Manifolds modeled on locally convex spaces

11 Sprays and geodesics

Sprays, the correspondence between quadratic sprays and symmetric connections, and geodesics on Banach manifolds: the geodesic equation in charts, local existence with uniform time, uniqueness, maximal geodesics, the exponential map and normal neighbourhoods, after Lang [ Lan99 , § § IV.3–4, VIII.5 ] . In general locally convex spaces there is no existence theorem for ordinary differential equations, so existence is only claimed for Banach models.

11.1 Sprays

A spray, or second order vector field, is a vector field \(\xi \) on \(TM\) such that \(T\pi \circ \xi = \mathrm{id}\). In a chart it has the form \((x, v) \mapsto (x, v; v, f(x, v))\), and it is determined by its local part \(f\). As we did for connections, we encode sprays through their local parts.

A spray on \(M\) assigns to each chart \(i\) of the atlas a smooth map \(f_i \colon i(U_i) \times E \to E\) such that, for every pair of charts \(i, j\), every \(y\) in the domain of \(\tau = j \circ i^{-1}\) and every \(v \in E\),

\[ f_j\bigl(\tau y, \mathrm{d}\tau (y)(v)\bigr) = \mathrm{d}\tau (y)\bigl(f_i(y, v)\bigr) + \mathrm{d}^{(2)}\tau (y; v, v). \]

The spray is quadratic if \(f_i(x, s v) = s^2 f_i(x, v)\) for all \(i\), \(x\), \(v\) and \(s \in \mathbb {R}\).

The change of charts law is the change of charts law of the double tangent bundle \(T(TM)\), restricted to the points of the form \((x, v; v, w)\). In [ Lan99 ] , §IV.3, a spray is by definition quadratic; we keep the two notions apart, since the quadratic condition is what is needed to pass to connections.

Every connection defines a quadratic spray, with local parts \(f_i(x, v) = \Gamma _i(x)(v, v)\).

Proof ▶

Restrict the change of charts law of \(\Gamma \) to the diagonal \(a = b = v\). Homogeneity of degree two follows from bilinearity.

The converse requires recovering a bilinear map from its quadratic form, and checking that the polarization obeys the change of charts law of connections. For the first point, the following identity holds without any regularity assumption.

Let \(g \colon E \to F\) satisfy \(g(s v) = s^2 g(v)\) for all \(s \in \mathbb {R}\), \(v \in E\). Then \(\mathrm{d}^{(2)} g(0; v, v) = 2\, g(v)\). Moreover, if \(B\) is symmetric and bilinear, then \(B(a + b, a + b) - B(a - b, a - b) = 4 B(a, b)\).

Proof ▶

By homogeneity, \(g(s v + t v) - g(s v) = \bigl((s + t)^2 - s^2\bigr) g(v)\), so the difference quotient can be computed exactly and \(\mathrm{d}g(s v)(v) = 2 s\, g(v)\) for every \(s\). Differentiating once more along the line \(s \mapsto s v\) at \(s = 0\) gives \(2 g(v)\). The polarization identity is immediate.

Let \(f\) be a quadratic spray. In each chart \(i\) put

\[ \Gamma _i(x)(a, b) = \tfrac 14\bigl(f_i(x, a + b) - f_i(x, a - b)\bigr) = \tfrac 12\, \mathrm{d}^{(2)}\bigl(f_i(x, \cdot )\bigr)(0; a, b). \]

Then the \(\Gamma _i\) are the symbols of a symmetric connection, whose spray is \(f\).

Proof ▶

By Lemma 11.3 applied to \(g = f_i(x, \cdot )\) and to the symmetric bilinear map \(\tfrac 12 \mathrm{d}^{(2)} g(0; \cdot , \cdot )\), the two expressions agree; the second one is bilinear and symmetric, and \(\Gamma _i(x)(v, v) = f_i(x, v)\). Smoothness of the symbols follows from the first expression. The change of charts law for \(\Gamma \) is obtained by polarizing the change of charts law of the spray, using that \(\mathrm{d}\tau (y)\) is linear and \(\mathrm{d}^{(2)}\tau (y; \cdot , \cdot )\) is symmetric and bilinear. This is [ Lan99 ] , Prop. IV.3.2 and §VIII.2.

Thus quadratic sprays and symmetric connections are the same object, through \(f(x, v) = \Gamma _x(v, v)\); in the rest of the chapter we work with connections.

11.2 Geodesics on Banach manifolds

In this section \(E\) is a Banach space. Then Michal–Bastiani and Fréchet derivatives agree, and in a chart \(i\) the geodesic equation is the second order equation

\[ c'' = \Gamma _i(c)(c', c') \]

on the open set \(\varphi _i(U_i) \subseteq E\), with \(\Gamma _i\) smooth and bilinear in the last two variables. We say that \(\alpha \) is a geodesic on an open set \(I \subseteq \mathbb {R}\) if it is continuous on \(I\) and satisfies this equation at every \(t \in I\) in the preferred chart at \(\alpha (t)\).

Theorem 11.5 The geodesic equation does not depend on the chart
✓

If \(\alpha \) is continuous at \(t\) and satisfies the geodesic equation at \(t\) in a chart around \(\alpha (t)\), it satisfies it in every chart around \(\alpha (t)\).

Proof ▶

With \(\tau = \varphi _j \circ \varphi _i^{-1}\) and \(c = \varphi _i \circ \alpha \), \((\tau \circ c)'' = D^2\tau (c', c') + D\tau (c'')\), and the change of charts law of the symbols reads \(\Gamma _j(\tau y)(D\tau \, a, D\tau \, b) = D\tau \, \Gamma _i(y)(a, b) + D^2\tau (a, b)\).

Theorem 11.6 Local existence of geodesics, with uniform time
✓

Let \(y_0\) be a point of the chart \(i\). There is \(r {\gt} 0\) such that for every \(K \ge 0\) there is \(\varepsilon {\gt} 0\) with the following property: for every \(y\) with \(\| y - y_0\| \le r\) and every \(w\) with \(\| w\| \le K\) there is a geodesic \(\alpha \) on \((-\varepsilon , \varepsilon )\), staying in the domain of \(i\), with \(\varphi _i(\alpha (0)) = y\) and \((\varphi _i \circ \alpha )'(0) = w\).

Proof ▶

Picard–Lindelöf for the first order system \((c, v)' = (v, \Gamma _i(c)(v, v))\), whose right hand side is \(C^{1}\), hence locally Lipschitz; the time is uniform for small velocities, and the homogeneity of the equation (\(t \mapsto c(st)\) is a solution with initial velocity \(s\, c'(0)\)) makes it uniform for velocities bounded by \(K\).

Theorem 11.7 Uniqueness of geodesics
✓

If \(M\) is Hausdorff, two geodesics on \((a, b)\) with the same position and velocity at some \(t_0 \in (a, b)\) agree on \((a, b)\).

Proof ▶

The set of \(t\) near which the two geodesics agree is open. It is also closed in \((a, b)\): at a limit point \(t\) the two curves meet, because \(M\) is Hausdorff; in a chart around this common point both solve the same equation, which has unique solutions since its right hand side is locally Lipschitz. By connectedness it is all of \((a, b)\).

Remark 11.8
#

The Hausdorff hypothesis cannot be dropped: on the line with two origins, the two copies of a line through the origins are geodesics of the flat connection with the same data at \(t = 1\), and they differ at \(t = 0\).

Theorem 11.9 Extension of geodesics
✓

Let \(M\) be Hausdorff, and let \(\alpha \) be a geodesic on \((a, b)\) such that \(\alpha (t) \to q\) as \(t \to b^-\) and the velocity of \(\alpha \), read in a chart around \(q\), stays bounded as \(t \to b^-\). Then \(\alpha \) extends to a geodesic on \((a, b')\) for some \(b' {\gt} b\).

Proof ▶

Let \(\varepsilon \) be the uniform time of Theorem 11.6 around \(\varphi _i(q)\). Take \(t_1 {\gt} b - \varepsilon /2\) with \(\alpha (t_1)\) close to \(q\), and the geodesic \(\mu \) with the position and velocity of \(\alpha \) at \(t_1\). By uniqueness \(\mu (t - t_1) = \alpha (t)\) where both are defined, so the two curves glue to a geodesic on \((a, t_1 + \varepsilon )\), and \(t_1 + \varepsilon {\gt} b\).

Theorem 11.10 Complete Finsler manifolds: geodesics extend [ Lar26 , Prop. 9.2.1 ]
✓

Let \(M\) be Hausdorff, with a Finsler length structure \(\mu \) bounded below, such that \(M\) is complete for the symmetrized distance \(\max (d_b(x, y), d_b(y, x))\). If \(\alpha \) is a geodesic on \((a, b)\) with \(\mu (\dot\alpha )\) and \(\mu (-\dot\alpha )\) bounded, then \(\alpha \) extends to a geodesic on an interval \((a', b') \supsetneq (a, b)\) at both ends.

Proof ▶

Reparametrizing arcs of \(\alpha \) by a smooth ramp gives \(C^{1}\) competitors for \(d_b\), so \(d_b(\alpha (s), \alpha (t)) \le L(t - s)\) and the same backwards; thus \(\alpha (t)\) is Cauchy as \(t \to b^-\). Its limit in \((M, d_b)\) is a limit in \(M\) because \(\mu \) is bounded below, and the lower bound also bounds the velocity in a chart around the limit point. Theorem 11.9 applies; the end \(a\) is treated by reversing time.

Remark 11.11
#

In [ Lar26 ] the metric is invariant under the parallel transport of the spray, so geodesics have constant speed; we only use bounded speed. Since \(\mu \) need not be reversible, both \(\mu (\dot\alpha )\) and \(\mu (-\dot\alpha )\) have to be bounded. With maximal geodesics, the statement “every geodesic is defined on \(\mathbb {R}\)” is Theorem 11.20.

Let \(M\) be Hausdorff and modeled on a Banach space, with a connection. The geodesics with initial velocity \(v \in TM\) glue to a maximal geodesic \(\gamma _v\), defined on an open interval \(J_v \ni 0\): every geodesic with initial velocity \(v\) on an open interval \(I \ni 0\) is the restriction of \(\gamma _v\) to \(I\). For \(s \neq 0\), \(J_{sv} = s^{-1}J_v\) and \(\gamma _{sv}(t) = \gamma _v(st)\); the maximal geodesic with zero initial velocity is constant and defined on \(\mathbb {R}\).

Proof ▶

As for the maximal flow (Theorem 5.4), from local existence and uniqueness on intervals; homogeneity because \(t \mapsto \alpha (st)\) is a geodesic with initial velocity \(s\, \dot\alpha (0)\). The maximal geodesic is built directly from the geodesics, not as the projection of the maximal flow of the geodesic spray on \(TM\); the two agree by Theorem 11.15. We do not claim this argument is new.

The exponential map is \(\exp v = \gamma _v(1)\), on \(D(\exp ) = \{ v : 1 \in J_v\} \), and \(\exp _p\) is its restriction to \(T_pM\). Then \(tv \in D(\exp )\) if and only if \(t \in J_v\), and \(\exp (tv) = \gamma _v(t)\); so the maximal geodesic is \(t \mapsto \exp (tv)\).

Let \(M\) be modeled on a Banach space, with a connection. The geodesic spray is the smooth vector field \(S\) on the manifold \(TM\) (modeled on \(E \times E\)) which, in the chart of \(TM\) over a chart \(i\) of \(M\), reads

\[ (x, w) \longmapsto (x, w;\ w, \Gamma _i(x)(w, w)). \]

It is well defined: under the transition maps of \(TTM\), the local formulas in two charts agree by the change of charts law of the symbols. It is a second-order vector field, \(\pi _* S(v) = v\).

The velocity \(\dot\alpha \) of a geodesic is an integral curve of \(S\); conversely, an integral curve \(\beta \) of \(S\) is the velocity of its projection \(\pi \beta \), which is a geodesic. Hence, if \(M\) is Hausdorff, the maximal interval \(J_v\) of the geodesic with initial velocity \(v\) is the maximal interval of the integral curve of \(S\) through \(v\), \(\gamma _v(t) = \pi (\mathrm{Fl}_S(t, v))\) and \(\mathrm{Fl}_S(t, v) = \dot\gamma _v(t)\).

Proof ▶

In the chart of \(TM\) over a chart \(i\) of \(M\), an integral curve of \(S\) is a curve \((x, w)\) with \(x' = w\) and \(w' = \Gamma _i(x)(w, w)\), which is the geodesic equation for \(x\).

Let \(M\) be Hausdorff and modeled on a Banach space, with a connection. The domain \(D(\exp )\) is open in \(TM\), contains the zero section and is star-shaped in the fibers, and \(\exp \colon D(\exp ) \to M\) is smooth. For each \(p \in M\), \(D(\exp _p)\) is an open star-shaped neighbourhood of \(0\) in \(T_pM\), and \(\exp _p\) is smooth on it.

Proof ▶

By Theorem 11.15, \(D(\exp ) = \{ v : (1, v) \in D(\mathrm{Fl}_S)\} \) and \(\exp = \pi \circ \mathrm{Fl}_S(1, \cdot )\); the domain of the maximal flow is open and the flow is smooth on it (Theorem 5.4, applied to the Hausdorff manifold \(TM\)). Star-shapedness is \(tv \in D(\exp ) \iff t \in J_v\). The library works with smooth manifolds, so \(\exp \) is smooth, rather than \(C^{p-2}\) on a \(C^p\) manifold as in [ Lan99 ] . That \(\exp _p\) is a local diffeomorphism at \(0\) is Theorem 11.17.

Let \(M\) be Hausdorff and modeled on a Banach space, with a connection, and let \(p \in M\). Read in the preferred chart \(\varphi \) at \(p\), the derivative at \(0\) of \(w \mapsto \varphi (\exp _p w)\) is the identity: \(d(\exp _p)(0_p) = \mathrm{id}\). Hence \(\exp _p\) restricts to a diffeomorphism from an open neighbourhood of \(0\) in \(T_pM\), contained in \(D(\exp _p)\), onto an open neighbourhood of \(p\).

Proof ▶

Since \(\exp _p(tw) = \gamma _w(t)\) for \(t \in J_w\) and \(\gamma _w'(0) = w\), the derivative of \(\varphi \circ \exp _p\) at \(0\) in the direction \(w\) is \(w\). Michal–Bastiani smooth maps between open sets of Banach spaces are Fréchet smooth, so the inverse function theorem applies to \(\varphi \circ \exp _p\); on the open set where its derivative is invertible the local inverse is smooth, and composing with \(\varphi ^{-1}\) gives the diffeomorphism onto an open set of \(M\).

Under the hypotheses of Theorem 11.17, the map \(G(v) = (\pi v, \exp v)\), \(D(\exp ) \to M \times M\), has differential \((a, b) \mapsto (a, a + b)\) at \(0_p\) in charts, and restricts to a diffeomorphism from an open neighbourhood \(V \subseteq D(\exp )\) of \(0_p\) onto an open neighbourhood of \((p, p)\). There is an open \(W \ni p\) with \(W \times W \subseteq G(V)\): any \(x, y \in W\) are joined by a unique geodesic \(t \mapsto \exp (tv)\), \(0 \le t \le 1\), with \(v \in V\), and \(v\) depends smoothly on \((x, y)\) [ Lar26 , Teo. 6.2.3 (1) ] , [ Lan99 , Cor. VIII.5.2 (1) ] .

Proof ▶

In the chart of \(TM\) over \(\varphi \) and the chart \(\varphi \times \varphi \) of \(M \times M\), \(G\) reads \((x, u) \mapsto (x, \varphi (\exp (x, u)))\). Since \(G(x, 0) = (x, x)\) and \(d(\exp _p)(0) = \mathrm{id}\) (Theorem 11.17), its differential at \((\varphi p, 0)\) is \((a, b) \mapsto (a, a + b)\), invertible; the inverse function theorem concludes. Uniqueness is injectivity of \(G\) on \(V\).

Under the hypotheses of Theorem 11.17, let \(\varphi \) be the preferred chart at \(p\) and read tangent vectors in the chart \(T\) of \(TM\) over \(\varphi \). There are an open \(W \ni p\) in the domain of \(\varphi \) and \(r {\gt} 0\) such that for every \(q \in W\) the vectors of \(T_qM\) with coordinates \((\varphi q, u)\), \(\| u\| {\lt} r\), lie in \(D(\exp )\), and \(u \mapsto \exp _q(T^{-1}(\varphi q, u))\) is a diffeomorphism of \(B_r(0) \subseteq E\) onto an open set containing \(W\).

Proof ▶

The local inverse of \(G\) in charts (Theorem 11.18) is defined on a neighbourhood of \((\varphi p, 0)\) in \(E \times E\), which contains a product of balls \(B_r(\varphi p) \times B_r(0)\); its image is an open neighbourhood of \((\varphi p, \varphi p)\), which contains \(B_\delta (\varphi p) \times B_\delta (\varphi p)\). Take \(W = \varphi ^{-1}(B_{\min (\delta , r)}(\varphi p))\). Since \(G\) preserves the first coordinate, slicing at \(x = \varphi q\) gives the diffeomorphism at \(q\). No local compactness is used: the radius \(r\) is uniform because it comes from the inverse function theorem applied to \(G\) at the single point \(0_p\). We do not claim this argument is new. The book also states the result for every \(r \le R\) and that the joining geodesic stays in \(\exp _x(B_R(0))\); these refinements are not formalized.

Under the hypotheses of Theorem 11.10, assume that every geodesic on a bounded open interval has bounded speed \(\mu (\dot\alpha )\) and bounded backward speed \(\mu (-\dot\alpha )\). Each assertion implies the next:

  1. \(M\) is complete for the symmetrized distance;

  2. every maximal geodesic is defined on \(\mathbb {R}\);

  3. \(M\) is geodesically complete: \(D(\exp ) = TM\);

  4. there is \(p \in M\) with \(D(\exp _p) = T_pM\).

Proof ▶

\(1 \Rightarrow 2\): if \(b = \sup J_v\) were finite, Theorem 11.10 would extend \(\gamma _v\) past \(b\), against maximality; boundedness below reduces to boundedness above through \(J_{-v} = -J_v\). \(2 \Rightarrow 3 \Rightarrow 4\) are immediate from Definition 11.13. We do not claim this argument is new.

Remark 11.21 What fails in infinite dimensions
#

In finite dimensions the Hopf–Rinow theorem makes 1–4 equivalent for Riemannian (and reversible Finsler) manifolds, and adds that \(\exp _p\) is onto. In infinite dimensions the last conclusion fails: Atkin constructs a complete Riemannian Hilbert manifold, the unit sphere of \(\mathbb {R}\times \ell ^2\) with the metric induced by \(|\xi |^2 + \langle Tx, x\rangle \) for a suitable strictly positive operator \(T\), whose two poles are joined by no geodesic [ Atk75 , (3), p. 262, and Lemma 6, p. 265 ] , so that \(\exp _p\) is not onto (see also [ Eke78 , p. 288 ] ). Even shortest curves need not exist: on the ellipsoid image of the unit sphere of a Hilbert space by \(A = \mathrm{diag}(1, 1 + 2^{-1}, 1 + 2^{-2}, \dots )\), a complete Riemannian Hilbert manifold, \(e_0\) and \(-e_0\) are at distance \(\pi \) but every curve joining them is strictly longer (McAlpin, as presented in [ Lar26 , Ejemplo 10.4.2, pp. 301–302 ] ). No local compactness is used in the proof above: the extension rests on the uniform existence time of geodesics with bounded initial velocity in a chart and on the lower bound of \(\mu \).