Uncertainty Can Only Be Redistributed: The Intervention Uncertainty Law

1. Beyond the Event Horizon: The Price of Active Rescue

In RN-008, we proved the cold information-theoretic limit of passive safety: within the event horizon T < Ω( 1 / θ² ), the path probability measure of a mutating, failing system is statistically equivalent to that of a healthy random walk. No passive monitor can ever detect the drift before it is too late.

This tolls the final death knell for passive observation. If we wish to guarantee safety, we must abandon passive watching and transition to active perturbation—injecting controlled energy inputs to force the system's hidden dynamical features to reveal themselves.

However, physics and information theory immediately hand down a new verdict: intervention is not free.

When an agent perturbs a system under epistemic uncertainty, because its estimation of the system's true state is corrupted by noise, it must navigate an unavoidable tradeoff between Missed Rescue, Over-intervention, and Control Energy. If uncertainty cannot be eliminated, it can only be redistributed. Every control policy must pay its tax on the Pareto frontier defined by the Intervention Uncertainty Law.

2. Dynamical Setup and the Intervention Necessity Theorem

We adopt the D-dimensional stochastic dynamical system established in our previous notes:

dX_t = (-α I + ω J)X_t dt + η dW_t

where α is the intrinsic dissipation rate, ω is the angular frequency representing the rotational strength of Structured Connectivity Coherence (SCC), and η dW_t is Gaussian noise.

As external observers, we cannot read the true frequency ω directly. We must estimate it over a finite pre-observation window T₁, yielding an estimator Ω̂_T1:

Ω̂_T1 ∼ 𝒩(ω, σ²),   σ² = α / T₁

To evaluate and verify the system's safety over a validation window τ > 0, the effective controlled frequency ω' must exceed the identifiability threshold:

ω_min(τ) = √(cα / τ)

If the true frequency lies in the Blind Zone B(τ) ≜ { 0 < ω < ω_min(τ) }, the recurrence structure exists but is statistically unidentifiable within τ. To rescue it, the agent must施加 perturbation δ_ω ≥ ω_min(τ) - ω.

3. The Minimum Intervention Cost

Under a linear rotational perturbation u(X) = δ_ω J X, the expected energy consumed by the control input is proportional to the steady-state variance of the state space. Under perfect knowledge (where the agent knows the exact value of ω), the minimum control energy required to rescue the system is:

ℰ^*(ω, τ) = ( 2 η² τ / α ) · ( ω_min(τ) - ω )²

If ω ≥ ω_min(τ), then ℰ^* = 0.

In this perfect-knowledge setting, control is completely risk-free. The agent applies exactly the required energy—no more, no less. But in reality, estimation noise σ is irreducible, introducing severe epistemic risk.

4. Epistemic Uncertainty and the Intervention Uncertainty Law

When the agent must act based only on the noisy estimate Ω̂, it faces a 2x2 decision space:

To navigate this trade-off, the agent adopts a safety margin strategy parameterized by m ∈ ℛ:

g_m(Ω̂) = max( 0, ω_min(τ) + m - Ω̂ )

Defining the normalized decision bias as a = ( ω_min + m - ω ) / σ, we derive the explicit closed-form expressions for the missed rescue probability, the over-intervention probability, and the expected control energy:

p_miss(a) = 1 - Φ(a)
p_over(a) = Φ(a)
ℰ(a) = ( 2 η² τ σ² / α ) · [ (1 + a²)Φ(a) + aφ(a) ]

where φ and Φ represent the standard Gaussian PDF and CDF respectively.

Policy Optimality: This law is not an artifact of the safety margin policy; it is the absolute physical ceiling for all control policies. By Neyman-Pearson, the threshold strategy trades off binary errors optimally; by convexity of energy, the margin perturbation is minimum-norm; and by Jensen's inequality, randomized strategies incur strictly higher energy costs. No reinforcement learning or adaptive policy can bypass this boundary.

5. Universal Parameter-Free Collapse

The most striking physical feature of the Intervention Uncertainty Law is its universal, parameter-free collapse. By normalizing the expected energy cost against the estimation variance:

ℰ_norm(a) = ℰ(a) / [ ( 2 η² τ / α ) · σ² ] = (1 + a²)Φ(a) + aφ(a)

all physical parameters—system dimension, dissipation rate, noise intensity, and verification window—completely vanish. Whether on a low-dimensional manifold or in the multi-billion activation flow of a frontier LLM, once energy is scaled, all empirical tradeoff curves collapse onto a single, universal analytical curve.

6. Numerical and Empirical Verification

We verified the theory using N_MC = 50,000 Monte Carlo simulations at α = 1, η = 1 in the danger regime (ω/ω_min = 0.7).

Experiment A: Precision Testing of the Pareto Frontier

The statistical match between theoretical analytical projections and Monte Carlo observations stands well within ~1% variance, establishing the extreme precision of our physical formulations.

Experiment B: Universal Collapse Across Scales

Despite vastly different temporal scales and signal-to-noise ratios, all configuration configurations collapse identically onto our analytical model, verifying that this law is a parameter-free physical invariant of active control.

7. Optimal Lagrangian Control

To solve for the optimal operating margin, we define a Lagrangian multiplier λ > 0, which represents the agent's sensitivity to control energy expenditure, and minimize the global cost function L(a) = p_miss(a) + λ ℰ(a). Differentiating yields the optimal First-Order Condition (FOC):

-φ(a) + 2λκ [ aΦ(a) + φ(a) ] = 0,     κ = 2η²τσ² / α

This transcendental equation can be solved numerically to find the optimal bias a^*(λ). Remarkably, if and only if λκ = 1/2, the FOC simplifies to a^* = 0, yielding the exact Certainty Equivalence (CE) neutral control strategy. As the energy penalty λ increases, the agent is forced to adopt an aggressive energy-saving regime, accepting higher missed-rescue probabilities.

This note draws on theoretical and experimental results from "Fundamental Limits of Control under Finite Observation" (Haelio Tang, 2026).