Natural selection subsumes and unites multiple theories of perceptual compression

Perceptual processing is thought to occur in two stages: compression of external world information, followed by inference of external world state. While inference is believed to be Bayes-optimal, the model for optimal compression is comparatively less well-established: multiple signal processing frameworks have been used to explain empirical outcomes and hypothesized to align with Darwinian fitness. Here I endeavor to constrain the model space by framing fitness as a function of the compression scheme and optimizing fitness directly. I find that signal processing and fitness objectives often deviate; for example, in a finite game where risk-avoidant bet-hedging beats maximization of average performance. However, when average performance determines fitness, then both are maximized by minimized cost-weighted signal distortion. Meanwhile, the signal processing metrics of estimation error and mutual information align with these latter objectives only under very specific circumstances. Taken together, my results establish a Darwinian (nested) hierarchy of perceptual compression objectives, which extends the existing understanding of optimal compression by incorporating not just signal processing but also economic frameworks, and reconciles the extant usages of the former by clarifying their respective domains of applicability. The evolutionary importance of the cost-weighted distortion objective is additionally intriguing because this objective is potentially congruent with a broad set of experimental findings.