Tangent Hyperplanes V: Contraction
Why Contraction Belongs in the Series
Parts 1-4 established the local geometry: tangent spaces carry exact first-order information, variational dynamics propagate perturbations along a reference trajectory, and finite perturbations create residuals whose size depends on curvature. Contraction theory asks a complementary question:
Can we prove that nearby trajectories forget their initial separation?
If the answer is yes, local tangent-space perturbations are not merely useful for computing a feedback law. They also become stability certificates.
This compact part distills the corrected core of the longer reference article Contraction Theory Meets Tangent Spaces. The source article remains available for derivations, examples, code sketches, and its companion critique. This page is the canonical reading-path version.
The Differential Object
For a smooth autonomous system
\[ \dot{x} = f(x), \]
the first variation satisfies
\[ \delta \dot{x} = A(x)\delta x, \qquad A(x) = \frac{\partial f}{\partial x}(x). \]
The perturbation vector \(\delta x\) lives in the tangent space. Contraction theory studies whether the norm of that perturbation decreases for every nearby trajectory in a region.
With a positive-definite metric \(M(x)\), define the differential length
\[ V(x, \delta x) = \delta x^\top M(x)\delta x. \]
The system is contracting with rate \(\lambda > 0\) in that metric if
\[ \dot{M} + A^\top M + M A \preceq -2\lambda M. \]
This is the tangent-space version of exponential forgetting: every infinitesimal separation shrinks at a certified rate.
A contraction inequality is only as broad as the region, metric, smoothness, and mode assumptions used to verify it. It does not make a nonlinear plant globally linear, and it does not remove the finite-scale residuals from Part 4. It certifies that the differential dynamics in a stated region are shrinking.
Stability-Optimality Duality
The useful bridge to control is that the same tangent objects appear in optimal control. LQR and local DDP/iLQR solve Riccati equations over the variational system
\[ \delta\dot{x} = A(t)\delta x + B(t)\delta u. \]
The value-function Hessian \(S(t)\) acts like a local metric on perturbations. When feedback makes
\[ A_{\text{cl}}(t) = A(t) - B(t)K(t) \]
contracting in that metric, optimality and stability are no longer separate stories. The controller is optimizing a local quadratic model while also shaping how perturbations decay.
The compact claim is conservative: Riccati and contraction analyses both live on the tangent bundle, and in favorable cases the Riccati matrix supplies a contraction-like metric for the closed-loop variational dynamics. The stronger claim that any optimal controller is automatically a useful contraction certificate needs hypotheses about controllability, costs, regularity, and the region of validity.
Contraction-DDP as a Design Pattern
A contraction-aware DDP loop can be read as a disciplined version of the standard residual-reduction loop:
- Linearize along the current trajectory.
- Solve the local quadratic control problem.
- Check whether the closed-loop tangent dynamics shrink perturbations in the chosen metric.
- Penalize or reject updates that improve cost while damaging the stability certificate.
- Re-integrate the nonlinear system and repeat.
The key engineering benefit is explicitness. Instead of hoping that the local controller is stable because the rollout improved, the method asks for a differential certificate at each step.
Coordinate-Free View
The metric \(M(x)\) should not be confused with a fixed Euclidean norm. In mechanical systems, natural coordinates, task-space coordinates, and configuration-space coordinates can imply different metrics. Pullback metrics make this explicit: a task-space stability objective can be transported back to joint space through the relevant Jacobian.
This is why contraction is a good fit for tangent-space language. The object being certified is not “distance” in a casual plotting sense; it is the length of a tangent vector under a chosen metric.
Biomechanics and Golf
The advanced manuscript uses golf-swing stability margins as a motivating example. The compact reading-path version keeps the claim modest:
- a swing model can define a nominal trajectory and local variational dynamics;
- feedback, muscle synergies, or impedance-like effects can be interpreted as shaping a local metric;
- stability margins are meaningful only when the model, measured states, and perturbation region are stated explicitly.
The contraction interpretation is a hypothesis generator for biomechanics, not a substitute for measurement. Human movement includes soft tissue, delays, contact changes, neural variability, and unmodeled actuation. A contraction margin computed from a simplified model should be reported as a model-local certificate, not as proof that the athlete’s nervous system optimizes that metric.
Reading Forward
Contraction theory is the smooth shrinking-perturbation story. Part 6 handles the case where the smoothness assumptions break at impacts, contacts, or controller mode switches.