There is a "proof" of S-duality in free abelian gauge theory. See, for example, Witten's On S-Duality in Abelian Gauge Theory or his IAS lecture. Morally, you should think of this as a Fourier transform. The non-Abelian version, of course, is much harder and requires N=4 SUSY. But this isn't a surprise as the non-Abelian Fourier transform is a tricky thing.
Now, at least from the math side, you should be thinking about Geometric Langlands which is pretty closely related and has the same distinction between the Abelian case (which I understand to reduce to some class field theory stuff) and the more difficult non-Abelian version. Long before Kapustin-Witten and all that has come after, this connection was first shown in Harvey, Moore and Strominger and Bershadsky, Johansen, Sadov and Vafa.
There is, of course, a long, long list of "coincidences", too.
This post has been migrated from (A51.SE)