10 Connections
Connections on a manifold \(M\) modeled on a locally convex space \(E\), given in charts by their Christoffel symbols: the covariant derivative of vector fields and of lifts of curves, the torsion and the curvature, connection forms and Cartan’s structure equations, and geodesics as the curves with zero covariant acceleration. All the objects are defined chart by chart, subject to a change of charts law, and the main point in each case is that the resulting object does not depend on the chart. We follow Lang [ Lan99 ] (§§IV.3, VIII.1–3, IX.1) and Chapter 6 of [ Lar26 ] .
The last section is closer to [ Lar26 ] : a smooth field of projectors onto the tangent spaces of an embedded submanifold of a locally convex space defines a symmetric connection, whose covariant derivative is obtained by differentiating in the ambient space and projecting, and whose geodesics are the curves with normal acceleration.
There is no good topology on spaces of continuous linear operators, so bilinear families such as \(x \mapsto \Gamma _x\) or \(p \mapsto E_p\) are encoded as maps of several variables, linear in some of them, with joint smoothness. Existence of geodesics is only proved on Banach manifolds, in Chapter 11.
10.1 Connections and the covariant derivative
Let \(\tau \) be a \(C^{2}\) map between open subsets of locally convex spaces. We write \(\mathrm{d}\tau (y)(a)\) for its derivative and \(\mathrm{d}^{(2)}\tau (y; a, b)\) for the derivative of \(y \mapsto \mathrm{d}\tau (y)(b)\) in the direction \(a\); by the symmetry of second derivatives the latter is symmetric in \((a, b)\). For two charts \(i, j\) of the atlas of \(M\) we write \(\tau = j \circ i^{-1}\) for the transition map, defined on \(i(U_i \cap U_j)\).
A connection on \(M\) assigns to each chart \(i\) of the atlas a family of maps \(\Gamma _i(x) \colon E \times E \to E\), \(x \in i(U_i)\), such that
each \(\Gamma _i(x)\) is bilinear, and \((x, a, b) \mapsto \Gamma _i(x)(a, b)\) is smooth on \(i(U_i) \times E \times E\);
(change of charts law) for every pair of charts \(i, j\), every \(y\) in the domain of \(\tau = j \circ i^{-1}\) and all \(a, b \in E\),
\[ \Gamma _j(\tau y)\bigl(\mathrm{d}\tau (y)(a), \mathrm{d}\tau (y)(b)\bigr) = \mathrm{d}\tau (y)\bigl(\Gamma _i(y)(a, b)\bigr) + \mathrm{d}^{(2)}\tau (y; a, b). \]
The maps \(\Gamma _i\) are the Christoffel symbols of the connection in the chart \(i\).
The symbols are not tensorial: they transform affinely, the defect being the second derivative of the transition map. Our convention is that the first argument of \(\Gamma \) is the direction in which the base point moves. Continuity of each bilinear map \(\Gamma _i(x)\) is a consequence of the joint smoothness, and no topology on the space of bilinear maps is needed.
Let \(\Gamma \) be a family of bilinear maps on an open set \(U \subseteq E\) and let \(X, Y \colon U \to E\). The local covariant derivative is
Let \(\tau \) be \(C^{2}\), and let \(X_i, Y_i\) and \(X_j, Y_j\) be related by \(\tau \), that is \(X_j \circ \tau = \mathrm{d}\tau \cdot X_i\) and \(Y_j \circ \tau = \mathrm{d}\tau \cdot Y_i\), with \(Y_i, Y_j\) of class \(C^{1}\). If \(\Gamma _i, \Gamma _j\) obey the change of charts law of Definition 10.1, then
Differentiating \(Y_j \circ \tau = \mathrm{d}\tau \cdot Y_i\) in the direction \(w\) gives, by the second-order chain rule, \(\mathrm{d}Y_j(\tau y)(\mathrm{d}\tau (y) w) = \mathrm{d}^{(2)}\tau (y; w, Y_i y) + \mathrm{d}\tau (y)(\mathrm{d}Y_i(y) w)\). Take \(w = X_i(y)\): the term \(\mathrm{d}^{(2)}\tau (y; X_i y, Y_i y)\) cancels with the one in the change of charts law of \(\Gamma \), and the rest is linear in \(\mathrm{d}\tau (y)\).
Let \(C\) be a connection on \(M\) and \(X, Y\) vector fields. The covariant derivative \(\nabla _X Y\) is the vector field whose value at \(x\) is the local covariant derivative of the local representatives \(\tilde X, \tilde Y\) of \(X, Y\), computed in the preferred chart at \(x\): \(\mathrm{d}\tilde Y \cdot \tilde X - \Gamma (\tilde X, \tilde Y)\).
Let \(Y\) be a \(C^{1}\) vector field and \(X\) any vector field. In every chart \(j\) of the atlas, the local representative of \(\nabla _X Y\) is \(\mathrm{d}\tilde Y_j \cdot \tilde X_j - \Gamma _j(\tilde X_j, \tilde Y_j)\).
Apply Lemma 10.3 to the transition map from the preferred chart to \(j\), which is smooth; the local representatives of vector fields are related by its derivative.
So the choice of the preferred chart in Definition 10.4 is only a matter of convenience. The usual algebraic rules follow at once from the local formula.
Let \(X, Y\) be vector fields with \(Y\) of class \(C^{1}\), and let \(\varphi \colon M \to \mathbb {R}\).
\(\nabla _{\varphi X} Y = \varphi \, \nabla _X Y\).
If \(\varphi \) is \(C^{1}\), then \(\nabla _X(\varphi Y) = X(\varphi )\, Y + \varphi \, \nabla _X Y\) (Leibniz rule).
If \(X\) is \(C^{k}\) and \(Y\) is \(C^{k+1}\), then \(\nabla _X Y\) is \(C^{k}\).
The first two items are computed in the preferred chart: \(\Gamma \) is linear in its first argument, and the derivative of \(\varphi \tilde Y\) in the direction \(\tilde X\) is \(\mathrm{d}\varphi (\tilde X)\, \tilde Y + \varphi \, \mathrm{d}\tilde Y(\tilde X)\). For the third, in each chart \(y \mapsto \mathrm{d}\tilde Y(y)(\tilde X y)\) and \(y \mapsto \Gamma (y)(\tilde X y, \tilde Y y)\) are compositions of \(C^{k}\) maps.
For vector fields \(X, Y\), the vector \(\nabla _X Y - \nabla _Y X - [X, Y]\) at \(x\) is, in the preferred chart, \(\Gamma (\tilde Y, \tilde X) - \Gamma (\tilde X, \tilde Y)\); in particular it only depends on \(X(x)\) and \(Y(x)\). We say that the connection is symmetric (torsion free) if its symbols are symmetric in every chart; in that case \(\nabla _X Y - \nabla _Y X = [X, Y]\).
The Lie bracket in a chart is \(\mathrm{d}\tilde Y \cdot \tilde X - \mathrm{d}\tilde X \cdot \tilde Y\), and the derivative terms cancel.
Symmetry of a connection is a chart-free condition: the change of charts law shows that if \(\Gamma _i\) is symmetric then so is \(\Gamma _j\), because \(\mathrm{d}^{(2)}\tau \) is symmetric and \(\mathrm{d}\tau (y)\) is invertible on the overlap. In the Lean code the definition quantifies over all the charts of the atlas, which avoids this argument.
10.2 Curvature
We use Lang’s sign for the curvature.
For vector fields \(X, Y, Z\), the curvature at \(x\) is
Locally, for a family of symbols \(\Gamma \) on an open set of \(E\), we write \(\mathrm{d}\Gamma _x(v)(a, b)\) for the derivative of \(x \mapsto \Gamma (x)(a, b)\) in the direction \(v\), and
Let \(X, Y, Z\) be \(C^{2}\) fields on an open set \(U \subseteq E\) and \(\Gamma \) a \(C^{1}\) family of bilinear maps. Then, pointwise on \(U\), \(\nabla _X \nabla _Y Z - \nabla _Y \nabla _X Z - \nabla _{[X, Y]} Z = R(X, Y) Z\) with \(R\) the trilinear expression of Definition 10.9.
Expand both covariant derivatives. The derivative of \(y \mapsto \Gamma (y)(A y, B y)\) is \(\mathrm{d}\Gamma _y(h)(A y, B y) + \Gamma _y(\mathrm{d}A(y) h, B y) + \Gamma _y(A y, \mathrm{d}B(y) h)\). All the terms containing second derivatives of \(Z\) cancel by the symmetry of second derivatives, and the terms containing first derivatives of \(X\), \(Y\), \(Z\) cancel with \(\nabla _{[X,Y]} Z\) and among themselves. This is Lang [ Lan99 ] , Prop. IX.1.2.
Let \(X, Y, Z\) be \(C^{2}\) vector fields. Then \(R(X, Y) Z\) at \(x\) is the trilinear expression \(R_{\tilde x}(\tilde X, \tilde Y)\tilde Z\) in the preferred chart, evaluated at the values of \(X, Y, Z\) at \(x\). In particular, it only depends on \(X(x)\), \(Y(x)\) and \(Z(x)\).
Near \(x\) the local representatives of \(\nabla _Y Z\), \(\nabla _X Z\) and \([X, Y]\) in the preferred chart at \(x\) are given by the local formulas (Theorem 10.5), and the local covariant derivative at a point only depends on the germ of the second field. Then apply Lemma 10.10. Compare [ Lan99 ] , Prop. IX.1.1–1.2.
For \(C^{2}\) vector fields, \(R(Y, X) Z = -R(X, Y) Z\). If the connection is symmetric, the first Bianchi identity holds: \(R(X, Y) Z + R(Y, Z) X + R(Z, X) Y = 0\).
Both identities are checked on the trilinear expression. Skew-symmetry is immediate. For Bianchi, if \(\Gamma \) is symmetric then so is \(\mathrm{d}\Gamma _x(v)\), and the twelve terms of the cyclic sum cancel in pairs.
10.3 Connection forms and the structure equations
Classically, the connection form of a connection on a vector bundle is the \(\mathrm{End}(F)\)-valued \(1\)-form \(A(y)(v) = \Gamma (y)(v, \cdot )\) in a local trivialization, so that \(\nabla = \mathrm{d}- A\), and the curvature is \(\Omega = -\mathrm{d}A + A \wedge A\). If the fiber \(F\) is not normable, no vector topology on \(L(F)\) makes the evaluation \(L(F) \times F \to F\) continuous [ Mai63 , Satz 1 ] , so \(A\) is not a differential form with values in a locally convex space of operators. We encode \(\mathrm{End}(F)\)-valued forms as families of \(F\)-valued forms depending linearly on an extra argument \(\xi \in F\). The local statements are proved for an arbitrary fiber \(F\); the tangent bundle is the case \(F = E\).
Let \(U \subseteq E\) be open and \(\Gamma \colon U \times E \times F \to F\) smooth and bilinear in the last two variables. For \(\xi \in F\), the connection form is the local \(1\)-form \(A^\xi (y)(v) = \Gamma (y)(v, \xi )\), and the curvature form is the local \(2\)-form
We write \((A \wedge A)(u, v)\xi = A(u)A(v)\xi - A(v)A(u)\xi \).
\(\Omega ^\xi = -\mathrm{d}A^\xi + (A \wedge A)\xi \). For the tangent bundle, \(\Omega \) is the local curvature of Lemma 10.10, so the curvature of a connection on \(M\), evaluated on vector fields, is computed in the preferred chart by \(-\mathrm{d}A + A \wedge A\).
\(\mathrm{d}A^\xi (y)(u, v) = \partial _u\Gamma (y)(v, \xi ) - \partial _v\Gamma (y)(u, \xi )\).
The covariant exterior derivative of the curvature vanishes, \(D\Omega = \mathrm{d}\Omega - [A \wedge \Omega ] = 0\): for \(u, v, t \in E\) and \(\xi \in F\),
where \([A(a), \Omega (b, c)]\xi = A(a)\Omega (b, c)\xi - \Omega (b, c)A(a)\xi \).
By the structure equation and \(\mathrm{d}\circ \mathrm{d}= 0\), \(\mathrm{d}\Omega ^\xi = \mathrm{d}((A \wedge A)\xi )\), which is computed with the product rule \(\partial _c(\Gamma (a, \Gamma (b, \xi ))) = \partial _c\Gamma (a, \Gamma (b, \xi )) + \Gamma (a, \partial _c\Gamma (b, \xi ))\). The terms with three factors \(\Gamma \) on the right cancel by the Jacobi identity for commutators.
Let \(\tau = j \circ i^{-1}\) and \(g(y) = \mathrm{d}\tau (y)\). The connection forms of a connection on \(M\) satisfy \(A_j(\tau y)(g v)(g w) = g\, A_i(y)(v) w + (\partial _v g)(y) w\), so \(A\) is not a tensor. The torsion \(T(u, v) = A(u)v - A(v)u = \Gamma (u, v) - \Gamma (v, u)\) is a tensor: \(T_j(\tau y)(g a, g b) = g\, T_i(y)(a, b)\).
The first statement is the change of charts law of the symbols. For the torsion, the inhomogeneous terms \(\mathrm{d}^2\tau (a, b)\) and \(\mathrm{d}^2\tau (b, a)\) cancel.
For a vector bundle with fiber \(F\) the same formalism applies to the symbols in local trivializations, with transition functions \(g(y) \in GL(F)\) in place of \(\mathrm{d}\tau (y)\); for non-normable \(F\) one can only require \((y, \xi ) \mapsto g(y)\xi \) to be smooth, not \(y \mapsto g(y)\). Vector bundles, their connections and the tensorial transformation law of \(\Omega \) are treated in Section 9.1.
10.4 Covariant derivative along curves and geodesics
We use the \(C^{k}\) curves of Definition 3.8 and the velocity of Definition 3.9. Let \(\alpha \colon \mathbb {R}\to M\) be a curve. A lift of \(\alpha \) is a map \(\beta \colon \mathbb {R}\to TM\) with \(\beta (t) \in T_{\alpha (t)} M\). In a chart \(i\) we write \(\tilde\alpha = i \circ \alpha \) and \(\mu _i(t)\) for the fiber coordinate of \(\beta (t)\).
A lift \(\beta \) of \(\alpha \) is \(C^{k}\) if, for every chart \(i\), its fiber coordinate \(\mu _i\) is \(C^{k}\) on \(\alpha ^{-1}(U_i)\).
For two charts \(i, j\) and \(\tau = j \circ i^{-1}\), we have \(\mu _j(t) = \mathrm{d}\tau (\tilde\alpha _i t)(\mu _i t)\) and \(\dot{\tilde\alpha }_j = \mathrm{d}\tau (\tilde\alpha _i)(\dot{\tilde\alpha }_i)\), and
The fiber coordinate of the velocity in any chart \(j\) is \(\dot{\tilde\alpha }_j\), and the velocity of a \(C^{k+1}\) curve is a \(C^{k}\) lift.
The first identity is the change of charts of the tangent bundle, and the others follow from the chain rule and the product rule applied to \(t \mapsto \mathrm{d}\tau (\tilde\alpha _i t)(\mu _i t)\).
Let \(C\) be a connection and \(\beta \) a lift of \(\alpha \). The covariant derivative of \(\beta \) along \(\alpha \) at \(t\) is, in the preferred chart at \(\alpha (t)\),
Let \(\alpha \) be a \(C^{1}\) curve and \(\beta \) a \(C^{1}\) lift. Read in any chart \(j\) around \(\alpha (t)\), the covariant derivative \(D_t\beta \) is \(\dot\mu _j - \Gamma _j(\tilde\alpha _j)(\dot{\tilde\alpha }_j, \mu _j)\).
By Lemma 10.19, \(\dot\mu _j\) contains the term \(\mathrm{d}^{(2)}\tau (\tilde\alpha _i; \dot{\tilde\alpha }_i, \mu _i)\), which cancels with the one in the change of charts law of \(\Gamma \) applied to \((\dot{\tilde\alpha }_i, \mu _i)\). The remaining terms are \(\mathrm{d}\tau (\tilde\alpha _i)\) applied to the expression in the chart \(i\). Compare [ Lan99 ] , §VIII.3.
A curve \(\alpha \) is a geodesic of the connection \(C\) if its velocity is parallel: \(D_t \alpha ' = 0\) for every \(t\).
A \(C^{2}\) curve \(\alpha \) is a geodesic if and only if, in every chart \(j\),
In terms of the spray of the connection (Proposition 11.2), the geodesic equation reads \(\ddot{\tilde\alpha } = f(\tilde\alpha , \dot{\tilde\alpha })\) in every chart, which is the classical definition of the geodesics of a spray.
We do not claim that geodesics exist, nor that they are unique, for given initial position and velocity. In a Banach space the geodesic equation is an ordinary differential equation with a smooth right hand side, and the Picard–Lindelöf theorem gives local existence and uniqueness; this is how the exponential map is constructed in [ Lan99 ] and [ Lar26 ] . In a general locally convex space, and even in Fréchet spaces, ordinary differential equations with smooth right hand side may have no solutions, or many [ Ham82 , Part I, 5.6.1 ] . Therefore existence of geodesics, the exponential map and parallel transport along curves are not part of the theory developed here; they would have to be proved case by case, or under additional hypotheses on \(E\). What we can do in general is to recognize geodesics, as in Theorem 10.33 below. When the model space is a Banach space, existence and uniqueness are proved in the next section.
10.5 Fields of projectors: derive and project
Let \(F\) be a locally convex space and \(X \subseteq F\) an embedded submanifold modeled on a closed subspace \(S \subseteq F\) (Definition 7.3). Recall that in a slice chart \(\varphi \) the tangent space at \(p\) is \(T_p X = \{ w \in F : \mathrm{d}\varphi (p)(w) \in S\} \), and that it does not depend on the slice chart (Definition 7.11). We write \(\psi = \varphi ^{-1}|_S\) for the local parametrization of \(X\) given by \(\varphi \); then \(\mathrm{d}\psi (y)\) maps \(S\) injectively onto \(T_{\psi (y)} X\).
A field of projectors onto the tangent spaces of \(X\) is a map \(E \colon X \times F \to F\), \((p, w) \mapsto E_p w\), such that
each \(E_p\) is linear, \(E_p^2 = E_p\), and the range of \(E_p\) is \(T_p X\);
for every slice chart \(\varphi \), the map \((y, w) \mapsto E_{\psi (y)} w\) is smooth on \(\varphi (U) \cap S \times F\).
In [ Lar26 ] the field of projectors is a smooth map \(X \to \mathcal B(F)\) into the bounded operators of a Banach space. For a locally convex space \(F\) there is no good topology on the space of continuous operators, so we replace it by a map of two variables, linear in the second one, with joint smoothness. Note also that no complement of \(S\) in \(F\) is assumed: the kernels \(\ker E_p\) are complements of the tangent spaces \(T_p X\), but they are not required to come from a global splitting of \(F\), nor to be related to each other.
\(E_p w \in T_p X\) for every \(w\), and \(E_p w = w\) if \(w \in T_p X\). In a slice chart, \(\mathrm{d}\psi (y)(s) \in T_{\psi (y)} X\) for \(s \in S\), hence \(E_{\psi (y)}\mathrm{d}\psi (y)(s) = \mathrm{d}\psi (y)(s)\), and \(\mathrm{d}\psi (y)\) is injective on \(S\).
The first two claims follow from \(E_p^2 = E_p\) and the description of the range. Since \(\mathrm{d}\varphi \) is a left and right inverse of \(\mathrm{d}\varphi ^{-1}\), we have \(\mathrm{d}\varphi (\mathrm{d}\psi (y) s) = s \in S\), which gives the rest.
In a slice chart \(\varphi \), with \(\psi = \varphi ^{-1}|_S\) and \(y \in \varphi (U) \cap S\), put
This vector lies in \(S\), because \(E_{\psi (y)}(\cdot )\) is tangent.
Equivalently, \(\mathrm{d}\psi (y)\bigl(\Gamma _y(a, b)\bigr) = -E\bigl(\mathrm{d}^{(2)}\psi (y; a, b)\bigr)\): the symbols are the tangent part of the second derivative of the parametrization, read in the chart.
The symbols \(\Gamma _y\) are bilinear and symmetric, they are smooth on the chart domain, and for two slice charts they obey the change of charts law of Definition 10.1.
Symmetry and bilinearity come from those of \(\mathrm{d}^{(2)}\psi (y; \cdot , \cdot )\) (Schwarz), and smoothness from the smoothness of \(E\) in slice charts and of \(\psi \). For the change of charts law let \(\tau \) be the transition map between two slice charts, so that \(\psi ' \circ \tau = \psi \) near \(y\). The second-order chain rule gives \(\mathrm{d}^{(2)}\psi = \mathrm{d}^{(2)}\psi '(\mathrm{d}\tau \, a, \mathrm{d}\tau \, b) + \mathrm{d}\psi '(\mathrm{d}^{(2)}\tau (a, b))\). The last term is tangent, so \(E\) fixes it; applying \(-\mathrm{d}\varphi ' \circ E\) yields the law.
A field of projectors \(E\) onto the tangent spaces of \(X\) defines a symmetric connection on \(X\), whose symbols in the chart induced by a slice chart \(\varphi \) are \(\Gamma _y(a, b) = -\mathrm{d}\varphi \bigl(E(\mathrm{d}^{(2)}\psi (y; a, b))\bigr)\).
The next result explains the title of the section: the covariant derivative is obtained by differentiating in \(F\) and projecting onto the tangent space.
Let \(A, B\) be fields on an open subset of the chart domain in \(S\), with \(B\) of class \(C^{1}\), and let \(\hat B(y) = \mathrm{d}\psi (y)(B y)\) be the corresponding tangent field read in \(F\). Then
In other words, \(\nabla _X Y = E(\mathrm{d}Y \cdot X)\) when the fields are read in the ambient space \(F\).
By the second-order chain rule, \(\mathrm{d}\hat B(y)(A y) = \mathrm{d}^{(2)}\psi (y; A y, B y) + \mathrm{d}\psi (y)(\mathrm{d}B(y)(A y))\). Apply \(E\): the second term is tangent and is fixed, and \(E(\mathrm{d}^{(2)}\psi (y; A y, B y)) = -\mathrm{d}\psi (y)(\Gamma _y(A y, B y))\) by Definition 10.28. Compare [ Lar26 , § 6.6 ] .
Let \(c\) be a \(C^{2}\) curve in the chart domain in \(S\) and \(\alpha = \psi \circ c\). Then
where \(\ddot\alpha \) is the acceleration of \(\alpha \) in \(F\). Moreover, a curve in \(X\) which is smooth as a curve in \(F\) is a \(C^{k}\) curve of \(X\) for every \(k\).
We have \(\dot\alpha = \mathrm{d}\psi (c)(\dot c)\) and, by the chain rule and the product rule, \(\ddot\alpha = \mathrm{d}^{(2)}\psi (c; \dot c, \dot c) + \mathrm{d}\psi (c)(\ddot c)\); apply \(E\) as in the proof of Theorem 10.31. For the last claim, the local representative in a slice chart is the corestriction to \(S\) of \(\varphi \circ \alpha \).
Let \(\alpha \) be a \(C^{2}\) curve in \(X\). Then \(\alpha \) is a geodesic of the connection defined by the field of projectors \(E\) if and only if its acceleration in \(F\) is normal:
By Theorem 10.23, \(\alpha \) is a geodesic if and only if \(\ddot c = \Gamma _c(\dot c, \dot c)\) in every chart, where \(c\) is the local representative. By Lemma 10.32 this is equivalent to \(E(\ddot\alpha ) = 0\), since \(\mathrm{d}\psi (c)\) is injective on \(S\) (Lemma 10.27). This is part of [ Lar26 , Teorema 6.6.1 ] .
For instance, for the unit sphere of a (complete) Hilbert space with the orthogonal projectors onto \(p^\perp \), the geodesics are the curves in the sphere with \(\ddot\alpha (t)\) proportional to \(\alpha (t)\). This example is not formalized.
Finally, we compare with the formulation of the book, where the connection is written in terms of the derivative of the field of projectors. We denote with \(E_{*p, v} w\) the derivative of \(q \mapsto E_q w\) at \(p\) in the direction \(v\), computed in the slice chart.
Let \(v = \mathrm{d}\psi (y)(a)\) and \(w = \mathrm{d}\psi (y)(b)\) be tangent vectors at \(p = \psi (y)\), with \(a, b \in S\). Then
In particular \(E_{*p, v}\, w = E_{*p, w}\, v\).
Since \(\mathrm{d}\psi (y')(b)\) is tangent, \(E_{\psi (y')}\mathrm{d}\psi (y')(b) = \mathrm{d}\psi (y')(b)\) for \(y'\) near \(y\). Differentiate this identity in the direction \(a\): by linearity of \(E\) in the second variable, the left side gives \(E_{*p, v} w + E_p\, \mathrm{d}^{(2)}\psi (y; a, b)\), and the right side gives \(\mathrm{d}^{(2)}\psi (y; a, b)\). The symmetry follows from the symmetry of \(\mathrm{d}^{(2)}\psi \).
In [ Lar26 ] the connection is presented through \(\Gamma _p(v, w) = E_{*p, v} w\), which by Proposition 10.34 is the normal part \((1 - E)\mathrm{d}^{(2)}\psi (a, b)\) of the second derivative, read in \(F\). The symbols of Theorem 10.30 are instead the tangent part, pulled back to \(S\) by the chart; the two descriptions are related through the decomposition \(\mathrm{d}^{(2)}\psi = E\, \mathrm{d}^{(2)}\psi + (1 - E)\mathrm{d}^{(2)}\psi \). In both cases the symmetry of the connection is a consequence of the symmetry of second derivatives of \(\psi \). In the Lean code the connection is defined by the tangent part, and the book’s formula enters only through Proposition 10.34.