Induced Acceleration Analysis: Superposition in the Biomechanics Literature
The preceding chapter develops a local fact: at one state, a fixed mechanical operator maps generalized forces to acceleration increments. Biomechanics uses this fact in induced acceleration analysis (IAA). The connection helps explain why a shoulder moment can affect wrist acceleration and why a large joint moment need not increase clubhead speed. This chapter connects the literature to the mathematics, including the conditions under which two apparently similar force ledgers answer the same question.
An induced-acceleration attribution is a term ledger for a declared model at a declared state. It is not a unique causal history. A publishable result must identify all of the following:
- Model and revision: segment and actuator definitions, parameter set, generalized coordinates, reference frame, output point or task metric, and the state at which the mass matrix is evaluated.
- Computation: engine, solver, and revision; constraint handling; contact model and active contact mode; integration or frozen-state procedure; and numerical tolerance plus closure error.
- Attribution convention: the complete force partition, including sign conventions, passive and applied terms, constraint reactions, and residual treatment. Reassigning a residual between two terms changes their reported contributions while leaving their sum unchanged.
- Identifiability contract: the measurements or assumptions that identify each generalized-force term and the output projection. An algebraic term contribution does not by itself identify anatomical source, neural intent, necessity, sufficiency, or intervention effect.
If any field is absent, label the result unsupported or unqualified. A cross-engine comparison must report an unavailable state as unsupported or unqualified; agreement on one output, or silence about an unsupported state, does not establish solver parity.
Counterexamples to Unique Attribution
- Coordinate representation: under a constant change of generalized coordinates \(q=Tz\), the same dynamics are represented by \(M_z=T^\mathsf{T}M_qT\) and \(Q_z=T^\mathsf{T}Q_q\). The component values of \(\ddot z=M_z^{-1}Q_z\) generally differ from those of \(\ddot q=M_q^{-1}Q_q\), even though \(T\ddot z=\ddot q\). A generalized- acceleration component is therefore not coordinate-invariant causal truth.
- Force partition: if \(Q=Q_A+Q_B\), then for any residual allocation \(r\), \(Q=(Q_A+r)+(Q_B-r)\). The reported terms \(M^{-1}Q_A\) and \(M^{-1}Q_B\) change, while the total acceleration does not. The partition must be declared and justified before a term is interpreted.
An intervention claim requires a separate governed counterfactual that re-solves constraints and contact and states what is held fixed. The frozen- state algebra alone supports neither a unique cause nor an anatomical or behavioral prescription.
Historical Context and Motivation
The review by (Zajac and Gordon 1989) is an early reference for relating muscle forces to motion in multi-articular systems. A muscle produces moments where it crosses joints; the coupled mechanics can transmit an acceleration response to other coordinates. That response need not be nonzero at every joint or in every posture. The force map, inertial geometry, constraints and selected output together determine it.
The literature includes walking (Zajac et al. 2002, 2003; Neptune et al. 2001; Kepple et al. 1997), throwing (Hirashima et al. 2007, 2008; Hirashima and Ohtsuki 2008), postural coupling (Challis 2011), and clinical movement models (Riley and Kerrigan 1999; Caruthers et al. 2016). These applications differ in how forces are estimated, how the body is supported, and what acceleration is reported. They demonstrate uses of a common calculation, not interchangeable human models or a single established control strategy.
Inverse dynamics supplies net generalized forces only after measurements, external loads and model assumptions have been specified. To assign terms to individual muscles requires additional force estimates, such as an optimization solution or an activation-driven model. IAA applies the forward operator to those terms. It does not supply the missing force-identification evidence.
The Formal Framework
Equations of Motion and the Forward Problem
Use independent coordinates \(q\) and velocities \(v=\dot q\) in a regular local chart. For an unconstrained model write
\[ M(q)\ddot q+C(q,v)v+g(q)=Q_u+Q_p+Q_e+Q_r. \tag{1}\]
Here \(g\) is the gravitational potential gradient; \(Q_u\) is applied actuation, \(Q_p\) passive constitutive force, \(Q_e\) other specified external load, and \(Q_r\) a declared residual. Each retained force belongs in exactly one category. The mass matrix is positive definite when the kinetic-energy metric is regular in these coordinates. Underactuation does not by itself make it singular. Constraints and non-coordinate velocities require the extensions below.
The acceleration solve is
\[ \ddot q=M^{-1}(Q_u+Q_p+Q_e+Q_r-Cv-g). \tag{2}\]
Define \(Q_v=-Cv\) and \(Q_g=-g\) without changing the meaning of \(g\). Then the right-hand-side force convention used in many IAA accounts becomes
\[ \ddot q=M^{-1}Q,\qquad Q=Q_u+Q_p+Q_e+Q_r+Q_v+Q_g. \tag{3}\]
The plus signs join signed vectors; they do not imply positive acceleration, positive work or a beneficial effect. Numerical implementations solve a factored linear system rather than explicitly constructing \(M^{-1}\).
Induced Acceleration Decomposition
At the same state, choose a complete partition \(Q=\sum_k Q_k\). Linearity gives
\[ \ddot q=\sum_k a_k,\qquad a_k=M^{-1}Q_k. \tag{4}\]
This equation is a closed accounting identity for that partition. A muscle-force estimate \(F_i\) needs its generalized moment-arm vector \(r_i(q)\), giving \(Q_i=r_iF_i\). A net joint moment, a multi-articular muscle force and an applied grip wrench are different input channels. Their mapped terms cannot be relabeled as one another.
The familiar direct/remote distinction is particularly simple for an unconstrained unit coordinate-torque channel: entries of the corresponding column of \(M^{-1}\) give its acceleration response. For a multi-articular actuator, several columns combine through \(r_i\). A positive moment at a joint can coexist with a negative acceleration there because of other entries in the same actuator vector.
Compatibility With Control-Affine Superposition
For \(Q=Q_0(q,v)+B(q,v)u\) and \(x=(q,v)\), the complete state equation is
\[ \dot x= \underbrace{\begin{bmatrix}v\\M^{-1}Q_0\end{bmatrix}}_{f(x)} +\underbrace{\begin{bmatrix}0\\M^{-1}B\end{bmatrix}}_{G(x)}u. \]
The kinematic block \(v\) belongs to drift. All retained autonomous forces belong to \(Q_0\), not just inertial velocity terms and gravity. Distributing the lower block over a declared force partition proves instantaneous closure; it does not prove unique attribution or trajectory superposition. With fixed equality constraints, replace the lower drift block by \(PQ_0+a_c\) and the lower input block by \(PB\), as derived below.
Input choice matters. If excitation \(u\) drives an activation state \(a\) through \(\dot a=(u-a)/T\) and a unit-mass coordinate obeys \(\ddot q=a\), changing \(u\) at fixed \(a\) produces no immediate acceleration jump. Starting from \(a=0\) with unit excitation gives \(a(t)=1-e^{-t/T}\) for constant \(T>0\). A direct force input would act in the acceleration equation immediately. Both descriptions can be control-affine while having different states and input columns. More elaborate actuator laws must be inspected before affinity in excitation is claimed.
Instantaneous Effects, Cumulative Effects, and the Drift Field
Instantaneous Effects
At a fixed state and within a feasible unchanged mode, changing an applied force by \(\Delta Q\) changes acceleration by \(M^{-1}\Delta Q\), or \(P\Delta Q\) in the constrained case. This is a precisely defined model intervention on an applied-force channel. Removing a muscle, suppressing excitation, or changing a controller can additionally alter passive mechanics, internal states or contact. Those are different experiments.
Cumulative Effects
The throwing literature uses cumulative effects to describe velocity-dependent terms that reflect motion established earlier (Hirashima et al. 2008; Hirashima 2011). Current velocity also depends on the initial state, gravity, external loads and other forces. Position changes the mass matrix, moment arms and gravitational terms. Thus history matters through the evolving state, not through an exclusive storage of prior muscle action inside \(Q_v\).
There is nevertheless a useful exact integral along one nominal trajectory when velocity is absolutely continuous over the interval:
\[ v(t_1)-v(t_0)=\sum_k\int_{t_0}^{t_1}a_k(q(t),v(t),t)\,dt. \]
Include any constraint baseline in this sum exactly once. Across impulsive impacts, also add the velocity jumps from the declared impact law; an ordinary acceleration integral between events omits those jumps. These signed integrals account for velocity change in the chosen coordinates. They are not trajectories generated by simulating each force alone: changing a force generally changes the state at which every other term is evaluated. In a trajectory sensitivity calculation, this interaction appears through
\[ \delta\dot x=\bigl(Df(x)+\sum_k u_k DG_k(x)\bigr)\delta x+G(x)\delta u \]
along a prescribed nominal input history. Feedback adds the derivatives of the specified policy. Contact events need their own sensitivity treatment.
Different histories can also reach the same state. For the dimensionless double integrator \(\ddot q=u\), initially at rest at zero, both \(u=0\) and \(u=6t^2-6t+1\) over \(0\le t\le1\) finish with \(q=\dot q=0\). Their motions and inputs differ in between. The current drift therefore cannot reconstruct a unique previous force history.
Mapping to the Drift-Control Decomposition
The literature’s instantaneous applied-force terms correspond to the acceleration block of \(G(x)u\) only when the force and control maps match. Its velocity-dependent terms can occupy part of the acceleration block of \(f(x)\). Gravity, passive forces, internal-state dynamics and constraint curvature can occupy other parts. Zero applied input need not mean zero muscle force, zero joint loading or motion determined only by velocity-dependent effects.
This distinction preserves the useful investigation: earlier inputs can establish a posture and velocity from which later motion develops favorably. To establish an improvement, compare feasible input histories or policies for a declared task, not just the magnitudes of two terms at one instant. A large drift term says nothing by itself about the remaining authority to redirect the club, the effort already spent preparing the state, or robustness to error.
Applications of Induced Acceleration Analysis
Walking: The Zajac–Neptune Framework
The two-part walking review (Zajac et al. 2002, 2003) connects dynamics with muscle coordination. A concrete example is (Neptune et al. 2001), whose simulation examined walking at 1.5 m/s. Its support and progression outputs were vertical and horizontal trunk acceleration, while swing initiation concerned energy delivered to the leg before toe-off. These definitions should not be silently changed to whole-body center-of-mass acceleration.
In that model, both soleus and gastrocnemius supported the trunk during single stance and pre-swing. Their energetic roles differed with phase: late-stance soleus output chiefly accelerated the trunk forward, whereas gastrocnemius supplied energy to the leg for swing initiation. The reported earlier-stance roles were not identical to these later roles. Consequently, “soleus supports, gastrocnemius swings” is an incomplete account of the paper.
The example also distinguishes two calculations. Acceleration attribution uses the forward dynamics and a specified output. Energy accounting uses power at consistent mechanical boundaries; generalized applied power is \(Q_k^{\mathsf T}v\), and rigid-body interface power is \(F\cdot v_P+\tau_P\cdot\omega\) at a common point \(P\). An acceleration contribution alone is neither a power nor an energy-transfer measurement. Net joint power also does not separate individual muscle and tendon work.
Overarm Throwing: The Hirashima Framework
The 2007 study used nonorthogonal torque decomposition (Hirashima et al. 2007); the 2008 coupled IAA study used three fast throws from each of six baseball players (Hirashima et al. 2008). The latter model had four segments and thirteen degrees of freedom, including trunk translation. Its generalized load vector contained three forces and ten moments, not thirteen anatomical muscle torques. These method and output distinctions matter when comparing the papers.
The 2008 analysis reported larger proximal joint-load contributions and larger distal velocity-dependent contributions for selected motions. It separated velocity terms into specified kinematic components and integrated acceleration terms along the reconstructed motion. Those operations provide a model-bounded account, not a unique muscle history or a measured golf mechanism.
Its methodological criticism concerned interaction terms that still contained other unknown joint accelerations. Treating such terms as independently specified inputs obscures the simultaneous coupled solve. The problem is not Newton–Euler equations written segment by segment: those equations are valid when solved consistently. Nor is discarding off-diagonal entries of \(M\) equivalent to discarding off-diagonal entries of \(M^{-1}\); in general \(1/M_{ii}\ne(M^{-1})_{ii}\). The review by (Hirashima and Ohtsuki 2008) provides related discussion, but neither citation establishes a general prohibition on nominal acceleration integrals.
Postural Control and the Induced Acceleration Index
(Challis 2011) introduced an induced-acceleration index and reported a more-than-twelvefold directional difference for its quiet-standing model. That statement remains attached to the study’s index and normalization. To explain why a normalized ratio can be asymmetric, define our own two-coordinate quantity explicitly. Let \(H=M^{-1}\) and perturb only the first generalized moment. Then
\[ R_{1\to2}=\left|\frac{\Delta\ddot q_2}{\Delta\ddot q_1}\right| =\left|\frac{H_{21}}{H_{11}}\right| =\left|\frac{M_{21}}{M_{22}}\right|. \]
This result follows from the unforced second row, \(M_{21}\Delta\ddot q_1+M_{22}\Delta\ddot q_2=0\). The reciprocal \(1/M_{22}\) is not the inverse-matrix element \(H_{22}\). For several unforced coordinates, replace scalar division by a block solve. With constraints, use the feasible constrained response instead. The ratio requires a nonzero denominator and is dimensionless only for commensurate coordinates with specified scaling.
A synthetic angular-coordinate example, with \(M\) in kg m\(^2\), is
\[ M=\begin{bmatrix}4&1\\1&1\end{bmatrix},\qquad H=\frac13\begin{bmatrix}1&-1\\-1&4\end{bmatrix}. \]
Moments \((3,0)\) and \((0,3)\) N m produce acceleration increments \((1,-1)\) and \((-1,4)\) rad/s\(^2\). The normalized ratios are \(1\) and \(1/4\), yet the equal-torque cross responses are the same: \(H_{12}=H_{21}=-1/3\). Directional index asymmetry therefore does not contradict symmetric mobility or prove a preferred neural strategy. A multi-articular force channel producing generalized moments \((1,2)\) N m instead gives \((-1/3,7/3)\) rad/s\(^2\), illustrating why a muscle channel is not a single matrix entry. These are manufactured mechanics examples, not fitted human parameters.
Clinical Applications
In (Riley and Kerrigan 1999), ten stroke participants with stiff-legged gait were compared with ten healthy controls. Subject-specific models gave differing hip-, knee- and ankle-related terms, and summed knee acceleration was compared with kinematic estimates. These model-reported patterns motivate individual hypotheses; the study does not establish that IAA-guided treatment improved outcomes.
(Caruthers et al. 2016) used static optimization to estimate muscle forces before calculating induced accelerations in sit-to-stand simulations of young, healthy adults. It reported forward center-of-mass contributions from gluteus maximus and upward/forward contributions from soleus, with an opposing quadriceps contribution to the specified forward output. “Opposing” here refers to that output component; it does not mean the quadriceps are unnecessary for standing up.
Electrical-stimulation cycling is another modeling application (Schutte et al. 1993). Its design questions involve stimulation and ergometer configuration. A model prediction about one configuration is distinct from a demonstrated rehabilitation benefit. Across these applications, estimating a plausible force distribution, reproducing motion and predicting a useful intervention are separate validation targets.
Methodological Considerations
Ground Contact and Constraint Forces
Rigid constrained support and deformable contact require different computations. For independent equality constraints in a smooth fixed mode, write
\[ M\ddot q=Q+A^{\mathsf T}\lambda,\qquad A\ddot q=b. \]
For stationary holonomic constraints, \(A\) is the constraint Jacobian and \(b=-\dot A v\) at a position- and velocity-compatible state. Assuming positive-definite \(M\) and full row rank of \(A\), define
\[ \begin{aligned} W&=AM^{-1}A^{\mathsf T},\\ \lambda&=W^{-1}(b-AM^{-1}Q),\\ P&=M^{-1}-M^{-1}A^{\mathsf T}W^{-1}AM^{-1},\\ a_c&=M^{-1}A^{\mathsf T}W^{-1}b. \end{aligned} \]
The resulting ledger is \(\ddot q=\sum_k PQ_k+a_c\). Count the baseline once: source increments satisfy \(APQ_k=0\), while \(Aa_c=b\). Adding complete isolated-force solves would repeat the curvature acceleration. \(P\) is an inverse-inertia operator, not generally an idempotent Euclidean projection. Dependent constraints need a declared rank treatment instead of inverting singular \(W\).
For an algebraic illustration, use the synthetic matrix above, \(A=(1,1)\) and \(b=2\). The homogeneous responses to \((3,0)\) and \((0,3)\) are \((1,-1)\) and \((-1,1)\), while \(a_c=(0,2)\). Their sum is \((0,2)\), not \((0,4)\). The latter answer would count the baseline twice. This example checks accounting, not a particular contact geometry.
For unilateral ground contact, reaction sign and friction limits must also hold. A unit mass on a horizontal floor with upward applied force \(F\) requires normal reaction \(\lambda=9.81-F\) N to stay at rest. Above \(F=9.81\) N, that solution requires adhesion; ordinary nonadhesive contact releases. A frozen sticking operator remains an algebraic map but no longer describes the feasible motion.
A compliant contact law can instead supply a force determined by deformation and velocity at the state. Its reaction may change only after a perturbation evolves the state. A finite-step force-removal experiment therefore needs its duration, integration scheme, contact law and comparison output; it is not automatically the same as \(P\Delta Q\). The contact methods and attribution conventions are part of the analysis discussed by (Silverman 2018). Measured ground loads used as prescribed inputs must not also be counted as separately solved reactions.
Physical Outputs and Coordinate Changes
For a physical point \(p(q)\) expressed in a fixed inertial frame, let \(J=\partial p/\partial q\). Its acceleration ledger is
\[ \ddot p=\sum_k JPQ_k+Ja_c+\dot Jv. \]
The last term accounts for the changing direction and leverage of the kinematic map. A clubhead moving at constant angular speed on a circle has inward acceleration even when the angular acceleration is zero. The acceleration component along the point velocity changes its speed; the perpendicular component changes direction. For nonzero speed \(s\), \(\dot s=\dot p^{\mathsf T}\ddot p/s\). At rest that unit-tangent expression is undefined.
A constant invertible coordinate change \(q=Tz\) transforms \(M_z=T^{\mathsf T}M_qT\), \(Q_z=T^{\mathsf T}Q_q\), \(A_z=A_qT\) and \(J_z=J_qT\). Consistently transformed force increments give the same physical point response. Coordinate dependence of components does not make physical predictions arbitrary. For a nonlinear chart \(q=h(z)\), include \(\ddot q=Dh\,\ddot z+\dot{(Dh)}\dot z\). The transport term belongs once in the transformed baseline. This chapter uses \(v=\dot q\); quaternion and other generalized-speed formulations require their explicit velocity map.
Model Dependence
Algebraic closure checks whether the declared terms sum correctly. It does not prove that the forces, inertia or contact model are accurate. A small force error can have a large acceleration effect in a low-inertia direction: with dimensionless \(M=\operatorname{diag}(1,0.001)\), a force perturbation \((0,0.001)\) changes acceleration by \((0,1)\) even though both solves close exactly. Parameter perturbations, measurement uncertainty and conditioning must be assessed separately, with units or a scaling metric stated for mixed coordinates.
Reintegrating estimated forces is useful numerical verification. Agreement is weaker evidence for anatomical force attribution when the same measured kinematics generated the forces through inverse dynamics. Stronger physical tests reserve independent measurements or interventions and report prediction error, contact consistency and sensitivity to alternative plausible models. For muscle attribution, a successful trajectory fit still leaves the force-estimation assumptions to test.
Synthesis: Two Communities, One Mathematical Structure
The connection between IAA and affine control is productive when it makes the investigation more precise. Posture changes the mechanical operator; prior inputs help establish that posture and motion; contact changes the admissible response; actuator dynamics determine how commands become force; and the output map decides whether a contribution changes clubhead speed, direction or another task variable. None of these roles can be inferred from the name of one term alone.
A golf comparison should first choose a common physical output and an inclusive golfer–club or club-only boundary. Then state the measured quantities, inferred forces, constraint mode and residuals; check signed ledger closure; and vary plausible parameters and input policies. If the question is energy transfer, add power and storage balances. If it is steering authority, impose actuator and contact limits. If it is human effort or injury, additional physiological outputs and independent evidence are necessary. Walking and throwing studies guide the questions and methods; they do not supply golf-specific answers.
Keep Three Questions Separate. A frozen-state ledger describes how declared terms combine now. Its nominal integrals account for a change along one trajectory. A forward intervention predicts a changed trajectory under specified inputs, state and contact rules. Each is useful, but evidence for one does not automatically establish the others.
References
Citation
@online{untitled,
author = {},
title = {Induced {Acceleration} {Analysis:} {Superposition} in the
{Biomechanics} {Literature}},
url = {https://affinedrift.com/articles/The_Geometry_of_Motion/quarto/ch03b_induced_acceleration_biomechanics.html},
langid = {en}
}