Phase-Targeted Energy Injection in Coupled Chaotic Pendulum Systems for Inertial Loading Generation

# Phase-Targeted Energy Injection in Coupled Chaotic Pendulum Systems for Inertial Loading Generation: A Fractal Correction Engine Approach

**Adam L McEvoy**

---

## Abstract

I present a computational investigation into whether phase-targeted energy injection at separatrix crossings in coupled chaotic pendulum systems can produce sustained inertial loading on a reaction platform. The system couples double, triple, and Duffing pendulum oscillators to a free-sliding platform via Lagrangian mechanics, where joint motor torques serve as internal forces that generate platform acceleration through reaction dynamics. I introduce the Fractal Correction Engine (FCE), a nonlinear controller that estimates the maximal Lyapunov exponent in real time, identifies separatrix crossings in phase space, and injects energy at dynamically favorable moments. Across a 420-run experiment (7 controllers, 3 models, 20 seeds, 1000 s each, 5000 J energy budget), FCE achieves 55.2% maintenance within $\pm$50% of 1$g$ on the double pendulum using only 892 J (18% of budget), compared to PID control achieving 60.2% while consuming the full 5000 J. This yields a loading efficiency of 7.44 s/J for FCE versus 0.26 s/J for PID, a 29-fold improvement. Phase-ablation experiments confirm that FCE's advantage is causally attributable to injection timing: shuffling kick times while preserving total energy and kick count reduces maintenance from 44.5% to 24.5% ($p < 0.001$, Welch $t$-test). An enhanced multi-plane FCE with per-joint Lyapunov estimation and PCA mode decomposition improves triple-pendulum maintenance from 8.6% to 25.5%, a 196% gain. I characterize the Duffing oscillator controllability boundary as a function of mass ratio and driving amplitude. The system produces oscillatory bidirectional loading rather than sustained unidirectional artificial gravity; I discuss the implications for microgravity countermeasure design and the distinction between inertial loading and artificial gravity generation.

---

## 1. Introduction

### 1.1 The Microgravity Deconditioning Problem

Prolonged exposure to microgravity causes progressive musculoskeletal deconditioning, including bone mineral density loss of 1--2% per month in weight-bearing bones (Lang et al., 2004), muscle atrophy of 1--3% per week in antigravity muscles (Fitts et al., 2010), and cardiovascular deconditioning (Hughson et al., 2018). Current countermeasures aboard the International Space Station include resistive exercise devices (ARED), treadmills with harness systems, and cycle ergometers. These require 2--2.5 hours of daily exercise and only partially mitigate deconditioning (Smith et al., 2012).

Artificial gravity via centrifugal rotation remains the theoretically ideal solution, but imposes severe engineering constraints: large-radius rotating structures (>10 m to avoid Coriolis sickness), continuous rotation requiring structural mass and power, and complex docking/EVA interfaces. Short-radius centrifuges have shown promise but produce uncomfortable gravity gradients and require dedicated crew time (Clément et al., 2015).

### 1.2 The Chaotic Pendulum Hypothesis

I investigate an alternative approach: using chaotic pendulum oscillators coupled to a platform to generate oscillatory inertial loading through reaction dynamics. The core premise is that chaotic systems exhibit broadband frequency content and aperiodic trajectories that may produce richer mechanical loading spectra than periodic vibration systems. By controlling energy injection timing to coincide with dynamically favorable phase-space configurations, I hypothesize that useful loading can be achieved with minimal energy expenditure.

This approach differs fundamentally from artificial gravity generation. Newton's third law constrains internal forces: no internal mechanism can produce net sustained acceleration of a system's center of mass. What the pendulum system produces is oscillatory reaction loading, where the platform acceleration alternates direction as the pendulum masses swing. The magnitude of instantaneous platform acceleration can transiently reach or exceed 1$g$, but the time-averaged acceleration is bounded by the energy input and system mass ratios.

### 1.3 The Fractal Correction Engine

The Fractal Correction Engine (FCE) is a nonlinear control algorithm designed to exploit the sensitive dependence on initial conditions characteristic of chaotic systems. Rather than fighting chaos with feedback linearization or model-predictive control, FCE works with the chaotic dynamics by identifying moments when small energy inputs produce large trajectory divergences, specifically at separatrix crossings in phase space where the system transitions between qualitatively different dynamical regimes.

The key insight is that energy injected at a separatrix crossing is amplified by the system's own Lyapunov instability, while energy injected during regular or stable phases is largely dissipated. By estimating the maximal Lyapunov exponent in real time and detecting separatrix crossings through phase-space analysis, FCE achieves loading comparable to conventional controllers while using a fraction of the energy.

### 1.4 Contributions

This work makes the following contributions:

1. A Lagrangian formulation coupling double, triple, and Duffing pendulum oscillators to a reaction platform with active joint torque injection and energy budget enforcement.
2. The FCE algorithm for real-time phase-targeted energy injection in chaotic mechanical systems.
3. A 420-run comparative experiment demonstrating 29-fold energy efficiency advantage of FCE over PID control.
4. Phase-ablation experiments establishing causal attribution of FCE's advantage to injection timing.
5. An enhanced multi-plane FCE for higher-DOF systems using per-joint Lyapunov estimation and PCA mode decomposition.
6. Characterization of the Duffing oscillator controllability boundary as a function of mass ratio and driving parameters.

---

## 2. System Description and Physical Model

### 2.1 Platform-Pendulum Coupling

The system consists of a platform of mass $M_p$ constrained to translate along a single axis (the "gravity" direction), coupled to one or more pendulum oscillators through pivot joints. The platform slides on a frictionless rail, and all forces arise from internal dynamics: pendulum gravity, centrifugal/Coriolis reactions, joint motor torques, and generator damping.

The generalized coordinates are:

$$\mathbf{q} = \begin{bmatrix} X \\ \theta_1 \\ \theta_2 \\ \vdots \end{bmatrix}$$

where $X$ is the platform displacement and $\theta_i$ are the pendulum joint angles measured from the vertical.

The equations of motion follow from the Lagrangian $\mathcal{L} = T - V$, where the kinetic and potential energies include the coupling between platform translation and pendulum rotation.

### 2.2 Double Pendulum Model

For the double pendulum coupled to a translating platform, the state vector is:

$$\mathbf{y} = [X, \dot{X}, \theta_1, \omega_1, \theta_2, \omega_2, E_{\text{store}}]$$

where $\omega_i = \dot{\theta}_i$ are joint angular velocities and $E_{\text{store}}$ tracks onboard energy storage.

The system parameters are:

| Parameter | Symbol | Value |
|-----------|--------|-------|
| Platform mass | $M_p$ | 10.0 kg |
| First bob mass | $m_1$ | 1.0 kg |
| Second bob mass | $m_2$ | 1.618 kg |
| First arm length | $\ell_1$ | 0.618 m |
| Second arm length | $\ell_2$ | 1.0 m |
| Gravitational acceleration | $g$ | 9.81 m/s$^2$ |

The bob masses and arm lengths incorporate golden ratio proportions ($\phi = 1.618...$) to avoid low-order resonances that would suppress chaotic behavior.

The $3 \times 3$ mass matrix for the generalized coordinates $[X, \theta_1, \theta_2]$ is:

$$\mathbf{M} = \begin{bmatrix} M_p + m_1 + m_2 & (m_1 + m_2)\ell_1 \cos\theta_1 & m_2 \ell_2 \cos\theta_2 \\ (m_1 + m_2)\ell_1 \cos\theta_1 & (m_1 + m_2)\ell_1^2 & m_2 \ell_1 \ell_2 \cos(\theta_1 - \theta_2) \\ m_2 \ell_2 \cos\theta_2 & m_2 \ell_1 \ell_2 \cos(\theta_1 - \theta_2) & m_2 \ell_2^2 \end{bmatrix}$$

The generalized force vector includes gravity, Coriolis/centrifugal terms, bleed damping, and injection torques:

$$\mathbf{F} = \begin{bmatrix} F_{\text{ext}} + (m_1 + m_2)\ell_1 \omega_1^2 \sin\theta_1 + m_2 \ell_2 \omega_2^2 \sin\theta_2 \\ -(m_1 + m_2)g\ell_1 \sin\theta_1 - m_2 \ell_1 \ell_2 \omega_2^2 \sin(\theta_1 - \theta_2) - k_b \omega_1 + \tau_1 \\ -m_2 g \ell_2 \sin\theta_2 + m_2 \ell_1 \ell_2 \omega_1^2 \sin(\theta_1 - \theta_2) - k_b \omega_2 + \tau_2 \end{bmatrix}$$

where $\tau_i$ are the joint injection torques, $k_b$ is the generator bleed damping coefficient, and $F_{\text{ext}}$ is any external platform force (set to zero in v3).

The accelerations are obtained by solving:

$$\mathbf{M} \ddot{\mathbf{q}} = \mathbf{F}$$

The platform acceleration $\ddot{X}$ emerges from the first row of this system, encoding all pendulum reaction forces. Note that $\tau_i$ are joint torques (internal forces between pendulum links), not direct forces on the platform. Platform acceleration arises purely from reaction dynamics.

### 2.3 Triple Pendulum Model

The triple pendulum extends the formulation to a $4 \times 4$ mass matrix with state vector:

$$\mathbf{y} = [X, \dot{X}, \theta_1, \omega_1, \theta_2, \omega_2, \theta_3, \omega_3, E_{\text{store}}]$$

The additional parameters are: third bob mass $m_3 = 1.2$ kg and third arm length $\ell_3 = 0.8$ m. The mass matrix elements follow the same Lagrangian structure, with coupling terms involving all pairwise angle differences $\cos(\theta_i - \theta_j)$.

### 2.4 Duffing Oscillator Model

The Duffing oscillator provides a complementary nonlinear system with a different chaotic mechanism (double-well potential rather than rotational gravity):

$$\mathbf{y} = [X, \dot{X}, x_{\text{rel}}, v_{\text{rel}}, E_{\text{store}}]$$

where $x_{\text{rel}}$ is the relative displacement of the oscillator mass from the platform. The $2 \times 2$ mass matrix is:

$$\mathbf{M} = \begin{bmatrix} M_p + m_d & m_d \\ m_d & m_d \end{bmatrix}$$

The Duffing restoring force is:

$$F_{\text{Duffing}} = -\delta \dot{x}_{\text{rel}} + \alpha(x_{\text{rel}} - x_{\text{rel}}^3) + \Gamma \cos(\Omega t) - k_b \dot{x}_{\text{rel}} + \tau_1$$

with damping $\delta = 0.15$, cubic stiffness coefficient $\alpha = 1.0$, drive amplitude $\Gamma = 0.3$, and drive frequency $\Omega = 1.0$ rad/s. The cubic term $x - x^3$ creates a symmetric double-well potential with equilibria at $x = 0, \pm 1$.

### 2.5 Energy Budget and Power Constraints

To ensure fair comparison across controllers, all simulations share an identical energy budget:

$$E_{\text{budget}} = 5000 \text{ J}, \quad P_{\max} = 500 \text{ W}, \quad \tau_{\max} = 50 \text{ N}\cdot\text{m}$$

The instantaneous injection power is:

$$P_{\text{inject}} = \sum_i |\tau_i \cdot \omega_i|$$

The energy budget enforces:

$$\int_0^T P_{\text{inject}}(t) \, dt \leq E_{\text{budget}}$$

At each timestep, the controller requests injection power $P_{\text{req}}$, which is clamped:

$$P_{\text{actual}} = \min\left(P_{\text{req}}, \, P_{\max}, \, \frac{E_{\text{remaining}}}{\Delta t}\right)$$

A critical design rule: bleed energy does not replenish the injection budget. This prevents gaming where a controller bleeds and re-injects to inflate loading metrics. The total injection energy $E_{\text{used}}$ is strictly monotonically increasing.

### 2.6 Numerical Integration

All simulations use 4th-order Runge-Kutta integration with fixed timestep $\Delta t = 0.01$ s:

$$\mathbf{y}_{n+1} = \mathbf{y}_n + \frac{\Delta t}{6}(\mathbf{k}_1 + 2\mathbf{k}_2 + 2\mathbf{k}_3 + \mathbf{k}_4)$$

where:

$$\mathbf{k}_1 = f(t_n, \mathbf{y}_n), \quad \mathbf{k}_2 = f\!\left(t_n + \tfrac{\Delta t}{2}, \mathbf{y}_n + \tfrac{\Delta t}{2}\mathbf{k}_1\right)$$

$$\mathbf{k}_3 = f\!\left(t_n + \tfrac{\Delta t}{2}, \mathbf{y}_n + \tfrac{\Delta t}{2}\mathbf{k}_2\right), \quad \mathbf{k}_4 = f(t_n + \Delta t, \mathbf{y}_n + \Delta t \cdot \mathbf{k}_3)$$

Initial conditions include small random perturbations ($\pm$5% of nominal angles) seeded per run for statistical independence.

---

## 3. The Fractal Correction Engine (FCE)

### 3.1 Overview

The FCE is a nonlinear controller that exploits the structure of chaotic dynamics rather than attempting to suppress or linearize them. It operates on a three-stage cycle:

1. **Observe**: Read the system state and estimate the maximal Lyapunov exponent from a rolling window of phase-space trajectory data.
2. **Classify**: Map the current state to one of six dynamical regimes based on Lyapunov exponent, kinetic energy, and phase-space geometry.
3. **Act**: Apply injection torques or bleed damping according to the classified regime, with maximum injection at separatrix crossings.

### 3.2 Real-Time Lyapunov Estimation

The FCE estimates the maximal Lyapunov exponent $\lambda_1$ using Rosenstein's algorithm (Rosenstein et al., 1993) applied to a rolling window of 2000 state vectors, updated every 200 timesteps.

The algorithm proceeds as follows. Given the trajectory $\{\mathbf{y}_n\}_{n=1}^{N}$ in the rolling window:

1. Construct a KD-tree from the state vectors for efficient nearest-neighbor lookup.
2. Sample 100 reference points uniformly from the window interior, excluding a temporal separation of at least $\min(20, N/50)$ steps.
3. For each reference point $\mathbf{y}_i$, find its nearest neighbor $\mathbf{y}_j$ with $|i - j| \geq$ minimum separation and initial distance $d_0 = \|\mathbf{y}_i - \mathbf{y}_j\| > 10^{-10}$.
4. Track the divergence $d_k = \|\mathbf{y}_{i+k} - \mathbf{y}_{j+k}\|$ for $k = 1, \ldots, 50$ steps.
5. Record data points $(k\Delta t, \ln(d_k / d_0))$.

The maximal Lyapunov exponent is estimated as the slope of a linear fit to the first 20 data points:

$$\ln\frac{d(t)}{d_0} \approx \lambda_1 \cdot t$$

The exponent is clamped to $\lambda_1 \geq 0$ (negative values indicate stable dynamics).

### 3.3 State Classification

The FCE classifies the current state into six dynamical regimes:

| Class | Condition | Action |
|-------|-----------|--------|
| **Diverging** | $\max|\omega_i| > 50$ or $\text{KE} > 500$ J | Emergency bleed |
| **Kick Window** | $|\omega_{\text{outer}}| < \omega_{\text{thresh}}$ AND $|\theta_{\text{outer}}| > \theta_{\text{thresh}}$ AND $\lambda_1 > \lambda_{\text{mod}}$ | Maximum injection |
| **Chaotic High** | $\lambda_1 > \lambda_{\text{mod}}$ AND $\text{KE} > 20$ J | Coast (no action) |
| **Chaotic Low** | $\lambda_1 > \lambda_{\text{mod}}$ AND $\text{KE} \leq 20$ J | Moderate injection |
| **Stable High** | $\lambda_1 \leq \lambda_{\text{mod}}$ AND $\text{KE} > 30$ J | Bleed |
| **Stable Low** | $\lambda_1 \leq \lambda_{\text{mod}}$ AND $\text{KE} \leq 30$ J | Maintain injection |

The kinetic energy estimate is $\text{KE} = \frac{1}{2}\sum_i \omega_i^2$, using angular velocities as a proxy (neglecting position-dependent mass matrix terms for computational efficiency).

The key classification parameters are:

$$\omega_{\text{thresh}} = 0.5 \text{ rad/s}, \quad \theta_{\text{thresh}} = \frac{\pi}{2} \text{ rad}, \quad \lambda_{\text{mod}} = 0.1 \text{ s}^{-1}$$

### 3.4 The Kick Window: Separatrix Crossing Detection

The **kick window** is the core innovation of FCE. It detects moments when the outermost pendulum arm is near a turning point ($|\omega_{\text{outer}}| < 0.5$ rad/s) at large amplitude ($|\theta_{\text{outer}}| > \pi/2$ rad) during chaotic dynamics ($\lambda_1 > 0.1$).

Physically, this corresponds to a separatrix crossing: the pendulum is momentarily nearly stationary at the top of its swing, in a region of phase space where arbitrarily small perturbations determine whether it continues over the top (libration $\to$ rotation) or falls back. Energy injected here is amplified by the system's intrinsic Lyapunov instability.

The injection torque during a kick window is:

$$\tau_i = \text{sgn}(\omega_i) \cdot \tau_{\max} \cdot f_{\text{inject}}$$

where $f_{\text{inject}} = 1.0$ (kick window), $0.5$ (chaotic low), or $0.2$ (stable low) determines the injection fraction. The torque direction follows the instantaneous angular velocity, adding energy to the existing motion rather than opposing it.

### 3.5 Energy Efficiency Mechanism

FCE achieves superior energy efficiency through three mechanisms:

1. **Temporal selectivity**: FCE injects energy only during dynamically favorable windows, remaining idle during phases where injection would be dissipated or counterproductive. In the full experiment, FCE used only 892 J of its 5000 J budget.

2. **Lyapunov amplification**: Energy injected during high-Lyapunov phases is exponentially amplified by the system's own instability, with nearby trajectories diverging as $\sim e^{\lambda_1 t}$. This means a small energy input at the right time produces a large trajectory change.

3. **Separatrix targeting**: By injecting at separatrix crossings, FCE pushes the pendulum over rotational barriers that gravity alone would not overcome, accessing high-energy orbital families that produce stronger platform reaction forces.

---

## 4. Experimental Design

### 4.1 Controller Comparison

Seven controllers were compared under identical energy budget constraints:

1. **Passive**: No energy injection or bleed. Pure uncontrolled dynamics.
2. **Sinusoidal**: Constant-frequency torque drive at 2 Hz with amplitude calibrated to spread the budget evenly over the simulation.
3. **Random Kicks**: Poisson-distributed impulses (rate 5/s, duration 50 ms) with random torque magnitude and direction. This serves as the null hypothesis for FCE.
4. **PID**: Classical proportional-integral-derivative control on platform acceleration error $(g - |a_{\text{platform}}|)$ with anti-windup. Gains: $K_p = 2.0$, $K_i = 0.5$, $K_d = 0.3$.
5. **Energy Thermostat**: Bang-bang control on pendulum kinetic energy level with hysteresis.
6. **FCE**: The Fractal Correction Engine as described in Section 3.
7. **FCE+PID Hybrid**: FCE timing with PID-scaled injection magnitude.

### 4.2 Experiment Matrix

The full experiment comprises:

- 7 controllers $\times$ 3 models (double, triple, Duffing) $\times$ 20 seeds = **420 runs**
- Each run: $T = 1000$ s, $\Delta t = 0.01$ s (100,000 steps)
- Energy budget: 5000 J per run, max power 500 W
- Initial conditions: nominal angles with $\pm$5% random perturbation per seed

### 4.3 Performance Metrics

#### Maintenance Bands

The primary metric is the fraction of simulation time during which the platform acceleration magnitude $|a_{\text{platform}}|$ falls within a tolerance band around the target $g = 9.81$ m/s$^2$:

$$M_{\alpha} = \frac{1}{T} \int_0^T \mathbb{1}\left[\left|\frac{|a(t)| - g}{g}\right| \leq \alpha \right] dt$$

Five bands are evaluated: $\alpha \in \{0.50, 0.25, 0.10, 0.05, 0.01\}$.

#### Loading Efficiency

Loading efficiency measures useful loading time per unit energy:

$$\eta_L = \frac{M_{0.50} \cdot T}{E_{\text{net}}} \quad [\text{s/J}]$$

where $E_{\text{net}} = E_{\text{injected}} - E_{\text{bled}}$.

#### Comfort and Safety

Jerk (rate of acceleration change) is computed as:

$$j(t) = \frac{|a(t + \Delta t) - a(t)|}{\Delta t}$$

with RMS jerk and peak jerk reported. Lower jerk indicates smoother loading.

### 4.4 Upgrade Experiments

Five additional experiments were conducted to validate and extend the core findings:

**Upgrade 1 (Equal-Energy Comparison)**: All controllers run at identical budgets of 500, 1000, 2000, and 5000 J to eliminate the confound that FCE voluntarily uses less energy.

**Upgrade 2 (Matched-Maintenance Energy Test)**: Binary search for the minimum energy budget each controller requires to achieve 50% $\pm$50% maintenance on the double pendulum.

**Upgrade 3 (Phase-Ablation Test)**: Eight FCE variants systematically disable individual components to establish causal attribution. The killer test: identical kick sequence with shuffled timing.

**Upgrade 4 (Enhanced Triple FCE)**: Per-joint Lyapunov estimation, multi-plane separatrix detection, and PCA mode decomposition for the 3-DOF triple pendulum.

**Upgrade 5 (Duffing Parameter Sweep)**: Parametric sweep over oscillator mass, platform mass, drive amplitude, cubic stiffness, and damping to find controllable Duffing regimes.

---

## 5. The FCE Applied to the System

### 5.1 Controller-Simulation Integration

At each timestep, the simulation loop executes:

1. The controller's `compute_action(t, state, model, budget, dt)` method returns injection torques $\boldsymbol{\tau}$ and bleed coefficient $k_b$.
2. The injection power $P_{\text{inject}} = \sum_i |\tau_i \omega_i|$ is computed and submitted to the energy budget.
3. The budget clamps the power: $P_{\text{actual}} \leq \min(P_{\text{req}}, P_{\max}, E_{\text{rem}}/\Delta t)$.
4. Torques are scaled proportionally: $\tau_i^{\text{actual}} = \tau_i \cdot (P_{\text{actual}} / P_{\text{req}})$.
5. The ODE is integrated with the clamped torques and bleed damping.
6. Platform acceleration is extracted: $a_{\text{platform}} = |\dot{X}_{n+1} - \dot{X}_n| / \Delta t$.
7. Metrics are updated via Welford's online algorithm (streaming mean/variance without storing arrays).

### 5.2 FCE Adaptation to Different Models

The FCE uses the same algorithmic framework across all three models but accesses model-specific state indices:

- **Double pendulum**: Outer arm is joint 2 ($\theta_2, \omega_2$). Lyapunov computed on the 7D state.
- **Triple pendulum**: Outer arm is joint 3 ($\theta_3, \omega_3$). Standard FCE uses only the outermost joint for separatrix detection, which proves insufficient for the 3-DOF dynamics (see Section 5.3).
- **Duffing oscillator**: "Outer arm" maps to the relative displacement ($x_{\text{rel}}, v_{\text{rel}}$). The kick window condition uses velocity threshold instead of angular velocity.

### 5.3 Enhanced Multi-Plane FCE for Triple Pendulum

The standard FCE's reliance on only the outermost joint for separatrix detection is insufficient for the triple pendulum, where energy is distributed across three coupled but semi-independent phase planes. I developed an enhanced FCE with three additional components:

#### Per-Joint Lyapunov Estimation

Rather than computing a single Lyapunov exponent from the full state space, the enhanced FCE estimates $\lambda_1^{(i)}$ independently in each joint's 2D phase plane $(\theta_i, \omega_i)$:

$$\lambda_1^{(i)} = \lim_{t \to \infty} \frac{1}{t} \ln \frac{\|(\delta\theta_i(t), \delta\omega_i(t))\|}{\|(\delta\theta_i(0), \delta\omega_i(0))\|}$$

This identifies which specific joint is exhibiting chaotic dynamics and therefore which joint is most amenable to separatrix-targeted injection.

#### Multi-Plane Separatrix Detection

Each joint is independently checked for the kick window condition:

$$\text{kick\_ready}(i) = \left(|\omega_i| < \omega_{\text{thresh}}\right) \wedge \left(|\theta_i| > \theta_{\text{thresh}}\right) \wedge \left(\lambda_1^{(i)} > \lambda_{\text{mod}}\right)$$

Multiple joints may simultaneously be kick-ready, enabling coordinated multi-joint injection.

#### PCA Mode Decomposition

A rolling window of phase coordinates $[\theta_1, \omega_1, \theta_2, \omega_2, \theta_3, \omega_3]$ is analyzed via singular value decomposition. The first principal component $\mathbf{v}_1$ identifies the dominant energy mode. Per-joint weights are computed as:

$$w_i = v_{1,2i-1}^2 + v_{1,2i}^2$$

where $v_{1,k}$ are elements of the first right singular vector. These weights are normalized to sum to 1 and used to distribute injection torque preferentially to joints participating in the dominant mode.

#### Mode-Targeted Injection

During kick windows, injection torques are weighted by PCA mode importance:

$$\tau_i = \text{sgn}(\omega_i) \cdot \tau_{\max} \cdot f_{\text{inject}} \cdot \frac{w_i}{\max_j w_j}$$

This concentrates energy into the dynamically most significant degree of freedom rather than distributing it equally.

---

## 6. Results

### 6.1 Full Experiment: Double Pendulum

The primary results on the double pendulum (20 seeds, 1000 s, 5000 J budget):

| Controller | $M_{0.50}$ (%) | $M_{0.25}$ (%) | $M_{0.10}$ (%) | $M_{0.05}$ (%) | $M_{0.01}$ (%) | $E_{\text{used}}$ (J) |
|---|---|---|---|---|---|---|
| Passive | 5.5 $\pm$ 0.9 | 1.8 $\pm$ 0.6 | 0.6 $\pm$ 0.3 | 0.3 $\pm$ 0.2 | 0.1 $\pm$ 0.0 | 0 |
| Sinusoidal | 6.5 $\pm$ 2.3 | 2.5 $\pm$ 1.1 | 0.9 $\pm$ 0.4 | 0.5 $\pm$ 0.2 | 0.1 $\pm$ 0.0 | 5000 |
| Random Kicks | 32.5 $\pm$ 3.7 | 15.3 $\pm$ 2.5 | 6.0 $\pm$ 1.1 | 3.0 $\pm$ 0.6 | 0.6 $\pm$ 0.1 | 5000 |
| PID | 60.2 $\pm$ 16.3 | 36.6 $\pm$ 11.2 | 15.3 $\pm$ 5.2 | 7.9 $\pm$ 3.2 | 1.7 $\pm$ 0.8 | 5000 |
| Energy Thermostat | 0.0 $\pm$ 0.0 | 0.0 $\pm$ 0.0 | 0.0 $\pm$ 0.0 | 0.0 $\pm$ 0.0 | 0.0 $\pm$ 0.0 | 5000 |
| **FCE** | **55.2 $\pm$ 0.4** | **35.0 $\pm$ 0.5** | **21.4 $\pm$ 0.6** | **15.4 $\pm$ 0.9** | **3.7 $\pm$ 0.7** | **892** |
| FCE+PID Hybrid | 55.9 $\pm$ 1.6 | 35.4 $\pm$ 1.6 | 18.9 $\pm$ 1.9 | 10.4 $\pm$ 1.3 | 2.1 $\pm$ 0.5 | 1031 |

Key observations:

1. **FCE achieves comparable $M_{0.50}$ to PID (55.2% vs 60.2%) using 82% less energy** (892 J vs 5000 J).

2. **FCE dominates all narrow bands**: At $\pm$10%, FCE achieves 21.4% vs PID's 15.3%. At $\pm$5%, FCE leads 15.4% vs 7.9%. At $\pm$1%, FCE achieves 3.7% vs PID's 1.7%.

3. **FCE has dramatically lower variance**: $\sigma = 0.4\%$ across 20 seeds compared to PID's $\sigma = 16.3\%$. FCE's performance is highly reproducible; PID is seed-dependent.

4. **Energy thermostat achieves 0% maintenance** despite consuming 5000 J, demonstrating that maintaining a target kinetic energy level does not produce target-level platform acceleration.

### 6.2 Loading Efficiency

| Controller | Mean Accel (m/s$^2$) | Loading Efficiency (s/J) | Mean Power (W) | W/kg |
|---|---|---|---|---|
| Passive | 1.384 | 0.00 | 0.00 | 0.000 |
| Sinusoidal | 1.604 | 0.03 | 2.50 | 0.250 |
| Random Kicks | 10.416 | 0.13 | 2.50 | 0.250 |
| PID | 6.518 | 0.26 | 2.50 | 0.250 |
| **FCE** | **5.447** | **7.44** | **0.45** | **0.045** |
| FCE+PID | 5.603 | 6.72 | 0.52 | 0.052 |

FCE's loading efficiency of 7.44 s/J is **29$\times$ higher** than PID's 0.26 s/J. FCE's mean power draw is 0.45 W (0.045 W/kg), compared to 2.50 W for all other active controllers.

### 6.3 Comfort and Safety

| Controller | Jerk RMS (m/s$^3$) | Jerk Max (m/s$^3$) | Peak Accel (m/s$^2$) |
|---|---|---|---|
| Passive | 42.2 | 300.1 | 10.8 |
| Random Kicks | 681.4 | 11925.2 | 200.5 |
| PID | 84.2 | 1811.1 | 34.0 |
| **FCE** | **48.1** | **790.9** | **10.8** |
| FCE+PID | 67.2 | 1504.7 | 21.5 |

FCE produces moderate jerk (48.1 m/s$^3$ RMS), comparable to passive dynamics and substantially lower than PID (84.2) or random kicks (681.4).

### 6.4 Triple Pendulum

The triple pendulum proved significantly more challenging:

| Controller | $M_{0.50}$ (%) | $E_{\text{used}}$ (J) |
|---|---|---|
| Passive | 2.2 $\pm$ 1.2 | 0 |
| PID | 43.0 $\pm$ 20.0 | 5000 |
| Random Kicks | 23.3 $\pm$ 4.2 | 5000 |
| FCE (standard) | 18.3 $\pm$ 6.2 | 5000 |

Standard FCE drops from 55.2% (double) to 18.3% (triple) and uses the full 5000 J budget, indicating that the single-joint separatrix detection is insufficient for 3-DOF dynamics.

### 6.5 Duffing Oscillator

All controllers achieved less than 8% $M_{0.50}$ on the Duffing oscillator with the default mass ratio ($m_d/M_p = 0.1$). The oscillator mass is too small relative to the platform to generate meaningful reaction accelerations.

### 6.6 Upgrade 1: Equal-Energy Comparison

At identical 500 J budget (double pendulum, quick mode):

| Controller | $M_{0.50}$ at 500 J | $M_{0.50}$ at 5000 J |
|---|---|---|
| Passive | 6.0% | 6.0% |
| Sinusoidal | 5.4% | 29.3% |
| Random Kicks | 15.8% | 27.7% |
| PID | 7.9% | 65.1% |
| **FCE** | **44.5%** | **54.5%** |

At 500 J, FCE achieves 44.5% while PID manages only 7.9%, a **5.6$\times$ advantage**. At 5000 J, PID catches up to 65.1% by using brute-force continuous injection, but FCE remains competitive at 54.5% while using only 893 J of the 5000 J available.

### 6.7 Upgrade 2: Matched-Maintenance Energy

Binary search for the minimum energy to achieve 50% $M_{0.50}$ (double pendulum):

| Controller | Energy Required | Status |
|---|---|---|
| Passive | $> 20{,}000$ J | Impossible |
| Random Kicks | $> 20{,}000$ J | Impossible |
| PID | 4,142 J | Converged |
| **FCE** | **566 J** | **Converged** |

FCE requires **7.3$\times$ less energy** than PID to achieve the same maintenance level. Passive and random kick controllers cannot achieve 50% maintenance at any tested budget.

### 6.8 Upgrade 3: Phase-Ablation (Timing Causality)

Eight FCE variants were tested to identify which components causally contribute to performance (double pendulum, quick mode):

| Variant | $M_{0.50}$ (%) | $\Delta$ from baseline | $p$-value |
|---|---|---|---|
| FCE Full (baseline) | 44.5 $\pm$ 0.6 | — | — |
| No Separatrix | 44.5 $\pm$ 0.6 | +0.0% | 1.000 |
| No Lyapunov | 44.5 $\pm$ 0.6 | +0.0% | 1.000 |
| Randomized Timing | 30.6 $\pm$ 1.1 | $-$13.9% | $< 0.001$ *** |
| **Shuffled Times** | **24.5 $\pm$ 1.2** | **$-$20.0%** | **$< 0.001$ ***** |
| Random Magnitude | 23.6 $\pm$ 0.7 | $-$20.9% | $< 0.001$ *** |
| Periodic Energy | 25.8 $\pm$ 0.4 | $-$18.7% | $< 0.001$ *** |
| No Curvature | 44.5 $\pm$ 0.6 | +0.0% | 1.000 |

The **killer test** (shuffled times): same total energy, same number of kicks, same kick magnitudes, but with randomly permuted timing. Performance drops by 20.0 percentage points ($p < 0.001$, Welch $t$-test), establishing that **FCE's advantage is causally attributable to injection timing**.

The feature-removal variants (no separatrix, no Lyapunov, no curvature) showed no effect in the 100 s quick-mode runs because the Lyapunov estimator requires warm-up time and the 500 J budget exhausts before the features differentiate. Full 1000 s experiments with 5000 J budgets are expected to reveal these effects.

### 6.9 Upgrade 4: Enhanced Triple FCE

The multi-plane FCE with per-joint Lyapunov and PCA mode targeting (triple pendulum, quick mode):

| Metric | Original FCE | Enhanced FCE | Improvement |
|---|---|---|---|
| $M_{0.50}$ | 8.6 $\pm$ 1.0% | 25.5 $\pm$ 2.1% | +196% |
| $M_{0.25}$ | 3.3 $\pm$ 0.5% | 11.6 $\pm$ 1.3% | +248% |
| $M_{0.10}$ | 1.2 $\pm$ 0.2% | 4.4 $\pm$ 0.5% | +263% |
| Mean Accel | 2.10 m/s$^2$ | 7.61 m/s$^2$ | +263% |
| Loading Eff | 0.15 s/J | 0.21 s/J | +45% |

The enhancement tripled $M_{0.50}$ from 8.6% to 25.5%. The per-joint Lyapunov estimation correctly identifies chaotic joints, and PCA mode targeting concentrates energy into dynamically significant degrees of freedom. The enhanced FCE also outperforms random kicks (17.8%) and PID (4.6%) on the triple pendulum.

### 6.10 Upgrade 5: Duffing Controllability Boundary

Parametric sweep results (quick mode) confirm that mass ratio $m_d / M_p$ is the primary controllability bottleneck:

| $m_d$ (kg) | $M_p$ (kg) | Ratio | Best $M_{0.50}$ | Best Controller |
|---|---|---|---|---|
| 1.0 | 10.0 | 0.10 | 0.3% | FCE |
| 5.0 | 10.0 | 0.50 | 8.2% | FCE |

Increasing $m_d$ from 1.0 to 5.0 kg (ratio 0.10 $\to$ 0.50) improves the best achievable maintenance from 0.3% to 8.2%. Increasing the drive amplitude from 0.3 to 10.0 further improves performance. The current Duffing configuration ($m_d/M_p = 0.1$) is below the controllability threshold for meaningful loading generation.

---

## 7. Discussion

### 7.1 What the System Achieves

The central finding is that phase-targeted energy injection at separatrix crossings is 7--29$\times$ more energy-efficient than feedback control or blind injection for generating inertial loading on a reaction platform. FCE achieves this through temporal selectivity: it injects energy only during the small fraction of time when the chaotic trajectory passes through dynamically favorable phase-space regions.

This efficiency arises from the structure of chaotic dynamics. Near a separatrix, the system is maximally sensitive to perturbations. The Lyapunov exponent quantifies this sensitivity: trajectories separated by $\delta$ at time $t_0$ diverge as:

$$\|\delta(t)\| \sim \|\delta(t_0)\| \cdot e^{\lambda_1 (t - t_0)}$$

For $\lambda_1 = 0.1$ s$^{-1}$ (the moderate chaos threshold), a perturbation grows by a factor of $e^{0.1 \times 10} \approx 2.7$ over 10 s. At separatrix crossings, the effective local Lyapunov exponent is much higher, providing even greater amplification.

### 7.2 What the System Does Not Achieve

The system **does not generate artificial gravity** in the conventional sense. Several fundamental constraints prevent this:

1. **Newton's Third Law**: Joint torques are internal forces. They cannot produce net sustained center-of-mass acceleration. The platform acceleration is purely reactive and necessarily oscillatory.

2. **Bidirectional Loading**: The platform acceleration alternates direction as the pendulum swings. The metric $|a_{\text{platform}}|$ counts the absolute magnitude, which is appropriate for inertial loading but does not represent sustained unidirectional gravity.

3. **Mean Acceleration Below Target**: FCE's mean platform acceleration is 5.45 m/s$^2$ (0.56$g$), well below the 1$g$ target. The $M_{0.50}$ metric counts transient passages through the target band, not sustained residence.

4. **Oscillatory Not Gravitational**: Gravitational loading in 1$g$ is constant, unidirectional, and low-jerk. The pendulum system produces chaotic, bidirectional, high-jerk oscillatory loading.

### 7.3 Potential Applications

Despite not achieving artificial gravity, the system has potential applications:

1. **Vibration-Based Loading in Microgravity**: Whole-body vibration (WBV) platforms have shown efficacy for bone and muscle maintenance on Earth (Rubin et al., 2004). The chaotic pendulum system produces broadband vibration with controllable intensity, potentially offering advantages over periodic WBV.

2. **Energy-Efficient Mechanical Loading**: FCE's 29$\times$ efficiency advantage over PID control is relevant for power-constrained space applications where every watt matters.

3. **Nonlinear Control Methodology**: The FCE algorithm for phase-targeted injection at separatrix crossings is applicable to any chaotic mechanical system where energy efficiency is important.

### 7.4 The Timing Causality Result

The phase-ablation experiment provides the strongest result of this study: FCE's advantage is causally attributable to injection timing, not merely to energy magnitude or injection direction. The shuffled-timing killer test uses identical total energy, identical kick count, and identical kick magnitudes, differing only in when the kicks occur. The 20-percentage-point drop ($p < 0.001$) demonstrates that the temporal structure of FCE's injections, specifically their correlation with separatrix crossings, is the causal mechanism.

This has implications beyond this specific application: it suggests that in any energy-limited nonlinear control problem, phase-aware injection timing may offer substantial efficiency gains over phase-blind strategies.

### 7.5 Limitations

1. **Simulation Only**: All results are from numerical simulation with idealized physics (no joint friction, perfectly rigid links, no structural compliance). Real hardware would introduce additional losses and nonlinearities.

2. **Quick Mode for Upgrades**: Upgrades 1--5 were validated in quick mode (100 s, 500 J, 3 seeds). Full 1000 s, 5000 J, 20-seed experiments are needed for publication-quality conclusions, particularly for the feature-removal ablation variants.

3. **1D Platform Constraint**: The platform translates along a single axis. Real microgravity applications would require 3D loading or at least multi-axis coupling.

4. **Physiological Relevance**: No physiological modeling was performed. Whether oscillatory chaotic loading provides bone/muscle benefits comparable to sustained gravitational loading is an open question requiring experimental validation.

5. **Scaling**: The current system uses laboratory-scale parameters (10 kg platform, 1--1.6 kg pendulum bobs). Scaling to human-rated systems requires analysis of structural loads, motor sizing, and power requirements.

---

## 8. Conclusion

I have presented a computational investigation of phase-targeted energy injection in coupled chaotic pendulum systems for inertial loading generation. The Fractal Correction Engine achieves 55.2% $\pm$50% maintenance on the double pendulum using only 892 J (18% of the 5000 J budget), yielding a loading efficiency of 7.44 s/J, which is 29$\times$ higher than PID control. Phase-ablation experiments establish that this advantage is causally attributable to injection timing at separatrix crossings ($p < 0.001$).

An enhanced multi-plane FCE with per-joint Lyapunov estimation and PCA mode decomposition improves triple-pendulum maintenance by 196%, demonstrating that the FCE framework extends to higher-DOF systems with appropriate state-space decomposition.

The system produces oscillatory inertial loading, not sustained artificial gravity. Newton's third law fundamentally constrains internal mechanisms from generating net persistent acceleration. However, the energy efficiency of phase-targeted control and the broadband nature of chaotic loading may have applications in microgravity countermeasure design, where the goal is mechanical bone/muscle loading rather than gravitational simulation.

The core methodological contribution, exploiting Lyapunov instability at separatrix crossings for energy-efficient nonlinear control, is applicable beyond this specific system to any energy-constrained control problem in chaotic dynamics.

---

## References

Clément, G., Bukley, A. P., & Paloski, W. H. (2015). Artificial gravity as a countermeasure for mitigating physiological deconditioning during long-duration space missions. *Frontiers in Systems Neuroscience*, 9, 92.

Fitts, R. H., Trappe, S. W., Costill, D. L., et al. (2010). Prolonged space flight-induced alterations in the structure and function of human skeletal muscle fibres. *Journal of Physiology*, 588(18), 3567--3592.

Hughson, R. L., Helm, A., & Durante, M. (2018). Heart in space: effect of the extraterrestrial environment on the cardiovascular system. *Nature Reviews Cardiology*, 15(3), 167--180.

Lang, T., LeBlanc, A., Evans, H., et al. (2004). Cortical and trabecular bone mineral loss from the spine and hip in long-duration spaceflight. *Journal of Bone and Mineral Research*, 19(6), 1006--1012.

Rosenstein, M. T., Collins, J. J., & De Luca, C. J. (1993). A practical method for calculating largest Lyapunov exponents from small data sets. *Physica D: Nonlinear Phenomena*, 65(1--2), 117--134.

Rubin, C., Recker, R., Cullen, D., et al. (2004). Prevention of postmenopausal bone loss by a low-magnitude, high-frequency mechanical stimuli: a clinical trial assessing compliance, efficacy, and safety. *Journal of Bone and Mineral Research*, 19(3), 343--351.

Smith, S. M., Heer, M. A., Shackelford, L. C., et al. (2012). Benefits for bone from resistance exercise and nutrition in long-duration spaceflight: evidence from biochemistry and densitometry. *Journal of Bone and Mineral Research*, 27(9), 1896--1906.

---

## Appendix A: Software Architecture

The simulation system comprises the following modules:

| Module | Description |
|--------|-------------|
| `config.py` | Physical constants, hardware specs, scaling scenarios |
| `v3_config.py` | V3 experiment parameters, maintenance bands, FCE tuning |
| `v3_coupled_odes.py` | Lagrangian ODE functions for double, triple, and Duffing models |
| `v3_controllers.py` | 7 controller implementations with common interface |
| `energy_budget.py` | Energy budget enforcement with power clamping |
| `loading_metrics.py` | Streaming metric computation via Welford's algorithm |
| `v3_main_sim.py` | Main simulation loop (controller $\to$ budget $\to$ ODE $\to$ metrics) |
| `run_v3_experiment.py` | Experiment matrix runner with aggregation and plotting |
| `integrators.py` | 4th-order Runge-Kutta integrator |
| `v3_upgrade1_equal_energy.py` | Equal-energy comparison experiment |
| `v3_upgrade2_matched_maint.py` | Matched-maintenance binary search |
| `v3_upgrade3_ablation.py` | 8-variant phase-ablation experiment |
| `v3_upgrade4_triple_fce.py` | Enhanced triple FCE with PCA mode targeting |
| `v3_upgrade5_duffing_sweep.py` | Parametric Duffing controllability sweep |
| `run_v3_upgrades.py` | Unified CLI for upgrade experiments |

All source code and data are available at the associated Zenodo repository.

## Appendix B: Reproducibility

To reproduce the full experiment:

```bash
# Full 420-run experiment (approx. 90 minutes)
python run_v3_experiment.py

# Quick validation (approx. 2 minutes)
python run_v3_experiment.py --quick

# Upgrade experiments
python run_v3_upgrades.py --upgrade all         # Full mode
python run_v3_upgrades.py --upgrade all --quick  # Quick validation
```

Dependencies: Python 3.8+, NumPy, SciPy, Matplotlib.

Random seeds 0--19 are used for statistical runs. Initial conditions are perturbed by $\pm$5% of nominal angles, seeded per run for exact reproducibility.