%! TEX root = ABF.tex % vim: tw=80 ft=tex % 17/03/2026 12PM \begin{fcdefn}[] Let $f : \Rbb^n \to \Rbb$ and let $\rho \in [-1, 1]$. Then $U_\rho f(x) = \Ebb_{y \sim N_\rho(x)} f(y)$. \end{fcdefn} \begin{remark*} If $x \sim N(0, 1)$ and $y \sim N_\rho(x)$, then $y \sim N(0, 1)$ and $x \sim N_\rho(y)$, from which it follows that $U_\rho$ is self-adjoint. If $y \sim N_\rho(x)$ and $z \sim N_\sigma(y)$, then there are independent Gaussians $g_1$, $g_2$ with $y \sim \rho x + \sqrt{1 - \rho^2} g_1$, $z \sim \sigma(\rho x + \sqrt{1 - \rho^2} g_1) + \sqrt{1 - \sigma^2} g_2$. But $\sigma^2 (1 - \rho^2) + 1 - \sigma^2 = 1 - \rho^2 \sigma^2$, so $z \sim N_{\rho \sigma}(x)$. From this, it follows easily that $U_\rho U_\sigma = U_{\rho \sigma}$, i.e. $\{U_\rho : \rho \in [-1, 1]\}$ forms a semigroup, called the \emph{Orstein--Uhlenbeck semigroup}. \end{remark*} We define, if $f : \Rbb^n \to \Rbb$, $\Stab_\rho f$ to be \[ \langle f, U_\rho f \rangle = \Ebb_{x \sim_\rho y} f(x) f(y) .\] \begin{fcthm}[Sheppard's formula] % Theorem 8.4 \label{thm:8.4} Let $A \subset \Rbb^n$ be a half space, i.e. a set of the form $\{x : \langle x, n \rangle \ge 0\}$ for some non-zero $u$. Then $\Stab_\rho \indicator{A} = \half - \frac{\cos^{-1} \rho}{2\pi}$. \end{fcthm} \begin{proof} We are interested in \[ \Ebb_{x \sim_\rho y} \indicator{A}(x) \indicator{A}(y) = \Pbb_{x \sim_\rho y} [\text{$\langle x, u \rangle \ge 0$ and $\langle y, u \rangle \ge 0$}] .\] Without loss of generality $u$ is a unit vector. Then $\langle x, u \rangle$ and $\langle y, u \rangle$ are $\rho$-correlated $1$-dimensional Gaussians (by rotational invariance, we can think of $u$ as just being $u = e_1$). So pick unit vectors $u, v \in \Rbb^2$, with $\langle v, w \rangle = \rho$, and consider $\langle v, g \rangle$, $\langle w, g \rangle$, $g \sim N(0, 1)^2$. Then draw a picture: \begin{center} \includegraphics[width=0.3\linewidth]{images/67569673e4b8438c.png} \end{center} From this we get \[ \Pbb[\text{$\langle v, g \rangle \ge 0$ and $\langle w, g \rangle \ge 0$}] = \frac{\pi - \cos^{-1} \rho}{2\pi} = \half - \frac{\cos^{-1} \rho}{2\pi} . \qedhere \] \end{proof} \begin{fcdefn}[Rotation sensitivity] Let $A \subset \Rbb^n$. The \emph{rotation sensitivity} $\RS_\delta(A)$ is \[ \Pbb_{x \sim_{\cos \delta} y} [\indicator{A}(x) \neq \indicator{A}(Y)] .\] \end{fcdefn} If $A$ is \emph{balanced} (i.e. has Gaussian measure $\half$), then $\Pbb[x \in A, y \in A] = \Pbb[x \notin A, y \notin A]$, so \[ \Pbb_{x \sim_{\cos \delta} y} [\indicator{A}(x) \neq \indicator{A}(y)] = 1 - 2 \Pbb_{x \sim_{\cos \delta} y} [\indicator{A}(x) = \indicator{A}(y)] = 1 - 2\Stab_{\cos \delta} \indicator{A}(x) .\] The statement $\Stab_{\cos \delta} \indicator{A}(x) \le \half - \frac{\delta}{2\pi}$ is equivalent to $\RS_\delta(A) \ge \frac{\delta}{\pi}$. \begin{fclemma}[Subadditivity of $RS$] Let $A$ be a balanced set in $\Rbb^n$. Then for any $\delta_1, \ldots, \delta_k \ge 0$ we have \[ \RS_{\delta_1 + \cdots + \delta_k}(A) \le \RS_{\delta_1}(A) + \cdots + \RS_{\delta_k}(A) .\] \end{fclemma} \begin{proof} For $i = 0, \ldots, k$, let $\theta_i = \delta_1 + \cdots + \delta_i$. Let $g$ and $g'$ be independent $n$-dimensional Gaussians and let $x_i = \cos \theta_i g + \sin \theta_i g'$ for each $i$. Then $x_0$ and $x_k$ are $\cos \theta_k$-correlated, so \[ \RS_{\delta_1 + \cdots + \delta_k}(A) = \Pbb[\indicator{A}(x_0) \neq \indicator{A}(x_k)] .\] Also, $x_{i - 1}$ and $x_i$ are ($\cos \theta_{i - 1} \cos \theta_i + \sin \theta_{i - 1} \sin \theta_i = \cos(\theta_i - \theta_{i - 1}) = \cos \delta_i$)-correlated. So the RHS equals $\sum_{i = 1}^{k} \Pbb[\indicator{A}(x_{i - 1}) \neq \indicator{A}(x_i)]$. The result now follows from a union bound. \end{proof} \begin{fccoro}[Special case of Borell's isoperimetric inequality] % Corollary 8.6 \label{coro:8.6} Let $A$ be balanced and let $k \in \Nbb$. Then $\RS_k(A) \ge \frac{1}{2k}$. \end{fccoro} \begin{proof} $\RS_{\frac{\pi}{2}}(A) = \half$ because $\cos \frac{\pi}{2}$-correlated variables are independent. Setting $\delta_1 = \cdots = \delta_k = \frac{\pi}{2k}$, we deduce that $k \RS_{\frac{\pi}{2k}} \ge \half$, hence $\RS_{\frac{\pi}{2k}} \ge \frac{1}{2k}$. \end{proof}