File size: 3,124 Bytes
8382c1e 43363ca 8382c1e dc22907 8382c1e a8e6d56 8382c1e a8e6d56 8382c1e 9b0020c 8382c1e dc22907 8382c1e 43363ca 8382c1e 43363ca 8382c1e 43363ca 8382c1e 43363ca 8382c1e 43363ca 8382c1e dc22907 9b0020c 8382c1e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 | {
"claim_title_fix": "2026-07-22T17:10:00Z",
"schema_version": 1,
"title": "Repro: High-accuracy sampling for diffusion models and log-concave distributions",
"emoji": "🎯",
"space_id": "snaykey/repro-high-accuracy-sampling",
"paper": {
"arxiv_id": "2602.01338",
"openreview_id": "GW3umRqsZZ"
},
"tags": [
"icml2026-repro",
"paper-GW3umRqsZZ"
],
"updated_at": "2026-07-19T11:37:20+00:00",
"root": {
"slug": "index",
"title": "Repro: High-accuracy sampling for diffusion models and log-concave distributions",
"file": "pages/index.md",
"children": [
{
"slug": "claim-1-diffusion-sampler-polylog-1-delta-steps-theorem-4-3",
"title": "The paper's First-Order Rejection Sampling (FORS) meta-algorithm (Theorem 3.1) produces samples with error δ using sample complexity bounded by 3Be^(2B)log(2/δ) with probability 1-δ (Theorem 3.1).",
"file": "pages/claim-1-diffusion-sampler-polylog-1-delta-steps-theorem-4-3/page.md",
"children": []
},
{
"slug": "claim-2-complexity-o-tilde-d-polylog-1-delta-theorem-4-3",
"title": "Under only a finite second-moment assumption (minimal assumptions), the diffusion sampler achieves query complexity O(d·log²(1/δ) + log³(1/δ)), giving polylog(1/δ) dependence rather than the poly(1/δ) of prior work (Theorem 4.1, Section 4).",
"file": "pages/claim-2-complexity-o-tilde-d-polylog-1-delta-theorem-4-3/page.md",
"children": []
},
{
"slug": "claim-3-intrinsic-dimension-d-star-complexity-corollary-4-4",
"title": "Under a non-uniform Lipschitz condition on the score (Assumption 4.3), a DDPM-like sampler achieves total-variation error controlled via chi-squared divergence with complexity O(√(dL_δ log(d/δ))·log(d/δ) + L_δ log²(d/δ)) (Theorem 4.4).",
"file": "pages/claim-3-intrinsic-dimension-d-star-complexity-corollary-4-4/page.md",
"children": []
},
{
"slug": "claim-4-non-uniform-lipschitz-refinement-theorem-4-9",
"title": "For distributions with low intrinsic dimension d★, an adaptive-step-size method attains complexity O(d★·log²((d+M₂²)/δ²)), replacing the ambient dimension d with d★ (Theorem 4.6).",
"file": "pages/claim-4-non-uniform-lipschitz-refinement-theorem-4-9/page.md",
"children": []
},
{
"slug": "claim-5-log-concave-sampling-via-gradient-queries-section-5",
"title": "Section 5 extends the FORS framework to sample from general log-concave distributions using only gradient evaluations (no density evaluations), giving the first polylog(1/δ) sampler in this setting (Section 5).",
"file": "pages/claim-5-log-concave-sampling-via-gradient-queries-section-5/page.md",
"children": []
},
{
"slug": "conclusion",
"title": "Conclusion",
"file": "pages/conclusion/page.md",
"children": []
}
]
},
"agent_view_tokens": 8606,
"revision": "1784461040769985400"
} |