Fix a measurable space , a -finite base measure on it, a finite-dimensional real vector space with dual and pairing , and a measurable sufficient statistic . The log-partition (cumulant) function

has as its effective domain the natural parameter space , assumed open and non-empty. The associated exponential family is

The potential is smooth and strictly convex on , so the family is regular.

Property. Dual flatness

The natural coordinates are affine for the exponential (-) connection; the expectation coordinates are affine for the mixture (-) connection, and is a diffeomorphism onto the open convex set . The free energy and its Legendre conjugate , (the negative entropy), satisfy

with Fisher metric .

Property. The maximum-entropy characterisation

Fixing the means and maximising the entropy forces the Gibbs form . This is an exponential family whose sufficient statistics are exactly the constrained , with the Lagrange multipliers as the natural parameters. Matching the means, , is the first-order condition of the convex dual . Equivalently it is the minimum Kullback–Leibler (KL) distribution to under the constraints, and KL restricted to is the Bregman divergence of . Maximum entropy, exponential families, Bregman divergence, and dual-flat geometry are thus four readings of one convex-duality picture.

Property. The two intrinsic geodesics are closed for free

The dual affine charts give two deterministic interpolations intrinsic to : the -geodesic (normalised geometric mean ) and the -geodesic (moment matching). Both stay in automatically because and are convex. This carries no content beyond dual flatness and should not be confused with a probabilistic operation on ; see the interpolation of distributions.

Example. Deformed ( ) generalization

Replacing by a strictly increasing (a deformed logarithm) and requiring gives the -deformed exponential families, dually flat in the -geometry. The endpoints and are the exponential and mixture families; Tsallis / Kaniadakis deformations give the -/-exponential families (Student-, -Gaussian).

Adding a dispersion index to a natural exponential family gives the exponential dispersion models, the reproductive/Tweedie convolution axis.