The free energy principle is a theoretical framework proposing that the organization and activity of certain self-maintaining systems can be described in terms of minimizing a quantity called variational free energy. Developed principally by Karl Friston, it connects ideas from neuroscience, statistics, and statistical mechanics. In its best-known application, the brain continually updates an internal model of the causes of sensory input, while actions alter the input it receives. Perception and action are thereby treated as complementary aspects of an inference process.
The principle is broader than the claim that brains predict their sensory inputs. It seeks to explain how organisms maintain characteristic patterns of organization despite continual exchanges with their surroundings. Its central mathematical quantity is an information-theoretic bound on the improbability of observations under a model, not ordinarily a directly measurable store of physical energy. The relationship between this variational quantity and thermodynamic free energy is important but requires additional assumptions.
The framework encompasses several distinguishable claims: mathematical identities about probability distributions, proposed descriptions of persistent stochastic systems, models of neural processing, and practical algorithms for adaptive behavior. These components differ in their evidential status. The underlying variational identities are established mathematics; whether particular biological systems implement the proposed mechanisms, and how broadly the physical interpretation applies, remain subjects of research and debate.
Origins and intellectual development
The free energy principle emerged from work on statistical inference in the brain rather than from a single experimental discovery. An important predecessor was Hermann von Helmholtz’s nineteenth-century account of perception as unconscious inference: sensory experience depends on interpreting ambiguous signals in light of previous experience. A retinal image, for example, does not uniquely determine the objects, illumination, and distances that produced it. An inferential account explains perception as the selection of plausible causes rather than the passive registration of stimulation.
During the twentieth century, this perspective became connected to Bayesian inference, in which prior expectations and incoming evidence jointly determine beliefs about hidden causes. Information theory, statistical learning, and neural-network research supplied mathematical tools for making such accounts explicit. Geoffrey Hinton and Richard Zemel’s 1994 work on minimum-description-length learning used a variational free-energy objective in connection with learning representations. Peter Dayan, Hinton, Radford Neal, and Zemel’s Helmholtz machine, described in 1995, likewise connected generative models with approximate inference. These developments were antecedents of the broader biological principle, not formulations of that principle in its later form.
Another important strand was predictive coding. In a widely discussed 1999 model, Rajesh Rao and Dana Ballard showed how hierarchical prediction and error correction could account for aspects of visual cortical responses. Friston’s 2005 article “A theory of cortical responses” developed related ideas using hierarchical empirical Bayesian models. His 2006 article “A free energy principle for the brain,” followed by the 2010 review “The free-energy principle: a unified brain theory?”, presented the framework as a candidate unifying account of brain function.
The scope subsequently expanded beyond sensory inference. Active inference treated action selection and bodily regulation as parts of the same inferential architecture. Later work placed greater emphasis on stochastic dynamics, statistical boundaries, and the conditions under which a system can be described as maintaining a model of its environment. This expansion connected neural theories with questions about organismic individuality and the persistence of biological organization.
The resulting literature is not a single fixed model. Some studies analyze abstract dynamical systems; others construct detailed models of eye movements, learning, or decision-making. Still others investigate philosophical consequences concerning representation, agency, and biological purpose. Historical continuity among these projects does not make their claims interchangeable: success in modeling a perceptual task does not, by itself, establish a universal principle of living organization.
Variational free energy and Bayesian inference
The central mathematical construction begins with observations , hidden states , and a generative model . The model specifies how hidden states and observations are probabilistically related. Its marginal likelihood, or model evidence, is
Calculating this quantity, and the corresponding posterior , can be difficult when hidden states are numerous or interact nonlinearly. Variational inference addresses this difficulty by introducing a tractable distribution that approximates the posterior. Here “belief” means a probability distribution or its parameters; it need not imply conscious judgment.
Variational free energy is defined as
Using Bayes’ rule, this becomes
where Kullback–Leibler divergence measures a discrepancy between distributions and is nonnegative. Consequently,
The quantity is called surprisal or self-information. For a fixed model and observation, minimizing over reduces the divergence between the approximate and exact posterior. The bound becomes exact when the two distributions coincide.
A second decomposition clarifies the trade-off involved:
Accuracy rewards explanations that make the observations likely. Complexity penalizes departures from prior beliefs. This does not mean that complicated organisms or rich representations are intrinsically disfavored. Rather, changes in beliefs must earn their statistical cost by improving explanatory accuracy. The negative of variational free energy is commonly called the evidence lower bound, or ELBO, in machine learning.
Several qualifications are essential. First, reducing free energy by changing does not necessarily change the probability of the observation: with the model fixed, it improves the approximation to an existing posterior. Learning model parameters is a separate optimization. Second, the achieved result depends on the allowed family of approximate distributions and the optimization procedure. Local minima, restrictive approximations, and limited computation can all prevent exact inference.
Finally, “surprise” here is a technical quantity, not necessarily a subjective feeling of astonishment. An ordinary event can have high surprisal under an unsuitable model, while a consciously startling event can be assigned substantial probability by a different model. For continuous observations, probability densities also depend on the measurement coordinates and reference measure. Statements about minimizing surprise therefore require a specified probabilistic setting rather than an unrestricted comparison of arbitrary signals.
Generative models, perception, and neural implementation
A generative model specifies how observations could arise from unobserved causes. In vision, hidden variables might include an object’s location, surface properties, and illumination. In bodily regulation, they might include blood chemistry, tissue state, or the causes of visceral sensations. A hierarchical model places variables at several levels: higher levels may describe relatively abstract or slowly changing causes, while lower levels describe more immediate sensory consequences.
Under common assumptions, variational inference in such a hierarchy can be implemented through prediction-error reduction. Higher levels generate predictions about lower-level activity; discrepancies between predicted and observed activity drive revisions to the inferred causes. For approximately Gaussian errors, part of the objective has the form
where is a vector of prediction errors and is a precision matrix, the inverse of a covariance matrix. The additional terms include contributions from uncertainty and normalization. This expression explains why free-energy minimization can resemble weighted error correction without being identical to minimizing raw sensory mismatch.
Precision is especially important. An error associated with a reliable signal should exert greater influence than an equally large error associated with a noisy signal. In perceptual models, precision can therefore regulate the relative influence of sensory evidence and prior expectations. Some neural accounts associate this regulation with changes in neuronal gain, attention, or neuromodulatory signaling. These are proposed implementations, not consequences forced by the variational identity alone.
Predictive-coding models often assign distinct functional roles to neuronal populations carrying predictions and prediction errors. Proposed cortical implementations involve reciprocal connections, with ascending signals conveying errors and descending signals conveying predictions. Real cortical anatomy includes recurrent, lateral, inhibitory, and thalamic interactions that complicate any simple correspondence. The identification of a recorded neuron as a “prediction-error neuron” consequently requires evidence beyond the observation that its response changes when a stimulus becomes expected.
Perception, learning, and uncertainty estimation can operate on different timescales. Fast changes in neural activity may update estimated states; slower changes in synaptic efficacy may alter model parameters; still slower processes may modify the model’s structure. This supplies a common mathematical vocabulary for processes that are biologically distinct.
Predictive coding is not synonymous with the free energy principle. Predictive coding is a family of proposed message-passing mechanisms, while the principle is a broader claim about the organization of adaptive systems. Variational inference can use algorithms other than predictive coding, and predictive-coding architectures can be studied without accepting the principle’s broader biological interpretation. Separating these levels is necessary when evaluating neural evidence.
Action as the control of sensory consequences
In active inference, action complements perception by changing the observations an agent receives. Perception adjusts beliefs to account for sensory input; action changes the agent–environment relationship so that subsequent input better accords with its model. This relationship is particularly clear in continuous motor-control formulations, where predictions about bodily states can guide movements through peripheral and spinal control mechanisms.
Suppose an organism predicts a particular joint configuration. Its actual proprioceptive input initially differs from that prediction. One way to reduce the discrepancy is to revise the predicted configuration. Another is to move the limb until sensory input approaches the predicted state. In simplified mathematical models, action can follow a gradient of free energy through its effects on observations:
where specifies an action-update rate or metric. This expression is a modeling choice whose biological adequacy depends on how motor commands, sensory consequences, delays, and bodily mechanics are represented.
The framework does not require an organism to possess an explicit symbolic description of every muscle or environmental consequence. Some formulations exploit relatively direct sensorimotor mappings, while others use learned transition models to predict longer sequences. Reflex pathways, biomechanics, and environmental constraints can contribute to implementing the resulting behavior. A model that generates appropriate trajectories nevertheless remains distinct from a demonstrated account of the physiological mechanism.
The control-theoretic connection is substantial. Conventional controllers compare a measured state with a target and use feedback to reduce deviations. Active-inference models express related operations in probabilistic terms, allowing uncertainty about states and targets to affect behavior. Under particular assumptions, their equations resemble familiar controllers. Such equivalences clarify shared structure rather than showing that all control theories are merely special cases in every relevant respect.
A frequently raised objection is the “dark-room problem”: if organisms minimize sensory surprise, why do they not remain in unchanging, unstimulating environments? The standard response is that surprise is evaluated relative to a model containing expectations about bodily viability and ordinary activity. Hunger, thirst, changing bodily signals, and learned expectations can make prolonged inactivity highly inconsistent with that model. In temporally extended models, information-seeking can also make exploration preferable to passive predictability.
This response places explanatory weight on how the model is specified. A theory cannot explain an organism’s behavior merely by assigning it whatever expectations reproduce that behavior after the fact. Useful action models must independently constrain physiological requirements, sensory mappings, learning histories, and uncertainty, then predict how behavior changes when those conditions are manipulated.
Expected free energy and the selection of policies
For decisions involving future outcomes, active-inference models commonly evaluate policies: possible sequences of actions or conditional action plans. A policy might specify where to look next, which route to take, or whether to sample an uncertain option before committing to a choice. Because future observations are not yet available, policy evaluation cannot simply minimize variational free energy for a currently observed outcome. It instead uses predicted future states and observations.
A common objective is expected free energy, often denoted for policy . Its precise expression depends on the model and on how preferred outcomes are introduced. In a widely used decomposition, it combines a preference-related term with an information-gain term:
Here denotes predictive beliefs, while assigns higher probability to preferred outcomes. This decomposition is exact for an appropriately defined objective with consistent predictive and conditional distributions; other formulations obtain related expressions using additional approximations. Presenting every expected-free-energy equation as interchangeable can conceal differences in assumptions.
The first term favors outcomes consistent with preferences. The second favors observations expected to improve knowledge of hidden states. Such knowledge-seeking is often called epistemic value. In a maze, for example, visiting a location that reveals which route contains food may initially be useful even if it does not itself provide food. Once the relevant uncertainty is resolved, repeated sampling of the same cue should lose epistemic value.
Some models extend information gain from hidden states to uncertain model parameters. This distinguishes learning where an object is from learning how reliable a cue is or how an environment changes. The latter can support exploration whose benefits emerge across several episodes. These distinctions matter because uncertainty reduction does not have one universal behavioral signature.
Policies are often assigned probabilities using a softmax-like rule,
where encodes prior policy tendencies and controls sensitivity to expected free-energy differences. Practical models may also incorporate evidence from observations already obtained. Habits, deliberation, and confidence can therefore enter through different components rather than being represented by a single reward signal.
These models overlap with reinforcement learning, Bayesian experimental design, and optimal control, but the objectives need not be identical. A distinctive emphasis is the joint treatment of preferred outcomes and information acquisition. This does not make exploration automatic or universally adaptive: its quality depends on the model’s accuracy, planning horizon, available actions, and representation of uncertainty.
Markov blankets and the boundaries of a system
The extension of the free energy principle beyond brains relies heavily on the concept of a Markov blanket. In a probabilistic description, a blanket is a set of variables that renders one group of variables conditionally independent of another. If internal states are , external states are , and blanket states are , the defining relation is
Knowing the blanket variables makes internal and external variables conditionally independent in the specified distribution. This is a statistical statement, not by itself a claim about a literal membrane, causal autonomy, or conscious individuality.
In common free-energy formulations, blanket states are partitioned into sensory and active states. Sensory states mediate the influence of external states on internal states, while active states mediate the system’s influence on its surroundings. Particular dynamical formulations impose restrictions on which state derivatives depend directly on which other variables. Those dynamical restrictions must be distinguished from conditional independences in a probability distribution: one does not establish the other without suitable assumptions.
A cell membrane provides an intuitive biological example because receptors and transport processes mediate interactions between a cell and its environment. However, the membrane is not automatically a Markov blanket in the technical sense. Its suitability depends on the chosen variables, temporal resolution, and treatment of unobserved influences. Similarly, the skin alone is not an adequate universal statistical boundary for every model of an organism.
The inferential interpretation introduces a further step. Internal states may parameterize a distribution over external states, so changes in internal states can be described as changes in beliefs about the environment. Establishing such a correspondence requires a mapping between physical states and statistical parameters. It is not enough merely to find a conditional-independence relation and then conclude that the system performs Bayesian inference in a substantive cognitive sense.
Nested boundaries motivate multiscale applications: cells within tissues, tissues within organisms, and organisms within social groups. But a successful boundary description at one scale does not guarantee an equivalent description at another. Coarse-graining can erase relevant dependencies, introduce memory effects, or change the form of the dynamics.
These distinctions make Markov blankets useful but demanding modeling devices. They offer a precise way to ask what information connects a system to its surroundings. Whether they identify biologically privileged individuals, support an inference interpretation, or yield an experimentally useful model must be assessed separately for each application.
Stochastic dynamics and the relation to thermodynamics
Biological systems persist far from thermodynamic equilibrium. They exchange matter and energy, consume resources, and dissipate heat. The free energy principle attempts to connect this persistence with the probability of occupying characteristic states. A living organism does not explore every physically possible configuration uniformly: it remains within a restricted, structured range for a finite period.
A common starting point is a stochastic differential equation,
where is a drift field, is a diffusion matrix, and is a Wiener process. Under suitable conditions, the associated probability density approaches a stationary distribution . The potential-like quantity
then measures stationary surprisal. Frequently occupied states have lower than rarely occupied states.
Stationarity does not imply equilibrium. A nonequilibrium steady state can maintain a time-independent density while supporting circulating probability currents. Its dynamics therefore need not be pure gradient descent on surprisal. Rotational, or solenoidal, components can move the system along contours of equal probability. Noise also allows individual trajectories to move temporarily toward less probable states. Descriptions that reduce all biological activity to monotonically descending a scalar objective omit these features.
Connections between such dynamics and variational inference require additional structure: a suitable partition into internal, external, and blanket variables; assumptions about dependencies; and a mapping between internal states and probability distributions. The existence of a stationary density alone does not show that a system’s physical trajectory implements a particular inference algorithm. Nor does every biologically interesting process remain stationary over development, injury, environmental change, or death.
The word “free energy” can cause confusion. Helmholtz free energy and Gibbs free energy are thermodynamic state functions with dimensions of energy. Variational free energy is ordinarily an information-theoretic quantity expressed in nats when natural logarithms are used. In equilibrium statistical mechanics, a relation such as permits a precise correspondence between negative log probability and physical energy scales. Outside that setting, the correspondence cannot simply be assumed.
Nothing in the principle exempts organisms from the second law of thermodynamics. Maintaining local organization requires exchanges that contribute to entropy production in the wider system. Minimizing a variational objective is not equivalent to minimizing metabolic expenditure, total entropy production, or every thermodynamic potential. Connections to those quantities must be derived for a specified physical model.
Homeostasis, adaptation, and biological organization
The biological motivation for the principle is persistence: organisms must keep important variables within viable ranges despite environmental disturbances. Homeostasis includes the regulation of variables such as body temperature, osmotic balance, and circulating metabolites. These variables need not remain perfectly constant; normal function often requires structured fluctuations and context-dependent changes.
In an inferential model, preferred or expected bodily states can constrain perception and action. If a model assigns high probability to physiologically viable states, deviations from those states can drive corrective behavior. Obtaining water, moving toward a suitable thermal environment, or altering food intake can then be modeled as actions that restore expected bodily conditions. The relevant expectations are not necessarily verbally represented beliefs. They may be embodied in neural circuitry, endocrine regulation, and the organization of peripheral control systems.
Allostasis emphasizes anticipatory regulation rather than correction only after a disturbance occurs. An organism may adjust circulation before exertion or seek resources before a deficit becomes severe. Generative models with temporal depth offer one way to formalize such anticipation. They also show why the probability of an outcome and its biological desirability should not always be identified: an organism can accurately predict a harmful event while acting to prevent it.
Applications to development and morphogenesis extend the same vocabulary to the acquisition and maintenance of anatomical form. Cells communicate through chemical and mechanical signals; their behavior depends on local context and developmental history. Some theoretical work describes collective pattern formation as inference toward expected tissue configurations. Such models can organize hypotheses about coordination, but they do not replace established mechanisms involving gene expression, molecular signaling, mechanical forces, and cellular metabolism. A high-level variational description must connect to these mechanisms to provide a biological explanation.
An evolutionary interpretation proposes that inherited organization constrains which states are expected or preferred. Natural selection can favor regulatory architectures that keep organisms within viable conditions. This does not mean that evolution literally optimizes one universal free-energy function, nor that an organism’s immediate objective is always the maximization of reproductive success. Evolutionary change, individual learning, and moment-to-moment regulation operate over different variables and timescales.
The distinction between persistence and adaptation is particularly important. A stable pattern can persist without being alive, learning, or serving a biological function. Conversely, living systems can undergo irreversible transformation rather than return to a fixed set point. Explanatory models must therefore specify what persists, over what interval, and under which interventions. Otherwise, “self-maintenance” risks becoming a retrospective label for whatever a system happened to do.
Computational applications and empirical investigation
The most developed applications are computational models of perception, action, learning, and decision-making. Active-inference agents have been constructed for visual sampling, navigation, reaching, and partially observable choice tasks. In these models, investigators specify hidden states, observations, transition probabilities, preferences, and inference rules, then examine the resulting behavior. The value of the exercise lies in making assumptions explicit and deriving consequences that can be compared with data.
Discrete-state implementations often use a partially observable Markov decision process. An agent cannot directly observe the state of the environment and must infer it from noisy observations. For example, a cue may indicate which of two locations contains a resource, while an action changes location or reveals the cue. Such a setup can separate uncertainty about the current state from uncertainty about the consequences of actions.
Continuous models are used for sensorimotor control and robotics. They can combine predictions about body position, visual input, and environmental dynamics. Their practical performance depends on numerical stability, model mismatch, sensory delays, actuator limitations, and computational cost. Calling a controller “active inference” does not remove these engineering constraints, and a successful demonstration is not evidence that biological motor systems use exactly the same implementation.
Neuroscientific tests often concern narrower hypotheses than the entire principle. Experiments may manipulate stimulus probability, sensory reliability, learned expectations, or the usefulness of additional information. Relevant measures include behavioral choices, reaction times, neuronal activity, and brain imaging. Evidence that expectation changes a response can support a predictive-processing account, but competing explanations must be evaluated. Reduced responses to repeated stimuli, for example, can reflect adaptation as well as changes in prediction.
A stronger test distinguishes models that make different quantitative predictions. An experiment can vary uncertainty independently of expected outcome value, then compare whether information-seeking follows expected free energy, a reinforcement-learning exploration rule, or another decision model. Parameter estimation should be accompanied by parameter recovery, checks that competing models can be distinguished, and prediction of held-out observations. Otherwise, apparent explanatory flexibility can arise from adjusting priors or precision parameters to fit each result.
Clinical applications have explored whether altered learning, prior beliefs, or precision could help describe differences in perception and behavior. These are research models, not established diagnostic instruments or general clinical explanations. Similar observed behavior can arise from several computational mechanisms, and estimated parameters depend on the task and model. Clinical relevance requires independent validation rather than an inference from the framework’s mathematical generality.
Explanatory scope, criticism, and open questions
The central methodological dispute concerns what kind of claim the free energy principle makes. A variational identity is mathematically true once its terms are defined. A proposed characterization of persistent stochastic systems is conditional on assumptions about dynamics and statistical structure. A hypothesis about cortical circuitry or behavior is empirical. These levels cannot be confirmed or refuted in the same way. Confusing them can make mathematical certainty appear to support biological claims that remain untested.
One criticism is that the principle can become too permissive. If any observed behavior can be explained by choosing suitable priors, preferences, or hidden states, the framework risks redescribing behavior rather than predicting it. The response is to constrain model structure independently and evaluate specific implementations against alternatives. A framework can support testable models without itself functioning as a single experiment-ready hypothesis, but its scientific contribution then depends substantially on the quality of those models.
A second issue concerns the passage from dynamics to inference. Researchers have examined whether proposed statistical partitions, conditional independences, and internal-state mappings are sufficient for the desired variational interpretation. Solenoidal flows, dependencies among variables, and the distinction between physical trajectories and gradients over probability distributions all matter. Results obtained for restricted classes of stochastic systems should not be generalized to arbitrary organisms without checking the relevant conditions.
A third issue is representational interpretation. One reading holds that internal states literally encode beliefs about external causes. Another treats inference as a useful mathematical description that need not imply explicit internal representations. These interpretations can lead to different research priorities: identifying neural codes, characterizing embodied control loops, or deriving relations between system dynamics and statistical models. The equations alone do not settle all questions in philosophy of mind concerning meaning, agency, or experience.
Computational feasibility is another open problem. Exact Bayesian inference and long-horizon policy evaluation are generally expensive. Biological systems have finite time, noisy components, and limited resources. Approximate distributions, restricted policy sets, amortized inference, habits, and hierarchical control can reduce these demands, but each introduces assumptions requiring investigation. An elegant global objective does not uniquely identify the local processes that could implement it.
The framework also faces the problem of selecting models and boundaries. Different descriptions of the same organism can assign different hidden states, sensory interfaces, and timescales. Determining which descriptions preserve causal and predictive relevance is especially difficult for developmental systems and interacting collectives. Progress depends on connecting formal assumptions to measurable variables, specifying where approximations fail, and identifying observations that would favor one proposed implementation over another.