THE CEREBELLAR VARIANCE PARADOX: ACHIEVING LEARNING STABILITY THROUGH ALEATORIC VARIANCE MINIMIZATION

The cerebellum utilizes a dual-system architecture for motor learning, employing a complex, high-capacity pathway through the cerebellar cortex and a simpler, direct pathway to the deep cerebellar nuclei (DCN). This study addresses the fundamental question of how the brain adjudicates between these systems using a normative framework based on a trade-off between bias, variance, and metabolic overhead. Empirical observations indicate that the cortex is preferred during the initial, data-limited phase of learning. However, our derivation reveals the "Cerebellar Variance Paradox": for the cortex to overcome its high metabolic overhead and be preferred early on, it must possess a lower intrinsic variance coefficient than the simpler DCN pathway. This contradicts the intuition that greater architectural complexity introduces more physical noise. To resolve this, we deconstruct variance into reducible epistemic uncertainty and irreducible aleatoric uncertainty, the latter comprising environmental and path-specific physical noise. We formalize the "Amplified Aleatoric Variance" (AAV) criterion, demonstrating that epistemic uncertainty is amplified by total aleatoric variance. Our biophysical modeling suggests that the cortex satisfies this criterion through two key strategies: ensemble averaging at the Purkinje cell-DCN synapse to cancel upstream noise, and sparse coding in the granule cell layer to reduce both multiplicative noise and effective system complexity. This framework explains how complex cerebellar circuits are exquisitely adapted for stable and efficient learning.