OK, blathering it is.

The key thing to understand in Yang-Mills theory (whether or not you're trying to be mathematically rigorous) is that *gauge symmetry is not a physical symmetry*. (Its job is to keep track of the overcounting in a redundant description.)

Classically, this means that the observables -- the numerical quantities which we can measure -- are *not* functions on the space $\mathcal{F}$ of connections but rather functions on the quotient space $\mathcal{F}/\mathcal{G}$ of connections modulo the action of the group of gauge transformations. So, for example, for any $G$-invariant polynomial $P$ on the Lie algebra $\mathfrak{g}$ of $G$ and any point $x \in \mathbb{R}^4$, there's an observable $\mathcal{O}_{P,x}$, which sends the equivalence class $[A]$ of a connection $A$ to $P(F_A(x))$, where $F_A$ is the curvature of $A$. But there isn't any physical way of measuring the value of a connection $A$ at the point $x$; this quantity is useful for intermediate steps in the computation, but it is not an observable.

In the Wightman framework, one describes a QFT in terms of a Hilbert space which carries a representation of the Poincare group and comes equipped with a collection of local observables. So for Yang-Mills theory, naively, what we'd like is an operator $\hat{\mathcal{O}}_{P,x}$ for every classical observable like $\mathcal{O}_{P,x}$. This doesn't quite work, however. You can't have an observable $\hat{\mathcal{O}}_{P,x}$ which measures the value of something precisely at $x$, for basically the same reason that you can't have a particle whose wave-function is supported precisely at $x$. Things have to get fuzzed out a little bit in quantum theories. So instead of a function $x \mapsto O_{P,x}$, what we get in the quantum theory is an operator-valued distribution, which sends any test function $f$ to the observable $\hat{\mathcal{O}}_{P,f}$ which measures the value of the smeared observable $\int P(F_A(x))f(x)dx$.

You will occasionally find claims (on wikipedia, for example) that the Wightman axioms fail for gauge theories because gauge theories involve states with negative norms. This isn't really correct. Negative norms show up at intermediate stages in some quantization procedures (but not all of them), but they should never be present in the final result, which is the only thing the Wightman axioms are concerned with.

Note that what I've written above isn't quite the usual statement of the Wightman axioms. People usually give them only for scalar field theory, where they can take a shortcut because the basic fields $\phi(x)$ that generate the observables are actually observable themselves. However, all one really needs for Wightman's setup is local observables whose classical analogues provide a complete set of coordinates on the space of classical fields.

Finally, the really interesting subtlety: If you're being careful, the group of gauge transforations is the subgroup of $Map(\mathbb{R}^4,G)$ consisting of $G$-valued functions which take the identity value at $\infty$. This would suggest that the values of invariant differential polynomials in the curvature tensor are close to but not quite a complete set of coordinates on $\mathcal{F}/\mathcal{G}$, since they're also invariant under the gauge transformations which are constant over $\mathbb{R}^4$. So why should these be enough? The answer is confinement. We expect that, in pure Yang-Mills theory, the only physically realizable states are color singlets, transforming trivially under the global $G$.

This post imported from StackExchange Physics at 2014-08-29 16:46 (UCT), posted by SE-user user1504