Conjecture on Grammians of independent symmetric rank-one tensors

Let a1,,aKa_1,\ldots,a_K be independent copies of the random vector aUnif(SD1)a \sim \operatorname{Unif}(\mathbb{S}^{D-1}), and let GnG_n be the Grammian associated with these vectors, with entries (Gn)ij=ai,ajn(G_n)_{ij}=\langle a_i,a_j\rangle^n. Fix ϵ>0\epsilon>0 arbitrarily.

Grammians conjecture. There exists a constant κn>0\kappa_n>0 and an increasing function γn:(0,κn)(0,1)\gamma_n:(0,\kappa_n)\to(0,1), both depending only on nn, such that γn(κ)1\gamma_n(\kappa)\to1 as κ0\kappa\to0, and whenever KκDnK\leq\kappa D^n for some κ<κn\kappa<\kappa_n,

P(Gn12γn(κ)ϵ)1asD.\mathbb{P}\left(\left\|G_n^{-1}\right\|_2\geq\gamma_n(\kappa)-\epsilon\right)\to1\quad\textrm{as}\quad D\to\infty.

In particular, if K=o(Dn)K=o(D^n), then Gn121\|G_n^{-1}\|_2\to1 as DD\to\infty. The conjecture concerns the well-posedness of inverting the Grammian for independent random symmetric rank-one tensors; the paper gives numerical evidence and notes that the stated proof is currently available only for tensor order n=2n=2, while the extension to n>2n>2 remains open.

Sources & referencesView supporting material

Primary source

Joe Kileel, Timo Klock and João M. Pereira, “Landscape analysis of an improved power method for tensor decomposition”, arXiv:2110.15821 (2021).

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.