1: \begin{abstract}%

2:   The mixing time $\tmix$ of an ergodic Markov chain measures the rate of 

3: convergence towards its stationary distribution $\bpi$.

4: We consider the problem of estimating $\tmix$ from one single trajectory 

5: of $m$ observations $(X_1, \dots, X_m)$, 

6: in the case where the transition kernel $\mc$ is unknown, a research program started

7: by \citet{hsu2015mixing}.

8: The community has so far focused primarily on leveraging spectral methods to estimate the

9: \emph{relaxation time} $\trel$ of a \emph{reversible} Markov chain as a proxy for $\tmix$.

10: Although these techniques have recently been extended to tackle non-reversible chains, 

11: this general setting remains much less understood.

12: Our new approach based on contraction methods is the first that aims at directly 

13: estimating $\tmix$ up to multiplicative small universal constants instead of $\trel$.

14: It does so by introducing a generalized version of Dobrushin's contraction coefficient $\tgencc$, which

15: is shown to control the mixing time regardless of reversibility.

16: We subsequently design fully data-dependent high confidence intervals around $\tgencc$ 

17: that generally yield better convergence guarantees and 

18: are more practical than state-of-the-art.

19: \end{abstract}