The digression on variability and its genetic control sets the stage to consider the issued of evolvability in a biological context. If the expression of genetic variation is itself under genetic control, then it is conceivable that species evolve "strategies" of how to allocate the phenotypic effects of genetic mutations. But what exactly is evolvability and what influences its degree?
In the field of evolutionary algorithms, it is essential to understand which algorithm will effectively improve by random variation and selection. Therefore most of the theory of evolvability has been developed in the context of evolutionary algorithm theory (Jones and Rawlins, 1993; Altenberg. 1994). A comprehensive concept of evolvability was published recently by Lee Altenberg (1994). It is based on a generalization of Price's covariance theorem of natural selection (Price, 1969).
Evolvability Theorem (Altenberg, 1994): the probability that a population generates individuals fitter than any existing is
where a is the maximal rate at which new genotypes are generated by mutation and/or recombination. R(wmax) is the probability that a random sampling of genotypes yields a genotype with fitness larger than the currently best, 
is the average search bias, the probability that the mutation and/or recombination of the given genotype produces better genotypes and 
is the covariance between the current fitness of genotypes and the probability to produce better ones by mutation and or recombination (for details see Altenberg, 1994).
In short this theorem says that evolvability depends on two main factors: the rate of production of genetic variation by what ever means, and a correlation between the current fitness of genotypes and the likelihood to obtain even better genotypes from the already good ones. This theorem defines what the intuitive notion of a "smooth" adaptive landscape suggests: it is easy to evolve by natural selection if the adaptive landscape is smooth, which means that the better genotypes are found in the mutational "neighborhood" of the good genotypes.
Another way of expressing this result is that adaptations are possible if their improvement can be achieved in a cumulative or stepwise fashion. This has been known for some time, but what are the structural features that make stepwise improvement possible? The key feature is the covariance term in Altenberg's theorem. On average, further improvements must not compromise past achievements. This is the essence of the so-called "building block hypothesis" to explain the performance of Genetic Algorithms (Holland, 1992; Forrest and Mitchell, 1993). In terms of the representation of phenotype in the "genomic code" this implies that independent biological functions shall be coded independently so that their improvement can be also be realized with minimal interference. This leads to the concept of modularity underlying the various explanations of complex adaptations offered by biologists.