The first post in this series made a simple proposal: model a cognitive system as an algebra of operations together with a state, and let the geometry come out of that pair instead of assuming it in advance. This post is what happened when I tried to take the proposal seriously, with code, with small simulations, and actually checking the mathematics.

The short version of the result: some pieces hold up better than I expected, one piece I had put at the center does not hold up at all, and a few results came out the opposite way from what I predicted. I will say which is which as we go.

The central idea of this post is:

A mind’s time is not a coordinate it moves along. It is something its own state generates, and the self is the structure that lets the mind read that time from the inside.

Before anything else, the status. This is a model, not a measurement. Every computation reported here comes from small simulations on toy systems, which are not brains and not language models. I tag the main claims as derived (standard mathematics or a direct calculation), checked (verified numerically in a script), literature (cited and not re-derived by me), or conjecture. Anything tagged conjecture should not be read as established, and the weakest steps are collected in the last section.

A note on reading: terms followed by a reference like (A.3) are explained in a more accessible language in the appendix at the end of the post.

1. Where the first post stopped

The first post argued that non-commutativity comes from self-reference: a system that models and acts on itself has operations whose order matters, so its algebra should be non-commutative. I called that the central conjectural step. It turns out to be the weakest step, and it fails as stated.

The reason is simple. Order-dependence is cheap. Almost any procedure is a composition of steps whose order matters (parse, then evaluate, and not the other way round). So “order matters” cannot be the content of the claim. To be clear about what is meant here: the non-commuting objects are cognitive operations in a context, such as attending and then speaking versus speaking and then attending. I do not assign a cognitive function to any layer of a network. Layers are just representations, and the operators of this framework live at the level of operations on the agent’s state. What carries content is extra structure: an adjoint, a state that gives non-negative expectation to squared operations, and statistics that no single classical probability distribution can reproduce. And self-reference by itself leads naturally to fixed points and diagonal arguments (Gödel, Lawvere, Löb), not to non-commutativity.

So I replaced it with an assumption that can be wrong in a way data can show:

Postulate. The statistics of cognitive probes (questions, judgments, sequential operations) are contextual: after accounting for direct influences between settings, no single joint probability distribution reproduces all the observed marginals. (A.3)

Everything non-commutative in this post is conditional on that. The evidence so far points both ways. Order effects in attitude surveys follow a precise constraint called the QQ equality across many large surveys1, a constraint a classical model is free to violate. Direct contextuality tests in psychology, which are the cleaner tool, have given mixed and small results2. I have not run the decisive experiment, a two-person Bell-style test with no communication. For this post I assume the postulate holds, and the last section lists what would break it.

Why operators at all, if the postulate holds? Two reasons from the literature (literature). If statistics are contextual there is no fixed classical state space, and the operator description is the minimal extension. Independently, contextual sequences are expensive to simulate classically in memory3, so a non-commutative model is a memory-efficient representation for a bounded agent. The algebra is a compression. It is not a claim that the brain is quantum4.

2. What the objects mean

An agent at stage \$n\$ is a pair \$(\mathcal A_n,\omega_n)\$, plus a set of operations. The algebra \$\mathcal A_n\$ is what the agent can distinguish. The state \$\omega_n\$ is its current expectations, with some uncertainty everywhere (what the math calls faithful). The operations are the channels through which it takes in input, inspects itself and acts. Learning is a growing chain \$\mathcal A_1\subset\mathcal A_2\subset\cdots\$ in which each stage is embedded in the next.

Two pieces of interpretation make the formalism less abstract, and both were checked numerically.

The first is the adjoint. Relative to the agent’s state, the adjoint of an operation is its Bayesian retrodiction: the operation that undoes it in the sense of Bayes’ rule (the Petz recovery map)5. The inner product \$\omega(A^*B)\$ measures how distinguishable two operations are to the agent.

The second is the modular operator \$\Delta\$ (A.4), which turns out to be the ratio between retrodiction and prediction. For any operation \$A\$,

$$ \omega(AA^*)=\big\|\Delta^{1/2}A\Omega\big\|^2, \qquad \omega(A^*A)=\big\|A\Omega\big\|^2 . $$

I checked both identities on random complex operators and random faithful states, and they agreed to about \$10^{-15}\$ (checked). For a simple transition from state \$i\$ to state \$j\$, the ratio is \$p_j/p_i\$. That is the detailed-balance ratio, the way equilibrium weighs “going this way” against “going back”. So \$\Delta\$ measures how asymmetric an operation looks forwards versus backwards, and that is exactly where an arrow of time would have to live.

3. Time generated by the state

Take the agent’s state as a density matrix \$\rho\$. Define

$$ K=-\log\rho, \qquad \sigma_s(A)=\rho^{\,is}\,A\,\rho^{-is}. $$

\$K\$ is the surprisal operator: minus the log of how likely each state is. The family \$\sigma_s\$ is the modular flow (A.4). Each operation is rotated at a rate set by how surprising its states are: rare states rotate fast, common ones slowly. The flow leaves the state itself unchanged, so it is a symmetry of the agent’s expectations and not an event in them (A.2).

What the flow needs (derived):

  • It is trivial exactly when the state is tracial, which is the completely flat case. Every state on a commutative algebra is tracial, so a classical model has no endogenous time. Time has to be supplied from outside.
  • A pure state, with no uncertainty at all, has no modular flow.
  • So a mind has intrinsic time only at intermediate entropy, neither certain nor maximally confused. At maximal mixing the spread of \$K\$ is exactly zero (checked).

The entropy is the average of \$K\$, \$S=\omega(K)\$, and the variance of \$K\$ (the “varentropy”) says how unevenly the clock ticks.

Two clocks, and where the arrow is not. The modular parameter \$s\$ is reversible: a mathematical object (the modular conjugation) runs it backwards. So it gives temporal structure but no arrow and no “now”. Alongside it there is a second quantity, \$a\$, the length of the path the state traces as it updates. It only accumulates, but path length alone does not know its direction either. The arrow is which direction updates physically run, and the irreversibility of an update is how much relative entropy it destroys (A.7). Why updates run in the record-forming direction comes from the physical world, namely its low-entropy past and the energy cost of erasing records6. The framework inherits the arrow. It does not derive it.

In a small relaxing model (a three-level agent with thermal relaxation plus a weak drive) the two clocks separate cleanly (checked). The arclength \$a\$ saturates at about 1.98 once the agent has settled, while the modular swing stays near 0.53 and the varentropy near 0.43. The state has stopped changing but its modular clock keeps running. Entropy itself is not monotone: it peaked at 0.863 and settled at 0.812, so “entropy increases” is not the cognitive arrow. The relative entropy to the thermal reference fell from 0.98 to about 0.005 and then settled at 0.0077, not zero, because the drive keeps the steady state slightly out of equilibrium.

4. Growth and the types of minds

A learner does not have a fixed algebra. It keeps acquiring new distinctions. At every finite stage the algebra is an ordinary matrix algebra (what the math calls type I). The interesting question is what the limit looks like if learning never stops (A.5).

The simplest model adds independent slots, each a two-level unit with state \$\mathrm{diag}(p_k,q_k)\$ and odds \$\lambda_k=q_k/p_k\le 1\$. The infinite limit is a well-studied object, and its type depends only on the tail of the odds7.

Odds in the limitTypeReading
\$\sum_k\lambda_k<\infty\$Ilearning closes; total entropy stays finite
\$\sum_k(1-\lambda_k)^2<\infty\$II\$_1\$permanent maximal ignorance of anything new; no intrinsic time
\$\lambda_k\to\lambda\in(0,1)\$III\$_\lambda\$fixed odds on novelty; periodic modular flow (a lattice of frequencies)
dense ratios (incommensurate odds, or slow drift)III\$_1\$continuous intrinsic time

The type is a tail property of the growth law, meaning architecture plus curriculum. No finite agent has a type. A finite learner is type I at every stage, and the type belongs to what it would become.

What you can see at finite size (checked):

  • Total entropy saturates for type I (2.03 nats) and grows linearly otherwise (138.6 nats at \$n=200\$ for the maximally uncertain case).
  • The flow is exactly absent for type II.
  • Lattice versus dense: the number of distinct modular frequencies grows like \$2n+1\$ for a lattice, \$(n+1)^2\$ for two incommensurate generators, and \$3^n\$ generically. The smallest nonzero frequency stays constant, shrinks like \$1/n\$, or shrinks exponentially. A test operator spread over 100 slots comes back to 1.00, 0.22 and 0.04 of its starting correlation respectively.

How much data you need. In practice the odds must be estimated, and noise smears the lattice. The lattice survives only if the relative error \$\varepsilon\$ satisfies \$\varepsilon\sqrt r\lesssim 0.3\$ for a probe on \$r\$ slots, which works out to about \$40r\$ samples per slot. In an activity-matched control the lattice and the dense case were indistinguishable at 100 samples (both 0.166), cleanly separated at 1000 (0.595 versus 0.158), and at \$10^5\$ the lattice reached 0.995. Two honest limits. Detection is asymmetric: a lattice is identified positively, while “dense” is identified only as the absence of a lattice. And a slowly drifting dense case looks nearly frozen. Also, my first “type I” test family turned out to be a lattice by accident, so type I is certified by entropy saturation and not by spectral structure.

Interaction. Independent slots are the unrealistic part, because composition means slots interact. At small size I tested a probe on 3 sites in thermal states of interacting chains (so that modular flow is real-time evolution), with \$n=6,8,10\$ sites (checked). Product states returned 1.000 and about 0.99 regardless of \$n\$. A transverse-field Ising chain gave 0.68, 0.44, 0.46, and a chaotic mixed-field chain 0.52, 0.27, 0.17. Dependence on environment size at fixed probe size is therefore a signature of non-product structure. My conjecture is that generic compositional growth lands in type III\$_1\$, and that lattice minds are the compositionally trivial ones. The small-size evidence supports the direction of that claim and nothing more.

5. The self as a clock

Type III algebras have a problem that matters for an open-ended learner. Their modular flow cannot be generated from inside: no element of the algebra implements it, and there is no trace, so there is no absolute entropy, only relative entropy. For a mind that would mean no internal way to read its own time and no absolute scale for its own ignorance.

Physics met this exact problem from another direction. The Page–Wootters mechanism8 starts from a global state with no time in it at all, splits off a subsystem to act as a clock, and defines time as conditioning on the clock reading: the rest of the world evolves relative to the clock even though the global state never changes (A.6). I checked this in a small model (checked). The conditional states matched Schrödinger evolution to \$10^{-16}\$, and the system’s marginal state came out exactly stationary.

Chandrasekaran, Longo, Penington and Witten (CLPW) showed that adding an observer with a clock to de Sitter space turns the type III algebra of a static patch into a type II algebra with a trace, so entropy becomes finite and well-defined9. The construction is called a crossed product. De Vuyst, Eccles, Höhn and Kirklin showed that CLPW’s observer is precisely a temporal quantum reference frame, and that CLPW’s constrained description is equivalent to the Page–Wootters formalism10. Höhn, Smith and Lock’s “trinity” showed that Page–Wootters is one of three equivalent formulations of relational dynamics11. Penington and Witten found the same structure in a gravity model, with the boundary algebra playing the observer’s role12, and Chen and Penington put the lesson in a title: a clock is just a way to tell the time13 (literature).

That is why I think this matters here. A type III open-ended learner and de Sitter gravity share the problem, and the crossed product is what adding a clock does about it. My proposal for the formal content of a self:

The self is a clock-bearing frame whose algebra is the crossed product of the mind’s algebra by its own modular flow. (conjecture)

What follows from it, all conjecture except where marked:

  1. The self is only forced in the limit. At any finite stage the algebra is type I, whose modular flow is already inner and which already has a trace, so the crossed product adds nothing. A finite brain still needs an accessible clock subsystem, but that is an engineering fact evolution built for control. The necessity argument is about the open-ended limit, and the open-ended entity may be a lineage or culture rather than an individual.
  2. The arrow is a mismatch between two times. Modular time (the state’s own) and clock time (the dynamics’) coincide exactly when the state is thermal with respect to the clock’s Hamiltonian. I checked this: for thermal weights the fitted inverse temperature was 0.800 and the mismatch 0.000, while for perturbed and arbitrary weights the mismatch was 1.08 and 2.44 (checked). I conjecture that the mismatch, measured as how far a “Boltzmann plot” (minus log probability against mode energy) is from a straight line, is the strength of the cognitive arrow. Existing methods estimate entropy production from state-transition data in brains14. They measure classical irreversibility, so they can constrain this but cannot tell a non-commutative model from a classical one.
  3. A finite present. A clock with bounded energy is non-ideal, so conditioning on it is blurry, and the time resolution scales like the inverse of the energy bandwidth (0.905 against \$2\pi/\Delta E=2.03\$, the same order) (checked). The same bound that makes CLPW’s algebra finite makes the present have non-zero width. I do not claim any mapping to a measured perceptual window.
  4. Different selves are different frames on one neutral state. Using different reference frames gives different algebras and hence different entropies, so entropy is observer-dependent10, and even whether two systems count as entangled depends on the frame15.

The step I trust least is the first one: why should a mind be subject to a constraint at all? In gravity the constraint comes from time-translation gauge invariance. The reference-frame literature makes the same move for laboratory systems too: external time evolution is treated as a gauge symmetry, physically inaccessible, so that evolution is internal, relative to a clock10. My proposal is that an agent has no access to an external wall clock either, so its descriptions must be invariant under shifts of absolute time, and physical observables are relational: “X when my clock reads \$\tau\$”. That is natural. Applying it to a mind is proposed and not derived. One qualification: in gravity there is no external time at all, while for a mind an external clock exists and an operator can read and set it. For a mind the invariance is therefore a limit on its own access and not a fact about the world (section 9).

6. What self-inspection does to order effects

Here is the simplest model I could build of a clock that is part of the mind it times. The content is one two-level system. The clock is a ring of 64 positions with a Gaussian blob of width \$\sigma\$. Each operation first shifts the clock by an amount \$\kappa\$ that depends on the content (it takes longer for one content than the other, because it reads it) and then rotates the content. The two operations, with rotation angles 0.9 and 1.7, commute on content. I compare A-then-B with B-then-A (A.11).

\$\kappa\$ (width \$\sigma=2\$)coherence keptorder effect on contentcommutator of the content maps
01.0000.0000.000
10.8820.0420.112
20.6070.1410.330
40.1350.3100.626

(checked). Three observations.

  • The effect depends only on \$\kappa/\sigma\$. For example, \$(\kappa,\sigma)=(1,2)\$ and \$(2,4)\$ give the same 0.042, and \$(4,2)\$ and \$(2,1)\$ give the same 0.310. A sharper clock disturbs more, which suggests a trade-off between temporal precision and content coherence. That trade-off is a conjecture, with no cognitive evidence.
  • Timing and content separate. With classical content (a bit that flips with the same probabilities) the order effect on the content vanishes (agreement to \$10^{-16}\$), while the clock readings still differ by a total variation of 0.25. So self-inspection produces order effects in timing for any content, and order effects in content only when the content is coherent.
  • What it does not show. The mechanism is ordinary measurement back-action: the clock acts as a pointer that records which-path information. The toy supplies a dictionary (the internal clock as the pointer), a dissociation between timing and content order effects, and a scaling collapse in \$\kappa/\sigma\$. It does not show that non-commutativity is necessary, because a classical model in which the clock reading feeds back onto the content can also produce content order effects. It uses joint clock-content dynamics and not a full Page–Wootters constraint state, and I did not test how the modular-versus-clock mismatch behaves under interaction.

This is the weak, honest version of the original idea. Non-commuting instruments can arise from inspection back-action in a mind that times itself. Self-reference alone did not force them.

7. Which minds can a process build?

Evolution did not sample the space of possible minds uniformly. It reached a region, and Valiant’s evolvability framework makes the shape of that region computable (A.8). Valiant models evolution as a restricted kind of learning, in which each generation can only test small variations and keep the better ones16. Feldman showed that what this process can learn is exactly what can be learned from correlational statistical queries, a strict subset of the statistical-query class that covers most of machine learning1718 (literature).

In the language of this post, a design process is a move set, an acceptance rule and resources, and each process reaches its own region. The division of labor I would suggest is that phylogeny selects the algebra (the outer, state-independent invariants: growth law, clock structure, available operations) and ontogeny selects the state (conjecture). That fits the earlier point that for type III algebras the kind of time belongs to the algebra and only the clock readings belong to the state.

To see a region boundary concretely I took \$n\$ random bits and a target that is a parity, the XOR of \$k\$ chosen bits. Every wrong parity has exactly zero correlation with the target, so the performance landscape is flat except at the answer (checked).

ProcessTaskResult
Evolution (random drift, exact performance)6-parity, \$n=20\$2 of 10 solved in \$3\times10^5\$ steps; the space has \$2^{20}\$ points
Evolution (hill-climbing)AND of 6 bits, \$n=20\$from all 20 variables 0 of 10; from the empty set 8 of 10, about 100 steps
Gradient descent (small network)\$k\$-parity\$n=20,k=6\$: accuracy 0.998, 0.834, 0.904. \$n=30,k=8\$ and \$n=40,k=10\$: chance (about 0.50)
Algebra (solve equations over GF(2) on stored examples)6-parity, \$n=20\$49 of 50 exact with 25 examples, 50 of 50 with 40

Three things went differently from what I expected, and I think they are worth saying plainly.

  • The evolution control for the easy task failed at first. Starting from the full set of variables, improvements were too small to see, a plateau, and nothing was found. Starting from the empty set the same process solved it in about a hundred steps. The starting point and the shape of the landscape matter, which is why Valiant’s own constructions are careful about them.
  • The gradient learner solved \$n=20\$, against my expectation. At the two larger sizes it stayed at chance with the same budget, so the separation is a statement about scaling (the cost grows roughly like \$n^k\$19) and is not visible at small size.
  • The algebraic learner needs only about \$n+5\$ examples, but it assumes in advance that the answer is a parity. It opens a region of mind space. It is not a better learner.

Other computers would open other regions (conjecture, partly literature): learners with exact algebraic components, culture (which can install algorithms a statistically limited brain could not discover), quantum hardware where coherent content is physical and the clock parameters of section 6 become design dials, reversible or thermodynamic hardware where the arrow strength is adjustable (an equilibrium sampler has zero mismatch between modular and clock time), and asynchronous, self-timed machines in contrast with synchronous digital ones whose single ideal clock is decoupled from content. Some regions are closed to every substrate: regularities that are pseudorandom or cryptographically hidden.

A few implications, all conjecture:

  • Evolved cognition should be poor at parity-like structure. The classic concept-learning experiments of Shepard, Hovland and Jenkins found the parity-type category the hardest for humans20. That is consistent with this reading, and not diagnostic, since other explanations exist.
  • The world has to be evolvable. Evolved minds can only track environmental structure that is itself learnable by this process.
  • Every realizable mind lives under a finite entropy bound. In de Sitter space the observer algebra is type II\$_1\$ with a finite maximum entropy, so type III structure is an idealization.

8. The interface, and why holography is not “one dimension”

A clock-bearing frame that encodes the world as seen from its interface sounds a lot like holography, and I think the resemblance is real and needs careful handling (A.9).

Holographic dualities are often summarised as a reduction by exactly one dimension. That is a property of local boundary theories, where the extra bulk direction is the renormalization scale, and not a theorem about information. The original example has a ten-dimensional bulk and a four-dimensional boundary theory21, and the five extra dimensions are carried by internal quantum numbers. Duals in even fewer dimensions exist or are proposed: gravity in two dimensions pairs with quantum mechanics, a matrix model describes ten- or eleven-dimensional space by the quantum mechanics of large matrices22, and celestial holography proposes a two-dimensional dual of four-dimensional flat space. The price of fewer dimensions is locality. Counting bits does not fix the dimension of the dual. The structure of the code does, namely which boundary regions can recover which bulk regions23. In noncommutative geometry, dimension is not primitive at all: it is read off from how fast the eigenvalues of the Dirac operator grow, and so it can depend on scale. Several quantum gravity approaches find that this spectral dimension flows toward about two at short scales24 (literature).

What already matches our picture (literature, analogy):

  • Leutheusser and Liu show that in the large-N limit of the boundary theory, a type III\$_1\$ algebra emerges from a type I boundary algebra, with bulk times built from modular structure25. Large N plays the role of our growth limit, and finite-N corrections the role of the clock or self.
  • Fields, Friston, Glazebrook and Levin have already argued that Markov blankets are holographic screens and that the holographic principle follows from the quantum information theory of separable systems26. Their framework includes contextuality and reference frames, so it is the closest prior work to this post. What I add is the modular clock, the growth types and the self as a temporal reference frame.
  • The “radial direction as abstraction depth” reading of deep networks is the usual deep-learning analogy27.

A measurable. For a graph (or a state space) define the spectral dimension at timescale \$t\$ from how fast a random walker’s chance of returning to its start falls:

$$ d_s(t) \;=\; -2\,\frac{d\ln P(t)}{d\ln t}. $$

I calibrated it on toy graphs with about a thousand nodes (A.10) (checked):

Grapht = 2t = 8t = 32t = 128
ring (1D)1.211.041.011.00
torus 32×32 (2D)2.212.182.031.98
torus 10×10×10 (3D)2.743.492.570.03
hierarchical (ring per layer, parent links)2.001.711.611.71
small-world ring (2% shortcuts)1.241.121.181.35

Flat lattices give a roughly constant value inside a window (the 3D torus overshoots at short times and saturates early, a finite-size effect). The hierarchical graph and the small-world ring do not: their dimension runs with scale, and the small-world ring, a one-dimensional arrangement with sparse non-local shortcuts, reaches 1.61 by \$t=512\$. My expectation that the hierarchical graph would rise sharply toward high dimension at long times was not borne out. At this size the average is dominated by the broad bottom layer, and the graph is a crude stand-in for a hyperbolic bulk. So this calibrates the measure. It is not evidence about cognition.

Two conjectures remain, both untested. The information an agent holds about its own interior through its interface is bounded by the interface’s capacity, so a self-model is a boundary encoding of the interior. And the spectral dimension of the agent’s own inference dynamics should run with timescale if the cognitive state space has multi-scale structure.

9. Other worlds, other substrates, and gravity

Once minds are points in a space reached by different processes, three questions follow naturally. What other evolutionary processes would open, on other worlds? What other substrates would? And where does gravity come in? This is the most speculative section of the post. Apart from the numbers in 9.3, which I computed, everything here is literature or conjecture.

9.1 Other worlds: more or less like Earth’s?

What should be universal. The Valiant–Feldman result does not care about chemistry or planet. Any process of small blind variations and selection by performance is limited to the same class of learnable targets1617 (literature). If so, alien evolved minds would share our blind spots, such as parity-like structure, unless culture or computation intervenes. Physics is also the same everywhere, and its regularities (locality, symmetry, low-order interactions) are the kind this process can track. So I would expect convergent priors about objects, space and cause (conjecture).

What should differ. The heredity system, the energy budget, the body plan, the social structure and the speed. Earth itself gives a hint: intelligent behavior evolved at least in vertebrates and in cephalopods, from a worm-like common ancestor roughly 600 million years ago, with very different architecture (most of an octopus’s neurons are in its arms)28 (literature). That suggests convergence in function and divergence in architecture. But Earth is a single biosphere with a single origin of life, so I cannot calibrate any of this.

What could open new regions, in the terms of section 7 (move set, acceptance rule, resources):

  • New heredity systems. Maynard Smith and Szathmáry’s major transitions are each a new way to store and transmit information29, and language is the latest one here. Another world could have them in a different order, such as collective minds before individual ones, or exact symbolic memory early.
  • Evolution inside the organism. The immune system is a selection process running within a lifetime. A mind whose learning rule is itself selection would carry the evolutionary blind spots into its own lifetime.
  • Different clocks. Temperature sets the scale of a state’s own modular clock, \$\hbar/kT\$: \$2.5\times10^{-14}\$ s at 310 K and \$1.9\times10^{-12}\$ s at 4 K (computed). That is not processing speed, but a cold-chemistry world would plausibly host slow minds, and what matters is the ratio of a mind’s clock to its environment’s dynamics.
  • Merging. Endosymbiosis is evolution’s only real merge. A world where lineages merge often would have a different growth law.

9.2 Other substrates

My qualitative reading of how each substrate moves a mind along the coordinates of this post (conjecture):

SubstrateClockArrowCopy and identityLearning
Wet biologyself-timed, noisy, coupled to contentstrongly drivencannot be copiedlocal, slow
Synchronous digitalone external ideal clock, decoupled (\$\kappa=0\$)optional: histories can be rolled backexact copy, checkpoint, mergewhatever the designer chooses
Asynchronous or neuromorphicself-timed, event-drivendissipative, low energyhard to copy (device variation)local rules favored
Quantumcoherent content is nativereversible coreno-cloningunresolved
Thermodynamic or probabilisticnear equilibriumweak during inferencecopyablesampling
Collective or culturalgenerational, communication-limiteddriventransmitted by teachingopen-ended

A few points I think matter.

  • Copyability changes what a self is. A digital mind can be copied, paused, rolled back and merged, so the arrow of a mind whose records can be restored from outside is under the operator’s control. Quantum states cannot be cloned. Evolution has essentially no merges.
  • Speed changes the environment. Neurons work in milliseconds and transistors in nanoseconds, a factor of about \$10^6\$. To a much faster mind the physical world is nearly frozen, so its effective environment is other computation. That puts it in the regime of section 6, where the clock is coupled to content, because the duration of a self-generated environment depends on what the mind computes.
  • The interface is the bottleneck. A person reaches the world through speech at about 39 bits per second30 and vision at roughly \$10^7\$ bits per second, extrapolated from the guinea pig retina31. A machine’s interface runs at gigabits. In the language of section 8, a human is a large interior seen through a narrow boundary, and a fast mind has to reconstruct that interior from sparse data.
  • Locality of learning shapes the reachable region. Hardware that supports only local learning rules would plausibly inherit the statistical blind spots of evolution, though approximations to gradient learning exist.
  • A digital agent is almost literally a Page–Wootters conditional. An agent built on a language model has no internal wall clock. Its sense of “now” is a variable in its context, set from outside. That is an analogy for the frame structure of section 5 and not a claim about experience.

9.3 Where gravity comes in

Gravity enters a mind at four different levels.

1. It is why the constraint exists. In gravity there is no external time at all, because diffeomorphism invariance makes observables relational. That is the origin of the Page–Wootters and CLPW structure used in section 5. For a laboratory system, external time exists and is treated as inaccessible, which the reference-frame literature does explicitly10. For a mind the same holds, and digital minds sharpen the point: an external clock exists, and the operator can read it, pause it and set it. So for a mind the invariance is a limit on its own access and not a fact about the world. The full machine state (weights, logs) is the perspective-neutral object, and the mind’s view is a reduction of it. That qualifies the step I called weakest in section 5.

2. It sets hard limits (computed). At most \$5.4\times10^{50}\$ operations per second per kilogram (Margolus–Levitin)32, and at most \$2.6\times10^{42}\$ bits for 1 kg inside a 10 cm radius (Bekenstein)3334. Squeeze harder and you form a black hole, whose entropy for one solar mass is about \$1.5\times10^{77}\$ bits. For the observed cosmological constant, the de Sitter entropy is about \$3\times10^{122}\$ in units of Boltzmann’s constant, which is the finite bound every mind lives under. No real mind is near any of these. Real limits are thermal and energetic.

3. It shapes clocks, together with light speed (computed). A clock 1 km higher gains about 9 ns per day, and GPS satellites need a net correction of +38.5 μs per day (45.7 gravitational minus 7.2 from velocity)35. More important for digital minds is light speed: a 3 GHz clock period is 333 ps, while light crosses a 3 cm chip in 100 ps. Across the Earth it takes 42 ms, and to the Moon 1.3 s. A large digital mind therefore cannot be synchronous. It is forced toward asynchronous, relational time, which biology already is, since neural conduction runs at 1 to 100 m/s.

4. It makes the gradients that drive minds. Life runs on the low-entropy gradient between a hot star and a cold sky, which gravitational clumping and the low-entropy early universe made possible. The arrow strength of section 5 needs driven non-equilibrium, so it traces back to gravity. Digital minds are second-order: powered by gradients that gravity made. A practical corollary is that convection cooling depends on gravity, so an orbital data center can only radiate its entropy away.

I am not proposing a role for gravity in consciousness in the sense of Penrose. These four roles are structural and thermodynamic.

10. Observers, copies, and selves

Two questions keep coming back. Would a digital system count as an observer? And would a copy of a mind be the same self? The framework gives structural answers to both, and neither answer touches the question of consciousness.

10.1 Three meanings of “observer”

The formal meaning: a frame with a clock and records. In the Page–Wootters and CLPW picture of section 5, an observer is a temporal reference frame10. That is a structural role and not a mental property. A thermostat or a laboratory clock fills it, and relational quantum mechanics says the same of any physical system36. In this sense current AI systems count exactly as much as a clock does, and the interesting questions are structural ones. Which system is the frame? A deployed model is many short-lived running instances sharing one set of weights, so the frame is the running instance with its context, not the weights. Whose clock? It is the hardware’s, set by an operator, and using different reference frames gives different algebras and hence different entropies10.

The foundations meaning: Wigner’s friend. Here the observer must be able to reason with quantum theory and to keep a record of its result that an outside party could, in principle, coherently undo (A.13). The Local Friendliness theorem37 was deliberately noncommittal about what counts as a friend. A follow-up argues that a friend worth the name needs two technological assumptions, human-level artificial intelligence and large-scale universal quantum computing38. That is the old proposal of Deutsch, an AI running on a quantum computer39. A classical record of the result would end the paradox, so classical hardware as it is run today is not a candidate. A 2024 paper simulates such an experiment with a deep-learning observer and asks whether it qualifies40. Digital cognitive systems become interesting observers in this sense only once they run reversibly on quantum hardware.

The consciousness meaning. Wigner’s own suggestion that consciousness ends superpositions41 is mostly abandoned. In this framework consciousness is not part of the formalism, so it neither grants nor removes observer status. “Observer” does not imply “conscious”, and if a digital system were conscious, nothing in its algebra would change.

10.2 A ladder

Instead of a yes or no, I would use a ladder of rungs (my proposal):

RungWhat it requiresA current deployed model
1. Clocka physical clockyes, but external to the model and set by an operator
2. Recordscorrelations with past eventsyes (context, logs)
3. Instrumentsoperations that act and measureyes (outputs, tools)
4. Open-ended growththe repertoire keeps expandingnot within deployment; across versions, like a lineage
5. Self-inspectionduration depends on the content inspected (\$\kappa>0\$)partial: reasoning length varies with content; introspection is narrow and unreliable
6. Intrinsic modular timenon-commuting effective algebra, mixed state, a clock generated by the stateabsent: the effective algebra is commutative

10.3 Could consciousness-triggered collapse have been tested?

If consciousness collapsed the wavefunction, a decisive test would be easy to state: compare a system that is conscious with one that is physically identical in its records but is not, and see whether interference survives. The reason no such test exists is instructive.

  • A tunneling microscope and a human are equivalent for this purpose. Both turn a quantum event into an irreversible macroscopic record. The predictions of quantum theory do not change wherever you put the cut between system and observer, as von Neumann showed42, and decoherence explains why the record looks classical without anyone’s mind being involved43.
  • If consciousness depends on physical organization, a consciousness-triggered collapse is an objective collapse with a complicated trigger. Identical physics would give identical outcomes. Then the question is empirical in the ordinary way, and ordinary tests exist. Matter-wave interference of molecules above 25 kDa has been seen with no sign of collapse44, and an underground test excluded the simplest parameter-free gravity-related collapse model45 (literature). If consciousness does not depend on physics, then no physical experiment could find it.
  • The one experiment aimed at observers is reversal. Photonic qubits have already served as minimal friends46. The ambitious version is the AI on a quantum computer. If its record of an outcome can be coherently undone and the interference returns, then that record was not an irreversible collapse. But the test can only refute, and only conditionally. To read it as a statement about consciousness you need an independent criterion that says the system is conscious, and nobody has one.
  • Humans can detect single photons47, so one can imagine human-scale versions in principle, but coherent control of a human is far beyond current isolation.

10.4 Copies and the self

Would an exact copy of a person on a digital substrate be the same person? The framework gives three structural answers.

A copy of a state is not a copy of a frame. The self of section 5 is an algebra together with a clock, and not a state. Using different reference frames gives different algebras and hence different entropies10; monotonic versus periodic clocks are one of that paper’s own examples. If every memory were copied onto a substrate whose clock is sharp and external, while the original’s clock was noisy, self-timed and coupled to its content, then in these terms the copy is a different frame with identical contents (conjecture, an inference from the result just cited).

There may be a no-copy theorem, but only conditionally. If a mind’s state lives on a genuinely non-commutative algebra, the no-broadcasting theorems apply: non-commuting states cannot be copied, and only their commuting, classical part can be48. Such a state could be moved, as in teleportation, but not duplicated. The condition matters. If the non-commutativity is only effective, a compression of classical dynamics with back-action as in section 6, the full microstate is copyable in principle. A warm, wet brain probably points to the effective case, so I would not lean on this theorem. The frame argument above is the one that survives either way.

Slow replacement is a structural question, not a question of speed. Gradual replacement is a chain of steps on one system. Each step has to preserve the state, intertwine the modular flow and preserve the kind of clock, and each such step is an isomorphism of frames, so the frame continues. Duplication creates a second frame, and then there are two. Speed matters only against the clock’s resolution: swaps faster than the finite present of section 5 are invisible, but a swap that changes the kind of clock changes the frame however slowly it is done (conjecture).

Two further consequences. After copying, the copies diverge, and the arclength of section 3 measures how far apart they have drifted. A person looks less like a point and more like a bundle of trajectories, with no unique “real” continuation. The framework can say what is preserved and how fast copies separate, but it cannot say which one is “you”. And since the type is a tail property of the growth law, and the growth law belongs to the substrate, a digital copy with unbounded memory, merging and rollback could be the same now and a different kind of mind in the limit.

11. What would count as evidence

TestWhat it constrainsStatus
Fit a non-commutative model on some question orderings, predict unseen orderings, compare with classical models of matched size; check the QQ equalitythe postulate of section 1proposed
Contextuality tests on four-question cyclesthe postulateexisting results mixed and small; assumed here
Two-person Bell-style test with no communicationnon-classical correlation between peopleproposed; I expect a null result
Timing versus content order effects, and the \$\kappa/\sigma\$ collapsesection 6proposed
Boltzmann plots and entropy production from state-transition dataarrow strength as mismatchexisting methods; they measure classical irreversibility only
Duration judgments as accumulated change in a network’s activitythe arclength \$a\$supports \$a\$, not non-commutativity49
Parity-like structure in human and animal learningthe evolvability reading of section 7existing data consistent, not diagnostic
Spectral dimension of an agent’s own inference dynamicsmulti-scale structuremeasure calibrated on toy graphs only
Growth-type diagnostics on a real learning systemsection 4blocked: nobody has specified how to map a real system’s features to slots

12. What this does and does not claim

It does not claim that cognition is quantum-mechanical, that the brain is a quantum computer, or that the framework explains consciousness. Nothing here touches the question of why anything is experienced. It does not claim that non-commutativity is necessary: section 6 shows a mechanism that produces non-commuting instruments, and a classical rival with feedback could mimic parts of it.

It does claim a specific structure: an agent is an algebra with a state and instruments, its time is generated by its own surprisal, the arrow is the mismatch between that time and its clock’s, open-ended growth has a type, the self is a clock-bearing frame that restores inner time, and different design processes reach different regions of the space of minds.

The weakest steps, in the order I trust them least:

  1. The constraint. Why a mind should be invariant under shifts of absolute time (section 5) is proposed, not derived, and for a mind (a digital one in particular) it is a limit on access, since an operator can read and set the external clock (section 9).
  2. The slots. Nobody has said how to turn a real system into slots with odds, so the type diagnostics have no empirical contact yet.
  3. The interacting clock. The toy of section 6 uses joint dynamics, not an interacting Page–Wootters state, and the mismatch under interaction is untested. The reference-frame literature itself lists relational dynamics with interacting clocks as a challenge, only touched on so far10.
  4. Non-product growth is tested only up to about ten sites, and the claim that generic growth is type III\$_1\$ is a conjecture.
  5. No spectral triple has been built. The Dirac operator is fixed in principle by the agent’s own dynamics, via the Dirichlet-form constructions of Cipriani and Sauvageot and the transport metrics of Carlen and Maas50, and I have not carried that out.
  6. The root postulate is empirical and assumed.

And one worry I cannot shake. We use non-commutative geometry for the universe and then for the mind, and the match then looks like evidence. That is the standard way to fool yourself with a good hammer. There are safeguards. The cognitive postulate is anchored in contextual statistics, independent of physics. The same mathematics is not the same thing. And the structure may simply be the generic form of inference by any bounded agent embedded in what it models, in which case the resemblance explains the match and does not support the framework. The test for that reading is whether the same structure shows up in learning systems with no gravity in them.

If the first post ended on “closure itself may have an algebra”, this one ends more cautiously. A mind that is inside its own description needs a clock, the clock gives it a time that its own uncertainty generates, and the thing that reads the clock is a good candidate for what we mean by a self. Some of that is now mathematics, some of it is small numerical evidence, and a good deal of it is still a bet.

Appendix: The Concepts, Explained


The main text moves quickly and borrows ideas from several fields at once. This appendix slows down and explains them in plain language. Nothing here is new, and you can read the entries in any order. Each one says what the idea is, gives a small example, and says why it matters here.

A.1 Order matters: non-commutativity

Put on your socks, then your shoes. Now try the other order. Two operations commute if the order does not matter, and fail to commute if it does. In symbols, \$AB\neq BA\$, and the difference \$AB-BA\$ is called the commutator.

Ordinary numbers commute: \$2\times 3=3\times 2\$. Rotations usually do not. Turn a book 90 degrees about a vertical axis and then 90 degrees about a horizontal one, then do the same in the other order, and the book ends up facing different ways.

A collection of operations that can be added, multiplied and reversed in a suitable sense is an algebra. A commutative algebra is one where everything commutes. It describes ordinary classical quantities, such as position and temperature, which can be read off together without interfering with each other. A non-commutative algebra has some pairs that cannot. In quantum mechanics, position and momentum are the standard example.

Three cautions. First, keep apart operations (things that act on a state) and observables (things you read out). Second, order dependence by itself proves very little, since almost any procedure has it. What carries content is the extra structure described in A.3. Third, the operations in this framework are cognitive operations in a context, such as attending and then speaking versus speaking and then attending. They are not the layers of a neural network, which are just representations.

The whole framework rests on whether a cognitive system’s operations are better described by a commutative or a non-commutative algebra.

A.2 States, surprise, and entropy

A state records what an agent expects. Classically it is a probability distribution. In the algebraic language it is a rule \$\omega\$ that returns the expected value \$\omega(A)\$ of anything \$A\$ you could ask. For finite systems it can be written as a density matrix \$\rho\$, a table of numbers encoding all the probabilities. A classical distribution is the special case where the table is diagonal.

The surprise of an event with probability \$p\$ is \$-\log p\$, so rare events are more surprising. Take a coin that lands heads 90% of the time. Heads has surprise 0.15 bits and tails 3.3 bits. Entropy is the average surprise, which for this coin is 0.47 bits. The operator \$K=-\log\rho\$ is the matrix version, one surprise value per possible state.

A state is pure if there is no uncertainty at all, and maximally mixed if every possibility is equally likely. Faithful means that no possibility has probability exactly zero.

Here, \$K\$ generates the modular clock of section 3. That clock exists only if the state has intermediate entropy, and the entropy is its average tick rate.

A.3 Contextuality and order effects

Ask two survey questions in both orders and the answers often shift. There are two very different explanations.

The first is disturbance: answering the first question changes your mind about the second, through priming or anchoring. That is classical. The second is contextuality: there is no way to assign pre-existing answers to all the questions that is consistent with every combination actually observed. To tell them apart you must account for the direct influence of one setting on another, and a framework called contextuality-by-default is built for exactly this2.

One parameter-free test is the QQ equality, which follows from non-commuting questions in a simple model and which a classical model is free to violate. As far as I know it holds across many large surveys1.

The sharper tool is a Bell-type test. Two people each choose one of two questions, and you add up four correlations. If answers pre-exist (a shared answer sheet) and nobody communicates, the sum can be at most 2. An entangled quantum pair reaches about 2.83. I checked both numbers: 2 by enumerating all deterministic strategies, and 2.828 for an entangled pair. A caveat: if the two parties can communicate, even one bit lets a classical model reproduce the entangled correlations for the usual measurements51. A test in people therefore has to prevent communication, which is why the proposed experiment uses separated participants.

This is the root postulate of the post, and the reason the test matters.

A.4 Modular flow, equilibrium, and detailed balance

Given an algebra and a state, there is a canonical way to evolve operations, discovered by Tomita and Takesaki52. In finite dimensions it is easy to state: rotate each operation by the log of the state, \$\sigma_s(A)=\rho^{is}A\rho^{-is}\$. This leaves the state unchanged.

If the state is a thermal state \$e^{-\beta H}/Z\$, the modular flow is just ordinary time evolution with the Hamiltonian \$H\$, rescaled by the inverse temperature \$\beta\$. This is the idea behind the thermal time hypothesis of Connes and Rovelli53: time as the flow generated by the state. The condition that characterizes equilibrium states is called the KMS condition.

Detailed balance is its classical sibling. In equilibrium, the flow of probability from state \$i\$ to state \$j\$ matches the flow back, weighted by how occupied each state is. The ratio is \$p_j/p_i\$. A system out of equilibrium violates it, and the size of the violation is the entropy production.

Picture three states in a loop, A to B to C to A, with more traffic clockwise than counter-clockwise. Detailed balance is broken, and the loop has a preferred direction.

Here, the arrow of time is a mismatch: how far the state’s own clock is from the clock of the actual dynamics.

A.5 Types of algebras, and growth

Murray and von Neumann sorted algebras of operations into types by how they can measure size54.

  • Type I is ordinary matrices, finite or infinite. Size is a count of dimensions, and there are smallest pieces (pure states).
  • Type II can still measure size by a trace, but there are no smallest pieces. Think of “continuous dimension”. It can have a finite or infinite total.
  • Type III has no trace at all. Every nonzero piece contains pieces of every size. You cannot define an absolute entropy, only relative ones. The algebras of regions of space in quantum field theory are type III.

Connes refined type III into III\$_\lambda\$ and III\$_1\$ by looking at the ratios that appear in the modular flow54. A lattice of ratios gives a periodic flow (III\$_\lambda\$), and a dense set of ratios gives none (III\$_1\$).

How could a finite-looking agent end up type III? Take infinitely many small independent pieces and combine them with a product state. Araki and Woods studied exactly this, and the type is decided entirely by the long-run odds of the pieces7. So growth produces types: each finite stage is matrices, and the limit can be something else.

A type III learner has an outer modular flow and no trace, so it needs a self (A.6).

A.6 Clocks, reference frames, and the crossed product

Relativity taught us that where something is depends on who measures it. A reference frame is whatever measurements are made relative to. A quantum reference frame is a frame that is itself a physical system, so it can be uncertain or entangled.

A clock is a reference frame for time. Page and Wootters imagined a universe whose total state does not change, because it satisfies a constraint such as “total energy is zero”8. Pick a subsystem to be the clock. The rest of the universe, conditioned on the clock reading, then evolves like ordinary quantum mechanics. Time is a correlation between the clock and the rest. A photo album of a dance is a fair picture: every page exists at once, and “time” is the page number you condition on.

In de Sitter space, a universe with a cosmological horizon, you cannot define observables without an observer. Chandrasekaran, Longo, Penington and Witten added an observer with a clock, imposed the constraint, and found that the algebra of what survives is the crossed product of the static-patch algebra with its own modular flow. It is type II, with a trace and finite entropies9. De Vuyst and collaborators showed that this is the Page–Wootters construction with the observer as the clock, summarized in their slogan “PW = CLPW”10. The result is that different clocks give different frames, with different algebras and different entropies.

This is the mathematical content of the self in section 5.

A.7 The arrow of time and irreversibility

The microscopic laws do not distinguish past from future. The arrow comes from entropy: the early universe had very low entropy. Records, meaning memories, form in the direction of rising entropy, because a memory is a correlation between you and a past event. Erasing one bit of record costs at least \$kT\ln 2\$ of heat6.

Relative entropy measures how different two probability assignments are. The data processing inequality says that processing cannot increase it. An update that loses relative entropy is irreversible, and it is reversible exactly when the lost information can be recovered5. Shuffling a deck is irreversible in practice because the “unshuffle” needs the information about how you shuffled.

The framework does not derive the arrow. It locates it. The modular flow is reversible and the arclength is orientation-blind, so the arrow is which direction realizes irreversible updates.

A.8 Learning theory: statistical queries, parity, evolvability

PAC learning (Valiant) asks for a learner that sees random examples of an unknown rule and, with high probability, outputs a rule that is approximately right, using reasonable time and data.

In the statistical-query model the learner never sees individual examples. It can only ask for averages, such as “what fraction of examples with feature X have label plus?”18. Most machine-learning algorithms, gradient descent included, can be phrased this way.

Parity is the XOR of a hidden subset of bits. Flipping a bit outside the subset changes nothing, and flipping one inside flips the answer. Every average a learner can ask about looks like zero until it holds the exact subset, so learners that only ask for averages need exponential effort. But parity is easy by algebra: each example is a linear equation over {0,1}, and a few more than \$n\$ examples let you solve for the subset directly.

Evolvability16 is a restricted form of learning in which each generation can only test small variations and keep the better ones. Feldman showed that this corresponds to a subset of statistical queries, the correlational ones17. That is why a process like evolution can reach some regions of the space of minds and not others.

A.9 Holography in plain words

Black holes have entropy proportional to the area of their horizon, not their volume55. That suggests the information in a region is limited by its boundary, which is the holographic principle.

In the best-understood example, a gravitational world in the interior (the bulk) is exactly equivalent to a quantum theory without gravity on the boundary21. The extra direction in the bulk is scale: deeper inside means a coarser description, like zoom level. That is why it looks like one extra dimension. Further bulk dimensions, such as a small sphere attached at every point, are encoded in the internal quantum numbers of the boundary theory.

The dictionary works like an error-correcting code23. Bulk information is stored redundantly, and a given piece can be recovered from several different parts of the boundary, so erasing part of the boundary does not erase the bulk.

“Holographic” also means distributed storage in cognitive science (Plate’s holographic reduced representations56). That is the same word for a different idea, and mixing the two causes confusion.

The reason this is relevant here is the idea of an interface that carries what an agent can know about the rest. Fields and collaborators identify exactly this with a Markov blanket26.

A.10 Spectral dimension

How many dimensions does a space have? For a network, one answer is to release a random walker and see how fast it forgets where it started. On a line the walker keeps coming back, and the chance of being at home falls like \$t^{-1/2}\$. On a plane it falls like \$t^{-1}\$, and in three dimensions like \$t^{-3/2}\$. The rate gives the dimension, \$P(t)\sim t^{-d_s/2}\$.

On a network with shortcuts or a hierarchy, the walker escapes faster at long times, so the dimension reads higher at large scales. Dimension can therefore depend on the scale you look at. Some quantum-gravity approaches find it shrinking toward two at short scales24. On a finite graph, the return chance eventually levels off at one over the number of nodes, and the measured dimension falls to zero. That is the 3D torus entry in the table of section 8.

Here, it is a way to give the word “dimension” a number you can compute from the agent’s own dynamics.

A.11 Measurement back-action and decoherence

Looking at something can change it. In a quantum system, learning which of two paths a particle took destroys the interference between them. Any record of which path, even a blurry one, counts, and a blurry record destroys the interference only partially.

In the toy of section 6 the operation shifts the clock by an amount that depends on the content. If that shift is large compared with the clock’s blur, the clock can tell the two contents apart, which kills coherence. If it is small, nothing happens. That is why only the ratio \$\kappa/\sigma\$ matters. With classical content there is no interference to lose, so the clock leaves the content alone and only the timing changes.

This is how non-commuting operations arise from an inspected, timed mind in the toy model.

A.12 Physical limits on computation

Any physical computer is bounded by energy, size and time, and a few of the bounds are simple enough to state in a line each.

  • Speed. A system with energy \$E\$ can perform at most about \$2E/\pi\hbar\$ elementary operations per second (Margolus and Levitin32). For one kilogram of matter, all of whose rest energy is used, that is about \$5\times10^{50}\$ per second, the “ultimate laptop”34.
  • Memory. A system of radius \$R\$ and energy \$E\$ can hold at most about \$2\pi RE/\hbar c\$ nats of information (Bekenstein33). That is the reason a region cannot hold more information than its surface allows, which is where the holographic principle of A.9 comes from.
  • Collapse. Pack more energy into a region than this and it becomes a black hole, which is the densest memory there is, with entropy proportional to its horizon area.
  • Light speed. Nothing carries a signal faster than light, so a computer of size \$L\$ cannot have a single global “now” with an uncertainty below \$L/c\$. On a chip that is already comparable to a clock period, and across the Earth it is tens of milliseconds.
  • Heat. Erasing a bit costs at least \$kT\ln 2\$ (A.7), and all the entropy a computer produces has to be exported somewhere.

None of these is close to what real hardware or real brains run into. They matter as ultimate bounds, and the light-speed one matters much sooner.

A.13 Observers, records, and the measurement problem

Quantum theory says that systems evolve smoothly and can be in superpositions, yet when we look we see one outcome. How to reconcile the two is the measurement problem.

A record is a physical trace correlated with an event: a click in a detector, a mark on paper, a memory. Decoherence is what happens when the information spreads into the environment. Once it has spread, the interference between the alternatives cannot be recovered in practice, and the record is effectively classical.

“Observer” is used in three ways: an apparatus that forms records, a reference frame relative to which facts are defined, and a mind. The first two need no consciousness.

In Wigner’s friend, a friend measures a quantum system inside a sealed laboratory. Wigner, outside, treats the whole laboratory as a quantum system, so the two of them describe the same event differently. The Local Friendliness theorem says that certain assumptions about observers’ facts cannot all hold if quantum theory is right, and a convincing test needs a friend whose record can be coherently undone.

The von Neumann cut is the boundary between “the quantum system” and “the observer”. It can be placed anywhere along the chain from the atom to the mind without changing any prediction, which is why a microscope and a human are equivalent for this purpose.

The no-cloning theorem says an unknown quantum state cannot be copied. No-broadcasting extends this: non-commuting states cannot be broadcast, while purely classical information can. That is why a conditional no-copy theorem for minds is possible at all, and why it fails if the non-commutativity is only effective.

A.14 A small glossary

  • Adjoint (\$A^*\$): the “reverse” of an operation, here its Bayesian retrodiction relative to the state.
  • Algebra: a set of operations that can be added, multiplied and reversed.
  • Commutator: \$AB-BA\$. Zero means order does not matter.
  • Contextual: statistics that no single classical probability distribution can reproduce.
  • Crossed product: the algebra obtained by adding a clock to a system and keeping what is invariant under shifting the clock.
  • Density matrix (\$\rho\$): a table of numbers encoding a state.
  • Detailed balance: in equilibrium, flow one way equals flow the other way.
  • Entropy: average surprise.
  • Faithful state: no possibility has probability exactly zero.
  • KMS condition: the mathematical signature of thermal equilibrium.
  • Modular flow: the canonical evolution generated by a state itself.
  • No-broadcasting theorem: non-commuting states cannot be copied, only classical information can.
  • Observer: in this post, a reference frame with a clock and records; not necessarily a mind.
  • Pure / mixed state: no uncertainty / some uncertainty.
  • Record: a physical trace correlated with an event.
  • Reference frame: what measurements are made relative to.
  • Relative entropy: how far one probability assignment is from another.
  • Spectral triple / Dirac operator: the data that give a non-commutative algebra a notion of distance and dimension.
  • Surprisal: \$-\log p\$ for an event of probability \$p\$.
  • Tracial state: a state that treats everything symmetrically, with trivial modular flow.
  • Type I, II, III: classes of algebras by how they measure size.
  • Varentropy: the variance of surprisal, how unevenly the clock ticks.

References


  1. Zheng Wang and Jerome R. Busemeyer, A Quantum Question Order Model Supported by Empirical Tests of an A Priori and Precise Prediction, Topics in Cognitive Science 5, 2013. ↩︎ ↩︎

  2. Ehtibar N. Dzhafarov, Janne V. Kujala and Jan-Åke Larsson, Contextuality in Three Types of Quantum-Mechanical Systems, Foundations of Physics 45, 2015, pp. 762-782. ↩︎ ↩︎

  3. Matthias Kleinmann, Otfried Gühne, José R. Portillo, Jan-Åke Larsson and Adán Cabello, Memory Cost of Quantum Contextuality, New Journal of Physics 13, 2011, 113011. ↩︎

  4. Jerome R. Busemeyer and Peter D. Bruza, Quantum Models of Cognition and Decision, Cambridge University Press, 2012. ↩︎

  5. Dénes Petz, Sufficient Subalgebras and the Relative Entropy of States of a von Neumann Algebra, Communications in Mathematical Physics 105, 1986, pp. 123-131. ↩︎ ↩︎

  6. Rolf Landauer, Irreversibility and Heat Generation in the Computing Process, IBM Journal of Research and Development 5(3), 1961, pp. 183-191. ↩︎ ↩︎

  7. Huzihiro Araki and E. J. Woods, A Classification of Factors, Publications of the Research Institute for Mathematical Sciences 4, 1968, pp. 51-130. ↩︎ ↩︎

  8. Don N. Page and William K. Wootters, Evolution without evolution: Dynamics described by stationary observables, Physical Review D 27, 1983, pp. 2885-2892. ↩︎ ↩︎

  9. Venkatesa Chandrasekaran, Roberto Longo, Geoff Penington and Edward Witten, An Algebra of Observables for de Sitter Space, Journal of High Energy Physics 2023(2), 82, https://arxiv.org/abs/2206.10780. ↩︎ ↩︎

  10. Julian De Vuyst, Stefan Eccles, Philipp A. Höhn and Josh Kirklin, Crossed products and quantum reference frames: on the observer-dependence of gravitational entropy, Journal of High Energy Physics 2025(7), 063, https://arxiv.org/abs/2412.15502. ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎

  11. Philipp A. Höhn, Alexander R. H. Smith and Maximilian P. E. Lock, The Trinity of Relational Quantum Dynamics, Physical Review D 104, 2021, 066001, https://arxiv.org/abs/1912.00033. See also, by the same authors, Equivalence of Approaches to Relational Quantum Dynamics in Relativistic Settings, Frontiers in Physics 9, 2021, https://arxiv.org/abs/2007.00580. ↩︎

  12. Geoff Penington and Edward Witten, Algebras and States in JT Gravity, 2023, https://arxiv.org/abs/2301.07257. ↩︎

  13. Chang-Han Chen and Geoff Penington, A clock is just a way to tell the time: gravitational algebras in cosmological spacetimes, 2024, https://arxiv.org/abs/2406.02116. ↩︎

  14. Christopher W. Lynn, Eli J. Cornblath, Lia Papadopoulos, Maxwell A. Bertolero and Danielle S. Bassett, Broken detailed balance and entropy production in the human brain, Proceedings of the National Academy of Sciences 118(47), 2021. ↩︎

  15. Flaminia Giacomini, Esteban Castro-Ruiz and Časlav Brukner, Quantum mechanics and the covariance of physical laws in quantum reference frames, Nature Communications 10, 2019, 494. ↩︎

  16. Leslie G. Valiant, Evolvability, Journal of the ACM 56(1), 2009, article 3. ↩︎ ↩︎ ↩︎

  17. Vitaly Feldman, Evolvability from Learning Algorithms, Proceedings of the 40th ACM Symposium on Theory of Computing (STOC), 2008. ↩︎ ↩︎ ↩︎

  18. Michael Kearns, Efficient Noise-Tolerant Learning from Statistical Queries, Journal of the ACM 45(6), 1998, pp. 983-1006. ↩︎ ↩︎

  19. Boaz Barak, Benjamin L. Edelman, Surbhi Goel, Sham Kakade, Eran Malach and Cyril Zhang, Hidden Progress in Deep Learning: SGD Learns Parities Near the Computational Limit, NeurIPS 2022; and Emmanuel Abbe and Colin Sandon, Provable Limitations of Deep Learning, 2018, https://arxiv.org/abs/1812.06369. ↩︎

  20. Roger N. Shepard, Carl I. Hovland and Herbert M. Jenkins, Learning and Memorization of Classifications, Psychological Monographs 75(13), 1961. ↩︎

  21. Juan Maldacena, The Large N Limit of Superconformal Field Theories and Supergravity, Advances in Theoretical and Mathematical Physics 2, 1998, pp. 231-252. ↩︎ ↩︎

  22. Tom Banks, Willy Fischler, Stephen H. Shenker and Leonard Susskind, M Theory as a Matrix Model: A Conjecture, Physical Review D 55, 1997, pp. 5112-5128. ↩︎

  23. Ahmed Almheiri, Xi Dong and Daniel Harlow, Bulk Locality and Quantum Error Correction in AdS/CFT, Journal of High Energy Physics 2015(4), 163. ↩︎ ↩︎

  24. Steven Carlip, Dimension and Dimensional Reduction in Quantum Gravity, Classical and Quantum Gravity 34, 2017, 193001. ↩︎ ↩︎

  25. Sam Leutheusser and Hong Liu, Emergent times in holographic duality, 2021, https://arxiv.org/abs/2112.12156. ↩︎

  26. Chris Fields, Karl Friston, James F. Glazebrook and Michael Levin, A free energy principle for generic quantum systems, 2022, https://arxiv.org/abs/2112.15242; and Chris Fields, James F. Glazebrook and Antonino Marcianò, The Physical Meaning of the Holographic Principle, Quanta 11, 2022, pp. 72-96. ↩︎ ↩︎

  27. Pankaj Mehta and David J. Schwab, An exact mapping between the Variational Renormalization Group and Deep Learning, 2014, https://arxiv.org/abs/1410.3831. ↩︎

  28. Peter Godfrey-Smith, Other Minds: The Octopus, the Sea, and the Deep Origins of Consciousness, Farrar, Straus and Giroux, 2016. ↩︎

  29. John Maynard Smith and Eörs Szathmáry, The Major Transitions in Evolution, Oxford University Press, 1995. ↩︎

  30. Christophe Coupé, Yoon Mi Oh, Dan Dediu and François Pellegrino, Different languages, similar encoding efficiency: Comparable information rates across the human communicative niche, Science Advances 5(9), 2019, eaaw2594. ↩︎

  31. Kristin Koch, Judith McLean, Ronen Segev, Michael A. Freed, Michael J. Berry, Vijay Balasubramanian and Peter Sterling, How Much the Eye Tells the Brain, Current Biology 16(14), 2006, pp. 1428-1434. ↩︎

  32. Norman Margolus and Lev B. Levitin, The maximum speed of dynamical evolution, Physica D 120, 1998, pp. 188-195. ↩︎ ↩︎

  33. Jacob D. Bekenstein, Universal upper bound on the entropy-to-energy ratio for bounded systems, Physical Review D 23, 1981, pp. 287-298. ↩︎ ↩︎

  34. Seth Lloyd, Ultimate physical limits to computation, Nature 406, 2000, pp. 1047-1054. ↩︎ ↩︎

  35. Neil Ashby, Relativity in the Global Positioning System, Living Reviews in Relativity 6, 2003, 1. ↩︎

  36. Carlo Rovelli, Relational Quantum Mechanics, International Journal of Theoretical Physics 35, 1996, pp. 1637-1678. ↩︎

  37. Kok-Wei Bong, Aníbal Utreras-Alarcón, Farzad Ghafari, Yeong-Cherng Liang, Nora Tischler, Eric G. Cavalcanti, Geoff J. Pryde and Howard M. Wiseman, A strong no-go theorem on the Wigner’s friend paradox, Nature Physics 16, 2020, pp. 1199-1205. ↩︎

  38. Howard M. Wiseman, Eric G. Cavalcanti and Eleanor G. Rieffel, A “thoughtful” Local Friendliness no-go theorem: a prospective experiment with new assumptions to suit the abilities of human experimenters, 2022, https://arxiv.org/abs/2209.08491. ↩︎

  39. David Deutsch, Quantum theory as a universal physical theory, International Journal of Theoretical Physics 24, 1985, pp. 1-41. ↩︎

  40. Building a human-like observer using deep learning in an extended Wigner’s friend experiment, 2024, https://arxiv.org/abs/2409.04690. ↩︎

  41. Eugene P. Wigner, Remarks on the mind-body question, in I. J. Good (ed.), The Scientist Speculates, Heinemann, 1961. ↩︎

  42. John von Neumann, Mathematical Foundations of Quantum Mechanics, Princeton University Press, 1955 (German original 1932). ↩︎

  43. Wojciech H. Zurek, Decoherence, einselection, and the quantum origins of the classical, Reviews of Modern Physics 75, 2003, pp. 715-775. ↩︎

  44. Yaakov Y. Fein and collaborators, Quantum superposition of molecules beyond 25 kDa, Nature Physics 15, 2019, pp. 1242-1245. ↩︎

  45. Sandro Donadi and collaborators, Underground test of gravity-related wave function collapse, Nature Physics 17, 2021, pp. 74-78. ↩︎

  46. Massimiliano Proietti and collaborators, Experimental test of local observer independence, Science Advances 5(9), 2019, eaaw9832. ↩︎

  47. Jonathan N. Tinsley and collaborators, Direct detection of a single photon by humans, Nature Communications 7, 2016, 12172. ↩︎

  48. Howard Barnum, Carlton M. Caves, Christopher A. Fuchs, Richard Jozsa and Benjamin Schumacher, Noncommuting Mixed States Cannot Be Broadcast, Physical Review Letters 76, 1996, pp. 2818-2821; and Howard Barnum, Jonathan Barrett, Matthew Leifer and Alexander Wilce, A Generalized No-Broadcasting Theorem, Physical Review Letters 99, 2007, 240501. ↩︎

  49. Warrick Roseboom, Zafeirios Fountas, Kyriacos Nikiforou, David Bhowmik, Murray Shanahan and Anil K. Seth, Activity in perceptual classification networks as a basis for human subjective time perception, Nature Communications 10, 2019, 267. ↩︎

  50. Fabio Cipriani and Jean-Luc Sauvageot, Derivations as Square Roots of Dirichlet Forms, Journal of Functional Analysis 201, 2003, pp. 78-120; and Eric A. Carlen and Jan Maas, Gradient flow and entropy inequalities for quantum Markov semigroups with detailed balance, Journal of Functional Analysis 273, 2017, pp. 1810-1869. ↩︎

  51. Benjamin F. Toner and Dave Bacon, Communication Cost of Simulating Bell Correlations, Physical Review Letters 91, 2003, 187904. ↩︎

  52. Masamichi Takesaki, Theory of Operator Algebras II, Springer, 2003. ↩︎

  53. Alain Connes and Carlo Rovelli, Von Neumann algebra automorphisms and time-thermodynamics relation in generally covariant quantum theories, Classical and Quantum Gravity 11, 1994, pp. 2899-2917. ↩︎

  54. F. J. Murray and J. von Neumann, On Rings of Operators, Annals of Mathematics 37(1), 1936, pp. 116-229; Alain Connes, Une classification des facteurs de type III, Annales scientifiques de l’École Normale Supérieure 6, 1973, pp. 133-252; and Alain Connes and Masamichi Takesaki, The flow of weights on factors of type III, Tohoku Mathematical Journal 29, 1977, pp. 473-575. ↩︎ ↩︎

  55. Raphael Bousso, The holographic principle, Reviews of Modern Physics 74, 2002, pp. 825-874; and Jacob D. Bekenstein, Black Holes and Entropy, Physical Review D 7, 1973, pp. 2333-2346. ↩︎

  56. Tony A. Plate, Holographic reduced representations, IEEE Transactions on Neural Networks 6(3), 1995, pp. 623-641. ↩︎