The synergistic interactions of low-dimensional brain modes
Neuroimaging techniques produce vast amounts of data, capturing brain activity in a high-dimensional space. However, brain dynamics are consistently shown to reside in a rather lower-dimensional space, which contains relevant information for cognition and behavior. This dimensionality reduction reflects distinct types of interactions between brain regions, such as redundancy -shared neural information distributed across regions- and synergy, where information emerges only when regions are considered collectively. Significant efforts have been devoted to developing linear and non-linear algorithms to reveal these low-dimensional dynamics, often termed "brain modes." Here, we apply various dimensionality reduction techniques to resting-state functional magnetic resonance imaging (fMRI) data from 100 healthy participants to examine how synergistic interactions in brain dynamics are preserved by these techniques. We first demonstrate that biologically informed brain parcellation modulates and preserves synergy-dominated interactions. Next, we show that synergy among low-dimensional modes enhances functional-connectivity reconstruction: nonlinear autoencoders not only achieve the lowest reconstruction error but also maximally preserve synergy, outperforming principal component analysis, diffusion maps, and Laplacian eigenmodes. Finally, we confirm previous results suggesting that global signal regression helps to identify synergistic interactions between regions. Our findings establish synergy preservation as a complementary criterion to reconstruction accuracy, highlighting autoencoders as a nonlinear tool for uncovering synergistic low-dimensional brain modes from high-dimensional neuroimaging data.