Stanford team cracks robust tensor completion with R-ItCUR

Researchers from Stanford University (Hanqin Cai, Longxiu Huang, Jing Qin, Chengyue Wu) introduced **R-ItCUR**, a novel algorithm for **robust low-tubal-rank tensor completion** under **cross-concentrated sampling (t-CCS)**. The method addresses a critical gap in existing approaches by handling partial observations contaminated with **sparse, arbitrarily large outliers**—a common real-world scenario in medical imaging and geophysics. R-ItCUR operates directly on the sampled cross structure, partitioning it into two exterior blocks and an intersection block, then applies **adaptive blockwise Welsch correction** to suppress outliers while updating the low-rank component via **projected blockwise gradient descent**.

The algorithm’s efficiency stems from avoiding full-tensor reconstruction during iterations, delivering **substantial memory and computational savings**. Validation on **synthetic tensors**, **cardiac MRI data**, and **3D seismic datasets** demonstrated robust recovery with high accuracy, even under severe corruption. The results underscore the importance of explicitly leveraging t-CCS structure for robust tensor completion—a breakthrough that could reshape how we process high-dimensional data corrupted by noise or anomalies.