the expectation threshold
The expectation threshold is a quantity that captures the simplest possible obstruction to a monotone property appearing in a random structure, and it sits at the centre of one of the deepest recent results in probabilistic combinatorics. For a monotone increasing property (one that, once true, stays true as you add more elements), there is a critical density p_c at which a random structure switches from typically lacking the property to typically having it — the threshold. The first moment method gives an immediate lower bound on this threshold: below the density at which the expected number of minimal witnesses for the property is around 1, the property is unlikely, simply because there is not even one witness on average.
Made precise, the expectation threshold q(F) of an increasing family F is, roughly, the least p for which one can place a 'cover' on F — a small collection of substructures whose presence is forced by F — whose total expected weight at density p is at least 1. It is by construction a lower bound for the true threshold p_c, because if even the cheapest cover is not expected to appear, the property cannot appear: q(F) <= p_c always. The natural question Kahn and Kalai asked in 2006 was how far apart these two thresholds can be. The expectation-threshold conjecture asserted that they are never more than a logarithmic factor apart: p_c <= O(log(size)) * q(F), so the trivial first-moment obstruction is, up to a log, the only obstruction.
This was proved in 2022 by Park and Pham (the fractional version of the Kahn-Kalai conjecture), one of the landmark theorems of the field, with a strikingly short proof. The practical upshot is enormous: to find the threshold of almost any monotone property up to a logarithmic factor, you no longer need a delicate second-moment or sharp-threshold analysis — you just compute the cheapest expected cover, a first-moment calculation. This unified and reproved a long list of threshold results (perfect matchings in hypergraphs, Hamilton cycles, bounded-degree spanning trees) almost mechanically. The honest caveat is the surviving logarithmic gap: for properties with a sharp threshold the log factor is real and the expectation threshold alone does not pin down the constant, only the order up to log.
For a perfect matching in the random hypergraph, the cheapest obstruction is having an isolated vertex: the expectation threshold is set by the density at which every vertex is expected to be covered, q ~ log(n)/something. Kahn-Kalai/Park-Pham then guarantees the true threshold is within a log factor, recovering the known answer without a custom second-moment argument — a single expectation calculation replaces a paper's worth of work.
The first-moment cover is, up to a log, the whole story: q(F) <= p_c <= O(log) q(F).
The Park-Pham theorem proves the fractional expectation-threshold version; the surviving log factor is genuine for sharp-threshold properties, so the expectation threshold gives the order of p_c but not its leading constant.