Research map / Learning theory and algorithms / Kernels and regression theory
The Optimal Ridge Penalty for Real-world High-dimensional Data Can Be Zero or Negative due to the Implicit Ridge Regularization
Dmitry Kobak, Jonathan Lomond, Benoit Sanchez
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use all p=3116 predictors instead of p=50 random predictors (ablation) | Ridge regression · liver.toxicity | CV MSE: lowest MSE at smallest lambda for five of ten responses | Mixed |
| Increase dimensionality to p=1000 (n=64) in spiked covariance model (ablation) | Ridge regression · Spiked covariance simulation | optimal lambda: 31 (p=50) -> -150 (p=1000) | Mixed |
| Allow negative ridge penalty in the high-dimensional regime (ablation) | Ridge regression · Spiked covariance simulation (p>=600) | expected MSE: lambda_opt negative for p of about 600 and above | Improved |
| Add more positive ridge penalty on top of minimum-norm OLS when n<<p (ablation) | Ridge regression · Spiked covariance simulation | expected MSE: minimum MSE at lambda=0 | No clear effect |
| Augment predictors with random predictors of variance lambda_opt/q (q to infinity) (ablation) | Minimum-norm OLS · Spiked covariance simulation (n=64, p=50) | MSE: approaches ridge with lambda_opt=31 around q=200 | Improved |
| Augment with q random predictors of variance 1 (ablation) | Minimum-norm OLS · Spiked covariance simulation (n=64, p=50) | MSE: MSE has a minimum at q_opt; larger q over-regularizes | Mixed |
| Use correlated predictors (rho=0.1) instead of spherical covariance (ablation) | Ridge regression · Spiked covariance simulation | optimal lambda: lambda_opt = p/alpha when rho=0; negative when rho=0.1 at large p | Mixed |
| Allow negative lambda with random Fourier features (ablation) | Ridge regression · MNIST (n=64, p=2000) | expected risk: lambda_opt of about -80 vs minimum at 0 for positive lambda | Improved |
Closest papers by meaning
- Optimal Ridge Regularization for Out-of-Distribution Prediction (ICML 2024 · similarity 0.89)
- Risk and cross validation in ridge regression with correlated samples (ICML 2025 · similarity 0.88)
- Ridge Boosting is Both Robust and Efficient (NeurIPS 2025 · similarity 0.87)
- An Agnostic View on the Cost of Overfitting in (Kernel) Ridge Regression (ICLR 2024 · similarity 0.87)
- Overfitting Behaviour of Gaussian Kernel Ridgeless Regression: Varying Bandwidth or Dimensionality (NeurIPS 2024 · similarity 0.87)
- High-Dimensional Analysis for Generalized Nonlinear Regression: From Asymptotics to Algorithm (AAAI 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
