Pith. sign in

REVIEW 2 major objections 5 minor 1 cited by

Equilibrium Propagation for Dissipative Dynamics

T0 review · 2 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Equilibrium propagation extends to damped linear mechanical and electrical networks via a time-reversed effective action.

desk verdict Solid extension of EqProp to damped linear systems; the O(β) concern is unfounded, but lack of code and error bars needs a revision. read the letter →

arxiv 2506.20402 v2 pith:4VDEFFOP submitted 2025-06-25 cond-mat.dis-nn cond-mat.soft

Marc Berneman , Daniel Hexner This is my paper
classification cond-mat.dis-nncond-mat.soft
keywords equilibriumpropagationdissipativedynamicseffectiveactionphysicallearningreciprocalsystemsRLCnetworksspringtemporalsignalclassification
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Equilibrium propagation is a scheme for computing gradients of a cost function from physical measurements, and until now it has mostly been used in quasistatic settings where the state minimizes an energy. This paper claims the scheme also works for damped linear reciprocal dynamics—networks of springs with friction and RLC circuits with resistance—even though a dissipative trajectory does not extremize an energy. The fix is an effective action built from a trajectory paired with its time-reversed copy; its extremum is exactly the equation of motion $Hx=f$. Under periodic driving or resting initial conditions, the boundary terms that would spoil the variational principle vanish, and the standard equilibrium-propagation gradient identity applies. The paper demonstrates the method by training disordered spring networks and by training an RLC network to classify spoken digits under periodic driving and to distinguish single pulses from rest.

What carries the argument

The load-bearing object is the effective action $E$, defined in the time domain by pairing the trajectory with its time-reversed copy: $E=\frac{1}{\tau}\int_{-\tau/2}^{\tau/2}\left(\frac{1}{2}x(-t)^{\top}Hx(t)-x(-t)^{\top}f(t)\right)dt$. Its stationarity condition is exactly $Hx=f$ because $H$ is reciprocal under integration by parts, and the boundary terms vanish under periodic boundary conditions or resting initial conditions. This converts a dissipative evolution, which has no conventional variational principle, into an extremization problem suitable for equilibrium propagation. The second piece of machinery is the nudged equation $Hx^{(\beta)}=f+2\beta\,\delta_T\,\mathcal{T}(x_D-x_T^{(0)})$, where $\mathcal{T}$ reverses time and the time-reversed nudged trajectory is replaced by the free trajectory; this makes the update computable while keeping the gradient unbiased to leading order in $\beta$. Concrete update rules, such as $\partial F/\partial \gamma_i=\frac{1}{2\tau}\int r_i(-t)^{\top}\dot r_i(t)\,dt$ for damping and $\partial F/\partial g_i=\frac{1}{2\tau}\int V_i(-t)\dot V_i(t)\,dt$ for conductance, follow from differentiating the action.

What would settle it

A direct numerical test on a small damped spring or RLC network would settle it: compute the gradient using the paper's replacement rule (equation 8) and compare it with the exact finite-difference gradient obtained by solving the fully time-reversed nudging equation (equation 7) across a range of $\beta$ values. If the two gradients differ by a term that does not vanish as $\beta\to 0$, or if gradient descent with the approximate estimator fails to minimize a known quadratic cost while the exact estimator succeeds, the central claim is wrong.

Watch

Extended reading notes

Core claim

The central claim is that for any linear reciprocal dynamical system with dissipation, written as $Hx=f$ with $H=M\frac{d^2}{dt^2}+\Gamma\frac{d}{dt}+K$, the trajectories that satisfy the dynamics are exactly the stationary points of the effective action $E=\frac{1}{\tau}\int (\frac{1}{2}x(-t)^{\top}Hx(t)-x(-t)^{\top}f(t))\,dt$. Because the action is stationary, the equilibrium-propagation identity with a central finite difference gives $\nabla_\theta C$ from two nudged trajectories, with nudging applied only at target nodes. The one nonlocal piece—the time-reversed nudged trajectory in the nudging equation—is replaced by the free trajectory, a step the authors state is a sub-leading $O(\beta)$ correction. The resulting update rules are local in space: a damping coefficient is updated from an integral involving that node's displacement and velocity at negative and positive times, a conductance from the corresponding voltage products, and a spring constant from the bond's parallel stretch. Demonstrated consequences include 92.8 percent test accuracy in classifying two spoken digits with a 50-node RLC network in periodic steady state, and agreement with backpropagation when training from resting initial conditions.

Load-bearing premise

The load-bearing assumption is that replacing the time-reversed, nudged trajectory by the free, un-nudged trajectory in the nudging equation leaves the finite-difference gradient estimator unbiased as $\beta\to 0$; the paper states this is a sub-leading $O(\beta)$ correction but does not prove the convergence of the gradient.

Editorial extensions

If this is right

  • Damped mechanical metamaterials and RLC circuits can be trained in place: each learning degree of freedom updates from local measurements of its own node or bond, with no global error signal.
  • Gradient descent can now shape frequency-dependent behavior, not just static response, so a passive network can become a trained filter or a temporal classifier.
  • Because the method holds for resting initial conditions as well as periodic steady states, a single wave packet is enough for inference, which suits passive sensors.
  • Nonlinearities degrade the gradient gradually with driving amplitude, so the linear method remains usable at small but finite amplitudes.
  • Classifying temporal signals such as spoken digits can be done by a simple RLC network fed raw time-domain data, without feature extraction.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The $O(\beta)$ replacement of the time-reversed nudged trajectory by the free trajectory could itself be treated as a fixed-point iteration; solving the full time-reversed coupling would yield an exact finite-$\beta$ estimator and would directly test the paper's convergence assumption.
  • A weakly nonlinear extension may be available by treating nonlinear forces as extra driving terms in the same action; the paper's amplitude-scaling results suggest a perturbative regime where the gradient bias is controlled by strain amplitude.
  • The frequency-by-frequency training simplification the authors sketch—where the update sign depends only on the phase between response and driving—could be turned into a hardware rule that measures phase differences rather than full time-reversed integrals.
  • The same time-reversed-action structure may unify recent time-reversal-based learning schemes for Hamiltonian and Lagrangian systems, since all of them obtain gradients from paired forward and backward trajectories.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper proposes an extension of equilibrium propagation (EqProp) to linear dissipative reciprocal dynamical systems. It introduces an effective action E (Eq. (5)) whose extremum coincides with the equation of motion Hx = f (Eq. (4)), and argues that the associated boundary terms vanish under periodic boundary conditions and under resting initial conditions. Adding a cost term yields the nudged equation (7), and the time-reversed state in the nudging force is replaced by the free trajectory in Eq. (8), justified by an order-counting argument. The central finite-difference estimator (2) is then applied to spring networks and RLC circuits, with local update rules in Eqs. (9)-(11). Simulations include spring networks in periodic steady state compared with backpropagation, an RLC network trained on audioMNIST, and an RLC network trained under resting initial conditions to classify two pulses.

Significance. If correct, the result broadens equilibrium propagation from quasistatic energy-minimizing systems to damped linear reciprocal systems, which is a substantive step toward physical learning in realistic mechanical and electrical networks. The paper is honest about the degradation of the method with nonlinearity and benchmarks the gradients against backpropagation simulations. The variational identification in Appendix C is sound, and the numerical demonstrations support the main claims. The principal weakness is not an evident error but a missing formal justification of the O(β) approximation in Eq. (8), which is load-bearing for the convergence of the gradient estimator. The paper also does not provide code, data, or error bars, which limits reproducibility.

major comments (2)
  1. [Algorithm, Eq. (8)] The central gradient claim depends on replacing x_T(β) in the nudging force of Eq. (7) with the free-state x_T(0) in Eq. (8). The manuscript justifies this only by saying that it is a 'sub-leading correction of order O(β)'. Since the finite-difference estimator in Eq. (2) divides by β, the relevant question is whether the approximation error in the state is O(β²), or otherwise vanishes after the central difference, and this is not shown. Please add an explicit first-order expansion x(β) = x(0) + βδx + O(β²), verify that Eq. (8) reproduces the first-order equation Hδx = 2δ_T T(x_D − x_T(0)) implied by the exact stationarity condition (7), and show that the error in the numerator of Eq. (2) is O(β²), so that the bias vanishes as β→0. Without this argument the convergence of the learned gradients to ∇θC is an assertion rather than a derivation.
  2. [Appendix C] The stationarity of the effective action is the foundation of the method, but the boundary-term discussion is too terse. For the second-order operator H = M d²/dt² + Γ d/dt + K, please write out the boundary terms from the integration by parts in Eq. (18) explicitly, and verify separately for the periodic case and for the resting-initial-condition case that all endpoint contributions vanish. In particular, the resting-case statement should identify which derivatives of x and δx at t = −τ/2 are required and why the endpoint at t = τ/2 is controlled by x(−τ/2). This will also clarify why the same update rules apply in both regimes.
minor comments (5)
  1. [Appendix D, Eq. (19)] The sign of the parenthetical expression is potentially confusing: the term written in Eq. (19) is +2β(x_T − x_D), while the nudging force in Eq. (7) is +2β(x_D − x_T). The derivation of Eq. (7) is correct after moving the term to the right-hand side, but the sentence 'The expression in the parentheses corresponds to the nudging in eq. (7)' should be rephrased to avoid an apparent sign error.
  2. [Simulations, Figs. 2-4] The figures report medians without error bars or the number of disorder realizations; please provide both, and consider stating the number of independent training runs used for each curve.
  3. [Figures 2 and 4] Several axis labels contain rendering artifacts (e.g., '10/axisshort5' and '10□2') that make the axes unreadable; the LaTeX captions or axis labels should be fixed.
  4. [Algorithm, effective action] The claim that the Fourier construction defines a 'unique' effective action is stronger than needed; specify the normalization and acknowledge the usual total-derivative ambiguities.
  5. [Resting initial conditions] The text says EqProp is 'equivalent' to BackProp for the resting-initial-condition task, but only a single training curve is shown; please quantify the comparison (e.g., final loss values) or soften the wording.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the effective action is an explicit variational construction, and the method is benchmarked against external backpropagation.

full rationale

The core derivation is self-contained and does not reduce to its inputs. The effective action in eq. (5) is explicitly constructed so that its variation yields the equations of motion Hx = f, as verified in Appendix C. This is an inverse variational reformulation, not a fitted prediction: the action is defined in terms of the known differential operator H and forcing f, and the stationarity condition is checked directly. The gradient estimator (eq. 2) is imported from the external equilibrium-propagation literature [1, 37], and the learning rules in eqs. (9)-(11) follow by differentiating the constructed action with respect to the learning parameters. The approximation in eq. (8), replacing the time-reversed nudged trajectory by the free trajectory, is a first-order-in-β simplification and is not circular: it is an approximation whose validity is checked numerically against backpropagation. The simulations compare EqProp gradients with backpropagation, an independent external benchmark, for both spring and RLC networks. The only self-citation is ref. [44], which is used for simulation details such as backpropagation implementation and relaxation-period estimates; it does not supply the core gradient derivation or the effective action. No fitted parameter is renamed as a prediction, and no load-bearing uniqueness claim is imported from the authors' prior work. The remaining concerns, such as the rigor of the O(β) replacement and the boundary-term treatment, are correctness or presentation issues rather than circularity.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The approach introduces no new physical entities or fitted constants. The effective action is a known mathematical construction (Bateman, Feshbach, Galley) applied to EqProp; the only ad hoc element is the free-trajectory substitution in the nudging force. The paper's claims rest on standard EqProp theory and the assumed linear reciprocal dynamics.

assumptions (3)
  • domain assumption The physical systems are linear and reciprocal, so M, Γ, and K are symmetric matrices and the equations of motion are Hx = f with H = M d²/dt² + Γ d/dt + K.
    This defines the class of systems studied. Reciprocity is used in the stationarity proof to combine the two halves of the variation (Appendix C).
  • standard math Equilibrium propagation's gradient formula (eq. 2) is valid for any energy function F whose extremum gives the system state, as established by Scellier and Bengio (ref 1) and Laborieux et al. (ref 37).
    The paper relies on this theorem to go from the effective action to the gradient estimator; it is cited, not reproved.
  • ad hoc to paper The approximation x(β)_T ≈ x(0)_T in eq. (8) introduces only a sub-leading O(β) correction that vanishes as β→0, so the central finite-difference gradient remains correct in the limit.
    The authors state "Since this offers a sub-leading correction of order O(β), the algorithm still converges for β→0", but do not provide a rigorous error bound. This is load-bearing for the causality of the nudging rule.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Equilibrium Propagation for Dissipative Dynamics." pith.science (2026). https://pith.science/paper/4VDEFFOP

@misc{pith2026250620402,
  author       = {Pith},
  title        = {Pith review of: Equilibrium Propagation for Dissipative Dynamics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4VDEFFOP}},
  note         = {Machine review of arXiv:2506.20402}
}
read the original abstract

Computing gradients of a cost function is central to design-based optimization and machine learning algorithms. Equilibrium propagation provides an exact method to compute gradients in hardware by exploiting the inherent physical laws. The locality of these algorithms, in conjunction with local updates, enables mechanical and electronic systems that autonomously learn a function. We extend these methods to damped dynamical systems operating in the linear regime, such as mechanical structures obeying damped Newtonian dynamics and RLC circuits. By introducing an effective action whose extremum corresponds to the underlying dynamics, we derive local learning rules. This approach applies both to problems with periodic boundary conditions and to those with resting initial conditions. We demonstrate the viability of our method in mechanical and electronic systems and explore novel functionality such as classifying temporal sound signals. Our work opens the door to intelligent materials that process dynamical signals, enabling temporal computations, passive and active sensors, and materials that act as frequency-dependent filters.

Figures

Figures reproduced from arXiv: 2506.20402 by the authors.

Figure 1
Figure 1. An illustration of the model and an example of the training task. The model is a network (a) of bonds that can either represent damped springs (d) or RLC elements (e). A time-dependent signal is applied to a source node (b), and the network outputs (c) a desired response. In this example, the network classifies a sound signal—the spoken words “zero” or “one”. Background Equilibrium propagation (EqProp) is a method f… view at source ↗
Figure 2
Figure 2. Training spring networks to assess the effect of nonlinear￾ities. A sinusoidal displacement is applied to a source to attain sinu￾soidal motion on the target with a phase difference of 90◦. (a),(b) The median error Cnorm as a function of the epochs using EqProp and BackProp, for different amplitudes A. In (a), we modify the damp￾ing coefficients γi and in (b) the spring constants kj . In (c) and (d), we decompose th… view at source ↗
Figure 3
Figure 3. RLC network operating in periodic steady state, trained on the audioMNIST dataset to classify the spoken digits “zero” and “one”. (a) The error C for the training and validation datasets. (b) The accuracy (percentage of correctly classified inputs) for the training and validation datasets. (c),(d) Probability densities of the normalized signal energy (SE) at the target bonds used to differentiate between “zero” (c) … view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: RLC network with resting initial conditions, trained to distinguish between two distinct pulses. (a) Error C, with the gra￾dient computed using EqProp and compared against BackProp as a reference. (b),(c) The external current IS applied to the source node, for the two …

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Equilibrium Propagation for Non-Conservative Systems

    cs.LG 2026-02 conditional novelty 6.0 of 10

    A modified Equilibrium Propagation with an antisymmetric-Jacobian correction computes exact cost gradients for non-conservative neural dynamics.

Reference graph

Works this paper leans on

54 extracted references · 43 canonical work pages · cited by 1 Pith paper

  1. [1]

    & Bengio, Y

    Scellier, B. & Bengio, Y. Equilibrium propagation: Bridging the gap between energy-based models and backpropagation. Frontiers in computational neuro- science 11, 24 (2017)

    work page 2017

  2. [2]

    Stern, M., Hexner, D., Rocks, J. W. & Liu, A. J. Su- pervised learning in physical networks: From machine learning to learning machines. Physical Review X 11, 021045 (2021)

    work page 2021

  3. [3]

    Hexner, D., Liu, A. J. & Nagel, S. R. Periodic train- ing of creeping solids. Proceedings of the National Academy of Sciences 117, 31690–31695 (2020)

    work page 2020

  4. [4]

    K., Xiao, T

    Vadlamani, S. K., Xiao, T. P. & Yablonovitch, E. Physics successfully implements Lagrange multiplier optimization. Proceedings of the National Academy of Sciences 117, 26639–26650 (2020)

    work page 2020

  5. [5]

    H., Mulder, E

    Lee, R. H., Mulder, E. A. & Hopkins, J. B. Mechani- cal neural networks: Architected materials that learn behaviors. Science Robotics 7, eabq7278 (2022)

    work page 2022

  6. [6]

    R., Scellier, B

    Anisetti, V. R., Scellier, B. & Schwarz, J. M. Learning by non-interfering feedback chemical signaling in phys- ical networks. Physical Review Research 5, 023024 (2023)

    work page 2023

  7. [7]

    Falk, M. J. et al. Learning to learn by using nonequilib- rium training protocols for adaptable materials. Pro- ceedings of the National Academy of Sciences 120, e2219558120 (2023)

    work page 2023

  8. [8]

    Self-learning mechanical circuits

    Patil, V. P., Ho, I. & Prakash, M. Self-learning mechanical circuits. arXiv preprint arXiv:2304.08711 (2023)

    work page Pith review arXiv 2023

Show all 54 references
  1. [9]

    & Murugan, A

    Stern, M. & Murugan, A. Learning without neurons in physical systems. Annual Review of Condensed Matter Physics 14, 417–441 (2023)

    2023

  2. [10]

    & Coulais, C

    Du, Y., Veenstra, J., van Mastrigt, R. & Coulais, C. Metamaterials that learn to change shape. arXiv preprint arXiv:2501.11958 (2025)

    2025

  3. [11]

    Pashine, N., Hexner, D., Liu, A. J. & Nagel, S. R. Directed aging, memory, and nature’s greed. Science advances 5, eaax4215 (2019)

    2019

  4. [12]

    Stern, M., Arinze, C., Perez, L., Palmer, S. E. & Muru- gan, A. Supervised learning through physical changes in a mechanical system. Proceedings of the National Academy of Sciences 117, 14843–14850 (2020)

    2020

  5. [13]

    Dillavou, S., Stern, M., Liu, A. J. & Durian, D. J. Demonstration of decentralized physics-driven learn- ing. Physical Review Applied 18, 014040 (2022)

    2022

  6. [14]

    Momeni, A. et al. Training of physical neural networks. arXiv preprint arXiv:2406.03372 (2024)

    2024 arXiv

  7. [15]

    H., Hinton, G

    Ackley, D. H., Hinton, G. E. & Sejnowski, T. J. A learning algorithm for Boltzmann machines. Cognitive science 9, 147–169 (1985)

    1985

  8. [16]

    Movellan, J. R. in Connectionist models 10–17 (Else- vier, 1991)

    1991

  9. [17]

    Stern, M., Liu, A. J. & Balasubramanian, V. Physi- cal effects of learning. Physical Review E 109, 024311 (2024)

    2024

  10. [18]

    Stern, M., Guzman, M., Martins, F., Liu, A. J. & Bala- subramanian, V. Physical networks become what they learn. Physical Review Letters 134, 147402 (2025)

    2025

  11. [19]

    & Scellier, B

    Kendall, J., Pantone, R., Manickavasagam, K., Bengio, Y. & Scellier, B. Training end-to-end analog neural networks with equilibrium propagation.arXiv preprint arXiv:2006.01981 (2020)

    2020 arXiv

  12. [20]

    Z., Durian, D

    Stern, M., Dillavou, S., Miskin, M. Z., Durian, D. J. & Liu, A. J. Physical learning beyond the quasistatic limit. Physical Review Research 4, L022037 (2022)

    2022

  13. [21]

    Dillavou, S. et al. Machine learning without a proces- sor: Emergent learning in a nonlinear analog network. Proceedings of the National Academy of Sciences 121, e2319718121 (2024)

    2024

  14. [22]

    R., Kandala, A., Scellier, B

    Anisetti, V. R., Kandala, A., Scellier, B. & Schwarz, J. Frequency propagation: Multimechanism learning in nonlinear physical networks. Neural Computation 36, 596–620 (2024)

    2024

  15. [23]

    E., Stern, M., Liu, A

    Altman, L. E., Stern, M., Liu, A. J. & Durian, D. J. Experimental demonstration of coupled learning in elastic networks. Physical Review Applied 22, 024053 (2024)

    2024

  16. [24]

    & Marquardt, F

    Lopez-Pastor, V. & Marquardt, F. Self-learning ma- chines based on Hamiltonian echo backpropagation. Physical Review X 13, 031020 (2023). 7

    2023

  17. [25]

    Mandal, R. et al. Learning dynamical behaviors in physical systems. arXiv preprint arXiv:2406.07856 (2024)

    2024 arXiv

  18. [26]

    & Ernoult, M

    Pourcel, G. & Ernoult, M. Learning long range de- pendencies through time reversal symmetry breaking. arXiv preprint arXiv:2506.05259 (2025)

    2025 arXiv

  19. [27]

    Equilibrium Propagation for Learning in Lagrangian Dynamical Systems

    Massar, S. Equilibrium Propagation for Learning in Lagrangian Dynamical Systems. arXiv preprint arXiv:2505.07363 (2025)

    2025 arXiv

  20. [28]

    & Gilra, A

    Pourcel, G., Basu, D., Ernoult, M. & Gilra, A. Lagrangian-based Equilibrium Propagation: general- isation to arbitrary boundary conditions & equiva- lence with Hamiltonian Echo Learning. arXiv preprint arXiv:2506.06248 (2025)

    2025 arXiv

  21. [29]

    A gradient estimator for time-varying electrical networks with non-linear dissipation

    Kendall, J. A gradient estimator for time-varying electrical networks with non-linear dissipation. arXiv preprint arXiv:2103.05636 (2021)

    2021 arXiv

  22. [30]

    & Dunkel, J

    Ronellenfitsch, H., Stoop, N., Yu, J., Forrow, A. & Dunkel, J. Inverse design of discrete mechanical meta- materials. Physical Review Materials 3, 095201 (2019)

    2019

  23. [31]

    Training nonlinear elastic functions: non- monotonic, sequence dependent and bifurcating

    Hexner, D. Training nonlinear elastic functions: non- monotonic, sequence dependent and bifurcating. Soft Matter 17, 4407–4412 (2021)

    2021

  24. [32]

    & Sheng, P

    Yang, Z., Mei, J., Yang, M., Chan, N. & Sheng, P. Membrane-type acoustic metamaterial with negative dynamic mass. Physical review letters 101, 204301 (2008)

    2008

  25. [33]

    & Dambre, J

    Hermans, M., Schrauwen, B., Bienstman, P. & Dambre, J. Automated design of complex dynamic systems. PloS one 9, e86696 (2014)

    2014

  26. [34]

    Dubˇ cek, T. et al. In-sensor passive speech classifica- tion with phononic metamaterials. Advanced Func- tional Materials 34, 2311877 (2024)

    2024

  27. [35]

    Kuo, F. F. Network Analysis and Synthesis, Second Edition isbn: 9780471511182 (John Wiley & Sons, Inc., 1966)

    1966

  28. [36]

    A Treatise on the Mathematical Theory of Elasticity, Fourth Edition (Cambridge University Press, 1927)

    Love, A.E.H. A Treatise on the Mathematical Theory of Elasticity, Fourth Edition (Cambridge University Press, 1927)

    1927

  29. [37]

    Laborieux, A. et al. Scaling equilibrium propagation to deep convnets by drastically reducing its gradient estimator bias. Frontiers in neuroscience 15, 633674 (2021)

    2021

  30. [38]

    On dissipative systems and related vari- ational principles

    Bateman, H. On dissipative systems and related vari- ational principles. Physical Review 38, 815 (1931)

    1931

  31. [39]

    Morse, P. M. & Feshbach, H. Methods of theoretical physics 298–299 (McGraw-Hill, 1953)

    1953

  32. [40]

    & Tikochinsky, Y

    Feshbach, H. & Tikochinsky, Y. Quantization of the damped harmonic oscillator. Transactions of the New York Academy of Sciences 38, 44–53 (1977)

    1977

  33. [41]

    & Vitiello, G

    Celeghini, E., Rasetti, M., Tarlini, M. & Vitiello, G. SU (1, 1) squeezed states as damped oscillators. Mod- ern Physics Letters B 3, 1213–1220 (1989)

    1989

  34. [42]

    & Vitiello, G

    Celeghini, E., Rasetti, M. & Vitiello, G. Quantum dis- sipation. Annals of Physics 215, 156–170 (1992)

    1992

  35. [43]

    Galley, C. R. Classical mechanics of nonconservative systems. Physical review letters 110, 174301 (2013)

    2013

  36. [44]

    & Hexner, D

    Berneman, M. & Hexner, D. Designing precise dy- namical steady states in disordered networks.Machine Learning: Science and Technology 6, 025073. https: //dx.doi.org/10.1088/2632- 2153/ade590 (June 2025)

    2025 doi

  37. [45]

    & Sun, K

    Lubensky, T., Kane, C., Mao, X., Souslov, A. & Sun, K. Phonons and elasticity in critically coordinated lattices. Reports on Progress in Physics 78, 073901 (2015)

    2015

  38. [46]

    Hagh, V. F. & Sadjadi, M. rigidpy: Rigidity analysis in python. Computer Physics Communications 275, 108306 (2022)

    2022

  39. [47]

    Becker, S. et al. AudioMNIST: Exploring Explainable Artificial Intelligence for audio analysis on a simple benchmark. Journal of the Franklin Institute. issn: 0016-0032. https : / / www . sciencedirect . com / science/article/pii/S0016003223007536 (2023)

    2023

  40. [48]

    & Hinton, G

    LeCun, Y., Bengio, Y. & Hinton, G. Deep learning. nature 521, 436–444 (2015)

    2015

  41. [49]

    S., Silbert, L

    O’hern, C. S., Silbert, L. E., Liu, A. J. & Nagel, S. R. Jamming at zero temperature and zero applied stress: The epitome of disorder. Physical Review E 68, 011306 (2003)

    2003

  42. [50]

    Runge–Kutta pairs of order 5 (4) sat- isfying only the first column simplifying assumption

    Tsitouras, C. Runge–Kutta pairs of order 5 (4) sat- isfying only the first column simplifying assumption. Computers & Mathematics with Applications 62, 770– 775 (2011)

    2011

  43. [51]

    Bradbury, J. et al. JAX: composable transformations of Python+NumPy programs version 0.3.13. 2018. http://github.com/google/jax

    2018

  44. [52]

    On Neural Differential Equations PhD the- sis (University of Oxford, 2021)

    Kidger, P. On Neural Differential Equations PhD the- sis (University of Oxford, 2021)

    2021

  45. [53]

    Schoenholz, S. S. & Cubuk, E. D. JAX M.D. A Frame- work for Differentiable Physics in Advances in Neu- ral Information Processing Systems 33 (Curran Asso- ciates, Inc., 2020). https://papers.nips.cc/paper/ 2020 / file / 83d3d4b6c9579515e1679aca8cbc8033 - Paper.pdf

    2020

  46. [54]

    zero”, y = 1 for “one

    Kingma, D. P. & Ba, J. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014). A Disordered networks We generate disordered networks from packings of polydis- perse soft spheres [49]. The packings are prepared with 50 particles, prior to the removal ...

    2014 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.