[论文解读] Rate of convergence of the Nesterov accelerated gradient method in the subcritical case $\alpha \leq 3$
本文在亚临界情形 α ≤ 3 下建立了Nesterov加速梯度法的收敛速率,证明了连续 (AVD)α 系统满足 Φ(x(t)) − min Φ = O(t⁻²ᵃ⁄³)。该结果被推广至惯性前向-后向算法,表明当 α ≤ 3 时,对所有 p < 2α/3,收敛速率为 O(k⁻ᵖ),并展示了对扰动的鲁棒性,完整描绘了收敛速率随阻尼参数 α 变化的连续变化图景。
(<em>Based on a joint work with Z. Chbani and H. Riahi</em>)<br> In a Hilbert space setting $\mathcal{H}$, given $\Phi: \mathcal H o \mathbb R$ a convex continuously differentiable function, and $\alpha$ a positive parameter, we first study the asymptotic behavior of the inertial system with Asymptotic Vanishing Damping $$ \mbox{(AVD)}_{\alpha} \quad \quad \ddot{x}(t) + \frac{\alpha}{t} \dot{x}(t) + abla \Phi (x(t)) =0, $$ and then the associated inertial algorithms.<br> Depending on the value of $ \alpha $ with respect to 3, and based on new Lyapunov functions, we give a complete picture of the convergence properties as $t o + \infty$ of the trajectories generated by $\mbox{(AVD)}_{\alpha}$. As shown by Su-Boyd-Cand&egrave;s, the case $\alpha = 3$ corresponds to a continuous version of the accelerated gradient method of Nesterov, with the convergence rate $\Phi (x(t))-\min \Phi = \mathcal O (t^{-2})$. Our main result concerns the subcritical case $\alpha \leq 3$, where we show that $\Phi (x(t))-\min \Phi = \mathcal O (t^{-\frac{2}{3}\alpha})$. When $\alpha > 3$, we find the recent results by May and A-Chbani-Peypouquet-Redont concerning the weak convergence of the trajectories, and the convergence of the values with the order $o\left(t^{-2} ight)$. This overall picture shows a continuous variation of the rate of convergence of the values $\Phi(x(t))-\min_\mathcal H \Phi= \mathcal O (t^{-p(\alpha)}) $ with respect to $\alpha >0$: the coefficient $p(\alpha)$ increases linearly up to 2 when $\alpha$ goes from $0$ to $3$, then displays a plateau. We also consider the convergence of trajectories in the critical case $ \alpha = $ 3, with a positive response in some particular cases.<br> Then we consider structured convex minimization problems of the form $\min \left\lbrace \Theta:= \Phi + \Psi ight brace$, with $\Phi$ smooth and $\Psi$ nonsmooth. As a major property, the Lyapunov analysis of the continuous dynamics serves as a guideline for the study of the associated forward-backward inertial algorithms. We obtain a similar convergence rate for the sequence of iterates $(x_k)$: for $\alpha < 3$ we have $\Theta (x_k)-\min \Theta = \mathcal O (k^{-p})$ for all $p <\frac{2\alpha}{3}$, for $\alpha = 3$ \ $\Theta (x_k)-\min \Theta = \mathcal O (k^{-2})$ (FISTA, Beck-Teboulle, 2009), and for $\alpha > 3$ \ $\Theta (x_k)-\min \Theta = o (k^{-2})$ (A-Peypouquet, 2016). We conclude this study by showing that the results are robust with respect to external perturbations.<br> <br> <dl> <dt>[1]</dt> <dd>H. Attouch, Z. Chbani, J. Peypouquet, P. Redont, <i>Fast convergence of inertial dynamics and algorithms with asymptotic vanishing damping</i>, to appear in Math. Program. DOI: 10.1007/s10107-016-0992-8.</dd> <dt>[2]</dt> <dd>H. Attouch, J. Peypouquet, <i>The rate of convergence of Nesterov's accelerated forward-backward method is actually faster than $\frac{1}{k^2}$</i>, SIAM J. Optim., 26 (2016), No. 3, pp. 1824-1834.</dd> <dt>[3]</dt> <dd> A. Beck, M. Teboulle, <i>A fast iterative shrinkage-thresholding algorithm for linear inverse problems</i>, SIAM J. Imaging Sci., 2 (2009), No. 1, pp. 183-202.</dd> <dt>[4]</dt> <dd> A. Chambolle, C. Dossal, <i>On the convergence of the iterates of the “Fast Iterative Shrinkage/Thresholding Algorithm”</i>, Journal of Optimization Theory and Applications, 166 (2015), pp. 968-982</dd> <dt>[5]</dt> <dd> Y. Nesterov, <i>Introductory lectures on convex optimization: A basic course</i>, volume 87 of Applied Optimization. Kluwer Academic Publishers, Boston, MA, 2004</dd> </dl>
研究动机与目标
- 刻画在阻尼参数 α ≤ 3 的亚临界区域中,Nesterov加速梯度法的收敛速率,此前该速率尚不明确。
- 将连续动力系统 (AVD)α 扩展至包含光滑与非光滑分量的结构化凸优化问题。
- 推导相应惯性前向-后向算法的收敛速率,并建立对扰动的鲁棒性。
- 完整描绘收敛速率随阻尼参数 α 的变化图景,展示从 α → 0 到 α = 3 的连续过渡。
提出的方法
- 分析希尔伯特空间 H 中的二阶微分方程 (AVD)α:¨x(t) + (α/t)˙x(t) + ∇Φ(x(t)) = 0。
- 引入一个定制的李雅普诺夫函数 E(t) = t²ᵖ[Φ(x(t)) − min Φ] + ∥λ(t)(x(t) − z) + tᵖ˙x(t)∥² 以研究能量衰减。
- 使用时变阻尼系数 γ(t) = α/t 来建模Nesterov方法的连续版本。
- 通过引入非光滑函数的次微分 ∂Ψ,将连续动力系统适配至前向-后向框架。
- 应用柯西-施瓦茨不等式与积分估计以控制扰动影响,证明在外部扰动下的鲁棒性。
- 通过将连续系统进行时间离散化,推导出离散收敛速率,得到惯性前向-后向算法 (IFB)α。
实验结果
研究问题
- RQ1当阻尼参数 α 小于或等于 3(临界值)时,Nesterov加速梯度法的收敛速率是什么?
- RQ2连续 (AVD)α 系统的收敛速率如何随 α > 0 连续变化?
- RQ3连续动力系统的收敛性质能否推广至结构化凸问题的惯性前向-后向算法?
- RQ4外部扰动如何影响 (AVD)α 系统及其离散对应物的收敛速率?
- RQ5在临界情形 α = 3 下,是否无需额外假设即可保证轨迹收敛到最优解?
主要发现
- 当 α ≤ 3 时,连续 (AVD)α 系统达到收敛速率 Φ(x(t)) − min Φ = O(t⁻²ᵃ⁄³),表明随着 α 从 3 降低,收敛速率持续减小。
- 在临界情形 α = 3 时,本文在不施加对 Φ 的限制性假设下证明了轨迹收敛到最优解,扩展了先前结果。
- 对于惯性前向-后向算法 (IFB)α,当 α ≤ 3 时,对所有 p < 2α/3,有 Θ(xk) − min Θ = O(k⁻ᵖ)。
- 当 α > 3 时,算法实现 o(k⁻²) 收敛,与已知的 FISTA 快速速率一致。
- 结果对扰动具有鲁棒性:若 ∫ₜ₀^∞ tᵖg(t)dt < ∞,则 Φ(x(t)) − min Φ = O(t⁻²ᵖ),其中 p = min(1, α/3)。
- 本文完整建立了收敛速率关于 α 的图景,其中 p(α) = 2α/3 从 α = 0 到 α = 3 线性增加,从 0 增至 2,之后在 α > 3 时保持为 2。
更好的研究,从现在开始
从阅读论文到最终审阅,大幅缩短您的研究时间。
无需绑定信用卡
本解读由 AI 生成,并经人工编辑审核。