<p>We propose several modifications of the Barzilai–Borwein (BB) step size in the variance reduction (VR) methods for finite-sum optimization problems. Our first approach relies on a scalar function, which we call the TaiL Function (TLF). The TLF maps the computed BB step size to some positive real number, which will be used as the step size instead. The computational overhead is almost negligible and the functional forms of TLFs in this work don’t involve any problem-dependent parameters. In the strongly convex setting, due to the undesirable appearance of the condition number <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="245_2025_10316_Article_IEq1.gif" Format="GIF" Height="10" Rendition="HTML" Resolution="72" Type="Linedraw" Width="13" /> </InlineMediaObject> <EquationSource Format="TEX">\(\kappa \)</EquationSource> </InlineEquation> in the linear convergence rate, the IFO complexity of VR methods with BB step size has the form <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="245_2025_10316_Article_IEq2.gif" Format="GIF" Height="19" Rendition="HTML" Resolution="72" Type="Linedraw" Width="155" /> </InlineMediaObject> <EquationSource Format="TEX">\(\mathcal {O}((n+\kappa ^a)\kappa \log (1/\epsilon ))\)</EquationSource> </InlineEquation>, <InlineEquation ID="IEq3"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="245_2025_10316_Article_IEq3.gif" Format="GIF" Height="17" Rendition="HTML" Resolution="72" Type="Linedraw" Width="51" /> </InlineMediaObject> <EquationSource Format="TEX">\(a\in \mathbb {R}_{+}\)</EquationSource> </InlineEquation>. With the utilization of the TLF, the aforementioned complexity is improved to <InlineEquation ID="IEq4"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="245_2025_10316_Article_IEq4.gif" Format="GIF" Height="20" Rendition="HTML" Resolution="72" Type="Linedraw" Width="145" /> </InlineMediaObject> <EquationSource Format="TEX">\(\mathcal {O}((n+\kappa ^{\tilde{a}})\log (1/\epsilon ))\)</EquationSource> </InlineEquation>, <InlineEquation ID="IEq5"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="245_2025_10316_Article_IEq5.gif" Format="GIF" Height="17" Rendition="HTML" Resolution="72" Type="Linedraw" Width="101" /> </InlineMediaObject> <EquationSource Format="TEX">\(\tilde{a}\in \mathbb {R}_{+}, \tilde{a}&lt;a\)</EquationSource> </InlineEquation>. In the non-convex setting, we improve <InlineEquation ID="IEq6"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="245_2025_10316_Article_IEq6.gif" Format="GIF" Height="20" Rendition="HTML" Resolution="72" Type="Linedraw" Width="91" /> </InlineMediaObject> <EquationSource Format="TEX">\(\mathcal {O}(n+n\epsilon ^{-1})\)</EquationSource> </InlineEquation> of SVRG-SBB to <InlineEquation ID="IEq7"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="245_2025_10316_Article_IEq7.gif" Format="GIF" Height="21" Rendition="HTML" Resolution="72" Type="Linedraw" Width="99" /> </InlineMediaObject> <EquationSource Format="TEX">\(\mathcal {O}(n+n^{\beta }\epsilon ^{-1})\)</EquationSource> </InlineEquation>, where <InlineEquation ID="IEq8"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="245_2025_10316_Article_IEq8.gif" Format="GIF" Height="17" Rendition="HTML" Resolution="72" Type="Linedraw" Width="54" /> </InlineMediaObject> <EquationSource Format="TEX">\(\beta \in \mathbb {R}_{+}\)</EquationSource> </InlineEquation> can take any value in (2/3,&#xa0;1). Specifically, the constant step size regime is recovered by taking the TLF as a constant function, whose function value relies on problem-dependent parameters. As a counterpart of the constant step size regime, we also propose a BB-based vibration technique to set step sizes for VR methods, leading to methods with novel one-parameter step sizes. These methods have the same complexities compared to their constant step size versions. Meanwhile, they are more robust w.r.t. the sole step size parameter empirically. Moreover, a novel analysis is proposed for SARAH-I-type methods in the strongly convex setting. Numerical tests corroborate the proposed methods.</p>

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

On the Improvement of the Barzilai–Borwein Step Size in Variance Reduction Methods

  • Hai Liu,
  • Yan Liu,
  • Tiande Guo,
  • Congying Han

摘要

We propose several modifications of the Barzilai–Borwein (BB) step size in the variance reduction (VR) methods for finite-sum optimization problems. Our first approach relies on a scalar function, which we call the TaiL Function (TLF). The TLF maps the computed BB step size to some positive real number, which will be used as the step size instead. The computational overhead is almost negligible and the functional forms of TLFs in this work don’t involve any problem-dependent parameters. In the strongly convex setting, due to the undesirable appearance of the condition number \(\kappa \) in the linear convergence rate, the IFO complexity of VR methods with BB step size has the form \(\mathcal {O}((n+\kappa ^a)\kappa \log (1/\epsilon ))\) , \(a\in \mathbb {R}_{+}\) . With the utilization of the TLF, the aforementioned complexity is improved to \(\mathcal {O}((n+\kappa ^{\tilde{a}})\log (1/\epsilon ))\) , \(\tilde{a}\in \mathbb {R}_{+}, \tilde{a}<a\) . In the non-convex setting, we improve \(\mathcal {O}(n+n\epsilon ^{-1})\) of SVRG-SBB to \(\mathcal {O}(n+n^{\beta }\epsilon ^{-1})\) , where \(\beta \in \mathbb {R}_{+}\) can take any value in (2/3, 1). Specifically, the constant step size regime is recovered by taking the TLF as a constant function, whose function value relies on problem-dependent parameters. As a counterpart of the constant step size regime, we also propose a BB-based vibration technique to set step sizes for VR methods, leading to methods with novel one-parameter step sizes. These methods have the same complexities compared to their constant step size versions. Meanwhile, they are more robust w.r.t. the sole step size parameter empirically. Moreover, a novel analysis is proposed for SARAH-I-type methods in the strongly convex setting. Numerical tests corroborate the proposed methods.