Tangent Hyperplane Framework: Reading Path

A guided route through exploratory notes on nonlinear control and geometric intuition.

A guided route through exploratory notes on nonlinear control and geometric intuition

Overview

This collection presents an exploratory geometric framework for understanding nonlinear dynamical systems through local tangent-space structure. The central technical point is narrower than many of the draft pages suggest: at a fixed state, the first variation gives a linear map for infinitesimal perturbations. Finite-step prediction, accumulated error, and controller performance require residual bounds and validation.

Target Audience: - Graduate students in control theory, robotics, biomechanics - Practitioners in aerospace, robotics, sports science - Researchers seeking geometric intuition for optimization algorithms - Anyone curious about why LQR, MPC, and DDP actually work

Prerequisites: - Linear algebra (matrices, eigenvalues, linear maps) - Multivariable calculus (partial derivatives, gradients) - Basic differential equations (first-order ODEs) - Introductory control (state-space representation helpful but not required)

Learning Tracks

We provide three learning paths depending on your goals:

Track 1: Conceptual Understanding (No Equations)

Goal: Understand “why” without “how” - Start with Layman’s Terms summaries - Read motivational sections only - Skip mathematical derivations - Time: 2-3 hours

Track 2: Applied Practitioner (Implement Algorithms)

Goal: Use methods in your work - Read main articles, skip proofs - Focus on algorithm pseudocode - Study case studies closely - Implement examples in Python/MATLAB - Time: 2-3 weeks

Track 3: Technical Review (Full Derivations and Caveats)

Goal: Understand the claims well enough to evaluate or extend them - Read the main pages, appendices, and critique pages together - Work through all derivations - Study critical reviews - Solve problem sets (future addition) - Time: 1-2 months

Part I: Foundations (Required for All)

1. Tangent_Hyperplanes_Unified_Thesis.qmd 📘 CORE DOCUMENT

Prerequisite: None (self-contained) Difficulty: ⭐⭐⭐ Moderate Time: 4-6 hours

What You’ll Learn: - Part I: Where tangent spaces provide exact infinitesimal structure - Part II: How integration preserves superposition of variations - Part III: How DDP/iLQR/MPC exploit this structure

Key Concepts: - Fréchet derivative as exact infinitesimal structure - State transition operators - Residuals as manifold curvature - Hamiltonian formulation - Worked DDP/iLQR algorithm sketches with case studies

Why Start Here: This is the broadest reference in the package. Read it together with the critique pages before treating any finite-time or control-design claim as established.

Supplementary Materials: - LAYMANS_TERMS_SUMMARY.qmd - Non-technical overview (read first if intimidated) - TECHNICAL_ASSESSMENT.md - Quality evaluation (for advisors/reviewers) - CRITICAL_REVIEW.md - Known weaknesses and defenses - CRITICS_CORNER.qmd - Responses to harsh criticisms

2. Tangent_Hyperplanes_Golf_Application.md 🏌️ OPTIONAL SUPPLEMENT

Prerequisite: Unified Thesis (Part I only) Difficulty: ⭐⭐ Easy Time: 1 hour

What You’ll Learn: - Application to biomechanics (golf swing) - Why coaching decomposition makes mathematical sense - Connection to sports science

Why Read: If you’re interested in human movement, this makes the abstract concrete.

Part II: Advanced Theory (For Deep Understanding)

After mastering Part I, these articles extend the framework in specific directions.

3. Residual-Aware_Control.qmd 📊 NEW

Prerequisite: Unified Thesis (Parts I-III) Difficulty: ⭐⭐⭐⭐ Advanced Time: 3-4 hours

What You’ll Learn: - Quantitative residual bounds with explicit constants - Real-time residual monitoring for control switching - Curvature-adaptive timesteps in DDP - Tube MPC with geometric residual bounds

Working contribution: Treats residuals as candidate diagnostics for when local models may be insufficient.

Applications: - Quadrotor: Hover (low curvature) vs. aerobatics (high curvature) - Humanoid gait: Foot strike = curvature spike - Spacecraft: Gimbal lock = curvature singularity

Why Read: Connects the local theory to a practical question: when is a linearized prediction no longer reliable enough for the intended task?

Supplementary: - Residual-Aware_Control_LAYMAN.qmd - Accessible explanation - Residual-Aware_Control_CRITIC.qmd - Technical objections answered

4. Contraction_Tangent_Unification.qmd 🔄 NEW

Prerequisite: Unified Thesis + familiarity with Lyapunov stability Difficulty: ⭐⭐⭐⭐⭐ Expert Time: 4-5 hours

What You’ll Learn: - Contraction theory (Lohmiller & Slotine) via tangent-bundle language - Conditional links between stability analysis and optimal-control calculations - Contraction metrics as Riemannian geometry on tangent spaces - How contraction constraints can be added to trajectory optimization, and what must be checked before claiming convergence

Working contribution: Frames contraction analysis and trajectory optimization as closely related local calculations, while leaving global or finite-time claims dependent on additional assumptions.

Applications: - Checking DDP convergence conditions for specific systems - Designing controllers with explicit local stability certificates - Understanding when LQR assumptions are strong enough for the intended claim

Why Read: If you know contraction theory, this shows a possible link to optimization. If you know DDP, this shows how stability certificates might be added and where the proof obligations enter.

Supplementary: - Contraction_Tangent_LAYMAN.qmd - Why trajectories “forget” initial conditions - Contraction_Tangent_CRITIC.qmd - Comparison to standard contraction literature

5. Hybrid_Tangent_Spaces.qmd 🔀 NEW

Prerequisite: Unified Thesis + basic knowledge of hybrid systems Difficulty: ⭐⭐⭐⭐ Advanced Time: 3-4 hours

What You’ll Learn: - Extending tangent space framework to discontinuous systems - Each mode has smooth dynamics (tangent space applies) - Guard conditions = tangent space jumps - Impact maps as instantaneous rotations in tangent bundle - DDP for hybrid systems (mode-aware optimization)

Key Innovation: Addresses “C¹ smoothness is unrealistic” criticism—shows framework extends to impacts, friction, switches.

Applications: - Bipedal walking (foot strike = impact → discrete jump in tangent space) - Robotic grasping (contact = mode switch) - Chemical reactors (phase transitions = guard crossings)

Why Read: Most real systems are hybrid (contact, switches, saturation). This makes the framework applicable to actual robots and biomechanics.

Supplementary: - Hybrid_Tangent_LAYMAN.qmd - Hopping on one foot (intuitive hybrid example) - Hybrid_Tangent_CRITIC.qmd - Comparison to hybrid automata literature

Concept Dependencies (What Builds on What)

Unified Thesis (Part I: Geometry)
    ├─→ Golf Application [optional sidebar]
    ├─→ Unified Thesis (Part II: Integration)
    │       └─→ Unified Thesis (Part III: Optimization)
    │               ├─→ Residual-Aware Control
    │               │       └─→ [Implements quantitative bounds]
    │               ├─→ Contraction Unification
    │               │       └─→ [Adds stability to optimization]
    │               └─→ Hybrid Systems
    │                       └─→ [Extends beyond C¹ assumption]
    │
    └─→ Contraction Unification
            └─→ [Can be read independently if familiar with Lyapunov theory]

Notation: A → B means “A is prerequisite for B”

Prerequisites by Document

Document Math Level Control Background Programming Time
Layman Summaries None None None 30 min each
Unified Thesis Calculus, Linear Algebra Helpful Optional 4-6 hrs
Golf Application Basic calculus None None 1 hr
Residual-Aware Multivariable calculus State-space Python recommended 3-4 hrs
Contraction Real analysis, ODEs Lyapunov theory MATLAB/Python 4-5 hrs
Hybrid Systems ODEs, Measure theory (basic) Hybrid automata (intro) Python + simulation 3-4 hrs

Key Equations Reference (Quick Lookup)

For when you need to remember “what was that formula?”

Concept Equation Document Section
Fréchet Derivative \(f(x_0 + \delta x) = f(x_0) + A\delta x + o(\|\delta x\|)\) Unified Thesis Part I, Ch. 1
Variational Dynamics \(\delta\dot{x} = A(t)\delta x + B(t)\delta u\) Unified Thesis Part I, Ch. 2
State Transition \(\delta x(t_1) = \Phi(t_1,t_0)\delta x(t_0) + \int \Phi(t_1,\tau)B(\tau)\delta u(\tau)d\tau\) Unified Thesis Part II, Ch. 6
Residual Scaling \(\|r\| = O(\epsilon^2)\) Unified Thesis Part I, Ch. 4
Quantitative Residual \(\|r(t_1)\| \leq \frac{1}{2}\|H\|_{\max}\int \|\delta x\|^2 dt\) Residual-Aware Part I, Theorem 1
Hamiltonian \(H = L + \lambda^T f\) Unified Thesis Part III, Ch. 9
DDP Q-function \(Q_{uu} = L_{uu} + B^T P B\) Unified Thesis Part III, Ch. 11
Contraction Metric \(\dot{V} \leq -\alpha V\) Contraction Part I, Def. 2
Impact Map \(\dot{x}^+ = \Delta(\dot{x}^-)\) Hybrid Systems Part II, Ch. 3

Glossary of Key Terms

Fréchet Derivative: The unique best linear approximation to a nonlinear function at a point. Not “an” approximation, but “the” exact derivative.

Tangent Space (\(T_x\mathcal{M}\)): The space of all velocity vectors at point \(x\) on manifold \(\mathcal{M}\). Locally looks like \(\mathbb{R}^n\), where dynamics are exactly linear.

Residual (\(r\)): The failure of superposition for finite perturbations. Quantifies how much tangent spaces vary. Scales as \(O(\epsilon^2)\) (quadratic in perturbation size).

State Transition Operator (\(\Phi(t_1,t_0)\)): Linear map transporting perturbations forward in time through varying tangent spaces. Fundamental solution of variational equation.

Costate (\(\lambda\)): Adjoint variable in optimal control. Represents marginal cost sensitivity. Evolves backward in time.

Contraction: Property where all trajectories exponentially converge to each other. Jacobian is uniformly negative definite.

Hybrid System: Combines continuous dynamics (flows) with discrete events (jumps). Example: walking robot (swing phase = flow, foot strike = jump).

DDP (Differential Dynamic Programming): Trajectory optimization via iterated local quadratic approximations. Newton’s method in function space.

Common Pitfalls and How to Avoid Them

Pitfall 1: “Linearization Is Approximate, so Results Are Approximate”

Why it’s wrong: The tangent space (derivative) is exact. What’s approximate is using it far from the linearization point. DDP/iLQR re-linearize constantly, maintaining exactness.

Where to learn more: Unified Thesis Ch. 1, CRITICS_CORNER.qmd Criticism #2

Pitfall 2: “This Only Works for Small Perturbations”

Why it’s misleading: A single linearization only works locally. But iterative methods (DDP, MPC) keep taking small steps with fresh linearizations. You can reach anywhere via many small steps.

Where to learn more: Unified Thesis Ch. 10-11, Residual-Aware Control Part II

Pitfall 3: “Residuals Are Errors I Need to Minimize”

Why it’s misleading: Residuals are geometric features (curvature), not modeling errors. They’re inevitable for nonlinear systems. Goal is to understand and account for them, not eliminate them.

Where to learn more: Unified Thesis Ch. 4, Residual-Aware Control Part I

Pitfall 4: “Contraction Is for Stability, Optimization Is Separate”

Why it’s wrong: Contraction and optimization are dual perspectives. Contraction = “trajectories converge,” Optimization = “cost decreases.” Both use Jacobians, both exploit tangent geometry.

Where to learn more: Contraction Unification, entire article

Pitfall 5: “Framework Fails for Impacts/Friction (Non-Smooth)”

Why it’s misleading: Framework applies within each smooth segment. Impacts/switches handled separately (impact maps, guard conditions). Hybrid Systems article shows how.

Where to learn more: Hybrid Systems, Unified Thesis “Scope of Validity”

FAQ: Choosing What to Read

Q: I’m a mechanical engineering undergrad. Where do I start? A: LAYMANS_TERMS_SUMMARY.qmd → Unified Thesis Parts I-II (skip proofs) → Golf Application

Q: I use MPC but don’t understand why it works. Help? A: Unified Thesis Part III (Chapters 10-11) → Residual-Aware Control Part II

Q: I need to prove my controller is stable. What do I read? A: Unified Thesis Part I (Ch. 3: Lyapunov) → Contraction Unification

Q: I’m implementing a walking robot with foot impacts. Relevant article? A: Unified Thesis Parts I-II → Hybrid Systems (all parts)

Q: I want to write a paper using this framework. What’s novel? A: Read CRITICAL_REVIEW.md and CRITICS_CORNER.qmd carefully. Novel aspects: geometric framing, residual-as-feature, unification across methods.

Q: I’m an advisor evaluating a student’s thesis using this. Quality assessment? A: TECHNICAL_ASSESSMENT.md (overall B+/A-) + CRITICAL_REVIEW.md (lists weaknesses)

Q: How does this compare to my textbook (Khalil, Sastry, Slotine)? A: CRITICS_CORNER.qmd Criticism #7 has detailed comparison table

Notation Conventions

We use consistent notation across all articles:

Symbol Meaning Dimensionality
\(x\) State vector \(\mathbb{R}^n\)
\(u\) Control input \(\mathbb{R}^m\)
\(f(x,u)\) Vector field (dynamics) \(\mathbb{R}^n \to \mathbb{R}^n\)
\(A = \frac{\partial f}{\partial x}\) State Jacobian \(\mathbb{R}^{n \times n}\)
\(B = \frac{\partial f}{\partial u}\) Input Jacobian \(\mathbb{R}^{n \times m}\)
\(\delta x\), \(\delta u\) Infinitesimal perturbations \(\mathbb{R}^n\), \(\mathbb{R}^m\)
\(\Phi(t_1, t_0)\) State transition operator \(\mathbb{R}^{n \times n}\)
\(\lambda\) Costate (adjoint) \(\mathbb{R}^n\)
\(H\) Hamiltonian \(\mathbb{R}\)
\(L\) Lagrangian (running cost) \(\mathbb{R}\)
\(r\) Residual (superposition failure) \(\mathbb{R}^n\)
\(\epsilon\) Perturbation magnitude \(\mathbb{R}_+\)
\(T_x\mathcal{M}\) Tangent space at \(x\) Vector space \(\cong \mathbb{R}^n\)

Calculus notation: - \(o(\cdot)\): Little-o (vanishes faster than argument) - \(O(\cdot)\): Big-O (grows no faster than argument) - \(\|\cdot\|\): Euclidean norm (2-norm) - \(\frac{\partial}{\partial x}\): Partial derivative - \(\nabla_x\): Gradient with respect to \(x\)

Contribution and Future Directions

This framework is actively evolving. Planned additions:

  1. Interactive Visualizations: Moving tangent spaces in 3D (WebGL)
  2. Problem Sets: Exercises with solutions for each article
  3. Code Repository: Full implementations of all case studies (Python/JAX)
  4. Video Lectures: Recorded walkthroughs of key concepts
  5. Application Notes: Domain-specific guides (aerospace, biomechanics, chemical eng.)

How to contribute: - GitHub: D-sorganization/AffineDrift - Issues: Report errors, request clarifications - Discussions: Ask questions, share applications

Acknowledgments

This framework synthesizes ideas from: - Differential geometry (Fréchet, Lee, Arnold) - Nonlinear control (Kalman, Jacobson, Mayne, Sastry, Khalil) - Contraction theory (Lohmiller, Slotine) - Trajectory optimization (Bryson, Ho, Tassa) - Biomechanics (McMahon, Alexander)

We stand on the shoulders of giants. The contribution is synthesis and geometric framing, not invention of new mathematics.

Final Recommendation: Where to Start

If you have 30 minutes: Read any LAYMANS summary If you have 2 hours: Read Unified Thesis Part I only If you have 1 day: Read the Unified Thesis with the critique pages open If you have 1 week: Read the Unified Thesis, one advanced article, and implement a small case study If you have 1 month: Read the full package, compare the critiques, and test one extension in your domain

Most important: Do not read the draft pages as finished evidence summaries. Use the table above to check prerequisites and keep the limitation notes alongside the main claims.

The goal is not to memorize formulas; it is to develop geometric intuition while keeping the mathematical scope conditions visible.

The useful habit is to separate exact local statements from finite-time conclusions that require residual bounds, model checks, and validation.

Document Version: 1.0 Last Updated: January 18, 2026 Maintainer: AffineDrift Team