Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Multilevel Stochastic Gradient Descent for Optimal Control Under Uncertainty

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper's central claim: multilevel Monte Carlo gradient estimation turns stochastic gradient descent into a linearly convergent method for optimal control under uncertainty, at MLMC-level cost instead of full fine-grid gradient cost.

desk verdict Solid engineering with a proof gap: the linear-convergence claim holds only for an oracle schedule, but the MLMC-in-SGD idea and the numerics are worth taking seriously. read the letter →

arxiv 2506.02647 v2 pith:3A52MSKG submitted 2025-06-03 math.OC cs.MScs.NAmath.NA

Niklas Baumgarten , David Schneiderhan This is my paper
classification math.OCcs.MScs.NAmath.NA MSC 49K2065K0565C0590C15
keywords multilevelMonteCarlostochasticgradientdescentoptimalcontrolunderuncertaintyPDE-constrainedoptimizationlinearconvergencelog-normaldiffusioncoefficientbudgetedadaptivestepsizes
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 proposes a stochastic gradient descent (SGD) method for PDE-constrained optimal control under uncertainty, in which the gradient is estimated by a multilevel Monte Carlo (MLMC) telescoping sum rather than a plain Monte Carlo batch. The claim is that this replacement preserves the fast linear convergence rate $O(\rho^K)$ usually associated with sample-average approximation, while paying only MLMC-level cost and never forming the full gradient on the finest discretization. If true, it means high-dimensional uncertain-control problems can be solved to a given accuracy with substantially less compute than standard batched SGD, and the method can be wrapped in a budgeted version that converts CPU-time and memory limits into explicit error bounds. The paper supports the claim with a convergence theorem, an error lemma controlling the gradient stochastic error by level-wise sample sizes, and numerical experiments on a two-dimensional elliptic diffusion problem with log-normal coefficients.

What carries the argument

The load-bearing object is the multilevel Monte Carlo gradient estimator $E^{ML}[g_L^{(k)}] = \lambda z_L^{(k)} - \sum_{\ell=0}^L P_L^\ell E_{M_\ell}^{MC}[p_\ell^{(k)}]$, a telescoping sum over finite-element levels of independent Monte Carlo averages of adjoint level differences $p_\ell^{(k)} = q_\ell^{(k)} - P_\ell^{\ell-1} q_{\ell-1}^{(k)}$. It is what lets the algorithm control the gradient error's sampling part through the level-wise sample counts $M_\ell$ and its bias part through the finest level $L$. The contraction in Theorem 5.2 comes from Lemma 5.1's error bound, which reduces the gradient estimation error to the two MLMC error terms plus the control-discretization inequality (2.15); choosing the per-step accuracy $\epsilon_k$ as a fixed fraction of the current error $e_k$ turns the descent estimate into $e_{k+1}^2 \le \rho_k^2 e_k^2$ with $\rho_k < 1$.

What would settle it

For a given PDE-control pair, run MLSGD on a sequence of refined meshes and measure both the per-step error ratio $e_{k+1}/e_k$ and the per-step cost: if no admissible sequence of batch sizes and step sizes keeps the ratio below 1 uniformly while cost follows Lemma 5.1's bounds, or if the exponent $\alpha_z$ estimated from (2.15) changes with the optimization step $k$, then Theorem 5.2 does not hold for that problem.

Watch

Extended reading notes

Core claim

The paper's central result, Theorem 5.2, states that for a strongly convex optimal control problem with Lipschitz gradients there exists a sequence of multilevel batch sizes $\{M_{k,\ell}\}$ and step sizes $\{t_k\}$ for which the MLSGD iterates satisfy $\|z_L^{(K)} - z^*\|_{L^2(\Omega,W)} = O(\rho^K)$ for some $\rho \in (0,1)$. Reaching error below $\epsilon$ then costs $C_\epsilon \sim \epsilon^{-2}$ when the variance-decay exponent $\beta$ exceeds the cost exponent $\gamma$, with the standard MLMC rates $\epsilon^{-2}(\log\epsilon)^2$ and $\epsilon^{-2-(\gamma-\beta)/\alpha}$ in the balanced and variance-dominated cases. This is the convergence rate of an SAA-style method at the cost of a multilevel Monte Carlo estimator, not the cost of a full fine-grid gradient evaluation. Corollary 5.4 extends the same linear rate to the objective error and the estimated gradient norm, and Corollary 6.2 states error bounds in terms of the CPU-time and memory budgets.

Load-bearing premise

The weakest assumption is inequality (2.15): at every optimization step the squared error of the discretized control must be bounded by a constant times the adjoint discretization error, with exponent and constants independent of the step $k$, and the proof additionally needs the oracle choice $\epsilon_k = \eta e_k$ so that the algorithm knows the very error it is bounding.

Editorial extensions

If this is right

  • MLSGD breaks the standard SGD trade-off: it converges linearly per optimization step while avoiding full-gradient cost on the finest mesh, making high-accuracy controls reachable with less compute.
  • For a prescribed accuracy $\epsilon$, total cost follows MLMC complexity: $\epsilon^{-2}$ when $\beta>\gamma$, $\epsilon^{-2}(\log\epsilon)^2$ when $\beta=\gamma$, and $\epsilon^{-2-(\gamma-\beta)/\alpha}$ when $\beta<\gamma$.
  • The budgeted version (BMLSGD) turns user-specified CPU-time and memory budgets into an error certificate, with convergence rate $\delta = \min\{1/2,\, \alpha/(2\alpha+(\gamma-\beta))\}$ with respect to resources.
  • Numerically, BMLSGD reaches BSGD-level accuracy roughly 18x faster and ends with errors about 5x smaller at equal cost, and it scales from 64 to 1024 CPUs in the reported experiments.
  • The methodology is not tied to the elliptic example; it only needs Assumptions 2.2 and 2.3, so it is intended to carry over to other PDE constraints and higher dimensions.

Reading between the lines

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

  • The proof relies on an oracle choice, $\epsilon_k = \eta e_k$, which requires knowing the very error that is being bounded; Algorithm 3 instead targets the estimated gradient norm, so proving that the gradient-norm criterion preserves the same complexity would be a direct test of how much of Theorem 5.2 survives in practice.
  • Because the gradient error bound only uses the MLMC estimator's bias-variance split, the same multilevel estimator could likely be attached to other stochastic optimizers (e.g. L-BFGS or ADAM), but the paper does not analyze those combinations; a convergence analysis would be needed before claiming the same rates.
  • If inequality (2.15) fails for a less regular PDE-control pair—say a hyperbolic or convection-dominated problem—the per-step contraction $\rho_k$ may not exist; estimating $\alpha_z$ and the $k$-dependence of $z_k$ online, as the paper does for its other exponents, would be a practical way to check whether the method remains applicable.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes a multilevel Monte Carlo (MLMC) version of stochastic gradient descent for PDE-constrained optimal control under uncertainty. It presents three algorithms: a batched SGD baseline, an MLSGD method using an MLMC gradient estimator, and a budgeted adaptive variant (BMLSGD) that respects CPU-time and memory constraints. The main theoretical claim is linear convergence in the number of optimization steps, with MLMC-level cost rather than the cost of full fine-grid gradients. The paper also contains extensive numerical experiments for a two-dimensional elliptic diffusion problem with log-normal coefficients, comparing MLSGD/BMLSGD with batched SGD and demonstrating speedups and parallel scaling.

Significance. If fully established, the main result would be a meaningful advance: it would combine the linear iteration complexity of sample-average approximation with the lower per-iteration cost of multilevel Monte Carlo gradient estimation. The paper also contributes a detailed, hardware-aware algorithmic framework and substantial HPC experiments, which are genuine strengths. However, as detailed below, the proof of the central linear-convergence claim currently relies on an oracle choice of error targets that is not implemented in the presented algorithms, so the significance of the result is conditional on closing that gap.

major comments (4)
  1. [Theorem 5.2, Eq. (5.8), Remark 5.3(2), Algorithm 3] The proof of linear convergence chooses epsilon_k = eta e_k, where e_k = ||z_L^(k) - z*|| is the very error being bounded. This makes the batch schedule in Theorem 5.2 depend on the unknown solution z* and on the trajectory; it is an existence result for a non-implementable oracle schedule, not a convergence proof for Algorithms 2 or 3. Algorithm 2 takes a fixed batch schedule as input, and Algorithm 3 sets epsilon_{k+1} = eta ||E_ML[g_L^(k)]||_W. To support the abstract's claim that 'the method exhibits linear convergence,' the authors must either prove that the implemented epsilon-rule yields a contraction (e.g., via a comparison lemma between ||E_ML[g_L^(k)]||_W and e_k) or explicitly restrict the claim to the oracle schedule and state that the implemented algorithm is heuristic.
  2. [Corollary 5.4] The proof of Corollary 5.4 is asserted rather than shown: it says the result 'follows from bounding ... with similar arguments as in Theorem 5.2.' In particular, the bound on ||E_ML[g_L^(K)]||_W is derived from the same oracle choice epsilon_k = eta e_k; no argument shows that the implemented update with epsilon_k = eta ||E_ML[g_L^(k-1)]||_W maintains the proportionality epsilon_k proportional to e_k. Without this, the recursion may plateau at a fixed point of size O(epsilon_k)/mu instead of contracting linearly. A complete proof of the corollary, or a clearly stated weaker claim, is needed.
  3. [Assumption 2.3, Eq. (2.15)] The inequality ||z_L^(k) - z^(k)||^2_{L^2_k(Omega,W)} <= c_G E_k[||q_L^(k) - q^(k)||^2_W] with k-independent c_G is the bridge that lets Lemma 5.1 bound the control-discretization component of the gradient error. It is asserted rather than derived, and it is not numerically verified: the experiments in Section 6.2 estimate the MLMC exponents alpha_q, beta_p, gamma_CT, and gamma_Mem, but not the constant c_G or the inequality (2.15) itself. Because Lemma 5.1 and Theorem 5.2 depend on this assumption, the paper should either prove it for the model problem, verify it computationally, or clearly state the failure mode if it does not hold.
  4. [Algorithm 3, global constants; Remark 5.3(2)] The theoretical contraction condition derived from (5.8) requires eta < mu/(sqrt(2)+1+sqrt(c_G) lambda). For the numerical example, lambda = 10^{-8} and the map z -> u is compact, so the strong convexity constant mu is of order lambda; this forces eta to be of order 10^{-8} for the theorem to apply. Algorithm 3, however, uses the 'tested choice' eta = 0.9. The numerical experiments therefore appear to be outside the parameter regime covered by Theorem 5.2. The paper should reconcile this discrepancy, for example by computing or estimating the relevant constants, or by explicitly presenting the choices in Algorithm 3 as a heuristic that is not covered by the convergence proof.
minor comments (4)
  1. [Eq. (5.8)] The coefficient of t_k^2 in (5.8) should be 4c_Lip^2(1+eta^2 c_G) + 4eta^2 rather than 4c_Lip^2(1+eta^2 c_G) + 2eta^2, because the term 2t_k^2||r_L^(k)||^2 contributes 4eta^2 when Lemma 5.1 is inserted. This appears to be a typo; the conclusion is unchanged.
  2. [Title page] The author name on the title page reads 'DA VID SCHNEIDERHAN'; this should be corrected to 'DAVID SCHNEIDERHAN'.
  3. [Corollary 6.2] The proof of Corollary 6.2 is a single sentence referring to arguments in [13] and [14]. Given that the corollary is used to frame the numerical convergence-rate estimates in Section 6.2, the proof sketch should be expanded or explicitly marked as a heuristic resource-accuracy relation.
  4. [Section 5.3, Theorem 5.2] The proof uses L^2_k(Omega,W) norms for the per-step errors and the final step uses L^2(Omega,W); the tower property is invoked but the transition is not written out. Making this step explicit would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity. The central convergence proof is self-contained under its stated assumptions; the oracle choice epsilon_k = eta e_k and the Corollary 5.4 replacement are correctness/rigor gaps, not definitional reductions.

full rationale

The derivation chain is not circular. Lemma 5.1 starts from the explicitly stated Assumption 2.3, including the MLMC bias and variance rates and the control-error bound (2.15), and from the standard MLMC error decomposition (4.2); it then selects batch sizes against a target epsilon_k and obtains the bounds (5.5) with the cost rates (5.6). Theorem 5.2 combines these bounds with the assumed strong convexity and Lipschitz continuity to derive the contraction inequality (5.8). The rates and cost exponents are genuine consequences of these inequalities: no parameter is fitted to data, and no quantity in the final error bound is set equal to an input by construction. The proof's choice epsilon_k = eta e_k is self-referential in that the schedule depends on the unknown error e_k, and the paper does not show that the computable target epsilon_k = eta ||E_ML[g]||_W used in Algorithm 3 satisfies the same condition. However, this is a non-constructive-existence and omitted-argument issue rather than an equivalence-by-construction. Corollary 5.4 is asserted with a sketch, so the practical Algorithm 3's linear convergence is not fully justified, but again this is a rigor gap, not circular reasoning. Self-citations [12,13,14] supply prior algorithmic tools and assumptions; the present theorem does not reduce to those citations, and it is not built on a uniqueness claim or a renamed known result. For these reasons, the appropriate circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central theorem rests on standard convexity and FE assumptions plus two paper-specific bridges: equation (2.15) linking control and adjoint errors, and the oracle availability of the current error for batch sizing. No new physical entities are introduced. The hand-chosen hyperparameters eta, theta, initial batches and t0 affect experimental behavior but not the existence proof.

free parameters (4)
  • eta (error target reduction factor) = 0.9
    Hand-specified in Algorithm 3 (global const eta=0.9). It controls the target gradient error epsilon_k = eta * ||E_ML[g]||_W. Not fitted to data; affects practical convergence speed.
  • theta (bias-variance tradeoff) = 0.5
    Hand-specified in Algorithm 3. Splits the allowed gradient error between sampling error and discretization bias in Lemma 5.1. Not fitted to data.
  • initial multilevel batch sizes (M_0,M_1,M_2) = 64, 16, 4 on levels h=2^-4,2^-5,2^-6
    Hand-chosen initial configuration for Algorithm 3 in Section 6.2; the algorithm adapts them later. Not fitted.
  • initial step size t0 = 200
    Hand-chosen for the adaptive step size rule (6.5) in Section 6.2; problem-specific.
assumptions (4)
  • domain assumption The objective J is strongly mu-convex and Lipschitz-smooth (Assumption 2.2).
    Standard convex optimization assumptions ensuring a unique minimizer and convergence of projected gradient steps; stated in Assumption 2.2.
  • domain assumption FE discretizations satisfy the MLMC error and cost estimates (2.11)-(2.14) with k-independent exponents and k-dependent constants.
    A priori FE error estimates assumed to hold uniformly along the optimization trajectory; used to bound the bias and sampling error in Lemma 5.1.
  • ad hoc to paper Control discretization error is dominated by adjoint discretization error: ||z_L^(k)-z^(k)||^2 <= c_G E_k[||q_L^(k)-q^(k)||^2] (2.15).
    This is a paper-specific bridge assumption encoding operator information in c_G; it makes the lambda(z_L-z) term controllable by the adjoint error. If it fails, Lemma 5.1 and Theorem 5.2 collapse.
  • ad hoc to paper The unknown current error e_k is available to choose epsilon_k = eta e_k in Theorem 5.2.
    The proof of Theorem 5.2 chooses batch sizes based on the error it is trying to bound; the implementable algorithms do not have access to e_k and use the gradient norm instead. This is an unstated oracle assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multilevel Stochastic Gradient Descent for Optimal Control Under Uncertainty." pith.science (2026). https://pith.science/paper/3A52MSKG

@misc{pith2026250602647,
  author       = {Pith},
  title        = {Pith review of: Multilevel Stochastic Gradient Descent for Optimal Control Under Uncertainty},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3A52MSKG}},
  note         = {Machine review of arXiv:2506.02647}
}
read the original abstract

We present a multilevel stochastic gradient descent method for the optimal control of systems governed by partial differential equations under uncertain input data. The gradient descent method used to find the optimal control leverages a parallel multilevel Monte Carlo method as stochastic gradient estimator. As a result, we achieve precise control over the stochastic gradient's bias, introduced by numerical approximation, and its sampling error, arising from the use of incomplete gradients, while optimally managing computational resources. We show that the method exhibits linear convergence in the number of optimization steps while avoiding the cost of computing the full gradient at the highest fidelity. Numerical experiments demonstrate that the method significantly outperforms the standard (mini-) batched stochastic gradient descent method in terms of convergence speed and accuracy. The method is particularly well-suited for high-dimensional control problems, taking advantage of parallel computing resources and a distributed multilevel data structure. Additionally, we evaluate and implement different step size strategies, optimizer schemes, and budgeting techniques. The method's performance is studied using a two-dimensional elliptic subsurface diffusion problem with log-normal coefficients and Mat\'ern covariance.

Figures

Figures reproduced from arXiv: 2506.02647 by the authors.

Figure 1
Figure 1. Left to right: Two GRF samples y (m) ℓ on different mesh diameters; Computed control z (k) ℓ after k = 10 and k = 100 iterations of Algorithm 1. 3.1 Algorithm At its core, BSGD generates, such as (2.8), a minimizing sequence {z (k) ℓ } K k=0 ⊂ Zℓ of approx￾imations to find the optima of Problem 2.1. We refer to Algorithm 1 for a high-level functional pseudocode generating this sequence. BSGD function. The algorithm … view at source ↗
Figure 2
Figure 2. Comparison of different batch sizes M. total computing time. A higher noise level in J MC M (z (k) ℓ ) is observed for smaller batch sizes M, yet all methods oscillate around the same value. The impact of the batch size on the gradient norm is even more pronounced: the optimality condition (2.6) is visibly better satisfied with larger batches, but this comes at the cost of an increased computational effort. Beyond t… view at source ↗
Figure 3
Figure 3. Comparison of MC and MLMC gradient estimation [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Comparison of different multilevel batch sizes {Mℓ} L ℓ=0. batch, the better is (2.6) satisfied due to the reduced variance, but this time also due to the reduced bias. Yet, the higher quality of the computational results naturally comes at increased cost. Discussion a…
Figure 5
Figure 5. Figure 5: Comparison of Algorithm 1 with Algorithm 3. the state (dash-dotted line) and the adjoint solutions (dashed line) in the left and center plots, respectively. The rightmost plot shows the measured computational cost exponents γ CT and γ Mem in a dual-axis format: the inc…
Figure 6
Figure 6. Figure 6: Comparison of tk ≡ 100 and tk ≡ 150 with the adaptive step size rule (6.5) [PITH_FULL_IMAGE:figures/full_fig_p023_6.png]
Figure 7
Figure 7. Figure 7: CPU-scaling experiment with |P| = 64, |P| = 256 and |P| = 1024. rate in the number of optimization steps. The total error is controlled via multilevel Monte Carlo estimation, as established in Lemma 5.1, Theorem 5.2, and Corollary 5.4. Our numerical experiments in Sect…
Figure 8
Figure 8. Figure 8: Outlook on three-dimensional random field sampling and optimal control. on [14], enables efficient usage of resource. It is thus applicable to three-dimensional problems, as illustrated by the outlook application in [PITH_FULL_IMAGE:figures/full_fig_p024_8.png]

Discussion (0). Sign in 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

  1. Multilevel Stochastic Gradient Descent for Risk-Averse PDE-Constrained Optimization

    math.OC 2026-06 unverdicted novelty 5.0 of 10

    Develops and analyzes multilevel SGD using adaptive MLMC gradients for risk-averse 3D elliptic PDE optimization, claiming better scalability, convergence rates, and complexity than standard batched SGD.

Reference graph

Works this paper leans on

49 extracted references · 46 canonical work pages · cited by 1 Pith paper

  1. [22]

    Robust optimization of pdes with random coefficients using a multilevel monte carlo method,

    A. Van Barel and S. Vandewalle, “Robust optimization of pdes with random coefficients using a multilevel monte carlo method,” SIAM/ASA Journal on Uncertainty Quantification, vol. 7, no. 1, pp. 174–202, 2019

    work page 2019

  2. [24]

    Multi-level stochastic approximation algorithms,

    N. Frikha, “Multi-level stochastic approximation algorithms,” The Annals of Applied Probability, vol. 26, no. 2, pp. 933 – 985, 2016

    work page 2016

  3. [1]

    Projected stochastic gradients for convex constrained problems in Hilbert spaces,

    C. Geiersbach and G. C. Pflug, “Projected stochastic gradients for convex constrained problems in Hilbert spaces,” SIAM J. Optim., vol. 29, no. 3, pp. 2079–2099, 2019

    work page 2019

  4. [2]

    A stochastic gradient method for a class of nonlinear pde-constrained optimal control problems under uncertainty,

    C. Geiersbach and T. Scarinci, “A stochastic gradient method for a class of nonlinear pde-constrained optimal control problems under uncertainty,” Journal of Differential Equations, vol. 364, pp. 635–666, 2023

    work page 2023

  5. [3]

    A stochastic gradient method with mesh refinement for PDE-constrained optimization under uncertainty,

    C. Geiersbach and W. Wollner, “A stochastic gradient method with mesh refinement for PDE-constrained optimization under uncertainty,” SIAM J. Sci. Comput., vol. 42, no. 5, pp. A2750–A2772, 2020

    work page 2020

  6. [4]

    Complexity analysis of stochastic gradient methods for PDE- constrained optimal control problems with uncertain parameters,

    M. Martin, S. Krumscheid, and F. Nobile, “Complexity analysis of stochastic gradient methods for PDE- constrained optimal control problems with uncertain parameters,” ESAIM Math. Model. Numer. Anal., vol. 55, no. 4, pp. 1599–1633, 2021

    work page 2021

  7. [5]

    A stochastic gradient algorithm with momentum terms for optimal control problems governed by a convection–diffusion equation with random diffusivity,

    S. C. Toraman and H. Y¨ ucel, “A stochastic gradient algorithm with momentum terms for optimal control problems governed by a convection–diffusion equation with random diffusivity,” Journal of Computational and Applied Mathematics, vol. 422, p. 114919, 2023

    work page 2023

  8. [6]

    Multilevel Monte Carlo methods and applications to elliptic PDEs with random coefficients,

    K. A. Cliffe, M. B. Giles, R. Scheichl, and A. L. Teckentrup, “Multilevel Monte Carlo methods and applications to elliptic PDEs with random coefficients,” Comput. Vis. Sci., vol. 14, no. 1, pp. 3–15, 2011

    work page 2011

Show all 49 references
  1. [7]

    Multilevel Monte Carlo methods,

    M. B. Giles, “Multilevel Monte Carlo methods,” Acta Numer., vol. 24, pp. 259–328, 2015. MULTILEVEL STOCHASTIC GRADIENT DESCENT 25

    2015

  2. [8]

    A multilevel stochastic collocation algorithm for optimization of PDEs with uncertain coef- ficients,

    D. P. Kouri, “A multilevel stochastic collocation algorithm for optimization of PDEs with uncertain coef- ficients,” SIAM/ASA J. Uncertain. Quantif., vol. 2, no. 1, pp. 55–81, 2014

    2014

  3. [9]

    A combination technique for optimal control problems constrained by random pdes,

    F. Nobile and T. Vanzan, “A combination technique for optimal control problems constrained by random pdes,” SIAM/ASA Journal on Uncertainty Quantification, vol. 12, no. 2, pp. 693–721, 2024

    2024

  4. [10]

    A quasi-Monte Carlo method for optimal control under uncertainty,

    P. A. Guth, V. Kaarnioja, F. Y. Kuo, C. Schillings, and I. H. Sloan, “A quasi-Monte Carlo method for optimal control under uncertainty,” SIAM/ASA J. Uncertain. Quantif., vol. 9, no. 2, pp. 354–383, 2021

    2021

  5. [11]

    Parabolic pde-constrained optimal control under uncertainty with entropic risk measure using quasi-monte carlo integration - numerische mathematik,

    P. A. Guth, V. Kaarnioja, F. Y. Kuo, C. Schillings, and I. H. Sloan, “Parabolic pde-constrained optimal control under uncertainty with entropic risk measure using quasi-monte carlo integration - numerische mathematik,” SpringerLink, Mar 2024

    2024

  6. [12]

    Baumgarten, A Fully Parallelized and Budgeted Multi-level Monte Carlo Framework for Partial Differential Equations

    N. Baumgarten, A Fully Parallelized and Budgeted Multi-level Monte Carlo Framework for Partial Differential Equations. PhD thesis, Karlsruher Institut f¨ ur Technologie (KIT), 2023

    2023

  7. [13]

    A fully parallelized and budgeted multilevel monte carlo method and the application to acoustic waves,

    N. Baumgarten, S. Krumscheid, and C. Wieners, “A fully parallelized and budgeted multilevel monte carlo method and the application to acoustic waves,” SIAM/ASA Journal on Uncertainty Quantification, vol. 12, no. 3, pp. 901–931, 2024

    2024

  8. [14]

    A budgeted multi-level monte carlo method for full field estimates of multi-pde problems,

    N. Baumgarten, R. Kutri, and R. Scheichl, “A budgeted multi-level monte carlo method for full field estimates of multi-pde problems,” 2025

    2025

  9. [15]

    Efficient mini-batch stochastic gradient descent with centroidal voronoi tessellation for pde-constrained optimization under uncertainty,

    L. Chen, M. Xiong, J. Ming, and X. He, “Efficient mini-batch stochastic gradient descent with centroidal voronoi tessellation for pde-constrained optimization under uncertainty,”Physica D: Nonlinear Phenomena, vol. 467, p. 134216, 2024

    2024

  10. [16]

    Optimization of piecewise smooth shapes under uncertainty using the example of navier-stokes flow,

    C. Geiersbach, T. Suchan, and K. Welker, “Optimization of piecewise smooth shapes under uncertainty using the example of navier-stokes flow,” arXiv preprint arXiv:2308.07742, 2023

    2023 arXiv

  11. [17]

    Dirichlet-neumann averaging: The dna of efficient gaussian process simulation,

    R. Kutri and R. Scheichl, “Dirichlet-neumann averaging: The dna of efficient gaussian process simulation,” arXiv preprint arXiv:2412.07929, 2024

    2024 arXiv

  12. [18]

    Caflisch and Y

    R. Caflisch and Y. Yang, Adjoint Monte Carlo Method, pp. 461–505. Cham: Springer Nature Switzerland, 2024

    2024

  13. [19]

    Multilevel monte carlo analysis for optimal control of elliptic pdes with random coefficients,

    A. A. Ali, E. Ullmann, and M. Hinze, “Multilevel monte carlo analysis for optimal control of elliptic pdes with random coefficients,” SIAM/ASA Journal on Uncertainty Quantification, vol. 5, no. 1, pp. 466–492, 2017

    2017

  14. [20]

    A multigrid solver for pde-constrained optimization with un- certain inputs,

    G. Ciaramella, F. Nobile, and T. Vanzan, “A multigrid solver for pde-constrained optimization with un- certain inputs,” Journal of Scientific Computing, vol. 101, no. 1, p. 13, 2024

    2024

  15. [21]

    Multilevel quasi-monte carlo for optimization under uncertainty,

    P. A. Guth and A. Van Barel, “Multilevel quasi-monte carlo for optimization under uncertainty,” Numerische Mathematik, vol. 154, no. 3-4, pp. 443–484, 2023

    2023

  16. [23]

    A Stochastic Approximation Method,

    H. Robbins and S. Monro, “A Stochastic Approximation Method,” The Annals of Mathematical Statistics, vol. 22, no. 3, pp. 400 – 407, 1951

    1951

  17. [25]

    Multilevel Monte Carlo path simulation,

    M. B. Giles, “Multilevel Monte Carlo path simulation,” Oper. Res., vol. 56, no. 3, pp. 607–617, 2008

    2008

  18. [26]

    Multilevel optimization for inverse problems,

    S. Weissmann, A. Wilson, and J. Zech, “Multilevel optimization for inverse problems,” in Conference on Learning Theory, pp. 5489–5524, PMLR, 2022

    2022

  19. [27]

    On the mean-field limit for stein variational gradient descent: stability and multilevel approximation,

    S. Weissmann and J. Zech, “On the mean-field limit for stein variational gradient descent: stability and multilevel approximation,” arXiv preprint arXiv:2402.01320, 2024

    2024 arXiv

  20. [28]

    Optimization methods for large-scale machine learning,

    L. Bottou, F. E. Curtis, and J. Nocedal, “Optimization methods for large-scale machine learning,” SIAM review, vol. 60, no. 2, pp. 223–311, 2018

    2018

  21. [29]

    Mini-batch gradient descent: Faster convergence under data sparsity,

    S. Khirirat, H. R. Feyzmahdavian, and M. Johansson, “Mini-batch gradient descent: Faster convergence under data sparsity,” in2017 IEEE 56th Annual Conference on Decision and Control (CDC), pp. 2880–2887, IEEE, 2017

    2017

  22. [30]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014

    2014 arXiv

  23. [31]

    Adaptive step sizes for preconditioned stochastic gradient descent,

    F. K¨ ohne, L. Kreis, A. Schiela, and R. Herzog, “Adaptive step sizes for preconditioned stochastic gradient descent,” 2024

    2024

  24. [32]

    Acceleration of stochastic approximation by averaging,

    B. Polyak and A. Juditsky, “Acceleration of stochastic approximation by averaging,” SIAM Journal on Control and Optimization, vol. 30, pp. 838–855, 07 1992

    1992

  25. [33]

    Stochastic gradient descent for non-smooth optimization: Convergence results and optimal averaging schemes,

    O. Shamir and T. Zhang, “Stochastic gradient descent for non-smooth optimization: Convergence results and optimal averaging schemes,” vol. 28, pp. 71–79, 17–19 Jun 2013

    2013

  26. [34]

    The parallel finite element system M++ with integrated multilevel preconditioning and multilevel Monte Carlo methods,

    N. Baumgarten and C. Wieners, “The parallel finite element system M++ with integrated multilevel preconditioning and multilevel Monte Carlo methods,” Comput. Math. Appl., vol. 81, pp. 391–406, 2021. 26 NIKLAS BAUMGARTEN AND DA VID SCHNEIDERHAN

    2021

  27. [35]

    Mpp 3.5.0,

    C. Wieners, D. Corallo, D. Schneiderhan, L. Stengel, H. D. N. Pham, and N. Baumgarten, “Mpp 3.5.0,” 2025

    2025

  28. [36]

    Hinze, R

    M. Hinze, R. Pinnau, M. Ulbrich, and S. Ulbrich, Optimization with PDE Constraints. Mathematical Modelling: Theory and Applications, Springer, 2009

    2009

  29. [37]

    G. J. Lord, C. E. Powell, and T. Shardlow, An Introduction to Computational Stochastic PDEs. Cambridge Texts in Applied Mathematics, Cambridge University Press, 2014

    2014

  30. [38]

    Pflug, Optimization of Stochastic Models

    G. Pflug, Optimization of Stochastic Models. The Kluwer international series in engineering and computer science, Kluwer Acad. Publ., 1996

    1996

  31. [39]

    Multilevel Monte Carlo Finite Element Method for elliptic PDEs with stochastic coefficients,

    A. Barth, C. Schwab, and N. Zollinger, “Multilevel Monte Carlo Finite Element Method for elliptic PDEs with stochastic coefficients,” Numerische Mathematik, vol. 119, no. 1, pp. 123–161, 2011

    2011

  32. [40]

    Finite element error analysis of elliptic PDEs with random coefficients and its application to multilevel Monte Carlo methods,

    J. Charrier, R. Scheichl, and A. L. Teckentrup, “Finite element error analysis of elliptic PDEs with random coefficients and its application to multilevel Monte Carlo methods,” SIAM J. Numer. Anal., vol. 51, no. 1, pp. 322–352, 2013

    2013

  33. [41]

    Further analysis of multilevel Monte Carlo methods for elliptic PDEs with random coefficients,

    A. L. Teckentrup, R. Scheichl, M. B. Giles, and E. Ullmann, “Further analysis of multilevel Monte Carlo methods for elliptic PDEs with random coefficients,” Numer. Math., vol. 125, no. 3, pp. 569–600, 2013

    2013

  34. [42]

    A continuation multilevel Monte Carlo algorithm,

    N. Collier, A.-L. Haji-Ali, F. Nobile, E. von Schwerin, and R. Tempone, “A continuation multilevel Monte Carlo algorithm,” BIT, vol. 55, no. 2, pp. 399–432, 2015

    2015

  35. [43]

    An explicit link between gaussian fields and gaussian markov random fields: the stochastic partial differential equation approach,

    F. Lindgren, H. Rue, and J. Lindstr¨ om, “An explicit link between gaussian fields and gaussian markov random fields: the stochastic partial differential equation approach,”Journal of the Royal Statistical Society Series B: Statistical Methodology, vol. 73, no. 4, pp. 423–498, 2011

    2011

  36. [44]

    Convergence analysis of multilevel monte carlo variance estimators and appli- cation for random obstacle problems,

    C. Bierig and A. Chernov, “Convergence analysis of multilevel monte carlo variance estimators and appli- cation for random obstacle problems,” Numerische Mathematik, vol. 130, pp. 579–613, 2015

    2015

  37. [45]

    On the limited memory bfgs method for large scale optimization,

    D. C. Liu and J. Nocedal, “On the limited memory bfgs method for large scale optimization,” Mathematical programming, vol. 45, no. 1, pp. 503–528, 1989

    1989

  38. [46]

    An exponential averaging process with strong convergence properties,

    F. K¨ ohne and A. Schiela, “An exponential averaging process with strong convergence properties,” arXiv preprint arXiv:2505.10605, 2025

    2025 arXiv

  39. [47]

    Risk-averse pde-constrained optimization using the conditional value- at-risk,

    D. P. Kouri and T. M. Surowiec, “Risk-averse pde-constrained optimization using the conditional value- at-risk,” SIAM Journal on Optimization, vol. 26, no. 1, pp. 365–396, 2016

    2016

  40. [48]

    Quantifying uncertain system outputs via the multi-level monte carlo method- distribution and robustness measures,

    Q. Ayoul-Guilmard, S. Ganesh, S. Krumscheid, and F. Nobile, “Quantifying uncertain system outputs via the multi-level monte carlo method- distribution and robustness measures,” International Journal for Uncertainty Quantification, vol. 13, no. 5, 2023

    2023

  41. [49]

    Multi-level monte carlo training of neural operators,

    J. Rowbottom, S. Fresca, P. Lio, C.-B. Sch¨ onlieb, and N. Boull´ e, “Multi-level monte carlo training of neural operators,” arXiv preprint arXiv:2505.12940, 2025

    2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.