Over the past couple of days I have been brushing up on introductory differential geometry. I’ve blogged about this subject a fair bit in the past, but I’ve never really had a good feel for it. I’d therefore like to make this post, and the next, a “big picture” one, rather than focusing on the technical details.

**1. Curvature of a connection**

Let be a manifold, and let be a vector bundle. Suppose given a connection on . This determines, and is equivalent to, the data of parallel transport along each (smooth) curve . In other words, for each such , one gets an isomorphism of vector spaces

with certain nice properties: for example, given a concatenation of two smooth curves, the parallel transport behaves transitively. Moreover, a homotopy of curves induces a homotopy of the parallel transport operators.

In particular, if we fix a point , we get a map

that sends a loop at to the induced automorphism of given by parallel transport along it. (Here we’ll want to take to consist of smooth loops; it is weakly homotopy equivalent to the usual loop space.)

There’s a homotopy theoretical analog of this. Given a fibration , and given any path , we get an identification between the fiber

which is well-defined up to coherent homotopy. We can think of this as the homotopy-theoretical analog of parallel transport. Another way to say it is that there is a coherently associative action of the loop space on a given fiber . In the theory of -categories, a very powerful way to express these ideas is given by the Grothendieck construction, which states that to give a fibration over is equivalent to giving a functor

where is the “fundamental -groupoid of .” In other words, a fibration over is determined by a space (the fiber) and homotopy coherent “parallel transport.”

What does this all have to do with curvature? The curvature of the vector bundle with connection is defined to be the operator (a -valued 2-form) with

for vector fields and sections of . It measures the failure of the differential operator to be symmetric: that is, for mixed partials to be equal.

If mixed partials are equal, that is if , then the connection is said to be **flat** and a consequence is that the infinitesimal holonomy is trivial. In other words, if the connection is flat and if is a contractible neighborhood of , then the composite

is trivial. In other words, if mixed partials are equal, then parallel transport between two points doesn’t depend on the choice of curve connecting them, as long as they’re close enough. Equivalently — the connection is flat precisely when two homotopic paths lead to equal parallel transports.

**2. Flat connections and local systems**

Let’s now say what flatness means homotopically. Given a vector bundle , we can think of that as a functor of -categories

where is, again, treated as an -groupoid and is the topological category of finite-dimensional real vector spaces and isomorphisms between them. This assertion is a restatement of the fact that is the classifying space for real vector bundles.

We can also say this in the following manner. To give a vector bundle is equivalent to giving:

- For each , a choice of finite-dimensional real vector space .
- For each path from to , an isomorphism .
- For each homotopy of paths between , a homotopy
- For each “homotopy of homotopies” of paths a “homotopy of homotopies” between the associated operators, and
- Coherence data for all this.

The choice of a connection — and it is not really a choice, since the space of connections is contractible — is one way to express the above data.

When a vector bundle comes with a flat connection, then any two homotopic paths lead to the same parallel transport data. That means that a flat connection determines a choice of factorization

where is the (ordinary!) groupoid of points on and homotopy classes of paths between them. This is also known as a local system on . In fact, one may state:

Theorem 1There is an equivalence of categories between local systems of -vector spaces on and pairs where is a vector bundle on and is a flat connection on .

There is another more explicit way of describing the correspondence. Given a vector bundle with a flat connection , we can obtain a local system (considered here as a locally constant sheaf on ) on as follows. Given an open subset , let

where a section is horizontal if .

In other words, curvature measures the fact that the vector bundle really depends on higher-order homotopical data, rather than simply the fundamental groupoid. There are more precise versions of this statement, for connections on principal -bundles, for instance the Ambrose-Singer theorem that states that the image of curvature is exactly the Lie algebra of the infinitesimal holonomy group.

In the next post, I’d like to describe some of the more quantitative versions of this statement, and specialize to Riemannian manifolds, where there are many additional things to say.

December 22, 2012 at 9:42 pm

The first paragraph in section 1 is correct, provided the vector bundle $V \to M$ have finite rank. For infinite-dimensional vector bundles, there are various generalizations of the notion of connection that no longer agree. One is the “holonomies” version that you are probably happiest with, in which one should give for every sufficiently regular path an isomorphism of fibers. Another is the “infinitesimal” version, in which one gives a map $V \to V \otimes \Omega^1_M$, where the second tensorand is the cotangent bundle. Without some regularity, you do not have integration and differentiation in infinite dimensions, and so these data are not equivalent.

December 23, 2012 at 12:13 am

Very interesting! I wasn’t aware of the difference for infinite-dimensional bundles.