Partial Differential Equations in Higher Categories
Waves, diffusions and potentials as functors of domain, symmetry and time
A treatment, in categorical language, of the material of Walter A. Strauss, Partial Differential Equations: An Introduction (2nd ed., Wiley, 2008): the three archetypal equations and the methods of an introductory course, namely characteristics, reflection, separation of variables, Green's functions, transforms, finite differences, and a sample of nonlinear phenomena. It reproduces none of the source's text, figures or exercises; a concordance maps its sections to this volume.
Format. Book, 98 pages, 15 chapters in five parts, 39 diagrams and figures (22 of them computed plots and schematics), 56 exercises with solutions to all of them, 75 references, status tags on every numbered statement (classical, literature, reformulation, observation, programme), and a companion script with 107 computational checks.
The organizing question
A differential equation is treated as an object and its solutions as the values of a functor $\mathrm{Sol}$ on a category of domains. The question asked of every method in the source is: for which maps is it natural? The first answer is the method of images. For a homogeneous wave, heat or Laplace equation, if a reflection group $W$ has fundamental chamber $C$ and preserves the equation, then for each character $\chi\colon W\to\{\pm1\}$
$$\begin{CD} \mathrm{Sol}(\mathbb R^n)^{\chi} @>{\text{evolve}}>> \mathrm{Sol}(\mathbb R^n)^{\chi} \\ @V{\text{restrict}}V{\cong}V @V{\cong}V{\text{restrict}}V \\ \mathrm{Sol}_\chi(C) @>>{\text{evolve}}> \mathrm{Sol}_\chi(C) \end{CD}$$
commutes, where $\mathrm{Sol}_\chi(C)$ has the Dirichlet condition on the walls whose reflection has $\chi=-1$ and the Neumann condition on the others. A boundary condition is a character; in general it is a Lagrangian plane of boundary values.
Results: qualitative statement, then its quantitative form
| Direction of functoriality | Qualitative | Quantitative (recomputed by the script) |
|---|---|---|
| Restriction | wave solutions on a diamond are determined by data on its waist; harmonic functions on an annulus carry a class in $H^1$ | changing data outside the interval of dependence changes nothing, to the last bit; periods $2\pi,0,0$ for $\log r$, $r^{\pm n}\cos n\theta$ |
| Reflection | boundary conditions are characters of a reflection group; images and eigenfunctions are exchanged by Pontryagin duality (Poisson summation) | for accuracy $10^{-10}$ on an interval: $0,1,4$ shells of images against $56,8,1$ eigenfunctions at $t=0.001,0.05,1$ |
| Time | waves: a group of symplectic maps; diffusions: a semigroup of Markov kernels | energy drift $5.5\times10^{-12}$; backward heat amplification $1.1$, $2.2\times10^{4}$, $1.2\times10^{39}$ for modes $n=1,10,30$ over time $0.1$ |
| Well-posedness | data are a restriction map; mismatched pairings fail | Hadamard: data $2\times10^{-9}$, solution $10^{162}$ |
| Boundary values | self-adjoint conditions are Lagrangian planes; intervals compose by transfer matrices | Robin eigenvalue $-2.3821$; Gibbs constant $0.58949$ (overshoot $8.95\%$) |
| Symmetry | eigenspaces are representations; special functions are matrix coefficients | hydrogen levels $-1/n^2$; Bessel zeros $2.4048, 3.8317, 5.1356, 5.5201$; nodal counts $1,2,2,4,3,3,6,4$ |
| Dimension | Huygens' principle is not inherited by invariants under a non-compact group | three dimensions: signal $0$ before and after; two dimensions: tail $1.29, 1.002, 1.0000$ times $1/t$ |
| Inclusion | Dirichlet eigenvalues are monotone in the domain, Neumann ones are not | Weyl counts $208, 869, 3531$ against $207.9, 865.9, 3532.5$ |
| Isospectrality | Sunada: equal spectra from isomorphic permutation representations | Gassmann triple of order 32; exactly two Laplacian spectra shared by connected six-vertex graphs with different degree sequences, none on five |
| Gluing | Dirichlet-to-Neumann maps and transfer matrices compose as Lagrangian relations | glued networks agree to ten digits; $\lvert R\rvert^2+\lvert T\rvert^2=1$ to twelve |
| Discretization | a scheme is a square that almost commutes; stability is positivity, causality, or a cochain map | growth $1.068$ per step at $s=0.52$; $1.8633$ beyond the CFL bound; sweeps $585, 298, 41$ for Jacobi, Gauss-Seidel, SOR; Betti numbers $(1,1,0)$ and $(1,3,3,1)$ |
| Morphisms of equations | Cole-Hopf and inverse scattering carry nonlinear flows to linear ones | breaking time $1.1658$; soliton shifts $1.8329$, $-2.9327$; eigenvalues $-0.6400$, $-0.2500$ conserved |
| Gauge | fields form a groupoid; the solution functor is a stack | three independent flat potentials on the discrete 3-torus |
Three points of contact with the source
1. The source describes a reflection formula for one-dimensional waves. For the reflection-invariant linear problems treated here, the method of images extends to diffusion and to the chambers of reflection groups, including half-spaces, slabs, wedges of angle $\pi/m$, boxes and three triangles. The construction uses the tiling of space by reflected chambers. 2. Its Gibbs constant "0.59" and "9 percent" are $0.58949$ and $8.95$ per cent. 3. In Section 13.1 the text gives the continuity equation with the opposite sign of $\nabla\cdot J$ to that in Exercise 1 of the same section; the exercise has the correct one.
How much is "higher"
Less than the title suggests, and the closing chapter says so chapter by chapter. Most of the volume is ordinary category theory (functors, naturality squares, relations), representation theory and sheaf theory. Cohomology enters in degree one: the annulus, the Aharonov-Bohm potentials, the Betti numbers of a mesh. The one place where the solution functor must take values in groupoids rather than sets is gauge theory, where gauge transformations are data; the source, which works with the fields $E$ and $B$, never needs it. The jet comonad and the category of differential equations (Marvan, Vinogradov, Khavkine-Schreiber) are stated and used as language, not developed.
What is and is not claimed
No new theorem about partial differential equations, and no existence theorem is proved: the analysis is cited. The observations of the volume are reformulations, among them the reading of the source's comparison table as symplectic group against Markov semigroup, the count of image shells against eigenfunctions, and the computed isospectral examples. Proofs marked as sketches are sketches. The source was read from its text layer; its figures were not consulted except for one page.
Files
partial-differential-equations-in-higher-categories.pdf— the book (98 pages).partial-differential-equations-in-higher-categories.tex— complete LaTeX source as a single file (pdflatex, three passes).verify-pde.py— 107 checks, Python standard library only, about five seconds; exits 0 if and only if all pass.
The companion paper Partial Differential Equations in Machine Learning and AI: Naturality, Well-Posedness, and Spectral Rates applies the framework to convolutional padding, diffusion models, graph networks, residual networks and physics-informed losses.
Status and authorship
Contribution. An exposition: known mathematics reorganized around naturality, with exact constants, solved exercises and reproducible computations.
Limitations. As above; in addition, the bibliography was assembled from memory except for two entries whose details were checked online on 17 November 2026, and nothing has been independently refereed. The script's checks are evidence that the numbers are stated correctly and are not proofs. Corrections are welcome.
AI assistance. The text was drafted and typeset with a large language model (Anthropic's Claude) under the direction of the submitting author, who is responsible for it. Generative AI assisted with drafting, literature recall, mathematical consistency checks, computation and typesetting.