Independent scholarship · Open accessThe monthly digest

Indexed text attachment

partial-differential-equations-in-higher-categories.tex

Source for Partial Differential Equations in Higher Categories

View paper · Download original file

Lines 981–1080 of 3921. Source text is displayed literally.

981\begin{exercise}\label{exer:par}982Prove the first half of \Cref{cor:par}.983\end{exercise}984 985\begin{exercise}\label{exer:neuwave}986For the half-line with the Neumann condition and data $u=\varphi$, $u_t=0$, write the solution in the region987$0<x<ct$.988\end{exercise}989 990\begin{exercise}\label{exer:period}991A string on $(0,l)$ is fixed at $0$ and free at $l$. Show from \Cref{thm:images} that every motion has period992$4l/c$.993\end{exercise}994 995\begin{exercise}\label{exer:duh}996Verify \Cref{thm:duhamel-wave} for $f(x,t)=\cos x$ by computing the integral.997\end{exercise}998%% ---- end ch/03-waves.tex ----999%% ---- begin ch/04-diffusions.tex ----1000\chapter{Diffusions}\label{ch:diffusions}1001 1002The diffusion equation $u_t=ku_{xx}$ has an explicit solution on the line as well, but the structures it carries1003are different in kind from those of \Cref{ch:waves}: order where there was a symplectic form, a semigroup where1004there was a group, probability where there was mechanics.1005 1006\section{A semigroup of Markov kernels}1007 1008Let $S(x,t)=(4\pi kt)^{-1/2}e^{-x^2/4kt}$, the source function, and $(T_t\varphi)(x)=\int S(x-y,t)\varphi(y)\,dy$.1009 1010\begin{theorem}\label{thm:semigroup}1011\Scl[source, \S2.4 and \S3.5] For bounded continuous $\varphi$, $u=T_t\varphi$ is the bounded solution of the1012diffusion equation with $u\to\varphi$ as $t\to0$. Moreover1013\begin{enumerate}[label=(\alph*)]1014\item $T_{t+s}=T_tT_s$ and $T_0=\id$;1015\item $T_t\varphi\ge0$ if $\varphi\ge0$, and $T_t1=1$;1016\item $\inf\varphi\le T_t\varphi\le\sup\varphi$.1017\end{enumerate}1018\end{theorem}1019 1020\noindent Properties (a) and (b) say that $t\mapsto T_t$ is a functor from the monoid $(\R_{\ge0},+)$, regarded as a1021category with one object, to the category $\Stoch$ of Markov kernels of \cite{ESCHC}. Property (c), the maximum1022principle on the line, is a consequence of (b): an average lies between the extremes. The kernel solves the1023equation, has unit mass and composes correctly to twelve digits; a unit step diffuses into1024$\tfrac12(1+\erf(x/\sqrt{4kt}))$. \chk{D1, D2}1025 1026\begin{theorem}[Maximum principle]\label{thm:max}1027\Scl[source, \S2.3] A solution in a space-time rectangle attains its maximum and its minimum on the initial side or1028the lateral sides. Hence the Dirichlet problem has at most one solution and depends continuously on its data in1029the uniform norm.1030\end{theorem}1031 1032\noindent The discrete form is exact: the explicit difference scheme is a substochastic matrix precisely when1033$k\Delta t/\Delta x^2\le\tfrac12$, and then it obeys the maximum principle; beyond that it has negative entries and1034the sawtooth grows. \chk{D3} Stability of that scheme, the first example in the source's chapter on computation, is1035membership in the Markov category (\Cref{ch:numerics}).1036 1037\begin{remark}[Generators]1038\Scl[\cite{Paz83}] A strongly continuous semigroup of contractions on a Banach space has a closed, densely defined1039generator $A=\lim(T_t-1)/t$, and the Hille--Yosida theorem characterizes the operators that arise: those with1040$\|(\lambda-A)^{-1}\|\le1/\lambda$ for $\lambda>0$. The diffusion equation is the statement that the generator of1041$(T_t)$ is $k\,\pd_x^2$. The resolvent is the Laplace transform of the semigroup (\Cref{ch:distributions}).1042\end{remark}1043 1044\section{Symmetry and the Gaussian}1045 1046The equation is invariant under translations, the reflection $x\mapsto-x$, and the scaling1047$(x,t)\mapsto(ax,a^2t)$ \src{\S2.4}. The last is not an isometry; it exists because the equation is of first order1048in time and second in space.1049 1050\begin{proposition}\label{prop:gauss}1051\Scl The source function is, up to a constant factor, the unique solution of the form1052$t^{-1/2}g(x/\sqrt t)$: the solution of unit mass fixed by the scaling action. \chk{D7}1053\end{proposition}1054 1055\noindent Convergence of $T_t\varphi$ to a multiple of $S$ as $t\to\infty$, for integrable $\varphi$, is the statement1056that this fixed point attracts; it is the central limit theorem in the language of \cite{ESCHC}.1057 1058\section{Irreversibility}1059 1060\begin{theorem}[Monotone quantities]\label{thm:mono}1061\Scl Along the diffusion flow of a probability density $u$:1062\begin{enumerate}[label=(\alph*)]1063\item $\frac d{dt}\int u^2=-2k\int u_x^2\le0$;1064\item $\frac d{dt}\bigl(-\int u\log u\bigr)=k\int u_x^2/u\ge0$ (entropy increases at the rate of the Fisher1065information);1066\item for two solutions, every relative entropy $\KL{u}{v}$ and every $f$-divergence is non-increasing.1067\end{enumerate}1068Moreover $T_t\varphi$ is smooth for every $t>0$, and positive everywhere if $\varphi\ge0$ is not zero.1069\end{theorem}1070 1071\noindent Statement (c) is the data-processing inequality: it holds for every morphism of $\Stoch$ \cite{ESCHC}. The1072invertible morphisms of $\Stoch$ are the deterministic bijections, and $T_t$ is not one of them; that is the1073categorical form of irreversibility. Entropy and square norm were tested under a doubly stochastic step on fifty1074random densities. \chk{D4} With $k=1$ the indicator function of $[-1,1]$ gives the positive value1075$2.3\times10^{-90}$ at $x=10$ after time $0.1$: nothing travels, and everything is everywhere at once. \chk{D8}1076 1077\begin{center}1078\begin{tikzpicture}1079\begin{scope}[shift={(0,0)}]1080\draw[ink2] (0,3)--(0,0)--(5.2,0);