Petty-Santalo Inequality

Abstract

Using results of K. Kiener and the Riesz-Sobolev convolution inequality we give a new proof of Petty's projection inequality. By the same method we also obtain a proof of Santalo's affine isoperimetric inequality.

Introduction and notation

We will show that both Petty's projection inequality ([P1]) and Santalo's affine isoperimetric inequality ([S]) can be derived from a rearrangement inequality for the distribution function of the convolution square of a convex body in $\R^d$. The proof is based on fomulas for the volume of the polar of the projection body and the affine surface area ([K], [Sm]) as well as on a consequence of the Riesz-Sobolev convolution inequality. We will also prove a distribution inequality for the convolution square.
To fix notation, let $C$ be a convex body in $\R^d$. If in addition $C$ is symmetric with respect to the origin we simply call $C$ a ball. In this note the letter $B$ always denote a ball and $\Vert.\Vert_B$ the Minkowski functional or gauge function of $B$. By $B_2^d$ we denote the standard $d$-dimensional euclidean ball. $\vol{d}$ (or simply $\vol{}$ if the dimension is clear from the context) will be ordinary Lebesgue measure in $\R^d$.
The convolution square $F_C$ of $C$ is defined as the convolution of the indicator function $I_C$ of $C$ and $I_{-C}$ i.e. $F_C\colon=I_C*I_{-C}$, and the distribution function $V_C(\d)$ of this convolution is defined by: \begin{eqnarray*} V_C(\d)\colon&=&\vol{}(F_C>\d)\colon= \vol{}\left(\left\{x\in\R^d:\,F_C(x)>\d\right\}\right)\\ &=&\vol{}\left(\left\{x\in\R^d:\vol{}(C\cap(C+x))>\d\right\}\right)~. \end{eqnarray*} $V_C$ is a continuous and strictly decreasing function on the interval $[0,\vol{}(C)]$, hence a homeomorphism onto $[0,\vol{}(C-C)]$. We will also use the distribution function $\wt V_C$ of the convolution $I_C*I_C$.
The Riesz Sobolev convolution inequality in its simplest form states that whenever $A_1,A_2$ and $A_3$ are measurable subsets of $\R^d$ and $B_1,B_2,B_3$ are multipels of the euclidean ball, such that for $i=1,2,3$: $\vol{}(A_i)=\vol{}(B_i)$, than $$ I_{A_1}*I_{A_2}*I_{A_3}(0)\leq I_{B_1}*I_{B_2}*I_{B_3}(0)~. $$ In the case $d=1$ his inequality is due to Riesz [R]. There is also a number of generalizations (cf. e.g. [BLL]). One of the applications (
Theorem below) of this inequality, given in [K], will be of essential importance in this paper. In order to state it we have to recall some additional notions: Let $f:\R^d\rar\R_0^+$ be an integrable function with distribution function $V_f(t)\colon=\vol{}(f>t)$, then we denote by $\bar f$ the right-continuous nonincreasing inverse of $V_f$, i.e. $$ \bar f(u)\colon=\inf\{t:\,V_f(t)\leq u\}~. $$ If $f$ and $g$ are two nonnegative integrable functions on $\R^d$ with distribution functions $V_f$ and $V_g$ respectively, then we say that $f$ is dominated by $g$ in the sense of Hardy and Polya if \begin{equation}\label{eq1}\tag{eq1} \forall t>0\ \int_0^t\bar f(u)\,du\leq\int_0^t\bar g(u)\,du \quad\mbox{and}\quad \int_0^\infty\bar f(u)\,du=\int_0^\infty\bar g(u)\,du. \end{equation} In this case we write $g\succ f$. In what is going to follow we simply write $V_2\succ V_1$ for two distribution functions $V_1$ and $V_2$, meaning that $g\succ f$ whenever $V_1=V_f$ and $V_2=V_g$.

Petty's Inequality

For all $x\in S^{d-1}$ the Minkowski functional of the polar of the projection body $\Pi^*C$ of a convex body $C$ is defined by $$ \Vert x\Vert_{\Pi^*C}\colon=\vol{d-1}\left(\Prn_x(C)\right) $$ where $\Prn_x$ denotes the orthogonal projection onto the subspace orthogonal to $x$. In this section we will present a new proof of Petty's projection inequality, which states that for all convex bodies $C$ in $\R^d$ such that $\vol{}(C)=\vol{}(B_2^d)$, we have: $\vol{}(\Pi^*C)\leq\vol{}(\Pi^*B_2^d)$. The following two theorems are due to K. Kiener (cf. [K]). The proof of the first one uses the Riesz Sobolev convolution inequality
Suppose $A_1$ and $A_2$ be measurable subsets of $\R^d$ with finite volume. Let $B_1$ and $B_2$ be multiples of the euclidean ball $B_2^d$ such that $\vol{}(A_1)=\vol{}(B_1)$ and $\vol{}(A_2)=\vol{}(B_2)$. Then $I_{A_1}*I_{A_2}$ is dominated by $I_{B_1}*I_{B_2}$ in the sense of Hardy and Polya.
This theorem will be generalized in the Appendix!
Let $C$ be a convex body in $\R^d$, then $$ \lim_{\d\uar\vol{}(C)}\frac{V_C(\d)}{(\vol{}(C)-\d)^d}=\vol{}(\Pi^*C)~. $$
We will also need the following elementary
Let $V_1:[0,a]\rar[0,b_1]$ and $V_2:[0,a]\rar[0,b_2]$ be homeomorphisms such that $V_1(a)=V_2(a)=0$. If $V_2\succ V_1$ then there exists a sequence $a_n\uar a$ such that for all $n\in\N$: $V_1(a_n)\leq V_2(a_n)$.
$\proof$ Draw a picture! Suppose not, than there exists an $\e>0$ such that for all $t > a-\e$: $V_1(t) > V_2(t)$; hence $$ \int_0^{V_2(a-\e)}V_1^{-1}(s)\,ds>\int_0^{V_2(a-\e)}V_2^{-1}(s)\,ds $$ which contradicts the assumtion that $V_2\succ V_1$. $\eofproof$
Using
Theorem, Theorem and Lemma we can now give a simple proof of Petty's projection inequality.
For all convex bodies $C$ in $\R^d$ satisfying $\vol{}(C)=\vol d(B_2^d)$ we have: $$ \vol{}\left(\Pi^*C\right)\leq\vol{}\left(\Pi^*B_2^d\right)~. $$
$\proof$ Put $a=\vol{}(C)$, $b_1=\vol{}(C-C)$ and $b_2=\vol{}(2C)$, then both $V_C:[0,a]\rar[0,b_1]$ and $V_{B_2^d}:[0,a]\rar[0,b_2]$ are homeomorphisms. Since $V_C(a)=V_{B_2^d}(a)=0$ it follows from Theorem and Lemma, that there exists a sequence $a_n\uar a$ such that for all $n\in\N$: $V_C(a_n)\leq V_{B_2^d}(a_n)$. Now the inequality is an immediate consequence of Theorem. $\eofproof$
Remark for the careful reader of [K]: This proof is somehow the reversed order of arguments given there.

Santalo's Inequality

Let $C$ be a convex body in $\R^d$ of class $C^2$ and denote by $k$ the Gauß-Kronecker curvature of $\pa C$. Then the affine surface area of $C$ is defined by: $$ \O(C)\colon=\int_{\pa C}k^{\frac1{d+1}}(x)\,\mu(dx) $$ where $\mu$ is the area measure of $\pa C$. $\O$ is called affine surface area, because it is invariant under volume-preserving affine transformations. Using the notion of a curvature function (cf. [Sn2] p.419) the affine surface area can also be defined for convex bodies wich have a curvature function.
By means of Steiner symmetrizations Santalo [S] proved that for all convex bodies of class $C^2$ satisfying $\vol{}(C)=\vol{}(B_2^d)$, the following inequality holds: $\O(C)\leq\O(B_2^d)$ with equality iff $C$ is an ellipsoid. This is called Santalo's affine isoperimetric inequality. Later on Petty (cf.[P2]) established this inequality, as well as the equality condition, for convex bodies which have a positive continuous curvature function. K. Leichtweiß [Le], E. Lutwak [Lu], C. Schütt and E. Werner [SW] extended the notion of an affine surface area to arbitrary convex bodies. In particular Lutwak's paper contains an extension of the affine isoperimetric inequality with the same equality conditions. For more information on the notion of the affine surface area we refer also to [Sn1], [Sn2].
In order to give another proof of Santalo's isoperimetric inequality in the general case, we use the following
Theorem, which was proved in [Sm] (there it is stated only for symmetric bodies, but, as was observed by S. Reisner, the proof given in [Sm] also works in the non symmetric case):
Let $C$ be a convex body in $\R^d$, then $$ \lim_{\d\to0}\frac{\wt V_C(0)-\wt V_C(\d)}{\d^{\frac2{d+1}}} =c(d)\O(C), $$ where $c(d)$ is a constant depending on the dimension $d$ only.
The proof of the following lemma is similar to that of Lemma
Let $V_1:[0,a_1]\rar[0,b]$ and $V_2:[0,a_2]\rar[0,b]$ be homeomorphisms such that $V_1(0)=V_2(0)=b$. If $V_2\succ V_1$ then there exists a sequence $\d_n\dar0$ such that for all $n\in\N$: $V_1(\d_n)\geq V_2(\d_n)$.
Suppose $C$ be a convex body in $\R^d$ such that $\vol{}(C)=\vol d(B_2^d)$. Then $$ \O(C)\leq \O(B_2^d)~. $$
$\proof$ Put $a_1=\max\{I_C*I_C(x)\}$, $a_2=\vol{}(C)$ and $b=\vol{}(2C)$, then again both $\wt V_C:[0,a_1]\rar[0,b]$ and $V_{B_2^d}:[0,a_2]\rar[0,b]$ are homeomorphisms and $\wt V_C(0)=V_{B_2^d}(0)=b$, hence Lemma implies the existence of a sequence $\d_n\dar0$ such that for all $n\in\N$: $\wt V_C(\d_n)\geq V_{B_2^d}(\d_n)$. Now we conclude by Theorem. $\eofproof$
In general symmetrization techniques are used to prove that the value of a certain expression decreases or increases when applying Steiner symmetrizations. In the above context this means the following: Let $C$ be a convex body in $\R^d$ and denote by $C_1$ the convex body obtained by symmetrizing $C$ about a hyperplane $H$. Let $F_C$ and $F_{C_1}$ be the convolution squares of $C$ and $C_1$ respectively. Then the following conjecture seems to be natural: $F_{C_1}\succ F_C$. An affirmative answer to this conjecture (cf. the Appendix) would imply that $$ \vol{}(\Pi^*C_1)\geq\vol{}(\Pi^*C) \quad\mbox{and}\quad \O(C_1)\geq\O(C)~. $$

An Inequality for Distributions

In this section we turn to the problem of estimating the distribution function of the convolution square of a ball $B$ in $\R^d$. This will give us an upper gaussian bound for the convolution square itself (cf. Corollary). We start by estimating the convolution square of the euclidean ball. For the convenience of the reader we include the proof of the following
Let $B_2^d(\e)$ be the set $\{x\in B_2^d:\,x_1>\e\}$. Then for some absolute constant $c$: $$ \tfrac c{\sqrt d}(1-\e^2)^{\frac{d+1}2}\leq \frac{\vol{}\left(B_2^d(\e)\right)}{\vol{}\left(B_2^d\right)}\leq \tfrac12(1-\e^2)^{\frac{d+1}2}\leq e^{-\frac12(d+1)\e^2}~. $$
$\proof$ Putting $1-t^2=(1-s^2)(1-\e^2)$ we get for all $0<\e<1$: \begin{eqnarray*} \vol{}\left(B_2^d(\e)\right)&=& \vol{d-1}\left(B_2^{d-1}\right)\int_\e^1(1-t^2)^{\frac{d-1}2}\,dt\\ &=&\vol{d-1}\left(B_2^{d-1}\right)(1-\e^2)^{\frac{d+1}2} \int_0^1(1-s^2)^{\frac{d-1}2}\tfrac st\,ds~. \end{eqnarray*} Since $1\geq s/t\geq s$ the result follows. $\eofproof$
Let $B$ be a ball in $\R^d$. Then for all $\d>0$ we have: $$ V_B(\d)^{\frac1d}\leq\tfrac83\sqrt{\tfrac2d \log\left(\tfrac{2\vol{}(B)}{\d}\right)}\,\vol{}(B)^{\frac1d}~. $$
$\proof$ Let $D$ be a multiple of the euclidean ball such that $\vol{}(D)=\vol{}(B)$. Since by Lemma \begin{equation}\label{eq2}\tag{eq2} \vol{}(D\cap(x+D)) =2\vol{}\left(D\cap\left[x=\tfrac12\Vert x\Vert_2^2\right]\right) \leq\vol{}(D)\exp\left(-\tfrac18d\Vert x\Vert_D^2\right) \end{equation} we get $$ V_D(\d)\leq\vol{}(D)\left( \tfrac8d\log\left(\tfrac{\vol{}(D)}{\d}\right)\right)^{\frac d2}~. $$ Hence $$ V_D^{-1}(\a)\leq\vol{}(D)\exp\left(-\tfrac18d\left( \tfrac{\a}{\vol{}(D)}\right)^{\frac2d}\right)~. $$ For all $\l\leq1$ we have by \eqref{eq1} and \eqref{eq2} \begin{eqnarray*} V_B^{-1}\left(\vol{}(D)t^d\right)&\leq& \frac1{\vol{}(D)t^d}\int_0^{\vol{}(D)t^d}V_D^{-1}(\a)\,d\a\\ &\leq& \frac{d\vol{}(D)}{t^d}\int_0^t s^{d-1}e^{-\frac18ds^2}\,ds\\ &\leq&\frac{d\vol{}(D)}{t^d}\left(\frac{\l^dt^d}d+ \frac{t^d(1-\l^d)}d\,e^{-\frac18d\l^2t^2}\right)\\ &\leq&\vol{}(D)\left(\l^d+e^{-\frac18d\l^2t^2}\right) \end{eqnarray*} Now we choose $\l$ in such a way that $\l^d\leq e^{-\frac12d\l^2}$ and put $c=8\l^{-2}$, then for all $t\leq2$ (which is the possible range of $t$): $$ V_B^{-1}\left(\vol{}(D)t^d\right) \leq2\vol{}(D)\exp\left(-\tfrac dct^2\right)~. $$ This implies: $$ V_B(\d)^{\frac1d}\leq\vol{}(D)^{\frac1d}\sqrt{\tfrac cd \log\left(\tfrac{2\vol{}(D)}{\d}\right)}~. $$ For $\l=\frac34$ we have $\l^2\leq-2\log\l$ and $c=\frac{128}9$. $\eofproof$
Remark: In the nonsymmetric case we have to choose $\l$ in such a way, that $\l^2\leq8t_0^{-2}\log\l$ where $$ t_0=\left(\frac{\vol{}(C-C)}{\vol{}(C)}\right)^{\frac1d}~. $$ By a theorem of Rogers and Shephard (cf. [RS]) this is bounded by $8$. Hence the upper bound of Proposition holds in the non symmetric case with a worse constant. We now turn to estimates for the convolution square itself. The following simple lemma is in constant use in the local theory of normed spaces (cf. e.g. [LM]).
Let $\s$ be the normalized Haar measure on the sphere $S^{d-1}$ and let $B$ be a ball such that for some $t<1$: $\vol{}(B)^{\frac1d}\leq t\vol{}(B_2^d)^{\frac1d}$. Then there exists a subset $A$ of $S^{d-1}$ such that $\s(A)\geq1-t^d$ and for all $x\in A$: $\Vert x\Vert_B\geq1$.
$\proof$ The formula $$ \frac{\vol{}(B)}{\vol{}(B_2^d)} =\int_{S^{d-1}}\Vert x\Vert_B^{-d}\,\s(dx) $$ and Chebyshev's inequality imply: $\s(\Vert.\Vert_B < 1) < t^d$. $\eofproof$
Before stating the next result we recall that the John ellipsoid of a convex body $C$ is the ellipsoid of maximal volume contained in $C$. If ${\cal E}_C$ denotes the John ellipsoid of a $d$ dimensional convex body $C$, then a famous result of F. John [J] states that $C\sbe d{\cal E}_C$. If in addition $C$ is symmetric than $C\sbe\sqrt{d}{\cal E}_C$.
There exists an absolute constant $c$ such that for any ball $B$ in $\R^d$ whose John ellipsoid is a multiple of the euclidean ball, we can find a subset $A$ of $S^{d-1}$ such that $\s(A)\geq1-d2^{-d}$ and for all $x\in A$ and all $r>0$: $$ F_B(rx)\leq2\vol d(B)\exp\left(-\frac1c\left(\frac r{\vol d(B)^{\frac1d}}\right)^2\right)~. $$
$\proof$ W.l.o.g. we may assume $\vol{}(B)=1$. Choose $c_d>0$ such that $\vol{}(c_d\sqrt{d}B_2^d)=1$; clearly: $c_d\leq1/2$. The assumption on the John ellipsoid of $B$ now implies: $B\sbe\frac12d B_2^d$.
For given $\d>0$ we choose $r(\d)>0$ such that \begin{eqnarray}\label{eq3} \vol{}(\tfrac12r(\d)B_2^d)^{\frac1d}&=& \tfrac83\sqrt{\tfrac2d\log\left(\tfrac{2}{\d}\right)}\nonumber\\ \mbox{i.e.}\quad r(\d)&=&\tfrac{16c_d}3\sqrt{2\log\left(\tfrac2{\d}\right)}~. \end{eqnarray} Let us denote by $B(\d)$ the set $[F_B>\d]$, then from
Proposition we conclude: $$ \tfrac12\vol{}(r(\d)B_2^d)^{\frac1d}\geq\vol{}(B(\d))^{\frac1d} $$ and from Lemma it follows that the set $$ A_\d\colon=\{x\in S^{d-1}:\,\Vert x\Vert_{B(\d)}\geq r(\d)^{-1}\} $$ is large i.e. $\s(A_\d)\geq1-2^{-d}$. For $n=1,\ldots,d$ define $\d_n$ such that $r(\d_n)=n$ and $A\colon=\bigcap_nA_{\d_n}$. Then $\s(A)\geq1-d2^{-d}$ and for all $x\in A$ and all $n=1,\ldots,d$ we have: $$ F_B(nx)\leq\d_n= 2\exp\left(-\tfrac12\left(\tfrac{3n}{16c_d}\right)^2\right)~. $$ Since for all $x\in S^{d-1}$ the function $t\mapsto F_B(tx)$ is decreasing and vanishes for $t>d$, we get the assertion of the corollary. $\eofproof$
In order to obtain an affinely invariant version of this corollary, we need to recall the classical notion of the volume ratio: Let $C$ be convex body in $\R^d$ and let ${\cal E}_C$ be the John ellipsoid of $C$, then the quantity $vr(C)\colon=(\vol{}(C)/\vol{}({\cal E}_C))^{\frac1d}$ is called the volume ratio of $C$.
Let $B$ be a ball in $\R^d$, ${\cal E}$ the John ellipsoid of $B$ and $\s$ the normalized Haar measure on the boundary $\pa{\cal E}$ of ${\cal E}$. Then there exists a subset $A$ of $\pa{\cal E}$ with $\s(A)\geq1-d2^{-d}$ such that and for all $x\in\R^d$ satisfying $x/\Vert x\Vert_{\cal E}\in A$: $$ F_B(x)\leq2\vol d(B)\exp\left(-\tfrac dc\left(\tfrac{\Vert x\Vert_{\cal E}} {vr(B)}\right)^2\right)~. $$
Remark: There is no chance to prove the above result for all directions $x$: take the unit ball $B_1^d$ of $\ell_1^d$ and choose for $x$ one of the basis vectors. Another even more symmetric counterexample is the set $$ \left\{(y,t)\in\R^{d-1}\times\R:\norm y_2+\tfrac{|t|}{\sqrt{d-1}}\leq1\right\} $$ and the direction $x=(0,\ldots,0,1)$.
Corollary has to be seen in connection with the following conjecture of V.D. Milman:
There exists an absolute constant $c$ such that for any ball $B$ in $\R^d$ with $\vol{}(B)=1$, we can find a vector $x\in S^{d-1}$ such that for all $r>0$: $$ \vol{}(B\cap[x>r])\leq\exp(-cr^2)~. $$
An affirmative solution of this 'deviation inequality', would solve the hyperplane problem (cf. e.g. [MP]): For all convex symmetric bodies $B$ in $\R^d$ there exists a hyperplane $H$ such that $\vol{d-1}(H\cap B)\geq c\vol{}(B)^{\frac{d-1}d}$, where $c$ denotes an absolute constant. A lower bound for $V_B$ can be obtained in a similar way: Lemma implies: $$ V_B^{-1}(\a)\geq\tfrac c{\sqrt d}\vol{}(D) \left(1-\left(\tfrac{\a}{2^d\vol{}(D)}\right)^{\frac 2d} \right)^{\frac{d+1}2} $$ where, as before, $D$ denotes a multiple of the euclidean ball with the same volume as $B$. Now we conclude again by \eqref{eq1}: \begin{eqnarray*} V_B^{-1}\left(\vol{}(D)t^d\right)&\geq& \tfrac1{(2^d-t^d)\vol{}(D)} \int_{t^d\vol{}(D)}^{2^d\vol{}(D)}V_d^{-1}(\a)\,d\a\\ &\geq&\tfrac c{\sqrt d}\vol{}(B) \int_{\frac{t^2}4}^1(1-s)^{\frac{d+1}2}\tfrac d2\tfrac{s^{\frac d2-1}} {1-(t^2/4)^{\frac d2}}\,ds\\ &=&\tfrac c{\sqrt d}\vol{}(B)H(t^2/4) \end{eqnarray*} where $H$ denotes the function $$ H(x)=\tfrac d2\b\left(1-x,\tfrac{d+3}2,\tfrac d2\right)\left(1-x^{\frac d2} \right)^{-1} $$ and $\b(x,p,q)$ the incomplete beta function. Taking the inverse we abtain: $$ V_B(\d)^{\frac1d}\geq 2\vol{}(B)^{\frac1d} \sqrt{H^{-1}\left(\tfrac{\d\sqrt{d}}{\vol{}(B)}\right)}~. $$ If $x$ is close to $0$ then $H^{-1}(x)\sim1-x^{\frac2{d+1}}$, actually: $$ \lim_{x\dar0}\tfrac{1-H^{-1}(x)}{x^{\frac2{d+1}}}=1~. $$ Finally, let $B$ be a ball in $\R^d$. For $0\leq\d\leq\frac12\vol{}(B)$ let $B_\d$ be the floating body of $B$ i.e. $B_\d$ is a subset of $B$ such that every tangent hyperplane of $B_\d$ cuts off a cap of volume $\d$ from $B$. K. Ball [B] and independently M. Meyer and S. Reisner [MR] proved that $B_\d$ is a convex symmetric body. This is not true in the general case. C. Schütt and E. Werner [SW] introduced the notion of the convex floating body, which coincides with the floating body provided that the latter is convex. Since $$ B_\d\spe\tfrac12[I_B*I_B>2\d] \quad\mbox{i.e.}\quad \vol{}(B_\d)\geq2^{-d}V_B(2\d) $$ a lower bound for $V_B$ implies a lower bound for $\vol{}(B_\d)$. We conclude this paper with two additional results on the distribution function. The first one being a characterization of ellipsoids
Let $B$ be a ball in $\R^d$ $(d\geq3)$ and let $B_\d$ denote the floating bodies. If for all $0<\d<\frac12\vol{}(B)$: $\vol d(B_\d)=2^{-d}V_B(2\d)$ then $B$ is an ellipsoid.
$\proof$ If $H$ is a hyperplane supporting $B_\d$ at $x_0$, then $$ H^+\cap B\cap(B+2x_0)=H^+\cap B $$ If this were false, we would have \begin{eqnarray*} \d=\vol{}(H^+\cap B)&=&\vol{}(H^+\cap B\cap(B+2x_0))\\ &&+\vol{}(H^+\cap B\cap(B+2x_0)^c)\\ &>&\tfrac12(2\d)=\d~. \end{eqnarray*} Hence for every hyperplane $H$ supporting $B_\d$ at $x_0$ $$ H\cap B\cap(B+2x_0)=H\cap B $$ i.e. $x_0$ is a center of symmetry for $H\cap B$. Since every hyperplane intersecting $B$ is supporting some $B_\d$ we conclude, that $H\cap B$ has a center of symmetry for all hyperplanes $H$. Hence, by a theorem of Brunn (cf. [A] p 123), $B$ is an ellipsoid. $\eofproof$
Remark: Using $\wt V_C$ instead of $V_C$ it is possible to get a characterization of ellipsoids as the only convex bodies $C$ satisfying $\vol{}(C_\d)=2^{-d}\wt{V}_C(2\d)$, where $C_\d$ denotes the convex floating body.
Let $V_B$ be as above, then $V_B(\d)$ is a convex function of $\d$.
$\proof$ Let $B(\d)$ be the set $[F_B>\d]$. For $h>0$ we have \begin{eqnarray*} \tfrac1h(V_B(\d+h)-V_B(\d))&=&-\tfrac1{dh}\int_{S^{d-1}} \Vert x\Vert^{-d}_{B(\d)} -\Vert x\Vert^{-d}_{B(\d+h)}\,dx\\ &=&-\tfrac1{dh}\int_{S^{d-1}} \Vert x\Vert^{-d}_{B(\d+h)}\sum_{k=1}^d\mbox{${d\choose k}$}\left(\D r \Vert x\Vert_{B(\d+h)}\right)^k\,dx \end{eqnarray*} where $\D r=\Vert x\Vert^{-1}_{B(\d)}-\Vert x\Vert^{-1}_{B(\d+h)}$. From [Sm] Lemma 4 it follows that $$ \tfrac1{\s_\d(x)}\leq\tfrac{\D r}h\leq\tfrac{1+\frac{\d}{h}}{\s_\d(x)} \log\left(1+\tfrac h\d\right)~. $$ where $\s_\d(x)$ denotes the $d-1$ dimensional volume of the orthogonal projection of the set $$ B\cap\left(B+\tfrac x{\Vert x\Vert_{B(\d)}}\right) $$ on the hyperplane orthogonal to $x$. Hence $$ \lim_{h\to0}\tfrac1h(V(\d+h)-V(\d)) =-\int_{S^{d-1}}\Vert x\Vert^{-n}_{B(\d)}\,\tfrac{ \Vert x\Vert_{B(\d)}}{\s_\d(x)}\,dx~. $$ Since $\s_\d(x)$ is a decreasing function of $\d$ the result follows. $\eofproof$

Appendix

The following affirmative answer to the problem stated above is due to A. Pruss, [P].
Suppose $A$ and $B$ be measurable subsets of $\R^d$ with finite volume. Let $A^\prime$ and $B^\prime$ be any Steiner symmetrization of $A$ and $B$ respectively. Then $I_{A}*I_{B}$ is dominated by $I_{A^\prime}*I_{B^\prime}$ in the sense of Hardy and Polya.
$\proof$ Let $U$ be any measurable subset of $\R$ of volume $2a$ and put $V=-B(x,\d)\times-U$. We then have by the Riesz-Sobolev inequality $$ F*I_V(0)=I_A*I_B*I_V(0)\leq I_{A^\prime}*I_{B^\prime}*I_{V^\prime} =G*I_{V^\prime}~. $$ Now, since $$ F*I_V(0)=\int F(y)I_{-V}(y)\,dy= \int_{B(x,\d)}\int_U F(z,t)\,dt\,dz $$ the above inequality implies: $$ \frac1{\vol{}(B(x,\d))}\int_{B(x,\d)}\int_U F(z,t)\,dt\,dz\leq \frac1{\vol{}(B(x,\d))}\int_{B(x,\d)}\int_{-a}^a G(z,t)\,dt\,dz $$ By taking the limit as $\d\rar 0$ we get: $$ \int_U F(x,t)\,dt\leq\int_{-a}^a G(x,t)\,dt $$ i.e. for all $x\in\R^{d-1}$ the function $F(x,.)$ is dominated by $G(x,.)$ in the sense of Hardy and Polya. Therefore for any increasing convex function $\vp:\R_0^+\rar\R_0^+$: $$ \int_\R\vp\circ F(x,t)\,dt\leq\int_\R\vp\circ G(x,t)\,dt $$ Integrating with respect to $x\in\R^{d-1}$ we obtain: $$ \int_{\R^d}\vp\circ F(y)\,dy\leq\int_{\R^d}\vp\circ G(y)\,dy $$ which is equivalent to Hardy Polya domination of $F$ by $G$. $\eofproof$

Bibliography

  1. [A] D. Amir. Charaterizations of Inner Product Spaces.
  2. [B]K. Ball, unpublished.
  3. [BLL] H.J. Brascamp, E.H. Lieb and J.M. Luttinger, A General Rearrangement Inequality for Multiple Integrals, J. Func. Anal. 17, 227-237, 1974.
  4. [J] F. John, Extremum Problems with Inequalities as Subsidiary Conditions, Courant Anniversary Volume, Interscience New York, 187-204, 1948.
  5. [K] K. Kiener, Extremalität von Ellipsoiden und die Faltungsungleichung von Sobolev, Arch. Math 46. 162-168, 1986.
  6. [L] K. Leichtwei\ss , Affinoberfläche Konvexer Körper, Manus. Math. 56, 429-464, 1986.
  7. [LM] J. Lindenstrauss and V.D. Milman, The Local Theory of Normed Spaces and its Applications to Convexity, IHES, 1991.
  8. [Lu] E. Lutwak, Extended Affine Surface Area, Adv. Math, 1991.
  9. [MO] A.W. Marshall and I. Olkin, Inequalities: Theory of Majorization and its Applications, 1979.
  10. [MP] V.D. Milman and A. Pajor, Isotropic Position and Inertia Ellipsoids and Zonoids of the Unit Ball of a Normed n-dimensional Space, GAFA 87/88, Springer LNM 1376.
  11. [MR] M. Meyer and S. Reisner, A Geometric Property of the Boundary of Symmetric Convex Bodies and Convexity of Floatation Surfaces, Geom. Ded. 37, 327-337, 1991.
  12. [P1] C.M. Petty, Isoperimetric Problems, Proc. Conf. Convexity and Combinatorial Geometry, Univ. of Oklahoma, 26-41, 1971.
  13. [P2] C.M. Petty, Affine Isoperimetric Problems in: Discrete Geometry and Convexity (eds J.E. Goodman, E. Lutwak, J. Malkevitch and R. Pollack), Ann. New York Acad. Sci. 440, 113-127, 1985.
  14. [P] A. Pruss, personal communication.
  15. [R] F. Riesz, Sur une Inegalite Integrale, J. L.M.S 5, 162-168, 1930.
  16. [RS] C.A. Rogers and G.C. Shephard, The Difference Body of a Convex Body, Arch. Math. 8, 220-233, 1957.
  17. [Sa] L.A. Santalo, Un invariante afin para los cuerpos convexos del espacio de $n$ dimensiones, Portugal Math. 8, 155-161, 1949.
  18. [Sm] M. Schmuckenschläger, The Distribution Function of the Convolution Square of a Convex Symmetric Body in $\R^n$, Isr. J. Math. 78, 309-334, 1992.
  19. [Sn1] R. Schneider, Boundary Structure and Curvature of Convex Bodies, Proc. Siegen 1978 ed. by J. Tölke and J. Wills, 13-59.
  20. [Sn2] R. Schneider, Convex Bodies: the Brunn-Minkowski Theory, Cambridge University Press, 1993.
  21. [SW] C. Schütt and E. Werner, The Convex Floating Body, Math. Scand. 66, 275-290, 1990.

<Home>
Last modified: Wed Nov 8 08:51:49 CET 2023