Bhattacharya, Deb and Mukherjee write the free-energy limit of multilinear Gibbs measures as an infinite-dimensional optimization, with sufficient conditions and counterexamples for replica symmetry

The limiting free energy admits a variational form: lim Z_n(θ)=sup_f{θG_W(f)−∫γ(β(f(x)))dx}, and the empirical measure L_n(X) converges in probability to the optimizer set Ξ(F_θ). Most prior mean-field results require compactly supported or log-concave base measures; this work allows a general base measure under moment condition (1.5) and cut-norm convergence, and links optimizers to the limit set of the empirical measure. Proposition 1.1 states the variational representation and weak compactness of Ξ(F_θ), building on the authors' earlier large-deviation results for inhomogeneous U-statistics.

Sufficient conditions for replica symmetry: if T[Sym[W]] is a.s. constant and θW is a.s. strictly positive, then all optimizers of (1.8) are constant functions provided either v is even or μ is stochastically non-negative; conversely, if T[Sym[W]] is not constant, no non-zero constant function is an optimizer. Extends the characterization of replica symmetry from quadratic Hamiltonians to multilinear forms, and uses three counterexamples to show that stochastic non-negativity of μ, θ>0 and θW>0 are necessary in the corresponding senses. Theorem 1.2 gives the sufficient conditions, with proofs via the first-order condition (1.13) and the equality case of Hölder's inequality; Examples 1.1–1.3 construct explicit graphs and measures as counterexamples.

The empirical measure of local fields m converges weakly to a set B_θ, the empirical measure of conditional means α′(θm_i) converges to B_θ^*, and this yields a universal weak law for contrasts: if Σc_i=o(n) and Σ|c_i|^r=O(n), then n^{-1}Σc_iX_i converges to 0 in probability. The weak law does not depend on the specific matrix sequence {Q_n} as long as the symmetrized tensor is regular; earlier universality results of this type were mainly for quadratic Ising models or log-concave base measures. Theorems 1.4 and 1.7 and Corollary 1.6 give the weak limits and the contrast result, with proofs relying on the stability lemma 2.5 and exponential tail bounds.

Local and global magnetizations satisfy exponential concentration: R_{n,θ}(n^{-1}Σ|X_i|^p≥K), R_{n,θ}(n^{-1}Σ|m_i|^q≥K) and R_{n,θ}(n^{-1}Σ|α′(θm_i)|^p≥K) are each bounded by 3exp(−nK/K_0); for compactly supported μ and higher-order interactions, there is a temperature threshold θ_c such that for θ<θ_c the unique optimizer is t=0, while for θ>θ_c t=0 is no longer a global optimizer. Extends the existence of a phase transition point from quadratic Hamiltonians to multilinear forms, and additionally provides exponential tail bounds of independent interest. Theorem 1.5 gives the tail bounds and uniform moment bounds; Theorem 1.10 establishes the existence of θ_c when μ is compactly supported on [−1,1] and α′(0)=0.

The results target multilinear Gibbs measures whose coupling matrices converge in cut norm and which satisfy moment conditions (1.5) and (1.7), covering tensor Ising and p-spin Curie-Weiss models in the dense graph limit; for compactly supported base measures and higher-order interactions, Theorem 1.10 establishes the existence of a temperature threshold θ_c. The authors note that future work could extend the techniques to more general Hamiltonians such as Potts models, to convergence topologies beyond cut norm (e.g. local weak convergence on bounded-degree graphs), and to more general tensor Hamiltonians not specified by a matrix Q_n.

The phase-transition result establishes existence of θ_c when μ is compactly supported on [−1,1] and α′(0)=0, but does not give an explicit expression for θ_c; the replica-symmetry sufficient conditions rely on T[Sym[W]] being constant and θW strictly positive, and verifying these in specific models still requires case-by-case work. The authors also note that finer questions such as central limit theorems and limit distributions remain open, and that extension to sparse graphs or general tensor Hamiltonians would require further development of cut-norm theory.