arXiv:2610.33321v1 Announce Type: new Abstract: Estimation and inference under boundary constraints arises in many areas of statistics, but the available asymptotic theory remains limited and fragmented relative to the classical case, where the parameter lies in the interior of the parameter space. In this paper we derive a boundary-constrained analogue of the standard asymptotic expansion for a minimiser of a loss function, in which the deviation of the estimator from the population parameter is decomposed over lower-dimensional subspaces determined by the directional derivatives at the minimiser. Using this decomposition we show that the limiting distribution of the estimator is a mixture of point masses and truncated skew-normal distributions, with mixture probabilities determined by the directional derivative conditions. A consequence of this result is that the constrained and unconstrained coordinates do not separate: the presence of parameters on or near the boundary affects the sampling distribution of the unconstrained parameter estimators. We apply the expansion to obtain the asymptotic distribution of the quasi log-likelihood ratio statistic under a range of null hypotheses, and since the resulting mixture is not pivotal in the quasi-likelihood setting, we propose a simulation-based procedure for approximating critical values. The framework unifies and simplifies several existing results. We illustrate the methodology by testing for interactions in exponential family graphical models, where the parameters are subject to compatibility constraints that ensure a joint distribution exists.