Glass--like transition described by toppling of stability hierarchy

Building on the work of Fyodorov (2004) and Fyodorov and Nadal (2012) we examine the critical behaviour of population of saddles with fixed instability index k in high dimensional random energy landscapes. Such landscapes consist of a parabolic confining potential and a random part in N 1 dimensions. When the relative strength m of the parabolic part is decreasing below a critical value mc, the random energy landscapes exhibit a glass-like transition from a simple phase with very few critical points to a complex phase with the energy surface having exponentially many critical points. We obtain the annealed probability distribution of the instability index k by working out the mean size of the population of saddles with index k relative to the mean size of the entire population of critical points and observe toppling of stability hierarchy which accompanies the underlying glasslike transition. In the transition region m = mc + δN −1/2 the typical instability index scales as k = κN and the toppling mechanism affects whole instability index distribution, in particular the most probable value of κ changes from κ = 0 in the simple phase (δ > 0) to a non-zero value κmax ∝ (−δ) in the complex phase (δ < 0). We also show that a similar phenomenon is observed in random landscapes with an additional fixed energy constraint and in the p-spin spherical model.

[1]  Mathematical Physics © Springer-Verlag 1996 Fredholm Determinants and the mKdV/ Sinh-Gordon Hierarchies , 1995 .

[2]  Guigang Zhang,et al.  Deep Learning , 2016, Int. J. Semantic Comput..

[3]  M. Mézard,et al.  Off-Equilibrium Glassy Dynamics: A Simple Case , 1994 .

[4]  A. Bray,et al.  Statistics of critical points of Gaussian fields on large-dimensional spaces. , 2006, Physical review letters.

[5]  Asymptotic form of the density profile for Gaussian and Laguerre random matrix ensembles with orthogonal and symplectic symmetry , 2005, math-ph/0508031.

[6]  Antonio Auffinger,et al.  Random Matrices and Complexity of Spin Glasses , 2010, 1003.1129.

[7]  Y. Fyodorov High-Dimensional Random Fields and Random Matrix Theory , 2013, 1307.2379.

[8]  Yan V Fyodorov Complexity of random energy landscapes, glass transition, and absolute value of the spectral determinant of random matrices. , 2004, Physical review letters.

[9]  A. Crisanti,et al.  The sphericalp-spin interaction spin glass model: the statics , 1992 .

[10]  Eliran Subag,et al.  The complexity of spherical p-spin models - a second moment approach , 2015, 1504.02251.

[11]  L. Laloux,et al.  Phase space geometry and slow dynamics , 1995, cond-mat/9510079.

[12]  Yan V Fyodorov,et al.  Critical behavior of the number of minima of a random landscape at the glass transition point and the Tracy-Widom distribution. , 2012, Physical review letters.

[13]  Jonas Gustavsson Gaussian fluctuations of eigenvalues in the GUE , 2004 .

[14]  Antonio Auffinger,et al.  The number of saddles of the spherical $p$-spin model , 2020, 2007.09269.

[15]  D. Thouless,et al.  Spherical Model of a Spin-Glass , 1976 .

[16]  Yan V Fyodorov,et al.  Replica Symmetry Breaking Condition Exposed by Random Matrix Calculation of Landscape Complexity , 2007, cond-mat/0702601.

[17]  Sean O’Rourke,et al.  Gaussian Fluctuations of Eigenvalues in Wigner Random Matrices , 2009, 0909.2677.

[18]  Yann LeCun,et al.  The Loss Surfaces of Multilayer Networks , 2014, AISTATS.

[19]  Andrew Lucas,et al.  Ising formulations of many NP problems , 2013, Front. Physics.

[20]  Yan V. Fyodorov,et al.  Counting equilibria of large complex systems by instability index , 2020, Proceedings of the National Academy of Sciences.

[21]  P. Forrester Log-Gases and Random Matrices , 2010 .

[22]  Folkmar Bornemann,et al.  On the Numerical Evaluation of Distributions in Random Matrix Theory: A Review , 2009, 0904.1581.

[23]  Antonio Auffinger,et al.  Complexity of random smooth functions on the high-dimensional sphere , 2011, 1110.5872.