Canonical connection

From Manifold Atlas
Jump to: navigation, search

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 Definition

Let M be a homogeneous space, that is a smooth manifold on which a Lie group G acts transitively by diffeomorphisms. Then M \cong G/H where H is the isotropy group of some base point p\in M, and the action map \pi : G \to M, g\mapsto gp becomes the canonical projection \pi : G \to G/H which is a principal bundle with structure group H. Let \textup{Ad} : G \to \Aut(\mathfrak{g}) denote the adjoint representation of G. Its restriction \textup{Ad}(H) clearly keeps \mathfrak{h} invariant. We will assume that the homogeneous space M = G/H is reductive: there is a vector space complement \mathfrak{m} of \mathfrak{h} in \mathfrak{g} which is also invariant under \textup{Ad}(H). Reductiveness is often fulfilled; in particular it holds if H is compact. Since d\pi_e : \mathfrak{g} \to T_pM has kernel \mathfrak{h}, it is an isomorphism on the complement \mathfrak{m}, and it carries the representation Ad(H)|_\mathfrak{m} into the isotropy representation of H on T_pM. Via d\pi_e we identify \mathfrak{m} with T_pM.

Using left translations L_g, g\in G, \mathfrak{m} defines a distribution \mathcal{H} on G (the ``horizontal distribution´´ \mathcal{H}_g = dL_g\mathfrak{m}) which is complementary to the vertical distribution \mathcal{V}_g = dL_g\mathfrak{h} and which is invariant under the right translations of H if \mathfrak{m} is a reductive complement. Thus \mathfrak{m} defines a connection on the H-principal bundle G \to G/H, called the canonical connection of the reductive homogeneous space (G/H,\mathfrak{m}).

The canonical connection determines a covariant derivative \nabla on the tangent bundle TM since this is associated to \pi:G\to M,

\displaystyle  (G\times\mathfrak{m})/H \buildrel \cong \over \longrightarrow TM,\ \ \ [g,X] \mapsto d\pi_g dL_g X,

where h \in H acts on G\times\mathfrak{m} by (g,X) \mapsto (gh^{-1},\textup{Ad}(h)X). The covariant derivative \nabla can be defined by its parallel vector fields. A curve g(t) in G is the parallel displacement for \nabla along the path p(t) = g(t)p in M if and only if it is horizontal, g'(t) \in \mathcal{H}_{g(t)} = dL_{g(t)}\mathfrak{m}. Thus for every X \in \mathfrak{m} = T_pM, the vector field X(t) = dg(t)_p X is parallel along the curve p(t).

Since G is a transformation group on M, its Lie algebra \mathfrak{g} also ``acts´´ on M by the action vector fields: To each X \in \mathfrak{g} we assign a vector field X^* on M by putting for each q = gp\in M

\displaystyle  X^*_q = \left.\frac{d}{dt}\right|_{t=0} x_tq = \left.\frac{d}{dt}\right|_{t=0} \pi(x_tq) = \left.\frac{d}{dt}\right|_{t=0} \pi(x_tgp) = d\pi_g dR_g X,

where t \mapsto x_t is a curve in G with x_0 = e and x'_0 = X, e.g. x_t = \exp(tX). Thus \mathfrak{g} is embedded into the Lie algebra of vector fields on M. However there is a sign change in the Lie bracket: Note that X^* is \pi-related to the right invariant vector field X^r on G since X^*_{gp} = d\pi_g dR_g X = d\pi_g X^r_g; thus [X^*,Y^*] is \pi-related to [X^r,Y^r] = -[X,Y]^r and therefore

(1)[X^*,Y^*] = -[X,Y]^* .
Now we can compute \nabla_{X} Y^* for any Y\in \mathfrak{g} and X \in \mathfrak{m}= T_pM. Putting g_t = \exp(tX) and Y(t) = \textup{Ad}(g_t^{-1})Y we have
\displaystyle  Y^*_{g_tp} = d\pi_{g_t} dR_{g_t}Y = d\pi_{g_t} dL_{g_t}Y(t) = (dg_t)_p d\pi_e Y(t).
Since g_t is the parallel transport along the curve t\mapsto g_tp in M, we obtain \nabla_{g_t'} Y^* = dg_td\pi_e Y'(t). Thus for t=0 we have
(2)\nabla_X Y^* = d\pi_e Y'(0) = -d\pi_e[X,Y].

From this we may compute the torsion tensor T for any X,Y \in \mathfrak{m}= T_pM by extending X,Y to the action vector fields X^*,Y^* on M, using (1), (2):

(3)
Tex syntax error

where Z_\mathfrak{m} denotes the \mathfrak{m}-component of any Z\in\mathfrak{g}=\mathfrak{h}\oplus\mathfrak{m}.

The \mathfrak{h}-valued curvature form R(X,Y) = \Omega(X,Y) for X,Y\in\mathfrak{m} is obtained from the Connections, (7.2):

(4)R(X,Y) = \Omega[X,Y] = -[X,Y]_\mathfrak{h},

where Z_\mathfrak{h} denotes the \mathfrak{h}-component of any Z\in\mathfrak{g} = \mathfrak{h}\oplus\mathfrak{m}. Since the connection on G is invariant under left translations, the action of G on M is affine, that is it preserves the covariant derivative \nabla, and the same is true for the torsion and the curvature tensors, T and R. In particular it is preserved under parallel displacements which are horizontal curves in G (in particular, the holonomy group of \nabla is contained in H). Thus these tensors are \nabla-parallel.

Vice versa, given any manifold M with a connection \nabla on TM with parallel torsion and curvature tensors, then M is a reductive locally homogeneous space, i.e. each point p\in M has an open neighborhood which can be identified to some set in a reductive homogeneous space G/H where \nabla becomes the canonical connection of G/H. In fact, let \mathfrak{m}= T_pM and H = \Aut(\mathfrak{m},T,R) the group of automorphisms of \mathfrak{m} preserving both the ``product´´ T : \mathfrak{m}\times\mathfrak{m}\to \mathfrak{m} and the ``triple product´´ R : \mathfrak{m}\times\mathfrak{m}\times\mathfrak{m}\to \mathfrak{m}. Then G is a Lie group with Lie algebra \mathfrak{g} = \mathfrak{h} \oplus \mathfrak{m} where \mathfrak{h}\subset\End\mathfrak{m} is the Lie algebra of H and where the remaining Lie brackets are given as follows:

(5)[A,X] = AX,\ \ \ [X,Y] = -(R(X,Y)+T(X,Y)) \in \mathfrak{h} \oplus \mathfrak{m}= \mathfrak{g}

for all A\in\mathfrak{h} and X,Y\in\mathfrak{m}.

When [\mathfrak{m},\mathfrak{m}] \subset \mathfrak{h}, these spaces are called locally symmetric. This happens if and only if the torsion tensor T = [\mathfrak{m},\mathfrak{m}]_\mathfrak{m} vanishes. Moreover, if \tilde M carries a G-invariant (semi-)Riemannian metric, the canonical connection \nabla preserves the metric and is torsion free, hence it is the Levi-Civita connection of \tilde M.

2 Example

Let M be the set of all real positive definite symmetric n\times n-matrices which is an open subset of the space of all symmetric matrices, S_n = \{X\in\R^{n\times n}: X^t = X\}. The group GL_n acts transitively on M by (g,p) \mapsto gpg^t. The isotropy group of p = I (unit matrix) is the orthogonal group H = O_n = \{g\in\R^{n\times n}: g^tg = I\} whose Lie algebra is the space of antisymmetric matrices, A_n = \{A\in \R^{n\times n}: A^t = -A\}. Thus M \cong GL_n/O_n, and the reductive splitting is \mathfrak{g} = \R^{n\times n} = A_n \oplus S_n with A_n = \mathfrak{h} and S_n = \mathfrak{m}. The vector space S_n is not a Lie algebra since for the commutator of any X,Y \in S_n we have [X,Y]\in A_n. But it is a Lie triple: Note that [A,X] \in S_n for all A\in A_n, X \in S_n (since (AX-XA)^t = X^tA^t-A^tX^t = -XA+AX), in particular [[X,Y],Z] \in S_n for all X,Y,Z\in S_n. The canonical connection is torsion free (i.e. M is symmetric) since [\mathfrak{m},\mathfrak{m}]\subset\mathfrak{h}, and its curvature tensor is R(X,Y)Z = - [[X,Y],Z]. The trace metric on S_n is invariant under H = O_n and extends to a GL_n-invariant Riemannian metric on M given as follows: For any p\in M and X,Y \in T_pM = S_n we put
\displaystyle  \langle X,Y\rangle _p := \textup{trace} p^{-1}Xp^{-1}Y
Thus the canonical connection is the Levi-Civita connection on M with respect to this metric. The sectional curvature is nonpositive; it is given by \langle R(X,Y)Y,X\rangle  = -\textup{trace}([[X,Y],Y]X) = \textup{trace}[X,Y]^2 \leq 0 (recall that [X,Y] \in A_n has imaginary eigenvalues, hence its square has nonpositive trace).

For further information, see [Kowalski1980].

3 References

Personal tools
Namespaces
Variants
Actions
Navigation
Interaction
Toolbox