The very essence of large-eddy simulation is that the computed flow field contains only eddies of size ≥Δ. This restriction determines the eddy viscosity ν e . Although the principle is quit apparent, it is not yet fully understood how to derive ν e from that. Therefore we address two related questions, namely how much eddy viscosity is needed to damp any disturbances having a scale smaller than Δ and when does eddy viscosity stop the production (by means of vortex stretching) of scales <Δ. From this we deduce that the ν e has to depend on the invariants q and r of the filtered strain tensor. The simplest model is then given by ν e =c 2 Δ2 r +/q.