Contribution
Critics had said the free energy principle is unfalsifiable or merely redescriptive. This paper isolates one construct, the Markov blanket, and traces how it changed meaning, from a device for simplifying mean-field variational inference to a claimed physical boundary of organisms and minds. Naming the two uses (Pearl and Friston blankets) and the two research programmes (inference with and within a model) gives the debate a vocabulary in which the critique can be stated precisely.
Key insight
A Markov blanket is a property of a graph, and a graph is a model. You can only find a blanket after choosing a graph and choosing which node you care about. So a blanket cannot tell you where a system ends unless you have already decided where it ends, or you add a metaphysics on which the graph is the world.
Assumptions
- Standard variational inference (§2): ELBO / variational free energy, mean-field factorisation; Pearl blankets reduce each factor’s update to expectations over its blanket (Eq. 20).
- Bayesian networks are directed acyclic graphs (§3.1); cyclic sensorimotor loops need unrolling in time (footnotes 8, 13).
- Scientific models are maps (§6.1), chosen for purposes and selected by parsimony (e.g. by free energy itself), so their structure is partly a modelling choice.
Key results
- Pearl blankets (§3.2). mb(x) = pa(x) ∪ ch(x) ∪ copa(x); worked on an 18-node network, they cut each mean-field update to a handful of nodes. Early active-inference papers used them this way, “if not slightly overzealous” (§4.1).
- The soup (§4.2, Figs. 3–5). Friston 2013b’s simulation, reproduced from its code: the adjacency matrix is built from electrochemical coupling only, “while other forms of influence included in the simulation (such as Newtonian forces) are ignored”; the eight most densely coupled nodes are declared internal; the blanket is then traced; the result is read as a membrane. Calling this a property of the system is “a clear example of the reification fallacy”. §6 adds that steady state is assumed when the simulation is stopped and that the number of internal clusters is “arbitrarily” assumed.
- Slippage in the literature (§4.3). Allen & Friston, Clark, Kirchhoff et al., Hohwy and Ramstead et al. are quoted moving between statistical, causal, spatial, epistemic and autopoietic boundaries.
- Friston blankets on an arbitrary graph (§5.1, Fig. 7). Labelling x10 or x9 internal yields different blankets; neither blanket’s sensory states include the network’s observed variables, which become “external”.
- Co-parents (§5.1, Fig. 8). In a patellar-reflex network a hammer strike is a co-parent of the spinal neurons with the cortical command; an external intervention has “exactly the same formal properties” as an internal cause. Whether co-parents are sensory, active or ignored varies by author.
- Granularity. Which cause is “most proximal” is model-relative (citing Anderson 2017), so the sensory/active boundary is too.
- Dilemma (§6.2, §7). “Inference with a model”: blankets on the scientist’s or agent’s map, legitimate, no ontological consequence. “Inference within a model”: the agent is the model and its blanket is its boundary, which needs metaphysical premises; Ramstead et al. 2019 are shown both deriving blankets from dynamics and “placing” them.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Pearl and Friston blankets are distinct constructs with different assumptions | strong | formal exposition §§3–5; worked graphs |
| C2 | The free-energy literature conflates them | moderate | quotations from six groups (§4.3, §6) |
| C3 | In Friston 2013b the blanket depends on unjustified modelling choices, so attributing it to the system is reification | moderate to strong | re-run of the published code; choices identified (§4.2, §6) |
| C4 | A Friston-blanket partition presupposes a chosen internal set and cannot itself define inside and outside | strong for arbitrary graphs | Fig. 7 |
| C5 | Realist (within-a-model) use needs metaphysical premises not in the formalism | moderate | conceptual argument §6 |
| C6 | Instrumental use cannot settle debates about the boundaries of mind or life | moderate | argument from model-relativity §6.1 |
Concepts
- Pearl blanket. The Markov blanket of a node in a Bayesian network: parents, children, co-parents.
- Friston blanket. A blanket read as a boundary in the world partitioning internal, sensory, active and external states, in recent work defined by sparse coupling at nonequilibrium steady state.
- Inference with a model. A scientist (or agent) uses a generative model to infer hidden states; blankets are features of the model.
- Inference within a model. A model contains both the generative process and an inferring subsystem; the blanket is claimed to bound that subsystem.
Connections
It builds on Biehl, Pollock & Kanai’s technical critique, Andrews and van Es on instrumentalism, and Baltieri, Buckley & Bruineberg’s Watt-governor example of pan-(active-)inferentialism. It cites Barandiaran, Di Paolo & Rohde (LIT-566) for “interactional asymmetry”. Raja et al. (LIT-598) cite this preprint.
Bearing on the record
- LIT-526 and NOTE-421. C3 is the basis of the LIT’s
corrects. It agrees with NOTE-421’s reading (k = 8 chosen by spectral clustering; the lemma as redescription) and adds two choices NOTE-421 did not flag: electrochemical-only adjacency, and steady state assumed at the stopping time. - Demarcation (LIT-192, NOTE-094). If blankets are model-relative, Friston’s exclusion of the candle flame (C7 in NOTE-421) is a fact about a model of a flame, which LIT-526 does not provide.
- THEORY filed as THEORY-073 (Proposed), with the formal part (C4) filed separately as THEORY-066 (Active): a Markov-blanket partition does not individuate a system; where the blanket falls depends on the graph and on which nodes are designated internal, so it presupposes the boundary it is used to find. Sources: this paper and LIT-598.
- No instruction for machine-learning practice.
Limitations
- Read in the version the authors say they rewrote; technical sections may be shorter in the BBS text, and the authors’ replies to commentaries are not read.
- The critique concerns the use of blankets, not the free energy principle’s mathematics, and the authors say so.
- Recent formal definitions of Friston blankets (sparse Hessians at nonequilibrium steady state) are noted, not analysed.
Open questions
- Is there a model-independent criterion that would make a Friston blanket “detected” rather than placed? The later sparse-coupling definitions are the candidate; the paper leaves them open.
- Does the BBS version keep the soup analysis, and how did Friston and colleagues answer it in the commentaries?