Negative Eigenvalues due to Mesh Refinement

I will keep my post brief to begin with.

The model consists of 7 rectangular interference fits (with 2 of the surfaces in direct interference). The model appears to be working perfectly fine at simulating the stresses induced due to the fits for a variety of mesh sizes. Once the mesh sizing of the slave components is reduced to below 0.4 the job fails due to experiencing negative eigenvalues.

To summarize - how can mesh refinement cause negative eigenvalues?