# Economic Applications of the Free Energy Principle ## Overview The Free Energy Principle offers a principled foundation for understanding economic behavior that goes beyond the classical rational agent model. Where neoclassical economics assumes agents maximize expected utility with perfect information, the FEP characterizes agents as **minimizing expected free energy** under a generative model with **bounded capacity** and **irreducible uncertainty**. This provides: - A formal account of **bounded rationality** (Simon) grounded in information theory - A natural explanation for **risk sensitivity** and **loss aversion** without ad hoc utility functions - A framework for understanding **market dynamics** as collective inference - Connections between **game theory** and multi-agent active inference - A mechanistic basis for **behavioral economics** phenomena - A principled approach to **institutional design** as collective free energy minimization The FEP does not replace economic theory but provides deeper foundations, explaining WHY agents exhibit the patterns that economists observe. ## Theoretical Framework ### Expected Utility vs. Expected Free Energy The central comparison between standard economics and FEP-based economics: ``` Expected Utility Theory (EUT): U(pi) = sum_o p(o|pi) * u(o) pi* = argmax_pi U(pi) Where: pi = policy (sequence of actions) u(o) = utility function over outcomes p(o|pi) = probability of outcome given policy Expected Free Energy (EFE): G(pi) = E_q(o,s|pi) [ln q(s|pi) - ln p(o, s)] pi* = argmin_pi G(pi) Decomposition: G(pi) = -D_KL[q(s|o,pi) || q(s|pi)] [epistemic value] + E_q[ln q(o|pi) - ln p(o)] [pragmatic value] ``` Key differences: | Feature | Expected Utility | Expected Free Energy | |---------|-----------------|---------------------| | Information | Assumes known probabilities | Maintains uncertainty, values information | | Exploration | Not addressed (separate problem) | Built in via epistemic value | | Risk | Encoded in utility curvature | Emerges from model uncertainty | | Complexity | No cost for complex beliefs | Complexity penalty: D_KL[q \|\| p] | | Bounded | Unbounded optimization | Bounded by generative model capacity | | Learning | Static or separate | Integrated belief updating | ### Bounded Rationality as Free Energy Minimization Herbert Simon's bounded rationality -- the idea that agents optimize subject to cognitive constraints -- is naturally formalized by the FEP: ``` Bounded rationality: Agent has a generative model m with limited capacity Inference is approximate: q(s) != p(s|o) in general Planning horizon is finite: policies extend T steps Policy space is limited: only K policies considered The agent is "rational" in the sense of minimizing F but "bounded" in the sense that F is computed under m Free energy decomposition reveals the boundedness: F = Accuracy - Complexity = E_q[ln p(o|s)] - D_KL[q(s) || p(s)] Complexity cost D_KL[q(s) || p(s)] penalizes deviation from prior -> More complex beliefs are "costly" in information-theoretic terms -> This is the formal basis for cognitive bounds ``` ### The Complexity Cost of Deliberation ``` Rate-distortion theory connection: Optimal belief updating trades off accuracy against complexity: q* = argmin_q {E_q[-ln p(o|s)] + (1/beta) * D_KL[q(s) || p(s)]} Where beta = "inverse temperature" (computational budget) beta -> infinity: Perfect Bayesian inference (unlimited computation) beta -> 0: Prior beliefs unchanged (no computation) Economic interpretation: beta encodes the agent's "computational budget" Higher beta = more deliberation = better decisions = higher cost Optimal beta balances decision quality against deliberation cost ``` This formally connects to: - **Satisficing** (Simon): Agents stop deliberating when F is "good enough" - **Rational inattention** (Sims): Agents pay an information cost for precise beliefs - **Dual-process theory** (Kahneman): System 1 (low beta, fast) vs. System 2 (high beta, slow) ## Bayesian Agents in Markets ### Individual Decision Making An economic agent under the FEP maintains a generative model of the economic environment: ``` Generative model of economic agent: Hidden states s: Market conditions, other agents' strategies, fundamentals Observations o: Prices, quantities, signals, news Actions a: Buy, sell, hold, invest, consume Prior preferences C: Desired wealth trajectory, consumption pattern Agent performs: 1. Perceptual inference: q(s) = beliefs about market state 2. Policy selection: pi* = argmin_pi G(pi) 3. Learning: Update generative model parameters over time ``` ### Asset Pricing Under the FEP Traditional asset pricing assumes risk-neutral or risk-averse utility maximizers. The FEP provides a richer foundation: ``` Asset price under FEP: P(asset) = E_q[sum_t discount^t * D_t] - risk_adjustment Where the risk adjustment comes from model uncertainty: risk_adjustment = f(H[q(returns)], D_KL[q || p]) H[q(returns)] = entropy of beliefs about future returns D_KL[q || p] = divergence between posterior and prior beliefs Key predictions: 1. Assets with more uncertain returns have lower prices (equity premium) 2. Price adjustments are slower when model uncertainty is high 3. Prices overshoot when precision is misallocated (bubbles) 4. Prices underreact when beliefs are too complex to update (stickiness) ``` ### Information and Market Efficiency ``` Efficient Market Hypothesis (EMH) under FEP: Strong EMH: All agents have the same generative model -> prices reflect all info This is impossible under FEP because: 1. Agents have different generative models (different priors, different capacity) 2. Belief updating is costly (complexity penalty) 3. Information has value ONLY because agents are uncertain Market efficiency is a LIMIT, not a description: As agents minimize free energy collectively, prices converge toward informationally efficient levels BUT the convergence is never complete because: - New information constantly arrives (non-stationarity) - Different models process information differently (heterogeneity) - Updating is costly (bounded rationality) ``` ## Risk Sensitivity and Loss Aversion ### Risk as Model Uncertainty Under the FEP, risk is not a property of outcomes but a property of the agent's **model uncertainty**: ``` Risk under EUT: Risk aversion encoded in concave utility: u''(x) < 0 -> Ad hoc: Why is utility concave? "People just are risk-averse." Risk under FEP: Risk emerges from uncertainty in the generative model G(pi) includes terms for: - Outcome uncertainty: H[q(o|pi)] (how unpredictable are the outcomes?) - Model uncertainty: H[q(m)] (how confident is the agent in its model?) "Risk-averse" behavior = high precision on prior preferences "Risk-seeking" behavior = high epistemic value of uncertain options The same agent can be risk-averse or risk-seeking depending on context: - Risk-averse when prior preferences are strong (protecting gains) - Risk-seeking when epistemic value is high (exploring losses) ``` ### Loss Aversion as Asymmetric Precision Kahneman and Tversky's prospect theory finding that "losses loom larger than gains" has a natural FEP account: ``` Loss aversion under FEP: Prior preference: C = ln p(o) centered on current state (reference point) Losses: o < reference -> surprisal increases sharply -> p(o) drops steeply below reference -> high free energy Gains: o > reference -> surprisal decreases gradually -> p(o) increases gradually above reference -> moderate free energy reduction Asymmetry arises from: 1. Precision is higher for deviations below reference (survival constraint) 2. The generative model predicts maintenance of current state 3. Losses represent prediction errors with high precision Loss aversion ratio (approx 2:1) reflects the precision asymmetry in the generative model's prior preferences. ``` ### Framing Effects ``` Framing effects under FEP: Different frames activate different generative models -> Different priors, different precisions, different inference "90% survival rate" vs. "10% mortality rate": Frame 1 activates: p(o | survival_model) -> emphasize survival Frame 2 activates: p(o | mortality_model) -> emphasize death Same information, different generative model -> different free energy -> Different decisions This is not "irrational" under FEP: It reflects the fact that different models are activated by different frames, and each model leads to locally optimal inference. ``` ## Game Theory and Multi-Agent Active Inference ### Games as Coupled Inference In multi-agent active inference, game-theoretic interactions arise naturally: ``` Two-agent game: Agent 1: Generative model m_1, beliefs q_1(s_1, s_2) Agent 2: Generative model m_2, beliefs q_2(s_1, s_2) Each agent's hidden states include the other agent's states -> Each agent infers what the other agent believes and will do -> This IS theory of mind / strategic reasoning Nash equilibrium emerges when: Both agents have converged to consistent beliefs: q_1(s_2) is consistent with agent 2's actual policy q_2(s_1) is consistent with agent 1's actual policy -> Neither agent can reduce free energy by changing policy ``` ### Cooperation and Defection ``` Prisoner's Dilemma under active inference: Classical analysis: Defection is dominant strategy Active inference analysis: Agent's generative model includes: - Beliefs about other agent's policy: q(pi_other) - Beliefs about future interactions: q(repeated?) - Prior preferences: C includes social outcomes Cooperation emerges when: 1. Agent models repeated interaction (shadow of the future) 2. Agent has social prior preferences (concern for other's outcomes) 3. Agent has accurate model of other's reciprocity 4. Epistemic value of cooperation is high (learning about the other) Defection emerges when: 1. Agent models one-shot interaction 2. Agent has purely self-interested prior preferences 3. Agent has high uncertainty about other's strategy 4. Pragmatic value of defection dominates epistemic value ``` ### Market Games ``` Market as multi-agent active inference: N agents, each minimizing their expected free energy Market price emerges from collective inference Price formation: P(t+1) = f(sum_i a_i(t)) Where a_i(t) = agent i's action (buy/sell) at time t Each agent: pi_i* = argmin_pi G_i(pi) Market dynamics: - Agents with similar generative models -> correlated behavior -> herding - Agents with different models -> diverse behavior -> market efficiency - Agents with high precision -> strong price impact -> large traders - Agents with low precision -> weak price impact -> noise traders ``` ## Organizational Behavior ### Firms as Markov Blankets The FEP provides a formal account of organizational boundaries: ``` Firm as a Markov blanket: Internal states mu: Employees, processes, intellectual property Sensory states s: Market signals, customer feedback, competitor actions Active states a: Products, services, marketing, pricing External states psi: Market environment, competitors, regulators The firm maintains itself by minimizing free energy: - Perception: Market research, competitive intelligence - Action: Product development, pricing, marketing - Learning: Process improvement, innovation, training - Model selection: Strategic pivots, reorganization ``` ### Institutional Design ``` Institutions as collective generative models: An institution provides: 1. Shared priors: Common beliefs about how the world works 2. Shared precision: Agreed-upon confidence in these beliefs 3. Shared policies: Coordinated action strategies 4. Communication channels: Mechanisms for sharing prediction errors Well-designed institutions minimize collective free energy: F_collective = sum_i F_i + F_coordination Where F_coordination = cost of maintaining shared generative model Trade-off: - Centralization: Low F_coordination, but model may be wrong for individuals - Decentralization: High F_coordination, but models fit individual contexts -> Optimal institutional structure depends on environmental complexity ``` ### Preference Learning and Revealed Preferences ``` Classical revealed preferences: Observe choices -> infer utility function Assumes stable, consistent preferences FEP-based preference learning: Observe choices -> infer generative model (including prior preferences) Allows for: - Context-dependent preferences (different models for different contexts) - Learning and updating preferences (Bayesian preference learning) - Precision dynamics (preferences can be strong or weak) - Epistemic vs. pragmatic choices (exploration vs. exploitation) Formal: p(C | choices, contexts) proportional_to p(choices | C, contexts) * p(C) -> Bayesian inference over prior preferences -> Richer than utility function estimation ``` ## Market Dynamics as Collective Inference ### Bubbles and Crashes ``` Bubble formation under FEP: 1. New asset class -> high model uncertainty 2. Early gains -> update toward optimistic model 3. Social contagion: Others' buying = evidence for bullish model -> Shared generative model develops (collective precision) 4. Precision on bullish model increases (confirmation bias) 5. Sensory precision (for negative signals) decreases -> Warning signs ignored 6. Price diverges from fundamentals (free energy accumulates) Crash: 1. Prediction error accumulates (price vs. fundamentals) 2. Some agents' models update (precision on negative signals rises) 3. Selling begins -> prediction error for bullish agents 4. Rapid model updating cascade (precision collapse) 5. Herding into pessimistic model (panic) 6. Price overshoots downward (symmetric to upward bubble) Mathematical: Bubble persistence time proportional_to: Pi_bullish / rate_of_contradicting_evidence Crash speed proportional_to: rate_of_precision_collapse * number_of_agents_updating ``` ### Information Cascades ``` Information cascade under FEP: Agent n observes: - Private signal s_n (personal information) - Actions of agents 1, ..., n-1 (social information) If social information has higher precision than private signal: Pi_social > Pi_private -> Agent follows the crowd regardless of private signal -> Cascade: All subsequent agents follow the crowd -> Fragile: A sufficiently strong private signal breaks the cascade FEP contribution: Precision weighting provides a QUANTITATIVE account: Cascade forms when Pi_social / Pi_private > threshold Cascade breaks when Pi_private > Pi_social for some agent ``` ### Market as Inference Engine ``` The market as a collective generative model: Prices = sufficient statistics of collective beliefs Trading = message passing between agents Liquidity = precision of the collective inference Volatility = uncertainty in the collective model Market "learns" through: 1. New information arrives (sensory input) 2. Agents update beliefs (perceptual inference) 3. Agents trade (active inference) 4. Prices adjust (collective belief update) 5. Market structure evolves (model selection) Market efficiency = how quickly the collective model updates Market failure = persistent divergence from optimal model ``` ## Current Research ### Computational Agent-Based Models ``` Building agent-based economic models with active inference agents: - Each agent is a full active inference system - Heterogeneous generative models (realistic diversity) - Emergent market dynamics from agent interactions - Can replicate: bubbles, crashes, herding, volatility clustering Advantage over classical ABMs: Agent behavior is principled (derived from FEP) rather than ad hoc (arbitrary behavioral rules) ``` ### Neuroeconomics Integration ``` Connecting brain-level FEP models to economic behavior: - fMRI studies of decision-making fit with active inference models - Dopamine as precision over economic policies - Ventral striatum encodes pragmatic value - Prefrontal cortex encodes epistemic value - Anterior insula encodes interoceptive prediction errors (gut feelings) Goal: Bridging neural mechanisms -> computational models -> economic behavior ``` ### Climate Economics and Long-Horizon Planning ``` Active inference for climate economics: - Deep temporal generative models for long-horizon planning - Epistemic value of climate research - Prior preferences encoding intergenerational welfare - Precision dynamics for uncertain climate projections Advantage: FEP naturally handles deep uncertainty (unlike expected utility which requires specifying probabilities) ``` ## Open Questions 1. **Aggregation**: How do individual-level FEP models aggregate to macroeconomic phenomena? The micro-macro bridge remains challenging. 2. **Computational tractability**: Can active inference models scale to realistic economic complexity with many agents and high-dimensional state spaces? 3. **Empirical validation**: Can FEP-based economic models make better predictions than standard models? Early results are promising but limited. 4. **Welfare economics**: How do we define social welfare when agents have different generative models? Classical welfare theorems assume common utility scales. 5. **Policy design**: Can governments use FEP insights to design better policies? Understanding how citizens minimize free energy could improve policy communication and implementation. 6. **Money and finance**: What role does money play in a free energy minimizing economy? Preliminary work suggests money reduces the complexity cost of multi-good exchange. ## References 1. Friston, K. J., Daunizeau, J., & Kiebel, S. J. (2009). Reinforcement learning or active inference? *PLoS One*, 4(7), e6421. 2. Ramstead, M. J. D., Constant, A., Badcock, P. B., & Friston, K. J. (2019). Variational ecology and the physics of sentient systems. *Physics of Life Reviews*, 31, 188-205. 3. Constant, A., Ramstead, M. J. D., Veissiere, S. P. L., & Friston, K. (2019). Regimes of expectations: an active inference model of social conformity and human decision making. *Frontiers in Psychology*, 10, 679. 4. Sims, C. A. (2003). Implications of rational inattention. *Journal of Monetary Economics*, 50(3), 665-690. 5. Kahneman, D. (2011). *Thinking, Fast and Slow*. Farrar, Straus and Giroux. 6. Simon, H. A. (1955). A behavioral model of rational choice. *Quarterly Journal of Economics*, 69(1), 99-118. 7. Tversky, A., & Kahneman, D. (1992). Advances in prospect theory: cumulative representation of uncertainty. *Journal of Risk and Uncertainty*, 5(4), 297-323. 8. Friston, K. J., & Ao, P. (2012). Free energy, value, and attractors. *Computational and Mathematical Methods in Medicine*, 2012, 937860. 9. Parr, T., Da Costa, L., & Friston, K. (2020). Markov blankets, information geometry and stochastic thermodynamics. *Philosophical Transactions of the Royal Society A*, 378(2164), 20190159. 10. Ortega, P. A., & Braun, D. A. (2013). Thermodynamics as a theory of decision-making with information-processing costs. *Proceedings of the Royal Society A*, 469(2153), 20120683. ## See Also - [[knowledge_base/free_energy_principle/mathematics/expected_free_energy|Expected Free Energy]] - [[knowledge_base/free_energy_principle/cognitive/decision_making|Decision Making]] - [[social_sciences|Social Science Applications]] - [[ai_safety|AI Safety]] - [[knowledge_base/free_energy_principle/philosophy/epistemology|Epistemology]] - [[knowledge_base/free_energy_principle/systems/complex_adaptation|Complex Adaptation]]