当前位置:
X-MOL 学术
›
arXiv.cs.DS
›
论文详情
Our official English website, www.x-mol.net, welcomes your
feedback! (Note: you will need to create a separate account there.)
Making mean-estimation more efficient using an MCMC trace variance approach: DynaMITE
arXiv - CS - Data Structures and Algorithms Pub Date : 2020-11-22 , DOI: arxiv-2011.11129 Cyrus Cousins, Shahrzad Haddadan, Eli Upfal
arXiv - CS - Data Structures and Algorithms Pub Date : 2020-11-22 , DOI: arxiv-2011.11129 Cyrus Cousins, Shahrzad Haddadan, Eli Upfal
The Markov-Chain Monte-Carlo (MCMC) method has been used widely in the
literature for various applications, in particular estimating the expectation
$\mathbb{E}_{\pi}[f]$ of a function $f:\Omega\to [a,b]$ over a distribution
$\pi$ on $\Omega$ (a.k.a. mean-estimation), to within $\varepsilon$ additive
error (w.h.p.). Letting $R \doteq b-a$, standard variance-agnostic MCMC
mean-estimators run the chain for $\tilde{\cal
O}(\frac{TR^{2}}{\varepsilon^{2}})$ steps, when given as input an (often loose)
upper-bound $T$ on the relaxation time $\tau_{\rm rel}$. When an upper-bound
$V$ on the stationary variance $v_{\pi} \doteq \mathbb{V}_{\pi}[f]$ is known,
$\tilde{\cal O}\bigl(\frac{TR}{\varepsilon}+\frac{TV}{\varepsilon^{2}}\bigr)$
steps suffice. We introduce the DYNAmic {Mcmc} Inter-Trace variance Estimation
(DynaMITE) algorithm for mean-estimation. We define the inter-trace variance
$v_{T}$ for any trace length $T$, and show that w.h.p., DynaMITE estimates the
mean within $\varepsilon$ additive error within $\tilde{\cal
O}\bigl(\frac{TR}{\varepsilon} + \frac{\tau_{\rm rel} v_{\tau\rm
rel}}{\varepsilon^{2}}\bigr)$ steps, without {a priori} bounds on $v_{\pi}$,
the variance of $f$, or the trace variance $v_{T}$. When $\epsilon$ is small, the dominating term is $\tau_{\rm rel} v_{\tau\rm
rel}$, thus the complexity of DynaMITE principally depends on the {\it a priori
unknown} $\tau_{\rm rel}$ and $v_{\tau\rm rel}$. We believe in many situations
$v_{T}=o(v_{\pi})$, and we identify two cases to demonstrate it. Furthermore,
it always holds that $v_{\tau\rm rel} \leq 2v_{\pi}$, thus the worst-case
complexity of DynaMITE is $\tilde{\cal O}(\frac{TR}{\varepsilon}
+\frac{\tau_{\rm rel} v_{\pi}}{\varepsilon^{2}})$, improving the dependence of
classical methods on the loose bounds $T$ and $V$.
中文翻译:
使用MCMC跟踪方差方法使均值估计更有效:DynaMITE
马尔可夫链蒙特卡罗(MCMC)方法已在文献中广泛用于各种应用,尤其是估计函数$ f:\ Omega的期望$ \ mathbb {E} _ {\ pi} [f] $到$ \ Omega $上的分布$ \ pi $上的[a,b] $(也就是均值估计),到$ \ varepsilon $附加误差(whp)以内。让$ R \ doteq ba $,与标准方差无关的MCMC均值估算器运行$ \ tilde {\ cal O}(\ frac {TR ^ {2}} {\ varepsilon ^ {2}})$步骤的链,当输入为松弛时间$ \ tau _ {\ rm rel} $的(通常是宽松的)上限$ T $时。当固定方差$ v _ {\ pi} \ doteq \ mathbb {V} _ {\ pi} [f] $的上限$ V $为已知时,$ \ tilde {\ cal O} \ bigl(\ frac {TR} {\ varepsilon} + \ frac {TV} {\ varepsilon ^ {2}} \ bigr)$个步骤就足够了。我们引入DYNAmic {Mcmc}迹间方差估计(DynaMITE)算法进行均值估计。我们定义任何迹线长度$ T $的迹线间方差$ v_ {T} $,并证明whp,DynaMITE估算$ \ tilde {\ cal O} \ bigl(\ frac {TR} {\ varepsilon} + \ frac {\ tau _ {\ rm rel} v _ {\ tau \ rm rel}} {\ varepsilon ^ {2}} \ bigr)$个步骤,没有对{ v _ {\ pi} $,方差$ f $或跟踪方差$ v_ {T} $。当$ \ epsilon $小时,主导项为$ \ tau _ {\ rm rel} v _ {\ tau \ rm rel} $,因此DynaMITE的复杂度主要取决于{\ it先验未知} $ \ tau_ { \ rm rel} $和$ v _ {\ tau \ rm rel} $。我们相信在许多情况下$ v_ {T} = o(v _ {\ pi})$,我们确定了两种情况来证明这一点。此外,它始终认为$ v _ {\ tau \ rm rel} \ leq 2v _ {\ pi} $,
更新日期:2020-11-25
中文翻译:
使用MCMC跟踪方差方法使均值估计更有效:DynaMITE
马尔可夫链蒙特卡罗(MCMC)方法已在文献中广泛用于各种应用,尤其是估计函数$ f:\ Omega的期望$ \ mathbb {E} _ {\ pi} [f] $到$ \ Omega $上的分布$ \ pi $上的[a,b] $(也就是均值估计),到$ \ varepsilon $附加误差(whp)以内。让$ R \ doteq ba $,与标准方差无关的MCMC均值估算器运行$ \ tilde {\ cal O}(\ frac {TR ^ {2}} {\ varepsilon ^ {2}})$步骤的链,当输入为松弛时间$ \ tau _ {\ rm rel} $的(通常是宽松的)上限$ T $时。当固定方差$ v _ {\ pi} \ doteq \ mathbb {V} _ {\ pi} [f] $的上限$ V $为已知时,$ \ tilde {\ cal O} \ bigl(\ frac {TR} {\ varepsilon} + \ frac {TV} {\ varepsilon ^ {2}} \ bigr)$个步骤就足够了。我们引入DYNAmic {Mcmc}迹间方差估计(DynaMITE)算法进行均值估计。我们定义任何迹线长度$ T $的迹线间方差$ v_ {T} $,并证明whp,DynaMITE估算$ \ tilde {\ cal O} \ bigl(\ frac {TR} {\ varepsilon} + \ frac {\ tau _ {\ rm rel} v _ {\ tau \ rm rel}} {\ varepsilon ^ {2}} \ bigr)$个步骤,没有对{ v _ {\ pi} $,方差$ f $或跟踪方差$ v_ {T} $。当$ \ epsilon $小时,主导项为$ \ tau _ {\ rm rel} v _ {\ tau \ rm rel} $,因此DynaMITE的复杂度主要取决于{\ it先验未知} $ \ tau_ { \ rm rel} $和$ v _ {\ tau \ rm rel} $。我们相信在许多情况下$ v_ {T} = o(v _ {\ pi})$,我们确定了两种情况来证明这一点。此外,它始终认为$ v _ {\ tau \ rm rel} \ leq 2v _ {\ pi} $,