Connections

The user responsible for this page is Jost Eschenburg. No other user may edit this page at present.

This page has not been refereed. The information given here might be incomplete or provisional.

1 Introduction

How can we differentiate a section $s$${{Authors|Jost Eschenburg}}{{Stub}} ==Introduction== <wikitex>; How can we differentiate a section $s$ in a vector bundle $E$ over a manifold $M$? It takes values in different vector spaces $E_p$ for each $p\in M$, but differentiating involves comparing values at different points $p$. This needs an extra structure on $E$ which connects the different vector spaces $E_p$, $p\in M$, among each other; it is therefore called ''connection''. It can be defined in two different ways. Either it is viewed as a linear map which assigns to each vector field $X$ on $M$ a so called ''covariant derivative'', a differential operator $\nabla_{\!X}$ on the space of sections on $E$. Or else it is given as a ''parallel displacement'' along curves in $M$: Given two points $p,q\in M$ and a curve $c$ connecting these points, a connection allows to move any element of $E_p$ to $E_q$ along $c$. Infinitesimally it is given by a distribution $\mathcal{H}$ on $E$ or on its frame bundle $FE$. Let us repeat: In ordinary analysis we differentiate functions with values in a constant vector space; differentiating functions with values in a variable vector space $E_p$ (a vector bundle) needs an extra structure called connection. The greater generality leads to a new notion: ''curvature''. In the covariant derivative model, curvature is the commutator $[\nabla_i,\nabla_j]$ of covariant derivatives in two coordinate directions. For ordinary derivatives $\nabla_i = \partial_i$ this quantity vanishes, and for covariant derivatives it is an algebraic quantity (a tensor) rather than a differential operator. In the parallel displacement model, curvature is just the non-integrability of the distribution, the tensor $[\mathcal{H},\mathcal{H}]^{\mathcal{H}^\perp}$. A case of particular importance is the tangent bundle, $E = TM$. A connection on $TM$ yields a second tensor quantity beneath curvature, the so called ''torsion''. This is explained best by passing to the affine frame bundle; therefore a connection on $TM$ is sometimes called ''affine connection''. </wikitex> ==Covariant derivatives== <wikitex>; A ''covariant derivative'' on a vector bundle $E$ over a smooth manifold $M$ is a directional derivative $\nabla$ for sections of $E$. It can be viewed as a bilinear map $\nabla : \Gamma TM \times \Gamma E \to \Gamma E$, $(X,s) \mapsto \nabla_Xs$ which is a tensor (linear over $C^\infty(M)$) in the first argument and a derivation in the second argument: {{equation|$\begin{matrix} \nabla_{(fX)}s &=& f\nabla_Xs\,\cr \nabla_{X}(fs) &=& (Xf)s + f\nabla_Xs \,, \end{matrix}$|1}} where $f$ is a smooth function and $X$ a vector field on $M$ and $s$ a section of $E$, and where $Xf = \partial_Xf = df.X$ is the ordinary derivative of the function $f$ in the direction of $X$. By these properties, $\nabla$ is defined locally and even pointwise regarding the first argument: For any $v\in T_pM$ we may define $\nabla_xs := (\nabla_Xs)_p$ where $X$ is any (local) vector field with $X_p = x$. </wikitex> ==Curvature== <wikitex>; A covariant derivative has all properties of the ordinary directional derivative for functions with exception of the commutativity: For functions $f$ we have $\partial_i\partial_jf = \partial_j\partial_if$ with $\partial_if = \frac{\partial(f\circ \phi)}{\partial u_i} = \partial_{\phi_i}f$, where $\phi : \R^n\to M$ is a local diffeomorphism (''local parametrization of $M$'') and $\phi_i := \partial_i\phi$ its $i$-th partial derivative. Instead, for covariant derivatives $\nabla_i = \nabla_{\phi_i}$ of a section $s$ on a vector bundle $E$, the quantity {{equation|$R_{ij}s := [\nabla_i,\nabla_j]s = \nabla_i\nabla_js - \nabla_j\nabla_is$|2}} is in general nonzero but just a tensor (rather than a differential operator): $R_{ij}(fs) = fR_{ij}s$. For arbitrary vector fields $X,Y$ with $X \circ \phi = \sum_i\xi^i\phi_i$ and $Y \circ \phi = \sum_j\eta^j\phi_j$ we put {{equation|$R(X,Y)s = \sum_{ij} \xi^i\eta^j R_{ij}s = [\nabla_X,\nabla_Y]s - \nabla_{[X,Y]}s$|3}} where $[X,Y] = \sum_i \left((X\eta^i) - (Y\xi^i)\right)\phi_i$ denotes the Lie bracket of vector fields. </wikitex> ==Moving frames== <wikitex>; <label>movingframe</label> A ''moving frame'' on $E$ is an $n$-tuple of sections $s_1,\dots,s_n \in \Gamma E|_{M_o}$ on some open subset $M_o \subset M$ such that for any $p\in M_o$ the vectors $s_1(p),\dots,s_n(p)$ form a basis of $E_p$. Denoting by $\nabla s_j$ the linear map $X \mapsto \nabla_Xs_j$, we have {{equation|$\nabla s_j = \sum_i\omega_{ij} s_i$|4}} for certain 1-forms $\omega_{ij}\in \Omega^1(M_o)$ called (''local'') ''connection forms''. Differentiating a second time we have $\nabla_X\nabla_Ys_j = \sum_i \left\{X\omega_{ij}(Y)s_i + \omega_{ij}(Y)\sum_{k}\omega_{ki}(X)s_k\right\}$ and therefore (interchanging the r\^oles of the indices $k$ and $i$ in the second term) {{equation|$R(.,.)s_j = \sum_i \left(d\omega_{ij} + \sum_k \omega_{ik}\wedge \omega_{kj}\right)s_i$|5}} where we have used $d\omega(X,Y) = X\omega(Y) - Y\omega(X) - \omega([X,Y])$ and $(\omega\wedge\psi)(X,Y) = \omega(X)\psi(Y)-\omega(Y)\psi(X)$ for arbitrary 1-forms $\omega,\psi \in \Omega^1(M_o)$. On the other hand we let {{equation|$R(.,.)s_j = \Omega_{ij}s_i$|6}} for some 2-forms $\Omega_{ji} \in \Omega^2(M_o)$ called (local) ''curvature forms'', and thus we obtain the following relation between connection and curvature forms: {{equation|$\Omega_{ij} = d\omega_{ij} + \sum_k \omega_{ik}\wedge \omega_{kj}$|7}} called ''Cartan Structure equations''. Using matrix notation $\Omega = (\Omega_{ij})$, $\omega = (\omega_{ij})$, we may write the latter equation in the short form {{equation|$\Omega = d\omega + \omega\wedge\omega.$|8}} </wikitex> ==Parallelity and connection== <wikitex>; A section $s$ defined along an injective smooth curve $c : I \to M$ is called ''parallel'' if $\nabla_{c'}s = 0$. This is a linear ODE for $s$ which is solvable along $c$ with arbitrary initial values. If $s_1,\dots,s_n$ is a basis of parallel sections and $s = \sum_i\sigma_is_i$ is an arbitrary section along $c$, then $\nabla_{c'}s = \sum_i \sigma_i's_i$. Thus a covariant derivative is determined by its notion of parallelity and vice versa. Parallelity is given by a distribution $\mathcal H$ on $E$, i.e.\ a subbundle $\mathcal{H} \subset TE$, where $\mathcal{H}_v$ for $v\in E_p$ consists of the initial vectors $s'(a)$ of parallel sections $s$ with $s(a) = v$ along curves $c : [a,b] \to M$ starting at $c(a) = p$. Stated differently, $\mathcal{H}_v = ds_p(T_pM)$ where $s$ is a local section defined near $p$ such that $s(p)=v$ and $\nabla_xs = 0$ for all $x\in T_pM$. Vice versa, a section $s$ along any curve $c$ (a smooth map $s : I \to E$ with $s(t) \in E_{c(t)}$ for all $t\in I$) is parallel iff $s'(t) \in \mathcal{H}_{s(t)}$ for all $t\in I$. Since this distribution ``connects'' the distinct fibres of $E$ among each other, it is called ``connection". For any piecewise smooth curve $c : [a,b] \to M$ from $p$ to $q$ and any initial value $v \in E_p$ we have a parallel section $s$ along $c$ with $s(a) = v$. The mapping $\tau_c : s(a) \mapsto s(b) : E_p \to E_q$ is an invertible linear map called ''parallel transport along the curve'' $c$. In general, the parallel transport depends on the curve $c$ itself, not only on the end points $p,q$, but it is independent of the parametrization of $c$. This dependence is measured by the ''holonomy group'' $\mathcal{H}ol_p$ at $p$ which is the set of parallel transports $\tau_c$ along all ''loops'' $c$ at $p$, i.e.piecewise smooth curves $c : [a,b] \to M$ with $c(a) = c(b) = p$. It is known by the Ambrose-Singer theorem \cite{Kobayashi&Nomizu1963|Theorem 8.1} that the connected component of $\mathcal{H}ol_p \subset GL(E_p)$ is a Lie subgroup and its Lie algebra is spanned by the linear maps $\tau_c^{-1}R(x,y)\tau_c$ for all curves $c : [a,b] \to M$ starting from $p$ and all $x,y \in T_qM$ where $q = c(b)$. </wikitex> ==Connection on the frame bundle== <wikitex>; Since it is useful to work with frames instead of single sections, we may replace $E$ with the linear frame bundle $FE$ whose fibre $FE_p$ over $p\in M$ is the set of all frames (bases) of the vector space $E_p$. Then parallelity of a frame $(s_1,\dots,s_n)$, i.e. parallelity of all sections $s_i$ in this frame, is expressed by a distribution on $FE$ which is also called $\mathcal H$. Together with the the ``vertical space" $\mathcal V_f$, the tangent space $T_f(FE_p)$ of the fibre through $f$, it yields a direct decomposition $T_fFE = \mathcal{H}_f \oplus \mathcal{V}_f$ and hence we have projections $\pi_{\mathcal H}$ and $\pi_{\mathcal V}$ of $TFE$ onto the two subbundles. $FE$ is a principal fibre bundle for the group $G = GL_n$, i.e.the fibres are the orbits of a free action of $G$ from the right given by $(f,g) \mapsto fg$ where $f = (f_1,\dots,f_n)$ is a frame and $g = (g_{ij})$ a matrix and where $fg$ is the line with $j$-th component $\sum_i f_ig_{ij}$. Fixing $f\in FE_p$, the action $\phi_f : G \to FE_p$, $g\mapsto fg$ is a diffeomorphism which is equivariant with respect to right translation, and the left invariant vector fields on $G$ are turned by $\phi_*$ into vector fields on $FE$ tangent to the fibres, so called ''fundamental vector fields''. In particular, the vertical space $\mathcal{V}_f$ is canonically isomorphic to the Lie algebra $\frak{g}$ of $G$, via the infinitesimal action $(\phi_f)_* = d(\phi_f)_e$. Using this identification, the projection $\pi_{\mathcal V}$ is a linear form on $TFE$ with values in $\frak{g}$; it will be called (''global'') ''connection form'' $\omega$. The form $\omega$ in Section \ref{movingframe} will be better called $\omega_{\textup{f}}$ since it depends on a moving frame $\textup{f}= (s_1,\dots,s_n) : M_o \to FE$. We have {{equation|$\omega_{\textup{f}} = \textup{f}^*\omega.$|9}} </wikitex> ==Curvature on the frame bundle== <wikitex>; We get the same Cartan structure equations as in Section \ref{movingframe} {{equation|$d\omega + \omega\wedge\omega = \Omega$|10}} <!--domega = 10 --> where the (''global'') ''curvature form'' $\Omega\in \Omega^2(FE)$ is given by {{equation|$\Omega(U,V) = -\omega([\pi_{\mathcal{H}}U,\pi_{\mathcal{H}}V])$|11}} <!--Omega = 11--> for all vector fields $U,V$ on $FE$. To prove {{eqref|10}} we work with two special kinds of vector fields on $FE$, the ''fundamental vertical fields'' $A,B$ etc.\ which are of type $A_f = (d\phi_f)_ea = \frac{d}{dt} (f\exp ta)|_{t=0}$ for some $a\in\frak{g}$, and the ''horizontal lifts'' $X,Y$ etc.\ which are horizontal vector fields projecting to a constant tangent vector on $M$ along every fibre; they are $\pi$-related to a vector field $\bar X$ on $M$. Choosing $U = A+X$ and similarly $V = B+Y$, we have $\omega(U) = a$ and $\omega(V) = b$ and hence $U\omega(V) = V\omega(U) = 0$ since $a,b$ are constant elements of $\frak{g}$. What remains is $-d\omega(U,V) = \omega([U,V]) = \omega([A,B] + [X,Y]) = [a,b] + \omega([X,Y]) = [\omega(U),\omega(V)] - \Omega(U,V) $, using {{eqref|11}}. Here we see the curvature form in a new r\^ole: it measures the non-integrability of the horizontal distribution: $\mathcal{H}$ is integrable $\iff$ $[X,Y]$ is horizontal $\iff$ $\Omega(X,Y) = -\omega([X,Y]) = 0$. </wikitex> ==Connections on general principal bundles== <wikitex>; More generally, let $P$ be a $G$-principal bundle over $M$: A manifold $P$ with a smooth submersion $\pi : P \to M$ and a free action of a Lie group $G$ on $P$ from the right such that the orbits are precisely the fibres, the preimages $\pi^{-1}(p)$, $p\in M$. A ''connection'' on $P$ is a $G$-invariant distribution $\mathcal{H}$ on $P$ (also called ``horizontal distribution'') which is complementary to the tangent spaces of the fibres forming the ``vertical distribution'' $\mathcal{V}$. As before, each vertical space can be identified with the Lie algebra $\frak{g}$ of $G$, and thus the vertical projection $\pi_\mathcal{V}$ can be viewed as a $\frak{g}$-valued 1-form $\omega \in \Omega(P,\frak{g})$, and the equations {{eqref|10}} and {{eqref|11}} hold accordingly. If $\Omega = 0$, then $P$ splits geometrically as $G\times M$ at least locally, and {{eqref|10}} becomes the Maurer-Cartan equation of the Lie group $G$. For any smooth action $\rho : G\times E_o \to E_o$ of $G$ on a smooth manifold $E_o$ we consider the ''associated bundle'' $E = (P\times E_o)/G$ where the action on $P\times E_o$ is given by $g(p,v) = (pg^{-1},\rho(g)v)$. This is a bundle over $M$ with fibre $E_o$, and since the distribution $\mathcal{H}$ is $G$-invariant, it can be transferred to $E$ via $P \subset P\times E_o \to E$, defining a connection on $E$. In the case $P = FE$ for a vector bundle $E$ and $E_o = \R^n$ with its linear $GL_n$-action we have $E \cong (PE\times\R^n)/GL_n$, using the map $PE \times \R^n \to E$, $(f,x) \mapsto fx = \sum_ix_if_i$. This map is obviously invariant under the diagonal $GL_n$-action on $PE\times \R^n$ since $fx = fg^{-1}gx$; it is the usual identification of $\R^n$ with the vector space $E_p$ by means of the basis $f = (f_1,\dots,f_n)$. </wikitex> ==Connections on the tangent bundle== <wikitex>; The tangent bundle $E = TM$ is somewhat special since it carries another 1-form $\theta$ besides $\omega$. In the moving frame language where a local frame $\textup{f}= (f_1,\dots,f_n)$ of $TM$ is given on an open subset $M_o \subset M$, any vector field $Y$ can be written as $Y= \sum \eta_i f_i$. The coefficients $\eta_i$ depend linearly on $Y$, and we may write $\eta_i = \theta_i(Y)$ where the 1-forms $\theta_1,\dots,\theta_n$ om $M_o$ form the dual basis of $(f_1,\dots,f_n)$, i.e.\ $\theta_i(f_j) = \delta_{ij}$. Thus {{equation|$Y = \sum_i \theta_i(Y)f_i$|12}} <!--theta = 12--> If we have a covariant derivative $\nabla$ on $TM$ and another vector field $X$, we obtain $$ \nabla_X Y = \sum_i \left\{X(\theta_i(Y))f_i + \theta_i(Y)\nabla_Xf_i\right\} = \sum_{ij} \left\{X(\theta_i(Y))f_i + \theta_i(Y)\omega_{ji}(X)f_j\right\}$$ from which we derive (interchanging the roles of $i$ and $j$ in the second term) $$ T(X,Y): = \nabla_XY-\nabla_YX - [X,Y] = \sum_{ij}\left(d\theta_i + \omega_{ij}\wedge\theta_j\right)(X,Y)f_i.$$ This tensor $T : \Lambda^2TM \to TM$ is called ''Torsion tensor''; letting {{equation|$T(.,,) = \sum\Theta_if_i$|13}} for some $\Theta_i \in \Omega^2(M_o)$ and putting $\Theta = (\Theta_1,\dots,\Theta_n)^T$ (called ''torsion form'') and $\theta = (\theta_1,\dots,\theta_n)^T$ (sometimes called ''soldering form''), we end up with the second Cartan structure equation {{equation|$\Theta = d\theta + \omega \wedge \theta.$|14}} The following section explains why beneath {{eqref|11}} a second equation occurs for $TM$. </wikitex> ==Affine connections== <wikitex>; An ''affine frame'' on $T_pM$ is a pair $(f,v)$ where $f = (f_1,\dots,f_n)$ is a frame of $T_pM$ and $v\in T_pM$. This is acted on from the right by the ''affine group'' $A_n$ which consists of the inhomogeneous linear transformations $x\mapsto Ax+a$ on $\R^n$ with $A\in GL_n$ and $a\in \R^n$: we let {{equation|$(f,v)(A,a) = (fA,v+f^{-1}a)$|15}} where the frame $f$ is considered as the isomorphism $f :\R^n\to T_pM$ mapping the standard basis vector $e_i\in\R^n$ onto $f_i\in T_pM$. This action turns the set $AM$ of affine frames on $TM$ into a $A_n$-principal bundle. A connection on the $A_n$-principal bundle $AM$ will be called ''generalized affine connection''. Its connection and curvature forms $\hat\omega$, $\hat\Omega$ are $\frak{a}_n$-valued where $\frak{a}_n$ is the Lie algebra of $A_n$. Since $\frak{a}_n = \R^{n\times n} \oplus \R^n$, the forms split accordingly into a matrix and a vector component. Now we consider the embedding $\gamma: FM \to AM$ with $\gamma(f) = (f,0)$. For the pull back forms on $FM$ we have the same splitting: {{equation|$\gamma^*\hat\omega = \omega \oplus \theta$|16}} {{equation|$\gamma^*\hat\Omega = \Omega \oplus \Theta$|17}} where the first term on the right takes values in $\R^{n\times n}$, the second on in $\R^n$. Moreover, the Cartan structure equations for the affine group are {{equation|$\begin{matrix} \Omega &=& d\omega + \omega\wedge \omega,\ \Theta &=& d\theta + \omega\wedge\theta, \end{matrix}$|18}} using agian the splitting $\frak{a}_n = \R^{n\times n}\oplus\R^n$. If $\theta = (\theta_1,\dots,\theta_n)$ has the special property {{eqref|12}}, we call the connection ''affine'', and $\Theta$ equals the torsion form as introduced in the last section. </wikitex> == References == {{#RefList:}} == External links == *The Encylopedia of Mathematics article on [http://www.encyclopediaofmath.org/index.php/Connection connections] *The Wikipedia page about [[Wikipedia:Connection_(mathematics)|connections]] [[Category:Theory]]s$ in a vector bundle $E$$E$ over a manifold $M$$M$? It takes values in different vector spaces $E_p$$E_p$ for each $p\in M$$p\in M$, but differentiating involves comparing values at different points $p$$p$. This needs an extra structure on $E$$E$ which connects the different vector spaces $E_p$$E_p$, $p\in M$$p\in M$, among each other; it is therefore called connection. It can be defined in two different ways. Either it is viewed as a linear map which assigns to each vector field $X$$X$ on $M$$M$ a so called covariant derivative, a differential operator $\nabla_{\!X}$$\nabla_{\!X}$ on the space of sections on $E$$E$. Or else it is given as a parallel displacement along curves in $M$$M$: Given two points $p,q\in M$$p,q\in M$ and a curve $c$$c$ connecting these points, a connection allows to move any element of $E_p$$E_p$ to $E_q$$E_q$ along $c$$c$. Infinitesimally it is given by a distribution $\mathcal{H}$$\mathcal{H}$ on $E$$E$ or on its frame bundle $FE$$FE$. Let us repeat: In ordinary analysis we differentiate functions with values in a constant vector space; differentiating functions with values in a variable vector space $E_p$$E_p$ (a vector bundle) needs an extra structure called connection. The greater generality leads to a new notion: curvature. In the covariant derivative model, curvature is the commutator $[\nabla_i,\nabla_j]$$[\nabla_i,\nabla_j]$ of covariant derivatives in two coordinate directions. For ordinary derivatives $\nabla_i = \partial_i$$\nabla_i = \partial_i$ this quantity vanishes, and for covariant derivatives it is an algebraic quantity (a tensor) rather than a differential operator. In the parallel displacement model, curvature is just the non-integrability of the distribution, the tensor $[\mathcal{H},\mathcal{H}]^{\mathcal{H}^\perp}$$[\mathcal{H},\mathcal{H}]^{\mathcal{H}^\perp}$. A case of particular importance is the tangent bundle, $E = TM$$E = TM$. A connection on $TM$$TM$ yields a second tensor quantity beneath curvature, the so called torsion. This is explained best by passing to the affine frame bundle; therefore a connection on $TM$$TM$ is sometimes called affine connection.

2 Covariant derivatives

A covariant derivative on a vector bundle $E$$E$ over a smooth manifold $M$$M$ is a directional derivative $\nabla$$\nabla$ for sections of $E$$E$. It can be viewed as a bilinear map $\nabla : \Gamma TM \times \Gamma E \to \Gamma E$$\nabla : \Gamma TM \times \Gamma E \to \Gamma E$, $(X,s) \mapsto \nabla_Xs$$(X,s) \mapsto \nabla_Xs$ which is a tensor (linear over $C^\infty(M)$$C^\infty(M)$) in the first argument and a derivation in the second argument:

$\begin{matrix} \nabla_{(fX)}s &=& f\nabla_Xs\,\cr \nabla_{X}(fs) &=& (Xf)s + f\nabla_Xs \,, \end{matrix}$$\begin{matrix} \nabla_{(fX)}s &=& f\nabla_Xs\,\cr \nabla_{X}(fs) &=& (Xf)s + f\nabla_Xs \,, \end{matrix}$(1)

where $f$$f$ is a smooth function and $X$$X$ a vector field on $M$$M$ and $s$$s$ a section of $E$$E$, and where $Xf = \partial_Xf = df.X$$Xf = \partial_Xf = df.X$ is the ordinary derivative of the function $f$$f$ in the direction of $X$$X$. By these properties, $\nabla$$\nabla$ is defined locally and even pointwise regarding the first argument: For any $v\in T_pM$$v\in T_pM$ we may define $\nabla_xs := (\nabla_Xs)_p$$\nabla_xs := (\nabla_Xs)_p$ where $X$$X$ is any (local) vector field with $X_p = x$$X_p = x$.

3 Curvature

A covariant derivative has all properties of the ordinary directional derivative for functions with exception of the commutativity: For functions $f$$f$ we have $\partial_i\partial_jf = \partial_j\partial_if$$\partial_i\partial_jf = \partial_j\partial_if$ with $\partial_if = \frac{\partial(f\circ \phi)}{\partial u_i} = \partial_{\phi_i}f$$\partial_if = \frac{\partial(f\circ \phi)}{\partial u_i} = \partial_{\phi_i}f$, where $\phi : \R^n\to M$$\phi : \R^n\to M$ is a local diffeomorphism (local parametrization of $M$$M$) and $\phi_i := \partial_i\phi$$\phi_i := \partial_i\phi$ its $i$$i$-th partial derivative. Instead, for covariant derivatives $\nabla_i = \nabla_{\phi_i}$$\nabla_i = \nabla_{\phi_i}$ of a section $s$$s$ on a vector bundle $E$$E$, the quantity

$R_{ij}s := [\nabla_i,\nabla_j]s = \nabla_i\nabla_js - \nabla_j\nabla_is$$R_{ij}s := [\nabla_i,\nabla_j]s = \nabla_i\nabla_js - \nabla_j\nabla_is$(2)

is in general nonzero but just a tensor (rather than a differential operator): $R_{ij}(fs) = fR_{ij}s$$R_{ij}(fs) = fR_{ij}s$. For arbitrary vector fields $X,Y$$X,Y$ with $X \circ \phi = \sum_i\xi^i\phi_i$$X \circ \phi = \sum_i\xi^i\phi_i$ and $Y \circ \phi = \sum_j\eta^j\phi_j$$Y \circ \phi = \sum_j\eta^j\phi_j$ we put

$R(X,Y)s = \sum_{ij} \xi^i\eta^j R_{ij}s = [\nabla_X,\nabla_Y]s - \nabla_{[X,Y]}s$$R(X,Y)s = \sum_{ij} \xi^i\eta^j R_{ij}s = [\nabla_X,\nabla_Y]s - \nabla_{[X,Y]}s$(3)

where $[X,Y] = \sum_i \left((X\eta^i) - (Y\xi^i)\right)\phi_i$$[X,Y] = \sum_i \left((X\eta^i) - (Y\xi^i)\right)\phi_i$ denotes the Lie bracket of vector fields.

4 Moving frames

A moving frame on $E$$E$ is an $n$$n$-tuple of sections $s_1,\dots,s_n \in \Gamma E|_{M_o}$$s_1,\dots,s_n \in \Gamma E|_{M_o}$ on some open subset $M_o \subset M$$M_o \subset M$ such that for any $p\in M_o$$p\in M_o$ the vectors $s_1(p),\dots,s_n(p)$$s_1(p),\dots,s_n(p)$ form a basis of $E_p$$E_p$. Denoting by $\nabla s_j$$\nabla s_j$ the linear map $X \mapsto \nabla_Xs_j$$X \mapsto \nabla_Xs_j$, we have

$\nabla s_j = \sum_i\omega_{ij} s_i$$\nabla s_j = \sum_i\omega_{ij} s_i$(4)

for certain 1-forms $\omega_{ij}\in \Omega^1(M_o)$$\omega_{ij}\in \Omega^1(M_o)$ called (local) connection forms. Differentiating a second time we have $\nabla_X\nabla_Ys_j = \sum_i \left\{X\omega_{ij}(Y)s_i + \omega_{ij}(Y)\sum_{k}\omega_{ki}(X)s_k\right\}$$\nabla_X\nabla_Ys_j = \sum_i \left\{X\omega_{ij}(Y)s_i + \omega_{ij}(Y)\sum_{k}\omega_{ki}(X)s_k\right\}$ and therefore (interchanging the r\^oles of the indices $k$$k$ and $i$$i$ in the second term)

$R(.,.)s_j = \sum_i \left(d\omega_{ij} + \sum_k \omega_{ik}\wedge \omega_{kj}\right)s_i$$R(.,.)s_j = \sum_i \left(d\omega_{ij} + \sum_k \omega_{ik}\wedge \omega_{kj}\right)s_i$(5)

where we have used $d\omega(X,Y) = X\omega(Y) - Y\omega(X) - \omega([X,Y])$$d\omega(X,Y) = X\omega(Y) - Y\omega(X) - \omega([X,Y])$ and $(\omega\wedge\psi)(X,Y) = \omega(X)\psi(Y)-\omega(Y)\psi(X)$$(\omega\wedge\psi)(X,Y) = \omega(X)\psi(Y)-\omega(Y)\psi(X)$ for arbitrary 1-forms $\omega,\psi \in \Omega^1(M_o)$$\omega,\psi \in \Omega^1(M_o)$. On the other hand we let

$R(.,.)s_j = \Omega_{ij}s_i$$R(.,.)s_j = \Omega_{ij}s_i$(6)

for some 2-forms $\Omega_{ji} \in \Omega^2(M_o)$$\Omega_{ji} \in \Omega^2(M_o)$ called (local) curvature forms, and thus we obtain the following relation between connection and curvature forms:

$\Omega_{ij} = d\omega_{ij} + \sum_k \omega_{ik}\wedge \omega_{kj}$$\Omega_{ij} = d\omega_{ij} + \sum_k \omega_{ik}\wedge \omega_{kj}$(7)

called Cartan Structure equations. Using matrix notation $\Omega = (\Omega_{ij})$$\Omega = (\Omega_{ij})$, $\omega = (\omega_{ij})$$\omega = (\omega_{ij})$, we may write the latter equation in the short form

$\Omega = d\omega + \omega\wedge\omega.$$\Omega = d\omega + \omega\wedge\omega.$(8)

5 Parallelity and connection

A section $s$$s$ defined along an injective smooth curve $c : I \to M$$c : I \to M$ is called parallel if $\nabla_{c'}s = 0$$\nabla_{c'}s = 0$. This is a linear ODE for $s$$s$ which is solvable along $c$$c$ with arbitrary initial values. If $s_1,\dots,s_n$$s_1,\dots,s_n$ is a basis of parallel sections and $s = \sum_i\sigma_is_i$$s = \sum_i\sigma_is_i$ is an arbitrary section along $c$$c$, then $\nabla_{c'}s = \sum_i \sigma_i's_i$$\nabla_{c'}s = \sum_i \sigma_i's_i$. Thus a covariant derivative is determined by its notion of parallelity and vice versa. Parallelity is given by a distribution $\mathcal H$$\mathcal H$ on $E$$E$, i.e.\ a subbundle $\mathcal{H} \subset TE$$\mathcal{H} \subset TE$, where $\mathcal{H}_v$$\mathcal{H}_v$ for $v\in E_p$$v\in E_p$ consists of the initial vectors $s'(a)$$s'(a)$ of parallel sections $s$$s$ with $s(a) = v$$s(a) = v$ along curves $c : [a,b] \to M$$c : [a,b] \to M$ starting at $c(a) = p$$c(a) = p$. Stated differently, $\mathcal{H}_v = ds_p(T_pM)$$\mathcal{H}_v = ds_p(T_pM)$ where $s$$s$ is a local section defined near $p$$p$ such that $s(p)=v$$s(p)=v$ and $\nabla_xs = 0$$\nabla_xs = 0$ for all $x\in T_pM$$x\in T_pM$. Vice versa, a section $s$$s$ along any curve $c$$c$ (a smooth map $s : I \to E$$s : I \to E$ with $s(t) \in E_{c(t)}$$s(t) \in E_{c(t)}$ for all $t\in I$$t\in I$) is parallel iff $s'(t) \in \mathcal{H}_{s(t)}$$s'(t) \in \mathcal{H}_{s(t)}$ for all $t\in I$$t\in I$. Since this distribution ``connects the distinct fibres of $E$$E$ among each other, it is called ``connection".

For any piecewise smooth curve $c : [a,b] \to M$$c : [a,b] \to M$ from $p$$p$ to $q$$q$ and any initial value $v \in E_p$$v \in E_p$ we have a parallel section $s$$s$ along $c$$c$ with $s(a) = v$$s(a) = v$. The mapping $\tau_c : s(a) \mapsto s(b) : E_p \to E_q$$\tau_c : s(a) \mapsto s(b) : E_p \to E_q$ is an invertible linear map called parallel transport along the curve $c$$c$. In general, the parallel transport depends on the curve $c$$c$ itself, not only on the end points $p,q$$p,q$, but it is independent of the parametrization of $c$$c$. This dependence is measured by the holonomy group $\mathcal{H}ol_p$$\mathcal{H}ol_p$ at $p$$p$ which is the set of parallel transports $\tau_c$$\tau_c$ along all loops $c$$c$ at $p$$p$, i.e.piecewise smooth curves $c : [a,b] \to M$$c : [a,b] \to M$ with $c(a) = c(b) = p$$c(a) = c(b) = p$. It is known by the Ambrose-Singer theorem [Kobayashi&Nomizu1963, Theorem 8.1] that the connected component of $\mathcal{H}ol_p \subset GL(E_p)$$\mathcal{H}ol_p \subset GL(E_p)$ is a Lie subgroup and its Lie algebra is spanned by the linear maps $\tau_c^{-1}R(x,y)\tau_c$$\tau_c^{-1}R(x,y)\tau_c$ for all curves $c : [a,b] \to M$$c : [a,b] \to M$ starting from $p$$p$ and all $x,y \in T_qM$$x,y \in T_qM$ where $q = c(b)$$q = c(b)$.

6 Connection on the frame bundle

Since it is useful to work with frames instead of single sections, we may replace $E$$E$ with the linear frame bundle $FE$$FE$ whose fibre $FE_p$$FE_p$ over $p\in M$$p\in M$ is the set of all frames (bases) of the vector space $E_p$$E_p$. Then parallelity of a frame $(s_1,\dots,s_n)$$(s_1,\dots,s_n)$, i.e. parallelity of all sections $s_i$$s_i$ in this frame, is expressed by a distribution on $FE$$FE$ which is also called $\mathcal H$$\mathcal H$. Together with the the ``vertical space" $\mathcal V_f$$\mathcal V_f$, the tangent space $T_f(FE_p)$$T_f(FE_p)$ of the fibre through $f$$f$, it yields a direct decomposition $T_fFE = \mathcal{H}_f \oplus \mathcal{V}_f$$T_fFE = \mathcal{H}_f \oplus \mathcal{V}_f$ and hence we have projections $\pi_{\mathcal H}$$\pi_{\mathcal H}$ and $\pi_{\mathcal V}$$\pi_{\mathcal V}$ of $TFE$$TFE$ onto the two subbundles. $FE$$FE$ is a principal fibre bundle for the group $G = GL_n$$G = GL_n$, i.e.the fibres are the orbits of a free action of $G$$G$ from the right given by $(f,g) \mapsto fg$$(f,g) \mapsto fg$ where $f = (f_1,\dots,f_n)$$f = (f_1,\dots,f_n)$ is a frame and $g = (g_{ij})$$g = (g_{ij})$ a matrix and where $fg$$fg$ is the line with $j$$j$-th component $\sum_i f_ig_{ij}$$\sum_i f_ig_{ij}$. Fixing $f\in FE_p$$f\in FE_p$, the action $\phi_f : G \to FE_p$$\phi_f : G \to FE_p$, $g\mapsto fg$$g\mapsto fg$ is a diffeomorphism which is equivariant with respect to right translation, and the left invariant vector fields on $G$$G$ are turned by $\phi_*$$\phi_*$ into vector fields on $FE$$FE$ tangent to the fibres, so called fundamental vector fields. In particular, the vertical space $\mathcal{V}_f$$\mathcal{V}_f$ is canonically isomorphic to the Lie algebra $\frak{g}$$\frak{g}$ of $G$$G$, via the infinitesimal action $(\phi_f)_* = d(\phi_f)_e$$(\phi_f)_* = d(\phi_f)_e$. Using this identification, the projection $\pi_{\mathcal V}$$\pi_{\mathcal V}$ is a linear form on $TFE$$TFE$ with values in $\frak{g}$$\frak{g}$; it will be called (global) connection form $\omega$$\omega$. The form $\omega$$\omega$ in Section 4 will be better called $\omega_{\textup{f}}$$\omega_{\textup{f}}$ since it depends on a moving frame $\textup{f}= (s_1,\dots,s_n) : M_o \to FE$$\textup{f}= (s_1,\dots,s_n) : M_o \to FE$. We have

$\omega_{\textup{f}} = \textup{f}^*\omega.$$\omega_{\textup{f}} = \textup{f}^*\omega.$(9)

7 Curvature on the frame bundle

We get the same Cartan structure equations as in Section 4

$d\omega + \omega\wedge\omega = \Omega$$d\omega + \omega\wedge\omega = \Omega$(10)

where the (global) curvature form $\Omega\in \Omega^2(FE)$$\Omega\in \Omega^2(FE)$ is given by

$\Omega(U,V) = -\omega([\pi_{\mathcal{H}}U,\pi_{\mathcal{H}}V])$$\Omega(U,V) = -\omega([\pi_{\mathcal{H}}U,\pi_{\mathcal{H}}V])$(11)

for all vector fields $U,V$$U,V$ on $FE$$FE$. To prove 10 we work with two special kinds of vector fields on $FE$$FE$, the fundamental vertical fields $A,B$$A,B$ etc.\ which are of type $A_f = (d\phi_f)_ea = \frac{d}{dt} (f\exp ta)|_{t=0}$$A_f = (d\phi_f)_ea = \frac{d}{dt} (f\exp ta)|_{t=0}$ for some $a\in\frak{g}$$a\in\frak{g}$, and the horizontal lifts $X,Y$$X,Y$ etc.\ which are horizontal vector fields projecting to a constant tangent vector on $M$$M$ along every fibre; they are $\pi$$\pi$-related to a vector field $\bar X$$\bar X$ on $M$$M$. Choosing $U = A+X$$U = A+X$ and similarly $V = B+Y$$V = B+Y$, we have $\omega(U) = a$$\omega(U) = a$ and $\omega(V) = b$$\omega(V) = b$ and hence $U\omega(V) = V\omega(U) = 0$$U\omega(V) = V\omega(U) = 0$ since $a,b$$a,b$ are constant elements of $\frak{g}$$\frak{g}$. What remains is $-d\omega(U,V) = \omega([U,V]) = \omega([A,B] + [X,Y]) = [a,b] + \omega([X,Y]) = [\omega(U),\omega(V)] - \Omega(U,V)$$-d\omega(U,V) = \omega([U,V]) = \omega([A,B] + [X,Y]) = [a,b] + \omega([X,Y]) = [\omega(U),\omega(V)] - \Omega(U,V)$, using 11. Here we see the curvature form in a new r\^ole: it measures the non-integrability of the horizontal distribution: $\mathcal{H}$$\mathcal{H}$ is integrable $\iff$$\iff$ $[X,Y]$$[X,Y]$ is horizontal $\iff$$\iff$ $\Omega(X,Y) = -\omega([X,Y]) = 0$$\Omega(X,Y) = -\omega([X,Y]) = 0$.

8 Connections on general principal bundles

More generally, let $P$$P$ be a $G$$G$-principal bundle over $M$$M$: A manifold $P$$P$ with a smooth submersion $\pi : P \to M$$\pi : P \to M$ and a free action of a Lie group $G$$G$ on $P$$P$ from the right such that the orbits are precisely the fibres, the preimages $\pi^{-1}(p)$$\pi^{-1}(p)$, $p\in M$$p\in M$. A connection on $P$$P$ is a $G$$G$-invariant distribution $\mathcal{H}$$\mathcal{H}$ on $P$$P$ (also called ``horizontal distribution) which is complementary to the tangent spaces of the fibres forming the ``vertical distribution $\mathcal{V}$$\mathcal{V}$. As before, each vertical space can be identified with the Lie algebra $\frak{g}$$\frak{g}$ of $G$$G$, and thus the vertical projection $\pi_\mathcal{V}$$\pi_\mathcal{V}$ can be viewed as a $\frak{g}$$\frak{g}$-valued 1-form $\omega \in \Omega(P,\frak{g})$$\omega \in \Omega(P,\frak{g})$, and the equations 10 and 11 hold accordingly. If $\Omega = 0$$\Omega = 0$, then $P$$P$ splits geometrically as $G\times M$$G\times M$ at least locally, and 10 becomes the Maurer-Cartan equation of the Lie group $G$$G$. For any smooth action $\rho : G\times E_o \to E_o$$\rho : G\times E_o \to E_o$ of $G$$G$ on a smooth manifold $E_o$$E_o$ we consider the associated bundle $E = (P\times E_o)/G$$E = (P\times E_o)/G$ where the action on $P\times E_o$$P\times E_o$ is given by $g(p,v) = (pg^{-1},\rho(g)v)$$g(p,v) = (pg^{-1},\rho(g)v)$. This is a bundle over $M$$M$ with fibre $E_o$$E_o$, and since the distribution $\mathcal{H}$$\mathcal{H}$ is $G$$G$-invariant, it can be transferred to $E$$E$ via $P \subset P\times E_o \to E$$P \subset P\times E_o \to E$, defining a connection on $E$$E$. In the case $P = FE$$P = FE$ for a vector bundle $E$$E$ and $E_o = \R^n$$E_o = \R^n$ with its linear $GL_n$$GL_n$-action we have $E \cong (PE\times\R^n)/GL_n$$E \cong (PE\times\R^n)/GL_n$, using the map $PE \times \R^n \to E$$PE \times \R^n \to E$, $(f,x) \mapsto fx = \sum_ix_if_i$$(f,x) \mapsto fx = \sum_ix_if_i$. This map is obviously invariant under the diagonal $GL_n$$GL_n$-action on $PE\times \R^n$$PE\times \R^n$ since $fx = fg^{-1}gx$$fx = fg^{-1}gx$; it is the usual identification of $\R^n$$\R^n$ with the vector space $E_p$$E_p$ by means of the basis $f = (f_1,\dots,f_n)$$f = (f_1,\dots,f_n)$.

9 Connections on the tangent bundle

The tangent bundle $E = TM$$E = TM$ is somewhat special since it carries another 1-form $\theta$$\theta$ besides $\omega$$\omega$. In the moving frame language where a local frame $\textup{f}= (f_1,\dots,f_n)$$\textup{f}= (f_1,\dots,f_n)$ of $TM$$TM$ is given on an open subset $M_o \subset M$$M_o \subset M$, any vector field $Y$$Y$ can be written as $Y= \sum \eta_i f_i$$Y= \sum \eta_i f_i$. The coefficients $\eta_i$$\eta_i$ depend linearly on $Y$$Y$, and we may write $\eta_i = \theta_i(Y)$$\eta_i = \theta_i(Y)$ where the 1-forms $\theta_1,\dots,\theta_n$$\theta_1,\dots,\theta_n$ om $M_o$$M_o$ form the dual basis of $(f_1,\dots,f_n)$$(f_1,\dots,f_n)$, i.e.\ $\theta_i(f_j) = \delta_{ij}$$\theta_i(f_j) = \delta_{ij}$. Thus

$Y = \sum_i \theta_i(Y)f_i$$Y = \sum_i \theta_i(Y)f_i$(12)

If we have a covariant derivative $\nabla$$\nabla$ on $TM$$TM$ and another vector field $X$$X$, we obtain

$\displaystyle \nabla_X Y = \sum_i \left\{X(\theta_i(Y))f_i + \theta_i(Y)\nabla_Xf_i\right\} = \sum_{ij} \left\{X(\theta_i(Y))f_i + \theta_i(Y)\omega_{ji}(X)f_j\right\}$ $$\displaystyle \nabla_X Y = \sum_i \left\{X(\theta_i(Y))f_i + \theta_i(Y)\nabla_Xf_i\right\} = \sum_{ij} \left\{X(\theta_i(Y))f_i + \theta_i(Y)\omega_{ji}(X)f_j\right\}$$

from which we derive (interchanging the roles of $i$$i$ and $j$$j$ in the second term)

$\displaystyle T(X,Y): = \nabla_XY-\nabla_YX - [X,Y] = \sum_{ij}\left(d\theta_i + \omega_{ij}\wedge\theta_j\right)(X,Y)f_i.$ $$\displaystyle T(X,Y): = \nabla_XY-\nabla_YX - [X,Y] = \sum_{ij}\left(d\theta_i + \omega_{ij}\wedge\theta_j\right)(X,Y)f_i.$$

This tensor $T : \Lambda^2TM \to TM$$T : \Lambda^2TM \to TM$ is called Torsion tensor; letting

$T(.,,) = \sum\Theta_if_i$$T(.,,) = \sum\Theta_if_i$(13)

for some $\Theta_i \in \Omega^2(M_o)$$\Theta_i \in \Omega^2(M_o)$ and putting $\Theta = (\Theta_1,\dots,\Theta_n)^T$$\Theta = (\Theta_1,\dots,\Theta_n)^T$ (called torsion form) and $\theta = (\theta_1,\dots,\theta_n)^T$$\theta = (\theta_1,\dots,\theta_n)^T$ (sometimes called soldering form), we end up with the second Cartan structure equation

$\Theta = d\theta + \omega \wedge \theta.$$\Theta = d\theta + \omega \wedge \theta.$(14)

The following section explains why beneath 11 a second equation occurs for $TM$$TM$.

10 Affine connections

An affine frame on $T_pM$$T_pM$ is a pair $(f,v)$$(f,v)$ where $f = (f_1,\dots,f_n)$$f = (f_1,\dots,f_n)$ is a frame of $T_pM$$T_pM$ and $v\in T_pM$$v\in T_pM$. This is acted on from the right by the affine group $A_n$$A_n$ which consists of the inhomogeneous linear transformations $x\mapsto Ax+a$$x\mapsto Ax+a$ on $\R^n$$\R^n$ with $A\in GL_n$$A\in GL_n$ and $a\in \R^n$$a\in \R^n$: we let

$(f,v)(A,a) = (fA,v+f^{-1}a)$$(f,v)(A,a) = (fA,v+f^{-1}a)$(15)

where the frame $f$$f$ is considered as the isomorphism $f :\R^n\to T_pM$$f :\R^n\to T_pM$ mapping the standard basis vector $e_i\in\R^n$$e_i\in\R^n$ onto $f_i\in T_pM$$f_i\in T_pM$. This action turns the set $AM$$AM$ of affine frames on $TM$$TM$ into a $A_n$$A_n$-principal bundle. A connection on the $A_n$$A_n$-principal bundle $AM$$AM$ will be called generalized affine connection. Its connection and curvature forms $\hat\omega$$\hat\omega$, $\hat\Omega$$\hat\Omega$ are $\frak{a}_n$$\frak{a}_n$-valued where $\frak{a}_n$$\frak{a}_n$ is the Lie algebra of $A_n$$A_n$. Since $\frak{a}_n = \R^{n\times n} \oplus \R^n$$\frak{a}_n = \R^{n\times n} \oplus \R^n$, the forms split accordingly into a matrix and a vector component. Now we consider the embedding $\gamma: FM \to AM$$\gamma: FM \to AM$ with $\gamma(f) = (f,0)$$\gamma(f) = (f,0)$. For the pull back forms on $FM$$FM$ we have the same splitting:

$\gamma^*\hat\omega = \omega \oplus \theta$$\gamma^*\hat\omega = \omega \oplus \theta$(16)

$\gamma^*\hat\Omega = \Omega \oplus \Theta$$\gamma^*\hat\Omega = \Omega \oplus \Theta$(17)

where the first term on the right takes values in $\R^{n\times n}$$\R^{n\times n}$, the second on in $\R^n$$\R^n$. Moreover, the Cartan structure equations for the affine group are

$\begin{matrix} \Omega &=& d\omega + \omega\wedge \omega,\\ \Theta &=& d\theta + \omega\wedge\theta, \end{matrix}$$\begin{matrix} \Omega &=& d\omega + \omega\wedge \omega,\\ \Theta &=& d\theta + \omega\wedge\theta, \end{matrix}$(18)

using agian the splitting $\frak{a}_n = \R^{n\times n}\oplus\R^n$$\frak{a}_n = \R^{n\times n}\oplus\R^n$. If $\theta = (\theta_1,\dots,\theta_n)$$\theta = (\theta_1,\dots,\theta_n)$ has the special property 12, we call the connection affine, and $\Theta$$\Theta$ equals the torsion form as introduced in the last section.

11 References

• [Kobayashi&Nomizu1963] S. Kobayashi and K. Nomizu, Foundations of differential geometry. Vol I, Interscience Publishers, a division of John Wiley & Sons, New York-London, 1963. MR1393940 (97c:53001a) Zbl 0508.53002

12 External links

• The Encylopedia of Mathematics article on connections
• The Wikipedia page about connections