pith. sign in

arxiv: 2606.25169 · v1 · pith:2U6IFGJTnew · submitted 2026-06-23 · 🧮 math.ST · cs.LG· stat.TH

Laplace--Fisher Gate Identities for Optimal Matrix-Gated Blended Score Estimation

Alois Duston , Tan Bui Tanh This is my paper

Pith reviewed 2026-06-25 21:44 UTC · model grok-4.3

classification 🧮 math.ST cs.LGstat.TH
keywords score estimationOrnstein-Uhlenbeck diffusionblended estimatorsLaplace-Fisher Gate IdentityBayesian inverse problemsmatrix gatesTweedie identitytarget score identity
0
0 comments X

The pith

The Laplace-Fisher Gate Identity supplies the variance-optimal matrix gate for blending Tweedie and target-score estimators.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper treats blended score estimation as a conditional risk minimization problem whose decision variables are matrix-valued blending coefficients called gates. Solving that problem produces an explicit formula for the optimal gate that involves the conditional expectation of the negative Hessian of the log target density. The construction preserves the estimator's expectation because the Tweedie-TSI disagreement has conditional mean zero, so only its variance changes. A reader would care because the resulting finite-reference estimator yields a normalized density surrogate from ordinary MCMC output and derivative information, which in turn supports evidence estimation and calibration checks in Bayesian inverse problems.

Core claim

Blended score estimation is cast as conditional risk minimization over matrix-valued blending coefficients, or gates, and the variance-optimal gate is derived as G*(y,t) = α_t² (α_t² I_d + γ_t E[H_0(X_0) | Y_t = y])^{-1}, where H_0 = -∇² log p_0, α_t = e^{-t} and γ_t = 1 - e^{-2t}. The formula is called the Laplace-Fisher Gate Identity. Because the Tweedie-TSI disagreement has conditional mean zero, the gate changes estimator variance without changing its expected value. Finite-reference consistency and stability bounds are proved for estimating the gate from weighted reference samples, and the estimator is applied to normalized posterior-density evaluation in Bayesian inverse problems.

What carries the argument

The Laplace-Fisher Gate Identity, which gives the optimal matrix gate as the scaled inverse of a matrix that regularizes the conditional expectation of the target Hessian.

If this is right

  • The optimal gate can be estimated consistently from finite weighted reference samples with proved stability bounds.
  • When MCMC pilot samples and derivative information are available, the gate produces a normalized posterior-density surrogate.
  • The surrogate supports posterior-energy evaluation, model-evidence estimation, and density-based diagnostics beyond sample-based methods.
  • On a PDE-constrained inverse-problem benchmark the method improves posterior-density calibration and sampling diagnostics relative to other tested score-estimator classes.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The matrix-valued gate naturally accommodates strongly anisotropic or singular targets that defeat scalar blending coefficients.
  • Because the gate is estimated from the same reference samples already used for score estimation, the overhead remains modest when derivative information is already computed.
  • The separation between variance reduction and expectation preservation may extend to blending other pairs of unbiased estimators whose difference has conditional mean zero.

Load-bearing premise

The disagreement between the Tweedie and target-score identities has conditional mean zero given the noisy observation.

What would settle it

A Monte Carlo check that computes the conditional expectation of the Tweedie-TSI difference given Y_t = y on a large sample and finds it statistically different from zero would falsify preservation of the estimator's expectation.

Figures

Figures reproduced from arXiv: 2606.25169 by Alois Duston, Tan Bui Tanh.

Figure 1
Figure 1. Figure 1: Reference-count score-RMSE sweep on the d = 8 misaligned singular-subspace GMM. The metric is the time-averaged noisy-score RMSE (Appendix section E, Def. E.7). LFGI remains below the other learned estimators across the displayed reference-bank sizes. 31 [PITH_FULL_IMAGE:figures/full_fig_p031_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Misaligned singular-subspace GMM, K = 8, d = 8, rank 3: two-dimensional marginal/PCA projections. Panels follow the method titles. 11.5 Experiment II: misaligned singular-subspace GMM in d = 24 The second GMM target keeps the same family and increases the ambient dimension to d = 24, with intrinsic rank 4, component radius 4.5, and normal scale σ⊥ = 0.035. The example remains a controlled PSD/pole-separate… view at source ↗
Figure 3
Figure 3. Figure 3: Neal funnel in d = 10: two-dimensional funnel-coordinate histograms. Panels follow the method titles; robust coordinate limits show the narrow neck and widening mouth simultaneously. Method Sliced KS ↓ MMD ↓ NLL ↓ Score RMSE proxy ↓ Tweedie 0.0721±0.0061 (3.90±0.57)×10−3 223.4±28.1 148.7±72.3 Uniform Scalar Blend 0.0688±0.0051 (3.80±0.86)×10−3 212.3±24.5 62.1±45.8 Scalar Blend 0.0902±0.0104 (6.70±2.30)×10−… view at source ↗
Figure 4
Figure 4. Figure 4: Same projection diagnostic as fig. 2, now for the d = 24, rank-4 target. Panels follow the method titles. 42 [PITH_FULL_IMAGE:figures/full_fig_p042_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Darcy-flow density-evaluation diagnostic on held-out MALA-EVAL samples. Left: true [PITH_FULL_IMAGE:figures/full_fig_p047_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Auxiliary gate-capture sweep on the d = 8 misaligned singular-subspace GMM. Columns use t = 0.04, 0.08, 0.16; rows show relative Frobenius gate error and risk-weighted gate error. LFGI has smaller error than Matrix Blend across the displayed gate-bank sizes in both geometries. H Auxiliary Pole Diagnostics H.1 Neal-funnel shifted-pole audit The pole diagnostic checks the finite-reference quantities that app… view at source ↗
read the original abstract

Sampling from an unnormalized target by reversing an Ornstein--Uhlenbeck diffusion requires the score of each noise-perturbed marginal. Tweedie's identity and a target-score identity give unbiased finite-reference estimators for this score. Scalar blends can reduce variance, but are too rigid for singular or strongly anisotropic targets. We cast blended score estimation as conditional risk minimization over matrix-valued blending coefficients, or gates, and derive the variance-optimal gate [ \Gstar(y,t)=\alphat^2\bigl(\alphat^2 I_d+\gammat,\E[H_0(X_0)\mid Y_t=y]\bigr)^{-1},\qquad H_0=-\nabla^2\log p_0 . ] Here (\alphat=e^{-t}) and (\gammat=1-e^{-2t}). We call this formula the \emph{Laplace--Fisher Gate Identity} (\LFGI{}). Since the Tweedie--TSI disagreement has conditional mean zero, the gate changes estimator variance without changing its expected value. We give the Gaussian special case and prove finite-reference consistency and stability bounds for estimating the gate from weighted reference samples. We apply the finite-reference LFGI estimator to normalized density evaluation for Bayesian inverse problems. When MCMC pilot samples and derivative information are available, LFGI uses these byproducts to construct a normalized posterior-density surrogate. The surrogate enables posterior-energy evaluation, model-evidence estimation, and density-based diagnostics beyond those available from samples alone. On a PDE-constrained inverse-problem benchmark, LFGI improves posterior-density calibration and sampling diagnostics relative to the other tested score-estimator classes, and known-evidence experiments check absolute calibration in Gaussian and non-Gaussian settings.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper casts blended score estimation for Ornstein-Uhlenbeck diffusion reversal as conditional risk minimization over matrix-valued gates and derives the variance-optimal Laplace-Fisher Gate Identity G*(y,t)=α_t²(α_t² I_d + γ_t E[H_0(X_0)|Y_t=y])^{-1} with α_t=e^{-t}, γ_t=1-e^{-2t}. It asserts that the Tweedie-TSI disagreement has conditional mean zero (so the gate is unbiased), proves finite-reference consistency and stability bounds when the gate is estimated from weighted reference samples, gives the Gaussian case, and applies the resulting estimator to construct normalized posterior-density surrogates for Bayesian inverse problems, reporting improved calibration and diagnostics on a PDE-constrained benchmark.

Significance. If the central claims hold, the matrix gate supplies a principled, variance-optimal blending rule that is more flexible than scalar blends for singular or anisotropic targets while preserving unbiasedness; the finite-reference consistency result and the use of MCMC byproducts for normalized density evaluation would be useful additions to the score-estimation and Bayesian-computation literature.

major comments (2)
  1. [Abstract (LFGI formula and following sentence)] Abstract (statement immediately following the LFGI formula): the claim that the Tweedie-TSI disagreement has conditional mean zero is invoked to guarantee that the gate changes only variance and not expectation, yet no derivation, reference, or explicit verification is supplied; this premise is load-bearing for the unbiasedness of the finite-reference estimator and must be established before the consistency bounds can be accepted.
  2. [Abstract (finite-reference consistency and stability bounds)] Abstract (finite-reference consistency claim): the gate formula depends on the conditional expectation E[H_0(X_0)|Y_t=y], which itself must be estimated from the same reference samples used to form the blended score; the dependence structure and any resulting bias in the plug-in estimator are not addressed in the abstract, so the stated consistency and stability bounds cannot yet be verified.
minor comments (1)
  1. [Abstract] Notation for α_t and γ_t is introduced only after the gate formula; moving the definitions to the first appearance of the formula would improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and the identification of points that require clarification in the abstract. We respond to each major comment below.

read point-by-point responses

  1. Referee: [Abstract (LFGI formula and following sentence)] Abstract (statement immediately following the LFGI formula): the claim that the Tweedie-TSI disagreement has conditional mean zero is invoked to guarantee that the gate changes only variance and not expectation, yet no derivation, reference, or explicit verification is supplied; this premise is load-bearing for the unbiasedness of the finite-reference estimator and must be established before the consistency bounds can be accepted.

    Authors: The conditional mean-zero property follows because both Tweedie's identity and the target-score identity are unbiased estimators of the same score function ∇log p_t(y); their difference therefore has conditional expectation zero given Y_t = y. This is shown by direct computation in Section 3.1. We will add a parenthetical reference to this section immediately after the claim in the revised abstract. revision: yes

  2. Referee: [Abstract (finite-reference consistency and stability bounds)] Abstract (finite-reference consistency claim): the gate formula depends on the conditional expectation E[H_0(X_0)|Y_t=y], which itself must be estimated from the same reference samples used to form the blended score; the dependence structure and any resulting bias in the plug-in estimator are not addressed in the abstract, so the stated consistency and stability bounds cannot yet be verified.

    Authors: The consistency and stability bounds (Theorem 4.1 and Corollary 4.2) are proved for the joint plug-in estimator in which both the blended score and the matrix gate (including the estimated conditional expectation E[H_0(X_0)|Y_t=y]) are formed from the same weighted reference samples. The proof uses a uniform concentration argument that accounts for the dependence. We will revise the abstract to state that the reported bounds apply to this joint estimation procedure. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation self-contained

full rationale

The paper derives the matrix gate explicitly by minimizing conditional risk over blending coefficients, producing the closed-form expression involving the conditional Hessian expectation. The subsequent statement that the Tweedie-TSI disagreement has conditional mean zero is presented as an independent premise that preserves unbiasedness; it is not obtained by substituting the gate formula back into itself or by any self-referential reduction. Finite-reference consistency and stability bounds are proved separately from the gate identity. No fitted parameter is relabeled as a prediction, no self-citation chain supplies a uniqueness theorem, and no ansatz is smuggled via prior work. The central identities therefore remain independent of their own outputs.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on the mathematical derivation of the optimal gate and on the claim that the Tweedie–TSI disagreement has conditional mean zero. No free parameters, invented entities, or additional axioms are stated in the abstract.

axioms (2)
  • domain assumption Tweedie identity and target-score identity supply unbiased finite-reference score estimators
    Invoked in the first paragraph of the abstract as the starting point for blended estimation.
  • domain assumption Tweedie–TSI disagreement has conditional mean zero
    Stated immediately after the gate formula to justify that the gate affects only variance.

pith-pipeline@v0.9.1-grok · 5849 in / 1447 out tokens · 34129 ms · 2026-06-25T21:44:16.337406+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

71 extracted references · 18 canonical work pages

  1. [1]

    Iterated denoising energy matching for sampling from boltzmann densities.arXiv preprint arXiv:2402.06121, 2024

    Tara Akhound-Sadegh, Jarrid Rector-Brooks, Avishek Joey Bose, Sarthak Mittal, Pablo Lemos, Cheng-Hao Liu, Marcin Sendera, Siamak Ravanbakhsh, Gauthier Gidel, Yoshua Bengio, Nikolay Malkin, and Alexander Tong. Iterated denoising energy matching for sampling from boltzmann densities.arXiv preprint arXiv:2402.06121, 2024. doi: 10.48550/ arXiv.2402.06121

    arXiv 2024

  2. [2]

    Brian D. O. Anderson. Reverse-time diffusion equation models.Stochastic Processes and their Applications, 12(3):313–326, 1982. doi: 10.1016/0304-4149(82)90051-5

    work page doi:10.1016/0304-4149(82)90051-5 1982

  3. [3]

    Kernel conditional exponential family

    Michael Arbel and Arthur Gretton. Kernel conditional exponential family. InProceedings of the 21st International Conference on Artificial Intelligence and Statistics (AISTATS), volume 84 ofProceedings of Machine Learning Research, pages 1337–1346. PMLR, 2018

    2018

  4. [4]

    Maximum mean discrepancy gradient flow

    Michael Arbel, Anna Korba, Adil Salim, and Arthur Gretton. Maximum mean discrepancy gradient flow. InAdvances in Neural Information Processing Systems (NeurIPS), 2019

    2019

  5. [5]

    Error bounds for flow matching methods.Transactions on Machine Learning Research, 2024

    Joe Benton, George Deligiannidis, and Arnaud Doucet. Error bounds for flow matching methods.Transactions on Machine Learning Research, 2024

    2024

  6. [6]

    Girolami

    Tan Bui-Thanh and Mark A. Girolami. Solving large-scale PDE-constrained Bayesian inverse problems with Riemann manifold Hamiltonian Monte Carlo.Inverse Problems, 30(11):114014,

  7. [7]

    doi: 10.1088/0266-5611/30/11/114014

    work page doi:10.1088/0266-5611/30/11/114014

  8. [8]

    Sequential controlled langevin diffusions

    Junhua Chen, Lorenz Richter, Julius Berner, Denis Blessing, Gerhard Neumann, and Anima Anandkumar. Sequential controlled langevin diffusions. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  9. [9]

    Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David K. Duvenaud. Neural ordinary differential equations. InAdvances in Neural Information Processing Systems, volume 31, 2018

    2018

  10. [10]

    The probability flow ODE is provably fast

    Sitan Chen, Sinho Chewi, Holden Lee, Yuanzhi Li, Jianfeng Lu, and Adil Salim. The probability flow ODE is provably fast. InAdvances in Neural Information Processing Systems, volume 36, 2023. 68

    2023

  11. [11]

    A kernel test of goodness of fit

    Krzysztof Chwialkowski, Heiko Strathmann, and Arthur Gretton. A kernel test of goodness of fit. InProceedings of the 33rd International Conference on Machine Learning (ICML), volume 48 ofProceedings of Machine Learning Research, pages 2606–2615. PMLR, 2016

    2016

  12. [12]

    Duncan, Sebastian Reich, and "O

    Paula Cordero-Encinar, Andrew B. Duncan, Sebastian Reich, and "O. Deniz Akyildiz. Sampling by averaging: A multiscale approach to score estimation. InAdvances in Neural Information Processing Systems, 2025

    2025

  13. [13]

    Sinkhorn distances: Lightspeed computation of optimal transport

    Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. InAdvances in Neural Information Processing Systems, volume 26, 2013

    2013

  14. [14]

    Target score matching.arXiv preprint arXiv:2402.08667, 2024

    Valentin De Bortoli, Michael Hutchinson, Peter Wirnsberger, and Arnaud Doucet. Target score matching.arXiv preprint arXiv:2402.08667, 2024. doi: 10.48550/arXiv.2402.08667

    work page doi:10.48550/arxiv.2402.08667 2024

  15. [15]

    Journal of the Royal Statistical Society: Series B (Statistical Methodology) 68, 411–436

    Pierre Del Moral, Arnaud Doucet, and Ajay Jasra. Sequential monte carlo samplers.Journal of the Royal Statistical Society: Series B (Statistical Methodology), 68(3):411–436, 2006. doi: 10.1111/j.1467-9868.2006.00553.x

    work page doi:10.1111/j.1467-9868.2006.00553.x 2006

  16. [16]

    Tweedie's formula and selection bias https://doi.org/10.1198/jasa.2011.tm11181

    Bradley Efron. Tweedie’s formula and selection bias.Journal of the American Statistical Association, 106(496):1602–1614, 2011. doi: 10.1198/jasa.2011.tm11181

    work page doi:10.1198/jasa.2011.tm11181 2011

  17. [17]

    Simulating normalizing constants: From importance sampling to bridge sampling to path sampling.Statistical Science, 13(2):163–185, 1998

    Andrew Gelman and Xiao-Li Meng. Simulating normalizing constants: From importance sampling to bridge sampling to path sampling.Statistical Science, 13(2):163–185, 1998. doi: 10.1214/ss/1028905934

    work page doi:10.1214/ss/1028905934 1998

  18. [18]

    Girolami and B

    Mark Girolami and Ben Calderhead. Riemann manifold langevin and hamiltonian monte carlo methods.Journal of the Royal Statistical Society: Series B (Statistical Methodology), 73(2):123–214, 2011. doi: 10.1111/j.1467-9868.2010.00765.x

    work page doi:10.1111/j.1467-9868.2010.00765.x 2011

  19. [19]

    Will Grathwohl, Ricky T. Q. Chen, Jesse Bettencourt, Ilya Sutskever, and David Duvenaud. FFJORD: Free-form continuous dynamics for scalable reversible generative models. In International Conference on Learning Representations, 2019

    2019

  20. [20]

    Stochas- tic localization via iterative posterior sampling

    Louis Grenioux, Maxence Noble, Marylou Gabrié, and Alain Oliviero Durmus. Stochas- tic localization via iterative posterior sampling. InProceedings of the 41st International Conference on Machine Learning, Proceedings of Machine Learning Research. PMLR, 2024

    2024

  21. [21]

    A kernel two-sample test.Journal of Machine Learning Research, 13:723–773, 2012

    Arthur Gretton, Karsten Borgwardt, Malte Rasch, Bernhard Schölkopf, and Alexander Smola. A kernel two-sample test.Journal of Machine Learning Research, 13:723–773, 2012

    2012

  22. [22]

    Havens, Benjamin Kurt Miller, Bing Yan, Carles Domingo-Enrich, Anuroop Sriram, Daniel S

    Aaron J. Havens, Benjamin Kurt Miller, Bing Yan, Carles Domingo-Enrich, Anuroop Sriram, Daniel S. Levine, Brandon M. Wood, Bin Hu, Brandon Amos, Brian Karrer, Xiang Fu, Guan- Horng Liu, and Ricky T. Q. Chen. Adjoint sampling: Highly scalable diffusion samplers via adjoint matching. InProceedings of the 42nd International Conference on Machine Learning, vo...

    work page doi:10.48550/arxiv.2504.11713 2025

  23. [23]

    Training neural samplers with reverse diffusive KL divergence

    Jiajun He, Wenlin Chen, Mingtian Zhang, David Barber, and José Miguel Hernández-Lobato. Training neural samplers with reverse diffusive KL divergence. InProceedings of The 28th International Conference on Artificial Intelligence and Statistics, Proceedings of Machine Learning Research. PMLR, 2025. 69

    2025

  24. [24]

    Zeroth-order sampling methods for non-log-concave distributions: Alleviating metastability by denoising diffusion

    Ye He, Kevin Rojas, and Molei Tao. Zeroth-order sampling methods for non-log-concave distributions: Alleviating metastability by denoising diffusion. InAdvances in Neural Information Processing Systems, volume 37, pages 71122–71161, 2024

    2024

  25. [25]

    Denoising diffusion prob- abilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion prob- abilistic models. InAdvances in Neural Information Processing Systems (NeurIPS), 2020. URL https://proceedings.neurips.cc/paper/2020/file/ 4c5bcfec8584af0d967f1ab10179ca4b-Paper.pdf

    2020

  26. [26]

    Reverse diffusion monte carlo

    Xunpeng Huang, Hanze Dong, Yifan Hao, Yian Ma, and Tong Zhang. Reverse diffusion monte carlo. InThe Twelfth International Conference on Learning Representations, 2024. URLhttps://openreview.net/forum?id=esc8PjUQ8e

    2024

  27. [27]

    Estimation of non-normalized statistical models by score matching.Journal of Machine Learning Research, 6:695–709, 2005

    Aapo Hyvärinen. Estimation of non-normalized statistical models by score matching.Journal of Machine Learning Research, 6:695–709, 2005

    2005

  28. [28]

    On a formula for the product-moment coefficient of any order of a normal frequency distribution in any number of variables.Biometrika, 12(1/2):134–139, 1918

    Leon Isserlis. On a formula for the product-moment coefficient of any order of a normal frequency distribution in any number of variables.Biometrika, 12(1/2):134–139, 1918

    1918

  29. [29]

    Unke, and Arnaud Doucet

    Khaled Kahouli, Romuald Elie, Klaus-Robert Müller, Quentin Berthet, Oliver T. Unke, and Arnaud Doucet. Control variate score matching for diffusion models.arXiv preprint arXiv:2512.20003, 2025. doi: 10.48550/arXiv.2512.20003

    work page doi:10.48550/arxiv.2512.20003 2025

  30. [30]

    Latent target score matching, with an application to simulation-based inference

    Joohwan Ko and Tomas Geffner. Latent target score matching, with an application to simulation-based inference. InMachine Learning and the Physical Sciences Workshop, NeurIPS, 2025

    2025

  31. [31]

    Kolmogorov

    Andrey N. Kolmogorov. Sulla determinazione empirica di una legge di distribuzione.Giornale dell’Istituto Italiano degli Attuari, 4:83–91, 1933

    1933

  32. [32]

    Liu, and Wing Hung Wong

    Augustine Kong, Jun S. Liu, and Wing Hung Wong. Sequential imputations and bayesian missing data problems.Journal of the American Statistical Association, 89(425):278–288, 1994

    1994

  33. [33]

    Kernel stein discrepancy descent

    Anna Korba, Pierre-Cyril Aubin-Frankowski, Szymon Majewski, and Pierre Ablin. Kernel stein discrepancy descent. InInternational Conference on Machine Learning (ICML), volume 139 ofProceedings of Machine Learning Research. PMLR, 2021

    2021

  34. [34]

    Liu.Monte Carlo Strategies in Scientific Computing

    Jun S. Liu.Monte Carlo Strategies in Scientific Computing. Springer Series in Statistics. Springer, New York, 2001

    2001

  35. [35]

    Stein variational gradient descent: A general purpose bayesian inference algorithm

    Qiang Liu and Dilin Wang. Stein variational gradient descent: A general purpose bayesian inference algorithm. InAdvances in Neural Information Processing Systems (NeurIPS), 2016. URLhttps://arxiv.org/abs/1608.04471

    arXiv 2016

  36. [36]

    A kernelized stein discrepancy for goodness-of-fit tests

    Qiang Liu, Jason Lee, and Michael Jordan. A kernelized stein discrepancy for goodness-of-fit tests. InProceedings of the 33rd International Conference on Machine Learning (ICML), volume 48 ofProceedings of Machine Learning Research, pages 276–284. PMLR, 2016

    2016

  37. [37]

    Maximum likelihood training for score-based diffusion ODEs by high-order denoising score matching

    Cheng Lu, Kaiwen Zheng, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Maximum likelihood training for score-based diffusion ODEs by high-order denoising score matching. InProceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 14429–14460. PMLR, 2022. 70

    2022

  38. [38]

    Wilcox, Carsten Burstedde, and Omar Ghattas

    James Martin, Lucas C. Wilcox, Carsten Burstedde, and Omar Ghattas. A stochastic Newton MCMC method for large-scale statistical inverse problems with application to seismic inversion.SIAM Journal on Scientific Computing, 34(3):A1460–A1487, 2012. doi: 10.1137/110845598

    work page doi:10.1137/110845598 2012

  39. [39]

    Effective sample size for importance sampling based on discrepancy measures.Signal Processing, 131:386–401, 2017

    Luca Martino, Víctor Elvira, and Francisco Louzada. Effective sample size for importance sampling based on discrepancy measures.Signal Processing, 131:386–401, 2017

    2017

  40. [40]

    Simulating ratios of normalizing constants via a simple identity: A theoretical exploration.Statistica Sinica, 6(4):831–860, 1996

    Xiao-Li Meng and Wing Hung Wong. Simulating ratios of normalizing constants via a simple identity: A theoretical exploration.Statistica Sinica, 6(4):831–860, 1996

    1996

  41. [41]

    Radford M. Neal. Annealed importance sampling.Statistics and Computing, 11(2):125–139,

  42. [42]

    doi: 10.1023/A:1008923215028

    work page doi:10.1023/a:1008923215028

  43. [43]

    Learned reference-based diffusion sampler for multi-modal distributions

    Maxence Noble, Louis Grenioux, Marylou Gabrié, and Alain Oliviero Durmus. Learned reference-based diffusion sampler for multi-modal distributions. InThe Thirteenth Interna- tional Conference on Learning Representations, 2025

    2025

  44. [44]

    Nocedal, S

    Jorge Nocedal and Stephen J. Wright.Numerical Optimization. Springer, New York, 2 edition, 2006. doi: 10.1007/978-0-387-40065-5

    work page doi:10.1007/978-0-387-40065-5 2006

  45. [45]

    Owen.Monte Carlo: Theory, Methods and Examples

    Art B. Owen.Monte Carlo: Theory, Methods and Examples. Stanford University, 2013. URLhttps://artowen.su.domains/mc/. Online book

    2013

  46. [46]

    Particle denoising diffusion sampler.arXiv preprint arXiv:2402.06320, 2024

    Angus Phillips, Hai-Dang Dau, Michael John Hutchinson, Valentin De Bortoli, George Deligiannidis, and Arnaud Doucet. Particle denoising diffusion sampler.arXiv preprint arXiv:2402.06320, 2024. doi: 10.48550/arXiv.2402.06320

    work page doi:10.48550/arxiv.2402.06320 2024

  47. [47]

    Improved sampling via learned diffusions

    Lorenz Richter and Julius Berner. Improved sampling via learned diffusions. InThe Twelfth International Conference on Learning Representations, 2024. URLhttps://openreview. net/forum?id=F2cS6SozN9

    2024

  48. [48]

    Robert and George Casella.Monte Carlo Statistical Methods

    Christian P. Robert and George Casella.Monte Carlo Statistical Methods. Springer, New York, 2 edition, 2004. ISBN 978-0387212395

    2004

  49. [49]

    CRC press, 1986

    Bernard W Silverman.Density Estimation for Statistics and Data Analysis. CRC press, 1986

    1986

  50. [50]

    2006, Bayesian Analysis, 1, 833 , doi: 10.1214/06-BA127

    John Skilling. Nested sampling for general bayesian computation.Bayesian Analysis, 1(4): 833–859, 2006. doi: 10.1214/06-BA127

    work page doi:10.1214/06-ba127 2006

  51. [51]

    Nikolai V. Smirnov. Table for estimating the goodness of fit of empirical distributions.The Annals of Mathematical Statistics, 19(2):279–281, 1948

    1948

  52. [52]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. InProceedings of the 32nd International Conference on Machine Learning, volume 37 ofProceedings of Machine Learning Research, pages 2256–2265. PMLR, 2015. URLhttps://proceedings.mlr.press/v37/ sohl-dickstein15.html

    2015

  53. [53]

    Maximum likelihood training of score-based diffusion models

    Yang Song, Conor Durkan, Iain Murray, and Stefano Ermon. Maximum likelihood training of score-based diffusion models. InAdvances in Neural Information Processing Systems, volume 34, pages 1415–1428, 2021. 71

    2021

  54. [54]

    Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. InInternational Conference on Learning Representations (ICLR), 2021. URLhttps:// openreview.net/forum?id=PxTIG12RRHS

    2021

  55. [55]

    Density estimation in infinite dimensional exponential families.Journal of Machine Learning Research, 18(57):1–59, 2017

    Bharath K Sriperumbudur, Kenji Fukumizu, Arthur Gretton, Aapo Hyvärinen, and Revant Kumar. Density estimation in infinite dimensional exponential families.Journal of Machine Learning Research, 18(57):1–59, 2017

    2017

  56. [56]

    A bound for the error in the normal approximation to the distribution of a sum of dependent random variables

    Charles Stein. A bound for the error in the normal approximation to the distribution of a sum of dependent random variables. InProceedings of the Sixth Berkeley Symposium on Mathematical Statistics and Probability, Volume 2: Probability Theory, pages 583–602. University of California Press, 1972

    1972

  57. [57]

    Andrew M. Stuart. Inverse problems: A bayesian perspective.Acta Numerica, 19:451–559,

  58. [58]

    doi: 10.1017/S0962492910000061

    work page doi:10.1017/s0962492910000061

  59. [59]

    Luke Tierney and Joseph B. Kadane. Accurate approximations for posterior moments and marginal densities.Journal of the American Statistical Association, 81(393):82–86, 1986. doi: 10.1080/01621459.1986.10478240

    work page doi:10.1080/01621459.1986.10478240 1986

  60. [60]

    Joel A. Tropp. User-friendly tail bounds for sums of random matrices.Foundations of Computational Mathematics, 12(4):389–434, 2012. doi: 10.1007/s10208-011-9099-z

    work page doi:10.1007/s10208-011-9099-z 2012

  61. [61]

    van der Vaart.Asymptotic Statistics

    Aad W. van der Vaart.Asymptotic Statistics. Cambridge University Press, 1998

    1998

  62. [62]

    Transport meets variational inference: Controlled monte carlo diffusions

    Francisco Vargas, Shreyas Padhy, Denis Blessing, and Nikolas N"usken. Transport meets variational inference: Controlled monte carlo diffusions. InThe Twelfth International Conference on Learning Representations, 2024. URLhttps://openreview.net/forum?id= PP1rudnxiW

    2024

  63. [63]

    Cambridge University Press, 2018

    Roman Vershynin.High-Dimensional Probability. Cambridge University Press, 2018

    2018

  64. [64]

    A connection between score matching and denoising autoencoders https://doi.org/10.1162/NECO_a_00142

    Pascal Vincent. A connection between score matching and denoising autoencoders.Neural Computation, 23(7):1661–1674, 2011. doi: 10.1162/NECO_a_00142

    work page doi:10.1162/neco_a_00142 2011

  65. [65]

    Wainwright.High-Dimensional Statistics

    Martin J. Wainwright.High-Dimensional Statistics. Cambridge University Press, 2019

    2019

  66. [66]

    Wenliang, Danica J

    Li K. Wenliang, Danica J. Sutherland, Heiko Strathmann, and Arthur Gretton. Learning deep kernels for exponential family densities. InInternational Conference on Machine Learning (ICML), volume 97 ofProceedings of Machine Learning Research, pages 6737–6746. PMLR, 2019

    2019

  67. [67]

    Naesseth, and John P

    Luhuan Wu, Yi Han, Christian A. Naesseth, and John P. Cunningham. Reverse diffusion sequential monte carlo samplers. InAdvances in Neural Information Processing Systems,

  68. [68]

    URLhttps://arxiv.org/abs/2508.05926

    arXiv

  69. [69]

    Deniz Akyildiz

    James Matthew Young, Paula Cordero-Encinar, Sebastian Reich, Andrew Duncan, and "O. Deniz Akyildiz. Diffusion path samplers via sequential monte carlo.arXiv preprint arXiv:2601.21951, 2026. URLhttps://arxiv.org/abs/2601.21951

    Pith/arXiv arXiv 2026

  70. [70]

    Path integral sampler: A stochastic control approach for sampling

    Qinsheng Zhang and Yongxin Chen. Path integral sampler: A stochastic control approach for sampling. InInternational Conference on Learning Representations, 2022. URLhttps: //openreview.net/forum?id=_uCb2ynRu7Y. 72

    2022

  71. [71]

    Nonparametric score estimators

    Yuhao Zhou, Jiaxin Shi, and Jun Zhu. Nonparametric score estimators. InInternational Conference on Machine Learning (ICML), volume 119 ofProceedings of Machine Learning Research, pages 11513–11523. PMLR, 2020. 73

    2020