QuaCE-Guard: Joint user-budget conformal certification for sequential pilot acquisition in MIMO-OFDM

This is the authors' abstract. We don't add key points for this paper.

Fixed pilot budgets impose identical channel-sounding cost on realizations with unequal estimation difficulty, while an additional statistical selection is introduced when the smallest apparently safe budget is chosen online. A leakage-free stop-or-escalate controller is presented in which an entire nested pilot-budget family is certified by one maximum-standardized conformal score before a stopping budget is selected. A user-by-budget certificate is then formulated by treating the complete user-by-budget tensor as the exchangeability unit, so simultaneous protection is retained without requiring independence among co-scheduled users. Across 810 repeated-split single-user outcomes, selected-budget coverage of 0.968 is obtained at a 0.063 final pilot fraction; predeclared noninferiority and pilot-saving criteria are passed relative to fixed 10% acquisition. Mean incremental protocol cost is 0.0615, corresponding to a 38.5% reduction, whereas full-pattern retransmission is separately measured at 0.1249. After estimator-specific source-side model refitting and source-side conformal recalibration, coverage of 0.9605 is retained with a learned estimator. In 2/4/8-user tests, empirical simultaneous-system coverage of 0.949–0.967 is obtained with orthogonal pilots and 0.909–0.962 under coherent shared-pilot contamination. Across ten simulated radio-frequency impairment arms, at least 0.953 coverage is obtained after impairment-aware source-side model adaptation and source-side conformal recalibration. A 128-receive-antenna spherical-wave evaluation yields 0.973 coverage versus 0.975 in the far field, while identical minimum-budget stopping identifies estimator fidelity as the dominant determinant of near-field acquisition sensitivity. Finally, measured DICHASUS testing is used to separate zero-shot transport from target-domain calibration: at 20 dB, frozen-source fine-bound coverage is 0.0005 on average across repeated evaluation cells, whereas measured q -only block recalibration raises mean per-cell coverage to 0.938 without refitting the estimator or point/scale predictors.