Pith. sign in

REVIEW 5 major objections 4 minor 40 references

Markov Blanket Density and Free Energy Minimization

T0 review · 5 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Markov blanket density governs free energy descent and makes the Free Energy Principle a local regime.

desk verdict The continuous blanket-density field is a real addition to FEP discussion, but the paper's central 'emergence' of free-energy minimization is built into its axioms, not derived. read the letter →

arxiv 2506.05794 v5 pith:RJATI33C submitted 2025-06-06 q-bio.NC cs.ITmath.IT

Luca M. Possati This is my paper
classification q-bio.NCcs.ITmath.IT MSC 62B1094A1792B20
keywords Markovblanketdensityfreeenergyprincipleactiveinferencevariationalinformationgeometrymutualepistemictrapsgradientdescent
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that the Free Energy Principle (FEP)—the idea that self-organizing systems behave as if they minimize variational free energy—is not a universal law but a local consequence of the spatial structure of information flow. The author defines a Markov blanket density $\rho(x)$ at every point in space, measuring how strongly a local boundary blocks information between internal and external states. The central assertion is that agents descend free energy only through regions where $\rho(x) < 1$, that this descent is throttled by a factor $(1 - \rho(x))$, and that free-energy minimization naturally carries the agent toward regions of lower density (stronger coupling). If correct, this reframes the FEP as an emergent regime rather than an imposed principle, with consequences for when and where inference is possible in heterogeneous or weakly bounded systems.

What carries the argument

The key mechanism is the Markov blanket density field $\rho(x)$ together with the throttled gradient descent $\dot{x} = -(1 - \rho(x))\nabla F(x)$. $\rho(x)$ is defined as 1 minus the ratio of conditional to unconditional mutual information between internal and external states given a local blanket, so it ranges from 0 (full coupling) to 1 (full insulation). The paper uses this field to build an information-theoretic geometry: a statistical manifold structure, a free-energy landscape via the log-transform of $\rho$, and an operational estimator (KSG k-nearest-neighbors) that turns sample data into local values of $\rho$. The axiomatic derivation in Section 12 makes $\rho$ the primary variable: after postulating symmetry, absence of tangential preference, and speed depending only on $\rho$, the dynamics is forced to be a gradient flow of an emergent free energy, recovering the classical FEP as a corollary.

What would settle it

Simulate Eq. (4) on a smooth domain with a generic $F$ and a generic $\rho$, and check whether trajectories ever move toward higher $\rho$ while the gradient-ratio condition holds; one such counterexample would refute Theorem 1. Decisively, any system in which free energy decreases at a point where $\rho(x) = 1$ would falsify the paper's claim that inference is informationally vacuous at full insulation.

Watch

Extended reading notes

Core claim

The paper's central object is the Markov blanket density $\rho(x) = 1 - I(I; E | B)/I(I; E)$, a scalar field on a spatial domain that measures, at each point, how strongly a local blanket screens internal from external states: 1 for perfect conditional independence (a full blanket), 0 for no effective separation. The author's claim is that the dynamics of an active-inference agent take the modulated gradient form $\dot{x} = -(1 - \rho(x))\nabla F(x)$, so that free-energy descent is throttled by the local density and halts entirely where $\rho(x) = 1$. Under axioms that force the dynamics to be symmetric, isotropic, and to stall at full screening, the paper proves (Theorem 17) that any admissible dynamics must be a gradient flow of an emergent potential $F(x) = \int^{\rho} 1/f(u)\,du + C$; consequently the Free Energy Principle is not a primitive law but a derivative structure that operates only in the regime $\rho < 1$. The paper further shows that free-energy descent aligns with movement toward lower $\rho$ only under a gradient-ratio condition, that violations create epistemic traps, and that an extended density with $\rho > 1$ inverts descent into ascent, which it reads as a formal model of pathological states.

Load-bearing premise

The load-bearing premise is that an agent's motion is the modulated gradient descent $\dot{x} = -(1 - \rho(x))\nabla F(x)$, which the paper postulates rather than derives from active inference; if the true dynamics do not have this mobility factor, the blocking, trapping, and emergence theorems do not follow.

Editorial extensions

If this is right

  • In any region where $\rho(x) = 1$, motion and belief updates stop entirely, so perfect statistical insulation acts as a wall that blocks free-energy reduction.
  • Free-energy descent moves an agent into lower-density (more coupled) regions only when the gradient-ratio condition holds; where it fails, epistemic traps form and exploration or stochasticity becomes necessary.
  • The temporal expected free energy along a trajectory is exactly the integral of $(1 - \rho(x(t)))F(x(t))$, so spatial coupling and belief updating are facets of one process.
  • If $\rho(x)$ exceeds 1, the same dynamics turn free-energy minimization into free-energy ascent, which the paper interprets as a formal picture of psychopathological inversion.
  • Operationally, estimating $\rho$ from finite samples delays convergence by an amount of order $N^{-1/(d+1)}$, so the feasibility of active inference is bounded by sample size and state-space dimension.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's empirical diagnostic for gradient alignment could be run on real behavioral or neural data: if trajectories consistently fail to enter low-$\rho$ regions even when such regions exist, the claimed coupling between free-energy descent and blanket density would be called into question.
  • Relaxing the isotropy axiom (Axiom 12) to allow tangential drift along level sets of $\rho$ would test whether the emergent-potential theorem survives less symmetric dynamics, a natural next step for anisotropic environments such as living tissues or structured workspaces.
  • The $\rho > 1$ inversion suggests an alternative reading of psychiatric symptoms as the outcome of an informational landscape that forces ascent; a concrete extension would look for measurable increases in variational free energy in patients exposed to specific high-density social contexts.
  • If the FEP is a local regime rather than a universal law, then debates about whether collectives or social systems have Markov blankets become empirical questions: one measures where $\rho < 1$ actually holds rather than assuming a blanket exists.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 4 minor

Summary. The paper introduces a scalar field ρ(x) = 1 − I(I;E|B)/I(I;E), called Markov blanket density, as a graded measure of the degree of conditional independence between internal and external states at each spatial point. It proposes that active inference agents follow the modulated gradient descent ẋ = −(1−ρ(x))∇F(x) (Eq. 4), so that free-energy minimization is throttled by local blanket strength. The paper states theorems claiming simultaneous descent of F and ρ under a gradient-ratio condition, blocking and slowing in high-ρ regions, stochastic descent under random ρ fields, a temporal expected free energy G(π)=∫(1−ρ)F dt, and an axiomatic derivation in Section 12 in which any admissible dynamics is a gradient flow of an emergent potential F(x)=∫^ρ 1/f(u)du + C. It also provides an operational KSG-based estimator, simulations, and a GitHub implementation.

Significance. If the derivation were sound, the framework would offer a spatially situated generalization of the Free Energy Principle, with a graded blanket measure, testable predictions about slowing and blocking, and an inversion regime for ρ>1. The paper's strengths include concrete algorithmic machinery for estimating ρ from data, reproducible simulations, and Section 13's candid listing of restrictive assumptions. However, the central 'emergence' claim is definitional rather than derived: Section 12's axioms force the gradient-flow form and define F as an integral of 1/f, so the result is a reparameterization of ρ rather than an independent free energy. The sign inconsistency in Theorem 1 and the stipulated Eq. (4) further undermine the core conclusions. The simulations impose Eq. (4) by construction and therefore do not test it.

major comments (5)
  1. [§12, Def. 12.2, Thm 17] The claimed emergence of the FEP is a construction, not a derivation. Axiom 12 forbids tangential motion and Axiom 14 makes the speed depend only on ρ, so Eq. (13), ẋ = −f(ρ)∇ρ, is the only admissible dynamics. Lemma 16 then defines w = (1/f(ρ))∇ρ = ∇(Ψ∘ρ), and Definition 12.2 sets F(x) = ∫^ρ 1/f(u) du + C; Theorem 17 states ẋ = −∇F. This holds by construction for any positive f, and the resulting F is a monotone function of ρ. The paper uses the same symbol F for this potential and for the variational free energy of Eq. (3) that bounds surprisal, but no connection between the two is established. The FEP is therefore not shown to emerge from the informational geometry; the axioms were chosen so that a gradient flow exists. Section 13's discussion of circularity addresses a different tautology and does not repair this definitional step.
  2. [§5, Eq. (4)] The load-bearing dynamics ẋ = −(1−ρ(x))∇F(x) is stipulated, not derived. Section 5 simply asserts the mobility factor M(x) = (1−ρ(x))I, and Theorems 1, 2, 4, 6, 7 and all simulations assume Eq. (4). Section 12's axiomatic argument does not repair this: it derives gradient flow in a potential constructed from ρ, not a given free energy F. Since blocking, slowing, epistemic traps, and the inversion regime all follow from the specific multiplicative factor (1−ρ), the paper's predictions rest on an unproven postulate about the agent's dynamics.
  3. [§6.1, Step 2; Theorem 1 conclusion] The proof of Theorem 1 contains a sign inconsistency. Step 2 computes ∇ρ_true·∇F = −[∇CMI·∇F]/MI + CMI[∇MI·∇F]/MI² and, under A(x)>0 and B(x)>0, concludes this is negative; the text then writes 'Equivalently, ∇ρ_true·∇F > 0'. These statements have opposite signs. The theorem's conclusion also states d/dt ρ_N = ∇ρ_N·∇F > 0, whereas Eq. (4) gives d/dt ρ_N = −(1−ρ_N)∇ρ_N·∇F, so positive alignment implies decreasing ρ_N. The Gradient-Ratio Condition may be able to fix the intended inequality, but as written the statement and proof are inconsistent.
  4. [§9.2, Theorem 5] The temporal expected free energy G(π) = ∫_0^τ (1−ρ(x(t)))F(x(t)) dt is presented as a theorem, but its proof only restates the definition of 'accessible free energy' (1−ρ)F and integrates it. No connection to the standard expected free energy functional of active inference is given. As a definition, Eq. (7) is acceptable; as a theorem asserting that the temporal EFE 'is exactly' this integral, it is circular.
  5. [Appendix B] The operational definition of ρ in Appendix B contradicts the main text. The main text defines ρ(x) = 1 − CMI/MI ∈ [0,1], but Appendix B defines ρ(x) := I(s_int; s_ext | s_blanket = x) and reports an estimated value of 1.44 nats for the simulated example. This value is not in [0,1] and would make the mobility factor 1−ρ negative; it also mismatches Eq. (1). If this appendix is meant to demonstrate the estimator, it must use the same definition as the main text.
minor comments (4)
  1. [§10, item 1] The 'original normalization' is misstated: Eq. (1) defines ρ = 1 − CMI/MI, not (CMI+ε)/(MI+ε). This should be corrected.
  2. [Appendix D] The final displayed equation under 'Gradient Conditions for Theorem 1' is incomplete; it ends with an unfinished expression. The appendix needs to be finished.
  3. [§1, §3.2] There are several typos, including 'bkanket' and 'extendind' in Section 1 and 'Morover' and missing spaces in 'MBdensityinitselfisnotaprobabilitydensity' in Section 3.2.
  4. [§4.1, §4.2.1] Equation (1) is numbered identically in Sections 4.1 and 4.2.1, which is confusing; renumber the second occurrence.

Circularity Check

3 steps flagged · score 8.0 of 10

Section 12's 'emergence' of the FEP is definitional: Theorem 17 defines F as a reparameterization of ρ, so Eq. (4) is in effect assumed; Theorem 5 labels a stipulated integral as temporal expected free energy.

  1. self definitional [Section 12.4, Axioms 14–15, Lemma 16, Definition 12.2, Theorem 17]
    "Combining Lemma 13, Axiom 14 and Axiom 15 we obtain the forced form of the dynamics: ẋ = v(x) = − f(ρ(x)) ∇ρ(x). Definition 12.2 (Emergent Free Energy). Fix a constant C ∈ R and define F(x) := Ψ(ρ(x)) + C = ∫_ρ(x) 1/f(u) du + C. Theorem 17 (Emergence of FEP). Under Axioms 8–15, the dynamics (13) is exactly a gradient flow of the potential F in (14): ẋ = −∇F(x)."

    Axioms 12 and 14 already force every admissible velocity to be v = −f(ρ)∇ρ. The potential F is then introduced as the integral of 1/f along ρ, so ∇F = (1/f(ρ))∇ρ and therefore v = −∇F holds identically. The 'emergence' of free-energy-minimizing gradient flow is thus a reparameterization of ρ, not a derivation from the variational free energy of Eq. (3) or from active inference. Reusing the symbol F for both the variational free energy and this constructed potential obscures that the conclusion is built into the definition.

  2. self definitional [Section 7 Eq. (5), Section 9.2 Theorem 5 and Remark]
    "Expected free energy can be redefined as a trajectory-dependent integral: G(π) = ∫_τ (1 − ρ(x_π(t))) F(x_π(t)) dt ... Theorem 5. Let π = { x(t)}^τ_{t=0} be any (piecewise-continuous) trajectory in Ω. Then the temporal expected free energy along π is G(π) = ∫_0^τ [ 1 − ρ(x(t))] F(x(t)) dt. ... Equation (7) recovers Eq. (5) verbatim and is exactly what is referred to as Theorem 5."

    Equation (5) had already stipulated that the temporal expected free energy is the integral of (1−ρ)F. Theorem 5 asserts the same object and its proof only re-integrates the already defined integrand. The dependence of the temporal EFE on the blanket density is therefore true by definition; no independent derivation, variational principle, or inference argument is supplied.

1 more flagged steps
  1. other [Section 12.5, Axiom 19 and Proposition 12.1]
    "Axiom 19 (informational alignment). There exists a continuous κ : [0,1] → [0,∞) with κ(1) = 0 and κ(ρ) > 0 for ρ < 1, such that ∇F(x) = κ(ρ(x)) ∇ρ(x). Proposition 12.1 (directional equivalence). Under Axiom 19, there exists a strictly increasing ϕ such that the gradient flow of F coincides with that of ϕ∘F; hence, after time-rescaling, the VFE-descent is dynamically equivalent to the emergent FEP of Theorem 17."

    The claimed link between the standard variational free energy and the emergent potential is not derived: Axiom 19 postulates that the VFE gradient is everywhere parallel to ∇ρ, which is exactly the alignment required to identify VFE descent with the gradient flow of Theorem 17. Proposition 12.1 then returns this postulate as a conclusion. Since no independent justification of Axiom 19 from Eq. (3) or from active inference is given, the sectional equivalence is assumption-as-conclusion rather than emergence.

full rationale

The central claim that the FEP 'emerges' from Markov blanket density is not supported by an independent derivation. In Section 12, Axioms 11–15 force the dynamics into the form ẋ = −f(ρ)∇ρ; Lemma 16 and Definition 12.2 then manufacture a potential F = ∫^(ρ) 1/f(u)du + C, for which the same dynamics is a gradient flow by construction. This is a reparameterization of ρ, not an emergence of the variational free energy of Eq. (3), and the paper reuses the symbol F for both objects. Theorem 5 is similarly definitional: Eq. (5) defines the temporal expected free energy as the integral of (1−ρ)F, and Theorem 5 proves exactly that stipulated identity. Section 12.5's bridge to standard VFE rests on Axiom 19, which assumes the desired gradient alignment. The paper candidly flags circularity in Section 13, but its rebuttal—that ρ is precomputed from external data—only addresses whether movement to low ρ is tautological; it does not repair the construction of F as a monotone function of ρ. The estimation algorithm, simulations, and conditional theorems (e.g., Theorem 1's alignment statement) are independent empirical content, but the headline 'FEP emergence' reduces to definitions and stipulated axioms. Hence a score of 8, not 10: substantial independent material exists, but the central derivation is forced by construction.

Assumptions & free parameters 3 free parameters · 7 assumptions · 2 invented entities

The central framework rests on a postulated dynamics, several strong regularity conditions, and axioms that force the gradient-flow form. The free-energy potential is then constructed from those axioms rather than derived from an independent physical or information-theoretic starting point.

free parameters (3)
  • r1, r2 (estimation radii) = r1=0.1, r2=0.2 in simulations
    Choose internal, blanket, and external partition around each point; central to the operational definition of ρ, not derived from data.
  • k, ε, δ (KSG and clipping hyperparameters) = k=5, ε=1e-6, δ=1e-3
    Set by hand in Algorithm 1; affect ρ estimates and thus the simulated dynamics.
  • speed function f(ρ) in Axiom 14 = f(ρ)=1−ρ+ε chosen later
    Axiom 14 only requires α=f(ρ); the classical free energy F=−log(1−ρ+ε) follows only for this ad hoc choice.
assumptions (7)
  • ad hoc to paper Dynamics postulate ẋ = −(1−ρ(x))∇F(x)
    Eq. (4), Section 5. Stated without derivation from active inference; all later theorems import it.
  • domain assumption Gradient alignment: A(x)>0, B(x)>0, and B/v > A/u
    Theorem 1 conditions (b) and (c); the author admits in Section 13 that this is extremely strong. Only under this condition does free-energy descent imply motion toward lower ρ.
  • domain assumption Constant covariance Cov(ρ(x), ||∇F||²) = C
    Theorem 4 assumption (iii); the paper itself calls it an artificial simplification.
  • ad hoc to paper Axioms 11-15 in Section 12: level-set symmetry, no tangential preference, level-invariant magnitude, stall at ρ=1
    These force v=−f(ρ)∇ρ and hence the gradient-flow structure; they are chosen to make the derivation work.
  • ad hoc to paper Axiom 19: informational alignment ∇F = κ(ρ)∇ρ
    Section 12.5; assumes the standard variational free energy gradient is aligned with the ρ gradient, making equivalence to the emergent F trivial.
  • domain assumption C1 smoothness and strict positivity of MI and CMI densities
    Sections 4.4 and 6.1; needed for C1 convergence of KSG estimators, but difficult to verify in high dimensions.
  • domain assumption Ergodicity of the true descent flow in Theorem 7
    Section 11.2, assumption 4; needed for parameter consistency and likely fails for generic free-energy landscapes.
invented entities (2)
  • Markov blanket density field ρ(x)
    purpose: Quantifies local informational shielding at every point and governs the throttle factor (1−ρ) in agent dynamics.
    The paper presents it as a modeling index rather than a physical quantity, and provides only synthetic simulations; no falsifiable prediction outside the model is made.
  • Extended blanket density allowing ρ>1
    purpose: Creates local free-energy ascent regions and limit cycles in Theorem 6.
    Introduced via arbitrary shift and weighted constructions; no empirical handle. It is a mathematical artifact of the extended definition.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Markov Blanket Density and Free Energy Minimization." pith.science (2026). https://pith.science/paper/RJATI33C

@misc{pith2026250605794,
  author       = {Pith},
  title        = {Pith review of: Markov Blanket Density and Free Energy Minimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RJATI33C}},
  note         = {Machine review of arXiv:2506.05794}
}
read the original abstract

This paper presents a continuous, information-theoretic extension of the Free Energy Principle through the concept of Markov blanket density, i.e., a scalar field that quantifies the degree of conditional independence between internal and external states at each point in space (ranging from 0 for full coupling to 1 for full separation). It demonstrates that active inference dynamics, including the minimization of variational and expected free energy, naturally emerge from spatial gradients in this density, making Markov blanket density a necessary foundation for the Free Energy Principle. These ideas are developed through a mathematically framework that links density gradients to precise and testable dynamics, offering a foundation for novel predictions and simulation paradigms.

Figures

Figures reproduced from arXiv: 2506.05794 by the authors.

Figure 1
Figure 1. Walking through Markov blankets. A schematic and intuitive representation of the path of an active inference agent (orange line) in a space “filled” with Markov blankets (MB) and touching points with different densities or porosities. Obviously, the agent also has its own Markov blanket, and therefore its movement is conditioned by the coupling with the other blankets and thus by the MB density. inference. These two… view at source ↗
Figure 2
Figure 2. Agent trajectories shaped by Markov blanket density. [PITH_FULL_IMAGE:figures/full_fig_p037_2.png] view at source ↗
Figure 3
Figure 3. MB density as an informational topology. [PITH_FULL_IMAGE:figures/full_fig_p037_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: When variational free energy minimization is obstructed by informational [PITH_FULL_IMAGE:figures/full_fig_p038_4.png]
Figure 5
Figure 5. Figure 5: Effect of Markov blanket density on agent movement across informational [PITH_FULL_IMAGE:figures/full_fig_p039_5.png]
Figure 6
Figure 6. Figure 6: Algorithm Agent Navigation in a Dynamic 3D Blanket-Density Field [PITH_FULL_IMAGE:figures/full_fig_p040_6.png]
Figure 7
Figure 7. Figure 7: Advanced Agent Navigation in Nonstationary 3D Environment [PITH_FULL_IMAGE:figures/full_fig_p041_7.png]
Figure 8
Figure 8. Figure 8: Inertial Agent Navigation in Ultra-Complex 3D Barriers [PITH_FULL_IMAGE:figures/full_fig_p042_8.png]
Figure 9
Figure 9. Figure 9: Simulation of Conditional Dependencies Mediated by a Markov Blanket [PITH_FULL_IMAGE:figures/full_fig_p050_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 36 canonical work pages

  1. [1]

    Friston, K., A Free energy Principle for a Particular Physics, arXiv preprint arXiv:1906.10184, 2019

    arXiv 1906

  2. [2]

    The Free-energy Principle: A Rough Guide to the Brain?

    Friston, K., “The Free-energy Principle: A Rough Guide to the Brain?”Trends in Cognitive Sciences13(7): 293–301. 2009

    work page 2009

  3. [3]

    Life as We Know It

    Friston, K., “Life as We Know It.”Journal of the Royal Society Interface10(86): 20130475. 2013

    work page 2013

  4. [4]

    Answering Schrödinger’s Question: A Free Energy Formulation

    Ramstead, M., Badcock, P., Friston, K., “Answering Schrödinger’s Question: A Free Energy Formulation.”Physics of Life Reviews(24):1–16. 2018

    work page 2018

  5. [5]

    The Markov Blankets of Life

    Kirchhoff, M., T. Parr, E. Palacios, K. Friston, Kiverstein, J., “The Markov Blankets of Life.”Journal of the Royal Society Interface15:20170792. 2018

    work page 2018

  6. [6]

    OnBayesianMechanics

    Ramstead, M., Sakthivadivel, D., Heins, C., Koudahl, M., Millidge, B., DaCosta, L., Klein, B., Friston, K., “OnBayesianMechanics.”Interface Focus13:20220029. 2023

    work page 2023

  7. [7]

    Dynamic Markov Blanket Detection for Macroscopic Physics Discovery

    Beck, J., Ramstead, M. "Dynamic Markov Blanket Detection for Macroscopic Physics Discovery"arXiv:2502.21217 [q-bio.NC] 2025

    arXiv 2025

  8. [8]

    Weak Markov Blankets in High-Dimensional, Sparsely- Coupled Random Dynamical Systems

    Sakthivadivel, D. “Weak Markov Blankets in High-Dimensional, Sparsely- Coupled Random Dynamical Systems” arXiv:2207.07620 2025

    arXiv 2025

Show all 40 references
  1. [9]

    The Free Energy Principle in Mind, Brain, and Behavior, MIT Press

    Parr, T., Pezzulo, G., Friston, K.Active Inference. The Free Energy Principle in Mind, Brain, and Behavior, MIT Press. 2022

    2022

  2. [10]

    Amari, S.Information Geometry and Its Applications, Springer. 2016

    2016

  3. [11]

    M., Thomas, J

    Cover, T. M., Thomas, J. A.Elements of Information Theory (2nd ed.), Wiley. 2006

    2006

  4. [12]

    H., Rishel, R

    Fleming, W. H., Rishel, R. W.Deterministic and Stochastic Optimal Control, Springer. 1975

    1975

  5. [13]

    P.Nonlinear Programming (2nd ed.), Athena

    Bertsekas, D. P.Nonlinear Programming (2nd ed.), Athena. 1999

    1999

  6. [14]

    Information theory and statistical mechanics

    Jaynes, E. T. "Information theory and statistical mechanics"Physical Review, 106(4), 620

  7. [15]

    Informationgeometryonhierarchyofprobabilitydistributions

    Amari, S."Informationgeometryonhierarchyofprobabilitydistributions" IEEE Transactions on Information Theory, 47(5), 1701–1711. 2001

    2001

  8. [16]

    Pearl, J.Causality: Models, Reasoning and Inference, Cambridge. 2009

    2009

  9. [17]

    Entropy production fluctuation theorem and the nonequilibrium work relation for free energy differences

    Crroks, G. "Entropy production fluctuation theorem and the nonequilibrium work relation for free energy differences"Physical Review E, 60(3), 2721. 1999

    1999

  10. [18]

    Natural gradient works efficiently in learning

    Amari, S. "Natural gradient works efficiently in learning"Neural Computation, 10(2), 251–276. 1998

    1998

  11. [19]

    Natural gradient via optimal transport

    Li, W., Montúfar, G. "Natural gradient via optimal transport"Information Ge- ometry, 1(2), 181–214. 2018

    2018

  12. [20]

    Generalised free energy and active inference

    Parr, T., Friston, K. "Generalised free energy and active inference."Biological Cybernetics, 113(5), 495–513. 2019

    2019

  13. [21]

    Estimating mutual information

    Kraskov, A., Stögbauer, H., Grassberger, P. "Estimating mutual information." Physical Review, 69(6), 066138. 2004

    2004

  14. [22]

    A Step-by-Step Tutorial on Active Inference and Its Application to Empirical Data

    Smith, R., Friston, K., Whyte, C. “A Step-by-Step Tutorial on Active Inference and Its Application to Empirical Data.”Journal of Mathematical Psychology, 107: 102632. 2022. 57

    2022

  15. [23]

    Estimation of entropy and mutual information

    Paninski, L. "Estimation of entropy and mutual information."Neural Computa- tion, 15(6), 1191–1253. 2003

    2003

  16. [24]

    Demystifying information-theoretic estima- tors

    Gao, W., Oh, S., Viswanath, P. "Demystifying information-theoretic estima- tors." arXiv:1708.00065. 2017

    2017 arXiv

  17. [25]

    JIDT: An information-theoretic toolkit for studying the dynamics of complex systems

    Lizier, J. T. "JIDT: An information-theoretic toolkit for studying the dynamics of complex systems."Frontiers in Robotics and AI, 1, 11. 2014

    2014

  18. [26]

    Bayesian mechanics for stationary processes

    Da Costa, L., Friston, K., Heins, C., Pavliotis, G.A. "Bayesian mechanics for stationary processes."Proceedings Royal Society, A 477: 20210518. 2021

    2021

  19. [27]

    Markov blankets, information geometry and stochastic thermodynamics

    Parr, T., Da Costa, L., Friston, K. "Markov blankets, information geometry and stochastic thermodynamics." Philosophical Transactions Royal Society A, 378:20190159. 2019

    2019

  20. [28]

    Thinking through Other Minds

    Veissière S., Constant A., Ramstead M., Friston K., Kirmayer L. “Thinking through Other Minds.”Behavioral and Brain Sciences, (43): 1–75. 2020

    2020

  21. [29]

    J.The ecological approach to visual perception(1st ed)

    Gibson, J. J.The ecological approach to visual perception(1st ed). Mifflin and Company, 1979

    1979

  22. [30]

    The Depressed Brain: An Evolutionary Systems Theory

    Badcock, P., Davey, C., Whittle, S., Allen, N., Friston, K. "The Depressed Brain: An Evolutionary Systems Theory."Trends in Cognitive Science, (21): 182-194. 2017

    2017

  23. [31]

    Active Infer- ence: A Process Theory

    Friston K, FitzGerald T, Rigoli F, Schwartenbeck P, Pezzulo G. "Active Infer- ence: A Process Theory."Neural Computations, 29(1):1-49. 2017

    2017

  24. [32]

    Partial mutual information for coupling analysis of mul- tivariate time series

    Frenzel, S., Pompe, B. “Partial mutual information for coupling analysis of mul- tivariate time series.”Physical Review Letters, 99:204101. 2007

    2007

  25. [33]

    Estimating mutual information for discrete–continuous mixtures

    Gao, W., Kannan, S., Oh, S., Viswanath, P. “Estimating mutual information for discrete–continuous mixtures.” Advances in Neural Information Processing Systems, 30:—. 2017

    2017

  26. [34]

    TSP—Infrastructure for the traveling salesperson prob- lem

    Hahsler, M., Hornik, K. “TSP—Infrastructure for the traveling salesperson prob- lem.” Journal of Statistical Software, 23:1–21. 2007

    2007

  27. [35]

    Estimation of the information by an adaptive par- titioning of the observation space

    Darbellay, G. A., Vajda, I. “Estimation of the information by an adaptive par- titioning of the observation space.”IEEE Transactions on Information Theory, 45:1315–1321. 1999

    1999

  28. [36]

    Exponential concentration of a den- sity functional estimator

    Singh, S., Poczos, B., Wasserman, L. “Exponential concentration of a den- sity functional estimator.”Neural Information Processing Letters and Reviews, 9:123–132. 2016

    2016

  29. [37]

    Clark, A.Surfing Uncertainty: Prediction, Action, and the Embodied Mind, Ox- ford University Press. 2016

    2016

  30. [38]

    The Emperor’s New Markov Blankets

    Bruineberg J, Dołęga K, Dewhurst J, Baltieri M. "The Emperor’s New Markov Blankets." Behavioral and Brain Sciences", 45:e183. 2022

    2022

  31. [39]

    Nothing but a useful tool? (F)utility and the free energy princi- ple

    Colombo, M. "Nothing but a useful tool? (F)utility and the free energy princi- ple." In Bruineberg J, Dołęga K, Dewhurst J, Baltieri M. "The Emperor’s New Markov Blankets."Behavioral and Brain Sciences", 45:e183. 2022

    2022

  32. [40]

    Path Integrals, Particular Kinds, and Strange Things

    Friston, K, Da Costa, L., Sakthivadivel, D., Heins, C., Pavliotis, G., Ramstead, M., Parr, T. "Path Integrals, Particular Kinds, and Strange Things."Physics of Life Reviews", 47, 2023. 58

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.