Compression Is Kinder to the Tail Than We Thought: Quantization Is Neutral, and Pruning's Tail Damage Is Recoverable

Post-training compression is a prerequisite for deploying time-series models under tight latency and memory budgets, and a well-known line of work reports that it disproportionately harms rare classes. This paper revisits that concern under standard deployment practice, on two long-tailed time-series datasets: RF modulation classification (RadioML 2018.01A) and wearable activity recognition (PAMAP2). Three pretraining conditions (class-reweighted supervised contrastive learning, unweighted supervised contrastive learning, and inverse-frequency balanced sampling) are crossed with four compression families over five seeds, with accuracy resolved per class-frequency bin and paired tests corrected for multiplicity. Three findings emerge, all of which soften rather than confirm the prevailing picture. First, INT8 quantization is tail-neutral on both domains and in every condition, and with per-channel scaling this neutrality extends to 3-bit weights; the catastrophic low-bit collapse reported under per-tensor quantization is substantially a quantizer artifact, and where it does occur it is not tail-specific, since head accuracy falls as far as tail accuracy. Second, the severe tail damage from structured pruning (up to 0.71 absolute on the tail bin) is almost entirely an artifact of one-shot pruning without recovery: ten epochs of fine-tuning with the sparsity mask held fixed reduce residual tail degradation to within 0.06 at 70% sparsity on both datasets, and to approximately zero at 30%. Third, we find no reliable interaction between pretraining-time rebalancing and compression: no condition contrast survives Holm correction across 36 paired tests, and an embedding-margin mechanism is rejected within both datasets once a pooling artifact is removed. Practically, static INT8 quantization is faster and roughly half the size of FP32 while leaving the tail intact, and pruning need not be avoided on fairness grounds provided a recovery pass is budgeted.