We introduce a minimum-spanning-tree-based regularization for self-supervised learning that provably improves representation uniformity and mitigates dimensional collapse.