Indexed text attachment
differential-topology-of-parametrized-neural-networks.tex
Source for Differential Topology of Parametrized Neural Networks
View paper · Download original file
Lines 694–793 of 1074. Source text is displayed literally.
694 S^1\times\mu_k \arrow[r,"h"] \arrow[d,"\operatorname{pr}_1"'] &695 B^+ \arrow[d,"\pi"]\\696 S^1 \arrow[r,"r"'] & P697\end{tikzcd}698\]699Here $h(\alpha,\zeta)=(\alpha^k,\alpha\zeta)$. Choosing a700fixed $k$th root of $-1$ gives the corresponding701trivialization for $B^-$. More generally, pullback along702$\alpha\mapsto\alpha^r$ for a positive integer $r$ replaces703the permutation by $j\mapsto j+r$. Its cycles have length704$k/\gcd(k,r)$, and there are $\gcd(k,r)$ of them. Thus the705cover with $r=k$ is the least positive winding cover that706permits a global labeling of all the sheets. In each fiber707the signed count is $k-k=0$. The transport still gives two copies of the cyclic708permutation representation. This persists on a range of709perturbed networks. Suppose a smooth perturbed score710$\widetilde s$ satisfies, for some $0<\eta<1$,711\[712 \|\widetilde s-s\|_\infty<\eta,\qquad713 \|\partial_\varphi(\widetilde s-s)\|_\infty714 < k\sqrt{1-\eta^2}.715\]716At a zero of any linear interpolation between the scores717we have $|s|<\eta$, hence718$|\partial_\varphi s|\geq k\sqrt{1-\eta^2}$.719Thus zero remains regular in the input direction and its720derivative sign is preserved at the interpolated zeros.721The relative zero set is a proper local diffeomorphism722over the cylinder of interpolation times and parameters.723Path lifting along the interpolation identifies the two724signed covers at the ends of the cylinder.725 726\subsection{Persistence over the parameter space}727 728Use the closed sublevels729$F_p(t)=\{z:s(p,z)\leq t\}$. At $t<-1$ the fiber is empty.730At $t=-1$ it consists of the $k$ minima. For $-1<t<1$731it is a disjoint union of $k$ closed arcs, each containing732one minimum. At $t=1$ it becomes the whole circle.733Over a field $\mathbb K$, the degree-zero barcode therefore734has one copy of $[-1,\infty)$ and $k-1$ copies of $[-1,1)$.735Degree one has one copy of $[1,\infty)$. These statements736follow because inclusions before $1$ preserve the components,737while the inclusion into the circle sends all component738classes to the same class. Each $p$ has the same barcodes.739 740The total sublevels741$\mathcal F(t)=\{(p,z):s(p,z)\leq t\}\to P$, with their742inclusions for increasing $t$, define a diagram of spaces743over $P$. A local lift $\alpha(p)^k=p$ simultaneously744trivializes the diagram by $z=\alpha(p)u$, since the score745becomes $\operatorname{Im}(u^k)$. The transitions rotate746$u$ by elements of $\mu_k$, and transport around the base747permutes the $k$ components for every $-1\le t<1$. Fiberwise748$H_0$ is therefore a local system with permutation module749$\mathbb K[\mu_k]$ in that range. At $t=1$ the structure map750is the augmentation $\mathbb K[\mu_k]\to\mathbb K$, where751$\mathbb K$ carries trivial transport. Its kernel gives752the local system of reduced degree-zero classes that die.753If the characteristic divides $k$, this augmentation has754no equivariant section. An invariant element is a multiple of755the sum of the $k$ basis vectors, whose augmentation is zero.756This extension is invisible in the barcode of each fiber,757where the augmentation is a surjective map of vector spaces.758Over $\mathbb C$, let $\omega=e^{2\pi i/k}$, with $e_j$ the759component basis of $H_0$. The character vector760$\sum_j\omega^{-\ell j}e_j$ acquires the factor761$\omega^\ell$ under monodromy. The trivial character survives762the merge at $1$ and each nontrivial character dies there.763The $k-1$ numerically equal finite bars consequently carry764distinct transport characters.765 766For parameters $p,q$, the chord distance $\delta=|p-q|$767bounds the uniform score change, since768\[769 |s(p,z)-s(q,z)|770 \leq |\overline p-\overline q|\,|z^k|=\delta.771\]772Thus $F_p(t)\subseteq F_q(t+\delta)$ and773$F_q(t)\subseteq F_p(t+\delta)$. The successive inclusion774maps are exactly the shift by $2\delta$, giving a775$\delta$-interleaving both before and after taking homology.776The same proof applies to any uniform score perturbation.777For q-tame persistent homology the stability theorems bound778the bottleneck distance by that uniform change779\cite{chazal-persistence,cohen-steiner-stability}.780In this rotation family the distance between fiber barcodes781is zero. The sublevel diagram over $P$ additionally records782which components are identified by going around a loop.783For example, the constant family of scores784$s_0(p,z)=\operatorname{Im}(z^k)$ has the same barcode at785every parameter. Its component local system is trivial,786whereas the generator of the parameter loop acts on the787components for $s$ by the nonidentity permutation.788The two filtered diagrams therefore cannot be equivalent789over $P$. For two general scores on the same parameter790and input spaces, a uniform bound on their difference791also gives the shifted inclusions of the \emph{total}792sublevels over $P$. This makes the sublevel diagram over793the parameter space a stable comparison object while