All-temperature pressure of orthogonally invariant Ising spin glasses

The explanation

Take two tiny magnets, called spins, each pointing up (+1) or down (−1). Give the pair an energy reward of 1 when they agree and a penalty of 1 when they disagree. There are four ways to set the two spins. Two agree and two disagree. Weight each setting by e raised to its reward, so agreeing settings count e and disagreeing ones count 1/e. Average these four weights, take the log, and divide by the number of spins. The result is the pressure. Here it is ½ × log(cosh 1), about 0.217.

A spin glass does the same with N spins and random rewards between every pair. The pressure is the key number. It sums up the system's behavior at one temperature, and its limit as N grows is what physicists want. For the classic Sherrington–Kirkpatrick model, Parisi introduced a layered order parameter in 1979Source 1. Guerra proved one half of it, and Talagrand proved the full formula in 2006Source 2,Source 3.

The rewards form a matrix. A matrix has eigenvalues, which say how strongly it acts in certain directions, the eigenvectors. This paper studies orthogonally invariant spin glasses. You fix the eigenvalues and spin the eigenvectors to a completely random direction. The question is whether the eigenvalues alone fix the limiting pressure.

The obstacle is a mismatch. The eigenvalues describe the matrix on all of space. But spins only take the values ±1, the corners of a cube, so the matrix is only ever tested thereSource 4. Earlier rigorous formulas for these models needed high temperatureSource 5. Bhattacharya and Sen proved one such formula with no outside fieldSource 6. At high temperature the spins behave simply and the hard structure never appears.

The paper proves a formula for the limiting pressure at every fixed temperatureSource 7. It assumes the spread of eigenvalues settles to a fixed shape on a bounded range. It also assumes the largest and smallest eigenvalues do not stray outside that range. The pressure converges on average, and also for almost every random draw, to a variational formulaSource 8,Source 9. A variational formula is one given as the best value over a family of trial choices. The eigenvalue spread need not have a density or any minimum varianceSource 10.

The formula keeps the layered overlap structure known from the Sherrington–Kirkpatrick modelSource 11. Overlaps measure how alike two random samples of the spins are. The paper also handles outside fields acting on each spinSource 12. By letting the temperature go to zero, it gets the lowest possible energy with no fieldSource 13. One example is a Hopfield model built from a random Gaussian matrix. There the formula also gives the matrix's largest and smallest stretch over ±1 vectorsSource 14.

The proof uses a tree-like overlap structure to describe all the overlaps at onceSource 15. The order of the minimum and maximum in the formula matters, and the proof never swaps themSource 16.

How sure is this? Lean, a computer proof checker, has formal versions of the main limits. They cover the pressure in expectation and almost surely, every positive temperature and the ground-state limit. They also cover outside fields, random spectra and the Gaussian-pattern case. No specialist review of the paper has been published.

What it doesn't show. This does not cover eigenvalue spreads with outliers outside the limiting range.