In Layman’s Terms: Tangent Hyperplane Framework
Target Audience: Non-technical readers, practitioners, students
Reading Path: Start with the Tangent-Space Series. The Full Reference Manuscript provides a longer treatment.
The Big Idea in One Sentence
A differentiable model tells us how very small changes affect motion near a chosen reference; we can use that information to plan and adjust a movement, while checking the errors, constraints and connections between its phases.
The derivative is an exact mathematical description of first-order sensitivity. Using it to predict a finite change is an approximation. A nonlinear system does not become linear merely because we look at one instant. Keeping that distinction clear makes the geometric viewpoint useful rather than misleading.
The Core Problem
Imagine driving along a winding mountain road. Your position, speed, steering and road conditions jointly affect what happens next. The road’s direction immediately ahead gives useful information, but continuing straight along that direction will not follow every bend.
A golf swing poses a related modeling problem. Body configuration, segment velocities, club motion and the available forces interact. A change in preparation can alter the state entering the downswing, the torques needed later and the eventual club delivery. A model should help us trace those connections.
In a linear model, responses to perturbations add and scale proportionally. A nonlinear model does not generally have that property over finite changes. Nonlinearity does not itself mean unpredictability: a nonlinear model may give accurate predictions within its validated conditions. Uncertainty in measurements, parameters and external forces is a separate problem.
Control engineering already uses linearization alongside nonlinear optimization, simulation and stability analysis. The tangent viewpoint helps explain those established tools and their limits. It does not replace them with a new guarantee of control.
The Key Insight: Compare Nearby Motions
The Winding-Road Analogy
A tangent line gives a curve’s direction at a point. It agrees with the curve to first order nearby, but it is not a finite straight section of the curve. A tangent plane provides the corresponding picture for a surface.
For motion, the important object is usually a map from the state and input to the rate at which the state changes. State means the variables needed to predict the future: for a simple mechanical model, positions and velocities; for a richer model, muscle activation, controller memory or flexible-shaft states may also matter. Input might mean joint torque, muscle excitation or an actuator command. Those are different choices, with different dynamics.
Linearizing this map asks: If we slightly change the state or input from a reference, how does the predicted rate change? It does not flatten the actual mechanism. Nor is a tangent space of possible state directions the same thing as the graph of the dynamics. The title’s hyperplane picture is easiest to visualize for a scalar map. For a vector-valued dynamics map, a derivative or Jacobian is the more precise general description: it collects the sensitivities of all the output components.
Small Changes Around a Reference
Let \(x\) denote state and \(u\) input. A model supplies the state rate \(\dot{x}=f(x,u)\). Near reference values \(\bar{x},\bar{u}\), its local prediction is
\[ f(\bar{x}+\delta x,\bar{u}+\delta u) =f(\bar{x},\bar{u})+A\delta x+B\delta u+r. \]
Here \(A\) and \(B\) describe first-order sensitivity, and \(r\) is the remainder. The prediction includes the reference rate: in absolute coordinates it is generally affine, meaning a linear response plus an offset. Proportionality concerns small deviations, not a claim that doubling the gas pedal doubles total acceleration.
If the reference itself follows the model, subtracting its rate gives a local equation for nearby motion: \(\delta\dot{x}\approx A(t)\delta x+B(t)\delta u\). The derivative of a smoothly parameterized family of solutions satisfies the corresponding linear equation exactly. This is called the variational equation: it describes how sensitivity evolves along the reference. Finite differences between trajectories still have a remainder. Repeatedly changing the reference does not remove that distinction.
There is a useful stronger statement in some mechanical models. At a fixed configuration and velocity, rigid-body acceleration can depend exactly affinely on applied joint torque when the inertia and any regular constraint equations are fixed. This means a proportional torque contribution plus a baseline contribution from terms such as gravity. It does not make the evolving swing linear, and it does not automatically hold for muscle excitation or through contact changes. A locally linear approximation and an exactly affine input dependence are different claims.
Why Does This Matter?
We Can Use Local Tools
Local models make some design problems easier. For example, LQR chooses feedback for specified linear dynamics and a quadratic objective. Applying it around a nonlinear reference gives a local controller whose actual performance still requires assessment. Its mathematical optimality is relative to its model, cost and assumptions; it is not a promise of globally optimal movement. Tedrake’s LQR treatment develops this connection.
A useful workflow is to choose a feasible reference, compute sensitivities, propose a change, and evaluate that change in the nonlinear model. Measurements then test whether the model adequately represents the real movement. Small modeling residuals do not compensate for an incorrect model of the athlete.
We Can Describe the Error Honestly
Differentiability means the remainder becomes small relative to the perturbation as that perturbation shrinks. It does not by itself promise a quadratic error bound. A continuously differentiable function is called \(C^1\); that regularity is still insufficient for a general quadratic claim.
For example, the dimensionless function \(f(z)=|z|^{3/2}\) is \(C^1\) and has derivative zero at zero. Its linear approximation there is zero, but the error is \(|z|^{3/2}\), not bounded by a fixed multiple of \(z^2\) arbitrarily close to zero. Halving a small positive perturbation reduces this error by a factor \(2^{3/2}\), not four.
A locally Lipschitz derivative means that the derivative itself changes at most proportionally to displacement within a neighborhood. It supplies a quadratic remainder bound; bounded second derivatives on the relevant neighborhood are a familiar sufficient condition. With a bound \(|r|\leq C|\delta|^2\), doubling the perturbation makes that bound four times larger while the same constant applies. The actual error need not grow exactly fourfold: cancellation, direction and higher-order terms matter. For states with different units, scales and a norm (a rule for measuring vector size) must be declared before comparing perturbation sizes.
The size of a state perturbation, a numerical integration time step and an optimization update are also different quantities. A quadratic local Taylor remainder does not automatically imply a quadratic accumulated trajectory error.
Local Errors and Earlier Choices Propagate
Consider the dimensionless model \(\dot{x}=x^2\), starting at \(x(0)=\varepsilon\). Its exact solution is \(x(t)=\varepsilon/(1-\varepsilon t)\) while the denominator remains positive. Around the reference \(x=0\), the linear sensitivity predicts \(x(t)\approx\varepsilon\). Their difference is
\[ \frac{\varepsilon^2t}{1-\varepsilon t}. \]
The sensitivity is correct to first order for fixed finite time as \(\varepsilon\) tends to zero, but the finite prediction error can grow substantially. Frequent recalculation does not by itself establish stability, feasibility or a reliable horizon. We must examine how dynamics amplify or attenuate disturbances and residuals over the interval of interest.
What This Means for a Golf Swing
Phases Are Useful Descriptions, but Remain Connected
A coach may focus attention on the backswing, transition or delivery. That can organize observation and practice, but linearization does not prove that these phases can be fixed independently. Each phase begins with the state produced by the preceding motion. Changes in position, momentum, activation or timing can affect what follows.
Even a perfectly linear system illustrates this connection. Take a dimensionless one-dimensional model with position rate equal to velocity and velocity rate equal to a commanded acceleration. Start at rest at position zero. Apply constant acceleration \(a\) during the first unit of time and \(b\) during the second. Direct integration gives
\[ v(2)=a+b,\qquad q(2)=\tfrac32a+\tfrac12b. \]
The first phase affects both final velocity and final position. Increasing \(a\) and decreasing \(b\) by the same amount preserves final velocity but changes final position. Thus even matching one endpoint quantity does not make two preparations equivalent. This is a mathematical counterexample to phase independence, not a validated model of a golfer or a coaching experiment.
Connect Preparation, Forces and Delivery
For a swing model, a useful question is: How does a specified change in preparation or input alter later club delivery, under stated constraints? The answer must distinguish an immediate effect at an identical complete state from the downstream effect after two motions diverge.
For instance, comparing two torque choices at the same instant and state isolates an instantaneous model response. Letting each choice evolve allows later states and forces to differ. Retuning the later inputs asks a third question: how well can the system compensate? These comparisons answer different causal questions and should not be mixed.
Kinematic constraints, actuator limits and contact conditions restrict which variations are feasible. A geometric direction is not necessarily achievable by the available muscles and contact forces. Clubhead speed alone also does not describe delivery: position, orientation, velocity direction and impact conditions can change together. A model can expose these tradeoffs, while measured outcomes determine whether they matter as predicted.
Segmenting practice may be useful, but its effectiveness requires evidence about people, learning and transfer to the full swing. Tangent geometry supplies neither a universal muscle sequence nor a proof that a coaching intervention works. The appropriate lesson is to study a phase in relation to the movement it inherits and the outcomes it changes.
Illustrative Engineering Applications
A Robot Reaching for a Cup
For a robot arm, inertia, gravity and velocity-dependent forces couple joint motion. A local model can estimate how changed torques influence deviations from a planned reach. A controller can then correct measured errors while respecting joint, torque and collision limits.
This is an illustrative design process, not a reported successful experiment. An arbitrary number of small solves cannot guarantee a successful grasp. Reference feasibility, sensing, contact modeling and robustness remain necessary parts of the assessment.
A Spacecraft Approaching a Dock
Local relative-motion models can help plan an approach, but position, velocity and attitude must be treated according to the chosen model. A thrust force through the center of mass need not produce rotation; an offset force can produce a moment. Attitude can affect the direction of thrust.
A sequence of locally optimized commands alone does not guarantee docking. The design must account for collision avoidance, actuator authority, navigation uncertainty and terminal conditions. This example illustrates a possible modeling strategy; no mission result or certification follows from it.
What Happens at Impact or a Switch?
Smooth sensitivity equations apply within a smooth mode. An impact may reset velocities; contact may change the allowed motion. A perturbation can also change when an event occurs, so joining two smooth models without accounting for the event can give the wrong sensitivity.
A hybrid model must specify its modes, event conditions and reset rules. At grazing contact or a change in event sequence, ordinary differentiability can fail. Shorter steps alone do not repair a missing impact model. Kong and colleagues’ treatment of hybrid sensitivity explains why event timing matters. For golf, interpreting pre-impact motion and predicting the ball’s response therefore require connected but distinct modeling stages.
What the Geometric Viewpoint Contributes
The contribution here is an explanatory organization of established mathematics: choose a reference, identify feasible small variations, propagate their effects and evaluate finite changes. Novel terminology does not establish a new theorem, better controller or effective coaching method.
For students, the viewpoint connects derivatives to motion. For engineers, it clarifies what a local prediction assumes and what must be tested. For golf practitioners, it motivates questions about how preparation, forces, timing and delivery interact. The system remains nonlinear; the value comes from understanding and checking the local information we use to reason about it.
Citation
@online{untitled,
author = {},
title = {In {Layman’s} {Terms:} {Tangent} {Hyperplane} {Framework}},
url = {https://affinedrift.com/articles/tangent-hyperplane-articles/LAYMANS_TERMS_SUMMARY.html},
langid = {en}
}