On a nonhierarchical generalization of the perceptron GREM (Q6171642)

The whole Parisi picture and its presumed universality for mean field (spin glasses) models remains mysterious, despite enormous progress over the last decades with a rich and challenging mathematical structure. The Sherrington-Kirkpatrick model [\textit{D. Sherrington} and \textit{S. Kirkpatrick}, Phys. Rev. Lett. 35, No. 26, 1792--1796 (1975; \url{doi:10.1103/PhysRevLett.35.1792})] still lacks of a rigorous treatment via large deviations techniques within the classical Gibbs-Boltzmann formalism, and a puzzle of the Parisi theory is the ultra-metricity, the emergence in the infinite volume limit of hierarchical structures closely related to GREM generalized random energy model) introduced by \textit{B. Derrida} [Phys. Rev. B (3) 24, No. 5, 2613--2626 (1981; Zbl 1323.60134); J. Phys. Lett. 46, No. 9, 401--407 (1985; \url{doi:10.1051/jphyslet:01985004609040100})]. The article under review introduces a random Hamiltonian neither hierarchical from the start nor given by linear superpositions of GREM-like model, and it builds a Sanov-type large deviation principle for empirical distributions of the random energy to derive a Boltzmann-Gibbs principle as well as a Parisi formula for the free energy in the thermodynamical limit. The first main result (Theorem 1) shows that the quenched infinite volume free energy exists and is given by the solution of a Boltzmann-Gibbs variational principle (BGVP for short later). Its proof follows quite closely to the Perceptron GREM in [\textit{E. Bolthausen} and \textit{N. Kistler}, Springer Proc. Math. 11, 425--442 (2012; Zbl 1246.82107)] with the main steps given in Section 3. The second main result (Theorem 2) solves the Parisi minimization (PVP) under the system admitting a unique minimal chain, and the PVP fulfills the consistent agreement with the Parisi theory for more sophisticated mean field spin glasses. Its proof is completed in Section 4. Section 2 states main theorems with some supplementary results. Section 3 presents a proof of the BGVP for Theorem 1. Proposition 3.1 shows that the purely first moment estimates with overwhelming probability (\(P (E_N) \ge 1 -e^{- \delta N}\) for \(N\) large enough), and there exists an open neighborhood such that the overwhelming probability zero over it. Proposition 3.2 shows that the logarithmic limit of the first moment exists over an open neighborhood with entropic conditions. By covering the subset through open sets, the partition function holds inequality estimates, Sanov's theorem consequence estimation together with Borel-Cantelli to get the upper bound of \(\lim_{N\to \infty} \sup \frac{1}{N} \log Z_N (\phi)\) with overwhelming probability. For the lower bound, the function \(\int \phi (x) \rho (dx) - H(\rho)\) of \(\rho\) is upper-semicontinuous and the intersection of the subset is compact, applying BGVP and Proposition 3.2 to guarantee strong concentration of the first moment for some neighborhood. First moment estimates and proof of Proposition 3.1 and second moment estimates and proof of Proposition 3.2 are given in subsections using a technical lemma (variance estimate). Section 4 completes the proof of Theorem 2, the Parisi variational principle (PVP), starting with the \(n=2\) case following the terminology from [\textit{E. Bolthausen} and \textit{N. Kistler}, Springer Proc. Math. 11, 425--442 (2012; Zbl 1246.82107)] in subsection 4.1; the generic case given in subsection 4.2 starts with how the uniqueness assumption for the minimal chain affects the signs of the derivatives of the associated Parisi function; the upper bounds is immediate, and the lower bound is the more focusing part. The authors introduce a family of generalized Gibbs measures for each chain, the measure satisfies all the entropic constraints appearing in BGVP for the chain, without loss of generality, taking the unique minimal chain for all Parisi variational principles, one can claim that \(\tilde{m} =\mathrm{arg} \min_{m\in \Delta_n}P^T (m) \to G^T_{\tilde{m}} \in \cap_{J \in P_J} C_J\), and the measure \(G^T_{\tilde{m}}\) is a viable candidate for the solution of the BGVP. The proof of Theorem 2 is completed further distinguishing two subcases by monotonicity property. In the appendix the authors prove technical lemmas from Section 2 which generalize results of \textit{E. Bolthausen} and \textit{N. Kistler} [Ann. Appl. Probab. 16, No. 1, 1--14 (2006; Zbl 1100.60026)] in Lemma 3.2.