Parisi Variational Principle

• The Parisi variational principle is a framework that determines the thermodynamic free energy in mean-field spin glass models by optimizing over probability measures.
• It uses a Parisi PDE and stochastic control representations to capture the complete replica symmetry breaking structure and ensure strict convexity of the functional.
• Its dual formulations and Legendre transformations reveal deep convexity and unique minimizers that correspond to physical order parameters in disordered systems.

The Parisi variational principle is the foundational framework that determines the thermodynamic limit of the free energy in mean-field spin glass models, most notably the Sherrington-Kirkpatrick (SK) and mixed $p$-spin models. It expresses the free energy as a variational problem over the space of probability measures on $[0,1]$, encoding the full replica symmetry breaking (RSB) structure through a nondecreasing order-parameter profile. The principle is formulated via a functional incorporating both energy and entropy-like terms and admits multiple dual and stochastic representations. The profound mathematical structure of the Parisi principle not only determines the equilibrium statistics of disordered systems but also uniquely characterizes the associated minimizer and reveals deep convexity and duality properties underlying the phase structure of spin glasses.

1. Formulation of the Parisi Variational Principle

The canonical setting is the mixed $p$-spin Ising model, defined for $N$ spins $\sigma\in\{-1,1\}^N$ with Hamiltonian

$H_N(\sigma) = H'_N(\sigma) + h\sum_{i}\sigma_i,$

where $H'_N$ is a centered Gaussian process with covariance $$\mathbb{E}H'_N(\sigma^1)H'_N(\sigma^2)=N\xi(R_{1,2})$$ for overlap $R_{1,2}=N^{-1}\sum_i \sigma_i^1 \sigma_i^2$ and $\xi(r)=\sum_{p\ge2}\beta_p^2 r^p$.

Let $[0,1]$0 be a probability measure on $[0,1]$1 (the overlap order parameter), and define the Parisi PDE (for $[0,1]$2, $[0,1]$3):

$[0,1]$4

The Parisi functional is given by

$[0,1]$5

and the thermodynamic free energy per spin is

$[0,1]$6

This variational problem selects the correct RSB pattern minimized by the physical free energy (Mourrat, 2023, Auffinger et al., 2014, Jagannath et al., 2015).

2. Stochastic Control and Dynamic Programming Representations

For fixed $[0,1]$7, $[0,1]$8 admits an exact stochastic control representation:

$[0,1]$9

where the supremum is over progressively measurable controls $p$0. The optimal control is given by $p$1 for $p$2 satisfying the associated stochastic differential equation. This perspective identifies $p$3 as the value function for a Hamilton–Jacobi–Bellman (HJB) equation and naturally unifies the Parisi PDE, stochastic control, and RSB structure (Mourrat, 2023, Jagannath et al., 2015, Chen, 2015).

The feedback control and control-theoretic approach permits alternative proofs of strict convexity, existence, regularity, and uniqueness for the Parisi functional, circumventing Ruelle probability cascades or Cole-Hopf techniques (Jagannath et al., 2015, Auffinger et al., 2014).

3. Convexity, Uniqueness, and Variational Structure

A fundamental property is the strict convexity of the Parisi functional $p$4. The strict convexity is established through the dynamic programming and optimal control representation of $p$5, leveraging the convexity of the terminal data ($p$6) and linearity of the quadratic penalty. This ensures that for any pair $p$7, the combination is strictly convex, so the variational principle admits a unique minimizer $p$8 for all temperatures and external fields (Auffinger et al., 2014, Jagannath et al., 2015).

The unique minimizer $p$9 corresponds to the physical overlap distribution and satisfies a self-consistency (Euler–Lagrange) equation, through which all equilibrium overlaps and higher-order statistics can be reconstructed. In the replica symmetric phase, $N$0 is a Dirac measure; in the full RSB phase, it is continuous, encoding the hierarchical ultrametric structure.

4. Duality and the "Inverted" Variational Principle

Unlike standard statistical mechanics, where the free energy emerges as a supremum over energy minus entropy, the Parisi principle features an infimum, producing what is known as the "inverted variational principle." Auffinger–Chen established a Legendre-type duality: the free energy is Legendre-conjugate to the so-called "squared interaction" (quadratic Hamiltonian) free energy (Auffinger et al., 2016).

Concretely, denoting $N$1 as the free energy at inverse temperature $N$2, and $N$3 as the free energy of the squared interaction model, one has:

$N$4

which is the Legendre duality in temperature squared. In the Parisi language, minimization over order-parameter profiles in the primal corresponds to maximization in the dual—a rigorous realization and resolution of the "inverted" variational principle (Auffinger et al., 2016, Mourrat, 2023).

Further, via convex duality and Legendre transformation, the Parisi infimum can be recast as a genuine supremum over martingales on Wiener space:

$N$5

where $N$6 is the Legendre transform of $N$7, and $N$8 is the set of $N$9-valued filtration martingales. This eliminates the infimum, restoring the "energy minus entropy" structure and facilitating new computational and analytical techniques (Mourrat, 2023).

5. Analytical Consequences, Extensions, and Applications

The Parisi variational principle's convex-geometric structure ensures robustness of the solution and applies broadly across models. For example:

• In models such as the Potts spin glass, the order parameter space forms a much higher-dimensional convex cone, but symmetry and synchronization reduce the principle to an analogous one-dimensional form (Bates et al., 2023).
• In the generalized random energy (perceptron-GREM) models, the principle can be viewed as a solution to a hierarchical entropy-constrained variational problem, again with unique minimizers (Bolthausen et al., 2010).
• The Parisi functional’s directional derivatives yield criteria for optimality, stability under perturbations, and provide characterization of physical phenomena such as overlap positivity, disorder chaos, and phase transitions (Chen, 2015, Kersting et al., 2019).

Computationally, the stochastic and supremal martingale reformulations suggest novel algorithms (e.g., approximate message passing, stochastic gradient ascent on the martingale landscape) for approximating thermodynamic quantities (Mourrat, 2023).

6. Physical Interpretation and Thermodynamic Implications

The Parisi principle demonstrates that in disordered mean-field models, the thermodynamic free energy is determined not by a macroscopic entropy-energy tradeoff, but by a subtle, infinite-dimensional optimization problem where the order-parameter profile encodes all RSB levels. Notably, contrary to the traditional Boltzmann–Gibbs maximization, the infimum appears due to the concavity of the finite-volume Gibbs potentials at infinite temperature—a direct manifestation of disorder-induced correlations (Kersting et al., 2019).

Phase transitions manifest as loss of concavity or divergence in the finite-$\sigma\in\{-1,1\}^N$0 Gibbs potential expansions, signaling the breakdown of current symmetry-breaking ansatzes and the emergence of new RSB phases.

Table: Key Mathematical Objects in the Parisi Principle

Object	Definition/Role	Reference
Order parameter $\sigma\in\{-1,1\}^N$1	Probability measure on $\sigma\in\{-1,1\}^N$2, overlap distribution	(Mourrat, 2023, Auffinger et al., 2014)
Parisi PDE $\sigma\in\{-1,1\}^N$3	Quasilinear backward equation for value function	(Jagannath et al., 2015, Mourrat, 2023)
Parisi functional $\sigma\in\{-1,1\}^N$4	Free energy functional to be minimized	(Auffinger et al., 2014)
Dual functional $\sigma\in\{-1,1\}^N$5	Legendre-dual "squared Hamiltonian" free energy	(Auffinger et al., 2016)
Martingale representation	Supremum over $\sigma\in\{-1,1\}^N$6-valued Wiener space martingales	(Mourrat, 2023)

The Parisi variational principle thus establishes a unique and mathematically explicit procedure for computing the equilibrium statistics in mean-field disordered systems, with deep consequences for the geometry of the solution space, the nature of phase transitions, and the development of both analytical tools and computational algorithms.