Disordered Systems and Neural Networks
See recent articles
- [1] arXiv:2407.13039 [pdf, html, other]
-
Title: Self-consistent theory for sound propagation in a simple model of a disordered, harmonic solidComments: 13 pagesSubjects: Disordered Systems and Neural Networks (cond-mat.dis-nn); Statistical Mechanics (cond-mat.stat-mech)
We present a self-consistent theory for sound propagation in a simple model of a disordered solid. The solid is modeled as a collection of randomly distributed particles connected by harmonic springs with strengths that depend on the interparticle distances, i.e the Euclidean random matrix model of Mezard et al. [Nucl. Phys. 559B, 689 (1999)]. The derivation of the theory combines two exact projection operator steps and a factorization approximation. Within our approach the square of the speed of sound is non-negative. The unjamming transition manifests itself through vanishing of the speed of sound.
- [2] arXiv:2407.13456 [pdf, html, other]
-
Title: Magnetic exponent for the long-range bond disordered Potts modelComments: 21 pages, 7 figures, 4 tablesSubjects: Disordered Systems and Neural Networks (cond-mat.dis-nn); High Energy Physics - Theory (hep-th)
We consider the critical behavior of two-dimensional Potts models in presence of a bond disorder in which the correlation decays as a power law. In some recent work the thermal sector of this theory was investigated by a renormalization group computation based on perturbed conformal field theory. Here we apply the same approach to study instead the magnetic sector. In particular we compute the leading corrections to the Potts spin scaling dimension. Our results include as a special case the long-range disorder Ising model. We compare our prediction to Monte-Carlo simulations. Finally, by studying the magnetization scaling function, we show a clear numerical evidence of a cross-over between the long-range and the short-range class of universality.
- [3] arXiv:2407.13655 [pdf, html, other]
-
Title: Can dissipation induce a transition between many-body localized and thermal states?Subjects: Disordered Systems and Neural Networks (cond-mat.dis-nn); Statistical Mechanics (cond-mat.stat-mech); Optics (physics.optics); Quantum Physics (quant-ph)
The many-body mobility edge (MBME) in energy, which separates thermal states from many-body localization (MBL) states, is a critical yet controversial concept in many-body systems. Here we examine the quasiperiodic $t_1-t_2$ model that features a mobility edge. With the addition of nearest-neighbor interactions, we demonstrate the potential existence of a MBME. Then we investigate the impact of a type of bond dissipation on the many-body system by calculating the steady-state density matrix and analyzing the transport behavior, and demonstrate that dissipation can cause the system to predominantly occupy either the thermal region or the MBL region, irrespective of the initial state. Finally, we discuss the effects of increasing system size. Our results indicate that dissipation can induce transitions between thermal and MBL states, providing a new approach for experimentally determining the existence of the MBME.
New submissions for Friday, 19 July 2024 (showing 3 of 3 entries )
- [4] arXiv:2407.12909 (cross-list from cond-mat.stat-mech) [pdf, html, other]
-
Title: State-dependent mobility edge in kinetically constrained modelsComments: 12+3 pages; 8+5 figuresSubjects: Statistical Mechanics (cond-mat.stat-mech); Disordered Systems and Neural Networks (cond-mat.dis-nn); Quantum Physics (quant-ph)
In this work, we show that the kinetically constrained quantum East model lies between a quantum scarred and a many-body localized system featuring an unconventional type of mobility edge in the spectrum. We name this scenario $\textit{state-dependent}$ mobility edge: while the system does not exhibit a sharp separation in energy between thermal and non-thermal eigenstates, the abundance of non-thermal eigenstates results in slow entanglement growth for $\textit{many}$ initial states, such as product states, below a finite energy density. We characterize the state-dependent mobility edge by looking at the complexity of classically simulating dynamics using tensor network for system sizes well beyond those accessible via exact diagonalization. Focusing on initial product states, we observe a qualitative change in the dynamics of the bond dimension needed as a function of their energy density. Specifically, the bond dimension typically grows $\textit{polynomially}$ in time up to a certain energy density, where we locate the state-dependent mobility edge, enabling simulations for long times. Above this energy density, the bond dimension typically grows $\textit{exponentially}$ making the simulation practically unfeasible beyond short times, as generally expected in interacting theories. We correlate the polynomial growth of the bond dimension to the presence of many non-thermal eigenstates around that energy density, a subset of which we compute via tensor network.
- [5] arXiv:2407.13444 (cross-list from q-bio.PE) [pdf, html, other]
-
Title: A symmetry-based approach to species-rich ecological communitiesComments: 12 pages, 3 figuresSubjects: Populations and Evolution (q-bio.PE); Disordered Systems and Neural Networks (cond-mat.dis-nn); Statistical Mechanics (cond-mat.stat-mech)
Disordered systems theory provides powerful tools to analyze the generic behaviors of highdimensional systems, such as species-rich ecological communities or neural networks. By assuming randomness in their interactions, universality ensures that many microscopic details are irrelevant to system-wide dynamics; but the choice of a random ensemble still limits the generality of results. We show here, in the context of ecological dynamics, that these analytical tools do not require a specific choice of ensemble, and that solutions can be found based only on a fundamental rotational symmetry in the interactions, encoding the idea that traits can be recombined into new species without altering global features. Dynamical outcomes then depend on the spectrum of the interaction matrix as a free parameter, allowing us to bridge between results found in different models of interactions, and extend beyond them to previously unidentified behaviors. The distinctive feature of ecological models is the possibility of species extinctions, which leads to an increased universality of dynamics as the fraction of extinct species increases. We expect that these findings can inform new developments in theoretical ecology as well as for other families of complex systems.
- [6] arXiv:2407.13767 (cross-list from cond-mat.mes-hall) [pdf, html, other]
-
Title: Topological insulators on fractal lattices: A general principle of constructionComments: 6 Pages, 3 Figures and 1 TableSubjects: Mesoscale and Nanoscale Physics (cond-mat.mes-hall); Disordered Systems and Neural Networks (cond-mat.dis-nn); Materials Science (cond-mat.mtrl-sci)
Fractal lattices, featuring the self-similarity symmetry, are often geometric descents of parent crystals, possessing all their discrete symmetries (such as rotations and reflections) except the translational ones. Here, we formulate three different general approaches to construct real space Hamiltonian on a fractal lattice starting from the Bloch Hamiltonian on the parent crystal, fostering for example strong and crystalline topological insulators resulting from the interplay between the nontrivial geometry of the underlying electronic wavefunctions and the crystal symmetries. As a demonstrative example, we consider a generalized square lattice Chern insulator model, and within the framework of all three methods we successfully showcase incarnations of strong and crystalline Chern insulators on the Sierpiński carpet fractal lattices. The proposed theoretical framework thus lays a generic foundation to build a tower of topological phases on the landscape of fractal lattices.
Cross submissions for Friday, 19 July 2024 (showing 3 of 3 entries )
- [7] arXiv:2310.14621 (replaced) [pdf, html, other]
-
Title: Spiking mode-based neural networksComments: 30 pages, 10 figures, submitted to PRESubjects: Neurons and Cognition (q-bio.NC); Disordered Systems and Neural Networks (cond-mat.dis-nn); Artificial Intelligence (cs.AI); Neural and Evolutionary Computing (cs.NE)
Spiking neural networks play an important role in brain-like neuromorphic computations and in studying working mechanisms of neural circuits. One drawback of training a large scale spiking neural network is that updating all weights is quite expensive. Furthermore, after training, all information related to the computational task is hidden into the weight matrix, prohibiting us from a transparent understanding of circuit mechanisms. Therefore, in this work, we address these challenges by proposing a spiking mode-based training protocol, where the recurrent weight matrix is explained as a Hopfield-like multiplication of three matrices: input, output modes and a score matrix. The first advantage is that the weight is interpreted by input and output modes and their associated scores characterizing the importance of each decomposition term. The number of modes is thus adjustable, allowing more degrees of freedom for modeling the experimental data. This significantly reduces the training cost because of significantly reduced space complexity for learning. Training spiking networks is thus carried out in the mode-score space. The second advantage is that one can project the high dimensional neural activity (filtered spike train) in the state space onto the mode space which is typically of a low dimension, e.g., a few modes are sufficient to capture the shape of the underlying neural manifolds. We successfully apply our framework in two computational tasks -- digit classification and selective sensory integration tasks. Our method accelerate the training of spiking neural networks by a Hopfield-like decomposition, and moreover this training leads to low-dimensional attractor structures of high-dimensional neural dynamics.
- [8] arXiv:2404.03750 (replaced) [pdf, html, other]
-
Title: Dynamical heterogeneity and large deviations in the open quantum East glass model from tensor networksComments: 5+11 pages, 4+6 figures, added SMSubjects: Statistical Mechanics (cond-mat.stat-mech); Disordered Systems and Neural Networks (cond-mat.dis-nn); Quantum Physics (quant-ph)
We study the non-equilibrium dynamics of the dissipative quantum East model via numerical tensor networks. We use matrix product states to represent evolution under quantum-jump unravellings for sizes beyond those accessible to exact diagonalisation. This allows us to demonstrate that dynamical heterogeneity accompanies slow relaxation, in analogy with what is seen in classical glassy systems. Furthermore, using variational matrix product operators we: (i) compute the spectral gap of the Lindbladian, and show that glassiness is enhanced in the presence of weak quantum fluctuations compared to the pure classical case, and (ii) obtain the dynamical large deviations by calculating the leading eigenvector of the tilted Lindbladian, and find clear evidence for a first-order active-inactive dynamical phase transition. We also show how to directly sample the rare quantum trajectories associated to the large deviations.
- [9] arXiv:2404.13631 (replaced) [pdf, html, other]
-
Title: Fermi-Bose Machine achieves both generalization and adversarial robustnessComments: 32 pages, 6 figures, a physics inspired machine without backpropagation yet with enhanced adversarial robustnessSubjects: Machine Learning (cs.LG); Disordered Systems and Neural Networks (cond-mat.dis-nn); Statistical Mechanics (cond-mat.stat-mech); Neural and Evolutionary Computing (cs.NE); Neurons and Cognition (q-bio.NC)
Distinct from human cognitive processing, deep neural networks trained by backpropagation can be easily fooled by adversarial examples. To design a semantically meaningful representation learning, we discard backpropagation, and instead, propose a local contrastive learning, where the representation for the inputs bearing the same label shrink (akin to boson) in hidden layers, while those of different labels repel (akin to fermion). This layer-wise learning is local in nature, being biological plausible. A statistical mechanics analysis shows that the target fermion-pair-distance is a key parameter. Moreover, the application of this local contrastive learning to MNIST benchmark dataset demonstrates that the adversarial vulnerability of standard perceptron can be greatly mitigated by tuning the target distance, i.e., controlling the geometric separation of prototype manifolds.