Robust high-order low-rank BUG integrators based on explicit Runge--Kutta methods

Fabio Nobile; S\'ebastien Riffaud

REVIEW 1 major objections 6 minor 1 cited by

Robust high-order low-rank BUG integrators based on explicit Runge--Kutta methods

T0 review · 1 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read One BUG step at each Runge–Kutta stage makes low-rank integrators keep the underlying scheme's order.

desk verdict A correct and useful generalization of BUG integrators to arbitrary explicit Runge–Kutta methods, with a clean convergence theorem that deserves real refereeing. read the letter →

arxiv 2502.07040 v7 pith:PUZ42XW4 submitted 2025-02-10 math.NA cs.NAmath.DS

Fabio Nobile , Sébastien Riffaud This is my paper · ORCID

classification math.NAcs.NAmath.DS MSC 65L0565L0665F55

keywords dynamicallow-rankapproximationmatrixdifferentialequationsBasis-Update&GalerkinintegratorsRunge–Kuttamethodshigh-orderconvergencerankadaptivityconservationlawsVlasov–Poisson

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 introduces a way to give dynamical low-rank integrators any order of accuracy. The central idea is to take the first-order BUG integrator and run one BUG step at every stage of an explicit Runge–Kutta method, building augmented bases that span each stage value and its tangent-space projection. The main theorem states that the resulting RK–BUG integrator inherits the underlying Runge–Kutta scheme's order $p$: the global error is bounded by $C'(\delta_r + \varepsilon_r + \gamma_r + h^p)$, where the three rank-dependent terms measure initial, projection, and truncation errors and vanish when the rank is full. A user can therefore pick a standard fourth-order scheme and get fourth-order low-rank time integration, with robustness to small singular values, forward-only updates, and rank adaptivity. The paper supports the claim numerically on Allen–Cahn, Lyapunov, discrete nonlinear Schrödinger, and Vlasov–Poisson problems.

What carries the argument

The central object is the RK–BUG step: at each Runge–Kutta stage, the current low-rank factors $(U,S,V)$ are augmented by the stage vector and its projected derivative, orthonormalized into wider bases $\hat{U}, \hat{V}$, a Galerkin projection computes $\hat{S}$, and the augmented solution is truncated back to rank $r$ by an SVD. The identity doing the work is the augmented-basis span property: the bases are constructed so that the projected discrete solution at each stage is exactly representable, so the only per-step losses are the tangent-space projection error $\varepsilon_r$, the truncation error $h\gamma_r$, and the Runge–Kutta local error $h^{p+1}$. Proposition 3 shows the Galerkin projection onto these augmented bases is no worse than the tangent-space projection used by projected Runge–Kutta methods, which is what lets the order bound transfer from the underlying scheme to the low-rank integrator.

What would settle it

Run RK–BUG based on a fourth-order Runge–Kutta scheme on a problem whose exact flow is only $C^1$ but whose right-hand side is Lipschitz, using a small step $h$ and a rank $r$ large enough that $\varepsilon_r$ and $\gamma_r$ are negligible; if the observed convergence slope is clearly below 4 rather than settling at a low-rank plateau, Theorem 2's order-inheritance claim fails. A direct comparison against the full-rank Runge–Kutta solution on the same problem would settle whether the deficit comes from the low-rank projection or from the flow regularity.

Watch

Extended reading notes

Core claim

The paper claims that the BUG construction, previously limited to first order, can be lifted to any explicit Runge–Kutta method without losing the method's convergence order. Theorem 2 proves that for fixed rank $r$ and a sufficiently small step size $h$, the global Frobenius-norm error satisfies $\|X(t_N)-Y_N\|_F \leq C'(\delta_r + \varepsilon_r + \gamma_r + h^p)$, where $\delta_r$ is the initial low-rank error, $\varepsilon_r$ bounds the tangent-space projection error, $\gamma_r$ bounds the per-step truncation error, and all three vanish when $r = \min\{n,m\}$. Thus the integrator converges at the underlying Runge–Kutta order until the low-rank truncation plateau dominates, and at full rank it reduces to the full Runge–Kutta method. The proof combines stagewise local error estimates with a Lady Windermere fan along the exact flow, using a Galerkin projection onto augmented bases at each stage. The authors also give a conservative variant that keeps linear invariants such as mass and momentum at machine precision, and a rank-adaptive strategy that keeps the truncation terms proportional to $h^{p+1}$.

Load-bearing premise

The load-bearing assumption is that the exact solution flow has $p$ continuous derivatives, the standard condition ensuring an explicit Runge–Kutta method has order $p$; if the flow is less regular, the underlying Runge–Kutta method itself loses its order and the theorem's inheritance claim has no ground to stand on.

Editorial extensions

If this is right

Any explicit Runge–Kutta scheme (in the paper: midpoint, Heun, SSP33, Heun3, RK4) yields a dynamical low-rank integrator of the same order, so users can choose a familiar time integrator rather than a dedicated low-rank one.
Before the plateau, halving $h$ reduces the error by about $2^p$; increasing the rank lowers the plateau, and at full rank RK–BUG is exactly the underlying Runge–Kutta method.
The rank-adaptive rule — truncate each augmented stage so the truncation error stays below $\alpha h^{p+1}$ — gives a practical, order-preserving way to choose rank automatically.
The conservative variant preserves linear invariants such as total mass and momentum to machine precision while keeping the convergence order, which matters for kinetic simulations such as Vlasov–Poisson.

Reading between the lines

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

Beyond the paper, the same stagewise augmentation should extend to implicit and exponential Runge–Kutta methods, because the proof uses only the RK coefficients through the stage combination; the authors list this as future work.
Because the truncation term $\gamma_r$ is independent of $h$, for a fixed rank the error floor is $O(\gamma_r)$ no matter how small $h$ is; the practical order a user sees is therefore capped by the singular-value decay of the exact solution.
The improved error bound over projected Runge–Kutta methods comes from explicitly tracking the truncation term $\gamma_r$, suggesting the analysis might transfer to other retraction or proximal truncation maps, not only SVD truncation.
Assumption 2 is not verified in the Vlasov–Poisson experiment, so the conservation and accuracy shown there do not directly test Theorem 2; a dedicated study on non-smooth kinetic flows would clarify how far order inheritance extends in practice.

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.

Desk Editor's Note

read the letter

Short version: this is the generalization of BUG to arbitrary explicit RK methods that the field needed, and the convergence theorem is the real contribution. I read the proof chain carefully (Proposition 3, Proposition 4, Lemma 4, Theorem 1, Theorem 2) and it is coherent. The gamma_r term cleanly separates truncation error from projection error, and the Lady Windermere's fan accumulation is standard. The construction itself—one BUG step per stage—is natural, but the general formulation and the error bound are not in the cited literature. The numerical tests back the claimed orders 2–4 and show the expected plateaus. Credit where due: Proposition 4's use of Horn–Johnson singular value inequalities to make gamma_r h-independent is a nice, and I believe correct, move; Proposition 3's Galerkin-versus-tangent projection inequality is also solid. This is more than a collection of experiments.

Soft spots, in proportion. The abstract says "superior accuracy compared to other existing dynamical low-rank integrators," but the Allen–Cahn results actually show midpoint BUG slightly more accurate for small ranks. Minor, but should be toned down. The rank-adaptive strategy has two tunable constants (alpha and beta), and alpha is set per benchmark; that is fine for a proof of concept, but the adaptivity section is heuristic rather than a complete strategy. No code or data is shipped, so reproducibility is limited; not fatal for a numerical analysis paper, but worth noting. The conservative variant is presented more briefly than the rest, and its preservation properties are demonstrated numerically rather than proven in the same detail. Assumption 2 is the standard RK smoothness assumption, so it is not a defect; it is not checked in the Vlasov test, but that test is only about conservation, so the theorem is unaffected.

Bottom line: this is a serious numerical analysis paper with a correct-looking central theorem and honest experiments. The flaws are small and fixable. I would send it to a referee, and I would cite it. Bring it to reading group if you want to see the Galerkin-versus-tangent projection argument in action.

Referee Report

1 major / 6 minor

Summary. The paper introduces high-order Basis-Update & Galerkin (BUG) integrators constructed by applying one BUG step at each internal stage of an explicit Runge-Kutta method (RK-BUG integrators). For a fixed rank r, the authors prove a local error bound of the form h(epsilon_r + gamma_r + h^p) (Theorem 1) and a global error bound delta_r + epsilon_r + gamma_r + h^p (Theorem 2), where epsilon_r is a tangent-space projection error, gamma_r is a truncation error, and all these terms vanish at full rank. The analysis is carried out under global Lipschitz and smoothness assumptions on the vector field, and the proof uses Galerkin projection accuracy (Proposition 3), a singular-value truncation estimate (Proposition 4), and Lady Windermere's fan accumulation. Numerical experiments on Allen-Cahn, Lyapunov, discrete nonlinear Schrodinger, and Vlasov-Poisson equations confirm second-, third-, and fourth-order convergence before a plateau, demonstrate rank adaptivity, and show that the conservative variant preserves mass and momentum.

Significance. If the main theorem is correct, the paper makes a substantial contribution to dynamical low-rank integration: it provides a simple, robust, fully forward-in-time, rank-adaptive family of integrators that provably retains the order of the underlying explicit Runge-Kutta method until the error reaches the low-rank plateau. The proof is self-contained and essentially parameter-free: epsilon_r and gamma_r are defined as suprema of projection and truncation errors, and Proposition 4 derives a concrete h-independent bound for gamma_r via Eckart-Young and Horn-Johnson singular-value inequalities. The numerical study is broad, covering several benchmark equations and comparing with midpoint BUG and PRK methods. The conservative extension is useful for kinetic applications. The main caveat concerns the smoothness assumption in the local-error argument, which is discussed below.

major comments (1)

[Section 3.1, Assumption 2 and Proposition 5] Assumption 2, as stated, asserts that the first p derivatives of the exact flow exist and are continuous. The standard local-error result for an order-p explicit Runge-Kutta method, cited as Theorem 3.1 in [33], requires the exact solution (the flow as a function of the time step) to be (p+1)-times continuously differentiable with a bounded (p+1)-st derivative; continuity of only the first p derivatives yields at best o(h^p), not O(h^{p+1}), for the local error. Since Proposition 5 supplies the O(h^{p+1}) local term used in Theorems 1 and 2, the main inheritance claim in Theorem 2 needs one additional order of smoothness. Please strengthen Assumption 2 (for example, require the flow to be (p+1)-times continuously differentiable with the relevant derivatives bounded on the compact neighborhoods used in the proof) or, alternatively, weaken Proposition 5 and Theorem 2 accordingly.

minor comments (6)

[Section 3.3, Theorem 2] The statement that the constant C' is independent of r is not fully justified: the local error constant C_L in Proposition 5 is obtained from bounds on derivatives of the flow on a neighborhood of the trajectory, and that neighborhood is V_r (or a subset thereof), so C_L may depend on r unless one assumes global boundedness of the relevant derivatives. The r-independence claim should be removed or justified.
[Section 2.3 and Section 4.4] The conservative variant is claimed to preserve local conservation laws, but no proof or precise invariant-preservation theorem is given; the numerical results in Section 4.4 are only empirical evidence. A formal statement of the conditions under which mass and momentum are conserved would strengthen the paper.
[Abstract and Section 5] The phrase 'superior accuracy compared to other existing dynamical low-rank integrators' is too broad: in the Allen-Cahn comparison (Figure 2a) the midpoint BUG integrator is slightly more accurate for small ranks, and for second-order schemes the RK-BUG and PRK errors are nearly identical (Figures 2b and 5b). The claim should be moderated to the cases where it is actually observed (high orders and small step sizes on some benchmarks).
[Section 3.2, proof of Lemma 2] In the estimate for the final stage, the basis \hat{U}_{ks} is used where \hat{U}_{k+1} would be more consistent; the inequality remains valid because \hat{U}_{ks} is contained in the span of \hat{U}_{k+1}, but the notation is confusing and should be clarified.
[Section 4] The reference solution computed by the Runge-Kutta-Fehlberg method is not described in terms of order, tolerance, or implementation; specifying these details would improve reproducibility of the reported errors.
[References] Reference [33] misspells the first author's name as 'Harrier'; it should be 'Hairer'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the RK-BUG convergence theorem is a self-contained derivation under stated assumptions, with error terms defined as suprema rather than fitted quantities.

full rationale

The paper's central claim, Theorem 2, is a parameter-free derivation under Assumptions 1 and 2. The error terms delta_r, epsilon_r, and gamma_r are not fitted to the target errors: epsilon_r is defined as the supremum of the tangent-space projection error over a neighborhood V_r, and gamma_r is defined so that h*gamma_r bounds the truncation error, with Proposition 4 giving an explicit bound via the Eckart-Young theorem and the Horn-Johnson singular-value inequality. The local error bound in Theorem 1 follows from Proposition 3 (Galerkin projection accuracy), Proposition 4 (truncation error), Lemma 4 (stagewise error propagation), and the standard explicit Runge-Kutta local error estimate quoted from the external reference [33]. The global bound in Theorem 2 uses the Lady Windermere fan and the Lipschitz continuity of the exact flow proven in Lemma 3. No fitted parameter is renamed as a prediction, and no self-citation is load-bearing: citations to earlier BUG and PRK works supply context and comparison, while the key supporting inequalities are proven in the paper itself. The Vlasov-Poisson experiment is used only to verify conservation, not to claim the order theorem, so the stated regularity assumption's unverified status in that test does not create circularity. The abstract's phrase about 'superior accuracy' is a numerical comparison, not a claim forced by the derivation. Consequently, the derivation is self-contained and there is no significant circularity.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The main theorem rests on standard ODE theory (Lipschitz flow, Gronwall, local RK error), the Eckart-Young theorem and singular value interlacing. No physical or mathematical entities are invented. The only hand-chosen constants are the rank-adaptive tolerances in the numerical experiments, which do not enter the fixed-rank convergence proof.

free parameters (3)

alpha (rank-adaptive tolerance scale) = 10 (Allen-Cahn); 10^5 (Lyapunov, p=2); 10^9 (Lyapunov, p>=3); 10^2 (DNLS)
Scale of the truncation tolerance alpha*h^(p+1) in the rank-adaptive strategy (Section 3.4); chosen empirically per benchmark to keep the adaptive rank small while preserving observed convergence order.
beta (relative tolerance floor) = 1e-14
Relative tolerance floor in the rank-adaptive truncation (Section 3.4), used because h^(p+1) can become very small for high orders.
initial rank r0 = Determined per test from fixed-rank experiments
Chosen to achieve a comparable error level in the adaptive experiments (Section 4); not fixed by a general protocol.

assumptions (6)

domain assumption Assumption 1: F(t,Z) is continuous in time and Lipschitz in Z with constant L (Section 3.1)
Guarantees existence and uniqueness of the exact solution and supplies the Lipschitz constants used in all error propagation bounds.
domain assumption Assumption 2: the first p derivatives of the exact flow exist and are continuous (Section 3.1)
Standard sufficient condition for the p-th order local error of explicit Runge-Kutta methods, inherited from [33, Theorem 3.1].
standard math Existence of a compact neighbourhood V_r containing both the exact and the RK-BUG solutions, on which F is bounded by B (Section 3.1, eq. (19)-(20))
Follows from Propositions 1-2 and continuity of F; needed to make epsilon_r and gamma_r independent of h.
standard math Eckart-Young theorem: the best rank-r approximation is given by the truncated SVD (Section 3.2, eq. (27))
Used to express the truncation error as the tail sum of squared singular values.
standard math Singular value interlacing inequality sigma_{i+j-1}(A+B) <= sigma_i(A) + sigma_j(B) (Theorem 3.3.16 of [35])
Used in Proposition 4 to bound the tail singular values of the augmented solution.
standard math Gronwall inequality
Used in Propositions 1-2 and Lemma 3 to derive the a priori bounds and flow Lipschitz continuity.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Robust high-order low-rank BUG integrators based on explicit Runge--Kutta methods." pith.science (2026). https://pith.science/paper/PUZ42XW4

@misc{pith2026250207040,
  author       = {Pith},
  title        = {Pith review of: Robust high-order low-rank BUG integrators based on explicit Runge--Kutta methods},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PUZ42XW4}},
  note         = {Machine review of arXiv:2502.07040}
}

read the original abstract

In this work, we introduce high-order Basis-Update & Galerkin (BUG) integrators based on explicit Runge-Kutta methods for large-scale matrix differential equations. These dynamical low-rank integrators extend the BUG integrator to arbitrary explicit Runge-Kutta schemes by performing a BUG step at each stage of the method. The resulting Runge-Kutta BUG (RK-BUG) integrators are robust with respect to small singular values, fully forward in time, and high-order accurate, while enabling conservation and rank adaptivity. We prove that RK-BUG integrators retain the order of convergence of the underlying Runge-Kutta method until the error reaches a plateau corresponding to the low-rank truncation error, which vanishes as the rank becomes full. This theoretical analysis is supported by several numerical experiments. The results demonstrate the high-order convergence of the RK-BUG integrator and its superior accuracy compared to other existing dynamical low-rank integrators.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

Iterative thresholding low-rank time integration
math.NA 2025-07 conditional novelty 7.0 of 10

Iterative soft thresholding is combined with collocation-based time stepping to produce low-rank solutions of evolution equations whose ranks stay quasi-optimal for the achieved accuracy.

Reference graph

Works this paper leans on

37 extracted references · 36 canonical work pages · cited by 1 Pith paper

[33]

Springer, Berlin (1993)

Harrier, E., Nørsett, S., Wanner, G.: Solving Ordinary Diﬀerentia l Equations I: Nonstiﬀ Problems. Springer, Berlin (1993)

work page 1993
[1]

BIT Numerical Mathematics 62(4), 1149–1174 (2022)

Ceruti, G., Kusch, J., Lubich, C.: A rank-adaptive robust integra tor for dynamical low-rank approximation. BIT Numerical Mathematics 62(4), 1149–1174 (2022)

work page 2022
[2]

Journal of Computational P hysics 373, 545–570 (2018)

Bernard, F., Iollo, A., Riﬀaud, S.: Reduced-order model for the B GK equation based on POD and optimal transport. Journal of Computational P hysics 373, 545–570 (2018)

work page 2018
[3]

SIAM Journal on Scientiﬁc Computing 40(5), 1330– 1360 (2018)

Einkemmer, L., Lubich, C.: A low-rank projector-splitting integra tor for the Vlasov–Poisson equation. SIAM Journal on Scientiﬁc Computing 40(5), 1330– 1360 (2018)

work page 2018
[4]

Journal of Computational Physics 403, 109063 (2020)

Einkemmer, L., Ostermann, A., Piazzola, C.: A low-rank projector -splitting inte- grator for the Vlasov–Maxwell equations with divergence correct ion. Journal of Computational Physics 403, 109063 (2020)

work page 2020
[5]

Journal of Computa tional Physics 443, 110495 (2021)

Einkemmer, L., Joseph, I.: A mass, momentum, and energy conse rvative dynam- ical low-rank scheme for the Vlasov equation. Journal of Computa tional Physics 443, 110495 (2021)

work page 2021
[6]

Journal of Computational Phys ics 470, 111590 (2022)

Coughlin, J., Hu, J.: Eﬃcient dynamical low-rank approximation for the Vlasov- Ampere-Fokker-Planck system. Journal of Computational Phys ics 470, 111590 (2022)

work page 2022
[7]

Journal of Computational Ph ysics 501, 112757 (2024)

Einkemmer, L.: Accelerating the simulation of kinetic shear Alfv´ en waves with a dynamical low-rank approximation. Journal of Computational Ph ysics 501, 112757 (2024)

work page 2024

Show all 37 references

[8]

Journal of Computational Physics, 114191 (2025)

Einkemmer, L., Kormann, K., Kusch, J., McClarren, R.G., Qiu, J.-M.: A review of low-rank methods for time-dependent kinetic simulations. Journal of Computational Physics, 114191 (2025)

2025
[9]

Physica D: Nonlinear Phenome na 238(23-24), 2347–2360 (2009)

Sapsis, T.P., Lermusiaux, P.F.: Dynamically orthogonal ﬁeld equatio ns for contin- uous stochastic dynamical systems. Physica D: Nonlinear Phenome na 238(23-24), 2347–2360 (2009)

2009
[10]

Journal of Computatio nal Physics 344, 303–319 (2017)

Babaee, H., Choi, M., Sapsis, T.P., Karniadakis, G.E.: A robust bi- orthogonal/dynamically-orthogonal method using the covariance pseudo-inverse with application to stochastic ﬂow problems. Journal of Computatio nal Physics 344, 303–319 (2017)

2017
[11]

Jour- nal of Computational Physics 354, 135–162 (2018)

Musharbash, E., Nobile, F.: Dual dynamically orthogonal approx imation of incompressible Navier Stokes equations with random boundary cond itions. Jour- nal of Computational Physics 354, 135–162 (2018)

2018
[12]

Siam Rev iew 60(3), 595– 625 (2018)

Feppon, F., Lermusiaux, P.F.: Dynamically orthogonal numerical schemes for 38 eﬃcient stochastic advection and Lagrangian transport. Siam Rev iew 60(3), 595– 625 (2018)

2018
[13]

BIT Num erical Mathematics 60(4), 1153–1201 (2020)

Musharbash, E., Nobile, F., Vidliˇ ckov´ a, E.: Symplectic dynamica l low rank approximation of wave equations with random parameters. BIT Num erical Mathematics 60(4), 1153–1201 (2020)

2020
[14]

Journal of Computational Physics 415, 109511 (2020)

Patil, P., Babaee, H.: Real-time reduced-order modeling of stoch astic partial diﬀerential equations via time-dependent subspaces. Journal of Computational Physics 415, 109511 (2020)

2020
[15]

Stoc hastics and Partial Diﬀerential Equations: Analysis and Computations 9(3), 603–629 (2021)

Kazashi, Y., Nobile, F.: Existence of dynamical low rank approxima tions for ran- dom semi-linear evolutionary equations on the maximal interval. Stoc hastics and Partial Diﬀerential Equations: Analysis and Computations 9(3), 603–629 (2021)

2021
[16]

Numerische Mathematik 149, 973–1024 (2021)

Kazashi, Y., Nobile, F., Vidliˇ ckov´ a, E.: Stability properties of a projector-splitting scheme for dynamical low rank approximation of random parabolic eq uations. Numerische Mathematik 149, 973–1024 (2021)

2021
[17]

Mathematics of Computation 94(353), 1335– 1375 (2025)

Kazashi, Y., Nobile, F., Zoccolan, F.: Dynamical low-rank approxim ation for stochastic diﬀerential equations. Mathematics of Computation 94(353), 1335– 1375 (2025)

2025
[18]

SIAM Journal on Matrix Analysis and A pplications 32(4), 1288–1316 (2011)

Kressner, D., Tobler, C.: Low-rank tensor Krylov subspace me thods for parametrized linear systems. SIAM Journal on Matrix Analysis and A pplications 32(4), 1288–1316 (2011)

2011
[19]

PAMM 18(1), 201800178 (2018)

Weinhandl, R., Benner, P., Richter, T.: Linear Low-Rank Paramet er-Dependent Fluid-Structure Interaction Discretization in 2D. PAMM 18(1), 201800178 (2018)

2018
[20]

ZAMM- Journal of Applied Mathematics and Mechanics/Zeitschrift f¨ ur An gewandte Mathematik und Mechanik 104(10), 202300562 (2024)

Benner, P., Richter, T., Weinhandl, R.: A low-rank method for par ameter- dependent ﬂuid-structure interaction discretizations with hyper elasticity. ZAMM- Journal of Applied Mathematics and Mechanics/Zeitschrift f¨ ur An gewandte Mathematik und Mechanik 104(10), 202300562 (2024)

2024
[21]

Journal of Scientiﬁc Computing 99(2), 34 (2024)

Riﬀaud, S., Fern´ andez, M.A., Lombardi, D.: A Low-Rank Solver fo r Parameter Estimation and Uncertainty Quantiﬁcation in Time-Dependent Syste ms of Partial Diﬀerential Equations. Journal of Scientiﬁc Computing 99(2), 34 (2024)

2024
[22]

SIAM Jo urnal on Matrix Analysis and Applications 29(2), 434–454 (2007)

Koch, O., Lubich, C.: Dynamical low-rank approximation. SIAM Jo urnal on Matrix Analysis and Applications 29(2), 434–454 (2007)

2007
[23]

BIT Numerical Mathematics 54(1), 171–188 (2014)

Lubich, C., Oseledets, I.V.: A projector-splitting integrator fo r dynamical low- rank approximation. BIT Numerical Mathematics 54(1), 171–188 (2014)

2014
[24]

SIAM Journal on Numerical Analysis 39 54(2), 1020–1038 (2016)

Kieri, E., Lubich, C., Walach, H.: Discretized dynamical low-rank ap proximation in the presence of small singular values. SIAM Journal on Numerical Analysis 39 54(2), 1020–1038 (2016)

2016
[25]

SIAM Journal on Scientiﬁc Computing 45(1), 1–24 (2023)

Kusch, J., Einkemmer, L., Ceruti, G.: On the stability of robust dy namical low-rank approximations for hyperbolic problems. SIAM Journal on Scientiﬁc Computing 45(1), 1–24 (2023)

2023
[26]

BIT Numerical Mathematics 62(1), 23–44 (2022)

Ceruti, G., Lubich, C.: An unconventional robust integrator fo r dynamical low- rank approximation. BIT Numerical Mathematics 62(1), 23–44 (2022)

2022
[27]

SIAM Journal on Scientiﬁc Computing 46(3), 205–228 (2024)

Ceruti, G., Kusch, J., Lubich, C.: A parallel rank-adaptive integr ator for dynam- ical low-rank approximation. SIAM Journal on Scientiﬁc Computing 46(3), 205–228 (2024)

2024
[28]

BIT Numerical Mathemat ics 64(3), 30 (2024)

Ceruti, G., Einkemmer, L., Kusch, J., Lubich, C.: A robust second -order low-rank BUG integrator based on the midpoint rule. BIT Numerical Mathemat ics 64(3), 30 (2024)

2024
[29]

BIT Numerical Mathematics 65(3), 31 (2025)

Kusch, J.: Second-order robust parallel integrators for dyn amical low-rank approximation. BIT Numerical Mathematics 65(3), 31 (2025)

2025
[30]

Computational Methods in App lied Mathematics 19(1), 73–92 (2019)

Kieri, E., Vandereycken, B.: Projection methods for dynamical low-rank approx- imation of high-dimensional problems. Computational Methods in App lied Mathematics 19(1), 73–92 (2019)

2019
[31]

arXiv preprint arXiv:2312.00172 ( 2023)

Carrel, B., Vandereycken, B.: Projected exponential method s for stiﬀ dynamical low-rank approximation problems. arXiv preprint arXiv:2312.00172 ( 2023)

2023 arXiv
[32]

Ava ilable at SSRN 4668132 (2023)

Einkemmer, L.D., Kusch, J., Schotth¨ ofer, S.: Conservation pr operties of the aug- mented basis update & Galerkin integrator for kinetic problems. Ava ilable at SSRN 4668132 (2023)

2023
[34]

Psychometrika 1(3), 211–218 (1936)

Eckart, C., Young, G.: The approximation of one matrix by anoth er of lower rank. Psychometrika 1(3), 211–218 (1936)

1936
[35]

Cambridge univ ersity press, Cambridge (1994)

Horn, R.A., Johnson, C.R.: Topics in Matrix Analysis. Cambridge univ ersity press, Cambridge (1994)

1994
[36]

SIAM Journal on Matrix Analysis and Applications 46(2), 1587–1615 (2025)

Lam, H.Y., Ceruti, G., Kressner, D.: Randomize Low-Rank Runge– Kutta Meth- ods. SIAM Journal on Matrix Analysis and Applications 46(2), 1587–1615 (2025)

2025
[37]

Journal o f Computational Physics 335, 283–310 (2017) 40

Mattsson, K.: Diagonal-norm upwind SBP operators. Journal o f Computational Physics 335, 283–310 (2017) 40

2017

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [33]

Springer, Berlin (1993)

Harrier, E., Nørsett, S., Wanner, G.: Solving Ordinary Diﬀerentia l Equations I: Nonstiﬀ Problems. Springer, Berlin (1993)

work page 1993

[2] [1]

BIT Numerical Mathematics 62(4), 1149–1174 (2022)

Ceruti, G., Kusch, J., Lubich, C.: A rank-adaptive robust integra tor for dynamical low-rank approximation. BIT Numerical Mathematics 62(4), 1149–1174 (2022)

work page 2022

[3] [2]

Journal of Computational P hysics 373, 545–570 (2018)

Bernard, F., Iollo, A., Riﬀaud, S.: Reduced-order model for the B GK equation based on POD and optimal transport. Journal of Computational P hysics 373, 545–570 (2018)

work page 2018

[4] [3]

SIAM Journal on Scientiﬁc Computing 40(5), 1330– 1360 (2018)

Einkemmer, L., Lubich, C.: A low-rank projector-splitting integra tor for the Vlasov–Poisson equation. SIAM Journal on Scientiﬁc Computing 40(5), 1330– 1360 (2018)

work page 2018

[5] [4]

Journal of Computational Physics 403, 109063 (2020)

Einkemmer, L., Ostermann, A., Piazzola, C.: A low-rank projector -splitting inte- grator for the Vlasov–Maxwell equations with divergence correct ion. Journal of Computational Physics 403, 109063 (2020)

work page 2020

[6] [5]

Journal of Computa tional Physics 443, 110495 (2021)

Einkemmer, L., Joseph, I.: A mass, momentum, and energy conse rvative dynam- ical low-rank scheme for the Vlasov equation. Journal of Computa tional Physics 443, 110495 (2021)

work page 2021

[7] [6]

Journal of Computational Phys ics 470, 111590 (2022)

Coughlin, J., Hu, J.: Eﬃcient dynamical low-rank approximation for the Vlasov- Ampere-Fokker-Planck system. Journal of Computational Phys ics 470, 111590 (2022)

work page 2022

[8] [7]

Journal of Computational Ph ysics 501, 112757 (2024)

Einkemmer, L.: Accelerating the simulation of kinetic shear Alfv´ en waves with a dynamical low-rank approximation. Journal of Computational Ph ysics 501, 112757 (2024)

work page 2024