Pith. sign in

REVIEW 3 major objections 3 minor 32 references

Relaxation of quenched structural glasses: descent in a stiffening caging potential over inflection-point 'speed bumps'

T0 review · 3 major / 3 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read The paper claims that power-law relaxation in structural glasses comes from an effective stiffening caging potential that diverges at a characteristic cage size, not from saddle points or marginal stability.

desk verdict A credible new caging-potential mechanism for GD power-law relaxation in glasses, but the central exponent relation rests on an unproven 1D reduction and a two-fit consistency check. read the letter →

arxiv 2607.25360 v1 pith:343CBN45 submitted 2026-07-28 cond-mat.dis-nn cond-mat.stat-mech

Ting Qu , Deng Pan , Yuliang Jin This is my paper · ORCID
classification cond-mat.dis-nncond-mat.stat-mech MSC 82D30
keywords glassrelaxationgradientdescentdynamicspower-lawcagingpotentialinherentstructurestatefollowingenergylandscapestructuralglasses
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

This paper aims to explain the slow, power-law energy decay seen when a quenched structural glass is relaxed by gradient descent. The authors argue that four existing explanations—saddle-point slowdown, marginal stability, phonon damping, and coarsening of localized excitations—do not fit their simulations. Their central claim is that each particle moves inside an effective stiffening cage whose potential diverges at a characteristic cage size, V_cage(x) ~ (x_c − x)^{−γ}, and that this single mechanism produces the observed decay with exponent β = (γ+1)/(γ+2). They further claim that the dynamics has no characteristic temperature, with state-following occurring at all initial temperatures near the inherent structure, in contrast to mean-field spin-glass theory. If correct, the picture reduces a many-body relaxation problem to a one-dimensional descent and unifies behavior across glass formers.

What carries the argument

The central object is the effective caging potential V_cage(x) for the average distance x to the inherent structure, with a divergent stiffening term (x_c − x)^{−γ}. This reduces the 3N-dimensional gradient-descent dynamics to a one-dimensional descent dx/dt = −V_cage'(x); the divergence at x_c generates the asymptotic law x_c − x(t) ∼ t^{1/(γ+2)} and hence v(t) ∼ t^{−(γ+1)/(γ+2)}. A minimal single-particle-in-fixed-cage model reproduces the power law, demonstrating that collective modes are not essential. Inflection-point events are identified through the eigenvalue of the instantaneous normal mode with largest overlap with the velocity field crossing zero, and these act as 'speed bumps' th

What would settle it

Measure, along a single gradient-descent trajectory, the instantaneous root-mean-square velocity v(t) and the average distance x(t) to the inherent structure, then check at every time whether v(t) equals |dV_cage/dx| evaluated at x(t) with the fitted parameters; a systematic violation anywhere in the power-law regime would refute the one-dimensional reduction.

Watch

Extended reading notes

Core claim

The discovery is that the asymptotic power-law relaxation v(t) ∼ t^{−β} in gradient-descent structural glasses is controlled by a stiffening caging potential. Tracking the average distance x(t) of the configuration to its inherent structure, the authors find the energy is well described by V_cage(x) = A[(x_c − x)^{−γ} − γ x_c^{−γ−1} x − x_c^{−γ}], which diverges at x_c, the cage size. In the limit x → x_c, integrating dx/dt = −V_cage'(x) gives x_c − x(t) ∼ t^{1/(γ+2)} and hence v(t) ∼ t^{−(γ+1)/(γ+2)}; the measured β matches (γ+1)/(γ+2). The same power law is reproduced by a minimal single-particle model with fixed neighbours, supporting the claim that collective relaxation modes are not ess

Load-bearing premise

The load-bearing premise is that the 3N-dimensional gradient descent can be reduced to a single collective coordinate x, the average distance to the inherent structure, whose motion obeys dx/dt = −V_cage'(x) with a fitted caging potential of the assumed divergent form; if the many-body dynamics is not slaved to x, the exponent relation β = (γ+1)/(γ+2) is coincidental.

Editorial extensions

If this is right

  • The power-law exponent in gradient-descent glass relaxation is set by the cage-potential exponent γ through β = (γ+1)/(γ+2), so measuring β gives a direct probe of the effective cage stiffness.
  • State-following is universal: for any initial temperature, sufficiently small perturbations near the inherent structure leave the system in the same glass basin, and no T_SF separates dynamical regimes.
  • Glass basins are stable, not marginally stable; the long-time relaxation is exponential and governed by the softest disordered mode of the inherent structure.
  • The caging mechanism is generic across several structural glass formers, with the caveat that jamming-adjacent granular systems may behave differently.
  • Inflection points act as speed bumps, producing velocity fluctuations but not controlling the exponent; no saddle points are encountered along the descent.

Reading between the lines

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

  • Inference: If the single-coordinate caging-potential description is the true mechanism, then deliberately changing the cage geometry or the short-range repulsion in simulations should shift γ and β together in the predicted way—a testable extension not reported in the paper.
  • Inference: The same reduction may apply to other conservative descent problems, such as training dynamics in neural networks, where effective 'cages' would be loss-landscape confinements; the paper hints at this in its final remark but does not develop it.
  • Inference: The absence of a T_SF suggests that the mean-field spin-glass state-following threshold is an artifact of infinite-range interactions, so finite-dimensional structural glasses may need a different spin-glass analogue; this could be checked by studying finite-size scaling of the escape probability.
  • Inference: The reported logarithmic coarsening of hot spots contradicts power-law coarsening found in jamming systems; a direct comparison of the same correlation function near jamming versus ordinary repulsive glasses would help locate the boundary of the caging mechanism's validity.
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

3 major / 3 minor

Summary. The paper studies zero-temperature gradient-descent relaxation in structural glasses (3D KALJ and amorphous silica, plus a 2D minimal cage model). It reports power-law decay of the mean velocity, v(t) ~ t^{-β}, and proposes that this decay is controlled by an effective stiffening cage potential V_cage(x) ~ (x_c - x)^{-γ}, where x is the RMS distance to the inherent structure. A single-coordinate integration argument gives β = (γ+1)/(γ+2), which is compared with fits to simulation data. The paper also attributes velocity fluctuations to crossings of inflection points ('speed bumps'), argues that saddle-point/marginal-stability, phonon, coarsening, and jamming mechanisms are not responsible, and claims state-following occurs at all studied temperatures below a perturbation-size boundary, implying no characteristic temperature T_SF.

Significance. If the central claim is correct, it constitutes a substantial advance: it replaces the mean-field saddle-point/marginal-stability picture with a simple local-caging mechanism and makes a testable prediction about the power-law exponent. The simulations are carefully done, and the exclusion of phonon, jamming, coarsening, and saddle-point scenarios is a useful contribution. However, the central analytical derivation is not yet established: it relies on an untested reduction of the many-body dynamics to a single collective coordinate, and on a phenomenological potential fitted to the same data whose relaxation it is used to explain. The paper's main claim therefore currently rests on a consistency check, not on an independent derivation.

major comments (3)
  1. [§3 (Effective stiffening caging potential), Eq. (4)] The derivation of β=(γ+1)/(γ+2) assumes that the 3N-dimensional gradient descent of Eq. (1) can be replaced by a single-coordinate equation dx/dt = -V_cage'(x), with x the RMS distance to the inherent structure and with the measured mean velocity v(t) identified as dx/dt. This reduction is asserted, not derived; the text near Eq. (4) states that the equation 'greatly simplifies' the description, but no argument or simulation test is given. In the power-law regime, Fig. 3c shows that the overlap of the velocity field with the softest mode of the inherent structure is small; if v(t) contains substantial components transverse to the cage-coordinate direction, the decay of v(t) need not be controlled by V_cage(x). A direct test is needed: compute |dx/dt| and v(t) from the same trajectories and show they are proportional, or at least show that the projection of the velocity onto the reaction
  2. [Eq. (4) and Fig. 2b-c] The potential V_cage is phenomenological: A, x_c, and γ are fitted to ΔE(x) from the same trajectories whose relaxation is being explained, and β is fitted independently to v(t). Therefore the relation β=(γ+1)/(γ+2) is a self-consistency check for the chosen fitting form, not an analytically derived prediction. To make the central claim load-bearing, the paper should either derive V_cage from the microscopic interactions or cage statistics, or validate out-of-sample: for example, fix γ from some systems or temperatures and predict β for held-out data without refitting. The minimal cage model is a separate, qualitative argument and does not by itself resolve this circularity.
  3. [Methods Appendix C and §3] The minimal cage model is an existence proof that a single particle in a fixed LJ cage can produce a similar power-law relaxation, but it does not establish that the many-body dynamics is governed by such single-particle cages. In the real system the cage itself relaxes and deforms during descent. The claim that 'collective relaxation modes are not essential' requires a quantitative comparison between the many-body effective coordinate and the MCM trajectories—for instance, the same relation between x(t), v(t), and the energy—rather than only similar exponents. As written, the MCM reproduces a power law in a simpler model but does not demonstrate that the original system's power law arises from that mechanism.
minor comments (3)
  1. [Fig. 2c] No error bars are shown for β and γ. Both exponents come from fits over finite time windows, so uncertainty estimates are needed to assess whether the scatter in Fig. 2c is statistically significant.
  2. [State-following section, Fig. 6] The threshold p_threshold = 0.3 is arbitrary. The collapse of ϵ*(x;T) in Fig. 6b should be checked for a range of thresholds (e.g., 0.1 and 0.5) to ensure the conclusion that no T_SF exists is not an artifact of this choice.
  3. [General notation, Figs. 3-4] The overlaps O_IS^0(t), O_+(t), and O_-(t) are defined somewhat tersely. Please state explicitly whether these are squared overlaps, whether they are sample-averaged or from a single representative sample, and how the 'largest overlap' mode is selected when multiple modes are nearly degenerate.

Circularity Check

1 steps flagged · score 6.0 of 10

The central exponent relation β=(γ+1)/(γ+2) is a consistency check between two fits to the same GD trajectories: γ is fitted from ΔE(x) via the phenomenological V_cage, and β is then read off from that fit form rather than predicted from the microscopic dynamics.

  1. fitted input called prediction [Section 'An effective stiffening caging potential for power-law relaxation', Eq. (4) and Fig. 2c]
    "The overall behavior of E(x) can be well described by a phenomenological potential (Fig. 2b): V_cage(x) = A[(x_c − x)^{−γ} − γ x_c^{−γ−1} x − x_c^{−γ}] (4), where A, x_c, and γ are fitting parameters. ... This relationship is confirmed by our simulation data, using the two exponents γ and β obtained from fits at different T (see Fig. 2c,d)."

    The analytically derived exponent β=(γ+1)/(γ+2) is obtained by integrating the assumed divergent form V_cage≈A(x_c−x)^{−γ} in the 1D equation dx/dt=−V_cage'(x). But γ is a free parameter fitted to ΔE(x) from the very same GD trajectories whose v(t) relaxation is being explained, and β is independently fitted from those same trajectories. Fig. 2c therefore compares a derived function of one fit parameter with another fit on the same dataset; Eq. (4) was chosen precisely because it fits those data, so the 'prediction' is largely an algebraic consequence of the fitted ansatz. The reduction of the 3N-dimensional dynamics to the single coordinate x is asserted ('Equation (4) greatly simplifies the description...'), and v(t) is implicitly identified with |dx/dt| without checking that transverse

full rationale

The paper's central quantitative claim is that the power-law exponent follows from a stiffening caging potential. That claim is partially circular in the specific sense defined here: V_cage(x) is not derived from the microscopic pair potential or from an independent measurement; it is a phenomenological form fitted to ΔE(x), with A, x_c and γ as fitting parameters. The subsequent relation β=(γ+1)/(γ+2) is then obtained by analytically integrating this fitted form. Since γ and β are extracted from the same simulation dataset, Fig. 2c is an internal consistency check rather than an independent, parameter-free prediction. This prevents the strongest claim from being fully self-contained. However, the paper is not merely renaming a known result: the toy minimum-cage model independently exhibits power-law relaxation, the comparison with amorphous silica and other models provides some generalization, and the saddle-point/inflection-point analyses do not reduce to the fitted potential. The self-citations to the authors' own prior work (e.g., Ref. [18] for the saddle-dynamics search) are not load-bearing enough to constitute circularity by themselves, and no uniqueness theorem is imported. Overall, the central exponent relation is partially circular because it reduces to fitting the same trajectories twice, but the separate toy model and empirical comparisons give the paper some independent content, yielding a score of 6 rather than higher.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central analytic 'derivation' of the power-law exponent rests on a three-parameter fit of an assumed functional form to the simulated ΔE(x), plus an unproven reduction to one-dimensional dynamics. The toy model is illustrative but adds no independent quantitative constraints. The free parameters (γ, x_c, A) are therefore the main content of the explanation.

free parameters (4)
  • γ (divergence exponent in V_cage) = varies with T; e.g., ~4.8–7.1 in MCM; for 3D KALJ values in Fig. 2d (not reported numerically)
    Fitted to the simulated ΔE(x) curves; directly controls the predicted exponent β=(γ+1)/(γ+2).
  • x_c (cage size) = 0.18–0.41 (3D KALJ), 0.66–0.91 (MCM)
    Fitted position of the divergence in Eq. (4), interpreted post hoc as cage size.
  • A (amplitude) = not reported
    Overall scale of the fitted V_cage; contributes to λ in the harmonic limit but not to the exponent.
  • p_threshold = 0.3 = 0.3
    Arbitrary threshold defining the state-following boundary ϵ*(x); the collapse in Fig. 6b depends on this choice.
assumptions (4)
  • ad hoc to paper The total energy as a function of the average distance to the inherent structure is well represented by Eq. (4), with a power-law divergence at x_c.
    This functional form is chosen because it fits the data; it is not derived from the microscopic pair potential. It is the load-bearing input for the analytic exponent.
  • domain assumption The 3N-dimensional GD can be reduced to a single coordinate x with dx/dt = -V_cage'(x), and the RMS velocity v(t) is identified with dx/dt.
    Introduced in the section 'An effective stiffening caging potential'; no derivation from the many-body equations. The exponent relation depends on this 1D overdamped dynamics.
  • domain assumption The initial configurations are equilibrium Boltzmann samples at temperature T, and GD with zero noise represents the quench dynamics.
    Standard simulation setup, stated in Methods.
  • standard math Instantaneous normal mode analysis (Hessian eigenvalues/eigenvectors) provides a valid description of the local landscape near the trajectory.
    Used to define inflection-point crossings and f_u(t).
invented entities (1)
  • Effective stiffening caging potential V_cage(x)
    purpose: To explain the power-law relaxation of the average energy/velocity in terms of a single collective coordinate.
    The potential is fitted to the same simulation data whose relaxation it purports to explain; no independent measurement of V_cage is provided. The toy cage model is a separate construction but not a direct measurement of V_cage.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Relaxation of quenched structural glasses: descent in a stiffening caging potential over inflection-point 'speed bumps'." pith.science (2026). https://pith.science/paper/343CBN45

@misc{pith2026260725360,
  author       = {Pith},
  title        = {Pith review of: Relaxation of quenched structural glasses: descent in a stiffening caging potential over inflection-point 'speed bumps'},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/343CBN45}},
  note         = {Machine review of arXiv:2607.25360}
}
read the original abstract

The slow energy relaxation in quenched glasses is a ubiquitous yet poorly understood phenomenon. Despite extensive study, the microscopic origin of the observed power-law decay remains debated, with proposed mechanisms ranging from saddle-point slowdown and marginal stability to coarsening of localized excitations and phonon dynamics. Here, by simulating gradient descent in archetypal structural glass formers, we show that none of these scenarios can account for our data. Instead, the power-law behavior emerges from a remarkably simple caging effect: each particle experiences an effective stiffening potential that arises from many-body confinement and diverges at a characteristic cage size. This mechanism analytically yields the observed power-law decay and is quantitatively reproduced by a minimal single-particle cage model with fixed neighbours, demonstrating that collective relaxation modes are not essential. The dynamics is punctuated by fluctuations as the system rolls through inflection points on the energy landscape, which act as `speed bumps' but do not affect the overall power-law behaviour. In contrast to mean-field spin glass theory, we find no characteristic temperature that separates distinct dynamical regimes; state following within a given glass basin occurs universally for all initial temperatures whenever the system is sufficiently close to the inherent structure. Our results establish a complete physical picture of gradient descent dynamics in typical structural glasses.

Figures

Figures reproduced from arXiv: 2607.25360 by the authors.

Figure 1
Figure 1. FIG. 1 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. a), which covers the onset temperature Tonset ≈ 0.7 (see Supplementary Section S2 for the estimation of Tonset). The power-law behavior over such a wide range of T simply cannot be explained by the phonon mecha￾nism. Next, we examine the coarsening mechanism (ii), and find that the coarsening length l ∗ grows logarithmically, l ∗ ∼ ln t, instead of following a power-law (see Supple￾mentary Section S3). In fact, l ∗ … view at source ↗
Figure 4
Figure 4. FIG. 4 [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (3 more)
Figure 3
Figure 3. Figure 3: a is analyzed in detail here. (a) Mean velocity v(t). (b) Largest overlap O + between vˆ(t) and the stable instantaneous normal mode xn∗ + . (c) Largest overlap O − between vˆ(t) and the unstable instantaneous normal mode xn∗ − . (d) Eigenvalue λ ∗ of the instantaneous…
Figure 5
Figure 5. Figure 5: FIG. 5 [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6 [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 3 linked inside Pith

  1. [1]

    C. A. Angell, Science267, 1924 (1995)

    1924

  2. [2]

    Baity-Jesi, L

    M. Baity-Jesi, L. Sagun, M. Geiger, S. Spigler, G. B. Arous, C. Cammarota, Y. LeCun, M. Wyart, and G. Biroli, inProceedings of the 35th International Con- ference on Machine Learning, Proceedings of Machine Learning Research, Vol. 80, edited by J. Dy and A. Krause (PMLR, 2018) pp. 314–323

    2018

  3. [3]

    Huang, L

    G. Huang, L. S. Chan, H. Yoshino, G. Zhang, and Y. Jin, arXiv preprint arXiv:2506.06789 (2025)

    arXiv 2025

  4. [4]

    Frauenfelder, S

    H. Frauenfelder, S. G. Sligar, and P. G. Wolynes, Science 254, 1598 (1991)

    1991

  5. [5]

    Kurchan and L

    J. Kurchan and L. Laloux, Journal of Physics A: Math- ematical and General29, 1929 (1996)

    1929

  6. [6]

    Folena, S

    G. Folena, S. Franz, and F. Ricci-Tersenghi, Physical Re- view X10, 031045 (2020)

    2020

  7. [8]

    Ikeda, T

    A. Ikeda, T. Kawasaki, L. Berthier, K. Saitoh, and T. Hatano, Physical review letters124, 058001 (2020)

    2020

  8. [10]

    Shimada, K

    M. Shimada, K. Shiraishi, H. Mizuno, and A. Ikeda, Soft Matter20, 1583 (2024)

    2024

Show all 32 references
  1. [11]

    Manacorda and F

    A. Manacorda and F. Zamponi, Journal of Physics A: Mathematical and Theoretical55, 334001 (2022)

    2022

  2. [12]

    Saitoh, T

    K. Saitoh, T. Hatano, A. Ikeda, and B. P. Tighe, Physical review letters124, 118001 (2020)

    2020

  3. [13]

    K. A. Interiano-Alberto, P. K. Morse, and R. S. Hoy, Physical Review E109, L062603 (2024)

    2024

  4. [14]

    Folena, A

    G. Folena, A. Manacorda, and F. Zamponi, Physica A: Statistical Mechanics and its Applications631, 128152 (2023)

    2023

  5. [15]

    Franz and G

    S. Franz and G. Parisi, Journal de Physique I5, 1401 (1995)

    1995

  6. [16]

    Rainone, P

    C. Rainone, P. Urbani, H. Yoshino, and F. Zamponi, Physical review letters114, 015701 (2015)

    2015

  7. [17]

    Zdeborov´ a and F

    L. Zdeborov´ a and F. Krzakala, Physical Review B—Condensed Matter and Materials Physics81, 224205 (2010)

    2010

  8. [18]

    S. Wu, Y. Xie, D. Pan, L. Zhang, and Y. Jin, Physical Review Letters136, 198202 (2026)

    2026

  9. [19]

    Y. Oku, K. Takae, and A. Ikeda, arXiv preprint arXiv:2004.02103 (2020)

    2004 arXiv

  10. [20]

    Doi,Introduction to polymer physics(Oxford univer- sity press, 1996)

    M. Doi,Introduction to polymer physics(Oxford univer- sity press, 1996)

    1996

  11. [21]

    D. Pan, F. Meng, and Y. Jin, PNAS nexus2, pgad047 (2023)

    2023

  12. [22]

    Y. Zhao, D. Pan, Y. Pang, J. Bar´ es, C. Xu, C. Liu, H. Hu, Y. Jin, and Q. Xu, arXiv preprint arXiv:2604.20437 (2026)

    2026 arXiv

  13. [23]

    Lerner and E

    E. Lerner and E. Bouchbinder, The Journal of chemical physics155(2021)

    2021

  14. [24]

    A. J. Liu and S. R. Nagel, Annu. Rev. Condens. Matter Phys.1, 347 (2010)

    2010

  15. [25]

    Charbonneau, J

    P. Charbonneau, J. Kurchan, G. Parisi, P. Urbani, and F. Zamponi, Nature communications5, 3725 (2014)

    2014

  16. [26]

    M¨ uller and M

    M. M¨ uller and M. Wyart, Annu. Rev. Condens. Matter Phys.6, 177 (2015)

    2015

  17. [27]

    Berthier, P

    L. Berthier, P. Charbonneau, Y. Jin, G. Parisi, B. Seoane, and F. Zamponi, Proceedings of the National Academy of Sciences113, 8397 (2016)

    2016

  18. [28]

    Jin and H

    Y. Jin and H. Yoshino, Nature communications8, 14935 9 (2017)

    2017

  19. [29]

    Y. Jin, P. Urbani, F. Zamponi, and H. Yoshino, Science advances4, eaat6387 (2018)

    2018

  20. [30]

    Pedone, G

    A. Pedone, G. Malavasi, M. C. Menziani, A. N. Cormack, and U. Segre, The Journal of Physical Chemistry B110, 11780 (2006)

    2006

  21. [31]

    J. F. Ziegler and J. P. Biersack, inTreatise on Heavy- Ion Science: Volume 6: Astrophysics, Chemistry, and Condensed Matter, edited by D. A. Bromley (Springer, Boston, MA, 1985) pp. 93–129. 10 Methods Appendix A: Simulation model and method for the KALJ model.The system consis...

    1985

  22. [32]

    Nishikawa, M

    Y. Nishikawa, M. Ozawa, A. Ikeda, P. Chaudhuri, and L. Berthier, Physical Review X12, 021001 (2022)

    2022

  23. [33]

    Banerjee, M

    A. Banerjee, M. K. Nandi, S. Sastry, and S. Maitra Bhattacharyya, The Journal of Chemical Physics147, 024504 (2017)

    2017

  24. [34]

    R. N. Chacko, P. Sollich, and S. M. Fielding, Physical review letters123, 108001 (2019)

    2019

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.