“…In particular, it is known that the spectral measure of any sequence of graphs converging locally to the d -regular tree converges to the Kesten-McKay law; see, for instance, [8]. Moreover, in [11], under an assumption on the number of small cycles (corresponding approximately to a locally treelike structure and satisfied with high probability by random regular graphs), it is proved that eigenvectors cannot be localized in the following sense: if for some`2-normalized eigenvector v D .v i / N i D1 a set B 1; N satisfies P i 2B jv i j 2 > " > 0, then jBj > N ı with high probability for some small ı / " 2 . In comparision, for a random d -regular graph with d > .log N / 4 , Corollary 1.2 implies that if a set B has`2-mass " > 0, then jBj > "N.log N / 4 with high probability, which is optimal up to the power of the logarithmic correction.…”