Loading [MathJax]/jax/output/HTML-CSS/jax.js

Monday, April 16, 2018

Conical stratifications via semialgebraic sets

The goal of this post is to describe a conical stratification of Rann(M)×R0 that refines the stratification previously seen (in "Exit paths, part 2," 2017-09-28, and "Refining stratifiations," 2018-03-11). Thanks to Shmuel Weinberger for the key observation that the strata under consideration are nothing more than semialgebraic sets, which are triangulable, and so admit a conical stratification via this triangulation.

Remark: Fix nZ>0, let M be a smooth, compact, connected, embedded submanifold in RN, and let Mn have the Hausdorff topology. We will be interested in Mn×R>0, though this will be viewed as the compact set Mn×[0,K]RnN+1 for some K large enough (for instance, larger than the diameter of M) when necessary. The point 0 is added for compactness.

Stratification of the Ran space by semialgebraic sets


We begin by stratifying Mn×R>0 by a poset A, creating strata based on the pairwise distance between points in each M component. Then we take that to a stratification of the quotient Rann(M)×R>0 via the action of the symmetric group Sn and overcounting of points.

Definition: Define a partial order on the set A={partitions of ({1,,n}2Δ)/S2 into 4 parts} of ordered 4-tuples of sets by (Q,R,S,T)(QQ, RQS, S(SS), TS), for all QQ and S,SS, with SS=.

The diagram to keep in mind is the one below, with arrows pointing from lower-ordered elements to higher-ordered elements. Once we pass to valuing the 4-tuple in simplicial complexes, moving between Q and R will not change the simplicial complex type (this comes from the definition of the Vietoris--Rips complex).

Lemma 1: The map f:Mn×R>0(A,) defined by ({P1,,Pn},t)({(i,j>i) : Pi=Pj}, {(i,j>i) : dM(Pi,Pj)<t},{(i,j>i) : dM(Pi,Pj)=t}, {(i,j>i) : dM(Pi,Pj)>t}) is continuous in the upset topology on (A,).

Proof: Choose (Q,R,S,T)A and consider the open set U=U(Q,R,S,T) based at (Q,R,S,T). Take (P,t)f1(U), which we claim has a small neighborhood still contained within f1(U). If we move a point Pi slightly that was exactly distance t away from Pj, then the pair (i,j) was in S, but is now in either R or T, and both (Q,R{(i,j)},S{(i,j)},T) and (Q,R,S{(i,j)},T{(i,j)}) are ordered higher than (Q,R,S,T), so the perturbed point is still in f1(U). If Pi=Pj in P and we move them apart slightly, since tR>0, the pair (i,j) will move from Q to R, and (Q,R,S,T)(Q{(i,j)},R{(i,j)},S,T), so the perturbed point is still in f1(U). For all pairs (i,j) in R or T, the distances can be changed slightly so that the pair still stays in R or T, respectively. Hence f is continuous.

This shows that Mn×R>0 is stratified by (A,), using Lurie's definition of a (poset) stratification, which just needs a continuous map to a poset. Our goal is to work with the Ran space of M, instead of the n-fold product of M, which are related by the natural projection map π:MnRann(M), taking P={P1,,Pn} to the unordered set of distinct elements in P. We also would like to stratify Rann(M)×R>0 by simplicial complex type, so we need the following map.

Definition: Let g:(A,)SC be the map into simplicial complexes that takes (Q,R,S,T) to the clique complex of the simple graph C on nk vertices, for |Q|=k(k+1)/2, defined as follows: 
  • V(C)={[i] : i=1,,n, [j]=[i]\ iff\ (i,j)Q},
  • E(C)={([i],[j]) : (i,j)RS}.
We require C to be simple, so if (i,j)Q and (i,),(j,)RS, we only add one edge ([i],[])=([j],[]) to C.

The map g induces a partial order on SC from the partial order on A, with CC in SC whenever there is (Q,R,S,T)g1(C) and (Q,R,S,T)g1(C) such that (Q,R,S,T)(Q,R,S,T) in A. Note that if CSC is not in the image of g, then it is not related to any other element of SC. By the universal property of the quotient and continuity of f and g (as A and SC are discrete), there is a continuous map h:Rann(M)×R>0(SC,) such that the diagram
commutes. Hence Rann(M)×R>0 is stratified by (SC,).

Remark: The map π can be thought of as a quotient by the action of the symmetric group Sn, followed by the quotient of the equivalence relation {P11,,P11,P12,,P22,P13,,Pkk}    {P11,,P111,P12,,P2+12,P13,,Pkk} on Mn, for all possible combinations 1++k=n and 1kn1, where Pim=Pjm for all 1i<jm.


Semialgebraic geometry


Next we move into the world of semialgebraic sets and triangulations, following Shiota. Here we come across a more restrictive notion of stratification of a manifold X, which requires a partition of X into submanifolds {Xi}. If Lurie's stratification f:XA gives back submanifolds {f1(a)}aA, then we have Shiota's stratification. Conversely, the poset ({Xi},), for XiXj iff Xicl(Xj) is always a stratification in the sense of Lurie.

Definition 2: A semialgebraic set in RN is a set of the form finite{xRN : f1(x)=0,f2(x)>0,,fm(x)>0}, for polynomial functions f1,,fm on RN. A semialgebraic stratification of a space XRN is a partition {Xi} of X into submanifolds that are semialgebraic sets.

Next we observe that the strata of Mn×R>0 are semialgebraic sets, with the preimage theorem and I.2.9.1 of Shiota, which says that the intersection of semialgebraic sets is semialgebraic. Take (Q,R,S,T)A  and note that f1(Q,R,S,T)={({P1,,Pn},t)Mn×R>0 : d(Pi,Pj)=0(i,j)Q,td(Pi,Pj)=0(i,j)S,td(Pi,Pj)>0(i,j)R,d(Pi,Pj)t>0(i,j)T.} Here d means distance on the manifold, and we assume the metric to be analytic. Alternatively, d could be Euclidean distance between points on the embedding of Mn×R>0, induced by the assumed embedding of M.

For his main Theorem II.4.2, Shiota uses cells, but we opt for simplices instead, and for cell complexes we use simplicial complexes. Every cell and cell complex admits a decomposition into simplicial complexes, even without introducing new 0-cells (by Lemma I.3.12), so we do not lose any generality.

Definition 3: Let X,Y be semialgebraic sets.
  • A map f:XY is semialgebraic if the graph of f is semialgebraic.
  • A semialgebraic cell triangulation of a semialgebraic set X is a pair (C,π), where C is a simplicial complex and π:|C|X is a semialgebraic homeomorphism for which π|int(σ) is a diffeomorphism onto its image.
  • A semialgebraic cell triangulation (C,π) is compatible with a family {Xi} of semialgebraic sets if π(int(σ))Xi or π(int(σ))Xi= for all σC and all Xi.

A semialgebraic cell triangulation (C,π) of X induces a stratification X(C0{π(int(σ))},), where the order is the one mentioned just before Definition 2. We use the induced stratification and the cell triangulation interchangeably, specifically in Proposition 4.

A compatible conical stratification


Finally we put everything together to get a conical stratification of Rann(M)×R>0. Unfortunately we have to restrict ourselves to piecewise linear manifolds, or PL manifolds, which are homeomorphic images of geometric realizations of simplicial complexes, as otherwise we cannot claim M is a semialgebraic set. We can also just let M=Rk, as the point samples we are given could be coming from an unknown space.

Proposition 4: Let M be a PL manifold embedded in RN. There is a conical stratification ˜h:Rann(M)×R>0(B,) compatible with the stratification h:Rann(M)×R>0(SC,).

Proof: (Sketch) The main lifting is done by Theorem II.4.2 of Shiota. Since M is PL, it is semialgebraic, and so Mn×R>0RnN+1 is semialgebraic, by I.2.9.1 of Shiota. Since the quotient π of diagram (1) is semialgebraic, the space Rann(M)×R>0 is semialgebraic, by Scheiderer. Similarly, {f1(a)}aA is a family of semialgebraic sets, where f is the map from Lemma 1.  Theorem II.4.2 gives that Rann(M)×R>0 admits a cell triangulation (K,τ) compatible with {h1(S)}SSC. By the comment after Definition ???, this means we have a stratification Rann(M)×R>0(K0{τ(int(σ))}σK,). Further, by Proposition A.6.8 of Lurie, we have a conical stratification |K|(B,). This is all described by the solid arrow diagram below.


The vertical induced map comes as the poset B has the exact same structure as the abstract suimplicial complex K. The diagonal induced map comes as the map |K|Rann(M)×R>0 is a homeomorphism, and so has a continuous inverse. Composing the inverse with the conical sratification of Lurie, we get a conical stratification of Rann(M)×R>0. Composing the vertical induced arrow and the maps to (SC,) show that there is a conical stratification of Rann×R>0 compatible with its simplicial complex stratification from diagram (1).

Shiota actually requires that the space that admits a triangulation be closed semialgebraic, and having R>0 violates that condition. Replacing this piece with R0, then applying Shiota, and afterwards removing the t=0 piece we get the same result.

Remark: Every (sufficiently nice) manifold admits a triangulation, so it may be possible to extend this result to a larger class of manifolds, but it seems more sophisticated technology is needed.

References: Shiota (Geometry of subanalytic and semialgebraic sets, Chapters I.2, I.3, II.4), Scheiderer (Quotients of semi-algebraic spaces), Lurie (Higher algebra, Appendix A.6)

1 comment: