When infinite percolation clusters coexist

← All digests

Family 214 · Probability and statistical mechanics
Importance 4/5 · Provisional editorial rating
Sources consulted 2026-10-11

1. The question

Imagine independently keeping each road in an infinite network with probability $p$. At small $p$, every connected region is finite. At larger $p$, infinite connected regions appear. Must the first infinite regions immediately join into a single giant component, or can infinitely many of them coexist?

The answer depends on the geometry of the network. A tree keeps producing fresh branches, whereas a Euclidean lattice has many routes that reconnect. The manuscript in this family claims that every sufficiently symmetric nonamenable graph has an interval of coexistence: infinite clusters exist, but none absorbs all the others.

Nonamenability means that every finite region has a boundary comparable to its volume. There is always substantial room through which different clusters can escape. Turning that picture into a theorem for every graph in this class is the longstanding difficulty. The manuscript claims a stronger operator estimate that also gives a common set of critical power laws.

2. Background and essential definitions

Let $G=(V,E)$ be an infinite connected graph. It is locally finite if every vertex has finitely many neighbors, and quasi-transitive if graph symmetries divide its vertices into finitely many types. A transitive graph has just one type. These assumptions imply a uniform bound on degrees.

For a finite nonempty set $A$ of vertices, let $\partial_V A$ be the vertices outside $A$ adjacent to it. The graph is nonamenable when

$$h_V(G)=\inf_{0<|A|<\infty}\frac{|\partial_V A|}{|A|}>0.$$

Thus no sequence of larger finite regions can make the boundary negligible relative to the interior. Large square boxes in the square lattice do make this ratio tend to zero; regular trees of degree at least three do not.

In Bernoulli bond percolation, every edge is independently open with probability $p$. The open cluster $C_v$ contains all vertices connected to $v$ by open edges. Write $N_\infty$ for the number of infinite clusters. Two thresholds matter:

$$\begin{aligned} p_c&=\inf\{p:\mathbb P_p(N_\infty\geq1)>0\},\\ p_u&=\inf\{p:\mathbb P_p(N_\infty=1)=1\}. \end{aligned}$$

The question is whether $p_c<p_u$. This says that existence and uniqueness occur at genuinely different densities.

A useful model is the regular tree with degree $d\geq3$. Away from its starting vertex, a cluster branches on average into $(d-1)p$ descendants. Its existence threshold is $p_c=1/(d-1)$, while $p_u=1$. A closed edge separates two branches permanently: there is no alternate route around it. General nonamenable graphs may have many cycles, so this tree argument does not settle the question.

The stronger claim uses the two-point function

$$T_p(x,y)=\mathbb P_p(x\leftrightarrow y).$$

View this array as a matrix acting on functions on the vertices. The space $\ell^2(V)$ consists of functions with $\sum_x|f(x)|^2<\infty$, using ordinary counting measure. The operator norm $\|T_p\|_{2\to2}$ measures how much this matrix can amplify that square-sum norm. Define $p_{2\to2}$ as the supremum of parameters at which the operator is bounded.

This is different from the expected cluster size, the row sum $\chi_v(p)=\sum_yT_p(v,y)$. An infinite row sum need not force an infinite square-sum operator norm on a rapidly branching graph.

3. Development history and previous partial results

Benjamini and Schramm’s 1996 paper made percolation on nonamenable graphs a systematic subject and posed the general nonuniqueness conjecture. The contrast with Euclidean lattices suggested that the boundary-to-volume ratio should control more than the mere existence of an infinite cluster. Their formulation helped connect probability with geometric group theory. Benjamini–Schramm

Several important classes were subsequently understood. Planar nonamenable graphs admit powerful duality arguments: open connections in the graph and closed connections in its planar dual constrain one another. Strong expansion and sufficiently large girth, meaning the absence of short cycles, also permit comparisons with branching processes. Other advances exploited harmonic functions or particular symmetry structures. These were substantial partial solutions, but none made nonamenability alone sufficient in the full quasi-transitive setting.

For graphs coming from groups, another distinction mattered. A Cayley graph joins each group element to its products with a chosen finite generating set. Earlier results could arrange a nonuniqueness phase by choosing suitable generators. The full conjecture asks for it on each prescribed nonamenable Cayley graph, so changing the generators does not solve that version.

Hutchcroft developed the operator threshold $p_{2\to2}$ as a way to obtain both nonuniqueness and quantitative critical behavior. His work on hyperbolic graphs established the desired gap for nonamenable quasi-transitive graphs with that extra geometric structure. Hyperbolicity says, roughly, that large geodesic triangles are thin; it is stronger than nonamenability. Percolation on hyperbolic graphs

Further work explained the consequences of $p_c<p_{2\to2}$, including critical exponents and estimates slightly above criticality. Results for nonunimodular symmetries and other group classes enlarged the territory where the conjecture was known. The remaining task was to establish the operator gap without those additional assumptions. The present manuscript explicitly builds on this earlier framework. The L² boundedness condition

4. What the new manuscripts claim

For every infinite, connected, locally finite, nonamenable, quasi-transitive undirected graph, the manuscript states

$$\sup_{0\leq p<p_c}\|T_p\|_{2\to2} =\|T_{p_c}\|_{2\to2}<\infty, \qquad p_c<p_{2\to2}\leq p_u.$$

The first assertion controls the operator uniformly all the way to criticality. The second pushes boundedness a positive distance beyond criticality and therefore separates existence from uniqueness. Applied to Cayley graphs, it concerns every fixed finite symmetric generating set of a nonamenable finitely generated group.

There is also a simultaneous statement. Assign each edge an independent uniform label in $[0,1]$, declaring it open at parameter $p$ when its label is at most $p$. This constructs all densities on one probability space. The manuscript gives a closed interval $[p_1,p_2]$ with

$$p_c<p_1<p_2<1$$

on which infinitely many infinite clusters exist simultaneously at every parameter, with probability one. This is stronger in its quantifiers than saying that each separately chosen parameter has the property almost surely.

The operator estimate implies the triangle condition:

$$\sum_{y,z\in V} T_{p_c}(v,y)T_{p_c}(y,z)T_{p_c}(z,v)<\infty.$$

This sum counts a weighted abundance of triangular connection patterns. Through earlier theorems, its finiteness yields mean-field critical powers. In particular, the manuscript records

$$\begin{aligned} \chi_v(p)&\asymp(p_c-p)^{-1} &&(p\uparrow p_c),\\ \mathbb P_p(|C_v|=\infty)&\asymp p-p_c &&(p\downarrow p_c),\\ \mathbb P_{p_c}(|C_v|\geq n)&\asymp n^{-1/2} &&(n\to\infty). \end{aligned}$$

Here $a\asymp b$ means that their ratio lies between two positive constants in the indicated regime. It does not assert an exact leading constant. Critical cluster-radius tails have order $n^{-1}$, both for distance measured along open paths and for distance in the original graph.

Finally, below $p_{2\to2}$, connection probabilities between specified vertices decay exponentially with their graph distance. This can coexist with infinite clusters: there are exponentially many possible distant vertices, even though connecting to any particular one is unlikely.

5. Why the advance matters

Importance: 4/5. The claim resolves a central conjecture linking large-scale geometry to the number of infinite clusters. Its scope includes arbitrary nonamenable Cayley graphs with fixed generators and extends beyond group graphs to finitely many vertex types.

The operator estimate gives the result additional reach. It connects a geometric assumption to an entire collection of critical laws, rather than only asserting an interval somewhere between two thresholds. The exponents match branching-process predictions even when the underlying graph has many cycles and no obvious tree decomposition.

The rating reflects that breadth within percolation and geometric probability. The theorem remains about a specified class of independent percolation models; it does not establish corresponding claims for every interacting spin system or for amenable geometries.

6. The main idea

The manuscript’s strategy is to convert the graph’s ability to disperse a random walk into control of cluster connections. Nonamenability makes a suitable symmetric lazy random walk contract square-sum norms by a factor strictly smaller than one. Repeating the walk disperses mass exponentially quickly. The challenge is that a percolation cluster can concentrate mass along rare, unusually large connected regions.

First obtain rough critical control

The first stage constructs ways of moving through sampled clusters while limiting the size of the pieces being explored. Independent cluster pieces can be spliced into a probability kernel. If critical clusters were too heavy-tailed, this kernel would spread mass efficiently at unexpectedly little cost.

Differential inequalities for expected cluster sizes turn that possibility into a contradiction. The initial conclusion is deliberately weaker than the final mean-field law: the expected size below criticality is bounded by a power arbitrarily close to $(p_c-p)^{-1}$. Related exploration estimates give correspondingly weak control of triangle sums. These estimates are stepping stones; the argument does not assume the desired exponent in advance.

Separate norm growth from total mass

The central construction modifies connections through narrow open corridors. A corridor is an open path whose internal vertices have no other open edges attached. Such a path can be especially influential: closing one of its essential edges destroys a particular connection.

The manuscript penalizes connections using selected corridors, with the corridors organized using random-walk paths. Expansion makes long walk segments small in operator norm. At the same time, the modification can substantially reduce row sums, which measure total connection mass.

Suppose the critical operator norm were unbounded. Near criticality, the argument would then force the modified connection matrix to retain a large square-sum norm while its total mass became too small to support that norm. The earlier rough estimates control the error terms in this comparison. This is the main contradiction behind boundedness at $p_c$.

The useful distinction is between two measurements of the same matrix. Row sums can become large because connections spread over many vertices; an operator norm detects whether they amplify a square-summable input. Nonamenable geometry supplies room for the first effect without the second.

Open a few more edges

Once critical boundedness is available, a controlled addition of open edges extends it slightly above $p_c$. Percolation arguments often call this addition sprinkling. The established relation between operator boundedness and uniqueness then places that interval below $p_u$.

The triangle condition follows directly from the bounded operator, and the critical exponents follow through the earlier conditional theory. Thus the final consequences depend on both the new estimate and the existing machinery connecting that estimate to percolation behavior.

7. Scope, limitations, and open questions

The assumptions are essential to the statement being explained: independent bond percolation, an undirected locally finite graph, finitely many symmetry types, and positive isoperimetric constant. The manuscript does not make the same universal assertion for arbitrary inhomogeneous graphs or dependent edge models.

It does not say $p_u<1$. Regular trees already show why that would be false. Nor does it claim strict inequality between $p_{2\to2}$ and $p_u$; the theorem only orders those two thresholds.

The critical powers compare quantities up to constants depending on the graph and, where relevant, the vertex. They do not identify a universal amplitude or an explicit numerical width for the coexistence interval on every input graph. Determining thresholds and quantitative constants for particular networks remains a different task.

Finally, coexistence and critical nonpercolation answer different questions. The former concerns a supercritical interval with many infinite clusters; the latter asks whether any infinite cluster exists exactly at $p_c$. The operator approach gives much finer information than a yes-or-no critical statement, and the manuscript’s argument uses the earlier theory for nonamenable graphs rather than requiring the separate new general critical-percolation claim.

8. Sources and further reading

The source below contains the nonuniqueness theorem, the stronger operator estimate, and the discussion of mean-field consequences. Its introduction also maps the earlier planar, group-theoretic, and symmetry-based partial results.

For the original question, read Benjamini and Schramm’s 1996 paper. For the framework that turns an operator gap into geometric and critical information, see Hutchcroft’s hyperbolic-graph theorem and L² boundedness paper. The operator notation is useful background, but the finite-region boundary condition is the main geometric idea to keep in mind.

Source manuscripts

• M1. OpenAI, Nonuniqueness of percolation on nonamenable quasi-transitive graphs (September 24, 2026).

Sources consulted on 11 October 2026 at repository revision fd4aeeb. Download the manuscript citations.