In the literature on electoral politics, full convergence of policy platforms is usually regarded as socially optimal. The reason is that risk-averse voters prefer a sure middle-of-the-road policy to a lottery of two extremes with the same expectation. In this paper, we study the normative implications of convergence in a simple model of electoral competition, in which parties are uncertain about voters' preferences. We show that, if political parties have incomplete information about voters' preferences, the voters may prefer some degree of policy divergence. The intuition is that policy divergence enables voters to correct policies that are based on a wrong perception of their preferences.