Finds the best block diagonal matrix approximation of a symmetric matrix. This can be exploited for divisive hierarchical clustering using singular vectors, named HC-SVD. The method is described in Bauer (202Xa) <doi:10.48550/arXiv.2308.06820>.