Two-Layer Linear Auto-Regressive Models Estimate Latent States — reproduction

arXiv:2606.12691v1 · OpenReview bMSnvqVWaB

ClaimVerdictHeadline
1 hidden repr = Kalman state up to SreproducedR² 0.9999 (H=5); shuffle control −0.25
2 Theorem 4 recovery boundreproducedslope −0.919 vs −1; h=2 control 91× worse
3 Theorem 2 in-sample O(H/T)reproducedslopes −0.919 / −0.962
4 Theorem 3 parameter errorreproducedslope −1.031; excitation control → 0
5 Proposition 1 truncation in Lreproducedidentity exact to 2.9e−13; base = ρ(Ā)
6 grid search picks h=4falsifiednested ranks ⇒ argmin h=10 in 10/10