Pith. sign in

REVIEW 3 major objections 6 minor 52 references

Dynamical low-rank filters can track high-dimensional SDE data-assimilation problems by jointly minimizing mean and covariance error while letting the reduced subspace react to observations.

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 →

T0 review · grok-4.5

2026-07-31 01:04 UTC pith:LFTJPT5V

load-bearing objection Solid methods paper: new joint mean-covariance DLRA filters with usable algorithms and supportive numerics; the observation-aware subspace claim rests on a modeling ansatz that needs clearer limits. the 3 major comments →

arxiv 2607.27432 v1 pith:LFTJPT5V submitted 2026-07-29 math.NA cs.NA

Dynamical Low-Rank Filters for Data Assimilation

classification math.NA cs.NA MSC 65C3065F5562M2093E11
keywords dynamical low-rank approximationdata assimilationstochastic differential equationsKalman–Bucy filterensemble methodsparticle filterreduced-order modeling
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.

Data assimilation for systems governed by stochastic differential equations is often blocked by dimension: the state is huge, the observations are noisy, and fixed reduced bases cannot keep up with turbulent dynamics. This paper builds filters whose state lives in a time-evolving low-rank factorization—mean plus orthonormal basis times stochastic coefficients—updated entirely on the fly. The central construction chooses each discrete update so that the joint root-mean-square error in mean and covariance is minimized, then passes to continuous-time equations that specialize to a reduced Kalman–Bucy filter for linear drift and to ensemble schemes for nonlinear drift. A complementary-space correction lets the reduced subspace feel the observation operator, not only the dynamics; a particle-style variant adds branching on likelihood weights for non-Gaussian settings. Numerics on linear advection–diffusion–reaction, Lorenz systems, and a two-layer quasi-geostrophic model show that small ensembles already match much larger full-order filters when the effective rank is modest.

Core claim

A discrete dynamical low-rank update that minimizes the joint Euclidean-plus-Frobenius error between the surrogate’s mean and covariance and the one-step filtered full-order moments yields continuous reduced equations for mean, basis, and stochastic coefficients. Those equations extend to a reduced Kalman–Bucy filter under linear drift, to ensemble methods under nonlinear drift, and—with a transported complementary measure—allow the basis itself to evolve under the observation operator; a particle version with optional complementary noise further handles nonlinear non-Gaussian regimes.

What carries the argument

The DLRA-JMCO (joint mean-covariance optimal) filter: at each time step the increments of mean, orthonormal basis, and stochastic coefficients are chosen to minimize the sum of squared mean error and squared Frobenius covariance error subject to the discrete gauge condition, producing the continuous system (28)–(31) and its prediction–analysis staggered discretization.

Load-bearing premise

Mass living outside the current low-rank subspace is assumed to be only the initial orthogonal piece transported forward in time; if the true law grows new orthogonal structure, the observation-aware basis update is no longer justified.

What would settle it

On a problem whose true filtered covariance rapidly develops mass in directions orthogonal to both the initial condition and the current dynamics (for example a sudden rotation of the observation operator onto a previously silent subspace), compare the JMCO complementary filter’s rank-k subspace and mean/covariance errors against a full-order particle or ensemble filter; systematic failure of the complementary update to capture the new directions would refute the claim.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Linear-drift continuous-observation problems admit a cheap reduced Kalman–Bucy system whose covariance lives only in the current range projector.
  • The same prediction–analysis split works for discrete-time observations by skipping the analysis step between assimilation times.
  • Ensemble and particle realizations of the filter remain interacting particle systems of size equal to the rank, not the ambient dimension.
  • When the observation operator is misaligned with the state subspace, the complementary correction supplies the missing directions without leaving the low-rank manifold.
  • Quasi-geostrophic and Lorenz-type experiments indicate that tens to a few thousand particles already recover full-order accuracy once the rank covers attractor plus noise dimension.

Where Pith is reading between the lines

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

  • The same joint-moment derivation could be repeated with Wasserstein or other geometry-aware costs, which the authors flag as future work and which would change the optimality conditions for the basis.
  • Merging the on-the-fly DLRA prediction step with triangular transport maps in the analysis step would give a fully online nonlinear reduced filter whose subspace still evolves with the data.
  • If the complementary ansatz is too strong, a cheap online rank-adaptive switch (monitor residual energy in the orthogonal complement) could restore consistency without abandoning the low-rank structure.

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 / 6 minor

Summary. The paper develops dynamical low-rank (DLRA) filters for continuous- and discrete-time data assimilation of SDEs. The central construction (DLRA-JMCO) obtains discrete updates for the mean, Stiefel basis U, and stochastic coefficients Y by minimizing the joint mean-plus-covariance RMSE subject to the discrete gauge condition, then passes to continuous-time equations (28)–(31). For affine drift this specializes to a reduced Kalman–Bucy system (41)–(42) that includes a diffusion contribution in the basis evolution. A complementary ansatz (Sec. 3.6) transports the initial orthogonal mass so that the observation operator enters the U-update. The same framework yields ensemble methods and a preliminary DLRA particle filter, optionally enriched by isotropic complementary noise scaled by a local projection residual ε. Algorithms are staggered projector-splitting schemes; numerics on a linear advection–diffusion–reaction problem, Lorenz-63/96, and a two-layer quasi-geostrophic system (d=2048) compare against particle and ensemble Kalman filters.

Significance. If the derivations and numerics hold, the work supplies a coherent on-the-fly reduced filtering toolkit that couples DLRA for SDEs with classical Kalman/particle analysis steps, including an observation-aware basis update and a high-dimensional QG demonstration. Strengths include fully written discrete-to-continuous derivations with appendix proofs (Lemma 3.1, Prop. 3.2, Prop. 3.6, Lemmas 4.1–4.2), a clear linear KBF reduction, and reproducible-style algorithmic statements (Algorithms 1–2). The contribution is incremental relative to the authors’ prior DLRA-SDE and DLRA-KBF line, but the joint mean–covariance optimality criterion and the complementary/particle extensions are concrete and of practical interest for high-dimensional DA.

major comments (3)
  1. [Section 3.6, Proposition 3.6] Sec. 3.6, Eqs. (43)–(45) and Prop. 3.6: the claim that the main subspace evolves under the observation operator rests on the ansatz that orthogonal mass is pure transport of the initial complement ξ_t = P^⊥_{U_t} ξ_0 for all time, so C_ξ remains P^⊥ C_ξ0 P^⊥. No analysis or diagnostic is given for when the true law develops new orthogonal structure (e.g., Lorenz-96 or QG turbulence). The observation-aware ΔU term and the adaptivity claim are therefore conditional on an untested modeling assumption. Either justify the ansatz (error bound, invariant-subspace argument) or clearly demote it to a heuristic and show a controlled experiment where the ansatz is violated.
  2. [Section 4.1, Assumption 1, Lemma 4.2] Sec. 4.1, Assumption 1, Lemma 4.2, Eq. (52): the complementary DLRA particle filter enriches the range by isotropic Gaussian noise of variance Λ_n built from a local ε residual. The local error bound (49) is an a priori comparison to Euler–Maruyama under the ε-assumption; it does not establish that isotropic noise of that scale recovers the missing observation-relevant directions, nor that the subsequent reduced SVD preserves filter consistency. Numerics (Figs. 8–11) show modest gains for DLR Compl. PF, but without a sensitivity study in ε/Λ or a non-isotropic alternative the enrichment mechanism remains heuristic. Tighten the claim to “preliminary heuristic enrichment” or add supporting analysis/diagnostics.
  3. [Sections 3.3–3.5, Remark 3.5] Well-posedness and stability of the continuous JMCO filter (28)–(31) and of the staggered Algorithm 1 are not established in this manuscript; Remark 3.5 defers to [37] for the linear case and the nonlinear/particle settings are left open. For a methods paper whose central claim is a new filter, at least a local well-posedness sketch or a clear pointer to which hypotheses of [20,21,37] carry over (invertibility of C_Y, non-degeneracy after analysis, discrete-observation jumps) is needed so that the continuous limit and the semi-implicit analysis step are not formal only.
minor comments (6)
  1. [Section 2] Notation for projectors and ranks is overloaded (P_U vs P_{U(X)}, St(k,d) defined with orthonormal rows while many DLRA papers use orthonormal columns). A short notation table would help.
  2. [Algorithm 1] Algorithm 1 line 3 writes E[˚A(...)] for the mean update but (33) and (28) use E[A]; centered vs uncentered drift should be consistent.
  3. [References] Several arXiv preprints in the bibliography carry future-dated identifiers (e.g., 2606.15843, 2601.21428). Ensure citation keys and years match the public record at submission.
  4. [Section 5] Figures 1–12 report relative averaged RMSE but do not state whether error bars or multiple independent observation paths were used; a brief reproducibility note (seeds, number of Z realizations) would strengthen the numerics section.
  5. Typos and grammar: “the following the following SDE” (Sec. 2); “defection of an additive noise” (Sec. 5.2); “Journal od Athmospheric” in [29]; occasional missing articles.
  6. [Remark 3.3] Remark 3.3 notes that U is independent of observations in JMCO; this is important and could be flagged earlier when motivating Sec. 3.6.

Circularity Check

0 steps flagged

No significant circularity: JMCO updates are constrained minimizers of mean/covariance mismatch; self-citations supply ambient DLRA calculus, not the filter identities.

full rationale

The central derivation (Sec. 3) obtains discrete DLRA updates by minimizing the joint mean-plus-covariance RMSE (21)–(22) between the rank-k surrogate and one Euler–Maruyama step of the full-order filter, subject to the discrete gauge and first-order truncation; continuous equations (28)–(31), the linear KBF specialization (41)–(42), and Algorithm 1 follow by taking Δt→0 and staggered discretization. That chain is variational and self-contained: it does not fit free constants to the reported errors, nor does it invoke a uniqueness theorem that forces the filter form. The particle-filter branch (Sec. 4) likewise uses standard likelihood weights and branching on a DLRA prediction. Heavy self-citation to the authors’ DLRA-for-SDEs series ([20]–[24], [37]) supplies the ambient DO/projector-splitting calculus and a comparison baseline, but those works do not encode the JMCO optimality conditions or the filter identities. The Sec. 3.6 transported-orthogonal-mass ansatz and the isotropic Λ_n noise from Assumption 1/Lemma 4.2 are explicit modeling choices that underwrite observation-aware subspace motion and complementary PF enrichment; they are assumptions, not circular reductions of a claimed prediction to its inputs. Numerics are external checks, not fitted targets renamed as predictions. Overall circularity is negligible.

Axiom & Free-Parameter Ledger

4 free parameters · 7 axioms · 3 invented entities

The work sits on standard SDE/filtering calculus plus the authors' DLRA-for-SDEs representation. Load-bearing extras are the joint RMSE objective, the discrete gauge, the transported-orthogonal-mass ansatz, and the ε-local-error isotropic complement for the particle filter. Rank k, sample size M, and assimilation step are chosen per experiment rather than derived.

free parameters (4)
  • reduced rank k = problem-dependent (2-32 in tests)
    Chosen per experiment (e.g. k=20 ADR, k=2 Lorenz-63, k=12 Lorenz-96, k=32 QG) to match noise rank or attractor dimension; central accuracy claims depend on this choice.
  • ensemble size M for DLRA filters = tens to thousands in plots
    User-chosen Monte Carlo size; convergence plots vs M are empirical.
  • assimilation step Δt_a / model step Δt = typically 10Δt
    Discrete-observation schedule (often 10Δt) chosen by hand; affects ESS and reported errors.
  • local ε_n / Λ_n complementary noise scale = computed each step from (50)
    Estimated on the fly from projected drift/diffusion residual (50-51); magnitude directly sets complementary PF exploration.
axioms (7)
  • domain assumption Itô SDE filtering model with additive state/observation noise and independent W, B, X_0 (Eq. 1).
    Standing problem class throughout; multiplicative noise only mentioned as extendable.
  • domain assumption DLRA/DO representation X = m + U^T Y with U in Stiefel rows, gauge ΔU U^T = 0, and McKean-Vlasov projectors (Section 2.1, Eqs. 4-6).
    Imported from prior DLRA-for-SDEs theory [20]; all filters live on this manifold.
  • standard math Euler-Maruyama one-step full-order predictor plus Gaussian conditional mean/covariance formulas yield the local targets m_{n+1}, C_{n+1} (Lemma 3.1).
    Classical Kalman update under local Gaussianity of EM increments.
  • ad hoc to paper Joint objective |Δm|^2 + ||ΔC||_F^2 with discrete gauge is the right optimality criterion for the reduced filter (Eq. 21-22).
    Design choice inspired by matrix DLRA [26]; not forced by Bayes optimality in general.
  • ad hoc to paper Orthogonal complement mass is pure transport of initial P^⊥_{U_0} X_0 for all time (Eq. 43-45).
    Section 3.6 ansatz enabling H-dependent ΔU; Remark 3.7 calls it reasonable for low-rank dynamics but unproved in general.
  • ad hoc to paper ε-error bound controlling projected drift/diffusion residual (Assumption 1) and isotropic Gaussian complement of variance Λ_n (Eq. 52).
    Used to justify cheap complementary PF enrichment; ε estimated locally, isotropy assumed without structure.
  • domain assumption Branching particle filter with likelihood weights converges at Monte Carlo rate in the full-order case [2]; reduced version inherits the same analysis step.
    Section 4 cites Bain-Crisan; reduced convergence not proved here.
invented entities (3)
  • DLRA-JMCO filter (joint mean-covariance optimal DLR filter) no independent evidence
    purpose: Name the reduced filter whose updates solve the discrete joint RMSE problem and its continuous limit.
    Central algorithmic object; defined by Prop. 3.2 and Algorithm 1.
  • Complemented / observation-aware DLRA subspace dynamics via transported ξ no independent evidence
    purpose: Let U evolve under H by carrying initial orthogonal covariance.
    Introduced in §3.6; efficacy only shown indirectly in numerics, not as a separately measured object.
  • Complementary DLRA particle filter with isotropic Λ_n noise no independent evidence
    purpose: Allow analysis-step basis change and better nonlinear performance.
    Section 4.1 construction; labeled preliminary; no external validation beyond in-paper RMSE.

pith-pipeline@v1.2.0-daily-grok45 · 46253 in / 4115 out tokens · 73386 ms · 2026-07-31T01:04:27.608785+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Dynamical Low-Rank Filters for Data Assimilation." pith.science (2026). https://pith.science/paper/LFTJPT5V

@misc{pith2026260727432,
  author       = {Pith},
  title        = {Pith review of: Dynamical Low-Rank Filters for Data Assimilation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LFTJPT5V}},
  note         = {Machine review of arXiv:2607.27432}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We propose dynamical low-rank (DLR) type filters for data-assimilation problems based on stochastic differential equations (SDEs). In detail, first we derive a DLRA filter for minimizing jointly the mean and covariance error, as well as a strategy to efficiently include the relevant orthogonal directions. This last approach allows the main subspace to evolve also according to the observation operator. Those procedures naturally extend to a Kalman-Bucy type filter when dealing with linear drift, and to ensemble methods, too, resulting also suitable for problems described by nonlinear drift and possible non-Gaussian distribution. Moreover, we further propose a preliminary particle-type DLRA filter that shows potentiality in nonlinear settings. Numerical simulations show the efficacy of these procedures in relevant applications, opening up to further studies in these filtering directions.

discussion (0)

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

Reference graph

Works this paper leans on

52 extracted references · 3 linked inside Pith

  1. [1]

    Existence of dynamical low-rank approximations to parabolic problems

    Markus Bachmayr, Henrik Eisenmann, Emil Kieri, and Andr´ e Uschmajew. “Existence of dynamical low-rank approximations to parabolic problems”. In:Mathematics of Computation90.330 (2021), pp. 1799–1830

  2. [2]

    Alan Bain and Dan Crisan.Fundamentals of stochastic filtering. Vol. 3. Springer, 2009

  3. [3]

    Exponential Convengence of DLRA for SDEs

    Jianhai Bao, Haitao Wang, and Yue Wu. “Exponential Convengence of DLRA for SDEs”. In:arXiv preprint arXiv:2606.15843(2026)

  4. [4]

    Cambridge University Press, 2005

    Andrew F Bennett.Inverse modeling of the ocean and atmosphere. Cambridge University Press, 2005

  5. [5]

    A strongly convergent numerical scheme from ensemble Kalman inversion

    Dirk Blomker, Claudia Schillings, and Philipp Wacker. “A strongly convergent numerical scheme from ensemble Kalman inversion”. In:SIAM Journal on Numerical Analysis56.4 (2018), pp. 2537– 2562

  6. [6]

    Stochastic dynamical low-rank approximation method

    Yu Cao and Jianfeng Lu. “Stochastic dynamical low-rank approximation method”. In:Journal of Computational Physics372 (2018), pp. 564–586

  7. [7]

    Data assimilation in the geosciences: An overview of methods, issues, and perspectives

    Alberto Carrassi, Marc Bocquet, Laurent Bertino, and Geir Evensen. “Data assimilation in the geosciences: An overview of methods, issues, and perspectives”. In:Wiley Interdisciplinary Reviews: Climate Change9.5 (2018), e535

  8. [8]

    Randomized methods for dynamical low-rank approximation

    Benjamin Carrel. “Randomized methods for dynamical low-rank approximation”. In:Journal of Computational Physics(2025), p. 114421

  9. [9]

    A robust second-order low-rank BUG integrator based on the midpoint rule

    Gianluca Ceruti, Lukas Einkemmer, Jonas Kusch, and Christian Lubich. “A robust second-order low-rank BUG integrator based on the midpoint rule”. In:BIT Numerical Mathematics64.3 (2024), p. 30

  10. [10]

    A rank-adaptive robust integrator for dy- namical low-rank approximation

    Gianluca Ceruti, Jonas Kusch, and Christian Lubich. “A rank-adaptive robust integrator for dy- namical low-rank approximation”. In:BIT Numerical Mathematics(2022), pp. 1–26

  11. [11]

    Interpolatory dynamical low-rank approximation for the 3+ 3d Boltzmann–BGK equation

    Alec Dektor and Lukas Einkemmer. “Interpolatory dynamical low-rank approximation for the 3+ 3d Boltzmann–BGK equation”. In:Journal of Computational Physics(2025), p. 114515

  12. [12]

    JHU press, 2013

    Gene H Golub and Charles F Van Loan.Matrix computations, 3rd Edition. JHU press, 2013

  13. [13]

    Cambridge university press, 2012

    Roger A Horn and Charles R Johnson.Matrix analysis. Cambridge university press, 2012

  14. [14]

    An Improved Adaptive Car-Following Model Based on the Unscented Kalman Filter for Vehicle Platoons’ Speed Control

    Caixia Huang, Wu Tang, Jiande Wang, and Zhiyong Zhang. “An Improved Adaptive Car-Following Model Based on the Unscented Kalman Filter for Vehicle Platoons’ Speed Control”. In:Machines 13.7 (2025), p. 569

  15. [15]

    Strong convergence of an explicit numerical method for SDEs with nonglobally Lipschitz continuous coefficients

    Martin Hutzenthaler, Arnulf Jentzen, and Peter E Kloeden. “Strong convergence of an explicit numerical method for SDEs with nonglobally Lipschitz continuous coefficients”. In:The Annals of Applied Probability22.4 (2012), pp. 1611–1641

  16. [16]

    Filtering in finance

    Alireza Javaheri, Delphine Lautier, and Alain Galli. “Filtering in finance”. In:Wilmott3 (2003), pp. 67–83

  17. [17]

    Cambridge university press, 2003

    Eugenia Kalnay.Atmospheric modeling, data assimilation and predictability. Cambridge university press, 2003

  18. [18]

    Extensive chaos in the Lorenz-96 model

    Alireza Karimi and Mark R Paul. “Extensive chaos in the Lorenz-96 model”. In:Chaos: An inter- disciplinary journal of nonlinear science20.4 (2010). 26

  19. [19]

    Stability properties of a projector-splitting scheme for dynamical low rank approximation of random parabolic equations

    Yoshihito Kazashi, Fabio Nobile, and Eva Vidliˇ ckov´ a. “Stability properties of a projector-splitting scheme for dynamical low rank approximation of random parabolic equations”. In:Numerische Mathematik149 (2021), pp. 973–1024

  20. [20]

    Dynamical low-rank approximation for stochastic differential equations

    Yoshihito Kazashi, Fabio Nobile, and Fabio Zoccolan. “Dynamical low-rank approximation for stochastic differential equations”. In:Mathematics of Computation94.353 (2025), pp. 1335–1375

  21. [21]

    Existence of Dynamical Low-Rank Approx- imation for SDEs with Locally Lipschitz Coefficients

    Yoshihito Kazashi, Fabio Nobile, and Fabio Zoccolan. “Existence of Dynamical Low-Rank Approx- imation for SDEs with Locally Lipschitz Coefficients”. In:arXiv preprint(2026)

  22. [22]

    Further Approaches of Dynamical Low-Rank Approximation for SDEs

    Yoshihito Kazashi, Fabio Nobile, and Fabio Zoccolan. “Further Approaches of Dynamical Low-Rank Approximation for SDEs”. In:arXiv preprint(2026)

  23. [23]

    Numerical Methods for Dynamical Low- Rank Approximations of Stochastic Differential Equations - Part II: Stochastic discretization

    Yoshihito Kazashi, Fabio Nobile, and Fabio Zoccolan. “Numerical Methods for Dynamical Low- Rank Approximations of Stochastic Differential Equations - Part II: Stochastic discretization”. In: arXiv preprint(2026)

  24. [24]

    Numerical Methods for Dynamical Low- Rank Approximations of Stochastic Differential Equations – Part I: Time discretization

    Yoshihito Kazashi, Fabio Nobile, and Fabio Zoccolan. “Numerical Methods for Dynamical Low- Rank Approximations of Stochastic Differential Equations – Part I: Time discretization”. In: (2026). arXiv:2601.21428 [math.NA].url:https://arxiv.org/abs/2601.21428

  25. [25]

    B Philipp Kellerhals.Financial pricing models in continuous time and Kalman filtering. Vol. 506. Springer Science & Business Media, 2013

  26. [26]

    Dynamical low-rank approximation

    Othmar Koch and Christian Lubich. “Dynamical low-rank approximation”. In:SIAM Journal on Matrix Analysis and Applications29.2 (2007), pp. 434–454

  27. [27]

    Estimation of market efficiency process within time-varying autoregressive models by extended Kalman filtering approach

    Maria V Kulikova and G Yu Kulikov. “Estimation of market efficiency process within time-varying autoregressive models by extended Kalman filtering approach”. In:Digital Signal Processing128 (2022), p. 103619

  28. [28]

    Data assimilation

    Kody Law, Andrew Stuart, and Kostas Zygalakis. “Data assimilation”. In:Cham, Switzerland: Springer214 (2015), p. 52

  29. [29]

    Deterministic Nonperiodic Flow

    Edward N Lorenz. “Deterministic Nonperiodic Flow”. In:Journal od Athmospheric Sciences. Vol. 20. AMS, 1963, pp. 130–141

  30. [30]

    Predictability: A problem partly solved

    Edward N Lorenz. “Predictability: A problem partly solved”. In:Proc. Seminar on predictability. Vol. 1. 1. Reading. 1996, pp. 1–18

  31. [31]

    Bayesian learning of stochastic dynamical models

    Peter Lu and Pierre FJ Lermusiaux. “Bayesian learning of stochastic dynamical models”. In:Physica D: Nonlinear Phenomena427 (2021), p. 133003

  32. [32]

    A projector-splitting integrator for dynamical low-rank approximation

    Christian Lubich and Ivan V Oseledets. “A projector-splitting integrator for dynamical low-rank approximation”. In:BIT Numerical Mathematics54.1 (2014), pp. 171–188

  33. [33]

    Dual dynamically orthogonal approximation of incom- pressible Navier Stokes equations with random boundary conditions

    Eleonora Musharbash and Fabio Nobile. “Dual dynamically orthogonal approximation of incom- pressible Navier Stokes equations with random boundary conditions”. In:Journal of Computational Physics354 (2018), pp. 135–162

  34. [34]

    Error analysis of the dynamically orthogonal approximation of time dependent random PDEs

    Eleonora Musharbash, Fabio Nobile, and Tao Zhou. “Error analysis of the dynamically orthogonal approximation of time dependent random PDEs”. In:SIAM Journal on Scientific Computing37.2 (2015), A776–A810

  35. [35]

    High-Order BUG Dynamical Low-Rank Integrators Based on Explicit Runge–Kutta Methods

    Fabio Nobile and S´ ebastien Riffaud. “High-Order BUG Dynamical Low-Rank Integrators Based on Explicit Runge–Kutta Methods”. In:Journal of Scientific Computing107.3 (2026), p. 102

  36. [36]

    Fabio Nobile, S´ ebastien Riffaud, and Thomas Trigo Trindade.Dynamical Low-Rank Ensemble Kalman filter for State/Parameter estimation. 2026. arXiv:2602.06614 [math.NA].url:https: //arxiv.org/abs/2602.06614

  37. [37]

    Fabio Nobile and Thomas Trigo Trindade.Dynamical Low-Rank Approximations for Kalman Fil- tering. 2025. arXiv:2509.11210 [math.NA].url:https://arxiv.org/abs/2509.11210

  38. [38]

    Transport map accelerated Markov chain Monte Carlo

    Matthew D Parno and Youssef M Marzouk. “Transport map accelerated Markov chain Monte Carlo”. In:SIAM/ASA Journal on Uncertainty Quantification6.2 (2018), pp. 645–682

  39. [39]

    K. B. Petersen and M. S. Pedersen.The Matrix Cookbook. 2012

  40. [40]

    Low-dimensional reduced-order models for statistical response and un- certainty quantification: Two-layer baroclinic turbulence

    Di Qi and Andrew J Majda. “Low-dimensional reduced-order models for statistical response and un- certainty quantification: Two-layer baroclinic turbulence”. In:Journal of the Atmospheric Sciences 73.12 (2016), pp. 4609–4639. 27

  41. [41]

    Interaction of additive noise and nonlinear dynamics in the double-gyre wind-driven ocean circulation

    Themistoklis P Sapsis and Henk A Dijkstra. “Interaction of additive noise and nonlinear dynamics in the double-gyre wind-driven ocean circulation”. In:Journal of physical oceanography43.2 (2013), pp. 366–381

  42. [42]

    Dynamically orthogonal field equations for continuous stochastic dynamical systems

    Themistoklis P Sapsis and Pierre FJ Lermusiaux. “Dynamically orthogonal field equations for continuous stochastic dynamical systems”. In:Physica D: Nonlinear Phenomena238.23-24 (2009), pp. 2347–2360

  43. [43]

    Blending modified Gaussian closure and non-Gaussian reduced subspace methods for turbulent dynamical systems

    Themistoklis P Sapsis and Andrew J Majda. “Blending modified Gaussian closure and non-Gaussian reduced subspace methods for turbulent dynamical systems”. In:Journal of Nonlinear Science23 (2013), pp. 1039–1071

  44. [44]

    Ren´ e L Schilling.Brownian Motion: A Guide to Random Processes and Stochastic Calculus. 3rd ed. De Gruyter, 2021

  45. [45]

    Data assimilation with Gaussian mixture models using the dynamically orthogonal field equations. Part I: Theory and scheme

    Thomas Sondergaard and Pierre FJ Lermusiaux. “Data assimilation with Gaussian mixture models using the dynamically orthogonal field equations. Part I: Theory and scheme”. In:Monthly Weather Review141.6 (2013), pp. 1737–1760

  46. [46]

    Data assimilation with Gaussian mixture models using the dynamically orthogonal field equations. Part II: Applications

    Thomas Sondergaard and Pierre FJ Lermusiaux. “Data assimilation with Gaussian mixture models using the dynamically orthogonal field equations. Part II: Applications”. In:Monthly Weather Review141.6 (2013), pp. 1761–1785

  47. [47]

    Coupling techniques for nonlinear en- semble filtering

    Alessio Spantini, Ricardo Baptista, and Youssef Marzouk. “Coupling techniques for nonlinear en- semble filtering”. In:SIAM Review64.4 (2022), pp. 921–953

  48. [48]

    A survey of feedback particle filter and related controlled interacting particle systems (CIPS)

    Amirhossein Taghvaei and Prashant G Mehta. “A survey of feedback particle filter and related controlled interacting particle systems (CIPS)”. In:Annual Reviews in Control55 (2023), pp. 356– 378

  49. [49]

    MIT Press, 2005

    Sebastian Thrun, Wolfram Burgard, and Dieter Fox.Probabilistic Robotics. MIT Press, 2005

  50. [50]

    Vallis.Atmospheric and Oceanic Fluid Dynamics: Fundamentals and Large-Scale Cir- culation

    Geoffrey K. Vallis.Atmospheric and Oceanic Fluid Dynamics: Fundamentals and Large-Scale Cir- culation. 2nd ed. Cambridge University Press, 2017

  51. [51]

    Dynamical low rank approximation for uncertainty quantification of time-dependent problems

    Eva Vidlickov´ a. “Dynamical low rank approximation for uncertainty quantification of time-dependent problems”. PhD thesis. Lausanne: EPFL, 2022, p. 222

  52. [52]

    complemented

    Curt Wells.The Kalman filter in finance. Vol. 32. Springer Science & Business Media, 2013. A. Proofs concerning DLRA-JMCO filter Lemma A.1.Consider the following matrices:R∈R h×h of full rank,H∈R h×d, and ˆ︁C∈R d×d of full rank. Then, discarding higher-order terms inO(∆t), we have the following chain of approximation (︁ R+H ˆ︁C H⊤ ∆t )︁−1 ≈R −1 −R −1H ˆ︁C...