\section*{How the inradius estimate depends on dimension} The stability estimate of Theorem~\ref{thm:radius} holds in every dimension $d\ge3$, with the coefficient $1/(d-2)$ in place of one. The example of two marks already shows why the coefficient changes: the inradius of the region of two marks does not depend on the dimension, but the fraction of the volume that is missing does. Place equal masses at $\pm t e_1$ with $00$. For a compactly supported probability measure $\mu$ on $\R^d$, define $G_\mu,E_R,\rho$ as in Theorem~\ref{thm:radius}, and put $D=1-|E_R|/(\omega_d R^d)$. Then \[ \lim_{h\downarrow0}\ \sup_{\mu:\,00$, there is a point $z_j$ such that for every $a$ \begin{equation}\label{eq:common-d} g_j(a)\ge\bigl(\alpha-O_d(\delta)-o_j(1)\bigr) \log(1+|a-z_j|^2). \end{equation} Here $o_j(1)\to0$ with $\delta$ held fixed. For $|a|\ge\delta$ we need the bound for all such $a$ at once, however far away they are. For $|a|\le\delta$ a second-derivative estimate will determine the center $z_j$. First let $H(s)=\int\log|\theta-se_1|\,d\sigma(\theta)$, with $e_1$ the first coordinate unit vector. For $01$ we use the identity $H(s)=\log s+H(1/s)$: the difference $H(s)-\alpha\log(1+s^2)$ equals the same difference at $1/s$ plus $(1-2\alpha)\log s$, and both terms are positive. Thus $H(s)>\alpha\log(1+s^2)$ for every $s>0$. If $a_j\to a\ne0$, then $\liminf g_j(a_j)\ge H(|a|)$. As in dimension three, this follows by truncating the logarithm and taking weak limits. Indeed for $0\alpha\log(1+s^2)$ and the annulus is compact, the argument by contradiction used in dimension three gives $g_j(a)\ge\alpha\log(1+|a|^2)$ on any fixed annulus $\delta\le|a|\le A$ for large $j$. Now let $B_2$ be the ball of radius two about zero. If $|a|>A>3$, then $|x-a|\ge|a|-2$ for $x$ in $B_2$, and splitting the integral over $B_2$ and its complement gives \[ g_j(a)\ge\nu_j(B_2)\log(|a|-2)-C_k\sqrt{\nu_j(B_2^c)} . \] The last term comes from the Cauchy--Schwarz inequality and $(\log^-s)^2\le C_k s^{-k}$, as in dimension three. Since $\nu_j(B_2)\to1$ and $\log(s-2)/\log(1+s^2)\to1/2>\alpha$, one can first choose $A$ and then $j$ large enough that $g_j(a)\ge\alpha\log(1+|a|^2)$ for every $|a|>A$. Next restrict $\int\phi_k(|x-a|)\,d\nu_j(x)$ to $1-\delta\le |x|\le1+\delta$ and call the result $F_j(a)$. For $|a|\le\delta<1/8$, its integrand has Hessian $k\theta\theta^{\mathsf T}+O_d(\delta)$, where $\theta=x/|x|$. Indeed the Hessian at distance $s=|x-a|$ has radial eigenvalue $\phi''_k(s)$ and tangential eigenvalues $\phi'_k(s)/s$. The projection onto the radial line differs from $\theta\theta^{\mathsf T}$ by $O(\delta)$ in operator norm, and $\phi'_k(1)=0$, $\phi''_k(1)=k$. The annular integral of $\theta\theta^{\mathsf T}$ tends to $I/d$ by rotational symmetry of $\sigma$ and weak convergence. Thus $\nabla^2 F_j\ge2b_j I$ with $b_j=\alpha-O_d(\delta)-o_j(1)>0$ for fixed small $\delta$ and large $j$. Let $z_j$ be a point where $F_j$ is smallest 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 to any point $a$ of the ball is nonnegative, and Taylor's theorem along that segment gives $g_j(a)\ge F_j(z_j)+b_j|a-z_j|^2$ throughout the ball. Taking $a=0$ gives $b_j|z_j|^2\le g_j(0)\to0$, so $|z_j|\to0$. Since $s^2\ge\log(1+s^2)$, this proves~\eqref{eq:common-d} for $|a|\le\delta$. For $|a|\ge\delta$, the function $L(a)=\log(1+|a|^2)$ is $1$-Lipschitz and $L(a)\ge\log(1+\delta^2)$, so $L(a-z_j)\le L(a)+|z_j|\le[1+|z_j|/\log(1+\delta^2)]\,L(a)$. Together with the bound $g_j(a)\ge\alpha L(a)$ for $|a|\ge\delta$, this proves~\eqref{eq:common-d} there as well. Write $b$ for the positive coefficient on the right of~\eqref{eq:common-d} at fixed $j,\delta$. Returning to the original coordinates multiplies all distances by $c_j\le1$. The cost at the trial point $c_j(y_j+z_j)$ therefore gives \[ J_j\le\int\log(1+c_j^2|a-z_j|^2)\,d\mu_j(a) \le b^{-1}\int g_j\,d\mu_j\le\beta_j/b . \] Divide by $\gamma_j$, take $j\to\infty$, and then let $\delta$ tend to zero. This gives $\limsup J_j/\gamma_j\le1/(d\alpha)=2/k$, and the bound $\rho_j\ge\sqrt{2-e^{J_j}}$ then gives $\limsup(1-\rho_j)/D_j\le1/k$. For the reverse bound, put equal masses at $\pm t e_1$. As shown before the theorem, $\rho=\sqrt{1-t^2}$ for $00$ the supremum is at least $1/(d-2)$, even when it is taken only over finite sets of equally weighted marks. \end{proof} At threshold one, Theorem~\ref{thm:all-dimensions} also gives the sharp lower bound for the inradius when the loss of volume is prescribed. Write $I_d(\varepsilon)$ for the infimum of $\rho$ over all distributions of marks with $0\le D\le\varepsilon$. We show that $I_d(\varepsilon)=1-\varepsilon/(d-2)+o(\varepsilon)$ as $\varepsilon$ tends to zero. First let $D=0$. Keep the distribution fixed, and choose compact subsets $K_j\subset E_1$ whose volumes tend to $\omega_d$ from below. In the proof above we may take $1/j$ as the tolerance in the choice of $K_j$, so that $0\le\beta_j\le(1/j+1/j^2)/d$. The estimate $J\le\beta_j/b$, in which $b$ stays bounded away from zero for each fixed small $\delta$, then gives $J=0$. The integrand in the definition of $J$ is zero only at the minimizing center and positive everywhere else. So the whole distribution sits at that point: it is a point mass, and $\rho=1$. Now let $00$. Indeed, by the radial volume formula their deficit $D(t)$ is a continuous function of $t$. It is zero at $t=0$ and exceeds $\varepsilon$ at $t=2\sqrt{\varepsilon/(d-2)}$, so $D(t)=\varepsilon$ for some $t$ in between. The two expansions above then give $1-\rho=\varepsilon/(d-2)+O(\varepsilon^2)$, with a constant that depends only on $d$. Hence $I_d(\varepsilon)=1-\varepsilon/(d-2)+o(\varepsilon)$. Two marks in the plane show why Theorem~\ref{thm:all-dimensions} requires $d\ge3$. In the plane, write a point in polar coordinates about the midpoint of the marks, with $u=\cos\theta$. The boundary of the region of two marks is then given by $r^2=t^2\cos(2\theta)+\sqrt{1-t^4\sin^2(2\theta)}$. The mean of $r^2$ over the angle $\theta$ is the area fraction, hence $D=t^4/4+O(t^8)$, while $1-\rho=t^2/2+O(t^4)$. Thus $1-\rho\sim\sqrt D$ as $t$ tends to zero. So already for two marks in the plane, the fractional loss of inradius is not bounded by a constant times the fractional loss of area.