From curvature to a catalogue of invariants
Guide 4 handed you a machine. Feed it a curvature form Omega — a Lie-algebra-valued 2-form measuring how a connection fails to be flat — and an invariant polynomial P, and Chern-Weil theory spits out a closed differential form P(Omega) whose de Rham cohomology class does not depend on which connection you chose. That last clause is the whole miracle: curvature is a local, connection-dependent quantity, yet certain symmetric functions of it integrate to a topological answer. This final guide stops admiring the machine and starts cataloguing what it produces — the named characteristic classes that organize the entire subject.
Why bother naming them? Because a characteristic class is an obstruction with a face you can recognize. It is a cohomology class attached functorially to every vector bundle (or principal bundle) E over M, vanishing exactly when some structure exists and refusing to vanish when it does not. Is the bundle trivial — a plain product M x R^k with no twisting? Then all its characteristic classes are zero. Find one that is nonzero and you have proven, with a single integral, that no global trivialization can exist. The three families below — Chern, Pontryagin, Euler — are the standard catalogue, each tuned to a different kind of bundle.
Three families: Chern, Pontryagin, Euler
Start with the cleanest. For a complex vector bundle E, feed Chern-Weil the elementary symmetric polynomials in the eigenvalues of (i/2 pi) Omega; the outputs are the Chern classes c_1(E), c_2(E), ..., living in even-degree cohomology H^2(M), H^4(M), and so on. The normalization by i/2 pi is exactly what forces the result into integer cohomology, which is the deep payoff: a quantity built from real-analytic curvature lands on the integer lattice. For a complex line bundle (rank 1) only c_1 survives, and it is the single integer that classifies the line bundle up to isomorphism — the same c_1 you met as the degree of a holomorphic line bundle.
Now the real case. A real vector bundle has no canonical i to insert, so the odd symmetric functions of curvature die by symmetry and only the even ones survive. These give the Pontryagin classes p_1(E), p_2(E), ... in H^4(M), H^8(M), .... The honest shortcut: complexify the real bundle to E ⊗ C and read p_k off its Chern classes, p_k(E) = (-1)^k c_{2k}(E ⊗ C). On a smooth manifold, applying this to the tangent bundle TM yields the Pontryagin numbers — diffeomorphism invariants rigid enough that, paired with the Atiyah-Singer index theorem, they detect exotic smooth structures.
Finally the Euler class e(E), the subtlest of the three. It is defined only for an oriented real bundle of even rank, and unlike the others it lives in the top-degree cohomology H^{rank}(M) and is not a polynomial in lower classes — it is a genuine square root. Chern-Weil builds it from the Pfaffian of the curvature, the natural square root of the determinant available only for skew-symmetric matrices. Its meaning is the most picturesque: e(E) is the obstruction to a nowhere-zero section. For the tangent bundle of a closed oriented surface, integrating e(TM) returns the signed count of zeros of any vector field — the Euler characteristic — which is why a hairy sphere cannot be combed flat but a torus can.
The splitting principle: pretend everything is a line
Computing with the symmetric polynomials directly is a slog. The splitting principle is the trick that makes characteristic classes almost arithmetic. It says: to verify an identity among characteristic classes of a rank-k bundle E, you may pretend E splits as a direct sum of line bundles L_1 ⊕ ... ⊕ L_k. You cannot really make E split — most bundles do not — but you can pull E back to a new base space (the flag bundle of E) where the pullback genuinely splits and, crucially, where the cohomology of the base injects into the cohomology upstairs. So any identity true upstairs, where everything is lines, is already true downstairs.
Here is the payoff in one move. If E splits as a sum of lines with first Chern classes x_1, ..., x_k (the Chern roots), then the total Chern class is simply the product c(E) = (1 + x_1)(1 + x_2) ... (1 + x_k). Multiply it out and the degree-j part is the j-th elementary symmetric polynomial in the roots — exactly c_j(E). Every messy Chern-Weil polynomial collapses into elementary algebra of the roots. The same idea gives the Chern character ch(E) = sum of e^{x_i}, the class that is additive on sums and multiplicative on tensor products and so turns the whole catalogue into a ring homomorphism into cohomology.
splitting principle: pretend E = L_1 (+) ... (+) L_k, x_i := c_1(L_i) (Chern roots)
total Chern class: c(E) = (1 + x_1)(1 + x_2) ... (1 + x_k) = 1 + c_1 + c_2 + ... + c_k
c_1 = x_1 + ... + x_k (elementary symmetric poly degree 1)
c_k = x_1 x_2 ... x_k (elementary symmetric poly degree k)
Chern character: ch(E) = e^{x_1} + ... + e^{x_k} = rank + c_1 + (1/2)(c_1^2 - 2 c_2) + ...
ch(E (+) F) = ch(E) + ch(F) ch(E (x) F) = ch(E) . ch(F)
Whitney sum: c(E (+) F) = c(E) . c(F) (multiply total classes)How the classes behave: functoriality and the Whitney sum
Two structural laws make characteristic classes usable. The first is naturality: if f: M -> N is smooth and you pull a bundle back along f, its characteristic classes pull back too, c(f* E) = f* c(E). This is what lets you compute on a model space and transport the answer. The second is the Whitney sum formula: the total Chern class of a direct sum is the product, c(E ⊕ F) = c(E) . c(F). The splitting principle makes this obvious — concatenate the two lists of Chern roots and multiply the two products. The Pontryagin and Stiefel-Whitney classes obey the same multiplicative law (up to the usual 2-torsion caveats).
Naturality has a converse worth knowing: every characteristic class is pulled back from one universal example. There is a classifying space BU(k) — concretely the infinite Grassmannian of k-planes — carrying a universal bundle, such that every rank-k bundle on M is f* of the universal one for some map f: M -> BU(k), unique up to homotopy. The cohomology of BU(k) is a polynomial ring on the universal Chern classes. So characteristic classes are not many separate inventions; they are the single cohomology of one classifying space, read back through classifying maps. That is the cleanest reason the catalogue is finite and complete.
Gauge theory: the same curvature, a different name
Here the geometry you built becomes physics almost word for word. In gauge theory a gauge field is exactly a connection on a principal bundle P with structure group G (the gauge group — U(1) for electromagnetism, SU(2) and SU(3) for the weak and strong forces). The field strength is exactly the curvature 2-form Omega. A gauge transformation is a change of trivialization, i.e. a vertical automorphism of P; the connection form changes by the familiar inhomogeneous rule while the curvature transforms by conjugation — covariantly. Maxwell's equations are the statement that the U(1) curvature is closed and co-closed; the source-free half, dF = 0, is nothing but the Bianchi identity from guide 3.
What singles out a physical connection? It should make the curvature as small as possible. The Yang-Mills functional YM(A) = integral over M of |Omega|^2 measures the total squared field strength, and its critical points are the Yang-Mills connections, solving the second-order equation that pairs with the Bianchi identity. Their absolute minima within a topological class are the instantons: self-dual connections with Omega = *Omega, which automatically minimize the energy because the energy is bounded below by a characteristic number — a Chern or Pontryagin number — of the bundle. The classes from the first half of this guide are precisely the topological charges the physics cannot change.
This loop closed back on geometry with stunning force. Counting instanton solutions on four-manifolds gave Donaldson his invariants, later streamlined by Seiberg-Witten theory, and these gauge-theoretic counts distinguish smooth structures that every classical tool calls identical — the source of phenomena like exotic R^4. Be honest about the level, though: setting up these moduli spaces rigorously, and proving the index and compactness theorems behind them, is a graduate course and a research literature, not a guide. We are stating where the road leads and showing that the milestones are the very characteristic classes you just learned to compute.
Putting the rung together
Trace the rung from end to end. Guide 1 glued a bundle from local pieces by transition functions; guide 2 added a connection so you could differentiate sections covariantly; guide 3 measured the connection's failure to commute as curvature and holonomy, governed by the Bianchi identity; guide 4 ran curvature through invariant polynomials to manufacture closed forms via Chern-Weil theory; and guide 5 named the resulting classes, computed them by the splitting principle, and watched the whole structure reappear as gauge theory. A local notion (curvature) became a global invariant (a class) became a physical charge — one idea wearing three coats.
Two honest closings. First on scope: we deliberately led with the torus and S^2, the line bundle, and SU(2), because one understood example outweighs a maximally general theorem you cannot picture; the full theory of K-theory, Bott periodicity, and the Atiyah-Singer index theorem deepens all of this and is its own long climb. Second on humility: characteristic classes detect obstructions but do not generally classify bundles by themselves, K-theory is a different and not a 'better' invariant than ordinary cohomology, and the hardest questions here — the smooth four-dimensional Poincaré conjecture among them — remain open. Carry the precise hypotheses, not the slogans, up to the next rung.