Corso, Myriam
(2026)
The Free Energy Principle and Nonlinear Filtering.
[Laurea magistrale], Università di Bologna, Corso di Studio in
Matematica [LM-DM270], Documento full-text non disponibile
Il full-text non è disponibile per scelta dell'autore.
(
Contatta l'autore)
Abstract
The free energy principle proposes that a living system, in contact with its environment through a limited number of channels, holds a belief about what causes its sensory readings, and that perception, action and learning all lower one functional of that belief. This thesis develops the principle inside a stochastic framework in which the spaces, the assumptions and the regime are fixed in advance, and checks each of its laws within the limits of that framework.
The first part develops nonlinear filtering for diffusion processes: the filtering problem, the innovation process, the Zakai and Kushner-Stratonovich equations, the three finite representations that make the recursion computable, and variational inference under the Laplace approximation. Both methods target the conditional law of the hidden state given the observations: exact filtering is in general infinite-dimensional, while variational inference is finite but approximate.
The second part states the principle as a single diffusion carrying an agent and its environment, whose four groups of states are separated by conditions on the coefficients. The principle is assumed rather than derived, and the assumptions give the system the structure that allows perception, action and learning. Perception is written twice, as a gradient descent on a variational free energy and as the exact filter of the first part applied to a state-space model of the agent, so that the two formulations can be compared term by term; action and learning follow only the variational route.
The thesis closes with the Bayesian thermostat: by reproducing the existing simulations, it computes and analyzes every quantity of the model, showing where perception comes to rest and what limits the action law. This law requires a quantity that the framework does not allow the agent to carry, and the thesis proposes a way to supply it without leaving the framework.
Abstract
The free energy principle proposes that a living system, in contact with its environment through a limited number of channels, holds a belief about what causes its sensory readings, and that perception, action and learning all lower one functional of that belief. This thesis develops the principle inside a stochastic framework in which the spaces, the assumptions and the regime are fixed in advance, and checks each of its laws within the limits of that framework.
The first part develops nonlinear filtering for diffusion processes: the filtering problem, the innovation process, the Zakai and Kushner-Stratonovich equations, the three finite representations that make the recursion computable, and variational inference under the Laplace approximation. Both methods target the conditional law of the hidden state given the observations: exact filtering is in general infinite-dimensional, while variational inference is finite but approximate.
The second part states the principle as a single diffusion carrying an agent and its environment, whose four groups of states are separated by conditions on the coefficients. The principle is assumed rather than derived, and the assumptions give the system the structure that allows perception, action and learning. Perception is written twice, as a gradient descent on a variational free energy and as the exact filter of the first part applied to a state-space model of the agent, so that the two formulations can be compared term by term; action and learning follow only the variational route.
The thesis closes with the Bayesian thermostat: by reproducing the existing simulations, it computes and analyzes every quantity of the model, showing where perception comes to rest and what limits the action law. This law requires a quantity that the framework does not allow the agent to carry, and the thesis proposes a way to supply it without leaving the framework.
Tipologia del documento
Tesi di laurea
(Laurea magistrale)
Autore della tesi
Corso, Myriam
Relatore della tesi
Correlatore della tesi
Scuola
Corso di studio
Indirizzo
CURRICULUM ADVANCED MATHEMATICS FOR APPLICATIONS
Ordinamento Cds
DM270
Parole chiave
Free energy principle,Active inference,Nonlinear filtering,Diffusion processes,Zakai and Kushner-Stratonovich equations,Variational inference,Markov blanket,Bayesian thermostat
Data di discussione della Tesi
25 Settembre 2026
URI
Altri metadati
Tipologia del documento
Tesi di laurea
(NON SPECIFICATO)
Autore della tesi
Corso, Myriam
Relatore della tesi
Correlatore della tesi
Scuola
Corso di studio
Indirizzo
CURRICULUM ADVANCED MATHEMATICS FOR APPLICATIONS
Ordinamento Cds
DM270
Parole chiave
Free energy principle,Active inference,Nonlinear filtering,Diffusion processes,Zakai and Kushner-Stratonovich equations,Variational inference,Markov blanket,Bayesian thermostat
Data di discussione della Tesi
25 Settembre 2026
URI
Gestione del documento: