Asymptotic approximation of population covariances by linearized recursions

About 2 years old · traced to

Let Ω\Omega, Φ\Phi, and Ψ\Psi be the population covariances appearing in Theorem, and let ΩLlin\Omega_L^\mathrm{lin}, ΦL~lin\Phi_{\tilde{L}}^\mathrm{lin}, and ΨL~lin\Psi_{\tilde{L}}^\mathrm{lin} be the final iterates of the linear recursions from Definition. The Frobenius-norm errors satisfy

\normΩ−ΩLlinF+\normΨ−ΨL~linF+\normΦ−ΦL~linF≲1.\norm{\Omega-\Omega_L^\mathrm{lin}}_F + \norm{\Psi-\Psi_{\tilde{L}}^\mathrm{lin}}_F + \norm{\Phi-\Phi_{\tilde{L}}^\mathrm{lin}}_F \lesssim 1.

Asymptotic covariance approximation conjecture. The population covariances can be asymptotically approximated by the last iterates of the corresponding linear recursions, with the displayed combined error bound. The source provides no resolution or further context establishing whether this claim is proved or remains open.

References

Primary source

Dominik Schröder, Daniil Dmitriev, Hugo Cui and Bruno Loureiro, “Asymptotics of Learning with Deep Structured (Random) Features”, arXiv:2402.13999 (2024).

Progress summary

Never refreshed

Nothing recorded yet. Refresh searches the literature and the public web for attempts on this problem, and writes the first summary here.

Solutions 0

No solutions have been posted yet.