Criteria for large deviations (Q4806464)

This paper addresses the relation between the existence of \(\Lambda(h)=\log\lim\mu^{t_\alpha}(e^{h/t_\alpha})\) for any bounded continuous Borel-measurable real-valued function \(h\) and that of a large deviation principle for \((\mu_\alpha)\) with powers \((t_\alpha)\). Here \((t_\alpha)\) is a net in \(]0,\infty[\) converging to 0, and \((\mu_\alpha)\) is a net of Borel probability measures on a topological space \(X\). From Varadhan's theorem, if \(X\) is regular, then the LDP with tight rate function implies \(\Lambda(h)\) exists for all suitable \(h\), including the class above, and there is an expression linking the two. This paper presents a new approach to this relationship based upon a variational representation of the functional \(\lim\sup \mu_\alpha^{t_\alpha}(e^{\cdot/t_\alpha})\), and a criterion of existence of \(\Lambda(\cdot)\) when \(X\) is normal. This leads to functional and set-theoretic large deviation criteria that permits the removal of tightness from the conditions of Varadhan's result. Indeed, there are no compactness arguments in the paper. Proposition 2.3 begins by providing a non-standard statement equivalent to the large deviation principle and which implies the existence of a minimal rate function. Expressions are given for the rate function. Under the (localized) tail condition of Varadhan's theorem, Theorem 3.1 obtains a general variational form for \(\lim\sup \mu_\alpha^{t_\alpha}(e^{h/t_\alpha}1_Y)\), and Theorem 3.3 obtains necessary (and sufficient for \(X\) normal) conditions for the existence of \(\Lambda\) for bounded continuous \(h\). Section 4 assumes \(X\) is normal, and develops (Theorem 4.1) necessary and sufficient conditions on \(\Lambda\) for a large deviation principle. As corollaries of Theorem 4.1, the paper obtains a number of results without the usual tightness conditions, for example the equivalence between the Laplace and large deviation principles, or the variational form of the rate function.