1: \begin{abstract}

2:     There is substantial empirical evidence about the success of dynamic

3:     implementations of Hamiltonian Monte Carlo (HMC), such as the No

4:     U-Turn Sampler (NUTS), in many challenging inference problems but

5:     theoretical results about their behavior are scarce.  The aim of

6:     this paper is to fill this gap. More precisely, we consider a

7:     general class of MCMC algorithms we call dynamic HMC. We show that

8:     this general framework encompasses NUTS as a particular case, implying

9:     the invariance of the target distribution as a by-product.  Second,

10:     we establish conditions under which NUTS is irreducible and

11:     aperiodic and as a corrolary ergodic.

12:   Under conditions similar to the ones existing for HMC, we also show that NUTS is geometrically ergodic.

13:   Finally, we improve existing convergence results for HMC showing that this method is ergodic without any

14:   boundedness condition on the stepsize and the number of leapfrog

15:   steps, in the case where the target is a perturbation of a Gaussian

16:   distribution. 

17: \end{abstract}

18: