Exact percolation laboratory

The Other Bunk Wins by One

On every small bunk you try, the nearer copy seems at least as easy to reach. Enumerate the hidden six-hyperedge core and the count reverses. Then operate the weighted graph gadget that carries that tiny reversal into an ordinary planar graph.

Every probability and count below is computed in this browser. The first instrument is a sandbox, not evidence for every graph. The second is an exact hypergraph result. The third shows the sufficient inequality used in the peer-reviewed ordinary graph proof.

Layer oneBuild a bunk that behaves

Choose a base graph, change its edges, choose the always-open posts, and set the endpoints. The browser enumerates every open or closed state of every horizontal edge on both levels.

Exact small-graph sandbox

Preset, then edit any edge or post.

Horizontal edges in each copy

Always-open vertical posts

same level P(u to v) computing live
other level P(u to v') computing live
same minus other computing live

Enumerating live.

This census fixes T = {2,3}, u = 1, v = 4, and p = 1/2, then generates every labelled connected graph on four vertices.

A result here is exact for this selected finite graph, posts, endpoints, and p. It does not establish a universal inequality.

Layer twoReveal the hidden core

The proof starts from Lawrence Hollom's alternative bunkbed hypergraph model. Each of six three-vertex hyperedges is placed on exactly one of the two levels by a fair bit. This is not ordinary bond percolation on a graph.

Six hyperedges, every state

Posts are fixed at u2, u7, and u9.

same-level successes
waiting
versus
cross-level successes
waiting

The exact count will be generated from the incidence list below.

Enumerate to inspect a state.

"Count-breaking" means the thirteenth cross success in ascending mask order. That ordering is a display convention, not part of the theorem.

The mechanism survivesTurn a hyperedge into a graph

For each triple {a,b,c}, the paper replaces the hyperedge with a fan graph Gn. Its spine edges open with probability p. Its star edges from a open with probability 1-p. Only at p = 1/2 is the gadget uniform.

Gadget terminal probabilities

Five mutually exclusive terminal partitions, computed from the paper's recurrences.

abccomputing
a|bccomputing
ab|ccomputing
ac|bcomputing
a|b|ccomputing
computing the robust inequality live

Computing.

computing gadget and graph sizes live

The base graph is duplicated only after this bookkeeping. Three always-open posts are then added between its two levels.

computing sampling burden live

This meter uses two independent binomial estimates and their worst-case one-standard-error bound. It is an order-of-magnitude diagnostic, not a proof method.

The check

These are fresh recomputations from the active controls. Published values are named separately.

Waiting for the optional four-vertex graph census.
Waiting for the exact hypergraph census.
Computing the gadget probabilities and inequality.
Computing the formal graph bookkeeping.
Computing the sampling comparison.
Uncertainty, choices, approximations, and conventions
  • The small sandbox depends freely on the base graph, edge set, post set T, endpoints u and v, and p. It uses vertex labels starting at 1 and independent horizontal edges on both levels. Posts are always open. One degenerate case worth naming, because the instrument will not warn you: if either endpoint is itself a post, the difference is exactly zero by definition rather than by evidence. If v is a post then the post edge is always open, so (v,0) and (v,1) are always in one component and the two events being compared are literally the same event. If u is a post then (u,0) and (u,1) are always in one component, and swapping the two levels is a symmetry of the remaining randomness, so the two probabilities are equal for that reason instead. Selecting every vertex as a post makes the whole comparison vacuous in this way, and a run of exact zeros there is not support for the conjecture. The presets deliberately use partial post sets.
  • The small sandbox and six-hyperedge census are exhaustive finite sums. They are not Monte Carlo. Decimal formatting rounds the displayed probabilities, while the state census remains integer-exact.
  • In the alternative hypergraph core, each base hyperedge has exactly one active level. That model is visually and mathematically separate from ordinary bond percolation.
  • For Gn at general p, spine edges use p and star edges use 1-p. The five displayed probabilities use IEEE 754 arithmetic. Extremely small a|bc values are shown in logarithmic form to avoid pretending that floating-point underflow is zero.
  • The exact normalized robust test is evaluated in a cancellation-free form. The simpler coefficient is only the paper's sufficient lower bound. The formal parameter is a free construction choice that clears the required threshold.
  • The formal theorem proves a negative gap whose magnitude is published only as less than 10-4331. That is an upper bound on magnitude, not the gap and not a lower bound.
  • The reported 82-vertex variant and its gap on the order of 10-47 are computer-assisted. This page does not present them as the formal theorem.
  • The page's publication date uses the build date because the supplied spec did not contain a separate LANDING date.
What is proved, and what remains open

The universal fixed-post bunkbed conjecture is false. Gladkov, Pak, and Zimin proved a counterexample at p = 1/2 on a connected planar base graph. The minimum size of an unweighted Bernoulli bond-percolation counterexample is not known. Which restricted graph families satisfy related inequalities is not fully classified. The monotonicity of percolation two-point connection probabilities with distance on grid-like graphs also remains unresolved.

Primary sources and review status