Indexed text attachment
partial-differential-equations-in-higher-categories.tex
Source for Partial Differential Equations in Higher Categories
View paper · Download original file
Lines 1301–1400 of 3921. Source text is displayed literally.
1301\end{center}1302\noindent \chk{F3} The last row is the source's example of conditions for which the method fails. The first three1303are \emph{local}: $V$ is a sum of a line at $a$ and a line at $b$, and the lines at one endpoint form a circle, with1304Dirichlet and Neumann two of its points and the Robin constant a coordinate on the rest. The periodic condition is1305non-local: it is the graph of the identity between the two ends, which is what it means to glue them.1306 1307\begin{center}1308\begin{tikzpicture}1309\draw[->,ink2] (-2.4,0)--(2.6,0) node[right]{\scriptsize $X(a)$};1310\draw[->,ink2] (0,-2.0)--(0,2.2) node[above]{\scriptsize $X'(a)$};1311\draw[ink2,very thin] (0,0) circle (1.3);1312\draw[very thick,fwd] (0,-1.9)--(0,1.9);1313\draw[very thick,bwd] (-2.2,0)--(2.2,0);1314\draw[thick,accent] (-1.9,-1.1)--(1.9,1.1);1315\draw[thick,accent,densely dashed] (-1.5,1.6)--(1.5,-1.6);1316\node[fwd,left] at (-0.05,1.75) {\scriptsize Dirichlet: $X=0$};1317\node[bwd,below] at (1.45,-0.03) {\scriptsize Neumann: $X'=0$};1318\node[accent,right] at (1.9,1.1) {\scriptsize Robin: $X'=aX$, $a>0$};1319\node[accent,right] at (1.5,-1.6) {\scriptsize $a<0$};1320\foreach \p in {(0,1.3),(1.3,0),(1.125,0.651),(0.889,-0.948)} \fill \p circle (1.6pt);1321\end{tikzpicture}1322\figcap{The boundary conditions at one end of an interval are the lines through the origin in the plane of $(X,X')$, a circle's worth. The Dirichlet and Neumann conditions are two points of it and the Robin conditions fill in the rest; a separated condition for the interval is a point of the product of two such circles.}1323\end{center}1324 1325\begin{remark}[Intervals compose]\label{rem:compose}1326For fixed $\lambda$ the solutions of $-X''=\lambda X$ on $[a,b]$ have boundary values forming a Lagrangian plane1327in $\R^4$, the graph of the transfer matrix from $(X(a),X'(a))$ to $(X(b),X'(b))$, which has determinant one. Joining1328$[a,b]$ to $[b,c]$ multiplies transfer matrices, that is, composes Lagrangian relations. An eigenvalue is a1329$\lambda$ for which the plane of the equation meets the plane of the boundary conditions. This compositional1330picture returns for networks in \Cref{ch:green} and for scattering in \Cref{ch:fields}.1331\end{remark}1332 1333\[\begin{tikzcd}[column sep=4.2em]1334B_a\arrow[r,hook] & V_a\arrow[r,"M_{[a,b]}(\lambda)"]\arrow[rr,bend right=22,"M_{[a,c]}(\lambda)"'] & V_b\arrow[r,"M_{[b,c]}(\lambda)"] & V_c & B_c\arrow[l,hook']1335\end{tikzcd}\]1336Here $V_x=\R^2$ holds the pair $(X(x),X'(x))$ and a separated boundary condition is a line $B$ in it. The number1337$\lambda$ is an eigenvalue exactly when the line $M_{[a,c]}(\lambda)B_a$ is the line $B_c$: the shooting method, read as1338a composite of relations.1339 1340\paragraph{The Robin condition.} With $X'-a_0X=0$ at $0$ and $X'+a_lX=0$ at $l$, the quadratic form of $A$ is1341$\int X'^2+a_0X(0)^2+a_lX(l)^2$, which is positive when both constants are positive (radiation) and can be1342negative when they are negative (absorption). The source finds a zero eigenvalue exactly when1343$a_0+a_l=-a_0a_ll$ \src{\S4.3}. For $a_0=a_l=-1$ that is $l=2$: there is one negative eigenvalue for $l<2$ and two for1344$l>2$; at $l=1$ it is $-2.3821$. \chk{F2} A negative eigenvalue $-\gamma^2$ means that the diffusion equation has1345the growing solution $e^{k\gamma^2t}X(x)$: the boundary feeds the interior faster than diffusion can spread it.1346 1347\section{Why sines and cosines}1348 1349Take periodic conditions on $(-l,l)$: the domain is the circle $\mathbb T=\R/2l\Z$, a compact abelian group. The1350operator $A$ commutes with translations, so its eigenspaces are representations of $\mathbb T$, and the1351irreducible ones are the characters $e^{in\pi x/l}$, $n\in\Z$.1352 1353\begin{theorem}[Fourier series as duality]\label{thm:pontryagin}1354\Scl The characters of $\mathbb T$ form the dual group $\Z$, and the map $f\mapsto(\hat f(n))_{n\in\Z}$ is a unitary1355isomorphism $L^2(\mathbb T)\to\ell^2(\Z)$ that carries every translation-invariant operator to a multiplication1356operator; $A$ becomes multiplication by $(n\pi/l)^2$.1357\end{theorem}1358 1359\noindent The reflection $x\mapsto-x$ acts on $\mathbb T$ and exchanges $n$ and $-n$. Its odd part is spanned by the1360sines and its even part by the cosines, and by \Cref{thm:images} these are the Dirichlet and Neumann problems on1361$(0,l)$: the interval is the circle of twice its length modulo reflection. The sine and cosine series of the1362source \src{\S5.1--5.2} are the two isotypic components of the full Fourier series. On the finite group $\Z/n$ the1363whole statement is a computation: the characters diagonalize the periodic difference Laplacian with eigenvalues1364$4\sin^2(\pi k/n)$, and their even and odd parts under reflection are the eigenvectors of the Neumann and Dirichlet1365matrices. \chk{F7} For the function $x$ on $(0,l)$ Parseval's identity gives $\sum n^{-2}=\pi^2/6$. \chk{F4}1366 1367\section{Completeness, and in which category}1368 1369The source distinguishes uniform, mean-square and pointwise convergence and proves a theorem for each1370\src{\S5.4}.1371 1372\begin{theorem}\label{thm:complete}1373\Scl[source, \S5.4, Theorems 2--6] For a symmetric boundary condition: \emph{(a)} the eigenfunction expansion of every1374$f\in L^2$ converges to $f$ in mean square, and Parseval's equality holds; \emph{(b)} if $f$ is $C^2$ and satisfies the1375boundary conditions the convergence is uniform; \emph{(c)} for the classical series, if $f$ and $f'$ are piecewise1376continuous the series converges at each point to the mean of the one-sided limits. The partial sum is the best1377approximation in mean square among combinations of the first $N$ eigenfunctions.1378\end{theorem}1379 1380\noindent Statement (a) says that the coefficient map is an isomorphism in the category of Hilbert spaces. It is not1381an isomorphism of the Banach space of continuous functions onto a sequence space: there are continuous functions1382whose Fourier series diverge at a point. Completeness, like well-posedness (\Cref{def:wellposed}), is a1383statement about a morphism in a specified category. The Gibbs phenomenon is the visible form of the difference:1384near a unit jump the partial sums overshoot to $\tfrac1\pi\int_0^\pi\frac{\sin\theta}{\theta}d\theta=0.58949$ where the1385function is $\tfrac12$, an excess of $8.95$ per cent of the jump that does not decrease with $N$ (the source's ``91386percent''); the peaks at $N=51$, $201$, $801$ are $0.58955$, $0.58949$, $0.58949$. \chk{F5} The series converges in1387$L^2$ and not uniformly.1388 1389\begin{center}1390\begin{tikzpicture}1391\begin{scope}[shift={(0,0)}]1392\draw[ink2] (0,4)--(0,0)--(8.6,0);1393\draw[ink2] (1.544,0)--(1.544,-0.07) node[below,inner sep=1.5pt]{\scriptsize $0$};1394\draw[ink2] (3.749,0)--(3.749,-0.07) node[below,inner sep=1.5pt]{\scriptsize $0.5$};1395\draw[ink2] (5.954,0)--(5.954,-0.07) node[below,inner sep=1.5pt]{\scriptsize $1$};1396\draw[ink2] (8.159,0)--(8.159,-0.07) node[below,inner sep=1.5pt]{\scriptsize $1.5$};1397\draw[ink2] (0,0.667)--(-0.07,0.667) node[left,inner sep=1.5pt]{\scriptsize $-0.5$};1398\draw[ink2] (0,2)--(-0.07,2) node[left,inner sep=1.5pt]{\scriptsize $0$};1399\draw[ink2] (0,3.333)--(-0.07,3.333) node[left,inner sep=1.5pt]{\scriptsize $0.5$};1400\node[below] at (4.3,-0.38) {\scriptsize $x$};