Independent scholarship · Open accessThe monthly digest

Indexed text attachment

differential-topology-in-higher-categories.tex

Source for Differential Topology in Higher Categories

View paper · Download original file

Lines 4941–5040 of 6119. Source text is displayed literally.

4941 4942 4943For compact $M$ the embedding locus is open in the $C^1$4944topology. To see this, let $f$ be an embedding. At each4945point choose coordinates on a small source ball and on4946the target such that projection of the coordinate image4947of $f$ to its first $m$ coordinates has invertible4948derivative at the center. Shrink the source ball until4949this derivative is uniformly close to its center value.4950For every sufficiently $C^1$-close map $g$, the projected4951coordinate derivative stays close to the same invertible4952linear map. Integrating along line segments in the ball4953then bounds the distance between the projected images of4954distinct points below by a positive multiple of their4955coordinate distance. This gives injectivity and full rank4956on that ball. Finitely many such balls cover $M$.4957Pairs sharing a ball constitute a neighborhood of the4958diagonal in $M\times M$. On the compact remaining set of4959pairs, the distances between their $f$-images have a4960positive minimum. A small uniform perturbation keeps4961those images distinct as well. This proves openness4962and explains how local derivative margins and global4963pair separation give different stability conditions.4964 4965\begin{proposition}[Generic embedding slices of a family]4966\label{prop:generic-embedding-slices}4967Let $M$ be compact of dimension $m$ and let4968$f:P\times M\to N^n$ be a smooth family, with $P,M,N$4969finite-dimensional, second-countable manifolds without4970boundary. Assume $n>2m$. Suppose the vertical first jet4971$(p,x)\mapsto j^1(f_p)(x)$ is transverse to each4972rank-deficient stratum in $J^1(M,N)$, and the map4973\[4974 (p,x,y)\longmapsto(f_p(x),f_p(y))\colon4975 P\times M^{[2]}\longrightarrow N\times N4976\]4977is transverse to $\Delta_N$. Then $f_p$ is an embedding4978outside a set of parameters of measure zero in local charts.4979\end{proposition}4980 4981\begin{proof}4982For each total map take the inverse image of the4983specified stratum or diagonal, a smooth submanifold by4984transversality. Sard's theorem applied to its projection4985to $P$ gives a full-measure set of regular parameters.4986Regularity of this projection is equivalent to4987transversality in the slice at that parameter. To check4988the equivalence, write $A,B$ for the parameter and source4989parts of the derivative into the normal quotient to the4990specified submanifold. Total transversality means that4991$\operatorname{im}A+\operatorname{im}B$ is that quotient.4992Projection of the kernel onto the parameter tangent4993space is onto exactly when every $A$-image lies in4994$\operatorname{im}B$, which under total transversality4995is precisely surjectivity of $B$.4996 4997There are finitely many rank strata. The stratum of rank4998$k<m$ in the space of linear maps $\mathbb R^m\to\mathbb R^n$4999has codimension $(m-k)(n-k)$: in a chart where a5000$k$-minor is invertible, the remaining block is prescribed5001by its Schur complement. This also gives the corresponding5002codimension in jet coordinates. It is at least $n-m+1$,5003which exceeds $m$ under the stated inequality. Therefore5004a transverse slice avoids every deficient stratum.5005The diagonal has codimension $n>2m$, so its transverse5006preimage in the $2m$-dimensional deleted product is also5007empty. The slice is an injective immersion, and compactness5008gives the embedding.5009\end{proof}5010 5011One can verify the diagonal hypothesis by showing that the5012parameter derivatives can move the difference of the two5013outputs in all normal directions at each collision.5014This requires independent control at the pair under5015consideration. A transversality theorem for the unrestricted5016smooth mapping space gives no such control for a prescribed5017family of finite dimension. For example the family5018\[5019 f_p(t)=\bigl(p\cos(2t),p\sin(2t)\bigr),5020 \qquad p\in\mathbb R\setminus\{0\},5021\]5022immerses the circle for every $p$, and it identifies $t$5023with $t+\pi$ for every $p$. Embedding the targets in a5024higher-dimensional Euclidean space by zero padding leaves5025this obstruction intact. Adding two output coordinates5026$a\cos t,a\sin t$ with $a\ne0$ yields an embedding, since5027the added coordinates separate points and already form5028an immersion. In an architecture using trigonometric5029features, the choice of features therefore changes the5030class of topological realizations before the weights5031are selected.5032 5033\section{The configuration stages}5034 5035Let $\mathcal O(U)$ be the poset of open subsets of an5036open manifold domain $U\subset M$, and let5037$\mathcal O_k(U)$ be its subposet of opens diffeomorphic to5038unions of at most $k$ disjoint open balls.5039Embedding calculus studies the restriction functor5040$V\mapsto\operatorname{Emb}(V,N)$. Its approximations