A Counterexample to the Tang Zhang Schatten Norm Conjecture and Sharp Positive Results

What the math gives to ML

The paper studies a dimension-free inequality for aggregating matrix summands, comparing the Schatten norm of a signed sum with the Schatten norm of the sum of positive operators |A_k|. The transferable asset is a principled budget for merging rank-one or low-rank parameter updates while accounting for noncommutative alignment and cancellation, rather than clipping each update independently. This suggests a stability mechanism for LoRA, adapter merging, and rank-one optimizer updates. The counterexample at p=3/2 is also practically useful: the closed-form constant should be used only in the proven p>=2 regime, with lower-p settings treated as empirical heuristics.