Rainbow percolation

Scatter points at random along a line and give each one a reach, a distance drawn from a heavy-tailed law. Most reaches are short and a few are very long. Join two points by an edge when each of them lies inside the reach of the other.

That is the model. The question is whether the resulting graph has an infinite connected component. The panels below simulate the model, its behaviour under rescaling, and the two structures the lower bound in the paper uses.

1 The model and its connection rule

Points of a Poisson process on the line, each with a reach R that is Pareto distributed with scale β. The reaches are drawn below the axis, the longest at the bottom, and the edges as arcs above it, coloured by connected component. Each point keeps its mark as β changes, so raising β only adds edges.

2.00

Hover over a point, or tap it, to see the connection rule. Its reach is highlighted and every point inside it is marked: green where that point reaches back, so that the edge is present, and red where it does not. An edge of length d requires two reaches of at least d, and each reach is shared by all the edges at its point, so edges meeting at a vertex are not independent. The lower plot measures the same sample on a wider window of 104 points, since a window of a few hundred points is spanned by a single point of large reach at values of β below the critical one.

The precise statement

The vertex set is a Poisson process of intensity λ on ℝ × (0,1). A vertex (x,t) is joined to (y,s) exactly when (t ∨ s)·|x − y| ≤ β, that is, when |x − y| ≤ min(Rx, Ry) for the reach R = β/t. Since t is uniform on (0,1), the reach is Pareto with scale β and tail exponent exactly one: P(R > r) = β/r. Averaging over the marks, two points at distance d > β are joined with probability β2/d2, the critical decay of one-dimensional long-range percolation.

The paper proves that the model has a non-degenerate phase transition: all components are finite once λβ < 1, and an infinite one exists once λβ ≥ 31. At intensity one this reads 1 ≤ βc ≤ 31, and the simulations reported in the paper place βc near 2.

The plotted quantity is the paper's θL(β), the fraction of the points of the central half of the window that lie in the largest component. This is one sample at one window size; the numerical study in the paper measures how it moves with L. The movement is already visible at small sizes. Taking one sample at L = 103, 104, 105 gives θ(1) = 0.61, 0.41, 0.28 and θ(0.5) = 0.16, 0.06, 0.01, falling towards zero, while θ(2) = 0.94, 0.98, 0.97 and θ(3) = 1.00, 0.995, 0.998 are stable. Curves that decay with L below the threshold and are stable above it locate βc. The paper runs this to L = 109, where every finite-size effect decays like 1/log L rather than like a power of L, because the reach tail exponent is exactly one.

2 The same picture at every scale

The same model, viewed through a window whose width you can set. Only the points whose reach is at least a fixed fraction of the window are drawn, since the others are too small to resolve at that width. Widen the window by a factor of a million and compare the picture and the counter with what they were.

1
2.00

The number of visible points, the length of the visible reaches and the density of the arcs stay the same as the window width runs through the decades. This is a property of the model and not of the drawing: rescaled to the unit window, the visible picture has the same law at every width. Drag inside the picture to pan, and scroll or pinch to zoom.

The rescaling

Fix a resolution ε and draw only the points whose reach exceeds εW, where W is the window width. Such a point has t ≤ β/(εW), so the visible points form a Poisson process of intensity λβ/(εW) per unit length. Rescale the picture to the unit window, u = x/W. The visible intensity becomes λβ/ε per unit u, and the rescaled reach ρ = R/W satisfies P(ρ > s) = ε/s for s ≥ ε. Neither depends on W. Since the connection rule |u − u'| ≤ min(ρ, ρ') involves no other scale, the rescaled picture has the same law at every window width, and the two parameters enter only through the product λβ.

Varying β or λ therefore moves only the constant and never the exponent, so the model cannot be compared to an off-critical regime, and no stochastic domination argument settles percolation. The paper also treats the other exponents: under the rule (t ∨ s)γ|x − y| ≤ β the reaches have tail exponent 1/γ, and the threshold is degenerate for every γ ≠ 1, infinite for γ < 1 and zero for γ > 1.

3 Rainbows and cut points

The lower bound in the paper is proved on a discrete skeleton of the model, with the points pinned to the integers. In the skeleton every component is finite for every admissible parameter, although almost surely infinitely many edges cross every site. The two tabs show the objects the proof uses.

0.60

A rainbow is a pair of edges {a,b} and {c,d} with a < c < d < b whose two diagonals {a,c} and {d,b} are absent and whose overhangs are balanced. Drag the four points along the axis, and drag the handles on the two inner reaches to change Rc and Rd. The checklist records which conditions of the definition hold. Extending an inner reach past the far outer endpoint produces a long diagonal, which the balance condition rules out.

The definition

Write ℓ = c − a, m = d − c, r = b − d. The outer arch forces Ra, Rb ≥ b − a, so the two absent diagonals constrain only the inner reaches: Rc ∈ [m, ℓ) and Rd ∈ [m, r), which already forces ℓ, r > m. Balance, |ℓ − r| ≤ m, then caps the inner reaches further, by Rc < m + r and Rd < m + ℓ, so the long diagonals are absent too.

A rainbow does not by itself separate the inner arch from the outer one, since a path through an overhang could still join them. The paper adds two things: the containment event that no interior vertex reaches an outer endpoint, and a cut point in each overhang. Containment has probability k(1 − xm/Dk), where Dk = min(ℓ+k, r+m−k) is the distance from the interior vertex c+k to the nearer outer endpoint, and it is bounded below by (4/9)2xm uniformly in , r and m. The bound does not depend on the scale, so the argument can be run at infinitely many independent scales.