Contribution
A scope argument, distinct from the unfalsifiability and map–territory critiques. It accepts the FEP’s derivation, then asks what the derivation needs: a NESS and a blanket. It argues that the blanket, not the dynamics, does the work, that it is chosen rather than found, and that it is chosen because it is what makes Bayesian inference applicable. It adds a separate argument against active inference as a process theory of perception and action.
Key insight
Variational inference needs three kinds of variable: hidden, observed, and the system’s own. A Markov blanket partition delivers exactly those three for any system you apply it to. So applying it turns any system into a Bayesian inferrer by construction; that is a modelling move, not a discovery about the system.
Assumptions
- The FEP as currently formulated (Friston 2019; Friston et al. 2020): ẋ = f(x) + ω (Eq. 1); Fokker–Planck with NESS solution f = (Q − Γ)·∇ℑ, ℑ = −ln p (Eqs. 2–3); blanket partition (Eqs. 4–5); flows of autonomous states as gradients on particular surprisal (Eqs. 6–7); internal states parametrising q_μ(η) (Eqs. 8–11); F = D_KL + ℑ(π) ≥ ℑ(π) (Eqs. 12–17); flows on F (Eq. 18).
- Empirical evidence for the FEP is out of scope and called “at best, scarce” (§1).
- No “Hegelian argument”: the authors do not argue against the FEP as a research programme.
Key results
- FEP holds given its commitments (§2, p. 17). Three commitments: a NESS; a blanket making internal states conditionally independent of external ones; and, combining them, surprisal minimisation readable as variational inference under a generative model “provided by the NESS of the thing”.
- §3.1 Every thing. Candle flames have no stable states to be blanket states (Friston 2013, 2019), though they are things; blankets whose states “wander away” (Friston 2019, p. 50) are not handled; so clouds and groups are doubtful. The FEP is “not a principle for a theory of everything”.
- §3.2 Placement. Brain (receptors as blanket), bacillus (membrane and cytoskeleton), coupled pendulums (beam as blanket; velocity sensory, position active, after Kirchhoff et al. 2019): “the Markov blanket partition is stipulated as being located exactly wherever happens to be convenient”. Formally the system x includes η, so the “thing” is organism-plus-environment; nested blankets do not help the pendulums, already explained by relative phase.
- §3.3 Outside the blanket. Relational properties: taller-than, heading (Fajen & Warren), affordances (step height to leg length, Warren 1984); redefining affordances as action-selection preferences is “a good demonstration of FEP’s inability to capture relational properties”. Constitutive self-organization: membranes are produced by cells; receptors are not produced by brains, nor beams by pendulums; footnote 17 on Friston 2013’s particles.
- §3.4 Fundamentality. Deriving physics from the FEP is matched by Jaynes’s MaxEnt and Wolfram’s hypergraphs, which share the lack of empirical yield.
- §3.5 The trick. Why blankets and not, say, causal blankets (Rosas et al. 2020)? Because blankets give the variational-inference structure. Historically F came first (Helmholtz machine, Hinton & van Camp, Dayan & Abbott 2001), the FEP after; the real logic runs from “model it as an autoencoder” to “therefore FEP holds”.
- §4 Active inference. With a T-maze generative model (Eqs. 19–20, after Friston et al. 2017 and Hesp et al. 2021), the A, B, C matrices, and even their dimensions, encode the experimenters’ rationalisation of the task; the cue’s perception and the movements’ control are assumed. Deriving priors from the NESS makes “hidden” states not hidden in x.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Given its commitments the FEP is true | stated | the authors’ own summary of the derivation (§2) |
| C2 | Markov blankets apply only to some things (not candle flames, doubtfully to clouds or groups) | moderate | FEP literature’s own exclusions; argument §3.1 |
| C3 | Blanket placement follows modelling convenience, not the system | moderate | three examples compared (§3.2) |
| C4 | The blanket formalism cannot represent relational properties or constitutive self-organization | moderate | examples and argument §3.3; footnote 17 on Friston 2013 |
| C5 | Blankets are adopted because they supply variational inference’s partition | moderate as an interpretation | historical priority of F; structural parallel (§3.5) |
| C6 | Active inference presupposes perception and action rather than explaining them | moderate | analysis of one generative-model example (§4.2) |
Concepts
- Thing. In the FEP, a random dynamical system with a NESS and a Markov blanket; the authors italicise it to keep it apart from things in general.
- Particular states. Internal and blanket states; autonomous states are internal and active.
- Markov blanket trick. Using the blanket partition to make any system modellable as variational inference, by analogy with the reparameterization trick.
- Constitutive self-organization. A system producing the boundary that distinguishes it from its environment.
Connections
Builds on Biehl, Pollock & Kanai (2021), Andrews (2020), van Es (2020), Bruineberg et al. (2020; LIT-603), Chemero’s radical embodied cognitive science and ecological psychology (Gibson; Warren). Its foil includes Kirchhoff et al. 2018, Ramstead et al. 2018–2020, Friston 2013 (LIT-526) and 2019. It uses HKB (LIT-588; read through LIT-601) as the explanation that makes a blanket idle for pendulums.
Bearing on the record
- LIT-526. C4 via footnote 17 is the basis of the LIT’s narrow
corrects(the autopoiesis claim). C2 accepts LIT-526’s flame claim and uses it against the FEP’s scope. - Organisational closure (LIT-582; LIT-570; LIT-536). The constitutive self-organization the paper says blankets miss is the organizational school’s subject; the paper cites autopoiesis rather than closure of constraints.
- THEORY filed as THEORY-073 (shared with LIT-603): a Markov-blanket partition presupposes rather than discovers a system’s boundary.
- LIT-577 (Sophisticated Inference) is open to C6 as stated.
- No instruction for machine-learning practice.
Limitations
- Conceptual throughout; no model is built to show a relational property or a self-produced boundary escaping a blanket description.
- The FEP’s derivation is summarised from secondary formal sources; the formal problems of Biehl et al. are reported, not used.
- The active-inference critique rests on one family of discrete-state examples.
Open questions
- Can a blanket formalism with time-varying blanket states (Ramstead et al. 2020a, noted in footnote 14) represent a self-produced membrane?
- Would causal blankets (Rosas et al.) or closure of constraints serve the same modelling role without the inferential reading?