The velocities used by the original, standard version of the Bryan-Cox mode
l to advect momentum are shown to be inconsistent with the main velocities
over steps in the bathymetry. In particular, the vertical shears of the two
sets of horizontal velocities disagree at the top edges of steps in the ba
thymetry. This is shown to be related to the way in which the velocities ar
e divided into baroclinic and barotropic parts when the barotropic Bow is r
epresented by a streamfunction. A list of desirable properties for the velo
cities at the faces of the velocity cells is suggested. Five schemes for ca
lculating them, including Webb's scheme and two new schemes (which are modi
fied versions of the original scheme), are discussed and their properties d
erived. Simple estimates are also made of the effect of grid-scale noise in
the streamfunction and thermal fields on these horizontal and vertical vel
ocities. The velocities from short integrations of a 1 degrees global model
using these schemes are compared. The horizontal velocities produced by th
e original scheme suffer from gridpoint noise in the streamfunction. One of
the new schemes represents these velocities well if the streamfunction is
smoothed appropriately. While the vertical velocities with this scheme are
less noisy than with the original scheme, they are still inferior to those
generated by Webb's scheme. It is argued that in any model with realistic s
tepped bathymetry Webb's scheme will be the best one to use.