Math 2.0

Statistical mechanics, structure of matter Disordered systems Theorem

Parisi formula

Builds on Sherrington–Kirkpatrick modelSpin glassProbability measure

Statement

Let β≥0\beta \ge 0 and h∈ℝh \in \R. For N≥2N \ge 2, let (gij)1≤i<j≤N(g_{ij})_{1 \le i < j \le N} be independent standard Gaussian random variables, and let FN=1N𝔼log⁡∑σ∈{−1,+1}Nexp⁡(βN∑i<jgijσiσj+h∑i=1Nσi)F_N = \frac{1}{N} \mathbb{E} \log \sum_{\sigma \in \set{-1,+1}^N} \exp\Big( \frac{\beta}{\sqrt{N}} \sum_{i < j} g_{ij} \sigma_i \sigma_j + h \sum_{i=1}^N \sigma_i \Big) be the free energy of the Sherrington–Kirkpatrick model. For a probability measure ζ\zeta on [0,1][0, 1], let α(q)=ζ([0,q])\alpha(q) = \zeta([0, q]), and let Φζ ⁣:[0,1]×ℝ→ℝ\Phi_\zeta \colon [0, 1] \times \R \to \R solve the Parisi equation ∂tΦζ=−β22(∂x2Φζ+α(t) (∂xΦζ)2),Φζ(1,x)=log⁡cosh⁡x.\partial_t \Phi_\zeta = -\frac{\beta^2}{2} \Big( \partial_x^2 \Phi_\zeta + \alpha(t) \, (\partial_x \Phi_\zeta)^2 \Big), \qquad \Phi_\zeta(1, x) = \log \cosh x. Define the Parisi functional 𝒫(ζ)=log⁡2+Φζ(0,h)−β22∫01q α(q) dq.\mathcal{P}(\zeta) = \log 2 + \Phi_\zeta(0, h) - \frac{\beta^2}{2} \int_0^1 q \, \alpha(q) \, dq. Then FNF_N converges as N→∞N \to \infty, and lim⁡N→∞FN=inf⁡ζ𝒫(ζ),\lim_{N \to \infty} F_N = \inf_{\zeta} \mathcal{P}(\zeta), where the infimum runs over all probability measures ζ\zeta on [0,1][0, 1].

Remarks

For ζ\zeta with finite support, α\alpha is a step function, and the Parisi equation is solved step by step with the Cole–Hopf transform. The general case follows by continuity in ζ\zeta. Parisi found the formula around 1980 with the non-rigorous replica trick. Guerra proved in 2003 the upper bound FN≤𝒫(ζ)+o(1)F_N \le \mathcal{P}(\zeta) + o(1) for every ζ\zeta by an interpolation argument. Talagrand proved the matching lower bound in 2006, which completed the proof.

Example: take h=0h = 0 and ζ=δ0\zeta = \delta_0, so that α≡1\alpha \equiv 1. The Cole–Hopf transform gives Φζ(0,x)=log⁡𝔼cosh⁡(x+βZ)=log⁡cosh⁡x+β2/2\Phi_\zeta(0, x) = \log \mathbb{E} \cosh(x + \beta Z) = \log \cosh x + \beta^2 / 2 with ZZ standard Gaussian. Then 𝒫(δ0)=log⁡2+β2/2−β2/4=log⁡2+β2/4\mathcal{P}(\delta_0) = \log 2 + \beta^2/2 - \beta^2/4 = \log 2 + \beta^2/4, the annealed free energy. This value is the true limit for β<1\beta < 1, while for β>1\beta > 1 the infimum is strictly smaller.

Going deeper

For a student

The infimum is attained at a unique measure ζ\zeta, the Parisi measure. In the physics picture, and in rigorous form for generic mixed pp-spin models, it is the limiting law of the overlap 1N∑iσi1σi2\frac{1}{N} \sum_i \sigma^1_i \sigma^2_i of two independent samples from the Gibbs measure. At high temperature it is a single point mass, which is called replica symmetry. At low temperature it is expected to have a continuous part, which is called full replica symmetry breaking. Panchenko later extended the formula to mixed pp-spin models, with a proof based on the ultrametric structure of the overlaps. The same variational form appears in many problems on random constraint satisfaction and on high-dimensional statistics.

Reading path

To understand this page from the ground up, read these 6 pages in order. The foundations come first.

Show the path as a graph
Each page uses the pages above it. The colour of the bar on a box shows how strong the evidence is, and the box also gives it as text. The list has the same pages.
  • no evidence level
  1. Ising model example
  2. Normal distribution concept
  3. Probability measure concept
  4. Spin glass concept
  5. Sherrington–Kirkpatrick model example
  6. Parisi formula statement

Further reading

  1. Web https://en.wikipedia.org/wiki/Spin_glass
About this page
Tag32S4Y36QNK
URIm2://32S4Y36QNK
Kindstatement, theorem
WritingW0 stub
MSC82B44, 60K35, 82D30
StatusThis copy of the site has no build data. Run m2 status 32S4Y36QNK.
AboutSherrington–Kirkpatrick model
UsesSherrington–Kirkpatrick model, Spin glass, Probability measure
Changed3 November 2026

Markdown · JSON · Edit on GitHub

What this page needs for W1

  • an inventory row: the row of the inventory that the node accounts for (0 of 1)
  • a source with a locator (0 of 1)

The writing levels say how complete the text of a page is. They do not measure its truth.

What links here

No page links to this page yet.