Topology of the Bend Loci of Convex Piecewise Linear Functions

As CW-complexes, $\mathbb{R}^{d}\cup\{pt\}$ is obtained from $X\cup\{pt\}$ by attaching $d$-cells. Since $\mathbb{R}^{d}\cup\{pt\}\cong S^{d}$ is $(d-1)$-connected, $X\cup\{pt\}$ is $(d-2)$-connected. This proves the first statement.

Let $\overline{X}$ be the polyhedral subdivision of $\mathbb{R}^{d}$ induced by $X$. Let $\overline{X}^{\text{cpt}}$ be the union of all compact faces of $\overline{X}$ and $X^{\text{cpt}}$ be the union of all compact faces of $X$. We claim that $\mathbb{R}^{d}$ deformation retracts onto $\overline{X}^{\text{cpt}}$. First, pick any point $x\in\overline{X}^{\text{cpt}}$ and let $B$ be a sufficiently large open ball centered at $x$ that contains $\overline{X}^{\text{cpt}}$. Since $\mathbb{R}^{d}$ deformation retracts onto $B$, it remains to show that $B$ deformation retracts onto $\overline{X}^{\text{cpt}}$. This can be done using an argument in [Hat02] (Proof of Proposition 0.16. See also Figure 1). For each unbounded face $F$ of $\overline{X}$, $F\cap B$ is a polyhedron truncated by a large sphere. It can be thought of as a glass filled with water, the sphere being the surface of the water. Pick a point $O$ that is in $F$ and sufficiently far from the sphere. Then the radial projection from $O$ deformation retracts the water onto the glass. We repeat this process for $F^{\prime}\cap B$ for all the unbounded lower dimensional faces $F^{\prime}$ of $F$, until all the unbounded faces of $F$ are contracted. The overall effect is that we get a deformation retraction of $F\cap B$ on to the compact faces of $F$. Apply the above construction to $F\cap B$ for all unbounded faces $F$. This deformation retracts $B$ onto $\overline{X}^{\text{cpt}}$. Hence, $\mathbb{R}^{d}$ deformation retracts onto $\overline{X}^{\text{cpt}}$, so $\overline{X}^{\text{cpt}}$ is contractible. Since $\overline{X}^{\text{cpt}}$ is obtained from $X^{\text{cpt}}$ by attaching $d$-cells, $X^{\text{cpt}}$ is $(d-2)$-connected. ∎

Take $X$ and $P$ as in the hypothesis of $C^{n}_{d}$. Let $\mathcal{H}$ be the minimal set of supporting hyperplanes of $P$. Consider the function

$$f:P\mapsto\mathbb{R},x\mapsto\prod_{H\in\mathcal{H}}d(x,H)$$

(4)

where $d(x,H)$ is the distance from $x$ to $H$. The genericity condition for $X$ and $P$ guarantees that $X$ is a Whitney stratified space and that $f$ is a Morse function on $X$. Let $x_{1},...,x_{r}$ be the finitely many critical points of $f_{X}$ and $t_{1},...,t_{r}$ be the corresponding distinct critical values. Set

$$X_{I}=f^{-1}(I)\cap X$$

(5)

for $I$ an interval or a single point in $\mathbb{R}_{\geq 0}$. We keep track of the topology change of $X_{[0,t]}$ as $t$ crosses the critical values. There are two cases (see Figure 2).

• •

  Case 1: $x_{k}$ is a vertex of $X$. Let $T_{x_{k}}X$ be the tangent fan of $X$ at $x_{k}$ and $H$ be the tangent hyperplane $T_{x_{k}}f^{-1}(t_{k})$. Translate $H$ slightly towards $T_{x_{k}}f^{-1}([0,t_{k}))$ and call the new hyperplane $H^{\prime}$. Let $A$ be the intersection of $H^{\prime}$ with $T_{x_{k}}X$. Note that $A$ is an intersection of $n$ generic polyhedral hypersurfaces in $\mathbb{R}^{d-1}$.

  Let $\epsilon>0$ be sufficiently small. By definition, the space $X_{[0,t_{k}]}$ is obtained from $X_{[0,t_{k}-\epsilon)}$ by attaching $X_{(t_{k}-\epsilon,t_{k}]}$ along $f^{-1}(t_{k}-\epsilon)\cap X$, which may have (finitely) many connected components. Suppose $X_{(t_{k}-\epsilon,t_{k}]}$ and $f^{-1}(t_{k}-\epsilon)\cap X$ decompose into disjoint unions of their connected components as follows

  $$X_{[t_{k}-\epsilon,t_{k}]}=\bigsqcup_{i}^{\ell}Y_{i},\quad f^{-1}(t_{k}-\epsilon)\cap X=\bigsqcup_{i}^{\ell}Z_{i},$$

  (6)

  such that $Y_{i}$ is glued along $Z_{i}$ for each $i$. Suppose $x_{k}$ is in $Y_{1}$. Then $(Y_{1},Z_{1})$ is homotopy equivalent to $(CA,A)$, and $(Y_{i},Z_{i})$ is homotopy equivalent to $(\mathbb{R}^{d-n-1}\times[0,\epsilon],\mathbb{R}^{d-n-1})$. By the induction hypothesis, $A$ is $(d-n-2)$-connected. By the homotopy long exact sequence for the pair $(CA,A)$, $(CA,A)$ is $(d-n-1)$-connected. Hence, the gluing data $(X_{[t_{k}-\epsilon,t_{k}]},f^{-1}(t_{k}-\epsilon))$ is $(d-n-1)$-connected. By definition, $(X_{[0,t_{k}-\epsilon]},f^{-1}(t_{k}-\epsilon))$ is path-connected. By homotopy excision ([Hat02], Theorem 4.23), $(X_{[0,t_{k}]},X_{[0,t_{k}-\epsilon]})$ is $(d-n-1)$-connected.

• •

  Case 2: $x_{k}$ is in the relative interior of a $k$-dimensional face $\sigma$ of $X$. In this case, the gluing data splits into normal Morse data and tangential Morse data. Similar to Case 1, the gluing data may have more than one connected components, and the only component that contributes to the topology change is the one where $x_{k}$ is. The treatment for multiple components is exactly the same as in the precious case. Therefore, it doesn’t hurt to assume that the gluing data has only one connected component.

  The tangential Morse data is $(\sigma,\partial\sigma)$. Let $N$ be the orthogonal complement of $T_{x_{k}}\sigma$. The normal Morse data is

  $$(N\cap T_{x_{i}}f^{-1}([0,t_{i}])\cap T_{x_{i}}X,N\cap H^{\prime}\cap T_{x_{i}}f^{-1}([0,t_{i}])\cap T_{x_{i}}X)$$

  (7)

  where $H^{\prime}$ is as in the previous case. Note that $N\cap T_{x_{i}}f^{-1}([0,t_{i}])\cap T_{x_{i}}X$ is the cone over $N\cap H^{\prime}\cap T_{x_{i}}f^{-1}([0,t_{i}])\cap T_{x_{i}}X$. The latter is a complete intersection of $n$ generic polyhedral hypersurfaces in $\mathbb{R}^{d-k-1}$. By the induction hypothesis and the homotopy long exact sequence for a pair, the normal Morse data is $(d-k-n-1)$-connected. Since the tangential Morse data $(\sigma,\partial\sigma)$ is $(k-1)$-connected, the total gluing data

  $$(\sigma,\partial\sigma)\times(N\cap T_{x_{i}}f^{-1}([0,t_{i}])\cap T_{x_{i}}X,N\cap H^{\prime}\cap T_{x_{i}}f^{-1}([0,t_{i}])\cap T_{x_{i}}X)$$

  (8)

  is $(d-n-1)$-connected. Now repeat the argument in the previous case, we conclude that $(X_{[0,t_{k}]},X_{[0,t_{k}-\epsilon]})$ is $(d-n-1)$-connected.

Observe that $X_{[0,t_{k+1}-\epsilon]}$ deformation retracts onto $X_{[0,t_{k}]}$, so to sum up, we have,

$$(X_{[0,t_{k+1}-\epsilon]},X_{[0,t_{k}-\epsilon]})\text{ is $(d-n-1)$-connected}$$

(9)

and

$$(X_{[0,t_{1}-\epsilon]},X_{0})\text{ is $(d-n-1)$-connected}$$

(10)

By applying homotopy long exact sequence to the triple $(X_{[0,t_{k+1}-\epsilon]},X_{[0,t_{k}-\epsilon]},X_{0})$, the above two connectedness properties imply that $(X_{[0,t_{k+1}-\epsilon]},X_{0})$ is $(d-n-1)$-connected. In other words, $(X_{[0,t)},X_{0})$ is $(d-n-1)$-connected for all $t>0$, which completes the proof. ∎

Let $X$ be a complete intersection of $n-1$ generic polyhedral hypersurfaces and $X_{n}$ be another generic hypersurface. Let $X^{\prime}=X\cap X_{n}$. Pick any connected component of $\mathbb{R}^{d}\backslash X_{n}$ and take its closure, which is a polyhedron $P$ with generic supporting hyperplanes. Applying Lemma 3 to $X$ and $P$, we know that $X$ is obtained from $X\cap\partial P$ be gluing cells of dimension no less than $d-n+1$. Therefore, $X$ is obtained from $X^{\prime}$ by gluing cells of dimension no less than $d-n+1$. Since by the induction hypothesis $X$ is $(d-n)$-connected, $X^{\prime}$ is $(d-n-1)$-connected. ∎