(sec-appendix-a-full-derivations)= # {ref}`Appendix A `: Full Derivations (Capacity-Constrained Curvature Functional) ## TLDR - This appendix records the **full derivations** behind the main geometric/information-theoretic claims in Volume 1. - It is reference material: use it when you want the complete steps, not first-pass intuition. - Most readers can skim the statements and return to details when implementing or auditing proofs. (sec-appendix-a-capacity-constrained-curvature-functional)= ## A.1 Capacity-Constrained Curvature Functional (Variational Principle) The capacity diagnostic is the operational postulate $I_{\text{bulk}}\le C_{\partial}$. A measured violation is rejected by the Sieve ({ref}`Section 3 `, Node 13). The curvature variation below is a separate identity for a curvature--risk functional. It does not impose the equality $I_{\text{bulk}}=C_{\partial}$; an active constraint would require an explicit multiplier term and a $G$-differentiable information model. :::{prf:definition} A.1.1 (Boundary capacity form) :label: def-a-boundary-capacity-form Define the boundary capacity $(n\!-\!1)$-form For a declared cutoff hypersurface $\partial_{\varepsilon}\mathcal Z$, use $$ \omega_{\partial} := \frac{1}{\eta_\ell}\, dA_G, $$ so that $C_{\partial}(\partial_{\varepsilon}\mathcal Z)= \oint_{\partial_{\varepsilon}\mathcal Z}\omega_{\partial}$ (Definition {prf:ref}`def-boundary-capacity-area-law-at-finite-resolution`). If no geometric cutoff is part of the model, replace this area expression by the interface-channel capacity. ::: :::{prf:definition} A.1.2 (Boundary-capacity constraint functional) :label: def-a-boundary-capacity-constraint-functional Define the diagnostic difference (not a variational constraint in the action) $$ \mathcal{C}[G,V] := \underbrace{\int_{\mathcal{Z}} \iota_{\mathrm{bulk}}\, d\mu_G}_{I_{\text{bulk}}} \;-\; \underbrace{\oint_{\partial_{\varepsilon}\mathcal Z}\omega_{\partial}}_{C_{\partial}}, $$ where $\iota_{\mathrm{bulk}}$ is the relative-information density in {prf:ref}`def-information-density-and-bulk-information-volume`. The realised shutter inflow is the separate rate $\lambda_{\mathrm{in}}=\mathbb{E}[I(X;K)]$ (Definition {prf:ref}`def-grounding-rate`). The coupling-window definition {prf:ref}`thm-information-stability-window-operational` supplies its admissible operating range. ::: :::{prf:definition} A.1.3 (Risk Lagrangian density) :label: def-a-risk-lagrangian-density Fix a smooth potential $V\in C^\infty(\mathcal{Z})$. A canonical risk Lagrangian density is the scalar-field functional $$ \mathcal{L}_{\text{risk}}(V;G) := \frac{1}{2}\,G^{ab}\nabla_a V\,\nabla_b V + U(V), $$ where $U:\mathbb{R}\to\mathbb{R}$ is a (possibly learned) on-site potential capturing non-gradient costs. (The sign convention is chosen for a Riemannian metric; see e.g. Lee, *Riemannian Manifolds*, 2018, for the variational identities used below.) ::: :::{prf:definition} A.1.4 (Curvature--risk functional with a cutoff penalty) :label: def-a-capacity-constrained-curvature-functional Let $R(G)$ be the scalar curvature of $G$ and let $\Lambda\in\mathbb{R}$ be a constant. Define the functional $$ \mathcal{S}[G,V] := \int_{\mathcal{Z}}\left(R(G)-2\Lambda - 2\kappa\,\mathcal{L}_{\text{risk}}(V;G)\right)d\mu_G \;-\; 2\kappa\oint_{\partial_{\varepsilon}\mathcal Z}\omega_{\partial}, $$ with coupling $\kappa\in\mathbb{R}$. The sign convention makes the positive Riemannian risk tensor below appear on the right-hand side of the stationarity equation. The cutoff term has no interior variation under clamping, and $\Lambda$ remains a free curvature offset rather than a multiplier determined by $C_{\partial}$. *Remark (why $\Lambda$ is allowed).* A constant term in the integrand is the simplest coordinate-invariant scalar density and produces a $\Lambda G_{ij}$ term in the metric Euler–Lagrange equation. Here $\Lambda$ plays the role of a baseline curvature / capacity offset. ::: (sec-appendix-a-first-variation)= ## A.2 First Variation (Expanded Derivation) We work in the standard calculus of variations on Riemannian manifolds with boundary. Assume: 1) $G$ is $C^2$ and $V$ is $C^2$ (so curvature and gradients are well-defined), 2) variations $\delta G^{ij}$ are smooth, symmetric, and compactly supported in $\mathcal{Z}$ or satisfy Dirichlet boundary conditions $\delta G^{ij}\vert_{\partial\mathcal{Z}}=0$ (the boundary is clamped by the sensorium; cf. Definition {prf:ref}`def-observation-inflow-form` / Theorem {prf:ref}`thm-generalized-conservation-of-belief`). Under these hypotheses, the first variation of $\mathcal{S}$ is well-defined as a distribution; the standard identities below can be found in standard differential-geometry references (e.g. {cite}`lee2018riemannian`). (sec-appendix-a-variation-of-the-volume-form)= ### A.2.1 Variation of the volume form Let $d\mu_G=\sqrt{|G|}\,dz^n$. The determinant identity gives $$ \delta \sqrt{|G|} = -\frac{1}{2}\sqrt{|G|}\,G_{ij}\,\delta G^{ij}, $$ equivalently $\delta d\mu_G = -\tfrac12\,G_{ij}\,\delta G^{ij}\, d\mu_G$. (sec-appendix-a-variation-of-the-curvature-term)= ### A.2.2 Variation of the curvature term Write the curvature functional as $\mathcal{S}_{\text{geo}}[G]:=\int_{\mathcal{Z}}R(G)\,d\mu_G$. The variation splits as $$ \delta(R\,d\mu_G) = (\delta R)\,d\mu_G + R\,\delta d\mu_G. $$ For the scalar curvature, use $$ R = G^{ij}R_{ij}, $$ hence $$ \delta R = R_{ij}\,\delta G^{ij} + G^{ij}\,\delta R_{ij}. $$ The Palatini identity gives $$ \delta R_{ij} = \nabla_k(\delta \Gamma^k_{ij})-\nabla_j(\delta\Gamma^k_{ik}), $$ and the Christoffel variation is $$ \delta\Gamma^k_{ij} = \frac12\,G^{k\ell}\left(\nabla_i \delta G_{j\ell}+\nabla_j \delta G_{i\ell}-\nabla_\ell \delta G_{ij}\right), $$ where $\delta G_{ij} = -G_{ia}G_{jb}\,\delta G^{ab}$. Substituting and collecting terms yields the standard decomposition $$ \delta\mathcal{S}_{\text{geo}} = \int_{\mathcal{Z}}\left(R_{ij}-\frac12 R\,G_{ij}\right)\delta G^{ij}\,d\mu_G + \oint_{\partial\mathcal{Z}} \mathcal{B}_{\text{curv}}(\delta G,\nabla\delta G), $$ where $\mathcal{B}_{\text{curv}}$ is an explicit boundary $(n\!-\!1)$-form built from $\delta\Gamma$ (equivalently from $\delta G$ and its first derivatives). For a well-posed Dirichlet variational problem one can add an appropriate boundary term to cancel $\mathcal{B}_{\text{curv}}$. On the declared cutoff we impose $\delta G\vert_{\partial_{\varepsilon}\mathcal Z}=0$ and the boundary term vanishes. (sec-appendix-a-variation-of-the-risk-term)= ### A.2.3 Variation of the risk term Let $\mathcal{S}_{\text{risk}}[G,V] := \int_{\mathcal{Z}}\mathcal{L}_{\text{risk}}(V;G)\,d\mu_G$. Define the (Riemannian-signature) risk tensor by $$ T_{ij} := \frac{2}{\sqrt{|G|}}\frac{\delta(\sqrt{|G|}\,\mathcal{L}_{\text{risk}})}{\delta G^{ij}}. $$ Holding $V$ fixed under $\delta G$ and using $\delta d\mu_G = -\tfrac12 G_{ij}\delta G^{ij} d\mu_G$ gives the standard identity $$ \delta \mathcal{S}_{\text{risk}} = \frac12 \int_{\mathcal{Z}} T_{ij}\,\delta G^{ij}\,d\mu_G. $$ For the risk Lagrangian $$ \mathcal{L}_{\text{risk}}=\tfrac12 G^{ab}\nabla_a V\nabla_b V + U(V), $$ the explicit computation yields $$ T_{ij} = \nabla_i V\,\nabla_j V - G_{ij}\left(\frac12\,G^{ab}\nabla_a V\nabla_b V + U(V)\right). $$ (sec-appendix-a-capacity-term-and-the-emergence-of)= ### A.2.4 Capacity (boundary) term and the emergence of $\Lambda$ The explicit boundary penalty $-2\kappa\oint_{\partial_{\varepsilon}\mathcal Z}\omega_{\partial}$ depends only on the induced cutoff metric through $dA_G$. Under the clamped boundary condition $\delta G\vert_{\partial_{\varepsilon}\mathcal Z}=0$, its first variation vanishes. The remaining constant $\Lambda$ in Definition A.1.4 is a free bulk curvature offset. It is not fixed by the interface capacity unless an additional multiplier equation is supplied. Formally, varying $-2\Lambda\int_{\mathcal{Z}} d\mu_G$ gives $$ \delta\left(-2\Lambda\int_{\mathcal{Z}} d\mu_G\right) = \int_{\mathcal{Z}} \Lambda G_{ij}\,\delta G^{ij}\,d\mu_G. $$ (sec-appendix-a-recovery-of-the-metric-stationarity-condition)= ## A.3 Recovery of the Metric Stationarity Condition :::{prf:lemma} A.3.1 (Divergence-to-boundary conversion) :label: lem-a-divergence-to-boundary-conversion For any sufficiently regular information flux field $\mathbf{j}$ on $\mathcal{Z}$, $$ \int_{\mathcal{Z}} \operatorname{div}_G(\mathbf{j})\, d\mu_G = \oint_{\partial \mathcal{Z}} \langle \mathbf{j}, \mathbf{n}\rangle\, dA_G, $$ which is the Riemannian divergence theorem underlying the global balance equation in Theorem {prf:ref}`thm-generalized-conservation-of-belief`. ::: :::{prf:theorem} A.3.2 (Capacity-consistency identity; proof of Theorem {prf:ref}`thm-capacity-constrained-metric-law`) :label: thm-a-capacity-consistency-identity-proof-of-theorem Under the hypotheses of Section A.2, stationarity of $\mathcal{S}[G,V]$ with respect to arbitrary variations $\delta G^{ij}$ that vanish on $\partial\mathcal{Z}$ implies the Euler–Lagrange equation $$ R_{ij} - \frac{1}{2}R\,G_{ij} + \Lambda G_{ij} = \kappa\, T_{ij}, $$ with $T_{ij}$ given by Section A.2.3. *Proof.* Combine Sections A.2.1–A.2.4: $$ \delta\mathcal{S} = \int_{\mathcal{Z}}\left[\left(R_{ij}-\frac12 R\,G_{ij}\right) + \Lambda G_{ij} - \kappa T_{ij}\right]\delta G^{ij}\,d\mu_G + \text{(boundary terms)}. $$ Boundary terms vanish under the clamped boundary condition (or after adding an appropriate boundary term). Because $\delta G^{ij}$ is arbitrary in the interior, the fundamental lemma of the calculus of variations implies the bracketed tensor must vanish pointwise almost everywhere, yielding the stated identity (see e.g. Evans, *Partial Differential Equations*, 2010, for the functional-analytic lemma). *Interpretation.* The Ricci curvature governs local volume growth. This stationarity identity is a consistency equation for the declared risk model; the separate capacity postulate tests whether the resulting belief is grounded at the interface. *Remark (regularizer).* The squared residual of this identity is the metric-law loss $\mathcal{L}_{\mathrm{EFE}}$ in {ref}`Appendix F `. ::: (sec-appendix-a-pitchfork-bifurcation-at-the-origin)= ## A.3b Radial Escape and Angular Selection at the Origin (Crossover calculation) This section records the radial and angular calculation used by Definition {prf:ref}`def-control-field-at-origin`. It establishes an outward radial drift and a finite-time direction-selection mechanism; it does not establish a pitchfork bifurcation or a critical temperature. **Setup.** Consider the Langevin equation on the Poincare disk $\mathbb{D}$ from Definition {prf:ref}`prop-so-d-symmetry-at-origin`: $$ dz_\tau = -\nabla_G U(z_\tau)\, d\tau + \sqrt{2T_c}\, G^{-1/2}(z_\tau)\, dW_\tau, $$ with $U(z) = -2\operatorname{artanh}(|z|)$ and initial condition $z(0) = 0$. **Step 1: Radial expansion near the origin.** For $r=|z|>0$, expand the radial potential near the origin: $$ U(r) = -2\operatorname{artanh}(r) = -2r - \frac{2}{3}r^3 + O(r^5). $$ The Euclidean gradient for $r>0$ is: $$ \nabla U(z) = -\frac{2z}{r(1-r^2)}. $$ The Riemannian gradient (with $G^{-1} = \frac{(1-r^2)^2}{4}I$) is: $$ \nabla_G U(z) = -\frac{1-r^2}{2r}z. $$ **Step 2: Generator and polar-coordinate caveat.** Away from $r=0$, the generator of the stated covariant Langevin model is $$ \mathcal{L}f=-\langle\nabla_G U,\nabla_G f\rangle_G+T_c\Delta_G f. $$ In polar coordinates the Laplace--Beltrami operator supplies the radial (Bessel-type) correction, and the coordinates are singular at $r=0$. A flat Cartesian linearization therefore cannot be used to infer a stationary law or a Landau normal form at the origin. **Step 3: Formal zero-flux density and its scope.** If one formally imposes zero flux for the conservative diffusion, the density relative to Riemannian volume would be $$ p_{\mathrm{formal}}(z)\propto\exp\!\left(-\frac{U(z)}{T_c}\right) =\exp\!\left(\frac{2\operatorname{artanh}(r)}{T_c}\right). $$ This is rotationally symmetric, but it is not normalizable on the full Poincare disk because it diverges as $r\uparrow1$. Consequently the displayed model has no invariant Gibbs probability without an outer boundary, a confining modification, or another integrability hypothesis. The detailed-balance calculation is therefore only a formal local statement: $$ \nabla_G U + T_c\,G^{-1}\,\nabla \log p_{\mathrm{formal}} = 0. $$ Substituting and solving: $$ \nabla \log p_{\mathrm{formal}} = -\frac{1}{T_c}\,G\,\nabla_G U = -\frac{1}{T_c}\,\nabla U. $$ Integrating: $$ p_{\mathrm{formal}}(z) \propto \exp\left(-\frac{U(z)}{T_c}\right) = \exp\left(\frac{2\operatorname{artanh}(r)}{T_c}\right). $$ The formal density is **rotationally symmetric** (depends only on $r$), but its non-normalizability limits the claim to local symmetry of the coefficients. **Step 4: Radial and angular drift.** Write $z = re^{i\theta}$ in polar coordinates. The effective potential in the radial direction is: $$ U_{\text{eff}}(r) = -2\operatorname{artanh}(r). $$ The radial force is $F_r = -\frac{dU_{\text{eff}}}{dr} = \frac{2}{1-r^2} > 0$ for all $r \in [0, 1)$. This means: - The origin $r = 0$ is an **unstable equilibrium** (force points outward). - There is no stable equilibrium in the interior; the "stable point" is at the boundary $r = 1$. - The angular direction $\theta$ is **neutral** (no restoring force). **Step 5: Symmetry breaking mechanism.** For small $\tau$: 1. The noise term dominates: $z(\tau) \approx \sqrt{2T_c}\int_0^\tau G^{-1/2} dW_\tau$ performs a random walk. 2. This random walk samples directions $\theta$ uniformly from $[0, 2\pi)$. 3. Once $|z|$ exceeds a threshold (order $\sqrt{T_c}$), the deterministic drift $-\nabla_G U$ takes over. 4. The trajectory then flows radially outward along the selected direction $\theta$. This is a finite-time radial escape with angular selection by noise and, when present, the policy drift. It is not a supercritical pitchfork: there is no pair of stable interior equilibria or cubic normal form in this SDE. **Step 6: Exit-time and direction statements.** For a declared stopping radius $R\in(0,1)$, define $$ \tau_R:=\inf\{\tau\ge0:r(\tau)\ge R\}. $$ Its expectation depends on $R$, the metric, the boundary condition, and the full radial drift/noise coefficients. No Kramers barrier formula follows here because the origin is unstable and $U$ has no interior barrier. If the policy field is zero and the noise and stopping rule are rotationally invariant, the exit angle satisfies $\theta^*\sim\mathrm{Uniform}[0,2\pi)$. A nonzero tangential policy or anisotropic boundary changes this law and must be analyzed from the angular SDE. This completes the finite-time radial/angular calculation supporting Definition {prf:ref}`def-control-field-at-origin`. (sec-appendix-a-overdamped-limit-via-singular-perturbation)= ## A.4 Overdamped Limit via Singular Perturbation (Proof of Theorem {prf:ref}`thm-overdamped-limit`) This section provides the full proof of Theorem {prf:ref}`thm-overdamped-limit` using singular perturbation theory. **Setup.** Work with the conservative second-order SDE in Theorem {prf:ref}`thm-overdamped-limit`, with $\beta_{\mathrm{curl}}=0$ and $u_\pi=0$: $$ m\,\ddot z^k+\gamma\,\dot z^k+G^{kj}\partial_j\Phi_{ \mathrm{eff}}+\Gamma^k_{ij}\dot z^i\dot z^j =\sqrt{2\gamma T_c}\,(G^{-1/2})^{kj}\,\xi^j. $$ The factor $\sqrt{\gamma}$ in the noise is required by the fluctuation--dissipation relation. Let $t$ be physical time and put $s=t/\gamma$ for the computation-time variable used in the theorem. **Step 1: Fast velocity scale.** The velocity equation has relaxation time $m/\gamma$. On every compact coordinate patch on which $G$, $G^{-1}$, $\Phi_{\mathrm{eff}}$, and the Christoffel symbols are bounded and locally Lipschitz, the velocity therefore has a fast Ornstein--Uhlenbeck layer while the position evolves on the slower time scale. The Smoluchowski--Kramers reduction gives, in physical time, $$ dz^k=-\frac1\gamma G^{k\ell}\partial_\ell\Phi_{\mathrm{eff}}\,dt -\frac{T_c}{\gamma}G^{ij}\Gamma^k_{ij}\,dt +\sqrt{\frac{2T_c}{\gamma}}\,(G^{-1/2})^{kj}\,dW_t^j +o(1). $$ The $o(1)$ term is in the chosen local topology as $m/\gamma\to0$; a global estimate additionally requires the confinement and boundary hypotheses of the theorem. **Step 2: Computation-time rescaling.** Since $dt=\gamma\,ds$ and $dW_t=\sqrt{\gamma}\,dW_s$, the preceding equation becomes $$ dz^k=\left[-G^{k\ell}\partial_\ell\Phi_{\mathrm{eff}} -T_cG^{ij}\Gamma^k_{ij}\right]ds +\sqrt{2T_c}\,(G^{-1/2})^{kj}\,dW_s^j. $$ This is the Ito equation stated in Theorem {prf:ref}`thm-overdamped-limit`. The contraction $G^{ij}\Gamma^k_{ij}$ is the coordinate correction for the chosen Riemannian reference measure; it is not the inertial geodesic term. **Step 3: Inertial and geodesic terms.** On the fast layer $\dot z$ is $O(\gamma^{-1})$ in physical time for a bounded force, so $\Gamma(\dot z,\dot z)$ is $O(\gamma^{-2})$ there. Its contribution vanishes in the singular limit after the declared rescaling. This argument is conservative: adding a curl mobility, policy forcing, or non-reversible jumps changes the reduced generator and is not covered by this appendix. **Step 4: Scope of the error statement.** The fast layer has width $O(m/\gamma)$, but the displayed exponential bound sometimes quoted for Euclidean constant-coefficient Langevin systems does not follow from the hypotheses above. A quantitative bound requires uniform derivative, confinement, and boundary estimates (and a choice of norm); none is asserted here. Under those additional standard Smoluchowski--Kramers hypotheses, the local reduction can be promoted to a finite-horizon error estimate. This completes the singular-perturbation proof sketch of Theorem {prf:ref}`thm-overdamped-limit`. $\square$ :::{prf:remark} Physical interpretation :label: rem-physical-interpretation The overdamped limit corresponds to: - **Information geometry:** The "friction" $\gamma$ represents the rate of information dissipation (forgetting). High friction means the system equilibrates quickly to the local gradient. - **Diffusion models:** Standard score-based diffusion models operate entirely in the overdamped regime, with $\gamma \to \infty$ implicitly. - **Neural network training:** The geodesic term $\Gamma(\dot{z},\dot{z})$ can be interpreted as a "momentum correction" that accounts for the curvature of the loss landscape. In standard gradient descent (overdamped), this term is ignored. ::: (sec-appendix-a-classification-as-relaxation)= ## A.5 Classification as Relaxation (Conditional derivation) :::{prf:proposition} Conditional Classification Relaxation :label: thm-classification-as-relaxation-a Under the conservative, deterministic overdamped dynamics with the smooth class-conditioned potential $V_y$: $$ dz = -G^{-1}(z)\nabla V_y(z)\,ds, \qquad T_c=0, $$ The limiting chart assignment satisfies $K(\lim_{s\to\infty}z(s))\in\mathcal{A}_y$ whenever the trajectory converges to a minimum in $K^{-1}(\mathcal{A}_y)$ and the initial condition lies in its basin. ::: (proof-thm-classification-as-relaxation-a)= :::{prf:proof} **Step 1: Lyapunov function.** Use the smooth potential itself, $$ L(z):=V_y(z). $$ The soft-router definition of $V_y$ is differentiable wherever the router weights are differentiable. No claim that a class-$y$ chart is a global minimum is made here; that is a training or modeling condition. **Step 2: Deterministic descent.** For the conservative overdamped equation, $$ \frac{dL}{ds}=dV_y(\dot z) =-\nabla V_y(z)^{\mathsf T}G^{-1}(z)\nabla V_y(z)\le 0. $$ Thus every trajectory that remains in a compact sublevel set approaches the largest invariant subset of $\{z:\nabla V_y(z)=0\}$ under the usual smoothness and completeness hypotheses. If the selected minimum is in $K^{-1}(\mathcal A_y)$, the limiting chart belongs to $\mathcal A_y$. This is a conditional deterministic statement. At $T_c>0$, noise permits barrier crossing and the basin claim becomes a finite-time or small-noise estimate; it is not almost-sure convergence to a class at fixed temperature. The same argument does not cover a hard chart index, nonzero curl, or an unspecified jump process. This completes the conditional derivation. \(\square\) ::: :::{prf:remark} Connection to Classification Accuracy :label: rem-connection-to-classification-accuracy The theorem provides a geometric interpretation of classification accuracy: a sample $x$ is correctly classified if and only if $\text{Enc}(x) \in \mathcal{B}_{y_{\text{true}}}$. Misclassification occurs when the encoder maps $x$ to the wrong basin—either due to encoder limitations or overlap between class distributions in observation space. ::: (sec-appendix-a-area-law)= ## A.6 Operational area-law normalization and conditional counting This section records the normalization used for the Causal Information Bound and the hypotheses of the associated counting and radial calculations. It does not prove a universal bulk-to-boundary area law or a physical identification with black-hole entropy. The $1/4$ value in the two-dimensional convention comes from the normalized Poincaré metric together with an explicit one-nat cell permit. **Setup.** Let $(\mathcal{Z}, G)$ be the latent Riemannian manifold. We seek the maximum bulk information $I_{\text{bulk}}$ that can be distinguished by an external observer through the boundary $\partial\mathcal{Z}$. **Scope.** Sections A.6.0–A.6.0h give local geometric facts and a conditional independent-cell counting model. Sections A.6.1–A.6.3 record a formal spherical ansatz and state the additional bulk-to-boundary permit that a field-theoretic derivation would require. Sections A.6.4–A.6.7 assemble the operational normalization and its dimension-dependent coefficient. The former field-theoretic route is retained for auditability, not as an established derivation. (sec-appendix-a-foundational-axioms)= ### A.6.0 Foundational Axioms for Microstate Counting This section states the information-theoretic conventions used by the conditional counting model. The black-hole microstate literature is a structural analogy only; no physical entropy identification is imported {cite}`strominger1996microscopic`. :::{prf:axiom} A.6.0a (Operational Distinguishability) :label: ax-a-operational-distinguishability Two probability distributions $p, q \in \mathcal{P}(\mathcal{Z})$ are **operationally distinguishable** if and only if: $$ D_{\text{KL}}(p \| q) \geq 1 \text{ nat}. $$ *Justification.* This is an **operational definition**, not a derived fact. The choice of 1 nat as the threshold is grounded in: 1. **Asymptotic error exponent.** For $n$ i.i.d. samples, the optimal Type II error probability at fixed Type I error decays as $\exp(-n \cdot D_{\text{KL}})$ (Stein's lemma). Thus $D_{\text{KL}} = 1$ nat corresponds to error decay rate $e^{-n}$. 2. **Information-theoretic meaning.** 1 nat = log(e) ≈ 1.44 bits represents a "natural unit" of information, where the likelihood ratio $p(x)/q(x)$ has expected log-value 1 under $p$. 3. **Dimensional analysis.** The nat is the natural unit when using natural logarithms; choosing 1 nat as the threshold makes the subsequent formulas dimensionally consistent. *Remark.* Alternative thresholds (e.g., 1 bit = ln 2 nats) would change the numerical coefficient in the Area Law but not its structure. ::: :::{prf:theorem} A.6.0b (Chentsov's Uniqueness Theorem) :label: thm-a-chentsov-uniqueness For a regular finite-dimensional family of strictly positive distributions, the **Fisher Information Metric** is, up to constant scaling, the unique Riemannian metric invariant under all Markov morphisms in the statistical category used by Chentsov's theorem. **Statement.** Let $\mathcal{M}$ be a statistical manifold parameterized by $\theta \in \Theta$. Any Riemannian metric $g$ on $\mathcal{M}$ satisfying: 1. **Markov invariance:** $g$ is preserved under the specified Markov morphisms (conditional expectations) 2. **Smoothness and regularity:** $g$ varies smoothly with $\theta$ and the model satisfies the regularity assumptions of the theorem is proportional to the Fisher Information Metric: $$ g_{ij}(\theta) = c \cdot \mathbb{E}_\theta\left[\frac{\partial \log p(x|\theta)}{\partial \theta^i} \frac{\partial \log p(x|\theta)}{\partial \theta^j}\right] $$ for some constant $c > 0$. *Proof.* This is the standard external uniqueness theorem; see Chentsov (1982) {cite}`chentsov1982statistical` and Campbell (1986) {cite}`campbell1986extended`. The present volume invokes it only as a regularity and normalization reference. It does not use the theorem to infer a finite cell-to-nat assignment or an area law. $\square$ *Scope.* Within that statistical category, Chentsov's theorem identifies the Fisher metric up to scale. The scale and the operational cell convention used below remain model permits. ::: ::::{admonition} Physics Isomorphism: Fisher Information Metric :class: note :name: pi-fisher-information **In Physics:** The Fisher Information Metric $\mathcal{F}_{ij}(\theta) = \mathbb{E}\left[\frac{\partial \log p}{\partial \theta^i}\frac{\partial \log p}{\partial \theta^j}\right]$ is the unique Riemannian metric on statistical manifolds invariant under sufficient statistics (Chentsov's Theorem) {cite}`chentsov1982statistical,amari1985differential`. **In Implementation:** The latent metric $G(z)$ combines value curvature with Fisher Information ({ref}`Section 2.5 `): $$ G_{ij}(z) = \nabla^2_{ij} V(z) + \lambda\,\mathcal{F}_{ij}(z) $$ where $\mathcal{F}_{ij} = \mathbb{E}_{a\sim\pi}[\partial_i \log\pi \cdot \partial_j \log\pi]$ is the state-space Fisher component $G_\pi$. **Correspondence Table:** | Information Geometry | Agent (Latent Metric) | |:---------------------|:----------------------| | Parameter space $\Theta$ | Latent state space $\mathcal{Z}$ | | Fisher metric $\mathcal{F}_{ij}$ | Base metric contribution | | Sufficient statistics | Macro-state $K$ (chart index) | | KL divergence $D_{KL}$ | Squared geodesic distance (locally) | | Natural gradient | Metric-aware policy gradient | **Significance:** Chentsov's theorem (Theorem {prf:ref}`thm-a-chentsov-uniqueness`) proves the Fisher metric is not a choice but a necessity—any geometry respecting statistical structure must be proportional to Fisher. :::: :::{prf:definition} A.6.0c (Computational Microstate) :label: def-a-computational-microstate A **computational microstate** at resolution $\ell$ is a complete specification of the agent's internal configuration $\mu = (\rho, K, \theta)$ where: - $\rho \in \mathcal{P}(\mathcal{Z})$ is the belief distribution over the latent manifold - $K \in \{1, \ldots, |\mathcal{K}|\}$ is the active chart assignment - $\theta$ are the model parameters discretized at the Levin Length scale: positions resolved to precision $\ell_L$, probabilities resolved to precision $e^{-1}$ in KL divergence. Two microstates $\mu_1, \mu_2$ are **boundary-distinguishable** if an external observer, receiving only boundary observations $\partial\mathcal{Z}$, can distinguish them with probability $> 1 - e^{-1}$. *Remark (Analogy to Physics).* In black hole thermodynamics, a microstate is a specific quantum configuration of the horizon degrees of freedom. Here, a microstate is a specific configuration of the agent's belief state. The boundary plays the role of the horizon: internal distinctions not visible at the boundary do not count toward the entropy. ::: (sec-appendix-a-microstate-counting)= ### A.6.0d Conditional microstate counting We now state the conditional counting model for boundary-distinguishable microstates. It uses the explicit channel permit below and does not derive a capacity law for an arbitrary latent model. :::{prf:lemma} A.6.0d (Geodesic Distance on the Probability Simplex) :label: lem-a-geodesic-distance-probability-simplex On the 1-simplex $\Delta^1 = \{(p, 1-p) : p \in [0,1]\}$ with Fisher Information Metric, the geodesic distance from the uniform distribution $(1/2, 1/2)$ to a vertex $(1, 0)$ is: $$ d_{\text{Fisher}}\left(\tfrac{1}{2}, 1\right) = \frac{\pi}{2}. $$ *Proof.* The Fisher metric on $\Delta^1$ is: $$ ds^2 = \frac{dp^2}{p(1-p)}. $$ Introduce the angular parameterization $p = \cos^2(\theta/2)$, so that $1-p = \sin^2(\theta/2)$ and: $$ dp = -\cos(\theta/2)\sin(\theta/2)d\theta = -\frac{1}{2}\sin\theta \, d\theta. $$ Then: $$ ds^2 = \frac{\frac{1}{4}\sin^2\theta \, d\theta^2}{\cos^2(\theta/2)\sin^2(\theta/2)} = \frac{\frac{1}{4}\sin^2\theta \, d\theta^2}{\frac{1}{4}\sin^2\theta} = d\theta^2. $$ The uniform distribution $(1/2, 1/2)$ corresponds to $\theta = \pi/2$. The vertex $(1, 0)$ corresponds to $\theta = 0$. The geodesic distance is: $$ d = \int_0^{\pi/2} d\theta = \frac{\pi}{2}. \quad \square $$ *Interpretation.* One bit of information (distinguishing "heads" from "tails") corresponds to geodesic distance $\pi/2$ in Fisher geometry. This is a derived quantity, not an assumption. ::: :::{prf:lemma} A.6.0e (Curvature Normalization and the Factor of 4) :label: lem-a-curvature-normalization-factor-4 The Poincare disk model with constant sectional curvature $K = -1$ has metric: $$ ds^2 = \frac{4(dx^2 + dy^2)}{(1-|z|^2)^2}. $$ The factor of 4 is uniquely determined by the curvature normalization. *Proof.* For a 2D Riemannian manifold with conformal metric $ds^2 = \lambda(z)(dx^2 + dy^2)$, the Gaussian curvature is {cite}`docarmo1992riemannian`: $$ K = -\frac{1}{2\lambda}\Delta(\log \lambda), $$ where $\Delta = \partial_x^2 + \partial_y^2$ is the flat Laplacian. For $\lambda = c/(1-r^2)^2$ where $r^2 = x^2 + y^2$ and $c > 0$: **Step 1:** Compute $\log \lambda = \log c - 2\log(1-r^2)$. **Step 2:** Compute the Laplacian. Let $f = \log(1-r^2)$. Then: $$ \partial_x f = \frac{-2x}{1-r^2}. $$ Applying the quotient rule to $\partial_x f = -2x \cdot (1-r^2)^{-1}$: $$ \partial_x^2 f = \frac{-2(1-r^2) - (-2x)(-2x)}{(1-r^2)^2} = \frac{-2 + 2r^2 - 4x^2}{(1-r^2)^2}. $$ Similarly for $y$. Adding: $$ \Delta f = \frac{(-2 + 2r^2 - 4x^2) + (-2 + 2r^2 - 4y^2)}{(1-r^2)^2} = \frac{-4 + 4r^2 - 4r^2}{(1-r^2)^2} = \frac{-4}{(1-r^2)^2}. $$ **Step 3:** Therefore $\Delta(\log \lambda) = -2\Delta f = \frac{8}{(1-r^2)^2}$. **Step 4:** The curvature is: $$ K = -\frac{1}{2\lambda} \cdot \frac{8}{(1-r^2)^2} = -\frac{(1-r^2)^2}{2c} \cdot \frac{8}{(1-r^2)^2} = -\frac{4}{c}. $$ **Step 5:** For $K = -1$, we require $c = 4$. $\square$ *Significance.* The choice $K = -1$ is canonical: it sets the "radius of curvature" to unity, making the hyperbolic distance formula $d(0,z) = 2\text{arctanh}|z|$ dimensionless. The factor of 4 in the metric is a *derived consequence* of the curvature normalization, not an assumption. ::: :::{prf:definition} A.6.0f (Fisher-coordinate cell convention) :label: prop-a-area-minimal-distinguishable-cell On a two-dimensional Poincaré chart with curvature normalization $K=-1$, $G(0)=4I$. If the declared coordinate resolution is $\ell_L$, a square coordinate cell of side $\ell_L$ has Riemannian area $$ A_{\mathrm{cell}}=4\ell_L^2. $$ Calling this cell one nat is an operational capacity permit. The Fisher metric and Chentsov's theorem fix the local metric up to scale; they do not, by themselves, identify a finite cell with one nat or prove an area law. This convention is a two-dimensional chart calculation and is not the $(D-1)$-dimensional boundary normalization used for the general operational capacity in {ref}`sec-causal-information-bound`. ::: :::{prf:proposition} A.6.0g (Conditional boundary-channel convention) :label: thm-a-boundary-channel-capacity Suppose a **two-dimensional boundary channel** is explicitly tiled by independent cells of Riemannian area $4\ell_L^2$, and suppose the channel permit assigns one nat to each such cell. Then its declared capacity is $$ C_\partial=\frac{A}{4\ell_L^2}\ \mathrm{nats}. $$ This is a conditional counting statement for a two-dimensional boundary (the boundary dimension is not the same as the $D=2$ Poincaré-disk bulk case). The general $D$-dimensional operational capacity is defined with $\nu_D$ and the dimensionally normalized Levin length in {ref}`sec-causal-information-bound`; no field equation or Fisher theorem supplies the cell-to-nat permit. ::: :::{prf:proposition} A.6.0h (Conditional microstate count) :label: thm-a-microstate-count-area-law Under the channel and achievability permit of Proposition {prf:ref}`thm-a-boundary-channel-capacity`, the number of distinguishable messages is bounded by $$ \Omega\le e^{C_\partial},\qquad \log\Omega\le \frac{A}{4\ell_L^2}. $$ Equality is an additional coding assumption: it requires an achievable independent-cell code and a specified input distribution. The data processing inequality supplies the upper-bound direction, but it does not prove equality for an arbitrary latent model. ::: (sec-appendix-a-holographic-reduction)= ### A.6.1 Conditional bulk-to-boundary route :::{prf:remark} A.6.1 (Conditional bulk-to-boundary permit) :label: lem-a-bulk-to-boundary-conversion A relation of the form $$ I_{\mathrm{bulk}}=\frac1\kappa\oint_{\partial\mathcal Z} \operatorname{Tr}(K)\,dA_G $$ may be adopted as an additional bulk-to-boundary permit for a specified stationary field theory. It is not a consequence of the contracted Bianchi identity: that identity gives a covariantly conserved Einstein tensor, while $\rho_I$ is an independently defined information density. A derivation of this permit must specify the action, source coupling, boundary term, and units. None is assumed by the Metric Law elsewhere in the volume. ::: (sec-appendix-a-saturation-geometry)= ### A.6.2 Formal spherical saturation ansatz For an isotropic manifold with spherical symmetry, we use the ansatz: $$ ds^2 = A(r) \, dr^2 + r^2 \, d\Omega_{n-1}^2, $$ where $d\Omega_{n-1}^2$ is the metric on the unit $(n-1)$-sphere. :::{prf:remark} A.6.2 (Formal spherical saturation ansatz) :label: prop-a-saturation-metric-solution For a selected spherical coordinate model one may study $$ ds^2=A(r)\,dr^2+r^2d\Omega_{n-1}^2, $$ and introduce a Schwarzschild-style denominator $$ A(r)^{-1}=1-\frac{2\mu(r)}{(n-2)r^{n-2}} -\frac{\Lambda_{\mathrm{eff}}r^2}{n(n-1)}. $$ This is an ansatz, not a solution of the capacity-constrained Metric Law. Uniform $T_{ij}=\sigma G_{ij}$ instead gives a constant-curvature source term after the field equation is written out; a mass term and the sign of $\Lambda_{\mathrm{eff}}$ require a separate boundary-value calculation. Consequently the functions $\mu$ and $\Lambda_{\mathrm{eff}}$ below are bookkeeping parameters for the ansatz, not an information mass derived from $\rho_I$. ::: (sec-appendix-a-horizon-condition)= ### A.6.3 Horizon condition for the ansatz :::{prf:definition} A.6.3 (Information Horizon) :label: def-a-information-horizon The **information horizon** $r_h$ is the smallest positive root of: $$ 1 - \frac{2\mu(r_h)}{(n-2)r_h^{n-2}} - \frac{\Lambda_{\text{eff}} r_h^2}{n(n-1)} = 0. $$ At this radius, $A(r_h) \to \infty$ and $G^{rr}(r_h) \to 0$. ::: For $n = 2$ (the Poincare disk case), the formula simplifies. The Poincare metric already encodes the horizon at $|z| = 1$: $$ G_{ij}(z) = \frac{4\delta_{ij}}{(1-|z|^2)^2} \xrightarrow{|z| \to 1} \infty. $$ (sec-appendix-a-fisher-normalization)= ### A.6.4 Fisher normalization and the local factor 4 The local factor $G(0)=4I$ fixes the coordinate-to-Riemannian-area conversion on the normalized Poincaré chart. It does not, by itself, fix a capacity coefficient or a bulk-to-boundary identity. :::{prf:remark} A.6.4a (Scope of the Fisher normalization) :label: rem-a-connection-microstate-counting The simplex distance and curvature normalization below justify the local metric convention used in Proposition {prf:ref}`prop-a-area-minimal-cell`. The statement that a cell carries one nat remains the explicit channel permit of Proposition {prf:ref}`thm-a-boundary-channel-capacity`; it is not a consequence of Chentsov's uniqueness theorem. ::: :::{prf:lemma} A.6.4 (Geodesic Distance on the Probability Simplex) :label: lem-a-geodesic-distance-simplex On the 1-simplex $\Delta^1 = \{(p, 1-p) : p \in [0,1]\}$ with the Fisher Information Metric, the geodesic distance between the uniform distribution $(1/2, 1/2)$ and a vertex $(1, 0)$ is: $$ d_{\text{Fisher}}\left(\frac{1}{2}, 1\right) = \frac{\pi}{2}. $$ *Proof.* See Lemma {prf:ref}`lem-a-geodesic-distance-probability-simplex` for the full derivation. $\square$ ::: :::{prf:proposition} A.6.5 (Poincaré chart area conversion) :label: prop-a-area-minimal-cell On the normalized two-dimensional Poincaré chart, a coordinate cell of side $\ell_L$ has Riemannian area $4\ell_L^2$ at the origin. This is the local geometric conversion used by the conditional cell-counting convention; it is not a statement about the entropy of an arbitrary data distribution. *Proof.* $G(0)=4I$, so $\sqrt{\det G(0)}=4$. $\square$ ::: (sec-appendix-a-assembly)= ### A.6.5 Assembly of the operational normalization :::{prf:definition} A.6.6 (Operational area-law normalization) :label: thm-a-complete-derivation-area-law For a declared latent dimension $D$, boundary measure, coefficient $\nu_D$, and Levin length $\ell_L$, define the operational capacity $$ I_{\max}:=\nu_D\, \frac{\operatorname{Area}(\partial\mathcal Z)}{\ell_L^{D-1}}. $$ This is the normalization used by the Causal Information Capacity in {ref}`sec-causal-information-bound`. A field-theoretic derivation would need, as separate hypotheses, a valid bulk-to-boundary identity, a solution of the chosen metric equation, a boundary extrinsic-curvature estimate, and a dimensionally consistent coupling. The former A.6.1--A.6.3 argument does not establish those hypotheses, so this definition carries no claim that the Metric Law generates the area law. ::: :::{prf:corollary} A.6.7 (Dimension-Dependent Coefficient) :label: cor-a-dimension-dependent-coefficient Under the operational capacity convention, a $D$-dimensional latent manifold with $(D-1)$-sphere boundary uses the dimension-dependent normalization: $$ I_{\max}(D) = \nu_D \cdot \frac{\text{Area}(\partial\mathcal{Z})}{\ell_L^{D-1}}, $$ where the Holographic Coefficient $\nu_D$ (Definition {prf:ref}`def-holographic-coefficient`) is: $$ \nu_D = \frac{(D-1)\Omega_{D-1}}{8\pi} = \frac{(D-1)\pi^{(D-2)/2}}{4\,\Gamma(D/2)}, $$ with $\Omega_{D-1} = 2\pi^{D/2}/\Gamma(D/2)$ the surface area of the unit $(D-1)$-sphere. **Explicit values:** | $D$ | $\Omega_{D-1}$ | $\nu_D$ | Numerical | |-----|----------------|------------|-----------| | 2 | $2\pi$ | $1/4$ | 0.250 | | 3 | $4\pi$ | $1$ | 1.000 | | 4 | $2\pi^2$ | $3\pi/4$ | 2.356 | | 5 | $8\pi^2/3$ | $4\pi/3$ | 4.189 | | 6 | $\pi^3$ | $5\pi^2/8$ | 6.169 | *Remark.* The coefficient $\nu_D$ is **not monotonic** in $D$: it increases from $D=2$ to a peak at $D \approx 9$ ($\nu_9 \approx 9.4$), then decreases toward zero. For typical latent dimensions ($3 \le D \le 20$), $\nu_D > \nu_2 = 1/4$, so using the 2D coefficient **underestimates** capacity. For very high dimensions ($D \gtrsim 22$), $\nu_D < 1/4$, so the 2D coefficient **overestimates** capacity—this is the dangerous case (false safety). Implementers should always use the dimension-appropriate coefficient. ::: :::{warning} :name: warning-dimension-dependent-node-56 **Implementation Note for Node 56 (CapacityHorizonCheck):** The saturation ratio $\eta_{\text{Sch}} = I_{\text{bulk}} / I_{\max}$ depends on the Holographic Coefficient $\nu_D$ (Definition {prf:ref}`def-holographic-coefficient`): $$ I_{\max}(D) = \nu_D \cdot \frac{\text{Area}(\partial\mathcal{Z})}{\ell_L^{D-1}}. $$ The default $\nu_2 = 1/4$ assumes a 2-dimensional latent manifold (Poincare disk). For $D$-dimensional latent spaces, use the appropriate $\nu_D$ from Corollary {prf:ref}`cor-a-dimension-dependent-coefficient`. Using the wrong coefficient leads to: - **$\nu > \nu_D$ (typical for $D > 21$):** **Dangerous.** False safety—agent enters super-saturated regime undetected. - **$\nu < \nu_D$ (typical for $3 \le D \le 20$):** Conservative—unnecessary fusion triggered, but safe. **Implementation Code:** ```python def holographic_coefficient(D: int) -> float: """Compute nu_D = (D-1) * Omega_{D-1} / (8 * pi)""" import math if D < 2: return 0.0 omega = 2 * (math.pi ** (D / 2)) / math.gamma(D / 2) return (D - 1) * omega / (8 * math.pi) ``` ::: :::{prf:remark} A.6.8 (Scope of curvature identities) :label: rem-a-gauss-bonnet-generalization The contracted Bianchi identity states $\nabla^iG_{ij}=0$; it is not the boundary-divergence identity $\int R\,d\mu_G=2\oint\operatorname{Tr}(K)\,dA_G$. The latter is not valid for a general manifold or dimension without additional curvature terms, boundary terms, and field equations. Classical Gauss--Bonnet identities have their own dimension and topology hypotheses. Therefore no such identity is used to prove the operational capacity in this volume. ::: :::{prf:remark} A.6.9 (Status of the area-law arguments) :label: rem-a-non-circularity The local Fisher/Poincaré calculation and the independent channel permit are useful normalization checks. They do not derive a universal area law: the cell-to-nat assignment is a coding assumption, and the former field-theoretic route requires additional lemmas that are not established here. The main text consequently presents $I_{\max}=\nu_D\operatorname{Area}(\partial\mathcal Z)/\ell_L^{D-1}$ as an operational capacity convention. Any comparison with the Bekenstein--Hawking formula is an explicitly labelled mathematical analogy, not a physical identification. ::: (sec-appendix-a-remark-bekenstein-hawking)= ### A.6.6 Remark: Connection to Bekenstein-Hawking The expression $A/(4\ell_P^2)$ is a physical result of gravitational thermodynamics. The agent quantity $\nu_D\operatorname{Area}(\partial\mathcal Z)/\ell_L^{D-1}$ is an operational normalization for a declared interface channel. Their similar form supports a mathematical analogy only; no field equation, entropy identification, or physical equivalence follows from the notation.