We formulate two classes of first-order algorithms more general than previously studied for minimizing smooth and strongly convex or, respectively, smooth and convex functions. We establish sufficient conditions, via new discrete Lyapunov analyses, for achieving accelerated convergence rates which match Nesterov's methods in the strongly and general convex settings. Next, we study the convergence of limiting ordinary differential equations (ODEs) and point out currently notable gaps between the convergence properties of the corresponding algorithms and ODEs. Finally, we propose a novel class of discrete algorithms, called the Hamiltonian assisted gradient method, directly based on a Hamiltonian function and several interpretable operations, and then demonstrate meaningful and unified interpretations of our acceleration conditions.
翻译:我们制定了两个比以前研究的更一般的一阶算法类别,用于最小化光滑、强凸或光滑、凸函数。我们通过新的离散李亚普诺夫分析,建立了足够的条件,以在强凸和一般凸设置中实现匹配Nesterov方法的加速收敛率。接下来,我们研究极限常微分方程(ODEs)的收敛性,并指出当前算法和ODEs的相应收敛性质之间的显著差距。最后,我们提出了一种新颖的离散算法类别,称为哈密顿助推梯度方法,直接基于哈密顿函数和几种可解释的操作,然后演示了我们加速条件的有意义和统一的解释。