Manifolds modeled on locally convex spaces

12 Metric geometry

Metric spaces in which the distance is the infimum of lengths of curves, after Gromov, as in Chapter 8 of [ Lar26 ] : the rectifiable distance, interior metric spaces, approximate midpoints, the metric Hopf–Rinow theorem, and their asymmetric counterparts. The lengths of curves are those of Mathlib, through the variation of a curve. Nothing in this chapter uses manifolds or the calculus; it is applied to Finsler manifolds in Chapter 13.

12.1 The rectifiable distance

Let \((X, d)\) be a pseudo-metric space, with distances in \([0, \infty ]\). The length of a continuous curve \(\gamma \colon [a, b] \to X\) is the supremum of the sums \(\sum _i d(\gamma (t_i), \gamma (t_{i+1}))\) over the partitions of \([a, b]\); in Mathlib this is the variation eVariationOn. Its basic properties are in Mathlib: refining a partition increases the sums, the length vanishes only for constant curves, it is additive under concatenation, invariant under monotone reparametrizations [ Lar26 , Obs. 8.1.1 ] , lower semicontinuous under uniform convergence [ Lar26 , Prop. 8.1.3 ] , and the arc length \(t \mapsto \ell (\gamma |_{[a, t]})\) is continuous [ Lar26 , Lema 8.1.4 ] .

The rectifiable distance \(d_\ell (x, y)\) is the infimum of the lengths of the continuous curves from \(x\) to \(y\), and \(\infty \) if there is none.

\(d_\ell \) is a pseudo-distance, and \(d \leq d_\ell \). In particular the topology of \(d_\ell \) is finer than that of \(d\) [ Lar26 , Obs. 8.1.7 ] .

Proof ▶

Constant curves, reversed curves and concatenations give the three properties of a distance. The trivial partition gives \(d(x, y) \leq \ell (\gamma )\).

Lemma 12.3 When the topologies agree
✓

The topologies of \(d\) and \(d_\ell \) coincide if and only if, for every \(x\) and every \(\varepsilon {\gt} 0\), the points close enough to \(x\) can be joined to \(x\) by curves of length \({\lt} \varepsilon \) [ Lar26 , Lema 8.1.8 ] .

Proof ▶

This is the continuity of the identity \((X, d) \to (X, d_\ell )\), written with balls.

Definition 12.4 Interior metric spaces
✓

\((X, d)\) is an interior metric space (a length space) if \(d = d_\ell \) [ Lar26 , § 8.1.1 ] .

Proposition 12.5 Approximate midpoints
✓

In an interior metric space, for all \(x, y\) and \(\varepsilon {\gt} 0\) there is \(z\) with \(\max \{ d(x, z), d(z, y)\} \leq \frac12 d(x, y) + \varepsilon \) [ Lar26 , Problema 8.III ] .

Proof ▶

Take a curve of length \({\lt} d(x, y) + 2\varepsilon \). By the continuity of the arc length and the intermediate value theorem it has a point at half its length.

An interior metric space satisfies the ball condition: for \(r_1 + r_2 \leq d(x, y)\), the closed balls \(\bar B_{r_1}(x)\) and \(\bar B_{r_2}(y)\) are at distance at most \(d(x, y) - r_1 - r_2\). The ball condition is equivalent to having approximate midpoints, and a complete space with approximate midpoints is interior. So for complete spaces the three properties are equivalent.

Proof ▶

The converse direction builds dyadic chains of approximate midpoints with steps at most \((d(x, y) + \varepsilon )/2^n\), and the curve as the uniform limit of the corresponding step functions, which exists by completeness and is \((d(x, y) + \varepsilon )\)-Lipschitz. This replaces the extension from dyadic rationals of [ Lar26 , Lema 8.1.9 ] .

12.2 The metric Hopf–Rinow theorem

In an interior metric space, for \(a, b \geq 0\) with \(a + b = d(x, y)\) and \(\varepsilon {\gt} 0\) there is \(z\) with \(d(x, z) \leq a + \varepsilon \) and \(d(z, y) \leq b + \varepsilon \).

Proof ▶

As in Proposition 12.5, with the point at arc length \(a\).

Lemma 12.8 Lipschitz reparametrization
✓

A curve of finite length \(\ell \) can be reparametrized on \([0, 1]\) as an \(\ell \)-Lipschitz curve [ Lar26 , Obs. 8.1.5, Cor. 8.1.6 ] .

Proof ▶

Compose with a right inverse of the normalized arc length; the curve is constant on the level sets of the arc length.

Let \(X\) be a complete interior metric space.

  1. If \(\bar B_r(x)\) is compact for all \(0 {\lt} r {\lt} \rho \), then \(\bar B_\rho (x)\) is compact.

  2. If \(X\) is locally compact, every closed ball is compact, and any two points are joined by a shortest curve, of length equal to their distance.

Proof ▶

(1) By Lemma 12.7 every point of \(\bar B_\rho (x)\) is close to the compact ball \(\bar B_{\rho - \delta }(x)\), so \(\bar B_\rho (x)\) is totally bounded, and it is complete. (2) The supremum of the radii of compact balls cannot be finite: by (1) the ball of that radius is compact, and a finite cover by balls with compact closure, together with a Lebesgue number, gives a larger compact ball. For the shortest curves, reparametrize almost minimizing curves as in Lemma 12.8; they lie in a compact ball and are uniformly Lipschitz, and a cluster point in the product topology (Tychonoff) is a Lipschitz curve of length at most the distance. Here the formal proof departs from [ Lar26 ] , which uses Arzelà–Ascoli and the lower semicontinuity of the length.

\(\gamma \colon [0, 1] \to X\) is a short geodesic of length \(L\) if \(d(\gamma (s), \gamma (t)) = L |s - t|\); its length is \(L\). \(X\) is geodesic if it is interior and any two points are joined by a short geodesic [ Lar26 , Def. 8.1.11 ] .

The sub-arcs of a curve realizing the distance realize the distance [ Lar26 , Obs. 8.1.13 ] ; normalized by arc length such a curve is a short geodesic. Hence a complete, locally compact interior metric space is geodesic [ Lar26 , Teorema 8.2.3 ] .

Proof ▶

Additivity of the length and the triangle inequality; for the normalization, the bounds \(d(\eta (s), \eta (t)) \leq L|s - t|\) and the triangle inequality through \(\eta (s), \eta (t)\) give the reverse inequality.

Remark 12.12
✓
#

A space whose closed balls are compact is complete [ Lar26 , Obs. 8.2.4 ] . Being geodesic and locally compact does not imply completeness: the interval \((0, 1)\).

Example 12.13 McAlpin: no shortest curves in infinite dimension
✓
#

Let \(A\) be the diagonal operator on \(\ell ^2\) with \(a_0 = 1\), \(a_n = 1 + 2^{-n}\), and \(M = A(S)\), \(S\) the unit sphere, with lengths measured by the norm of \(\ell ^2\). Then the distance in \(M\) between \(e_0\) and \(-e_0\) is \(\pi \), but every piecewise \(C^{1}\) curve in \(M\) joining them is strictly longer than \(\pi \) [ Lar26 , Ejemplo 10.4.2 ] .

Proof ▶

The curves \(A(\cos (\pi t) e_0 + \sin (\pi t) e_n)\) have length at most \(\pi a_n \to \pi \). Every curve on \(S\) from \(e_0\) to \(-e_0\) has length at least \(\pi \), by the triangle inequality for angles. If \(\beta = A\alpha \) is a curve in \(M\), \(\| \beta '\| = \| A\alpha '\| \geq \| \alpha '\| \), with equality only when \(\alpha ' \in \mathbb {R}e_0\), which forces \(\alpha \) to stay at \(e_0\).

In finite dimension this cannot happen: a complete Riemannian manifold has shortest curves [ Lar26 , Cor. 10.4.3 ] , and Corollary 13.55 gives the Finsler version.

12.3 Asymmetric distances

For a non-reversible Finsler structure \(d_b\) is not symmetric. Following Mennucci [ Men13 ] , an asymmetric distance \(b\) satisfies \(b(x, x) = 0\) and the triangle inequality, and the topology is that of the symmetrized distance \(\max \{ b(x, y), b(y, x)\} \); forward and backward balls alone may generate different topologies [ Men13 , Example 4.1 ] . Lengths are forward lengths: suprema of \(\sum b(\gamma (t_i), \gamma (t_{i+1}))\), which change under reversal of the curve [ Men13 , Remark 3.5 ] .

An asymmetric metric space carries \(b\) and the topology of its symmetrization. The forward length of a curve is its forward variation; \(b_\ell \) is the infimum of forward lengths of continuous curves, and \(X\) is forward interior if \(b_\ell = b\) (g-intrinsic in [ Men13 , Def. 3.16 ] ).

Forward lengths are additive and invariant under increasing reparametrizations; the running length of a continuous curve of finite forward length is continuous [ Men13 , Prop. 3.9 ] ; \(b \leq b_\ell \) and \(b_\ell \) satisfies the triangle inequality. Forward interior spaces have approximate intermediate points, and curves of finite forward length admit forward Lipschitz reparametrizations on \([0, 1]\).

Proof ▶

The proofs of the symmetric case adapt, with care for the order of the arguments of \(b\); the continuity of the running length follows [ Men13 , Appendix B.2 ] .

Let \(X\) be a Hausdorff forward interior space, \(b(x, y) {\lt} \infty \), with compact closed forward ball \(\{ z : b(x, z) \leq b(x, y) + 1\} \) on which \(b(w, z) \leq C\, b(z, w)\). Then there is a continuous curve from \(x\) to \(y\) of forward length \(b(x, y)\).

Proof ▶

As in Theorem 12.9, with forward Lipschitz reparametrizations. The bounded asymmetry makes the limit curve Lipschitz for the symmetrized distance, hence continuous; without it the limit need not be continuous [ Men13 ] .

Proof ▶

As Theorem 13.45, with forward sums; reversed chains give the bound on the asymmetry, and Lemma 13.50 makes \((M, d_b)\) Hausdorff.

Proof ▶

Pointwise comparisons of the integrands pass to lengths of chains and to distances.

So for the topology, bounded asymmetry reduces to the reversible case; lengths and shortest curves of \(d_b\) and \(d_s\) are, however, different, and without a reversibility constant the topology of \(d_s\) may be strictly finer.