Karl Friston (2013), Journal of the Royal Society Interface 10(86):20130475 (received 27 May 2013, accepted 12 June 2013, published 6 September 2013; CC BY 3.0).
The brief’s citation (J. R. Soc. Interface 10(86):20130475) is correct.
Key takeaways
- The lemma. “Any ergodic random dynamical system that possesses a Markov blanket will appear to actively maintain its structural and dynamical integrity.” With the flow written in the Helmholtz form f = (Γ + R)·∇ln p(x|m), the flow of internal and active states averaged over external states climbs the marginal ergodic density of the blanket and internal states (Eq. 2.6). Choosing the variational density q(ψ|λ) equal to the posterior over external states makes that same flow a descent on the variational free energy F (Eqs. 2.7–2.8).
- The inversion. Earlier arguments said living systems must minimize free energy or their sensory entropy would grow without bound. This paper turns that round: “any system that exists will appear to minimize free energy and therefore engage in active inference.” The word “appear” carries the claim.
- Four marks of the lifelike. Ergodicity, a Markov blanket, active inference and autopoiesis, used as tests on the simulation.
- The simulation. 128 subsystems with Lorenz “electrochemical” states and Newtonian motion, coupled within distance one. The principal Markov blanket of the time-averaged adjacency matrix encloses 8 internal subsystems. Internal dynamics predict the motion of 5 of the 82 external subsystems above the largest value of a time-reversed null. Lesioning active, sensory or internal subsystems makes the structure disperse (“oscillator death”).
- Demarcation. “A candle flame cannot possess a Markov blanket, because any pattern of molecular interactions is destroyed almost instantaneously by the flux of gas molecules from its surface.” Then, in the conclusion: “there are probably an uncountable number of Markov blankets in the universe. Does this mean there is lifelike behaviour everywhere?” The answer offered is a conjecture: the blankets of living systems have minimum entropy, but not so little as “a petrified stone”.
Standing in the record
Filed on 2026-10-02 at the owner’s request, as part of filling out the record’s coverage of the “life as dissipation” literature and the philosophy that tries to mark living systems off from merely dissipative ones. No anthology topic holds it. The anthology does not hold this paper or any paper on the free energy principle, and the paper carries no instruction for machine-learning practice. It borrows the variational bound from machine learning (Hinton & van Camp, Beal, the Helmholtz machine) but uses it to describe organisms. So it is not filed there, and ADR-013 dual holding does not arise.
NOTE-421 is the close reading of 2026-10-02, and it placed the work: Active. It is the text later Markov-blanket accounts of life cite, and it is short and complete. As an argument it is weaker than its title. The “proof” is a redescription that holds because the variational density is set equal to the posterior. The posterior depends on sensory and active states as well as internal ones, which the paper’s q(ψ|λ) does not allow for. The four marks of life are read off one simulation run.
Where it bears on what the record holds:
- The demarcation question (LIT-192, NOTE-094). Nahas & Sachs report that closure-of-constraints accounts face the candle-flame problem, and that Deacon & Cashman fault autopoiesis for lacking tools to tell living teleology from dissipative self-organization. Friston addresses the candle flame by name and says it is excluded. His criterion is that a Markov blanket needs a set of states whose couplings “change slowly with respect to the states per se”. That is a principled line in form, statistical persistence of a conditional-independence structure, but the paper asserts it of the flame in one sentence and shows it for nothing else. His own conclusion then gives the line away: blankets are everywhere, and what marks life is a conjecture about blanket entropy.
- Autopoiesis. Bruckner (LIT-120, NOTE-093) uses a loose, behavioural autopoiesis to ground welfare subjecthood. Here autopoiesis is operationalized as dispersion after lesioning blanket states in a simulation. That is a third use, closer to a test than either Bruckner’s or the organizational accounts NOTE-094 surveys.
- Seth (LIT-135). Seth’s biological naturalism runs from predictive processing through the free energy principle to autopoiesis. It takes thermodynamic and variational free energy to be “in some physical sense … the same thing” (NOTE-169). This paper, which Seth’s line of argument draws on, computes no thermodynamic quantity. Its free energy is the variational bound of Eq. (2.7), and its only contact with the second law is the introduction’s sentence that bounding sensory entropy “enables biological systems to resist the second law”. So this paper does not supply the equivalence Seth relies on (his C5).
- Still et al. (LIT-327) and THEORY-026. Both tie a system’s internal states to prediction of what drives it. Still makes this a physical identity, dissipated work equals nonpredictive memory, for a system driven without feedback. Friston’s systems are built on feedback: active states change external states, a “circular causality”. That is the regime where THEORY-026 says (from LIT-041) prediction is no longer what efficiency requires. Friston’s prediction is also not priced in work; his “inference” is a description of how the dynamics look. The two should not be read as one result in two vocabularies.
- Levin (LIT-439) and nested selves. Friston’s remark that “the Markov blanket of an animal encloses the Markov blankets of its organs, which enclose Markov blankets of cells” is the same nesting Levin’s Selves have (NOTE-341). It also bears on the anti-nesting question in THEORY-043, though Friston does not take it up.
- Artime & De Domenico (LIT-021). That review cites this paper for its claim that life is emergent in any random dynamical system with a Markov blanket, given a “heuristic theoretical and computational proof”. NOTE-007 left the claim unassessed (its C8). This reading assesses it: the claim is the paper’s own, stated as “(almost) inevitable”, and its support is the lemma (a redescription) plus one simulation run.
- “What Lives?” (LIT-211). Friston’s account sits across two of its clusters (NOTE-109), Cognitive Autonomy (apparent inference, action) and Dissipative Self-Organizing Systems (the soup “is dissipative at two levels”).