Skip to main content

8  A thermodynamically consistent body: Langevin noise

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

The noise of Part I is added to the readout, independently at each sample. A body that is physically noisy is driven by thermal forces matched to its own dampers, and its noise is shaped by its dynamics. This chapter builds that model and asks whether the conclusions of Part I survive.

The model. The equation of motion becomes \[ H\ddot q+D\dot q+Kq=J^\top f(t)+\xi(t),\qquad \E[\xi(t)\xi(s)^\top]=2\sigma_T^2D\,\delta(t-s), \] started from the Gibbs law \(x_q(0)\sim\mathcal N(0,\sigma_T^2\Em^{-1})\) on the clamped bench. By Proposition 2.1 the noise is stationary at covariance \(P=\sigma_T^2\Em^{-1}\). Sampling with step \(\Delta t\) gives \(x_{k+1}=A_dx_k+B_df_k+w_k\) with \(w_k\) i.i.d. \(\mathcal N(0,Q_d)\) and \(Q_d=P-A_dPA_d^\top\); this is gibbs_pair, scaled by \(\sigma_T^2\). We set \(\sigma_T^2=\sigma^2=10^{-4}\) J, so a snapshot of the noise has exactly the covariance used in Part I. Only its correlation over time differs.

Proposition 8.1 (Full observation: an exact formula) If the full state is observed without readout noise, the information about the force profile is \[ \bits=\tfrac12\log_2\det\!\big(I_d+\sigma_w^2\,g\,\bar\Phi^\top\bar\Phi\big), \qquad g=B_d^\top Q_d^{-1}B_d , \] where \(\bar\Phi\) holds the force basis at the samples that drive the state. The design enters only through the scalar \(g\), the per-sample signal-to-noise ratio of a unit force against one thermal kick.

Proof. Step 1: change variables. The map from \((x_0,x_1,\dots,x_N)\) to \((x_0,e_0,\dots,e_{N-1})\), with \(e_k=x_{k+1}-A_dx_k\), is invertible. So the two carry the same information (Lemma 2.1(a)).

Step 2: what the new variables contain. \(x_0\) is independent of the force, and \(e_k=B_df_k+w_k\), where the \(w_k\) are i.i.d., independent of \(x_0\), and Gaussian with covariance \(Q_d\). This is a Gaussian channel with noise covariance \(I\otimes Q_d\).

Step 3: apply the Gaussian formula. Whiten the noise by \(Q_d^{-1/2}\) and apply Lemma 1.1 with \(f_k=\sum_jw_j\bar\Phi_{kj}\). The Gram matrix is \(\sigma_w^2\sum_k\bar\Phi_k^\top(B_d^\top Q_d^{-1}B_d)\bar\Phi_k\), where \(\bar\Phi_k\) is row \(k\) of \(\bar\Phi\). Since \(g\) is a scalar, this equals \(\sigma_w^2g\,\bar\Phi^\top\bar\Phi\).

Proposition 8.2 (Only the dissipation at the contact matters) Suppose \(J^\top\) lies in the range of \(D\). As \(\Delta t\to0\), \[ g\;\to\;\frac{\Delta t}{2\sigma_T^2}\,J D^+J^\top . \] With the full state observed, the information therefore depends only on \(JD^+J^\top\): the damping seen at the point where the force acts. For the ground model used here, \(JD^+J^\top=1/b_i\) for the rigid transmission and for every SEA, so every design gives the same bits.

Proof (Proof sketch). Continuous time. In the continuous-time limit, observing the state exactly is a white-Gaussian-noise channel for the drift \(Bf\), with noise intensity \(GG^\top=\diag(0,\,2\sigma_T^2H^{-1}DH^{-1})\). In the low-signal limit, where the causal estimate of the signal is negligible, Duncan’s theorem (Duncan 1970) gives the information rate \(\frac12\E[(Bf)^\top(GG^\top)^+(Bf)]\) nats per second.

The rate. Substituting \(B=[0;H^{-1}J^\top]\) gives \[ \tfrac12\E[f^2]\,JH^{-1}\big(2\sigma_T^2H^{-1}DH^{-1}\big)^+H^{-1}J^\top =\frac{\E[f^2]}{4\sigma_T^2}\,JD^+J^\top . \]

Matching to the discrete channel. Over one sample, the discrete channel carries \(\frac12\log(1+g\,\E[f^2])\approx\frac12g\,\E[f^2]\) nats about a force sample. The continuous rate gives \(\Delta t\,\E[f^2]\,JD^+J^\top/(4\sigma_T^2)\) over the same interval. Equating the two gives the stated limit for \(g\), which does not depend on the signal amplitude.

The design-independence.

  • For the rigid leg, \(D=b_ir^2\) and \(J=-r\), so \(JD^+J^\top=1/b_i\).
  • For the SEA, \(D=\left[\begin{smallmatrix}b_s+b_ir^2&-b_s\\-b_s&b_s\end{smallmatrix}\right]\) has \(\det D=b_sb_ir^2\), so \((D^{-1})_{11}=b_s/(b_sb_ir^2)=1/(b_ir^2)\). With \(J=(-r,0)\), this gives \(JD^{-1}J^\top=1/b_i\), independently of \(k_s\) and \(b_s\).

What is not proved here. We do not prove that the discrete position-block terms leave the constant unchanged. The numbers below confirm it.

NoteNumerical check
  • \(2\sigma_T^2 g/(\Delta t\,JD^+J^\top)\) is \(1.0000\) at 4, 16 and 64 kHz for the rigid leg and for SEAs at \(k_s=70\), \(10^4\) and \(10^6\) (shown as \(g\sigma_T^2 f_s/JD^+J^\top=0.5000\) by check_langevin_dt_limit). It deviates only for coarse sampling (0.19 at 100 Hz).
  • The exact formula of Proposition 8.1 and a direct computation with the full stacked noise covariance agree, 37.74216 bits (check_langevin_innovations).
  • At the settings of Section 1.2, every design gives 37.742 bits with the full state observed.
TipWhat this means physically

This is the thermal limit of force sensing. The force must be read against the fluctuating force of the damper at the contact, whose strength is set by the fluctuation-dissipation theorem (the mechanical analogue of Johnson-Nyquist noise (Nyquist 1928)). If the whole state is observed, the body’s inertia and stiffness are irrelevant: no morphology can beat the contact-damper limit.

The body’s design matters only when the observation is partial, which is always the case for a real robot.

Partial observation: what morphology does. With noise injected through the dampers, observing coordinates \(C\) with readout noise \(\varepsilon\) gives the Gram matrix \(\sigma_w^2M_C^\top(C\Sigma C^\top+\varepsilon^2I)^{-1}M_C\). Here \(\Sigma\) is the stacked noise covariance, with blocks \(\Cov(x_k,x_l)=A_d^{k-l}P\) for \(k\ge l\) (langevin_bits). By data processing, observing a subset can never exceed the full-state value. So 37.742 bits is a hard ceiling for every design and every sensor.

Table 8.1: Bits about the force profile (\(B_w=100\) Hz) under the Langevin body for several readouts, and under the readout-noise models of Part I (check_langevin_designs). Efficiency is motor bits divided by full-state bits.
Design Langevin: full state Langevin: motor Efficiency Langevin: motor + encoder Readout noise: energy Readout noise: motor
rigid 37.742 37.695 0.999 35.707 71.723 66.836
SEA, \(k_s=70\) 37.742 2.266 0.060 1.637 69.280 24.379
SEA, \(10^3\) 37.742 18.628 0.494 17.925 70.267 65.892
SEA, \(3\times10^3\) 37.742 29.038 0.769 28.194 71.084 72.780
SEA, \(10^4\) 37.742 35.067 0.929 33.820 71.773 69.414
SEA, \(10^5\) 37.742 37.242 0.987 35.631 71.742 67.104
SEA, \(10^6\) 37.742 37.537 0.995 35.704 71.725 66.863

Three findings (Table 8.1; Figure 8.1):

  1. The rigid transmission routes almost everything to its motor encoder. Through the motor angle alone it captures 99.9% of the thermal ceiling. The SEA’s efficiency rises monotonically with stiffness: 6% at the prototype’s \(k_s=70\), 49% at \(10^3\), 93% at \(10^4\), 99.5% at \(10^6\). This is the argument of Wensing et al. (2017) for proprioceptive actuation, derived here from first principles: the rigid transmission delivers contact information to the actuator’s own sensor without loss.

  2. The readout-noise model can rank designs wrongly. With fixed encoder noise added to the readout, the SEA at \(k_s=3\times10^3\) beats the rigid leg by 5.9 bits through the motor encoder, and at \(10^4\) by 2.6 bits. Under Langevin noise this cannot happen: the rigid leg is already within 0.05 bits of the ceiling, so no design can beat it by more.

    The difference is resonance. A passive resonance amplifies the force signal at the motor, and with readout-only noise that amplification looks like free SNR. Under fluctuation-dissipation the same resonance amplifies the thermal force noise by exactly the same factor, so it gains nothing. This is the duplicate-readout artifact again, now in the frequency domain.

  3. With full-state readout, the two models agree qualitatively. Langevin noise plus energy readout noise gives 36.2 to 36.4 bits for every design. The energy-whitened readout model of Part I gives 69.3 to 71.8. Both say the full state hardly separates the designs, whereas the motor encoder separates them strongly.

Figure 8.1: (a) SEA minus rigid bits through different noise models. Langevin noise with the full state observed gives zero, to within \(10^{-5}\) bits, for every design. Through the motor encoder, Langevin noise gives a monotone climb to the rigid leg, while readout-only noise lets the SEA overtake the rigid leg near \(3\times10^3\). (b) Routing efficiency under Langevin noise: the fraction of the thermal ceiling that reaches the motor encoder.

Caveats.