Pith. sign in

REVIEW 3 major objections 3 minor 57 references

Reduced-Order Modeling of Cyclo-Stationary Time Series Using Score-Based Generative Methods

T0 review · 3 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read A clock-augmented generative model turns seasonally forced climate data into an autonomous surrogate that reproduces the original statistics.

desk verdict A useful engineering extension that overclaims its theoretical guarantee; worth peer review but needs the pinned-clock proof fixed. read the letter →

arxiv 2508.19448 v2 pith:WKETUR56 submitted 2025-08-26 nlin.CD stat.ML

classification nlin.CDstat.ML
keywords cyclo-stationarytimeseriesreduced-ordermodelingscore-basedgenerativemodeldenoisingscorematchingFokker-Planckdecompositionstate-spaceaugmentationclimatesurrogatePlaSim
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that cyclo-stationary time series—systems whose statistics vary periodically, like climate under the annual solar cycle—can be reduced to a generative model that preserves the original probabilities without storing time-dependent parameters. The method's trick is to append two harmonic 'clock' coordinates (sine and cosine of the annual cycle) to the physical state, turning the seasonally forced problem into a higher-dimensional autonomous one. A score function is learned by denoising score matching, and a single constant drift matrix is fitted by moment matching after splitting the drift into gradient and circulatory parts. The central claim, Eq. (25), is that if the clock is pinned to true time and only physical coordinates are integrated, the simulated law at every instant is exactly the learned conditional density given the clock phase, making the output cyclo-stationary with the data's periodicity. On the 20 leading principal components of PlaSim surface temperature, the paper reports that marginal and joint PDFs, autocorrelations up to roughly 20 days, and seasonally varying non-Gaussian SST distributions are reproduced.

What carries the argument

The load-bearing object is the augmented state x_aug = (sin ωt, cos ωt, x_phys) ∈ R^{D+2}, carrying a deterministic harmonic 'clock.' In this space the paper estimates the stationary score s = ∇ log pS via denoising score matching and a constant matrix Φ = S + A via moment matching, then pins the clock, integrates only the physical block, and claims the conditional law pS(x_phys | c(t)) is preserved. The proof leverages the Fokker-Planck identity that an antisymmetric drift times the score has zero divergence, so any density whose score is used in both drift and diffusion is stationary.

What would settle it

Run the pinned-clock ROM for many independent 1,000-year trajectories and compute, for each season, the Kolmogorov-Smirnov distance between the model's season-conditioned principal-component distributions and the original PlaSim seasons, plus lag-30 to lag-100 autocorrelations. If the conditional-law claim holds, all should fall within sampling noise; a systematic seasonal bias or a spectral bump would falsify Eq. (25).

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is the state-space augmentation plus pinned-clock integration recipe. In Eqs. (22)–(25), after learning the augmented score and drift, the physical coordinates evolve by dx_phys = Φ_phys s(x_phys, t) dt + √2 Σ_phys dW, with the clock coordinates evaluated deterministically at the true phase; the paper asserts this process has marginal law p(x_phys(t)) = pS(x_phys | c(t)) for all t, so the ROM is cyclo-stationary with the same periodicities as the data. Consequently the surrogate is claimed to inherit the invariant measure conditional on season—hence non-Gaussian seasonal statistics—and, through matching short-lag cross moments, autocorrelations up to t

Load-bearing premise

Integrating only the physical coordinates while freezing the clock at its true phase leaves the conditional distribution pS(x_phys | c(t)) exactly invariant; the paper shows this for the full augmented process, not for the pinned projected process.

Editorial extensions

If this is right

  • If correct, the method gives a parameter-free route to seasonally conditioned generative surrogates: no time-dependent matrices, just one augmented autonomous SDE.
  • Additional periodicities can be stacked by adding more sine/cosine pairs, so combined annual and diurnal forcing is a direct extension.
  • The 20-PC PlaSim ROM reproduces non-Gaussian pointwise SST distributions at all seasons, including bimodality near sea-ice margins, not just the reduced-coordinate statistics.
  • Multi-millennial synthetic integrations are reported stable, with no drift, supporting long-run ensemble and extreme-event studies.
  • Computational cost shifts from weeks of climate simulation to minutes of ROM generation after training.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same augmentation should apply to any periodically forced system—circadian biological rhythms, seasonal economics, rotating machinery—since the method never uses the governing equations.
  • A direct test would be whether the pinned-clock projected SDE actually obeys the conditional Fokker-Planck equation; if cross-terms with the frozen clock coordinates matter, the conditional law may be only approximately preserved, with errors growing with coupling strength.
  • The reported autocorrelation agreement at roughly 20-day lags suggests the score captures the slow manifold; probing multi-year lags or spectral coherence at sub-annual frequencies would show where the approximation decays.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The paper proposes a reduced-order modeling framework for cyclo-stationary time series. It augments the physical state with harmonic 'clock' coordinates to convert the non-autonomous problem into an autonomous one, estimates the score of the augmented stationary density by denoising score matching, and learns a constant drift/diffusion pair by matching one-step cross-moments. In simulation, the clock is pinned to its true harmonic phase and only the physical block is integrated. The method is applied to the 20 leading PCs of seasonally detrended PlaSim surface temperature anomalies, with claims of accurate reproduction of marginal and joint PDFs, autocorrelations up to ~20 days, seasonal modulation, and spatial SST distributions. The central theoretical guarantee is Eq. (25), which asserts that the pinned-clock projected process has the learned conditional law as its marginal at all times.

Significance. If Eq. (25) and the accompanying moment-matching arguments are valid, the paper provides a useful, computationally cheap extension of the score-based decomposition framework of [19] to periodically forced high-dimensional systems. The demonstration on a 20-dimensional climate problem is a meaningful scalability step, and the reported speed-up (centuries of synthetic data in minutes) is attractive. The paper also contains a fairly extensive visual validation, including physical-space reconstruction of non-Gaussian SST distributions. However, the central theoretical guarantee for the actual simulated process is not established, and some of the reported PDF agreement is expected by construction because the score is trained on the data and the SDE drift form is designed to preserve the learned density. The non-trivial empirical content lies in the temporal autocorrelations and spatial reconstructions, which are encouraging but need quantitative backing.

major comments (3)
  1. [Section 2.5, Eq. (25) and Appendix A] The claim that pinning the clock and integrating only the physical block (22) yields p(xphys(t)) = pS(xphys | c(t)) for all t is asserted without proof. Appendix A proves only that the full augmented autonomous SDE (32) has pS as a stationary solution; it does not analyze the non-autonomous, projected SDE (22). The two generators are different: the full generator contains derivatives with respect to the clock coordinates, while the pinned-clock generator has c(t) as a prescribed function. For rho_t(p)=pS(p|c(t)) to solve the projected Fokker-Planck equation, the explicit time derivative of rho_t must be balanced by p-space drift/diffusion terms; no general identity ensures this. A linear-Gaussian example with pS(p|c)=N(alpha c, 1-alpha^2) and Phi=S=I gives an OU process whose mean lags alpha c(t), so Eq. (25) fails exactly. The authors should either prove the conditional invariance under
  2. [Section 2.3, Eqs. (12)-(14)] The moment-matching step is more direct than presented. Using the Stein identity E[s(x)x^T] = -I for densities with vanishing boundary terms, Eq. (13) reduces to Phi = -M = -E[dx/dt x^T], so the drift matrix is essentially fixed by the empirical cross-covariance, and the score does not constrain Phi. Combined with the SDE form (7), any Phi whose symmetric part is S yields the learned density as invariant measure for the full augmented autonomous process, up to score-estimation error. Thus the agreement of marginal and joint PDFs in Section 3.4 is partly by construction and should not be presented as an independent validation. The paper should acknowledge this and place the evidential weight on autocorrelations, spectral properties, and physical-space reconstructions.
  3. [Section 2.5, Eq. (26) and surrounding text] The statement that matching the one-step cross-moment E[dxphys/dt x_aug^T] = Phi_phys E[s_aug x_aug^T] 'yields agreement of the discrete-time power spectra' via the Wiener-Khinchin theorem is too strong. Matching a lag-0 cross-moment does not determine the autocovariance sequence at positive lags for a nonlinear score; the ACF is a property of the full transition kernel. The empirical ACF agreement in Figures 1-2 is encouraging, but the theoretical sentence should be softened to refer to short-lag consistency of the Euler-Maruyama linearization, not full spectral agreement.
minor comments (3)
  1. [Appendix B, Eqs. (45)-(46)] The sign and scaling convention for the denoiser is consistent with Eq. (11) but should be harmonized explicitly, since the notation f_theta vs. b_s can confuse readers. Also, the standardization in Eq. (43) uses M and S without defining them as mean and standard deviation.
  2. [Section 3.4 and Appendix C] The validation is entirely visual. Please report quantitative metrics (e.g., Kolmogorov-Smirnov distances, Jensen-Shannon divergences, RMS ACF errors, and confidence intervals) for the PDF and ACF comparisons. This is particularly important because the claims in the abstract are stated in absolute terms.
  3. [Author list and references] The surname 'Noguiera' should be 'Nogueira'. Several references use placeholders such as 'and co-authors' and 'and colleagues'; these should be completed in the final version.

Circularity Check

3 steps flagged · score 6.0 of 10

Cyclo-stationary claim rests on a self-definitional invariant measure and an unproved slice-invariance assertion; PDFs are reproduced by construction, ACFs only partly fitted.

  1. self definitional [Section 2.1 (Eq. 7); Appendix A (Eqs. 35-37)]
    "Using (34) and s = ∇z log pS, the gradient part cancels and we obtain 0 = −∇z · (A∇zpS). Since ∇z · (A∇zpS) = tr(A∇2zpS) = 0, because the Frobenius inner product of a skew-symmetric matrix with a symmetric matrix is zero, (36) holds identically. Thus, pS is a stationary solution for Approach A provided the learned score approximates the full ∇z log pS."

    The SDE drift is defined as Φs(x) with Φ=S+A and the diffusion as ΣΣ^T=S. For any density pS, substituting s=∇log pS makes the stationary Fokker-Planck equation reduce to tr(A∇²pS)=0, which is an algebraic identity for every skew-symmetric A. Hence the learned density is stationary by construction, not by any dynamical property of the original system. The later validation that the ROM reproduces marginal and joint PDFs is therefore a check of the score estimator against the training density, not an independent prediction of the reduced-order dynamics.

  2. other [Section 2.5, Eq. (25)]
    "In the conditional integration (20)–(22) we pin the clock to the true phase c(t) = (sin ω1t, cos ω1t, . . . ,sin ωN t, cos ωN t) and evolve only the physical block with the full augmented score evaluated on that slice. Consequently, the pair (c(t), xphys(t)) follows the augmented dynamics restricted to {c = c(t)}, and the physical marginal at each time is exactly the learned conditional: p(xphys(t)) = pS(xphys | c(t)), for all t,"

    This is the load-bearing bridge from the augmented autonomous SDE to cyclo-stationarity. Appendix A proves stationarity only for the full augmented SDE (32), where the clock coordinates are dynamical. Equation (22) is a different, non-autonomous SDE for xphys alone with c(t) frozen; its generator is the projection of the full generator onto the physical coordinates. No identity shows that the conditional density pS(p|c(t)) solves the projected Fokker-Planck equation. The phrase 'follows the augmented dynamics restricted to {c=c(t)}' is an assertion, not a derivation. The paper's own Appendix A, in the 'Approach B' paragraph, states that the analogous projected-score/time-dependent strategy 'does not, in general, recover the joint stationary density pS(c,p)', which directly undermines Eq. (

1 more flagged steps
  1. fitted input called prediction [Section 2.3, Eq. (13); Section 2.5, Eq. (26)]
    "This results in the linear system M ≈ ΦV⊤, which is solved for Φ in the least-squares sense using the Moore-Penrose pseudoinverse: bΦ = M (V⊤)+."

    The drift matrix is the least-squares solution matching the data's cross-moment M = E[ẋx^T]. Thus the short-time conditional drift of the reduced model is fitted to the same time series whose autocorrelations are later reported as successful agreement. The paper states the cross-moment identity 'holds by construction from (16)–(17) and (21)', meaning the one-step correlation statistics are inputs, not outputs. Consequently, the short-lag part of the reported autocorrelation agreement is a restatement of the fit; only longer-lag autocorrelations and spatial coherence carry genuinely independent predictive content.

full rationale

The mathematical core of the method guarantees, via the identity tr(A∇²pS)=0, that any density whose score is inserted into the drift is the stationary law of the augmented SDE. The marginal/joint PDF matches in Sections 3.4 and 3.5 are therefore in-sample checks of the score estimator, not independent confirmations of reduced-order physics. The drift matrix is additionally solved from the data cross-moment via Eq. (13), so short-lag correlation statistics are fitted inputs. The pivotal cyclo-stationarity claim Eq. (25) is asserted rather than proved: Appendix A establishes stationarity only for the full augmented autonomous SDE, while the pinned-clock SDE (22) is a projected, non-autonomous process; Appendix A's own 'Approach B' paragraph warns that such projected/time-dependent strategies do not in general recover pS(c,p). Thus the cyclic conditional law is not derived, and the 'by construction' language transfers the trivial invariant-measure property to a different generator without proof. Overall the paper is partially circular: the PDF reproduction is definitional, the cyclic conditional law is assumed, and only the longer-lag autocorrelation and spatial-coherence results provide independent empirical content.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The method rests on the score-linear SDE ansatz from [19], the constant-circulation approximation, and an unproven projected-invariance assertion. The score and drift are fitted to the same data, so the ledger is small in invented entities but substantial in fitted assumptions.

free parameters (3)
  • Number of principal components retained = 20 (about 60% of anomaly variance)
    State dimension of the reduced model; no sensitivity analysis reported.
  • Score estimation noise level sigma_G = 0.05 (standardized units)
    Controls the accuracy of the denoising score match and the scale of the smoothed density; no sensitivity study.
  • Score network and training hyperparameters = MLP [22,128,64,22], 40 epochs, batch 32, Adam lr 2e-4
    Affect score fidelity and therefore all generated statistics; reported but not varied.
assumptions (5)
  • domain assumption The true dynamics can be approximated by an SDE with drift linear in the exact score, F(x) approximately (Sigma Sigma^T + Phi_A) grad log pS(x)
    Section 2.1, Eqs. (6)-(7), inherited from [19].
  • domain assumption The state-dependent circulatory tensor R(x) can be replaced by a constant antisymmetric matrix Phi_A
    Section 2.1, after Eq. (5).
  • ad hoc to paper The pinned-clock projected SDE has pS(xphys | c(t)) as an invariant measure
    Eq. (25); Appendix A proves the full augmented SDE invariant, not the projected process.
  • domain assumption Denoising score matching with sigma_G = 0.05 yields a score close to the true grad log pS
    Section 2.2 and Appendix B.
  • standard math Time averages over the 1000-year run converge to expectations under the stationary law
    Used in Eqs. (12)-(13).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reduced-Order Modeling of Cyclo-Stationary Time Series Using Score-Based Generative Methods." pith.science (2026). https://pith.science/paper/WKETUR56

@misc{pith2026250819448,
  author       = {Pith},
  title        = {Pith review of: Reduced-Order Modeling of Cyclo-Stationary Time Series Using Score-Based Generative Methods},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WKETUR56}},
  note         = {Machine review of arXiv:2508.19448}
}
read the original abstract

Many natural systems exhibit cyclo-stationary behavior characterized by periodic forcing such as annual and diurnal cycles. We present a data-driven method leveraging recent advances in score-based generative modeling to construct reduced-order models for such cyclo-stationary time series. Our approach accurately reproduces the statistical properties and temporal correlations of the original data, enabling efficient generation of synthetic trajectories. We demonstrate the performance of the method through application to the Planet Simulator (PlaSim) climate model, constructing a reduced-order model for the 20 leading principal components of surface temperature driven by the annual cycle. The resulting surrogate model accurately reproduces the marginal and joint probability distributions, autocorrelation functions, and spatial coherence of the original climate system across multiple validation metrics. The approach offers substantial computational advantages, enabling generation of centuries of synthetic climate data in minutes compared to weeks required for equivalent full model simulations. This work opens new possibilities for efficient modeling of periodically forced systems across diverse scientific domains, providing a principled framework for balancing computational efficiency with physical fidelity in reduced-order modeling applications.

Figures

Figures reproduced from arXiv: 2508.19448 by the authors.

Figure 1
Figure 1. Validation of the PlaSim ROM for principal components 1-10. Each row corresponds [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. Validation of the PlaSim ROM for principal components 11-20. The layout is identical [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. SST PDFs — 4 points × 4 seasons. Columns correspond to four grid points (titles give (i, j)); rows correspond to seasons (DJF, MAM, JJA, SON). Each panel shows KDEs of SST reconstructed from: Obs (all PCs) (black), Obs (20 PCs) (red, dashed), and Model (20 PCs) (blue). Observed reconstructions use the day-of-year climatology plus spatial mean plus EOF back-projection; model curves use the same back-projection applie… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Bivariate PDFs between principal components 1–10 (rows) and principal components 1, 2, and 3 (columns). Each pair of panels shows the ROM-generated dis￾tributions (blue) and the PlaSim distributions (red). The strong agreement across all modes demonstrates that the red…
Figure 5
Figure 5. Figure 5: Bivariate PDFs between principal components 11–20 (rows) and principal components 1, 2, and 3 (columns). Layout is the same as [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 46 canonical work pages

  1. [19]

    Giorgini

    Ludovico T. Giorgini. Data-driven decomposition of conservative and non-conservative dynamics in multiscale systems. arXiv preprint arXiv:2505.01895 , 2025

  2. [1]

    Perez, and Michael Tippett

    Matthew Alexander, Alejandro J. Perez, and Michael Tippett. Impact of annual cycle on enso variability and predictability. Journal of Climate , 34(1):183–202, 2021. doi: 10.1175/ JCLI-D-19-0291.1

  3. [2]

    Alexander, Ludmila Matrosova, C´ ecile Penland, James D

    Michael A. Alexander, Ludmila Matrosova, C´ ecile Penland, James D. Scott, and Prashant Chang. Forecasting pacific ssts: Linear inverse model predictions of the pdo. Journal of Climate, 21:1401–1417, 2008. doi: 10.1175/2007JCLI1849.1

  4. [3]

    Anchukaitis and co-authors

    Kevin J. Anchukaitis and co-authors. Large ensemble particle filter for spatial climate re- constructions using a linear inverse model. Journal of Advances in Modeling Earth Systems, 15:e2022MS003094, 2023. doi: 10.1029/2022MS003094

  5. [4]

    Potential in stochastic differential equations: novel construction

    Ping Ao. Potential in stochastic differential equations: novel construction. Journal of Physics A: Mathematical and General , 37(3):L25–L30, 2004

  6. [5]

    Understanding causation via corre- lations and linear response theory

    Marco Baldovin, Fabio Cecconi, and Angelo Vulpiani. Understanding causation via corre- lations and linear response theory. Physical Review Research, 2(4):043436, 2020. 15 Figure 4: Bivariate PDFs between principal components 1–10 (rows) and principal components 1, 2, and 3 (columns). Each pair of panels shows the ROM-generated dis- tributions (blue) and t...

  7. [6]

    Extracting causation from millennial-scale climate fluctuations in the last 800 kyr

    Marco Baldovin, Fabio Cecconi, Antonello Provenzale, and Angelo Vulpiani. Extracting causation from millennial-scale climate fluctuations in the last 800 kyr. Scientific Reports, 12(1):15320, 2022

  8. [7]

    Kolmogorov modes and linear re- sponse of jump-diffusion models: Applications to stochastic excitation of the enso recharge oscillator

    Micka¨ el D Chekroun, Niccol` o Zagli, and Valerio Lucarini. Kolmogorov modes and linear re- sponse of jump-diffusion models: Applications to stochastic excitation of the enso recharge oscillator. arXiv preprint arXiv:2411.14769 , 2024

Show all 57 references
  1. [8]

    Climate sensitivity via a nonparametric fluctuation–dissipation theorem

    Fenwick C Cooper and Peter H Haynes. Climate sensitivity via a nonparametric fluctuation–dissipation theorem. Journal of the Atmospheric Sciences, 68(5):937–953, 2011

  2. [9]

    Cox, Chris Huntingford, and Mark S

    Peter M. Cox, Chris Huntingford, and Mark S. Williamson. Emergent constraint on equi- 16 Figure 5: Bivariate PDFs between principal components 11–20 (rows) and principal components 1, 2, and 3 (columns). Layout is the same as Figure 4. Agreement across these higher-order compo...

  3. [10]

    The heterogeneous multiscale method: A review

    Weinan E, Bj¨ orn Engquist, Xiantao Li, Weiqing Ren, and Eric Vanden-Eijnden. The heterogeneous multiscale method: A review. Communications in Computational Physics , 2(3):367–450, 2007

  4. [11]

    Falasca, A

    F. Falasca, A. Basinski, L. Zanna, and M. Zhao. A fluctuation-dissipation theorem per- spective on radiative responses to temperature perturbations. Journal of Climate , 38: 3239–3260, 2025. doi: https://doi.org/10.1175/JCLI-D-24-0479.1

  5. [12]

    Neural models of multiscale systems: conceptual limitations, stochastic parametrizations, and a climate application

    Fabrizio Falasca. Neural models of multiscale systems: conceptual limitations, stochastic parametrizations, and a climate application. arXiv preprint arXiv:2506.22552 , 2025. 17

  6. [13]

    Data-driven dimensionality reduction and causal inference for spatiotemporal climate fields

    Fabrizio Falasca, Pavel Perezhogin, and Laure Zanna. Data-driven dimensionality reduction and causal inference for spatiotemporal climate fields. Physical Review E , 109(4):044202, 2024

  7. [14]

    The planet simulator: towards a user friendly model

    Klaus Fraedrich, Edilbert Kirk, Ute Luksch, and Frank Lunkeit. The planet simulator: towards a user friendly model. Meteorologische Zeitschrift, 14(3):299–304, 2005. doi: 10. 1127/0941-2948/2005/0043

  8. [15]

    Modes of variability and trends in atmospheric constituents

    Hella Garny, Thomas Birner, Veronika Eyring, and Peter Braesicke. Modes of variability and trends in atmospheric constituents. Journal of Climate , 33:2927–2945, 2020. doi: 10.1175/JCLI-D-19-0542.1

  9. [16]

    The evolving butterfly: Statis- tics in a changing attractor

    Gosha Geogdzhayev, Andre N Souza, and Raffaele Ferrari. The evolving butterfly: Statis- tics in a changing attractor. Physica D: Nonlinear Phenomena , page 134107, 2024

  10. [17]

    Boris Gershgorin and Andrew J. Majda. A test model for fluctuation-dissipation theorems with time-periodic statistics. Physica D: Nonlinear Phenomena , 239(17):1741–1757, 2010. ISSN 0167-2789. doi: https://doi.org/10.1016/j.physd.2010.05.009. URL https://www. sciencedirect.com...

  11. [18]

    The physics of climate variability and climate change

    Michael Ghil and Valerio Lucarini. The physics of climate variability and climate change. Reviews of Modern Physics , 92(3):035002, 2020

  12. [20]

    Giorgini, Andre N

    Ludovico T. Giorgini, Andre N. Souza, and Peter J. Schmid. Reduced markovian models of dynamical systems. Physica D: Non-linear Phenomena , 470:134393, 2024

  13. [21]

    Giorgini, T

    Ludovico T. Giorgini, T. Bischoff, and Andre N. Souza. Kgmm: A k-means clustering approach to gaussian mixture modeling for score function estimation. arXiv preprint arXiv:2503.18054, 2025

  14. [22]

    Predicting forced responses of probability distributions via the fluctuation-dissipation theorem and generative modeling

    Ludovico T Giorgini, Fabrizio Falasca, and Andre N Souza. Predicting forced responses of probability distributions via the fluctuation-dissipation theorem and generative modeling. arXiv preprint arXiv:2504.13333 , 2025

  15. [23]

    Learning dissipation and instability fields from chaotic dynamics

    Ludovico T Giorgini, Andre N Souza, Domenico Lippolis, Predrag Cvitanovi´ c, and Peter Schmid. Learning dissipation and instability fields from chaotic dynamics. arXiv preprint arXiv:2502.03456, 2025

  16. [24]

    Non-gaussian stochastic dynamical model for the el ni˜ no southern oscillation.Physical Review Research, 4(2):L022065, 2022

    Ludovico Theo Giorgini, Woosok Moon, Nan Chen, and John S Wettlaufer. Non-gaussian stochastic dynamical model for the el ni˜ no southern oscillation.Physical Review Research, 4(2):L022065, 2022

  17. [25]

    Response theory via generative score modeling

    Ludovico Theo Giorgini, Katherine Deck, Tobias Bischoff, and Andre Souza. Response theory via generative score modeling. Physical Review Letters, 133(26):267302, 2024

  18. [26]

    Subseasonal prediction with an icosahedral atmospheric model

    Michael Godwin and Timothy DelSole. Subseasonal prediction with an icosahedral atmospheric model. Monthly Weather Review , 149:2851–2869, 2021. doi: 10.1175/ MWR-D-20-0266.1

  19. [27]

    Gritsun and G

    A. Gritsun and G. Branstator. Climate response using a three-dimensional operator based on the fluctuation–dissipation theorem. Journal of The Atmospheric Science , page 2558–2575, 2007. doi: https://doi.org/10.1175/JAS3943.1. 18

  20. [28]

    On some aspects of the response to stochas- tic and deterministic forcings

    Manuel Santos Guti´ errez and Valerio Lucarini. On some aspects of the response to stochas- tic and deterministic forcings. Journal of Physics A: Mathematical and Theoretical, 55(42): 425002, 2022

  21. [29]

    Hasselmann

    K. Hasselmann. Stochastic climate models part i. theory. Tellus, 28:473, 1976

  22. [30]

    Holden, Neil R

    Philip B. Holden, Neil R. Edwards, Klaus Fraedrich, Edilbert Kirk, Frank Lunkeit, and Xiuhua Zhu. Plasim-genie v1.0: a new intermediate complexity aogcm. Geoscientific Model Development, 9:3347–3361, 2016. doi: 10.5194/gmd-9-3347-2016

  23. [31]

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

    Aapo Hyv¨ arinen. Estimation of non-normalized statistical models by score matching.Jour- nal of Machine Learning Research , 6:695–709, 2005

  24. [32]

    Some extensions of score matching

    Aapo Hyv¨ arinen. Some extensions of score matching. In Conference on Independent Com- ponent Analysis and Signal Separation , pages 88–95, 2007

  25. [33]

    N. D. B. Keyes, Ludovico T. Giorgini, and J. S. Wettlaufer. Stochastic paleoclimatology: Modeling the epica ice core climate records. Chaos, 33:093132, 2023

  26. [34]

    Bivariate decomposition of irreversible diffusion processes

    Xing Liang et al. Bivariate decomposition of irreversible diffusion processes. Journal of Mathematical Physics, 63:053302, 2022. doi: 10.1063/5.0092468

  27. [35]

    E. N. Lorenz. Deterministic nonperiodic flow. Journal of the Atmospheric Sciences , 20(2): 130–141, 1963

  28. [36]

    The challenge of non-markovian energy balance models in climate

    Shaun Lovejoy and colleagues. The challenge of non-markovian energy balance models in climate. Chaos, 34(7):072105, 2024. doi: 10.1063/5.0197663

  29. [37]

    Lucarini and M.D

    V. Lucarini and M.D. Chekroun. Theoretical tools for understanding the climate crisis from Hasselmann’s programme and beyond. Nat Rev Phys (2023) , 2023. URL https: //doi.org/10.1038/s42254-023-00650-8

  30. [38]

    Interpretable and equation-free response theory for complex systems

    Valerio Lucarini. Interpretable and equation-free response theory for complex systems. arXiv preprint arXiv:2502.07908 , 2025

  31. [39]

    Storm-track organization and variability in a simplified atmospheric global circulation model

    Frank Lunkeit, Silke Borth, Klaus Fraedrich, and Edilbert Kirk. Storm-track organization and variability in a simplified atmospheric global circulation model. Quarterly Journal of the Royal Meteorological Society, 137:1051–1063, 2011. doi: 10.1002/qj.807

  32. [40]

    A. J. Majda and J. Harlim. Filtering Complex Turbulent Systems . Cambridge University Press, 2012

  33. [41]

    A. J. Majda, I. Timofeyev, and E. Vanden-Eijnden. Models for stochastic climate predic- tion. Proc. Natl. Acad. Sci. USA , 96:14687, 1999

  34. [42]

    A. J. Majda, R. V. Abramov, and M. J. Grote. Information Theory and Stochastics for Multiscale Nonlinear Systems . CRM Monograph Series, American Mathematical Society, 2005

  35. [43]

    An applied mathematics perspective on stochastic modelling for climate

    Andrew J Majda, Christian Franzke, and Boualem Khouider. An applied mathematics perspective on stochastic modelling for climate. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences , 366(1875):2427–2453, 2008

  36. [44]

    Moon and J

    W. Moon and J. S. Wettlaufer. Coupling functions in climate. Phil. Trans. R. Soc. A, 377: 20190006, 2019. 19

  37. [45]

    Compo, and Michael A

    Matthew Newman, Gilbert P. Compo, and Michael A. Alexander. Enso-forced variability of the pacific decadal oscillation. Journal of Climate , 16:3853–3857, 2003. doi: 10.1175/ 1520-0442(2003)016⟨3853:EVOTPD⟩2.0.CO;2

  38. [46]

    T. N. Palmer. Predicting uncertainty in forecasts of weather and climate. Reports on Progress in Physics, 63(2):71–116, 2000

  39. [47]

    Sardeshmukh

    C´ ecile Penland and Prashant D. Sardeshmukh. The optimal growth of tropical sea sur- face temperature anomalies. Journal of Climate , 8:1999–2024, 1995. doi: 10.1175/ 1520-0442(1995)008⟨1999:TOGOTS⟩2.0.CO;2

  40. [48]

    The role of the annual cycle in tropi- cal cyclone seasonal forecasting

    Kunj Salvi, Sulagna Ray, and Anupam Sobel. The role of the annual cycle in tropi- cal cyclone seasonal forecasting. Journal of Climate , 35:4593–4605, 2022. doi: 10.1175/ JCLI-D-21-0922.1

  41. [49]

    Schneider, S

    T. Schneider, S. Lan, A. Stuart, and J. Teixeira. Earth system modeling 2.0: A blueprint for models that learn from observations and targeted high-resolution simulations. Geophysical Research Letters, 44(24):12–396, 2017

  42. [50]

    Generative modeling by estimating gradients of the data distribution

    Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems , 32, 2019

  43. [51]

    What’s the score? automated denoising score matching for nonlinear diffusions

    Yang Song et al. What’s the score? automated denoising score matching for nonlinear diffusions. arXiv preprint arXiv:2407.07998 , 2024

  44. [52]

    Andre N. Souza. Representing turbulent statistics with partitions of state space. part 1. theory and methodology. Journal of Fluid Mechanics , 997:A1, 2024. doi: 10.1017/jfm. 2024.658

  45. [53]

    Statistical non-locality of dynamically coherent structures

    Andre N Souza, Tyler Lutz, and Glenn R Flierl. Statistical non-locality of dynamically coherent structures. Journal of Fluid Mechanics , 966:A44, 2023

  46. [54]

    A sta- tistical emulator design for averaged climate fields

    Andre Nogueira Souza, Gosha Geogdzhayev, Raffaele Ferrari, and Glenn R Flierl. A sta- tistical emulator design for averaged climate fields. Authorea Preprints, 2024

  47. [55]

    Modeling climate diversity, tidal dynamics and the fate of volatiles on trappist-1 planets

    Martin Turbet, Guillaume Boucher, Carl-Fredrik Karlen, Gabriella Gilli, Franck Selsis, Franck Montmessin, and Fran¸ cois Forget. Modeling climate diversity, tidal dynamics and the fate of volatiles on trappist-1 planets. Astronomy & Astrophysics , 654:A52, 2021. doi: 10.1051/0...

  48. [56]

    A connection between score matching and denoising autoencoders

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

  49. [57]

    A proof of concept for scale-adaptive parametriza- tions: The case of the Lorenz’96 model

    Gabriele Vissio and Valerio Lucarini. A proof of concept for scale-adaptive parametriza- tions: The case of the Lorenz’96 model. Quarterly Journal of the Royal Meteorological Society, 144(710):63–75, 2018. 20

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.