Pairwise network measures (e.g. correlations) are represented in matrices.
However, classical statistical tests typically compare distributions, ignoring aspects of their structure.
So, we construct the statistical "eigenangle test" to evaluate structural matrix similarity.
2/5

Imagine the correlation matrices for two sets of N spike trains.
When they are similar, their respective eigenvectors (ev) are more aligned, i.e, the angles between ev pairs are small.
Thus we can quantify similarity by angle-smallness, compared to random angles in N-dim space 📐
3/5

We demonstrate how the similarity score correctly indicates correlated subgroup structures between network realizations.
The test complements classical tests (e.g. Kolmogorov-Smirnov) because it compares the organization of correlations in a network instead of just the amount.
4/5