The free energy principle (FEP), formulated by Karl Friston (2006, 2010), proposes that biological systems — from single cells to entire brains — exist in states that minimize variational free energy, an information-theoretic quantity that bounds the surprise (negative log evidence) of sensory observations. This ambitious framework attempts to unify perception, action, learning, and attention under a single imperative: minimize the divergence between the organism's internal model and the statistical structure of its environment.
Variational Free Energy
Equivalently:
F = D_KL[q(θ) ‖ p(θ|x)] + (−ln p(x))
Since D_KL ≥ 0:
F ≥ −ln p(x) = surprise
Minimizing F → minimizing surprise & improving the internal model
Here x represents sensory data, θ represents the hidden causes of sensory data, p(x, θ) is the generative model (the organism's model of how the world produces sensory signals), and q(θ) is a recognition density — the brain's approximate posterior belief about hidden causes. The free energy F decomposes into two terms: the KL divergence between the approximate and true posterior (a measure of inferential accuracy) plus the surprise of the sensory data (negative log model evidence). Since KL divergence is always non-negative, free energy is always an upper bound on surprise.
Perception and Active Inference
Minimizing free energy with respect to the recognition density q(θ) — holding sensory data fixed — corresponds to approximate Bayesian inference: the brain updates its beliefs to match the posterior distribution over causes. This is perceptual inference or predictive coding. Minimizing free energy with respect to action — changing sensory data by moving in the world — corresponds to active inference: the organism acts to make its sensory input conform to its predictions. This dual optimization provides a unified account of perception and action.
Predictive coding, the hierarchical neural architecture in which higher cortical areas send predictions downward and lower areas send prediction errors upward (Rao & Ballard, 1999), can be derived as the neural process that minimizes variational free energy. Prediction errors are the gradients of free energy with respect to neural activity, and their propagation up the hierarchy implements a form of gradient descent on free energy. This connection provides a neurobiological implementation for the abstract principle and has generated specific predictions about laminar cortical circuitry.
Scope and Criticism
The FEP has been applied to an extraordinarily wide range of phenomena: interoception and emotion, psychopathology (aberrant precision-weighting in schizophrenia and autism), habit formation, curiosity (minimizing expected free energy about future states), morphogenesis, and even the behavior of simple organisms. Friston argues that the principle is not merely a model of brain function but a fundamental property of any system that maintains itself in a nonequilibrium steady state — a claim that connects it to the physics of self-organization.
Critics have raised several concerns. Some argue that the principle is too general to be falsifiable — that almost any behavior can be redescribed as free energy minimization with an appropriate choice of generative model. Others contend that the mathematical formalism, while elegant, requires strong assumptions (Gaussian approximations, particular factorizations of the posterior) that may not hold in biological systems. The relationship between variational free energy and thermodynamic free energy remains debated. Nevertheless, the FEP has been enormously influential in computational neuroscience, providing a unifying language for disparate research programs and generating novel empirical predictions.