Independent scholarship · Open accessThe monthly digest

Indexed text attachment

algebraic-and-differential-topology-in-machine-learning-and-ai.tex

Source for Algebraic and Differential Topology in Machine Learning and AI: Obstructions, Invariants, and Counts

View paper · Download original file

Lines 101–200 of 1055. Source text is displayed literally.

101 102\section{Introduction}103 104The companion volume \cite{ADTHC} retells two monographs of Pontryagin: one on the homology of polyhedra, ending105with the fixed point theorem of Lefschetz and Hopf, and one on the classification of maps of spheres by framed106manifolds. This paper collects what those theories say about learning systems. The connection is direct. A107feed-forward network is a continuous map; with rectifier activations it is piecewise linear, a simplicial map108after subdivision \cite{ABMM18}. Data are assumed to lie near a manifold or a polyhedron109\cite{Car09,EdHa10}. Latent codes, rotation heads, feature fields on meshes and recurrent attractors are maps to110and from spaces whose topology is known. Theorems about continuous maps then become statements about111architectures, and they have a particular character: they hold for every choice of weights.112 113We sort the consequences by the kind of statement.114 115\emph{Obstructions} say that something cannot be done continuously: a manifold cannot be coded in too few116dimensions, a rotation cannot be named by four numbers, a sphere cannot be combed, a narrow network cannot117enclose a region \cite{ZBLYL19,CWKW19,FdH18,Joh19}.118 119\emph{Invariants} are what a functor computes. Simplicial and Hodge-theoretic networks120\cite{EDS20,BFW21,Lim20,PBB24} compute homology with real coefficients; persistent homology121\cite{CEH07,NSW08} computes it across scales, where the complexes themselves have topology of their own122\cite{AdAd17}.123 124\emph{Counts} are forced numbers: of points where classes meet, of fixed points of a recurrent map, of preimages125of a value.126 127\paragraph{What is and is not claimed.} Every statement carries a status tag: classical, literature,128reformulation, observation or programme. The theorems are classical, and several of the applications are in the129literature cited; where a formulation or a constant seems not to be standard we say ``observation''. Tags such as130\chk{O4} name checks in the companion script \texttt{verify-topology-learning.py} (\cref{app:checks}), which uses131only the Python standard library, exact arithmetic for homology, and a fixed seed. The checks are instances and132not proofs. The models are small and exact; no trained network appears. The references were assembled from133memory, except \cite{ZBLYL19,AdAd17,PBB24}, which were checked online on 19 November 2026.134 135\section{Setting}\label{sec:setting}136 137A \term{network} is a continuous map $f_w\colon\R^a\to\R^b$ depending continuously on weights $w$. Data lie on a138compact subset $M\subset\R^a$. A \term{code} for $M$ in dimension $m$ is a continuous \emph{encoder} $e\colon M\to\R^m$;139a \term{decoder} is any map $d\colon\R^m\to\R^a$, and the \term{worst-case error} of the pair is140$\sup_{x\in M}|d(e(x))-x|$. A \term{representation} of a space $Y$ in $\R^m$ is a continuous map $g\colon\R^m\supseteq141D\to Y$ onto $Y$; it is \term{continuous} in the sense of \cite{ZBLYL19} if it has a continuous right inverse142$Y\to D$, so that a target in $Y$ can be turned into a continuous regression target.143 144The tools are the functors of \cite{ADTHC}: homology $H_*$, which sends a continuous map to a homomorphism and a145homotopy to an equality; the degree of a map between spheres; the fundamental group; and the Lefschetz number146$L(f)=\sum_k(-1)^k\tr(f_*\mid H_k)$. An obstruction is a diagram that such a functor forbids. The simplest is the147one that proves Brouwer's theorem \cite[Thm.~4.2]{ADTHC}:148\[\begin{tikzcd}[column sep=small]149S^{n-1}\arrow[r,hook]\arrow[rr,bend right=28,"\id"'] & D^n\arrow[r,"r"] & S^{n-1} &[-0.6em]\leadsto&[-0.6em]150\Z\arrow[r]\arrow[rr,bend right=28,"\id"'] & 0\arrow[r] & \Z151\end{tikzcd}\]152A retraction of the ball onto its boundary would factor the identity of $\Z$ through $0$.153 154\section{Obstructions}\label{sec:obstructions}155 156\subsection{Bottlenecks}157 158\begin{theorem}[Bottleneck bound]\label{thm:bottleneck}159\Srf[Borsuk--Ulam \cite{Mat03}] Let $M$ be the unit sphere $S^n\subset\R^{n+1}$, $e\colon S^n\to\R^n$ continuous and160$d\colon\R^n\to\R^{n+1}$ arbitrary. Then161\[ \sup_{x\in S^n}|d(e(x))-x|\ \ge\ 1, \]162and equality holds for $e$ the projection to the first $n$ coordinates and $d$ the inclusion.163\end{theorem}164\begin{proof}165By the Borsuk--Ulam theorem there is $x$ with $e(x)=e(-x)$. Let $y$ be the common reconstruction. Then166$|y-x|+|y+x|\ge|2x|=2$, so one of the two errors is at least $1$. For the projection the error is167$|x_{n+1}|\le1$.168\end{proof}169 170\begin{figure}[t]171\centering172\begin{tikzpicture}173\begin{scope}[shift={(1.5,2.0)}]174\draw[ink2,thick] (0,0) circle (1.25);175\draw[ink2,densely dotted] (-1,-0.75)--(-1,0.75);176\fill[ink2] (-1,-0.75) circle (0.9pt); \fill[ink2] (-1,0.75) circle (0.9pt);177\draw[ink2,densely dotted] (-0.5,-1.146)--(-0.5,1.146);178\fill[ink2] (-0.5,-1.146) circle (0.9pt); \fill[ink2] (-0.5,1.146) circle (0.9pt);179\draw[ink2,densely dotted] (0.5,-1.146)--(0.5,1.146);180\fill[ink2] (0.5,-1.146) circle (0.9pt); \fill[ink2] (0.5,1.146) circle (0.9pt);181\draw[ink2,densely dotted] (1,-0.75)--(1,0.75);182\fill[ink2] (1,-0.75) circle (0.9pt); \fill[ink2] (1,0.75) circle (0.9pt);183\draw[accent,thick] (0,-1.25)--(0,1.25);184\fill[accent] (0,1.25) circle (1.6pt) node[above,inner sep=2pt]{\scriptsize $x$}; \fill[accent] (0,-1.25) circle (1.6pt) node[below,inner sep=2pt]{\scriptsize $-x$};185\fill[bwd] (0,0.28) circle (1.6pt) node[right,inner sep=2pt]{\scriptsize $y$};186\draw[ink2,-{Stealth[length=3pt]}] (-1.35,-1.87)--(1.5,-1.87);187\node[right,inner sep=1pt] at (1.52,-1.87) {\scriptsize $\R$};188\fill[ink2] (-1,-1.87) circle (1.1pt);189\fill[ink2] (-0.5,-1.87) circle (1.1pt);190\fill[accent] (0,-1.87) circle (1.1pt);191\fill[ink2] (0.5,-1.87) circle (1.1pt);192\fill[ink2] (1,-1.87) circle (1.1pt);193\node[left,inner sep=1pt] at (-1.37,-1.87) {\scriptsize $e$};194\node at (0,-2.33) {\scriptsize (a)};195\end{scope}196\begin{scope}[shift={(4.3,0.75)}]197\draw[ink2] (0,2.7)--(0,0)--(3.6,0);198\draw[ink2] (0,0)--(0,-0.07) node[below,inner sep=1.5pt]{\scriptsize $0$};199\draw[ink2] (3.6,0)--(3.6,-0.07) node[below,inner sep=1.5pt]{\scriptsize $\pi$};200\node[below] at (1.8,-0.38) {\scriptsize $\theta$};