Indexed text attachment
partial-differential-equations-in-higher-categories.tex
Source for Partial Differential Equations in Higher Categories
View paper · Download original file
Lines 2501–2600 of 3921. Source text is displayed literally.
2501\node[fwd] at (2.167,0.943) {\scriptsize $0.3$};2502\node[ink2] at (0.917,0.571) {\scriptsize $1$};2503\end{scope}2504\end{tikzpicture}2505\figcap{Left: the Poisson kernel $P_y(x)=y/\pi(x^2+y^2)$ of the half-plane, tending to the delta function as $y\to0$. Right: its Fourier transform $e^{-y|\xi|}$, tending to the constant $1$. Convergence of distributions on one side is pointwise convergence of multipliers on the other.}2506\end{center}2507 2508\begin{remark}[Hyperfunctions]\label{rem:hyper}2509\Sli[\cite{Sat59}] Sato defines a generalized function on the line as a pair of holomorphic functions above and2510below it, modulo functions holomorphic across it: the space is $\OO(\C\setminus\R)/\OO(\C)$, the first2511cohomology of the plane with support in the line and coefficients in $\OO$. Distributions embed in it, $\delta$ as2512the class of $-1/2\pi iz$. This is the second place in the volume, after \Cref{prop:annulus}, where a cohomology2513group is the object of study and not a tool; in several variables it is a higher cohomology group, and the theory2514of linear equations built on it is derived from the start \cite{KaSc90}.2515\end{remark}2516 2517\section{The Laplace transform}2518 2519\begin{proposition}\label{prop:resolvent}2520\Scl[source, \S12.5; \cite{Paz83}] For a contraction semigroup with generator $A$ and $s>0$,2521$(s-A)^{-1}=\int_0^\infty e^{-st}T_t\,dt$.2522\end{proposition}2523 2524\noindent So the Laplace transform in time, the source's last technique, turns an evolution problem into a family2525of stationary ones, and the three presentations of one object, semigroup, generator and resolvent, determine one2526another. \chk{T6} Together with \Cref{thm:wick} this closes a circle: elliptic problems $(s-A)u=f$, parabolic2527problems $u_t=Au$ and, for $A\le0$ self-adjoint, hyperbolic problems $u_{tt}=Au$ are three functions of one2528operator.2529 2530\section*{Notes}2531 2532The source's Chapter 12 is followed. For distributions and the kernel theorem see \cite{Sch57,Hor83}; for the2533impossibility of multiplication \cite{Sch54}; for fundamental solutions \cite{Mal56,Ehr54}; for hyperfunctions2534\cite{Sat59,KaSc90}.2535 2536\section*{Exercises}2537 2538\begin{exercise}\label{exer:absx}2539Show that $|x|''=2\delta$ in the sense of distributions.2540\end{exercise}2541 2542\begin{exercise}\label{exer:fund1}2543Find fundamental solutions of $d^2/dx^2$ and of $-d^2/dx^2+1$ on the line.2544\end{exercise}2545 2546\begin{exercise}\label{exer:fts}2547Derive the source function of the diffusion equation from $\hat S=e^{-k^2t}$.2548\end{exercise}2549 2550\begin{exercise}\label{exer:poisson-sg}2551Show that the Poisson kernels of the half-plane form a semigroup under convolution.2552\end{exercise}2553%% ---- end ch/11-distributions.tex ----2554%% ---- begin ch/12-numerics.tex ----2555\chapter{Discretization}\label{ch:numerics}2556 2557The source devotes a chapter to computation and opens it with a warning: a scheme that looks reasonable can2558produce nonsense \src{\S8.1}. The categorical view of a numerical method is that it is a functor on a small2559category of grids that approximately commutes with the solution functor; the warning is that approximate2560commutation of one step says nothing about many.2561 2562\section{Squares that almost commute}2563 2564Let $E(t)$ be the solution operator of a well-posed linear evolution problem, $R_h$ the restriction of functions to2565a grid of mesh $h$, and $E_h$ one step of a scheme with time step $\Delta t$ tied to $h$.2566\[ \begin{tikzcd}[column sep=large]2567X\ar[r,"E(\Delta t)"]\ar[d,"R_h"'] & X\ar[d,"R_h"]\\2568X_h\ar[r,"E_h"'] & X_h2569\end{tikzcd} \]2570The scheme is \term{consistent} of order $p$ if the square commutes up to $O(\Delta t\,h^p)$ on smooth functions;2571\term{stable} if the powers $E_h^n$ are bounded uniformly for $n\Delta t\le T$; \term{convergent} if2572$E_h^nR_h-R_hE(n\Delta t)\to0$.2573 2574\begin{theorem}[Lax equivalence]\label{thm:lax}2575\Scl[\cite{LaRi56}] For a consistent scheme, stability is equivalent to convergence.2576\end{theorem}2577\begin{proof}[Idea]2578$E_h^nR_h-R_hE^n=\sum_{j<n}E_h^{n-1-j}(E_hR_h-R_hE)E^j$: the defect of one square, transported by the remaining2579steps. Stability is exactly what keeps the transport from amplifying it.2580\end{proof}2581 2582\noindent The centred three-level scheme for the heat equation is consistent of second order and its amplification2583factor exceeds one for every mesh ratio, $1.48$ at ratio $0.1$: it converges for no data but the trivial. \chk{N10}2584 2585\section{Diffusion}2586 2587For the explicit scheme with $s=k\Delta t/\Delta x^2$ the amplification factor of the mode $e^{ikx}$ is2588$1-2s(1-\cos k\Delta x)$ \src{\S8.2}.2589 2590\begin{theorem}\label{thm:explicit}2591\Scl[source, \S8.2] The explicit scheme is stable if and only if $s\le\tfrac12$, which is also the condition for its2592matrix to have non-negative entries.2593\end{theorem}2594 2595\noindent At $s=0.52$ the computed growth per step equals the predicted $1.0684$; at $s=0.5$ nothing grows.2596\chk{N1} The second clause is \Cref{thm:max}: stability, here, is that the scheme is a morphism of the Markov2597category, and the source's observation that ``none of the coefficients are negative'' predicts the right2598condition for that reason. The error is of second order in $\Delta x$, and of fourth order at the special ratio2599$s=\tfrac16$, where the leading error terms in time and space cancel: halving the mesh divides the error by2600$4.06$ and by $16.1$. \chk{N2} The Crank--Nicolson scheme, the Cayley transform of the generator, has amplification