A theoretical analysis is made of a cyclical selection model exhibiting 'marginal overdominance' (in the sense of geometric mean overdominance) in small populations. When the fitnesses are additive, geometric mean overdominance produces similar results as constant overdominance: fixation is retarded most effectively at central equilibrium gene frequencies and can be accelerated (as compared to the neutral case) at extreme equilibrium frequencies (Robertson, 1962). However, with increasing degree of dominance with respect to fitness, the maximal retardation of fixation is shifted towards low equilibrium frequencies of the 'recessive' allele. From a comparison between the cyclical selection models and appropriate constant selection models it is concluded that for low equilibrium gene frequencies and (some degree of) dominance, geometric mean overdominance resulting from cyclical selection is more effective in maintaining genetic variation than constant overdominance, while at intermediate equilibrium frequencies the reverse holds. The results are discussed in relation to the idea of using theoretical distribution patterns of heterozygosity under various assumptions in order to discriminate among different hypotheses on the maintenance of enzyme polymorphisms.