核心概念
Optimization algorithms' convergence rates are formally analyzed using the Lean4 theorem prover, enhancing mathematical representation.
摘要
The article formalizes gradient and subgradient for convex functions, proving properties crucial for algorithmic convergence. It introduces first-order algorithms like Gradient Descent, Subgradient Descent, and Proximal Gradient methods with convergence analysis.
Introduction:
- First-order optimization algorithms are fundamental in addressing challenges in machine learning and engineering.
- Theoretical foundations ensuring efficacy demand rigorous formalization.
Mathematical Preliminaries:
- Definitions of subgradient, proximal operator, and differentiability in normed spaces.
- Properties of convex functions formalized in mathlib library.
Gradient and Subgradient in Lean:
- Definition of gradient using Riesz Representation Theorem on Hilbert space.
- Formalization of subgradient properties for non-smooth convex functions.
Properties of Convex Functions in Lean:
- Theorems establishing equivalence between convexity conditions and gradient monotonicity.
- Strongly convex function definitions and related theorems formalized.
Proximal Operator in Lean:
- Proximal property definition leading to unique proximal set under certain conditions.
- Theorems connecting proximal operator with subgradient for closed convex functions.
Convergence of First Order Algorithms in Lean:
Gradient Descent Method:
- Class structure defining general and fixed step size gradient descent methods.
- Convergence theorems for convex and strongly convex functions derived.
Subgradient Descent Method:
- Class structure defining subgradient descent method with diminishing step size.
- Convergence theorem for subgradient descent method under specific assumptions.
Proximal Gradient Method:
- Class structure defining proximal gradient method for composite optimization problems.
- Convergence rate theorem formalized under Assumption 2 with fixed step size.