Resampling Methods for Markov Processes with No Mixing Constraints
Date: Thu, November 30, 2017
Time: 3:00pm -
Location: Holmes Hall 389
Speaker: Kevin Oshiro, candidate for MS, advisor: Dr. Narayana Prasad Santhanam
Jackknife and bootstrap are resampling procedures that can be used to reduce the bias or estimate the variance of a statistic. These methods are useful because they perform well and are simple to implement, but an important assumption for their good performance is that of i.i.d. sampling. Previous analysis of these techniques for processes with memory generally require constraints on the memory or the mixing.
In this work, we adapt the jackknife and bootstrap procedures to estimate the variance of conditional probability estimates when we have unbounded memory and make no assumptions on the mixing of the process. We only require that the process satisfies a continuity condition, which says that the incremental value of a bit in the past diminishes with increasing distance. We then analyze the procedures to provide bounds on the bias of the estimates.