What you are looking at
A random field is laid over a square grid. Crossing each point costs the exponential of the field there, so some regions are cheap and others expensive. From the source, Dijkstra’s algorithm finds the cheapest distance to every one of the grid’s points, along the eight neighbouring directions. The height of the surface is that distance. The gold lines are the cheapest routes themselves. As disorder rises, they stop running straight and crowd into a few cheap channels, like streams joining a river.
h discrete Gaussian free field on the torus, heat-smoothed by τ cᵢ = exp(ξ hᵢ) / mean(exp(ξ h)) wᵢⱼ = ℓᵢⱼ (cᵢ + cⱼ) / 2 eight neighbours, ℓ = 1 or √2 lattice steps D(v) = min over paths p of Σ wₑ
The mathematics behind it
This is a finite companion to Metric-measure limits of subcritical FK and spanning-tree planar maps (24 September 2026), one of the manuscripts released in OpenAI’s mathematics repository and produced by an internal OpenAI model. The paper shows that, after rescaling, distances on large random planar maps converge to Liouville quantum gravity spheres. Random geometry of this kind is studied by distorting distance with the exponential of a Gaussian free field, which is what this page does on a small grid.
What it does not show
This is eight-neighbour first-passage percolation on a regularized field: a Liouville first-passage percolation analogue, not the paper’s object. The disorder ξ is controlled independently and is not identified with the paper’s parameters, and no continuum limit is computed or claimed. The model follows the GLOBAÏA Atlas of Mathematical Forms (plate “Geodesic estuary”) and runs entirely in your browser. Its random-number generator differs from the atlas’s Python, so a seed gives a statistically equivalent landscape, not the identical one.