Multi-kernel spectral clustering: Entrywise eigenvector perturbation bounds and exact recovery

What the math gives to ML

The paper offers a transferable multi-scale similarity construction: instead of choosing one kernel bandwidth, it selects several bandwidths from empirical quantiles of pairwise squared distances. This is useful for neural representation learning because a single contrastive temperature or RBF scale can miss clusters with heterogeneous density and variance. Its row-wise spectral perturbation perspective also motivates preserving individual embedding locations, not merely the global span of a representation. The most practical adaptation is a quantile-calibrated multi-scale spectral regularizer for minibatch embeddings.