The original definition of epistasis was the suppression of a gene's expression by the action of a second gene. It is important to note that epistasis occurs between different loci, and not between alternative alleles at a single locus. Though loci may be as close together as base-pairs on a single gene, they are conventionally thought of as separate genes.
Epistasis as a concept has been since modified to accomodate any sort of interaction between loci that will cause a departure from the sum total of their effects from additive expectations. To illustrate, consider two loci A and B each at which there exist two alternative alleles, a1, a2, b1 and b2. Suppose that A is responsible for some phenotypic character that we will express as some real number from 0 to 1. Listing haploid genotypes and their corresponding phenotypic values:
a1, b1 = 0; a2, b1 = 1; a1, b2 = 1;
(additive) a2, b2 = 2
(epistatic) a2, b2 = 1
Here, the effect of changing the allele a1 to a2 or b1 to b2 is +1. We would expect that having both a2 and b2 would cause the effects to sum. However, epistatic interactions between the two alleles may cause the phenotypic value to remain at +1.
Such epistatic interactions may be mediated by one of many possible genetic, biochemical or physiological mechanisms. For example, if the a2 allele causes the loss of function in the glycolysis pathway, then having the b2 allele is redundant.
When the paired effects of two alleles is greater than their expected sum, the interaction is called synergistic. When it is less, the interaction is antagonistic. Interactions have also been classified as positive and negative given the direction that it deflects the effect on some polarised scale. The above example may be considered to be one of negative epistasis. This latter terminology is usually reserved for discussion of the epistatic effects on fitness.
Epistasis remains one of the most difficult biological properties to measure, despite having a tremendous importance in evolutionary genetic theory.