mixtures as exponential families

Something I had not realised earlier and that came to me when answering a question on X validated about the scale parameter of a Gamma distribution. Following an earlier characterisation by Dennis Lindley, Ferguson has written a famous paper characterising location, scale and location-scale families within exponential families. For instance, a one-parameter location exponential family is necessarily the logarithm of a power of a Gamma distribution. What I found surprising is the equivalent for one-parameter scale exponential families: they are necessarily mixtures of positive and negative powers of Gamma distributions. This is surprising because a mixture does not seem to fit within the exponential family representation… I first thought Ferguson was using a different type of mixtures. Or of exponential family. But, after checking the details, it appears that the mixture involves a component on ℜ⁺ and another component on ℜ⁻ with a potential third component as a Dirac mass at zero. Hence, it only nominally writes as a mixture and does not offer the same challenges as a regular mixture. Not label switching. No latent variable. Having mutually exclusive supports solves all those problems and even allows for an indicator function to permeate through the exponential function… (Recall that the special mixture processed in Rubio and Steel also enjoys this feature.)