Optimization: Thermodynamic Governor and Stable Updates#
TLDR#
We formalize the optimizer as a controlled dynamical system with a Lyapunov objective, consistent with the Universal Governor (Theory of Meta-Stability: The Universal Governor as Homeostatic Controller).
We derive trust-region (“Mach limit”) clipping, alignment damping, SNR gating, varentropy cooling, and thermal conduction from explicit assumptions and show the resulting descent or stability guarantees.
Guarantees are conditional: they hold under stated smoothness, bounded-noise, and metric-approximation assumptions. No claim is made about global optimality.
Roadmap#
Formal setup and assumptions.
Preconditioned descent and trust-region (Mach limit) guarantee.
Varentropy brake as rigorous annealing control.
Alignment damping and oscillation suppression.
SNR gating for stochastic stability.
Thermal conduction across param-group learning rates.
Summary of guarantees and implementation surrogates.
1. Formal Setup: Controlled Optimization#
We optimize a Lyapunov objective that combines task loss with constraint penalties as in the Governor framework (Theory of Meta-Stability: The Universal Governor as Homeostatic Controller). Let
and define the Lyapunov objective
We write the controlled update as a preconditioned step
where \(M_t\) is a symmetric positive definite (SPD) preconditioner. In practice, \(M_t\) is a bounded surrogate that implements the trust-region regulation implied by the state-space metric \(G\) (see Second-Order Sensitivity: Value Defines a Local Metric), without equating parameter-space and state-space geometry. Its surrogate construction is justified by the approximation doctrine (Infeasible Implementation Replacements).
We treat the learning rate as a temperature-scaled control input,
where \(\tau_t\) is a dimensionless temperature scale driven by the Governor (Theory of Meta-Stability: The Universal Governor as Homeostatic Controller), and \(\eta_0\) is the base step size.
Assumptions (Explicit)#
We state the assumptions required for the guarantees in this chapter.
A1 (Smoothness). \(\mathcal{V}\) is \(L\)-smooth: for all \(x,y\), \(\mathcal{V}(y) \le \mathcal{V}(x) + \nabla\mathcal{V}(x)^\top (y-x) + \frac{L}{2}\|y-x\|^2\).
A2 (Preconditioner bounds). \(M_t\) is SPD with eigenvalues in \([m_{\min}, m_{\max}]\) for all \(t\), with \(0 < m_{\min} \le m_{\max} < \infty\). For the stochastic gate, \(M_t\) is measurable with respect to the past information \(\sigma(\theta_t,\hat g_{t-1},\hat g_{t-2},\ldots)\) and is therefore independent of the current noise conditional on \(\theta_t\).
A3 (Stochastic gradients). The observed gradient is \(\hat g_t = g_t + \xi_t\) with \(\mathbb{E}[\xi_t \mid \theta_t] = 0\) and \(\mathbb{E}[\|\xi_t\|^2 \mid \theta_t] \le \sigma^2\).
A4 (Temperature mapping). The temperature scale \(\tau_t\) controls the effective step size, and changes in \(\tau_t\) are slow relative to parameter updates (timescale separation).
A5 (Conduction ordering). Parameter groups are ordered along the data-flow chain of a single physical subsystem (Encoder \(\to\) Latent Dynamics \(\to\) Decoder). Cross-system conduction is disallowed.
These are standard in stability analyses and make the guarantees precise.
2. Preconditioned Descent and Trust-Region (Mach Limit)#
Definition 30 (Preconditioned Update)
The preconditioned update is
with \(M_t\) SPD and \(g_t = \nabla\mathcal{V}(\theta_t)\).
Theorem 3 (Preconditioned Descent (Sufficient Condition))
Under A1–A2, if
then the update in Definition Definition 30 satisfies
In particular, \(\mathcal{V}\) decreases whenever \(g_t \ne 0\). If \(\eta_t = 2 m_{\min} / (L m_{\max}^2)\), the right-hand side yields nonincrease.
Proof
By \(L\)-smoothness (A1), for \(d_t := \eta_t M_t g_t\),
Since \(M_t\) is SPD with eigenvalues in \([m_{\min}, m_{\max}]\) (A2), we have \(g_t^\top M_t g_t \ge m_{\min}\|g_t\|^2\) and \(\|M_t g_t\|^2 \le m_{\max}^2\|g_t\|^2\). Substituting yields
The right-hand side is strictly smaller than \(\mathcal{V}(\theta_t)\) whenever \(\eta_t < 2 m_{\min} / (L m_{\max}^2)\) and yields nonincrease at equality. \(\square\)
Definition 31 (Relative Trust Region (Mach Limit))
A relative trust region is the constraint
with \(\kappa \in (0,1)\) and \(\epsilon_\theta \ge 0\). The case \(\epsilon_\theta = 0\) recovers the strict relative form, while \(\epsilon_\theta > 0\) avoids degeneracy at \(\|\theta_t\| = 0\).
Lemma 3 (Trust-Region Scaling Preserves Descent)
Let \(d_t = \eta_t M_t g_t\) with \(\eta_t\) satisfying Theorem Theorem 3. Define the scaled step
Then \(\mathcal{V}(\theta_t - \tilde d_t) \le \mathcal{V}(\theta_t)\).
Proof
If \(d_t = 0\), then \(\theta_{t+1} = \theta_t\) and the inequality holds trivially. Otherwise, for \(s \in (0,1]\), \(\mathcal{V}(\theta_t - s d_t)\) satisfies the smoothness bound
The right-hand side minus \(\mathcal{V}(\theta_t)\) equals \(s\,(-g_t^\top d_t+\tfrac{L}{2}s\|d_t\|^2)\). The bracket is nondecreasing in \(s\) and is nonpositive at \(s=1\) by Theorem Theorem 3, hence nonpositive for all \(s\in[0,1]\). \(\square\)
Interpretation. The Mach limit is a trust-region enforcement consistent with the Zeno/step-size constraint (Using Scaling Exponents to Gate Updates and Tune Step Sizes). It is rigorous provided the descent condition holds and the scaling is applied by \(s \le 1\).
3. Varentropy Brake: Rigorous Annealing Control#
The Varentropy Brake is stated and proved in Corollary 12 and E.10 Proof of Corollary cor-varentropy-brake. We restate a discrete-time version.
Proposition 6 (Discrete Varentropy Brake)
Let \(T_t > 0\) be the cognitive temperature and \(V_H(\theta_t)\) the varentropy. Define
with \(\eta_T \in (0,1)\) and \(\gamma > 0\). If a relaxation constant \(C>0\) is available and \(\eta_T\le 2C\sqrt{\gamma}\), then
\(0 < T_{t+1} \le T_t\) (temperature is positive and nonincreasing), and
\(|T_{t+1} - T_t| \le C T_t / \sqrt{V_H(\theta_t)}\) whenever \(V_H(\theta_t)>0\) (cooling is slowed when \(V_H\) is large).
Proof
Since \(\eta_T \in (0,1)\) and \(V_H \ge 0\), the multiplier lies in \((0,1]\), proving positivity and monotone nonincrease. The update gives the first bound, and the condition \(\eta_T\le2C\sqrt\gamma\) together with \(1+\gamma V_H\ge2\sqrt{\gamma V_H}\) gives the second. \(\square\)
Implementation note. When we identify \(\eta_t = \eta_0 \tau_t\) with \(\tau_t \propto T_t\), this gives an admissible annealing schedule when the displayed \(C\)-bound holds and A4’s slow-change condition has been checked. Governor-initiated reheating is outside this monotone-brake proposition.
4. Alignment Damping: Oscillation Suppression#
Alignment damping is a step-size shrinkage rule. It preserves the local descent bound when the underlying update is already a descent step; it does not by itself prove a momentum-oscillation bound.
Definition 32 (Gradient-Momentum Alignment)
Let \(g_t\) be the gradient and \(m_t\) a momentum estimate. Define the alignment score
Proposition 7 (Alignment-Triggered Step Damping)
Let \(a_t := g_t^\top m_t\) and fix a damping factor \(\rho \in (0,1]\). Define
If \(\eta_t\) satisfies Theorem Theorem 3, then the damped step with \(\eta_t^{+}\) also decreases \(\mathcal{V}\).
Proof
The update with \(\eta_t^{+}\) is a scaled version of the step with \(\eta_t\), with scaling factor in \((0,1]\). By the same smoothness inequality used in the proof of Theorem Theorem 3, any reduction in step size preserves nonincreasing \(\mathcal{V}\) as long as the original step was in the descent regime. Equivalently, this is a special case of the scaling argument in Lemma Lemma 3. \(\square\)
Interpretation. When the alignment trigger fires, reducing \(\eta_t\) preserves the descent bound of the underlying gradient step. A separate momentum analysis is required to claim oscillation suppression.
5. SNR Gate: Stochastic Stability#
We derive a conservative learning-rate gate from the stochastic update model (A3), where the update uses \(\hat g_t\) instead of \(g_t\).
Proposition 8 (SNR-Gated Step Size)
Under A1–A3 and the stochastic update \(\theta_{t+1} = \theta_t - \eta_t M_t \hat g_t\), assume \(g_t \ne 0\). The expected Lyapunov change satisfies
Consequently, a sufficient condition for expected descent is
with \(\mathrm{SNR} := \|g_t\|^2 / \sigma^2\). If \(g_t = 0\), the sufficient bound reduces to \(\eta_t = 0\).
Proof
Apply the smoothness bound with \(d_t = \eta_t M_t \hat g_t\) and take conditional expectations. The measurability clause in A2 makes \(M_t\) independent of the current noise, so \(\mathbb{E}[g_t^\top M_t\xi_t\mid\theta_t]=0\). Use \(\mathbb{E}[\hat g_t] = g_t\) and \(\mathbb{E}[\|\hat g_t\|^2] = \|g_t\|^2 + \mathbb{E}[\|\xi_t\|^2] \le \|g_t\|^2 + \sigma^2\). Then apply the eigenvalue bounds from A2 and solve for \(\eta_t\) to make the coefficient negative. \(\square\)
Interpretation. When SNR is low, the gate shrinks \(\eta_t\) to protect expected descent. This is a rigorous stochastic analogue of the “cool when signal is weak” rule.
6. Thermal Conduction Across Param Groups#
Thermal conduction smooths learning rates across adjacent blocks inside a single physical system (A5). We implement conduction in log space to preserve positivity.
Definition 33 (Log-LR Conduction Update)
Let \(\eta_i > 0\) be per-group learning rates and \(x_i := \log(\eta_i)\). Define
with Neumann boundary conditions \(x_0 = x_1\), \(x_{n+1} = x_n\) and conductivity \(k \in [0,1]\).
Proposition 9 (Conduction Contracts LR Disparities)
Let \(L\) be the path-graph Laplacian on \(n\) groups and define the energy
The update in Definition Definition 33 is gradient descent on \(E\) with step size \(k/2\), and for \(k \in [0,1]\) it satisfies \(E(x^{+}) \le E(x)\).
Proof
We have \(\nabla E(x) = L x\) for the path graph. The update is \(x^{+} = x - (k/2) L x\), i.e. gradient descent with step size \(k/2\). Because \(L\) is symmetric positive semidefinite with eigenvalues bounded by \(\lambda_{\max} \le 4\) for a path graph, step size \(k/2 \le 1/2\) ensures \((1 - (k/2)\lambda)^2 \le 1\) for all eigenvalues. Therefore \(E(x^{+}) = \tfrac{1}{2} x^\top (I - (k/2)L)^\top L (I - (k/2)L)x \le E(x)\). \(\square\)
Interpretation. Conduction is a provably contractive smoothing of log learning rates. It enforces coherence without mixing separate physical subsystems.
7. Combined Guarantees (Under Assumptions)#
Theorem 4 (Thermodynamic Governor Stability (Conditional))
Under A1–A5, with updates that apply in the order: conduction on log learning rates, SNR/alignment gates, the preconditioned step, and trust-region clipping last, and with \(M_t\) block-diagonal across parameter groups, assume:
preconditioned descent (Theorem Theorem 3),
trust-region scaling (Lemma Lemma 3),
alignment-triggered step damping (Proposition Proposition 7),
varentropy brake (Proposition Proposition 6),
SNR gating (Proposition Proposition 8), and
log-LR conduction (Proposition Proposition 9),
the optimizer produces a nonincreasing Lyapunov objective in the deterministic case. The expected-descent claim under noise applies only to the unscaled stochastic step in Proposition Proposition 8; trust-region clipping and alignment damping are covered there only in the deterministic regime. Learning rates remain positive and coherent across adjacent groups, and the temperature schedule obeys the adiabatic constraint during annealing phases with no Governor-initiated heating.
Remark 3
These guarantees are conditional and local. They ensure stability and controlled descent, not global optimality. Violations of A1–A5 (e.g., non-smooth losses, unbounded noise, misordered groups) void the guarantees.
8. Implementation Surrogates (Rigorous, but Practical)#
The theory uses metric-aware updates and diagnostic signals. Practical implementations may use the following surrogates, which are explicitly permitted in Infeasible Implementation Replacements:
Use a lagged Adam second-moment statistic \(M_t=\operatorname{diag}(1/(\sqrt{\hat v_{t-1}}+\epsilon))\), clipped to \([m_{\min},m_{\max}]\), as a diagonal proxy for the trust-region regulator. The usual current-step Adam direction lies outside the measurability class analysed by Proposition Proposition 8.
Estimate \(\mathrm{SNR}\) with running averages of gradient norms.
Replace exact varentropy with a windowed variance of the chosen metric signal (loss or grad RMS).
Implement conduction on param groups ordered by data-flow to satisfy A5.
Each surrogate preserves the theoretical intent: bounded step sizes, adiabatic annealing, step-size damping, noise-aware cooling, and cross-layer coherence.
9. Summary Checklist#
Verify A1–A5 for your setting (or document violations).
Enforce trust-region scaling when relative step sizes exceed \(\kappa\).
Apply varentropy brake to \(T_c\) (and thus \(\eta_t\)) for safe annealing.
Use alignment-triggered step damping as a descent-preserving shrinkage rule; verify any momentum claim separately.
Gate \(\eta_t\) by SNR to maintain expected descent.
Apply log-LR conduction only within a single physical subsystem.
This chapter connects the Governor theory to local optimizer guarantees under the stated assumptions and scope.