Paperlane

no. 17 · Wednesday 18 November 2026

Smoothing Bias and Direction Recovery at Boundary Minima of the Nuclear Norm along Matrix Lines

Haocheng Wang, Yuanqi Zhang and Yi Xu

math.NA · 43 pp

Abstract

We study optimal matrix directions associated with minimizing $\|G+λΘ\|_*$, the nuclear norm along a fixed affine line of real matrices with scalar multiplier $λ$. At a rank-deficient minimum, the compact polar factor can violate the tangency constraint even at the minimizing multiplier. Spectral smoothing gives a continuous scalar equation for tangency. We quantify its bias at a unique boundary minimum, where exactly one one-sided derivative vanishes. An orthogonal block decomposition shows that pseudo-Huber smoothing with scale $δ$ shifts the multiplier by exact order $δ^{2/3}$ as $δ\to0$. The Frobenius error relative to the minimum-Frobenius-norm optimal direction has exact order $δ^{1/3}$ when some singular values vanish at exactly second order in the multiplier displacement, and $δ^{2/3}$ otherwise. A sharp rank bound rules out the slower rate when $Θ$ has rank at most two. Huber smoothing gives an error of exact order $δ$ for the same limiting direction. Convergent expansions in powers of $δ^{1/3}$ also yield a three-scale extrapolation with $O(δ)$ error after restoring feasibility.

comments
43 pages, 1 figure
fields
Numerical Analysis
pdf ↗ page arXiv:2611.22963 ↗