3 The dynamics in whitened coordinates
In whitened coordinates the equations of motion separate into a part that conserves energy and a part that dissipates it. That structure is behind several later results: the information-preservation property of Proposition 7.1, the stationarity used for \(\mathrm{MC}_W\) in Section 5.2, and the uniform energy bound in the proof of Proposition 4.2.
Theorem 3.1 (Port-Hamiltonian form) Let \(K\succ0\), and let \(z=\Em^{1/2}x_q\). Then \(\dot z=\At z+\tilde Bf\) with \[ \At=\Em^{1/2}A\,\Em^{-1/2} =\underbrace{\begin{bmatrix}0&\Omega\\-\Omega^\top&0\end{bmatrix}}_{\Jt\;=\;-\Jt^\top} -\underbrace{\begin{bmatrix}0&0\\0&H^{-1/2}DH^{-1/2}\end{bmatrix}}_{\Rt\;=\;\Rt^\top\;\succeq\;0}, \qquad \Omega=K^{1/2}H^{-1/2},\qquad \tilde B=\begin{bmatrix}0\\H^{-1/2}J^\top\end{bmatrix}. \tag{3.1}\] Consequently:
- Power balance: \(\frac{d}{dt}\frac12|z|^2=-\dot q^\top D\dot q+(J\dot q)\,f\). The rate of change of energy is the power delivered at the foot minus the power dissipated.
- Contraction: \(\|e^{\At t}\|_2\le1\) for \(t\ge0\). Free motion never increases the whitened norm.
- Natural frequencies: the singular values of \(\Omega\) are the undamped natural frequencies \(\omega_i\), the square roots of the eigenvalues of \(K\phi=\omega^2H\phi\).
- Modes are rotations: if \(D=0\), each mode is a rotation at rate \(\omega_i\) in a two-dimensional plane of \(z\)-space. The radius of the circle is \(\sqrt{2\times(\text{modal energy})}\).
Proof. Computing \(\At\). With \(\Em^{1/2}=\diag(K^{1/2},H^{1/2})\), multiply out \(\Em^{1/2}A\Em^{-1/2}\) block by block using Equation 1.2:
- the \((1,1)\) block is \(K^{1/2}\cdot0\cdot K^{-1/2}=0\);
- the \((1,2)\) block is \(K^{1/2}\cdot I\cdot H^{-1/2}=\Omega\);
- the \((2,1)\) block is \(H^{1/2}(-H^{-1}K)K^{-1/2}=-H^{-1/2}K^{1/2}=-\Omega^\top\);
- the \((2,2)\) block is \(H^{1/2}(-H^{-1}D)H^{-1/2}=-H^{-1/2}DH^{-1/2}\).
Also \(\tilde B=\Em^{1/2}B=[0;H^{1/2}H^{-1}J^\top]=[0;H^{-1/2}J^\top]\). The matrix \(\Rt\) is positive semidefinite because it is congruent to \(D\succeq0\).
Write \(z=(z_q,z_v)\), so \(z_v=H^{1/2}\dot q\). Then \(\frac{d}{dt}\frac12|z|^2=z^\top\At z+z^\top\tilde Bf\). Now \(z^\top\Jt z=0\) because \(\Jt\) is skew. The remaining terms are \(-z_v^\top H^{-1/2}DH^{-1/2}z_v=-\dot q^\top D\dot q\) and \(z_v^\top H^{-1/2}J^\top f=(J\dot q)f\).
Take \(f=0\) in (a): \(\frac{d}{dt}|z|^2=-2z^\top\Rt z\le0\). So \(|e^{\At t}z_0|\le|z_0|\) for every \(z_0\).
\(\Omega\Omega^\top=K^{1/2}H^{-1}K^{1/2}=K^{1/2}(H^{-1}K)K^{-1/2}\) is similar to \(H^{-1}K\). The eigenvalues of \(H^{-1}K\) are the \(\omega_i^2\).
Take the singular value decomposition \(\Omega=P\Sigma U^\top\). In the coordinates \((P^\top z_q,\,U^\top z_v)\) the matrix \(\Jt\) becomes \(\begin{bmatrix}0&\Sigma\\-\Sigma&0 \end{bmatrix}\). This decouples into \(2\times2\) blocks \(\begin{bmatrix}0&\omega_i\\-\omega_i&0\end{bmatrix}\), each generating a rotation at rate \(\omega_i\). The squared radius in block \(i\) is that mode’s share of \(|z|^2\), which is twice its energy by Lemma 2.2.
For the rigid leg and SEAs with \(k_s\in\{70,10^3,10^4,10^5,10^6\}\):
- the symmetric part of \(\At\) equals \(-\Rt\) to \(7\times10^{-16}\) relative error;
- the largest eigenvalue of \(\At+\At^\top\) is \(\le 3\times10^{-14}\);
- the singular values of \(\Omega\) equal \(2\pi\times\)
modes_hz(). For example, at \(k_s=10^4\) they are \(482.087\) and \(1197.605\) rad/s (76.7 and 190.6 Hz).
(check_port_hamiltonian.)
Relation to the literature.
- Port-Hamiltonian systems. Equation 3.1 is the input-state-output port-Hamiltonian form \(\dot z=[\Jt-\Rt]\,\partial\mathcal H/\partial z+gu\), \(y=g^\top\partial\mathcal H/\partial z\) of Schaft and Jeltsema (2014, Eq. (4.10)), with Hamiltonian \(\mathcal H=\frac12|z|^2\) and \(g=\tilde B\). Its natural output, \(y=\tilde B^\top z=J\dot q\), is the foot velocity: the power conjugate of the unknown force.
- Energy whitening is the change of coordinates that turns the storage function into the identity.
- The storage function \(\frac12x_q^\top\Em x_q\) is the classical one of dissipativity theory (Willems 1972).
- Modal analysis and sensor placement (Li et al. 2007). The velocity block \(z_v=H^{1/2}\dot q\) is the mass weighting behind the modal kinetic energy (MKE) criterion for sensor placement. The eigenvectors of \(\Omega^\top\Omega=H^{-1/2}KH^{-1/2}\) are the mass-normalized modes \(H^{1/2}\phi_i\). The input in modal coordinates is the modal participation \(\phi_i^\top J^\top f\).
- As we understand it from secondary sources (Appendix A), the cited work relates effective independence (EfI) and MKE through iteration and re-orthonormalization, without showing that the two coincide under mass whitening.
- EfI is the Fisher/D-optimal criterion of Kammer (1991). It assumes isotropic noise in physical coordinates.
- The analogy is with MKE’s mass weighting. It covers the kinetic block only, not the stiffness block.
- Gramians and balancing (Moore 1981). Moore (1981) observed that principal components of the controllability Gramian depend on coordinates: under \(x\mapsto Sx\) the Gramian transforms as \(W\mapsto SWS^\top\). He removed the dependence by balancing, i.e. choosing coordinates in which the controllability and observability Gramians are equal.
- Energy whitening removes the same dependence differently. By Proposition 2.1, it is the coordinate system in which the controllability Gramian of the thermal input is isotropic.
- The Gram matrix Equation 1.5 is a finite-horizon, input-restricted, Gramian-like object whose output inner product is fixed by \(\Em\).
- Balanced and energy coordinates generally differ. One does not justify the other.