Peak Inference for Gaussian Random Fields on a Lattice

TL;DR

This paper introduces a Monte Carlo method for accurately estimating the distribution of local maxima in Gaussian random fields on lattices, especially effective for datasets with low smoothness, and demonstrates its practical application in fMRI analysis.

Contribution

The paper presents a novel Monte Carlo approach for peak inference in Gaussian fields on lattices, extending existing methods and improving accuracy for low-smoothness data.

Findings

The method provides valid peak-based inference in low-smoothness datasets.

It outperforms existing formulas derived for continuous domains.

Application to fMRI data demonstrates practical utility.

Abstract

In this work we develop a Monte Carlo method to compute the height distribution of local maxima of a stationary Gaussian or Gaussian-related random field that is observed on a regular lattice. We show that our method can be used to provide valid peak based inference in datasets with low levels of smoothness, where existing formulae derived for continuous domains are not accurate. We also extend the methods in Worsley (2005) and Taylor et al. (2007) to compute the peak height distribution and compare them with our approach. Lastly, we apply our method to a task fMRI dataset to show how it can be used in practice.