The operation of complementary distribution can be performed upon the new elements as well as upon the old. Theoretically, therefore, we could break the allophones into components and then do all the complementary grouping on the components. Actually, it is more efficient to group the complementary allophones into tentative phonemes, and to analyze these tentative phonemes into components. We can then try to group the components by complementary distribution in order to get fewer components, each having wider coverage. If certain limitations of occurrence exist for some components, we may even try to express their limitations in turn by a second extraction of components, on much the same grounds that we used in expressing phonemic limitations by components, in order to obtain the most general and least limited set of elements.
- Harris (1944), a pag.202