In Karmarkar's interior-point linear programming algorithm,
the potential function is not convex.  This paper proves that its
multiplicative version is strictly convex when the feasible
region is bounded.  This fact can be used in the line search stage.