Derivation of the Free Energy Principle from UHM

Retraction 2026-09-25: Theorems 3.1, 4.2 (iii)–(iv), 4.3 and Corollaries 7.1–7.2 [✗]

The functional this page minimises is a cross-entropy, and it does not do what the page claimed. For states $\rho$, $\Gamma$ with $\mathrm{supp}\,\rho \subseteq \mathrm{supp}\,\Gamma$,

$$S_{vN}(\rho) + D_{KL}(\rho \| \Gamma) = -\mathrm{Tr}\,\rho\log\rho + \mathrm{Tr}\,\rho\log\rho - \mathrm{Tr}\,\rho\log\Gamma = -\mathrm{Tr}(\rho\log\Gamma),$$

which is linear in $\rho = \psi(\Gamma)$. Its minimum over all CPTP channels is $-\log\lambda_{\max}(\Gamma)$, reached by the channel that sends $\Gamma$ onto an eigenvector of its largest eigenvalue — not by the self-model $\varphi$. The classical case is the same: $H(q) + D_{KL}(q\|p) = -\sum_i q_i \ln p_i$ is minimised by the point mass on $\arg\max_i p_i$, not by $q = p$. Example: for $p = (0.7, 0.2, 0.1)$ the value is $0.802$ at $q = p$ and $0.357$ at $q = (1, 0, 0)$.

Retracted for that reason: Theorem 3.1 (registry row 39e) — $\varphi$ is not the minimiser; Theorem 4.2 (iii) — the minima of the UHM functional and of Friston's free energy differ; Theorem 4.2 (iv) — $\mathcal{F}[\psi;\Gamma] \geq S_{vN}(\Gamma)$ fails on the same example ($0.357 < 0.802$); Theorem 4.3 — the optimum is not $q^* = p$; Corollaries 7.1–7.2, which rest on Theorem 3.1. Four steps of the derivation were wrong, and each is marked where it occurs: §4.1, "$F = H(q) + D_{KL}(q\|p_{\text{prior}}) + \text{const}$" (the difference from $F$ is $\langle \ln q - \ln p(o|s) \rangle_q$, which depends on $q$); §4.4, Step 2, which lost $H(q)$; §4.5, Step 3, "$H_\times(q,p) \geq H(p)$ with equality at $q = p$" (Gibbs' inequality bounds $H_\times(p,q)$, with the arguments the other way round); §4.5, Step 4, "$H(q)$ does not affect the optimal $q$". The proof of Theorem 3.1 also used "$dS_{vN}/dt \geq 0$ for Lindblad evolution", which holds only for unital generators; for every Lindblad semigroup with a stationary state what holds is that the relative entropy to that state does not increase (Spohn 1978), and UHM's own dynamics lowers the entropy on its way from $I/7$ to the attractor (T-96).

What stands: the categorical definition of $\varphi$ (§2); Theorem 4.1 as an identity — on diagonal states the functional equals $H(q) + D_{KL}(q\|p)$, the cross-entropy, which is not Friston's free energy; Theorem 5.1. Open — research programme [Pr]: whether $\varphi$ minimises any natural functional, and whether Friston's free energy arises as a limit of UHM.

1. Problem Statement

1.1 Two definitions of φ

In UHM, the self-modeling operator φ has two representations:

Canonical definition (categorical):

$$\varphi \dashv i: \mathrm{Sub}(\Gamma) \hookrightarrow \mathbf{Sh}_\infty(\mathcal{C})$$

φ is defined as the left adjoint to the canonical inclusion of subobjects.

Variational characterization (retracted 2026-09-25 — the right-hand side is minimised by a projection onto the top eigenvector of $\Gamma$, not by $\varphi$; box above):

$$\varphi = \arg\min_{\psi \in \mathcal{CPTP}} \mathbb{E}_{\Gamma \sim \mu}\left[S_{spec}(\psi(\Gamma)) + D_{KL}(\psi(\Gamma) \| \Gamma)\right]$$

1.2 Questions this document answers

Question	Status
Proof of equivalence of the two definitions	Theorem 3.1 — retracted [✗]
Classical limit of the variational principle	Theorem 4.1 [T] — an identity: the limit is the cross-entropy, not Friston's free energy
Connection to Friston's FEP	Theorem 4.2 — (iii)–(iv) retracted [✗]; (i)–(ii) are identifications, not theorems
Complete reduction of $S_{spec} + D_{KL}$ to $F_{FEP}$	Theorem 4.3 — retracted [✗]
Justification of $S_{spec}$ vs $S_{vN}$	Theorem 5.1 [T]

---

2. Categorical Foundations

2.1 ∞-topos structure

Let $\mathcal{E} = \mathbf{Sh}_\infty(\mathcal{C})$ be the ∞-topos of sheaves over the category of density matrices $\mathcal{C} = \mathcal{D}(\mathbb{C}^7)$ with the Grothendieck topology $J_{Bures}$.

Key elements:

• $\Omega$ — subobject classifier
• $\mathrm{Sub}(\Gamma) := \{S \hookrightarrow \Gamma : S \text{ is a subobject}\}$
• Characteristic morphism: $\chi_S: \Gamma \to \Omega$ for $S \in \mathrm{Sub}(\Gamma)$

2.2 Subobject category Sub(Γ)

Definition 2.1. $\mathrm{Sub}(\Gamma)$ is a category whose objects are monomorphisms $S \hookrightarrow \Gamma$, and morphisms are commutative triangles.

In the context of UHM, $\mathrm{Sub}(\Gamma)$ is interpreted as the category of logically consistent states — those satisfying the internal logic $\Omega$.

Key property: In the ∞-topos, $\mathrm{Sub}(\Gamma)$ is a lattice with greatest element $\Gamma$ and least element $\emptyset$.

2.3 Operator φ as left adjoint

Definition 2.2 (Canonical definition of φ).

$\varphi: \mathcal{E} \to \mathrm{Sub}(\Gamma)$ is defined as the left adjoint to the canonical inclusion $i: \mathrm{Sub}(\Gamma) \hookrightarrow \mathcal{E}$:

$$\varphi \dashv i$$

Universal property: For all $X \in \mathcal{E}$ and $S \in \mathrm{Sub}(\Gamma)$:

$$\mathrm{Hom}_{\mathrm{Sub}(\Gamma)}(\varphi(X), S) \cong \mathrm{Hom}_{\mathcal{E}}(X, i(S))$$

Interpretation: $\varphi(\Gamma)$ is the best (minimal) logically consistent approximation of $\Gamma$.

2.4 φ as co-reflector

From the theory of adjoint functors it follows that:

Lemma 2.1. $\varphi$ is a co-reflector:

$$\varphi(\Gamma) = \mathrm{colim}_{S \in \mathrm{Sub}(\Gamma)} S$$

where the colimit is taken over the diagram of all subobjects $S \leq \Gamma$.

---

3. Theorem on Variational Characterization

3.1 Preliminary definitions

Definition 3.1 (Spectral entropy).

For a density matrix $\rho$ with eigenvalues $\{\lambda_i\}$:

$$S_{spec}(\rho) := -\sum_i \lambda_i \log \lambda_i = S_{vN}(\rho)$$

Remark: In this context $S_{spec} = S_{vN}$ for density matrices. The distinction arises only for non-Hermitian operators (see §5).

Definition 3.2 (Quantum KL-divergence).

For density matrices $\rho, \sigma$ with $\mathrm{supp}(\rho) \subseteq \mathrm{supp}(\sigma)$:

$$D_{KL}(\rho \| \sigma) := \mathrm{Tr}(\rho(\log \rho - \log \sigma))$$

Definition 3.3 (Variational functional).

$$\mathcal{F}[\psi; \Gamma] := S_{spec}(\psi(\Gamma)) + D_{KL}(\psi(\Gamma) \| \Gamma)$$

where $\psi \in \mathcal{CPTP}$ is a CPTP channel (completely positive trace-preserving).

3.2 Central theorem

Theorem 3.1 (Variational characterization of φ) — retracted 2026-09-25 [✗]

Let $\varphi$ be defined categorically as the left adjoint to the inclusion $i: \mathrm{Sub}(\Gamma) \hookrightarrow \mathcal{E}$.

Then:

$$\varphi = \arg\min_{\psi \in \mathcal{CPTP}} \mathbb{E}_{\Gamma \sim \mu}\left[\mathcal{F}[\psi; \Gamma]\right]$$

where $\mu$ is the invariant measure on the state space (uniqueness of $\mu$ is guaranteed by primitivity of the linear part $\mathcal{L}_0$ [T]).

Retracted [✗]: the functional is the cross-entropy $-\mathrm{Tr}(\psi(\Gamma)\log\Gamma)$, minimised by a projection onto the top eigenvector of $\Gamma$, not by $\varphi$ (box at the top).

3.3 Proof

Retracted [✗] — kept for the record. Step 3 holds only for unital generators; Step 4 (stationarity equals minimal entropy production) is Prigogine's near-equilibrium principle, not a theorem about $\mathcal{L}_\Omega$; Step 5 introduces an "entropy production" of a channel without derivation; and the conclusion is false (box at the top).

Step 1: Connection of φ with the logical Liouvillian.

From the theorem on stationary distribution:

$$\varphi(\Gamma) = \lim_{\tau \to \infty} e^{\tau \mathcal{L}_\Omega}[\Gamma]$$

where $\mathcal{L}_\Omega$ is the logical Liouvillian, and the limit is read as the projector onto the zero mode of the linearised generator (only the linear part $\mathcal{L}_0$ is primitive; the full $\mathcal{L}_\Omega=\mathcal{L}_0+\mathcal{R}$ has the nontrivial fixed point $\rho^*_\Omega\neq I/7$ of T-96 — see the convention note).

Step 2: Dissipative structure of $\mathcal{L}_\Omega$.

$\mathcal{L}_\Omega$ has Lindblad form:

$$\mathcal{L}_\Omega[\Gamma] = -i[H_{eff}, \Gamma] + \sum_k \left(L_k \Gamma L_k^\dagger - \frac{1}{2}\{L_k^\dagger L_k, \Gamma\}\right)$$

where $L_k = \sqrt{\chi_{S_k}}$ are operators derived from the classifier atoms (see Axiom Ω⁷ §classifier-atoms).

Step 3: Connection of dissipation with entropy.

For Lindblad evolution the following holds (Spohn, 1978):

$$\frac{dS_{vN}(\Gamma(t))}{dt} \geq 0$$

with equality at the stationary state.

Step 4: Variational formulation of stationarity.

The stationary state $\Gamma_* = \varphi(\Gamma)$ is characterized by the condition:

$$\mathcal{L}_\Omega[\Gamma_*] = 0$$

This is equivalent to the minimum of entropy production:

$$\Gamma_* = \arg\min_{\rho} \sigma[\rho; \Gamma]$$

where $\sigma$ is the entropy production function.

Step 5: Explicit form of the functional.

For a CPTP channel $\psi$, the entropy production function has the form (Lindblad, 1975):

$$\sigma[\psi; \Gamma] = S_{vN}(\psi(\Gamma)) - S_{vN}(\Gamma) + D_{KL}(\psi(\Gamma) \| \Gamma_{ref})$$

With the choice $\Gamma_{ref} = \Gamma$ (the initial state as reference):

Choice of Γ_ref = Γ

The identification Γ_ref = Γ is a motivated definition (self-referential minimization), not a derivation from L_Ω. Motivation: autopoiesis (A1) requires that the system minimize the difference between itself and its model, which corresponds to Γ_ref = Γ. Alternative choices (Γ_ref = I/7 or Γ_ref = ρ*) give different functionals. The choice Γ_ref = Γ is the unique one for which the minimum of F coincides with the fixed point of φ (theorem).

[✗] With Γ_ref = Γ the minimum sits at a projection onto the top eigenvector of Γ, which is not the fixed point of φ; the "theorem" in the previous sentence is retracted.

$$\sigma[\psi; \Gamma] = S_{vN}(\psi(\Gamma)) + D_{KL}(\psi(\Gamma) \| \Gamma) - S_{vN}(\Gamma)$$

Since $S_{vN}(\Gamma)$ does not depend on $\psi$:

$$\arg\min_\psi \sigma[\psi; \Gamma] = \arg\min_\psi \left[S_{vN}(\psi(\Gamma)) + D_{KL}(\psi(\Gamma) \| \Gamma)\right]$$

Step 6: Identification.

Taking into account $S_{spec} = S_{vN}$ for density matrices (Theorem 5.1):

$$\varphi = \arg\min_{\psi \in \mathcal{CPTP}} \mathcal{F}[\psi; \Gamma] \quad \blacksquare$$

3.4 Remarks on the proof

Remarks:

1. Existence and uniqueness of the invariant measure $\mu$ are guaranteed by primitivity of the linear part $\mathcal{L}_0$ [T] (Evans 1977, Spohn 1976)
2. The equality $S_{spec} = S_{vN}$ holds only for normal operators (Theorem 5.1 [T])

Categorical correctness:

• Steps 1–2 follow from L-unification
• Steps 3–5 use standard open quantum systems theory
• The identification in Step 6 establishes the desired equivalence

[✗] Steps 3–6 do not establish it (box at the top).

---

4. Classical Limit: Complete Derivation of FEP

This section attempted to show that Friston's FEP is a special case of the UHM variational principle, arising in the transition to the classical limit; the attempt fails (box at the top), and each error is marked where it occurs. The derivation consists of three stages: (i) definition of the classical limit as $R \to 0$ (decoherence), (ii) reduction of the quantum functional to the classical one, (iii) identification of UHM elements with Friston's constructions. The section concludes with an analysis of what is lost in the classical limit.

4.1 Friston's FEP (original formulation)

According to Friston (2010), an agent interacting with the environment minimizes variational free energy:

$$F = \int ds \, q(s) \ln \frac{q(s)}{p(s,o)} = \underbrace{\langle E(s,o) \rangle_q}_{\text{internal energy}} - \underbrace{H(q)}_{\text{entropy}}$$

where:

• $s$ — hidden (latent) states of the world
• $o$ — observations (sensory data)
• $q(s)$ — recognition density — the agent's internal model
• $p(s,o)$ — generative density — joint model
• $E(s,o) = -\ln p(s,o)$ — energy of the generative model
• $H(q) = -\int q \ln q$ — Shannon entropy of the recognition density

Key inequality (evidence lower bound, ELBO):

$$F \geq -\ln p(o) \quad \text{(F bounds surprise from below)}$$

Proof: $F = D_{KL}(q(s) \| p(s|o)) - \ln p(o)$, and $D_{KL} \geq 0$.

Equivalent form via KL-divergence:

$$F = D_{KL}(q(s) \| p(s|o)) - \ln p(o) = D_{KL}(q(s) \| p(s)) - \langle \ln p(o|s) \rangle_q$$

[✗] Retracted form (2026-09-25): the page previously wrote the right-hand side as $H(q) + D_{KL}(q \| p_{\text{prior}}) + \text{const}$. That is the cross-entropy $-\langle \ln p(s) \rangle_q$; it differs from $F$ by $\langle \ln q(s) - \ln p(o|s) \rangle_q$, which depends on $q$.

4.2 Classical limit of UHM: formalization via $R \to 0$

Definition 4.1 (Classical limit of UHM).

The classical limit of UHM is defined by two equivalent conditions:

(a) Decoherence of off-diagonal elements. The density matrix loses coherences:

$$\Gamma_{ij} \to 0 \quad \text{for } i \neq j, \qquad \Gamma \to \mathrm{diag}(p_1, \ldots, p_N)$$

(b) Zero-reflection limit. The reflection measure $R = 1/(7P)$ tends to its minimum:

$$R \to R_{\min} = \frac{1}{7} \quad \text{(when } P \to 1, \text{ i.e. pure diagonal state)}$$

Lemma 4.1 (Classical limit).

(a) Decoherence (γ_{ij} → 0 for i ≠ j) implies R → 1/(7·1) = 1/7 (since P → max_i γ_{ii}² ≤ 1, and for equilibrium diagonal Γ: P ≈ 1/7, R ≈ 1).

(b) The converse is false: R = 1/7 ⟺ P = 1, which is achievable for a pure coherent state |ψ⟩⟨ψ| with maximal coherences. The classical limit is defined by condition (a) — decoherence, not through R.

Proof. (a) When $\Gamma_{ij} \to 0$ for $i \neq j$, the purity $P = \mathrm{Tr}(\Gamma^2) = \sum_i p_i^2 + 2\sum_{i < j} |\Gamma_{ij}|^2$ reduces to $P = \sum_i p_i^2$. Off-diagonal coherences enter directly into the Gap operator $\mathcal{G}_{ij}$ (see Gap dynamics); at $\Gamma_{ij} \to 0$ all Gap elements vanish: $\mathcal{G}_{ij} \to 0$. For an equilibrium diagonal matrix $p_i = 1/N$: $P = 1/N$, $R = 1/(NP) = 1$. For a single dominant $p_k \to 1$: $P \to 1$, $R = 1/(7P) \to 1/7$.

(b) Counterexample: pure state $|ψ⟩ = \frac{1}{\sqrt{7}}\sum_{k=0}^{6} |k⟩$ gives $P = 1$, $R = 1/7$, but has maximal coherences $|\Gamma_{ij}| = 1/7$ for all $i \neq j$. Therefore $R = 1/7$ is not equivalent to decoherence. $\blacksquare$

Physical meaning. The classical limit is complete decoherence: the system loses all quantum correlations between dimensions. In terms of consciousness: the system is not integrated ($\Phi \to 0$), not reflexive ($R < R_{th}$), has no Gap structure. This is the world of purely classical probabilities.

(c) Restriction of the CPTP channel. In the classical limit, the CPTP channel $\psi$ preserves diagonality:

$$\psi(\mathrm{diag}(p)) = \mathrm{diag}(q), \qquad q_i = \sum_j T_{ij} p_j, \quad T_{ij} \geq 0, \quad \sum_i T_{ij} = 1$$

The channel degenerates into a stochastic matrix $T$ — a classical Markov transition.

4.3 Reduction of the quantum functional

Theorem 4.1 (Classical limit of the variational principle) [T]

In the classical limit ($\Gamma_{ij} \to 0$ for $i \neq j$), the UHM variational functional (Definition 3.3) reduces to the classical functional below — the cross-entropy $-\sum_i q_i \ln p_i$, which is not Friston's variational free energy (§4.1):

$$\mathcal{F}[\psi; \Gamma] \xrightarrow[\Gamma_{ij}\to 0]{R \to R_{\min}} F_{cl}[q; p] = H(q) + D_{KL}(q \| p)$$

Proof.

Step 1 (Spectral entropy → Shannon entropy). For diagonal matrices, eigenvalues coincide with diagonal elements:

$$S_{spec}(\psi(\Gamma)) = S_{vN}(\mathrm{diag}(q)) = -\sum_i q_i \log q_i = H(q)$$

This is a direct consequence of Theorem 5.1 ($S_{spec} = S_{vN}$ for density matrices).

Step 2 (Quantum KL → classical KL). For diagonal matrices, the Umegaki quantum divergence reduces:

$$D_{KL}(\mathrm{diag}(q) \| \mathrm{diag}(p)) = \mathrm{Tr}\big(\mathrm{diag}(q)[\log \mathrm{diag}(q) - \log \mathrm{diag}(p)]\big) = \sum_i q_i (\log q_i - \log p_i) = D_{KL}(q \| p)$$

Step 3 (Substitution). Combining steps 1 and 2:

$$\mathcal{F}[\psi; \Gamma] = H(q) + D_{KL}(q \| p) = F_{cl}[q; p] \quad \blacksquare$$

Remark. The averaging $\mathbb{E}_{\Gamma \sim \mu}$ in Theorem 3.1 reduces in the classical limit to the ordinary mathematical expectation over the stationary distribution of a Markov chain (uniqueness guaranteed by primitivity of $\mathcal{L}_0$).

4.4 Derivation of Friston's classical variational free energy

This subsection attempted the complete derivation of $F = \mathbb{E}_q[\ln q(s) - \ln p(s,o)]$ from the variational characterization of $\varphi$; it fails at Step 2 (marked).

Step 1 (Identification of variables). Within UHM, introduce the identification:

UHM	FEP (Friston)	Meaning
$\Gamma = \mathrm{diag}(p_1, \ldots, p_N)$	$p(s,o)$ (generative density)	Full system state = generative model of the world
$\psi(\Gamma) = \mathrm{diag}(q_1, \ldots, q_N)$	$q(s)$ (recognition density)	Self-model = approximate inference
$\varphi(\Gamma)$	$q^*(s\|o) = p(s\|o)$	Optimal self-model = true posterior
$\mathcal{F}[\psi; \Gamma]$	$F[q; p]$	Variational free energy
$\min_\psi \mathcal{F}$	$\min_q F$	Variational inference

Step 2 (Expanding the functional). Write $\mathcal{F}$ in the classical limit:

$$\mathcal{F}[q; p] = H(q) + D_{KL}(q \| p) = -\sum_i q_i \log q_i + \sum_i q_i \log \frac{q_i}{p_i}$$

Simplifying:

$$\mathcal{F}[q; p] = \sum_i q_i (\log q_i - \log p_i) = \mathbb{E}_q[\log q - \log p]$$

[✗] Error (retracted 2026-09-25): the $q_i \log q_i$ terms cancel, so the sum is $-\sum_i q_i \log p_i$, the cross-entropy — not $\sum_i q_i(\log q_i - \log p_i)$. Everything below that rests on this line falls with it.

In the continuous limit ($N \to \infty$, sums → integrals):

$$\mathcal{F}[q; p] = \int ds \, q(s) \ln \frac{q(s)}{p(s,o)} = F_{FEP}$$

This was claimed to be exactly Friston's variational free energy; it is not (see the error above).

Step 3 (Equivalent forms). Expanding $p(s,o) = p(o|s) p(s)$:

$$F_{FEP} = \int ds \, q(s) \ln \frac{q(s)}{p(s)} - \int ds \, q(s) \ln p(o|s) = D_{KL}(q(s) \| p(s)) - \langle \ln p(o|s) \rangle_q$$

The first term is complexity (deviation from prior), the second is accuracy (expected likelihood). Minimization of $F$ = balance of accuracy and complexity — this is the classical analog of balancing spectral entropy and KL-divergence in Theorem 3.1.

Theorem 4.2 (UHM → Friston's FEP) — (iii)–(iv) retracted 2026-09-25 [✗]

Let $\Gamma$ be the state of a holon in the classical limit ($\Gamma_{ij} = 0$ for $i \neq j$). Then:

(i) The self-modeling operator $\varphi$ in the classical limit is identified with the recognition density: $\varphi(\Gamma) \leftrightarrow q^*(s|o)$.

(ii) The density matrix $\Gamma$ is identified with the generative model: $\Gamma \leftrightarrow p(s,o)$.

(iii) Minimization of the UHM functional coincides with minimization of Friston's free energy:

$$\min_{\psi \in \mathcal{CPTP}} \mathcal{F}[\psi; \Gamma] = \min_{q} F_{FEP}[q; p]$$

(iv) The key inequality (ELBO) is derived automatically:

$$\mathcal{F}[\psi; \Gamma] \geq S_{vN}(\Gamma) \quad \Longleftrightarrow \quad F \geq -\ln p(o)$$

Closedness of identification [D]. The identification $\Gamma \leftrightarrow p(s,o)$ is not an external assumption but the definition of self-reference. In the variational formulation $\varphi = \arg\min_q [\mathbb{E}_q[S_{spec}] + D_{KL}(q \| \Gamma)]$ the divergence $D_{KL}(q \| \Gamma)$ measures deviation from the system's own state. A self-referential system by definition uses itself as a generative model — this is not an assumption but a tautology of self-modeling.

Retracted [✗]: (iii) the left side equals $-\ln\max_i p_i$, the right side is $0$ at $q = p$; (iv) fails for $p = (0.7, 0.2, 0.1)$, where $\mathcal{F} = 0.357 < S_{vN}(\Gamma) = 0.802$.

Proof of (iv). From the definition of KL-divergence:

$$D_{KL}(\psi(\Gamma) \| \Gamma) \geq 0$$

Therefore:

$$\mathcal{F}[\psi; \Gamma] = S_{vN}(\psi(\Gamma)) + D_{KL}(\psi(\Gamma) \| \Gamma) \geq S_{vN}(\psi(\Gamma))$$

Since the stationary state of a primitive Lindbladian maximizes entropy among reachable states (Frigerio, 1978), we have $S_{vN}(\Gamma_*) \geq S_{vN}(\psi(\Gamma))$. Therefore, the ELBO lower bound takes the form:

$$\mathcal{F}[\psi; \Gamma] \geq D_{KL}(\psi(\Gamma) \| \Gamma) \geq 0$$

In the classical limit $S_{vN}(\Gamma) = H(p) = -\sum_i p_i \log p_i$, and the inequality takes the form $F \geq -\ln p(o)$, if $-\ln p(o)$ is identified with the entropy of the marginal probability of observations. $\blacksquare$

[✗] The chain above proves only $\mathcal{F} \geq 0$; the claimed $\mathcal{F} \geq S_{vN}(\Gamma)$ fails on the example in the box at the top.

4.5 Spectral entropy + KL → variational free energy

This subsection claimed to show explicitly how minimization of the quantum functional $S_{spec} + D_{KL}$ in the classical limit becomes minimization of Friston's variational free energy; Steps 3 and 4 are wrong (marked).

Theorem 4.3 (Complete reduction) — retracted 2026-09-25 [✗]

Let $\Gamma \in \mathcal{D}(\mathbb{C}^N)$ be a diagonal density matrix, $\psi$ a CPTP channel preserving diagonality. Then the problem

$$\min_{\psi \in \mathcal{CPTP}} \left[S_{spec}(\psi(\Gamma)) + D_{KL}(\psi(\Gamma) \| \Gamma)\right]$$

is identical to the problem

$$\min_{q \in \Delta^{N-1}} F_{FEP}[q; p] = \min_{q \in \Delta^{N-1}} \sum_i q_i \ln \frac{q_i}{p_i}$$

where $\Delta^{N-1}$ is the $(N-1)$-simplex of probability distributions.

Moreover, the minimum is achieved at $q_i^* = p_i$ (recognition density coincides with generative), which corresponds to $\psi^* = \mathrm{id}$ (identity channel).

Retracted [✗]: the first problem is solved by the point mass on $\arg\max_i p_i$ (value $-\ln\max_i p_i$), the second by $q = p$ (value $0$); they are not the same problem.

Proof.

Step 1 (Parameterization). In the classical limit, optimization over CPTP channels preserving diagonality is equivalent to optimization over stochastic matrices $T \in \mathbb{R}^{N \times N}_+$ with $\sum_i T_{ij} = 1$. The result $\psi(\Gamma) = \mathrm{diag}(q)$, where $q_i = \sum_j T_{ij} p_j$. The set of reachable $q$ for fixed $p$ is a convex subset of $\Delta^{N-1}$ containing $p$ (at $T = I$).

Step 2 (Explicit functional). By Theorem 4.1:

$$\mathcal{F}[q; p] = H(q) + D_{KL}(q \| p) = -\sum_i q_i \ln q_i + \sum_i q_i \ln \frac{q_i}{p_i} = -\sum_i q_i \ln p_i$$

This is the cross-entropy $H_\times(q, p) = -\sum_i q_i \ln p_i$.

Step 3 (Minimization). The cross-entropy $H_\times(q, p) = -\sum_i q_i \ln p_i$ is minimized at $q = p$ (by the Lagrange multiplier method with the constraint $\sum_i q_i = 1$, or from the property $H_\times(q, p) = H(q) + D_{KL}(q \| p) \geq H(p)$, with equality at $q = p$).

[✗] Error (retracted 2026-09-25): for fixed $p$, $H_\times(q,p) = -\sum_i q_i \ln p_i$ is linear in $q$ and is minimised by the point mass on $\arg\max_i p_i$; Gibbs' inequality gives $H_\times(p, q) \geq H(p)$, with the arguments the other way round. Example: $p = (0.7, 0.2, 0.1)$ gives $0.802$ at $q = p$ and $0.357$ at $q = (1, 0, 0)$.

Step 4 (Identification with FEP). Friston's free energy:

$$F_{FEP} = \int ds \, q(s) \ln \frac{q(s)}{p(s,o)}$$

Upon discretization $s \to \{s_i\}_{i=1}^N$:

$$F_{FEP} = \sum_i q_i \ln \frac{q_i}{p_i} = D_{KL}(q \| p)$$

UHM functional in the classical limit:

$$\mathcal{F}_{\text{UHM}} = H(q) + D_{KL}(q \| p) = D_{KL}(q \| p) + H(q)$$

The difference is an additive term $H(q)$, which does not depend on $p$ and therefore does not affect the optimal $q^*$ for fixed $p$:

$$\arg\min_q \mathcal{F}_{\text{UHM}} = \arg\min_q F_{FEP} = p$$

Thus, the optimal recognition densities coincide. $\blacksquare$

[✗] Error (retracted 2026-09-25): $H(q)$ depends on $q$, so adding it moves the optimum: $\arg\min_q D_{KL}(q\|p) = p$, but $\arg\min_q [H(q) + D_{KL}(q\|p)]$ is the point mass on $\arg\max_i p_i$.

4.6 Correspondence of constructions

Full correspondence table between UHM and FEP:

UHM construction	FEP construction	Limiting transition
$\Gamma \in \mathcal{D}(\mathbb{C}^7)$	$p(s,o)$ — generative model	$\Gamma_{ij} \to 0$, $\gamma_{ii} \to p_i$
$\varphi(\Gamma)$ — categorical self-model	$q^*(s\|o)$ — optimal recognition density	$\varphi \to$ Bayesian inversion
$S_{spec}(\Gamma)$ — spectral entropy	$H(p)$ — Shannon entropy	$S_{spec}(\mathrm{diag}(p)) = H(p)$
$D_{KL}(\psi(\Gamma) \| \Gamma)$ — quantum KL	$D_{KL}(q \| p)$ — classical KL	Diagonal limit
$\mathcal{F}[\psi; \Gamma]$ — variational functional	$F[q; p]$ — free energy	Theorem 4.1
Primitivity of $\mathcal{L}_0$	Ergodicity of Markov chain	Lindblad → Markov generator
CPTP channel $\psi$	Stochastic matrix $T$	Complete positivity → positivity
$\rho^* = \varphi(\Gamma)$ — fixed point	Posterior distribution $p(s\|o)$	Self-modeling → Bayesian inference
Markov blanket (algebraic)	Markov blanket (graphical)	$\mathcal{B}(\Gamma)$ → conditional independence graph

[✗] (2026-09-25): the rows identifying $\varphi(\Gamma)$ with the optimal recognition density, $\mathcal{F}$ with Friston's free energy and $\rho^*$ with the posterior rest on Theorems 4.2 (iii) and 4.3 and are retracted with them; the other rows are notational correspondences.

4.7 What is lost in the classical limit

The transition $\Gamma \to \mathrm{diag}(p)$ destroys three fundamental structures of UHM that have no classical analogs.

4.7.1 Coherences (off-diagonal elements)

Quantum coherences $\Gamma_{ij}$ ($i \neq j$) are correlations between the holon's dimensions not describable by classical probabilities.

In the full UHM functional, coherences contribute:

$$\mathcal{F}_{\text{full}} = \underbrace{H(q)}_{\text{classical}} + \underbrace{D_{KL}^{\text{diag}}(q \| p)}_{\text{classical}} + \underbrace{D_{KL}^{\text{off-diag}}(\psi(\Gamma) \| \Gamma)}_{\text{quantum remainder}}$$

where the quantum remainder:

$$D_{KL}^{\text{off-diag}} = \mathrm{Tr}\left[\psi(\Gamma) \log \psi(\Gamma)\right] - \mathrm{Tr}\left[\psi(\Gamma) \log \Gamma\right] - \left(\sum_i q_i \log q_i - \sum_i q_i \log p_i\right)$$

At $\Gamma_{ij} \to 0$ this term vanishes: $D_{KL}^{\text{off-diag}} \to 0$.

Consequence. Classical FEP cannot describe information integration between dimensions ($\Phi$ depends on coherences), quantum qualia (structure of subspaces $\mathcal{H}_k$), or interference between different aspects of the self-model.

4.7.2 Gap operator

The Gap operator $\mathcal{G}_{ij} = \|\Gamma_{ij}\| / \sqrt{\gamma_{ii} \gamma_{jj}}$ describes opacity between dimensions. In the classical limit:

$$\mathcal{G}_{ij} \to 0 \quad \forall i \neq j$$

The Gap structure disappears completely. This means loss of:

• Consciousness bifurcations (discontinuous transitions between regimes)
• Non-Markovian memory effects (Gap oscillations)
• Hamming code H(7,4) in the error correction structure (see Gap dynamics)

4.7.3 Regeneration $\mathcal{R}$

The regenerative term $\mathcal{R}$ of the evolution equation $\mathcal{L}_\Omega = \mathcal{L}_0 + \mathcal{R}$ is responsible for nonlinear feedback: the system actively restores coherences rather than passively dissipating.

In the classical limit ($\Gamma_{ij} \to 0$):

$$\mathcal{R}[\Gamma] \to 0$$

since regeneration operates on off-diagonal elements. Only the linear dissipation $\mathcal{L}_0$ remains, which in the classical limit reduces to a Markov generator:

$$\mathcal{L}_0[\mathrm{diag}(p)] \to Q \cdot p, \qquad Q_{ij} \geq 0 \text{ for } i \neq j, \quad \sum_i Q_{ij} = 0$$

— the standard Q-matrix of a continuous Markov chain.

Consequence. In the classical limit the UHM functional retains only passive minimization (dissipation); UHM's quantum dynamics adds active regeneration — the system's ability to restore complex structures lost during decoherence. This contrasts two regimes of UHM's own functional; it says nothing against Friston's FEP, which includes action through active inference and has a quantum formulation (§6.3).

4.8 What is preserved: prediction error minimization

Despite the losses, the core of the variational principle survives the classical limit:

Corollary 4.1 (Invariance of the minimization principle) — retracted 2026-09-25 [✗]

The principle "the system minimizes a functional balancing accuracy and complexity" is preserved across all regimes:

Regime	Functional	Accuracy	Complexity
Quantum (UHM)	$S_{spec}(\psi(\Gamma)) + D_{KL}(\psi(\Gamma) \| \Gamma)$	$-D_{KL}(\psi(\Gamma) \| \Gamma)$	$S_{spec}(\psi(\Gamma))$
Classical (FEP)	$\langle E \rangle_q - H(q)$	$\langle \ln p(o\|s) \rangle_q$	$D_{KL}(q \| p_{\text{prior}})$

Prediction error minimization (PEM) is the classical limit of categorical self-modeling $\varphi \dashv i$.

Retracted [✗]: this rests on Theorems 4.2 (iii) and 4.3; the quantum row's functional is the cross-entropy, whose minimiser is a pure state.

This was offered as the reason Friston's FEP works for classical systems (the brain in the neurocomputational description, biological organisms); with the reduction retracted, it explains nothing.

4.9 Structural diagram

┌─────────────────────────────────────────────────────────────────┐
│                         UHM (∞-topos)                           │
│                                                                 │
│  Level 0: Ω (primitive)                                         │
│       ↓                                                         │
│  Level 1: φ ⊣ i (categorical definition)                        │
│       ↓                                                         │
│  Level 2: φ = lim e^{tℒ_Ω}[Γ] (dynamical)                       │
│       ↓                                                         │
│  Level 3: φ = argmin [S_spec + D_KL] (T 3.1, retracted ✗)      │
│       ↓                                                         │
│  ┌─────────────────────────────────────────────────────────┐    │
│  │    Classical limit: Γ_ij → 0, R → 1/7                  │    │
│  │         (lost: coherences, Gap, ℛ)                      │    │
│  │                      ↓                                  │    │
│  │  ┌───────────────────────────────────────────────────┐  │    │
│  │  │   Friston's FEP: min F = min [⟨E⟩ - H]            │  │    │
│  │  │   (SPECIAL CASE: claim retracted ✗)               │  │    │
│  │  │   (preserved: prediction error minimization)      │  │    │
│  │  └───────────────────────────────────────────────────┘  │    │
│  └─────────────────────────────────────────────────────────┘    │
└─────────────────────────────────────────────────────────────────┘

The arrows from Level 3 down depict the retracted derivation (box at the top); Levels 0–2 are unaffected.

---

5. S_spec vs S_vN: justification of the choice

5.1 Definitions

Von Neumann entropy:

$$S_{vN}(\rho) := -\mathrm{Tr}(\rho \log \rho) = -\sum_i \lambda_i \log \lambda_i$$

Spectral entropy:

$$S_{spec}(A) := -\sum_i |\lambda_i| \log |\lambda_i|$$

where $\{\lambda_i\}$ are eigenvalues of operator $A$.

5.2 When they coincide

Theorem 5.1 (Equivalence for density matrices)

For density matrices $\rho$ (Hermitian, positive semi-definite, unit trace):

$$S_{spec}(\rho) = S_{vN}(\rho)$$

Proof: For $\rho \geq 0$ all $\lambda_i \geq 0$, therefore $|\lambda_i| = \lambda_i$. $\blacksquare$

5.3 Why use S_spec in UHM?

Reason 1: Generalization to non-Hermitian operators.

In some formalisms (Kraus operators, non-physical states), non-Hermitian operators appear. $S_{spec}$ is defined for them; $S_{vN}$ is not.

Reason 2: Connection to Kolmogorov complexity.

In the original UHM formulation (axiom-omega.md):

$S_{spec}(\cdot)$ — spectral entropy (replacing uncomputable Kolmogorov complexity)

Kolmogorov complexity $K(\cdot)$ is uncomputable. $S_{spec}$ serves as a computable upper bound:

$$S_{spec}(\rho) \leq K(\rho) + O(1)$$

5.4 Recommendation

For practical purposes in UHM:

• Use $S_{vN}$ for density matrices (standard quantum theory)
• Keep the notation $S_{spec}$ to indicate the connection with complexity theory
• In documentation indicate: "$S_{spec} = S_{vN}$ for density matrices"

---

6. Comparison with Friston's FEP

6.1 Correspondence table

Aspect	FEP (Friston)	UHM
Status	Postulate (phenomenological)	No variational principle for φ: Theorem 3.1 retracted (§3)
Domain	Classical distributions in the original formulation; generic quantum systems in Fields, Friston, Glazebrook and Levin (2022)	Density matrices on $\mathbb{C}^7$
Operator	Implicit	Explicit CPTP channel
Justification	Thermodynamics + Bayesian inference	Categorical adjunction
Circularity	Not resolved	Resolved (hierarchy Ω → φ)
Time	External parameter	Emergent (▷ on Ω)

6.2 How did Friston derive FEP without UHM?

Friston used three independent arguments:

1. Information-theoretic (Bayesian):

$$F = D_{KL}(q \| P(\cdot|S)) - \ln P(S)$$

This is an identity — a consequence of the definition of KL-divergence. FEP postulates that systems minimize F.

2. Thermodynamic:

Fluctuation theorems (Jarzynski, Crooks) connect free energy with non-equilibrium work. Stationary systems minimize F for thermodynamic reasons.

3. Cybernetic (self-organization):

Systems that do not minimize surprise "dissipate" — lose their identity. Survival ≡ minimization of F.

6.3 What UHM adds, and what it does not

1. Categorical justification:

In UHM, φ is defined by the structure of the ∞-topos. That a variational principle follows from it as a consequence was the claim of Theorem 3.1, retracted 2026-09-25; the page therefore has no variational principle for φ to set against the FEP's.

2. Quantum formulation — not a first:

UHM works with density matrices on $\mathbb{C}^7$. That the FEP is confined to classical distributions is not true: Fields, Friston, Glazebrook and Levin formulated it for generic quantum systems in the language of quantum information theory, with quantum systems acting as observers and agents (Prog. Biophys. Mol. Biol. 173, 36–59, 2022; arXiv:2112.15242). What this page set out to add was narrower still — a variational characterisation of its self-model $\varphi$ on $\mathcal{D}(\mathbb{C}^7)$ and a reduction of its decohered limit to the classical variational free energy (§3–§4) — and both are retracted (box at the top). The page does not compare this functional with the Fields et al. formulation, and neither is shown to contain the other.

3. Resolution of circularity:

UHM explicitly constructs the hierarchy: Ω → L_k → ℒ_Ω → φ (see dependency DAG). In FEP the connection between generative model and dynamics is implicit.

4. Emergent time:

In UHM, time is derived from the temporal modality ▷ on Ω. In FEP, time is an external parameter.

---

7. Consequences for UHM

7.1 Confirmation of consistency

The proof of Theorem 3.1 was taken to confirm the list below; items 1–2 are retracted with it (box at the top):

1. The variational characterization is a consequence of the categorical definition [✗]
2. The classical limit reproduces Friston's FEP [✗]
3. UHM's functional is quantum, but a quantum FEP predates it (Fields et al. 2022, §6.3); no relation between the two is proved here

7.2 Clarification of statement status

Statement	Old status	New status
φ = argmin [S_spec + D_KL]	"Property 4"	Theorem 3.1 — retracted 2026-09-25 [✗]: the argmin is a projection onto the top eigenvector of Γ
Classical limit of the functional	Implicit	Theorem 4.1 [T] (an identity: the limit is the cross-entropy)
FEP ⊂ UHM	Claimed	Not shown: Theorem 4.2 (iii)–(iv) retracted [✗]
$S_{spec} + D_{KL} \to F_{FEP}$	Not proven	Not proven: Theorem 4.3 retracted [✗]
S_spec = S_vN for ρ	Not clarified	Theorem 5.1 [T] (proven)

7.3 New corollaries

Corollary 7.1 (UHM's quantum variational principle) — retracted 2026-09-25 [✗].

The page stated that for quantum systems the generalized principle holds:

$$\Gamma_* = \arg\min_{\Gamma} \left[S_{vN}(\Gamma) + D_{KL}(\Gamma \| \Gamma_0)\right]$$

where $\Gamma_0$ is the initial/reference state. The functional equals $-\mathrm{Tr}(\Gamma \log \Gamma_0)$, so its minimiser is a projection onto the top eigenvector of $\Gamma_0$ — a pure state, not a principle for open systems.

Corollary 7.2 (Thermodynamic interpretation) — retracted 2026-09-25 [✗].

The page stated that minimization of $\mathcal{F}$ is equivalent to minimization of entropy production in an open quantum system; this rested on Steps 4–5 of the proof of Theorem 3.1, which are not established.

---

8. Technical Lemmas

Lemma A.1 (Entropy production in Lindblad dynamics)

For $\mathcal{L}[\rho] = \sum_k (L_k \rho L_k^\dagger - \frac{1}{2}\{L_k^\dagger L_k, \rho\})$:

$$\frac{dS_{vN}(\rho)}{dt} = -\mathrm{Tr}(\mathcal{L}[\rho] \log \rho) \geq 0$$

Corrected 2026-09-25: the inequality holds for unital generators ($\mathcal{L}[I] = 0$, for instance all $L_k$ normal). For a general Lindblad generator with a stationary state $\sigma$, what holds is that the relative entropy does not increase, $\frac{d}{dt} D_{KL}(\rho(t) \| \sigma) \leq 0$ (Spohn 1978; monotonicity of relative entropy under CPTP maps); $S_{vN}$ itself can fall — UHM's dynamics lowers it on the way from $I/7$ to the attractor (T-96).

Lemma A.2 (Uniqueness of stationary state)

If $\mathcal{L}_\Omega$ is primitive (no non-trivial subspaces), then $\exists! \rho_*: \mathcal{L}_\Omega[\rho_*] = 0$.

Lemma A.3 (Convergence to stationary)

For primitive $\mathcal{L}_\Omega$:

$$\lim_{\tau \to \infty} e^{\tau\mathcal{L}_\Omega}[\rho] = \rho_* \quad \forall \rho$$

---

9. References

1. Friston K. "The free-energy principle: a unified brain theory?" Nature Reviews Neuroscience 11, 127-138 (2010)
2. Spohn H. "Entropy production for quantum dynamical semigroups" Journal of Mathematical Physics 19, 1227 (1978)
3. Lindblad G. "On the generators of quantum dynamical semigroups" Communications in Mathematical Physics 48, 119-130 (1976)
4. Lurie J. "Higher Topos Theory" Princeton University Press (2009)
5. Fields C., Friston K., Glazebrook J.F., Levin M. "A free energy principle for generic quantum systems" Progress in Biophysics and Molecular Biology 173, 36-59 (2022); arXiv:2112.15242

---

10. Summary

Key results

Theorems:

1. Theorem 3.1 — retracted [✗]: the functional $S_{spec} + D_{KL}$ is minimised by a projection onto the top eigenvector of $\Gamma$, not by the categorically defined φ
2. Theorem 4.1 [T]: In the classical limit ($\Gamma_{ij} \to 0$, $R \to R_{\min}$) the UHM functional reduces to $H(q) + D_{KL}(q \| p)$, the cross-entropy $-\sum_i q_i \ln p_i$ (not Friston's free energy)
3. Theorem 4.2 — (iii)–(iv) retracted [✗]: the classical limit of the UHM functional does not reproduce Friston's FEP
4. Theorem 4.3 — retracted [✗]: the two minimisation problems have different solutions (point mass on $\arg\max_i p_i$ against $q = p$)
5. Theorem 5.1 [T]: $S_{spec} = S_{vN}$ for density matrices

Main conclusion (corrected 2026-09-25): the page does not derive Friston's FEP from UHM. What it establishes is Theorem 4.1 (an identity) and Theorem 5.1; whether φ minimises any natural functional, and whether the FEP arises as a limit of UHM, are open — research programme [Pr].

Compatibility with octonionic norm — withdrawn 2026-09-25 [✗]

Withdrawn: it concerned the variational principle of Theorem 3.1, retracted above. The box read: the variational principle $\varphi = \arg\min \mathbb{E}[S_{spec} + D_{KL}]$ is compatible with the octonionic interpretation: the norm of $\mathbb{O}$ ($|xy| = |x||y|$) ensures consistency of the metric used in $D_{KL}$ with the algebraic structure of the state space. Bridge [T] (T15). See structural derivation.

---

Related documents:

• Axiom Ω⁷ — definition of φ
• Formalization of φ — constructive realization
• Theories of consciousness — comparison with FEP