On the metric Kollár-Pardon problem

We argue by induction on \dim X. If \dim X = 0 there is nothing to prove. In general, from Theorem 2.1, (ii) and Remark 1 it follows that we can find a \mathcal{C}^k- (and \mathcal{C}^{\infty}-, if \Sigma is of QA type) cellular decomposition

Let I=\{i_1, \ldots, i_q\} be the set of indices of cells of non-maximal dimension, that is \dim X_{i_j}<\dim X if and only if i_j \in I. Put S_{r-1}:= \bigcup_{i\in I} X_i. Then V_r:=X \setminus S_{r-1}=\bigsqcup_{j \notin I} X_j is a union of non-interesecting definable \mathcal{C}^k- (respectively \mathcal{C}^{\infty}-) manifolds. At the same time, \dim S_{r-1}< \dim X, so the induction assumption applies to it.

Let U_{\alpha} \in \xi. By Theorem 2.1, (ii) there exists a finite covering of U_{\alpha} by open \Sigma-definable simply connected V^{\alpha}_{\beta}, \ U_{\alpha} = \bigcup_{\beta} V^{\alpha}_{\beta}. The cover p \colon \widetilde{X} \to X splits over V^{\alpha}_{\beta}, that is, p^{-1}(V^{\alpha}_{\beta})\simeq V^{\alpha}_{\beta} \times \{1, \ldots, k\}. Here we use that p^{-1}(V^{\alpha}_{\beta}) has finitely many connected components since \pi_1(X) is finite and p is a finite cover. Each of these connected components is homeomorphic to V^{\alpha}_{\beta}. The connected components of preimages of V^{\alpha}_{\beta} for all \alpha and \beta give rise to a definable atlas \widetilde{\xi} on \widetilde{X} with local coordinates \phi_{\alpha} \circ p. The projection p is definable since in a finite collection of definable charts V^{\alpha}_{\beta} it coincides with the projection V^{\alpha}_{\beta} \times \{1, \ldots, k\} \to V^{\alpha}_{\beta}.

To prove (ii) notice that the graph \Gamma_{\widetilde{f}} \subset Y \times \widetilde{X} is nothing but the preimage of the graph \Gamma_f \subset Y \times X under the (definable) map id \times p \colon Y \times \widetilde{X} \to Y \times X. Therefore, \widetilde{f} is definable whenever f is.

Without loss of generality, we may assume that \phi_{\alpha} are all open embeddings. Then d\phi_{\alpha}^{-1}(\partial/\partial x_i) defines trivialisations of TM|_{U_{\alpha}}. The rest of the Proposition follows from the fact that the derivatives of definable smooth functions are definable and computations in the local coordinates.

(i). Let s_{\alpha}=s|_{U_{\alpha}}. Then \nabla(s_{\alpha})=ds_{\alpha}+P_{\alpha}s_{\alpha}. Both summands are definable: the second one because of the definability of P_{\alpha}, and the first one because the derivative of a definable function is definable (Theorem 2.1, (iv)) and ds_{\alpha} can be expressed via partial derivatives of the components of s_{\alpha} with respect to local definable trivialisations.

(ii). Choose a definable chart U with definable coordinates (x^1, \ldots, x^n). On this chart, \nabla can be written as d+\Gamma, where the elements of the tensor \Gamma are Christoffel symbols \Gamma_{i,j}^k. Those are expressed in terms of the partial derivatives of g along the basis vector fields:

The functions g^{a,b} are definable with respect to x^c and a derivation of a definable function is definable. Hence \nabla is definable.

The item (i) follows from the finite cell decomposition for \widetilde{N} (Theorem 2.1, (iii)). To see (ii) note that (i) implies that \widetilde{N} has finitely many connected components and the connected components of \widetilde{N} are in bijection with \pi_1(M)/i_*(\pi_1(N)).

Let W_{\widetilde{M}} to be the Console-Olmos map of (\widetilde{M}, \widetilde{g}). Let K \subseteq \mathbb R^k be the image of W_{\widetilde{M}}. By Corollary 2.7 the map W_{\widetilde{M}} is definable. In particular, K is definable in \mathbb R^k and the fibres of W_{\widetilde{M}} are definable subspaces of \widetilde{M}.

Notice that W_{\widetilde{M}} coincides with the pullback \pi^*W_{M} of the Console-Olmos map of (M,g). Indeed, \pi \colon (\widetilde{M}, \widetilde{g}) \to (M, g) is a local isometry, therefore each point x \in \widetilde{M} has a neighbourhood U, such that W_{\widetilde{M}}|_U = (\pi^*W_M)|_U. Hence K= W_M(M) is compact since M is compact.

Let p \colon \widetilde{K} \to K be the universal cover of K and \Phi \colon \widetilde{M} \to \widetilde{K} be the canonical lifting of W_{\widetilde{M}}. We get a diagram:

Since W_{\widetilde{M}} is definable, its fibres have finitely many connected components. At the same time, W_{\widetilde{M}}=p \circ \Phi, so the fibres of p have finitely many connected components as well. This means that \pi_1(K) is finite and by Lemma 2.3 the topological space \widetilde{K} can be endowed with a definable structure in such a way that \Phi \colon \widetilde{M} \to \widetilde{K} is definable. Since \widetilde{K} is a finite cover of K and K is compact, \widetilde{K} is compact as well. The rest of the claim follows immediately from Console-Olmos theorem (Theorem 2.9)

The proof is done by induction on \dim M. If \dim M=0, the statement is clear. Otherwise, let W_M \colon M \to \mathbb R^n be the Console-Olmos map of (M,g) and K=W_M(M). This map exists (see Remark 2) and its regular fibres are locally homogeneous. It is moreover bi-definable with respect to the map \widetilde{M} \to M since the same argument as in the proof of Theorem 3.2 applies.

Let K^{\circ} \subset K be the set of non-critical values of W_M. Clearly, K^{\circ} is an open dense definable subset and M^{\circ}:=W_M^{-1}(K^{\circ}) is an open dense definable submanifold of M. In fact, M^{\circ} is strongly bi-definable since \pi^{-1}(M^{\circ})=W_{\widetilde{M}}^{-1}(K^{\circ}) is definable in \widetilde{M}.

Let N:= M \setminus M^{\circ}. This is a closed strongly bi-definable subset of M. Notice that \dim N < \dim M. In fact, N=S_{r-1} will be the first step of our stratification.

By Lemma 2.2 we can find a stratification of N by definable locally closed subsets N_i, \ \varnothing=N_0 \subset \ldots \subset N_s =N.

We claim that in fact, we can choose all N_i to be strongly bi-definable. Indeed, one can choose N_{s-1} to be the singular locus of N. Clearly \pi^{-1}(N_{s-1}) is the singular locus of the definable subset \pi^{-1}(N) \subset \widetilde{M}, hence N_{s-1} is strongly bi-definable and we proceed by induction on \dim N.

Now for every i the set U_i:=N_i\setminus N_{i-1} is a definable Riemannian \mathcal{C}^k-manifold with definable cover \pi_i \colon \widetilde{U_i} = \pi^{-1}(U_i) \to U_i and \pi_i^*g=\widetilde{g}|_{\widetilde{U_i}} is definable. Finally, notice that \dim U_i \le \dim N < \dim M. Therefore, we may apply the induction assumption to U_i and for each U_i get stratifications

with the required properties. It is left to gather all nested stratifications S^i_j into a single stratification S_p of M. For this one can put S_p:= \bigcup_{\dim S^i_j =n_p} S^i_j, where \{n_p\} is the set of natural numbers appearing as dimensions of the strata of S^i_j.

For each (V_i:= S_i \setminus S_{i-1}, g|_{V_i}) the Console-Olmos map f_i:=W_{V_i} \colon V_i \to K_i has no critical fibres, therefore defines a locally trivial fibration with locally homogeneous fibres.

Take the union of all the fibres of the maps f_i, where f_i are as in Proposition 3.3.

Let M:=U/\Gamma and p \colon U \to M the projection. The metric g descends to a Riemannian metric \overline{g} on M. Since U \to M is a \Gamma-Galois cover, we have a surjective homomorphism \pi_1(M) \to \Gamma.

Let w_M \colon M \to K be the Console-Olmos map on (M, \overline{g}) and F \subset M a connected component of a non-critical fibre of w_M. Denote by \Gamma_F the image of the composition homomorphism \pi_1(F) \to \pi_1(M) \to \Gamma. The connected components of p^{-1}(F) are in bijection with \Gamma/\Gamma_F. On the other hand, p^{-1}(F) consists of several connected components of a fibre of the Console-Olmos map w_U \colon (U, g) \to K', hence definable. It follows that \pi_0(p^{-1}(F)) = \Gamma/\Gamma_F is finite. Therefore, \Gamma is quasi-isometric to \Gamma_F.

Let \widetilde{F} be a connected component of p^{-1}(F). The group \Gamma_F acts on (\widetilde{F}, g|_{\widetilde{F}}) by isometries and the action is free and properly discontinuous with a compact quotient F=\widetilde{F}/\Gamma_F. By Milnor-Švarc Lemma ([BH], Proposition 8.19) this implies \Gamma_F is quasi-isometric to (\widetilde{F}, g|_{\widetilde{F}}). By the Console-Olmos Theorem (Theorem 2.9) (\widetilde{F}, g|_{\widetilde{F}}) is locally homogeneous.

Note that we do not assume \widetilde{N} to be connected. Nevertheless, \widetilde{N} has only finitely many connected components since its cohomology algebra is finitely generated, and

is finite (here i\colon N \hookrightarrow M is the tautological embedding).

Let \Gamma:=\pi_1(M). Since \widetilde{N} is a \Gamma-Galois cover of N, the action of \pi_1(N) on H^{\bullet}(\widetilde{N}, \mathbb Z/2\mathbb Z) factors through the natural homomorphism \pi_1(N) \to \pi_1(M)=\Gamma. At the same time, if H^{\bullet}(\widetilde{N}, \mathbb Z/2\mathbb Z) is finitely generated, it is a finite-dimensional \mathbb Z/2\mathbb Z-algebra and the group \operatorname{Aut}(H^{\bullet}(\widetilde{N}, \mathbb Z/2\mathbb Z) is finite. Let G be the image of \pi_1(N) in \operatorname{Aut}(H^{\bullet}(\widetilde{N}, \mathbb Z/2\mathbb Z)). Consider the composition of homomorphisms of groups

Let q \colon M' \to M be the finite G-Galois cover of M associated to the group homomorphism \pi_1(M)=\Gamma \to G \subseteq \operatorname{Aut}(H^{\bullet}(\widetilde{N}, \mathbb Z/2\mathbb Z)). Let N':=q^{-1}(N). Notice that N' is connected since both \pi_1(N) and \pi_1(M) map to G surjectively. Of course the universal cover \widetilde{M} \xrightarrow{\pi} M factors through M' \xrightarrow{q} M and \widetilde{N} \to N factors through N' \to N. At the same time, the action of \pi_1(N') on H^{\bullet}(\widetilde{N}, \mathbb Z/2\mathbb Z) is trivial.

First of all, notice that we can replace M and N with M' and N' as in Proposition 3.6 so that the action of \pi_1(N) on H^{\bullet}(\widetilde{N}, \mathbb Z/2\mathbb Z) is trivial. Indeed, M' will be still aspherical, will have the same universal cover, and N'=M' if and only if N=M. Therefore, starting from now we assume that \pi_1(N) acts trivially on the cohomology algebra of \widetilde{N}.

Since \widetilde{N} is a \Gamma-Galois cover of N, the \mathbb Z/2\mathbb Z-cohomology groups of N can be computed using the Cartan-Leray spectral sequence ([McCl], Theorem 8^{bis}.9):

Since we assume that the H^{j}(\widetilde{N}, \mathbb Z/2\mathbb Z)’s are trivial \pi_1(N)-modules and the action of \pi_1(N) on \widetilde{N} factors through \Gamma, each H^j(\widetilde{N}, \mathbb Z/2\mathbb Z) is a finite-dimensional \mathbb Z/2\mathbb Z-vector space with trivial \Gamma-action and H^i(\Gamma, H^j(\widetilde{N}, \mathbb Z/2\mathbb Z))=H^i(\Gamma, \mathbb Z/2\mathbb Z) \otimes_{\mathbb Z/2\mathbb Z}H^j(\widetilde{N}, \mathbb Z/2\mathbb Z). Let n:=\dim M. The compact manifold M is a K(\Gamma, 1) space, thus H^n(\Gamma, \mathbb Z/2\mathbb Z)=\mathbb Z/2\mathbb Z and H^{n+i}(\Gamma, \mathbb Z/2\mathbb Z)=0 for i>0.

Let r be maximal integer such that H^r(\widetilde{N}, \mathbb Z/2\mathbb Z) is non-zero. Consider

It is non-zero, since it equals

At the same time

and

Therefore E_2^{n,r} survives in the spectral sequence and H^{n+r}(N, \mathbb Z/2\mathbb Z) is non-zero. We assumed that N is a closed manifold, so H^p(N, \mathbb Z/2\mathbb Z) vanishes for p>\dim N. It follows that n+r \le \dim N. Of course \dim N \le \dim M=n and it follows that n+r=\dim N, that is, \dim N=\dim M and r=0. Since M was assumed to be connected and N is a closed submanifold, we deduce that N=M.

If N \subseteq M is strongly be-definable, \pi^{-1}(N)=:\widetilde{N} \subseteq \widetilde{M} is definable. In particular, H^{\bullet}(\widetilde{N}, \mathbb Z/2\mathbb Z) is finitely generated by Theorem 2.1, (ii). The Corollary then follows from Lemma 3.7.

Let W_{M} \colon M \to \mathbb R^k be the Console-Olmos map. Let N \subseteq M be a connected component of a non-critical level set of W_M. Then N is a smooth submanifold of M and in the same time, it is strongly bi-definable: \pi^{-1}(N) is a union of several connected components of a level set of the Console-Olmos map W_{\widetilde{M}} of (\widetilde{M}, \widetilde{g}). However, by Corollary 3.8 this implies N=M and the map W_M is constant. Since (\widetilde{M}, \widetilde{g}) is complete and simply connected, it is homogeneous by Prüfer-Tricerri-Vanhecke theorem (Theorem 2.8).

The universal cover \widetilde{M} is contractible by Cartan-Hadamard Theorem ([BH], Ch. 4), so the statement follows from Theorem 3.9.

Let \widetilde{h}=\pi^*h and \widetilde{h}=\widetilde{g}+\sqrt{-1}\widetilde{\omega}. Here \widetilde{g} is a Riemannian metric on \widetilde{X} which is the real part of \widetilde{h}, and \widetilde{\omega} is a symplectic form which is the imaginary part of \widetilde{h}. Being the real part of a definable tensor, \widetilde{g} is definable. Theorem 3.9 implies that (\widetilde{X}, \widetilde{g}) is homogeneous as a Riemannian manifold. Let \xi be a Killing vector field on \widetilde{X}. Since the form \widetilde{\omega} is harmonic with respect to \widetilde{g}, its Lie derivative along \xi vanishes. This implies that any one-parameter subgroup of diffeomorphisms of \widetilde{X} that preserves \widetilde{g}, preserves also \widetilde{\omega}. Hence it preserves the complex structure operator J_{\widetilde{X}} as well. We deduce that if G is a Lie group that acts by isometries and transitively on \widetilde{X}, its action preserves the complex structure. Therefore, \widetilde{X} is homogeneous not only as Riemannian but also as a complex (and as a Kähler) manifold.