Abstract:In this paper, we study the convergence properties of the Stochastic Gradient Descent (SGD) method for finding a stationary point of a given objective function $J(\cdot)$. The objective function is not required to be convex. Rather, our results apply to a class of ``invex'' functions, which have the property that every stationary point is also a global minimizer. First, it is assumed that $J(\cdot)$ satisfies a property that is slightly weaker than the Kurdyka-Lojasiewicz (KL) condition, denoted here as (KL'). It is shown that the iterations $J(\boldsymbol{\theta}_t)$ converge almost surely to the global minimum of $J(\cdot)$. Next, the hypothesis on $J(\cdot)$ is strengthened from (KL') to the Polyak-Lojasiewicz (PL) condition. With this stronger hypothesis, we derive estimates on the rate of convergence of $J(\boldsymbol{\theta}_t)$ to its limit. Using these results, we show that for functions satisfying the PL property, the convergence rate of both the objective function and the norm of the gradient with SGD is the same as the best-possible rate for convex functions. While some results along these lines have been published in the past, our contributions contain two distinct improvements. First, the assumptions on the stochastic gradient are more general than elsewhere, and second, our convergence is almost sure, and not in expectation. We also study SGD when only function evaluations are permitted. In this setting, we determine the ``optimal'' increments or the size of the perturbations. Using the same set of ideas, we establish the global convergence of the Stochastic Approximation (SA) algorithm under more general assumptions on the measurement error, compared to the existing literature. We also derive bounds on the rate of convergence of the SA algorithm under appropriate assumptions.

A Certain Class of Statistical Probability Convergence and Its Applications to Approximation Theorems

On Convergence to Stochastic Integrals

Statistical Convergence of Random Variables

Deferred Riesz statistical convergence via power series method

Statistical convergence of integral form of modified Szász–Mirakyan operators: an algorithm and an approach for possible applications

The Convergence of the Sums of Independent Random Variables under the Sub-Linear Expectations

Statistical Convergence and Convergence in Statistics

Convergence Rates for Stochastic Approximation: Biased Noise with Unbounded Variance, and Applications

Approximation by a new kind of -Bernstein–Kantorovich operators

Strong Approximation Theorems for Sums of Random Variables when Extreme Terms Are Excluded

Approximation by Szász-integral type operators

Rates of convergence in the strong law of large numbers for weighted averages of nonidentically distributed random variables

Certain Approximation Results for Kantorovich Exponential Sampling Series

On approximation of functions, conjugate to the functions of several variables belonging to weighted Lipschitz and Zygmund classes

Statistical convergence of complex uncertain sequences defined by Orlicz function

Statistical and rough statistical convergence in an S-metric space

Convergence of Density Approximations for Stochastic Heat Equation

Concentration of Randomized Functions of Uniformly Bounded Variation

Formalization of a Stochastic Approximation Theorem

Rate of Convergence of Space Time Approximations for Stochastic Evolution Equations

Functional weak convergence of stochastic integrals for moving averages and continuous-time random walks