How to Train Your Intercept: A Comparison of Intercept Update Strategies in Coordinate Descent

Coordinate descent
Optimization
Regression
Regularization

How intercept updates affect coordinate descent for regularized generalized linear models, and why Newton steps help with imbalanced responses.

Authors

Johan Larsson

Frederik Fabricius-Bjerre

Published

13 October 2026

Details

arXiv, arXiv:2610.18043

Links
Abstract

The way a coordinate descent solver updates the intercept can determine how quickly it fits a regularized generalized linear model, particularly when responses are imbalanced. We compare updates using a global curvature bound, a safeguarded Newton step, and exact minimization over the intercept. Our analysis explains how conservative updates slow convergence by failing to resolve the coupling between the intercept and the coefficients. Experiments with simulated data, real data, and production solvers show that one Newton step often achieves progress similar to exact minimization at a lower cost. We also examine quadratic approximations and multinomial models, and give recommendations for handling the intercept in each setting.

 

Citation

BibTeX citation:
@misc{larsson2026,
  author = {Larsson, Johan and Fabricius-Bjerre, Frederik},
  publisher = {arXiv},
  title = {How to {Train} {Your} {Intercept:} {A} {Comparison} of
    {Intercept} {Update} {Strategies} in {Coordinate} {Descent}},
  number = {arXiv:2610.18043},
  date = {2026-10-13},
  url = {https://arxiv.org/abs/2610.18043},
  doi = {10.48550/arXiv.2610.18043},
  langid = {en},
  abstract = {The way a coordinate descent solver updates the intercept
    can determine how quickly it fits a regularized generalized linear
    model, particularly when responses are imbalanced. We compare
    updates using a global curvature bound, a safeguarded Newton step,
    and exact minimization over the intercept. Our analysis explains how
    conservative updates slow convergence by failing to resolve the
    coupling between the intercept and the coefficients. Experiments
    with simulated data, real data, and production solvers show that one
    Newton step often achieves progress similar to exact minimization at
    a lower cost. We also examine quadratic approximations and
    multinomial models, and give recommendations for handling the
    intercept in each setting.}
}
For attribution, please cite this work as:
Larsson, Johan, and Frederik Fabricius-Bjerre. 2026. “How to Train Your Intercept: A Comparison of Intercept Update Strategies in Coordinate Descent.” arXiv:2610.18043. Preprint, arXiv, October 13. https://doi.org/10.48550/arXiv.2610.18043.