{
  "schema_version": 1,
  "title": "Repro: Markov Chain Monte Carlo without Evaluating the Target",
  "emoji": "chart",
  "space_id": "snaykey/repro-speedbench",
  "paper": {
    "openreview_id": "dDkl5ZcyTl"
  },
  "tags": [
    "icml2026-repro",
    "paper-dDkl5ZcyTl"
  ],
  "updated_at": "2026-07-26T00:00:00+00:00",
  "root": {
    "slug": "index",
    "title": "Repro: Markov Chain Monte Carlo without Evaluating the Target",
    "file": "pages/index.md",
    "children": [
      {
        "slug": "executive-summary",
        "title": "Executive summary",
        "file": "pages/executive-summary/page.md",
        "children": []
      },
      {
        "slug": "claim-1-unification-common-procedure",
        "title": "Section 2 unifies the exchange algorithm (Algorithm 1), PoissonMH (Algorithm 2), and TunaMH (Algorithm 3) as instances of one common auxiliary-variable procedure (Section 2).",
        "file": "pages/claim-1-unification-common-procedure/page.md",
        "children": []
      },
      {
        "slug": "claim-2-meta-algorithm-two-aux-vars",
        "title": "The proposed meta-algorithm (Algorithm 4) introduces two auxiliary variables, ω1 guiding the proposal and ω2 enabling target-ratio estimation, and Proposition 2 proves it preserves the target distribution Π as stationary (Algorithm 4, Proposition 2).",
        "file": "pages/claim-2-meta-algorithm-two-aux-vars/page.md",
        "children": []
      },
      {
        "slug": "claim-3-proposition-1-reversibility",
        "title": "Proposition 1 gives a reversibility condition on the auxiliary-variable estimator, R_{θ→θ'}(ω)π(θ|x)P_{θ→θ'}(ω) = π(θ'|x)P_{θ'→θ}(ω), that guarantees the resulting kernel targets the correct posterior (Proposition 1).",
        "file": "pages/claim-3-proposition-1-reversibility/page.md",
        "children": []
      },
      {
        "slug": "claim-4-two-new-algorithms",
        "title": "The framework yields two new algorithms: Locally Balanced PoissonMH (Algorithm 5), which uses gradient-informed locally balanced proposals with balancing functions such as Barker's g(t)=t/(1+t) or the MALA-type g(t)=sqrt(t), and TunaMH with an SGLD proposal (Algorithm 6), which uses independent minibatches for gradient estimation and acceptance-ratio evaluation (Algorithm 5, Algorithm 6).",
        "file": "pages/claim-4-two-new-algorithms/page.md",
        "children": []
      },
      {
        "slug": "claim-5-doubly-intractable-and-tall-data",
        "title": "The framework is designed to apply to both doubly-intractable distributions, where the likelihood f_θ(x)/Z(θ) has an intractable normalizing constant, and tall-data settings with millions to billions of observations, without needing to evaluate the full target density (Section 1).",
        "file": "pages/claim-5-doubly-intractable-and-tall-data/page.md",
        "children": []
      },
      {
        "slug": "conclusion",
        "title": "Conclusion",
        "file": "pages/conclusion/page.md",
        "children": []
      }
    ]
  }
}