Curvature as the failure of transport to close up
From guide 2 you can already covariantly differentiate sections and parallel-transport them along a path. Here is the question that births curvature: if you carry a vector around a small closed loop and come back to where you started, do you get the same vector? On the flat plane, yes — transport is path-independent and every loop returns you unchanged. On a curved surface, no: walk a tangent vector around a spherical triangle and it comes back rotated. Curvature is exactly the infinitesimal measure of that rotation per unit area enclosed.
Make it algebraic. The covariant derivative gives operators nabla_X for each direction X. On the plane these commute; on a curved bundle they do not, and the defect is the curvature operator. For a connection on a vector bundle one defines R(X, Y) = nabla_X nabla_Y - nabla_Y nabla_X - nabla_[X, Y], acting on sections. The first two terms are "go along X then Y" versus "Y then X" — the two ways around an infinitesimal parallelogram — and the bracket term [X, Y] subtracts off the part of the discrepancy that is just the directions failing to commute, leaving the genuinely geometric twist.
The curvature 2-form: dA + A wedge A
In guide 2 the connection was encoded locally by a connection 1-form A — a matrix-valued (more precisely, Lie-algebra-valued) 1-form recording the Christoffel-type coefficients in a chosen frame. The curvature 2-form F is then given by the clean structural equation F = dA + A ^ A. The first term dA is the naive "how does the connection vary" piece; the wedge term A ^ A is the nonlinear correction that appears precisely because matrices do not commute. Antisymmetry of the wedge does not kill A ^ A here — for matrix-valued forms the wedge interleaves a matrix product, and that product is non-commutative.
Why is F a 2-form and not a 1-form? Because curvature is born from a loop, and the smallest loop is a parallelogram spanned by two directions — you must feed it two tangent vectors, X and Y, and it returns the infinitesimal transport defect F(X, Y), an endomorphism of the fiber. For a principal bundle with structure group G, both A and F take values in the Lie algebra of G, and the same formula F = dA + (1/2)[A, A] holds with the Lie bracket replacing the matrix commutator. One equation, two readings: matrix product for a vector bundle, Lie bracket for a principal bundle.
vector bundle: F = dA + A ^ A (A, F are matrix-valued forms; ^ includes matrix product) principal bundle: F = dA + (1/2)[A, A] (A, F valued in Lie(G); bracket = Lie bracket) link to the operator: F(X, Y) s = R(X, Y) s = (nabla_X nabla_Y - nabla_Y nabla_X - nabla_[X,Y]) s flat connection <=> F = 0 <=> transport is path-independent locally
Holonomy: the global memory of every loop
Curvature is local — it is what an infinitesimal loop sees. Holonomy is its global counterpart: fix a point p, and for each loop based at p, parallel-transport once around. The result is a linear map of the fiber over p back to itself, an element of the structure group G. Letting the loop range over all loops at p, these maps form a subgroup of G called the holonomy group. It is the complete record of how the connection twists, integrated over loops of every size — exactly what curvature only knew infinitesimally.
The bridge between the two is the Ambrose-Singer theorem: the Lie algebra of the holonomy group is spanned by all the curvature values F(X, Y), transported around to the base point. Read in plain words — holonomy is curvature accumulated, and curvature is holonomy differentiated. A flat connection (F = 0) on a simply connected base has trivial holonomy: every loop bounds a disk you can sweep transport across with no defect. But on a base with nontrivial fundamental group, a flat connection can still have nontrivial holonomy purely from loops that cannot be contracted — that part is topological, not curvature-driven.
- Pick a base point p and a fiber over it; you will track where a frame goes after a round trip.
- For a chosen loop gamma based at p, solve the parallel-transport equation along gamma — this is an ODE, so transport exists and is unique once the connection is given.
- Record the resulting fiber automorphism; varying gamma over all loops gives a set of automorphisms closed under composition and inverse — the holonomy group.
- By Ambrose-Singer, the infinitesimal generators of that group are exactly the curvature values F(X, Y) carried back to p, sealing the curvature-holonomy correspondence.
The Bianchi identity: curvature is never free
Curvature cannot be any 2-form you like; it obeys a universal differential constraint, the second Bianchi identity. In the cleanest form it reads dF + [A, F] = 0, usually abbreviated as the covariant exterior derivative D_A F = 0. In words: the curvature is covariantly closed. The proof is a one-line miracle of the structure equation — differentiate F = dA + A ^ A, use d^2 = 0 from the exterior derivative, and watch the cross terms reorganize into exactly -[A, F]. No geometry is assumed beyond having a connection; this identity is automatic.
Why care about a constraint that holds automatically? Because it is the engine of the next two guides. The Bianchi identity is exactly what makes the Chern-Weil construction work: when you feed F into an invariant polynomial to build a characteristic form, D_A F = 0 is the reason the resulting form is closed, hence defines a de Rham cohomology class. Strip the Bianchi identity away and characteristic classes simply would not be well-defined. It is also the geometric origin of conservation laws in gauge theory — the homogeneous Maxwell equations are literally the Bianchi identity for the electromagnetic connection.
Tiny worked picture and an honest map forward
Anchor everything on the round 2-sphere with its Levi-Civita connection. Parallel-transport a tangent vector around a geodesic triangle with angles summing to (pi + excess); the vector returns rotated by exactly that angle excess. That is holonomy you can see by hand, and the Gauss-Bonnet machinery says the excess equals the integral of Gaussian curvature over the enclosed region — holonomy = integrated curvature, made utterly concrete. The infinitesimal version of that same statement, shrunk to a point, is the curvature 2-form F; the global version, summed over a closed loop, is the holonomy. Same phenomenon, two zoom levels.
Be honest about what we did and did not prove. We defined curvature, identified it with the structure-equation 2-form, sketched (not fully proved) Ambrose-Singer, and gave the one-line proof of the second Bianchi identity. We did not classify holonomy groups — that is Berger's theorem, a deep classification with its own list (the exceptional cases like G2 and Spin(7) holonomy are genuine research territory). And every sign here depends on conventions: some books write F = dA - A ^ A, or put the 1/2 elsewhere, or orient loops oppositely. Pick a source, hold its conventions fixed, and translate carefully when you switch.
With curvature, holonomy, and Bianchi in hand, the road forward is short and steep. Guide 4 feeds F into invariant polynomials — the Chern-Weil homomorphism — to manufacture closed forms whose cohomology classes do not depend on the connection at all, turning a geometric quantity into a topological invariant. Guide 5 names the resulting classes (Chern, Pontryagin, Euler) and connects them to gauge theory. Everything downstream rests on the two facts you now own: curvature is dA + A ^ A, and it is covariantly closed.