Bi-refringence versus bi-metricity

In this article we carefully distinguish the notion of bi-refringence (a polarization-dependent doubling in photon propagation speeds) from that of bi-metricity (where the two photon polarizations ``see'' two distinct metrics). We emphasise that these notions are logically distinct, though there are special symmetries in ordinary (3+1)-dimensional nonlinear electrodynamics which imply the stronger condition of bi-metricity. To illustrate this phenomenon we investigate a generalized version of (3+1)-dimensional nonlinear electrodynamics, which permits the inclusion of arbitrary inhomogeneities and background fields. [For example dielectrics (a la Gordon), conductors (a la Casimir), and gravitational fields (a la Landau--Lifshitz).] It is easy to demonstrate that the generalized theory is bi-refringent: In (3+1) dimensions the Fresnel equation, the relationship between frequency and wavenumber, is always quartic. It is somewhat harder to show that in some cases (eg, ordinary nonlinear electrodynamics) the quartic factorizes into two quadratics thus providing a bi-metric theory. Sometimes the quartic is a perfect square, implying a single unique effective metric. We investigate the generality of this factorization process.