Gaussian Equivalence for Self-Attention: Asymptotic Spectral Analysis of Attention Matrix
This work addresses a foundational problem in machine learning theory for researchers and practitioners by offering new insights into the behavior of attention mechanisms, though it is incremental in advancing theoretical analysis.
The authors tackled the limited theoretical understanding of self-attention layers by providing a rigorous asymptotic analysis of the singular value spectrum of the attention matrix, establishing the first Gaussian equivalence result and showing that the distribution deviates from the previously believed Marchenko-Pastur law.
Self-attention layers have become fundamental building blocks of modern deep neural networks, yet their theoretical understanding remains limited, particularly from the perspective of random matrix theory. In this work, we provide a rigorous analysis of the singular value spectrum of the attention matrix and establish the first Gaussian equivalence result for attention. In a natural regime where the inverse temperature remains of constant order, we show that the singular value distribution of the attention matrix is asymptotically characterized by a tractable linear model. We further demonstrate that the distribution of squared singular values deviates from the Marchenko-Pastur law, which has been believed in previous work. Our proof relies on two key ingredients: precise control of fluctuations in the normalization term and a refined linearization that leverages favorable Taylor expansions of the exponential. This analysis also identifies a threshold for linearization and elucidates why attention, despite not being an entrywise operation, admits a rigorous Gaussian equivalence in this regime.