Pith. sign in

REVIEW 2 major objections 7 minor 3 cited by

Two staggered 'projector splitting' schemes for low-rank approximations of SDEs converge with constants independent of the smallest singular value, removing the restrictive time-step condition of DLR Euler–Maruyama.

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

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

Three time-stepping schemes for dynamically-orthogonal low-rank SDE surrogates are introduced; two projector-splitting variants converge with bounds independent of the smallest singular value.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection First rigorous time-discretization analysis for DLRA-SDEs, with useful staggered schemes, but one of the two central convergence theorems has a bound that does not actually imply convergence. the 2 major comments →

arxiv 2601.21428 v2 pith:E4A4EZOY submitted 2026-01-29 math.NA cs.NA

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

Yoshihito Kazashi , Fabio Nobile , Fabio Zoccolan This is my paper
classification math.NA cs.NA MSC 65C3060H3565L20
keywords dynamical low-rank approximationstochastic differential equationsDynamically Orthogonal methodEuler–Maruyama methodprojector splittingGramian singular valuestime discretizationmean-square stability
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper studies how to advance in time a low-rank approximation of a high-dimensional stochastic differential equation, where the solution is written as a product of a small number of time-dependent deterministic modes and stochastic modes. It proposes three discretizations. The straightforward DLR Euler–Maruyama converges with the usual strong rate, but only under a time-step bound that depends on the smallest singular value of the Gramian of the stochastic modes, and its error constants can blow up when that singular value is tiny. The two staggered schemes, named DLR Projector Splitting for Euler–Maruyama and DLR Projector Splitting for SDEs, alternate updates of the two mode families and are proven stable and convergent with constants independent of that singular value. This matters because near-degenerate stochastic modes, a common source of failure in low-rank methods, no longer force impractically small time steps; the price is that the analysis assumes exact expectations, with Monte Carlo sampling deferred to Part II.

Core claim

The central result is that the DLR Projector Splitting for Euler–Maruyama satisfies √E[sup_{0≤s≤t}|X_ns - X_true(s)|^2] ≤ C(√E|X_0 - X_true_0|^2 + Δt^{min{1/2,α}} + ε + ε/√Δt), with C independent of the smallest singular value of the covariance matrices of the stochastic modes. The DLR Projector Splitting for SDEs enjoys a similar bound with the ε/√Δt term removed, because only the drift is projected onto the tangent space in the deterministic-mode update, making the scheme adapted to the natural filtration. The paper also shows that, under non-degenerate noise, the stochastic-mode Gramian stays positive definite with an explicit lower bound for all three schemes, and that the two staggered

What carries the argument

The Dynamically Orthogonal / dynamical low-rank ansatz X = U^T Y, with orthonormal deterministic modes U and stochastic modes Y, leads to the Gramian C_Y = E[Y Y^T] whose inverse appears in the equation for the deterministic modes. The staggered schemes replace the simultaneous forward update by: first evolve Y with Euler–Maruyama using the current U, then compute the new deterministic modes using the freshly computed stochastic modes and a tangent-space projector P_{U^T Y~}, followed by QR orthonormalization. This projector-splitting structure is what removes the dependence on the smallest singular value of the Gramian.

Load-bearing premise

The proofs assume that every expectation inside Algorithms 1–3 is computed exactly at each time step; the numerical experiments instead use Monte Carlo sample averages, and the analysis does not control that replacement.

What would settle it

Repeat the Section 7 experiments for the staggered schemes on problem (70) with σB=10^-19 so the stochastic-mode Gramian hovers near machine zero; if the strong error, or the Δt needed to reach a fixed tolerance, grows like 1/σ_min as the smallest singular value decreases, the claimed independence from the singular value is wrong.

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

If this is right

  • DLR Euler–Maruyama is usable when Δt meets the singular-value-dependent condition; under non-degenerate noise the Gramian remains full rank with a lower bound matching the continuous case.
  • The two staggered schemes converge strongly to the true SDE solution with order min{1/2,α} plus a low-rank projection error ε; for DLR Projector Splitting for SDEs the bound does not include the ε/√Δt term.
  • Both staggered schemes are mean-square stable for linear SDEs under the same sufficient condition as standard Euler–Maruyama, independent of the smallest singular value of the stochastic Gramian.
  • No time-step restriction is needed for the L2 norm bound of the staggered schemes, so near-rank-deficient stochastic modes do not halt the computation.
  • Algorithm 3's update is adapted to the filtration, making it the natural target for the Monte Carlo discretization treated in Part II.

Where Pith is reading between the lines

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

  • If Part II justifies Monte Carlo replacement of expectations with rigorous error control, the projector-splitting schemes become immediately usable in large-scale uncertainty quantification; until then the proven rates apply only to the idealized exact-expectation algorithms.
  • The ε term in the error bound is a computable measure of how far the drift and diffusion are from the tangent space of rank-k processes, so it could be used adaptively to grow or shrink the rank during a simulation.
  • Because the deterministic-mode update in linear-drift cases decouples from the Gramian, the staggered schemes may parallelize across stochastic samples or combine with rank-adaptive integrators without triggering step-size crashes.
  • The experiments with nearly singular Gramians suggest the staggered schemes may tolerate stochastic modes whose Gramian becomes numerically singular, a regime where classical DLRA integrators for deterministic problems typically require regularization.
Share X Bluesky LinkedIn Reddit HN

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

2 major / 7 minor

Summary. The paper proposes and analyzes three time-discretization schemes for the Dynamically Orthogonal (DO) formulation of dynamical low-rank approximation (DLRA) for high-dimensional SDEs. Algorithm 1 (DLR Euler–Maruyama) is shown to converge under a time-step restriction tied to the smallest singular value of the Gramian; Algorithms 2 and 3 are staggered projector-splitting variants designed to remove that restriction. The paper proves moment bounds, non-degeneracy of the Gramian under non-degenerate noise, mean-square stability for linear SDEs, and strong convergence estimates comparing the numerical DLR solution both to the continuous DLRA solution and to the exact SDE solution. Numerical experiments with Monte Carlo expectations are provided for low-dimensional SDEs and for discretized SPDEs.

Significance. If the results are correct, this is a useful contribution: it extends the DLRA time-discretization theory from deterministic/random PDEs to general SDEs, gives explicit convergence rates, identifies the singular-value dependence of the standard Euler–Maruyama approach, and proposes staggered schemes with better robustness. The paper is careful to state that all expectations are exact in the theory and to defer Monte Carlo sampling to Part II. The proofs are detailed and are derived from stated assumptions with no fitted parameters. However, the central convergence claim for Algorithm 2 is not established as stated because the error bound in Theorem 5.5 contains an ε/√Δt term that diverges as Δt→0 for fixed modeling error ε>0. This is a load-bearing defect in one of the paper's headline contributions.

major comments (2)
  1. [Theorem 5.5 / Proposition 5.3] The advertised convergence of Algorithm 2 is not established. In the proof of Proposition 5.3, the diffusion term T2 is bounded by repeatedly applying Cauchy–Schwarz and then using E[|Y~_{n+1}^T C_{Y~_{n+1}}^{-1} Y~_{n+1}|]=k for each of the n_t summands. This yields the displayed bound ε²Δt n_t E[Σ |C_{Y~_{n+1}}^{-1/2}Y~_{n+1}|²] ≤ ε²T²/Δt. Passing to the square root gives the ε/√Δt term in Theorem 5.5. For any fixed positive modeling error ε>0, this term diverges as Δt→0, so the right-hand side has no finite limit and the theorem does not imply convergence in time. This is not a Monte Carlo issue; it fails at the exact-expectation level. The estimate discards the martingale-difference structure of the Brownian increments: a sharper bound exploiting the summation structure is needed to obtain O(ε) rather than O(ε/√Δt). As stated, Theorem 5.5 cannot support the abstract's claim that the
  2. [Section 6, Theorem 6.3] Theorem 6.3 is stated without proof. This theorem is a central convergence result for Algorithm 3 and is the source of the clean ε dependence that contrasts with Algorithm 2. The section says only that the results are obtained 'via similar proofs,' but no proof or even a detailed sketch is provided. In particular, the reader cannot verify how the ε/√Δt factor is avoided. Please provide the proof or a precise sketch, especially the treatment of the diffusion term that leads to the O(ε) bound.
minor comments (7)
  1. [Proposition 4.1] The vector v is first taken in R^d ('Let us take a unit vector v ∈ R^d') but the Gramian CY_{n+1} is k×k and later the infimum is over v∈R^k. This should be corrected to avoid confusion.
  2. [Lemma 4.2 proof] The treatment of T5 is incomplete: after 'For T4 one obtains' there is no displayed bound for T5, and the final line labels the term as T4+T5 without showing the T5 estimate. Please add the missing derivation or state explicitly that it is analogous to T4.
  3. [Theorem 4.7 proof] In the line after (46), the factor is written 'expmax(A,B)T ·2(C + D)((Δt)^{min{1,2α}})' which is ambiguous. It should be exp(max{A,B}T)·2(C+D)(Δt)^{min{1,2α}} (or similar).
  4. [Section 5] The text says 'This result is subsequently applied in Proposition 5.1' but the result referred to is Proposition 5.2. The cross-reference should be corrected.
  5. [Section 7.3] The first occurrence of 'Figures 14, 15, and 16' in Section 7.3 should instead refer to Figures 10–12; the same figure numbers are reused in Section 7.4. Please fix the figure references.
  6. [Conclusion / Numerical experiments] All theorems assume exact expectations, while Section 7 replaces them with Monte Carlo sample averages. The conclusion acknowledges this and defers the analysis to Part II, but a sentence in Section 7 clarifying that the numerical plots validate the combined time-and-sample discretization (not the time discretization alone) would help prevent over-reading of the empirical convergence rates.
  7. [Theorem 5.5] The theorem says 'Suppose Assumptions (1)-(2) hold' but the Assumptions are numbered 1–3; please update the reference.

Circularity Check

0 steps flagged

No significant circularity: the convergence estimates are new derivations; reliance on the authors' prior continuous DLRA theory is normal, non-load-bearing self-citation, though Theorem 5.5's ε/√Δt term and the exact-expectation/Monte-Carlo gap are non-circular limitations.

full rationale

I found no circular step that makes a prediction equivalent to its inputs by construction. The DO equations are taken from the authors' prior work [18] and the projector-splitting perspective from [16,17,27], but all new convergence bounds (Theorems 4.7, 4.10, 5.3, 5.5, 6.3, 6.4) are obtained from the stated Lipschitz, linear-growth, uniform-ellipticity, Hölder-in-time, and rank-k projection-error assumptions via Gronwall/Doob/Itô arguments. The quantity ε in (47) is a best-approximation projection error, not a fitted parameter, so its appearance in the error bounds is not circular. The paper explicitly flags the exact-expectation restriction in the conclusion: 'all the analyses done in this first part always assumes exact expectations for Algorithms 1, 2, and 3', while Section 7 uses Monte Carlo averages; this is an acknowledged gap, not a disguised reuse of inputs. The self-citations to [16,18] supply well-posedness and technical lemmas for the continuous DLRA; they are parameter-free results with stated assumptions and are not the target discretization claims, so they do not create a self-citation chain that forces the present results. One non-circular correctness caveat should be flagged: in Proposition 5.3, the T2 estimate is bounded by replacing a sum of n_t terms with a Cauchy–Schwarz factor n_t, yielding 3ε²T²/Δt, so Theorem 5.5 contains the term ε/√Δt, which diverges as Δt→0 for fixed ε>0. Thus Algorithm 2's advertised convergence is not established by that bound; this is a proof-strength issue, not circularity, and Algorithm 3's Theorem 6.3 (with plain ε) is not affected. Overall, no load-bearing circular step; only minor, non-load-bearing self-citation.

Axiom & Free-Parameter Ledger

0 free parameters · 8 axioms · 0 invented entities

The central claims rest on standard stochastic-analysis tools and on the authors' prior well-posedness theory for continuous DLRA-SDEs. There are no fitted free parameters in the theory; the main idealization is exact expectation computation. The non-degenerate-noise assumption and the projection-error bound ε are the most consequential domain assumptions.

axioms (8)
  • domain assumption Drift and diffusion are uniformly Lipschitz continuous (Assumption 1, Eq. (3)).
    Required for strong well-posedness of the SDE (2) and used throughout all error estimates.
  • domain assumption Drift and diffusion satisfy linear growth bound (Assumption 2, Eq. (4)).
    Used to control second moments of the numerical and continuous solutions.
  • domain assumption Initial condition is square integrable and F0-measurable (Assumption 3).
    Standard assumption for strong solutions and for the moment bounds used in the proofs.
  • domain assumption Initial Gramian CY0 is invertible (Assumption 4).
    Ensures the DLR initial data has k nontrivial components and allows the inverse Gramian in the update equations.
  • domain assumption Non-degenerate diffusion: b b^T ⎷ σB I (Assumption 5).
    Used to prove uniform lower bounds on the Gramian for all three schemes (Propositions 4.1, 5.2, 6.2).
  • domain assumption Hölder continuity in time of drift and diffusion (Eq. (41)).
    Needed for the temporal error terms, giving rate α in the convergence theorems.
  • domain assumption Bounded rank-k projection error ε (Eq. (47)).
    Theorems comparing the numerical DLR solution to the true SDE solution require finite maximal tangent-space projection error of drift and diffusion on the low-rank manifold.
  • domain assumption All expectations in Algorithms 1–3 are computed exactly.
    The theoretical convergence and stability results assume exact expectations; Monte Carlo approximation is explicitly deferred to Part II.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Numerical Methods for Dynamical Low-Rank Approximations of Stochastic Differential Equations -- Part I: Time discretization." pith.science (2026). https://pith.science/paper/E4A4EZOY

@misc{pith2026260121428,
  author       = {Pith},
  title        = {Pith review of: Numerical Methods for Dynamical Low-Rank Approximations of Stochastic Differential Equations -- Part I: Time discretization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E4A4EZOY}},
  note         = {Machine review of arXiv:2601.21428}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

In this work (Part I), we study three time-discretization schemes for the Dynamical Low-Rank Approximation (DLRA) of high-dimensional stochastic differential equations (SDEs). Specifically, we consider the Dynamically Orthogonal (DO) method for DLRA proposed and analyzed in arXiv:2308.11581v4, which approximates the true solution by a linear combination of few products between deterministic orthonormal modes and stochastic modes, both time-dependent. The first scheme considered consists in a forward discretization in time of both deterministic and stochastic components, in a Euler-Maruyama style. Its convergence is proven subject to a time-step restriction dependent on the smallest singular value of the Gram matrix associated to the stochastic modes, which, on its turn, is shown to be always positive, provided that the SDE under study is driven by a non-degenerate noise. The second and the third schemes, on the other hand, are staggered ones, alternating updates of the deterministic and the stochastic modes in half steps, and have a projector splitting nature. We show stability of the second scheme and prove convergence with constants independent of the smallest singular value. The third scheme works better in practice, although our theoretical convergence bounds are worse than those for the second one. Computational experiments support our theoretical results. In this work we do not consider the discretization in probability, which will be the topic of Part II.

Figures

Figures reproduced from arXiv: 2601.21428 by Fabio Nobile, Fabio Zoccolan, Yoshihito Kazashi.

Figure 1
Figure 1. Figure 1: For ∆t = 0.1, 0.05, 0.02, k = 2, M = 10000, σB = 10−8 for (68). (Top) Smallest singular values of E[YnY ⊤ n ] for the DLR Euler-Maruyama, (Center) for the DLR Projector Splitting for EM, (Bottom) for the DLR Projector Splitting for SDEs. Lower bounds of σk refer to (37), (54), and (65) for all the three algorithms, respectively [PITH_FULL_IMAGE:figures/full_fig_p037_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: (Left) Singular values for the process Xtrue(t), solution of equation (68). (Right) Relative errors for the DLR Euler-Maruyama, DLR Projector Splitting for EM, and the DLR Projector Splitting for SDEs, with respect to the true solution Xtrue and the continuous DLRA of rank k = 2. M = 10000 paths are used in all methods to approximate expectations. to 2, but close to be rank deficient. Furthermore, we choos… view at source ↗
Figure 3
Figure 3. Figure 3: For ∆t = 0.1, 0.05, 0.02, k = 2, M = 10000, σB = 10−19. (Top) Smallest singular values of E[YnY ⊤ n ] for the DLR Euler-Maruyama, (Center) for the DLR Projector Splitting for EM, (Bottom) for the DLR Projector Splitting for SDEs for Problem (69). singular values of the stochastic basis does not affect the convergence of any of the algorithms. Moreover, one can notice a convergence in time of order o(∆t); t… view at source ↗
Figure 4
Figure 4. Figure 4: Relative errors for the DLR Euler-Maruyama, DLR Projector Splitting for EM, [PITH_FULL_IMAGE:figures/full_fig_p038_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: For ∆t = 0.1, 0.05, 0.02, k = 2, M = 10000, σB = 10−19. For Problem (70) (Top) smallest singular values of E[YnY ⊤ n ] for the DLR Euler-Maruyama, (Center) for the DLR Projector Splitting for EM, (Bottom) for the DLR Projector Splitting for SDEs [PITH_FULL_IMAGE:figures/full_fig_p039_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Relative errors for the DLR Euler-Maruyama, DLR Projector Splitting for EM, [PITH_FULL_IMAGE:figures/full_fig_p040_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: For ∆t = 0.0907, for equation (71) (Left) Mean-square norm of the numerical solution Xn, i.e. E[|Xn| 2 ], for the DLR Euler-Maruyama, (Center) for the DLR Projector Splitting for EM, (Right) for the DLR Projector Splitting for SDEs, M = 2000 paths. 7.3 Additive Noise: Stochastic Advection-Diffusion-Reaction PDE in 1D In this section, we provide a numerical example consisting of a one-dimensional advection￾… view at source ↗
Figure 8
Figure 8. Figure 8: For ∆t = 0.909, for equation (71) (Left) Mean-square norm of the numerical solution Xn, i.e. E[|Xn| 2 ], for the DLR Euler-Maruyama, (Center) for the DLR Projector Splitting for EM, (Right) for the DLR Projector Splitting for SDEs, M = 2000 paths [PITH_FULL_IMAGE:figures/full_fig_p041_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: For ∆t = 0.0911, for equation (71) (Left) Mean-square norm of the numerical solution Xn, i.e. E[|Xn| 2 ], for the DLR Euler-Maruyama, (Center) for the DLR Projector Splitting for EM, (Right) for the DLR Projector Splitting for SDEs, M = 2000 paths. time step based on the smallest singular values of the Gramian of the stochastic basis Yn. The spatial domain is [0, L] with L = 1, where we consider Neumann bo… view at source ↗
Figure 10
Figure 10. Figure 10: Condition (74) related to the square root of the singular values [PITH_FULL_IMAGE:figures/full_fig_p042_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Condition (74) related to the square root of the singular values [PITH_FULL_IMAGE:figures/full_fig_p042_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Condition (74) related to the square root of the singular values [PITH_FULL_IMAGE:figures/full_fig_p042_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Rel. errors in L 2 − sup t∈[0,T] norm for equation (73), k = 18, M = 2000 paths, (Blue (−)) error between the true solution and the DLRA algorithms and between the continuous DLR and the DLRA algorithms 7.4 Multiplicative noise: Stochastic Laplacian in 1D In this numerical example we consider a stochastic Laplacian problem in 1D with Dirichlet boundary conditions under multiplicative colored noise. Like t… view at source ↗
Figure 14
Figure 14. Figure 14: Condition (74) related to the square root of the singular values [PITH_FULL_IMAGE:figures/full_fig_p044_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Condition (74) related to the square root of the singular values [PITH_FULL_IMAGE:figures/full_fig_p045_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Condition (74) related to the square root of the singular values [PITH_FULL_IMAGE:figures/full_fig_p045_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Rel. errors in L 2 − sup t∈[0,T] norm for the Problem (75), k = 14, M = 2200 paths, (Blue (−)) error between the true solution and the DLRA algorithms and between the continuous DLRA and the DLRA algorithms and continue to satisfy the previously stated lower-boundedness property of the Gramian. Their similar convergence behavior in time is further supported by numerical simulations. Among possible directi… view at source ↗

discussion (0)

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

Forward citations

Cited by 3 Pith papers

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

  1. Dynamical Low-Rank Filters for Data Assimilation

    math.NA 2026-07 conditional novelty 6.0

    Dynamical low-rank filters minimize joint mean-covariance error for SDE data assimilation and extend to Kalman-Bucy, ensemble, and particle forms for nonlinear problems.

  2. Long-time Behaviour of DLRA for SDEs

    math.PR 2026-06 unverdicted novelty 5.0

    Proves existence of invariant measures for strong DO systems approximating SDEs via moment estimates and fixed-point arguments under standard dissipativity assumptions.

  3. Dynamical Low-Rank Smoothing

    math.NA 2026-07 conditional novelty 4.0

    DLRA-JMCO filtering is extended via RTS recursion to a reduced-order smoother, with a Kalman–Bucy form for affine drift, cutting cost and storage while keeping an adaptive low-rank basis.

Reference graph

Works this paper leans on

43 extracted references · cited by 3 Pith papers

  1. [1]

    Modeling with Itˆ o stochastic differential equations

    Edward Allen. Modeling with Itˆ o stochastic differential equations. Vol. 22. Springer Science & Business Media, 2007

    2007

  2. [2]

    Stochastic differential equations

    Ludwig Arnold. Stochastic differential equations. John Wiley, New York (NY), 1974

    1974

  3. [3]

    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 Computation 90.330 (2021), pp. 1799–1830

    2021

  4. [4]

    Interpola- tory Dynamical Low-Rank Approximation: Theoretical Foundations and Algorithms

    Benjamin Carrel, Daniel Kressner, Hei Yin Lam, and Bart Vandereycken. “Interpola- tory Dynamical Low-Rank Approximation: Theoretical Foundations and Algorithms”. In: arXiv preprint arXiv:2510.19518 (2025)

    arXiv 2025

  5. [5]

    A rank-adaptive robust integra- tor for dynamical low-rank approximation

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

    2022

  6. [6]

    An unconventional robust integrator for dy- namical low-rank approximation

    Gianluca Ceruti and Christian Lubich. “An unconventional robust integrator for dy- namical low-rank approximation”. In:BIT Numerical Mathematics 62.1 (2022), pp. 23– 44

    2022

  7. [7]

    Stochastic equations in infinite dimensions

    Giuseppe Da Prato and Jerzy Zabczyk. Stochastic equations in infinite dimensions . Vol. 152. Cambridge university press, 2014

    2014

  8. [8]

    A low-rank projector-splitting integrator for the Vlasov–Poisson equation

    Lukas Einkemmer and Christian Lubich. “A low-rank projector-splitting integrator for the Vlasov–Poisson equation”. In: SIAM Journal on Scientific Computing 40.5 (2018), B1330–B1360

    2018

  9. [9]

    A robust and con- servative dynamical low-rank algorithm

    Lukas Einkemmer, Alexander Ostermann, and Carmela Scalone. “A robust and con- servative dynamical low-rank algorithm”. In: Journal of Computational Physics 484 (2023), p. 112060

    2023

  10. [10]

    On the Dirac–Frenkel vari- ational principle on tensor Banach spaces

    Antonio Falc´ o, Wolfgang Hackbusch, and Anthony Nouy. “On the Dirac–Frenkel vari- ational principle on tensor Banach spaces”. In: Foundations of computational mathe- matics 19 (2019), pp. 159–204

    2019

  11. [11]

    Low-order stochastic mode reduction for a realistic barotropic model climate

    Christian Franzke, Andrew J Majda, and Eric Vanden-Eijnden. “Low-order stochastic mode reduction for a realistic barotropic model climate”. In: Journal of the atmo- spheric sciences 62.6 (2005), pp. 1722–1745

    2005

  12. [12]

    Non-equilibrium transitions in multiscale systems with a bifurcating slow manifold

    Tobias Grafke and Eric Vanden-Eijnden. “Non-equilibrium transitions in multiscale systems with a bifurcating slow manifold”. In:Journal of Statistical Mechanics: Theory and Experiment 2017.9 (2017), p. 093208

    2017

  13. [13]

    Certified reduced basis methods for parametrized partial differential equations

    Jan S Hesthaven, Gianluigi Rozza, Benjamin Stamm, et al. Certified reduced basis methods for parametrized partial differential equations . Vol. 590. Springer

  14. [14]

    A-stability and stochastic mean-square stability

    Desmond J Higham. “A-stability and stochastic mean-square stability”. In: BIT Nu- merical Mathematics 40 (2000), pp. 404–409

    2000

  15. [15]

    Matrix analysis

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

    2012

  16. [16]

    Existence of dynamical low rank approximations for random semi-linear evolutionary equations on the maximal interval

    Yoshihito Kazashi and Fabio Nobile. “Existence of dynamical low rank approximations for random semi-linear evolutionary equations on the maximal interval”. In:Stochastics and Partial Differential Equations: Analysis and Computations 9.3 (2021), pp. 603– 629

    2021

  17. [17]

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

    Yoshihito Kazashi, Fabio Nobile, and Eva Vidliˇ ckov´ a. “Stability properties of a projector- splitting scheme for dynamical low rank approximation of random parabolic equa- tions”. In: Numerische Mathematik 149 (2021), pp. 973–1024

    2021

  18. [18]

    Dynamical low-rank approxi- mation for stochastic differential equations

    Yoshihito Kazashi, Fabio Nobile, and Fabio Zoccolan. “Dynamical low-rank approxi- mation for stochastic differential equations”. In: Mathematics of Computation 94.353 (2025), pp. 1335–1375. doi: 10.1090/mcom/3999

    doi:10.1090/mcom/3999 2025

  19. [19]

    Numerical Methods for Dynam- ical Low-Rank Approximations of Stochastic Differential Equations - Part II: Stochas- tic discretization

    Yoshihito Kazashi, Fabio Nobile, and Fabio Zoccolan. “Numerical Methods for Dynam- ical Low-Rank Approximations of Stochastic Differential Equations - Part II: Stochas- tic discretization”. In preparation. 2025. 47

    2025

  20. [20]

    Stochastic stability of differential equations

    Rafail Khasminskii. Stochastic stability of differential equations . Vol. 66. Springer Sci- ence & Business Media, 2011

    2011

  21. [21]

    Discretized dynamical low-rank ap- proximation in the presence of small singular values

    Emil Kieri, Christian Lubich, and Hanna Walach. “Discretized dynamical low-rank ap- proximation in the presence of small singular values”. In: SIAM Journal on Numerical Analysis 54.2 (2016), pp. 1020–1038

    2016

  22. [22]

    Projection methods for dynamical low-rank ap- proximation of high-dimensional problems

    Emil Kieri and Bart Vandereycken. “Projection methods for dynamical low-rank ap- proximation of high-dimensional problems”. In: Computational Methods in Applied Mathematics 19.1 (2019), pp. 73–92

    2019

  23. [23]

    Numerical Solution of Stochastic Differential Equations

    Peter E Kloeden and Eckhard Platen. Numerical Solution of Stochastic Differential Equations. Vol. 23. Springer, Berlin, Heidelberg, New York, 1992

    1992

  24. [24]

    Dynamical low-rank approximation

    Othmar Koch and Christian Lubich. “Dynamical low-rank approximation”. In: SIAM Journal on Matrix Analysis and Applications 29.2 (2007), pp. 434–454

    2007

  25. [25]

    Dynamical Low-Rank Approximation for Burgers’ Equations with Uncertainty

    Jonas Kusch, Gianluca Ceruti, Lukas Einkemmer, and Martin Frank. “Dynamical Low-Rank Approximation for Burgers’ Equations with Uncertainty”. In: International Journal for Uncertainty Quantification 12.5 (2022)

    2022

  26. [26]

    Data assimilation

    Kody Law, Andrew Stuart, and Kostas Zygalakis. Data assimilation. Vol. 214. Springer, Cham, Switzerland, 2015, p. 52

    2015

  27. [27]

    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 Mathematics 54.1 (2014), pp. 171–188

    2014

  28. [28]

    Models for stochastic climate prediction

    Andrew J Majda, Ilya Timofeyev, and Eric Vanden Eijnden. “Models for stochastic climate prediction”. In: Proceedings of the National Academy of Sciences 96.26 (1999), pp. 14687–14691

    1999

  29. [29]

    Stochastic differential equations and applications

    Xuerong Mao. Stochastic differential equations and applications . Elsevier, 2007

    2007

  30. [30]

    Stochastic numerics for mathematical physics

    Grigori N Milstein and Michael V Tretyakov. Stochastic numerics for mathematical physics. Vol. 39. Springer, 2004

    2004

  31. [31]

    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 Computing 37.2 (2015), A776–A810

    2015

  32. [32]

    Stochastic processes and applications

    Grigorios A Pavliotis. Stochastic processes and applications. Springer, New York, NY, 2014

    2014

  33. [33]

    On-the-fly reduced order modeling of passive and reactive species via time-dependent manifolds

    Donya Ramezanian, Arash G Nouri, and Hessam Babaee. “On-the-fly reduced order modeling of passive and reactive species via time-dependent manifolds”. In: Computer Methods in Applied Mechanics and Engineering 382 (2021), p. 113882

    2021

  34. [34]

    Low-dimensional approxima- tions of high-dimensional asset price models

    Martin Redmann, Christian Bayer, and Pawan Goyal. “Low-dimensional approxima- tions of high-dimensional asset price models”. In: SIAM Journal on Financial Math- ematics 12.1 (2021), pp. 1–28

    2021

  35. [35]

    Mean-square stability of numerical schemes for stochastic differential systems

    Yoshihiro Saito and Taketomo Mitsui. “Mean-square stability of numerical schemes for stochastic differential systems”. In: Vietnam J. Math 30 (2002), pp. 551–560

    2002

  36. [36]

    Stability Analysis of Numerical Schemes for Stochastic Differential Equations

    Yoshihiro Saito and Taketomo Mitsui. “Stability Analysis of Numerical Schemes for Stochastic Differential Equations”. In: SIAM J. Numer. Anal. 33.6 (1996), pp. 2254– 2267

    1996

  37. [37]

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

    Themistoklis P Sapsis and Henk A Dijkstra. “Interaction of additive noise and nonlin- ear dynamics in the double-gyre wind-driven ocean circulation”. In:Journal of physical oceanography 43.2 (2013), pp. 366–381

    2013

  38. [38]

    Dynamically orthogonal field equa- tions for continuous stochastic dynamical systems

    Themistoklis P Sapsis and Pierre FJ Lermusiaux. “Dynamically orthogonal field equa- tions for continuous stochastic dynamical systems”. In:Physica D: Nonlinear Phenom- ena 238.23-24 (2009), pp. 2347–2360

    2009

  39. [39]

    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 Science 23 (2013), pp. 1039–1071

    2013

  40. [40]

    Brownian Motion: A Guide to Random Processes and Stochastic Calculus

    Ren´ e L Schilling. Brownian Motion: A Guide to Random Processes and Stochastic Calculus. 3rd ed. De Gruyter, 2021. 48

    2021

  41. [41]

    Stochastic calculus for finance II: Continuous-time models

    Steven E Shreve et al. Stochastic calculus for finance II: Continuous-time models . Vol. 11. Springer, 2004

    2004

  42. [42]

    Data-driven structure-preserving model reduction for stochastic Hamil- tonian systems

    T Tyranowski. “Data-driven structure-preserving model reduction for stochastic Hamil- tonian systems”. In: Journal of Computational Dynamics 11.2 (2024), pp. 220–255

    2024

  43. [43]

    Coupling parameter and particle dynamics for adaptive sampling in Neural Galerkin schemes

    Yuxiao Wen, Eric Vanden-Eijnden, and Benjamin Peherstorfer. “Coupling parameter and particle dynamics for adaptive sampling in Neural Galerkin schemes”. In: Physica D: Nonlinear Phenomena 462 (2024), p. 134129. 49

    2024

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.