Manifolds modeled on locally convex spaces

13 Metrics

This chapter collects the metric structures on a manifold \(M\) modeled on a locally convex space \(E\). There are two families, and it is worth separating them from the start. A Finsler structure assigns to each tangent space a (possibly asymmetric) seminorm; it measures lengths of curves and gives a distance. A bilinear structure assigns to each tangent space a symmetric bilinear form, of any signature; it gives Riemannian, Lorentzian and pseudo-Riemannian metrics. The two families meet in exactly one place: a Riemannian metric is a Finsler metric, through \(v \mapsto \sqrt{g_x(v,v)}\). Lorentzian and pseudo-Riemannian metrics are not.

In infinite dimensions each family splits further, according to two different questions. For Finsler structures the question is how \(\mu _p\) depends on \(p\): continuously on the tangent bundle, or only along curves. For bilinear structures the question is which topology \(g_x\) induces on \(T_x M\): the one of the model (strong) or a coarser one (weak). We keep a separate word for each question, and we never use one for the other.

13.1 Finsler seminorms and two continuity conditions

A Finsler seminorm on a topological vector space \(E\) is a continuous function \(\mu \colon E \to [0, \infty )\) that is subadditive, \(\mu (v + w) \le \mu (v) + \mu (w)\), and positively homogeneous, \(\mu (c v) = c\, \mu (v)\) for \(c \ge 0\). It is a Finsler norm if \(\mu (v) = 0\) only for \(v = 0\), and it is reversible if \(\mu (-v) = \mu (v)\).

This is the notion of [ Lar19 , Def. 2.2 ] . A Finsler seminorm need not be symmetric: a reversible one is a continuous seminorm in the usual sense. Nothing is lost by allowing asymmetry at this stage.

A field of Finsler seminorms on \(M\) is a function \(\mu \colon TM \to [0,\infty )\) whose restriction \(\mu _x\) to each tangent space \(T_x M\) is a Finsler seminorm.

No condition relates \(\mu _x\) and \(\mu _y\) for \(x \neq y\). There are two natural ones, and the whole point of this section is that they are different.

Definition 13.3 Finsler metric
✓
#

A Finsler metric is a field of Finsler seminorms \(\mu \) that is continuous as a function \(\mu \colon TM \to \mathbb {R}\) on the tangent bundle.

Definition 13.4 Finsler length structure
✓

A Finsler length structure is a field of Finsler seminorms \(\mu \) such that \(t \mapsto \mu (\dot\gamma (t))\) is continuous for every \(C^{1}\) curve \(\gamma \) in \(M\).

These are, respectively, the “continuous” Finsler metrics and the Finsler metrics “continuous along paths” of [ Lar19 , Def. 2.3 ] . We reserve the name Finsler metric for the first notion, since this is what the term means for any reader, and we give the second one its own name: a Finsler length structure is exactly what is needed to measure lengths of curves (compare the length structures of [ BBI01 , § 2.1.1 ] ), and nothing more.

Remark 13.5 Two words we avoid
#

We do not speak of weak Finsler metrics: the word weak is used below for bilinear metrics, and there it refers to the topology of the fibers, not to the continuity on \(TM\). We do not use Finsler structure alone either: Palais [ Pal66 , Def. 2.1 and 2.6 ] uses it for a much stronger notion, in which each \(\mu _x\) is a norm inducing the topology of the fiber, locally uniformly equivalent to \(\mu _{x_0}\). That notion only makes sense for normable models and is not formalized here.

Theorem 13.6 Velocities are continuous
✓

The velocity \(t \mapsto \dot\gamma (t)\) of a \(C^{1}\) curve is a continuous curve in \(TM\). Consequently, every Finsler metric is a Finsler length structure.

Proof ▶

In a chart \(i\) around \(\gamma (t_0)\) the velocity reads \(s \mapsto (i(\gamma (s)), (i \circ \gamma )'(s))\), which is continuous because \(i \circ \gamma \) is \(C^{1}\). The corresponding chart of \(TM\) is a homeomorphism onto its image, so the velocity is continuous near \(t_0\).

The converse fails, already on \(\mathbb {R}^2\) (Section 19.2). The natural source of Finsler length structures that are not Finsler metrics are quotient metrics (Section 13.3).

13.2 Length and distance

The following notions only use the continuity of \(\mu \) along curves. We state them for Finsler length structures, so that they apply to quotient metrics; for a Finsler metric they are obtained through Theorem 13.6.

The length of a \(C^{1}\) curve \(\gamma \) on \([a, b]\) is \(L_\mu (\gamma ) = \int _a^b \mu \bigl(\dot\gamma (t)\bigr)\, dt\).

Lengths are nonnegative, additive on adjacent intervals, and constant curves have length zero.

Proof ▶

The integrand is continuous and nonnegative; additivity is that of the integral. The velocity of a constant curve vanishes and \(\mu _x(0) = 0\) by positive homogeneity.

A chain of \(C^{1}\) arcs from \(x\) to \(y\) of length \(L\) is either empty (\(x = y\), \(L = 0\)) or an arc \(\gamma |_{[a, b]}\) of a \(C^{1}\) curve with \(\gamma (a) = x\), followed by a chain from \(\gamma (b)\) to \(y\); its length is the sum of the lengths of the arcs. Chains concatenate, and their lengths add.

A chain is a piecewise \(C^{1}\) path, cut at its corners. Encoding the paths in this way makes concatenation a formality: there is no need to smooth the curve at the junctions, which on a manifold would require reparametrizations with vanishing velocity at the end points.

The distance \(d_\mu (x, y) \in [0, \infty ]\) is the infimum of the lengths of the chains of \(C^{1}\) arcs from \(x\) to \(y\), and \(\infty \) if there is none. It satisfies \(d_\mu (x, x) = 0\).

This is [ Lar19 , Def. 2.7 ] , where the infimum is taken over piecewise \(C^{1}\) curves.

Theorem 13.11 Triangle inequality
✓

\(d_\mu (x, z) \leq d_\mu (x, y) + d_\mu (y, z)\) for all \(x, y, z \in M\). Hence \(d_\mu \) is an extended, possibly asymmetric, pseudo-distance.

Proof ▶

If there is no chain from \(x\) to \(y\) or from \(y\) to \(z\), the right-hand side is \(\infty \). Otherwise concatenate a chain from \(x\) to \(y\) with a chain from \(y\) to \(z\), and take infima.

The reversed curve \(t \mapsto \gamma (-t)\) of a \(C^{1}\) curve is \(C^{1}\), with velocity \(-\dot\gamma (-t)\). If \(\mu \) is reversible, reversing an arc does not change its length, so chains can be run backwards and \(d_\mu (x, y) = d_\mu (y, x)\).

Proof ▶

In a chart, the reversed curve is the composition of \(\gamma \) with \(t \mapsto -t\), and the chain rule gives the velocity. The length follows by the change of variables \(t \mapsto -t\).

Remark 13.13 Degenerate distances
#

The distance may vanish between distinct points even when every \(\mu _x\) is a norm: Michor and Mumford proved that the geodesic distance of the \(L^2\) metric on the diffeomorphism group of a manifold vanishes [ MM05 , Thm. 5.2 ] . So \(d_\mu \) is in general only a pseudo-distance.

13.3 Lifting curves versus lifting neighborhoods

Let \(\pi \colon N \to M\) be a \(C^{1}\) map and \(\mu \) a field of Finsler seminorms on \(M\). Typically \(\mu \) is defined through \(\pi \), as a quotient metric \(\mu (w) = \inf \{ \mu _N(v) : T\pi \, v = w\} \), and what one controls is the function \(\mu \circ T\pi \) on \(TN\). The two continuity conditions then correspond to two lifting properties of \(\pi \).

Lemma 13.14 Chain rule for velocities
✓

If \(\pi \) is \(C^{1}\) and \(\Gamma \) is a \(C^{1}\) curve in \(N\), then \(T\pi \bigl(\dot\Gamma (t)\bigr) = (\pi \circ \Gamma )^{\boldsymbol \cdot }(t)\). The velocity at \(t\) only depends on the germ of the curve at \(t\).

Proof ▶

In charts, this is the chain rule of Michal–Bastiani calculus [ LCA ] for the composition of the local representative of \(\pi \) with the local representative of \(\Gamma \).

Theorem 13.15 Lifting curves gives a Finsler length structure
✓

Suppose that every \(C^{1}\) curve of \(M\) has, near each time, a \(C^{1}\) lift to \(N\), and that \(\mu \circ T\pi \) is continuous along the velocities of \(C^{1}\) curves of \(N\). Then \(\mu \) is a Finsler length structure.

Proof ▶

Near \(t_0\), \(\gamma = \pi \circ \Gamma \), so \(\dot\gamma = T\pi (\dot\Gamma )\) by Lemma 13.14, and \(\mu (\dot\gamma ) = (\mu \circ T\pi )(\dot\Gamma )\) is continuous.

Theorem 13.16 Lifting neighborhoods gives a Finsler metric
✓

If \(T\pi \colon TN \to TM\) is a quotient map and \(\mu \circ T\pi \) is continuous on \(TN\), then \(\mu \) is a Finsler metric.

Proof ▶

This is the universal property of quotient maps.

Remark 13.17 Submersions and homogeneous spaces
#

For Banach manifolds, a submersion (surjective differential with complemented kernel) has local smooth sections through every point, by the inverse function theorem. Then \(T\pi \) is open, hence a quotient map, and \(C^{1}\) curves lift locally (compose them with a local section): the hypotheses on \(\pi \) of both theorems hold. On locally convex manifolds there is no inverse function theorem and neither property is automatic; a curve may have \(C^{1}\) lifts while \(T\pi \) is not a quotient map. This is the situation of homogeneous spaces \(G/K\) of infinite-dimensional Lie groups in [ Lar19 , Lemma 3.19 ] : under a uniformity assumption on the metric of \(G\), the quotient metric is continuous along paths that have piecewise smooth lifts, and continuous on the tangent bundle when \(\pi _*\) is a quotient map.

13.4 Bilinear metrics

A smooth field of symmetric bilinear forms on \(M\) assigns to each \(x\) a symmetric bilinear form \(g_x\) on \(T_x M\), such that the local representatives \((y, a, b) \mapsto g_{i^{-1}(y)}(a, b)\) are smooth in every chart \(i\). Each \(g_x\) is then jointly continuous, and \(v \mapsto g_x(v,v)\) is continuous on \(TM\). The field is weakly nondegenerate if \(g_x(v, \cdot ) = 0\) forces \(v = 0\).

13.4.1 Riemannian metrics

Definition 13.19 Weak Riemannian metric
✓

A weak Riemannian metric is a smooth field of positive definite symmetric bilinear forms. It is strong at \(x\) if the sets \(\{ v : g_x(v,v) {\lt} \varepsilon \} \) form a basis of neighborhoods of \(0\) in \(T_x M\).

This is [ Sch23 , Definition 4.1 ] . Weak refers to the topology of the fibers: \(g_x\) is continuous on \(T_x M\), but the norm \(\sqrt{g_x(v,v)}\) may induce a strictly coarser topology than the one of the model. A strong Riemannian metric forces the model to be a pre-Hilbert space; it is a Hilbert space only if the model is complete, and we say so only in that case. On a Fréchet model that is not normable, such as the space \(C^\infty _{\mathrm{per}}\) of Chapter 18, every Riemannian metric is weak.

If \(g\) is a weak Riemannian metric, \(\mu _x(v) = \sqrt{g_x(v, v)}\) is a Finsler metric, and each \(\mu _x\) is a reversible Finsler norm.

Proof ▶

Subadditivity is the triangle inequality, which follows from the Cauchy–Schwarz inequality \(g_x(v,w)^2 \le g_x(v,v)\, g_x(w,w)\); the latter is the nonpositivity of the discriminant of \(t \mapsto g_x(v + t w, v + t w) \ge 0\). Continuity on \(TM\) is that of \(v \mapsto g_x(v,v)\).

So lengths and distances of Riemannian metrics are those of the previous sections.

13.4.2 Fundamental decompositions and the index

Definition 13.21 Fundamental decomposition
✓

Let \(B\) be a symmetric bilinear form on a topological vector space \(E\). A fundamental decomposition of \(B\) is a splitting \(E = E_+ \oplus E_-\) given by a continuous projection, \(B\)-orthogonal, with \(B\) positive definite on \(E_+\) and negative definite on \(E_-\).

A negative definite subspace has (algebraic) dimension at most \(\dim E_-\). In particular, two fundamental decompositions of \(B\) have negative parts of the same dimension.

Proof ▶

A negative definite subspace \(N\) meets \(E_+\) only at \(0\), so the projection onto \(E_-\) along \(E_+\) is injective on \(N\). Apply this to the negative part of each decomposition.

Remark 13.23 Algebraic versus Hilbert dimension
#

The argument works for any cardinal and needs no completeness. For a Krein space one usually measures \(E_-\) by its Hilbert dimension; for infinite index the two dimensions differ, and the invariance of the Hilbert dimension uses the completeness of \((E_+, B)\) and \((E_-, -B)\) [ Bog74 , Corollary IV.7.4 ] . That version is formalized in lc-analysis (KreinDecomposition.intrinsicDim_neg_eq and KreinDecomposition.intrinsicDim_neg_eq_card).

Proposition 13.24
✓

A form with a fundamental decomposition is weakly nondegenerate.

Proof ▶

If \(B(v, \cdot ) = 0\), write \(v = a + b\) with \(a \in E_+\), \(b \in E_-\). By orthogonality \(B(a,a) = B(v,a) = 0\) and \(B(b,b) = B(v,b) = 0\), so \(a = b = 0\).

13.4.3 Pseudo-Riemannian and Lorentzian metrics

A weak pseudo-Riemannian metric is a smooth field of symmetric bilinear forms admitting a fundamental decomposition of each tangent space. Its index at \(x\) is the dimension of the negative part of any such decomposition; it may be any cardinal.

Index \(0\) gives back weak Riemannian metrics; finite index corresponds to tangent spaces of Pontryagin type, infinite index to tangent spaces of Krein type. We keep the infinite index in the theory on purpose: Krein spaces appear naturally in operator theory.

Proof ▶

Take \(E_+ = E\), \(E_- = 0\) and the identity as projection.

Definition 13.27 Weak Lorentzian metric
✓

A weak Lorentzian metric is a weak pseudo-Riemannian metric of index \(1\) at every point.

Let \(g\) be a smooth field of symmetric bilinear forms such that each \(T_x M\) contains a vector \(v\) with \(g_x(v,v) {\lt} 0\) whose \(g_x\)-orthogonal is positive definite. Then \(T_x M = v^\perp \oplus \mathbb {R}v\) is a fundamental decomposition, and \(g\) is a weak Lorentzian metric.

Proof ▶

The projection onto \(v^\perp \) is \(w \mapsto w - \frac{g_x(w, v)}{g_x(v, v)}\, v\), which is continuous because \(g_x(v, \cdot )\) is. The remaining conditions are direct computations, and \(\mathbb {R}v\) has dimension \(1\).

So the decomposition of a Lorentzian tangent space is automatically topological: no closedness or complementation has to be assumed.

Remark 13.29 Lorentzian metrics are not Finsler metrics
#

Since \(g_x(v,v)\) takes negative values, \(\sqrt{g_x(v,v)}\) is not defined, and \(\sqrt{|g_x(v,v)|}\) is not subadditive. The theory of lengths of a Lorentzian metric (timelike curves, the reverse triangle inequality) is different and is not developed here.

Remark 13.30 What is not formalized
#

We do not prove that the index is locally constant, nor that fundamental decompositions can be chosen continuously in \(x\). These are genuine infinite-dimensional questions when the topology of \(g_x\) is weaker than that of the model. The Krein-space conditions (completeness of \((E_+, g)\) and \((E_-, -g)\), and the topology of the majorant \(g(P_+\cdot , P_+\cdot ) - g(P_-\cdot , P_-\cdot )\)) are not formalized either.

13.5 The Levi-Civita connection

In this section \(g\) is a smooth field of symmetric bilinear forms of any signature (Definition 13.18). In a chart \(i\) we write \(G(y)(a, b)\) for its local representative and \(\partial G(y)(u)(a, b)\) for the derivative of \(G\) in the base point, in the direction \(u\). Recall that, with the conventions of Chapter 10, the covariant derivative is \(\nabla _X Y = \mathrm{d}\tilde Y \cdot \tilde X - \Gamma (\tilde X, \tilde Y)\) in a chart.

A connection is metric for \(g\) if in every chart

\[ \partial G(y)(u)(a, b) = -G(y)\bigl(\Gamma _y(u, a), b\bigr) - G(y)\bigl(a, \Gamma _y(u, b)\bigr). \]

This is the chart expression of \(X\, g(Y, Z) = g(\nabla _X Y, Z) + g(Y, \nabla _X Z)\): differentiate \(y \mapsto G(y)(\tilde Y(y), \tilde Z(y))\) in the direction \(\tilde X\) and use the bilinearity of \(G(y)\).

If the connection is symmetric and metric for \(g\), then in every chart

\[ G(y)\bigl(\Gamma _y(u, a), b\bigr) = \tfrac 12 \bigl(-\partial G(y)(u)(a, b) - \partial G(y)(a)(u, b) + \partial G(y)(b)(u, a)\bigr). \]
Proof ▶

Write the compatibility condition for the three cyclic permutations of \((u, a, b)\), add the first two and subtract the third; the symmetry of \(\Gamma _y\) and of \(G(y)\) cancels four of the six terms.

The proof is pure algebra at a point: it uses neither completeness, nor nondegeneracy, nor the sign of \(g\).

If \(g\) is weakly nondegenerate, two symmetric metric connections have the same symbols in every chart, hence the same covariant derivative. In particular this holds for weak pseudo-Riemannian metrics of any index, finite or infinite, and for weak Riemannian and weak Lorentzian metrics.

Proof ▶

By Theorem 13.32, \(G(y)(\Gamma _y(u, a) - \Gamma '_y(u, a), b) = 0\) for every \(b\). The local representative \(G(y)\) is weakly nondegenerate because \(g_x\) is and the coordinate changes are invertible.

A symmetric connection is metric for \(g\) if and only if its spray \(F_y(v) = \Gamma _y(v, v)\) satisfies, in every chart,

\[ G(y)\bigl(F_y(v), z\bigr) = -\partial G(y)(v)(v, z) + \tfrac 12\, \partial G(y)(z)(v, v). \]
Proof ▶

On the diagonal \(u = a = v\) the Koszul formula is the formula of the statement. Conversely, polarizing the formula of the statement in \(v\) gives back the Koszul formula, since \(\Gamma _y\) is symmetric and bilinear and \(\partial G(y)\) is additive in each of its three arguments; and the Koszul formula gives back the compatibility with \(G\). The additivity of \(\partial G(y)(u)(a, b)\) in \(a\) and \(b\) follows by differentiating the identity \(G(y + tu)(a + a', b) = G(y + tu)(a, b) + G(y + tu)(a', b)\), valid for small \(t\).

In the notation of [ Lar26 , Teorema 10.1.1 ] , where \(G(y)(a, b) = \langle a, g(y) b\rangle \) and \(\partial G(y)(z)(a, b) = \langle a, g_{*y,z}\, b\rangle \), the formula of the statement is \(\langle F_y(v), g(y) z\rangle = -\langle g_{*y,v}\, v, z\rangle + \frac12 \langle g_{*y,z}\, v, v\rangle \), and the Koszul formula is its polarization; see also [ Lan99 , Ch. VIII, Theorem 4.2, MS 1 ] , whose bilinear map \(B_U\) is our \(\Gamma \).

Proposition 13.35 The obstruction to existence
✓

Let \(B\) be a symmetric, weakly nondegenerate bilinear form and \(D(u)(a, b)\) symmetric in \((a, b)\). Symbols \(\Gamma \) that are symmetric and compatible with \(B\) exist if and only if every Koszul functional \(b \mapsto \frac12(-D(u)(a, b) - D(a)(u, b) + D(b)(u, a))\) is of the form \(B(c, \cdot )\).

Proof ▶

One direction is Theorem 13.32. For the other, choose \(\Gamma (u, a)\) representing the Koszul functional; it is symmetric by weak nondegeneracy, and compatible by the algebra of the previous proof.

Remark 13.36 Existence is not automatic
#

For a strong Riemannian metric on a Hilbert manifold every continuous functional is represented (Riesz), and the Levi-Civita connection exists [ Lan99 , Ch. VIII, Theorem 4.1 ] . For weak metrics, and every metric on a Fréchet manifold that is not normable is weak, the Koszul functional need not be represented: Schmeding gives a weak Riemannian metric on \(\ell ^2\) with no metric spray [ Sch23 , Example 4.22 ] . Even when it is represented at every point, the symbols so obtained have to be smooth and satisfy the change of charts law; neither follows from the pointwise statement. Existence is therefore a property of the metric at hand, to be checked in examples.

The following example, after Schmeding, shows that the obstruction of Proposition 13.35 does occur.

Let \(H\) be a real inner product space, \(T\) a continuous, symmetric and positive definite operator on \(H\), and \(v \in H\). Then \(g_p(x, y) = e^{\langle v, p\rangle } \langle Tx, y\rangle \) is a weak Riemannian metric on \(H\), and if \(v \notin T(H)\) it has no symmetric metric connection. On \(\ell ^2\), the operator \(Tx = (x_n / (n + 1))_n\) and the vector \(v = (1 / (n + 1))_n\) satisfy these hypotheses.

Proof ▶

The derivative in the base point is \(\partial G(p)(w)(x, y) = \langle v, w\rangle g_p(x, y)\). By Theorem 13.34, a symmetric metric connection would give \(c \in H\) with

\[ \langle Tc, b\rangle = -\langle v, x\rangle \langle Tx, b\rangle + \tfrac 12 \langle Tx, x\rangle \langle v, b\rangle \qquad (b \in H), \]

that is, \(T(c + \langle v, x\rangle x) = \frac12 \langle Tx, x\rangle v\). For \(x = v \neq 0\) this puts \(v\) in \(T(H)\). On \(\ell ^2\), \(Tc = v\) forces \(c_n = 1\) for every \(n\), and the constant sequence \(1\) is not in \(\ell ^2\).

When \(T\) is invertible, \(g\) is a strong metric and the Levi-Civita connection exists; the example needs a metric \(\langle T\cdot , \cdot \rangle \) whose topology is strictly coarser than that of \(H\). Schmeding’s original example [ Sch23 , Example 4.22 ] , the metric \(e^{-\| p\| ^2} \sum _n x_n y_n / n^3\) of [ Sch23 , Example 4.12 ] (after Magnani and Tiberio), uses the weight \(n^{-3}\) and the factor \(e^{-\| p\| ^2}\); the obstruction is the same.

13.6 Submanifolds: orthogonal projectors in any signature

Let \(\beta \) be a continuous symmetric bilinear form on \(F\), of any signature: an inner product, or the indefinite form of a Krein or a Pontryagin space. Let \(X \subseteq F\) be an embedded submanifold with a field of projectors \(P\) onto its tangent spaces (Section 10.5). In a slice chart, with \(\psi = \varphi ^{-1}|_S\), the induced form is \(G(y)(a, b) = \beta (\mathrm{d}\psi _y a, \mathrm{d}\psi _y b)\). We say that \(P\) is \(\beta \)-orthogonal if \(\beta (P_p w, t) = \beta (w, t)\) for every \(w \in F\) and every \(t \in T_p X\).

Lemma 13.38 Orthogonal projectors need nondegenerate ranges
✓

If \(\beta \) is weakly nondegenerate and \(Q\) is a \(\beta \)-orthogonal projector, no nonzero vector of the range of \(Q\) is \(\beta \)-orthogonal to the whole range.

Proof ▶

If \(t \in \operatorname {ran} Q\) is orthogonal to \(\operatorname {ran} Q\), then \(\beta (w, t) = \beta (Qw, t) = 0\) for every \(w \in F\), so \(t = 0\).

If \(P\) is \(\beta \)-orthogonal, the connection of \(P\) is metric for the induced form in every slice chart: \(\partial G(y)(u)(a, b) = -G(y)(\Gamma (u, a), b) - G(y)(a, \Gamma (u, b))\).

Proof ▶

By the product rule, \(\partial G(y)(u)(a, b) = \beta (\mathrm{d}\psi \, a, \mathrm{d}^2\psi (u, b)) + \beta (\mathrm{d}^2\psi (u, a), \mathrm{d}\psi \, b)\). On the other hand \(\mathrm{d}\psi \, \Gamma (u, a) = -P(\mathrm{d}^2\psi (u, a))\), and since \(\mathrm{d}\psi \, b\) is tangent, \(\beta (P \mathrm{d}^2\psi (u, a), \mathrm{d}\psi \, b) = \beta (\mathrm{d}^2\psi (u, a), \mathrm{d}\psi \, b)\). The same holds with \(a\) and \(b\) exchanged.

No sign condition on \(\beta \) is used: the computation is the one of the Riemannian case [ Lar26 , § 10.1.1 ] .

If \(\beta \) is weakly nondegenerate and \(P\) is \(\beta \)-orthogonal, then in every slice chart the induced form is weakly nondegenerate, and every symmetric family of symbols compatible with it coincides with the symbols of \(P\).

Proof ▶

By Lemma 13.38 the tangent spaces are nondegenerate, and \(\mathrm{d}\psi _y\) maps \(S\) injectively onto \(T_p X\); so \(G(y)\) is weakly nondegenerate. Then Theorem 13.39 and the uniqueness given by the Koszul formula (Theorem 13.33) apply.

The induced forms of the slice charts define a smooth field of symmetric bilinear forms on \(X\), the induced metric, whose local representative in each slice chart is \(\beta (\mathrm{d}\psi _y a, \mathrm{d}\psi _y b)\). If \(\beta \) is positive definite, it is a weak Riemannian metric.

Proof ▶

Define the form at \(x\) in the preferred slice chart. If \(\tau \) is the change from a slice chart \(\varphi \) to the preferred chart \(\varphi '\), then \(\psi ' \circ \tau = \psi \) near \(y\), and the chain rule gives \(\mathrm{d}\psi '_{\tau y}(\mathrm{d}\tau _y v) = \mathrm{d}\psi _y v\). So the local representative in the chart of \(\varphi \) is \(\beta (\mathrm{d}\psi _y a, \mathrm{d}\psi _y b)\), which is smooth by the product rule. Positivity follows from the injectivity of \(\mathrm{d}\psi _y\).

If \(P\) is \(\beta \)-orthogonal, the connection of \(P\) is metric for the induced metric. If moreover \(\beta \) is weakly nondegenerate, the induced metric is weakly nondegenerate, and every symmetric connection which is metric for it has the covariant derivative of \(P\): the Levi-Civita connection of \(X\) is obtained by deriving in \(F\) and projecting.

Proof ▶

In each chart of the atlas of \(X\) the local representative is the induced form, so Theorem 13.39 gives the compatibility. Uniqueness is Theorem 13.33.

Remark 13.43 The indefinite case
#

For an inner product on a Hilbert space every closed subspace has an orthogonal projector. For an indefinite \(\beta \) this fails, and Lemma 13.38 shows the first obstruction: a degenerate tangent space, such as that of a lightlike hypersurface of a Lorentzian space, admits no \(\beta \)-orthogonal projector, and the description of the Levi-Civita connection by “derive and project” is lost. In a Krein space a closed subspace has a \(\beta \)-orthogonal projector exactly when it is itself a Krein space with the induced form (Bognár, Theorem V.3.4, in lc-analysis); being closed, nondegenerate and decomposable is not enough [ Bog74 , Example V.3.3 ] (in lc-analysis).

The theory of ortho-complemented subspaces of Krein spaces used below (Bognár, Theorems V.3.1, V.3.4, V.3.5 and Example V.3.3), together with the majorant topology, the intrinsic dimension and continuous families of Krein spaces, is formalized in lc-analysis and documented in the chapter on Krein spaces of its blueprint.

Let \(X\) be a submanifold of a real Hilbert space with \(\beta (x, y) = \langle Jx, y\rangle \). A \(\beta \)-orthogonal field of projectors has ortho-complemented tangent spaces, which are Krein spaces, and continuous projectors. Conversely, a closed tangent space which is a Krein space has a continuous \(\beta \)-orthogonal projector; whether these can be chosen smoothly in \(p\) is not addressed.

Proof ▶

Bognár’s Theorem V.3.4 [ Bog74 , Theorem V.3.4 ] , formalized in lc-analysis (chapter on Krein spaces of its blueprint), items 2, 3 and 4.

13.7 \((M, d_b)\) is an interior metric space

Let \(\mu \) be a reversible Finsler length structure on \(M\) (Definition 13.4). Its distance \(d_b\) (Definition 13.10) is then a pseudo-distance, and \((M, d_b)\) is a pseudo-metric space; without reversibility \(d_b\) is only a quasi-distance and the symmetric theory of the previous section does not apply as such.

Proof ▶

\(d_b(\gamma (s), \gamma (t)) \leq |L_b(\gamma |_{[s, t]})|\) gives the continuity. For a partition, \(\sum d_b(\gamma (t_i), \gamma (t_{i+1})) \leq \sum L_b(\gamma |_{[t_i, t_{i+1}]}) \leq L_b(\gamma )\). Hence \(d_\ell \) is bounded by the length of every chain of \(C^{1}\) arcs, so \(d_\ell \leq d_b\); the other inequality is Proposition 12.2.

13.8 Bounded Finsler structures and the topology

In this section the model \(E\) is a normed space; from Lemma 13.49 on, it is complete and \(M\) is Hausdorff.

\(\mu \) is bounded below if every point has an open neighborhood \(U\) in the domain of a chart \(\varphi \) and \(C {\gt} 0\) with \(\| \varphi _{*x} v\| \leq C \mu _x(v)\) for \(x \in U\); it is bounded above if \(m \mu _x(v) \leq \| \varphi _{*x} v\| \) for some \(m {\gt} 0\). A Finsler structure bounded above gives a weak Finsler manifold; bounded above and below, a strong one.

Remark 13.47 Locally convex models
#

If the model is not normable, no Finsler norm \(\mu _x\) can be bounded above and below by a norm inducing the topology of the model, since that topology would then be a norm topology. Strong Finsler manifolds are a phenomenon of normed models.

Lemma 13.48 The integrated bound
✓

If \(\mu \) is bounded below on \(U\) with constant \(C\) and \(E\) is complete, a \(C^{1}\) curve that stays in \(U\) satisfies \(\| \varphi (\gamma (t)) - \varphi (\gamma (s))\| \leq C L_b(\gamma |_{[s, t]})\) [ Lar26 , (9.2) ] .

Proof ▶

Integrate the derivative of \(\varphi \circ \gamma \) and bound its norm by \(C \mu (\dot\gamma )\).

Let \(\mu \) be bounded below on \(U\) with constant \(C\), and let \(\bar B(\varphi (x_0), r)\) be the image of a part of \(U\). A chain of \(C^{1}\) arcs of length \(L\) starting at \(x_0\) either has \(C L \geq r\), or ends at a point \(y \in U\) with \(\| \varphi (y) - \varphi (x_0)\| \leq C L\).

Proof ▶

For a single arc, let \(t_0\) be the first time at which the arc leaves \(U \cap \varphi ^{-1}(B(\varphi (x_0), r))\). If \(C L_b(\gamma |_{[a, t_0]}) {\lt} r\), the points \(\varphi (\gamma (t))\), \(t {\lt} t_0\), form a Cauchy family in a closed ball inside \(\varphi (U)\); by completeness they converge, and since \(M\) is Hausdorff the limit is \(\varphi (\gamma (t_0))\), which is then inside the ball: a contradiction. Chains follow by induction.

If \(\mu \) is bounded below, every point \(x_0\) has a chart and \(C, r {\gt} 0\) such that \(d_b(x_0, y) {\lt} r / C\) implies that \(y\) is in the chart and \(\| \varphi (y) - \varphi (x_0)\| \leq C d_b(x_0, y)\). Consequently \(d_b\) separates points, and, for a reversible \(\mu \), every open set of \(M\) is open for \(d_b\).

Proof ▶

Apply Lemma 13.49 to chains whose length approaches \(d_b(x_0, y)\).

For every incomplete normed space \(E_0\) there is a Hausdorff manifold \(M\) modeled on \(\mathbb {R} \times E_0\) with a reversible, nondegenerate Finsler metric, bounded above and below, such that the identity \((M, d_b) \to M\) is not continuous. This applies to \(c_{00}\) with the \(\ell ^2\) norm. So the completeness of the model cannot be dropped from Theorem 13.50 nor from Corolario 9.1.9 of [ Lar26 ] .

Proof ▶

Every incomplete normed space carries a smooth curve \(c \colon (-\infty , 1) \to E_0\) with \(\| c'\| \le K\), \(c'(s) \to 0\) as \(s \to 1^-\) and no cluster point in \(E_0\) as \(s \to 1^-\): move along a non-convergent sequence with fast decaying increments by smooth steps on the intervals \([1 - 2^{-k}, 1 - 2^{-k-1}]\). Glue two copies of \(E = \mathbb {R} \times E_0\) along \(\{ s {\lt} 1\} \) by \(\tau (s, v) = (s, v + c(s))\). The result is Hausdorff because \(c\) has no cluster point. The norm of \(E\) read in the two charts is a Finsler metric, continuous because \(c' \to 0\) and bounded above and below because \(\| c'\| \le K\). For \(p = (1, 0)\) in the first copy and \(q_\delta = (1, c(1 - \delta ))\) in the second, going down the first copy to \((1 - \delta , 0)_1 = (1 - \delta , c(1 - \delta ))_2\) and up the second gives \(d_b(p, q_\delta ) \le (2 + K)\delta \), while no \(q_\delta \) lies in the first copy, an open neighborhood of \(p\). The exit argument of Lemma 13.49 fails exactly where completeness is used: the limit lies in the completion of \(E\). We found no source for this counterexample; we do not claim it is new.

If \(\mu \) is bounded above with constant \(m\), then \(d_b(x, y) \leq \| \varphi (y) - \varphi (x)\| / m\) near each point [ Lar26 , (9.3) ] , so the topology of \(d_b\) is coarser than that of \(M\). For a strong Finsler manifold (reversible, complete model, \(M\) Hausdorff) the identity is a homeomorphism between \(M\) and \((M, d_b)\) [ Lar26 , Cor. 9.1.9 ] .

Proof ▶

The segment from \(\varphi (x)\) to \(\varphi (y)\), reparametrized by a smooth transition function so that it is a \(C^{1}\) curve on all of \(\mathbb {R}\), has Finsler length at most \(\| \varphi (y) - \varphi (x)\| / m\). The second statement combines this with Theorem 13.50.

Proof ▶

Near \((\varphi (p), 0)\) the Lipschitz condition gives \(\mu _q(w) \leq K(\| \varphi (q) - \varphi (p)\| + \| \varphi _{*q} w\| )\); applied to a multiple of \(w\) of fixed small norm and combined with homogeneity, it gives \(\mu _q(w) \leq 2K \| \varphi _{*q} w\| \) for \(q\) close to \(p\).

If the model is finite-dimensional, every Finsler metric whose seminorms are norms is strong.

Proof ▶

In a chart, \((y, v) \mapsto \mu (y, v)\) is continuous on the tangent bundle, hence attains a positive minimum and a maximum on the compact set \(\bar B \times S\), closed ball times unit sphere. Homogeneity gives the bounds for every vector.

Let \(M\) be a Hausdorff manifold modeled on a finite-dimensional normed space, with a reversible Finsler metric whose seminorms are norms and with finite distance \(d_b\). If \((M, d_b)\) is complete, any two points are joined by a continuous curve whose length is their distance.

Proof ▶

By Theorem 13.54 the metric is strong; by Theorems 13.50 and 13.52, \(d_b\) is a distance and \((M, d_b)\) is homeomorphic to \(M\), hence locally compact. It is an interior space by Theorem 13.45, and Theorem 12.9 applies.

Whether these shortest curves can be taken piecewise \(C^{1}\) is a separate question [ Lar26 , Obs. 9.1.11 ] .