Research map / Learning theory and algorithms / Sample complexity
Composable Sketches for Functions of Frequencies: Beyond the Worst Case
Edith Cohen, Ofir Geri, Rasmus Pagh
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 ℓ2 sampling instead of ℓ1 sampling to emulate high-moment sampling | Sampling sketches · SO out, SO in, AOL, CAIDA, and UGR | emulation overhead for 3rd and 10th moments: SO out: 124.30 -> 3.74 (3rd); 600.44 -> 18.04 (10th); SO in: 299.80 -> 4.12 (3rd); 677.56 -> 9.30 (10th); AOL: 34.81 -> | Improved |
| Use ℓ2 sampling for universal emulation rather than ℓ1 or concave-sublinear sampling | Sampling sketches · SO out, SO in, AOL, CAIDA, and UGR | universal emulation overhead: SO out: 624.58 vs 1.25 × 10^5 and 1672.60; SO in: 681.72 vs 4.20 × 10^4 and 1628.36; AOL: 92.92 vs 2.94 × 10^5 and 170.8 | Hurt |
| Use sampling by advice rather than ℓ1 PPSWOR for estimating high moments with small samples | Sampling-by-advice sketch · AOL and Stack Overflow | NRMSE for high frequency moments | Mixed |
| Use with-replacement ℓ2 sampling rather than sampling by advice | High-moment estimation · AOL and Stack Overflow | NRMSE for high frequency moments | Improved |
| Restrict inputs to distributions with an ℓq heavy hitter | ℓq sampling · Frequency distributions; target moments p ≥ q | emulation overhead: 1/φ | Improved |
| Use ℓ1 sampling for the universal objective instead of ℓ2 sampling | Universal sampling · SO out, SO in, AOL, CAIDA, and UGR | universal emulation overhead: SO out: 1.25 × 10^5 -> 624.58; SO in: 4.20 × 10^4 -> 681.72; AOL: 2.94 × 10^5 -> 92.92; CAIDA: 2.65 × 10^5 -> 301.28; UG | Improved |
| Use multi-objective concave-sublinear sampling for universal emulation rather than ℓ2 sampling | Universal sampling · SO out, SO in, AOL, CAIDA, and UGR | universal emulation overhead: SO out: 1.25 × 10^5 -> 1672.60; SO in: 4.20 × 10^4 -> 1628.36; AOL: 2.94 × 10^5 -> 170.84; CAIDA: 2.65 × 10^5 -> 846.15; | Improved |
Closest papers by meaning
- Sampling Sketches for Concave Sublinear Functions of Frequencies (NeurIPS 2019 · similarity 0.91)
- WOR and $p$'s: Sketches for $\ell_p$-Sampling Without Replacement (NeurIPS 2020 · similarity 0.88)
- Improved Frequency Estimation Algorithms with and without Predictions (NeurIPS 2023 · similarity 0.87)
- Putting the “Learning" into Learning-Augmented Algorithms for Frequency Estimation (ICML 2021 · similarity 0.86)
- Learning-Augmented Frequent Directions (ICLR 2025 · similarity 0.84)
- A Separation in Heavy-Tailed Sampling: Gaussian vs. Stable Oracles for Proximal Samplers (NeurIPS 2024 · similarity 0.84)
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.
