Rigorous inequalities between length and time scales in glassy systems (Q863507)

The authors consider spin models on rather general graphs. Each spin variable \(x_i\) can take only a finite number of values. No kind of translation invariance is required on the graph or on the interaction, and only a uniform bound on the graph degree is assumed, plus mild conditions on the interactions. The results apply for instance to diluted, mean field spin glass models on sparse random graphs of constant connectivity. For each site on the graph one defines a correlation length \(\ell_i\) as the minimal distance \(D\) such that the variable \(x_i\) is essentially independent of the set of all variables \(x_j\) such that the (graph) distance between \(i\) and \(j\) is at least \(D\), and a ``time scale'' \(\tau_i\) after which the correlation between \(x_i(t)\) and \(x_i(0)\) is negligible (the dynamics being started from the equilibrium measure). The main result is a general bound of the type \[ \ell_i \leq \tau_i \leq e^{V(i,\ell_i)} \] where \(V(i,r)\) is the volume of the ball of radius \(r\) centered at \(i\) (e.g., \(V(i,r)\simeq r^d\) if the graph is the regular lattice \(Z^d\).) Bounds of the same type (although not for \(\ell_i\) and \(\tau_i\) defined in exactly the same way) were known before in the literature (references in this sense are disseminated along the paper). One of the original points of this work is to define relaxation to equilibrium via quantities like \(\tau_i\) which are easily accessible in (real or numerical) experiments, rather than through (say) the spectral gap or the log-Sobolev constant. In Section 4, a concrete example is considered: the \(p\)-spin model (\(p\geq3\)) on random regular hypergraphs. It is proven in particular that a purely dynamical phase transition occurs at a certain temperature: above this temperature \(\tau_i\) is of order one, while below it is exponentially large in the system size for a positive fraction of sites \(i\). At the same temperature there is no equilibrium phase transition.