1: \begin{abstract}

2: We consider a bivariate Markov chain $Z=\{Z_k\}_{k \geq

3: 1}=\{(X_k,Y_k)\}_{k \geq 1}$ taking values on product space $\Z=\X

4: \times \Y$, where $\X$ is possibly uncountable space and

5: $\Y=\{1,\ldots, |\Y|\}$ is a finite state-space.  The purpose of the

6: paper is to find sufficient conditions that guarantee the

7: exponential convergence of smoothing, filtering and predictive

8: probabilities: $$\sup_{n\geq t}\|P(Y_{t:\infty}\in

9: \cdot|X_{l:n})-P(Y_{t:\infty}\in \cdot|X_{s:n}) \|_{\rm TV} \leq K_s

10: \alpha^{t}, \quad \mbox{a.s.}$$ Here $t\geq s\geq l\geq 1$, $K_s$

11: is $\sigma(X_{s:\infty})$-measurable finite random variable and

12: $\alpha\in (0,1)$ is fixed. In the second part of the paper, we

13: establish two-sided versions of the above-mentioned convergence. We

14: show that the desired convergences hold under fairly general

15: conditions. A special case of above-mentioned very general model is

16: popular hidden Markov model (HMM). We prove that in HMM-case, our

17: assumptions are more general than all similar mixing-type of

18: conditions encountered in practice, yet relatively easy to verify.

19: \end{abstract}