A FREE ENERGY PRINCIPLE FOR A

PARTICULAR PHYSICS
Karl Friston

The Wellcome Centre for Human Neuroimaging, UCL Queen Square Institute of Neurology, London, UK WC1N
3AR. Email: [email protected]

(This work is under consideration for publication by The MIT Press)

Abstract
This monograph attempts a theory of every ‘thing’ that can be distinguished from other ‘things’ in a statistical
sense. The ensuing statistical independencies, mediated by Markov blankets, speak to a recursive composition of
ensembles (of things) at increasingly higher spatiotemporal scales. This decomposition provides a description of
small things; e.g., quantum mechanics – via the Schrödinger equation, ensembles of small things – via statistical
mechanics and related fluctuation theorems, through to big things – via classical mechanics. These descriptions
are complemented with a Bayesian mechanics for autonomous or active things. Although this work provides a
formulation of every ‘thing’, its main contribution is to examine the implications of Markov blankets for self-
organisation to nonequilibrium steady-state. In brief, we recover an information geometry and accompanying free
energy principle that allows one to interpret the internal states of something as representing or making inferences
about its external states. The ensuing Bayesian mechanics is compatible with quantum, statistical and classical
mechanics and may offer a formal description of lifelike particles.

Key words: self-organisation; nonequilibrium steady-state; active inference; active particles; free energy;
entropy; random dynamical attractor; autopoiesis; Markov blanket; Bayesian; variational.

Contents

Abstract................................................................................................................................................................... 1

Introduction ............................................................................................................................................................ 4

Part One: the setup .................................................................................................................................................. 7

Something or nothing ......................................................................................................................................... 7

Some preliminaries........................................................................................................................................................... 8

Nonequilibrium steady states ......................................................................................................................................... 10

Fluctuations and information length ............................................................................................................................... 13

Random dynamical systems and Markov blankets ......................................................................................................... 16

Markov blankets and marginal flows ............................................................................................................................. 17

Summary ........................................................................................................................................................................ 18

Symmetry breaking and self-organisation ....................................................................................................... 19

Self-organization and self-evidencing ............................................................................................................................ 24

Self-organisation, frustration and supersymmetry .......................................................................................................... 24

Self-organisation and information length ....................................................................................................................... 28

Summary ........................................................................................................................................................................ 31

Synthetic soups and active matter .................................................................................................................... 32

An active soup ................................................................................................................................................................ 33

A random dynamical attractor and its Markov blankets ................................................................................................. 35

The Markov blanket ....................................................................................................................................................... 35

The emergence of order.................................................................................................................................................. 36

Summary ........................................................................................................................................................................ 37

States, particles and fluctuations ...................................................................................................................... 38

Starting at the end .......................................................................................................................................................... 39

The Markovian partition................................................................................................................................................. 41

The adiabatic reduction .................................................................................................................................................. 44

Elimination and renormalisation .................................................................................................................................... 47

Summary ........................................................................................................................................................................ 50

Part Two: some special cases ............................................................................................................................... 53

A theory of small things – quantum mechanics ............................................................................................... 53

The Schrödinger equation from first principles .............................................................................................................. 55

Wave particle duality and the de Broglie hypothesis ..................................................................................................... 56

Heisenberg uncertainty principle.................................................................................................................................... 58

Inference, measurement and wave function collapse? .................................................................................................... 58

Summary ........................................................................................................................................................................ 60

A theory of lots of little things – statistical mechanics .................................................................................... 64

Stochastic thermodynamics ............................................................................................................................................ 65

Stochastic energetics ...................................................................................................................................................... 68

Fluctuation theorems ...................................................................................................................................................... 72

Summary ........................................................................................................................................................................ 75
2
The free energy principle

A theory of big things – classical mechanics ................................................................................................... 76

Conservative systems ..................................................................................................................................................... 77

Random fluctuations and generalised motion................................................................................................................. 79

Summary ........................................................................................................................................................................ 80

Part Three: a particular case ................................................................................................................................. 84

A theory of autonomous things – Bayesian mechanics.................................................................................... 84

Risk and ambiguity ........................................................................................................................................................ 88

Inference and measurement ............................................................................................................................................ 92

Information geometry..................................................................................................................................................... 93

Variational Bayes ........................................................................................................................................................... 94

Summary ...................................................................................................................................................................... 100

Simulating sentience ...................................................................................................................................... 102

The representation of order .......................................................................................................................................... 103

Summary ...................................................................................................................................................................... 106

Active inference and self-evidencing ............................................................................................................. 107

Active inference with continuous states ....................................................................................................................... 108

Active inference with discrete states ............................................................................................................................ 110

Deep inference: gradient flows or least action?............................................................................................................ 113

Summary ...................................................................................................................................................................... 117

The thermodynamics of inference.................................................................................................................. 118

Potentials and surprisal ................................................................................................................................................ 118

Ensemble free energies ................................................................................................................................................ 120

Summary ...................................................................................................................................................................... 120

Discussion........................................................................................................................................................... 122

Conclusion ..................................................................................................................................................... 122

Appendix A: Stratonovich path integrals ............................................................................................................ 124

Appendix B: lemmas and proofs ........................................................................................................................ 125

Appendix C: nonequilibrium steady-state energy functions ............................................................................... 129

Appendix D: the Fokker Planck operator ........................................................................................................... 130

Appendix E: generalised motion ........................................................................................................................ 131

Appendix F: discrete state-space models ............................................................................................................ 134

References .......................................................................................................................................................... 140

Introduction

This monograph attempts a theory of every ‘thing’ – in a tongue in cheek way – starting from the premise that a
‘thing’ is distinguishable from something else and from no ‘thing’. Its ambition is to validate a formulation of
dynamical systems by appealing to constructs in physics (e.g., quantum, statistical and classical mechanics) and
then use the ensuing formulation to derive an account of self-organisation within the same framework 1. Our
starting point is a definition of things in terms of systems that possess an invariant measure; namely, weakly
mixing systems that possess an attracting set. The description of such systems usually starts using the formalism
of random dynamical systems; for example, the flow or dynamics of systemic states based on random differential
equations (e.g., a Langevin equation). This is where the current treatment starts – and then stops. It stops by asking
some obvious questions; like, what are states and where do random fluctuations come from? These questions lead
to even simpler questions; namely, if we are dealing with the states of something, what is the thing that possesses
those states – and how does one distinguish anything from something else? The answers to these questions lead
to a theory of everything in a literal sense.

To address the nature of things, we start by asking how something can be distinguished from everything else. In
pursuing a formulation of self-organisation, we will call on the notion of conditional independence as the basis of
this separation. More specifically, we assume that for something to exist it must possess (internal or intrinsic)
states that can be separated statistically from (external or extrinsic) states that do not constitute the thing. This
separation implies the existence of a Markov blanket; namely, a set of states that render the internal and external
states conditionally independent. The existence of things (i.e., internal states and their blanket) further implies a
partition of the Markov blanket into active and sensory states – that are not influenced by external and internal
states, respectively. This may sound a bit arbitrary; however, this is the minimal set of conditional independencies
– and implicit partition of states – that licenses talk about things (that possess states). Specifically, it provides a
partition that constitutes the ‘self’ in self-organisation. The subsequent sections tackle the next obvious question:
what are things? At this point, we deploy the Langevin formulation of random dynamical systems as an ansatz
that is recursively self-verifying, when considered in the light of Markov blankets. In brief, the formulation on
offer says that the states of things (i.e., particles) comprise mixtures of blanket states, where the Markov blanket
surrounds things at a smaller scale. Effectively, this eludes the question “what is a thing?” by composing things
from the Markov blanket of smaller things. By induction, we have Markov blankets all the way down, which
means one never has to specify the nature of things.

1
This paper was written as an autodidactic exercise to ensure the author’s intuitions played out over complementary
formulations in statistical physics. The result is a long, over inclusive paper that tries to adopt conventions from different fields
(which the author is not expert in), while emphasizing common themes.

4
The free energy principle

More specifically, we will see that the Langevin formulation of dynamics – at any given spatiotemporal scale –
can be decomposed into an ensemble of Markov blankets. These blanket states have a dynamics at a higher scale
with exactly the same (Langevin) form as the dynamics of the original scale. When lifting the dynamics from one
scale to the next, internal states are effectively eliminated, leaving only slow, macroscopic dynamics of blanket
states. These become the states of things at the next level, which have their own Markov blankets and so on. The
endpoint of this formalism is a description of everything at progressively higher spatial and temporal scales. The
implicit separation of temporal scales is used in subsequent sections to examine the sorts of dynamics, physics or
mechanics of progressively larger things.

This monograph comprises 12 sections organised into three parts. The first part establishes some basic results, the
second part applies these results to limiting cases of dynamical systems to recover quantum, statistical and
classical mechanics. The third part considers the special case of active or autonomous systems, in terms of a
Bayesian mechanics for particles with internal states that ‘matter’ for their behaviour.

Part One: The first section is a foundational treatment that introduces some constraints on the dynamics of
Markov blankets that possess measurable characteristics. The constraint of measurability – or possessing an
invariant measure over sufficiently long periods of time – allows one to express the flow of states as a function of
their non-equilibrium steady-state (NESS) density2. The relationship between flow and the NESS density follows
in a straightforward way from the Fokker Planck formulation of density dynamics and, in particular, its
eigensolution. The interesting results here are the dependencies – implicit in the system’s equations of motion –
that inherit from Markov blankets at nonequilibrium steady-state. The ensuing, relatively straightforward lemma
and corollaries concerning marginal flows and conditional independencies then form the basis for emergent
behaviours in subsequent sections. The second section looks at various ways in which one can characterise density
dynamics in terms of symmetry breaking and self-organisation. This section uses information theory and geometry
to characterise different sorts of self-organisation to nonequilibrium steady-state. The third section provides an
illustration of self-organisation, using numerical analyses of a particular system (a synthetic primordial soup based
on ensemble of Lorenz systems). This system is used throughout the monograph to illustrate how one can take
complementary perspectives on the same dynamics. The fourth section considers the behaviour of this (Langevin)
formulation of Markov blankets at nested scales. In brief, we assume that as one ascends to higher scales, random
and intrinsic fluctuations are progressively suppressed, resulting in a move from dissipative dynamics – that are
dominated by random fluctuations – through to large systems whose conservative dynamics are dominated by
divergence-free flow.

2NESS could also be an acronym for Nearly Ergodic Steady-State in weakly mixing systems. As observed by my young
colleague Brennan Klein, nonequilibrium steady-state puts the “ness” in “thingness” (From Middle English -nes, -nesse:
appended to adjectives to form nouns meaning “the state of being”). We will argue later that any ‘state of being’ rests upon a
NESS.

5
The free energy principle

Part Two: Section 5 considers the very small in terms of quantum mechanics. This section derives the Schrödinger
wave equation using the relationship between a particle’s flow and the NESS density established in the first
section. The trick here is to express or factorise the NESS density in terms of (complex) roots that play the role of
a wave function. Section 6 then considers the collective behaviour of small things in terms of ensemble dynamics
and stochastic thermodynamics. Our focus here is on linking the dissipative dynamics of ensembles to established
results in statistical mechanics; namely, the laws of thermodynamics and related fluctuation theorems such as the
Jarzynski equality. We then turn to the physics of big things in the limit of small amplitude random fluctuations.
This limit allows us to write down equations of motion in terms of a classical Lagrangian or Hamiltonian, leading
to classical mechanics, Newtonian laws of motion and Maxwell's equations.

Part Three: Having cast quantum, statistical and classical mechanics as limiting cases of the density dynamics
of inert particles, we turn to the ontology of big things – whose internal states cannot be ignored – that show
autonomous behaviour (e.g., large active particles like ourselves). Section 8 asks why one might attribute
representational or inferential capacities to biological self-organisation. In other words, how notions like the good
regulator theorem (Conant and Ashby, 1970) and the Bayesian brain hypothesis (Helmholtz, 1878 (1971); Knill
and Pouget, 2004) could be substantiated in terms of a sentient physics. The argument here is fairly
straightforward: namely, that the internal states of a system encode probabilistic beliefs about external states that
cause sensory impressions on the Markov blanket – and are caused by the influence of active states on external
states. This section provides a formal basis for an information geometry and attending free energy principle that
describes autonomous things (e.g., cells or brains) as inferring the causes of actively sampled sensations. Here,
we pursue a variational theme by showing how variational Bayes (Beal, 2003) is an emergent property of certain
kinds of particles, leading to a form of Bayesian mechanics. Section 9 illustrates a particular inference using
numerical analyses of the synthetic soup from Part One (and a virus like denizen). Section 10 then considers active
states and agency in terms of corollaries of the free energy principle based upon an integral fluctuation theorem
and expected free energy. The penultimate section considers the ensuing active inference in light of previous
(thermodynamic) treatments. We conclude with a brief discussion of the relationships between quantum,
stochastic, classical and Bayesian mechanics.