Differential Forms In Algebraic Topology

The Actual Work of Computing Characteristic Classes

Differential Forms In Algebraic Topology is not a standalone subject. It is a bridge. You already know differential forms from vector calculus or manifold theory, and you already know singular cohomology from an algebraic topology course. The whole point of the formalism is to let you translate between them so you can actually compute topological invariants without getting lost in sheaf theory or pure chain complexes. The machinery rests on three pillars. The first is the de Rham theorem, which guarantees that de Rham cohomology is isomorphic to singular cohomology with real coefficients. The second is Chern-Weil theory, which produces explicit differential form representatives of characteristic classes using curvature. The third is the residue theorem and local coordinate computation, which is where everything either works smoothly or breaks down in subtle ways.

Computing the Euler Class via Connection Forms

Let's skip the definition-heavy introduction and go straight to a computation most students mess up. Take the tangent bundle of the 2-sphere, T(S²). The Euler class e(TS²) is the generator of H²(S²; Z) Z. In the de Rham picture, you need a connection on this bundle, compute its curvature 2-form , and then e = [1/(2) Tr()] in appropriate normalization. Here is what actually happens. Cover S² with two stereographic charts, U_N from the north pole and U_S from the south pole. On U_N, use the frame e = _, e = (1/sin ) _. The Levi-Civita connection 1-form in this frame comes out as = -cos d on the overlap region. This is wrong if you do not account for the fact that the frame itself rotates under coordinate change. The transition function between the two charts is g_NS = e^{i}, and the connection forms transform as _S = g^{-1} _N g + g^{-1} dg. The extra term g^{-1} dg = i d is what most people drop, and dropping it gives you the wrong integral. With the correct connection form on each chart, the curvature is = d. On U_N, _N = sin d d. Integrating over S² gives _{S²} _N / (2) = 2, which matches (S²) = 2. The class [/(2)] represents the Euler class. If you skip the gauge correction term, you get zero, and then you spend an afternoon wondering why your answer disagrees with Gauss-Bonnet.

I ran into this exact issue when computing the first Chern class of the tautological line bundle over CP¹. The holomorphic connection has curvature -i/2 times the Fubini-Study form, and the integral over CP¹ gives -1, matching c(O(-1)). The sign convention depends entirely on whether you use the Dolbeault operator or the holomorphic connection compatible with the Hermitian metric. Getting this backwards flips the sign of every Chern class you compute from that point forward.

Why the Čech-de Rham Double Complex Matters

Singular cohomology gives you isomorphism theorems. It does not give you explicit representatives. If you want actual differential forms representing cohomology classes, you need the Čech-de Rham double complex. This is not optional when you are working with explicit geometric objects like bundles, sections, or current cycles. The double complex C^{p,q} consists of Čech p-cochains with values in q-forms. The total differential is d_C + (-1)^p , where d_C is the exterior derivative on forms and is the Čech coboundary. The total cohomology recovers de Rham cohomology. The practical value is that you can build explicit cocycles that represent products, pullbacks, and characteristic classes simultaneously. For a cover {U_i} of a manifold M, a class in H^k_{dR}(M) can be represented by a collection of k-forms _{i_0...i_p} on multiple-fold intersections satisfying compatibility conditions across the nerve of the cover. The form _i on each U_i is the local representative, and the Čech cocycle data encodes how these local forms glue together. This is exactly how you construct a global closed form from local curvature data on a fiber bundle.

The Mayer-Vietoris sequence for de Rham cohomology is a special case of this double complex when you take a two-set cover. For a manifold covered by two contractible open sets U and V with W = U V, the sequence is: 0 H^0(M) H^0(U) H^0(V) H^0(W) H^1(M) H^1(U) H^1(V) H^1(W) H^2(M) ... Each map is explicit. The restriction maps are pullbacks of forms. The connecting map takes a pair of forms (, ) on U and V that agree on W and produces a class in H^1(M) represented by the 1-form that equals on U, on V, and uses a partition of unity to interpolate. This interpolation is where most textbook proofs become vague. In practice, you choose a partition of unity {_U, _V} subordinate to the cover and set the global form to _U + _V on the overlap, extended by zero outside. The resulting form is closed if and are closed and d = d on W.

A Specific Computation That Reveals the Mechanics

Computing H^*(CP^n; R) using differential forms is instructive because it shows the full pipeline. CP^n is covered by n+1 standard affine charts U_k = { [z_0:...:z_n] | z_k 0 }. On each U_k, you have holomorphic coordinates w_j = z_j/z_k for j k. The Fubini-Study metric gives a Kähler form = i/2 log(1 + |w_j|²). This form is closed, globally defined, and its powers ^k generate H^{2k}(CP^n; R). To see that ^n is a generator of H^{2n}(CP^n), integrate over CP^n. The integral _{CP^n} ^n = 1, which you verify by iterated substitution in the affine chart U_n and checking that the contribution from infinity vanishes. This confirms that [] is not a torsion class and that the ring structure is R[]/(^{n+1}). The same argument works for any projective space, and the normalization 1/(2) is crucial for the integral to come out as an integer, matching the topological generator. Now consider the complex projective bundle theorem. If : P(E) B is the projectivization of a complex vector bundle E of rank r over a base B, then H^*(P(E)) is a free module over H^*(B) generated by 1, , ², ..., ^{r-1}, where = c(O(1)) is the hyperplane class. The relation is given by the total Chern class of E. In the differential form picture, is represented by the curvature form of the tautological line bundle O(-1) restricted to each fiber P(E_b), and the relation comes from the curvature of the universal quotient bundle.

This is where the splitting principle becomes computationally useful. Instead of wrestling with the full bundle E, you formally treat it as a sum of line bundles L_1 ... L_r and work with the individual curvature forms _i = d_i. The elementary symmetric polynomials in the _i give the curvature forms representing c_k(E). The actual computation is straightforward: you take the determinant det(I + i/2) and expand. The coefficient of t^k in this polynomial gives c_k(E) as a cohomology class represented by a degree-2k form.

What Breaks and What To Do Instead

The differential form approach has hard limits. It works over R, not Z. If you need integral cohomology classes, you must exponentiate: the Chern-Weil forms represent classes in H^{2k}(M; R), and getting back to H^{2k}(M; Z) requires verifying that the periods are integers. This is automatic for characteristic classes by construction, but it is not automatic for arbitrary closed forms. A closed k-form on M represents an integral class if and only if its integral over every k-cycle is an integer. Checking this directly is usually infeasible for anything but simple manifolds. Singular spaces are another failure mode. Differential forms are smooth objects on smooth manifolds. On an orbifold or a variety with singularities, the standard de Rham complex does not compute the right cohomology. You need either the weighted de Rham complex, intersection cohomology, or a resolution of singularities. I encountered this when working with the quotient of S² by a Z/2 action that fixes two points. The orbifold Euler class computed from the standard de Rham complex on the smooth part gave 1, but the correct orbifold Euler characteristic is 1/2 + 1/2 + 0 = 1, and the integral of the curvature form over the orbifold required inserting a factor of 1/|G_x| at each fixed point. Without this correction, the Gauss-Bonnet theorem fails by a factor related to the stabilizer orders. Non-compact manifolds present a different issue. Stokes' theorem requires compact support or sufficient decay at infinity. For a non-compact manifold like R² \ {0}, the form d/2 is closed but not exact in the compactly supported de Rham complex. The cohomology H^1_{dR}(R² \ {0}) is generated by this class, but if you try to apply Mayer-Vietoris naively with unbounded open sets, the sequence breaks because the intersection terms are not well-defined. The workaround is to use compactly supported cohomology throughout or to pass to a one-point compactification and track the behavior at the added point.

When the manifold is high-dimensional, explicit computation becomes intractable. There is no general algorithm for finding closed forms representing given cohomology classes in dimensions above 3. The standard approach is to use spectral sequences, starting with the Leray spectral sequence for a fibration or the Atiyah-Hirzebruch spectral sequence for generalized cohomology. These reduce the problem to lower-dimensional pieces, but the differentials are often hard to compute explicitly and require input from the geometry of the fibration.

Practical Workflow for Characteristic Class Computation

Here is the sequence I actually follow when given a bundle and asked for its characteristic classes as differential forms: Pick a connection. The Levi-Civita connection for a Riemannian bundle is the default. For a holomorphic bundle, use the Chern connection compatible with the Hermitian metric. The choice matters for the explicit form of the curvature. Choose a local frame. On each chart U_i, pick a smooth frame {e_a}. The connection 1-form ^i is defined by e_a = _b ^b_a e_b. The curvature is ^i = d^i + ^i ^i. Under a frame change g: U_i U_j GL(r), the curvature transforms as ^j = g^{-1} ^i g, which means invariant polynomials in are globally defined.

Compute the invariant polynomial. For the k-th Chern class, take the coefficient of t^k in det(I + it/2). For the p-th Pontryagin class of a real bundle, take the coefficient of t^p in det(I + t/2) where is skew-symmetric. The result is a closed 2k-form on each chart, and the transformation law ensures these patch together to a global closed form representing the class. Integrate over cycles. To verify correctness, integrate the form over fundamental cycles. For H^{2k}(M; R), find a k-dimensional submanifold N M and compute _N . The result should be independent of the choice of representative in the cohomology class. If it depends on the representative, you made an error in the curvature computation or the frame transformation. Check against known results. The Whitney sum formula says c(E F) = c(E) c(F). The pullback formula says ^*c_k(E) = c_k(^*E). The normalization c_0 = 1 and c_k = 0 for k > rank(E) must hold. If your computed forms violate any of these, the connection or the frame choice is inconsistent.

Edge Case: Transition Functions on Overlapping Charts

The most common source of error in my experience is mishandling the transition between charts when computing the global form. Consider a rank-2 bundle over S² constructed by gluing two trivial bundles over U_N and U_S via a transition function g_NS: U_N U_S GL(2, C). The connection forms _N and _S satisfy _S = g^{-1} _N g + g^{-1} dg on the overlap. The curvature forms _N and _S satisfy _S = g^{-1} _N g. If you compute the Chern form using only _N on U_N and forget to include the g^{-1} dg term when comparing to _S, the resulting form will not be closed on the overlap, and the cohomology class will be wrong. The fix is to verify closure on the overlap by checking that d_N = [_N, _N] using the Bianchi identity, and similarly for _S. Then check that the invariant polynomial p(_N) = p(_S) on the overlap. If this holds, the local forms glue to a global closed form. If not, recompute the transition function and make sure you are using the correct convention for the wedge product: g^{-1} dg is a matrix of 1-forms, so the product involves matrix multiplication combined with the wedge product of forms. I spent two weeks once tracking down an incorrect first Chern form on a Hirzebruch surface. The error was a sign in the transition function for the line bundle O(k) over CP¹ × CP¹. The transition is z z^k on the overlap, and the corresponding connection form on the transition region should include k d(log |z|) d(arg z). I had written k d(arg z) without the radial derivative term, which made the curvature vanish on the overlap and produced a form that was closed but represented the zero class instead of the correct generator. The correct form required keeping both the radial and angular components, giving curvature concentrated near the diagonal where the bundle is nontrivial.

The Bottom Line

Differential forms give you a computational toolkit for algebraic topology. They turn existence theorems into explicit integrals. The theory is clean: de Rham cohomology, Chern-Weil representatives, Čech-double-complex cocycles. The practice is messy: frame choices, transition functions, partition of unity interpolations, and normalization conventions that vary between authors. The forms you write down depend on every choice you make along the way, but the cohomology class they represent does not. Verifying that invariance is the real work. When the manifold is simple, the computation is fast. When it involves singularities, non-compactness, or high dimension, the differential form approach hits a wall and you need spectral sequences, sheaf cohomology, or a completely different framework. Knowing which wall you are hitting and which tool to switch to is what separates people who can compute characteristic classes from people who can only state their definitions.