异步系统与二元对角随机矩阵:一个证明及收敛速度
Asynchronous Systems and Binary Diagonal Random Matrices: A Proof and Convergence Rate
摘要:
在$n$节点的同步网络中,每个节点将基于给定迭代中的系统状态来更新其参数。众所周知,如果$n \times n$迭代矩阵${\bf{F}}$的最大绝对特征值(谱半径)小于1(即$ρ({\bf{F}})<1$),则更新可以收敛到不动点。然而,如果在迭代中仅节点的子集更新它们的参数(由于延迟或陈旧反馈),则这有效地使迭代矩阵的谱半径为1。我们考虑矩阵的单位谱半径产生的${\bf{F}}$由于随机延迟的更新。我们表明,如果每个节点更新至少一次,在每$T$迭代,那么这些迭代对应的随机矩阵(联合谱半径)的产品是小于一。然后,我们使用这个属性来证明异步迭代系统的收敛性。最后,我们证明了这样一个系统的收敛速度是$ρ({\bf{F}})\frac{(1-(1-γ)^T)^n}{T}$,假设遍历,$γ$是一个节点将在任何给定的迭代更新的概率的最低限度。
英文摘要:
In a synchronized network of $n$ nodes, each node will update its parameter based on the system state in a given iteration. It is well-known that the updates can converge to a fixed point if the maximum absolute eigenvalue (spectral radius) of the $n \times n$ iterative matrix ${\bf{F}}$ is less than one (i.e. $ρ({\bf{F}})<1$). However, if only a subset of the nodes update their parameter in an iteration (due to delays or stale feedback) then this effectively renders the spectral radius of the iterative matrix as one. We consider matrices of unit spectral radii generated from ${\bf{F}}$ due to random delays in the updates. We show that if each node updates at least once in every $T$ iterations, then the product of the random matrices (joint spectral radius) corresponding to these iterations is less than one. We then use this property to prove convergence of asynchronous iterative systems. Finally, we show that the convergence rate of such a system is $ρ({\bf{F}})\frac{(1-(1-γ)^T)^n}{T}$, where assuming ergodicity, $γ$ is the lowest bound on the probability that a node will update in any given iteration.
机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。