We study the denoising of low-rank matrices by singular value shrinkage. Recent work of Gavish and Donoho constructs a framework for finding optimal singular value shrinkers for a wide class of loss functions. We use this framework to derive the optimal shrinker for operator norm loss. The optimal shrinker matches the shrinker proposed by Gavish and Donoho in the special case of square matrices, but differs for all other aspect ratios. We precisely quantify the gain in accuracy from using the optimal shrinker. We also show that the optimal shrinker converges to the best linear predictor in the classical regime of aspect ratio zero.
View on arXiv