[Paper Review] Rapid Acceleration of the Permutation Test via Slow Random Walks in the Permutation Group

Moo K. Chung, Yixian Wang+2 more|arXiv (Cornell University)|Dec 17, 2018

Bayesian Methods and Mixture Models4 citations

TL;DR

This paper proposes a novel slow random walk strategy on the permutation group to accelerate exact permutation tests in neuroimaging without approximation. By exploiting the algebraic structure of symmetric groups, the method achieves full permutation testing in under an hour on a laptop—over 100× faster than standard methods—enabling large-scale studies on HCP and ADNI datasets with 274 females and 182 males, and twin correlation analysis with 138 MZ and 79 DZ pairs.

ABSTRACT

The permutation test is an often used test procedure in brain imaging. Unfortunately, generating every possible permutation for large-scale brain image datasets such as HCP and ADNI with hundreds images is not practical. Many previous attempts at speeding up the permutation test rely on various approximation strategies such as estimating the tail distribution with known parametric distributions. In this study, we show how to rapidly accelerate the permutation test without any type of approximate strategies by exploiting the underlying algebraic structure of the permutation group. The method is applied to large number of MRIs in two applications: (1) localizing the male and female differences and (2) localizing the regions of high genetic heritability in the sulcal and gyral pattern of the human cortical brain.

Motivation & Objective

• To overcome the computational bottleneck of exact permutation tests in large-scale neuroimaging studies with hundreds of subjects.
• To eliminate reliance on approximate methods such as parametric tail estimation or limited resampling.
• To enable exact inference in large-sample studies (e.g., HCP, ADNI) where full permutation enumeration is infeasible.
• To provide a theoretically grounded, non-approximate framework for permutation testing using group-theoretic random walks.
• To apply the method to detect sex differences and genetic heritability in cortical sulcal and gyral patterns.

Proposed method

• The method uses a slow random walk over the symmetric group $\mathbb{S}_{m+n}$ to explore permutations systematically, avoiding the need to enumerate all $ (m+n)! $ permutations.
• It leverages the algebraic structure of the permutation group to ensure coverage of the space with high probability using far fewer samples than full enumeration.
• The random walk is constructed via composition of transpositions and cycles, ensuring ergodicity and convergence to the uniform distribution over permutations.
• Test statistics (e.g., t-statistic, correlation) are computed on each permuted sample, and p-values are estimated as the fraction of permutations yielding more extreme values.
• The approach is applied to diffusion maps on cortical surfaces, where heat diffusion measures local geometric similarity between sulcal and gyral patterns.
• For twin correlation analysis, the method computes average correlations over 10,000 random walks to estimate heritability indices without approximation.

Experimental results

Research questions

• RQ1 Can exact permutation testing be accelerated significantly without approximation in large neuroimaging datasets?
• RQ2 How can the algebraic structure of the symmetric group be exploited to generate representative permutations efficiently?
• RQ3 Can the method enable exact inference in large-scale studies (e.g., 274 females, 182 males) where full permutation is computationally infeasible?
• RQ4 How accurately can twin correlation and heritability be estimated using this non-approximate random walk framework?
• RQ5 Can the method be applied to complex, non-monotonic test statistics such as those derived from diffusion maps on cortical surfaces?

Key findings

• The method computed 500,000 permutations in 40 minutes on a laptop—equivalent to over 18 days using standard methods—achieving a 100× speedup.
• The approach enabled the first known exact permutation test on a dataset of 274 females and 182 males, localizing significant sex differences in the temporal lobe (p < 0.05, corrected).
• The method computed twin correlation maps over 10,000 random walks in 97 seconds, achieving convergence within 3 decimal places.
• Heritability index (HI) maps revealed that 70% of highly heritable regions (HI > 0.7) were localized in central gyral and sulcal areas.
• The maximum t-statistic for sex differences reached 7.08, and the minimum was -6.44, with corrected p-values below 0.05 at thresholds of ±4.27 and 4.48.
• The method demonstrated that slow random walks on the permutation group can achieve full statistical inference without approximation, even in high-dimensional brain imaging data.