Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Learning to Integrate

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 claims that expectations over arbitrary target distributions can be computed by transporting Smolyak sparse Gauss–Hermite nodes through a learned normalizing flow, without changing Gaussian weights.

desk verdict Solid, honest experimental study of learned sparse-grid quadrature; useful as a practical recipe, but the convergence theory is missing and the monomial tests show the key assumption can fail. read the letter →

arxiv 2506.11801 v1 pith:FYEHLMJJ submitted 2025-06-13 math.NA cs.NAmath.PR

Oliver G. Ernst , Hanno Gottschalk , Toni Kowalewitz , Patrick Krüger This is my paper
classification math.NAcs.NAmath.PR MSC 65C3065D4068T07
keywords uncertaintyquantificationsparsegridquadratureSmolyakGauss–HermitenormalizingflowsflowmatchingLévyrandomfieldsKarhunen–Loèveexpansionlearned
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

The paper sets out a recipe for computing expected values of expensive simulation outputs when the input distribution is neither Gaussian nor factorized: train a normalizing flow that maps the input distribution to a standard multivariate normal, then evaluate the simulation at the images of Smolyak sparse Gauss–Hermite quadrature nodes, keeping the Gaussian weights unchanged. The one equation that carries the method is (1.4), which replaces the intractable integral over the target density with a weighted sum of quantity-of-interest evaluations at mapped nodes. On monomials within the exactness range of the unmapped rules, and on the flux through an outflow boundary of a stationary diffusion equation with exponentiated Levy random-field coefficients in 9 and 25 modes, the learned quadrature matches a $10^{5}$-sample Monte Carlo estimate within its 95% confidence interval for smooth fully supported targets such as bigamma. The paper also documents where the recipe breaks: gamma distributions (support on a cone) and Poisson distributions (discrete components) degrade the learned maps and the quadrature. If the recipe holds up, it gives a data-driven integration tool for uncertainty quantification with complex inputs and costly PDE solves.

What carries the argument

The load-bearing object is the composed integrand $Q(\hat g(\xi))$ on standard Gaussian space. Here $\hat g$ is the inverse of a learned normalizing map, trained by affine coupling flows, conditional flow matching, or optimal-transport conditional flow matching; the quadrature is the Smolyak sparse Gauss–Hermite rule, whose nodes $\xi_j$ and weights $w_j$ for the standard normal are fixed. The mechanism: map the sparse-grid nodes through $\hat g$ to obtain learned quadrature nodes $\hat\eta_j = \hat g(\xi_j)$, keep the Gaussian weights, and sum $w_j Q(\hat\eta_j)$. The Smolyak construction makes the node count grow mildly with dimension, and its polynomial exactness is what the monomial tests exploit; after the map, that exactness is no longer guaranteed, which the paper identifies as a source of error.

What would settle it

Take the bigamma test integrand with exactly computable moments, train an OT-CFM flow, and tabulate the Smolyak quadrature error at levels 1 through 4: if the error does not decrease with level for a fixed training set while a $10^{5}$-sample Monte Carlo reference stays narrower than the gap, the bounded-mixed-derivative premise is violated.

Watch

Extended reading notes

Core claim

The central claim is that a learned generative transport map turns a Gaussian sparse-grid rule into a rule for the target distribution at no extra quadrature cost: for $\eta \sim p$ and a generative map $g$ with $g(\xi)\sim p$ when $\xi\sim\varphi$, the expectation $\mathbb{E}_{\eta\sim p}[Q(\eta)]$ is approximated by $\sum_j w_j Q(\hat g(\xi_j))$, where $(\xi_j,w_j)$ are Smolyak sparse Gauss–Hermite nodes and weights. The weights are untouched; only the nodes are transported. In the numerical core of the paper this is demonstrated for multivariate monomials and for the outflow flux of a random diffusion equation whose conductivity is an exponentiated smoothed Lévy field, in dimensions 9 and 25, with the largest rule using 22,201 nodes. For smooth distributions with full support, such as bigamma, the learned quadrature lands inside the 95% confidence interval of a $10^{5}$-sample Monte Carlo reference; for gamma and Poisson targets the result degrades, as the paper explains by loss of smoothness or support restrictions on the transport map.

Load-bearing premise

The trained map must be smooth enough that the quantity of interest composed with it has the bounded mixed derivatives required by Smolyak sparse Gauss–Hermite error bounds, but the paper offers no regularity guarantee for the neural-network map, and its own monomial tests show the unmapped rule's exactness is destroyed after transport.

Editorial extensions

If this is right

  • For any target distribution that is continuous and vanishes nowhere, a learned transport map plus a fixed Smolyak sparse Gauss–Hermite rule gives a quadrature whose weights never need recomputation.
  • In the tested diffusion problem, the learned rule reaches Monte Carlo-quality accuracy with up to 22,201 PDE solves instead of 10^5, an order-of-magnitude reduction in simulation cost for a level-4 rule.
  • The method extends the reach of sparse-grid uncertainty quantification from Gaussian and separable inputs to data-defined distributions, as long as the flow is trained well.
  • Because the paper only needs samples of the input distribution, the same pipeline applies when the density is unknown and only observational data are available.

Reading between the lines

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

  • Beyond the paper, a natural next stress test is a high-dimensional posterior from a Bayesian inverse problem, where the density is known only up to a constant; the same pipeline would deliver a quadrature surrogate for the posterior expectation without MCMC mixing diagnostics.
  • Beyond the paper, the monomial results hint that the effective regularity of the composed integrand, not the training loss, controls convergence; monitoring that composed function's smoothness at the mapped nodes could serve as a cheap diagnostic for when to trust a learned rule.
  • Beyond the paper, combining the mapped sparse grid with control variates or multilevel estimators could reduce the remaining Monte Carlo reference cost, since the learned nodes already concentrate where the target density has mass.
  • Beyond the paper, if the flow is trained on a finite dataset, the quadrature error inherits a statistical component; the paper's training-size experiments suggest the practical rule of growing the dataset and quadrature level together, but the interaction deserves a theoretical bound.
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 manuscript proposes a numerical quadrature method ('learning to integrate') for expectations with respect to a target distribution p that is available only through samples. A normalizing flow or flow-matching model is trained to approximate a transport map g from the multivariate standard normal to p; Smolyak sparse Gauss–Hermite quadrature nodes are then mapped through the learned map, and the standard Gauss–Hermite weights are used to approximate E_p[Q] via Eq. (1.4). The method is tested on monomials of total degree up to 3 in 9 and 25 dimensions and on the flux QoI of a stationary diffusion equation with exponentiated smoothed Lévy random-field coefficients. Errors are compared with 10^5-sample Monte Carlo estimates; experiments vary quadrature level, normalizing-flow architecture (ACF, CFM, OT-CFM), target distribution (normal, bigamma, gamma, Poisson), modal truncation, finite-element mesh, and training-set size.

Significance. If the central claim were established, this would be a valuable practical extension of sparse-grid quadrature to non-Gaussian, non-tensor-product distributions, with obvious use in PDE uncertainty quantification. The paper has clear strengths: the workflow is fully described, the experiments are extensive and honest (failures are reported), the error sources are decomposed carefully, and the implementation is made publicly available. The main limitation is theoretical: no regularity or approximation estimate is provided for the composed integrand Q∘g_hat, and the paper's own monomial experiments in §5.4.3 show that convergence with sparse-grid level can fail in a smooth, full-support setting. The significance therefore currently rests on selected favorable PDE examples rather than on a general convergence mechanism.

major comments (4)
  1. [§5.4.3, Figure 6] The monomial test for the 9-mode bigamma model shows no reduction of the quadrature error when the sparse-grid level is increased from 1 to 4, although monomials are smooth and, before composition, are exactly the integrands for which the unmapped Smolyak rule has the stated exactness. The explanation given in the text—that the transported integrand is 'no longer monomials of that degree, most likely not even polynomials'—is precisely the statement that the learned map destroys the polynomial structure on which the rule's exactness rests. Since the same mechanism is invoked to justify Eq. (1.4) for the PDE examples, this negative result must be addressed, either by a regularity analysis of Q∘g_hat or by a clear and testable restriction of the applicability claim.
  2. [§3, Eq. (3.3) and following remark] The only theoretical error bound cited for the method is the Smolyak bound for integrands with bounded mixed derivatives of order r, quoted from [30] for bounded domains. The paper explicitly disclaims the extension to Gauss–Hermite weights ('beyond the scope of this work'), and no weighted Sobolev regularity is established for the neural-network map g_hat or for Q∘g_hat. Without such an estimate, Eq. (3.3) cannot be used to support convergence of the learned quadrature in the PDE examples; this is load-bearing for the central claim and should be fixed, or the convergence claim should be explicitly weakened to an empirical observation.
  3. [§5.4.5 and §5.4.6, Figure 7] The stated scope is broader than the evidence. The abstract and introduction suggest a 'generic input distribution', but Fig. 7 shows that for the continuous gamma distribution only CFM comes close to the Monte Carlo confidence interval, and for Poisson all three models fail. These failures are discussed, but they are not merely pathological: they show that the method's applicability is restricted to distributions that are sufficiently close to normal in a sense that is never made precise. In addition, the statement in §5.4.5 that a bijective map transforming R^M into a cone is 'impossible' is mathematically inaccurate (R is diffeomorphic to (0,∞) via exp), and the actual obstruction—the constrained structure of the modal coefficients and its effect on the regularity of the transported integrand—should be formulated correctly.
  4. [§5.4, Eq. (5.1)] The success criterion 'error lies within the 95% confidence interval of a 10^5-sample Monte Carlo estimator' is weak: the Monte Carlo estimator itself has random error of order σ/√N, and for integrands with small variance the confidence interval can be narrow, while for heavy-tailed integrands it can be wide enough to mask systematic quadrature error. For the monomial tests, exact or highly accurate reference moments could be obtained by standard quadrature or by a much larger Monte Carlo sample, which would allow a sharper assessment of whether Eq. (1.4) actually converges.
minor comments (4)
  1. [Abstract, §3, §4, §5.1] There are several typographical errors: 'multivatiate' in the Abstract, 'methids' in Section 3, 'modeal expansion' in Section 4, and 'sub-functionss, twe chose' in Section 5.1.
  2. [§2.1] The notation η=η1, ξ=η0 and then η1=... in Eq. (2.2) is confusing; the distinction between the data-space variable and the latent-space variable should be marked more clearly throughout the section.
  3. [§5.4.3] The sentence 'For the monomials, the quadrature error does not decrease for the 9-mode model' should specify the relevant panel and describe whether the error is flat across all levels or erratic; a reader cannot easily infer this from the current text alone.
  4. [Appendix A.1] The attribution of the SparseGrid library to 'J. Ympa' should be checked; the usual spelling of the author is 'J. Ypma'.

Circularity Check

0 steps flagged · score 2.0 of 10

There is no load-bearing circularity: the learned quadrature weights are fixed, the map is trained on the target distribution, and the reported integrals are not fitted.

full rationale

Walking the derivation chain, Eq. (1.3)-(1.4) applies the change-of-variables formula and replaces the exact generative map g by a learned map g_hat. The quadrature weights w_j are the standard Smolyak sparse Gauss-Hermite weights and are not re-fit; only the nodes are transported by g_hat. The flow g_hat is trained to match the target distribution (via MMD for ACF and flow-matching losses for CFM/OT-CFM), not to match the target integral. Thus the reported integrals are not fitted constants. The monomial experiments compare learned quadrature against a 10^5-sample Monte Carlo estimator, an external benchmark, and the PDE experiments likewise compare against Monte Carlo. The sparse-grid error bound (3.3) is cited to Novak and Ritter [30], an external source, and the paper explicitly states that the analogous Gauss-Hermite analysis is beyond its scope; no uniqueness or convergence theorem by the authors is invoked to force the choice. The only overlapping self-citations, [13] and [33], concern the construction of the smoothed Levy random-field test case and the modal expansion; they are background ingredients, not grounds for the quadrature claim. The conclusion honestly states that the theoretical status is unresolved and that the results are experimental. The Section 5.4.3 observation that mapped monomials are no longer polynomials is a regularity limitation, not a circular reduction. Overall there is no significant circularity; at most a minor non-load-bearing self-citation, which is reflected in the score of 2.

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

The method adds no fitted constants to the target integrals. It assumes standard optimal-transport existence, universal approximation of the flow architectures, a non-proved sparse-Gauss-Hermite convergence extension, and prior results on Levy fields that the paper cites. It invents no new entities.

assumptions (5)
  • standard math There exists a smooth bijective transport map from the target distribution p to the standard normal whenever p has positive density on R^M.
    Invoked in Section 1 after Eq. (1.2), citing optimal transport existence results [35,43]. Needed for the change of variables (1.3).
  • domain assumption The chosen flow architectures (ACF, CFM, OT-CFM) are expressive enough to approximate the transport map to within the accuracy needed.
    Invoked in Section 1 (universal approximation cited [39,34]) and relied on throughout Section 5; no quantitative approximation-error bound for the specific flows is given.
  • domain assumption Smolyak sparse Gauss-Hermite quadrature converges for integrands with bounded mixed derivatives, and the univariate Gauss-Hermite version of the Novak-Ritter bound (3.3) holds.
    Section 3, Eq. (3.3); the paper notes the GH extension 'is beyond the scope of this work', so the error bound is assumed rather than proved for the actual rule used.
  • domain assumption The Levy random field modal expansion (4.8), circulant embedding, and well-posedness of the diffusion equation with Levy coefficients hold as stated.
    Section 4, based on prior work [13,33] by the same authors; the paper relies on these results for the PDE test case.
  • domain assumption For the transport map to exist and be effectively learned, the target distribution must be continuous and vanish nowhere on R^M.
    Section 5.4.5: gamma supported on a positive half-line and Poisson with atoms cause poor flow approximations and quadrature failures.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning to Integrate." pith.science (2026). https://pith.science/paper/FYEHLMJJ

@misc{pith2026250611801,
  author       = {Pith},
  title        = {Pith review of: Learning to Integrate},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FYEHLMJJ}},
  note         = {Machine review of arXiv:2506.11801}
}
read the original abstract

This work deals with uncertainty quantification for a generic input distribution to some resource-intensive simulation, e.g., requiring the solution of a partial differential equation. While efficient numerical methods exist to compute integrals for high-dimensional Gaussian and other separable distributions based on sparse grids (SG), input data arising in practice often does not fall into this class. We therefore employ transport maps to transform complex distributions to multivatiate standard normals. In generative learning, a number of neural network architectures have been introduced that accomplish this task approximately. Examples are affine coupling flows (ACF) and ordinary differential equation-based networks such as conditional flow matching (CFM). To compute the expectation of a quantity of interest, we numerically integrate the composition of the inverse of the learned transport map with the simulation code output. As this map is integrated over a multivariate Gaussian distribution, SG techniques can be applied. Viewing the images of the SG quadrature nodes as learned quadrature nodes for a given complex distribution motivates our title. We demonstrate our method for monomials of total degrees for which the unmapped SG rules are exact. We also apply our approach to the stationary diffusion equation with coefficients modeled by exponentiated L\'evy random fields, using a Karhunen-Lo\`eve-like modal expansions with 9 and 25 modes. In a series of numerical experiments, we investigate errors due to learning accuracy, quadrature, statistical estimation, truncation of the modal series of the input random field, and training data size for three normalizing flows (ACF, conditional Flow Matching and Optimal transport Flow Matching) We discuss the mathematical assumptions on which our approach is based and demonstrate its shortcomings when these are violated.

Figures

Figures reproduced from arXiv: 2506.11801 by the authors.

Figure 1
Figure 1. Overview of the learning to integrate method. The target distribution is normalized with the learned flow map f. For integration with respect to an i.i.d normal density, Smolyak sparse Gauss–Hermite rules allow for efficient quadrature. The associated quadrature nodes are transformed via the generative map g = f −1 and then fed into the simulation code which computes the quantity of interest (QoI). The results, weig… view at source ↗
Figure 2
Figure 2. Images of Smolyak sparse Gauss–Hermite nodes under the transport map for the bigamma distribution generated by the ACF model for M = 9 modes displayed in all pairs of distinct coordinate directions. the necessary background, we first discuss Lévy noise fields and noise fields smoothed by pseudodifferential operators following [2, 13]. These are then introduced into the diffusion equation as coefficients. We also dis… view at source ↗
Figure 3
Figure 3. Realizations of smoothed Lévy noise fields with different distributions: from the left Gaussian, Poisson, gamma and bigamma interpolated to the coarse FE mesh vertices (green points) and FE quadrature points (grey points). 4.2. Lévy noise fields. Random noise fields cannot be represented as functions with values defined pointwise. They can be treated using the more comprehensive theory of generalized random fields, … view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: The three meshes employed in the finite element solution consisting of 48, 798, and 8510 elements and referred to in the following as coarse, medium and fine, respectively. The random conductivity a is represented by a smoothed Lévy random field Zk, for which we consid…
Figure 5
Figure 5. Figure 5: Discretization and truncation errors of the QoI a random realization of the conductivity following four different probability distributions. Left: error against finite element mesh size (full field without truncation). Right: error against modes retained in Mercer seri…
Figure 6
Figure 6. Figure 6: Quadrature error comparison in the bigamma case for monomials as well as for the quantity of interest of random PDEs for the best models, for varying degrees and highest training size. Left: 9 modes. Right: 25 modes. 5.4.4. Regularity constraints for sparse quadrature.…
Figure 7
Figure 7. Figure 7: Quadrature error of sparse quadrature rule learned by three different methods (9 modes). We compared the model with the highest amount of training data to a Monte Carlo estimator with 105 samples on the finest mesh. we can train longer without overfitting, resulting in…
Figure 8
Figure 8. Figure 8: Statistical error comparison in the bigamma case for monomials as well as for the quantity of interest of random PDEs for the best models for varying training sizes and highest quadrature degree. Left: 9 modes. Right: 25 modes. pling from the complex distribution is as…

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. Optimality of quasi-Monte Carlo methods and suboptimality of the sparse-grid Gauss--Hermite rule in Gaussian Sobolev spaces

    math.NA 2025-09 conditional novelty 6.0 of 10

    Sparse-grid Gauss-Hermite quadrature is provably suboptimal at N^{-alpha/2}, while some QMC methods with change of variables achieve the optimal N^{-alpha} (ln N)^{(d-1)/2} rate in Gaussian Sobolev spaces.

Reference graph

Works this paper leans on

47 extracted references · 31 canonical work pages · cited by 1 Pith paper

  1. [30]

    Nov ak and K

    E. Nov ak and K. Ritter , Simple cubature formulas with high polynomial exactness, Constructive Approximation, 15 (1999), pp. 499–522, https://doi.org/10.1007/s003659900119

    doi:10.1007/s003659900119 1999

  2. [1]

    M. S. Albergo and E. V anden-Eijnden, Building normalizing flows with stochastic interpolants, arXiv preprint arXiv:2209.15571, (2022)

    arXiv 2022

  3. [2]

    Albeverio, H

    S. Albeverio, H. Gottschalk, and J.-L. Wu , Convoluted generalized white noise, schwinger func- tions and their analytic continuation to wightman functions, Reviews in Mathematical Physics, 8 (1996), pp. 763–817

    work page 1996

  4. [3]

    Applebaum, Lévy processes and stochastic calculus, Cambridge university press, 2009

    D. Applebaum, Lévy processes and stochastic calculus, Cambridge university press, 2009

    work page 2009

  5. [4]

    Ardizzone, T

    L. Ardizzone, T. Bungert, F. Draxler, U. Köthe, J. Kruse, R. Schmier, and P. Sorren- son, Framework for Easily Invertible Architectures (FrEIA), 2018-2022, https://github.com/vislearn/ FrEIA

    work page 2018

  6. [5]

    Babuška, F

    I. Babuška, F. Nobile, and R. Tempone ,A stochastic collocation method for elliptic partial differential equations with random input data, SIAM Review, 52 (2010), pp. 317–355, https://doi.org/10.1137/ 050645142

    work page 2010

  7. [6]

    Bachmayr, A

    M. Bachmayr, A. Cohen, and G. Migliorati , Representations of Gaussian random fields and ap- proximation of elliptic pdes with lognormal coefficients, Journal of Fourier Analysis and Applications, 24 (2018), pp. 621–649

    work page 2018

  8. [7]

    R. Chan, S. Penquitt, and H. Gottschalk , Lu-net: Invertible neural networks based on matrix factorization, in 2023 International Joint Conference on Neural Networks (IJCNN), IEEE, 2023, pp. 1–10

    work page 2023

Show all 47 references
  1. [8]

    R. T. Chen, Y. Rubanov a, J. Bettencourt, and D. K. Duvenaud , Neural ordinary differential equations, Advances in neural information processing systems, 31 (2018)

    2018

  2. [9]

    M. F. Daniele Boffi, Franco Brezzi , Mixed Finite Element Methods and Applications, vol. 44, Springer Series in Computational Mathematics, 2013

    2013

  3. [10]

    L. Dinh, J. Sohl-Dickstein, and S. Bengio , Density estimation using real nvp, arXiv preprint arXiv:1605.08803, (2016)

    2016 arXiv

  4. [11]

    D~ ung, V

    D. D~ ung, V. K. Nguyen, C. Schw ab, and J. Zech , Analyticity and sparsity in uncertainty quan- tification for PDEs with Gaussian random field inputs, vol. 2334 of Lecture Notes in Mathematics, springer, cham, 2023, https://doi.org/10.1007/978-3-031-38384-7

    2023 doi

  5. [12]

    Ern and J.-L

    A. Ern and J.-L. Guermond , Finite Elements I: Approximation and Interpolation, Texts in Applied Mathematics, Springer Cham, 2021, https://doi.org/10.1007/978-3-030-56341-7

    2021 doi

  6. [13]

    O. G. Ernst, H. Gottschalk, T. Kalmes, T. Kow alewitz, and M. Reese , Integrability and approximability of solutions to the stationary diffusion equation with Lévy coefficient, 2021, https: //arxiv.org/abs/2010.14912v3

    2021 arXiv

  7. [14]

    O. G. Ernst, B. Sprungk, and L. Tamellini , Convergence of sparse collocation for functions of countably many Gaussian random variables (with application to elliptic PDEs), SIAM Journal on LEARNING TO INTEGRATE 25 Numerical Analysis, 56 (2018), pp. 877–905, https://doi.org/10....

    2018 doi

  8. [15]

    Finlay, J.-H

    C. Finlay, J.-H. Jacobsen, L. Nurbekyan, and A. Oberman , How to train your neural ODE: the world of Jacobian and kinetic regularization, in International conference on machine learning, PMLR, 2020, pp. 3154–3164

    2020

  9. [16]

    Gautschi , Orthogonal Polynomials: Computation and Approximation, Oxford University Press, Oxford, UK, 2004, https://doi.org/10.1093/oso/9780198506720.001.0001

    W. Gautschi , Orthogonal Polynomials: Computation and Approximation, Oxford University Press, Oxford, UK, 2004, https://doi.org/10.1093/oso/9780198506720.001.0001

    2004

  10. [17]

    Gholami, K

    A. Gholami, K. Keutzer, and G. Biros , Anode: Unconditionally accurate memory-efficient gradients for neural odes, arXiv preprint arXiv:1902.10298, (2019)

    2019 arXiv

  11. [18]

    Gretton, K

    A. Gretton, K. Borgw ardt, M. Rasch, B. Schölkopf, and A. Smola , A kernel method for the two-sample-problem, Advances in neural information processing systems, 19 (2006)

    2006

  12. [19]

    K. He, X. Zhang, S. Ren, and J. Sun , Deep residual learning for image recognition, in 2016 IEEE Con- ference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, 2016, pp. 770– 778, https://doi.org/10.1109/CVPR.2016.90

    2016 doi

  13. [20]

    Heiss and V

    F. Heiss and V. Winschel , Likelihood approximation by numerical integration on sparse grids, Journal of Econometrics, 144 (2008), pp. 62–80, https://doi.org/10.1016/j.jeconom.2007.12.004

    2008 doi

  14. [21]

    Itô, Foundations of Stochastic Differential Equations in Infinite Dimensional Spaces, SIAM, 1984

    K. Itô, Foundations of Stochastic Differential Equations in Infinite Dimensional Spaces, SIAM, 1984

    1984

  15. [22]

    Kallenberg, Random measures, Akademie-Verlag Berlin, 1983

    O. Kallenberg, Random measures, Akademie-Verlag Berlin, 1983

    1983

  16. [23]

    Kazashi, Y

    Y. Kazashi, Y. Suzuki, and T. Goda , Suboptimality of gauss–hermite quadrature and optimality of the trapezoidal rule for functions with finite smoothness, SIAM Journal on Numerical Analysis, 61 (2023), pp. 1426–1448, https://doi.org/10.1137/22M148027

    2023 doi

  17. [24]

    D. P. Kingma and P. Dhariw al , Glow: Generative flow with invertible 1x1 convolutions, Advances in neural information processing systems, 31 (2018)

    2018

  18. [25]

    Krueger, H

    P. Krueger, H. Gottschalk, B. Werdelmann, and W. Krebs , Generative design of a gas turbine combustor using invertible neural networks, Journal of Engineering for Gas Turbines and Power, 147 (2025)

    2025

  19. [26]

    LeCun, Y

    Y. LeCun, Y. Bengio, and G. Hinton , Deep learning, nature, 521 (2015), pp. 436–444

    2015

  20. [27]

    Lipman, R

    Y. Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le , Flow matching for generative modeling, arXiv preprint arXiv:2210.02747, (2022)

    2022 arXiv

  21. [28]

    X. Liu, C. Gong, and Q. Liu , Flow straight and fast: Learning to generate and transfer data with rectified flow, arXiv preprint arXiv:2209.03003, (2022)

    2022 arXiv

  22. [29]

    Nov ak and K

    E. Nov ak and K. Ritter , High dimensional integration of smooth functions over cubes, Numerische Mathematik, 75 (1996), pp. 79–97, https://doi.org/10.1007/s002110050231

    1996 doi

  23. [31]

    Onken and L

    D. Onken and L. Ruthotto , Discretize-optimize vs. optimize-discretize for time-series regression and continuous normalizing flows, 2020, https://arxiv.org/abs/2005.13420

    2020 arXiv

  24. [32]

    L. S. Pontryagin, Mathematical theory of optimal processes, Routledge, 2018

    2018

  25. [33]

    M. Reese , Partial Differential Equations and Spatial Structures of Lévy Type: : Uncertainty Quan- tification and Optimization, PhD thesis, Bergische Universität Wuppertal, Fakultät für Mathematik und Naturwissenschaften, 2021, https://elekpub.bib.uni-wuppertal.de/ubwhsmig/con...

    2021

  26. [34]

    Rochau, R

    D. Rochau, R. Chan, and H. Gottschalk , New advances in universal approximation with neural networks of minimal width, arXiv preprint arXiv:2411.08735, (2024)

    2024

  27. [35]

    Santambrogio, Optimal transport for applied mathematicians, vol

    F. Santambrogio, Optimal transport for applied mathematicians, vol. 87, Springer, 2015

    2015

  28. [36]

    Singleton, Mixed radix fast fourier transforms, in programs for digital signal processing, 1979

    R. Singleton, Mixed radix fast fourier transforms, in programs for digital signal processing, 1979

    1979

  29. [37]

    A. H. Stroud and D. Secrest , Gaussian Quadrature Formulas, Prentice-Hall series in automatic computation, Prentice-Hall, Englewood-Cliffs, NJ, 1966

    1966

  30. [38]

    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., 125 (2013), pp. 569– 600, https://doi.org/10.1007/s00211-013-0546-4

    2013 doi

  31. [39]

    Teshima, I

    T. Teshima, I. Ishika w a, K. Tojo, K. Oono, M. Ikeda, and M. Sugiyama , Coupling-based invert- ible neural networks are universal diffeomorphism approximators, Advances in Neural Information Processing Systems, 33 (2020), pp. 3362–3373

    2020

  32. [40]

    A. Tong, K. F ATRAS, N. Malkin, G. Huguet, Y. Zhang, J. Rector-Brooks, G. Wolf, 26 O. G. ERNST, H. GOTTSCHALK, T. KOWALEWITZ, AND P. KRÜGER and Y. Bengio , Improving and generalizing flow-based generative models with minibatch optimal transport, Transactions on Machine Learnin...

    2024

  33. [41]

    L. N. Trefethen , Exactness of quadrature formulas, SIAM Review, 64 (2022), pp. 132–150, https: //doi.org/10.1137/20M1389522

    2022 doi

  34. [42]

    v an den Berg and G

    C. v an den Berg and G. Forst , Potential theory on locally compact abelian groups, vol. 87, Springer Science & Business Media, 2012

    2012

  35. [43]

    Villani et al

    C. Villani et al. , Optimal transport: old and new, vol. 338, Springer, 2008

    2008

  36. [44]

    W asilkowski and H

    G. W asilkowski and H. Woźniakowski , Explicit cost bounds of algorithms for multivariate tensor product problems, Journal of Complexity, 11 (1995), pp. 1–56, https://doi.org/10.1006/jcom.1995. 1001

    1995 doi

  37. [45]

    Westny, A

    T. Westny, A. Mohammadi, D. Jung, and E. Frisk , Stability-informed initialization of neural ordinary differential equations, arXiv preprint arXiv:2311.15890, (2023)

    2023 arXiv

  38. [46]

    H. Xia, V. Suliafu, H. Ji, T. Nguyen, A. Bertozzi, S. Osher, and B. W ang , Heavy ball neu- ral ordinary differential equations, Advances in Neural Information Processing Systems, 34 (2021), pp. 18646–18659

    2021

  39. [47]

    Zheng, Z

    H. Zheng, Z. Yang, W. Liu, J. Liang, and Y. Li , Improving deep neural networks using softplus units, in 2015 International joint conference on neural networks (IJCNN), IEEE, 2015, pp. 1–4. LEARNING TO INTEGRATE 27 Appendix A. Supplementary Material.In this appendix, we provid...

    2015

Pith tools

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