Gaussian Free Field Fluctuations

• Gaussian Free Field fluctuations are defined by their covariance structure given by the Green's function with Dirichlet or mixed boundary conditions, essential for modeling random surfaces and spatial correlations.
• Scaling limits and explicit conformal mappings enable rigorous derivations of GFF behavior in discrete models such as lozenge tilings, interlacing particle systems, and random matrix ensembles.
• Extensions including inhomogeneous and generalized GFFs capture non-universal fluctuations in disordered media and interacting models, with implications for SLE, reconstruction problems, and extreme value statistics.

The Gaussian Free Field (GFF) is a paradigmatic model for random surfaces and spatially correlated Gaussian processes in probability theory, mathematical physics, statistical mechanics, and random matrix theory. GFF fluctuations describe asymptotic spatial correlations and universal scaling limits in a variety of discrete and continuous models, with Dirichlet or mixed boundary conditions. This article surveys rigorous developments in the theory of GFF fluctuations, including explicit scaling limits, covariance structure, universality, and deviations in non-classical settings.

1. Definition and Covariance Structure

The Gaussian Free Field on a domain $D \subset \mathbb{R}^d$ is the centered Gaussian generalized function (or random distribution) $h$ with covariance kernel given by the Green’s function for the Laplacian (or elliptic operator) subject to prescribed boundary conditions,

$$\mathbb{E}[h(f)h(g)] = \int_{D\times D} f(x) G_D(x,y) g(y)\, dx\,dy,$$

where $f,g$ are test functions and $G_D(x,y)$ solves $-\Delta_x G_D(x,y) = \delta_y(x)$ on $D$, with $G_D(x,\cdot) = 0$ for $x$ on the boundary. In two dimensions with Dirichlet boundary conditions on a simply connected domain, the covariance reduces to

$$G_D(z,w) = -\frac{1}{2\pi}\log\left|\frac{z-w}{z-\overline{w}}\right|$$

(after conformal mapping to the upper half-plane, for instance). The GFF can be equivalently characterized as the canonical centered Gaussian process with the Dirichlet inner product

$h$0

in its Cameron-Martin space (for $h$1, $h$2 is a distribution rather than a function due to logarithmic roughness).

2. GFF Fluctuations in Discrete and Continuum Models

2.1 Random Surface and Dimer/Tiling Models

In a broad class of random interface and dimer models, macroscopic height fluctuations around the deterministic limit shape converge to the GFF in the critical ("liquid") region. For example, large uniformly random lozenge tilings of lattice polygons (Petrov, 2012) and stepped random surfaces arising from interlacing particle systems (Kuan, 2011, Duits, 2011) exhibit rescaled and centered height functions $h$3 that converge in law, as random distributions, to the pullback of the GFF on a planar domain. The limiting covariance kernel is the Dirichlet Green's function, typically after an explicit conformal or algebraic mapping from physical coordinates to the upper half-plane. The determinantal structure of the correlation kernel and double-contour integral representations are crucial for the analysis. A canonical expression is

$h$4

where $h$5 is the domain mapping (Petrov, 2012).

2.2 Random Matrix Ensembles and GFF Limits

Certain eigenvalue statistics in random matrix theory, notably in multi-level Jacobi or Wishart ensembles, exhibit height fluctuations converging to the GFF. For the multilevel $h$6-Jacobi corners process, global fluctuations of the empirical counting function, after centering and scaling, converge to the GFF under a nonlinear pullback from the configuration to the upper half-plane (Borodin et al., 2013). In overlapping Wishart matrix models with stochastically evolving entries, centered traces of powers of Wishart matrices (linear eigenvalue statistics) admit a multidimensional CLT whose covariance kernel is explicit in terms of contour integrals,

$h$7

where each $h$8 is a double contour integral involving overlap parameters, the time-correlation functions $h$9, and universal logarithmic factors. These encode a bona fide three-dimensional GFF structure indexed by matrix aspect ratios and time (Kuan et al., 2021).

2.3 Interacting Six-Vertex and Inhomogeneous GFF

Beyond free-fermion (determinantal) models, in interacting vertex models (e.g., the six-vertex model with domain-wall boundary conditions), the height fluctuations in the critical region (inside the "arctic curve") are governed by a GFF whose coupling constant $$\mathbb{E}[h(f)h(g)] = \int_{D\times D} f(x) G_D(x,y) g(y)\, dx\,dy,$$0 becomes spatially dependent. The effective action takes the inhomogeneous form

$$\mathbb{E}[h(f)h(g)] = \int_{D\times D} f(x) G_D(x,y) g(y)\, dx\,dy,$$1

and the covariance is the Green’s function for the generalized Laplacian $$\mathbb{E}[h(f)h(g)] = \int_{D\times D} f(x) G_D(x,y) g(y)\, dx\,dy,$$2 with Dirichlet boundary (Granet et al., 2018). This "inhomogeneous GFF" describes non-universal, position-dependent variance, with singularities emerging at special coupling values (e.g., $$\mathbb{E}[h(f)h(g)] = \int_{D\times D} f(x) G_D(x,y) g(y)\, dx\,dy,$$3).

3. Non-Standard, Generalized, and Disordered GFF Fluctuations

3.1 GFFs in Disordered and Inhomogeneous Media

In media with random or correlated inclusions, such as metallic nanoparticles governed by an Ising model or percolation-diluted environments, the GFF framework persists at large scales, but amplitudes of logarithmic fluctuations and geometric exponents are modified. For example, in GFFs on backgrounds with percolation-tuned "metallic islands," the logarithmic amplitude of the variance and the fractal dimension of iso-potential level sets are reduced, scaling with the correlation length of the disorder (Cheraghalizadeh et al., 2018, Cheraghalizadeh et al., 2018). Two fixed points arise: a UV (critical percolation) regime with non-Gaussian exponents, and an IR regime with standard GFF statistics, the latter acting as the stable fixed point under RG flow.

3.2 Generalized GFF and Breakdown of the Markov Property

Generalized GFFs are constructed as solutions $$\mathbb{E}[h(f)h(g)] = \int_{D\times D} f(x) G_D(x,y) g(y)\, dx\,dy,$$4 to $$\mathbb{E}[h(f)h(g)] = \int_{D\times D} f(x) G_D(x,y) g(y)\, dx\,dy,$$5 with $$\mathbb{E}[h(f)h(g)] = \int_{D\times D} f(x) G_D(x,y) g(y)\, dx\,dy,$$6 and $$\mathbb{E}[h(f)h(g)] = \int_{D\times D} f(x) G_D(x,y) g(y)\, dx\,dy,$$7 symmetric positive-definite matrices and $$\mathbb{E}[h(f)h(g)] = \int_{D\times D} f(x) G_D(x,y) g(y)\, dx\,dy,$$8 a $$\mathbb{E}[h(f)h(g)] = \int_{D\times D} f(x) G_D(x,y) g(y)\, dx\,dy,$$9-white-noise vector field. Spatial Markovity is lost unless $f,g$0, in which case only the classical GFF is recovered. In stochastic homogenization problems, limits of fluctuations (the "corrector") are often generalized GFFs with non-proportional $f,g$1 and $f,g$2, resulting in nonlocal covariance kernels and non-Markovian structure (Gu et al., 2016).

4. Extreme Value Fluctuations and Fluctuations on Fractals

4.1 Maxima of Discrete GFF and Fractal Graphs

The behavior of the recentered maximum of GFFs depends crucially on the recurrence/transience of the underlying graph. For two-dimensional lattices, the recentered maximum is tight (order $f,g$3 fluctuations), but on one-dimensional segments and a variety of recurrent fractal graphs (e.g., Sierpinski carpets), the recentered maximum is not tight and exhibits fluctuations of the same order as the maximal pointwise standard deviation, scaling with effective resistance to the boundary (Kumagai et al., 2013). This is encoded in the resistance entropy, growth exponents, and covering properties of the graph.

4.2 Integer-Valued GFFs, Vortex Fluctuations, and BKT Transition

For integer-valued GFFs (IV-GFF) and related models (Villain, Coulomb gas), vortex-induced fluctuations lead to substantial deviations from the standard GFF maxima and two-point function. Explicit quantitative bounds on the maximum,

$f,g$4

demonstrate a reduction of the fluctuation amplitude compared to the standard case due to topological defects (Garban et al., 2020). The effective GFF stiffness is renormalized by vortex statistics, consistent with RG predictions for the Kosterlitz-Thouless transition (BKT).

5. GFF Fluctuations in Dynamical and Higher-Dimensional Systems

5.1 Three-Dimensional and Space-Time GFF Limits

Multi-parameter models such as growing random surfaces, random walks on quantum groups, or overlapping Wishart matrices with time evolution give rise to Gaussian field limits indexed by both "space" and "time" parameters. The covariance structure involves contour-integral representations and time-dependent weightings, generalizing the static GFF covariance to a bona fide 3D GFF structure (Kuan et al., 2021, Zhou, 2022, Kuan, 2016). For example, in the fluctuating height statistics of noncommutative random walks on $f,g$5, one recovers space-like GFF correlations along suitable paths, with explicit distinctions from overlapping Wishart models in terms of the multi-time covariance structure.

5.2 Fluctuations of SPDEs and Internal Aggregation Models

For stochastic PDEs such as the (mollified) KPZ equation in $f,g$6, small disorder fluctuations about the stationary solution converge, after rescaling, to a GFF evolved by the linear heat equation, with initial field given by spatial GFF data (Comets et al., 2019). Similarly, in higher-dimensional internal DLA (diffusion-limited aggregation), appropriately normalized fluctuations of the difference from mean shape converge to GFFs or augmented GFFs, reflecting the integration over harmonic polynomials and the geometry of the aggregation front (Jerison et al., 2011).

6. Reconstruction Problems, SLE Level Lines, and Conformal Geometry

6.1 Information Reconstruction and the KT Transition

In statistical reconstruction problems, the ability to recover macroscopic GFF observables from wrapped or modulo fields ($f,g$7) exhibits a Kosterlitz-Thouless (KT) type phase transition: for small $f,g$8, full recovery is possible, while for $f,g$9, delocalization (as in integer-valued GFFs) prevents macroscopic information extraction. The critical regime is governed by a percolation argument and relates to the structure of integer-valued GFFs in inhomogeneous environments (Garban et al., 2020).

6.2 SLE and Level-Line Structure

The zero contour lines (level lines) of the two-dimensional GFF with Dirichlet boundary conditions, under scaling, converge to Schramm-Loewner Evolution (SLE$G_D(x,y)$0) curves, revealing the conformal invariance and deep connections to conformal field theory (Miller, 2010, Ivanov et al., 2015). Modifications of the GFF coupling (background charge, drift terms) correspond to SLE variants traced by stochastic holomorphic flows. Explicit martingale observables are constructed as ratios of conformal field theory correlation functions involving inserted primary fields (vertex operators) corresponding to SLE drift and background charge structure.

7. Summary Table: Key Settings and Limiting GFF Covariances

Setting / Model	Fluctuating Quantity	GFF/Covariance Structure
Random lozenge tilings/polygons (Petrov, 2012)	Centered height in "liquid"	Pullback of GFF on $G_D(x,y)$1: $G_D(x,y)$2
Multilevel Jacobi/Wishart (Borodin et al., 2013, Kuan et al., 2021)	Matrix eigenvalue trace stats	Covariance via double contour integrals, explicit scaling kernels
Interlacing particles/surfaces (Kuan, 2011, Duits, 2011)	Rescaled height fluctuations	GFF via pushforward under critical-point mapping
Integer-valued GFF/Villain (Garban et al., 2020)	Maximum, vortex fluctuations	Reduced prefactor, effective stiffness via vortex statistics
Inhomogeneous/Disordered GFF (Cheraghalizadeh et al., 2018, Cheraghalizadeh et al., 2018)	Electrostatic potential	Covariance with amplitude reduced by disorder, geometric exponents modified
Interacting 6-vertex (Granet et al., 2018)	Height inside critical region	GFF for generalized Laplacian $G_D(x,y)$3

The theory of GFF fluctuations thus underpins the universal scaling behavior spanning random combinatorics, matrix models, interface growth, and critical statistical mechanics, with explicit fluctuational forms and quantitative universality results realized in both the "standard" and generalized regimes.