The Asymmetrical Architecture of Ultradian REM Sleep: Direction, NREM Currency, and a Finite Post-REM Predictive Horizon in Mice

Study Objectives: Competing accounts of ultradian REM sleep place the controlling variable either before or after the REM episode. We reanalysed 24-hour recordings from 25 mice to test directionality, the structure of post-REM suppression, the currency of the interval, and the macro-organization of REM output. Methods: REM episodes and inter-REM intervals were reconstructed from 15-s vigilance scoring, yielding 1,955 valid intervals. Analyses used mixed-effects models, lower-quantile regression with leave-one-mouse-out validation and within-mouse permutation, discrete-time hazard models, and held-out-mouse comparison of accumulated NREM against elapsed time. Results: Accumulated NREM did not predict subsequent REM duration (rho=0.015, p=0.50), whereas preceding REM duration (REMpre) predicted subsequent NREM accumulation (rho=0.506), positively in all 25 animals; the log-log elasticity was 0.56 +- 0.03. A convex REMpre-dependent lower envelope, approximated by 1 min NREM per min REMpre, survived leave-one-mouse-out validation (6.1% held-out violations against 10.2% under permutation, p{approx}0.0001) but was not an absolute floor. Suppression relaxed over several minutes, and the REMpre effect became undetectable beyond approximately 10 min of accumulated NREM. An accumulated-NREM hazard model outperformed an equivalent elapsed-time model in held-out mice. Episode-termination hazard rose with episode duration (Weibull shape 1.48), and inter-animal variability in episode number exceeded that in mean duration about fourfold. Conclusions: Ultradian REM organization is strongly asymmetrical. Preceding REM imposes an early constraint expressed in accumulated NREM, neither equivalent to a strict refractory floor nor sustained across the whole interval: a graded refractory regime with a lower envelope, a several-minute recovery, and a finite predictive horizon near 10 min.