[Paper Review] Decoupling inequalities for the Ginzburg-Landau $ abla \varphi$ models

Pierre‐François Rodriguez|arXiv (Cornell University)|Dec 7, 2016

Markov Chains and Monte Carlo Methods16 references3 citations

TL;DR

This paper establishes decoupling inequalities for massless gradient Gibbs measures in the Ginzburg-Landau $\nabla\varphi$ model in dimensions $d \geq 3$, extending results from the Gaussian case to non-Gaussian, uniformly convex, and non-convex perturbations. By leveraging a Helffer-Sjöstrand random walk representation and a novel sprinkling technique based on boundary condition perturbations in the Poisson equation, the authors prove exponential decay of connectivity in level sets at high levels, generalizing results on percolative and non-percolative phases.

ABSTRACT

We consider a class of of massless gradient Gibbs measures, in dimension greater or equal to three, and prove a decoupling inequality for these fields. As a result, we obtain detailed information about their geometry, and the percolative and non-percolative phases of their level sets, thus generalizing results obtained in arXiv:1202.5172, to the non-Gaussian case. Inequalities of similar flavor have also been successfully used in the study of random interlacements, see arXiv:1010.1490, arXiv:1212.1605. A crucial aspect is the development of a suitable sprinkling technique, which relies on a particular representation of the correlations in terms of a random walk in a dynamic random environment, due to Helffer and Sj\\"{o}strand. The sprinkling can be effectively implemented by studying the Dirichlet problem for the corresponding Poisson equation, and quantifiying in how far a change in boundary condition along a sufficiently "small" part of the boundary affects the solution. Our results allow for uniformly convex potentials, and extend to non-convex perturbations thereof.

Motivation & Objective

• To extend decoupling inequalities—previously known only for Gaussian fields—to non-Gaussian, uniformly convex, and non-convex gradient interface models in $d \geq 3$.
• To establish a rigorous framework for analyzing long-range correlations and percolation transitions in non-Gaussian $\nabla\varphi$ fields.
• To generalize the sprinkling technique used in random interlacements to the anharmonic setting by using a Helffer-Sjöstrand representation of correlations.
• To quantify the decay of connectivity in level sets of $\varphi$-fields at high levels, proving that high-level level sets are typically non-percolating.

Proposed method

• Utilizes the Helffer-Sjöstrand representation to express correlations as expectations over a random walk in a dynamic random environment.
• Applies an interpolation argument via Proposition 2.4 to compare the field under different boundary conditions.
• Implements a sprinkling technique by perturbing boundary conditions on a small part of the boundary and analyzing the resulting change in harmonic solutions.
• Quantifies the error in conditional expectations using the Dirichlet problem for the Poisson equation, bounding the effect of boundary perturbations.
• Employs the Brascamp-Lieb inequality to control error terms arising from non-convex perturbations.
• Reduces the problem to a comparison with the Gaussian free field by analyzing the deviation of the actual field from harmonic extensions.

Experimental results

Research questions

• RQ1 Can decoupling inequalities be established for non-Gaussian $\nabla\varphi$ models in $d \geq 3$, beyond the Gaussian case?
• RQ2 How can the sprinkling technique be adapted in the absence of the Markov property, as in the Gaussian free field?
• RQ3 What is the rate of decay of connectivity in level sets of $\varphi$-fields at high levels under non-convex perturbations?
• RQ4 To what extent can the Gaussian free field serve as a reference for approximating the anharmonic model?
• RQ5 Under what conditions does the level set percolation transition disappear at high levels in the non-Gaussian case?

Key findings

• For uniformly convex potentials, the paper establishes a decoupling inequality of the form $\mathbb{E}_{\mu}[f^{h}g] \leq \mathbb{E}_{\mu}[f^{h-\varepsilon}]·\mathbb{E}_{\mu}[g] + \|g\|_{L^{\infty}} \cdot \delta_{S,S'}(\varepsilon)$, with $\delta_{S,S'}(\varepsilon)$ decaying as $e^{-L^{\alpha}}$ when $\varepsilon > R^{-\beta}$.
• The error term $\delta_{S,S'}(\varepsilon)$ can be made exponentially small in $L$ under suitable conditions on the boundary perturbation, confirming the effectiveness of the sprinkling method.
• For non-convex perturbations satisfying $\sqrt{\beta}\|g''\|_{L^1(\mathbb{R})} < c_{24}$, the level set connectivity decays exponentially: $\mu(\widetilde{E}_{L,x}^{h}) \leq ce^{-c' L^{\varepsilon}}$ for large $h$.
• The high-level connectivity decay implies that $h^+(\mu_{\beta}) < \infty$, meaning that percolation in level sets disappears at sufficiently high levels.
• The method allows for finite-range multi-body interactions under a random walk representation condition, extending beyond nearest-neighbor interactions.
• The results generalize the percolation phase transition picture from the Gaussian case to a broad class of non-Gaussian models, including non-convex perturbations.