Pith. sign in

REVIEW 3 major objections 5 minor 41 references

Hand-derived multigrid adjoint operators make PDE-constrained Bayesian inference practical

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

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

A hand-coded adjoint multigrid solver, wrapped in JAX, enables memory-efficient variational inference for a 3D tissue-imaging inverse problem.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Useful integration of a hand-adjointed multigrid solver with variational inference, but the central adjoint correctness claim is never directly verified — fix that and this is a solid methods paper. the 3 major comments →

arxiv 2608.00760 v1 pith:HTSDV7OL submitted 2026-08-01 math.NA cs.NAphysics.med-phstat.CO

Variational Inference Using a Differentiable Multigrid Linear Solver

Andr\'es Ram\'irez , Philipp Haim , David F\"oger , Torsten A. En{\ss}lin , Philipp Frank , Philipp Gschwandtner , Dominik J\"ustel , Philipp Mertsch
show 4 more authors
Vo Hong Minh Phan Laurin S\"oding Hanieh Zandinejad Ralf Kissmann
This is my paper
classification math.NA cs.NAphysics.med-phstat.CO MSC 65N5562F1565K10
keywords multigrid methodsadjoint operatorsautomatic differentiationvariational inferencediffusion-absorption equationradiative transferBayesian inverse problemsPDE-constrained inference
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Gradient-based Bayesian inference needs cheap Jacobian and adjoint-Jacobian products of the forward model, but multigrid solvers—near-optimal for elliptic PDEs—are rarely exposed in a differentiable form. This paper tries to close that gap by deriving the adjoint of a multigrid solver explicitly: since the solver is a composition of linear operations, its pull-back is the composition of the component adjoints in reverse order. The authors implement this in a memory-efficient solver wrapped for an automatic-differentiation framework, and use it inside metric-based variational inference to reconstruct an effective radiative source in 3D tissue from Monte Carlo–simulated data. The reconstruction achieves reduced chi-squared about 1.1 and generalizes to 32 independent validation datasets, while benchmarks against an autodiff-native multigrid show comparable runtime and consistently lower peak memory. If right, this makes differentiable multigrid a practical building block for variational inference in PDE-constrained inverse problems.

Core claim

At the center is the identity S-dagger = gbar_1 composed with gbar_2 composed with ... composed with gbar_l: the adjoint of the multigrid solver is obtained by taking the adjoints of the individual components—Red–Black Gauss–Seidel smoothing, defect computation, restriction, and prolongation—and composing them in the reverse order of the forward sweep. Because the solver is linear in the source for fixed coefficients and boundary conditions, the Jacobian is the solver itself and the vector–Jacobian product is exactly this reversed adjoint. The paper constructs each component adjoint algebraically, including the factor-2^n relation between restriction and prolongation that makes the transfer

What carries the argument

The load-bearing object is the reversed-order adjoint multigrid operator, S-dagger = gbar_1 composed with gbar_2 composed with ... composed with gbar_l (Eq. 18), built from analytic adjoints of each solver component. For Red–Black Gauss–Seidel, the adjoint maps a cotangent field through the same stencil with divided coefficients; for the defect step it adds the defect back into source and fluence cotangents; for transfers, full-weighting restriction and interpolation are adjoint up to the factor 2^n, so their pull-backs are simple scalings of the forward transfer. Matching the number of V-cycles in forward and reverse modes removes any dependence on the stopping criterion and lets the push-f

Load-bearing premise

The diffusion approximation to radiative transfer, combined with the 0.2 cm edge mask, assumes all non-diffusive (ballistic) light-transport error stays inside the excluded boundary layer; if it leaks into the unmasked volume, the forward model is wrong and the fit no longer validates the solver.

What would settle it

Run the reconstruction on the same data with the edge mask removed or reduced; if residuals spike near the boundary and the inferred source migrates into the newly exposed region, the model error is in the diffusion approximation, not the adjoint. Separately, on a small grid, compare the hand-derived adjoint against a finite-difference or automatic-differentiation Jacobian-transpose by checking the inner-product identity for random fields and cotangent vectors; any discrepancy beyond solver tolerance falsifies the reversed-order adjoint as exact.

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

If this is right

  • Metric-based variational inference can treat multigrid as an off-the-shelf differentiable component, so high-dimensional PDE-constrained posteriors become accessible without writing bespoke adjoint code for each solver.
  • Hand-derived adjoints cut peak memory relative to autodiff implementations, with reverse-mode overhead under 1% in 3D, which matters when grid size is the bottleneck.
  • The reconstruction's chi-squared about 1.1 and agreement across 32 validation datasets show the solver-plus-adjoint pair can drive an inference loop to a statistically consistent fit.
  • Deviations concentrate near the boundary mask, confirming that the forward-model error budget is controlled by the diffusion approximation, not by the adjoint construction.

Where Pith is reading between the lines

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

  • The same adjoint-reversal recipe should extend to nonlinear forward models or to differentiation with respect to optical coefficients, since each component's Jacobian can be derived independently; the paper lists these as future work.
  • Because the adjoint is only exact when forward and reverse iteration counts match, an adaptive-stopping forward solve would require checkpointing the iteration count or a reversible solver; otherwise the gradient would be of a different operator than the one evaluated.
  • A direct self-adjointness test—checking inner-product equality on random fields—would make the correctness of the reversed-order adjoint independently verifiable, and is a natural cheap addition to the benchmark suite.
  • The validation logic could be stress-tested by moving the edge mask inward or outward: if chi-squared degrades sharply as the mask thins, the diffusion approximation's boundary layer is the real constraint, and a hybrid transport-diffusion forward model would be a testable remedy.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper develops DMGS, a C++ multigrid solver for the steady-state diffusion-absorption equation, and derives its adjoint operations by reversing the computational graph of the solver. The forward solver is a fixed 15-V-cycle multigrid with red-black Gauss-Seidel smoothing, and the adjoint is expressed as a composition of adjoint component operations (Eq. 18). The solver is wrapped with JAXbind for use in JAX and NIFTy.re, and is integrated into the MGVI/geoVI variational-inference pipelines. The method is applied to a 3D diffuse radiative transfer problem in tissue, where the authors reconstruct an effective radiative source from Monte Carlo-simulated data and report a reduced chi-squared of about 1.1 on the main dataset, plus generalization to 32 validation datasets. The paper also benchmarks DMGS against a JAX-native autodiff multigrid solver (JMGS), showing comparable runtimes and lower reverse-mode memory usage.

Significance. If the adjoint correctness is established, this is a useful practical contribution: explicit adjoints for a full multigrid hierarchy give memory-efficient reverse-mode differentiation of a near-optimal linear PDE solver, and the integration into metric-based VI is a relevant demonstration. The 3D reconstruction and 32-dataset check provide a nontrivial application. The paper is careful in attributing the Dataset 17 outlier to diffusion-approximation failure rather than to the solver, and the benchmark comparison is informative. However, the central claim—that the hand-derived S† is the exact transpose of the implemented fixed-iteration solver—is not directly verified: no dot-product test, finite-difference check, or comparison with AD is reported, and no code is released. The derivation is also stated to be illustrated only in 1D, leaving the delicate red-black Gauss-Seidel reverse pass and its 3D extension without a numerical check. The fixed 15-V-cycle schedule is asserted without a convergence study. These gaps mean the reconstruction and generalization results cannot by themselves validate the adjoint solver, but they are fixable with added numerical verification and convergenc

major comments (3)
  1. [§2.3–2.4, Eq. (18)] The central claim—that the hand-derived composition of adjoint component operations is the exact transpose of the fixed-iteration forward solver—is not verified numerically. The derivation is explicitly presented only for 1D, and the delicate red-black Gauss-Seidel recursion (Eqs. (9)–(11)) is exactly the kind of operation where an ordering mistake in the reverse pass is easy to make. The paper reports no dot-product test, no finite-difference gradient check, and no comparison with jax.vjp or an AD implementation; the benchmarks in §3.3 measure runtime and memory only. Without such a check, the χ²≈1.1 reconstruction and the 32-dataset generalization cannot discriminate a correct S† from an approximately correct one. Please add a standalone verification (e.g., ⟨w,S(q+εv)−S(q)⟩≈ε⟨S†w,v⟩ at several resolutions and levels) and, if possible, release the solver code.
  2. [§2.2/§2.4, fixed 15 V-cycles] The manuscript asserts that 15 V-cycles ensure 'sufficiently high precision' but gives no convergence study. Because DMGS is used in truncated-solver mode, the forward model is S(q), not L^{-1}(q); if 15 cycles leave a non-negligible residual at the finest grid, the reconstruction target is altered. This is especially relevant for Dataset 17, where the diffusion approximation is anyway stressed. Add a convergence diagnostic (residual norm versus V-cycle count, and the resulting χ² or RMSE dependence) to justify the fixed schedule.
  3. [§3.1, Eq. (26)] The reported χ²≈1.1 is computed on the same data used to fit the posterior; it is a goodness-of-fit statistic, not an independent check of the adjoint solver. The generalization to 32 datasets in §3.2 is a stronger check, but the high-RMSE outlier (Dataset 17) is attributed to forward-model breakdown, so the validation separates poorly between solver and model error. Please add a diagnostic that directly tests the solver's pull-back on this dataset (e.g., compare with JMGS predictions or with a direct L† solve) or at least quantify the fraction of the RMSE attributable to the diffusion approximation.
minor comments (5)
  1. [§2.4, Eq. (18)] The bar notation (¯g_i) is used without a formal definition of the adjoint of a linear map in finite-dimensional spaces; define it explicitly as ⟨g x, y⟩ = ⟨x, g† y⟩ for the relevant inner product.
  2. [§2.3.1, Eq. (9)] The red-black update is written for one color only; the black half-sweep and its ordering with respect to the updated red points is not displayed. Clarify the two half-sweeps so the reverse order in Eqs. (10)–(11) is unambiguous.
  3. [§2.5.3, Eq. (22)] Table 1 lists prior means and standard deviations for parameters that are said to have log-normal priors. Clarify whether the reported values refer to the parameters themselves or to their logarithms; otherwise the posterior summary is hard to interpret.
  4. [§3.3, Fig. 13] JMGS is not described beyond 'a JAX-native implementation relying on automatic differentiation' and is said to be presented in a future publication. The benchmark comparison would be more reproducible if the JMGS computation graph and the exact V-cycle schedule were summarized in an appendix or supplementary note.
  5. [§3.2, Fig. 12] The color bars and residual scales in Fig. 12 are small and the min-max rescaling complicates quantitative comparison. State the units of the normalized L_e values and the normalization used for the residuals in the caption.

Circularity Check

0 steps flagged

No significant circularity; adjoint derivation is self-contained.

full rationale

The derivation of the adjoint multigrid operator S† (Eq. 18) is self-contained: it follows from the linearity of the forward composition S = g_l ∘ ... ∘ g_1 and the explicit adjoint formulas for the RB-GS relaxation (Eqs. 10–11), defect (Eqs. 15–16), and transfer operations (Eq. 17). These formulas are derived directly from the forward definitions and do not invoke the NIFTy/PICARD framework. The reconstruction χ²≈1.1 is a fit quality statistic, not presented as a prediction, and the 32-dataset generalization is an independent forward-model check against Monte Carlo-simulated data with fixed illumination but varying optical parameters. Self-citations to NIFTy, MGVI, geoVI, and PICARD provide software components and inference algorithms but are not load-bearing for the adjoint correctness claim. Therefore no circular step is exhibited.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The solver derivation itself introduces no fitted constants; the 15-cycle count and the 0.2 cm mask are the two hand-chosen numbers that shape the practical demonstration. The 11 prior hyperparameters (Table 1) are actively inferred by VI, so they are part of the statistical model rather than free parameters of the method. No new physical entities are proposed.

free parameters (2)
  • fixed V-cycle count = 15
    Hand-chosen number of multigrid iterations used identically in forward and adjoint modes (Section 2.2: 'the cycle is repeated 15 times to ensure a solution with sufficiently high precision'). No convergence or residual study is reported, so the choice is a free numerical parameter.
  • edge mask depth = 0.2 cm
    Hand-chosen depth for excluding boundary voxels from the likelihood (Section 2.5.1). The reconstruction's chi-squared about 1.1 and validation RMSE are computed only on unmasked voxels, so this choice directly affects the headline results.
axioms (4)
  • domain assumption Steady-state diffusion-absorption equation (1) approximates the Monte Carlo radiative transfer in the unmasked tissue volume.
    Invoked in Section 2.5; the validity of the forward model determines whether chi-squared and RMSE compare against the true data-generating process.
  • standard math The prolongation operator satisfies I = 2^n R^T because R is full-weighting (Eq. 17).
    Used in Section 2.3.3 to express the adjoint transfer operations; holds for the stated full-weighting stencil.
  • ad hoc to paper A fixed schedule of 15 V-cycles is sufficiently converged that the adjoint of the truncated solver is a reliable pull-back for VI.
    Section 2.4 selects the fixed-iteration mode; no residual threshold, convergence test, or comparison against a fully converged solve is provided.
  • standard math Dirichlet boundary conditions and fixed optical coefficients make the solver S linear in q, so J = S.
    Stated in Section 2.3; follows from the linearity of the discretized operations, and is the justification for the Jacobian-equals-solver claim.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Variational Inference Using a Differentiable Multigrid Linear Solver." pith.science (2026). https://pith.science/paper/HTSDV7OL

@misc{pith2026260800760,
  author       = {Pith},
  title        = {Pith review of: Variational Inference Using a Differentiable Multigrid Linear Solver},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HTSDV7OL}},
  note         = {Machine review of arXiv:2608.00760}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Gradient-based Bayesian inference methods require efficient access to Jacobian and adjoint-Jacobian operators of high-dimensional forward models. While multigrid solvers provide near-optimal complexity for elliptic partial differential equations, they are rarely available in forms compatible with automatic differentiation (AD). We develop a differentiable multigrid solver for steady-state diffusion-absorption problems and derive its adjoint operations analytically through the full multigrid hierarchy. The resulting solver, DMGS, is implemented in C++ and interfaced with JAX to provide efficient Jacobian-vector and vector-Jacobian products for variational inference in the NIFTy framework. We validate the approach on a 3D inverse problem involving diffuse radiative transfer in tissue, reconstructing an effective radiative source from Monte Carlo-simulated data. The reconstruction reproduces the data at a reduced chi-squared of 1.1 and generalizes to 32 independent validation datasets. Benchmarks against a JAX-native multigrid implementation show comparable runtimes and consistently lower peak memory for the hand-derived adjoint, with modest reverse-mode overhead. These results establish differentiable multigrid solvers as practical building blocks for variational inference in PDE-constrained problems.

discussion (0)

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

Reference graph

Works this paper leans on

41 extracted references · 26 canonical work pages · 4 internal anchors

  1. [1]

    PeerJ Computer Science9(e1516) (2023)

    Abril-Pla, O., Andreani, V ., Carroll, C., Dong, L., Fonnesbeck, C.J., Kochurov, M., Kumar, R., Lao, J., Luhmann, C.C., Martin, O.A., Osthege, M., Vieira, R., Wiecki, T., Zinkov, R.: PyMC: A modern and comprehensive probabilistic programming framework in Python. PeerJ Computer Science9(e1516) (2023). DOI 10.7717/peerj-cs.1516

    doi:10.7717/peerj-cs.1516 2023

  2. [2]

    Arras, P., Baltac, M., Ensslin, T.A., Frank, P., Hutschenreuter, S., Knollmueller, J., Leike, R., Newrzella, M.N., Platz, L., Reinecke, M., Stadler, J.: NIFTy5: Numerical Information Field Theory v5 (2019)

    work page 2019

  3. [3]

    Journal of Machine Learning Research18(2018)

    Baydin, A.G., Pearlmutter, B.A., Radul, A.A., Siskind, J.M.: Automatic differentiation in machine learning: a survey. Journal of Machine Learning Research18(2018). URLhttps://www.jmlr.org/papers/volume18/17-468/17-468.pdf

    work page 2018

  4. [4]

    Science Advances11(34), eadu7319 (2025)

    Berger, C., Kim, M., Scheel-Platz, L., Eigenberger, A., Prantl, L., Liu, P., Gujrati, V ., Ntziachristos, V ., J ¨ustel, D., Pleitez, M.A.: Bayesian reconstruction of rapidly scanned mid-infrared optoacoustic signals enables fast, label-free chemical mi- croscopy. Science Advances11(34), eadu7319 (2025). DOI 10.1126/sciadv.adu7319

    work page doi:10.1126/sciadv.adu7319 2025

  5. [5]

    Computers & Fluids (2023)

    Bezgin, D.A., et al.: Jax-fluids: A fully-differentiable high-order computational fluid dynamics package. Computers & Fluids (2023). URLhttps://www.sciencedirect.com/science/article/abs/pii/S0010465522002466

    work page 2023

  6. [6]

    Journal of Machine Learning Research (2019)

    Bingham, E., et al.: Pyro: Deep universal probabilistic programming. Journal of Machine Learning Research (2019). URL https://jmlr.org/papers/v20/18-403.html VI Using a Differentiable Multigrid Linear Solver 17

    work page 2019

  7. [7]

    URLhttp: //github.com/jax-ml/jax

    Bradbury, J., Frostig, R., Hawkins, P., Johnson, M.J., Leary, C., Maclaurin, D., Necula, G., Paszke, A., VanderPlas, J., Wanderman-Milne, S., Zhang, Q.: JAX: composable transformations of Python+NumPy programs (2018). URLhttp: //github.com/jax-ml/jax

    work page 2018

  8. [8]

    SIAM (2000)

    Briggs, W.L., Henson, V .E., McCormick, S.F.: A Multigrid Tutorial, 2 edn. SIAM (2000). URLhttps://epubs.siam.org /doi/book/10.1137/1.9780898719505

    doi:10.1137/1.9780898719505 2000

  9. [9]

    Journal of Statistical Software76(1), 1–32 (2017)

    Carpenter, B., Gelman, A., Hoffman, M.D., Lee, D., Goodrich, B., Betancourt, M., Brubaker, M., Guo, J., Li, P., Riddell, A.: Stan: A probabilistic programming language. Journal of Statistical Software76(1), 1–32 (2017). DOI 10.18637/jss.v076.i01. URLhttps://www.jstatsoft.org/index.php/jss/article/view/v076i01

    doi:10.18637/jss.v076.i01 2017

  10. [10]

    IEEE Journal of Quantum Electronics 26(12), 2166–2185 (1990)

    Cheong, W., Prahl, S., Welch, A.: A review of the optical properties of biological tissues. IEEE Journal of Quantum Electronics 26(12), 2166–2185 (1990). DOI 10.1109/3.64354

    work page doi:10.1109/3.64354 1990

  11. [11]

    Edenhofer, G., Frank, P., Roth, J., Leike, R.H., Guerdi, M., Scheel-Platz, L.I., Guardiani, M., Eberle, V ., Westerkamp, M., Enßlin, T.A.: Re-Envisioning Numerical Information Field Theory (NIFTy.re): A Library for Gaussian Processes and Varia- tional Inference (2024)

    work page 2024

  12. [12]

    A&A685, A82 (2024)

    Edenhofer, Gordian, Zucker, Catherine, Frank, Philipp, Saydjari, Andrew K., Speagle, Joshua S., Finkbeiner, Douglas, Enßlin, Torsten A.: A parsec-scale galactic 3d dust map out to 1.25 kpc from the sun. A&A685, A82 (2024). DOI 10.1051/0004-6 361/202347628. URLhttps://doi.org/10.1051/0004-6361/202347628

    doi:10.1051/0004-6 2024

  13. [13]

    Reviews of Modern Physics97, 015005 (2025)

    Englert, J.M., J ¨ustel, D., Ntziachristos, V .: The need for optoacoustic microscopy. Reviews of Modern Physics97, 015005 (2025). DOI 10.1103/RevModPhys.97.015005

    work page doi:10.1103/revmodphys.97.015005 2025

  14. [14]

    Automated derivation of the adjoint of high-level transient finite element programs

    Farrell, P.E., Ham, D.A., Funke, S.W., Rognes, M.E.: Automated derivation of the adjoint of high-level transient finite element programs. SIAM Journal on Scientific Computing (2013). URLhttps://arxiv.org/abs/1204.5577

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  15. [15]

    Entropy23(7), 853 (2021)

    Frank, P., Leike, R., Enßlin, T.A.: Geometric variational inference. Entropy23(7), 853 (2021). URLhttps://arxiv.org/ abs/2105.10470

    Pith/arXiv arXiv 2021

  16. [16]

    Flow, Turbulence and Combustion65, 393–415 (2000)

    Giles, M.B., Pierce, N.A.: An introduction to the adjoint approach to design. Flow, Turbulence and Combustion65, 393–415 (2000). URLhttps://people.maths.ox.ac.uk/gilesm/files/ftc00.pdf

    work page 2000

  17. [17]

    Journal of the American Statistical Association117(540), 2182–2193 (2022)

    Goldman, J.V ., Sell, T., Singh, S.S.: Gradient-based markov chain monte carlo for bayesian inference with non-differentiable priors. Journal of the American Statistical Association117(540), 2182–2193 (2022). DOI 10.1080/01621459.2021.1909600. URLhttps://doi.org/10.1080/01621459.2021.1909600

    arXiv 2022

  18. [18]

    SIAM (2008)

    Griewank, A., Walther, A.: Evaluating Derivatives: Principles and Techniques of Algorithmic Differentiation, 2 edn. SIAM (2008). URLhttps://epubs.siam.org/doi/book/10.1137/1.9780898717761

    doi:10.1137/1.9780898717761 2008

  19. [19]

    DOI 10.48550/arXiv.2601.09439

    Haim, P., Ntziachristos, V ., Enßlin, T.A., J¨ustel, D.: DeepLight: A Sobolev-trained Image-to-Image Surrogate Model for Light Transport in Tissue (2026). DOI 10.48550/arXiv.2601.09439

    work page doi:10.48550/arxiv.2601.09439 2026

  20. [20]

    arXiv e-prints arXiv:2512.02204 (2025)

    Harth-Kitzerow, J., Gerland, U., Enßlin, T.A.: MoRSAIK: Sequence Motif Reactor Simulation, Analysis and Inference Kit in Python. arXiv e-prints arXiv:2512.02204 (2025). DOI 10.48550/arXiv.2512.02204

    work page doi:10.48550/arxiv.2512.02204 2025

  21. [21]

    Differentiable Matrix Elements with MadJax

    Heinrich, L., et al.: Differentiable matrix elements with madjax. arXiv preprint (2022). URLhttps://arxiv.org/abs/22 03.00057. ArXiv:2203.00057

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  22. [22]

    Hoffman, M.D., Blei, D.M., Wang, C., Paisley, J.: Stochastic variational inference. J. Mach. Learn. Res.14(1), 1303–1347 (2013)

    work page 2013

  23. [23]

    Computing in Science & Engineering9(3), 90–95 (2007)

    Hunter, J.D.: Matplotlib: A 2d graphics environment. Computing in Science & Engineering9(3), 90–95 (2007). DOI 10.1109/MCSE.2007.55

    doi:10.1109/mcse.2007.55 2007

  24. [24]

    Astroparticle Physics55, 37–50 (2014)

    Kissmann, R.: Picard: A novel code for the galactic cosmic ray propagation problem. Astroparticle Physics55, 37–50 (2014). DOI https://doi.org/10.1016/j.astropartphys.2014.02.002. URLhttps://www.sciencedirect.com/science/article/ pii/S0927650514000140

    doi:10.1016/j.astropartphys.2014.02.002 2014

  25. [25]

    arXiv preprint (2019)

    Knollm ¨uller, J., Enßlin, T.A.: Metric gaussian variational inference. arXiv preprint (2019). URLhttps://arxiv.org/ab s/1901.11033. ArXiv:1901.11033

    Pith/arXiv arXiv 2019

  26. [26]

    Journal of Machine Learning Research (2017)

    Kucukelbir, A., Tran, D., Ranganath, R., Gelman, A., Blei, D.M.: Automatic differentiation variational inference. Journal of Machine Learning Research (2017). URLhttps://jmlr.org/papers/volume18/16-107/16-107.pdf

    work page 2017

  27. [27]

    JAX-DIPS: Neural bootstrapping of finite discretization methods and application to elliptic problems with discontinuities

    Mistani, P., et al.: Jax-dips: Neural bootstrapping of finite discretization schemes. arXiv preprint/J. Comp. Sci. (2022). URL https://arxiv.org/abs/2210.14312. ArXiv:2210.14312

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  28. [28]

    arXiv preprint (2024)

    Newbury, R.: A review of differentiable simulators. arXiv preprint (2024). URLhttps://arxiv.org/abs/2407.05560. ArXiv:2407.05560

    Pith/arXiv arXiv 2024

  29. [29]

    arXiv preprint (2019)

    Phan, D., Pradhan, N., Jankowiak, M.: Composable effects for flexible and accelerated probabilistic programming in numpyro. arXiv preprint (2019). URLhttps://arxiv.org/abs/1912.11554. ArXiv:1912.11554

    Pith/arXiv arXiv 2019

  30. [30]

    JOSS (2023)

    Quera-Bofarull, A., et al.: Blackbirds: Black-box inference for differentiable simulators. JOSS (2023). URLhttps://www. theoj.org/joss-papers/joss.05776/10.21105.joss.05776.pdf

    work page 2023

  31. [31]

    Reeb, N., Hutschenreuter, S., Zehetner, P., Ensslin, T., Albert, A., Alves, S., Andr´e, M., Anghinolfi, M., Anton, G., Ardid, M., Aubert, J.J., Aublin, J., Baret, B., Basa, S., Belhorma, B., Bendahman, M., Bertin, V ., Biagi, S., Bissinger, M., Boumaaza, J., Bouta, M., Bouwhuis, M.C., Br ˆanzas,, H., Bruijn, R., Brunner, J., Busto, J., Caiffi, B., Capone,...

    work page 2023

  32. [32]

    Journal of Open Source Software9(98), 6532 (2024)

    Roth, J., Reinecke, M., Edenhofer, G.: Jaxbind: Bind any function to jax. Journal of Open Source Software9(98), 6532 (2024). DOI 10.21105/joss.06532. URLhttps://doi.org/10.21105/joss.06532

    work page doi:10.21105/joss.06532 2024

  33. [33]

    A versatile PYTHON library for signal inference

    Selig, M., Bell, M.R., Junklewitz, H., Oppermann, N., Reinecke, M., Greiner, M., Pachajoa, C., Enßlin, T.A.: NIFTY - Numerical Information Field Theory. A versatile PYTHON library for signal inference. A&A554, A26 (2013). DOI 10.1051/0004-6361/201321236

    doi:10.1051/0004-6361/201321236 2013

  34. [34]

    A&A693, A139 (2025)

    S ¨oding, L., Edenhofer, G., Enßlin, T.A., Frank, P., Kissmann, R., Phan, V .H.M., Ram ´ırez, A., Zandinejad, H., Mertsch, P.: Spatially coherent 3D distributions of HI and CO in the Milky Way. A&A693, A139 (2025). DOI 10.1051/0004-6361/2024 51361

    work page doi:10.1051/0004-6361/2024 2025

  35. [35]

    Annalen der Physik531(3), 1800290 (2019)

    Steininger, T., Dixit, J., Frank, P., Greiner, M., Hutschenreuter, S., Knollm ¨uller, J., Leike, R., Porqueres, N., Pumpe, D., Reinecke, M., Sraml, M., Varady, C., Enßlin, T.: NIFTy 3 - Numerical Information Field Theory: A Python Framework for Multicomponent Signal Inference on HPC Clusters. Annalen der Physik531(3), 1800290 (2019). DOI 10.1002/andp.20180 0290

    work page doi:10.1002/andp.20180 2019

  36. [36]

    Applied Optics44(6), 876 (2005)

    Tarvainen, T., Vauhkonen, M., Kolehmainen, V ., Kaipio, J.P.: Hybrid radiative-transfer–diffusion model for optical tomogra- phy. Applied Optics44(6), 876 (2005). DOI 10.1364/ao.44.000876

    work page doi:10.1364/ao.44.000876 2005

  37. [37]

    Computer Methods in Applied Mechanics and Engineering447, 118401 (2025)

    Thiagarajan, P., Zaki, T.A., Shields, M.D.: Accelerating hamiltonian monte carlo for bayesian inference in neural networks and neural operators. Computer Methods in Applied Mechanics and Engineering447, 118401 (2025). DOI https://doi.org/10 .1016/j.cma.2025.118401. URLhttps://www.sciencedirect.com/science/article/pii/S0045782525006735

    work page arXiv 2025

  38. [38]

    arXiv preprint arXiv:1610.09787 (2016)

    Tran, D., Kucukelbir, A., Dieng, A.B., Rudolph, M., Liang, D., Blei, D.M.: Edward: A library for probabilistic modeling, inference, and criticism. arXiv preprint arXiv:1610.09787 (2016)

    Pith/arXiv arXiv 2016

  39. [39]

    Academic Press (2001)

    Trottenberg, U., Oosterlee, C.W., Sch ¨uller, A.: Multigrid. Academic Press (2001). URLhttps://www.elsevier.com/boo ks/multigrid/trottenberg/978-0-08-047956-9

    work page 2001

  40. [40]

    A&A690, A102 (2024)

    Tsouros, Alexandros, Bendre, Abhijit B., Edenhofer, Gordian, Enßlin, Torsten, Frank, Philipp, Mastorakis, Michalis, Pavlidou, Vasiliki: Nonparametric bayesian reconstruction of galactic magnetic fields using information field theory - the inclusion of line-of-sight information in ultrahigh-energy cosmic-ray backtracing. A&A690, A102 (2024). DOI 10.1051/00...

    doi:10.1051/0004-6361/20 2024

  41. [41]

    Bringing PDEs to JAX with forward and reverse modes automatic differentiation

    Yashchuk, I.: Bringing pdes to jax with forward and reverse modes of automatic differentiation. arXiv preprint (2023). URL https://arxiv.org/abs/2309.07137. ArXiv:2309.07137

    work page internal anchor Pith review Pith/arXiv arXiv 2023

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.