<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Acceleration on Chen Kai Blog</title><link>https://www.chenk.top/en/tags/acceleration/</link><description>Recent content in Acceleration on Chen Kai Blog</description><generator>Hugo</generator><language>en</language><lastBuildDate>Tue, 20 Sep 2022 09:00:00 +0000</lastBuildDate><atom:link href="https://www.chenk.top/en/tags/acceleration/index.xml" rel="self" type="application/rss+xml"/><item><title>Optimization (5): Acceleration Beyond Nesterov</title><link>https://www.chenk.top/en/optimization-theory/05-acceleration-beyond-nesterov/</link><pubDate>Tue, 20 Sep 2022 09:00:00 +0000</pubDate><guid>https://www.chenk.top/en/optimization-theory/05-acceleration-beyond-nesterov/</guid><description>&lt;p>Article 02 introduced Nesterov acceleration and showed it improves the per-iteration cost from &lt;span class="math-inline">$\kappa$&lt;/span>
 to &lt;span class="math-inline">$\sqrt{\kappa}$&lt;/span>
. This article asks the deeper questions:&lt;/p>
&lt;ul>
&lt;li>&lt;strong>Why &lt;span class="math-inline">$\sqrt{\kappa}$&lt;/span>
 and not faster?&lt;/strong> We prove a matching lower bound — no first-order method can do better.&lt;/li>
&lt;li>&lt;strong>Is Nesterov the only way?&lt;/strong> Polyak&amp;rsquo;s Heavy-Ball method achieves the same rate using a completely different update rule.&lt;/li>
&lt;li>&lt;strong>Can we accelerate any solver?&lt;/strong> The Catalyst framework wraps a black-box optimizer to gain the accelerated rate, at the cost of solving a regularized subproblem.&lt;/li>
&lt;/ul>
&lt;p>The unifying tool is a &lt;strong>Lyapunov potential&lt;/strong> — a non-negative quantity that the algorithm decreases at every step. Both Nesterov and Heavy-Ball admit Lyapunov proofs, and the lower bound essentially says no Lyapunov decrease can happen faster.&lt;/p></description></item></channel></rss>