pith. sign in

arxiv: 2605.18745 · v1 · pith:YGQREDCWnew · submitted 2026-05-18 · 📊 stat.ML · cs.LG· cs.NA· math.NA· math.PR· q-fin.MF· stat.CO

SURGE: Approximation-free Training Free Particle Filter for Diffusion Surrogate

Lifu Wei , Yinuo Ren , Naichen Shi , Yiping Lu This is my paper

Pith reviewed 2026-05-20 07:46 UTC · model grok-4.3

classification 📊 stat.ML cs.LGcs.NAmath.NAmath.PRq-fin.MFstat.CO
keywords diffusion modelsGirsanov theoremparticle filtersequential Monte Carloimportance samplinginference-time guidanceunbiased resampling
0
0 comments X

The pith

URGE performs unbiased path-wise resampling for diffusion guidance by attaching Girsanov multiplicative weights to trajectories and resampling periodically without any score or gradient evaluations.

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

The paper introduces URGE as a derivative-free method that improves sample quality in diffusion generative models at inference time by using a Girsanov change of measure to reweight entire simulated trajectories. Instead of computing gradients or scores for particle weights as in prior work, it attaches a simple multiplicative factor to each path and resamples at intervals. The authors prove an equivalence showing that the path weight admits a backward conditional expectation recovering the exact particle-level weights from sequential Monte Carlo. This guarantees both schemes produce the same unbiased terminal law. Readers would care because the approach eliminates repeated derivative computations while preserving theoretical unbiasedness and delivering better empirical results on benchmarks.

Core claim

The central claim is that path-wise importance reweighting via the Girsanov change of measure is equivalent to particle-wise sequential Monte Carlo for diffusion processes: the Girsanov path weight admits a backward conditional expectation that recovers the previous particle-level weights exactly, so that both schemes produce the same unbiased terminal law. This equivalence underpins URGE, which requires no score, Hessian, or PDE evaluation and is implemented by attaching multiplicative weights to trajectories followed by periodic resampling.

What carries the argument

The Girsanov path weight under a change of measure on diffusion trajectories, which supplies multiplicative importance weights that admit a backward conditional expectation recovering particle weights.

If this is right

  • URGE produces the same unbiased terminal distribution as gradient-based particle filters while requiring only trajectory simulation and simple multiplicative weighting.
  • No score, Hessian, or PDE solves are needed at inference time, removing the main sources of bias and overhead in prior guidance methods.
  • The method applies to any diffusion satisfying the regularity conditions and can be combined with mixture-of-experts or drift adjustments for task-specific objectives.
  • Empirical tests show improved generation quality over existing inference-time baselines on both synthetic tasks and standard diffusion-model benchmarks.

Where Pith is reading between the lines

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

  • The equivalence may allow swapping path-wise and particle-wise implementations interchangeably in other stochastic sampling settings where Girsanov weights can be computed.
  • Because the method is fully gradient-free, it could be integrated into black-box simulators or non-differentiable forward models that still admit a Girsanov representation.
  • Extensions could explore whether approximate Girsanov weights (e.g., via learned estimators) preserve unbiasedness up to controllable error in high dimensions.

Load-bearing premise

The diffusion process and the Girsanov change of measure must satisfy regularity conditions such as the Novikov condition so that the path weights are well-defined and the backward conditional expectation recovers the particle weights exactly.

What would settle it

Running both URGE and a standard particle-wise SMC sampler on the same diffusion model and target objective, then checking whether their empirical terminal distributions differ in total variation or in any moment that the theory predicts must match.

Figures

Figures reproduced from arXiv: 2605.18745 by Lifu Wei, Naichen Shi, Yinuo Ren, Yiping Lu.

Figure 1
Figure 1. Figure 1: Many modern systems come with complemen [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Conceptual description of SURGE for Data As [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Performance comparison between baseline meth [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Performance comparison between baseline meth [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of vorticity field recon [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative and quantitative comparison of [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: More trajectory wise comparison between baselines and SURGE on Lorenz system. The blue SURGE’s trajectory [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: More trajectory wise comparison between baselines and SURGE on Lorenz system. The blue SURGE’s trajectory [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Failure case analysis of the Lorenz system under partial observation. When the trajectory is initialized near [PITH_FULL_IMAGE:figures/full_fig_p023_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Illustration of SURGE behavior under unstable and erroneous predictions from the diffusion surrogate, and [PITH_FULL_IMAGE:figures/full_fig_p023_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: More trajectory wise comparison between baselines and SURGE on Navier-stokes flow in terms of Energy [PITH_FULL_IMAGE:figures/full_fig_p024_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: More trajectory wise comparison between baselines and SURGE on Navier-stokes flow in terms of Energy [PITH_FULL_IMAGE:figures/full_fig_p024_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Impact of Ensemble Averaging. Individual particles (left) exhibit high stochastic variance, whereas the ensemble [PITH_FULL_IMAGE:figures/full_fig_p025_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: More trajectory wise comparison between baselines and SURGE on weather forecasting in terms of VIL [PITH_FULL_IMAGE:figures/full_fig_p025_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: More trajectory wise comparison between baselines and SURGE on weather forecasting in terms of VIL [PITH_FULL_IMAGE:figures/full_fig_p026_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: More trajectory wise comparison between baselines and SURGE on weather forecasting in terms of VIL [PITH_FULL_IMAGE:figures/full_fig_p027_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: More trajectory wise comparison between baselines and SURGE on weather forecasting in terms of VIL [PITH_FULL_IMAGE:figures/full_fig_p028_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Ablation of the guidance term. Without guidance (right), the trajectory fails to correct drift and degrades to [PITH_FULL_IMAGE:figures/full_fig_p028_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Ablation of the reward term. Without reward (right), the trajectory fails to correct drift and degrades to the [PITH_FULL_IMAGE:figures/full_fig_p029_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Ablation of SURGE weight computing and resampling. The trajectory is same as FlowDAS predicted. [PITH_FULL_IMAGE:figures/full_fig_p029_20.png] view at source ↗
read the original abstract

Diffusion-based generative models increasingly rely on inference-time guidance, adding a drift term or reweighting mixture of experts, to improve sample quality on task-specific objectives. However, most existing techniques require repeated score or gradient evaluations, introducing bias, high computational overhead, or both. We introduce \texttt{URGE}, Unbiased Resampling via Girsanov Estimation, a derivative-free inference-time scaling algorithm that performs path-wise importance reweighting via a Girsanov change of measure. Instead of computing gradient-based particle weights in previous work, \texttt{URGE} attaches a simple multiplicative weight to each simulated trajectory and periodically resamples. No score, no Hessian, and no PDE evaluation is required. We establish an equivalence between path-wise and particle-wise SMC: the Girsanov path weight admits a backward conditional expectation that recovers the previous particle-level weights, guaranteeing that both schemes produce the same unbiased terminal law. Empirically, \texttt{URGE} outperforms existing inference-time guidance baselines on synthetic tests and diffusion-model benchmarks, achieving better generation quality, while being significantly simpler to implement and fully gradient-free.

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

1 major / 2 minor

Summary. The manuscript introduces SURGE (also referred to as URGE), an approximation-free, training-free particle filter for diffusion surrogates. It performs path-wise importance reweighting via a Girsanov change of measure, attaching multiplicative weights to simulated trajectories and resampling periodically without any score, gradient, or PDE evaluations. The central claim is an equivalence between path-wise and particle-wise SMC: the Girsanov path weight admits a backward conditional expectation that recovers the previous particle-level weights, guaranteeing both schemes produce the same unbiased terminal law. Empirical results show outperformance over existing inference-time guidance baselines on synthetic tests and diffusion-model benchmarks.

Significance. If the equivalence holds, the work provides a simple gradient-free alternative for inference-time scaling in diffusion models, eliminating bias and overhead from repeated score evaluations. The derivation from established stochastic calculus and the empirical gains are strengths that could influence practical generative modeling pipelines.

major comments (1)
  1. [Theoretical equivalence derivation (continuous-time Girsanov application)] The equivalence between Girsanov path weights and particle-wise SMC weights is derived in the continuous-time semimartingale setting. Diffusion sampling uses discrete-time schemes (Euler–Maruyama or similar) with finite steps; the manuscript does not show that the discrete Radon–Nikodym derivative equals the backward conditional expectation of the continuous Girsanov exponential or bound the resulting O(Δt) discrepancy. This is load-bearing for the exact unbiasedness and approximation-free claims.
minor comments (2)
  1. [Title and abstract] Title uses SURGE while abstract introduces URGE; ensure acronym consistency and expand it on first use.
  2. [Abstract and introduction] Hyphenate 'Training Free' as 'training-free' and check for similar compound-adjective issues throughout.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their careful reading and constructive feedback. Below we address the single major comment point by point, with a commitment to strengthen the presentation of the discrete-time case.

read point-by-point responses

  1. Referee: [Theoretical equivalence derivation (continuous-time Girsanov application)] The equivalence between Girsanov path weights and particle-wise SMC weights is derived in the continuous-time semimartingale setting. Diffusion sampling uses discrete-time schemes (Euler–Maruyama or similar) with finite steps; the manuscript does not show that the discrete Radon–Nikodym derivative equals the backward conditional expectation of the continuous Girsanov exponential or bound the resulting O(Δt) discrepancy. This is load-bearing for the exact unbiasedness and approximation-free claims.

    Authors: We appreciate the referee identifying this important clarification. The continuous-time derivation is presented to exploit standard results from stochastic calculus and to make the connection to Girsanov’s theorem transparent. In the discrete-time setting actually used for sampling, the path-wise weight is exactly the product, over Euler–Maruyama steps, of the Radon–Nikodym derivatives between the two Gaussian transition kernels. This product is the natural discrete counterpart of the continuous Girsanov exponential. By the tower property of conditional expectation, the backward conditional expectation of these discrete weights recovers the particle-wise weights exactly (no additional approximation). The only O(Δt) discrepancy appears when one compares the discrete weights to their continuous-time limit; however, because the underlying diffusion sampler itself is already an O(Δt) approximation, the terminal measure produced by URGE remains unbiased relative to the discrete particle filter that would be obtained by direct particle-wise reweighting. We will add a short subsection (or appendix paragraph) that (i) states the discrete Radon–Nikodym form explicitly, (ii) verifies the tower-property equivalence in discrete time, and (iii) notes that any remaining discretization error is of the same order as the numerical scheme already employed by all competing methods. This revision will be included in the next manuscript version. revision: yes

Circularity Check

0 steps flagged

No circularity: equivalence derived from external Girsanov theorem and conditional expectation identity

full rationale

The paper's central derivation establishes an equivalence between path-wise Girsanov reweighting and particle-wise SMC by invoking the standard Girsanov change of measure and a backward conditional expectation that recovers prior particle weights. This step relies on established stochastic calculus results (Girsanov theorem under Novikov-type regularity) rather than any self-definitional loop, fitted parameter renamed as prediction, or load-bearing self-citation. No equations reduce the claimed unbiased terminal law to the paper's own inputs by construction; the argument is self-contained against external mathematical benchmarks and does not smuggle ansatzes or rename known empirical patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The method rests on the applicability of Girsanov's theorem to the diffusion paths and the existence of the backward conditional expectation that equates path and particle weights. No free parameters or new entities are introduced in the abstract description.

axioms (1)
  • domain assumption The underlying stochastic differential equation satisfies the regularity conditions required for Girsanov's theorem to define a valid change of measure.
    This is needed for the path weights to be well-defined and for the equivalence to hold.

pith-pipeline@v0.9.0 · 5750 in / 1340 out tokens · 39287 ms · 2026-05-20T07:46:38.818768+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

30 extracted references · 30 canonical work pages · 9 internal anchors

  1. [1]

    Albergo, M. S. and Vanden-Eijnden, E. Building nor- malizing flows with stochastic interpolants. arXiv preprint arXiv:2209.15571,

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    Albergo, M. S. and Vanden-Eijnden, E. Nets: A non-equilibrium transport sampler. arXiv preprint arXiv:2410.02711,

    work page arXiv

  3. [3]

    The Ensemble Schr{\"o}dinger Bridge filter for Nonlinear Data Assimilation

    Bao, F. and Sun, H. The ensemble schrödinger bridge filter for nonlinear data assimilation. arXiv preprint arXiv:2512.18928,

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    A score-based non- linear filter for data assimilation

    Bao, F., Zhang, Z., and Zhang, G. A score-based non- linear filter for data assimilation. arXiv preprint arXiv:2306.09282,

    work page arXiv

  5. [5]

    and Han, J

    Bruna, J. and Han, J. Posterior sampling with de- noising oracles via tilted transport. arXiv preprint arXiv:2407.00745,

    work page arXiv

  6. [6]

    Cardoso, G., Idrissi, Y. J. E., Corff, S. L., and Moulines, E. Monte carlo guided diffusion for bayesian linear inverse problems. arXiv preprint arXiv:2308.07983,

    work page arXiv

  7. [7]

    R., Ying, L., and Izzo, Z

    Chen, H., Ren, Y., Min, M. R., Ying, L., and Izzo, Z. Solving inverse problems via diffusion-based priors: An approximation-free ensemble sampling approach. arXiv preprint arXiv:2506.03979, 2025a. Chen, S., Jia, Y., Qu, Q., Sun, H., and Fessler, J. A. Flowdas: A stochastic interpolant-based framework for data assimilation. arXiv preprint arXiv:2501.16642,...

    work page arXiv

  8. [8]

    Split gibbs discrete diffusion posterior sampling

    Chu, W., Wu, Z., Chen, Y., Song, Y., and Yue, Y. Split gibbs discrete diffusion posterior sampling. arXiv preprint arXiv:2503.01161,

    work page arXiv

  9. [9]

    Domingo-Enrich, C., Drozdzal, M., Karrer, B., and Chen, R. T. Adjoint matching: Fine-tuning flow and diffusion generative models with memo- ryless stochastic optimal control. arXiv preprint arXiv:2409.08861,

    work page arXiv

  10. [10]

    Discrete feynman-kac correctors

    Hasan, M., Ohanesian, V., Gazizov, A., Bengio, Y., Aspuru-Guzik, A., Bondesan, R., Skreta, M., and Neklyudov, K. Discrete feynman-kac correctors. arXiv preprint arXiv:2601.10403,

    work page arXiv

  11. [11]

    K., Yan, B., Domingo-Enrich, C., Sriram, A., Wood, B., Levine, D., Hu, B., Amos, B., Karrer, B., et al

    Havens, A., Miller, B. K., Yan, B., Domingo-Enrich, C., Sriram, A., Wood, B., Levine, D., Hu, B., Amos, B., Karrer, B., et al. Adjoint sampling: Highly scal- able diffusion samplers via adjoint matching. arXiv preprint arXiv:2504.11713,

    work page arXiv

  12. [12]

    RNE: plug-and-play diffusion inference-time control and energy-based training

    He, J., Hernández-Lobato, J. M., Du, Y., and Vargas, F. Rne: a plug-and-play framework for diffusion density estimation and inference-time control. arXiv preprint arXiv:2506.05668, 2025a. 10 Title Suppressed Due to Excessive Size He, J., Jeha, P., Potaptchik, P., Zhang, L., Hernández- Lobato, J. M., Du, Y., Syed, S., and Vargas, F. Crepe: Controlling diff...

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    Flow Matching for Generative Modeling

    Lipman, Y., Chen, R. T., Ben-Hamu, H., Nickel, M., and Le, M. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747,

    work page internal anchor Pith review Pith/arXiv arXiv

  14. [14]

    K., and Chen, R

    Liu, G.-H., Choi, J., Chen, Y., Miller, B. K., and Chen, R. T. Adjoint schr \” odinger bridge sampler. arXiv preprint arXiv:2506.22565,

    work page arXiv

  15. [15]

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow

    Liu, X., Gong, C., and Liu, Q. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003,

    work page internal anchor Pith review Pith/arXiv arXiv

  16. [16]

    URL https://journals.ametsoc.org/view/journals/ atsc/20/2/1520-0469_1963_020_0130_dnf_2_0_ co_2.xml

    doi: 10.1175/1520-0469(1963)020<0130:DNF>2.0.CO;2. URL https://journals.ametsoc.org/view/journals/ atsc/20/2/1520-0469_1963_020_0130_dnf_2_0_ co_2.xml. Ma, N., Tong, S., Jia, H., Hu, H., Su, Y.-C., Zhang, M., Yang, X., Li, Y., Jaakkola, T., Jia, X., et al. Inference-time scaling for diffusion mod- els beyond scaling denoising steps. arXiv preprint arXiv:2...

    work page doi:10.1175/1520-0469(1963)020 1963

  17. [17]

    GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models

    Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., and Chen, M. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741,

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [18]

    Large Language Diffusion Models

    Nie, S., Zhu, F., You, Z., Zhang, X., Ou, J., Hu, J., Zhou, J., Lin, Y., Wen, J.-R., and Li, C. Large language diffusion models. arXiv preprint arXiv:2502.09992,

    work page internal anchor Pith review Pith/arXiv arXiv

  19. [19]

    M., and Han, J

    Ren, Y., Gao, W., Ying, L., Rotskoff, G. M., and Han, J. Driftlite: Lightweight drift control for inference- time scaling of diffusion models. arXiv preprint arXiv:2509.21655, 2025a. Ren, Y., Rotskoff, G. M., and Ying, L. A unified ap- proach to analysis and design of denoising markov models. arXiv preprint arXiv:2504.01938, 2025b. Robinson, M., Evans, J....

    work page arXiv

  20. [20]

    S., Domingo-Enrich, C., Boffi, N

    Sabour, A., Albergo, M. S., Domingo-Enrich, C., Boffi, N. M., Fidler, S., Kreis, K., and Vanden-Eijnden, E. Test-time scaling of diffusions with flow maps. arXiv preprint arXiv:2511.22688,

    work page arXiv

  21. [21]

    A general framework for inference-time scaling and steering of diffusion models.arXiv preprint arXiv:2501.06848, 2025

    Singhal, R., Horvitz, Z., Teehan, R., Ren, M., Yu, Z., McKeown, K., and Ranganath, R. A general frame- work for inference-time scaling and steering of diffu- sion models. arXiv preprint arXiv:2501.06848,

    work page arXiv

  22. [22]

    Feynman-kac correctors in diffusion: Annealing, guidance, and product of experts

    Skreta, M., Akhound-Sadegh, T., Ohanesian, V., Bondesan, R., Aspuru-Guzik, A., Doucet, A., Brekelmans, R., Tong, A., and Neklyudov, K. Feynman-kac correctors in diffusion: Annealing, guidance, and product of experts. arXiv preprint arXiv:2503.02819,

    work page arXiv

  23. [23]

    Score-Based Generative Modeling through Stochastic Differential Equations

    Song, Y., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., and Poole, B. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456,

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  24. [24]

    Improving and generalizing flow-based generative models with minibatch optimal transport

    Tong, A., Fatras, K., Malkin, N., Huguet, G., Zhang, Y., Rector-Brooks, J., Wolf, G., and Bengio, Y. Im- proving and generalizing flow-based generative mod- els with minibatch optimal transport. arXiv preprint arXiv:2302.00482,

    work page internal anchor Pith review Pith/arXiv arXiv

  25. [25]

    Inference-time alignment in diffusion models with reward-guided generation: Tutorial and review

    Uehara, M., Zhao, Y., Wang, C., Li, X., Regev, A., Levine, S., and Biancalani, T. Inference-time alignment in diffusion models with reward-guided generation: Tutorial and review. arXiv preprint arXiv:2501.09685,

    work page arXiv

  26. [26]

    Training-free adaptation of diffusion mod- els via doob’s h-transform

    Zhu, Q., Ye, Z., Liu, H., Wang, Z., and Chen, M. Training-free adaptation of diffusion mod- els via doob’s h-transform. arXiv preprint arXiv:2602.16198,

    work page arXiv

  27. [27]

    and Lu, Y

    Zhu, Y. and Lu, Y. On the power of (approximate) reward models for inference-time scaling. arXiv preprint arXiv:2602.01381,

    work page arXiv

  28. [28]

    Then for any integrable test function ϕ, E " 1 N NX i=1 ϕ( ˜X (i)) {(X (j), ˜w(j))}N j=1 # = NX j=1 ˜w(j) ϕ(X (j))

    Let { ˜X (i)}N i=1 be obtained by multinomial resampling from P i ˜w(i)δX (i) and assigning equal weights 1/N. Then for any integrable test function ϕ, E " 1 N NX i=1 ϕ( ˜X (i)) {(X (j), ˜w(j))}N j=1 # = NX j=1 ˜w(j) ϕ(X (j)). Proof. Conditioned on the current weighted particles, the resampling indices are i.i.d. with P(A(i) = j) = ˜w(j) and ˜X (i) = X (A...

    work page 1963

  29. [29]

    SURGE consistently improves both SDA and FlowDAS backbones

    Full results on the Lorenz 1963 experiment. SURGE consistently improves both SDA and FlowDAS backbones. Method RMSE ↓ W1 ↓ BPF (N=20) 0.0625 0 .0448 DM 0.0766 0 .0549 EnKF 0.0624 0 .0448 SDA 0.0589 0 .0426 + SURGE 0.0555 0 .0396 FlowDAS 0.0545 0 .0388 FlowDAS A VG 0.0923 0 .0698 + SURGE 0.0502 0 .0363 Here, E(k) represents the kinetic energy spectrum, cal...

    work page 1963

  30. [30]

    Ensemble Kalman Filter (EnKF) maintains a finite ensemble and applies a Kalman-style update under a Gaussian approximation of the forecast distribution( Evensen, 2003)

    Diffusion Model (DM) refers to a plain diffusion sampler that generates trajectories from the learned prior without observation guidance, included to isolate the contribution of guidance. Ensemble Kalman Filter (EnKF) maintains a finite ensemble and applies a Kalman-style update under a Gaussian approximation of the forecast distribution( Evensen, 2003). ...

    work page 2003