Passive Stabilization in Parallel Loops

What Needs to Be Stabilized?

A golfer must tolerate variability while delivering an accurate clubhead state at impact. Geometry, muscle activation, tissue elasticity, and feedback all affect that task. The useful question is how these mechanisms share the work of rejecting a specified disturbance, at what energetic and loading cost. Neither maximum stiffness nor maximum compliance is a universal answer.

Bosch’s attractor-fluctuation framework offers a way to discuss robust features of movement (Bosch 2020). Its application to golf is a hypothesis to operationalize. A repeatable impact configuration is not automatically an attractor of an autonomous mechanical system, and an energy minimum is not automatically evidence of reduced muscle activity.

Constraints, Impedance, and Activation

For ideal stationary loop constraints $\phi(q)=0$, let $A=\partial\phi/\partial q$. The equations are \[ \begin{aligned} M(q)\ddot q+h(q,\dot q)&=Bu+A^T\lambda,\\ A\dot q&=0,\\ A\ddot q+\dot A\dot q&=0. \end{aligned} \] The generalized reaction $A^T\lambda$ is determined jointly by geometry, motion, applied forces, and the closure conditions. It depends on actuation even though it is not an independently commanded input. For these ideal stationary constraints, $\dot q^TA^T\lambda=0$. Reactions can transfer energy between connected bodies while doing zero net work on the complete system. Real ground contacts also require friction, unilateral contact, and possible slip or separation. Pelvis–thorax separation is a measured kinematic relation, not by itself an extra rigid loop constraint.

For small perturbations around a fixed operating condition, a local model is \[ \begin{aligned} M\delta\ddot q+D\delta\dot q+K\delta q&=B\delta u+J^T\delta w,\\ Z_q(s)&=Ms^2+Ds+K. \end{aligned} \] Here the local equation uses independent perturbation coordinates; its matrices are the corresponding reduced matrices when rigid loops are present. If $q$ instead retains redundant loop coordinates, retain the linearized constraint and reaction equations. Constraint reduction is detailed below. Here $Z_q$ maps displacement to generalized force. Mechanical impedance is also often defined relative to velocity, in which case it is $Z_q(s)/s$. State the convention. Along a swing these coefficients vary; delayed reflex feedback generally gives a frequency-dependent contribution rather than a constant viscous damper. Neural delay can reduce stability margins.

Intrinsic muscle response begins without waiting for a new neural command. That does not make activated muscle metabolically passive. Existing activation, short-range stiffness, force–length and force–velocity behavior, tendon compliance, and contraction history affect the response. They cannot all be represented by an activation-independent spring (Rack and Westbury 1974). The descending limb of the active force–length curve does not guarantee restoring stiffness, and a shortened muscle does not abruptly lose all force as soon as it is below optimal length.

How Muscle Stiffness Becomes Joint Stiffness

Let muscle length be $l_i(q)$, tension $F_i$, and $r_i=-\partial l_i/\partial q$ be its signed moment-arm vector. The generalized muscle force is $\tau=\sum_i r_iF_i$. For this quasistatic constitutive approximation, take $F_i=F_i(l_i)$ with activation and internal states held fixed, and no independent shortening-velocity variation. The restoring tangent stiffness is \[ \begin{aligned} K_q&=-\frac{\partial\tau}{\partial q}\\ &=\sum_i\left[k_i r_i r_i^T-F_i\frac{\partial r_i}{\partial q}\right],\\ k_i&=\frac{\partial F_i}{\partial l_i}. \end{aligned} \] The first term includes squared moment arms; without them a muscle stiffness in force per length cannot become joint stiffness in torque per angle. The second is a geometric or prestress contribution and need not be positive semidefinite. This local formula needs modification for tendon states, muscle coupling, history-dependent response, and neural feedback. During motion, changing geometry can also change shortening velocity at fixed joint velocity. A force–velocity law then contributes additional derivatives; it cannot be included merely by calling $k_i$ the force–length slope.

Balanced antagonist forces can yield small net torque while both muscles carry large tension. Co-contraction commonly raises intrinsic joint stiffness and loading; it does not define a low-impedance mode. Finite stiffness changes compliance, not the rank of the kinematic Jacobian or the number of exact degrees of freedom. Treating a stiff direction as locked is an approximation that needs a time scale and a tolerated deflection.

At an unloaded static operating point with $K_q\succ0$ and task displacement $\delta y=J\delta q$, compliance is \[ C_y=J K_q^{-1}J^T,\qquad K_y=C_y^{-1} \] when the task compliance is invertible. Only for square nonsingular $J$ does this become $K_y=J^{-T}K_qJ^{-1}$. The familiar $J^TK_yJ$ maps task stiffness back to joint coordinates; it is not the general forward map. External preload adds geometric terms, and rigid constraints require a reduced-coordinate or constrained solve.

For $J=[1\;1]$ and $K_q=\operatorname{diag}(2,3)$ in consistent units, $C_y=1/2+1/3=5/6$. Multiplying both joint stiffnesses by ten divides compliance by ten and leaves the rank of $J$ unchanged. Geometry and stiffness interact without being the same mathematical object.

Loaded Loops: Geometry Enters the Restoring Stiffness

To compute the preload contribution, consider a smooth conservative static equilibrium $q_*$ with fixed activation and fixed spring parameters. Include all conservative applied loads in $U(q)$ and use regular stationary constraints $\phi(q)=0$, with $A$ of full row rank. With the reaction sign used above, $\nabla U(q_*)=A(q_*)^T\lambda_*$. Let the columns of $N$ form a basis of $\ker A(q_*)$, so that $\delta q=N\delta\eta$ to first order. Then \[ \begin{aligned} K_r&=N^TH_LN,\\ H_L&=U_{qq}-\sum_i\lambda_{*,i}\phi_{i,qq}. \end{aligned} \tag{1}\]

All derivatives are evaluated at $q_*$. This is the Hessian of $U-\lambda_*^T\phi$ restricted to admissible directions, with the multiplier held fixed when taking the Hessian. It is the equality-constrained second-variation test (Freund 2004). The minus sign follows from this chapter’s reaction convention; changing the multiplier convention changes both signs.

The constraint-curvature term is essential. For any feasible curve $q(s)$, differentiating $\phi(q(s))=0$ twice gives $Aq''+\{q'^T\phi_{i,qq}q'\}_i=0$. Substitution of $\nabla U=A^T\lambda_*$ into the second derivative of $U(q(s))$ produces the displayed $K_r$. Merely projecting $U_{qq}$ omits the work of the applied preload along the curved feasible path. Ideal reactions still do zero net work; this geometric effect creates no energy.

If $K_r\succ0$, a small additional task load has static compliance $C_y=J_rK_r^{-1}J_r^T$, where $J_r=JN$. Its inverse defines task stiffness only when this compliance is nonsingular. This result concerns local static response, not finite-frequency impedance. A change of tangent basis must transform both $J_r$ and $K_r$; their combined task compliance is unchanged. Nonconservative follower loads, delayed feedback, moving constraints, and a moving swing require their own linearized dynamics, not this potential test.

A Preload Example With Unchanged Degrees of Freedom

Consider a manufactured point constrained to a circle, with Cartesian coordinates $q=(x,y)$ in metres: \[ \begin{aligned} \phi(q)&=\tfrac12(x^2+y^2-R^2)=0,\\ U(q)&=\tfrac12k_x(x-a)^2+\tfrac12k_y y^2. \end{aligned} \tag{2}\]

At $q_*=(R,0)$ choose $N=(0,1)^T$, so $\eta$ measures tangent displacement in metres. Then $\lambda_*=k_x(R-a)/R$ and $K_r=k_y-\lambda_*$. For $R=1\,\mathrm{m}$, $k_x=100\,\mathrm{N/m}$, and $k_y=40\,\mathrm{N/m}$, choosing $a=0.8\,\mathrm{m}$ gives $K_r=20\,\mathrm{N/m}$ and tangent compliance $0.05\,\mathrm{m/N}$. Choosing $a=0.5\,\mathrm{m}$ instead gives $K_r=-10\,\mathrm{N/m}$: the equilibrium is unstable for a positive-mass conservative particle even though the ambient Hessian $\operatorname{diag}(100,40)$ is positive definite. Its inverse is not a stable static compliance measurement. The constraint rank and tangent dimension have not changed.

The exact path $q(\theta)=(R\cos\theta,R\sin\theta)$ checks the result: \[ \left.\frac{d^2U(q(\theta))}{d\theta^2}\right|_{\theta=0} =R^2(k_y-\lambda_*). \tag{3}\]

Dividing the angular energy curvature by $R^2$ gives tangent-displacement stiffness. These numbers illustrate a mechanism, not human tissue parameters or a golf posture recommendation. They show why activation-induced loading and loop geometry must be assessed together rather than ranking postures by muscle stiffness alone.

Energy Minima, Stability, and Attractors

For a conservative mechanical model, a strict potential minimum with positive reduced Hessian supplies restoring forces and local Lyapunov stability under the usual regularity and positive-inertia assumptions. Without dissipation, nearby trajectories can oscillate forever; they need not converge to that minimum. An attracting equilibrium requires an additional dissipation argument. An attractor can also be a periodic orbit or another invariant set, rather than an equilibrium. Address, transition, and impact must not all be called potential wells merely because they are recognizable landmarks.

For $M\ddot e+D\dot e+Ke=0$ with scalar $M>0$, $K>0$, and $D>0$, \[ \begin{aligned} \omega_n&=\sqrt{K/M},\qquad \zeta=\frac{D}{2\sqrt{KM}},\\ \lambda_{1,2}&=\frac{-D\pm\sqrt{D^2-4MK}}{2M}. \end{aligned} \] The equilibrium is asymptotically stable. At $D=0$ it is stable but not attracting. Increasing $K$ with fixed $M,D$ raises natural frequency and lowers damping ratio. Consequently, “more stiffness reduces bandwidth” is not a general law. Disturbance rejection depends on frequency and on which transfer function is being measured.

For a constrained equilibrium, stability uses admissible perturbations and the reduced Hessian of the constrained potential, including preload terms. Along a moving swing, assess the time-varying variational dynamics. Stable eigenvalues of every frozen-time matrix alone do not establish stability of the full time-varying system. A change in stiffness is not evidence of a bifurcation; one must identify the solution family and demonstrate a change of stability.

What Pre-Tension Changes in an Affine Model

A mechanical state includes position and velocity, not configuration alone: $x=(q,v,a,z)$ may also include activation $a$ and internal tissue states $z$. For a specified input and contact mode, a model may take the form \[ \dot x=f(x)+G(x)u. \] If $u$ denotes residual torque around a maintained baseline activation, that baseline remains in the zero-input dynamics. If $u$ is neural excitation, setting it to zero does not instantaneously erase activation or tendon force. If total actuator torque is set to zero, the intervention removes a different quantity. A zero-torque counterfactual must specify which model is used.

Pre-tension changes state, elastic forces, and local tangent dynamics. In a fixed-geometry rigid model with direct torque input, changing a spring does not add a stiffness matrix to $M$ or directly change $M^{-1}B$. Through later changes of posture, constraints, activation, or actuator model, it can change the input-to-task response. This is why same-state acceleration authority and finite-duration control performance must be assessed separately.

A zero-input branch near a target suggests a potentially useful initial state; it does not establish low total effort. Earlier actions created that state, maintained activation may remain, and small errors in difficult directions may require substantial correction. Compare the complete task and preparation cost under matched input limits and measurement uncertainty.

Why a Drift Ratio Is Not an Efficiency Measurement

A ratio of drift magnitude to a specified bounded-input magnitude summarizes model terms in a chosen norm. It is not a work fraction, an energy source fraction, or a measurement of metabolic efficiency. Norms of sums contain cross terms; the norm of one arbitrarily signed sum of actuator columns need not equal maximum available control. State the admissible input set and use its actual optimization when making an authority claim.

Magnitude also omits direction. Fast motion away from the target can have a large drift ratio and poor performance. Use task error, sensitivity over the remaining horizon, and explicit work or energetic costs to evaluate whether the drift is helpful.

Stiffness Modulation Has an Energy Budget

For a spring with $e=q-q_0(t)$ and storage $V_s=\tfrac12k(t)e^2$, \[ \dot V_s=ke\dot q-ke\dot q_0+\tfrac12\dot k e^2. \] Changing stiffness or rest position can exchange energy with the mechanism. A scheduled stiffness change is not automatically a passive operation. Report where this energy goes in the actuator or tissue model.

Isometric activated muscle can consume metabolic energy at zero shortening velocity. Thus $(F_{\rm ago}+F_{\rm ant})v$ is not a general metabolic-power model. Joint power is $\tau^T\dot q$; individual muscle work, tendon storage, activation and maintenance heat, and joint work differ. Net torque can conceal large opposing forces and energetic expenditure.

Fascia and other connective tissues can carry load and store energy, but a tensegrity analogy does not quantify their contribution to club speed or prove a universal low-cost posture. A model needs tissue properties, load paths, boundary conditions, and validation. Likewise, reduced modeled loading is not by itself clinical evidence of reduced injury risk.

Testing Coordination With the Uncontrolled Manifold

For task output $y=h(x)$, the nonlinear task-equivalent set is $\{x:h(x)=y^*\}$. At a regular point its tangent is $\ker J$, where $J=\partial h/\partial x$. This is not the stationary set of a scalar cost. If the task is clubhead speed, the state must include velocities: joint angles alone do not determine speed. The UCM method originated in task-specific coordination analysis (Scholz and Schöner 1999); that study examined sit-to-stand, not golf. Its framework does not by itself validate clubhead-speed regulation.

Declare characteristic scales before combining position, velocity and other state components. For constant nonsingular $S$ and $\xi=S^{-1}(x-x_*)$, use $J_\xi=J_xS$ and $\Sigma_\xi=S^{-1}\Sigma_xS^{-T}$. In the following formulas, $J$ and $\Sigma$ mean these consistently scaled quantities. With the Euclidean metric in $\xi$, define $P_\parallel=I-J^+J$, $P_\perp=J^+J$, and sample covariance $\Sigma$. With $r=\operatorname{rank}J$ and $0<r<n$, variance per dimension is \[ \begin{aligned} V_\parallel&=\frac{\operatorname{tr}(P_\parallel\Sigma)}{n-r},\\ V_\perp&=\frac{\operatorname{tr}(P_\perp\Sigma)}{r}. \end{aligned} \] Parallel means within the task-equivalent tangent; perpendicular means changing the task to first order. Test $V_\parallel>V_\perp$ with an appropriate subject-level statistical design. Coordinate scaling, measurement noise, phase alignment, and nonlinear curvature affect the result. Such covariance structure supports a task-stabilization interpretation but does not distinguish passive mechanics from feedback or prove a neural algorithm.

An Experimental Program for the Golf Swing

The earlier chapter asserted a universal elite-golfer stiffness pulse, minimal late-downswing activation, and lower metabolic cost without identifying studies that established those claims. Treat these as proposed comparisons, not results. No numerical co-contraction schedule is prescribed here.

Define the task: impact location, clubhead velocity, face orientation, or their joint distribution. Specify the tested club, intent, and skill groups.
Synchronize kinematics, external forces, and relevant EMG. EMG is an electrical measurement; it is not a direct force or stiffness measurement.
Identify impedance using controlled perturbations and a model separating inertia, damping, stiffness, delays, and measurement error. A dynamic torque/displacement ratio alone conflates these contributions.
Compare preparation, swing, and recovery costs. A baseline posture measurement does not determine stiffness throughout a transient swing.
Distinguish ordinary force-plate loading from ball-contact transmission. Do not assume a universal ground-force spike exactly at impact without synchronized measurements and adequate bandwidth.
Test alternative explanations: geometry, starting speed, net torque, passive tissue loading, activation, and predictive or delayed feedback. Report unfavorable outcomes as well as favorable ones.

Optimizing a Stiffness Strategy

Let $\alpha$ parameterize a specified activation or stiffness mechanism and let $e_k=h_{\rm task}(x_k)-y_k^*$ be the declared task error. For step size $h>0$, use a piecewise-linear parameter schedule with $\nu_k=(\alpha_{k+1}-\alpha_k)/h$, or an explicit rate input with its own dynamics. A finite-horizon model comparison can minimize \[ \begin{aligned} J={}&\ell_T(x_N)+\sum_{k=0}^{N-1}h\,L_k,\\ L_k={}&e_k^TQe_k+u_k^TRu_k\\ &+c_\alpha(x_k,\alpha_k)+c_{\rm rate}(\nu_k). \end{aligned} \] subject to state dynamics, contact and loop constraints, input bounds, and admissible activation/stiffness rates. Every running-cost term carries the time step. The quantities $c_\alpha$ and $c_{\rm rate}$ are stated penalties unless calibrated as physical energetic costs.

Use a state containing velocities and activation dynamics where required. LQR applies to a linearization with a quadratic objective; a finite horizon uses a differential or discrete Riccati recursion, not automatically an algebraic steady-state solution. Joint optimization of state, inputs, and stiffness is generally nonlinear and constrained. Check feasibility, derivative accuracy, integration convergence, stationarity, and sensitivity to starting guesses. Penalty terms alone do not exactly enforce loop closure.

The optimal profile is an output of this specified comparison. It need not peak at transition or decline with increasing load. Those are hypotheses to test, not conditions to build into the answer while calling it a prediction.

How the Mechanisms Fit Together

Geometry directs forces and couples joints. Activation changes force capacity and local response. Elastic tissues store and return part of earlier work. Feedback and prediction manage deviations that mechanics alone does not reject. Impact precision selects which deviations matter. The technical argument for coordination is strongest when these contributions are computed in one consistent model and tested with measurements, including the cost of preparing the state. Apparent effortlessness alone cannot identify any of them.

Exercises

For $M=0.02$, $D=0.5$, and $K=100$ in rotational SI units, compute natural frequency, damping ratio, and static deflection under a $5\,\mathrm{N\,m}$ torque. Double $K$ and explain which quantities change.
Show that the undamped spring conserves energy and therefore cannot attract all nearby nonzero-energy states to its equilibrium.
Verify the task-compliance example and preserve the Jacobian rank while scaling stiffness. Derive the circle example from the exact constrained energy and explain why projecting the ambient Hessian gives the wrong answer.
Derive the variable-spring energy identity, including moving rest length. Identify the terms an actuator must supply or absorb.
For $J=[1\;1]$ and a specified covariance, compute both UCM variances per dimension. Repeat after rescaling one coordinate and discuss the metric.
Design a perturbation study that can distinguish intrinsic stiffness from delayed feedback without interpreting EMG amplitude as force.

On This Site

Passive Distributed Control: Frans Bosch Framework — Self-organization, impedance, and elastic stabilization in loops
Nonlinear Control Insights — Stability, passivity, and underactuated coordination

Cite this page

References

Bosch, Frans. 2020. Anatomy of Agility: Movement Analysis in Sport. 20/10 Publishers.

Freund, Robert M. 2004. Optimality Conditions for Constrained Optimization Problems. MIT OpenCourseWare, 15.084J Nonlinear Programming. https://ocw.mit.edu/courses/15-084j-nonlinear-programming-spring-2004/21073fc1bb9e5f68b3319bec0895683f_lec6_constr_opt.pdf.

Rack, P. M. H., and D. R. Westbury. 1974. “The Short Range Stiffness of Active Mammalian Muscle and Its Effect on Mechanical Properties.” The Journal of Physiology 240 (2): 331–50. https://doi.org/10.1113/jphysiol.1974.sp010613.

Scholz, John P., and Gregor Schöner. 1999. “The Uncontrolled Manifold Concept: Identifying Control Variables for a Functional Task.” Experimental Brain Research 126 (3): 289–306. https://doi.org/10.1007/s002210050738.

Citation

BibTeX citation: Download .bib

@online{untitled,
  author = {},
  url = {https://affinedrift.com/articles/The_Physics_of_Golf/quarto/ch09b_passive_stabilization.html},
  langid = {en}
}

For attribution, please cite this work as:

n.d. https://affinedrift.com/articles/The_Physics_of_Golf/quarto/ch09b_passive_stabilization.html.

--- description: "How geometry, activation, elasticity, and feedback jointly affect swing stability, with explicit mechanical assumptions and testable hypotheses." --- # Passive Stabilization in Parallel Loops {#sec-09b_passive_stabilization} ## What Needs to Be Stabilized? A golfer must tolerate variability while delivering an accurate clubhead state at impact. Geometry, muscle activation, tissue elasticity, and feedback all affect that task. The useful question is how these mechanisms share the work of rejecting a specified disturbance, at what energetic and loading cost. Neither maximum stiffness nor maximum compliance is a universal answer. Bosch's attractor-fluctuation framework offers a way to discuss robust features of movement [@bosch2020agility]. Its application to golf is a hypothesis to operationalize. A repeatable impact configuration is not automatically an attractor of an autonomous mechanical system, and an energy minimum is not automatically evidence of reduced muscle activity. ## Constraints, Impedance, and Activation For ideal stationary loop constraints $\phi(q)=0$, let $A=\partial\phi/\partial q$. The equations are $$ \begin{aligned} M(q)\ddot q+h(q,\dot q)&=Bu+A^T\lambda,\\ A\dot q&=0,\\ A\ddot q+\dot A\dot q&=0. \end{aligned} $$ The generalized reaction $A^T\lambda$ is determined jointly by geometry, motion, applied forces, and the closure conditions. It depends on actuation even though it is not an independently commanded input. For these ideal stationary constraints, $\dot q^TA^T\lambda=0$. Reactions can transfer energy between connected bodies while doing zero net work on the complete system. Real ground contacts also require friction, unilateral contact, and possible slip or separation. Pelvis--thorax separation is a measured kinematic relation, not by itself an extra rigid loop constraint. For small perturbations around a fixed operating condition, a local model is $$ \begin{aligned} M\delta\ddot q+D\delta\dot q+K\delta q&=B\delta u+J^T\delta w,\\ Z_q(s)&=Ms^2+Ds+K. \end{aligned} $$ Here the local equation uses independent perturbation coordinates; its matrices are the corresponding reduced matrices when rigid loops are present. If $q$ instead retains redundant loop coordinates, retain the linearized constraint and reaction equations. Constraint reduction is detailed below. Here $Z_q$ maps displacement to generalized force. Mechanical impedance is also often defined relative to velocity, in which case it is $Z_q(s)/s$. State the convention. Along a swing these coefficients vary; delayed reflex feedback generally gives a frequency-dependent contribution rather than a constant viscous damper. Neural delay can reduce stability margins. Intrinsic muscle response begins without waiting for a new neural command. That does not make activated muscle metabolically passive. Existing activation, short-range stiffness, force--length and force--velocity behavior, tendon compliance, and contraction history affect the response. They cannot all be represented by an activation-independent spring [@rack1974short]. The descending limb of the active force--length curve does not guarantee restoring stiffness, and a shortened muscle does not abruptly lose all force as soon as it is below optimal length. ## How Muscle Stiffness Becomes Joint Stiffness Let muscle length be $l_i(q)$, tension $F_i$, and $r_i=-\partial l_i/\partial q$ be its signed moment-arm vector. The generalized muscle force is $\tau=\sum_i r_iF_i$. For this quasistatic constitutive approximation, take $F_i=F_i(l_i)$ with activation and internal states held fixed, and no independent shortening-velocity variation. The restoring tangent stiffness is $$ \begin{aligned} K_q&=-\frac{\partial\tau}{\partial q}\\ &=\sum_i\left[k_i r_i r_i^T-F_i\frac{\partial r_i}{\partial q}\right],\\ k_i&=\frac{\partial F_i}{\partial l_i}. \end{aligned} $$ The first term includes squared moment arms; without them a muscle stiffness in force per length cannot become joint stiffness in torque per angle. The second is a geometric or prestress contribution and need not be positive semidefinite. This local formula needs modification for tendon states, muscle coupling, history-dependent response, and neural feedback. During motion, changing geometry can also change shortening velocity at fixed joint velocity. A force--velocity law then contributes additional derivatives; it cannot be included merely by calling $k_i$ the force--length slope. Balanced antagonist forces can yield small net torque while both muscles carry large tension. Co-contraction commonly raises intrinsic joint stiffness and loading; it does not define a low-impedance mode. Finite stiffness changes compliance, not the rank of the kinematic Jacobian or the number of exact degrees of freedom. Treating a stiff direction as locked is an approximation that needs a time scale and a tolerated deflection. At an unloaded static operating point with $K_q\succ0$ and task displacement $\delta y=J\delta q$, compliance is $$ C_y=J K_q^{-1}J^T,\qquad K_y=C_y^{-1} $$ when the task compliance is invertible. Only for square nonsingular $J$ does this become $K_y=J^{-T}K_qJ^{-1}$. The familiar $J^TK_yJ$ maps task stiffness back to joint coordinates; it is not the general forward map. External preload adds geometric terms, and rigid constraints require a reduced-coordinate or constrained solve. For $J=[1\;1]$ and $K_q=\operatorname{diag}(2,3)$ in consistent units, $C_y=1/2+1/3=5/6$. Multiplying both joint stiffnesses by ten divides compliance by ten and leaves the rank of $J$ unchanged. Geometry and stiffness interact without being the same mathematical object. ### Loaded Loops: Geometry Enters the Restoring Stiffness To compute the preload contribution, consider a smooth conservative static equilibrium $q_*$ with fixed activation and fixed spring parameters. Include all conservative applied loads in $U(q)$ and use regular stationary constraints $\phi(q)=0$, with $A$ of full row rank. With the reaction sign used above, $\nabla U(q_*)=A(q_*)^T\lambda_*$. Let the columns of $N$ form a basis of $\ker A(q_*)$, so that $\delta q=N\delta\eta$ to first order. Then $$ \begin{aligned} K_r&=N^TH_LN,\\ H_L&=U_{qq}-\sum_i\lambda_{*,i}\phi_{i,qq}. \end{aligned} $$ {#eq-loaded-loop-stiffness} All derivatives are evaluated at $q_*$. This is the Hessian of $U-\lambda_*^T\phi$ restricted to admissible directions, with the multiplier held fixed when taking the Hessian. It is the equality-constrained second-variation test [@Freund2004ConstrainedOptimality]. The minus sign follows from this chapter's reaction convention; changing the multiplier convention changes both signs. The constraint-curvature term is essential. For any feasible curve $q(s)$, differentiating $\phi(q(s))=0$ twice gives $Aq''+\{q'^T\phi_{i,qq}q'\}_i=0$. Substitution of $\nabla U=A^T\lambda_*$ into the second derivative of $U(q(s))$ produces the displayed $K_r$. Merely projecting $U_{qq}$ omits the work of the applied preload along the curved feasible path. Ideal reactions still do zero net work; this geometric effect creates no energy. If $K_r\succ0$, a small additional task load has static compliance $C_y=J_rK_r^{-1}J_r^T$, where $J_r=JN$. Its inverse defines task stiffness only when this compliance is nonsingular. This result concerns local static response, not finite-frequency impedance. A change of tangent basis must transform both $J_r$ and $K_r$; their combined task compliance is unchanged. Nonconservative follower loads, delayed feedback, moving constraints, and a moving swing require their own linearized dynamics, not this potential test. ### A Preload Example With Unchanged Degrees of Freedom Consider a manufactured point constrained to a circle, with Cartesian coordinates $q=(x,y)$ in metres: $$ \begin{aligned} \phi(q)&=\tfrac12(x^2+y^2-R^2)=0,\\ U(q)&=\tfrac12k_x(x-a)^2+\tfrac12k_y y^2. \end{aligned} $$ {#eq-loaded-loop-example} At $q_*=(R,0)$ choose $N=(0,1)^T$, so $\eta$ measures tangent displacement in metres. Then $\lambda_*=k_x(R-a)/R$ and $K_r=k_y-\lambda_*$. For $R=1\,\mathrm{m}$, $k_x=100\,\mathrm{N/m}$, and $k_y=40\,\mathrm{N/m}$, choosing $a=0.8\,\mathrm{m}$ gives $K_r=20\,\mathrm{N/m}$ and tangent compliance $0.05\,\mathrm{m/N}$. Choosing $a=0.5\,\mathrm{m}$ instead gives $K_r=-10\,\mathrm{N/m}$: the equilibrium is unstable for a positive-mass conservative particle even though the ambient Hessian $\operatorname{diag}(100,40)$ is positive definite. Its inverse is not a stable static compliance measurement. The constraint rank and tangent dimension have not changed. The exact path $q(\theta)=(R\cos\theta,R\sin\theta)$ checks the result: $$ \left.\frac{d^2U(q(\theta))}{d\theta^2}\right|_{\theta=0} =R^2(k_y-\lambda_*). $$ {#eq-loaded-loop-curve} Dividing the angular energy curvature by $R^2$ gives tangent-displacement stiffness. These numbers illustrate a mechanism, not human tissue parameters or a golf posture recommendation. They show why activation-induced loading and loop geometry must be assessed together rather than ranking postures by muscle stiffness alone. ## Energy Minima, Stability, and Attractors For a conservative mechanical model, a strict potential minimum with positive reduced Hessian supplies restoring forces and local Lyapunov stability under the usual regularity and positive-inertia assumptions. Without dissipation, nearby trajectories can oscillate forever; they need not converge to that minimum. An attracting equilibrium requires an additional dissipation argument. An attractor can also be a periodic orbit or another invariant set, rather than an equilibrium. Address, transition, and impact must not all be called potential wells merely because they are recognizable landmarks. For $M\ddot e+D\dot e+Ke=0$ with scalar $M>0$, $K>0$, and $D>0$, $$ \begin{aligned} \omega_n&=\sqrt{K/M},\qquad \zeta=\frac{D}{2\sqrt{KM}},\\ \lambda_{1,2}&=\frac{-D\pm\sqrt{D^2-4MK}}{2M}. \end{aligned} $$ The equilibrium is asymptotically stable. At $D=0$ it is stable but not attracting. Increasing $K$ with fixed $M,D$ raises natural frequency and lowers damping ratio. Consequently, "more stiffness reduces bandwidth" is not a general law. Disturbance rejection depends on frequency and on which transfer function is being measured. For a constrained equilibrium, stability uses admissible perturbations and the reduced Hessian of the constrained potential, including preload terms. Along a moving swing, assess the time-varying variational dynamics. Stable eigenvalues of every frozen-time matrix alone do not establish stability of the full time-varying system. A change in stiffness is not evidence of a bifurcation; one must identify the solution family and demonstrate a change of stability. ## What Pre-Tension Changes in an Affine Model A mechanical state includes position and velocity, not configuration alone: $x=(q,v,a,z)$ may also include activation $a$ and internal tissue states $z$. For a specified input and contact mode, a model may take the form $$ \dot x=f(x)+G(x)u. $$ If $u$ denotes residual torque around a maintained baseline activation, that baseline remains in the zero-input dynamics. If $u$ is neural excitation, setting it to zero does not instantaneously erase activation or tendon force. If total actuator torque is set to zero, the intervention removes a different quantity. A zero-torque counterfactual must specify which model is used. Pre-tension changes state, elastic forces, and local tangent dynamics. In a fixed-geometry rigid model with direct torque input, changing a spring does not add a stiffness matrix to $M$ or directly change $M^{-1}B$. Through later changes of posture, constraints, activation, or actuator model, it can change the input-to-task response. This is why same-state acceleration authority and finite-duration control performance must be assessed separately. A zero-input branch near a target suggests a potentially useful initial state; it does not establish low total effort. Earlier actions created that state, maintained activation may remain, and small errors in difficult directions may require substantial correction. Compare the complete task and preparation cost under matched input limits and measurement uncertainty. ### Why a Drift Ratio Is Not an Efficiency Measurement A ratio of drift magnitude to a specified bounded-input magnitude summarizes model terms in a chosen norm. It is not a work fraction, an energy source fraction, or a measurement of metabolic efficiency. Norms of sums contain cross terms; the norm of one arbitrarily signed sum of actuator columns need not equal maximum available control. State the admissible input set and use its actual optimization when making an authority claim. Magnitude also omits direction. Fast motion away from the target can have a large drift ratio and poor performance. Use task error, sensitivity over the remaining horizon, and explicit work or energetic costs to evaluate whether the drift is helpful. ## Stiffness Modulation Has an Energy Budget For a spring with $e=q-q_0(t)$ and storage $V_s=\tfrac12k(t)e^2$, $$ \dot V_s=ke\dot q-ke\dot q_0+\tfrac12\dot k e^2. $$ Changing stiffness or rest position can exchange energy with the mechanism. A scheduled stiffness change is not automatically a passive operation. Report where this energy goes in the actuator or tissue model. Isometric activated muscle can consume metabolic energy at zero shortening velocity. Thus $(F_{\rm ago}+F_{\rm ant})v$ is not a general metabolic-power model. Joint power is $\tau^T\dot q$; individual muscle work, tendon storage, activation and maintenance heat, and joint work differ. Net torque can conceal large opposing forces and energetic expenditure. Fascia and other connective tissues can carry load and store energy, but a tensegrity analogy does not quantify their contribution to club speed or prove a universal low-cost posture. A model needs tissue properties, load paths, boundary conditions, and validation. Likewise, reduced modeled loading is not by itself clinical evidence of reduced injury risk. ## Testing Coordination With the Uncontrolled Manifold For task output $y=h(x)$, the nonlinear task-equivalent set is $\{x:h(x)=y^*\}$. At a regular point its tangent is $\ker J$, where $J=\partial h/\partial x$. This is not the stationary set of a scalar cost. If the task is clubhead speed, the state must include velocities: joint angles alone do not determine speed. The UCM method originated in task-specific coordination analysis [@Scholz1999]; that study examined sit-to-stand, not golf. Its framework does not by itself validate clubhead-speed regulation. Declare characteristic scales before combining position, velocity and other state components. For constant nonsingular $S$ and $\xi=S^{-1}(x-x_*)$, use $J_\xi=J_xS$ and $\Sigma_\xi=S^{-1}\Sigma_xS^{-T}$. In the following formulas, $J$ and $\Sigma$ mean these consistently scaled quantities. With the Euclidean metric in $\xi$, define $P_\parallel=I-J^+J$, $P_\perp=J^+J$, and sample covariance $\Sigma$. With $r=\operatorname{rank}J$ and $0<r<n$, variance per dimension is $$ \begin{aligned} V_\parallel&=\frac{\operatorname{tr}(P_\parallel\Sigma)}{n-r},\\ V_\perp&=\frac{\operatorname{tr}(P_\perp\Sigma)}{r}. \end{aligned} $$ Parallel means within the task-equivalent tangent; perpendicular means changing the task to first order. Test $V_\parallel>V_\perp$ with an appropriate subject-level statistical design. Coordinate scaling, measurement noise, phase alignment, and nonlinear curvature affect the result. Such covariance structure supports a task-stabilization interpretation but does not distinguish passive mechanics from feedback or prove a neural algorithm. ## An Experimental Program for the Golf Swing The earlier chapter asserted a universal elite-golfer stiffness pulse, minimal late-downswing activation, and lower metabolic cost without identifying studies that established those claims. Treat these as proposed comparisons, not results. No numerical co-contraction schedule is prescribed here. 1. Define the task: impact location, clubhead velocity, face orientation, or their joint distribution. Specify the tested club, intent, and skill groups. 1. Synchronize kinematics, external forces, and relevant EMG. EMG is an electrical measurement; it is not a direct force or stiffness measurement. 1. Identify impedance using controlled perturbations and a model separating inertia, damping, stiffness, delays, and measurement error. A dynamic torque/displacement ratio alone conflates these contributions. 1. Compare preparation, swing, and recovery costs. A baseline posture measurement does not determine stiffness throughout a transient swing. 1. Distinguish ordinary force-plate loading from ball-contact transmission. Do not assume a universal ground-force spike exactly at impact without synchronized measurements and adequate bandwidth. 1. Test alternative explanations: geometry, starting speed, net torque, passive tissue loading, activation, and predictive or delayed feedback. Report unfavorable outcomes as well as favorable ones. ## Optimizing a Stiffness Strategy Let $\alpha$ parameterize a specified activation or stiffness mechanism and let $e_k=h_{\rm task}(x_k)-y_k^*$ be the declared task error. For step size $h>0$, use a piecewise-linear parameter schedule with $\nu_k=(\alpha_{k+1}-\alpha_k)/h$, or an explicit rate input with its own dynamics. A finite-horizon model comparison can minimize $$ \begin{aligned} J={}&\ell_T(x_N)+\sum_{k=0}^{N-1}h\,L_k,\\ L_k={}&e_k^TQe_k+u_k^TRu_k\\ &+c_\alpha(x_k,\alpha_k)+c_{\rm rate}(\nu_k). \end{aligned} $$ subject to state dynamics, contact and loop constraints, input bounds, and admissible activation/stiffness rates. Every running-cost term carries the time step. The quantities $c_\alpha$ and $c_{\rm rate}$ are stated penalties unless calibrated as physical energetic costs. Use a state containing velocities and activation dynamics where required. LQR applies to a linearization with a quadratic objective; a finite horizon uses a differential or discrete Riccati recursion, not automatically an algebraic steady-state solution. Joint optimization of state, inputs, and stiffness is generally nonlinear and constrained. Check feasibility, derivative accuracy, integration convergence, stationarity, and sensitivity to starting guesses. Penalty terms alone do not exactly enforce loop closure. The optimal profile is an output of this specified comparison. It need not peak at transition or decline with increasing load. Those are hypotheses to test, not conditions to build into the answer while calling it a prediction. ## How the Mechanisms Fit Together Geometry directs forces and couples joints. Activation changes force capacity and local response. Elastic tissues store and return part of earlier work. Feedback and prediction manage deviations that mechanics alone does not reject. Impact precision selects which deviations matter. The technical argument for coordination is strongest when these contributions are computed in one consistent model and tested with measurements, including the cost of preparing the state. Apparent effortlessness alone cannot identify any of them. ## Exercises 1. For $M=0.02$, $D=0.5$, and $K=100$ in rotational SI units, compute natural frequency, damping ratio, and static deflection under a $5\,\mathrm{N\,m}$ torque. Double $K$ and explain which quantities change. 1. Show that the undamped spring conserves energy and therefore cannot attract all nearby nonzero-energy states to its equilibrium. 1. Verify the task-compliance example and preserve the Jacobian rank while scaling stiffness. Derive the circle example from the exact constrained energy and explain why projecting the ambient Hessian gives the wrong answer. 1. Derive the variable-spring energy identity, including moving rest length. Identify the terms an actuator must supply or absorb. 1. For $J=[1\;1]$ and a specified covariance, compute both UCM variances per dimension. Repeat after rescaling one coordinate and discuss the metric. 1. Design a perturbation study that can distinguish intrinsic stiffness from delayed feedback without interpreting EMG amplitude as force. ::: {.callout-note} ## On This Site - **[Passive Distributed Control: Frans Bosch Framework](../../passive-distributed-control.html)** — Self-organization, impedance, and elastic stabilization in loops - **[Nonlinear Control Insights](../../nonlinear-control-insights.html)** — Stability, passivity, and underactuated coordination :::