arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2301.03576math.OC
收藏

统一Nesterov加速梯度法求解凸和强凸目标函数:从连续时间动态到离散时间算法

Unifying Nesterov's Accelerated Gradient Methods for Convex and Strongly Convex Objective Functions: From Continuous-Time Dynamics to Discrete-Time Algorithms

Jungbin Kim, Insoon Yang

摘要:

虽然Nesterov的加速梯度(NAG)方法已经从不同的角度进行了研究,但仍然不清楚为什么最流行的NAG形式必须分别处理凸和强凸目标函数。出于这种不一致性,我们提出了一个NAG方法,统一现有的凸和强凸的情况下。首先,我们设计了一个拉格朗日函数,连续扩展的第一布雷格曼拉格朗日的强凸设置。作为一个特定的情况下,欧拉-拉格朗日方程的拉格朗日量,我们推导出一个常微分方程(ODE)模型,我们称之为统一的NAG ODE,桥梁之间的差距ODE模型NAG凸和强凸目标函数。然后,我们设计了统一的NAG,一种新的动量方法,连续时间限制对应于统一的常微分方程。在强凸性参数$[0, +\infty)$上,统一NAG和统一ODE的系数和收敛速度是连续的。与现有的强凸目标函数的算法和常微分方程不同,对于非强凸目标函数,统一NAG和统一NAG常微分方程总是比已知的算法和常微分方程有更好的收敛性保证.当考虑具有小的$μ$的强凸目标函数时,该性质在实际观点中是有益的。此外,我们扩展我们的统一动力学和算法的高阶设置。最后,我们提出了一种新的极小化强凸目标函数梯度范数的常微分方程模型--统一NAG-G常微分方程。我们的统一拉格朗日框架是至关重要的,在这个过程中构造常微分方程。有趣的是,使用我们的新工具,称为微分核,我们观察到,统一的NAG常微分方程和统一的NAG-G常微分方程有一个反转置关系。

英文摘要:

Although Nesterov's accelerated gradient (NAG) methods have been studied from various perspectives, it remains unclear why the most popular forms of NAG must handle convex and strongly convex objective functions separately. Motivated by this inconsistency, we propose an NAG method that unifies the existing ones for the convex and strongly convex cases. We first design a Lagrangian function that continuously extends the first Bregman Lagrangian to the strongly convex setting. As a specific case of the Euler--Lagrange equation for this Lagrangian, we derive an ordinary differential equation (ODE) model, which we call the unified NAG ODE, that bridges the gap between the ODEs that model NAG for convex and strongly convex objective functions. We then design the unified NAG, a novel momentum method whereby the continuous-time limit corresponds to the unified ODE. The coefficients and the convergence rates of the unified NAG and unified ODE are continuous in the strong convexity parameter $μ$ on $[0, +\infty)$. Unlike the existing popular algorithm and ODE for strongly convex objective functions, the unified NAG and the unified NAG ODE always have superior convergence guarantees compared to the known algorithms and ODEs for non-strongly convex objective functions. This property is beneficial in practical perspective when considering strongly convex objective functions with small $μ$. Furthermore, we extend our unified dynamics and algorithms to the higher-order setting. Last but not least, we propose the unified NAG-G ODE, a novel ODE model for minimizing the gradient norm of strongly convex objective functions. Our unified Lagrangian framework is crucial in the process of constructing this ODE. Fascinatingly, using our novel tool, called the differential kernel, we observe that the unified NAG ODE and the unified NAG-G ODE have an anti-transpose relationship.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑