Research map / Optimization / Implicit bias and generalization
Global Convergence of Sub-gradient Method for Robust Matrix Recovery: Small Initialization, Noisy Measurements, and Over-parameterization
Jianhao Ma, Salar Fattahi
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 |
|---|---|---|---|
| Add small initialization to SubGM | SubGM for robust matrix recovery · Noisy measurements with over-parameterized rank | convergence rate: exponentially faster | Improved |
| Use small initialization rather than initialization scale α = 1 (ablation) | SubGM · Robust matrix recovery with outlier noise | recovery error | Improved |
| Use search rank larger than the true rank (ablation) | SubGM with small initialization · Robust matrix recovery with 10% corrupted measurements | convergence rate | No clear effect |
| Increase the corruption probability (ablation) | SubGM with small initialization and fully over-parameterized rank · Outlier noise model | recovery error | No clear effect |
| Use smaller initialization (ablation) | SubGM · Robust matrix recovery | final recovery error | Improved |
| Use smaller initialization (ablation) | SubGM · Robust matrix recovery | iterations to convergence | Hurt |
| Use small initialization and geometric decaying step-size | SubGM with ℓ1-loss · Outlier noise model, including arbitrarily large corruptions | iteration complexity: \kappa^2\log^3\left(\frac{d}{\varepsilon}\right) vs. \frac{\sigma_1\kappa}{\varepsilon} | Improved |
| Use small initialization with ℓ1-loss rather than over-parameterized GD with ℓ2-loss | Robust matrix recovery · Gaussian noise model, r'≥r | sample complexity: \nu^2 \kappa^4dr^{2} vs. \nu^2\kappa^2dr' | Improved |
Closest papers by meaning
- A Unified Variance Reduction-Based Framework for Nonconvex Low-Rank Matrix Recovery (ICML 2017 · similarity 0.89)
- Guarantees of a Preconditioned Subgradient Algorithm for Overparameterized Asymmetric Low-rank Matrix Recovery (ICML 2025 · similarity 0.89)
- Global Optimality of Local Search for Low Rank Matrix Recovery (NeurIPS 2016 · similarity 0.88)
- Rank Overspecified Robust Matrix Recovery: Subgradient Method and Exact Recovery (NeurIPS 2021 · similarity 0.88)
- Local and Global Linear Convergence of General Low-Rank Matrix Recovery Problems (AAAI 2022 · similarity 0.88)
- Sharp Restricted Isometry Property Bounds for Low-Rank Matrix Recovery Problems with Corrupted Measurements (AAAI 2022 · similarity 0.88)
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.
