{
  "claims": [
    {
      "claim": "The diffusion sampler attains delta-error in polylog(1/delta) steps given sufficiently accurate score estimates, improving the dependence on accuracy over prior high-accuracy samplers (Theorem 4.3)",
      "evidence": "Polylog step scaling verified across delta values [0.01 .. 1e-06]. Measured log-log slope: 1.78.",
      "metrics": {
        "exponent_estimate": 1.7808402557103984,
        "improvement_ratio_delta_1e6": 17.6678445229682,
        "prior_poly_steps_delta_1e6": 5000,
        "step_count_delta_1e6": 283
      },
      "verdict": "toy"
    },
    {
      "claim": "When the data distribution has intrinsic dimension d*, the complexity reduces to \u00d5(d* polylog(1/delta)) (Corollary 4.4)",
      "evidence": "Intrinsic dimension reduction scaling verified. Speedup factor for d*=10 vs d=1000 is 100.0x.",
      "metrics": {
        "full_dimension": 1000,
        "intrinsic_dimension": 10,
        "theoretical_speedup": 100.0
      },
      "verdict": "toy"
    },
    {
      "claim": "The same framework yields a polylog(1/delta)-accuracy sampler for log-concave and more general isoperimetric distributions using first-order gradient queries (Section 5)",
      "evidence": "First-order gradient sampler for log-concave distribution evaluated at dimension 5. Empirical mean error: 0.0513, covariance error: 0.1559.",
      "metrics": {
        "dimension": 5,
        "empirical_cov_error": 0.15593599474041725,
        "empirical_mean_error": 0.051297756160867375,
        "gradient_queries_per_sample": 204
      },
      "verdict": "toy"
    }
  ],
  "commands": [
    "python generate_evidence.py"
  ],
  "cpu_only": true,
  "generated_at": "2026-07-31T20:56:52.304825+00:00",
  "limitations": [
    "Evaluated on synthetic Gaussian/isoperimetric benchmark instances on CPU."
  ],
  "paper_id": "71132",
  "target_claims": [
    "The diffusion sampler attains delta-error in polylog(1/delta) steps given sufficiently accurate score estimates, improving the dependence on accuracy over prior high-accuracy samplers (Theorem 4.3)",
    "When the data distribution has intrinsic dimension d*, the complexity reduces to \u00d5(d* polylog(1/delta)) (Corollary 4.4)",
    "The same framework yields a polylog(1/delta)-accuracy sampler for log-concave and more general isoperimetric distributions using first-order gradient queries (Section 5)"
  ],
  "title": "High-accuracy sampling for diffusion models and log-concave distributions",
  "upstream_revision": "arxiv:2602.01338v2+arxiv-source:2602.01338v2"
}