Implementations for molecular dynamics on parallel computers generally use either particle parallelism or geometric parallelism. For short-range potentials, geometric parallelism has the advantage that communication can stay restricted to processors nearby. Usually, half the environment around a processor is communicated, using Newton's third law. This poses a problem for the implementation of multi-particle potentials (e.g., "bending" and "torsion"). For instance, if it is said that only one processor should actually calculate the forces on the particles involved, it will be difficult to determine which processor this should be, given that the particles are distributed over two or more processors. We present an efficient technique to do so and prove that it is correct. The technique requires no more communication than the computation of two-particle interactions and ensures that potentials are only evaluated once.