A fundamental tenet of statistical mechanics is that the rate of collision of two objects is related to the expectation value of their relative velocities. In pioneering work by Saffman and Turner [J. Fluid Mech.1, 16 (1956)], two different formulations of this tenet are used to calculate the collision kernel &Ggr; between two arbitrary particle size groups in a turbulent flow. The first or spherical formulation is based on the radial componentwrof the relative velocitywbetween two particles:&Ggr;sph=2&pgr;R2⟨|wr|⟩,wherewr=w⋅R/R,Ris the separation vector, andR=|R|.The second or cylindrical formulation is based on the vector velocity itself:&Ggr;cyl=2&pgr;R2⟨|w|⟩,which is supported by molecular collision statistical mechanics. Saffman and Turner obtained different results from the two formulations and attributed the difference to the form of the probability function ofwused in their work. A more careful examination reveals that there is a fundamental difference between the two formulations. An underlying assumption in the second formulation is that the relative velocity at any instant is locally uniform over a spatial scale on the order of the collision radiusR, which is certainly not the case in turbulent flow. Therefore, the second formulation is not expected to be rigorously correct. In fact, both our analysis and numerical simulations show that the second formulation leads to a collision kernel about 25&percent; larger than the first formulation in isotropic turbulence. For a simple uniform shear flow, the second formulation is about 20&percent; too large. The two formulations, however, are equivalent for treating the collision rates among random molecules and the gravitational collision rates. ©1998 American Institute of Physics.