A Bernstein-von Mises Theorem for Generalized Fiducial Distributions

Abstract: We prove a Bernstein-von Mises result for generalized fiducial distributions following the approach based on quadratic mean differentiability in Le Cam (1986); van der Vaart (1998). Building on their approach, we introduce only two additional conditions for the generalized fiducial setting. While asymptotic normality of generalized fiducial distributions has been studied before, particularly in Hannig (2009) and Sonderegger and Hannig (2014), this work significantly extends the usefulness of such a result by the much more general condition of quadratic mean differentiability. We demonstrate the applicability of our result with two examples that necessitate these more general assumptions: the triangular distributions and free-knot spline models.