Understanding the star formation process is a key issue in astronomy. Since direct observation provide only very limited information, this issue is investigated by models. Such models need to take into account complex physical processes while following the gas dynamics, so that simulations need a lot of time to run and do not follow the star formation process for longer than 0.2 Myr. The best known observational results concerns the field population, evolved open clusters or younger clusters or associations, which are between 1 Myr and a few Gyr old. Therefore in order to compare the results from models to known observations, we need to bridge the gap between the two. Star formation appears to produce groups of stars from the collapse of turbulent molecular clouds. As stars form, the gas is progressively ejected from the cluster, and the evolution is dominated by gravitational interactions. Following the dynamical evolution of a group of star using N-Body codes is a standard way used to constraint the models and understand the origin of the different populations. Star formation may produce sub-structure or small groups that merge to form bigger entities, or end up as loose association. In my thesis I focused on the dynamics of small groups, that have not been investigated as thoroughly as 1000 or 10^4 star groups. I performed N-Body simulations of small stellar groups, with N<100, in order to study their dynamics using a statistical approach, made possible by running a large number of simulations, and to find some observational signatures of given initial conditions. This approach enable to take full account of stochastic effects due to dynamical interactions. Using a large number of initial configurations (with N=20, 50, 100, a typical radius from 0.025 pc to 1 pc) and a sample of 500 simulations per configuration, I looked at equal mass groups as well as groups having a mass spectrum, without any binary initially. Such small groups show similar evolution to bigger groups, but with faster and less pronounced collapse phase. I described the average behaviour of slow expansion of the cluster, and an alternative evolution, occurring with 17% probability, that ended in the complete dissolution of the group due to ejection of a central binary. Searching for a way to identify the initial configuration from observational measure, I looked at the complementarity of density and 3D velocity and was able to show that it could be sufficient in some cases to determine the initial density. Further investigations are needed to take into account the information on the binary population and will be used to investigate the formation of known associations or young regions. Finally, I ran a large number of simulations, aiming at reproducing the observed state of the eta Chamaeleontis from standard initial conditions and pure dynamical evolution. This association properties are consistent with a dynamical evolved cluster, namely low-mass object poor and having only tight binaries. I showed that these properties cannot be reproduced with pure dynamical evolution from standard initial mass function and binary population, meaning that its particular features must have been pristine.
