% vim: tw=50 % 15/03/2022 11AM \begin{proof} $X = (X_1, X_2)$ independent. If $m_X((u_1, u_2))$ splits as a product $f_1(u_1)f_2(u_2)$. In our setting: \begin{align*} \exp (u^\top \mu) &= \exp(u_1 \mu_2) \exp(u_2 \mu_2) \\ \exp \left( \half u^\top V u \right) &= \exp(u_1^2 \sigma_1^2) \exp(u_2^2 \sigma_2^2) \exp(2u_1u_2 \mathrm{Cov}(X_1, x_2)) \end{align*} So it splits as a product if and only if $\mathrm{Cov} = 0$. \end{proof} \myskip Motivation: $\mathrm{Cov}(100X_1, X_2) = 100\mathrm{Cov}(X_1, X_2)$ so ``\emph{large} covariance'' doesn't imply ``\emph{very} dependent''. \begin{definition*} \emph{Correlation} of $X$, $Y$ is \[ \mathrm{Corr}(X, Y) = \frac{\mathrm{Cov}(X, Y)}{\sqrt{\mathrm{Var}(X)\mathrm{Var}(Y)}} \] (It is a fact that this is always $\in [-1, 1]$) \end{definition*} \begin{proposition*} If $(X, Y)$ Gaussian, then $Y = aX + Z$ where $Z$ is Gaussian, and $(X, Z)$ independent. \end{proposition*} \begin{proof} Define $Z = Y - aX$ for $a \in \RR$. \begin{claim*} $(X, Z)$ is Gaussian. \end{claim*} \begin{proof} \[ u_1 X + u_2 Z = u_1 X + u_2 (Y - aX) = (u_1 - au_2) X + u_2 Y .\] \end{proof} \myskip Goal: find $a$ such that $\mathrm{Cov}(X, Z) = 0$. \[ \mathrm{Cov}(X, Z) = \mathrm{Cov}(X, Y - aX) = \mathrm{Cov}(X, Y) - a\mathrm{Var}(X) \] so take \[ a = \frac{\mathrm{Cov}(X, Y)}{\mathrm{Var}(X)} \] Then $\mathrm{Cov}(X, Z) = 0$ so $X, Z$ independent. \end{proof} \subsection{Two Historical Models} \subsubsection*{Bertrand's Paradox} Goal: choose a uniform chord of circle. Two methods: \begin{center} \includegraphics[width=0.6\linewidth] {images/40ae266aa45211ec.png} \end{center} \begin{enumerate}[(i)] \item $A$, $B$ uniform on circumference. \item midpoint $M$ uniform on disc. \end{enumerate} Conclusion: Gives different distributions. (Completely unsurprising?) \myskip \ul{Method (i)} \begin{center} \includegraphics[width=0.6\linewidth] {images/6d2bf11ca45311ec.png} \end{center} $\theta \sim \mathrm{Unif} \left[ 0, \frac{\pi}{2} \right]$ then $|AB| = 2r \sin \theta$. Note $|OM| = r\cos\theta$, so $\PP(|OM| \le \eps r) \approx r\eps$ when $\eps \to 0$. \myskip \ul{Method (ii)} \\ $\PP(|OM| \le \eps r) = \frac{\pi(\eps r)^2}{\pi r^2} = \eps^2$. \subsubsection*{Buffon's Needle} \begin{center} \includegraphics[width=0.6\linewidth] {images/96ace8f2a45311ec.png} \end{center} \begin{itemize} \item Lines spaced $L$ apart. \item Needle length $L$ dropped ``uniformly'' \item Observe whether intersects a line. \end{itemize} We work ``modulo $L$'': \begin{center} \includegraphics[width=0.6\linewidth] {images/fbc5e2fca45311ec.png} \end{center} \[ \text{$X$ centre} \sim \mathrm{Unif}[0, L) \] \[ \text{Angle }\theta \sim \mathrm{Unif}[0, \pi) \] Density of $(X, \theta)$ constant $= \frac{1}{L\pi}$. Crosses line if \[ X \le \frac{L}{2} \sin \theta \] or \[ L - X \le \frac{L}{2} \sin\theta \] \begin{align*} \PP(\text{crosses line}) &= \PP \left( \min(X, L - X) \le \frac{L}{2} \sin\theta \right) \\ &= 2\PP \left( X \le \frac{L}{2} \sin \theta \right) \\ &= 2 \int_{\theta = 0}^{\pi} \int_{x = 0}^{\frac{L}{2} \sin\theta} \frac{1}{L\pi} \dd x \dd \theta \\ &= 2 \int_{\theta = 0}^{\pi} \frac{1}{2\pi} \sin\theta \dd \theta \\ &= \frac{2}{\pi} \\ &\approx 0.64 \end{align*} What's the point? Calculate $\pi$ experimentally. \\ Efficiency? Try $n$ times. Number of intersections: $S_n \sim \mathrm{Bin} \left( n, \frac{\pi}{2} \right)$. \\ Proportion $\hat{p}_n$ of intersections $= \frac{S_n}{n}$. By CLT: \[ \hat{p}_n = p + \sqrt{\frac{p(1 - p)}{n}} Z \] so \[ \hat{p}_n - p \approx \sqrt{\frac{p(1 - p)}{n}} Z .\] Estimate: \[ \hat{\pi}_n = \frac{2}{\hat{p}_n} \] Taylor expanding: \begin{align*} \hat{\pi}_n &= \frac{2}{\hat{p}_n} \\ &\approx \frac{2}{p} - (\hat{p}_n - p) \frac{2}{p^2} \end{align*} so \[ \hat{\pi}_n - \pi \approx -\frac{\pi^2}{2} \sqrt{\frac{p(1 - p)}{n}} Z \approx \frac{-2.4}{\sqrt{n}} Z \] So if you seek \[ \hat{\pi}_n - \pi \approx O(10^{-k}) \] (correct to $k$ decimal places) then we need $n \approx 10^{2k}$. \begin{itemize} \item Historical interest. \item Not computationally efficient. \item Detailed calculation of \emph{sampling errors} in other settings on problem sheet. \end{itemize}