Electromagnetic-Twin Channel Estimation: From Sparse Pilots to Persistent CSI

Abstract

Each repeated pilot block is typically treated as a separate channel-estimation problem, although much of the angle--delay state persists across updates. This paper defines an electromagnetic twin (ET) as a recursively synchronized signal-processing state containing channel state information (CSI), coefficientwise uncertainty, support confidence, and age. A channel knowledge map may initialize this state; subsequent pilots predict, correct, validate, and store the state. We formulate an anchored-window inverse problem that yields online uncertainty-weighted correction and fixed-lag smoothing as two convex operating modes of a persistent-state estimator. Uncertainty-aware ET FISTA (UET-FISTA) gives a zero-delay update, while an alternating direction method of multipliers (ET-ADMM) jointly refines a short window when delayed CSI is acceptable. Held-out pilots choose update, complete refresh, or hold under a prescribed family-wise false-acceptance probability, preventing an inaccurate stored state from being reused blindly. Conditional Bayesian Cramér--Rao bounds quantify the information supplied by the stored state and by future pilot blocks. With 20\% pilots, UET-FISTA reaches $-11.55$ dB normalized mean-square error, compared with $-3.42$ dB for static FISTA. Over 200 independent 50-epoch trajectories, three-epoch smoothing improves matched-state NMSE from $-10.28$ to $-13.41$ dB. On 40 off-grid QuaDRiGa trajectories, online UET-FISTA improves static recovery from $-1.37$ to $-1.88$ dB and effective rate from 1.688 to 1.919 bit/s/Hz. These results identify when persistent CSI reduces pilot demand and when the receiver should instead refresh or preserve its state.