Fluctuations of the free energy in the high temperature Hopfield model (Q2485781): Difference between revisions

The authors consider a doubly indexed family \(\xi_i^k\) (\(i,k\geq 1\)) of independent Bernoulli variables defined on some probability space. Let \(\xi ^k=(\xi_1^k,\dots,\xi_N^k)\) be the \(k\)th pattern, and for \(\sigma \in \{-1,1\}^N\), \(\sigma\cdot \xi^k =\sum_{i=1}^N \sigma_i \xi_i ^k.\) The Hamiltonian of the Hopfield model is defined as \[ H_N (t,\sigma)=\frac{N}{2}\sum_{k=1}^{[tN]}(\sigma\cdot\xi ^k/N)^2 \] with \(H_N (t,\sigma)=0\) when \(0\leq t< 1/N\). The partition function at the inverse temperature \(\beta >0\) is given by \(Z_N ^{\beta} (t)= E_{\sigma} \exp| \beta H_N (t,\sigma)|.\) The main result of this paper is the following: Let \(\beta_0 \in[0,1)\) and \(t_0>0\) be such that \(\beta(1+\sqrt{t})<1\) holds. Then the normalized partition function converges in distribution as \(N \rightarrow \infty\) on the Skorokhod space \(D([0,t_0[,R^+)\) to the process \(\exp\{M_{\infty} ^{\beta_0} (t)-\frac{1}{2}\Gamma (\beta_0 ,t)\}\), where \(M_{\infty} ^{\beta_0 (t)}\) is an independent increments Gaussian process on \([0,t_0 [\) with continuous paths, mean zero and variance \(\Gamma(\beta_0 ,t).\) The function \(\Gamma(\beta ,t)\) is defined as \[ \Gamma(\beta,t)=-\tfrac{1}{2}\ln (1-t(\beta/(1-\beta))^2 )- \tfrac{t}{2} (\beta/(1-\beta))^2 . \]