A Fisher Information Theory of Aesthetic Preference for Complexity

When evaluating sensory stimuli, people tend to prefer those with not too little or not too much complexity. A recent theoretical proposal for this phenomenon is that preference has a direct link to the Observed Fisher Information that a stimulus carries about the environment. To make this theory complete, one must specify the model that the brain has about complexities in the world. Here, we develop this model by first obtaining the distributions of three indices of complexity measured as normalized Shannon Entropy in real-world images from seven environments. We then search for a parametric model that accounts for these distributions. Finally, we measure the Observed Fisher Information that each image has about the parameters of this model. The results show that with few exceptions, the distributions of image complexities are unimodal, have negative skewness, and are leptokurtotic. Moreover, the sign and magnitude of the skewness varies systematically with the location of the mode. After investigating tens of models for these distributions, we show that the Logit-Losev function, a generalization of the hyperbolic-secant distribution, fits them well. The Observed Fisher Information for this model shows the inverted-U-shape behavior of complexity preference. Finally, we discuss ways to test our Fisher-Information theory.