Skip to main content

7  Why energy

Status: draft · Updated: 2026-11-21 · Checked by: tests/test_energy_whitening.py

Part I took the energy metric as given and showed what it does. This chapter asks whether it is the right choice for measuring mechanical intelligence, and the next asks what it leaves out.

7.1 What maximum entropy does and does not decide

The proof of Theorem 2.2 never uses the fact that \(\Em\) is an energy. For any positive-definite form \(W\), the same three steps show that the maximum-entropy law under \(\E[\nu^\top W\nu]=m\sigma^2\) is \(\mathcal N(0,\sigma^2W^{-1})\). So maximum entropy turns a choice of what is held fixed into a noise law. It does not make that choice.

For example, constraining the mean action, i.e. each mode’s energy divided by its frequency, gives the action-weighted law. That is the shape of quantum zero-point noise (Remark 7.2). The real modeling assumption is “resolution is measured in energy.” The rest of this section justifies that assumption from properties a noise model should have.

7.2 Five requirements, and a uniqueness theorem

Definition 7.1 (Mechanisms and noise rules) A mechanism is assembled from single-parameter elements:

  • springs: stiffness \(k_e\) acting on a relative coordinate \(s_e=a_e^\top q\);
  • inertias: a positive-definite inertia matrix \(M_m\) acting on a velocity \(u_m=B_m\dot q\);
  • dampers: coefficient \(b_e\) acting on \(\dot s_e\).

Hence \(K=\sum_ek_ea_ea_e^\top\), \(H=\sum_mB_m^\top M_mB_m\) and \(D=\sum_eb_ea_ea_e^\top\). A noise rule assigns to every mechanism a positive-semidefinite quadratic form \(W(q,\dot q)\), the precision of the body’s resolution: noise \(\mathcal N(0,\sigma^2W^{-1})\).

The requirements, each with the reason it matters:

  • (U1) Coordinate-free. \(W\) transforms with the coordinates (\(\tilde W=\mathcal T^\top W\mathcal T\) under \(q=T\tilde q\)), including rescalings of each element’s own coordinate. Otherwise relabeling changes the answer (Theorem 2.1).
  • (U2) Local and universal. \(W\) is a sum of one term per element. Each term depends only on that element’s parameter and kinematics, through a function fixed for each element type. Adding a component somewhere must not change how precisely another component’s state is known; this is what makes components compose, stiff limits behave (Theorem 4.1) and meshes converge (Theorem 6.1). Universality says the rule cannot know which spring is which.
  • (U3) No outside scales. No dimensional constant other than the element’s own parameter may appear: no reference frequency, length or time. Otherwise the result depends on an arbitrary choice; the "kinetic" readout’s \(\omega_{\rm ref}\) can flip the sign of the SEA-rigid comparison (test_the_ground_decides_whether_series_compliance_helps_or_hurts).
  • (U4) Time-reversal symmetric. \(W(q,-\dot q)=W(q,\dot q)\). Resolution should not depend on the direction of motion.
  • (U5) Unchanged by free motion. For every mechanism without dampers, \(W\) is constant along every solution of \(H\ddot q+Kq=0\). Proposition 7.1 shows this is exactly the requirement that free motion never creates information.

Theorem 7.1 (Uniqueness of the energy metric) A noise rule satisfies (U1) to (U5) if and only if \(W=c\,(q^\top Kq+\dot q^\top H\dot q)=c\,x_q^\top\Em x_q\) for a constant \(c>0\), i.e. it is the energy metric up to the overall noise scale.

Proof. The energy metric satisfies all five.

  • (U1) is Theorem 2.1.
  • (U2) to (U4) hold by inspection: \(k_es_e^2\) and \(u_m^\top M_mu_m\) are energies.
  • (U5) holds because energy is conserved without damping: \(\frac{d}{dt}(q^\top Kq+\dot q^\top H\dot q)=2\dot q^\top(Kq+H\ddot q)=0\).

Conversely, step 1: a spring. By (U2) its term is a quadratic form in its own kinematics \((s,\dot s)\) with coefficients depending only on \(k\): \(\alpha_1(k)s^2+\alpha_2(k)s\dot s+\alpha_3(k)\dot s^2\).

  • (U4) gives \(\alpha_2=0\).
  • Rescaling the spring’s coordinate, \(s=\lambda\tilde s\), turns the stiffness into \(\tilde k=\lambda^2k\). By (U1) the term must keep its value: \(\alpha_1(\lambda^2k)\tilde s^2=\alpha_1(k)\lambda^2\tilde s^2\) for every \(\lambda>0\), so \(\alpha_1(k)=\alpha\,k\) for a dimensionless constant \(\alpha\). Likewise \(\alpha_3(k)\propto k\).
  • But \(k\dot s^2\) has units of energy times time squared. Turning it into an energy needs a time scale, which (U3) forbids. So \(\alpha_3=0\), and the spring’s term is \(\alpha\,k_es_e^2\).

Step 2: an inertia. Its kinematics is a velocity \(u\), and a position-dependent term would need a time scale (U3). So its term is \(u^\top F(M)u\) for some matrix function \(F\).

  • (U1), applied to a change of the element’s velocity coordinates \(u=T\tilde u\) (which sends \(M\mapsto T^\top MT\)), requires \(F(T^\top MT)=T^\top F(M)T\) for every invertible \(T\).
  • Taking \(T=M^{-1/2}\) gives \(F(M)=M^{1/2}F(I)M^{1/2}\).
  • Taking \(T\) orthogonal gives \(F(I)=T^\top F(I)T\) for every orthogonal \(T\), so \(F(I)=\beta I\) (a matrix that commutes with all rotations is a multiple of the identity).
  • Hence the term is \(\beta\,u_m^\top M_mu_m\).

Step 3: a damper. The same rescaling argument makes any term proportional to \(b\). The candidates are \(b\,s\dot s\), \(b\,s^2\) and \(b\,\dot s^2\):

  • \(b\,s\dot s\) is odd in time, which (U4) forbids;
  • \(b\,s^2\) has units of energy times time, and \(b\,\dot s^2\) is a power; turning either into an energy needs a time scale, which (U3) forbids.

Dampers therefore contribute nothing.

Step 4: free motion fixes the ratio. By universality ((U2)) the constants \(\alpha\) and \(\beta\) are the same for every element, so \(W=\alpha\,q^\top Kq+\beta\,\dot q^\top H\dot q\). Along any solution of \(H\ddot q=-Kq\), \[ \tfrac{d}{dt}W=2\alpha\,\dot q^\top Kq+2\beta\,\dot q^\top H\ddot q =2(\alpha-\beta)\,\dot q^\top Kq . \] For any mechanism with a spring, some state has \(\dot q^\top Kq\neq0\), so (U5) forces \(\alpha=\beta=:c\).

Remark 7.1 (What the axioms do and do not exclude).

  • Elements are single constitutive laws. A Kelvin-Voigt element (spring and damper as one element) has an internal time scale \(b/k\). If it were treated as one element, (U3) would allow a term \((b^2/k)\dot s^2\). Treating springs and dampers as separate elements, as their additive force laws suggest, removes this.
  • Universality matters. Without it, (U5) still forces \(\alpha=\beta\), but only within each mechanically connected component. Two disconnected bodies could carry different noise scales.
  • Dissipation drops out by itself. No assumption singled out damping; the axioms exclude it from the metric.
NoteNumerical check

The identity \(\frac{d}{dt}W=2(\alpha-\beta)\dot q^\top Kq\) holds to \(4\times10^{-14}\) on random SEA states (check_uniqueness_flow). Over 20 ms of undamped SEA motion, the deviation of \(\Phi^\top W\Phi\) from \(W\) is \(4\times10^{-13}\) for the energy and the frequency-weighted forms of Section 7.4, but \(6.4\) for per-coordinate noise and \(1.7\) for the kinetic metric (relative to \(\max|W|\); check_flow_invariance).

7.3 Free motion cannot create information

The fifth requirement looks like a technicality. It is the most important one for attributing computation to a body.

Proposition 7.1 (Information is created only by work) Store information about a random variable \(Y\) in the body’s state, \(x(0)\), and let the body move freely (\(f=0\)). Observe \(x(t)\) through physical noise \(\nu\sim\mathcal N(0,\sigma^2W^{-1})\) independent of \(Y\), and let \(I_t=I(Y;x(t)+\nu)\). Write \(\Phi(t)\) for the state-transition matrix.

  1. If \(D=0\) and \(\Phi^\top W\Phi=W\) (for example \(W=\Em\)), then \(I_t=I_0\) for all \(t\).
  2. If \(W=\Em\) and \(D\succeq0\), then \(I_t\) is non-increasing in \(t\).
  3. If \(D=0\) and \(\Phi(t)^\top W\Phi(t)\neq W\) for some \(t\), then there is a Gaussian scalar \(Y\) with \(I_t>I_0\): free motion creates apparent information.

Consequently, among quadratic noise models, (U5) holds if and only if free conservative motion never creates information.

Proof. Work in the coordinates \(z=W^{1/2}x\), where the noise is \(\mathcal N(0,\sigma^2I)\) and the flow is \(\tilde\Phi=W^{1/2}\Phi W^{-1/2}\). Note that \(\tilde\Phi^\top\tilde\Phi=W^{-1/2}\Phi^\top W\Phi W^{-1/2}\).

  1. If \(\Phi^\top W\Phi=W\), then \(\tilde\Phi\) is orthogonal. The observation \(\tilde\Phi z(0)+n\) is an invertible transformation of \(z(0)+\tilde\Phi^\top n\), and \(\tilde\Phi^\top n\sim\mathcal N(0,\sigma^2I)\) has the same law as \(n\). So the two observations carry the same information about \(Y\) (Lemma 2.1(a)).

  2. With \(W=\Em\), \(\tilde\Phi(s)=e^{\At s}\) is a contraction, \(\|C\|_2\le1\), by Theorem 3.1(b). Fix \(t\) and \(s\ge0\), and let \(C=e^{\At s}\), so that \(z(t+s)+n=Cz(t)+n\). Let \(n_1\sim\mathcal N(0,\sigma^2I)\) and \(n_2\sim\mathcal N(0,\sigma^2(I-CC^\top))\) be independent; the second covariance is valid because \(\|C\|\le1\). Then \(C(z(t)+n_1)+n_2\) has noise covariance \(\sigma^2(CC^\top+I-CC^\top)=\sigma^2I\), so it has the same joint law with \(Y\) as \(Cz(t)+n\). It is obtained from the earlier observation \(z(t)+n_1\) by further processing plus independent noise. By the data-processing inequality, \(I_{t+s}\le I_t\).

  3. Let \(R=\tilde\Phi^\top\tilde\Phi\neq I\); it is symmetric positive definite. Liouville’s theorem gives \(\det\Phi(t)=e^{t\,\tr A}=1\), because \(\tr A=-\tr(H^{-1}D)=0\) when \(D=0\). Hence \(\det R=\det(\Phi)^2=1\). A positive-definite matrix with determinant 1 that is not the identity has an eigenvalue \(\lambda_{\max}>1\) (eigenvalues whose product is 1 cannot all be \(\le1\) unless all equal 1). Let \(u\) be its eigenvector, and store \(Y\sim\mathcal N(0,\sigma_Y^2)\) as \(x(0)=W^{-1/2}u\,Y\). Then \[ I_0=\tfrac12\log\!\Big(1+\tfrac{\sigma_Y^2}{\sigma^2}|u|^2\Big) <\tfrac12\log\!\Big(1+\tfrac{\sigma_Y^2}{\sigma^2}\lambda_{\max}|u|^2\Big)=I_t . \]

TipWhat this means physically

Under the energy metric, a body’s information about anything it has stored can change in only two ways: the force does work on it, which can add information, or the dampers dissipate it, which can only remove information. Conservative motion only moves information around, as a rotation in whitened coordinates.

Under any other quadratic noise model, a free oscillation that rotates the state into better-resolved directions creates bits from nothing. A measure of mechanical computation should credit the body for what the force delivered and debit it for what was dissipated, not for coordinate artifacts of free motion. This is the same accounting as in the thermodynamics of computation (Landauer 1961; Bennett 1982): conservative dynamics are reversible and lose nothing, and dissipation is the irreversible step.

Figure 7.1: Information about a signal stored in the SEA’s state (\(k_s=10^4\)), under free motion, for four noise models. Each starts from the direction it resolves worst, scaled to 2 bits. (a) Undamped: energy and action weighting stay at exactly 2 bits. Per-coordinate noise reaches 3.58 bits and the kinetic metric 3.22 bits, created by free motion alone. (b) With the leg’s damping, energy can only decrease (Proposition 7.1(ii)). Per-coordinate noise still rises to 3.08 bits first (check_free_motion).

Remark 7.2 (Classical versus quantum: why energy and not action). The quantum fluctuation-dissipation theorem (Callen and Welton 1951) gives each mode of frequency \(\omega\) a fluctuation energy \(\frac{\hbar\omega}2\coth\frac{\hbar\omega}{2k_BT}\). In whitened coordinates this is the noise variance per mode.

  • When \(k_BT\ll\hbar\omega\) it tends to \(\frac12\hbar\omega\): noise energy grows with frequency, which is action weighting (precision \(\propto1/\omega\)).
  • When \(k_BT\gg\hbar\omega\) it tends to \(k_BT\): equal energy per mode, the energy metric.

For a 200 Hz mode at room temperature, \(\hbar\omega/k_BT\approx3\times10^{-11}\), so macroscopic mechanisms sit deep in the classical regime. Physics and the locality axiom therefore agree that, for robots, resolution should be measured in energy.

Energy is also the additive conserved quantity associated with time-translation symmetry (Noether), which is why it can be assembled element by element.

7.4 The alternatives compared

Table 7.1 lists the candidate noise models, the requirements each violates, and how it behaves in the two limits of Part I. The frequency-weighted family is \[ W_p=\sum_i\omega_i^p\,(\text{modal energy form } i) =\Em^{1/2}\diag\!\big((\Omega\Omega^\top)^{p/2},(\Omega^\top\Omega)^{p/2}\big)\Em^{1/2}, \] with \(\Omega=K^{1/2}H^{-1/2}\) as in Theorem 3.1. Here \(p=0\) is the energy metric and \(p=-1\) is action weighting. (For numerical comparison each \(W_p\) is divided by \(\omega_{\rm ref}^p\) so that it is measured in joules; this constant cancels in every design gap.)

Table 7.1: Candidate noise models. The stiff-limit column is SEA minus rigid bits at \(k_s=10^{10}\) (\(B_w=100\) Hz). The mesh column is the change from \(N=32\) to \(N=64\) elements on the rod of Theorem 6.1.
Rule Violates Stiff limit (bits) Mesh, \(N=32\to64\) (bits)
energy, \(W_0=\Em\) none 0.0000 +0.017, converges
action, \(W_{-1}\) U2 0.0000 +0.010, converges
\(W_{+1}\) U2 0.0007, converging slowly +2.96, keeps growing
\(W_{+2}\) U2 1.489, no limit
\(W_{+3}\) U2 60.1, diverges
kinetic, \(\diag(\omega_{\rm ref}^2H,H)\) U3, U5 0.0000 +0.018, converges
per-coordinate ("state") U1, U2, U3, U5 8.4716 +33.3, keeps growing
data-driven whitening not built from the body erases the body (Proposition 7.2)
Gramian or balanced coordinates not built from the body
encoder on the motor ("motor") answers a different question (the sensor)

The frequency-weighted family. Every \(W_p\) is coordinate-free (spread \(<10^{-13}\) over the reparametrizations of Figure 2.1) and flow-invariant (check_flow_invariance). Each is non-local, however: it is a matrix function of \(H^{-1}K\), so every element influences every entry.

  • Locality, concretely. Changing the ground stiffness \(k_i\) from \(10^5\) to \(10^7\) changes the action metric’s motor-velocity entry by a factor of 2.2. The energy metric’s entry stays exactly at \(I_{\rm rot}\) (check_locality).
  • The stiff limit decides \(p<2\). By Proposition 4.2 the fast mode’s energy is \(O(k_s^{-1})\), while its frequency, and so its weight, grows like \(k_s^{p/2}\). Its weighted contribution therefore behaves like \(k_s^{p/2-1}\), which vanishes if and only if \(p<2\). This is the argument of Theorem 4.1 with a weighted norm; we state it as a prediction rather than a theorem, because the slow-mode part is only checked numerically. The numbers confirm it exactly (Figure 7.2(a)):
    • \(p\le0\) converges;
    • \(p=1\) converges at the predicted rate: the gap falls tenfold, from \(7.1\times10^{-3}\) to \(7\times10^{-4}\), per factor 100 in \(k_s\);
    • \(p=2\) stalls at \(1.49\) bits;
    • \(p=3\) diverges.
  • The mesh decides \(p\le0\). On the rod, \(p=-1\) and \(p=0\) converge, and \(p=+1\) keeps growing by about 3 bits per doubling (Figure 7.2(b)).

So action weighting is the one serious competitor: it passes both limit tests. What excludes it is locality and, physically, the classical regime (Remark 7.2).

Figure 7.2: (a) Stiff limit of the frequency-weighted metrics \(W_p\): SEA minus rigid bits against \(k_s\). The gap vanishes for \(p<2\), stalls at \(p=2\) and diverges at \(p=3\), as predicted. (b) Mesh refinement of the rod, relative to \(N=64\). Energy, action and the kinetic metric converge; \(p=+1\) does not (check_alternatives_sea, check_alternatives_rod).

Proposition 7.2 (Data-driven whitening erases the body) Suppose the stacked response \(y=Mw\) (\(M\) of full column rank \(d\)) is whitened by its own covariance, \((\sigma_w^2MM^\top)^{+1/2}\) on the range of \(M\), and isotropic noise \(\sigma_n\) is then added. Then the information is \(\frac d2\log_2(1+\sigma_n^{-2})\) for every body.

Proof. Write the thin SVD \(\sigma_wM=U\Sigma V^\top\). The whitener is \(U\Sigma^{-1}U^\top\), and the whitened signal is \(U\Sigma^{-1}U^\top U\Sigma V^\top(w/\sigma_w)=UV^\top(w/\sigma_w)\). Its Gram matrix is \(\sigma_n^{-2}VU^\top UV^\top=\sigma_n^{-2}I_d\). By Lemma 1.1, the bits are \(\frac12\log_2\det((1+\sigma_n^{-2})I_d)\).

TipWhat this means physically

Whitening by the data, i.e. \(z\)-scoring or ZCA on the observed response, makes every direction equally informative by construction, so the body’s contribution vanishes. The check gives 56.59 bits for the rigid leg and for SEAs at 70 and \(10^4\), equal to the prediction (check_data_driven_whitening). A noise model has to be built from the body.