1: \begin{abstract}\noindent

2: The work treats dynamical systems given by ordinary differential equations

3: in the form $\frac {dX^\ve(t)}{dt}=\ve B(X^\ve(t),Y^\ve(t))$ where fast motions

4: $Y^\ve$ depend on the slow motion $X^\ve$ (coupled with it) and they are either

5: given by another differential equation $\frac {dY^\ve(t)}{dt}=b(X^\ve(t),

6: Y^\ve(t))$ or perturbations of an appropriate parametric family of Markov

7: processes with freezed slow variables.

8: In the first case we assume that the fast motions are hyperbolic for each 

9: freezed slow variable and in the second case we deal with Markov processes

10: such as random evolutions which are combinations of diffusions and continuous

11: time Markov chains. First, we study large deviations of the slow motion 

12: $X^\ve$ from its averaged (in fast variables $Y^\ve$) approximation $\brX^\ve.$

13: The upper large deviation bound justifies the averaging approximation on the 

14: time scale of order $1/\ve$, called the averaging principle, in the sense of 

15: convergence in measure (in the first case) or in probability (in the second

16: case) but our real goal is to obtain both the upper and the lower large 

17: deviations bounds which together with some Markov property type arguments 

18: (in the first case) or with the real Markov property (in the second case) 

19: enable us to study (adiabatic) behavior of the slow motion on the much longer 

20: exponential in 

21: $1/\ve$ time scale, in particular, to describe its fluctuations in a vicinity

22:  of an attractor of the averaged motion and its rare (adiabatic) transitions 

23: between neighborhoods of such attractors. When the fast motion $Y^\ve$ does

24:  not depend on the slow one we arrive at a simpler averaging setup studied in

25:  numerous papers but the above fully coupled case, which better describes real

26: phenomena, leads to much more complicated problems.

27: \end{abstract}

28: