The Mézard–Parisi formula for diluted spin glasses

The explanation

Take 1000 tiny magnets, each pointing up (+1) or down (−1). Draw about 1500 random pairs among them. Give each pair a coupling of +1 or −1 by a coin flip. A +1 coupling rewards the pair for agreeing, and −1 for disagreeing. Each magnet then has on average 3 partners, not 999. This is a diluted spin glass: a finite mean number of interactions per spin, however large the systemSource 1.

Physicists want the pressure, the expected log of the partition function, per spinSource 2. The partition function adds a weight for every one of the 2^1000 states. A quick upper bound comes from averaging before taking the log. At inverse temperature 1 it gives log 2 + 1.5 × log cosh 1, about 1.344. That bound ignores the frustration of competing couplings, which can make the true value smaller.

The cavity method does better. Remove one spin, look at the fields its few partners exert, and put it back. When many competing states matter, the order parameter is a hierarchy of distributions of these fieldsSource 3. Mézard and Parisi built this picture for sparse spin glasses in 2001. Franz and Leone then turned some of it into rigorous bounds. Panchenko and Talagrand proved that every finite level of the hierarchy gives an upper boundSource 4. Whether the best such bound is exact was left as a conjecture.

This paper proves that the infimum of these bounds is the limiting pressureSource 5. The infimum is the lowest value the bounds approach. The result needs only first moments of the interactions and external field, meaning their averages must be finiteSource 6. It covers the Viana–Bray model, diluted even-spin models with symmetric couplings, and soft random even-K satisfiability at every positive temperatureSource 7. Satisfiability here means a random logic puzzle, softened so each broken clause costs energy rather than being forbidden.

The main obstacle is how several samples from the same system share spins. Call each sample a replica. The chance of a joint pattern depends on products over every subset of replica spins. Their spatial averages are multioverlaps, and pair overlaps alone do not fix themSource 8. The proof controls all of these at once, but only on average over the depths of a branching treeSource 9. That averaged control is enough, because the cavity step itself averages over finite trees.

The result has limits. Every interaction must involve an even number of spins. The couplings must satisfy Panchenko and Talagrand's factorization and positivity conditions. The formula is an infimum over trial laws, so it gives no closed form for any specific model.

How sure is this? A computer proof checker called Lean has checked a formal version of the main theorem. It covers even arity from 2 up, positive interaction density, and the class's stated assumptions. The paper's consequences for named models and their ground states are not listed in that check. No specialist review has been published.

What it doesn't show. This does not cover odd-size interactions, or models outside the stated factorization and positivity conditions.