We study the accumulation of errors in cosmological N-body algorithms that are caused by representing the continuous distribution of matter by massive particles, comparing the PPPM and Adaptive Multigrid codes. We use for this a new measure of two-body relaxation suitable for nonstationary gravitating systems. This measure is based on accumulated deflection angles of particle orbits and it does not saturate during evolution. We have found that the role of gravitational collisions in standard PPPM-models is rather high, and in order to avoid collision effects we recommend to use large comoving softening parameters epsilon>=1.0 (in grid units).

We have developed an adaptive multigrid code for solving the Poisson equation in gravitational simulations. Finer rectangular subgrids are adaptively created in locations where the density exceeds a local level-dependent threshold. We describe the code, test it in cosmological simulations, and apply it to the study of the birth and evolution of a typical pancake singularity. The initial conditions for the pancake are generated on the basis of the theory of Lagrangian singularities; we follow its evolution for a few collapse times, finding a rich substructure in the final object. We achieve a spatial resolution of 1/1024 of the size of the overall computational cube in the central parts of the pancake, with computing time comparable to that of the FFT-solvers.