Pith. sign in

REVIEW 3 major objections 4 minor 32 references

MANGO: An Autodiff Neutrino Oscillation Engine for Differentiable Analysis Pipelines

T0 review · 3 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read MANGO makes neutrino oscillation probabilities exactly differentiable with respect to every input—geometry, Earth-shell densities, and BSM couplings—at a constant 2.5–3x forward cost, and carries derivatives through the inverse Fisher…

desk verdict A genuinely useful, well-validated JAX oscillation engine whose one load-bearing marketing claim—singularity-free exact derivatives at degenerate spectra—is not actually demonstrated. read the letter →

arxiv 2608.13429 v1 pith:QYQR2SWM submitted 2026-08-13 hep-ex hep-ph

Pierre Granger This is my paper · ORCID
classification hep-exhep-ph
keywords neutrinooscillationsautomaticdifferentiationmattereffectsPREMEarthmodelFisherinformationbeyond-standard-modelsearchesexperimentaldesigncomputationalmethods
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

Neutrino oscillation probabilities are routine to compute, but their exact derivatives are not part of any standard engine. This paper presents MANGO, a neutrino oscillation engine built from automatic-differentiation primitives, and claims that every output is exactly differentiable with respect to every input: the six mixing parameters, trajectory geometry such as zenith angle and detector depth, each of the 369 density, electron-fraction, and shell-radius parameters of a layered PREM Earth, and beyond-standard-model couplings. The practical payoff is that reverse-mode differentiation evaluates all 369 Earth sensitivities in a single pass at only 2.5–3 times the forward cost, and gradients continue downstream through flux weighting, binning, detector response, and even the inverse Fisher matrix. The paper demonstrates this on a stylized Earth-tomography analysis in which the marginalized core-density uncertainty is itself differentiated with respect to detector angular resolution, a design metric that analytic probability formulas cannot produce.

What carries the argument

The load-bearing object is a divided-difference evaluation of the matrix exponential: instead of diagonalizing with eigenvectors, whose derivatives contain $1/(\lambda_i-\lambda_j)$ singularities, the propagator $\exp(-iHL)$ is expanded via Newton divided differences of the exponential over the eigenvalues of $H$, which is exact in the confluent limit. For three flavors the eigenvalues come from a closed-form trigonometric cubic, and the whole graph is built from automatic-differentiation primitives so forward and reverse modes traverse the same code. The chord-geometry construction maps zenith angle, production height, and detector depth to the segment-length vector in closed form, which is what makes derivatives with respect to geometry and individual Earth shells well defined.

What would settle it

Set up an exactly degenerate spectrum in the default divided-difference backend (for example, $\Delta m^2_{41}=\Delta m^2_{31}$ for a 3+1 model, or an exactly degenerate three-flavor point) and compare the automatic-differentiation derivative against central finite differences refined over several step sizes; if the error does not converge to the finite-difference truncation level, the smoothness claim is falsified. The paper already reports a concrete failure for its eigenvector fallback at such a point ($-67.8$ versus $-5.54$), so the same test applied to the production backend would settle the question.

Watch

Extended reading notes

Core claim

The central claim is that a composable, JAX-native engine can make neutrino oscillation pipelines differentiable end to end. For a piecewise-constant layered Earth, propagation is an ordered product of segment propagators $\exp(-iH_k L_k)$, with segment lengths determined by a closed-form chord construction, so all probability derivatives arise from exact automatic differentiation rather than hand-derived formulas. The paper verifies three-flavor and layered-Earth probabilities against external benchmarks to $10^{-9}$–$10^{-5}$ and validates every differentiation path against finite differences, with relative deviations typically $10^{-10}$–$10^{-9}$ and worst case $7\times10^{-7}$. In the worked example, one forward-mode sweep builds the full Fisher matrix; one more nested automatic-differentiation pass through its inverse yields a design sensitivity of $\partial\sigma(\ln\rho_{\mathrm{core}})/\partial\sigma_{\cos\theta_z}=1.186$ and a curvature of $-3.64$, reproducing an explicit scan to 0.1%.

Load-bearing premise

The entire exact-gradient program depends on the eigenvalue solve returning a true derivative when two neutrino masses become exactly equal; the paper verifies the forward probability there but does not report a derivative-level test for its default backend, and its fallback backend demonstrably gives the wrong derivative at an exact degeneracy.

Editorial extensions

If this is right

  • At 2.5–3x forward cost, a single reverse pass returns gradients for all six standard oscillation parameters, 42 geometry parameters, and 369 layered-Earth matter parameters, where central finite differencing would need 738 forward evaluations.
  • Analysis objectives—expected counts, likelihoods, and Fisher information—remain differentiable past the probability stage, since flux weighting, binning, and detector migration are themselves automatic-differentiation functions.
  • Beyond-standard-model sensitivities (NSI $\varepsilon$, sterile angles, Lindblad decoherence $\gamma$, and non-unitary $\alpha$) are obtained exactly at the Standard-Model point, where limit-setting fits linearize, without manual derivation.
  • Differentiating through the inverse Fisher matrix gives experimental-design derivatives such as $\partial\sigma/\partial\sigma_{\cos\theta_z}=1.186$, with second curvature $-3.64$, enabling detector-design optimization in a single pass.
  • Validation matches external codes to $10^{-9}$–$10^{-5}$ in probability and finite differences to $10^{-10}$–$10^{-5}$ in derivatives, so the engine can serve as a drop-in forward and sensitivity calculator.

Reading between the lines

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

  • If the degeneracy-smoothness claim survives a derivative-level test, gradient-based sampling (HMC, variational inference) over hundreds of Earth and nuisance parameters becomes practical for global oscillation fits—a route the paper gestures at but does not itself execute.
  • The same nested differentiation through an inverse Fisher matrix generalizes to optimizing bin edges, detector energy resolution, and exposure allocation for any oscillation measurement, not just core tomography.
  • Extending the divided-difference construction to arbitrary $N$ flavors would remove the paper's remaining eigenvector fallback, whose derivative fails at exact degeneracy ($-67.8$ versus $-5.54$); until then, $3+N$ sterile fits at level crossings need care.
  • The square-root cusps at shell-grazing zenith angles imply gradient-based optimizers should clip or smooth near critical $\cos\theta_z$; the paper exposes a function for those critical angles, so the practical impact can be tested directly.
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

3 major / 4 minor

Summary. The paper introduces MANGO, a JAX-based neutrino oscillation engine that computes forward probabilities for vacuum, constant-density, layered-PREM, arbitrary-profile, and adiabatic solar propagation, and that provides automatic differentiation of every output with respect to all inputs, including oscillation parameters, geometry, Earth-shell densities, and BSM couplings. The central technical claim is that a divided-difference matrix-exponential backend avoids eigenvector-derivative singularities and thus supplies C^infty smooth, exact derivatives even at degenerate spectra. The paper reports reverse-mode Jacobian costs of 2.5-3x the forward pass, validates probabilities against OscProb and NuFast/NuFast-Earth to 1e-9-1e-5, checks derivatives against finite differences, and demonstrates a full differentiation chain through a binned Poisson Fisher matrix and its inverse for Earth tomography, including derivatives of the marginalized uncertainty with respect to detector angular resolution.

Significance. If the claims hold, MANGO would fill a genuine gap: existing oscillation codes do not provide input-agnostic exact derivatives, and CHIC covers only constant-density three-flavor propagation. The paper's strengths are its broad validation program (external codes, analytic limits, finite differences, and cross-backend consistency), its open-source release with reproducible benchmark scripts and a Zenodo archival DOI, and its demonstration that reverse-mode cost is essentially independent of input dimension, which makes 369-parameter sensitivity calculations practical. The end-to-end Fisher-matrix differentiation and the functional solar-tomography kernel are interesting and potentially useful for gradient-based inference and experimental design. However, the marquee singularity-free differentiability claim at exact spectral degeneracies is not supported by the evidence presented, and one validation-table entry appears to conflict with the text; these issues need to be resolved before the central claim can be accepted.

major comments (3)
  1. [Section III, item 1; Section VII.d; Section IX] The advertised C^infty, singularity-free differentiability at exact mass level-crossings is not established for the default divided-difference backend. Section III acknowledges that the trigonometric cubic eigenvalue solve passes through arccos(r)/3, whose derivative diverges as r approaches +/-1, and that an exactly degenerate pair is resolved only to about 5e-7 of the spectral radius. A forward-propagator error of 5e-14 does not bound the derivative error, because AD differentiates the implemented approximate function; near r=+/-1 the derivative of that approximation can contain large spurious components. No AD-versus-finite-difference test at an exactly degenerate three-flavor spectrum is reported, and the only degenerate derivative test reported anywhere, Section VII.d for the eigh fallback in a 3+1 model, fails badly (-67.8 versus -5.54). The authors should supply a derivative-level validation at exact degeneracy for the default backend, ideally with a scan approaching degeneracy, or explicitly restrict the singularity-free claim.
  2. [Table III] The validation table contains a row reading 'AD vs. FD 3+1 sterile partial P/partial Delta m^2_41 at Delta m^2_41 = Delta m^2_31: 12x(see text)' that is incomplete and appears to contradict Section VII.d, which reports an 'unreliable AD derivative of -67.8' at exactly this degeneracy. Either the metric and numerical value should be stated explicitly, or the row should be removed; as written, the table implies that exact-degeneracy derivatives are validated while the text says the opposite.
  3. [Section IV; Section VII.e] The unconditional statement in Section IV that automatic differentiation 'yields exact derivatives of any output with respect to any input' and the abstract's claim that every computed quantity is differentiable are too broad in light of the square-root cusps of partial P/partial cos theta_z at shell-grazing angles acknowledged in Section VII.e and the degenerate-spectrum failure in Section VII.d. The central differentiability claims should be stated with explicit measure-zero caveats so that 'exact' is understood to mean 'exact away from the enumerated non-differentiable sets'.
minor comments (4)
  1. [Sections V and IX; Acknowledgements] The text mentions 'Nuwaves' as a validation or comparison tool, but no Nuwaves reference appears in the bibliography; a reference should be added or the name removed.
  2. [Section II.B.e, Eq. (10)] The production weights w_k are used in the expression for F_i(r) before they are defined in the following line; reordering the definitions would improve readability.
  3. [Table IV caption] The one-sided-difference convergence sequence (8e-4, 3e-6, 3e-8, 7e-10) is informative, but the step sizes h=1e-4 through 1e-7 should be shown alongside the values in the caption rather than only in prose after the table.
  4. [Section VII.d] The sentence claiming that the divided-difference backend 'guarantees exact, smooth derivatives across degenerate points' repeats the unsupported central claim; even if the authors choose to retain the claim, it needs a reference to the new derivative-level validation requested above.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: MANGO's AD claims are definitional properties of the JAX graph, and its probability and derivative claims are checked against external benchmarks and finite differences rather than against the claims themselves.

full rationale

MANGO's central claims are self-contained engineering claims about automatic differentiation through a JAX computational graph, not physical derivations with fitted inputs. The 'exact derivatives by construction' statement in Section IV is the standard chain-rule property of AD over the implemented primitives, and the paper does not attempt to derive oscillation probabilities from assumptions that already contain the target derivatives. Forward probabilities are benchmarked against external codes (OscProb, NuFast-Earth), analytic limits, and its own NuFast port under an external MIT license; AD derivatives are checked against finite differences of the same forward model, which is the appropriate consistency check for AD correctness rather than a reimport of the claim. The only self-citation ([11], Gasiorowski et al., on which Granger is a coauthor) appears in the introduction as contextual motivation for differentiable programming and is not load-bearing. The admitted limitation in Section VII.d—the eigh fallback's derivative of -67.8 versus -5.54 at an exact 3+1 degeneracy, and the absence of a derivative-level test at exact degeneracy for the default divided-difference backend—is a numerical robustness and support concern, not a circularity: no equation in the paper reduces to its own input, and no fitted parameter is renamed as a prediction.

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

The engine's central derivability claim does not rest on any fitted constants in this paper. Oscillation parameters are fixed to the external NuFIT 5.1 normal-ordering best fit, and PREM is taken from the literature. The illustrative Fisher analysis introduces design parameters (six zone scales, flux nuisances, theta23, Delta m2_31) that are inputs of the demonstration, not fitted outputs. The main assumptions are the physical model choices (Hamiltonian, PREM, adiabatic solar MSW) and the reliability of JAX differentiation; the contested degeneracy assumption is listed separately.

assumptions (6)
  • standard math Standard three-flavor Hamiltonian in Eq. (1) with PMNS parametrization and VCC = sqrt(2) GF Ne.
    The starting physics model; accepted and externally benchmarked.
  • standard math Ordered product of segment propagators S = exp(-i H_k L_k) for piecewise-constant matter.
    Requires piecewise-constant Hamiltonian within each shell segment; standard for layered Earth propagation.
  • domain assumption PREM spherical-shell Earth with two-zone electron fraction and straight-line chord geometry.
    Standard model for oscillation codes but ignores lateral heterogeneities and non-spherical structure; Section VII.c lists 3D models as future work.
  • domain assumption JAX reverse-mode AD returns the derivative of the implemented numerical function.
    The entire differentiation machinery depends on JAX's correctness and on float64 precision; the paper enables float64 on import.
  • domain assumption Adiabatic MSW solar propagation with incoherent production weights, Eq. (10).
    Approximation that decoheres mass eigenstates on the Sun-Earth baseline; validated to 6e-15 against a numerical MSW reference.
  • ad hoc to paper Divided-difference backend remains AD-safe at exactly degenerate spectra.
    The paper asserts this to support the singularity-free claim, but reports only forward error at degeneracy and notes the arccos derivative diverges; contradicted by the eigh fallback test.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MANGO: An Autodiff Neutrino Oscillation Engine for Differentiable Analysis Pipelines." pith.science (2026). https://pith.science/paper/QYQR2SWM

@misc{pith2026260813429,
  author       = {Pith},
  title        = {Pith review of: MANGO: An Autodiff Neutrino Oscillation Engine for Differentiable Analysis Pipelines},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QYQR2SWM}},
  note         = {Machine review of arXiv:2608.13429}
}
read the original abstract

Computing neutrino oscillation probabilities is a solved problem; computing their derivatives is not. We present MANGO (MANGO: A Neutrino Gradient Oscillator), a composable oscillation engine in which every computed quantity is differentiable with respect to all inputs, including propagation geometry, detector depth, and individual Earth-shell densities. None of these quantities appear in traditional analytic probability formulas. The engine supports vacuum, constant-density, layered-PREM, arbitrary-profile, and adiabatic solar propagation, alongside front-ends for non-standard interactions, 3+N sterile states, decoherence, and non-unitary mixing. Because reverse-mode cost depends on output rather than input dimension, evaluating sensitivities incurs a constant 2.5-3x forward-pass overhead. As a result, calculating sensitivities for all 369 density, electron-fraction, and shell-radius parameters of a layered Earth costs no more than the 6 standard oscillation parameters. By maintaining exact sensitivity signals, MANGO allows gradients to flow continuously past the probability stage and through detector response, event weighting, binning, and likelihoods. We demonstrate this on a stylized Earth-tomography analysis. In a single pass, MANGO computes the marginalized uncertainty on a six-zone radial density model and, by differentiating through the inverse Fisher matrix, evaluates its sensitivity with respect to detector angular resolution. This provides an experimental-design metric unreachable using traditional analytic probability formulas. Three-flavor and layered-Earth probabilities match external benchmarks (OscProb, NuFast-Earth) to within 10^-9 to 10^-5, while all forward models, BSM limits, and differentiation paths are verified against exact analytic solutions and finite differences.

Figures

Figures reproduced from arXiv: 2608.13429 by the authors.

Figure 1
Figure 1. FIG. 1. The chord construction of equation (2). A neutrino [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. Atmospheric oscillograms computed with [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3. Muon-neutrino survival probability through the PREM Earth (top left) and exact derivatives spanning three standard [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: FIG. 4. Derivatives with respect to the beyond-standard-model parameters themselves, evaluated at the standard-model point, [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5. Solar neutrino oscillation Automatic Differentiation benchmark and functional core tomography kernel. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6. The differentiation chain of section VI. (a) Derivative of the expected bin counts with respect to the bulk-core density, in [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 21 canonical work pages

  1. [1]

    Wendell R (for the Super-Kamiokande Collaboration) Prob3++: three-flavor neutrino oscillation proba- bility calculatorhttps://github.com/rogerwendell/ Prob3plusplus

    work page

  2. [2]

    [×103 eV 2] 0.05 0.10 0.15 0.20 0.25 0.30 0.35 solar radial position r/R 10 1 100 101 neutrino energy E [MeV] pp end 7Be hep end (b) functional core tomography kernel via VJP 0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 sensitivity P/ m2 21 [103 eV 2] 0.00 0.02 0.04 0.06 0.08 0.10 0.12 functional kernel | Pee ln Ne(r) | FIG. 5. Solar neutrino oscillation Autom...

    work page

  3. [3]

    cross-code

    (red, right axis) acrossE ν ∈[0.1,30] MeV, featuring the sharp MSW resonance peak atE= 4.41 MeV (1.68×10 3 eV−2). Overlaid markers show two-point central finite-difference evaluations matching AD values to 1.3×10 −10 relative precision.Panel (b): 2D functional solar core tomography kernel|δP ee/δlnN e(r)|evaluated across 80 radial shellsr∈[0.01,0.35]R ⊙ i...

    work page

  4. [4]

    Argu¨ elles C A, Salvado J and Weaver C N 2022 Comput. Phys. Commun.277108346 [arXiv:2112.13804]

    arXiv 2022

  5. [5]

    Denton P B and Parke S J 2024 Phys. Rev. D110073005 [arXiv:2405.02400]

    work page Pith review arXiv 2024

  6. [6]

    Coelho J A B, Stocco D, Domi A, Pestes R, Loup Met al.OscProb: neutrino oscillation probability calculator, Zenodo,10.5281/zenodo.6347002(concept DOI; latest version v2.4.0,10.5281/zenodo.21335289, 2026)

    work page doi:10.5281/zenodo.6347002(concept 2026

  7. [7]

    Huber P, Lindner M and Winter W 2005 Comput. Phys. Commun.167195 [hep-ph/0407333]

    arXiv 2005

  8. [8]

    Dorigo T, Giammanco A, Vischia Pet al.2023Rev. Phys. 10100085

    work page

Show all 32 references
  1. [9]

    de Castro P and Dorigo T 2019 Comput. Phys. Commun. 244170 [arXiv:1806.04743]

    2019 arXiv

  2. [10]

    Open Source Softw.62823

    Heinrich L, Feickert M, Stark G and Cranmer K 2021J. Open Source Softw.62823

  3. [11]

    Phys.: Conf

    Simpson N and Heinrich L 2023J. Phys.: Conf. Ser. 2438012105

  4. [12]

    Astrophys.615 [arXiv:2302.05163]

    Campagne J-E, Lanusse F, Zuntz Jet al.2023Open J. Astrophys.615 [arXiv:2302.05163]

    arXiv

  5. [13]

    Iacovelli F, Mancarella M, Foffa S and Maggiore M 2022 Astrophys. J. Suppl.2632 [arXiv:2207.06910]

    2022 arXiv

  6. [14]

    Learn.: Sci

    Strong G Cet al.2024Mach. Learn.: Sci. Technol.5 035002 [arXiv:2309.14027]

    arXiv

  7. [15]

    Learn.: Sci

    Gasiorowski S, Chen Y, Nashed Y, Granger P, Mironov Cet al.2024Mach. Learn.: Sci. Technol.5025012 [arXiv:2309.04639]

    arXiv

  8. [16]

    com/jax-ml/jax

    Bradbury Jet al.2018 JAX: composable transforma- tions of Python+NumPy programshttps://github. com/jax-ml/jax

    2018

  9. [17]

    Phys.B908250 [arXiv:1511.05154]

    Winter W 2016Nucl. Phys.B908250 [arXiv:1511.05154]

    arXiv

  10. [18]

    Desdoigts L, Pope B J S, Dennis J and Tuthill P G 2023J. Astron. Telesc. Instrum. Syst.9028007 [arXiv:2301.05602]

    arXiv

  11. [19]

    Fern´ andez-Men´ endez P 2026 J. Phys. G53075002 [arXiv:2512.16427] 15

    2026 arXiv

  12. [20]

    Kallenborn F, Hundt C, B¨ oser S and Schmidt B 2019 Comput. Phys. Commun.234235

    2019

  13. [21]

    Beyond scalar parameter derivatives, reverse-mode au- tomatic differentiation enables functional tomography of continuous matter profiles

    exhibits a prominent MSW resonance peak reaching 1.68×10 3 eV−2 atE= 4.41 MeV, demon- strating AD’s ability to differentiate smoothly through MSW level-crossings without numerical instability. Beyond scalar parameter derivatives, reverse-mode au- tomatic differentiation enable...

  14. [22]

    High Energy Phys.05187 [arXiv:2110.00003]

    Kelly K J, Machado P A N, Martinez-Soler I and Perez-Gonzalez Y F 2022J. High Energy Phys.05187 [arXiv:2110.00003]

    arXiv

  15. [23]

    Phys.1537 [arXiv:1803.05901]

    Donini A, Palomares-Ruiz S and Salvado J 2019Nat. Phys.1537 [arXiv:1803.05901]

    arXiv

  16. [24]

    J.621L85 [astro-ph/0412440]

    Bahcall J N, Serenelli A M and Basu S 2005Astrophys. J.621L85 [astro-ph/0412440]

    arXiv

  17. [25]

    High Energy Phys.09178 [arXiv:2007.14792]; NuFIT 5.1 (2021)http://www

    Esteban I, Gonz´ alez-Garc ´ ıa M C, Maltoni M, Schwetz T and Zhou A 2020J. High Energy Phys.09178 [arXiv:2007.14792]; NuFIT 5.1 (2021)http://www. nu-fit.org

    2021 arXiv

  18. [26]

    Earth Planet

    Dziewonski A M and Anderson D L 1981Phys. Earth Planet. Inter.25297

  19. [27]

    com/patrick-kidger/diffrax

    Kidger P J 2021On Neural Differential Equations PhD thesis Mathematical Institute, University of Ox- ford [arXiv:2202.02435];diffrax:https://github. com/patrick-kidger/diffrax

    arXiv

  20. [28]

    Methods17261

    Virtanen Pet al.2020Nat. Methods17261

  21. [29]

    Denton P B and Parke S J 2026 Phys. Rev. D113053002 [arXiv:2511.04735]

    2026

  22. [30]

    Harris C Ret al.2020Nature585357

  23. [31]

    in a 3+1 sterile model leaves forward probabilities unaffected but yields an unreliable AD derivative of−67.8 compared to the true finite-difference value of−5.54.mangoresolves this for standard 3-flavor oscillations by providing the divided-difference backend as its default p...

  24. [32]

    Hunter J D 2007Comput. Sci. Eng.990

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.