Gaussian Free Field (GFF) Overview

• The Gaussian Free Field is a canonical random field that generalizes Brownian motion to higher dimensions and exhibits conformal invariance and the domain Markov property.
• Its covariance structure, defined via the Green’s function for the Laplacian, uniquely determines the spatial correlations and scaling limit properties.
• Discrete approximations and scaling limits of the GFF connect probability theory, statistical mechanics, and quantum field theory through diverse applications in random surfaces.

The Gaussian Free Field (GFF) is a canonical object in probability theory, statistical physics, and mathematical field theory, generalizing Brownian motion to higher-dimensional spaces and providing a rigorous framework for random surfaces and fields with Gaussian fluctuations. In two dimensions, the GFF stands as the archetype of a conformally invariant random distribution and is the universal scaling limit for a wide class of discrete random surface models. Its axiomatic characterization—conformal invariance, the domain Markov property, and mild moment assumptions—rigidly determines its law up to scalar multiples. The GFF has deep connections with SLE, CLE, Liouville quantum field theory, percolative phase transitions, stochastic homogenization, and integrable models.

1. Mathematical Definition and Covariance Structure

Let $D \subset \mathbb{C}$ be a simply connected domain. The mean-zero Dirichlet Gaussian Free Field $h^D_{\mathrm{GFF}}$ is a centered Gaussian random distribution indexed by test functions $\phi \in C^\infty_c(D)$ such that

$$E\left[ (h^D_{\mathrm{GFF}},\phi_1) (h^D_{\mathrm{GFF}},\phi_2) \right] = \iint_{D^2} \phi_1(x)\, \phi_2(y)\, G^D(x,y)\, dx\, dy,$$

where $G^D(x,y)$ is the Green’s function for the Laplacian in $D$ with Dirichlet boundary conditions. Equivalently, $$(h^D_{\mathrm{GFF}}, \phi) \sim N\big(0, \iint_{D^2} G^D(x,y)\, \phi(x)\phi(y)\, dx\,dy \big)$$ (Berestycki et al., 2018).

The covariance structure defines the GFF uniquely among centered Gaussian fields on distributions. The Green’s function $G^D(x,y)$ encodes the spatial correlations and inherits the conformal invariance of the underlying Laplace operator in two dimensions.

2. Domain Markov Property and Conformal Invariance

A defining property of the GFF is its domain Markov property: for any subdomain $D' \Subset D$,

$$h^D_{\mathrm{GFF}} = h^{D'}_{\mathrm{GFF}} + \varphi,$$

where

• $h^D_{\mathrm{GFF}}$0 is a zero-boundary GFF in $h^D_{\mathrm{GFF}}$1, extended by zero outside $h^D_{\mathrm{GFF}}$2,
• $h^D_{\mathrm{GFF}}$3 is almost surely harmonic in $h^D_{\mathrm{GFF}}$4, agrees with $h^D_{\mathrm{GFF}}$5 on $h^D_{\mathrm{GFF}}$6, and vanishes outside $h^D_{\mathrm{GFF}}$7,
• $h^D_{\mathrm{GFF}}$8 and $h^D_{\mathrm{GFF}}$9 are independent (Berestycki et al., 2018).

The GFF is also conformally invariant: if $\phi \in C^\infty_c(D)$0 is conformal, the law of $\phi \in C^\infty_c(D)$1 pushes forward to that of $\phi \in C^\infty_c(D)$2 via

$\phi \in C^\infty_c(D)$3

This property pins down the GFF as the two-dimensional Gaussian field invariant under conformal coordinate changes, provided the Dirichlet boundary and moment assumptions are satisfied (Berestycki et al., 2018).

3. Classification and Rigidity Theorem

The central classification result states: any family $\phi \in C^\infty_c(D)$4 of random distributions in two dimensions, each mean-zero, jointly continuous in the sense above, satisfying the domain Markov property, conformal invariance, and a mild fourth-moment bound, must be a scalar multiple of the Dirichlet GFF: $\phi \in C^\infty_c(D)$5 for some $\phi \in C^\infty_c(D)$6 (Berestycki et al., 2018). This places the 2D GFF alongside SLE among random structures characterized entirely by symmetry and locality properties. The proof relies on the analysis of circle averages and the Dubins–Schwarz theorem to identify the limiting Gaussianity of the corresponding martingales.

4. Gaussian Free Field in Discrete and Scaling Limits

Discrete approximations of the GFF arise naturally on finite graphs. The discrete GFF on a finite subset $\phi \in C^\infty_c(D)$7 of $\phi \in C^\infty_c(D)$8, with Dirichlet boundary, is the centered Gaussian field $\phi \in C^\infty_c(D)$9 with covariance matrix given by the inverse of the (combinatorial) Laplacian. The maximum of the discrete 2D GFF, properly centered, is tight and obeys universal logarithmic corrections: $$E\left[ (h^D_{\mathrm{GFF}},\phi_1) (h^D_{\mathrm{GFF}},\phi_2) \right] = \iint_{D^2} \phi_1(x)\, \phi_2(y)\, G^D(x,y)\, dx\, dy,$$0 for a box of size $$E\left[ (h^D_{\mathrm{GFF}},\phi_1) (h^D_{\mathrm{GFF}},\phi_2) \right] = \iint_{D^2} \phi_1(x)\, \phi_2(y)\, G^D(x,y)\, dx\, dy,$$1 (Bramson et al., 2010). These asymptotics reflect the log-correlated structure that remains robust under rescaling, and the centered maxima are universally tight across other log-correlated models.

In scaling limits of a large class of random surface and height models, including dimer models, random tilings, interlacing particle systems, and certain growth models, properly centered fluctuations of the height field converge to the GFF on appropriate domains, as shown in determinantal and interacting models (Kuan, 2011, Duits, 2011, Granet et al., 2018). In particular, universality persists even in inhomogeneous or non-determinantal settings: inside the arctic curve of the interacting six-vertex model, the fluctuation field is a GFF governed by a generalized Laplacian $$E\left[ (h^D_{\mathrm{GFF}},\phi_1) (h^D_{\mathrm{GFF}},\phi_2) \right] = \iint_{D^2} \phi_1(x)\, \phi_2(y)\, G^D(x,y)\, dx\, dy,$$2, where the coupling constant $$E\left[ (h^D_{\mathrm{GFF}},\phi_1) (h^D_{\mathrm{GFF}},\phi_2) \right] = \iint_{D^2} \phi_1(x)\, \phi_2(y)\, G^D(x,y)\, dx\, dy,$$3 is position-dependent (Granet et al., 2018).

5. Structural Properties: Excursions, Local Sets, and Loop Ensembles

The GFF exhibits a rich geometric and probabilistic structure, including decomposition into local sets and connections to random loop ensembles.

• The excursion decomposition in 2D expresses the GFF as an infinite sum of signed Minkowski content measures supported on the disjoint clusters of a critical Brownian loop soup, with independent Rademacher signs. This two-dimensional analogue of the excursion decomposition of Brownian motion illuminates the Markov property and the spatial organization of fluctuations (Aru et al., 2023):

$$E\left[ (h^D_{\mathrm{GFF}},\phi_1) (h^D_{\mathrm{GFF}},\phi_2) \right] = \iint_{D^2} \phi_1(x)\, \phi_2(y)\, G^D(x,y)\, dx\, dy,$$4

where $$E\left[ (h^D_{\mathrm{GFF}},\phi_1) (h^D_{\mathrm{GFF}},\phi_2) \right] = \iint_{D^2} \phi_1(x)\, \phi_2(y)\, G^D(x,y)\, dx\, dy,$$5 are Minkowski content measures and $$E\left[ (h^D_{\mathrm{GFF}},\phi_1) (h^D_{\mathrm{GFF}},\phi_2) \right] = \iint_{D^2} \phi_1(x)\, \phi_2(y)\, G^D(x,y)\, dx\, dy,$$6 are independent signs.

• The first passage set (FPS) at level $$E\left[ (h^D_{\mathrm{GFF}},\phi_1) (h^D_{\mathrm{GFF}},\phi_2) \right] = \iint_{D^2} \phi_1(x)\, \phi_2(y)\, G^D(x,y)\, dx\, dy,$$7, defined as the maximal region reachable from the boundary without the field dropping below $$E\left[ (h^D_{\mathrm{GFF}},\phi_1) (h^D_{\mathrm{GFF}},\phi_2) \right] = \iint_{D^2} \phi_1(x)\, \phi_2(y)\, G^D(x,y)\, dx\, dy,$$8, is a local set with boundary Minkowski content measure proportional to $$E\left[ (h^D_{\mathrm{GFF}},\phi_1) (h^D_{\mathrm{GFF}},\phi_2) \right] = \iint_{D^2} \phi_1(x)\, \phi_2(y)\, G^D(x,y)\, dx\, dy,$$9 in the scaling limit, analogous to first-hitting times for Brownian motion (Aru et al., 2017).
• The Markov property underpins the local set structure; the field conditioned on the complement of a local set decomposes into an independent GFF in each component plus a harmonic function determined by the boundary (Berestycki et al., 2018, Werner et al., 2020).
• Level lines and interfaces of the GFF couple naturally to Schramm–Loewner Evolution (SLE$G^D(x,y)$0), and the collection of these level lines forms a conformal loop ensemble (CLE$G^D(x,y)$1), encoding the fractal geometry of the field’s level sets and interfaces (Wang et al., 2014, Werner et al., 2020).

6. Generalizations, Martingale Characterization, and Connections

Extensions of the GFF framework cover generalized covariance structures and fractional Laplacians. A generalized GFF driven by $G^D(x,y)$2 may lack the Markov property unless the fluctuation tensor $G^D(x,y)$3 is proportional to the conductivity tensor $G^D(x,y)$4; in stochastic homogenization, limiting fluctuation fields may be non-classical GFFs without spatial Markovianity (Gu et al., 2016).

A martingale-type characterization of the GFF, valid in any dimension, asserts that any random distribution possessing a local resampling (martingale) property, fourth-moment bounds, and appropriate scaling must be (up to scale) the GFF. This characterization links the GFF to the stationary solution of the stochastic heat equation. Analogous statements hold for fractional Laplacians and their fields (Aru et al., 2024).

7. Physical and Mathematical Significance

The GFF is the foundational object in Liouville quantum gravity, where it provides the basic random geometry underlying two-dimensional quantum field theories (Berestycki et al., 2018). Its centrality is amplified by its universality as the scaling limit in diverse lattice models (random tilings, dimers, interacting vertex models), its rigidity under symmetry and locality properties, and its deep connections to stochastic interface dynamics, scaling limits, and conformal structures.

The study of GFF extremes, thick points, and field extrema contributes to the theory of Gaussian multiplicative chaos, interfaces of random surfaces, and critical phenomena in statistical mechanics (Bramson et al., 2010, Arguin et al., 2015). In percolation, sharp phase transitions for the level-set percolation of the discrete GFF have been characterized, revealing exponential decay below and long-range connectivity above a critical threshold (Rigas, 2023).

The GFF thus forms a bridge between probability, statistical physics, complex analysis, and field theory, with an axiomatic characterization that cements its role as the universal conformally invariant random distribution in two dimensions.