Answers are grounded in this paper and cite their sources.
Let Y=σ(WX/√ d), where W is an m × d matrix with independent standard Gaussian entries, the N columns of X are independent and uniform on the sphere of radius √ d, and σ is a fixed polynomial with Eσ(Z)=E[Zσ(Z)]=0 for a standard Gaussian Z. Let m,d,N→∞ proportionally with m≠ N asymptotically, and let K_P = YPY^T/N, where P removes the mean of each row of Y over the N data points. We prove that the largest eigenvalue of K_P has Tracy–Widom fluctuations of order N^-2/3 around v(1+√m/N)², where v=Eσ(Z)². The same holds for the largest eigenvalue of YY^T/N when a_2^2N/d stays below v√N/m, where a_2 is the second Hermite coefficient of σ. The proof applies, conditionally on the data, the edge universality theorem of Chen and Lyu for sample covariance matrices whose sample vectors satisfy a bound on quadratic forms. For polynomial features this bound follows from the Gaussian Poincaré inequality and hypercontractivity, with constants independent of the dimension.
No takes yet. Share an insight, caveat, or question.
Yitzchak Shmalo (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: