For an arbitrary dataset, finite or infinite, it is shown that the following three conditions are equivalent. (i) D satisies GARP; (ii) D can be rationalized by a utility function; (iii) D can be rationalized by a strictly increasing, quasiconcave utility function. Examples of infinite datasets satisfying GARP are provided for which every utility rationalization fails to be lower semicontinuous, upper semicontinuous, or concave. Thus condition (iii) cannot be substantively improved upon.