The Gaussian Free Field on a graph is the canonical random surface: a Gaussian random variable at each vertex, correlated according to the graph's Green function. Constructing it typically involves inverting the graph Laplacian or sampling from the appropriate multivariate Gaussian. These methods work but they're tied to the specific graph and its spectral properties.
Hedenmalm, Mozolyako, and Panov construct the GFF differently — through Hadamard's variational formula. They expand the Green function dynamically along an expanding family of subgraphs, building the field one vertex at a time as the subgraph grows. At each step, the variation of the Green function under the addition of a new vertex determines the field's value there.
The construction is completely dimension-free. It makes no reference to the ambient dimension of the graph, requires no smoothness, and uses only the discrete topology (edges) and metric information (conductances). The same formula works on a path graph, a lattice in any dimension, or an arbitrary weighted graph.
The dimension-free property means the construction can pass to continuum limits on fractal geometries — domains with arbitrary Hausdorff dimension where standard PDE-based constructions of the GFF fail because the notions of “Laplacian” and “boundary” become problematic. The Hadamard formula doesn't need these concepts; it only needs the expanding family of subgraphs and the conductances.
The Gaussian Free Field, rebuilt from the ground up. Not through the Laplacian. Through the growth of the graph itself.