Numerical renormalization of glassy dynamics
Johannes Lang, Subir Sachdev, Sebastian Diehl · 2025
Read Numerical renormalization of glassy dynamics by Johannes Lang, Subir Sachdev, Sebastian Diehl (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-06849.txt.
Introduction
Johannes Lang and co-authors present a faster way to simulate glassy systems that never settle, in a 2025 paper. When a glass is quenched, it ages: it keeps changing on time scales that grow with its own age. It never reaches a steady state. That endless drift makes the math costly, with effort growing as the cube of the time span.
Their method stores the system's response curves on a two-dimensional grid and fills gaps by smart interpolation. The cost then grows more slowly than the time span itself. They ran it on a classic spin glass model and reached times a thousand times longer than before. At those spans, a sharp change point showed up between glasses that forget their past well and those that do not. Oddly, the change traits vary with the setup rather than staying fixed.
Catalog record
| CREATOR | Johannes Lang, Subir Sachdev, Sebastian Diehl |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 33 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-06849
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Numerical renormalization of glassy dynamics Johannes Lang, Subir Sachdev, Sebastian Diehl Abstract The quench dynamics of glassy systems are challenging. Due to aging, the system never reaches a stationary state but instead evolves on emergent scales that grow with its age. This slow evolution complicates field-theoretic descriptions, as the weak long-term memory and the absence of a stationary state hinder simplifications of the memory, always leading to the worst-case scaling of computational effort with the cubic power of the simulated time. Here, we present an algorithm based on two-dimensional interpolations of Green’s functions, which resolves this issue and achieves sublinear scaling of computational cost. We apply it to the quench dynamics of the spherical mixed p p -spin model to establish the existence of a phase transition between glasses with strong and weak ergodicity breaking at a finite temperature of the initial state. By reaching times three orders of magnitude larger than previously attainable, we determine the critical exponents of this transition. Interestingly, these are continuously varying and, therefore, non-universal. While we introduce and validate the method in the context of a glassy system, it is equally applicable to any model with overdamped excitations. Introduction — Optimizing complex, high-dimensional landscapes is a central challenge in computer science, inspiring influential algorithms [ 1 , 2 , 3 ] . This challenge has gained renewed significance with the rise of artificial intelligence, where deep learning models exhibit remarkable generalization despite being massively overparameterized. Yet, why and how neural networks consistently find solutions that generalize well remains one of the most profound open questions in computer science [ 4 ] . In the absence of a fundamental theory, many successful learning strategies remain heuristic, often drawing inspiration from the physics of disordered systems [ 5 , 6 , 7 , 8 ] . Indeed, it has been shown that stochastic gradient descent, the backbone of deep learning, maps onto the relaxational dynamics of mean-field spin glasses [ 9 , 10 , 11 , 12 ] . This connection highlights the importance of understanding dynamical phase transitions in spin glasses, as jamming transitions are associated with learnability thresholds, while strong ergodicity breaking corresponds to underparameterization, where the network fails to explore all relevant configurations [ 13 , 8 , 12 , 10 ] . While analytical solutions exist for specific mean-field models [ 14 , 15 ] , addressing these questions generally requires numerically evolving thermodynamically large systems over extremely long timescales. The efficient simulation of spin glass dynamics is therefore crucial not only for statistical physics but also for advancing our theoretical understanding of deep learning. The aging dynamics of mean-field spin glasses arises from marginal stability, whereby the evolution in high-dimensional configuration space gets stuck in saddles, not local minima. Consequently, all mean-field spin glasses are (self-organized) critical [ 16 , 17 , 18 ] , as seen in the emergent scale invariance of correlations and response functions [ 19 ] . The latter is broken on a scale set by the age of the system t , making it natural to address aging and the approach to marginal stability from the vantage point of the renormalization group (RG). Although the interaction constant is not renormalized, due to the weak long-term memory characteristic of glassy systems, all coarse features of the correlation and response functions of scale \sim t are relevant, whereas most finer details of the distant past are forgotten and thus irrelevant in the RG sense. In this work, we introduce an efficient algorithm that exploits this RG flow by systematically discarding irrelevant details of the past through adaptive rescaling of the discretization. Crucially, this rescaling is performed at a fixed cost per time step, ensuring efficiency even for arbitrarily long memory effects. We apply it to the dynamics of the mixed spherical p -spin model after a quench to zero temperature from a finite temperature state, resolving the previously inaccessible phase diagram and identifying, for the first time, the critical exponents of the weak-to-strong ergodicity breaking transition. Keldysh approach to quench dynamics — The time-evolution of any interacting system can be described in terms of correlations C(t,t^{\prime}) and response functions R(t,t^{\prime})=\bar{R}(t^{\prime},t) . These Green’s functions satisfy the Kadanoff-Baym equations [ 20 ] \displaystyle\begin{split}R_{0}^{-1}\star C=&\Sigma^{R}\star C+\Sigma^{C}\star\bar{R}\,,\\ R_{0}^{-1}\star R=&\Sigma^{R}\star R+\delta(t-t^{\prime})\,.\end{split} (1) Here, R_{0} is the response function of the non-interacting theory and \Sigma^{C/R} are so-called self-energies that can be calculated from two-particle irreducible Feynman diagrams [ 21 , 22 ] . Furthermore, A\star B=\int_{s}A(t,s)B(s,t^{\prime}) represents a convolution. For simplicity, we neglect the possibility of spatial dependence. Although formally exact, Eqs. ( 1 ) typically require some approximation for the self-energies. The same problem occurs in equilibrium and can be solved in many cases (see e.g. [ 23 , 24 ] ). Thermalization requires energy redistribution and thus self-energies that are non-local in time. Since R_{0}^{-1} can be written as derivative operators, out of equilibrium the equations of motion ( 1 ) take the form of partial integro-differential equations. The computation effort to evolve the system to an age t with a fixed time step and without further approximations scales at least as t^{3} [ 25 , 26 , 27 , 28 ] . As glasses break ergodicity, equilibration and accordingly time-translation invariance is dynamically unachievable [ 19 , 29 ] . Consequently, standard approximations like memory integral truncation [ 30 ] , generalized Kadanoff-Baym [ 31 ] , or the Wigner approximation [ 32 ] are not applicable to glassy dynamics. This calls for an adaptive approach to time discretization–one that captures all relevant scales while remaining computationally efficient. Dynamical renormalization algorithm — To solve the Kadanoff-Baym equations we aim to evolve two-point functions A(t,t^{\prime}) in time. Even at times much greater than any microscopic scale intrinsic to the system t\gg\tau_{\text{micro}} , they strongly depend on the relative time \tau=t-t^{\prime} for \tau\lesssim\tau_{\text{micro}} . Furthermore, if the dynamics are initiated by a quench, there may also be a strong dependence on t^{\prime} for t^{\prime}\lesssim\tau_{\text{micro}} , regardless of t . We therefore parametrize the A(t,t^{\prime}) as \mathcal{A}(t,\theta)\equiv A(t,\theta t) and choose a discretization t^{\prime}_{i}=t\theta_{i} , where \theta_{i}\in[0,1] is a fixed irregular grid of length N that is dense near 0 and 1. The precise choice of the \theta_{i} is irrelevant as long as \displaystyle\theta_{i+1}-\theta_{i}\ll\frac{\tau_{\text{micro}}}{t}\quad\forall i:\min(\theta_{i},1-\theta_{i})\lesssim\frac{\tau_{\text{micro}}}{\min{(t^{\prime},\tau)}} (2) for all simulated times t . The specific grid used in our calculations, together with some further details of the implementation, are given in the supplemental material. In principle, this grid could be optimized further by making it adaptive. However, this has turned out to be unnecessary for all our simulations, which use an adaptive discretization exclusively of t . From the RG perspective, keeping the grid length N fixed under a scale transformation/time evolution t\to bt corresponds to a coarse graining step. Due to condition ( 2 ), it retains sufficient resolution of microscopic scales for \tau\lesssim\tau_{\text{micro}} and t^{\prime}\lesssim\tau_{\text{micro}} . Continuing in the RG logic, as the \theta_{i} are dimensionless, they require no rescaling, which leaves the renormalization step, which is performed by the explicit propagation of the Kadanoff-Baym equations. Aging of classical glasses is then characterized by C\to C and R\to b^{-1}R for t^{\prime}\sim t and R\to R for \tau\lesssim\tau_{\text{micro}} . In the numerical implementation,