| [ | |
| "Theorem 1 (informal) shows the adaptive sensing algorithm reaches constant-level alignment with the true eigenvector after O(λ₁λ₂d²/Δ²) iterations, after which the sine-squared alignment error decays as O(λ₁λ₂d²/(Δ²t)) (Theorem 1).", | |
| "Theorem 2 (formal) specifies a warmup phase of t₀ = (4S+1)log(d/2) iterations after which the expected squared sine alignment satisfies E[1-(ūᵀu_{t₀})²] ≤ 0.5, followed by a distinct local convergence phase (Theorem 2).", | |
| "The paper's rate matches the minimax lower bound Ω(λ₁λ₂/Δ² · d/t) from Li et al. (2018) up to an extra factor of d, which is attributed to the cost of compressive (two-measurement) sampling (Section 3, Theorem 2).", | |
| "Section 5.1 ('Tracking a Moving Eigenvector') derives a closed-form optimal step size η̂ = √(V/S) and fixed point x* = V + √(VS) for the non-stationary tracking setting (Section 5.1).", | |
| "Figure 1 empirically validates the theoretical convergence rate of Algorithm 1 using d=10, Δ=1 across 20 trials, reporting 20th/80th percentile error bars (Figure 1)." | |
| ] | |