CORE-CSI: Minimal-Pilot CSI Prediction via Complex One-Step Residual Evolution for Mobile QPSK Links

Taipe C.V.C.; Luna J.C.; Tamo A. · 2026 IEEE Colombian Conference on Communications and Computing, COLCOM 2026 - Conference Proceedings (2026) · DOI: 10.1109/colcom71387.2026.11703492

Resumen

Channel aging degrades coherent demodulation in mobile links and makes frequent pilot-based CSI updates costly. This paper proposes CORE-CSI, a lightweight, task-specific, world-model-inspired predictor for one-step complex CSI evolution. Rather than directly regressing the next channel state, CORE-CSI predicts a sample-adaptive complex residual ratio from recent CSI history and reconstructs H[t+1]=H[t]⊙ R[t+1]. The residual uses log-amplitude and circular sine/cosine phase components to preserve multiplicative channel structure and avoid phase-wrap discontinuities; user motion and Sionna RT geometry proxies are evaluated as candidate side information through ablation. We evaluate uncoded OFDM/QPSK with onetap equalization under LOS and NLOS fixed-path Doppler conditions initialized in the same Munich scene. Condition-specific models use five training seeds and chronological non-overlapping splits. At 10 dB, CORE-CSI reduces stale-CSI BER from 6.734 × 10-2 to (4.940 ± 0.034) × 10-3 in LOS and from 4.718 × 10-1 to (2.618 ± 0.079) × 10-3 in NLOS, corresponding to relative reductions of 92.66% and 99.45%. At 20 dB, no errors are observed over 147,456 bits per seed; the Wilson upper bound is 2.61 × 10-5. The statistically competitive AR(1) baseline indicates that the evaluated fixed-path Doppler regime contains strong first-order temporal structure, while ablation does not resolve incremental gains from motion or geometry. Condition-wise unit-energy normalization isolates CSI dynamics and excludes the NLOS path-loss penalty. CORE-CSI requires no additional pilot subcarriers at the target snapshot, whereas the evaluated CPE configuration reserves 25%; this comparison characterizes a zeroadditional-target-pilot operating point rather than a neural-only throughput gain.

Instituciones: Universidad Nacional de San Agustín de Arequipa