Waldorp's paper shows MDL beats Lasso for high-dimensional network selection
New proof: MDL cuts false positives where Lasso, AIC, and BIC fail
Lourens Waldorp's new arXiv preprint (2608.13171) revisits a core problem in network analysis: how to reliably pick out which nodes are connected when the number of parameters far exceeds the number of observations. The classic toolkit—Lasso, AIC, BIC—tends to produce spurious edges, especially when the true underlying model is nonlinear but is approximated by a linear one. Waldorp shows that the ridge penalty plays a surprising role: tuning it reduces test variance and pushes mean squared error into a 'double descent' curve, where adding more parameters actually improves generalization. That insight connects network selection to modern machine-learning theory, but the practical takeaway is stark: conventional penalties are too permissive in high dimensions because they ignore the volume of the model space.
The paper's central contribution is a theoretical proof that minimum description length (MDL) fixes this. MDL naturally incorporates model-space volume into its penalty, allowing it to keep false positive rates low in both correctly specified and misspecified settings. In experiments, MDL-based neighborhood selection matches or outperforms Lasso, AIC, and BIC even when the true model is nonlinear—a realistic scenario in genomics, neuroscience, and social network analysis. Waldorp's work offers a rigorous alternative for researchers who want conservative, reproducible network estimates without sacrificing sensitivity, and it reframes classic model-selection debates through the lens of double descent and high-dimensional geometry.
- MDL achieves provably lower false positive rates than Lasso, AIC, and BIC in high-dimensional neighborhood selection
- Ridge parameter controls mean squared error and triggers double descent, explaining why overly complex models can generalize well
- Theoretical guarantees hold for misspecified nonlinear models, making the method robust for real-world network data
Why It Matters
For genomics, neuroscience, and social network researchers, MDL could drastically reduce spurious edges and improve reproducibility without sacrificing true connections.