1: \begin{abstract}

2: \vspace{-3pt}

3: The problem of drawing samples from a discrete distribution can be

4: converted into a discrete optimization problem \cite{Papandreou11,

5: tarlow2012randomized,hazan2012partition,ermon2013embed}.  In this

6: work, we show how sampling from a continuous

7: distribution can be converted into an optimization problem over 

8: continuous space.

9: Central to the method 

10: is a stochastic process recently described in mathematical statistics

11: that we call the

12: \emph{\ourprocess}. We present a new construction of the \ourprocess~and \emph{\OurAlgorithm},

13: a practical generic sampling algorithm that searches for the maximum of a Gumbel process

14: using \astar~search. 

15: We analyze the correctness and convergence time of \OurAlgorithm~and demonstrate empirically that it

16: makes more efficient use of bound and likelihood evaluations than the most closely

17: related  adaptive rejection sampling-based algorithms.

18: \end{abstract}

19: