A1299
Title: Inference for low-rank models without rank estimation
Authors: Hyukjun Kwon - Rutgers University (United States) [presenting]
Yuan Liao - Rutgers University (United States)
Jungjun Choi - Rutgers University (United States)
Abstract: A new debiasing procedure for linear low-rank models is introduced, where the parameter of interest is a high-dimensional matrix coefficient. The procedure achieves asymptotic normality without requiring knowledge of the true rank of the parameter matrix. The key feature of this approach is the use of diversified weights. An intermediate estimator is projected onto low-rank linear spaces that are estimated using these weights. Notably, this projection is robust to the rank misspecification. However, the estimated projection matrices are inconsistent with the true projections, creating new challenges in characterizing the asymptotic distribution of the debiased estimator. Nonetheless, the proposed debiasing procedure successfully addresses these issues. Lastly, the procedure does not require sample splitting.