Contribution
Before this paper, the free energy principle was argued as a necessity: systems that did not minimize free energy would not keep their sensory entropy bounded, so living systems must. This paper inverts the argument. Any ergodic system with a Markov blanket can be described as minimizing a variational free energy, so inference-like and self-maintaining behaviour is “(almost) inevitable” wherever blankets form. It adds a simulation that operationalizes four marks of the lifelike: ergodicity, a blanket, predictive coupling across the blanket, and dispersion under lesion.
Key insight
If internal states are shielded from external states by a blanket, their time-averaged flow can depend on external states only through the posterior given the blanket. So their motion can be read as if they encoded beliefs about what lies outside. Self-maintenance is then the same flow seen as bounding the entropy of the blanket and internal states.
Assumptions
- Dynamics (Eq. 2.1). ẋ = f(x) + ω, with the partition x = (ψ, s, a, λ): external, sensory, active, internal. Sensory and external flows depend on (ψ, s, a); active and internal flows depend on (s, a, λ).
- Ergodicity. The system converges to a random global attractor with an ergodic density p(x|m) solving the Fokker–Planck equation (Eq. 2.2).
- Helmholtz form (Eq. 2.3). f = (Γ + R)·∇(−G), with Γ the diffusion tensor and R antisymmetric. Eq. (2.4) then takes p = exp(−G) as the stationary solution “using this standard form”. For state-dependent Γ or R that solution needs a correction term the paper does not write; it holds as written when they are constant.
- Eq. (2.6) moves (Γ + R) outside the integral over ψ. That requires the relevant blocks of Γ + R not to depend on ψ. This is not stated.
- The variational density is parametrized by internal states only, q(ψ|λ), but the proof sets it equal to the posterior p(ψ|s, a, λ), which in general depends on s and a too. The step needs the posterior to depend on the blanket only through λ, or q to be read as q(ψ|s, a, λ). This is not stated.
- Short-range coupling makes blankets (almost) inevitable. Asserted: “if the coupling between dynamical systems can be neglected … the intervening systems will necessarily form a Markov blanket”. Not shown for any physical system.
- A blanket’s couplings must change slowly relative to the states. This is introduced to exclude the candle flame and used again in the conclusion.
Key results
- Eqs. (2.5)–(2.6). Internal and active flows are a “circuitous gradient ascent” on ln p(s, a, λ|m), the marginal over blanket and internal states, once averaged over external states.
- Lemma 2.1 (Eqs. 2.7–2.8). F(s, a, λ) = E_q[G] − H[q] = −ln p(s, a, λ|m) + D_KL[q(ψ|λ) ‖ p(ψ|s, a, λ)]. Eq. (2.6) “requires the gradients of the divergence to be zero”, so q equals the posterior, D_KL = 0, and the flows are −(Γ + R)·∇F. Given the assumptions above this is consistent, but it is a redescription: F equals −ln p(v) by construction at that point, so “minimizing free energy” says no more than Eq. (2.6) already said.
- Eq. (2.9). F ≥ −ln p(s, a, λ|m), so the time average of F upper-bounds the entropy H[p(s, a, λ|m)]. Read as action “plac[ing] an upper bound on their dispersion”. The entropy is fixed by the ergodic density the system already has; nothing in Eq. (2.9) is controlled.
- Simulation (§3). 128 subsystems, forward Euler with step 1/512 s,
unit-variance noise. Ergodic behaviour is reported “for most values of
the parameters”, after about 1000 s. The principal eigenvector of B = A
- Aᵀ + AᵀA, with A taken over the last 256 s of 2048 s, picks k = 8 internal subsystems; their blanket is recovered as [B·χ]. Functionally closed subsystems end up at the periphery, and “no simulation ever produced a functionally closed internal state”.
- Prediction across the blanket (§3.4, Fig. 4). Canonical variates analysis of 32 eigenvariates of lagged internal states against the motion of each external subsystem; 5 of 82 exceed the largest χ² of a time-reversed null. Reported p = 0.00052.
- Lesions (§3.5, Fig. 5). Making active, sensory or internal subsystems functionally closed makes the cluster disperse within 512 s. The intact cluster holds its “quasicrystalline arrangement”.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | In an ergodic system with a Markov blanket, internal and active flows can be written as gradient descent on a variational free energy | moderate | Lemma 2.1, given unstated conditions on Γ, R and q (Assumptions); a redescription, not a new dynamical fact |
| C2 | Such systems therefore appear to perform Bayesian inference about external states | interpretation of C1 | Remarks 2.2; “appear” is the paper’s own hedge |
| C3 | Action places an upper bound on the entropy of the blanket and internal states, i.e. self-maintenance | weak as a causal claim | Eq. (2.9) is an inequality about a fixed ergodic density |
| C4 | Markov blankets are (almost) inevitable under short-range coupling, so lifelike behaviour is (almost) inevitable | assertion | no argument beyond the sentence; the conclusion concedes it does not address “the conditions … necessary for the emergence of ergodic Markov blankets” |
| C5 | A simulated soup shows a blanket whose internal states predict external motion | moderate, for one run | Fig. 4; 5/82 above a single time-reversed null; the p-value is not reproduced (Corrections) |
| C6 | Lesioning blanket or internal states destroys structural integrity (autopoiesis) | weak to moderate | Fig. 5; one run per lesion type, assessed by eye |
| C7 | A candle flame cannot possess a Markov blanket | assertion | one sentence (§2); no model of a flame |
| C8 | Minimum-entropy Markov blankets may characterize biological systems | conjecture | §4, labelled “speculative” |
Concepts
- Markov blanket. A set of states separating internal from external states in the conditional-independence sense: parents, children and parents of children (Pearl). Here it is split into sensory states (children of external states) and active states (not).
- Ergodic density. The occupancy measure of the random attractor, read as the probability of finding the system in a state.
- Gibbs energy G. −ln p(x|m), the potential in the Helmholtz form.
- Variational free energy F. E_q[G] − H[q], the bound from variational Bayes (Feynman; Hinton & van Camp; Beal). It is not a thermodynamic free energy, and the paper does not claim it is.
- Active inference. Internal states appear to infer external causes; active states, through the blanket, change the external states that cause sensations.
- Autopoiesis (operational). Self-organized dynamics are necessary to maintain structural integrity, tested by lesion.
Connections
- Schrödinger (1944), the paper’s epigraph and ref. [1]. It takes Schrödinger’s question, how events “within the spatial boundary of a living organism” are accounted for by physics. Its answer locates the boundary in statistics rather than in a membrane. Filed beside it in this batch as LIT-530 (Deferred).
- Dissipative structures. §3 compares the soup with Turing instabilities, Bénard cells and the Belousov–Zhabotinsky reaction, and calls the ensemble “dissipative at two levels”: viscous friction, and non-divergence-free functional dynamics. Neither sense is entropy production.
- Conant & Ashby, the good regulator theorem (ref. [13]). Invoked as “exactly consistent”. Not held in the record.
- Predictive information (Bialek, Nemenman & Tishby; Ay et al.; refs. [22–24]). Named as related; the information bottleneck lineage is LIT-338.
- Still et al., LIT-327 (NOTE-295), and THEORY-026. See “Bearing on the record”.
- Seth, LIT-135 (NOTE-169). Seth’s route from the free energy principle to autopoiesis passes through a claimed equivalence of variational and thermodynamic free energy. This paper does not provide it.
- Levin, LIT-439 (NOTE-341). Blankets nest (animal, organ, cell, nucleus), as Levin’s Selves do.
Bearing on the record
- THEORY-026 (Proposed): a different and weaker claim, not a second source. THEORY-026 is a physical identity, dissipated work equals nonpredictive memory, for a system driven without feedback. Friston’s internal states “predict” external states in a system defined by feedback (active states change external states). His “prediction” carries no work or heat. Where it touches the THEORY, it falls on the feedback side, where THEORY-026 (from LIT-041) says prediction is no longer what efficiency requires. This paper should not be cited as thermodynamic support for “living systems are efficient predictors”.
- THEORY-043 (Proposed). Nested blankets, each inducing “active (Bayesian) inference”, are a case of minded parts inside a minded whole. The paper takes no position on anti-nesting, and its conjecture (C8) would pick one blanket out by entropy. That is a candidate architectural criterion of the kind THEORY-043 says is not yet principled. It is conjectured here, not argued.
- LIT-192 / NOTE-094, the demarcation question. The only text in this batch that names the candle flame and draws a line against it (C7). The line, persistence of a conditional-independence structure, is principled in form but asserted, not shown, and the conclusion’s “uncountable number of Markov blankets” puts the burden back on a conjecture (C8).
- LIT-021 / NOTE-007 C8. NOTE-007 recorded, secondhand and unassessed, the claim that life is emergent in any random dynamical system with a Markov blanket. Assessed here: it is the paper’s claim, and its support is C1 (a redescription) plus C5–C6 (one simulation).
- No ML instruction. The paper uses variational Bayes to describe organisms; it recommends nothing for building or training models. The anthology does not hold it.
Limitations
- The inference reading depends on the choice q(ψ|λ) = posterior. Any ergodic system with the stated sparsity can be redescribed this way, so the description does not separate living from non-living systems with blankets. The paper concedes this in §4.
- Γ and R are taken as constant without saying so, and the q(ψ|λ) versus p(ψ|s, a, λ) mismatch is unaddressed (Assumptions).
- The free energy is variational. The paper computes no heat, work or entropy production. Its contact with thermodynamics is the introduction’s appeal to the fluctuation theorem.
- The simulation is one parameter setting, with internal states fixed at k = 8 by the analyst, a blanket recovered from an adjacency matrix aggregated over 256 s, and lesion effects judged visually.
- Ergodicity is assumed for living systems, and the conclusion concedes that real systems are “only locally ergodic”.
Open questions
- Is there a formal criterion that separates a candle flame from a cell on the blanket account, and does a model of a flame actually lack one?
- Does the minimum-entropy blanket conjecture (C8) pick out the systems we call living in any worked case?
- Under what conditions on the coupling does a sparse coupling graph give a blanket in the stationary density (which the lemma needs), and not only in the instantaneous dynamics?
Corrections
- none to a seeded skim (there was no seed)
- The reported p-value does not follow from the stated null. The Fig. 4 legend says the largest null statistic protects against false positives “at a level of 1/82”, and that five true statistics above it give p = 0.00052. Under that per-element rate, P(≥ 5 of 82) is binomial: I compute 0.0034, about six times larger. The result is still small, but the figure printed is not reproduced by the stated procedure, and the paper does not say how 0.00052 was obtained.