Research map / Learning theory and algorithms / Matrix and tensor methods
New Insights for the Multivariate Square-Root Lasso
Aaron J. Molstad
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 cross-validation instead of theory-based quantile tuning of lambda (ablation) | MSR · Model 1-3, L1 penalty, M1 beta | squared Frobenius norm error: directly tuned estimators tend to perform substantially worse than MSR-CV, including MSR-Or | Improved |
| add refitting (SUR-type likelihood re-estimation) to theory-tuned MSR (ablation) | MSR-q95/q85/q75/q50 · Model 1-3, L1 penalty, M1 beta | relative squared Frobenius norm error to MSR-CV: performance relative to MSR-CV much improved with refitting | Improved |
| use theory-based tuning (q50) with group lasso penalty (ablation) | MSR-q50 · Model 1-3, group lasso, M2 beta | relative squared Frobenius norm error to MSR-CV: error less than 1.25 that of MSR-CV almost always | No clear effect |
| use multivariate square-root lasso instead of MRCE-Approx | MSR-CV · Model 2 and 3, L1 penalty | squared Frobenius norm error: MSR-CV outperforms MRCE-Approx in nearly every setting where the precision matrix is nonsparse | Improved |
| use MSR-CV instead of MRCE-Approx under group lasso | MSR-CV · Model 1-3, group lasso, M2 beta | squared Frobenius norm error: MSR-CV outperforms MRCE-Approx in every setting | Improved |
| use MSR-CV instead of MRCE-Approx on Model 1 xi=0.3 | MSR-CV · Model 1 xi=0.3, normal errors | squared Frobenius norm error: MRCE-Approx slightly outperforms MSR-CV at xi=0.3 | Hurt |
| use ADMM instead of CVX solver | MSR with L1 penalty · Model 1 normal errors | average computing time (s): 0.61 vs 82.89 seconds at xi=0.3 | Improved |
| use ADMM instead of MRCE-Approx | MSR-CV · Model 1 normal errors | implementation time: MSR-CV under a minute versus MRCE-Approx over 250 seconds | Improved |
| use MSR-CV instead of PLS/Calibrated on Model 2 with small n | MSR-CV · Model 2, (n,p,q)=(50,500,60) | squared Frobenius norm error: MSR-CV and MRCE-Approx worse than PLS and Calibrated | Hurt |
| use theory-based lambda (MSR-q95) instead of cross-validation (ablation) | MSR · Model 1 xi=0.3, L1 penalty | true positive rate: 0.812 -> 0.487 | Hurt |
| use theory-based lambda (MSR-q95) instead of cross-validation (ablation) | MSR · Model 1 xi=0.3, L1 penalty | false positive rate: 0.043 -> 0.000 | Improved |
| use theory-based lambda (MSR-q95) instead of cross-validation (ablation) | MSR · Model 1 xi=0.3, nuclear norm | average estimated rank: 14.860 -> 2.590 | Improved |
Closest papers by meaning
- A Unified Semismooth Newton Approach to Multitask and Multivariate Square-Root Lasso Problems (NeurIPS 2026 · similarity 0.88)
- A Cluster Elastic Net for Multivariate Regression (JMLR 2018 · similarity 0.87)
- Regression Learning with Limited Observations of Multivariate Outcomes and Features (ICML 2024 · similarity 0.87)
- Sharp Oracle Inequalities for Square Root Regularization (JMLR 2017 · similarity 0.87)
- Modular Regression: Improving Linear Models by Incorporating Auxiliary Data (JMLR 2023 · similarity 0.87)
- A convex optimization formulation for multivariate regression (NeurIPS 2020 · similarity 0.86)
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.
