\section{The largest inscribed ball near maximum volume}\label{sec:inradius} Theorem~\ref{thm:volume} gives the exact maximum volume in every dimension $d\ge3$. We now ask how large a ball must still fit inside a region whose volume is just below that maximum. In this section we prove the following, in three dimensions. If the volume of the region falls short of the maximum $\omega_3R^3$ by a small fraction, then the region still contains a ball whose radius falls short of $R$ by at most the same fraction, to first order. The coefficient one is sharp. The expanded paper includes this three-dimensional proof and then extends the estimate to higher dimensions, where the sharp coefficient is $1/(d-2)$~\cite[Theorems~2 and~3]{expanded}. The finitely many weighted marks may be replaced by any compactly supported Borel probability measure $\mu$ on $\R^3$. We call $\mu$ a \emph{distribution of marks}. The finite configuration in Theorem~\ref{thm:volume} corresponds to $\mu=\sum_j w_j\delta_{a_j}$, where $\delta_a$ is the unit mass at $a$, and $G_\mu$ below is then the weighted geometric mean $G$ of that theorem. Set \[ G_\mu(x)=\exp\!\int\log|x-a|\,d\mu(a),\qquad E_R=\{x:G_\mu(x)\le R\},\qquad \exp(-\infty)=0 . \] The integral $\int\log|x-a|\,d\mu(a)$ is the decreasing limit, as $M\to\infty$, of the continuous functions $\int\max(\log|x-a|,-M)\,d\mu(a)$. Thus $G_\mu$ is a measurable function and $E_R$ is a measurable set. The set $E_R$ is bounded because $\mu$ has compact support. Write $\rho$ for the \emph{inradius} of $E_R$, the supremum of the radii of balls contained in $E_R$, taking $\rho=0$ if none fits. Write $|A|$ for the volume of a set $A$, and define the fractional volume deficit by $D=1-|E_R|/(\omega_3 R^3)$, where $\omega_3=4\pi/3$. Thus $D$ is the fraction of the maximal volume that is missing, and $1-\rho/R$ is the fraction of the maximal radius that is missing. At threshold one, the argument with compact subsets given below shows that $|E_1|\le\omega_3$. Rescaling does not change $D$. Hence $D\ge0$. \begin{theorem}\label{thm:radius} For each $R>0$, \[ \lim_{h\downarrow0}\ \sup_{\mu:\,00$, every such region with sufficiently small positive deficit satisfies $1-\rho/R\le(1+\varepsilon)D$. In each fixed dimension $d>3$ the same statement holds with the coefficient $1/(d-2)$ in place of one~\cite[Theorem~3]{expanded}. Under the dilation $x\mapsto x/R$ of space and of the marks, every distance is divided by $R$. Thus $G_\mu$ and $\rho$ are divided by $R$, and $D$ stays fixed. We work at threshold one from now on. \subsection*{Two marks approaching one another} We begin with the example that shows why the coefficient one cannot be lowered. Move two equally weighted marks toward each other. Their region approaches the unit ball. We first find the largest ball that fits inside the region, and then compare the fractional loss of radius with the fractional loss of volume. Figure~\ref{fig:ball} shows one position along the way, with the marks $1.2$ units apart. Put mass $1/2$ at each of $\pm t e_1$, where $e_1=(1,0,0)$ and $01-t^2$, then $b^2>t^2$ since $t^2<1/2$. Then the terms $2s^2(b^2-t^2)$ and $s^4$ are nonnegative, so the expression is at least $(b^2+t^2)^2$, which exceeds one because $b^2+t^2>1$. Thus all of $E_1$ lies in the cylinder of radius $\sqrt{1-t^2}$ about the line through the marks, and no ball inside that cylinder, wherever its center lies, has a larger radius. This proves $\rho=\sqrt{1-t^2}$. \begin{figure}[htbp] \centering \includegraphics[width=.90\linewidth]{figures/inscribed_ball_width.pdf} \caption{Two marks $1.2$ units apart ($t=0.6$). The region consists of the points whose distances to the two marks have product at most one. A ball of radius $0.8$ fits inside it (left). The whole region fits inside a cylinder of diameter $1.6$ about the line through the marks (right). Any larger ball would be wider than this cylinder, wherever its center lay, so $0.8$ is the largest possible radius.}\label{fig:ball} \end{figure} \FloatBarrier Jasion and Magnucki~\cite[(2.2)--(2.4), (3.1)--(3.3)]{jasion} computed the volume of this region and its meridian curve, that is, its outline in a plane through both marks (Figure~\ref{fig:ball}, left). The leading term of the volume loss can also be found from the equation of the boundary in spherical coordinates. For a unit vector $\theta$ put $u=\theta\cdot e_1$, and let $r$ be the distance from the origin to the boundary point in the direction $\theta$. Then $(r^2+t^2)^2-4r^2t^2u^2=1$. This is a quadratic equation in $r^2$ with exactly one positive root, namely \[ r^2=t^2(2u^2-1)+\sqrt{1-4t^4u^2(1-u^2)} =1+t^2(2u^2-1)+O(t^4) \] uniformly in $u$. Integrating in spherical coordinates, along each ray from $0$ to $r(\theta)$, shows that the volume fraction $|E_1|/\omega_3$ is the mean of $r(\theta)^3$ over the unit sphere. By the expansion above, $r(\theta)^3=1+\tfrac32t^2(2u^2-1)+O(t^4)$. The mean of $u^2$ over the unit sphere is $1/3$, so the mean of $2u^2-1$ is $-1/3$, and hence $D=t^2/2+O(t^4)$. Also $1-\rho=1-\sqrt{1-t^2}=t^2/2+O(t^4)$. Their ratio therefore tends to one as $t\to0$, which gives the lower bound for the supremum in Theorem~\ref{thm:radius}. \begin{figure}[htbp] \centering \includegraphics[width=.90\linewidth]{figures/deficit_ratio.pdf} \caption{Two equally weighted marks at distance $2t$ from each other, at threshold one. The curve shows the missing fraction of the radius, $1-\rho$, divided by the missing fraction of the volume, $D$. Both are $t^2/2+O(t^4)$, so their ratio tends to one as the marks approach each other (open circle).} \label{fig:ratio} \end{figure} \FloatBarrier \subsection*{A center for a large ball inside the region} We now prove the upper bound in Theorem~\ref{thm:radius}. To find a large ball inside $E_1$ we need a good center. We choose it by minimizing, over all points $z\in\R^3$, the \emph{cost} $\int\log(1+|z-a|^2)\,d\mu(a)$. We write $J$ for the smallest value of the cost: \[ J=\min_{z\in\R^3}\int\log(1+|z-a|^2)\,d\mu(a). \] The cost is a continuous function of $z$ and tends to infinity as $|z|\to\infty$, so it attains its minimum. Let $z$ be a point where the cost attains its minimum. The gradient of the cost vanishes there, so $\int(z-a)/(1+|z-a|^2)\,d\mu(a)=0$. Put $M=\int(1+|z-a|^2)^{-1}d\mu(a)$. Jensen's inequality, applied to the exponential function, gives $M\ge e^{-J}$. Now take any $v$ with $|v|\le1$, and regard the quotient $|z-a+v|^2/(1+|z-a|^2)$ as a function of the mark $a$. Since the cost at $z$ equals $J$, the geometric mean of this quotient with respect to $\mu$ is $e^{-J}G_\mu(z+v)^2$. The inequality between the geometric and the arithmetic mean bounds it by the arithmetic mean of the quotient. The numerator equals $|z-a|^2+2(z-a)\cdot v+|v|^2$. After division by $1+|z-a|^2$, the middle term integrates to zero, because the gradient vanishes at $z$. We obtain \[ e^{-J}G_\mu(z+v)^2 \le\int\frac{|z-a+v|^2}{1+|z-a|^2}\,d\mu(a) =1-(1-|v|^2)M . \] Since $M\ge e^{-J}$, this gives $G_\mu(z+v)^2\le e^J-1+|v|^2$. The right side is at most one when $|v|^2\le2-e^J$. So when $J<\log 2$, the closed ball of radius $\sqrt{2-e^J}$ about $z$ is contained in $E_1$. We now show that nearly maximal volume forces $J$ to be so small that this ball gives $1-\rho\le D+o(D)$, with the same coefficient one in front of $D$ as in the two-mark example. It suffices to prove the following bound for every sequence of compactly supported distributions of marks with positive deficits $D_j\to0$, where $J_j$ denotes the smallest cost of the $j$th distribution: \begin{equation}\label{eq:cost} \limsup_{j\to\infty}\frac{J_j}{\gamma_j}\le2,\qquad \gamma_j=-\log(1-D_j). \end{equation} Indeed, \eqref{eq:cost} forces $J_j\to0$, and the ball above then gives $1-\rho_j\le1-\sqrt{2-e^{J_j}}=J_j/2+O(J_j^2)$, where $\rho_j$ is the inradius for the $j$th distribution. Since also $\gamma_j/D_j\to1$, this gives $\limsup_j(1-\rho_j)/D_j\le1$, the upper bound in Theorem~\ref{thm:radius}. The ball just constructed has its center at $z$, a point where the cost is smallest. Its radius $\sqrt{2-e^J}$ depends only on the smallest cost $J$, so we never need to know where $z$ lies. We only need to show that $J$ is small. Since $J$ is the smallest value of the cost, the cost at any point we choose is at least $J$. If some point has cost $C<\log2$, then $J\le C$, and the ball about $z$ has radius at least $\sqrt{2-e^C}$. We therefore evaluate the cost at a well-chosen \emph{trial point}. The proof works with a sequence of distributions of marks, indexed by $j$, and constructs the trial point in two steps. In the first step we change coordinates. We take a compact subset of the region of nearly full volume, and divide all coordinates by $c_j$, the radius of the ball that has the same Newtonian capacity as this subset. We then choose a point $y_j$ about which the logarithmic moment of the equilibrium measure of the subset is small, and translate $y_j$ to the origin. A point with original position $b$ then has the coordinate $a=b/c_j-y_j$. In these coordinates the equilibrium measures concentrate near the unit sphere. In the second step we choose a point $z_j$ near the new origin about which one quadratic lower bound for the logarithmic moment holds for all marks near the origin. The trial point is the point with coordinate $z_j$, that is, the point with original position $c_j(y_j+z_j)$. \subsection*{Equilibrium measure of a nearly maximal region} Let $\mu_j^0$ be any sequence of such distributions with positive deficits $D_j\to0$, and put $\gamma_j=-\log(1-D_j)$. The $j$th region has volume $\omega_3(1-D_j)=\omega_3e^{-\gamma_j}$, which is larger than $\omega_3e^{-\gamma_j-\gamma_j^2}$. So, by inner regularity of Lebesgue measure, we may choose a compact subset $K_j$ of the $j$th region with $|K_j|\ge\omega_3 e^{-\gamma_j-\gamma_j^2}$. Let $c_j$ be the radius of the ball with the same Newtonian capacity as $K_j$, and let $\nu_j^0$ be the Newtonian equilibrium probability measure of $K_j$. With the kernel $|x-a|^{-1}$, the equilibrium potential is at most $1/c_j$ everywhere~\cite[Appendix A, Lemma A.1]{clark}. Jensen's inequality then gives $\int\log|x-a|\,d\nu_j^0(x)\ge\log c_j$ for every $a$. The supports of $\mu_j^0$ and $\nu_j^0$ are bounded for each $j$, and the negative part of $\log r$ is at most $1/r$. The potential bound therefore allows us to exchange the order of integration by Fubini's theorem. Averaging over the marks gives $\log c_j\le\int\log G_{\mu_j^0}\,d\nu_j^0\le0$, where the last inequality holds because $K_j$ lies in the region. Thus $c_j\le1$. On the other hand, the isocapacitary inequality gives $c_j\ge (|K_j|/\omega_3)^{1/3}$. Together with $c_j\le1$ and the choice of $K_j$, this gives \[ 0\le\beta_j=-\log c_j\le(\gamma_j+\gamma_j^2)/3 . \] The same argument applies to any compact $K\subset E_1$ of positive volume: the ball with the same capacity as $K$ has radius at most one. Thus $|K|\le\omega_3$ by the isocapacitary inequality, and inner regularity gives $|E_1|\le\omega_3$ for any compactly supported distribution of marks. Write $\mu_j,\nu_j$ for the images of $\mu_j^0$ and $\nu_j^0$ under $x\mapsto x/c_j$. In these coordinates $x^*=x/c_j$ and $a^*=a/c_j$ satisfy \[ |x^*-a^*|^{-1}=c_j|x-a|^{-1},\qquad \log|x^*-a^*|=\log|x-a|+\beta_j . \] From now on we drop the stars. Then $\int|x-a|^{-1}\,d\nu_j(x)\le1$, and the function $g_j(a)=\int\log|x-a|\,d\nu_j(x)$ satisfies $g_j\ge0$ and $\int g_j\,d\mu_j\le\beta_j$. Since the average of $g_j$ over the marks is at most $\beta_j$, there is a point $y_j$ with $g_j(y_j)\le\beta_j+\gamma_j^2$. Translate both measures by $-y_j$, and keep the names $\mu_j$, $\nu_j$ and $g_j$. From now on a point with original position $b$ has the coordinate $a=b/c_j-y_j$. The point $y_j$ is the origin, and a mark with coordinate $a$ has original position $c_j(a+y_j)$. The function $\phi(s)=\log s+s^{-1}-1$ is nonnegative, vanishes only at one, and tends to infinity at both ends of $(0,\infty)$. Since $\int\phi(|x-a|)\,d\nu_j(x)=g_j(a)+\int|x-a|^{-1}\,d\nu_j(x)-1$ and the middle integral is at most one, for each $a$ \[ 0\le \int\phi(|x-a|)\,d\nu_j(x)\le g_j(a). \] At $a=0$ the right side is at most $\beta_j+\gamma_j^2$, which tends to zero. Outside any fixed annulus about the unit sphere, $\phi(|x|)$ is bounded below by a positive constant. The mass of $\nu_j$ outside that annulus therefore tends to zero. The sequence $(\nu_j)$ is tight. Each weak limit is supported on the sphere, and, because the Newtonian kernel is lower semicontinuous, its potential is at most one everywhere. Its energy is therefore at most one. Among probability measures supported on the unit sphere, the uniform probability $\sigma$ uniquely minimizes the energy, and its energy is one~\cite[Appendix A]{clark}. Hence $\nu_j$ converges weakly to $\sigma$. \subsection*{Logarithmic moments about a common center} The limiting sphere is the unit sphere about the point $y_j$, which is now the origin. We now construct the point $z_j$. With $z_j$ we want one lower bound for $g_j(a)$, in terms of the distance from $a$ to $z_j$, that holds both for marks $a$ near the origin and for distant marks, so that we can average the bound over all the marks. We first treat marks whose distance from the origin is at least a fixed positive number. Write $H(s)=\int\log|\theta-se_1|\,d\sigma(\theta)$. For every $s>0$, \begin{equation}\label{eq:sphere} H(s)>\tfrac16\log(1+s^2). \end{equation} Indeed, regard $H(|x|)$ as a radial function of $x\in\R^3$. Inside the unit sphere its Laplacian is $\int|\theta-se_1|^{-2}\,d\sigma(\theta)$, where $s=|x|$. By Jensen's inequality this is at least $1/(1+s^2)$, since $\int|\theta-se_1|^2\,d\sigma(\theta)=1+s^2$. The radial Laplacian of the right side of~\eqref{eq:sphere} equals $1/(1+s^2)-(2/3)s^2/(1+s^2)^2$. The two functions and their first derivatives vanish at zero, and the Laplacian of $H$ exceeds the other Laplacian by at least $(2/3)s^2/(1+s^2)^2$. Integrating twice, using $\Delta f=s^{-2}(s^2 f')'$, shows that the difference $H(s)-\frac16\log(1+s^2)$ is positive and strictly increasing for $01$, the identity $H(s)=\log s+H(1/s)$ gives \[ H(s)-\tfrac16\log(1+s^2) =H(1/s)-\tfrac16\log(1+s^{-2})+\tfrac23\log s>0 . \] Let $(a_j)$ be a sequence of points, one for each $j$, that converges to a point $a\ne0$; these $a_j$ are not the marks of Theorem~\ref{thm:volume}. Then $\liminf_j g_j(a_j)\ge H(|a|)$. Weak convergence alone does not give this, because the logarithm is unbounded below near its pole $a_j$. So fix $0<\eta<1$, write $s=|x-a_j|$, and replace $\log s$ by the truncated function $\max(\log s,\log\eta)$. This function is continuous and bounded below, and the measures $\nu_j$ translated by $-a_j$ converge weakly to $\sigma$ translated by $-a$. Hence the lower limit of the truncated integrals is at least the truncated integral for $\sigma$, which is at least $H(|a|)$. The truncation changes the integrand only where $s<\eta$, and there the change is at most $|\log s|\le C/\sqrt{s}\le C\sqrt{\eta}/s$. Since the inverse distance has integral at most one against $\nu_j$, the truncation raises each integral by at most $C\sqrt{\eta}$, uniformly in $j$ and in the point $a_j$. Letting $\eta$ tend to zero proves the claim. Fix $0<\deltaA$ when $A$ is large. Indeed $\nu_j(B_2)\to1$, where $B_2$ is the radius-two ball about zero. For $|a|>3$, \[ g_j(a)\ge\nu_j(B_2)\log(|a|-2)-C\sqrt{\nu_j(B_2^c)} . \] Here the first term comes from $B_2$, where $|x-a|\ge|a|-2$. The second term bounds the contribution of the complement of $B_2$: by the Cauchy--Schwarz inequality and $(\log^-s)^2\le C/s$, where $\log^-s=\max(0,-\log s)$, the negative part of the logarithm contributes at most $C\sqrt{\nu_j(B_2^c)}$. As $|a|\to\infty$, the ratio $\log(|a|-2)/\log(1+|a|^2)$ tends to $1/2$, which is larger than $1/6$, while $\nu_j(B_2)\to1$ and the error term tends to zero. This proves the bound for all $|a|\ge\delta$ once $j$ is sufficiently large. For marks close to zero, we integrate $\phi(|x-a|)$ against $\nu_j$ only over the shell $1-\delta\le|x|\le1+\delta$, and we denote this restricted integral by $F_j(a)$. Since $\phi\ge0$, restricting the integral can only lower it, so $0\le F_j\le g_j$. On this shell the Hessian of the integrand is close to the orthogonal projection onto the radial direction. Averaging these projections over the sphere and applying Taylor's formula will give the coefficient $1/6$. For $|a|\le\delta<1/8$ the integrand has Hessian in $a$ equal to $\theta\theta^{\mathsf T}+O(\delta)$ in operator norm, with $\theta=x/|x|$. Indeed, put $s=|x-a|$ and $u=(x-a)/s$. The Hessian has eigenvalue $\phi''(s)$ in direction $u$ and $\phi'(s)/s$ in the orthogonal directions. On this shell $|s-1|\le2\delta$ and $u=\theta+O(\delta)$. The stated estimate follows from $\phi'(1)=0$ and $\phi''(1)=1$. The integral of $\theta\theta^{\mathsf T}$ against $\nu_j$ over the shell tends to its integral against $\sigma$, because $\nu_j$ converges weakly to $\sigma$ and $\sigma$ gives the boundary of the shell zero mass. Rotational symmetry makes this average a scalar matrix and its trace is one, so the average is $I/3$. The quadratic term of Taylor's formula carries the factor $1/2$, which turns $1/3$ into $1/6$. Consequently, with $\delta$ fixed first and $j$ then tending to infinity, \[ \nabla^2F_j\ge2b_j I\quad\text{on the ball }|a|\le\delta,\qquad b_j=\tfrac16-O(\delta)-o_j(1)>0 . \] Choose $z_j$ to minimize $F_j$ on the closed ball $|a|\le\delta$. Since $z_j$ is a minimizer, the one-sided derivative of $F_j$ at $z_j$ along the segment from $z_j$ to any point $a$ of the ball is nonnegative, even when $z_j$ lies on the boundary of the ball. Taylor's theorem along that segment gives $F_j(a)\ge F_j(z_j)+b_j|a-z_j|^2$, and so, since $g_j\ge F_j$, $g_j(a)\ge F_j(z_j)+b_j|a-z_j|^2$. Since $F_j(z_j)\ge0$, taking $a=0$ gives $b_j|z_j|^2\le g_j(0)\to0$. At fixed small $\delta$, $b_j$ stays bounded below by a positive constant, so $z_j\to0$. For $|a|\le\delta$ we therefore have $g_j(a)\ge b_j|a-z_j|^2\ge b_j\log(1+|a-z_j|^2)$. Outside the ball, put $L(a)=\log(1+|a|^2)$. This function is $1$-Lipschitz, so for $|a|\ge\delta$, \[ L(a-z_j)\le L(a)+|z_j|\le(1+q_j)L(a), \qquad q_j=\frac{|z_j|}{\log(1+\delta^2)} . \] Thus, for $|a|\ge\delta$ and all sufficiently large $j$, the bound $g_j(a)\ge\frac16L(a)$ proved above for $|a|\ge\delta$, which we call the \emph{exterior bound}, gives $g_j(a)\ge L(a-z_j)/[6(1+q_j)]$. For fixed $\delta$, $q_j\to0$. Set \[ k_j=\min\bigl\{b_j,\ 1/[6(1+q_j)]\bigr\}. \] The interior and exterior bounds together give, for every $a$, \[ g_j(a)\ge k_j\log(1+|a-z_j|^2). \] For fixed small $\delta$ the lower limit of $k_j$ as $j\to\infty$ is at least $1/6-C\delta$, because $b_j=1/6-O(\delta)-o_j(1)$ and $q_j\to0$. Let $J_j$ be the smallest cost for the original distribution $\mu_j^0$. We bound it by the cost at the trial point $c_j(y_j+z_j)$. A mark with coordinate $a$ lies at distance $c_j|a-z_j|$ from the trial point in the original coordinates, and $c_j\le1$. Therefore \[ J_j\le\int\log(1+c_j^2|a-z_j|^2)\,d\mu_j(a) \le k_j^{-1}\int g_j\,d\mu_j\le\beta_j/k_j. \] Since $\beta_j\le(\gamma_j+\gamma_j^2)/3$, dividing by $\gamma_j$ and taking the upper limit in $j$ at fixed $\delta$ gives $\limsup_j J_j/\gamma_j\le(1/3)/(1/6-C\delta)$. Letting $\delta$ tend to zero gives~\eqref{eq:cost}. The ball constructed above shows that the inradius $\rho_j$ for $\mu_j^0$ is at least $\sqrt{2-e^{J_j}}$ once $J_j<\log2$. Together with~\eqref{eq:cost} this gives $\limsup_j(1-\rho_j)/D_j\le1$. In the example of two approaching marks the ratio $(1-\rho)/D$ tends to one as $D\to0$, so for every $h>0$ the supremum over $0