IPS and related theories — a map of convergences and divergences

Neuroscience, thermodynamics, topology, and cybernetics each offer a partial answer to the same question: why do some patterns keep existing while others dissolve? The information-persisting system (IPS) framework does not replace those answers. It assembles them into one accounting identity — the Fractal Persistence Equation (FPE) — and states a hard survival condition: persistence ratio \(\mathcal{R} \ge 1\).

This essay is a comparative map. For each adjacent theory we ask three questions:

1. What does it measure?
2. Where does it appear inside IPS?
3. What does IPS add that the theory alone does not supply?

Canonical symbols: glossary.md. Formal derivation: information_persisting_systems.md.

The IPS claim in one paragraph

An IPS is a bounded, far-from-equilibrium subsystem with (i) a Markov blanket \(\partial\Sigma\), (ii) non-zero drive, (iii) an internal model \(q_\mu\), and (iv) identity preserved over \(\tau \gg \tau_{\mathrm{relax}}\). Identity is not a microstate; it is a macrostate trajectory \(m(t) = \mathcal{G}(\sigma(t))\) in a finite alphabet \(\mathcal{M}\).

The FPE compares usable power income \(P_{\mathrm{in}}\) (after coupling efficiency \(\eta\)) to the debit of maintaining that trajectory:

• \(\mathcal{D}_{\mathrm{KL}}\) — delusion divergence: gap between internal model and resolved reality.
• \(\Gamma\) — structural fatigue: backlog of unresolved damage.
• \(\Phi\) — substrate integrity: graph-indexed health of level-\((L{-}1)\) components (min on critical paths, softer pool when redundant). Not a mean. substrate_shelter_and_redundancy.md
• \(\Psi\) — shelter coefficient: product over independent shelter channels from level-\((L{+}1)\); \(\le \min\) when buffers overlap.

\(\mathcal{R} \ge 1\) is necessary for long-horizon persistence (Theorem 5.1). Violation implies dissolution in bounded time. The theory is substrate-neutral: any system satisfying the four clauses qualifies — cell, eddy, firm, or polity.

Three kinds of relation

Most confusion comes from treating limits as competitors or from symbol collisions (Tononi’s \(\Phi\) vs IPS substrate integrity \(\Phi\)).

Thermodynamics and non-equilibrium physics

Second law and open-system balances

What it measures: Entropy production in a subsystem; the impossibility of spontaneous order in a closed box.

IPS reading: Equations (3.1)–(3.3) of the foundation paper. An IPS in a non-equilibrium steady state (NESS) must export at least as much entropy as it produces internally. Power income \(P_{\mathrm{in}}\) is the free-energy flux that pays for that export.

What IPS adds: A dimensionless balance sheet (\(\mathcal{R}\)) and an explicit debit ledger (noise floor, delusion tax, fatigue) rather than a qualitative “negentropy import.”

Prigogine — dissipative structures

What it measures: Ordered patterns maintained by continuous dissipation far from equilibrium (Bénard cells, reaction–diffusion waves).

IPS reading: The limit \(\mathcal{D}_{\mathrm{KL}} = 0\), \(\Gamma = 0\). Pure dissipative structure: income must exceed the Landauer–Bennett floor \(\omega\mathcal{E}_\Sigma\).

What IPS adds: Internal models and delusion as first-class debits. A Bénard cell has no \(q_\mu\); a cell, firm, or agent does — and wrong models pay exponentially (Theorem 5.2).

Schrödinger — What Is Life?

What it measures: Life as order maintained by feeding on negative entropy through a boundary.

IPS reading: Definition 2.1 is Schrödinger’s criterion formalised: drive + blanket + memory (internal model) + identity over time. The “aperiodic crystal” is one substrate realisation of \(\Sigma_{\mathrm{int}}\).

What IPS adds: Quantitative lifetime bounds and fractal composition (substrate \(\Phi\), shelter \(\Psi\)) — not only “negative entropy” but how much, at what efficiency, and with what model error.

Landauer, Bennett, and stochastic thermodynamics

What it measures: Minimum heat dissipated per logically irreversible bit erasure; fluctuation theorems linking dissipation to information.

IPS reading: Equations (3.4)–(3.9). Internal noise floor \(\mathcal{E}_\Sigma\) sets the per-bit maintenance cost; delusion divergence \(\mathcal{D}_{\mathrm{KL}}\) incurs excess dissipation per model-update cycle (Crooks/Jarzynski structure).

What IPS adds: Landauer is not an isolated bound — it is one term in \(P_{\mathrm{out}}^{\mathrm{req}}\), combined with surprise and senescence. Prediction (P4) in the paper: no IPS persists below \(\omega\mathcal{E}_\Sigma\) regardless of model quality.

Bylicka et al. (2015) — non-Markovian memory and extractable work

Source: arXiv:1504.06533; local copy references/1504.06533v1.pdf.

What it measures: In open quantum systems, non-Markovian environment memory can produce revivals of extractable work — temporary recovery of usable free energy via information back-flow, even when Markovian (divisible) dynamics would monotonically erode it.

IPS reading: The default FPE ledger treats environmental coupling in a Markovian / NESS idiom: debits accumulate unless paid for by \(P_{\mathrm{in}}\). Non-Markovian revivals are the quantum-thermodynamic case where history feeds back useful information, temporarily improving the income side or reducing effective erasure cost. This does not overturn Theorem 5.1 (\(\mathcal{R} < 1\) still dissolves in finite time) but explains transient recoveries after shock.

What IPS adds: A macroscopic dimensionless balance (\(\mathcal{R}\)) and explicit delusion/fatigue debits across scales; 1504 stays in quantum open-system thermodynamics. Open problem: non-Markovian analogue of the Delusion Tax lifetime bound (to_publish/04_delusion_tax.md §6.4).

Bützbach — continuity as thermodynamic memory; Mamun (2026) review

What it measures: Bützbach (2026) treats persistence as active thermodynamic memory maintenance — not static endurance but recursive self-refreshment against entropy. The VP4/VP5 continuity functional combines metabolic power \(E_{\mathrm{met}}\), Landauer-scaled informational maintenance power \(P_{\mathrm{info}}\), and a recursive continuity coefficient \(C\) (re-entry probability after perturbation). Continuity power \(P_\Omega = (E_{\mathrm{met}} + P_{\mathrm{info}})\,C\) yields a normalised life-likeness index \(\Omega = P_\Omega / P_{\mathrm{ref}}\) (microbial baseline). Mamun’s critical review extends the framework to quantum decoherence and LLM entropic recursion budgets.

IPS reading: Parallel formalism, not a derived limit. Bützbach’s triad — recursion, constraint load \(K\), meaning-path integrity — maps onto IPS clauses (iii)–(iv): the internal model must refresh coherently or pathways fragment. Symbol correspondences:

Mamun’s LLM extension (\(C_{\mathrm{LLM}}\) from attention recursion, \(P_{\mathrm{info}}\) from gradient erasures) is the same architectural move as cognitive_processor.md: score a runtime by whether its books balance, not by task reward alone.

What IPS adds:

Independent convergence on Landauer-grounded persistence metrics (2026) supports falsifiability; IPS remains distinct in fractal composition, delusion/fatigue ledger, and operational scoring in the Aion stack.

Fagan (2026) — Conservation-Congruent Encoding (CCE)

Source: arXiv:2601.00021; local copy references/2601.00021v1.pdf; internal note papers/fagan.md.

What it measures: Intelligence as a substrate-neutral physical efficiency: goal-directed work per nat of irreversible information processed (\(\chi\)); an operational consciousness efficiency \(\kappa\) (work per nat preserved). Irreversible distinction collapse exports a generalised Landauer cost across conserved channels.

IPS reading: Parallel formalism, not a derived limit. Port factorisation (agent–environment conditional independence) is the same structural move as equation (2.1). Measurement as active coarse-graining aligns with the quantum commitment bridge. Lindblad dynamics as dissipative exhaust of recording a bit supports the commitment interpretation at \(L_0\).

What IPS adds: A single necessary survival inequality (\(\mathcal{R} \ge 1\)) with finite dissolution time, explicit delusion divergence and structural fatigue, fractal \(\Phi\)/\(\Psi\) composition (Theorem 5.3), and operational scoring in the Aion stack. Fagan’s cosmological extension is not part of our conservative physics claim.

Inference, surprise, and the Markov blanket

Pearl — graphical models and conditional independence

What it measures: When internal and external variables are independent given a separating set (the blanket).

IPS reading: Equation (2.1): \(p(\sigma, b, e) = p(\sigma \mid b)\, p(e \mid b)\, p(b)\). The blanket is not metaphorical — it is the factorisation condition for being a distinct system.

What IPS adds: Thermodynamic cost of maintaining that factorisation under noise and model error. A statistical separation that cannot be paid for dissolves.

Friston — variational free energy and active inference

What it measures: Variational free energy \(\mathcal{F}[\mu]\) as an upper bound on surprise; systems minimise \(\mathcal{F}\) by updating \(\mu\) or acting on the world.

IPS reading: \(\mathcal{D}_{\mathrm{KL}}^{(\Sigma)}\) in the FPE is the complexity–accuracy residue of FEP, but treated as a thermodynamic tax, not only an optimisation objective. Section 7.1 of the foundation paper: with \(\Gamma = 0\) and varying \(\mathcal{D}_{\mathrm{KL}}\), IPS reduces to FEP embedded in an explicit energy budget.

What IPS adds:

Active inference (action to change sensations) is the behavioural half of lowering \(\mathcal{F}\). IPS scores the same move in bits and trust in the Aion stack: forecasts that reduce \(\mathcal{D}_{\mathrm{KL}}\) earn power income for the next loop.

A generic persistence score \(S(\eta)\)

Writings outside IPS sometimes use a phenomenological form:

with reversibility efficiency \(\eta\), dissipation \(Q\), and environmental volatility \(T\).

IPS reading: Same shape as the income/debit ratio, but IPS derives terms rather than fitting \(\alpha\):

• \(\eta \to \eta_I(\mathcal{D}_{\mathrm{KL}})\) — efficiency depends on model quality (equation 4.2).
• \(Q/T \to \omega\mathcal{E}_\Sigma(1 + \mathcal{D}_{\mathrm{KL}} + \Gamma) / P_{\mathrm{in}}\) — structural complexity times noise floor, modulated by delusion and fatigue, relative to income.

What IPS adds: A threshold (\(\mathcal{R} \ge 1\)) instead of a smooth probability; composition across scales; falsifiable lifetime laws (P1–P5 in the paper).

Topology and structural stability

Persistent homology and topological entropy

What it measures: Which features of a point cloud or network persist across scales of mathematical noise; topological entropy \(H_T\) quantifies the complexity of persistence diagrams.

IPS reading: Complementary, not competing. IPS asks which macrostate identities \(m \in \mathcal{M}\) survive as trajectories under real thermodynamic noise, not only which homological features survive in a filtered complex.

What IPS adds: A mechanism (drive, blanket, model, debit) explaining why a topological signature stops persisting — not only that it did.

Critical transitions and tipping points (Scheffer et al.)

What it measures: Sudden regime shifts when a control parameter crosses a threshold; early-warning signals (critical slowing down).

IPS reading: The transition \(\mathcal{R} > 1 \to \mathcal{R} < 1\) (open problem 3 in Section 7.4). Ecosystem fold catastrophes (Section 6.6) are FPE phase transitions as \(\Phi\) degrades.

What IPS adds: Identifies which control parameters matter (\(\Phi\), \(\Psi\), \(\mathcal{D}_{\mathrm{KL}}\), \(\Gamma\)) and predicts percolation-like collapse (Theorem 5.3, prediction P2) rather than smooth degradation.

Biology, cognition, and enaction

Autopoiesis (Maturana & Varela)

What it measures: Living systems as self-producing networks that continuously regenerate their own components and boundary.

IPS reading: Autopoiesis is a biological specialisation of Definition 2.1: the blanket is the membrane; drive is metabolism; the internal model is implicit in regulatory architecture; identity is the reproduced organisation. Enactivism’s “structurally coupled” organism maps to blanket-mediated exchange.

What IPS adds: Substrate-neutrality — autopoiesis is sufficient for many biological IPS, not necessary for the class (nuclei, firms, eddies qualify without self-production in the chemical sense). Explicit delusion and fatigue terms absent from classical autopoiesis.

Homeostasis and allostasis

What they measure: Maintenance of regulated variables (homeostasis); predictive adjustment of setpoints under expected demand (allostasis).

IPS reading: Control around a macrostate class \([m_0]\). Internal model \(\mu\) encodes setpoints; delusion divergence is setpoint tracking error; structural fatigue \(\Gamma\) is unresolved allostatic load.

What IPS adds: The control-theoretic reading in Section 7.1 of the foundation paper and aion_core_control_theory.md: \(\mathcal{R} \ge 1\) is the persistence setpoint averaged over a regulation window; momentary \(\mathcal{R} < 1\) is out-of-band excursion, not instant death.

Eigen — error threshold and quasi-species

What it measures: Maximum per-copy mutation rate before a replicating information string delocalises in sequence space.

IPS reading: Direct instance of Theorem 5.2. Population distance from the master sequence is \(\mathcal{D}_{\mathrm{KL}}\); above threshold, \(\mathcal{R} < 1\) and the quasi-species dissolves into noise.

What IPS adds: Same mathematics applied uniformly to firms (bad market models), ideologies, and neural nets — not only RNA replication.

Information, consciousness, and what IPS is not

Integrated Information Theory (Tononi)

What it measures: \(\Phi_{\mathrm{IIT}}\) — integrated information, proposed as correlate of consciousness.

IPS reading: Orthogonal on phenomenology. IPS substrate integrity \(\Phi\) is a compositional factor (health of sub-nodes), not Tononi’s consciousness scalar. Symbol collision is deliberate disambiguation in the glossary.

What IPS adds: Nothing to IIT’s consciousness claims — and claims nothing about qualia. Phenomenology is developed separately in books/book1/. IPS supplies the thermodynamic exterior any experiencing system must satisfy.

Computational functionalism and the abstraction fallacy

What it claims: Symbol manipulation or functional isomorphism suffices for mind or persistence-like dynamics.

IPS reading: Theorem 5.3 and Proposition 5.4: symbol manipulation without a substrate satisfying (i)–(iv) is not an IPS. The mapmaker must be a thermodynamically bounded pattern, not an abstract decoder — consistent with abstraction fallacy arguments.

What IPS adds: A necessary-condition theorem, not only a philosophical objection.

Anthropic reasoning

What it addresses: Why observers find themselves in universes compatible with complexity.

IPS reading: Orthogonal. IPS states what a persisting pattern must do, not which universes contain such patterns.

Organisation, markets, and engineered systems

Cybernetics (Wiener, Ashby)

What it measures: Regulation, feedback, requisite variety, homeostatic control.

IPS reading: Feedback is how \(\mu\) is updated to lower \(\mathcal{D}_{\mathrm{KL}}\). Ashby’s “requisite variety” maps to complexity \(\omega\) — the constraint tax on the blanket.

What IPS adds: Energy and bit costs of variety matching; trustworthiness scoring when models are disputed (see trustworthiness_token.md, ips_economics.md).

Prediction markets and KL scoring

What they measure: Collective forecasts; proper scoring rules (log score → KL divergence).

IPS reading: Implementation layer, not separate physics. Markets score delusion divergence after outcomes resolve; trust paid in bits is power income \(P_{\mathrm{in}}\) for the next cognitive loop. See cognitive_processor.md and society_of_aion_nodes.md.

What IPS adds: Explains why KL scoring is the correct currency: each nat of sustained delusion divides lifetime by \(e\) (Theorem 5.2).

Control theory (aion-core)

What it measures: Setpoints, observers, actuators, stability margins.

IPS reading: Full mapping in aion_core_control_theory.md — band telemetry, cascade promotion, anti-windup as engineering realisations of \(\mathcal{R}\) regulation.

What IPS adds: The persistence objective that classical control leaves implicit: survive while tracking.

Engram (2026) — conditional memory as a second sparsity axis

Source: arXiv:2601.07372; internal note papers/engram.md; experiments in aion-llm/dev/LOG.md.

What it measures: Conditional memory — O(1) hash lookup of static N-gram patterns — as a complement to MoE conditional computation. A U-shaped law trades expert capacity against memory table capacity.

IPS reading: Section 6.8 of the foundation paper: sparse routing is the minimal architecture respecting blanket factorisation (2.1) at every compositional level. Engram adds a second sparse coupling:

Nodes off the MoE path receive zero gradient (substrate isolation); Engram tables similarly isolate stereotyped local patterns from full-backbone recomputation. Together they instantiate the fractal MoE programme and prediction P5 (forgetting \(\propto\) router overlap, not total steps). Local prototype: BigramEmbed engram-lite in nanochat / aion-llm.

What IPS adds: The why — persistence ratio, Theorem 5.3 critical subgraphs, and falsifiable continual-learning predictions — not only an architecture benchmark.

Quantum measurement (downward extension)

The quantum IPS bridge extends IPS to measurement without replacing Hilbert-space dynamics.

What it reframes: Wavefunction “collapse” as commitment — forced tokenization at the blanket: the observer must select \(m \in \mathcal{M}\), update \(q_\mu\), and pay Landauer erasure before entering the \(\mathcal{R}\) loop.

Relation to other theories:

Summary table

What IPS uniquely contributes

Collectively, the adjacent theories supply partial derivatives — surprise, reversibility, shape, feedback, integration. IPS supplies the integrated identity:

1. One inequality — \(\mathcal{R} \ge 1\), not three incompatible scores.
2. Thermodynamic price tags — surprise, erasure, and wrong models cost watts.
3. Fractal composition — substrate \(\Phi\) and shelter \(\Psi\) make persistence a graph property (Theorem 5.3).
4. Necessary condition with time bound — not metaphorical decay (Theorem 5.1).
5. Substrate neutrality — same law for nuclei, cells, firms, polities, and designed MoE networks (Section 6.8).
6. Operational scoring — Aion implements the same variables at runtime.

IPS is conservative physics (no new fundamental constants) with a consequential rearrangement: persistence is a quantity in the currency of bits and joules, and the patterns that stay are exactly those whose books balance.

Further reading

Companion essay. For vocabulary rules see CONTRIBUTING.md.