1: \begin{abstract}

2: In a regular full exponential family,

3: the maximum likelihood estimator (MLE) need not exist in the traditional

4: sense, but

5: the MLE may exist in the Barndorff-Nielsen completion of the

6: family.   Existing algorithms for finding the MLE in the 

7: Barndorff-Nielsen completion solve many linear programs;

8: they are slow in small problems and too slow for large problems.

9: We provide new, fast, and scalable methodology for finding the MLE in the 

10: Barndorff-Nielsen completion based on approximate null eigenvectors of

11: the Fisher information matrix.  Convergence of Fisher information follows

12: from cumulant generating function convergence, conditions for which are

13: given.

14: \end{abstract}

15: