betaFeedback

WorksMathematics

The Zaporozhets-Tarasov Conjecture on Mean Distances

Abstract

For a convex body K⊂RdK \subset \mathbb R^d let Δ(K)Δ(K) be the expected distance between two independent uniform points of KK, and let θ(K)θ(K) be the corresponding expectation for normalized surface measure on ∂K\partial K. The Zaporozhets-Tarasov conjecture asserts Δ(K)≤θ(K)Δ(K) \le θ(K). We prove this conjecture in case d=2d=2. In addition, we give a six-vertex convex polytope in R3\mathbb R^3 for which the reverse strict inequality holds, and obtain counterexamples in every dimension d≥3d \ge 3 by taking products with segments.

Introduction

Let K⊂RdK \subset\mathbb{R}^{d} be a compact convex body. The conjecture of Zaporozhets and Tarasov asks whether

Δ(K):=E∣X1−X2∣≤E∣B1−B2∣=:θ(K)\Delta(K) := \mathbb{E}|X_{1}-X_{2}| \le\mathbb{E}|B_{1}-B_{2}| =: \theta(K)

for every convex body K⊂RdK \subset\mathbb{R}^{d}, where X1,X2X_{1}, X_{2} are independent uniform in KK with respect to the normalized Lebesgue measure and B1,B2B_{1}, B_{2} are independent uniform on ∂K\partial K with respect to the normalized (d−1)(d-1)-dimensional Hausdorff measure. The conjecture was proposed by D. N. Zaporozhets and A. S. Tarasov in 2019 through informal communication. To the best of our knowledge, its first explicit formulation in the literature is due to Tokmachev [3] (Eq. (3) on p. 3), who attributes it to them; see also discussion before Conjecture 1.1 in the work [2].

Lotnikov proved the planar centrally symmetric case and an analogue for sufficiently high moments in arbitrary dimension [2]. Tokmachev proved the conjecture, in fact a convex-order statement, for planar bodies whose area and boundary centroids coincide and for a class of circumscribed polytopes [4]. Related extremal questions for Δ\Delta and θ\theta were studied in [1, 3]. In particular, Bonnet, Gusakova, Thäle and Zaporozhets proved the sharp inequalities

760<Δ(K)per⁡K<16\frac{7}{60} < \frac{\Delta(K)}{\operatorname{per} K} < \frac{1}{6}

Email address: [email protected] (Eric Shen) for planar convex bodies [1], where per⁡K\operatorname{per} K stands for perimeter of KK. It is also shown that the upper bound is approached by thin rectangles converging to a doubled segment. According to Tokmachev [4, Eq. (5)], Gusakova and Zaporozhets suggested an analogous pair of inequalities for θ\theta

16≤θ(K)per⁡K≤2π2.\frac{1}{6} \le\frac{\theta(K)}{\operatorname{per} K} \le\frac{2}{\pi^2}.

The upper bound, with equality only for the disk, was proved by Tokmachev [3]; the complementary lower bound was established by the author after the initial submission of the present paper, see Section 5 in [8]. Combined with the upper bound for Δ(K)\Delta(K) from [1], it yields another proof of the planar Zaporozhets–Tarasov conjecture.

We prove the Zaporozhets-Tarasov conjecture in the case d=2d = 2, see Corollary 3.1. The result is mostly self-contained, in particular, it does not use the aforementioned inequalities on Δ(K)\Delta(K) from [1] and is obtained as a corollary of a more general statement, see Theorem 2.1. We also prove an analogous inequality for the second moment, see Corollary 3.2. In addition to that, we show that the conjecture is false in every dimension d≥3d \ge3, see Example 4.2, Figure 1, and Proposition 4.5.

The polytope $K_{L,\varepsilon}$ in rescaled coordinates

Figure 1. The polytope KL,εK_{L,\varepsilon} in the rescaled coordinates (t,x,y/ε)(t,x,y/\varepsilon). One transverse width grows with tt, while the other decreases.

Definition 1.1. For a probability measure ρ\rho on R\mathbb{R}, write

D(ρ)=∬R2∣x−y∣ dρ(x)dρ(y).D(\rho) = \iint_{\mathbb{R}^2} \lvert x-y\rvert\,d\rho(x)d\rho(y).

This is the Gini mean difference of ρ\rho.

We slightly abuse this notation throughout the text, writing D(X)D(X) with XX being a random variable or a probability density. In either case, D(X)D(X) is understood as D(ρX)D(\rho_X), where ρX\rho_X denotes the corresponding probability measure.

We use the following standard identities for the Gini mean difference; see also [6 Sec. 2.1.1].

Lemma 1.2. Let ρ\rho, σ\sigma be probability measures on R\mathbb{R} with finite first moments, and let F,GF,G be their distribution functions. Then

D(ρ)=2∫RF(x)(1−F(x)) dx.(1)D(\rho) = 2\int_{\mathbb{R}} F(x)(1-F(x))\,dx. \tag*{(1)}

Moreover, for 0≤λ≤10 \le\lambda\le1,

D(λρ+(1−λ)σ)=λD(ρ)+(1−λ)D(σ)+2λ(1−λ)∫R(F−G)2 dx.(2)D(\lambda\rho+(1-\lambda)\sigma) = \lambda D(\rho)+(1-\lambda)D(\sigma) +2\lambda(1-\lambda)\int_{\mathbb{R}}(F-G)^2\,dx. \tag*{(2)}

Finally, if XX has an atomless distribution ρ\rho, then

D(X)=4Cov⁡(X,F(X)).(3)D(X)=4\operatorname{Cov}(X,F(X)). \tag*{(3)}

Proof. Let XX, X′X' be independent with law ρ\rho. Since

∣X−X′∣=∫R∣1{X≤x}−1{X′≤x}∣ dx,\lvert X-X'\rvert=\int_{\mathbb{R}}\left|\mathbf{1}_{\{X\le x\}}-\mathbf{1}_{\{X'\le x\}}\right|\,dx,

taking expectations and applying Tonelli’s theorem gives (1.1). The distribution function of λρ+(1−λ)σ\lambda\rho+ (1-\lambda)\sigma is H=λF+(1−λ)GH = \lambda F + (1-\lambda)G. The identity

H(1−H)=λF(1−F)+(1−λ)G(1−G)+λ(1−λ)(F−G)2H(1-H) = \lambda F(1-F) + (1-\lambda)G(1-G) + \lambda(1-\lambda)(F-G)^2

then proves (1.2).

Finally, if ρ\rho is atomless, then F(X)F(X) is uniform on [0,1][0,1]. By independence and symmetry,

2Cov⁡(X,F(X))=2E[X1{X′≤X}]−EX=E[(X−X′)1{X′≤X}]=12E∣X−X′∣.2\operatorname{Cov}(X,F(X)) = 2\mathbb{E}[X\mathbf{1}_{\{X'\leq X\}}] - \mathbb{E}X = \mathbb{E}[(X-X')\mathbf{1}_{\{X'\leq X\}}] = \frac{1}{2}\mathbb{E}|X-X'|.

This proves (1.3).

The main result immediately follows from a stronger directional inequality.

Theorem 1.3. Let K⊂R2K \subset\mathbb{R}^{2} be a convex body. Fix a unit vector uu. Let pup_u be the pushforward of normalized area measure on KK under x↦⟨x,u⟩x \mapsto\langle x,u\rangle, and let quq_u be the corresponding pushforward of normalized arclength measure on ∂K\partial K. Then

D(qu)≥D(pu).(4)\mathcal{D}(q_u) \geq\mathcal{D}(p_u). \tag*{(4)}

The inequality is strict if uu is not perpendicular to any nondegenerate segment contained in ∂K\partial K.

Plan. After rotating and rescaling, the projection of KK is [0,1][0,1] and

K={(x,y):0≤x≤1, g(x)≤y≤f(x)},K = \{(x,y): 0 \leq x \leq1,\ g(x) \leq y \leq f(x)\},

where ff is concave and gg is convex. Put

W=f−g,a=f′,b=g′.W = f-g,\qquad a=f',\qquad b=g'.

The projected area density is proportional to WW, while the projected boundary density is proportional to

1+a2+1+b2.\sqrt{1+a^2}+\sqrt{1+b^2}.

The proof has two steps. First, we find two densities φ\varphi and ψ\psi, nonincreasing and nondecreasing, respectively, such that the projected arclength density qq satisfies q=φ+ψ2q=\frac{\varphi+\psi}{2} while the projected area density pp is proportional to the difference of the CDFs of φ\varphi and ψ\psi. Then we apply a one-dimensional theorem saying that an endpoint mixture is more dispersed than the normalized occupation measure between the endpoints. The only substantial step is the second one, to which we now turn.

The monotone-endpoint theorem

Theorem 2.1. Let φ\varphi be a nonincreasing probability density and ψ\psi a nondecreasing probability density on [0,1][0,1]. Let

Φ(x)=∫0xφ(s) ds,Ψ(x)=∫0xψ(s) ds.\Phi(x)=\int_{0}^{x}\varphi(s)\,\mathrm{d}s,\qquad\Psi(x)=\int_{0}^{x}\psi(s)\,\mathrm{d}s.

Put

H=Φ−Ψ,A=∫01H(x) dx,H=\Phi-\Psi,\qquad A=\int_{0}^{1}H(x)\,\mathrm{d}x,

and assume A>0A > 0. Define

p(x)=H(x)A,q(x)=φ(x)+ψ(x)2.p(x) = \frac{H(x)}{A}, \qquad q(x) = \frac{\varphi(x) + \psi(x)}{2}.

Then

D(q)>D(p).\mathcal{D}(q) > \mathcal{D}(p).

Since Φ\Phi is concave and Ψ\Psi is convex, Φ(x)≥x≥Ψ(x)\Phi(x) \ge x \ge\Psi(x). Thus H≥0H \ge0, and pp is indeed a probability density.

Remark 2.2. There is also a geometric reformulation. Let x(u)=Φ−1(u)x(u) = \Phi^{-1}(u) and y(u)=Ψ−1(u)y(u) = \Psi^{-1}(u). If U∼Unif⁡[0,1]U \sim\operatorname{Unif}[0,1], then x(U)∼φx(U) \sim\varphi, y(U)∼ψy(U) \sim\psi. Consider

L={(u,z):0≤u≤1, x(u)≤z≤y(u)}.L = \{(u,z): 0 \le u \le1,\ x(u) \le z \le y(u)\}.

Since xx is convex and yy is concave, LL is convex. Choosing one of the two endpoints of the (uniformly distributed) vertical slice [x(U),y(U)][x(U),y(U)] with equal probability gives the law qq. Schematically,

q(z) dz=12∫01(δx(u)+δy(u))(dz) du.q(z)\,dz = \frac{1}{2}\int_0^1 \left(\delta_{x(u)} + \delta_{y(u)}\right)(dz)\,du.

On the other hand, the length of the horizontal section of LL at height zz is

∣{u:x(u)≤z≤y(u)}∣=Φ(z)−Ψ(z)=H(z).\left|\{u:x(u) \le z \le y(u)\}\right| = \Phi(z) - \Psi(z) = H(z).

Consequently, pp is the second-coordinate distribution of a uniform point of LL, that is, choose a slice with a probability proportional to its length and then a random point in it. Explicitly,

p(z) dz=∫011{x(u)≤z≤y(u)} du∫01(y(u)−x(u)) du dz.p(z)\,dz = \frac{\int_0^1 \mathbf{1}_{\{x(u)\le z\le y(u)\}}\,du}{\int_0^1 (y(u)-x(u))\,du}\,dz.

Every nonincreasing probability density on [0,1][0,1] is a mixture of uniform densities on intervals [0,t][0,t], as shown by the following Proposition. This is a one-sided form of Khintchine’s representation theorem for unimodal distributions, see [5].

Proposition 2.3. Let φ\varphi be a probability density on [0,1][0,1] admitting a nonincreasing representative, which we also denote by φ\varphi. There is a random variable T∈[0,1]T \in[0,1] such that, for U∼Unif⁡[0,1]U \sim\operatorname{Unif}[0,1] independent of TT, the variable

X=TUX = TU

has density φ\varphi. Likewise, for every nondecreasing density ψ\psi, there is S∈[0,1]S \in[0,1] such that

Y=1−SV,V∼Unif⁡[0,1] independent of S,Y = 1 - SV,\qquad V \sim\operatorname{Unif}[0,1]\ \text{independent of }S,

has density ψ\psi.

Proof. As the law of TT one may take the Stieltjes mixing measure

ρ(dt)=−t dφ(t)+φ(1−)δ1(dt),\rho(dt)=-t\,d\varphi(t)+\varphi(1-)\delta_1(dt),

for which

φ(x)=∫[x,1]1tρ(dt)for a.e. x∈(0,1).\varphi(x)=\int_{[x,1]} \frac{1}{t}\rho(dt)\quad\text{for a.e. }x\in(0,1).

As φ\varphi is nonincreasing, ρ\rho is nonnegative and the Stieltjes integration by parts gives

ρ([0,1])=∫01φ=1.\rho([0,1])=\int_0^1\varphi=1.

We first record two consequences of Proposition 2.3.

Lemma 2.4. Let XX have a nonincreasing density supported on [0,1][0,1], and let YY have a nondecreasing one. Then

D(X)≥2Var⁡X+23EX(1−EX),D(Y)≥2Var⁡Y+23EY(1−EY).(5)\begin{aligned} \mathcal{D}(X)&\ge2\operatorname{Var}X+\frac{2}{3}\mathbb{E}X(1-\mathbb{E}X),\\ \mathcal{D}(Y)&\ge2\operatorname{Var}Y+\frac{2}{3}\mathbb{E}Y(1-\mathbb{E}Y). \tag*{(5)} \end{aligned}

Moreover,

E(X−EX)3≥0,E(Y−EY)3≤0.\mathbb{E}(X-\mathbb{E}X)^3\ge0,\qquad\mathbb{E}(Y-\mathbb{E}Y)^3\le0.

Proof. By Proposition 2.3, there are mutually independent T,S∈[0,1]T,S\in[0,1] and U,V∼Unif⁡[0,1]U,V\sim\operatorname{Unif}[0,1] such that X=TUX=TU and Y=1−SVY=1-SV. Since UU is uniform on [0,1][0,1] and independent of TT,

EX=ET2,EX2=ET23,EX3=ET34.\mathbb{E}X=\frac{\mathbb{E}T}{2},\qquad\mathbb{E}X^2=\frac{\mathbb{E}T^2}{3},\qquad\mathbb{E}X^3=\frac{\mathbb{E}T^3}{4}.

Consequently,

4E(X−EX)3=ET3−2ETET2+(ET)3=E[T(T−ET)2]≥0.4\mathbb{E}(X-\mathbb{E}X)^3=\mathbb{E}T^3-2\mathbb{E}T\mathbb{E}T^2+(\mathbb{E}T)^3=\mathbb{E}\left[T(T-\mathbb{E}T)^2\right]\ge0.

Applying the same argument to 1−Y=SV1-Y=SV gives E(Y−EY)3≤0\mathbb{E}(Y-\mathbb{E}Y)^3\le0.

Take U,U′∼Unif⁡[0,1]U,U'\sim\operatorname{Unif}[0,1] independent. Direct integration gives

E∣tU−sU′∣=t+s6+(t−s)23max⁡(t,s),\mathbb{E}|tU-sU'|=\frac{t+s}{6}+\frac{(t-s)^2}{3\max(t,s)},

with the second term read as zero when s=t=0s=t=0. Since max⁡(s,t)≤1\max(s,t)\le1,

E∣tU−sU′∣≥t+s6+(t−s)23.\mathbb{E}|tU-sU'|\ge\frac{t+s}{6}+\frac{(t-s)^2}{3}.

Now let (T′,U′)(T',U') be an independent copy of (T,U)(T,U). Then

D(X)=E∣TU−T′U′∣=ET,T′E(∣TU−T′U′∣∣T,T′)≥E(T+T′6+(T−T′)23).\mathcal{D}(X)=\mathbb{E}|TU-T'U'|=\mathbb{E}_{T,T'}\mathbb{E}(|TU-T'U'|\mid T,T')\ge\mathbb{E}\left(\frac{T+T'}{6}+\frac{(T-T')^2}{3}\right).

Using that TT and T′T' are independent gives

D(X)≥ET3+23Var⁡T.\mathcal{D}(X)\ge\frac{\mathbb{E}T}{3}+\frac{2}{3}\operatorname{Var}T.

Using

Var⁡T=3Var⁡X−(EX)2,ET=2EX,\operatorname{Var}T=3\operatorname{Var}X-(\mathbb{E}X)^2,\qquad\mathbb{E}T=2\mathbb{E}X,

we obtain the first inequality in (5). Applying the same argument to 1−Y1-Y gives the second one.

Lemma 2.5. Let φ\varphi, ψ\psi, pp and AA be as in Theorem 2.1, let XX, YY, ZZ have densities φ\varphi, ψ\psi, and pp, respectively, and put m=(EX+EY)/2m = (\mathbb{E}X + \mathbb{E}Y)/2. Then

E(Z−m)2≤Var⁡X+Var⁡Y2+A212.\mathbb{E}(Z-m)^2 \le\frac{\operatorname{Var} X+\operatorname{Var} Y}{2}+\frac{A^2}{12}.

Proof. Recall that we denote by Φ\Phi and Ψ\Psi the CDFs corresponding to φ\varphi and ψ\psi, respectively. Integrating the CDFs gives

A=∫01(Φ−Ψ)=EY−EXA=\int_0^1(\Phi-\Psi)=\mathbb{E}Y-\mathbb{E}X

Since H′=φ−ψH'=\varphi-\psi and H(0)=H(1)=0H(0)=H(1)=0, integration by parts yields

3AE(Z−m)2=∫01((x−m)3)′H(x) dx=−∫01(x−m)3(φ(x)−ψ(x)) dx=E(Y−m)3−E(X−m)3.\begin{aligned} 3A\mathbb{E}(Z-m)^2=\int_0^1\left((x-m)^3\right)'H(x)\,\mathrm{d}x=-\int_0^1(x-m)^3(\varphi(x)-\psi(x))\,\mathrm{d}x \\ &=\mathbb{E}(Y-m)^3-\mathbb{E}(X-m)^3. \end{aligned}

Now

X−m=(X−EX)−A2,Y−m=(Y−EY)+A2.X-m=(X-\mathbb{E}X)-\frac{A}{2},\qquad Y-m=(Y-\mathbb{E}Y)+\frac{A}{2}.

Expanding the two cubes and using E(X−EX)=E(Y−EY)=0\mathbb{E}(X-\mathbb{E}X)=\mathbb{E}(Y-\mathbb{E}Y)=0, we obtain

3AE(Z−m)2=E(Y−EY)3−E(X−EX)3+3A2(Var⁡X+Var⁡Y)+A34.3A\mathbb{E}(Z-m)^2=\mathbb{E}(Y-\mathbb{E}Y)^3-\mathbb{E}(X-\mathbb{E}X)^3+\frac{3A}{2}(\operatorname{Var}X+\operatorname{Var}Y)+\frac{A^3}{4}.

The first two terms have nonpositive sum by Lemma 2.4. Dividing by 3A3A proves the claim.

Proof of Theorem 2.1. Let X,Y,Z,mX,Y,Z,m be as in Lemma 2.5. Let PP denote the CDF of ZZ. Since ZZ has a density, P(Z)P(Z) is uniform on [0,1][0,1]. Hence the Gini covariance identity 1.3 and 2uv≤u2+v22uv\le u^2+v^2 give

D(p)=4Cov⁡(Z,P(Z))=4E[(Z−m)(P(Z)−12)]≤2E(Z−m)2+2E(P(Z)−12)2=2E(Z−m)2+16≤Var⁡X+Var⁡Y+A26+16,\begin{aligned} D(p)&=4\operatorname{Cov}(Z,P(Z))=4\mathbb{E}\left[(Z-m)\left(P(Z)-\frac{1}{2}\right)\right]\\ &\le2\mathbb{E}(Z-m)^2+2\mathbb{E}\left(P(Z)-\frac{1}{2}\right)^2\\ &=2\mathbb{E}(Z-m)^2+\frac{1}{6}\le\operatorname{Var}X+\operatorname{Var}Y+\frac{A^2}{6}+\frac{1}{6}, \end{aligned}

where the last inequality follows from Lemma 2.5.

On the other hand, since qq is an equal-probability mixture of laws of XX and YY, equation (2) from Lemma 1.2 gives

D(q)=D(X)+D(Y)2+12∫01H(x)2 dx.(6)D(q)=\frac{D(X)+D(Y)}{2}+\frac{1}{2}\int_0^1H(x)^2\,\mathrm{d}x. \tag*{(6)}

Now, (5) yields

D(q)≥Var⁡X+Var⁡Y+EX(1−EX)+EY(1−EY)3+12∫01H(x)2 dx.(7)D(q)\ge\operatorname{Var}X+\operatorname{Var}Y+\frac{\mathbb{E}X(1-\mathbb{E}X)+\mathbb{E}Y(1-\mathbb{E}Y)}{3}+\frac{1}{2}\int_0^1H(x)^2\,\mathrm{d}x. \tag*{(7)}

By Cauchy–Schwarz, ∫01H2≥A2\int_{0}^{1} H^{2} \ge A^{2}. In fact, the inequality is strict. Indeed, HH is continuous, H(0)=H(1)=0H(0)=H(1)=0, and A>0A>0. Thus HH cannot be constant almost everywhere. Subtracting the upper bound for D(p)\mathcal{D}(p), and using A=EY−EXA=\mathbb{E}Y-\mathbb{E}X, gives

D(q)−D(p)≥EX(1−EX)+EY(1−EY)3+12∫01H(x)2 dx−A26−16=2EX(1−EX)+2EY(1−EY)+2A2−16+12(∫01H(x)2 dx−A2)=(1−2EX)(2EY−1)6+12(∫01H(x)2 dx−A2)>0.\begin{aligned} \mathcal{D}(q)-\mathcal{D}(p) \ge\frac{\mathbb{E}X(1-\mathbb{E}X)+\mathbb{E}Y(1-\mathbb{E}Y)}{3}+\frac{1}{2}\int_{0}^{1}H(x)^{2}\,dx-\frac{A^{2}}{6}-\frac{1}{6} \\ &=\frac{2\mathbb{E}X(1-\mathbb{E}X)+2\mathbb{E}Y(1-\mathbb{E}Y)+2A^{2}-1}{6}+\frac{1}{2}\left(\int_{0}^{1}H(x)^{2}\,dx-A^{2}\right) \\ &=\frac{(1-2\mathbb{E}X)(2\mathbb{E}Y-1)}{6}+\frac{1}{2}\left(\int_{0}^{1}H(x)^{2}\,dx-A^{2}\right)>0. \end{aligned}

In the right-hand side of the final inequality the former summand is non-negative, as EX=ET/2≤1/2\mathbb{E}X=\mathbb{E}T/2\le1/2 and EY=1−ES/2≥1/2\mathbb{E}Y=1-\mathbb{E}S/2\ge1/2, where TT and SS are as in Proposition 2.3 and Lemma 2.4. □\square

The analogous result for the second moment is immediate.

Corollary 2.6. In the setting of Theorem 2.1, let QQ be a random variable with density qq, and let ZZ, as before, have density pp. Then Var⁡(Q)>Var⁡(Z)\operatorname{Var}(Q)>\operatorname{Var}(Z).

Proof. Since qq is an equal-probability mixture of XX and YY,

EQ=m,Var⁡Q=Var⁡X+Var⁡Y2+A24.\mathbb{E}Q=m,\qquad\operatorname{Var}Q=\frac{\operatorname{Var}X+\operatorname{Var}Y}{2}+\frac{A^{2}}{4}.

By Lemma 2.5,

Var⁡Z≤E(Z−m)2≤Var⁡X+Var⁡Y2+A212<Var⁡Q.\operatorname{Var}Z\le\mathbb{E}(Z-m)^{2}\le\frac{\operatorname{Var}X+\operatorname{Var}Y}{2}+\frac{A^{2}}{12}<\operatorname{Var}Q.

□\square

Proof of Theorem 1.3

Fix a direction uu that is not perpendicular to any nondegenerate segment of ∂K\partial K, normalize its projection to [0,1][0,1], and use the notation from the plan:

W=f−g,a=f′,b=g′.W=f-g,\qquad a=f',\qquad b=g'.

The functions ff, gg are absolutely continuous and their derivatives are understood almost everywhere. The functions aa and bb are respectively nonincreasing and nondecreasing, while

W′=a−b,W(0)=W(1)=0.W'=a-b,\qquad W(0)=W(1)=0.

The normalized projected area density is

p(x)=W(x)A,A=∫01W(x) dx.(8)p(x)=\frac{W(x)}{A},\qquad A=\int_{0}^{1}W(x)\,dx. \tag*{(8)}

The unnormalized projected arclength boundary density is

r(x)=1+a(x)2+1+b(x)2.r(x)=\sqrt{1+a(x)^{2}}+\sqrt{1+b(x)^{2}}.

Let

R=∫01r(x) dx.R=\int_{0}^{1}r(x)\,dx.

so that the normalized projected boundary density is q(x)=r(x)Rq(x)=\frac{r(x)}{R}. Define

φ(x)=1+a(x)2+a(x)+1+b(x)2−b(x)R\varphi(x)=\frac{\sqrt{1+a(x)^2}+a(x)+\sqrt{1+b(x)^2}-b(x)}{R}

and

ψ(x)=1+a(x)2−a(x)+1+b(x)2+b(x)R.\psi(x)=\frac{\sqrt{1+a(x)^2}-a(x)+\sqrt{1+b(x)^2}+b(x)}{R}.

Clearly,

q=φ+ψ2.q=\frac{\varphi+\psi}{2}.

Since aa is nonincreasing and bb is nondecreasing, it follows that φ\varphi is nonincreasing and ψ\psi is nondecreasing. Both are nonnegative. Moreover,

∫01(a−b) dx=W(1)−W(0)=0,\int_0^1(a-b)\,dx=W(1)-W(0)=0,

and hence

∫01φ(x) dx=1R∫01(r(x)+a(x)−b(x)) dx=1,\int_0^1\varphi(x)\,dx=\frac{1}{R}\int_0^1(r(x)+a(x)-b(x))\,dx=1,

and similarly

∫01ψ(x) dx=1.\int_0^1\psi(x)\,dx=1.

Thus φ\varphi and ψ\psi are probability densities of the type appearing in Theorem 2.1. Let Φ\Phi and Ψ\Psi be their distribution functions. Then

Φ(x)−Ψ(x)=2R∫0x(a(s)−b(s)) ds=2W(x)R.\Phi(x)-\Psi(x)=\frac{2}{R}\int_0^x(a(s)-b(s))\,ds=\frac{2W(x)}{R}.

Consequently,

∫01(Φ(x)−Ψ(x)) dx=2AR,\int_0^1(\Phi(x)-\Psi(x))\,dx=\frac{2A}{R},

and therefore

Φ(x)−Ψ(x)∫01(Φ−Ψ)=W(x)A=p(x).\frac{\Phi(x)-\Psi(x)}{\int_0^1(\Phi-\Psi)}=\frac{W(x)}{A}=p(x).

Since q=(φ+ψ)/2q=(\varphi+\psi)/2, Theorem 2.1 gives

D(q)>D(p).\mathcal{D}(q)>\mathcal{D}(p).

There are at most countably many directions perpendicular to nondegenerate segments of ∂K\partial K. For these exceptional directions, the non-strict inequality follows by continuity in uu from the inequality already proved for all other directions. Indeed, for a fixed random vector WW, the dominated convergence theorem implies that the function

fW(u)=E∣⟨W,u⟩∣f_W(u)=\mathbb{E}|\langle W,u\rangle|

is continuous in uu, and it remains to note that D(qu)\mathcal{D}(q_u) and D(pu)\mathcal{D}(p_u) are exactly of this form. This completes the proof.

Corollary 3.1. For every planar convex body KK,

E∣X1−X2∣<E∣B1−B2∣,\mathbb{E}|X_1-X_2|<\mathbb{E}|B_1-B_2|,

where X1,X2X_1,X_2 are uniform in KK and B1,B2B_1,B_2 are uniform on ∂K\partial K.

Proof. Indeed, for normalized angular measure σˉ\bar{\sigma} on S1S^1,

∣z∣=π2∫S1∣⟨z,u⟩∣ dσˉ(u).(9)|z|=\frac{\pi}{2}\int_{S^1}|\langle z,u\rangle|\,d\bar{\sigma}(u). \tag*{(9)}

See e.g. [7]. Integrating (9) against the two-point area and boundary distributions, and then using (4), gives the result. The inequality is strict as (4) is strict for almost every uu.

Corollary 3.2. For every planar convex body KK,

E∣X1−X2∣2<E∣B1−B2∣2,\mathbb{E}|X_1-X_2|^2<\mathbb{E}|B_1-B_2|^2,

where X1,X2X_1,X_2 are uniform in KK and B1,B2B_1,B_2 are uniform on ∂K\partial K.

Proof. The proof of Theorem 1.3 shows that, for almost every u∈S1u\in S^1, the projected area and boundary densities are obtained from the same pair of nonincreasing and nondecreasing densities as in Theorem 2.1. Corollary 2.6 gives

Var⁡⟨X1,u⟩<Var⁡⟨B1,u⟩.\operatorname{Var}\langle X_1,u\rangle<\operatorname{Var}\langle B_1,u\rangle.

Observe that

E⟨X1−X2,u⟩2=2Var⁡⟨X1,u⟩,\mathbb{E}\langle X_1-X_2,u\rangle^2=2\operatorname{Var}\langle X_1,u\rangle,

and similarly

E⟨B1−B2,u⟩2=2Var⁡⟨B1,u⟩.\mathbb{E}\langle B_1-B_2,u\rangle^2=2\operatorname{Var}\langle B_1,u\rangle.

Therefore

E⟨X1−X2,u⟩2<E⟨B1−B2,u⟩2\mathbb{E}\langle X_1-X_2,u\rangle^2<\mathbb{E}\langle B_1-B_2,u\rangle^2

for almost every u∈S1u\in S^1. For normalized angular measure σˉ\bar{\sigma} on S1S^1,

∣z∣2=2∫S1⟨z,u⟩2 dσˉ(u).|z|^2=2\int_{S^1}\langle z,u\rangle^2\,d\bar{\sigma}(u).

Integrating the preceding directional inequality over uu therefore gives

E∣X1−X2∣2<E∣B1−B2∣2.\mathbb{E}|X_1-X_2|^2<\mathbb{E}|B_1-B_2|^2.

A long-body observation

In this section, we construct a three-dimensional counterexample to the Zaporozhets-Tarasov conjecture, see Example 4.2. This example is then used to construct counterexamples in higher dimensions, see Proposition 4.5. The main observation, discussed in Remark 4.4, is that in dimension three the area and perimeter profiles of sections orthogonal to a fixed direction can be controlled almost independently. We exploit this freedom to make the projected Gini mean difference strictly larger for the volume measure than for the surface measure. A sufficiently large stretch in the projection direction then yields the required counterexample.

Lemma 4.1. Suppose a random point has the form Z=(LT,U)∈R×RmZ=(LT,U)\in\mathbb{R}\times\mathbb{R}^{m}, where T∈[0,1]T\in[0,1] and the support of UU has diameter at most DD. If Z1,Z2Z_1,Z_2 are independent copies, then

LE∣T1−T2∣≤E∣Z1−Z2∣≤LE∣T1−T2∣+D.L\mathbb{E}|T_1-T_2|\leq\mathbb{E}|Z_1-Z_2|\leq L\mathbb{E}|T_1-T_2|+D.

Proof. The lower bound follows by projection onto the first coordinate. For the upper bound, use

∣(L(T1−T2),U1−U2)∣≤L∣T1−T2∣+∣U1−U2∣.\left|(L(T_1-T_2),U_1-U_2)\right|\leq L|T_1-T_2|+|U_1-U_2|.

Thus, after division by LL, mean distances in a long body converge to mean absolute differences of the corresponding axial coordinates.

Example 4.2. For L>0L>0 and ε>0\varepsilon>0, define

KL,ε:={(Lt,x,y):0≤t≤1, ∣x∣≤t, ∣y∣≤ε(1−2t3)}.K_{L,\varepsilon}:=\left\{(Lt,x,y):0\leq t\leq1,\ |x|\leq t,\ |y|\leq\varepsilon\left(1-\frac{2t}{3}\right)\right\}.

This is the convex hull of (0,0,±ε)(0,0,\pm\varepsilon) and the four points (L,±1,±ε/3)(L,\pm1,\pm\varepsilon/3).

Theorem 4.3. For L=1000L=1000 and ε=1/100\varepsilon=1/100, one has

Δ(KL,ε)>θ(KL,ε).\Delta(K_{L,\varepsilon})>\theta(K_{L,\varepsilon}).

Proof. The area of the section at level tt equals 4εt(1−2t3)4\varepsilon t\left(1-\frac{2t}{3}\right). Hence the axial coordinate of a uniform interior point has density

pI(t)=185t(1−2t3).p_I(t)=\frac{18}{5}t\left(1-\frac{2t}{3}\right).

and distribution function

FI(t)=9t2−4t35.F_I(t)=\frac{9t^2-4t^3}{5}.

Using D(p)=2∫01F(t)(1−F(t)) dt\mathcal{D}(p)=2\int_0^1F(t)(1-F(t))\,\mathrm{d}t, we obtain

D(pI)=246875=0.281142857….\mathcal{D}(p_I)=\frac{246}{875}=0.281142857\ldots.

Lemma 4.1 therefore gives

Δ(KL,ε)L≥246875.\frac{\Delta(K_{L,\varepsilon})}{L}\geq\frac{246}{875}.

We next compute the axial distribution of normalized surface measure. The two faces x=±tx=\pm t have combined area density

4ε(1−2t3)L2+1,4\varepsilon\left(1-\frac{2t}{3}\right)\sqrt{L^2+1},

while the two faces y=±ε(1−2t/3)y=\pm\varepsilon(1-2t/3) have combined area density

4tL2+4ε29.4t\sqrt{L^2+\frac{4\varepsilon^2}{9}}.

The end face t=1t=1 has area 4ε/34\varepsilon/3. Put

a=εL2+1,b=L2+4ε29−2ε3L2+1,a=\varepsilon\sqrt{L^2+1},\qquad b=\sqrt{L^2+\frac{4\varepsilon^2}{9}}-\frac{2\varepsilon}{3}\sqrt{L^2+1},
D=a+b2+ε3,α=aD,β=b2D.D=a+\frac{b}{2}+\frac{\varepsilon}{3},\qquad\alpha=\frac{a}{D},\qquad\beta=\frac{b}{2D}.

For 0≤t<10\leq t<1, the boundary axial distribution function is

F∂(t)=αt+βt2;F_{\partial}(t)=\alpha t+\beta t^2;

the remaining mass is an atom at t=1t=1. Consequently its Gini mean difference is exactly

gL,ε=α+2β3−2α23−αβ−2β25.g_{L,\varepsilon}=\alpha+\frac{2\beta}{3}-\frac{2\alpha^2}{3}-\alpha\beta-\frac{2\beta^2}{5}.

The diameter of the projection of KL,εK_{L,\varepsilon} onto x,yx,y coordinates is at most 21+ε22\sqrt{1+\varepsilon^2}, so Lemma 4.1 gives

θ(KL,ε)L≤gL,ε+21+ε2L.\frac{\theta(K_{L,\varepsilon})}{L}\leq g_{L,\varepsilon}+\frac{2\sqrt{1+\varepsilon^2}}{L}.

For the values suggested in the statement, g1000,1/100=0.2692731635…g_{1000,1/100}=0.2692731635\ldots and hence

θ(K1000,1/100)1000≤0.2712732636…<0.2811428571…≤Δ(K1000,1/100)1000\frac{\theta(K_{1000,1/100})}{1000}\leq0.2712732636\ldots<0.2811428571\ldots\leq\frac{\Delta(K_{1000,1/100})}{1000}

Remark 4.4. Consider a family of convex bodies

KL={(Lt,z1,…,zd−1):0≤t≤1, ∣zi∣≤ai(t)},K_L=\{(Lt,z_1,\ldots,z_{d-1}):0\leq t\leq1,\ |z_i|\leq a_i(t)\},

where the functions aia_i are nonnegative, concave and Lipschitz, and positive on (0,1)(0,1). The pushforward of volume measure under (Lt,z)↦t(Lt,z)\mapsto t has density

2d−1L∏i=1d−1ai(t).2^{d-1}L\prod_{i=1}^{d-1}a_i(t).

On the other hand, the pushforward of surface measure restricted to the part of ∂KL\partial K_L with 0<t<10<t<1 has density

2d−1∑i=1d−1L2+ai′(t)2∏j≠iaj(t).2^{d-1}\sum_{i=1}^{d-1}\sqrt{L^2+a_i'(t)^2}\prod_{j\ne i}a_j(t).

The two end sections contribute only O(1)O(1) to the surface area. Consequently, as L→∞L\to\infty, the normalized volume and surface measures, projected onto the tt-axis, converge to distributions with densities proportional to

∏i=1d−1ai(t)and∑i=1d−1∏j=1j≠id−1aj(t),\prod_{i=1}^{d-1}a_i(t)\qquad\text{and}\qquad\sum_{i=1}^{d-1}\prod_{\substack{j=1\\j\ne i}}^{d-1}a_j(t),

respectively. In dimension three this is the difference between a product a1a2a_1a_2 and a sum a1+a2a_1+a_2. Choosing one width nondecreasing and one nonincreasing can make the normalized surface density more concentrated along the long axis than the normalized volume density. In particular, choosing a1(t)=ta_1(t)=t and a2(t)=ε(1−2t/3)a_2(t)=\varepsilon(1-2t/3) gives precisely the contrasting axial distributions used in Example 4.2. In dimension two the second expression is the empty product.

Proposition 4.5. There are counterexamples in every dimension d≥3d\geq3.

Proof. The case d=3d=3 was established above. Put KL:=KL,1/100K_L:=K_{L,1/100}. We first show that Δ(KL)−θ(KL)\Delta(K_L)-\theta(K_L) grows linearly with LL. By Lemma 4.1 and the computations in the proof of Theorem 4.3,

lim⁡L→∞Δ(KL)L=246875,\lim_{L\to\infty}\frac{\Delta(K_L)}{L}=\frac{246}{875},

whereas

lim⁡L→∞θ(KL)L=4(19ε2+33ε+9)15(4ε+3)2∣ε=1/100=93319346560.\lim_{L\to\infty}\frac{\theta(K_L)}{L}=\left.\frac{4(19\varepsilon^2+33\varepsilon+9)}{15(4\varepsilon+3)^2}\right|_{\varepsilon=1/100}=\frac{93319}{346560}.

Consequently,

lim⁡L→∞Δ(KL)−θ(KL)L=71992760648000>0.\lim_{L\to\infty}\frac{\Delta(K_L)-\theta(K_L)}{L}=\frac{719927}{60648000}>0.

Thus there exists c>0c > 0 such that

Δ(KL)−θ(KL)≥cL\Delta(K_L) - \theta(K_L) \ge cL

for all sufficiently large LL.

Now let d=3+md = 3 + m, where m≥1m \ge1, and put

CM:=[−M,M]m,K~L,M:=KL×CM.C_M := [-M,M]^m,\qquad\widetilde{K}_{L,M} := K_L \times C_M.

The boundary of K~L,M\widetilde{K}_{L,M} is the union

(∂KL×CM)∪(KL×∂CM).(\partial K_L \times C_M) \cup(K_L \times\partial C_M).

The two sets intersect along ∂KL×∂CM\partial K_L \times\partial C_M, which has zero (d−1)(d-1)-dimensional Hausdorff measure. Hence their surface measures add. The first part has surface measure

H2(∂KL)(2M)m,\mathcal{H}^2(\partial K_L)(2M)^m,

whereas the second has surface measure

∣KL∣Hm−1(∂CM)=2m∣KL∣(2M)m−1.|K_L|\mathcal{H}^{m-1}(\partial C_M) = 2m|K_L|(2M)^{m-1}.

Here by ∣KL∣|K_L| we denote the volume of KLK_L. Elementary computations show

∣KL∣=L90,H2(∂KL)≥2L.|K_L| = \frac{L}{90},\qquad\mathcal{H}^2(\partial K_L) \ge2L.

Observe that the surface area of KL×∂CMK_L \times\partial C_M normalized on the total surface area of K~L,M\widetilde{K}_{L,M} is

ηL,M=2m∣KL∣(2M)m−1H2(∂KL)(2M)m+2m∣KL∣(2M)m−1=m∣KL∣MH2(∂KL)+m∣KL∣≤m180M.\eta_{L,M} = \frac{2m|K_L|(2M)^{m-1}}{\mathcal{H}^2(\partial K_L)(2M)^m + 2m|K_L|(2M)^{m-1}} = \frac{m|K_L|}{M\mathcal{H}^2(\partial K_L) + m|K_L|} \le\frac{m}{180M}.

Let μL,M\mu_{L,M} be normalized surface measure on ∂KL×CM\partial K_L \times C_M. It is the product of normalized surface measure on ∂KL\partial K_L and normalized volume measure on CMC_M. If Z1,Z2Z_1, Z_2 are independent with law μL,M\mu_{L,M}, then, similarly to the proof of Lemma 4.1, by the triangle inequality

E∣Z1−Z2∣≤θ(KL)+2Mm.\mathbb{E}|Z_1-Z_2| \le\theta(K_L) + 2M\sqrt{m}.

For L≥2L \ge2, the diameter of K~L,M\widetilde{K}_{L,M} is

DL,M=L2+1+15625+4mM2.D_{L,M} = \sqrt{L^2 + 1 + \frac{1}{5625} + 4mM^2}.

Replacing normalized surface measure on ∂K~L,M\partial\widetilde{K}_{L,M} by μL,M\mu_{L,M} changes the corresponding mean distance by at most

2ηL,MDL,M≤m90ML2+1+15625+4mM2=O(LM+1).2\eta_{L,M}D_{L,M} \le\frac{m}{90M}\sqrt{L^2 + 1 + \frac{1}{5625} + 4mM^2} = O\left(\frac{L}{M}+1\right).

Indeed, normalized surface measure on ∂K~L,M\partial\widetilde{K}_{L,M} is a mixture with weight 1−ηL,M1-\eta_{L,M} on μL,M\mu_{L,M} and weight ηL,M\eta_{L,M} on the other boundary component. In the corresponding two-point distribution, the contribution from μL,M⊗μL,M\mu_{L,M}\otimes\mu_{L,M} has weight (1−ηL,M)2(1-\eta_{L,M})^2. The remaining contribution has weight 2ηL,M−ηL,M2≤2ηL,M2\eta_{L,M}-\eta_{L,M}^{2}\le2\eta_{L,M}, and all distances lie in [0,DL,M][0,D_{L,M}]. Consequently,

θ(K~L,M)≤θ(KL)+O(M)+O(LM+1).\theta(\widetilde{K}_{L,M})\le\theta(K_L)+O(M)+O\left(\frac{L}{M}+1\right).

Choose M=M(L)M=M(L) so that

M⟶∞,ML⟶0(L⟶∞),M\longrightarrow\infty,\qquad\frac{M}{L}\longrightarrow0\qquad(L\longrightarrow\infty),

say, M=LM=\sqrt{L}, then the error terms above are o(L)o(L). On the other hand, projection onto the first three coordinates does not increase distances, so

Δ(K~L,M)≥Δ(KL).\Delta(\widetilde{K}_{L,M})\ge\Delta(K_L).

Combining all inequalities above, we obtain

Δ(K~L,M)−θ(K~L,M)≥cL−o(L)>0\Delta(\widetilde{K}_{L,M})-\theta(\widetilde{K}_{L,M})\ge cL-o(L)>0

for all sufficiently large LL. Thus K~L,M\widetilde{K}_{L,M} is a counterexample in dimension dd.

Note added

After the first version of our work was posted on arXiv, two related independent works became publicly available. Kukushkin [9] gave a proof of Theorem 2.1 and the planar Zaporozhets–Tarasov conjecture using a quite different approach. Lotnikov [10], whose work is based on his bachelor’s thesis defended in June 2026, constructed centrally symmetric counterexamples in every dimension d≥3d\ge3, for every moment of the distance. These results overlap, respectively, with our Theorem 1.3 and the higher-dimensional counterexamples.

Acknowledgements

The author thanks A. Lotnikov and D. Zaporozhets for useful discussions.

Declaration of AI use

The author used ChatGPT with the GPT-5.6 Sol model during the preparation of this work. The first proof of the planar case of Zaporozhets-Tarasov conjecture, namely Theorem 1.3, was obtained by the author without the use of AI tools; it was based on KKT conditions and an exhaustive case analysis and was substantially more technical. The argument was subsequently simplified through discussions with ChatGPT. The author also used ChatGPT to search for the explicit construction in Example 4.2 from the author’s underlying idea described in Remark 4.4 (that includes generation of relevant code). The author reviewed and verified all AI-assisted output and takes full responsibility for the content of the article.

References

  1. [1]G. Bonnet, A. Gusakova, C. Thäle, and D. Zaporozhets, Sharp inequalities for the mean distance of random points in convex bodies, Adv. Math. 386 (2021), 107813.DOI
  2. [2]A. S. Lotnikov, Mean distance between random points in centrally symmetric convex bodies, Zap. Nauchn. Sem. POMI 544 (2025), 211–235.
  3. [3]A. S. Tokmachev, The Mean Distance Between Random Points on the Boundary of a Convex Figure, J. Math. Sci. 286 (2024), 798–806.DOI
  4. [4]A. S. Tokmachev, Inequalities for convex functions of random points inside and on the boundary of convex bodies, arXiv:2607.08869, 2026.arxiv.org/abs/2607.08869
  5. [5]A. Ya. Khintchine, On unimodal distributions, Izv. Nauchno-Issled. Inst. Mat. Mekh. Tomsk. Gos. Univ. 2 (1938), 1–7. (In Russian.)
  6. [6]S. Yitzhaki, Gini’s mean difference: a superior measure of variability for non-normal distributions, Metron 61 (2003), no. 2, 285–316.
  7. [7]R. Schneider, Convex Bodies: The Brunn–Minkowski Theory, 2nd expanded ed., Cambridge University Press, Cambridge, 2014.
  8. [8]E. Shen, The Zaporozhets–Tarasov Conjecture on Mean Distances, arXiv:2608.06470v3, 2026.arxiv.org/abs/2608.06470
  9. [9]M. Kukushkin, The Zaporozhets–Tarasov inequality for an arbitrary planar convex body, arXiv:2608.14833, 2026.arxiv.org/abs/2608.14833
  10. [10]A. S. Lotnikov, Mean distance between points inside and on the boundary of a convex body, arXiv:2608.14848, 2026.arxiv.org/abs/2608.14848

Paper details

vlnc:W_rmamea92
Version
v2 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Mathematics

GPT-5.6 Sol (bare)

View statement
Looking for related papers…
Contents