Scholarlink

<p>In this paper, we develop a unified theory of IPA estimators for steady-state performance characteristics. Our goal is to clarify how infinitesimal perturbation analysis can be systematically connected to a measure-valued differentiation framework for the underlying Markov chain dynamics. Traditionally, the asymptotic behavior of gradient estimators has been studied by analyzing the estimators themselves. Here, we take a different perspective. We show that by clearly separating the level at which arguments are made, either at the sample-path level, where IPA is usually formulated, or at the operator/measure level, where Markov chain differentiation is naturally expressed, one can combine the strengths of both approaches. Our framework helps bridge two established streams of research and contributes a more general understanding of IPA for steady-state problems.</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

IPA for stationary problems

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月27日

查看全文
  • Bernd Heidergott

摘要

In this paper, we develop a unified theory of IPA estimators for steady-state performance characteristics. Our goal is to clarify how infinitesimal perturbation analysis can be systematically connected to a measure-valued differentiation framework for the underlying Markov chain dynamics. Traditionally, the asymptotic behavior of gradient estimators has been studied by analyzing the estimators themselves. Here, we take a different perspective. We show that by clearly separating the level at which arguments are made, either at the sample-path level, where IPA is usually formulated, or at the operator/measure level, where Markov chain differentiation is naturally expressed, one can combine the strengths of both approaches. Our framework helps bridge two established streams of research and contributes a more general understanding of IPA for steady-state problems.