linear-regression

This paper establishes a sharp Loewner envelope for coefficient covariance under volume-sampled least squares, using feature-only geometry to analyze tightness conditions.

This paper proposes an analytical framework for estimating the optimal early stopping time in neural network training using Rademacher complexity with L1-norm, avoiding probabilistic assumptions and applicable to both linear and nonlinear models.

This paper proposes a method using reservoir computing to recycle computational processes of dynamic programming for combinatorial optimization problems, achieving improved approximation accuracy and reduced computation time on traveling salesman and subset sum problems.

This paper proposes a robust variant of smart predict-then-optimize that accounts for feature perturbations, providing a convex surrogate with theoretical guarantees and demonstrating superior performance over standard methods.

This paper constructs a transformer with linear self-attention that performs in-context learning of the closed-form least squares solution for simple linear regression, using layer normalization to approximate the analytical solution rather than gradient descent.

Proves a tight approximation ratio for the greedy algorithm in myopic Bayesian active learning for linear regression, identifying the maximum initial leverage score as a key quantity.

This paper presents an algorithm for group distributionally robust least squares regression using block Lewis weights, achieving improved complexity over interior point methods. It also provides interpolating algorithms between average and robust losses.

This paper introduces a new differential privacy sketching mechanism based on fast transforms that achieves state-of-the-art privacy guarantees and improved runtime, and applies it to DP linear regression to obtain the first fast method for DP ordinary least squares.

This paper derives batch scaling laws for sketched linear regression under power-law spectra, analyzing one-pass and multi-pass mini-batch SGD. It provides explicit risk decompositions showing how batch size affects bias, variance, and fluctuation terms, and establishes that without-replacement sampling yields lower noise than with-replacement.