Samplinglib
Lean gate not recorded for this source state main · ed709a7f320d

Midpoint Langevin

Dimension-dependent upper bounds and separate oracle lower bounds.

Main paper results remain open in Lean. Verified building blocks are listed separately below.

Suggested reading order
  1. Read exact v1 source and the full assumptions below; source informal overview does not replace precise main theorem contracts.
  2. Keep deterministic Heun/exponential midpoint and randomized/Poisson methods as separate kernels.
  3. Inspect exact source Section3/4 upper bounds and Section5/6 lower bounds/initialization; source proofs remain open in Lean.
  4. Reusable Gaussian/transport/convexity background belongs to existing canonical shared nodes, only genuinely relevant textbook chapters; no blanket seven-route mapping.
Paper, notation and source conventions

Sharp dimensional analysis of midpoint methods for Langevin sampling

Source: 2610.06308v1. Pinned source version

\[d\ge1,\quad V\in C^2(\mathbb R^d),\quad 0<\alpha I\preceq\nabla^2V(x)\preceq\beta I,\quad \kappa=\beta/\alpha,\quad \pi(dx)=Z^{-1}e^{-V(x)}dx\]
  • The Hessian inequalities hold at every point; \(\alpha\) and \(\beta\) are positive real numbers. Separate Hessian-to-chord and Gibbs-integrability proofs support the compiled normalized PBPS augmentation certificate: positive Gibbs integral, probability and exact product-volume joint density. Consult linked cells for independent-review and admission states. Conditional kernels, the sampling process and its guarantees remain separate.
  • \(\operatorname{TV}(P,Q)\) means sup over measurable events of \(|P(A)-Q(A)|\). For probability laws with densities it is half their \(L^1\) distance.
  • \(R_2(P\Vert Q)=\log\int\left(\frac{dP}{dQ}\right)^2\,dQ\) when P is absolutely continuous with respect to Q, and infinity otherwise. The PBPS paper writes \(D_2\) and calls the target \(\mu\); here the common target is \(\pi\).
  • Expected oracle queries are not worst-case runtime. Reference-point preparation, proximal access and gradient-only reductions must be charged under the particular theorem's model.

1. Main statements

RED · source theorem known · local proof open

Deterministic guarantees: source Theorem1.1 (informal)

Source: 2610.06308v1. Theorem1.1; precise Theorems3.1/4.1 and source Section2 algorithms

Open this theorem in the interactive research workspace

Statement

For \(0<\varepsilon\le\sqrt d/2\), Heun initialized at the unique minimizer \(x_\star\) reaches \(\sqrt{\alpha}\,W_2\le\varepsilon\) with the displayed gradient-query bound. Exponential midpoint with \(\text{friction }2\sqrt{\beta}\), \(X_0=x_\star\) and independent \(P_0\sim\mathcal N(0,I_d)\) has the second bound for its position law. \(\widetilde O\) suppresses logarithms. This faithfully restates the labelled informal theorem; exact schedules/universal constants of3.1/4.1 remain separate source obligations.

  • \(d\ge1\), V globally \(C^2\) on real d-space, \(0\prec\alpha I\preceq\nabla^2 V(x)\preceq\beta I\) at every x; \(\kappa=\beta/\alpha\); normalized target \(\pi\) proportional \(e^{-V}\).
  • No bounds on third or higher derivatives are assumed. \(W_2\) is true Wasserstein distance; accuracy is \(\sqrt{\alpha}\,W_2\).
  • Use the actual source midpoint algorithms in Section2, not an arbitrary integrator or abstract-error certificate.
  • \(0<\varepsilon\le\sqrt d/2\). Heun starts at \(x_\star\); exponential midpoint has \(X_0=x_\star\), independent standard Gaussian momentum and \(\text{friction }2\sqrt{\beta}\).
\[Q_{\rm Heun}=\widetilde O(\kappa^{4/3}d^{1/3}\varepsilon^{-2/3}),\qquad Q_{\rm ExpMid}=\widetilde O(\kappa^{5/4}d^{1/4}\varepsilon^{-1/2}),\qquad\sqrt\alpha W_2\le\varepsilon.\]
Lean statement — not formalized yet

A future declaration must bind the probability laws, normalization, regularity, source algorithm and oracle model explicitly. The formula is a source theorem contract, not Lean code.

No corresponding ASTIS declaration is asserted. Search locations below are candidates, not established dependencies or copied library proofs.

Proof architecture and calculations · 2 steps

This is a source-linked proof route, not a complete reconstruction of all cited lemmas. The dependencies below remain separate formalization tasks.

  1. Preserve the genuine transition and stationary error

    \[\nu_N^{\rm H}=\nu_0^{\rm H}(K_h^{\rm H})^N,\quad\boldsymbol\nu_N^{\rm E}=\boldsymbol\nu_0^{\rm E}(K_h^{\rm E})^N,\quad\nu_N^{\rm E}=(\boldsymbol\nu_N^{\rm E})_x,\quad\pi_h^{\rm E}=(\boldsymbol\pi_h^{\rm E})_x;\qquad W_2(\nu_N,\pi)\le W_2(\nu_N,\pi_h)+W_2(\pi_h,\pi).\]

    Heun uses the position Markov kernel \(K_h^H\). Exponential midpoint uses the true position-momentum kernel \(K_h^E\) and its phase-space law; \(\nu_N^E\) and \(\pi_h^E\) are position marginals before the position \(W_2\) triangle inequality. The position marginal is not assumed Markov. Actual invariant laws and finite-step contraction must be proved, not supplied as oracle premises.

    Source: 2610.06308v1. Sections2-4; AppendicesA/B

    Track or question this step in Live Formalization

    Lean step — not formalized yet

    This mathematical step is still open in Lean. The future declaration must expose exactly the objects and side conditions named above; no source-cited wrapper is counted as a proof.

    No corresponding ASTIS declaration is asserted. Search locations below are candidates, not established dependencies or copied library proofs.

  2. Use cancellation under only C2 Hessian bounds

    \[\alpha I\preceq\nabla^2V\preceq\beta I.\]

    Heun uses a discrete Poisson corrector, Malliavin–Stein and Gaussian smoothing. Exponential midpoint uses weighted negative-Sobolev divergence cancellation and transport. Missing regularity details must be supplied without adding higher derivative bounds to the source theorem.

    Source: 2610.06308v1. AppendicesA/B/E

    Track or question this step in Live Formalization

    Lean step — not formalized yet

    This mathematical step is still open in Lean. The future declaration must expose exactly the objects and side conditions named above; no source-cited wrapper is counted as a proof.

    No corresponding ASTIS declaration is asserted. Search locations below are candidates, not established dependencies or copied library proofs.

Lean proof — not formalized yet

Formalization will first match the named technology interfaces, then assemble this source theorem. No placeholder proof or source-cited wrapper has been added.

No corresponding ASTIS declaration is asserted. Search locations below are candidates, not established dependencies or copied library proofs.

Strict boundary

Source-pinned phase3 plan only; no new midpoint Lean theorem, actual kernel, invariant bias, schedule, convergence or complexity certificate is claimed.

Further source statements · 1

RED · source theorem known · local proof open

Stationary and finite-iteration lower bounds: source Theorem1.2 (informal)

Source: 2610.06308v1. Theorem1.2/Table1; Proposition5.1,Theorem5.2,Propositions5.3/6.1/6.2,Corollary6.3

Open this theorem in the interactive research workspace

Statement

For each Table1 method and every sufficiently small normalized step h, there is a strongly convex smooth target of bounded condition number and a position initialization with expected potential gap at most C*d. For every sufficiently small \(\varepsilon\), attaining \(\sqrt\alpha\,W_2(\text{position law after }N\text{ steps},\pi)\le\varepsilon\) requires the Table1 lower-bound number of queries up to logarithms. For underdamped methods the initial momentum is independent standard Gaussian. This is the source-labelled informal overview; stationary bias and finite-iteration claims keep their separate precise source contracts.

  • \(d\ge1\), V globally \(C^2\) on real d-space, \(0\prec\alpha I\preceq\nabla^2 V(x)\preceq\beta I\) at every x; \(\kappa=\beta/\alpha\); normalized target \(\pi\) proportional \(e^{-V}\).
  • No bounds on third or higher derivatives are assumed. \(W_2\) is true Wasserstein distance; accuracy is \(\sqrt{\alpha}\,W_2\).
  • Use the actual source midpoint algorithms in Section2, not an arbitrary integrator or abstract-error certificate.
  • Existential target and initialization, bounded condition number, sufficiently small normalized h and \(\varepsilon\); \(\mathbb E[V(X_0)-V(x_\star)]\le C d\). Underlying momentum is standard Gaussian and independent of X0.
  • Poisson midpoint uses at least2 grid points and constant expected gradients per batch. Exact stationary/finitestep quantifiers and source constants remain in the named precise results.
\[Q_{\rm Heun}\gtrsim d^{1/3}\varepsilon^{-2/3},\quad Q_{\rm ExpMid}\gtrsim d^{1/4}\varepsilon^{-1/2},\quad Q_{\rm RLMC,PLMC,RULMC,PULMC}\gtrsim d^{1/3}\varepsilon^{-1/3}.\]
Lean statement — not formalized yet

A future declaration must bind the probability laws, normalization, regularity, source algorithm and oracle model explicitly. The formula is a source theorem contract, not Lean code.

No corresponding ASTIS declaration is asserted. Search locations below are candidates, not established dependencies or copied library proofs.

Proof architecture and calculations · 2 steps

This is a source-linked proof route, not a complete reconstruction of all cited lemmas. The dependencies below remain separate formalization tasks.

  1. Keep deterministic bias and finite initialization distinct

    \[\mathbb E[V(X_0)-V(x_\star)]\le Cd.\]

    Oscillatory targets yield the deterministic Heun obstruction; Gaussian/anisotropic Gaussian targets give exponential-midpoint bounds. Stationary bias alone does not establish every finite-step initialization claim.

    Source: 2610.06308v1. Section5; Corollary6.3

    Track or question this step in Live Formalization

    Lean step — not formalized yet

    This mathematical step is still open in Lean. The future declaration must expose exactly the objects and side conditions named above; no source-cited wrapper is counted as a proof.

    No corresponding ASTIS declaration is asserted. Search locations below are candidates, not established dependencies or copied library proofs.

  2. Track shared randomized covariance fluctuations

    \[X\mid\mathcal H\sim\mathcal N(0,\Sigma(\mathcal H)).\]

    For the actual randomized methods, shared quadrature choices induce common covariance fluctuations. The source Section6 calculations and finite-step passage must be formalized under their actual randomness; independent coordinates must not be invented unconditionally.

    Source: 2610.06308v1. Section6; AppendixD

    Track or question this step in Live Formalization

    Lean step — not formalized yet

    This mathematical step is still open in Lean. The future declaration must expose exactly the objects and side conditions named above; no source-cited wrapper is counted as a proof.

    No corresponding ASTIS declaration is asserted. Search locations below are candidates, not established dependencies or copied library proofs.

Lean proof — not formalized yet

Formalization will first match the named technology interfaces, then assemble this source theorem. No placeholder proof or source-cited wrapper has been added.

No corresponding ASTIS declaration is asserted. Search locations below are candidates, not established dependencies or copied library proofs.

Strict boundary

Source-pinned phase3 plan only; neither a lower-bound oracle certificate nor finite-step proof is claimed. Upper-bound and hardness lanes stay separate until method/accuracy/initialization/query contracts match.

3. Verified work

No declaration-level publication is attached to this page yet. A source proof route is not a Lean verification.

Compare proof mechanisms and prerequisites

Reusable prerequisites · What changes in the proof?