Pith. sign in

REVIEW 3 major objections 3 minor 2 cited by

Blade samples calibrated posterior distributions for inverse problems using only black-box forward evaluations, without ever computing a gradient.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 10:12 UTC pith:EOTNAUEB

load-bearing objection New derivative-free ensemble sampler with strong empirical results on a hard nonlinear inverse problem; the calibration claim is partly tuned and the main theorem relies on an assumption that is unverified and arguably false as stated. the 3 major comments →

arxiv 2510.10968 v3 pith:EOTNAUEB submitted 2025-10-13 cs.LG stat.ML

Blade: A Derivative-free Bayesian Inversion Method using Diffusion Priors

Hongkai Zheng , Austin Wang , Zihui Wu , Zhengyu Huang , Ricardo Baptista , Yisong Yue This is my paper
classification cs.LG stat.ML MSC 62F1562M2065C3068T07
keywords Bayesian inversionderivative-freediffusion priorsensemble Kalman samplingsplit Gibbs samplingstatistical linearizationuncertainty calibrationNavier-Stokes
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper introduces Blade, a Bayesian inversion method that produces accurate, well-calibrated posterior samples when the forward model can only be queried through evaluations, not differentiated. It combines the split Gibbs sampling framework with an ensemble of interacting particles: a likelihood step that approximates the forward model by a statistically linearized surrogate, and a prior step that runs a diffusion model in reverse. The authors prove a non-asymptotic convergence bound showing the time-averaged Fisher divergence to the reference process decays as O(1/K) up to the statistical-linearization error and the diffusion score-approximation error. Empirically, on a nonlinear Navier-Stokes inverse problem, Blade achieves spread-skill ratios near one and the best CRPS among derivative-free baselines, whereas competing methods collapse to over-confident point estimates.

Core claim

The central claim is that accurate and well-calibrated posterior sampling can be achieved for high-dimensional, nonlinear inverse problems without gradient or adjoint access to the forward model by alternating between an ensemble covariance-preconditioned likelihood step and a denoising diffusion prior step. The likelihood step replaces the true forward model gradient with a statistical linearization estimated from the current ensemble, and each particle samples its own split-Gibbs conditional, allowing multi-modal posteriors that shared-potential ensemble methods cannot capture. The theoretical result (Theorem 2) bounds the time-averaged Fisher divergence between the approximate and referen

What carries the argument

The algorithm alternates between two steps within a split Gibbs sampler. The likelihood step evolves particles under covariance-preconditioned Langevin dynamics in which the forward model gradient is replaced by a least-squares linear surrogate estimated from the ensemble (statistical linearization), and a correction term (n+1)/J times the centered particle preserves the invariant measure in the finite-particle system. The prior step implements the reverse process of a pre-trained diffusion model, mapping each z to x via denoising with coupling strength ρ set to the diffusion noise level. The convergence bound is carried by a generalized Fisher-divergence contraction inequality (Lemma 2) app

Load-bearing premise

The convergence guarantee rests on Assumption 3, which requires the Radon-Nikodym derivative of the reference distribution with respect to the approximate one to be constant along the null space of the ensemble covariance; with an ensemble of 512 particles and state dimension 16,384 this rank-deficiency condition is not verified and the bound in Eq. (12) may not apply.

What would settle it

A direct check would be to compute, for a low-rank ensemble on a toy linear inverse problem, whether the left-hand side of Eq. (12) actually respects the predicted O(1/K) decay; a violation in a regime where ϵ_model and ϵ_score are both small but the ensemble covariance is rank-deficient would indicate that Assumption 3 is load-bearing. Alternatively, constructing an explicit counterexample where Assumption 3 fails but the algorithm nevertheless produces miscalibrated samples would refute the universal applicability of the bound.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the bound in Theorem 2 holds, more split-Gibbs iterations decrease the posterior approximation error at rate O(1/K) until the irreducible linearization and score errors dominate, and the method does not drift or blow up under those errors.
  • Blade can serve as a practical posterior sampler for scientific problems such as data assimilation, geophysical inversion, and fluid reconstruction where forward-model derivatives are unavailable or prohibitively expensive.
  • Because the likelihood step uses only ensemble evaluations, the method can be wrapped around any black-box simulator or legacy numerical code without modification.
  • The distinction between Blade (main) and Blade (diag) gives practitioners a principled choice: full covariance for calibrated uncertainty, diagonal approximation for sharper point estimates at the cost of under-dispersion.
  • Scaling the ensemble size and number of interactions improves both accuracy and calibration, matching the theoretical prediction and the empirical scaling plots.

Where Pith is reading between the lines

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

  • The statistical linearization error term ϵ_model is likely the dominant bottleneck in practice: when the ensemble spans a low-dimensional subspace, the linear surrogate is accurate only in that subspace, so the bound's dependence on λ* suggests performance degrades in exactly the regime where finite ensembles meet high-dimensional state spaces.
  • One testable extension is to use localization or covariance tapering to restore a full-rank effective preconditioner when J << n, which would directly address Assumption 3's rank-deficiency failure mode and may improve calibration in the Navier-Stokes setting.
  • The proof technique, which compares two processes through a common preconditioner geometry, could extend to other preconditioned ensemble methods (e.g., consensus-based samplers) by identifying a suitable common metric structure.
  • Blade's data-scaling plot suggests the diffusion prior itself benefits from more data, so coupling Blade with foundation-model priors trained on large physical datasets could broaden its reach beyond the benchmarks shown.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 3 minor

Summary. The paper proposes Blade, a derivative-free Bayesian inversion algorithm built on split Gibbs sampling. It alternates between an ensemble-based likelihood step that uses statistical linearization of the forward model and a prior step implemented with a pretrained diffusion model. The authors provide a non-asymptotic convergence bound (Theorem 2) that characterizes the time-averaged Fisher divergence between the approximate process and a reference process in terms of the number of iterations, the statistical-linearization error, and the score-estimation error. Empirically, they evaluate Blade on a linear-Gaussian problem, a Gaussian-mixture problem, and a high-dimensional Navier-Stokes inverse problem, reporting CRPS, spread-skill ratio (SSR), rank histograms, and relative L2 error. The central claim is that Blade produces accurate and well-calibrated posterior samples using only forward evaluations of the model, which is argued both theoretically and empirically.

Significance. If the claims hold, Blade would be a useful contribution: it extends derivative-free ensemble Kalman methods to multimodal posteriors by coupling them with diffusion priors, and it provides an explicit convergence analysis that accounts for both statistical linearization and score error. The theoretical framework is a natural extension of existing analyses, and the paper is transparent about many approximations. The empirical results on the Navier-Stokes benchmark are encouraging, and the comparison with strong baselines (including paired-data conditional diffusion models) is a strength. However, the central theoretical guarantee hinges on an assumption that is unverified and likely violated in the high-dimensional experiments, and the headline calibration result is partly manufactured by tuning a key hyperparameter. These issues materially affect the paper's main claims and require revision.

major comments (3)
  1. [§3.4, Eq. (12), Assumption 3] Theorem 2's bound rests on Assumption 3 and on λ*>0. In the Navier-Stokes experiments n=16384 and J=512, so the instantaneous covariance C_t has rank at most 511 and a null space of dimension >15000. The remark after Theorem 2 claims Assumption 3 holds if μ_t≪μ̃_t with continuously differentiable log-densities; this implication is false, because differentiability does not imply the density ratio is constant along Ker(C_t). Fig. 3 tracks the rank of the accumulated matrix [X_0,Z_0,...], not the instantaneous C_t appearing in the theorem. If Assumption 3 fails, ∇log(dμ_t/dμ̃_t) has components in Ker(C_t) and the λ* lower bound in Lemma 2 is invalid, so Eq. (12) is not established. λ* is also unquantified; for a rank-511 covariance in 16384 dimensions it can be arbitrarily small, making the 1/λ* factor vacuous.
  2. [Appendix B.1] The effective observation noise σ̃_y is, in the authors' words, 'tuned so that spread-skill ratio is close to 1.' Since SSR is one of the two headline calibration metrics in Table 1, the claim of producing 'well-calibrated posterior samples' is partly enforced by construction: adjusting a single global spread parameter to match the average RMSE will mechanically give SSR≈1. The rank histogram and CRPS provide some independent evidence, but the paper should report a fixed σ̃_y (or tune on a separate validation set and then evaluate on unseen test data) to make the calibration claim non-circular.
  3. [§3.4, Theorem 2] The reference process is defined as starting from the stationary distribution πXZ, but the proof concludes 'μ0 = πX', the marginal of πXZ. Lemma 2 is applied to the conditional distribution of z given x^{(j)}, yet the KL term in Eq. (12) is written as DKL(πX || μ̃0). It is therefore unclear whether the theorem concerns the joint distribution, the x-marginal, or a conditional distribution over the active variable. This ambiguity needs to be resolved for the theorem to be a well-posed statement.
minor comments (3)
  1. [Eq. (14)] The diffusion coefficient δ(t) is defined as sqrt(2σ̇σ)+sqrt(2β), but the diffusion coefficient in Eq. (2) is sqrt(2σ̇σ+2β). This is inconsistent and should be corrected; the derivation uses δ(t)² as the coefficient of the Fisher-divergence term.
  2. [Table 1] The SSR values for Blade (main) are all very close to unity, but no uncertainty or significance information is given. Reporting confidence intervals over test cases would strengthen the calibration claim.
  3. [§4.1 / Fig. 1] The Gaussian-mixture experiment is presented only qualitatively. Adding quantitative distributional metrics (e.g., sliced Wasserstein distance or KL divergence, as used in the linear-Gaussian case) would make the multimodal claim more robust.

Circularity Check

1 steps flagged

SSR calibration metric is also the tuning objective for the effective observation noise; the theoretical bound itself is self-contained.

specific steps
  1. fitted input called prediction [Appendix B.1 'Effective observation noise'; Algorithm 2; Table 1/Sec. 4.2 results]
    "In practice, we treat ˜σy as a hyperparameter and tune it so that spread-skill ratio is close to 1."

    The paper's headline calibration evidence for Blade(main) is SSR≈0.95 (Table 1). The same SSR is the explicit objective used to select σ̃_y, the hyperparameter that directly scales the ensemble-likelihood drift (Alg. 2 line 4 uses 1/σ̃_y² and hence controls ensemble spread). Thus the report 'its SSR remains close to one' is not an independent model output for that metric; it is the fitted target. This is a partial empirical circularity. It does not touch Theorem 2, since ϵ_model and ϵ_score are defined independently and CRPS, rank histograms, and point accuracy are separate, unfitted evidence; but the SSR-based calibration claim is partly constructed by the tuning procedure.

full rationale

Theorem 2's derivation is not circular: the bound (Eq. 12) is expressed in terms of independently defined errors ϵ_model (Assumption 2) and ϵ_score (Assumption 1), and Lemma 2 is proven in the appendix rather than imported from the overlapping-author citation Wu et al. (2024). The reference process is simply the stationary split-Gibbs target, so convergence to it is a well-defined stability statement. The main non-circular caveat is Assumption 3: the paper states that absolute continuity plus C^1 log-densities suffices for constancy of dµ_t/dµ̃_t along Ker(C_t), but that implication is not proved and appears mathematically insufficient; this is a rigor/correctness concern about whether Eq. (12) applies (especially at n=16384, J=512), not a circularity. The one genuinely circular element is the empirical calibration loop: σ̃_y is tuned to make the spread-skill ratio close to 1 and the same SSR is then presented as evidence that Blade is well-calibrated. Because this is a disclosed validation-set hyperparameter choice and because CRPS, rank histograms, and the theoretical analysis provide independent content, the overall circularity is minor rather than derivation-level.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 0 invented entities

The method rests on several tuned hyperparameters, the most consequential being σ̃_y, which is fit to the calibration metric. The theoretical guarantees require technical conditions (Assumption 3, positivity of λ*) that are not verified in the evaluated high-dimensional regime. No new physical entities are postulated.

free parameters (8)
  • Effective observation noise \tilde{\sigma}_y = 0.03 (main, noise=0), 0.17 (noise=1), 0.3 (noise=2); diag: 0.001–0.25
    Tuned per noise level and mode so that the spread-skill ratio is close to 1; directly controls the posterior spread and the reported calibration.
  • Step size scale \gamma = 20–35
    Controls the Euler step size in the likelihood SDE; chosen by hand/validation.
  • Minimum coupling strength \rho_{\min} = 0.06–0.08
    Endpoint of the annealing schedule; chosen by validation.
  • Maximum coupling strength \rho_{\max} = 4.8
    Start of annealing; chosen by validation.
  • Coupling annealing schedule = linear for main, concave for diag
    Selected per mode based on ablation.
  • Ensemble size J = 512
    Set from ablation study (Fig. 11) where performance plateaus.
  • Number of split-Gibbs iterations K = 25
    Set from scaling study; gains plateau beyond ~20 iterations.
  • Likelihood discretization steps N = 50
    Number of Euler steps per likelihood step; chosen by hand.
axioms (6)
  • domain assumption Bounded score approximation error of the diffusion model (Assumption 1)
    Standard in diffusion sampling theory; assumed to hold for the pretrained prior, but not verified.
  • domain assumption Bounded statistical linearization error ϵ_model (Assumption 2)
    Requires smoothness of the forward model and uniformly bounded ensemble covariance; not established for the Navier-Stokes setting.
  • ad hoc to paper Assumption 3: dµ_t/dµ˜_t is constant along the null space of C_t
    Technical condition introduced to apply Lemma 2; unlikely to hold when C_t is rank-deficient (n=16384, J=512).
  • domain assumption λ* = inf λ_t^* > 0 and δ > 0
    The ensemble covariance must retain positive eigenvalues and the diffusion noise must be nonzero; not guaranteed in finite-particle high-dimensional settings.
  • standard math Reverse SDE of the diffusion model exactly samples the conditional p(x0|xt)
    Underlying diffusion model theory (Karras et al. 2022); used in Eq. (10) for the prior step.
  • standard math Statistical linearization is exact for linear forward models
    Definition of A_t in Eq. (7); exact when G is linear, but only approximate for nonlinear G.

pith-pipeline@v1.3.0-alltime-deepseek · 29260 in / 13986 out tokens · 120122 ms · 2026-08-04T10:12:09.910900+00:00 · methodology

0 comments
read the original abstract

Derivative-free Bayesian inversion arises in science and engineering applications, particularly when forward model is costly or infeasible to differentiate through. Existing derivative-free methods collapse the posterior to a point estimate or return severely over-confident uncertainty on high-dimensional, nonlinear problems. We introduce Blade, which produces accurate and well-calibrated posteriors using an ensemble of interacting particles. Blade leverages diffusion models as data-driven priors, and only queries the forward model through forward evaluations (i.e., derivative-free). Theoretically, we show the convergence and stability of Blade under forward model approximation and prior score estimation error. Empirically, on nonlinear fluid dynamics, Blade produces well-calibrated posterior samples that existing derivative-free methods cannot, as measured by CRPS, the spread-skill ratio, and the rank histogram. Its accuracy and calibration improve consistently with more iterations and particles, backed by our convergence and stability analysis and empirical experiments.

Figures

Figures reproduced from arXiv: 2510.10968 by Austin Wang, Hongkai Zheng, Ricardo Baptista, Yisong Yue, Zhengyu Huang, Zihui Wu.

Figure 1
Figure 1. Figure 1: (a): Results on linear Gaussian and Gaussian mixture problems. Blue samples are from [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustrative depiction of Blade (see Sec. 3). Algorithm 1 Blade method for derivative-free Bayesian inversion using diffusion priors Require: initial ensemble X(0) = {x (j) ∈ R n} J j=1, number of iterations K, {ρk} K k=0, observed data y ∈ R m, pre-trained diffusion model sθ 1: for k ∈ {0, . . . , K − 1} do 2: Z (k) ← Ensemble-likelihood-step(X(k) , ρk) ▷ Algorithm 2 3: X(k+1) ← Ensemble-prior-step(Z (k+1… view at source ↗
Figure 3
Figure 3. Figure 3: Evolution of the rank of the space spanned by ensemble par￾ticles during Blade iterations. Putting it all together. We provide pseudocode for the com￾plete sampling algorithm in Algorithm 1. The method operates by iteratively updating an ensemble of particles, alternating between the likelihood and prior steps discussed above. At the same time, the parameter ρ follows an annealing schedule that gradually d… view at source ↗
Figure 4
Figure 4. Figure 4: (a): Effect of different hyperparameters. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Illustration of the three annealing schedules. Each curve visualizes how the coupling [PITH_FULL_IMAGE:figures/full_fig_p023_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Effect of different hyperparameters of the [PITH_FULL_IMAGE:figures/full_fig_p024_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Effect of different hyperparameters of the [PITH_FULL_IMAGE:figures/full_fig_p025_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Ablation study on the effect of annealing schedules. [PITH_FULL_IMAGE:figures/full_fig_p026_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative and probabilistic assessment on Navier-Stokes inverse problem at three observa [PITH_FULL_IMAGE:figures/full_fig_p032_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Scaling of different methods’ performance with training data size across three measurement [PITH_FULL_IMAGE:figures/full_fig_p033_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Effect of ensemble size. At each observation noise level, we study how the performance [PITH_FULL_IMAGE:figures/full_fig_p033_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Effect of initialization. At each observation noise level, we compare the performance of [PITH_FULL_IMAGE:figures/full_fig_p034_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Effect of resampling strategy. At each observation noise level, we compare the performance [PITH_FULL_IMAGE:figures/full_fig_p034_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Runtime comparison of different algorithms on Navier-Stokes inverse problem, measured [PITH_FULL_IMAGE:figures/full_fig_p034_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Additional image generation results. 35 [PITH_FULL_IMAGE:figures/full_fig_p035_15.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Flow Annealing Posterior Sampling for Function-Space Regression and Inverse Problems

    stat.ML 2026-06 unverdicted novelty 7.0

    FAPS is a new function-space posterior sampling method built on flow-matching priors that unifies stochastic-process regression and PDE inverse problems while avoiding explicit prior density evaluation.

  2. Provable Diffusion Posterior Sampling for Bayesian Inversion

    stat.ML 2025-12 conditional novelty 6.0

    A diffusion posterior sampler using Monte Carlo Langevin score estimation and warm start is proven to converge in Wasserstein-2 distance under semi-log-concavity and sub-Gaussian assumptions, and outperforms DPS/TV on...

Reference graph

Works this paper leans on

24 extracted references · 14 linked inside Pith · cited by 2 Pith papers

  1. [1]

    0.85 0.51 0.94 linear concave edm 0.6 0.8 1.0 1.2 0.95 0.72 1.19 linear concave edm 0.6 0.8 1.0 1.2 1.05 0.80 1.24 Figure 8: Ablation study on the effect of annealing schedules. 26 Preprint Using Bayes theorem, the posterior distribution is given by p(x|y) = p(y|x)p(x) p(y) .(34) The likelihoodp(y|x)reads p(y|x) =N(y;Hx,Σ ϵ).(35) Therefore, p(x|y) = PK i=...

    2007

  2. [2]

    Further, if the covariance matrix is positive definite, the stationary distribution is unique

    as a stationary distribution. Further, if the covariance matrix is positive definite, the stationary distribution is unique. Proof. This result has been proved in various forms in the literature (Ma et al., 2015; Garbuno-Inigo et al., 2020a), we provide a simple proof of our use case for ease of understanding. Suppose µt(z) is the probability density of z...

    2015

  3. [3]

    We discretize the denoising diffusion process with the standard Euler method

    We set σ(t) =t for simplicity and employ the Euler ODE sampler for faster sampling. We discretize the denoising diffusion process with the standard Euler method. Following Karras et al. (2022), we use the following step size: ti = t1/7 max + i N−1 (t1/7 min −t 1/7 max) 7 , i= 0, . . . , N−1. B.3 ANNEALING SCHEDULE In our experiments, we explored three dif...

    2022

  4. [7]

    Fourier neural operator for parametric partial differential equations

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. arXiv preprint arXiv:2010.08895,

    Pith/arXiv arXiv 2010

  5. [9]

    A general framework for inference-time scaling and steering of diffusion models.arXiv preprint arXiv:2501.06848,

    Raghav Singhal, Zachary Horvitz, Ryan Teehan, Mengye Ren, Zhou Yu, Kathleen McKeown, and Rajesh Ranganath. A general framework for inference-time scaling and steering of diffusion models.arXiv preprint arXiv:2501.06848,

    Pith/arXiv arXiv

  6. [10]

    Solving inverse problems with latent diffusion models via hard data consistency.arXiv preprint arXiv:2307.08123,

    Bowen Song, Soo Min Kwon, Zecheng Zhang, Xinyu Hu, Qing Qu, and Liyue Shen. Solving inverse problems with latent diffusion models via hard data consistency.arXiv preprint arXiv:2307.08123,

    Pith/arXiv arXiv

  7. [11]

    Evaluation of probabilistic prediction systems

    Olivier Talagrand. Evaluation of probabilistic prediction systems. InWorkshop Proceedings” Workshop on Predictability”, 20-22 October 1997, ECMWF , Reading, UK,

    1997

  8. [13]

    Prin- cipled probabilistic imaging using diffusion models as plug-and-play priors.arXiv preprint arXiv:2405.18782,

    Zihui Wu, Yu Sun, Yifan Chen, Bingliang Zhang, Yisong Yue, and Katherine L Bouman. Prin- cipled probabilistic imaging using diffusion models as plug-and-play priors.arXiv preprint arXiv:2405.18782,

    Pith/arXiv arXiv

  9. [14]

    Provably robust score-based diffusion posterior sampling for plug-and- play image reconstruction.arXiv preprint arXiv:2403.17042,

    Xingyu Xu and Yuejie Chi. Provably robust score-based diffusion posterior sampling for plug-and- play image reconstruction.arXiv preprint arXiv:2403.17042,

    Pith/arXiv arXiv

  10. [15]

    Fast sampling of diffusion models with exponential integrator

    Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. arXiv preprint arXiv:2204.13902,

    Pith/arXiv arXiv

  11. [17]

    (6) reads ∂µt ∂t =∇ ·(µtCt∇Φ(z)) +∇ ·(Ct∇µt), 15 Preprint which can be rewritten as ∂µt ∂t =∇ ·(µtCt(∇Φ(z) +∇logµ t)).(19) Letµ ∞ denote the stationary distribution of Eq

    The corresponding Fokker-Planck equation for Eq. (6) reads ∂µt ∂t =∇ ·(µtCt∇Φ(z)) +∇ ·(Ct∇µt), 15 Preprint which can be rewritten as ∂µt ∂t =∇ ·(µtCt(∇Φ(z) +∇logµ t)).(19) Letµ ∞ denote the stationary distribution of Eq. (19). We have 0 =∇ ·(µtCt(∇Φ(z) +∇logµ ∞)). If the particle distribution is not Dirac, Ct ̸= 0 due to Lemma 2.1 in Garbuno-Inigo et al. ...

    1996

  12. [18]

    Proof.Fort∈[T k, Tk +t †], k= 0,

    Assuming that DKL(πX ||µ0)<+∞and Assumption 3 holds, forKsplit Gibbs iterations, we have 1 TK Z TK 0 DFI(µt||˜µt)dt≤ 4 min(λ∗, δ) DKL(πX ||˜µ0) K(t † +t ∗) + t†ϵmodel +t ∗ϵscore t† +t ∗ (12) where DFI and DKL are Fisher divergence and KL divergence respectively, defined in Appendix A.1. Proof.Fort∈[T k, Tk +t †], k= 0, . . . , K−1, we apply Lemma 2 to the...

    2019

  13. [21]

    Intuitively, it measures the distance between a predicted distribution and the single observed ground truth x∗ that actually occurred

    is a standard probabilistic metric to assess the quality of the entire predicted distribution for inverse problems, which is defined as CRPS =E|x−x ∗| −1 2 E|x−x ′|, where x,x ′ are independent random predictions andx∗ is the single observed ground truth. Intuitively, it measures the distance between a predicted distribution and the single observed ground...

    2024

  14. [23]

    We fix the ensemble size to 1024 and optimize the other hyperparameters using Bayesian optimization

    is a method based on interacting particle system, which can be viewed as a generalization of SCG (Huang et al., 29 Preprint 2024). We fix the ensemble size to 1024 and optimize the other hyperparameters using Bayesian optimization. C.3 IMAGE RESTORATION C.3.1 PROBLEM SETUP We evaluated our method on image restoration tasks with the FFHQ256 dataset. Our ev...

    2024

  15. [24]

    At each observation noise level, we compare the performance ofBladewith and without the resampling strategy

    0.0 0.2 0.4 0.6 0.8 1.0Metric value noise = 2 Perfect calibration Resample No resample Figure 13: Effect of resampling strategy. At each observation noise level, we compare the performance ofBladewith and without the resampling strategy. Blade (ours) EnKG SCG DPG EKI EKS+DM Algorithms 0 20 40 60 80 100Runtime (minutes) 48.0 62.0 59.0 114.0 60.5 20.0 Figur...

    arXiv

  16. [1999]

    assesses ensemble calibration by comparing the truth to the empirical distribution formed by the ensemble. For each grid point in each test case, the ensemble members are sorted, and the rank of the true value within this ordering (0 means below all, J means above all, or an intermediate integer) is recorded. Pooling these ranks over all points and cases ...

    2022

  17. [2007]

    Efficient derivative- free bayesian inference for large-scale inverse problems.Inverse Problems, 38(12):125006, 2022a

    Daniel Zhengyu Huang, Jiaoyang Huang, Sebastian Reich, and Andrew M Stuart. Efficient derivative- free bayesian inference for large-scale inverse problems.Inverse Problems, 38(12):125006, 2022a. Daniel Zhengyu Huang, Tapio Schneider, and Andrew M Stuart. Iterated kalman methodology for inverse problems.Journal of Computational Physics, 463:111262, 2022b. ...

    Pith/arXiv arXiv

  18. [2015]

    Note on interacting langevin diffusions: gradient struc- ture and ensemble kalman sampler by garbuno-inigo, hoffmann, li and stuart.arXiv preprint arXiv:1908.10890,

    12 Preprint Nikolas N ¨usken and Sebastian Reich. Note on interacting langevin diffusions: gradient struc- ture and ensemble kalman sampler by garbuno-inigo, hoffmann, li and stuart.arXiv preprint arXiv:1908.10890,

    Pith/arXiv arXiv 1908

  19. [2017]

    10 Preprint Ross N Bannister

    URLhttps://arxiv.org/abs/1701.04862. 10 Preprint Ross N Bannister. A review of operational methods of variational and ensemble-variational data assimilation.Quarterly Journal of the Royal Meteorological Society, 143(703):607–633,

    Pith/arXiv arXiv

  20. [2020]

    Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions.arXiv preprint arXiv:2209.11215,

    Sitan Chen, Sinho Chewi, Jerry Li, Yuanzhi Li, Adil Salim, and Anru R Zhang. Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions.arXiv preprint arXiv:2209.11215,

    Pith/arXiv arXiv

  21. [2022]

    Derivative-free diffusion manifold-constrained gradient for unified xai.arXiv preprint arXiv:2411.15265,

    Won Jun Kim, Hyungjin Chung, Jaemin Kim, Sangmin Lee, Byeongsu Sim, and Jong Chul Ye. Derivative-free diffusion manifold-constrained gradient for unified xai.arXiv preprint arXiv:2411.15265,

    Pith/arXiv arXiv

  22. [2023]

    Florentin Coeurdoux, Nicolas Dobigeon, and Pierre Chainais

    URL https://openreview.net/forum? id=OnD9zGAGT0k. Florentin Coeurdoux, Nicolas Dobigeon, and Pierre Chainais. Plug-and-play split gibbs sampler: embedding deep generative priors in bayesian inference.arXiv preprint arXiv:2304.11134,

    Pith/arXiv arXiv

  23. [2024]

    Diffusion probabilistic modeling of protein backbones in 3d for the motif- scaffolding problem.arXiv preprint arXiv:2206.04119,

    Brian L Trippe, Jason Yim, Doug Tischer, David Baker, Tamara Broderick, Regina Barzilay, and Tommi Jaakkola. Diffusion probabilistic modeling of protein backbones in 3d for the motif- scaffolding problem.arXiv preprint arXiv:2206.04119,

    Pith/arXiv arXiv

  24. [2025]

    Elucidating the design space of diffusion- based generative models.arXiv preprint arXiv:2206.00364,

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion- based generative models.arXiv preprint arXiv:2206.00364,

    Pith/arXiv arXiv