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 301–400 of 3921. Source text is displayed literally.

301(\Cref{thm:sunada}).302\item \emph{Gluing.} Dirichlet-to-Neumann maps and transfer matrices compose as Lagrangian relations303(\Cref{thm:gluing}, \Cref{thm:transfer}).304\item \emph{Discretization.} A scheme is a square that almost commutes; its stability is positivity for diffusion,305causality for waves, and compatibility with the de Rham complex for elliptic problems (\Cref{ch:numerics}).306\item \emph{Morphisms of equations.} The solvable nonlinear examples of the source are images of linear equations307(\Cref{thm:colehopf}, \Cref{thm:ist}).308\end{itemize}309 310Quantitative statements accompany the qualitative ones throughout: amplification factors, exponents, constants,311eigenvalues, rates, counts of terms. All of them are recomputed by the companion script, which performs 107 checks.312 313How much of this is ``higher'' is assessed in \Cref{ch:coda}. In brief: most of it is ordinary category theory and314representation theory; cohomology enters in degree one; and gauge fields are the one place where the solution315functor must take values in higher groupoids (\Cref{rem:stack}). Nothing here is a new theorem about partial316differential equations, and no existence theorem is proved. A companion paper \cite{PDEML} applies the framework to317machine learning.318 319The volume belongs to a series with \cite{GIHC}, on game theory, \cite{ESCHC}, on information theory, \cite{DTHC}, on320differential topology, and \cite{EGHC}, on evolutionary games, and uses the second and third.321 322\tableofcontents323%% ---- end front.tex ----324\mainmatter325\part{Equations}326%% ---- begin ch/01-equations.tex ----327\chapter{What an Equation Is}\label{ch:equations}328 329The source begins by solving a few equations and only then asks what kind of thing has been solved330\src{\S1.1--1.2, \S1.6}. We reverse the order, because the answer decides what the rest of the volume can say.331A differential equation will be an object of a category. Its solutions will be the values of a functor. The332statements of the following chapters are then of one form: such-and-such a map of domains, of equations or of333function spaces induces a map of solutions.334 335\section{Equations in jets}336 337Let $M\subseteq\R^n$ be open and let $u$ be a real function on $M$. The \term{$k$-jet} of $u$ at $x$ is the list338of its derivatives of order at most $k$ there; jets form a bundle $J^kM\to M$ with coordinates339$(x,u_\alpha)_{|\alpha|\le k}$, and $u$ determines a section $j^ku$.340 341\begin{definition}\label{def:pde}342A \term{differential equation of order $k$} on $M$ is a subset $\EE\subseteq J^kM$, typically the zero set of a343function $F(x,u,u_i,u_{ij},\dots)$. A \term{solution} is a function $u$ whose jet $j^ku$ takes values in $\EE$.344\end{definition}345 346Differentiating $F=0$ gives equations of higher order that every solution also satisfies; adjoining all of347them gives the \term{prolongation} $\EE^\infty\subseteq\Jet M$. A point of $\EE^\infty$ over $x$ is a348\term{formal solution} at $x$: a Taylor series that satisfies the equation to all orders.349 350\begin{theorem}[Equations are coalgebras]\label{thm:jet}351\Sli[\cite{Mar86,KhSc17}] The infinite jet construction is a comonad $\Jet$ on bundles over $M$. Its coalgebras are352the (formally integrable) differential equations in the sense of Vinogradov, and a morphism of coalgebras is a353differential operator that carries solutions to solutions. In synthetic differential geometry $\Jet$ is base change354along the unit $M\to\Im M$ of the infinitesimal shape, and in the cases of interest equations over $M$ are355bundles over $\Im M$.356\end{theorem}357 358\noindent We use two consequences and none of the machinery. There is a category $\PDE$ whose morphisms are359differential substitutions; the Cole--Hopf and Miura transformations of \Cref{ch:nonlinear} are morphisms in it.360And solutions are functorial: $\EE\mapsto\Sol(\EE)$ sends a morphism of equations to a map of solution sets. Since361solutions restrict to open subsets and glue, $\Sol(\EE)$ is a sheaf on $M$.362 363\section{Linear equations and their modules}364 365Let $\DD$ be the ring of linear differential operators on $M$ with smooth coefficients. A linear equation $Pu=0$366is recorded by the left $\DD$-module $\DD/\DD P$, and for any left $\DD$-module $\FF$ of functions367\[ \Sol(P;\FF)=\Hom_\DD(\DD/\DD P,\FF)=\{u\in\FF:Pu=0\}. \]368This is functorial in both variables. In $\FF$ it is covariant: smooth functions sit inside distributions, and369classical solutions inside weak ones (\Cref{ch:distributions}). In the equation it is contravariant: if $QR=SP$ then370$u\mapsto Ru$ maps $\Sol(P)$ to $\Sol(Q)$.371 372\begin{proposition}[Factorization]\label{prop:factor}373\Srf For operators $P,Q$ there is an exact sequence374\[ 0\longrightarrow\Sol(Q)\longrightarrow\Sol(PQ)\xrightarrow{\ Q\ }\Sol(P). \]375If $P$ and $Q$ commute, then $\Sol(P)+\Sol(Q)\subseteq\Sol(PQ)$, with equality as soon as $Q$ maps $\Sol(P)$ onto376itself: for $u\in\Sol(PQ)$ choose $w\in\Sol(P)$ with $Qw=Qu$, and then $u-w\in\Sol(Q)$.377\end{proposition}378 379\noindent The wave operator is the case $P=\pd_t-c\pd_x$, $Q=\pd_t+c\pd_x$, checked on polynomials \chk{E5}; d'Alembert's380solution in \Cref{ch:waves} is the equality $\Sol(PQ)=\Sol(P)+\Sol(Q)$.381 382In a diagram: for commuting $P$ and $Q$ the square of inclusions383\[\begin{tikzcd}[column sep=large]384\Sol(P)\cap\Sol(Q)\arrow[r,hook]\arrow[d,hook] & \Sol(P)\arrow[d,hook]\\385\Sol(Q)\arrow[r,hook] & \Sol(PQ)386\end{tikzcd}\]387is a pullback of vector spaces, and it is also a pushout exactly when $\Sol(P)+\Sol(Q)=\Sol(PQ)$. For the wave388operator the corner is the constants, and d'Alembert's theorem says that the square is bicartesian: a solution is389a right-moving and a left-moving wave, glued along what they have in common.390 391For constant coefficients $\DD=\R[\pd_1,\dots,\pd_n]$ is a polynomial ring, $P=p(\pd)$, and the module is392$\R[\xi]/(p)$: the equation \emph{is} the hypersurface $V(p)=\{p(\xi)=0\}\subseteq\C^n$. Each point $\xi\in V(p)$ gives the393exponential solution $e^{\xi\cdot x}$.394 395\begin{theorem}[Fundamental principle]\label{thm:ehrenpreis}396\Sli[\cite{Ehr70,Pal70}] On a convex domain every smooth solution of a constant-coefficient equation $p(\pd)u=0$ is an397integral, over $V(p)$, of exponential solutions with polynomial coefficients.398\end{theorem}399 400\noindent Separation of variables is thus not a trick that happens to work: for constant coefficients it is the