In Layman’s Terms: Contraction Theory Meets Tangent Spaces

A non-technical explanation of contraction theory and tangent spaces in control theory.

Target Audience: Non-technical readers, practitioners, students Source: Contraction_Tangent_Unification.qmd Date: January 18, 2026

The Big Idea in One Sentence

“In some local control problems, the same calculation that prices errors can also help design feedback that reduces nearby errors.”

The Core Problem

Imagine you’re designing a robot arm to paint intricate patterns on pottery. You face two completely different challenges:

Optimality: Plan an efficient motion for the brushstroke
Stability: Ensure tiny vibrations or bumps don’t ruin the painting

Traditionally, these were treated as separate problems requiring separate solutions. Control engineers would: - First optimize the trajectory - Then add stabilizing feedback (make it “robust”)

The useful observation: In certain linearized or LQR-like settings, these problems can share the same mathematical object. That does not make them identical in every nonlinear system, but it gives engineers a way to design and check them together.

This article explains that local relationship and why its assumptions matter.

What Is Contraction Theory?

Before we can talk about the unification, we need to understand what “contraction” means.

The Rubber Band Analogy

Think of a rubber band stretched between your fingers. What happens when you release one end?

It snaps back. The rubber band wants to return to its natural, unstretched state. The farther you stretch it, the harder it pulls back.

Contraction theory says: Some dynamical systems behave like rubber bands—errors naturally shrink over time.

Everyday Examples

1. Your car’s cruise control: - If you go a bit too fast, the system reduces throttle - If you go a bit too slow, it increases throttle - Either way, you converge back to the target speed - The system is “contracting” toward the setpoint

2. A swinging pendulum with friction: - No matter where you release it, it eventually stops at the bottom - All trajectories converge to the same resting state - The friction makes the system contracting

3. A thermostat: - Room too hot? Turn on AC - Room too cold? Turn on heat - Temperature naturally drifts toward the setpoint

The Mathematical Definition (Simplified)

A system is contracting if: - Any two nearby states get closer together over time - This happens exponentially fast—like compound interest in reverse - The rate of convergence can be bounded when the required assumptions hold

Physical meaning: In a certified region, small errors can shrink at a rate estimated from the local dynamics and metric.

What Are Tangent Spaces?

Now the second piece of the puzzle.

The Curved Road Analogy (Revisited)

From the main thesis, you learned that nonlinear systems admit an exact linear approximation at any instant — the dynamics follow simple, linear rules on a tangent plane to first order (with error growing as the square of the step size).

This article’s new insight: That tangent plane isn’t just a crude approximation. It’s the natural geometry where both optimality and stability live — and in that geometry, contraction and optimality coincide. (The approximation caveat from the main thesis still applies: finite-step predictions on the tangent plane incur curvature error; contraction theory handles this by ensuring errors shrink rather than accumulate.)

The Optimal Control Connection

When you solve an optimal control problem (like planning a robot’s motion), you get two things:

A trajectory: The path the robot should follow
Feedback gains: How hard to push if you drift off the path

The feedback gains define a metric—a way of measuring “distance” from the optimal path. And here’s the kicker:

That metric can sometimes also serve as a contraction metric.

In other words: - The optimal controller makes errors expensive (optimality) - Under the right assumptions, the same controller can make local errors shrink exponentially (stability)

These can be two compatible readings of the same local calculation.

The Duality Revealed

Here’s the heart of the unification:

Two Perspectives, One Reality

Optimality perspective: - “If I deviate from the plan, it costs me this much (measured by the Riccati matrix)” - Focus: Minimizing cost

Stability perspective: - “If I deviate from the plan, I’ll converge back this fast (measured by the contraction metric)” - Focus: Guaranteeing convergence

The key point: In LQR-like cases, the Riccati matrix from optimal control can be used as a contraction metric for the local closed-loop system.

That relationship is useful, but it depends on the model, boundary conditions, and the region where the local approximation remains valid.

The Rubber Band Is Optimal

Going back to the rubber band:

Stability view: The rubber band pulls you back (contraction rate)
Optimality view: The rubber band’s tension measures how far off you are (cost-to-go)

Physical interpretation: The “springiness” of the optimal feedback is exactly what’s needed to make the system self-correcting.

Why This Matters: Real-World Applications

1. Robot Arms That Self-Stabilize

The old way: 1. Plan optimal trajectory 2. Design separate stabilizing controller 3. Hope they don’t interfere with each other

The new way (contraction-aware control): 1. Plan trajectory with contraction constraints built in 2. Check whether the resulting feedback gives a certified local region of attraction

What would need to be shown: A credible robot-arm example should report the model, disturbance tests, baseline controller, sample size, and whether the contraction certificate predicts the observed tracking behavior.

Why it could work: The controller does not just steer toward the goal; it is also checked for local error-shrinking behavior.

2. Golf Swings With Safety Margins

The problem: Professional golfers swing faster than amateurs but maintain control better. Why?

What contraction analysis could test: - Whether a measured swing has regions where perturbations are locally damped - Whether stability margins differ across golfers, clubs, or phases of the swing - Whether the model remains valid near impact and contact transitions

Interpretation: A measured stability margin would be evidence about a specific model and data set, not a general proof that professional swings are globally contracting.

Training implication: Any coaching claim would need direct data showing that a drill improves measured robustness, not only path accuracy.

3. Spacecraft Docking With Guarantees

Challenge: Docking with the ISS requires centimeter-level precision. Orbital mechanics are nonlinear. Fuel is limited.

Contraction-based approach: 1. Plan trajectory using DDP (differential dynamic programming) 2. Add contraction penalty to ensure exponential convergence 3. Result: A candidate basin of attraction that still needs verification against model error and disturbances

Benefit: If the disturbance stays inside the certified region and the model assumptions remain valid, contraction analysis can support a recovery claim.

Safety-critical systems need explicit margins, independent verification, and failure-mode analysis; contraction theory can be one part of that evidence.

The Geometric Picture

Curved Space and Metrics

Here’s where it gets beautiful (and a bit abstract).

Traditional view: State space is flat. Distance is measured with a ruler (Euclidean metric).

Contraction view: State space is curved. Distance is measured with a Riemannian metric—like measuring distance on Earth’s surface using latitude/longitude.

The Riccati matrix defines that curvature.

What This Means Physically

Imagine you’re walking on a trampoline: - Near the center (the goal), the trampoline curves steeply downward - The curvature creates a “bowl” that pulls you toward the center - The steeper the bowl, the faster you slide to the center

The optimal controller shapes the “bowl” (the metric) to make errors decay at the desired rate.

Analogy to gravity: Optimal control doesn’t push you toward the goal. It bends spacetime so that the goal is “downhill” from everywhere.

The Math (Simplified)

You don’t need to understand the equations to appreciate the beauty, but here’s the executive summary:

The Riccati Equation

What it is: A differential equation that describes how the “cost-to-go” (value function) changes along a trajectory.

Standard form:

-dS/dt = A^T S + S A - S B R^(-1) B^T S + Q

What it says: The value function S(t) satisfies a nonlinear differential equation.

The Contraction Condition

What it is: A requirement that ensures errors shrink exponentially.

Standard form:

A^T M + M A < -2λ M

What it says: If the metric M satisfies this inequality, trajectories converge with rate λ.

The Duality

The conditional fact: In the linear-quadratic setting, setting M = S (metric equals value Hessian) can satisfy both the optimal-control and contraction interpretations.

In other words: - Solving the optimal control problem gives you S - That S can be tested as a contraction metric - The contraction rate comes from the eigenvalues of the closed-loop system

The connection is useful, but it still has to be checked outside the linear-quadratic case.

Contraction-Constrained Optimization

Once the local link between optimality and stability is explicit, we can design algorithms that optimize a trajectory and check its stability region.

The Idea: Optimize With Stability Guarantees

Standard trajectory optimization: - Minimize cost (fuel, time, error) - Hope the result is stable

Contraction-aware trajectory optimization: - Minimize cost subject to contraction constraints - Get a stability certificate as part of the solution, if the verification conditions hold

Mathematical form:

Minimize: trajectory cost
Subject to:
  - Dynamics constraints
  - Contraction rate ≥ λ (user-specified)
  - Control limits

Output: - Optimal trajectory - Feedback gains - Estimated or certified basin of attraction–the region where convergence has been verified under the stated assumptions

The Cartpole Example

Problem: Swing up an inverted pendulum on a cart from hanging down to upright.

Comparison:

Metric	Classical DDP	Contraction-DDP	Improvement
Success rate (σ=0.1)	94%	100%	+6%
Success rate (σ=0.2)	71%	98%	+38%
Success rate (σ=0.3)	42%	89%	+112%
Basin of attraction	14.3 units³	28.7 units³	2× larger
Convergence consistency	High variance	Low variance	More reliable

σ = initial state perturbation (noise in starting position/velocity)

Takeaway: By explicitly enforcing contraction, you get: - Larger basin of attraction (more robust to disturbances) - More consistent convergence (predictable behavior) - Modest computational cost (+60% with warm-starting)

Applications to Biomechanics

This is where the control-theory framework becomes a biological hypothesis rather than an established result.

Muscle Synergies Explained

The puzzle: Humans have many shoulder, elbow, wrist, and hand muscles, but the task-level movement often has far fewer spatial degrees of freedom. This redundancy raises a modeling question: which muscle combinations are used, and why?

Old hypothesis: “The brain simplifies by grouping muscles into synergies.”

Contraction-based hypothesis: “Some measured synergies may be organized to improve local error recovery or impedance in task-relevant directions.”

What This Means

If this hypothesis is correct, measured synergies should align with contraction or impedance directions predicted by a specified musculoskeletal model. That is a testable claim, not something established by the equations alone.

Analogy: Imagine you’re steering a ship with 50 ropes. You could pull them randomly, but it’s smarter to: - Group ropes by their effect (e.g., “left turn bundle”, “right turn bundle”) - Pull bundles, not individual ropes

Contraction theory would predict: The useful groupings are those that help the system self-correct when perturbed, after the model, task, and stability metric are specified.

Evidence status: Studies showing that muscle synergies change with task demands are compatible with this view, but they are not direct evidence that the nervous system optimizes a contraction metric. Direct support would require perturbation, EMG, impedance, and motion-capture data compared against model-predicted metrics.

Impedance Control

What it is: The nervous system adjusts muscle stiffness (impedance) to control motion.

Contraction view: Impedance can sometimes be interpreted as a task-dependent weighting of errors. Treating it as a contraction metric requires a model that links muscle activation, endpoint stiffness, and error decay.

Example: Holding a full coffee cup: - High stiffness in vertical direction (don’t spill!) - Low stiffness in horizontal direction (allow smooth motion)

The cautious interpretation is that the nervous system may shape stiffness and damping to prioritize stability where errors are costly. Calling that a contraction metric is a model claim that needs measurement.

The Algorithms: How to Actually Use This

Contraction-DDP Algorithm

Input: - Initial trajectory - Desired contraction rate λ - Penalty weight μ (how much to prioritize stability)

Repeat until convergence:

Forward pass: Simulate dynamics
Metric optimization: Solve for optimal contraction metric M (convex optimization)
Backward pass: Compute feedback gains using an augmented cost or constraint set that includes a contraction term
Line search: Update trajectory with step size α

Output: - Optimal trajectory - Time-varying feedback gains - Verified or estimated contraction-rate bound, if the assumptions hold

Computational cost: Metric optimization adds solver cost. The size of that cost and the strength of the resulting certificate are implementation-dependent.

Practical Implementation (JAX)

Why JAX? - Automatic differentiation (compute derivatives of the implemented model, within numerical precision) - JIT compilation (fast execution) - Batching (vectorize over time steps)

Key steps: 1. Define dynamics function 2. Use jacfwd to compute linearization matrices A, B 3. Solve Riccati equation backward in time 4. Verify contraction condition

Example output (pendulum swing-up):

Solving with classical DDP...
Converged in 12 iterations

Solving with contraction-DDP...
Converged in 18 iterations
Final contraction rate: 1.94 Hz (target: 2.0 Hz)

Trade-off: More iterations and solver complexity, with a local stability certificate only when the contraction condition, model assumptions, and numerical tolerances have been verified.

Common Questions

“Isn’t This Just LQR?”

Almost, but with a twist.

LQR (Linear Quadratic Regulator) gives optimal feedback for linear systems. For nonlinear systems, you need to: - Linearize at every point along the trajectory (time-varying LQR) - This is what DDP does

What’s useful: In linear-quadratic settings, the Riccati matrix can serve both as a value-function Hessian and as a local Lyapunov or contraction candidate. That connection can support a stability certificate, but only after the closed-loop conditions and bounds are checked.

“Does This Work for Discontinuous Systems?”

Not directly. Standard contraction arguments require smooth dynamics (at least C^1–continuously differentiable) within the region being analyzed.

For systems with: - Impacts (e.g., walking, bouncing) - Switches (e.g., mode changes) - Friction (stick-slip)

You need hybrid system techniques. But: You can use contraction within each mode, then analyze switching separately.

“What’s the Computational Bottleneck?”

Metric optimization. Finding a contraction metric M may require solving a semidefinite program (SDP) or another constrained optimization problem at each time step.

Scaling: The cost depends on the metric parameterization and solver. Dense interior-point SDP methods can become expensive quickly, while structured or fixed-metric approaches trade certificate strength for tractability.

Mitigation: - Warm-starting (use previous solution as initial guess) - Sparsity (exploit block structure) - Approximation (use fixed metric M = I, lose optimality but retain stability)

Practical limit: There is no single state-count threshold. Real-time feasibility depends on model dimension, sparsity, warm starts, solver choice, hardware, and how strong a certificate is required.

Why Two Fields Seemed Separate

Historical Accident

Optimal control (1960s): Focus on minimizing cost functionals, Pontryagin’s maximum principle, dynamic programming.

Stability theory (1890s-present): Focus on Lyapunov functions, convergence analysis, robustness.

Different communities, different journals, different notation.

Example: The Riccati equation appears in both fields but was called: - “Value function propagation” in optimal control - “Lyapunov function candidate” in stability theory

The same Riccati and Lyapunov objects appear in both traditions, but the conditions under which they carry the same interpretation must be stated.

The Conceptual Barrier

Optimal control asks: “What’s the best thing to do?”

Stability theory asks: “Will the system behave predictably?”

These can be related questions in restricted problem classes. In linear-quadratic systems, value functions, feedback gains, and Lyapunov functions can share the same matrices. Outside those settings, the relationship is conditional and may be non-unique.

The Broader Impact

1. Unified Design Tools

Before: Design trajectory with one tool, analyze stability with another, iterate until both work.

After: A single optimization can include both performance objectives and stability constraints, if the assumptions needed for the certificate are satisfied.

Example: Satellite trajectory planning: - Old: Optimize fuel, then check if stable, then re-optimize, repeat - New: Optimize fuel with a contraction constraint, then verify the resulting local stability certificate

2. Interpretability

Machine learning controllers can perform well, but their closed-loop behavior is often difficult to certify without additional analysis.

Contraction-aware control: - Transparent: The metric gives an auditable stability argument when the assumptions are met - Certifiable: The basin of attraction may be estimated or bounded under specified conditions - Reviewable: The certificate can be compared with safety requirements rather than treated as empirical performance alone

Use case: Safety-critical autonomy needs explicit evidence about failure modes and operating regions, not only aggregate success rates.

3. Biomechanical Insights

Neuroscience puzzle: How does the brain control 600+ muscles without getting tangled?

Contraction hypothesis: Some motor strategies may favor robust, self-correcting motions rather than only energetic efficiency.

Testable prediction: Brain-damaged patients should show: - Preserved optimality (can still reach targets) - Degraded contraction (unstable motions, more corrections)

This is a proposed experimental test. Support would require comparing patient movement data with model-predicted contraction or impedance structure, not only observing that movements are unstable.

The Key Equations (For the Curious)

If you want to dig deeper, here are the core equations:

1. Contraction Condition

A^T M + M A + dM/dt ≤ -2λ M

Says: Metric M makes errors decay with rate λ.

2. Differential Riccati Equation

-dS/dt = Q + A^T S + S A - S B R^(-1) B^T S

Says: Value function S propagates backward in time.

3. Duality

M = S

Says: In the restricted linear-quadratic case, the metric can be identified with the value-function Hessian.

4. Closed-Loop Dynamics

dx/dt = (A - B K) x,   K = R^(-1) B^T S

Says: Under appropriate conditions, optimal feedback can render the local closed-loop system contracting.

5. Contraction Rate

λ = (1/2) λ_min(S B R^(-1) B^T S)

Says: A contraction-rate bound can be related to eigenvalue margins after the assumptions and units are specified.

The point is not that these equations are universally interchangeable. It is that, in specific problem classes, they expose the same local stability and cost geometry.

Philosophical Takeaway

When Cost and Stability Align

In a well-specified control problem: - The chosen cost can reward stable behavior - Stability constraints can be part of the design objective - The “cost of error” and “rate of error decay” may be represented by related local quadratic forms

Engineering lesson: Do not assume that performance and robustness are automatically aligned. State the cost, constraints, model class, and certificate before claiming that an optimal design is robust.

Who Should Care?

Researchers

Unified framework for analyzing nonlinear control
New connections between geometry, optimization, and dynamics

Engineers

Design controllers with verifiable local stability certificates
Compute basins of attraction for safety certification

Neuroscientists

Test whether contraction or impedance metrics explain measured motor-control structure
Compare predicted synergy structure with task-specific data

Coaches/Trainers

Treat mechanical stability as a hypothesis to compare with measured technique
Use model predictions to identify which training claims need empirical validation

What’s Next?

Open Problems

Stochastic systems: Extend to noisy dynamics (partially done, see stochastic Riccati equation)
Learning metrics: Use machine learning to discover optimal contraction metrics from data
Hybrid systems: Handle impacts and mode switches
High dimensions: Scale to 100+ states (requires faster SDP solvers)

Practical Tools

All code examples available at:

https://github.com/AffineDrift/contraction-tangent-unification

Includes: - JAX implementation of Contraction-DDP - CVXPY metric optimization - Benchmark problems (cartpole, pendulum, robot arm) - Visualization tools

The Bottom Line

Optimal control and contraction theory are not the same theory in general, but they can share the same local mathematical objects in important restricted settings.

In linear-quadratic and closely related local problems, the Riccati matrix can function as both a cost-to-go Hessian and a stability metric. Extending that interpretation to nonlinear, stochastic, hybrid, or biological systems requires additional assumptions and validation.

Practical impact: - Design controllers whose stability claims can be checked - Compare basin estimates under explicitly stated model assumptions - Convergence-rate bounds when the model and certificate assumptions hold - A design methodology that keeps performance and stability checks in the same workflow

Conceptual impact: - Optimality and stability can share local quadratic structure in restricted settings - Dynamics may require geometry-aware local models - The chosen metric determines which errors are emphasized

This is a useful way of seeing some control problems: performance and robustness may be linked by the same local model, but the link must be checked rather than assumed.

The practical value is the discipline it imposes: state the local model, state the metric, and state the region where the certificate applies.

<div class="laymans-terms-inner">
  <p class="laymans-terms-intro">
    Here is a simplified breakdown of how local planning and local error correction can be linked.
  </p>

  <div class="laymans-item">
    <h3>Self-Correcting Systems</h3>
    <p>This is about machines or motions whose nearby errors shrink under a verified feedback law. If a small bump stays inside the certified region, the system is designed to move back toward the reference path.</p>
    <div class="analogy">

Think of it like: Dropping a marble near the bottom of a smooth bowl. Close to the bottom, the slope points the marble back inward; far away, the shape may be different.

</div>

  <div class="laymans-item">
    <h3>Efficient Plans vs. Checked Plans</h3>
    <p>Finding an efficient path and proving local stability are different tasks, but they can share calculations in some control problems. The point is to design the path and the feedback law together, then verify the region where the certificate applies.</p>
    <div class="analogy">

Think of it like: Riding a bicycle at a steady pace. Some corrections that keep you efficient also help you stay balanced, but the claim depends on speed, steering, road conditions, and rider response.