entropy

Article
Fusions of Consciousness
Donald D. Hoffman 1 , Chetan Prakash 2,3 and Robert Prentner 3,4, *

1 Department of Cognitive Sciences, University of California, Irvine, CA 92697, USA

2 Department of Mathematics, California State University, San Bernadino, CA 92407, USA
3 Association for Mathematical Consciousness Science, D-80539 Munich, Germany
4 Munich Center for Mathematical Philosophy, LMU Munich, D-80539 Munich, Germany
* Correspondence: robert.prentner@amcs.science

Abstract: What are conscious experiences? Can they combine to form new experiences? What are
conscious subjects? Can they combine to form new subjects? Most attempts to answer these questions
assume that spacetime, and some of its particles, are fundamental. However, physicists tell us that
spacetime cannot be fundamental. Spacetime, they say, is doomed. We heed the physicists, and drop
the assumption that spacetime is fundamental. We assume instead that subjects and experiences
are entities beyond spacetime, not within spacetime. We make this precise in a mathematical theory
of conscious agents, whose dynamics are described by Markov chains. We show how (1) agents
combine into more complex agents, (2) agents fuse into simpler agents, and (3) qualia fuse to
create new qualia. The possible dynamics of n agents form an n(n − 1)-dimensional polytope with
nn vertices—the Markov polytope Mn . The total fusions of n agents and qualia form an (n − 1)-
dimensional simplex—the fusion simplex Fn . To project the Markovian dynamics of conscious agents
onto scattering processes in spacetime, we define a new map from Markov chains to decorated
permutations. Such permutations—along with helicities, or masses and spins—invariantly encode
all physical information used to compute scattering amplitudes. We propose that spacetime and
scattering processes are a data structure that codes for interactions of conscious agents: a particle in
spacetime is a projection of the Markovian dynamics of a communicating class of conscious agents.

Keywords: consciousness; qualia; subjective experience; hard problem of consciousness;

panpsychism; combination problem; conscious agents; interface theory of perception; decorated
permutations; amplituhedron; positive Grassmannian; Markov chains; Markov polytopes; fusion
simplex

Citation: Hoffman, D.D.; Prakash, C.;

1. Introduction
Prentner, R. Fusions of
Consciousness. Entropy 2023, 25, 129.
Consciousness is perplexing, even for expert researchers. Witness the recent plethora
https://doi.org/10.3390/e25010129
of conflicting theories [1]. Even their core ideas are at odds: quantum states of neuronal
microtubules [2–4], causal architectures that integrate information [5–8], neuronal global
Received: 24 November 2022 workspaces [9–12], user illusions and attentional schemas [13–15], panpsychism [16,17],
Revised: 2 January 2023
and various forms of dualism [18,19]. However, most of these theories of consciousness
Accepted: 3 January 2023
agree on a key assumption: spacetime, and some of its particles, are fundamental, i.e., on-
Published: 9 January 2023
tologically primitive, irreducible, and non-emergent. For example, physicalist theories
assume this and nothing more, while many panpsychists likewise assume this but would
add that the “intrinsic nature” [20] of such particles is nothing other than consciousness.
Copyright: © 2023 by the authors.
It is natural to assume that spacetime is fundamental. Indeed, one might argue that
Licensee MDPI, Basel, Switzerland. any theory must assume this to respect some variety of naturalism [21]. So why bother to
This article is an open access article point it out? Because spacetime and its particles are doomed, as a fundamental posit in any
distributed under the terms and theory of reality. Current research in physics tells us so. We sketch, in Section 2, how the
conditions of the Creative Commons interaction of gravity and quantum theory leads physicists to this stunning conclusion.
Attribution (CC BY) license (https:// Remarkably, evolution by natural selection agrees: spacetime and objects are not
creativecommons.org/licenses/by/ fundamental. We explain, in Section 3, how evolutionary games reveal that our perceptions
4.0/). of space, time, and objects are no more than a user interface that guides adaptive action.

Entropy 2023, 25, 129. https://doi.org/10.3390/e25010129 https://www.mdpi.com/journal/entropy

Entropy 2023, 25, 129 2 of 40

An interface? To what? Both physics and evolutionary theory are silent. We propose,
in Section 4, that conscious agents are fundamental. Some agents interact with others via
an interface whose format is spacetime and objects located in spacetime. We offer a precise
definition of conscious agent.
When agents interact, what happens? We demonstrate, in Section 5, that they can
combine into a complex agent, or fuse into a simpler agent with a novel conscious ex-
perience. If n agents interact, their possible combinations form an n(n − 1)-dimensional
polytope with nn vertices—the Markov polytope Mn . Their possible fusions form an (n − 1)-
dimensional simplex—the fusion simplex Fn .
If agents are fundamental and spacetime is their interface, how precisely do agents
create a spacetime interface? As it happens, theoretical physicists have recently peered
beyond spacetime and discovered new structures beyond spacetime, such as the ampli-
tuhedron [22] and cosmological polytope [23]. They generate spacetime and quantum
theory by projection. The essence of these structures, their invariant physical content, can
be derived from what is known as “decorated permutations” (in non-supersymmetric
theories helicities, or masses and spins, are also required [24]). This is discussed briefly in
Section 6 but leaves open the question: what beyond spacetime is permuted, and why? We
submit, in Section 7, that decorated permutations are a convenient précis of the dynamics
of conscious agents, and we conjecture how to relate spacetime-physics to the combination
and fusion of conscious agents.
Agents are no ephemerals in spacetime; spacetime is a data structure for compact
representation of agent dynamics. For this reason, spacetime is not fundamental. It is an
interface. However, if spacetime is an interface, then its objects, such as neurons and brains,
are icons in the interface—useful fictions. Neurons have no causal powers. Standard inter-
pretations for neural correlates of consciousness assume otherwise. We discuss, in Section 8,
how these interpretations must be revised. We conclude by noting limitations of the theory
of conscious agents, and scouting new directions for its development.

2. Spacetime Is Doomed
Most theories of consciousness take spacetime as fundamental. Most theoretical physi-
cists do not. The disconnect is striking, and is a strike against most theories
of consciousness.
Physicists tell us that spacetime lacks operational meaning for distances smaller
than the Planck length, roughly 10−33 centimeters, or durations shorter than the Planck
time, roughly 10−43 s [25]. They recognize that “classical spacetime is not a fundamental
ingredient of the world, but a construction consisting of more fundamental degrees of
freedom” [26]. This view is not exotic. On the contrary, many physicists believe this,
no matter what particular approach to spacetime they prefer. However, despite this
consensus that spacetime is not fundamental, most theories of consciousness in cognitive
neuroscience, artificial intelligence, and philosophy of mind still include spacetime among
their foundational entities. In this section, we review some arguments from physics for the
doom of spacetime [25]. There are many more, but for the sake of our discussion we will
limit ourselves to the following two.
The first argument starts with a simple fact: to measure smaller objects we need
light, or other radiation, with shorter wavelengths. Quantum theory tells us that if the
wavelength, λ, of the radiation decreases then its energy, E, increases,

hc
E= (1)
λ
where c is the speed of light and h is Planck’s constant. This is why particle colliders, such
as the Large Hadron Collider, use higher energies to probe smaller particles.
Entropy 2023, 25, 129 3 of 40

Without gravity, there would be no obstruction. We could in principle generate

endlessly higher energies to probe ever smaller particles. However, gravity spoils the party.
Einstein taught us that mass and energy are equivalent,

E = mc2 , (2)

where m is mass. Einstein also taught us that mass curves spacetime, with greater mass
creating greater curvature. So, as the wavelength of the radiation shrinks, the concentration
of mass-energy grows, creating greater curvature in spacetime. As the wavelength nears
the Planck scale, the curvature of spacetime becomes so great that radiation cannot escape.
A black hole is born, and destroys the object we want to observe. If we persist, and use
even higher energies, the black hole just becomes bigger. Thus, no operational meaning
can be given to distances and durations below the Planck scale.
A second argument turns on the quantum theory of measurement. Suppose we have
a room of fixed size that contains a measuring device and a particle to be measured. Every
measuring device is a physical system. As such, it is subject to the quantum uncertainty for
energy and time, and to the uncertainty for position, x, and momentum, p,

∆E∆t ≥ h/(4π ), (3)

∆x∆p ≥ h/(4π ). (4)

The uncertainty principle dictates that to make the device more accurate we must add
degrees of freedom. As we add them, we add more parts to the device, and it becomes more
massive. Eventually, the device collapses into a black hole and destroys the measurement.
Arguments like these have prompted many physicists to conclude that spacetime is
not fundamental. David Gross, for instance, says “I believe that space for sure, and probably
time as well, will be emergent” [25]. In this same paper, Gross quotes Ed Witten saying,
“Space and time may be doomed”; Nathan Seiberg, “I am almost certain that space and time
are illusions”; and Andrew Strominger, “The notion of spacetime is clearly something we’re
going to have to give up.” In his 2010 Cornell Messenger Lecture [27], Nima Arkani-Hamed
says, “the very notion of spacetime is not a fundamental one. Spacetime is doomed. There is
no such thing as spacetime fundamentally in the actual underlying description of the laws
of physics.” Arkani-Hamed [28] also argues that quantum theory itself is not fundamental,
and will arise with spacetime from some deeper structure: “So there’s some other structure
that we’re looking for, and some way of thinking about interpreting this structure will let
us see spacetime and quantum mechanics emerge simultaneously and joined at the hip.”
If spacetime is not fundamental, neither are its particles, which are irreducible repre-
sentations of the Poincaré symmetries of spacetime. Nor are macroscopic objects, such as
neurons and brains, made of particles. A theory of consciousness that starts with spacetime,
particles, neurons, or brains has little chance. This is particularly problematic for versions
of panpsychism that assume spacetime is the stage where the drama of consciousness plays
out [29].

3. Evolutionary Games
Quantum field theory and gravity together tell us that spacetime and objects in
spacetime are doomed. However, how is this consistent with another pillar of modern
science, the theory of evolution by natural selection? Evolution describes how certain
objects called organisms evolve in time and forage for resources distributed in space.
Usually, this is taken to imply that evolution by natural selection provides us with
a grip on the true structure of reality. Surely organisms that see reality more accurately
are necessarily more fit—more likely to pass on their genes. We are here today because
our ancestors saw reality more accurately than their competition. For example, according
to the philosopher and psycholinguist Jerry Fodor, “there is nothing in the ‘evolutionary’,
or the ‘biological’, or the ‘scientific’ worldview that shows, or even suggests, that the proper
function of cognition is other than the fixation of true beliefs” [30].
Entropy 2023, 25, 129 4 of 40

It may seem obvious that perceiving the truth must make you more fit. However, this
is false, and flatly contradicted by the mathematics of evolution itself. As the philosopher
Patricia Churchland puts it, “[t]he principle chore of brains is to get the body parts where
they should be in order that the organism may survive. Improvements in sensorimotor
control confer an evolutionary advantage: a fancier style of representing [the world]
is advantageous so long as it is geared to the organism’s way of life and enhances an
organism’s chances for survival. Truth, whatever that is, takes the hindmost” [31].
The cognitive scientist Steven Pinker agrees: “Our minds evolved by natural selection
to solve problems that were life-and-death matters to our ancestors, not to commune with
correctness” [32]. Pinker scouts several counterexamples to the idea that perceiving truth
enhances fitness [33]. For instance, computing the truth takes valuable resources, such as
time and energy; this can favor the evolution of heuristics that cut corners on truth and
resources. The desire for social acceptance can lead one to adopt false beliefs as a form of
virtue signaling. Strategic lies can be advantageous; the best liars are those who are not
aware they are lying.
Pinker goes on, however, to admit that, “The idea that the mind is designed for truth is
not completely wrong. We do have some reliable notions about the distribution of middle-
sized objects around us ....” [33] Churchland agrees. She assumes that certain middle-sized
objects, such as brains and body parts, are not fictions.
However, middle-sized objects, it turns out, are not exempt. This becomes clear once
one distinguishes between informal accounts of biological evolution and mathematical
treatments based on evolutionary game theory. It is only the former that appear to assume
the reality of spacetime and middle-sized objects.
Simulations and theorems using the tools of evolutionary game theory reveal a coun-
terintuitive result: the probability is zero that any sensory system has ever been shaped to
report any true structures of objective reality [34–36]. Our senses do not show us truths
about objective reality. They simply guide adaptive action.
A key insight from these studies centers on the “fitness payoffs” that govern evolution.
One can think of evolution as like a video game in which one must collect points and
avoid death to reach the next level of the game. Fitness payoffs in evolutionary theory
are like those points. Organisms that reap more fitness payoffs send their genes to the
next generation.
Fitness payoffs thus shape the evolution of sensory systems. Are the senses shaped to
report truths about the structures of objective reality? No, for the simple reason that fitness
payoff functions are, almost surely, not homomorphisms of the structures of objective
reality [37]. That is, fitness payoffs lack information about the structures of objective reality,
and hence cannot shape the senses to perceive those structures.
However, does this argument not refute itself? It uses math and logic to prove that
human cognitive capacities, such as math and logic, are unreliable. Not at all. The argu-
ments here only target the evolution of sensory systems, not the evolution of all cognitive
capacities. Each capacity must be studied separately. In the case of math and logic, there
are selection pressures for some proficiency: we can reap fitness advantages by accurate
reasoning about fitness payoffs. For instance, two bites of a pear offer more fitness payoffs
than one bite. The pear is not objective reality; it is a representation of fitness payoffs that
can occur for different actions, such as biting it. Reasoning accurately about these fitness
payoffs can enhance fitness, even if our sensory systems do not reveal objective reality.
If our senses are not a window on reality, then what are they? A useful metaphor is
to think of our senses as providing a user interface to an unknown reality, much like the
desktop interface on a laptop ([34,38–47]; see also [48–50] for von Uexküll’s related idea
of an Umwelt). The icons on the desktop let you control voltages in circuits of the laptop
without you having to know anything about these circuits. If you had to toggle voltages to
craft an email, no one would hear from you. Knowing the truth will not make you more
fit; it will make you extinct. Evolution shaped our senses to guide adaptive action, not to
know the truth. This is called the “interface theory of perception” (ITP).
Entropy 2023, 25, 129 5 of 40

So, evolution by natural selection agrees with physics that spacetime and objects,
as reported by our senses, are not fundamental reality. They are simply data structures
that guide adaptive action [51], i.e., that help us reap fitness payoffs. There are countless
fitness payoff functions relevant to our actions. We hierarchically cluster these payoff
functions and compress the clusters into convenient units that we call macroscopic objects.
The actions and payoffs appropriate to a pear are quite different from those appropriate
to poison ivy. The tens of thousands of perceptual units that we call objects are the way
our senses handle the plethora of fitness payoffs that we must deal with. We do not deal
with this plethora by seeing the truth. We deal with it by creating thousands of objects as
compact data structures that guide context-appropriate actions.
This sensory process of clustering fitness payoffs into objects can be “cognitively
impenetrable”: its inner workings can proceed independently of higher cognitive states,
such as goals and beliefs. Moreover, the construction of objects can occur simultaneously
in multiple sensory systems, such as vision and touch; this creates concurrent mappings
from a single stimulus to multiple perceived objects, such as a visual kitten and a tactile
kitten that are perceived simultaneously when one sees and pets a kitten. (For some people,
the simultaneous visual object and tactile object are not even recognized as arising from the
same stimulus [52]). One’s cognitive goals can guide attention among the objects delivered
by a sensory system, and this attention can trigger the sensory system to provide more
details on an object of interest. This mapping of a single stimulus into distinct objects in
concurrent sensory systems contradicts a recent claim: “Critically, as an implementation
of cognitive impenetrability, the agent must always use the same mapping from stimulus
(resource) to percept (color), regardless of what the current payoff function is” [53]. This
mistaken claim has been used to justify a further mistaken claim that sensory systems,
if forced to handle many payoff functions, will be shaped by natural selection to construct
veridical perceptions [53].
Objects are not mind-independent entities. They are data structures that we construct
and employ as needed, and then garbage-collect when, for the moment, we no longer need
them. We create each object with a glance and delete it with a blink, or a glance away.
Objects do not exist when they are not observed. Thus, they have no definite values of
physical properties, such as position or momentum when they are not observed. This aligns
with the conclusion of quantum physics that local realism and non-contextual realism are
both false [54–56].
However, what about, say, a high-speed train bulleting down a track? If you think it is
not there when you do not look, why not step in front of it and simply look away?
The answer is that evolution shaped our senses to guide adaptive behavior. We must
take them seriously. However, that does not entail that we are entitled to take them literally.
To compare, I must take a blue, rectangular icon on my desktop seriously. If I drag it to the
trash, I could lose my file. However, the file is not literally blue or rectangular. Similarly,
when I see a train I am interacting with some reality that I must take seriously, and that
reality exists whether I look or not. However, that reality is not a train. Indeed, as we will
discuss, that reality is beyond space and time.
This applies to neurons and brains. They do not exist when they are not observed.
We create them when we look and delete them when we look away. Thus, neurons create
none of our behaviors or conscious experiences. Theories of consciousness that assume
otherwise contradict the clear implications of physics and evolution.
However, is this not too hasty? Do we not have good evidence that brains exist and
cause our conscious experiences? Consider the long list of neural correlates of conscious-
ness that have been found. Activity in area V4 of the visual cortex, for instance, is correlated
with our conscious experience of color. If a transcranial magnetic stimulation (TMS) device
is used to inhibit area V4 in the left hemisphere of the brain of an awake person, then
that person will lose all color experience in their right visual field. Everything looks like
a black and white photograph. If the TMS device is removed, then the person will see
color flow back into their right visual field. Intervening on the brain causes changes in
Entropy 2023, 25, 129 6 of 40

conscious experiences. So, apparently, brains must exist and cause our conscious experi-
ences. As Gerald Edelman put it, “There is now a vast amount of empirical evidence to
support the idea that consciousness emerges from the organization and operation of the
brain ... The question then becomes: What features of the body and brain are necessary and
sufficient for consciousness to appear?” [57].
However, this conclusion does not follow. Consider a virtual-reality tennis game. I
hit a virtual tennis ball with my virtual tennis racket. I intervene with my virtual racket,
and it causes the virtual ball to move. However, this does not entail that the virtual racket
exists when I do not perceive it. In fact, the racket does not exist when it is not perceived.
Additionally, it has no causal powers. In this metaphor, what has causal powers is some
complicated supercomputer, and what I am really doing is toggling millions of voltages in
that computer. If I had to toggle them explicitly, I would be overwhelmed. The racket is
just a useful fiction that lets me play the game. The same is true of physical objects, such
as neurons.
If neurons do not exist when they are not perceived, if they are just useful fictions,
should the science of consciousness ignore neuroscience? Not at all. On the contrary, we
need more funding for neuroscience, not less. The relationship between neural activity and
conscious experience is far more complex than is usually imagined. We need to reverse
engineer neurons to understand the deeper reality behind the useful fiction.
We can think of ITP in terms of a distinction that some philosophers have drawn
between primary and secondary qualities. John Locke argued that “primary qualities” of
objects, such as their “bulk, figure, or motion” exist when unperceived, but that “secondary
qualities” of objects, such as their “colors and smells” do not. He then claimed that “... the
ideas of primary qualities of bodies are resemblances of them, and their patterns do really
exist in the bodies themselves, but the ideas produced in us by these secondary qualities
have no resemblance of them at all” [58]. ITP says that all qualities are secondary qualities.
ITP thus agrees with Immanuel Kant who proposed to “go farther, and for weighty reasons
rank as mere appearances the remaining qualities of bodies also, which are called primary,
such as extension, place, and in general space” [59].

4. Conscious Agents
Quantum theory and gravity tell us that spacetime is doomed. Spacetime and objects
are not fundamental reality. Evolution by natural selection agrees.
What, then, is fundamental? These theories are silent. Our theory of spacetime tells us
the limits of spacetime, but it cannot reveal what lies beyond. Our theory of evolution tells
us that our sensory systems, which show us objects in space and time, are not a window on
truth. They are a user interface. However, an interface to what? What reality lies beyond
our spacetime interface? Evolution cannot say.
So, we must take a leap. We must propose some reality beyond spacetime. We must
then propose precisely how this deeper reality maps onto spacetime. Additionally, we must
show that this deeper reality looks like quantum field theory and evolution by natural
selection when it is projected onto spacetime.
We propose that consciousness is fundamental, and can be modeled as a network of
interacting “conscious agents”. In this section, we briefly motivate and present a definition
of conscious agents. A more detailed development is presented elsewhere [60,61].
Here, we seek minimal posits about consciousness that permit the construction of a
general theory. This makes our approach unusual. Most scientific theories of consciousness
posit physical systems, or functional properties of physical systems. Some propose that
these give rise to consciousness; for instance, global workspace [10,62,63], integrated
information [8,64], and orchestrated objective reduction [3]. Others propose that these give
rise to the illusion of consciousness; for instance, illusionism [14] and attention schema [15].
All of the above falsely assume that spacetime physics is fundamental. This is not the
case, for example, in the system of G.W. Leibniz [65] who proposed that simple substances
(“monads”) are the ultimate constituents of the universe and that physics, as it was known
Entropy 2023, 25, 129 7 of 40

back then, would result from the dynamics of a network of such monads (Other related
views, sometimes referred to as “objective idealism”, are discussed in [66]) Our own view
has some similarities with Leibniz’, but also some differences.
We start with two posits about consciousness: (1) there are conscious experiences; and
(2) there are probabilistic relations among conscious experiences. These posits lead us to
the notion of a conscious agent. Our posits for the notion of a conscious agent mirror G.W.
Leibniz’s posits for his notion of a simple substance: “there is nothing besides perceptions
and their changes to be found in the simple substance. Additionally, it is in these alone
that all the internal activities of the simple substance can consist” [65]. Leibniz further
introduced the notion of “appetitions” to describe the monad’s capacity to bring about
changes in its (internal) state. In the theory of conscious agents, this is mirrored by the
decisions an agent could take. Leibniz believed that any monad perceptually mirrors the
whole universe. We slightly adapt Leibniz view here by introducing actions an agent could
take on the world. In turn, the world could “perceive” the agent, analogously as the agent
perceives its world.
Informally, as shown in Figure 1, a conscious agent has a set of possible experiences.
It also has a set of possible actions. It is embedded in a world, which we assume to
be a network of conscious agents (the thesis of “conscious realism”, [60]). Based on its
current experience, the conscious agent decides what action to take. It then acts to affect
the experiences of conscious agents in the network. It then perceives a new experience,
influenced by the network.

Figure 1. Informal picture of a conscious agent, who acts on a network of other agents, based on
its experiences.

We formalize these posits using (1) measurable spaces for conscious experiences and
(2) Markovian kernels for probabilistic relations among conscious experiences. Recall that
a measurable space, ( X, X ), specifies the elementary outcomes, X, and possible events, X ,
for a probabilistic experiment. If, for instance, the experiment is one toss of a coin, then
the elementary outcomes are X = { H, T }, where H denotes “heads” and T denotes “tails”.
The possible events are X = {{ H }, { T }, X, ∅}. The set of events, X , is a σ-algebra: in
standard probability theory it is closed under complement and countable union, and thus
also under countable intersection. A set of elementary outcomes can have more than one
σ-algebra. In our coin-toss example, for instance, one could let X = { X, ∅} (either we
observe anything or nothing at all), but not X = {{ H }, ∅} (not closed under complement)
or X = {{ H }, { T }, ∅} (not closed under countable union).
With these tools, we can construct a precise definition of conscious agent. A conscious
agent, C, has a set of potential conscious experiences, X, that form a measurable space. It has
a measurable space of potential actions, G, that it may take in response to its experiences.
Its actions influence the experiences of a network of conscious agents. We assume the
network as a whole includes the agent in question and has a well-defined set W of states,
which itself forms a measurable space. In the following, we will refer to this set W, of states
of the world network, simply as “the world”.
Entropy 2023, 25, 129 8 of 40

Additionally, recall that a Markovian kernel can be represented, in the finite case, as a
matrix in which (1) all entries are non-negative and (2) the entries in each row sum to 1.
For instance, a kernel, K, relating two experiences to three actions would be a 2 × 3 matrix,
such as  
.1 .3 .6
K= . (5)
.4 .4 .2
When agent C has a particular conscious experience x ∈ X, it chooses an action g ∈ G. This
choice is described by a Markovian kernel, D, the “decision kernel”. When a particular
action g ∈ G is taken, it influences the experiences of the agents in the network W. This
influence is described by a Markovian kernel, A, the “action kernel”. The network, W,
of conscious agents in turn influences the conscious experiences of C via a Markovian
kernel, P, the “perception kernel”.
Formally, a conscious agent, C, is a 6-tuple:

C = (( X, X ), ( G, G), (W, W ), P, D, A), (6)

where X , G , W are σ-algebras on the sets X, G, and W, respectively, so that ( X, X ), ( G, G),

(W, W ) are measurable spaces, and

P : W × X → [0, 1], (7)

D : X × G → [0, 1], (8)
A : G × W → [0, 1], (9)

are Markovian kernels (cf. Figure 2).

Figure 2. Conscious agent diagram.

One can think of the experience space, ( X, X ), as pointing to an aware subject, which is
aware whether or not it is having a specific experience. When there is no specific experience,
then the aware subject enjoys awareness without content.
The set X points to the potential of this aware subject in having specific experiences.
How this potential is actualized in specific experiences is not specified by the current theory
of conscious agents. This is an important limitation to be addressed in future versions of
the theory.
The σ-algebra X points to the conceptual representation of this aware subject about its
own potential for having specific experiences. Since the σ-algebra X can be much smaller
than the power set, P ( X ), the conceptual representation of the aware subject about its own
potential can be simpler, even infinitely simpler, than its actual potential.
This conceptual representation of the aware subject need not be a self, but it may be
among the building blocks of a self. The basic definition of a conscious agent does not
include the notion of a self. We propose that a self must be constructed by networks of
interacting conscious agents.
The set G points to the potential of the aware subject for acting. However, G is not
necessarily experienced directly by the aware subject, because its experiences are limited to
the set X.
Entropy 2023, 25, 129 9 of 40

For notational simplicity, we will not always explicitly mention the σ-algebras, writing,
e.g., X to mean the pair ( X, X ) whenever obvious from context.
It is straightforward to show that networks of conscious agents are computationally
universal. Anything that can be computed by neural networks or universal Turing machines
can be computed by networks of conscious agents [67,68]. Thus, we expect to build
networks of conscious agents that can construct a (simplified) model of themselves. Indeed,
we expect to build networks for most items in the laundry list that started this section,
including learning, memory, problem solving, intelligence, a self, free will, attention,
combinations of qualia, combinations of subjects, morality, levels of awareness, altered
states of consciousness, semantics, understanding, comprehension, the notion of a physical
object, spacetime, quantum theory, and, finally, the relationship between physics and
consciousness (including the hard problem of consciousness).
In the next Section 5, we look at one specific problem, namely the combination problem
of consciousness. We have chosen this problem because it presents itself as arguably the
most difficult problem for any theory in which consciousness is fundamental. We then
move on to discussing the problem of how to conceive of physical objects within the theory
of conscious agents (Section 6).
Those problems, we argue, are deeply related. A theory of combination (and “fusion”
as we will introduce and discuss) defines, precisely, the basic ingredients we need to make
sense of the notion of “physical objects in spacetime”. Our experience is immersive. We feel
that we are inside spacetime and inside our bodies (which is also an object in spacetime).
This feeling is not captured by the user interface metaphor. It is better captured if we switch
to a virtual reality metaphor: we each wear a VR headset that presents us as an avatar
immersed in a spacetime and interacting with objects. This allegory of the headset is just
Plato’s allegory of the cave, with technology updated from fire to VR.
Before we turn to the combination problem, we briefly address a couple confusions
that often arise about this program.
The interface theory of perception and the theory of conscious agents are distinct
theories [69]. One might accept one and not the other (or reject both). However, if we
accept both theories, they together entail that the distinction we make between conscious
and unconscious objects is not principled, but is instead an artifact of the limitations of our
perceptual interface. We say that a human being is conscious but a rock is not. However, if
our perceptions of spacetime and objects are just an interface to a network of conscious
agents beyond spacetime, then we are always interacting with conscious agents, no matter
what object we see. In the case of a human being, when I look at a person’s face and
body, I get some insight into their consciousness—are they happy, sad, fearful, relaxed,
interested, distracted, and so on. When I look at a rock I get, say, hardness and brittleness,
but minimal insights into consciousness. That doesn’t mean that I am not interacting with
conscious agents. I always am. It’s just that my interface is giving me little insight into
those agents. This is no surprise. The whole point of a user interface is to present a much
simpler description, with much of the details of reality deleted, and to provide an adaptive
representation of functional aspects of the interaction, so that one is not overwhelmed with
too much information.
This entails, a fortiori, that the distinction we make between living and non-living
objects is not principled. It too is merely an artifact of the limitations of our perceptual
interface, and not an insight into the nature of reality. It should be no surprise, then, that
attempts by scientists and philosophers to give a principled definition of life have so far
failed. As the Stanford Encyclopedia of Philosophy says, in its entry on Life, “The literature
on the definition of life is vast, repetitive, and utterly inconclusive. Philosophers have
disagreed as to the ultimate source of the lack of consensus, citing unstated assumptions in
either the definer’s approach or the question itself” [70].
The combination of the interface theory of perception with the theory of conscious
agents has caused some to worry that they might be self-refuting. As the philosopher Philip
Goff put it, “If we ought to doubt the testimony of our senses, then we ought similarly
Entropy 2023, 25, 129 10 of 40

to doubt the testimony of our evolved capacity for forming judgements concerning the
mental states of others. We are hardwired to judge the emotions of others on the basis of
their behavior and facial expressions. However, if this hardwired capacity was evolved for
survival rather than truth, and if this is sufficient for us to reject the deliverances of our
sensory perception, then we ought likewise to reject the deliverances of our judgements
about other minds. We ought to think others are zombies, or at least have no faith in our
judgments that crying indicates sadness. I consider this a reductio ad absurdum of the
‘Fitness Beats Truth’ argument.” Similar reductio arguments are offered in [71].
The resolution of this apparent reductio ad absurdum comes from recognizing the two
separate steps in our argument. The first step is a theorem about the theory of evolution
by natural selection. That theory, for better or worse, entails that sensory systems are
user interfaces, not windows on reality. However, that theory cannot answer the question,
“Interfaces to what? What is the reality beyond our user interfaces?” To answer this
question, we must take a new step, entirely separate from and beyond the theory of
evolution by natural selection. We propose that the reality beyond our user interfaces is a
reality composed of conscious agents.
On our proposal, our senses are interfaces to a network of conscious agents. This
proposal goes beyond the confines of evolution by natural selection. In the definition
of a conscious agent in Equation (6), we have introduced the σ-algebra X to stand for
the conceptual representation of an aware subject. It is here where ITP demands caution.
Almost certainly, X will not mirror any true structure of reality, such as when we sort what
we experience into the categories of living/non-living. However, this does not mean that
we do not undergo experiences X. In fact, we are having experiences every moment. Nor
does it mean that experience is not being expressed in behavior. In fact, we do propose that
X leads to G in our formalism. Hence, we should not expect others to be zombies or to not
experience sadness when crying, as long as they (roughly) share our evolutionary history.
It would probably be detrimental to our survival if we believed that others are zombies.
In short, it is a theorem that our perceptions of objects in spacetime is not a veridical
perception of real objects in real spacetime; it is a user interface to a realm that is utterly
unlike objects in spacetime. That realm happens to be a realm of conscious agents. It is
no problem then to conclude that our interface gives real access to some of the conscious
experiences of these agents.

5. Combine and Fuse

Can conscious subjects combine to form new conscious subjects? Can conscious
experiences combine to form new conscious experiences? These are central questions of
the combination problem of consciousness, famously raised in 1895 by William James in The
Principles of Psychology:
“Where the elemental units are supposed to be feelings, the case is in no wise al-
tered. Take a hundred of them, shuffle them and pack them as close together as you can
(whatever that may mean); still each remains the same feeling it always was, shut in its
own skin, windowless, ignorant of what the other feelings are and mean. There would
be a hundred-and-first feeling there, if, when a group or series of such feelings were set
up, a consciousness belonging to the group as such should emerge. Additionally, this
101st feeling would be a totally new fact; the 100 original feelings might, by a curious
physical law, be a signal for its creation, when they came together; but they would have no
substantial identity with it, nor it with them, and one could never deduce the one from the
others, or (in any intelligible sense) say that they evolved it.” Further:
“Take a sentence of a dozen words, and take twelve men and tell to each one word.
Then stand the men in a row or jam them in a bunch, and let each think of his word as
intently as he will; nowhere will there be a consciousness of the whole sentence. We talk of
the ‘spirit of the age,’ and the ‘sentiment of the people,’ and in various ways we hypostatize
‘public opinion’. However, we know this to be symbolic speech, and never dream that the
spirit, opinion, sentiment, etc., constitute a consciousness other than, and additional to, that
Entropy 2023, 25, 129 11 of 40

of the several individuals whom the words ‘age,’ ‘people,’ or ‘public’ denote. The private
minds do not agglomerate into a higher compound mind”.
This combination problem has recently received substantial philosophical interest,
propelled largely by the development of panpsychist theories of consciousness [20,72–80].
Seager describes the combination problem as “the problem of explaining how the myriad
elements of ‘atomic consciousness’ can be combined into a new, complex and rich con-
sciousness such as that we possess” [72]. We use the theory of conscious agents to propose
two mathematical approaches to the combination problem: combination and fusion.
To understand how conscious agents and qualia combine to create new agents and
qualia, we study interactions among agents and how these interactions affect their experi-
ences. To this end, we compose the kernels P, D, and A of an agent to create its “qualia
kernel”, Q from X to itself (Using the definitions in Equations (7)–(9), the combined
R kernel
Q = DAP is given, for x ∈ X and a measurable set B ∈ X , by Q( x, B) = g∈G D ( x, dg)
R

w∈W A ( g, dw ) P ( w, B ) ):
Q = DAP : X × X → [0, 1]. (10)
Q describes the sequential experiencing of an agent, without reference to its other internal
aspects, such as its decisions or actions, or to the nature of the network it is interacting with.
For kernels that have a matrix representation, composition is simply matrix multipli-
cation. In this case, Q is a matrix that maps X to X.
The qualia kernel, Q, thus expresses the relation that conscious experience has to
itself. This might also account for the elusive “what-it-is-likeness” of conscious experience.
For the simplest agent, X has just one point x ∈ X, i.e., just one conscious experience.
The qualia kernel would then give a Dirac measure on this point:

Q( x, ·) = δx (·). (11)

We call an agent with n distinct qualia an n-agent. Qualia can differ from one n-agent
to another. We posit a large, perhaps countably infinite, set of 1-agents, each with a
unique quale.
Now consider two 1-agents, C1 and C2 , with qualia kernels

Q1 : X1 × X1 → [0, 1], (12)

and
Q2 : X2 × X2 → [0, 1], (13)
respectively. Let X1 = { x1 } and X2 = { x2 }, where xi denotes an experience.
If we take x1 to be the quale red and x2 to be green, then we can depict the action of
their qualia kernels as shown in Figure 3.

Figure 3. The qualia kernels of two 1-agents, one that experiences red and one that experiences green.
In this diagram, the agents do not interact, and their qualia do not combine.

How might these agents interact and combine? To study this, consider the qualia
kernel, Q, for the possible experiences of a 2-agent. This kernel is a 2 × 2 matrix whose
entries are the following conditional probabilities:
 
p ( x1 | x1 ) p ( x2 | x1 )
Q ( x1 , x2 ) = , (14)
p ( x1 | x2 ) p ( x2 | x2 )
Entropy 2023, 25, 129 12 of 40

We interpret the entry p( x1 | x2 ) in this matrix as the probability that the next experience of
the pair of agents will be red given that the current experience is green.
Suppose we have two agents

C1 = ( X1 , G1 , W1 , P1 , D1 , A1 ); C2 = ( X2 , G2 , W2 , P2 , D2 , A2 ). (15)

Observe that the 6-tuple

C1 × C2 = ( X1 × X2 , G1 × G2 , W1 × W2 , P1 ⊗ P2 , D1 ⊗ D2 , A1 ⊗ A2 ), (16)

satisfies the definition of being a single conscious agent (cf. Equation (6) and [60]). We are
assuming that the σ-algebras on the sets X1 × X2 , etc., are the product algebras. The Marko-
vian kernel P1 ⊗ P2 : W1 × W2 → X1 × X2 is defined by P1 ⊗ P2 (w1 , w2 ; dx1 , dx2 ) =
P1 (w1 ; dx1 ) P2 (w2 ; dx2 ). D1 ⊗ D2 : X1 × X2 → G1 × G2 and A1 ⊗ A2 : G1 × G2 → W1 × W2
are defined similarly, and, thus, must be considered to be a single agent by the theory of
conscious agents.
Hence, this juxtaposition of two simple agents leads to a more complex agent. The new
agent then has two potential experiences, red and green. However, this new agent has a
very simple Q kernel, namely Q1 ⊗ Q2 , which has a block-diagonal form. In particular,
the off-diagonal terms, or cross terms, in Equation (14) are both zero, and we can think of
agents C1 and C2 as “non-interacting”.
For now, we take it as an axiom that two agents can combine with arbitrary cross terms
(as long as they fulfil the Markov criterion), as expressed in Equation (14). We conjecture
that this axiom is not needed, and that for every legal set of cross terms, a network W can
be constructed that would generate those cross terms.
If we let

x = p ( x2 | x1 ), (17)
y = p ( x1 | x2 ), , (18)

then we can simplify (14) to

 
1−x x
Q( x, y) = , (19)
y 1−y

because each row needs to sum to 1 to satisfy the requirement of being Markovian.
If x = y = 0, then this matrix is the identity matrix
 
1 0
Q( x, y) = . (20)
0 1

In this special case, where cross terms are 0, the two agents do not “interact.” Never-
theless, as we demonstrated in Equation (16), a pair of such non-interacting 1-bit agents,
C1 , C2 , can be thought of as a single (combined) 2-bit agent. We will now show that two
1-agents cannot only be combined into a more complex 2-agent, but they can also “fuse”
into a simpler 2-agent.
If either cross term is greater than 0, then the agents interact in a non-trivial dynamic.
The maximum value of either cross term is 1. Thus, the set of kernels can be represented
by points ( x, y) in the unit square bounded by {(0, 0), (0, 1), (1, 0), (1, 1)}. This unit square,
which contains all kernels with 2 states, we call the Markov polytope, M2 . The collection of
all kernels with n states is a unit cube in n-dimensional real space and is here called the
Markov polytope Mn .
Entropy 2023, 25, 129 13 of 40

The kernel Q describes a single interaction of the agents. To describe two consecutive
interactions, we need the kernel Q2 , where

(1 − x )2 + xy x (2 − x − y) 1 − x0 x0
   
2
Q = QQ = = (21)
y(2 − x − y) (1 − y)2 + xy y0 1 − y0

The difference quotient of these two kernels is the (in general non-Markovian) matrix
Q2 − Q. We restrict attention to the off-diagonal terms of Q2 − Q, since they fix the diagonal
terms. The off-diagonal terms define a discrete “kernel derivative,” dQ/dτ, with respect to a
discrete-step parameter τ. This derivative, at a point ( x, y) in the Markov polytope, M2 , is

dQ dx dy
=( , ) = ( x (1 − x − y), y(1 − x − y)). (22)
dτ dτ dτ
The derivatives show the direction of “kernel flows” through M2 , as illustrated by the
vector field in Figure 4. The “flow” of a Markovian kernel Q is the Markovian kernel
P = limn→∞ Qn .

Figure 4. The flow of Markovian kernels on M2 , the Markov polytope of two states. This polytope is
a unit square with coordinates ( x, y), where x = p( x2 | x1 ) and y = p( x1 | x2 ). Arrowheads point in the
direction of the local flow. Arrow length and color indicate the speed of flow.

The lower left corner of M2 , the point (0, 0) corresponding to the identity matrix, is
the single source for this vector field. The upper right corner, the point (1, 1) corresponding
to the NOT operator, does not flow but represents instead a periodic kernel with period 2.
There is a line of sinks, the line y = 1 − x, depicted by the diagonal red line. Along this line,
the kernels    
1−x x 1−x x
Q( x, y) = = . (23)
y 1−y 1−x x
Thus, when kernels reach this line, they drop from rank 2 to rank 1. As we discuss later,
this signals the fusion of agents and creation of new qualia.
Each kernel Q( x, y) has a unique combination of colors that it leaves invariant. We
can represent a combination of colors by a probability measure
 
µ = α 1 − α , 0 ≤ α ≤ 1, (24)
Entropy 2023, 25, 129 14 of 40

where α is the proportion of red and 1 − α is the proportion of green. The measure µ is
invariant for the kernel Q if
µQ = µ. (25)
We can use this to find all kernels Q that have the same invariant measure.
   
  1−x x α
α 1−α = . (26)
y 1−y 1−α

Solving this we find the lines

α
y= x. (27)
1−α
Thus, for each value of α there is a line of kernels that follows the flow of the vector field.
Two examples are shown in Figure 5, depicted as green lines.

Figure 5. Invariant measures on the Markov polytope M2 . Kernels with the same invariant measure
lie on a straight line passing through the origin. Two examples are shown as green lines. The lower
line corresponds to α = 31 . The upper line corresponds to α = 23 .

It is helpful to visualize the invariant measure associated to each kernel in the Markov
polytope M2 . This is completed in Figure 6, which depicts the mixture of red and green
dictated by the invariant measure for each kernel in M2 . A stationary kernel is any kernel
Q that is idempotent, i.e., that satisfies QQ = Q. These kernels correspond to the line of
sinks shown in Figure 4. Figure 6 shows the relationship of these kernels to the mix of red
and green in the invariant measures.
The stationary kernels lie on the line y = 1 − x. Along this line, the kernels can be
written as:    
1−x x 1−x x
Q( x, y) = = . (28)
y 1−y 1−x x
Thus, when kernels reach this line of stationary kernels, they drop from rank 2 to rank 1,
and are a function of just one parameter, say

α = x = 1−y (29)
Entropy 2023, 25, 129 15 of 40

This signals the fusion of agents. The two agents, Q1 and Q2 first combine to form a two-
parameter family, Q( x, y). All combined agents have the original two qualia: red and green.
Now, with the drop in rank of the qualia kernel, the combined agents fuse to single agents
with just one quale. There is a one-parameter family, indeed a unit 1-simplex, of new fused
agents, Q(α), each with its own new quale, as illustrated in Figure 7.

Figure 6. Invariant measures and stationary kernels in the Markov polytope M2 . The color at
each point ( x, y) indicates the mix of red and green in the invariant measure for the kernel Q( x, y)
associated to that point by Equation (26). The blue diagonal line depicts the stationary kernels Q,
satisfying QQ = Q, except for the identity matrix, which is trivially stationary.

Figure 7. The unit 1-simplex of fusions for two conscious agents.

This pattern continues. If n conscious agents interact, their possible fusions form
a unit (n − 1)-simplex, the fusion simplex Fn , corresponding to the possible stationary
kernels. For instance, if we have three agents, Q1 , Q2 , Q3 , with qualia red, green, and blue,
respectively, then their possible fusions form a unit 2-simplex, illustrated in Figure 8.
Any collection of agents is itself an agent. Thus, there is ultimately one agent. A
similar stance has been advocated by the physicist Erwin Schrödinger in his essay “Mind
and Matter” [81]. Yet, the exploration of consciousness through its possible combinations
and fusions appears to be endless. For this reason, a theory of consciousness cannot start
with a theory of the “Ultimate One Consciousness.” However, if we start with a countable
infinity, ℵ0 , of agents, then the number of possible combinations, 2n , and their fusions is a
Entropy 2023, 25, 129 16 of 40

larger infinity, ℵ1 , and the number of possible new combinations of these combinations is
yet a larger infinity, ℵ2 and so on through Cantor’s hierarchy [82].
Viewed this way, the smaller units are perhaps nothing but projections of the One. If we
could grasp the deep connection between these units and the One, we might at once resolve
Schrödinger’s “arithmetic paradox” (the world is one, but the subject seems to be many, [81])
and also see the theory of conscious agents as a formal model of Leibniz’s monadology to
document the “pre-established harmony” [65] between monads. However, for now, we
must start with humble beginnings, and crawl up Cantor’s hierarchy. At each step in this
process, our theory itself points to its inherent limitations, and to the infinite work undone.
This antidote to dogmatism offers infinite job security for students of consciousness.

Figure 8. The unit 2-simplex of fusions for three conscious agents.

In the meantime, a next step is to study the possible combinations and fusions of three
agents. We begin this study in the Appendix A to this paper.

6. Spacetime and Decorated Permutations

This section and the next propose how to project the dynamics of conscious agents
down to spacetime, using structures called “decorated permutations.” We start by develop-
ing some necessary background.
How is consciousness related to spacetime and physical objects? Physicalist theo-
ries of consciousness usually assume spacetime and physical objects to be fundamental,
and consciousness to be somehow emergent. By contrast, many panpsychist theories take
consciousness to be the intrinsic nature of physical objects, and some seem to suggest that
spacetime is fundamental. However, physics itself and evolution entail that spacetime is
not fundamental, ruling out these theories.
The theory of conscious agents starts with a dynamics of agents that is, by hypothesis,
outside of spacetime. So this theory must explain how spacetime and objects arise entirely
from the dynamics of agents. This is a colossal project.
As Goff puts it, “It’s been hard enough to get the equations of physics we already
have. Coming up with a whole new level of mathematical structure underneath that, which
yields precisely the same predictions, is rather a large challenge. Moreover, I can’t see the
motivation for taking on that challenge” [83].
Entropy 2023, 25, 129 17 of 40

“Physics” here is usually taken to refer to the physics of spacetime and objects, namely
quantum field theory and general relativity. However, we do not locate the network of
conscious agents inside spacetime precisely because physics and evolution tell us that these
structures are not fundamental, and because physics itself has found new structures beyond
spacetime, such as amplituhedra and their associated decorated permutations. So instead
of merely re-describing conventional physics with conscious agents, we aim to show how
decorated permutations, and other structures that physicists have found beyond spacetime,
arise as a projection of a deeper theory of conscious agents.
An early hint of structures beyond spacetime came in 1986. Physicists study elementary
particles by smashing them together at high energies and seeing what sprays out. Quantum
field theory provides formulas, called scattering amplitudes, for computing the probabilities
of various outcomes. These formulas derive from Feynman diagrams that model scattering
as quantum processes occurring in spacetime, that is, as satisfying locality and unitarity.
However, the formulas are complex. The formula for two gluons smashing to produce four
gluons requires hundreds of pages of algebra. Then, in 1986, two mathematicians, Parke
and Taylor, discovered a formula for gluon scattering that required only one term [84]. It did
not model scattering as a process in spacetime, but pointed to a realm beyond spacetime.
The Parke–Taylor formula was soon followed by others that magically simplified
the mathematics and pointed to a world beyond spacetime. Then in 2013, many of these
results were unified into a single structure, a geometric object beyond spacetime, called
the “amplituhedron” [22]. The amplitude of a particular scattering process is obtained by
computing the volume of its corresponding amplituhedron. The amplituhedra for various
scattering processes are faces of an infinite-dimensional master amplituhedron.
The amplituhedron is obtained via a linear mapping from a positive Grassman-
nian, [22,24,85]. Recall that the real Grassmannian G (k, n) is the space of all k-dimensional
subspaces in an n-dimensional vector space. An element of G (k, n) can be represented,
non-canonically, by any of its bases and therefore by a full-rank k × n matrix C. The
Plücker coordinates for C are the determinants of all k × k minors of C. The positive Grass-
mannian, G + (n, k), is the subset of the real Grassmannian where all Plücker coordinates
are non-negative, i.e., all subspaces have non-negative slope (see [85,86] for an informal
discussion).
Remarkably, the invariant physical content of the positive Grassmanian is combi-
natorial, described by “decorated permutations”. Recall that a permutation s on the set
n := {1, . . . , n} is a bijection from the set to itself. We denote such a permutation by
s = [s(1), s(2), . . . , s(n)].
An ordinary permutation s, say, can be decorated to yield mappings σ : n → 2n =
{1, . . . , 2n} in the following ways: if s( a) > a, set σ( a) = s( a). If s( a) < a, set σ( a) =
s( a) + n. If s( a) = a, set σ ( a) to be either a or a + n. This suggests the

Definition 1. A decorated permutation is a mapping σ : n → 2n such that, for any a,

a ≤ σ( a) ≤ a + n (30)

and
σ ( a) mod(n) (31)
is an ordinary permutation.

Notice that a decorated permutation thus defined is an injective mapping and is indeed
a decoration of the (unique) ordinary permutation σ ( a) mod(n). Moreover, we see that if
an ordinary permutation has exactly k fixed points, there are 2k decorated permutations
corresponding to it. In particular, there are 2n decorations of the identity.
Decorated permutations can be represented by diagrams that mathematicians call
“plabic graphs” and physicists call “on-shell diagrams” [22,87]. For instance, for the dec-
orated permutation [3, 4, 5, 6], an on-shell diagram is shown in Figure 9. The numbers to
Entropy 2023, 25, 129 18 of 40

be permuted are arranged around a circle clockwise. To read off the permutation from
the diagram, follow the line from a number into the diagram. If the line hits a white dot,
turn left. If it hits a black dot, turn right. For instance, the line inward from 1 hits a white
dot, so we turn left and hit a black dot, so we turn right and hit a white dot, so we turn
left and arrive at 3. Thus, 1 is permuted to 3. Similarly, the line inward from 2 hits a black
dot, so we turn right and hit a white dot, so we turn left and hit a black dot, so we turn
right and arrive at 4. Thus, 2 is permuted to 4. If we start with 3, we will end up at 1.
However, observe that, for the corresponding decorated permutation, we now have to add
4, because a = 3 > σ ( a) = 1, so we end up with 3 → 5. Analogously, we find 4 → 6.

Figure 9. On-shell diagram for the decorated permutation σ = [3, 4, 5, 6].

In the other direction, for a given decorated permutation, there are many correspond-
ing on-shell diagrams, of varying complexity. The diagram of relevance to physicists
computing scattering amplitudes is a “reduced” diagram. It can be obtained by decom-
posing the decorated permutation into a minimal sequence of “adjacent” transpositions
by the following simple algorithm, which can then be used to generate a corresponding
on-shell diagram that is reduced [22]. For any decorated permutation σ that is not already
a decoration of the identity permutation, find the lexicographically first pair of numbers
in {1, . . . , n}, such that (1) a < c, (2) σ ( a) < σ (c), and (3) the numbers between a and c
are all mapped by σ to themselves, or to themselves plus n. Let ( ac) denote the (ordinary)
transposition of a and c: if it satisfies the three conditions above is called an adjacent transpo-
sition. Then we can decompose σ as σ = ( ac) ◦ σ0 . Repeat this process on the next available
adjacent transposition. We can see that once all adjacent transpositions are exhausted, we
will be left with the remaining permutation being a decoration of the identity. The final step
from the product of adjacent transpositions to the minimal on-shell diagram is as follows:
list the numbers 1, . . . , n above a rectangle. For the first transposition, say ( ac), drop a line
from a ending in a white circle and draw a horizontal line to a line dropped from c, ending
in a black circle. Notice that, by our rules for the circles, one traces a path from a down to
the white circle and turns left. Now there are two possibilities: (i) if c appears in a later
transposition, extend the line downwards from the black circle (upon reaching the black
circle one turns right): to a white circle if c is the lower number in the new transposition,
a black circle if the higher number; (ii) if the state c is never again represented in one of the
transpositions to follow, draw a line from the black circle back up to c. Continue until all
transpositions are accounted for. Finally, if a number b is in no transposition, then σ(b) = b
or σ (b) = b + n. In the first instance drop a line to a white circle and in the second to a
black one. Finally, we simplify the diagram by eliminating bipartite circles.
Entropy 2023, 25, 129 19 of 40

An example of this process is shown in Figure 10, where we see that the decorated
permutation [3, 5, 4, 6, 7] is the decorated version of the ordinary product of adjacent trans-
positions: (12)(23)(24)(12)(25).
Note that we can recover all consistent decorated permutations from the diagram,
by following these rules: starting at a state i, at any bipartite node, ignore it and continue
straight through. At any tripartite node, turn left if the node is white and right if black.
This path will end at some state. Then σ(i ) equals this state if it is greater than i; if it is less
than i, add n to obtain σ (i ); and if σ (i ) = i we can choose i or i + n.

Figure 10. Construction of a reduced on-shell diagram in the middle corresponding to the decorated
permutation [3, 5, 4, 6, 7]. On the right is a minimal version.

This description of structures beyond spacetime is just an overview, and for brevity
omits important structures, such as the associahedron and cosmological polytope [23,88].
If the theory of conscious agents is to explain how spacetime and objects arise entirely from
the dynamics of agents, their combination and fusion, then it should somehow connect
with these new structures beyond spacetime. This would allow the new physics to do some
of the heavy lifting in building a bridge from consciousness to spacetime. Two key insights
from physics guide our proposed connection.
One key insight is this: the deepest structure beyond spacetime that distills physics
is the decorated permutation. From decorated permutations, one can construct reduced
on-shell diagrams. Differential forms on these on-shell diagrams give rise to the scattering
amplitudes. (Without supersymmetry one also needs helicities, or masses and spins [24]).
A second key insight is that any on-shell diagram arises from combining diagrams
containing single three-legged black or white dots. They are the only diagrams for three-
particle interactions, which are sufficient for computing all interactions [22].
These two insights give a clear target for the theory of conscious agents: its Markov-
chain dynamics must map to decorated permutations and spins. With that map, we can
propose a precise correspondence between (1) the Markov polytopes that describe all
possible agent dynamics and (2) the on-shell diagrams that generate scattering amplitudes.
So, what is the map from Markov chains to decorated permutations? There is recent
work relating Markov chains, positive Grassmannians, and amplituhedra [85]. How-
ever, the map from Markov chains to decorated permutations has been an open problem.

7. Correspondence between Agent Dynamics and Physical Particles

In a Markov chain with kernel Q, we say that state a communicates with state b if
there is a positive probability that the chain starting at a reaches b in finite time: we
have Q j ( a, b) > 0 for some natural number j ≥ 0. Given a state a ∈ n̄ := {1, . . . , n}, its
communicating class [ a] consists of all states b such that a and b communicate with each
other. Mutual communication is an equivalence relation, so the communicating classes
of a given kernel partition the state space. Since the state a communicates with itself (set
j = 0 above), the singleton { a} is always a subset of the communicating class [ a]. When
[ a] = { a}, i.e., when Q( a, a) = 1, we say that a is absorbing: the chain, starting at a, stays at
a. More generally, if the chain returns to a infinitely often with positive probability, we say
that a is recurrent. If a state is not recurrent, it is transient: the probability of its returning to a
infinitely often is zero, and this is equivalent to having ∑∞ j=0 Q ( a, a ) < ∞ (cf. Theorem 1.5.3
j

of [89]).
Entropy 2023, 25, 129 20 of 40

Definition 2. Markov Decorated Permutations. Given a Markov kernel on n̄, we define the
decorated permutation σ : n → 2n as follows:
• If a is transient, set σ ( a) = a.
• If a is recurrent, let σ( a) be the first element b > a of {1, . . . , 2n}, such that the sequence
( a, a + 1 . . . , b) between a and b contains the communicating class [ a] (recall that a number
c > n represents the state c mod(n)).
In both instances, a ≤ σ( a) ≤ a + n; moreover σ ( a) mod(n) is bijective, so we indeed have a
decorated permutation. Notice, also, that when a is absorbing, we have σ ( a) = a + n.

For instance, a Markov chain on 9 states in which the cycles (including cycles of
period 1) are (158), (2), (34), (6), and (79) would be represented by the decorated permu-
tation [8, 11, 4, 12, 10, 15, 9, 14, 16]. If state 2 were transient, then the permutation would be
[8, 2, 4, 12, 10, 15, 9, 14, 16]. We note, in passing, that the assignment of decorated permuta-
tions to Markov chains is easily generalized to an assignment of decorated permutations to
arbitrary graphs, a generalization that may prove useful for the analysis of networks.

Definition 3. Decorated Permutations of Arbitrary Graphs. Given an arbitrary graph whose

set of nodes is n̄, define the decorated permutation for each a ∈ n̄ as follows: (1) if a has no links,
then set σ ( a) = a; (2) if a only links with itself, then set σ ( a) = a + n; and (3) otherwise, let σ ( a)
be the first element b > a of (1, . . . , 2n), such that ( a, . . . , b) contains all nodes in the strongly
connected component of the graph containing a. (Recall that a number c > n represents the node
c mod(n), and that a strongly connected graph has a path from each node to every other node).

The Markov polytope Mn is a cell complex. The cell Sσ is the set of all Markov kernels
K ∈ Mn which give rise to the decorated permutation σ. For instance, M2 has four cells,
as shown in Figure 11. The identity kernel, depicted by a red disk, is a cell by itself, with
σ = [3, 4], a decoration of the identity indicating that each state maps to itself. The NOT-
kernel, depicted by a light blue disk, and the entire interior of M2 is a cell, with σ = [2, 3],
indicating a communicating class of size 2. The segment y = 0 minus the origin, depicted
in yellow, is a cell, with σ = [3, 2]. The segment x = 0 minus the origin, depicted in dark
blue, is a cell, with σ = [1, 4]. Figure 11 also shows on-shell diagrams associated to each
cell in M2 .
Decorated permutations code for subspaces of the positive Grassmannian as follows.
For a k ≤ n, let C be a k × n matrix of full rank k, representing an element of the positive
Grassmannian G ≥0 (k, n) (Recall that the Grassmannian G (k, n) is the set of k-dimensional
subspaces of Rn . Any such subspace may be represented by such a matrix C, whose k rows
are a basis of that subspace: at least one k × k minor is non-zero. The positive Grassmannian
G ≥0 (k, n) consists of those subspaces that can be represented by a matrix C all of whose
k × k minors are non-zero and of the same sign. All representative matrices of such a
subspace will then share this property). Repeat the n columns of C, in order, to produce a
k × 2n matrix C ∗ . The decorated permutation σC associated to C assigns to each column
number a ∈ {1, . . . , n} the value σC ( a) > a as follows: if c a 6= ~0, σC ( a) is the first column b
of the expanded matrix C ∗ , such that c a is in the span of {c a+1 , . . . , cb }. If, however, c a = ~0,
define cσ ( a) = a.
Clearly, the set of full rank k × n matrices with a given decorated permutation corre-
sponds to a unique subspace of Rn and vice versa. Thus, a decorated permutation that
codes for a communicating class of size l of a Markov chain also codes for a subspace
of dimension l − 1 in the positive Grassmannian. Since communicating classes dictate
the asymptotic behavior of Markov chains, this correspondence between communicating
classes and subspaces may underlie the unexpected finding that the stationary distribu-
tion of the asymmetric simple exclusion process depends on the combinatorics of cells
of the positive Grassmannian [90], and that regular soliton solutions of the KP equation,
governing interaction patterns of shallow waves, depend on the cell of the positive Grass-
Entropy 2023, 25, 129 21 of 40

mannian that the starting point lies in [91]. In turn, the positive Grassmanian maps into an
amplituhedron, on whose faces the computations of physical interactions are then made.
This has guided our search for a correspondence between the dynamics of conscious
agents and the dynamics of scattering in spacetime.

Figure 11. The cell complex for M2 (colored disks, edges, and interior of the square) and the
associated on-shell diagrams (black and white dots).

Recalling Figure 4, notice that, with the exception of the NOT corner, the cells of
Figure 11 are precisely the segmentation of Mn into its different possible flow patterns.
This suggests that we think of the map from Mn to its cells, which are communicating
classes, as a projection via its dynamics to their associated decorated permutations. Physics,
as we have noted above, finds fundamental significance in decorated permutations: they
lead from their on-shell diagrams to faces of the amplituhedron and the differential forms
on those faces that yield the scattering amplitudes. These last are a complete description
of any physical energetic exchange at the micro-level. Physical scattering events are then
explained in terms of “particles” in spacetime, so these ideas suggest the conjecture that this
procession, or “projection,” from at least some of the cells of Mn , have meaning as particle
interactions in physics: if so, let us call such projections physical. That is, we conjecture an
agent–particle correspondence: a particle (in spacetime) is an aspect of a physical projection of
the dynamics of a communicating class of conscious agents to a face of an amplituhedron.
The smallest non-trivial communicating class is a single conscious agent. So the
Markov polytope M j , describing all possible dynamics of the conscious j-agent, is the
smallest Markov polytope that may have projections onto the dynamics of j-particle scat-
tering in spacetime. All Markov polytopes Mk , with k > j, may also have projections onto
the dynamics of j-particle scattering.
As a specific example of this agent-particle correspondence, consider the Markov
polytope M3 , which describes all dynamics of conscious 3-agents. Appendix B presents
the 17 distinct decorated permutations for the 27 vertices of M3 . The adjacency graph for
M3 is shown in Figure 12. Appendix A presents some of the geometric structure of M3 .
In this case, M3 is the smallest Markov polytope with projections onto the two possible
on-shell diagrams for three-particle interactions shown in Figure 13. We conjecture that the
fusion simplex thereby defines two types of flows that correspond to the pair of 3-particle
Entropy 2023, 25, 129 22 of 40

amplitudes of scattering theory. It will be intriguing to see how physical properties such as
mass, momentum, and spin arise as projections from Markov polytopes.

Figure 12. The adjacency graph for the Markov polytope M3 . Vertices are numbered to correspond
with Appendix B. The grey triangle represents the total fusion simplex. Blue dots are vertices of the
partial fusion complex.

Figure 13. Flows in the Markov polytope M3 . Computer simulations indicate that, analogous to
the situation in M2 , there are two types of flow on either side of the total fusion simplex coming
from the identity (6) on one hand and the two maximal derangements (16 and 20) on the other. We
conjecture that these two types of flows are in the relation indicated (by the pointing hands) to the
pair of 3-particle amplitudes of scattering theory.
Entropy 2023, 25, 129 23 of 40

8. Discussion
Newtonian physics is a beautiful theory, with practical applications that are explored
and exploited to this day. However, it is not fundamental. In 1905, with the publication
of Einstein’s paper on special relativity, its reign ended. For those seeking a fundamental
theory, it was time to move on.
However, scientific theories have their own momentum. In 1922, when Einstein
received his Nobel Prize, the committee noted that the award was given “without taking
into account the value that will be accorded your relativity and gravitation theories after
these are confirmed in the future” [92].
Quantum field theory and Einstein’s theory of gravity are beautiful theories, with prac-
tical applications that are likely to be explored and exploited for centuries to come. How-
ever, they, and spacetime itself, are not fundamental. For those seeking a fundamental
theory, it is time to move on. Physics has indeed moved on, proposing, e.g., string the-
ory [93], loop quantum gravity [94], causal sets [95], amplituhedra [22], and cosmological
polytopes [23,88].
However, scientific theories of consciousness, and the combination problem, have
not moved on. Physicalist, dualist, and panpsychist theories assume that spacetime is
fundamental. Functionalist theories could, in principle, eschew spacetime and seek in-
stantiation elsewhere [96]. However, functionalists tacitly, and often explicitly, assume an
instantiation within spacetime. We propose that, as long as functionalists seek instantiation
within spacetime, the question why and how it feels for something to have a mind will
forever elude them.
Spacetime recidivism has consequences. No theory in the scientific study of conscious-
ness, to date, accounts for any specific conscious experience. What integrated information
must be the taste of chocolate and could not be the taste of vanilla? What orchestrated
collapse of quantum states must be the smell of coffee and could not be the smell of co-
conut? What state of a global workspace must be the sound of a harp and could not be
the sound of a flute? No answer has yet been given. This failure was foreseen by Leibniz
in his famous Mill Argument: “It must be confessed, however, that Perception, and that
which depends upon it, are inexplicable by mechanical causes, that is to say, by figures
and motions. Supposing that there was a machine whose structure produced thought,
sensation, and perception, we could conceive of it as increased in size with the same
proportions until one was able to enter into its interior, as he would into a mill. Now,
on going into it he would find only pieces working upon one another, but never would
he find anything to explain Perception. It is accordingly in the simple substance, and not
in the composite nor in a machine that the Perception is to be sought. Furthermore, there
is nothing besides perceptions and their changes to be found in the simple substance.
Additionally, it is in these alone that all the internal activities of the simple substance can
consist” [65]. We agree with Leibniz. Any theory that reduces consciousness to mechanisms
involving states, configurations, or processes of objects will fail to account for any specific
conscious experience.
Steven Pinker acknowledges this failure: “Our best science tells us that consciousness
consists of a global workspace representing our current goals, memories, and surroundings,
implemented in synchronized neural firing in fronto-parietal circuitry. However, the last
dollop in the theory—that it subjectively feels like something to be such circuitry—may
have to be stipulated as a fact about reality where explanation stops” [97]. Indeed, current
theories of consciousness do not explain conscious experience, they stipulate it.
Pinker explains why. “This should not be entirely surprising. As Ambrose Bierce
noted in The Devil’s Dictionary, the mind has nothing but itself to know itself with, and it
may never feel satisfied that it understands the deepest aspect of its own existence, its
intrinsic subjectivity” [97]. Indeed, it may be that subjective experience cannot be explained,
and must be stipulated. If so, then let us stipulate it and, with a nod to William of
Ockham, nothing else. Stipulate a dynamics of experiences and derive, rather than stipulate
spacetime and objects as a projection of the dynamics. This is the project of the theory of
Entropy 2023, 25, 129 24 of 40

conscious agents. To that end, this paper proposes a map from agent dynamics through the
amplituhedron into spacetime.
This approach agrees with Pinker’s next point. “Whatever we make of the hard
problem of consciousness, positing an immaterial soul is of no help at all. For one thing, it
tries to solve a mystery with an even bigger mystery. For another, it falsely predicts the
existence of paranormal phenomena” [97]. Indeed, positing an immaterial soul offers no
formal theory, and thus no help to a science of consciousness. It introduces a dualism,
and thus entails paranormal phenomena. The theory of conscious agents, by contrast,
is monistic and precise. Consciousness is no ghost in the machine. Instead, the laws of
physics and the special sciences are themselves a projection of the dynamics of conscious
agents. New laws may emerge from the study of conscious agents. However, these will be
formal advances with testable consequences—such as the move from classical to quantum
theorie—not ad hoc posits.
In this paper, we sketched how spacetime and particles may arise as a projection of the
dynamics of conscious agents. By itself, this does not tell us anything about the experience
of time, which flows from a definite past to a present moment towards an open future. It is
instructive therefore to see how an entropic, i.e., a forward-directed, time might similarly
arise in the theory of conscious agents.
It is straightforward to construct a homogeneous Markovian dynamics, X, of conscious
agents with constant entropy, H ( Xn ). That is,

H ( Xn ) = H ( Xn−1 ), ∀n. (32)

Here, n refers to the number of updating steps in the dynamics of conscious agents (cf.
Equation (6)), and not to the number of agents involved.
This dynamics has no preferred direction. However, any projection of this dynamics
via conditional probability induces an entropic arrow of time [98]. That is, if we condition,
say, on X1 , then
H ( Xn | X1 ) ≥ H ( Xn−1 | X1 ), ∀n. (33)

Proof. The proof is straightforward. We note that

H ( X n | X1 ) ≥ H ( X n | X1 , X2 ) . (34)

(Conditioning reduces uncertainty). However, by the Markov property we have

H ( X n | X1 , X2 ) = H ( X n | X2 ) , (35)

and by homogeneity,
H ( X n | X2 ) = H ( X n − 1 | X1 ) . (36)
The last three equations imply that Equation (33) holds.

Thus, the theory of conscious agents may find that entropic time is not a fundamental
feature of reality, but merely an artifact of projection. To explain the “arrow of time”
as an emergent property of entropic systems is not a new idea, e.g., [99,100]. However,
in contrast to most approaches, we do not assume that our experience of an arrow of
time somehow mirrors physical evolution (in fact, evolution by natural selection tells
us this is most unlikely to happen), but that it arises as a projection of a dynamics of
consciousness that underlies physical evolution. The same formalism (i.e., Markovian
networks of conscious agents) that could explain scattering events in spacetime could also
shed light on experienced temporality.
In evolution, organisms compete for resources to counter the ravages of entropy.
The fundamental limited resource is time. Could it be that this entire evolutionary frame-
work is an artifact of projection from the dynamics of conscious agents in which there
are no limited resources and no competition? This becomes a technical question in the
theory of conscious agents, requiring a specification of the dynamics of agents and the
Entropy 2023, 25, 129 25 of 40

precise mapping of this dynamics into spacetime. We would require to obtain evolution as
a projection of agent dynamics with the same rigor in which we get, say, classical physics
as a limiting case of quantum.

Author Contributions: Conceptualization, D.D.H., C.P. and R.P.; methodology and formal analysis,
D.D.H. and C.P.; investigation, D.D.H., C.P. and R.P.; writing—original draft preparation: D.D.H.;
writing—Appendix A: C.P.; writing—Appendix B: D.D.H.; writing—review and editing, D.D.H., C.P.
and R.P. All authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Acknowledgments: We thank Joe Arpaia, Ben Knepper, and Manish Singh for comments and
additions to an earlier draft, and R.P. thanks the Munich Center for Mathematical Philosophy and
the Association for Mathematical Consciousness Science for institutional support. We also thank the
anonymous reviewers for their comments.
Conflicts of Interest: The funders had no role in the design of the study; in the collection, analyses,
or interpretation of data; in the writing of the manuscript; or in the decision to publish the results.

Abbreviations
The following abbreviations are used in this manuscript:

ITP The Interface Theory of Perception

Appendix A. Geometry and Dynamics of Markov Polytopes

In order to understand agent fusion and its possible connections with particle physics,
we need to study the geometry and dynamics of Markov polytopes Mn , of arbitrary order
n. In this Appendix, we make a start towards such a study. We give some definitions and
facts for general n, but mostly focus, as in the body of this paper, on the instances M2 and
M3 , respectively, of two and three interacting conscious agents. Proofs of some assertions
are provided in the last subsection.
Our main goal here is to demonstrate how enormously richer, already, is the 3-agent
theory than that for 2-agents; the analysis here is inspired by the considerations in the body
of this article. Here, we take a naïve approach to the analysis; we anticipate in future work
a deeper study, for arbitrary n, applying the general theory of convex polytopes [101–103].
In addition to delineating aspects of the geometry of M3 , we have made a start in identify-
ing the connection between the geometry and asymptotic dynamics. We hope to announce
later further results on this, both by use of computer simulation and by theoretical analysis.

Appendix A.1. Geometric Structure of Markov Polytopes

An n × n matrix is stochastic, a Markov kernel or a transition matrix, if its entries are real
numbers between 0 and 1 and its rows sum to 1. We can think of a stochastic matrix as the
transition matrix of a Markov chain, on the state space n := {1, . . . , n}. For a given n, the set
2
Mn of all such matrices forms a compact, convex set in Rn under the usual operations
of matrix addition and scalar multiplication. The extreme points, or vertices, of a convex
set are those of its elements which are not convex combinations of any others. In Mn the
vertices are those stochastic matrices each of whose rows is a unit row: a row with a single 1
entry (and the remaining entries all 0). It can be seen that Mn is the set of the set of convex
combinations of its vertices. Hence, we call Mn the Markov polytope of degree n (As the
vertices have all integral coordinates, it is an integral polytope, and we will see that it is also
simple: each of its vertices is connected by boundary edges to a number of other vertices
equal to the dimension n(n − 1) of Mn ).
Entropy 2023, 25, 129 26 of 40

For instance, elements of M2 and M3 can be described, respectively, by matrices of

the form  
  1−a−b a b
1−a a
,  d 1−c−d c , (A1)
b 1−b
e f 1−e− f
where all entries are between 0 and 1. For a 2 × 2 stochastic matrix there are two parameters
to be specified; M2 is thus the image, under a 1:1 affine map, of a 2-dimensional domain
manifold: the Cartesian product of 2 copies, one for each row, of the standard simplex S1
of dimension 1. M2 is thus a 2-dimensional manifold in R2×2 ' R4 . Similarly, M3 is the
image, under a 1:1 affine map, of a 6-dimensional domain: the Cartesian product of 3 copies,
one for each row, of the standard simplex S2 of dimension 2. Thus, M3 is a 6-dimensional
manifold in R3×3 ' R9 . For convenience, Figure A1 reproduces the depiction of M2 given
earlier in Figure 4.

Figure A1. Parameter space for M2 with fusion polytope, Birkhoff polytope, periodic points,
and dynamics.

Consider the general polytope Mn . For i, j denoting integers in n we define the

parameter space Πn := {( xij )i6= j | xij ≥ 0; ∀i, ∑nj=1,j6=i xij ≤ 1}, which is a subset of <n(n−1) .
A bijective mapping Πn → Mn is then given by

(1 − ∑ j6=1 x1j ) x21 xn1 −1 xn1

 
... ... ...

 ... ... ... ... ... ... ... 

( xij )i,j
n
=1, i 6= j 7→ 
 x1i x2i ... (1 − ∑ j6=i xij ) ... xni −1 xni 

 
 ... ... ... ... ... ... ... 
x1n x2n ... ... ... xnn−1 (1 − ∑ j6=n x nj )

In general, Mn is thus an affinely embedded manifold of dimension n(n − 1) in

2
Rn × n ' Rn . Again, it is the convex hull of its extreme points, or vertices. These are
the matrices with all unit rows, i.e., rows with a single unit entry: thus, there are nn
vertices in Mn . The 4 vertices of M2 are shown in Figure A1. The 27 vertices of M3 are
arrayed in Figure A2 (The numbering was assigned by a program in ™MATHEMATICA to
systematically generate them starting from No. 1; we have arrayed and colored the matrix
numbers by “distance” from the identity, i.e., the number of rows in the identity that need
to be changed to reach the new matrix: blue means one row, green means two, red means
three rows. This numbering turns out to be somewhat prescient: see Proposition 2 and the
example before it).
Entropy 2023, 25, 129 27 of 40

Figure A2. Vertices of M3 .

Given a set X of points in a vector space, we will say that its convex hull C( X ) is
generated by that set. As Mn is a convex polytope, its boundary, denoted ∂Mn , consists
of faces: maximal flat regions generated by a subset of vertices. These are polytopes in
their own right and their faces are subsets of those of the mother polytope. Amongst the
faces, those of 0 dimension are the vertices, or 0-faces, 1-dimensional faces are edges, or 1-
faces, those of 1 dimension less than the maximum are the ridges, while those of maximal
dimension are termed facets. The dimension of a polytope is 1 more than the dimension of
its facets (This notion of dimension is equivalent to the usual definition of dimension: a
convex polytope is of dimension r if it is homeomorphic to the r-dimensional unit ball).
Two vertices are adjacent if they are connected by an edge of the polytope: the line
segment generated by them is not contained in the interior of any other face; equivalently,
that none of its points are convex combinations of any other pair of vertices. Each face is
also a polytope; its own faces are amongst the faces of Mn . If two faces share a common
(lower-dimensional) boundary face, we say that they abut one another.

Proposition A1. Two vertices are adjacent if they differ only in the position of a unit entry in the
same row, the other rows being identical.

The set of vertices adjacent to a given vertex V will be termed the cap CV of V. As in
Figure A1 the caps in M2 have 2 vertices each and we will see that the caps of M3 have
6 vertices each. By Proposition A1, the caps of Mn have n(n − 1) vertices each.
The description above of a polytope as the convex hull of its vertices is termed the
V -description. An equivalent description is the H-description, in terms of the set of half-
spaces bounding the polytope. Each such half-space is a linear inequality in the coordinates
of the parameter space (for Mn , this is Rn(n−1) ) and the faces of the polytope are given by
specifying certain subsets of these as equalities, with the proviso that the resulting set is
bounded. For example, recalling Equation (A1), M2 can be described as the set bounded
by four half-spaces, which we can separate into two groups of two inequalities each:

a ≥ 0, 1 − a ≥ 0

and
b ≥ 0, 1 − b ≥ 0
We get a face of M2 by converting, into equalities, one of the inequalities in at least one
group, while keeping the remaining inequalities as is. If this is performed in both groups,
Entropy 2023, 25, 129 28 of 40

we get one of four different 0-faces; if in exactly one group, we get one of four different
1-faces, the facets of the 2-dimensional polytope M2 . This is depicted in Figure A1.
To find the faces of M3 , we follow the same principle on its three bounding groups
of inequalities, by converting into equalities, within at least one group, no more than
two of its inequalities (since if n − 1 entries in a row of a stochastic n × n matrix are zero,
the remaining entry has to be 1). Recalling the second matrix of Equation (A1), the inequality
groupings are
a ≥ 0, b ≥ 0, 1 − a − b ≥ 0
c ≥ 0, d ≥ 0, 1 − c − d ≥ 0 (A2)
e ≥ 0, f ≥ 0, 1 − e − f ≥ 0
So now we can distinguish the following possibilities for the different types of faces of
each dimension in M3 . Figure A3 depicts examples of each type. The types are exemplified
in Example A1 and Proposition A2 gives details of the shapes and numbers of each type.

Figure A3. Examples of the types of r-faces of M3 of dimension r = 1, . . . , 5. Numbering as in

Figure A2. There are two 4-Faces, with 6 and 8 vertices, respectively: their own faces can be seen
to include the prior 3-faces, additionally surrounded by and connected to, respectively, an outer
“enveloping” triangle; or two prisms connected by a square. The 5-facet is generated by the 12-vertex
4-face with two purple enveloping triangles (13-14-15 and 22-23-24) and their connections to the
4-face and each other.

Example A1. Faces of M3

 
0 0 1
• C{25} is of the form 0 0 1, which generates a 0-face or vertex
1 0 0
 
0 0 1
• C{25, 27} is of the form 0 0 1 , which generates a 1-face, or edge
e 0 1−e
 
0 0 1
• C{25, 26, 27} is of the form 0 0 1 , which generates a 2-face triangle
e f 1−e− f
 
0 0 1
C{25, 26, 20, 19} is of the form 1 − c 0 c , which generates a 2-face square
e 1−e 0
Entropy 2023, 25, 129 29 of 40

 
0 0 1
• C{25, 26, 27, 19, 20, 21} is of the form d
 0 1 − d , which generates a 3-face triangu-
e f 1−e− f
lar prism
 
0 a 1−a
• C{25, 26, 10, 11, 16, 17, 19, 20} is of the form 1 − c 0 c , which generates a
e 1−e 0
3-face cube
 
0 0 1
• C{19, 20, 21, 22, 23, 24, 25, 26, 27} is of the form d 1 − c − d c , which gener-
e f 1−e− f
ates a 9-vertex 4-face
 
0 a 1−a
• C{25, 26, 27, 10, 11, 12, 16, 17, 18, 19, 20, 21} is of the form d 0 1 − d , which gen-
e f 1−e− f
erates a 12-vertex 4-face
• C{10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27} is of the form
 
0 a 1−a
d 1 − c − d c , which generates an 18-vertex 5-face, or facet.
e f 1−e− f

These examples suggest how to count the face vector, the vector whose components
count the number of faces f i of dimension i, and more: the numbers of vertices of each type
of face and how many of each type:

Proposition A2. The faces of M3 are as follows:

• 0-faces or vertices: the sets of matrices in M3 with 6 zero entries in common. These are
singletons and there are f 0 = 27 vertices.
• 1-faces, or edges: these are the sets of matrices in M3 with 5 zero entries in common. These
are line segments between adjacent vertices; there are f 1 = 81 edges.
• 2-faces: these are the sets of matrices with 4 zero entries in common. There are two types of
such faces:
Triangles: three vertices will generate a triangular face if there is a single row in which they
differ: the vertices of a triangular face have all four zeros in two of the rows. Each vertex is
attached, via its cap, to three distinct triangular faces. There are 27 triangles, no two of which
abut one another.
Squares: these are generated by four vertices and have a single zero each in two rows, and the
remaining row with two zeros. There are 81 squares, so f 2 = 108.
Equivalently, any two vertices which differ in exactly two rows will lie on a unique square.
Every vertex is on 12 squares.
• 3-faces: these are the sets of matrices with 3 zero entries in common. There are two types of
such faces:
Triangular Prisms are generated by 6 vertices, and with their zeros in two of the rows; there
are 54 triangular prisms.
Cubes are generated by 8 vertices, with a single zero in each of the three rows; there are 27 cubes
and f 3 = 81.
• 4-faces, or ridges: there are two such types of face, with 2 zero matrix entries in common.
9-vertex faces of matrices in M3 with 2 zero entries, in the same row. There are 9 such faces.
12-vertex faces of matrices with 2 zero entries, in two different rows. There are 27 such faces,
so f 4 = 36.
• 5-faces, or facets: the sets, generated by 18 vertices, of matrices with 1 zero entry in common.
There are f 5 = 9 such faces.

Note that, as with any convex polytope, each face is again a polytope, whose faces are,
in turn, themselves lower dimensional faces of M3 . For example, as in Figure A3, we can
Entropy 2023, 25, 129 30 of 40

see that the 6 vertex 3-face is generated by two triangular faces “translated” along three
square faces (or vice-versa), and the 8 vertex cube is a square translated along four square
faces. The 9-vertex 4-face is a “quadra-triangular prism,” in that it consists of four triangular
prisms wrapped together and the other 4-face has two triangular prisms joined by a cube.
Similarly, the 5-facet is an extension of some of the previous faces along some others.
Note also that our 6-dimensional polytope M3 embedded in Rd with d = 9 satisfies,
as it should, the Euler–Poincaré relation

f 0 − f 1 + . . . (−1)d−1 f d−1 = 1 + (−1)d (A3)

With a view towards understanding the Markovian asymptotics, we now identify

certain distinguished subsets of Mn .
The Birkhoff Polytope and Periodic Sets. An n × n matrix Q is doubly stochastic if all
entries are non-negative and each row and each column sums to 1. The subset of all doubly
stochastic matrices constitutes the Birkhoff polytope Bn . This is the sub-polytope of Mn
generated by a subset of n! of the nn vertices of Mn : 2 for M2 and 6 for M3 . The vertices of
the Birkhoff polytope are, under matrix multiplication, isomorphic to the symmetric group
Sn of permutations on the set n of n objects. For example, in the standard correspondence,
the transposition (13) ∈ S3 , which exchanges objects 1 and 3, corresponds to the matrix
 
0 0 1
0 1 0.
1 0 0
Since the vertices of Bn form a finite group, each of its elements is cyclic. As transition
matrices for Markov chains, then, the vertex matrices of Bn are periodic: an n × n matrix
Q being periodic if, for some integer k, Qk = I, the least such k being the period of Q.
The periodic set Pn consists of all periodic stationary stochastic n × n matrices, and the
k-periodic set Pnk consists of the period k matrices in P . For all n, there is only one period 1
matrix: Pn1 = { I }. Define a derangement in Sn , to be a permutation leaving no object of n in
its place. Then, Pnn corresponds to the set of derangements in Sn .  
0 1
In M2 , there is only one derangement matrix: the “NOT” operator ∈ P22 and
1 0
B2 is the line segment between NOT and the identity matrix (as in Figure A1).
In M3 , there are two vertices in P33 : the ones corresponding to the two derangements
in S3 . There are three other vertices in P32 , which we call the mirror matrices. Together with
the identity, we have identified all six vertices of B3 , numbered as in Figure A2:
Vertices of B3
   
0 1 0 0 0 1
P33 : EVEN 16 : 0 0 1, ODD 20 : 1 0 0 (A4)
1 0 0 0 1 0
     
0 0 1 0 1 0 1 0 0
P32 : Mirrors 22 : 0 1 0, 12 : 1 0 0, 8 : 0 0 1 (A5)
1 0 0 0 0 1 0 1 0
 
1 0 0
P31 : Identity 6 : 0 1 0 (A6)
0 0 1
Periodic matrices, being invertible, are of full rank and therefore every periodic vertex is in
Bn . Periodicity in M2 and M3 only happens at vertices: no convex combination of periodic
stochastic matrices can be periodic. We conjecture that this is generally true for all n:
Conjecture A1. Pn is the vertex set of Bn , the set of all periodic stochastic matrices: Pnn corre-
sponds to the set of n-cyclic derangements. For k = 1, . . . , n, Pnk consists of all Birkhoff matrices
corresponding to permutations s ∈ Sn of the following kind: let s = c1 c2 . . . cl , where the ci are,
up to order, the unique decomposition of s into its cycles. Let k s be the l.c.m. of the lengths of the
Entropy 2023, 25, 129 31 of 40

cycles in s. Then, Pnk consists of those Birkhoff matrices corresponding to permutations s ∈ Sn ,

such that k s = k (For n = 2, 3 this comports with the fact that the number of derangements of
k things is !k = [k!/e], where e is the base of the natural logarithm and [ x ] is the integer nearest to x).

The Stationary Complex and Fusion Complex. We will refer to an idempotent

stochastic matrix A, i.e., one satisfying A2 = A, as stationary (The term “stationary” refers
to the fact that a Markov chain with such a transition matrix A is stationary: if the chain has
an initial state given by µ, a measure on its state space, subsequent states are constant and
equal to µA). The stationary complex Sn consists of all stationary matrices in Mn . The next
proposition says that the rank 1 matrices in Mn constitute a polytope in Sn : this is the fusion
polytope Fn1 . When n = 2 this, together with the identity, fills up the stationary manifold:
S2 = F21 ∪ { I } (see Figure A1). Its intersection with the Birkhoff polytope is a singleton
consisting of the matrix each of whose rows is its stationary distribution

Figure A4. The stationary stochastic elements of M3 . Solid lines are edges; dotted lines are segments
internal to a square face of M3 . The fusion polytope F31 is a 2-dimensional simplex while F32 ,
the partial fusion complex, is the union of 6 segments.

In M3 , the fusion polytope F31 is shown in Proposition A3 below to be a 2-simplex.

Except for its vertices, this polytope lies entirely in the interior of M3 , and its intersection
with the Birkhoff
 polytope again consists of one matrix, each of whose rows is its stationary
distribution 1/3 1/3 1/3 . There are now additional stationary matrices, namely the
partial fusion complex Fn2 which consists of the rank 2 matrices in the stationary manifold
S3 . The collection of stationary stochastic matrices is the union of these disjoint sets:
S3 = F31 ∪ F32 ∪ { I }. This is depicted in Figure A4, where solid lines are (boundary) edges
of M3 and dotted lines are in the interior (the numbering of vertices is as in Figure A2).
The figure shows it as a 1-dimensional curve, made up of 6 segments between its
vertices. Moreover, it is connected in paired edges to the (full) fusion polytope F31 , a 2-
dimensional simplex. Note, as in Proposition A2, the three triangular faces issuing from
the identity. Additionally, 3 of the 12 square faces involving the identity actually connect
the identity, as shown in the figure, to the full fusion manifold.

Proposition A3.
(i) The fusion polytope Fn1 is an (n − 1)-dimensional simplex: the convex hull of the n vertex
matrices {ei }, where ei has all 1’s in the ith column and 0’s everywhere else. Other than its
vertices, the fusion polytope lies entirely in the interior of Mn .
(ii) The vertices of the partial fusion complex F32 constitute the cap of the identity and F32 is
the boundary of the hexagon made up of 6 linear segments between these (three of which are
Entropy 2023, 25, 129 32 of 40

edges of M3 and three are internal to it). F32 has exactly three intersections with the Birkhoff
polytope, each lying (as they should) in the middle of the interior segments joining the pairs of
vertices (3,15), (5,9), and (4,24), respectively.
Similarly, for k = 2, . . . , n − 1, the stationary matrices of rank k constitute the rank k
partial fusion complex Fnk (We are not including in the fusion complex Fnn , which would be
just { I } and would involve no fusions).
We conjecture that the following facts, true for n = 2 and 3, are generally true for all n.

Conjecture A2.
(i) The vertices of Fnk , for k = 1, . . . , n − 1, consists of stochastic matrices of the following form:
k columns, say j1 , . . . , jk have a single 1 in their diagonal entries. One of these columns, say
column j1 , has zeroes in the rows corresponding to the other diagonal 1’s (i.e., in the ji th rows,
i 6= 1) and 1’s otherwise; and the remaining columns of the matrix are all zeroes. There are,
therefore, k (nk) partial fusion vertices in Fnk .
(ii) Fnn−1 is the boundary of a figure made up of n(n − 1) simplices of dimension n − 2, whose
vertices constitute the cap of the identity
It is straightforward to see that the type of matrix in Conjecture A2(i) is indeed
idempotent and of rank k. The conjecture will be proved upon showing that these are the
only such stochastic matrices. The structure of the intermediate Fnk requires further study.
In identifying the Birkhoff polytope and stationary complex of M3 , we have exhausted
15 of its 27 vertices. The remaining 12 vertices are the members of the caps of derangement
vertices 16 and 20, as in Figure A2. There, these two caps are just the columns of 6 vertices
next to each of these two cyclic derangements, on the left and right of the figure, respectively.
We will see later that it is useful to distinguish two further subsets each in each cap. With “E”
for 16 and “O” for 20, these are:

Cap of 16 E1 :{10, 17, 21} E2 :{7, 13, 18} (A7)

Cap of 20 O1 :{11, 19, 26} O2 :{2, 21, 23} (A8)

Appendix A.2. Dynamics in Markov Polytopes

We wish to study the flows of Markov chains on finite state spaces [89]; Mn is the set
of their transition kernels. Such a Markov chain is specified by giving both the kernel and a
stochastic vector as its starting probability distribution. This is a 1 × n matrix µ = (µ1 , . . . , µn ),
such that 0 ≤ µi ≤ 1, for all i and ∑in=1 µi = 1. The stochastic vector µ is a stationary
distribution for the stochastic matrix Q if µQ = µ.

Lemma A1. Every stochastic n × n matrix Q has a stationary distribution.

In general, there may be more than one stationary distribution. However, if Q is

irreducible and aperiodic, the stationary measure is unique.
By applying µQ = µ to any matrix in the fusion polytope Fn1 we see, as in Proposition
A3(i), that each row of Q is the same as µ. So the fusion polytope has unique stationary
distributions. For a matrix in the partial fusion complex F32 , however, one coordinate of
the stationary distribution is fixed at 0 (This is the coordinate corresponding to the column
with one diagonal 1 and otherwise 0 s). Thus, the set of stationary distributions for a partial
fusion matrix form a 1-simplex.

Conjecture A3. The sets of stationary distributions of matrices in the partial fusion complexes
Fnk (k = 1, . . . , n − 1), are simplices of dimension n − k.

 First, weestablish Figure A1 as representing the dynamics in M2 . Let Q( x, y) =

1−x x
be the 2 × 2-stochastic-matrix-valued function on the parameter space,
y 1−y
Entropy 2023, 25, 129 33 of 40

 
the unit square 0 ≤ x, y ≤ 1. Then a 2 × 1 matrix µ = a b is a stochastic vector if
0 ≤ a, b, ≤ 1, a + b = 1. The following applies Proposition A3 to n = 2: here the fusion
complex is just the fusion polytope:

Lemma A2. If Q ∈ M2 is stationary, then either:

(i) Q is in the fusion polytope and has equal rows: it is rank 1 and each row of Q is the unique
stationary vector for Q, or
(ii) Q = I and every stochastic vector is stationary for Q.

Theorem A1.
(i) The fusion polytope F21 corresponds to the segment {( x, y)| x, y ∈ [0, 1]; x + y = 1}, includ-
ing end-points.
(ii) The Birkhoff polytope B2 corresponds to the segment  x = y, including
 end-points. Its
intersection with the fusion polytope is the singleton { 1/2 1/2 }.
(iii) The periodic set P22 corresponds to thegreen upper right corner Q(1, 1). Its powers do not
converge, but its stationary measure is 1/2 1/2 }.
(iv) The powers of any matrix other than the identity and the periodic point converge to the fusion
complex, monotonically along the line between itself and the identity. In particular, the powers
of Q( x, y), for ( x, y) 6= (0, 0) and ( x, y) 6= (1, 1), converge to the stationary matrix given by
" y #
x
n x y x +y x +y
lim Q( x, y) = Q( , )= y x (A9)
n→∞ x+y x+y x +y x +y

We would like to study what aspects of the patterns we see in M2 persist at higher
dimensions and what new kinds of behavior emerge. Here, we make a bare start at
this study. Consider Figure A5, which indicates some of the dynamics in M3 where,
for simplicity, we have indicated no adjacencies between the identity and its cap F32 , nor
those between F32 and F31 . The assertions made in Figure A5 are easily checked:

Figure A5. Caps and their dynamics in M3 . Solid lines and arrows are edges of M3 , dashed ones
are not. Single solid arrows indicate that powers of the tail matrix becomes the head matrix. For
example, 13n = 14, n ≥ 2. Double-dashed arrows indicate that powers of the cap matrices of 16 and
20 alternate between itself and the partial fusion matrix at the other end. For example, 112n+1 = 11
and 112n = 4.

Theorem A1. Dynamics between the caps and the fusion polytope are as in Figure A5.
Entropy 2023, 25, 129 34 of 40

Appendix A.3. Proofs

Proof of Proposition A1. Suppose two vertices differ in the position of a unit entry in
some row. Then it is straightforward to see a given convex combination of these two cannot
be also a convex combination of any other pair of vertices. Thus, their convex hull is a
1-face, or edge, of the Markov polytope.
Conversely, suppose two vertices V1 , V2 differ in two rows. WLOG, we can take these
to be rows 1 and 2. Let the jth row of Vi , i, j = 1, 2 be the 1 × 3 row vector rij and assume
r11 6= r21 and r12 6= r22 . Let V3 be the matrix equal to V1 , except in its first row and let V4
be the matrix equal to V1 , except in its second row (an example is in Equations (A2) and
(A3) above). Let V be the convex combination V = xV1 + (1 − x )V2 , with x ≥ 1/2. Then,
V is also in the convex hull of V1 , V3 , and V4 : solving for V = x 0 V1 + y0 V3 + (1 − x 0 − y0 )V2 ,
we get y0 = 1 − x and x 0 = 2x − 1. Our assertion will be verified if we can show that
0 ≤ x 0 , y0 ≤ 1. However, x ≥ 1/2 =⇒ x 0 ≥ y ≥ 0 and x ≤ 1 =⇒ x 0 ≤ 1. A similar
consideration applies to the instance x ≤ 1/2, in which case we can see that V is now also
a convex combination of V2 , V3 , and V4 . In either case, V is not on an edge.

Proof of Proposition A2. (added new line before bullet)

• 0-faces or vertices: a matrix in M3 has 6 zero entries if it has three unit rows, each with
three possibilities for its unit entry: there are 27 vertices.
• 1-faces, or edges: the 5 zero entries can be distributed by choosing a row with a single
zero in 3 ways, choosing that zero entry’s position in 3 ways, and then choosing places
in the two other rows for their unit entry in 3 ways each: there are 81 edges.
• 2-faces: one way a set of vertices can have 4 zero entries in common is if they all
have two fixed zeros each of two rows (3 ways to choose this pair of rows). The zero
placement, equivalently the unit entry, in each of these rows can be chosen in 3 ways
each. Thus, there are 33 = 27 triangular faces.
Once this choice is made, the set generates a Triangular face, as the unit entry in the
remaining row can be chosen in exactly three ways and the resulting vertices are
all adjacent, by Proposition A1. Each vertex is adjacent to three distinct triangular
faces, one for each row where the unit entry is shifted. Since there are three such
distinguished rows, there are three distinct triangles including that vertex. Thus, no
two triangles can abut one another (Another way to see that there are 81 triangles is to
note that each of the 27 vertices is on 3 triangles, and all triangles are distinct).
Squares: the other way a set of vertices will 4 zero entries in common is if they all have
two rows ((32) = 3 ways to choose this pair of rows) with a single zero each (3 ways
to position the zero in each row), and the remaining row with two zeros (3 ways to
position these zeros). There are, thus, 81 such faces.
Each of the square faces will be the convex hull of the four vertices that are obtained
as follows: call one of the vertices V1 and let V2 be a vertex which differs from V1 in
both rows. Pick one of the distinguished rows with a single zero and let V3 be the
vertex obtained by shifting V1 ’s unit entry in that row, to its position in V2 . Let V4 be
the vertex obtained from V1 by shifting its unit entry in the other distinguished row to
that of V2 . There are, by Proposition A1, edges between V1 and V3 , V1 and V4 , V3 and
V2 , and V4 and V2 and no other edges: the resulting figure, which (see Example A1)
has two free parameters, is indeed a 2-dimensional square.
Given an arbitrary vertex, there are 3 ways to choose two of its rows, and 4 ways to
change the position of one or both unit entries in those two rows. Thus, any given
vertex will lie on 12 squares.
• 3-faces: sets of vertices with 3 zero entries in common can happen in two ways:
Their common 3 zeros lie in two of the rows: one with 2 zeros and the other with
one. Pick the 2-zero row in 3 ways and position the 2 zeros in three ways. Then,
pick the 1-zero row in two ways and the position of the zero in three ways; there are
54 such faces.
Entropy 2023, 25, 129 35 of 40

Such a set will have 6 vertices, as there are two possible positions for the unit entry in
the single-zero row, and three in the row with no shared zeros. Example 1 shows how
to see that the resulting figure generated by these 6 vertices is a Triangular Prism, as in
Figure A3.
Similarly, cubes are generated by 8 vertices, with a single zero in each of the three rows;
there are 3 × 3 × 3 = 27 cubes.
• 4-faces, or ridges have 2 zero entries in common: when these are in the same row, the re-
maining two rows are free, so there are 3 × 3 = 9 vertices in such a face. Pick the row
with 2 zero entries in common in 3 ways, and place those zeros in 3 ways: there are 9
such faces.
When the 2 zero entries in common are in two different rows, we get a 12 vertex face,
since there are 2 × 2 × 3 = 12 ways to distribute unit entries in the rows. Pick the two
distinguished rows in 3 ways and distribute the two common zeros in 3 ways each.
There are, thus, 27 such faces.
• The 5-faces, or facets are those 18-vertex faces (18 = 2 × 3 × 3) with 1 zero entry in
common. There are 3 × 3 = 9 such faces.

Proof of Proposition A3.

(i) A rank 1 stochastic matrix Q is one whose rows are all the same stationary distribution,
say µ = (µ1 , . . . , µn ). Thus Q = ∑in=1 µi ei . Since each row of Q is equal to µ and its jth
column is µ j 1, where 1 is the column of all 1’s, we obtain

n
( Q2 )ij = µ · (µ j 1) = µ j ∑ µi = µ j
i =1

In other words, Q2 = Q: rank 1 stochastic matrices are idempotent and so constitute Fn1 .
(ii) It can be verified from Figure A2 that F32 , the idempotent rank 2 vertices, constitute
the cap of the identity and that it is the boundary of the hexagon made up of 6 linear
segments between these (three of which are edges of M3 and three are internal to it)
(A quick way to verify that any segment between idempotent matrices A, B is itself
idempotent is to prove that this is so if their anticommutator { A, B} := AB + BA is
equal to A + B). 

Proof of Lemma A1. Q can be viewed as a linear (and so continuous) operator on Rn ,

which restricts to a continuous function from the convex set of probability measures on
n := {1, . . . , n} to itself. Brower’s fixed point theorem then says that there is a fixed point
of the continuous map Q of the convex set of probability measures on n := {1, . . . , n} to
itself: this fixed point is what is meant by a stationary distribution.

Proof of Theorem A1. Items (i), (ii), and (iii) are straightforward. Recall that the stochastic
matrix with parameters x, y is given by
 
1−x x
Q( x, y) =
y 1−y

with x, y ∈ [0, 1].

(iv) Fix p0 = ( x0 , y0 ) in the unit square, except for the corners (0, 0) and (1, 1). Points on the
line between the origin and p0 can be parameterized as p a = ( ax0 , ay0 ). A calculation
shows that


Q( p a ) Q( p0 ) = Q x0 (1 + a − a( x0 + y0 )), y0 (1 + a − a( x0 + y0 ))

Thus Q( p a ) Q
( p0 ) = Q( p1 ), where p1 = ( x1 , y1 ) = (1 + a − a( x0 + y0 )) x0 , (1 + a −

a( x0 + y0 ))y0 is again a point on that same line. Iterating the above formula starting
at p0 , we see that the sequence ( Q( p0 )n )∞
n=1 stays on that line. Suppose p0 is to the
Entropy 2023, 25, 129 36 of 40

right of the cross-diagonal x + y = 1. In terms of our parameter the cross-diagonal

1 1
is intersected at a = x0 + y0 . This means that if a > x0 +y0 , our formula shows that
x1 < ax0 and y1 < ay0 : p1 is shifted from p0 a positive distance towards the origin.
1
Similarly, when p0 is to the left of x + y = 1, so that a < x0 + y0 , then p1 is shifted away
n ∞
from the origin. In either case, the sequence ( Q( p0 ) )n=1 is strictly monotonic towards
the cross-diagonal, unless it is actually on it, where the matrix Q( x, y) is stationary.

Appendix B. Decorated Permutations for the 27 Vertices of the Markov Polytope M3

Number Matrix Decorated Permutation

 
1 0 0
1 1 0 0 [4, 2, 3]
1 0 0
 
1 0 0
2 1 0 0 [4, 2, 3]
0 1 0
 
1 0 0
3 1 0 0 [4, 2, 6]
0 0 1
 
1 0 0
4 0 1 0 [4, 5, 3]
1 0 0
 
1 0 0
5 0 1 0 [4, 5, 3]
0 1 0
 
1 0 0
6 0 1 0 [4, 5, 6]
0 0 1
 
1 0 0
7 0 0 1 [4, 2, 3]
1 0 0
 
1 0 0
8 0 0 1 [4, 3, 5]
0 1 0
 
1 0 0
9 0 0 1 [4, 2, 6]
0 0 1
 
0 1 0
10 1 0 0 [2, 4, 3]
1 0 0
 
0 1 0
11 1 0 0 [2, 4, 3]
0 1 0
 
0 1 0
12 1 0 0 [2, 4, 6]
0 0 1
Entropy 2023, 25, 129 37 of 40

Number Matrix Decorated Permutation

 
0 1 0
13 0 1 0 [1, 5, 3]
1 0 0
 
0 1 0
14 0 1 0 [1, 5, 3]
0 1 0
 
0 1 0
15 0 1 0 [1, 5, 6]
0 0 1
 
0 1 0
16 0 0 1 [3, 4, 5]
1 0 0
 
0 1 0
17 0 0 1 [1, 3, 5]
0 1 0
 
0 1 0
18 0 0 1 [1, 2, 6]
0 0 1
 
0 0 1
19 1 0 0 [3, 2, 4]
1 0 0
 
0 0 1
20 1 0 0 [3, 4, 5]
0 1 0
 
0 0 1
21 1 0 0 [1, 2, 6]
0 0 1
 
0 0 1
22 0 1 0 [3, 5, 4]
1 0 0
 
0 0 1
23 0 1 0 [1, 5, 3]
0 1 0
 
0 0 1
24 0 1 0 [1, 5, 6]
0 0 1
 
0 0 1
25 0 0 1 [3, 2, 4]
1 0 0
 
0 0 1
26 0 0 1 [1, 3, 5]
0 1 0
 
0 0 1
27 0 0 1 [1, 2, 6]
0 0 1
Entropy 2023, 25, 129 38 of 40

References
1. Signorelli, C.M.; Szczotka, J.; Prentner, R. Explanatory profiles of models of consciousness-towards a systematic classification.
Neurosci. Conscious. 2021, 2021, niab021. [CrossRef] [PubMed]
2. Hameroff, S.R.; Penrose, R. Consciousness in the universe. A review of the ‘Orch OR’ theory. Phys. Life Rev. 2014, 11, 39–78.
[CrossRef] [PubMed]
3. Hameroff, S.R. Quantum computation in brain microtubules? The Penrose-Hameroff “Orch OR” model of consciousness. Philos.
Trans. R. Soc. A 1998, 356, 1869–1896. [CrossRef]
4. Penrose, R. Shadows of the Mind: A Search for the Missing Science of Consciousness; Oxford University Press: New York, NY,
USA, 1994.
5. Oizumi, M.; Albantakis, L.; Tononi, G. From the Phenomenology to the Mechanisms of Consciousness: Integrated Information
Theory 3.0. PLoS Comput. Biol. 2014, 10, e1003588. [CrossRef] [PubMed]
6. Tononi, G.; Boly, M.; Massimini, M.; Koch, C. Integrated information theory: From consciousness to its physical substrate. Nat.
Rev. Neurosci. 2016, 17, 450–461. [CrossRef]
7. Haun, A.; Tononi, G. Why Does Space Feel the Way it Does? Towards a Principled Account of Spatial Experience. Entropy 2019,
21, 1160. [CrossRef]
8. Barbosa, L.S.; Marshall, W.; Albantakis, L.; Tononi, G. Mechanism integrated information. Entropy 2021, 23, 362. [CrossRef]
9. Baars, B.J. A Cognitive Theory of Consciousness; Cambridge University Press: New York, NY, USA, 1988.
10. Baars, B.J. In the Theater of Consciousness; Oxford University Press: New York, NY, USA, 1997.
11. Dehaene, S.; Naccache, L. Towards a cognitive neuroscience of consciousness: Basic evidence and a workspace framework.
Cognition 2001, 79, 1–37. [CrossRef]
12. Mashour, G.A.; Roelfsema, P.; Changeux, J.P.; Dehaene, S. Conscious Processing and the Global Neuronal Workspace Hypothesis.
Neuron 2020, 105, 776–798. [CrossRef]
13. Dennett, D.C. Consciousness Explained; Little, Brown & Company: Boston, MA, USA, 1991.
14. Frankish, K. (Ed.) Illusionism as a Theory of Consciousness; Imprint Academic: Exeter, UK, 2017.
15. Graziano, M.S.A. Rethinking Consciousness: A Scientific Theory of Subjective Experience; W & W Norton: New York, NY, USA, 2019.
16. Strawson, G. Realistic Monism: Why Physicalism Entails Panpsychism. In Consciousness and Its Place in Nature. Does Physicalism
Entail Panpsychism? Freeman, A., Ed.; Imprint Academic: Exeter, UK, 2006; pp. 3–31.
17. Chalmers, D.J. Panpsychism and Panprotopsychism. In Consciousness in the Physical World: Perspectives on Russellian Monism;
Alter, T., Nagasawa, Y., Eds.; Oxford University Press: New York, NY, USA, 2015; pp. 246–276.
18. Popper, K.R.; Eccles, J.C. The Self and its Brain: An Argument for Interactionism; Springer: Berlin, Germany, 1977.
19. Beck, F.; Eccles, J.C. Quantum aspects of brain activity and the role of consciousness. Proc. Natl. Acad. Sci. USA 1992, 89,
11357–11361. [CrossRef]
20. Goff, P. Consciousness and Fundamental Reality; Oxford University Press: Oxford, UK, 2017.
21. Prentner, R. The Natural Philosophy of Experiencing. Philosophies 2018, 3, 35. [CrossRef]
22. Arkani-Hamed, N.; Bourjaily, J.; Cachazo, F.; Goncharov, A.; Postnikov, A.; Trnka, J. Grassmanian Geometry of Scattering Amplitudes;
Cambridge University Press: Cambridge, UK, 2016.
23. Arkani-Hamed, N.; Benincasa, P. On the emergence of Lorentz invariance and unitarity from the scattering facet of cosmological
polytopes. arXiv 2018, arXiv:1811.01125.
24. Arkani-Hamed, N.; Huang, T.C.; Huang, Y.T. Scattering Amplitudes For All Masses and Spins. arXiv 2021, arXiv:1709.04891.
https://doi.org/10.48550/arXiv.1709.04891.
25. Gross, D. Einstein and the search for unification. Curr. Sci. 2005, 89, 2035–2040.
26. Musser, G. Spacetime is Doomed. In Space, Time and the Limits of Human Understanding; Wuppuluri, S., Ghirardi, G.C., Eds.;
Springer: Cham, Switzerland, 2017; pp. 217–227.
27. Arkani-Hamed, N. The Future of Fundamental Physics, 2010. Cornell Messenger Lectures. Available online: https://www.
youtube.com/watch?v=SWWBuHszyD8 (accessed on 8 January 2023).
28. Arkani-Hamed, N. Lecture # 1, Spacetime & Quantum Mechanics, Total Positivity & Motives, 2019. Lecture Series at Harvard
University. Available online: https://www.youtube.com/watch?v=Sn0W_mwA7Q0 (accessed on 8 January 2023).
29. Schneider, S. Spacetime Emergence, Panpsychism and the Nature of Consciousness; Scientific American: New York, NY, USA, 2018.
30. Fodor, J. The Mind Doesn’t Work That Way: The Scope and Limits of Computational Psychology; MIT Press: Cambridge, MA, USA, 2000.
31. Churchland, P.S. Epistemology in the age of neuroscience. J. Philos. 1987, 84, 544–555. [CrossRef]
32. Pinker, S. How the Mind Works; W.W. Norton: New York, NY, USA, 1999.
33. Pinker, S. So how does the mind work? Mind Lang. 2005, 20, 1–24. [CrossRef]
34. Hoffman, D.D.; Singh, M.; Prakash, C. The interface theory of perception. Psychon. Bull. Rev. 2015, 22, 1480–1506. [CrossRef]
35. Mark, J.T.; Marion, B.B.; Hoffman, D.D. Natural selection and veridical perceptions. J. Theor. Biol. 2010, 266, 504–515. [CrossRef]
36. Prakash, C.; Stepehens, K.D.; Hoffman, D.D.; Singh, M.; Fields, C. Fitness Beats Truth in the Evolution of Perception. Acta Biotheor.
2021, 69, 319–341. [CrossRef]
37. Prakash, C.; Fields, C.; Hoffman, D.D.; Prentner, R.; Singh, M. Fact, Fiction, and Fitness. Entropy 2020, 22, 514. [CrossRef]
38. Fields, C. This boundary-less world. In Brain, Mind, Cosmos; Chopra, D., Ed.; Chopra: La Costa, CA, USA, 2014; Chapter 13.
39. Hoffman, D.D. Visual Intelligence: How We Create What We See; W.W. Norton: New York, NY, USA, 1998.
Entropy 2023, 25, 129 39 of 40

40. Hoffman, D.D. The interface theory of perception. In Object Categorization: Computer and Human Vision Perspectives; Dickinson, S.,
Tarr, M., Leonardis, A., Schiele, B., Eds.; Cambridge University Press: New York, NY, USA, 2009; pp. 148–165.
41. Hoffman, D.D. The construction of visual reality. In Hallucinations: Theory and Practice; Blom, J., Sommer, I., Eds.; Springer:
New York, NY, USA, 2011; pp. 7–15.
42. Hoffman, D.D. The sensory desktop. In This Will Make you Smarter: New Scientific Concepts to Improve Your Thinking; Brockman, J.,
Ed.; Harper Perennial: New York, NY, USA, 2012; pp. 135–138.
43. Hoffman, D.D. Public Objects and Private Qualia: The Scope and Limits of Psychophysics. In Handbook of Experimental
Phenomenology; Albertazzi, L., Ed.; John Wiley & Sons: Hoboken, NJ, USA, 2013; pp. 71–90.
44. Hoffman, D.D.; Singh, M.; Mark, J.T. Does evolution favor true perceptions? In Human Vision and Electronic Imaging XVIII;
Rogowitz, B.E., Pappas, T.N., de Ridder, H., Eds.; International Society for Optics and Photonics, SPIE: Bellingham, WA, USA,
2013; Volume 8651, p. 865104. [CrossRef]
45. Koenderink, J.J. World, environment, Umwelt, and innerworld: A biological perspective on visual awareness. In Human Vision
and Electronic Imaging XVIII; Rogowitz, B.E., Pappas, T.N., de Ridder, H., Eds.; International Society for Optics and Photonics,
SPIE: Bellingham, WA, USA, 2013; Volume 8651, p. 865103. [CrossRef]
46. Koenderink, J.J. The all seeing eye? Perception 2014, 43, 1–6. [CrossRef] [PubMed]
47. Mausfeld, R. The physicalist trap in perception theory. In Perception and the Physical World: Psychological and Philosophical Issues in
Perception; Heyer, D., Mausfeld, R., Eds.; Wiley: New York, NY, USA, 2002; pp. 75–112.
48. von Uexküll, J. Umwelt und Innenwelt der Tiere; Springer: Berlin, Germany, 1909.
49. von Uexküll, J. Theoretical Biology; Hartcourt, Brace & Co.: New York, NY, USA, 1926.
50. von Uexküll, J. A stroll through the worlds of animals and men: A picture book of invisible worlds. In Instinctive Behavior:
Development of a Modern Concept; Hallmark Press: New York, NY, USA, 1957.
51. Fields, C.; Hoffman, D.D.; Prakash, C.; Prentner, R. Eigenforms, interfaces and holographic encoding. Constr. Found. 2017,
12, 265–291.
52. Held, R.; Ostrovosky, Y.; de Gelder, B.; Gandhi, T.; Ganesh, S.; Mathur, U.; Sinha, P. The newly sighted fail to match seen with felt.
Nat. Neurosci. 2021, 14, 551–553. [CrossRef]
53. Berke, M.; Walter-Terill, R.; Jara-Ettinger, J.; Scholl, B. Flexible goals require that inflexible perceptual systems produce veridical
representations: Implications for realism as revealed by evolutionary simulations. Cogn. Sci. 2022, 46, e13195. [CrossRef]
[PubMed]
54. Giustina, M.; Versteegh, M.A.M.; Wengerowsky, S.; Handsteiner, J.; Hochrainer, A.; Phelan, K.; Steinlechner, F.; Kofler, J.; Larsson,
J.A.; Abellán, C.; et al. Significant-loophole-free test of Bell’s Theorem with entangled photons. Phys. Rev. Lett. 2015, 115, 250401.
[CrossRef] [PubMed]
55. Gröblacher, S.; Paterek, T.; Kaltenbaek, R.; Brukner, Č.; Żukowski, M.; Aspelmeyer, M.; Zeilinger, A. An experimental test of
non-local realism. Nature 2007, 446, 871–875. [CrossRef] [PubMed]
56. Hensen, B.; Bernien, H.; Dréau, A.; Reiserer, A.; Kalb, N.; Blok, M.S.; Ruitenberg, J.; Vermeulen, R.F.L.; Schouten, R.N.; Abellán,
C.; et al. Loophole-free Bell inequality violation using electron spins separated by 1.3 kilometres. Nature 2015, 526, 682–686.
[CrossRef]
57. Edelman, G. Wider than the Sky: The Phenomenal Gift of Consciousness; Yale University Press: New Haven, CT, USA, 2004.
58. Locke, J. An Essay Concerning Human Understanding; Tho Basset: London, UK, 1690.
59. Kant, I. Prolegomena to Any Future Metaphysics. With Selections from the Critique of Pure Reason; Cambridge University Press:
Cambridge, UK, 1997.
60. Hoffman, D.D.; Prakash, C. Objects of consciousness. Front. Psychol. 2014, 5, 577. [CrossRef]
61. Fields, C.; Hoffman, D.D.; Prakash, C.; Singh, M. Conscious agent networks: Formal analysis and application to cognition. Cogn.
Syst. Res. 2018, 47, 186–213. [CrossRef]
62. Baars, B.J. The conscious access hypothesis: Origins and recent evidence. Trends Cogn. Sci. 2002, 6, 47–52. [CrossRef]
63. Dehaene, S. Consciousness and the Brain: Deciphering How the Brain Codes Our Thoughts; Viking: New York, NY, USA, 2014.
64. Tononi, G. Conciousness as Integrated Information: A Provisional Manifesto. Biol. Bull. 2008, 215, 216–242. [CrossRef]
65. Leibniz, G.W. Discourse on Metaphysics and The Monadology; Dover: Mineola, NY, USA, 2005.
66. Atmanspacher, H. The Pauli–Jung Conjecture and Its Relatives: A Formally Augmented Outline. Open Philos. 2020, 3, 527–549.
[CrossRef]
67. Ceccherini-Silberstein, T.; Coornaert, M. Cellular Automata and Groups; Springer: Berlin, Germany, 2010. [CrossRef]
68. Cook, M. Universality in elementary cellular automata. Complex Syst. 2004, 15, 1–40.
69. Prentner, R. Dr Goff, Tear Down This Wall! The Interface Theory of Perception and the Science of Consciousness. J. Conscious.
Stud. 2021, 28, 91–103. [CrossRef]
70. Mariscal, C. Life. In The Stanford Encyclopedia of Philosophy, Winter 2021 ed.; Zalta, E.N., Ed.; Metaphysics Research Lab, Stanford
University: Stanford, CA, USA, 2021.
71. Bagwell, J.N. Ordinary Object Beliefs and Scientific Theory. Ph.D. Thesis, University of California, Santa Barbara, CA, USA, 2022.
Available online: https://escholarship.org/uc/item/9d8413x3 (accessed on 8 January 2023).
72. Seager, W. Consciousness, information and panpsychism. J. Conscious. Stud. 1995, 2, 272–288.
73. Goff, P. Why Panpsychism doesn’t Help Us Explain Consciousness. Dialectica 2009, 63, 289–311. [CrossRef]
Entropy 2023, 25, 129 40 of 40

74. Seager, W.E. Panpsychism, aggregation and combinatorial infusion. Mind Matter 2010, 8, 167–184.
75. Coleman, S. The real combination problem: Panpsychism, micro-subjects, and emergence. Erkenntnis 2014, 79, 19–44. [CrossRef]
76. Mørch, H.H. Panpsychism and Causation: A New Argument and a Solution to the Combination Problem. Ph.D. Thesis,
University of Oslo, Oslo, Norway, 2014. Available online: https://philpapers.org/archive/HASPAC-2.pdf (accessed on 8
January 2023).
77. Chalmers, D.J. The Combination Problem for Panpsychism. In Panpsychism: Philosophical Essays; Brüntrup, G., Jaskolla, L., Eds.;
Oxford University Press: Oxford, UK, 2016; pp. 19–47.
78. Goff, P. Galileo’s Error. Foundations for a New Science of Consciousness; Vintage Books: New York, NY, USA, 2019.
79. Roelofs, L. Combining Minds: How to Think about Composite Subjectivity; Oxford University Press: Oxford, UK, 2019.
80. D’Ariano, G.M.; Faggin, F. Hard problem and free will: An information-theoretic approach. arXiv 2021, arXiv:2012.06580v2.
81. Schrödinger, E. Mind and Matter; Cambridge University Press: Cambridge, UK, 1958.
82. nLab authors. Cantor’s Theorem. 2022. Available online: http://ncatlab.org/nlab/show/Cantor%27s%20theorem (accessed on
8 January 2023).
83. Goff, P. Putting Consciousness First: Replies To Critics. J. Conscious. Stud. 2021, 40, 289–328. [CrossRef]
84. Parke, S.; Taylor, T. Amplitude for n-gluon scattering. Phys. Rev. Lett. 1986, 56, 2459. [CrossRef] [PubMed]
85. Williams, L. The positive Grassmannian, the amplituhedron, and cluster algebras. arXiv 2022, arXiv:2110.10856v2.
86. Hartnet, K. A Mathematician’S Unanticipated Journey through the Physical World; Quanta Magazine: New York, NY, USA, 2020.
87. Postnikov, A.; Speyer, D.; Williams, L. Matching polytopes, toric geometry, and the totally non-negative Grassmannian. arXiv
2008, arXiv:0706.2501v3.
88. Arkani-Hamed, N.; Maldacena, J. Cosmological Collider Physics. arXiv 2015, arXiv:1503.08043v1.
89. Norris, J. Markov Chains; Cambridge University Press: Cambridge, UK, 1997.
90. Corteel, S.; Williams, L. Tableaux combinatorics for the asymmetric exclusion process. Adv. Appl. Math. 2007, 39, 293–310.
[CrossRef]
91. Kodama, Y.; Williams, L. KP solitons and total positivity for the Grassmannian. Invent. Math. 2014, 198, 637–699. [CrossRef]
92. Clark, S. Why Einstein never received a Nobel prize for relativity. The Guardian, 8 October 2012.
93. Becker, K.; Becker, M.; Schwarz, J. String Theory and M-Theory: A Modern Introduction; Cambridge University Press: New York, NY,
USA, 2007.
94. Rovelli, C. Quantum Gravity; Cambridge University Press: New York, NY, USA, 2004.
95. Dowker, F. Introduction to causal sets and their phenomenology. Gen. Relativ. Gravit. 2013, 45, 1651–1667. [CrossRef]
96. Prentner, R. A Framework for Critical Materialists. Mind Matter 2014, 12, 93–118.
97. Pinker, S. Enlightenment Now: The Case for Reason, Science, Humanism, and Progress; Viking: New York, NY, USA, 2018.
98. Cover, T.M.; Thomas, J. Elements of Information Theory, 2nd ed.; Wiley: Hoboken, NJ, USA, 2006.
99. Price, H. Time’s Arrow and Archimedes’ Point: New Directions for the Physics of Time; Oxford University Press: Oxford, UK, 1996.
100. Rovelli, C. Is Time’s Arrow Perspectival? In The Philosophy of Cosmology; Chamcham, K., Silk, J., Barrow, J.D., Saunders, S., Eds.;
Cambridge University Press: Cambridge, UK, 2017; pp. 285–296.
101. Grünbaum, B. Convex Polytopes; Springer: New York, NY, USA, 2003.
102. Ziegler, G.M. Lectures on Polytopes; Springer: New York, NY, USA, 1995.
103. Henk, M.; Richter-Gebert, J.; Ziegler, G. Basic properties of convex polytopes. In Handbook of Discrete and Computational Geometry,
3rd ed.; Goodman, J., O’Rourke, J., Tóth, C.D., Eds.; CRC Press: Boca Raton, FL, USA, 2017; Chapter 15.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.