Indexed text attachment
partial-differential-equations-in-higher-categories.tex
Source for Partial Differential Equations in Higher Categories
View paper · Download original file
Lines 2681–2780 of 3921. Source text is displayed literally.
2681\fill[ink2] (-1,1.8) circle (1pt);2682\fill[ink2] (-0.5,1.8) circle (1pt);2683\fill[ink2] (0,1.8) circle (1pt);2684\fill[ink2] (0.5,1.8) circle (1pt);2685\fill[ink2] (1,1.8) circle (1pt);2686\fill[ink2] (1.5,1.8) circle (1pt);2687\fill[ink2] (2,1.8) circle (1pt);2688\fill[fwd!20,opacity=0.6] (0,1.8)--(-1.5,0)--(1.5,0)--cycle;2689\draw[thick,fwd] (-1.5,0)--(0,1.8)--(1.5,0);2690\draw[thick,bwd,densely dashed] (-0.9,0)--(0,1.8)--(0.9,0);2691\draw[very thick,bwd] (-0.9,-0.12)--(0.9,-0.12);2692\fill (0,1.8) circle (1.6pt);2693\node at (0,-0.6) {\scriptsize $c\,\Delta t<\Delta x$: stable};2694\end{scope}2695\begin{scope}[xshift=6cm]2696\fill[ink2] (-2,0) circle (1pt);2697\fill[ink2] (-1.5,0) circle (1pt);2698\fill[ink2] (-1,0) circle (1pt);2699\fill[ink2] (-0.5,0) circle (1pt);2700\fill[ink2] (0,0) circle (1pt);2701\fill[ink2] (0.5,0) circle (1pt);2702\fill[ink2] (1,0) circle (1pt);2703\fill[ink2] (1.5,0) circle (1pt);2704\fill[ink2] (2,0) circle (1pt);2705\fill[ink2] (-2,0.6) circle (1pt);2706\fill[ink2] (-1.5,0.6) circle (1pt);2707\fill[ink2] (-1,0.6) circle (1pt);2708\fill[ink2] (-0.5,0.6) circle (1pt);2709\fill[ink2] (0,0.6) circle (1pt);2710\fill[ink2] (0.5,0.6) circle (1pt);2711\fill[ink2] (1,0.6) circle (1pt);2712\fill[ink2] (1.5,0.6) circle (1pt);2713\fill[ink2] (2,0.6) circle (1pt);2714\fill[ink2] (-2,1.2) circle (1pt);2715\fill[ink2] (-1.5,1.2) circle (1pt);2716\fill[ink2] (-1,1.2) circle (1pt);2717\fill[ink2] (-0.5,1.2) circle (1pt);2718\fill[ink2] (0,1.2) circle (1pt);2719\fill[ink2] (0.5,1.2) circle (1pt);2720\fill[ink2] (1,1.2) circle (1pt);2721\fill[ink2] (1.5,1.2) circle (1pt);2722\fill[ink2] (2,1.2) circle (1pt);2723\fill[ink2] (-2,1.8) circle (1pt);2724\fill[ink2] (-1.5,1.8) circle (1pt);2725\fill[ink2] (-1,1.8) circle (1pt);2726\fill[ink2] (-0.5,1.8) circle (1pt);2727\fill[ink2] (0,1.8) circle (1pt);2728\fill[ink2] (0.5,1.8) circle (1pt);2729\fill[ink2] (1,1.8) circle (1pt);2730\fill[ink2] (1.5,1.8) circle (1pt);2731\fill[ink2] (2,1.8) circle (1pt);2732\fill[fwd!20,opacity=0.6] (0,1.8)--(-1.5,0)--(1.5,0)--cycle;2733\draw[thick,fwd] (-1.5,0)--(0,1.8)--(1.5,0);2734\draw[thick,bwd,densely dashed] (-2,0)--(0,1.8)--(2,0);2735\draw[very thick,bwd] (-2,-0.12)--(2,-0.12);2736\fill (0,1.8) circle (1.6pt);2737\node at (0,-0.6) {\scriptsize $c\,\Delta t>\Delta x$: cannot converge};2738\end{scope}2739\end{tikzpicture}2740\figcap{The Courant--Friedrichs--Lewy condition. The value at the top grid point is computed from the grid points in the shaded triangle; the solution there depends on the data in the interval cut out by the characteristics (dashed). If that interval is not inside the triangle, the scheme never sees data on which the answer depends.}2741\end{center}2742 2743\section{Laplace's equation}2744 2745The five-point scheme says that each value is the mean of its four neighbours: the discrete mean value property,2746with the random walk of \Cref{thm:kakutani} behind it \src{\S8.4}. The classical iterations solve it by running a2747process to equilibrium.2748 2749\begin{proposition}\label{prop:iter}2750\Scl[\cite{You50}] On an $N\times N$ grid the Jacobi iteration contracts the error by $\cos(\pi/N)$ per sweep, the2751Gauss--Seidel iteration by $\cos^2(\pi/N)$, and successive over-relaxation with the optimal parameter2752$\omega=2/(1+\sin(\pi/N))$ by $\omega-1$.2753\end{proposition}2754 2755\noindent For $N=16$ a random error is reduced by $10^{-6}$ in $585$, $298$ and $41$ sweeps. \chk{N6} Jacobi iteration is2756the explicit diffusion scheme at $s=\tfrac14$ run to its steady state, as the source notes; Gauss--Seidel is ``about2757twice as fast'', exactly so asymptotically; and the choice of $\omega$, which the source leaves as an art, has for2758this model problem the closed form above.2759 2760\section{Finite elements}2761 2762By \Cref{thm:dirichlet} the Galerkin solution in a space of trial functions is the energy projection of the true2763solution \src{\S8.5}. For piecewise linear functions on a uniform mesh of the interval the stiffness matrix is the2764difference Laplacian divided by $h$; the Galerkin solution is exact at the nodes, because the Green's function2765$x(1-y)$ is itself piecewise linear; and the energy error halves with $h$. \chk{N7} On a uniform triangulation of2766the square, linear elements give the five-point scheme. \chk{N8} The method earns its keep on domains where2767differences do not fit, and there the corner exponents of \Cref{prop:wedge} limit the rate.2768 2769\section{Discretize the complex}2770 2771\Cref{prop:tonti} suggests discretizing the two factors of the Laplacian separately: the exterior derivative2772exactly, on cochains of a mesh, and the Hodge star approximately.2773 2774\begin{theorem}[Finite element exterior calculus]\label{thm:feec}2775\Sli[\cite{AFW06}] On a triangulated domain the Whitney forms are a finite-dimensional subcomplex of the de Rham2776complex, isomorphic to simplicial cochains, and there are projections onto it that are bounded and commute with2777$\dd$. Consequently the discrete Hodge Laplacian is stable and convergent in every degree, and its harmonic forms2778have the dimensions of the cohomology of the domain.2779\end{theorem}2780