We offer a theoretical explanation for the sparse learning observed in our recent work: on-policy sampling leads to small diagonal Fisher and small expected gradients, connecting sparsity across RL and on-policy distillation.