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 15–114 of 1055. Source text is displayed literally.

15\usepackage{microtype}16\usepackage{xcolor}17\definecolor{accent}{HTML}{7A1F2B}18\definecolor{ink2}{HTML}{55595F}19\definecolor{fwd}{HTML}{1F4E79}20\definecolor{bwd}{HTML}{5B3A7A}21\usepackage{tikz,tikz-cd}22\usetikzlibrary{arrows.meta,positioning}23\usepackage{booktabs,array}24\usepackage{enumitem}25\usepackage[font=small,labelfont={bf,sf}]{caption}26\usepackage[colorlinks=true,linkcolor=accent,citecolor=accent,urlcolor=accent,27  pdftitle={Algebraic and Differential Topology in Machine Learning and AI: Obstructions, Invariants, and Counts},28  pdfauthor={Anonymous},29  pdfsubject={Topological obstructions, invariants and counting results for neural networks},30  pdfkeywords={algebraic topology, machine learning, neural networks, Borsuk-Ulam theorem, persistent homology, Hodge Laplacian, topological degree}]{hyperref}31\usepackage[nameinlink,capitalise,noabbrev]{cleveref}32\theoremstyle{plain}33\newtheorem{theorem}{Theorem}[section]34\newtheorem{proposition}[theorem]{Proposition}35\newtheorem{corollary}[theorem]{Corollary}36\newtheorem{lemma}[theorem]{Lemma}37\theoremstyle{definition}38\newtheorem{definition}[theorem]{Definition}39\newtheorem{example}[theorem]{Example}40\theoremstyle{remark}41\newtheorem{remark}[theorem]{Remark}42\newcommand{\stag}[2]{{\footnotesize\sffamily\upshape\color{#1}[#2]}\enspace\ignorespaces}43\newcommand{\Scl}[1][]{\stag{ink2}{classical\if\relax\detokenize{#1}\relax\else; #1\fi}}44\newcommand{\Sli}[1][]{\stag{fwd}{literature\if\relax\detokenize{#1}\relax\else; #1\fi}}45\newcommand{\Srf}[1][]{\stag{accent}{reformulation\if\relax\detokenize{#1}\relax\else; cf.\ #1\fi}}46\newcommand{\Sob}{\stag{accent}{observation}}47\newcommand{\Spg}{\stag{bwd}{programme}}48\newcommand{\chk}[1]{{\footnotesize\sffamily\upshape\color{accent}[#1]}}49\newcommand{\cat}[1]{\mathsf{#1}}50\newcommand{\R}{\mathbb{R}}51\newcommand{\Z}{\mathbb{Z}}52\newcommand{\E}{\mathbb{E}}53\newcommand{\one}{\mathbf{1}}54\newcommand{\T}{\mathsf{T}}55\newcommand{\Lap}{\Delta}56\newcommand{\pd}{\partial}57\newcommand{\KL}[2]{D(#1\,\Vert\,#2)}58\newcommand{\Var}{\operatorname{Var}}59\newcommand{\relu}{\operatorname{relu}}60\newcommand{\SO}{\mathrm{SO}}61\newcommand{\RP}{\mathbb{RP}}62\newcommand{\Fp}{\mathbb{F}}63\newcommand{\lk}{\operatorname{lk}}64\newcommand{\tr}{\operatorname{tr}}65\newcommand{\id}{\mathrm{id}}66\newcommand{\sinc}{\operatorname{sinc}}67\newcommand{\term}[1]{\textbf{\itshape #1}}68\setlength{\parskip}{0pt}69\frenchspacing70\raggedbottom71\title{\textbf{Algebraic and Differential Topology\\ in Machine Learning and AI:\\[2pt] Obstructions, Invariants, and Counts}}72\author{Anonymous}73\date{Higher Categorical Semantics \quad$\cdot$\quad Version 1.0, 19 November 2026}74\begin{document}75\twocolumn[76\maketitle77\begin{center}\begin{minipage}{0.86\textwidth}\small78\textbf{Abstract.} A neural network is a continuous map, usually piecewise linear, and the classical theorems of79algebraic and differential topology are statements about what continuous maps can and cannot do. We organize80their consequences for machine learning under three headings, pairing each qualitative statement with a computed81one. \emph{Obstructions}: by the Borsuk--Ulam theorem an autoencoder for data on an $n$-sphere with an82$n$-dimensional bottleneck has worst-case reconstruction error at least the radius, and this is attained; a83network output that names a rotation by a unit quaternion jumps by $2$ somewhere on every non-contractible loop84of rotations, because $\pi_1(\SO(3))=\Z/2$, while a six-dimensional code does not; every tangent feature field on a85sphere-like mesh has singularities of total index $2$; and a closed piecewise linear curve with $N$ pieces winds86fewer than $N/2$ times, while composition multiplies degrees, so that depth $L$ reaches degree $2^L$ with a87fixed number of pieces per layer; and a network no wider than its input, with an injective activation, cannot88enclose a bounded class.89\emph{Invariants}: diffusion with the Hodge Laplacian of a complex converges to a space whose dimension is a90Betti number, at the rate of the spectral gap, and with real coefficients it cannot tell the projective plane91from a disc; the Vietoris--Rips complex of twelve points on a circle is a circle up to a third of a turn, then92three $2$-spheres, then a $5$-sphere; and persistence diagrams move by at most the size of a perturbation, which93the count of their points does not. \emph{Counts}: a classifier of the $n$-cube with small decision regions has points where $n+1$ of them meet ($13$ to $18$ triple points for random prototype classifiers of the square);94the fixed points of an attractor network on a circle or a torus have total index $1-d$ and $0$; the degree of95a normalized output changes only when the output passes through zero. A scalar odd network on an origin-centred96circle has $2,6,10,\dots$ zeros if all of them are simple. All numerical statements are recomputed97by a companion script with $25$ checks. No new theorem is claimed.98\end{minipage}\end{center}99\vspace{1.2em}100]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