interval. After an affine reparametrization of Ω\Omega by (0,1)(0,1), normalized Lebesgue measure on Ω\Omega becomes the constant weight, which is precisely the weight ℓn−2≡1\ell^{n-2}\equiv1 from § 5. Extending aa and bb continuously to [0,1][0,1] and applying Proposition 5.5, we obtain

∣V(a)−V(b)∣≤2CA(h).|V(a)-V(b)|\le2CA(h).

Combining this with (25), we obtain

∣1∣Ω∣∫Ωm dy−∫Ωhm dy∫Ωh dy∣=∣A(m)−A(hm)A(h)∣≤C.\left|\frac{1}{|\Omega|}\int_{\Omega}m\,dy-\frac{\int_{\Omega}hm\,dy}{\int_{\Omega}h\,dy}\right|=\left|A(m)-\frac{A(hm)}{A(h)}\right|\le C.

This proves (5) when n=2n=2.

We may therefore assume n≥3n\ge3 in the localization argument below. To reach a contradiction, suppose that

A(m)>A(hm)A(h)+C.A(m)>\frac{A(hm)}{A(h)}+C.

Then the continuous functions

φ=m−A(hm)A(h)−C,ψ=h(m−A(hm)A(h))\varphi=m-\frac{A(hm)}{A(h)}-C,\qquad\psi=h\left(m-\frac{A(hm)}{A(h)}\right)

satisfy

∫Ωφ>0,∫Ωψ=0.\int_{\Omega}\varphi>0,\qquad\int_{\Omega}\psi=0.

By the localization lemma of Kannan, Lovász, and Simonovits [3 Corollary 2.4], there exist a line segment I⊂ΩI\subset\Omega and a nonnegative affine function ℓ\ell on II such that

∫Iφ(x)ℓn−2(x) dx>0,∫Iψ(x)ℓn−2(x) dx=0.(26)\int_I\varphi(x)\ell^{n-2}(x)\,dx>0,\qquad\int_I\psi(x)\ell^{n-2}(x)\,dx=0. \tag*{(26)}

where dxdx denotes length measure along II.

Let ν\nu denote the normalized measure on II with density proportional to ℓn−2\ell^{n-2}. Using (26) together with the definitions of φ\varphi and ψ\psi, we get

Aν(m)>A(hm)A(h)+C,Aν(hm)Aν(h)=A(hm)A(h).A_\nu(m) > \frac{A(hm)}{A(h)} + C,\qquad\frac{A_\nu(hm)}{A_\nu(h)} = \frac{A(hm)}{A(h)}.

Here Aν(h)>0A_\nu(h) > 0 because h>0h > 0 on Ω\Omega. After an affine parametrization of II by [0,1][0,1], the measure ν\nu has a density of the form (14). The restrictions of aa and bb remain respectively convex and concave and satisfy 0≤a≤b≤10 \leq a \leq b \leq1. Thus Proposition 5.5 applies.

The same algebra as in (25), now applied to ν\nu, together with Proposition 5.5, yields

∣Aν(m)−Aν(hm)Aν(h)∣=∣Vν(a)−Vν(b)∣2Aν(h)≤C,\left|A_\nu(m)-\frac{A_\nu(hm)}{A_\nu(h)}\right|=\frac{\left|V_\nu(a)-V_\nu(b)\right|}{2A_\nu(h)}\leq C,

a contradiction.

To obtain the opposite inequality, apply the same argument to the functions

a~=1−b,b~=1−a,\widetilde{a}=1-b,\qquad\widetilde{b}=1-a,

where a~\widetilde{a} is convex, b~\widetilde{b} is concave, and 0≤a~≤b~≤10 \leq\widetilde{a} \leq\widetilde{b} \leq1. For this pair,

h~=b~−a~=h,m~=a~+b~2=1−m.\widetilde{h}=\widetilde{b}-\widetilde{a}=h,\qquad\widetilde{m}=\frac{\widetilde{a}+\widetilde{b}}{2}=1-m.

The inequality already proved therefore gives

A(m~)≤A(h~m~)A(h~)+C.A(\widetilde{m})\leq\frac{A(\widetilde{h}\widetilde{m})}{A(\widetilde{h})}+C.

Since

A(m~)=1−A(m),A(h~m~)A(h~)=1−A(hm)A(h),A(\widetilde{m})=1-A(m),\qquad\frac{A(\widetilde{h}\widetilde{m})}{A(\widetilde{h})}=1-\frac{A(hm)}{A(h)},

we obtain

A(1−m)≤1−A(hm)A(h)+C,A(1-m)\leq1-\frac{A(hm)}{A(h)}+C,

or equivalently

A(m)≥A(hm)A(h)−C.A(m)\geq\frac{A(hm)}{A(h)}-C.

This proves the opposite inequality and hence (5). □\square

Acknowledgments. The project was made possible by SQuaRE at the American Institute of Mathematics. The authors thank AIM for providing a supportive and mathematically rich environment.

AI Disclaimer. ChatGPT was used in exploratory discussions and in the preparation of this manuscript.

S. Myroshnychenko, Department of Mathematics and Statistics, University of the Fraser Valley, Abbotsford, BC V2S 7M8, Canada

Email address: [email protected]

D. Ryabogin, Department of Mathematics, Kent State University, Kent, OH 44242, USA

Email address: [email protected]

K. Tatarko, Department of Pure Mathematics, University of Waterloo, Waterloo, ON N2L 3G1, Canada

Email address: [email protected]

V. Yaskin, Department of Mathematical and Statistical Sciences, University of Alberta, Edmonton, AB T6G 2G1, Canada

Email address: [email protected]

References

  1. [1]H. T. Croft, K. J. Falconer, and R. K. Guy, Unsolved Problems in Geometry, Problem Books in Mathematics, Springer-Verlag, New York, 1991.
  2. [2]H. Federer, Geometric Measure Theory, Grundlehren der mathematischen Wissenschaften, vol. 153, Springer-Verlag, New York, 1969.
  3. [3]R. Kannan, L. Lovász, and M. Simonovits, Isoperimetric problems for convex bodies and a localization lemma, Discrete Comput. Geom. 13 (1995), 541–559.DOI
  4. [4]S. Myroshnychenko, K. Tatarko, and V. Yaskin, How far apart can the projection of the centroid of a convex body and the centroid of its projection be?, Math. Ann. 390 (2024), no. 1, 1155–1169.arxiv.org/abs/2212.14456
  5. [5]F. Nazarov, D. Ryabogin, and V. Yaskin, On the maximal distance between the centers of mass of a planar convex body and its boundary, Discrete Comput. Geom. 73 (2025), no. 4, 1016–1036.arxiv.org/abs/2305.15646

Paper details

vlnc:W_fpbjdt3p
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

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

  • Mathematics

ChatGPT

View statement
Looking for related papers…
Contents