The Setting
The zero-temperature Parisi formula expresses the ground-state energy of the Sherrington-Kirkpatrick model as the minimum of the Parisi functional over order parameters , where is computed from the value function of a stochastic control problem. Given an order parameter and another one, , the question is how changes to first order when moves a small step toward , and what this says when is a minimizer.
Order Parameters and the Parisi Functional
The extension by zero of a real function on the time interval is the function on equal to for and to elsewhere.✦ An order parameter is a real function on that is nonnegative, nondecreasing and right-continuous, and whose extension by zero is Lebesgue integrable.✦ Wherever an order parameter is evaluated outside , its extension by zero is meant.
For order parameters , and a weight , the chord is the order parameter whose value at is .✦ The variation step of index is ✦, and the chord perturbation is the chord ✦.
A Brownian system on a measurable space is a probability measure on , its law, together with a family of random variables on that is a standard Brownian motion under .✦ Expectation under is written . The increment filtration of after a time is the filtration in which is generated by the increments , .✦ A bounded control from time is a real process, indexed by the time elapsed since , that is progressively measurable for this filtration and satisfies for all and .✦
The value function of an order parameter on a Brownian system is the function of a time and a position given by (1)✦ the supremum running over bounded controls from time ; the quantity in the brackets is the control payoff of . The value gradient is the derivative of at .✦
The Parisi functional on assigns to an order parameter the number (2)✦ An order parameter is a Parisi minimizer when for every order parameter .
The Diffusion and the Function
The left mesh time of level of a time is , the last point of the grid of mesh not after .✦ The progressive version of is the process , the limit taken at each sample point.✦ The driving Brownian motion of is the process .✦
A bounded Lipschitz drift is a jointly measurable function of a time and a position , with a bound and a Lipschitz constant , such that and for all , , .✦ Along a real process , the correction iterates of are and ✦, and its drift correction is their limit as ✦. The drift equation solution of along is the Picard solution of .✦
For a horizon , write for the clamp of a time to the strip .✦ The clamped drift is the product of the clamped coefficient and the clamped gradient; for it is the drift .✦ The strip drift with horizon is the bounded Lipschitz drift given by the clamped drift.✦
The horizons , , increase from toward .✦ The localized diffusion is the drift equation solution along for the strip drift with horizon , so that .✦ The raw diffusion limit is .✦ The gradient cut at time is the function of a time and a position equal to for and to for , and the raw feedback is the value of the gradient cut at .
The diffusion of is the process (3)✦ the integrated form of , , with the gradient read along . The second moment of the gradient along is the function of a time given by (4)✦
The Claim
Theorem 1. Let be a Brownian system on , let be an order parameter, and let be the second moment of the gradient along the diffusion of . For every order parameter , the chord perturbations from toward satisfy (5)✦ If is a Parisi minimizer, then every order parameter satisfies (6)✦
In the paper, the result reads: if and, for some , is an order parameter for , then the one-sided derivative of at is .
What Is Checked
The limit (5) and the inequality (6) are stated over a Brownian system and order parameters and , through the Parisi functional (2) built on the value function (1), the chord perturbations with steps , the condition that be a Parisi minimizer, and the function of (4) along the diffusion (3).
About these constructions the library proves that each variation step lies in ✦ and each horizon in ✦, and that for a horizon the clamped drift is jointly measurable, bounded in absolute value by and Lipschitz in the position.