Indexed text attachment
partial-differential-equations-in-higher-categories.tex
Source for Partial Differential Equations in Higher Categories
View paper · Download original file
Lines 601–700 of 3921. Source text is displayed literally.
601$\dd$ exactly and $\star$ approximately (\Cref{ch:numerics}).602 603With signs suppressed, the proposition is the square604\[\begin{tikzcd}[column sep=huge,row sep=large]605\text{potential }u\arrow[r,"\dd"]\arrow[d,dashed,"\Lap"'] & \text{gradient }\dd u\arrow[d,"\star\ \text{(material)}"]\\606\text{source }f & \text{flux }q\arrow[l,"\dd"]607\end{tikzcd}\]608in which the horizontal arrows are natural for all smooth maps and the right-hand one only for isometries. Every609linear equation of the volume is a way round a square of this shape, sometimes with time as one of the variables.610 611\section{Data as a restriction map}612 613An equation has many solutions; a problem adds conditions \src{\S1.4}. Initial conditions prescribe $u$, and for614second order in time also $u_t$, on $\{t=0\}$. Boundary conditions prescribe $u$ (Dirichlet), the normal derivative615(Neumann) or a combination $\pd_nu+au$ (Robin) on the boundary of the spatial domain. In every case there is a616linear map617\[ r\colon\Sol(P;X)\longrightarrow Y \]618from solutions in a chosen function space $X$ to data in a chosen space $Y$.619 620\begin{definition}[Hadamard]\label{def:wellposed}621The problem $(P,r,X,Y)$ is \term{well-posed} if $r$ is an isomorphism of topological vector spaces: surjective622(existence), injective (uniqueness), with continuous inverse (stability).623\end{definition}624 625\noindent The definition is the source's \src{\S1.5} with one addition that the categorical phrasing makes626unavoidable: stability is a property of a morphism in a category of topological vector spaces, and it depends on627the category. A problem may be well-posed for one pair $(X,Y)$ and ill-posed for another.628 629\begin{example}[Kernel and cokernel]\label{ex:fredholm}630For the Neumann problem $-\Lap u=f$, $\pd_nu=0$, the map from $u$ to $f$ is not an isomorphism: constants are in631the kernel and a solution exists only if $\int f=0$. It is Fredholm of index zero, with an exact sequence632$0\to\R\to X\to Y\to\R\to0$. On a path with six vertices the rank is $5$ and the system is solvable exactly for633sources of mean zero. \chk{W5} The source's underdetermined and overdetermined problems are those with non-zero634kernel and cokernel; its first example, $u''+u=0$ with $u(0)=u(L)=0$, has a kernel exactly when $\sin L=0$. \chk{W6}635\end{example}636 637\begin{example}[Hadamard]\label{ex:hadamard}638The functions $u_n(x,y)=n^{-1}e^{-\sqrt n}\sin nx\sinh ny$ are harmonic, vanish on $y=0$, and have normal639derivative $e^{-\sqrt n}\sin nx$ there, which tends to zero with all its derivatives. At $y=1$ their size is640$e^{-\sqrt n}\sinh n/n$.641\begin{center}\small642\begin{tabular}{@{}lccc@{}}643\toprule644$n$ & $25$ & $100$ & $400$\\645\midrule646size of the data & $6.7\times10^{-3}$ & $4.5\times10^{-5}$ & $2.1\times10^{-9}$\\647size of the solution at $y=1$ & $9.7\times10^{6}$ & $6.1\times10^{36}$ & $1.3\times10^{162}$\\648\bottomrule649\end{tabular}650\end{center}651The Cauchy problem for Laplace's equation has at most one solution, and its inverse is not continuous in any652topology of finitely many derivatives. \chk{W1}653\end{example}654 655\begin{example}[Time's arrow]\label{ex:backward}656For $u_t=u_{xx}$ on $(0,\pi)$ the mode $\sin nx$ evolves by the factor $e^{-n^2t}$. Forwards this is a contraction.657Backwards over a time $0.1$ the modes $n=1,10,30$ are amplified by $1.1$, $2.2\times10^4$ and $1.2\times10^{39}$.658\chk{W2} The backward problem is solvable on a dense set of data and its solution map is unbounded: the heat659flow is a semigroup that is not a group (\Cref{ch:diffusions}).660\end{example}661 662\section{Uniqueness by energy}663 664\begin{theorem}[Two energy laws]\label{thm:energy}665\Scl[source, \S2.2--2.3] For the wave equation $u_{tt}=c^2u_{xx}$ on the line, with data of compact support, the energy666$E=\tfrac12\int(u_t^2+c^2u_x^2)\,dx$ is constant. For the diffusion equation $u_t=ku_{xx}$ with Dirichlet or Neumann667conditions, $\tfrac{d}{dt}\tfrac12\int u^2\,dx=-k\int u_x^2\,dx\le0$. In both cases a solution with zero data is zero.668\end{theorem}669 670\noindent On a grid of twenty points the wave energy drifted by $5.5\times10^{-12}$ in relative terms over five time671units, and the square norm under the diffusion flow decreased at every step. \chk{W4} The two laws are the first672appearance of the dichotomy of \Cref{ch:diffusions}: a conserved quadratic form makes the wave flow a group of673isometries; a decreasing one makes the heat flow a semigroup of contractions.674 675\section{Which problem for which type}676 677\begin{center}\small678\begin{tabular}{@{}llll@{}}679\toprule680type & archetype & well-posed problem & solution map\\681\midrule682elliptic & $\Lap u=0$ & values on the boundary & a projection (minimizer)\\683hyperbolic & $u_{tt}=c^2\Lap u$ & $u,u_t$ at one time & a group\\684parabolic & $u_t=k\Lap u$ & $u$ at one time, forwards & a semigroup\\685\bottomrule686\end{tabular}687\end{center}688 689\begin{center}690\begin{tikzpicture}691\begin{scope}692\draw[very thick,accent,fill=fwd!8] plot[smooth cycle,tension=0.8] coordinates {(0.2,0.3) (1.6,0.0) (2.8,0.6) (2.9,1.9) (1.7,2.5) (0.4,2.0)};693\node at (1.55,1.3) {\scriptsize $\Lap u=0$};694\node[accent] at (1.5,-0.4) {\scriptsize values on all of $\partial D$};695\node at (1.5,3.0) {\scriptsize elliptic};696\end{scope}697\begin{scope}[xshift=4.2cm]698\fill[fwd!8] (0,0) rectangle (3,2.5);699\draw[very thick,accent] (0,0)--(3,0);700\draw[fwd] (1.5,1.6)--(0.3,0) (1.5,1.6)--(2.7,0);