arxiv: 2605.02252 · v1 · submitted 2026-05-04 · 💻 cs.RO · cs.MS

Recognition: 3 theorem links

· Lean Theorem

Exact Higher-Order Derivatives for SE(3) via Analytical/AD Methods

Frank O. Kuehnel

Authors on Pith no claims yet

Pith reviewed 2026-05-08 18:51 UTC · model grok-4.3

classification 💻 cs.RO cs.MS

keywords SE(3)automatic differentiationHessianLie groupsnegative log-likelihoodrobotics estimationexact derivatives

0 comments

The pith

A hybrid analytical-automatic differentiation seam at the point-action interface delivers exact SE(3) Hessians and higher-order tensors five times faster than finite-differencing the gradient.

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

The paper establishes a compact recipe for moving from a single generic gradient expression for an SE(3) negative log-likelihood to exact Hessians and higher derivative tensors without writing second derivatives by hand or tuning finite-difference steps. The writer places closed-form Lie-group Jacobians up to the transformation y equals Tx and lets automatic differentiation handle everything after that point. The same source code is then run with ordinary scalars for first derivatives, vector-seeded dual numbers for Hessians in one forward pass, and nested dual numbers for tensors. On a 6-DoF five-landmark benchmark this produces machine-precision agreement with a full nested-AD reference while running roughly five times faster than finite-differencing the AD gradient, at the cost of about seventy extra lines of analytical Jacobian code plus a fix for a removable singularity in the standard rotation basis.

Core claim

By writing the NLL gradient once as a generic function over scalar type and applying closed-form Lie-group Jacobians only up to the y equals Tx interface, the identical source can be instantiated with floating-point scalars, vector-seeded dual numbers, or nested dual numbers to obtain gradients, exact Hessians, or higher-order tensors. On the representative 6-DoF five-landmark benchmark the seeded-Hessian instantiation matches a nested-AD oracle to machine precision and runs approximately five times faster than finite-differencing the AD gradient, while adding roughly seventy lines of analytical code over a pure-AD baseline and correcting a removable singularity in the SO(3)/SE(3) scalar map

What carries the argument

The analytical/AD seam placed at the point-action interface y equals Tx, where closed-form Lie-group Jacobians are used up to that point and automatic differentiation is applied only beyond it, allowing the same generic gradient code to be reused for exact higher-order results via seeded dual numbers.

If this is right

Exact Newton steps become available for SE(3) objectives without additional manual derivative work.
Observed-information matrices for covariance estimation can be obtained directly from the same gradient source.
Higher-order derivative tensors needed for covariance correction or advanced error analysis are produced in a single forward pass.
Development of new SE(3) estimation objectives requires only the first-order gradient plus the seventy-line Jacobian supplement.

Where Pith is reading between the lines

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

The same seam placement could be tested on other Lie groups such as SE(2) or the special Euclidean group in higher dimensions to see whether the fivefold speedup persists.
The stabilized basis correction may remove similar NaN issues in differentiation pipelines that reuse the standard SO(3) scalar representation.
Existing robotics libraries could adopt the pattern to expose exact second-order information to users who already write first-order objectives.

Load-bearing premise

The closed-form Lie-group Jacobians supplied up to the transformation interface must be correctly implemented, and the 6-DoF five-landmark benchmark must be representative of the SE(3) negative log-likelihood problems that arise in practice.

What would settle it

Implement the hybrid method on the stated 6-DoF five-landmark NLL benchmark, compare its Hessian values and runtime against both finite-differencing of the AD gradient and a nested-AD oracle, and check whether the values agree to machine precision while the runtime is approximately five times lower.

read the original abstract

Fast prototyping of new SE(3) estimation objectives remains awkward in practice. Modern Lie-group frameworks -- GTSAM, manif, Sophus, SymForce, Ceres -- target first-order workloads through different code-generation and automatic-differentiation strategies, each optimized for a particular seam between hand-derived geometry and generic differentiation. The remaining gap is a compact, AD-safe path from these first-order primitives to exact Hessians, observed-information matrices, and higher-order derivative tensors: the quantities needed for exact Newton steps, observed-information covariance estimates, and covariance correction. This paper presents a hybrid analytical/AD recipe for SE(3) negative log-likelihoods. The practitioner writes the NLL gradient once, generic over a scalar type, and places the analytical/AD seam at the point-action interface y = Tx. Closed-form Lie-group Jacobians are used up to this interface; AD is applied only beyond it. The same source is then instantiated with ordinary floating-point scalars for gradients, vector-seeded dual numbers for exact Hessians in a single forward-mode pass, and nested dual numbers for higher-order derivative tensors. On a representative 6-DoF, 5-landmark SE(3) NLL, the advocated seeded-Hessian path is approximately 5x faster than finite-differencing the AD gradient on this benchmark while matching a nested-AD oracle to machine precision. The implementation adds roughly 70 lines of analytical-Jacobian code over an AD-only baseline. We also identify and fix a removable singularity in the standard SO(3)/SE(3) scalar basis that would otherwise produce NaNs at the origin under seeded AD, and we audit which Lie-group derivative tensors require this stabilized basis. The result is a practical path from rapidly written SE(3) objectives to exact higher-order derivatives, with predictable runtime and no finite-difference tuning.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This paper gives a compact hybrid recipe for exact SE(3) Hessians by splitting at the y=Tx interface and seeding dual numbers, with a reported 5x speedup on one benchmark and a fix for the origin singularity.

read the letter

The main point is that the authors show how to reuse a single generic gradient expression for SE(3) negative log-likelihoods to get exact first-order, Hessian, and higher tensors. They put the analytical/AD seam right at the point-action y = Tx, run closed-form Lie-group Jacobians up to there, and apply automatic differentiation only after. Vector-seeded dual numbers handle the Hessian in one forward pass, and they also stabilize the scalar basis to remove NaNs at the identity under AD seeding. On their 6-DoF 5-landmark test case the seeded path runs about 5 times faster than finite-differencing the AD gradient while matching nested AD to machine precision, and the whole addition is roughly 70 lines over a pure-AD baseline. That combination of seam placement, seeding strategy, and singularity fix is the concrete new piece relative to the first-order libraries they cite. The approach is internally consistent and the benchmark numbers are stated plainly. The soft spot is the narrow experimental scope. All timing and accuracy claims rest on a single representative NLL instance, so it remains unclear how the speedup holds up on larger maps, different landmark counts, or objectives with more complex measurement models. The method also assumes the closed-form Jacobians up to the seam are implemented correctly, which is standard but still a potential source of silent error. No other load-bearing gaps appear in the argument as given. This is aimed at people already working with SE(3) estimation code who need reliable second-order information for Newton steps or covariance work without hand-deriving everything or tuning finite differences. A practitioner who wants a low-overhead path from first-order primitives to exact Hessians will get direct value. The paper deserves a serious referee because the core construction is explicit, the performance claim is quantified, and the singularity fix is a useful detail even if more varied test cases would strengthen the case for broader adoption.

Referee Report

0 major / 2 minor

Summary. The paper presents a hybrid analytical/automatic-differentiation recipe for exact higher-order derivatives (gradients, Hessians, and higher tensors) of SE(3) negative log-likelihoods. The method writes the NLL gradient once over a generic scalar type, places the analytical/AD seam at the point-action interface y=Tx, applies closed-form Lie-group Jacobians up to that point, and uses vector-seeded dual numbers for single-pass exact Hessians or nested dual numbers for higher-order tensors. On a 6-DoF 5-landmark benchmark NLL, the seeded-Hessian path is reported to be approximately 5x faster than finite-differencing the AD gradient while matching a nested-AD oracle to machine precision; the approach adds roughly 70 lines of analytical Jacobian code and includes a fix for a removable singularity in the standard SO(3)/SE(3) scalar basis that otherwise produces NaNs at the origin under seeded AD.

Significance. If the benchmark results and singularity fix hold, the work offers a practical, low-overhead bridge between existing first-order Lie-group libraries (GTSAM, manif, Sophus, etc.) and the exact Hessians/observed-information matrices needed for Newton steps and covariance estimation. The hybrid seam, dual-number seeding strategy, and explicit audit of affected derivative tensors constitute reusable engineering contributions that reduce finite-difference tuning and support rapid prototyping of SE(3) objectives. The machine-precision match to nested AD and the scoped performance claim on a representative instance are concrete strengths.

minor comments (2)

[Abstract] Abstract: the reported 'approximately 5x faster' claim would be strengthened by including the precise timing ratio, hardware platform, and exact baseline implementation details (e.g., which AD library and finite-difference step size) in the results section.
The singularity fix and the audit of which Lie-group derivative tensors require the stabilized basis are valuable; a short table or enumerated list of the affected tensors (with before/after behavior) would improve clarity for readers implementing the method.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive review and the recommendation of minor revision. The assessment that the hybrid analytical/AD seam, dual-number seeding, and singularity fix constitute reusable engineering contributions is appreciated. No specific major comments were raised in the report.

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper describes a direct computational recipe that places an analytical/AD seam at the y=Tx interface, using standard closed-form Lie-group Jacobians up to that point and off-the-shelf AD beyond it. The central performance claim is scoped to runtime and precision measurements on one concrete 6-DoF 5-landmark NLL benchmark; these are empirical outcomes of executing the described procedure rather than quantities defined in terms of themselves or reduced to prior self-citations. No load-bearing step invokes a uniqueness theorem, fitted parameter renamed as prediction, or ansatz smuggled via citation. The singularity fix and tensor audit are internal consistency checks on the same construction. The derivation chain is therefore self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the existence and correctness of standard closed-form Lie-group Jacobians for SE(3) actions up to the point transformation interface; no free parameters, new entities, or ad-hoc axioms are introduced.

axioms (1)

domain assumption Closed-form expressions for the Lie-group Jacobians of SE(3) point actions exist and are stable once the identified singularity is removed.
The method invokes these Jacobians without re-deriving them.

pith-pipeline@v0.9.0 · 5640 in / 1479 out tokens · 69173 ms · 2026-05-08T18:51:53.505283+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

Cost/FunctionalEquation.lean (J(x)=½(x+x⁻¹)−1 ratio-symmetric coordinates) washburn_uniqueness_aczel echoes

?

echoes
ECHOES: this paper passage has the same mathematical shape or conceptual pattern as the Recognition theorem, but is not a direct formal dependency.

Work in s=ω⊤ω and implement D̃(s) ≡ D'(θ)/θ = 2 dD/ds = 1/360 + s/7560 + s²/201600 + ...

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

70 extracted references · 3 canonical work pages

[1]

IEEE Transactions on Automatic Control , volume =

Barrau, Axel and Bonnabel, Silvère , title =. IEEE Transactions on Automatic Control , volume =. 2017 , doi =

2017
[2]

Symmetry-Preserving Observers , journal =

Bonnabel, Silv\`. Symmetry-Preserving Observers , journal =
[3]

Invariant Extended

Bonnabel, Silv\`. Invariant Extended. Proceedings of the 48th IEEE Conference on Decision and Control (CDC) , pages =
[4]

A Code for Unscented

Brossard, Martin and Barrau, Axel and Bonnabel, Silv\`. A Code for Unscented. Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) , pages =. 2020 , doi =

2020
[5]

2022 , howpublished =

2022
[6]

2025 , howpublished =

Riseborough, Paul and. 2025 , howpublished =

2025
[7]

2016 , howpublished =

Riseborough, Paul , title =. 2016 , howpublished =

2016
[8]

2020 , howpublished =

Riseborough, Paul , title =. 2020 , howpublished =

2020
[9]

, title =

Pittelkau, Mark E. , title =. Journal of Guidance, Control, and Dynamics , volume =. 2003 , doi =

2003
[10]

2023 , howpublished =

Bresch, Mathieu and. 2023 , howpublished =

2023
[11]

2018 , howpublished =

Riseborough, Paul and. 2018 , howpublished =

2018
[12]

Proceedings of Robotics: Science and Systems (RSS) , address =

Martiros, Hayk and Miller, Aaron and Bucki, Nathan and Solliday, Bradley and Kennedy, Ryan and Zhu, Jack and Dang, Tung and Pattison, Dominic and Zheng, Harrison and Tomic, Teo and Henry, Peter and Cross, Gareth and VanderMey, Josiah and Sun, Alvin and Wang, Samuel and Holtz, Kristen , title =. Proceedings of Robotics: Science and Systems (RSS) , address ...

2022
[13]

, title =

Kuehnel, Frank O. , title =. 2026 , howpublished =

2026
[14]

and Price, Charles F

Deyst, John J. and Price, Charles F. , title =. IEEE Transactions on Automatic Control , volume =. 1968 , note =

1968
[15]

Proceedings of the 21st IFAC Symposium on Mathematical Theory of Networks and Systems (MTNS) , year =

Mahony, Robert and Trumpf, Jochen , title =. Proceedings of the 21st IFAC Symposium on Mathematical Theory of Networks and Systems (MTNS) , year =
[16]

Proceedings of the 54th IEEE Conference on Decision and Control (CDC) , year =

Trumpf, Jochen and Mahony, Robert and Hamel, Tarek , title =. Proceedings of the 54th IEEE Conference on Decision and Control (CDC) , year =
[17]

Proceedings of the 9th IFAC Symposium on Nonlinear Control Systems (NOLCOS) , year =

Mahony, Robert and Trumpf, Jochen and Hamel, Tarek , title =. Proceedings of the 9th IFAC Symposium on Nonlinear Control Systems (NOLCOS) , year =
[18]

IEEE Transactions on Automatic Control , volume =

Mahony, Robert and Hamel, Tarek and Pflimlin, Jean-Michel , title =. IEEE Transactions on Automatic Control , volume =. 2008 , doi =

2008
[19]

IEEE Transactions on Automatic Control , volume =

Lageman, Christian and Trumpf, Jochen and Mahony, Robert , title =. IEEE Transactions on Automatic Control , volume =
[20]

, booktitle =

Ge, Yixiao and van Goor, Pieter and Mahony, Robert E. , booktitle =. A Note on the Extended. 2023 , organization =

2023
[21]

IEEE Control Systems Letters , volume =

Ge, Yixiao and van Goor, Pieter and Mahony, Robert , title =. IEEE Control Systems Letters , volume =
[22]

The geometry of extended kalman filters on manifolds with affine connection,

Ge, Yixiao and van Goor, Pieter and Mahony, Robert , title =. arXiv preprint arXiv:2506.05728 , year =

work page arXiv
[23]

Control Engineering Practice , volume =

Ge, Yixiao and Delama, Giulio and Scheiber, Martin and Fornasier, Alessandro and van Goor, Pieter and Weiss, Stephan and Mahony, Robert , title =. Control Engineering Practice , volume =. 2026 , doi =

2026
[24]

, title =

Chirikjian, Gregory S. , title =
[25]

Continuous-Discrete Extended

Bourmaud, Guillaume and M\'. Continuous-Discrete Extended. Journal of Mathematical Imaging and Vision , volume =
[26]

Discrete Extended

Bourmaud, Guillaume and M\'. Discrete Extended. Proceedings of the European Signal Processing Conference (EUSIPCO) , pages =
[27]

, title =

Ye, Jianing and Chirikjian, Gregory S. , title =. arXiv preprint arXiv:2401.03425 , year =

work page arXiv
[28]

, title =

Wang, Yunfeng and Chirikjian, Gregory S. , title =. International Journal of Robotics Research , volume =
[29]

International Journal of Computer Vision , volume =

Pennec, Xavier and Thirion, Jean-Philippe , title =. International Journal of Computer Vision , volume =
[30]

and Ghaffari, Maani and Vasudevan, Ram and Eustice, Ryan M

Mangelson, Joshua G. and Ghaffari, Maani and Vasudevan, Ram and Eustice, Ryan M. , title =. 2019 , eprint =

2019
[31]

and Bruno, Marcelo G

Labsir, Samy and El Bouch, Sara and Bordin Jr., Claudio J. and Bruno, Marcelo G. S. , title =. Signal Processing , volume =. 2026 , doi =

2026
[32]

and Moskalev, Anatoli N

Varshalovich, Dmitri A. and Moskalev, Anatoli N. and Khersonskii, Valery K. , title =
[33]

, title =

Edmonds, Alan R. , title =. 1957 , note =

1957
[34]

2021 , note =

Sakurai, Jun John and Napolitano, Jim , title =. 2021 , note =

2021
[35]

and Rockmore, Daniel N

Kostelec, Peter J. and Rockmore, Daniel N. , title =. Journal of Fourier Analysis and Applications , volume =. 2008 , note =

2008
[36]

Sol\`. A Micro. arXiv preprint arXiv:1812.01537 , year =

work page arXiv
[37]

, title =

Barfoot, Timothy D. , title =
[38]

Proceedings of the American Mathematical Society , volume =

Wei, James and Norman, Edward , title =. Proceedings of the American Mathematical Society , volume =
[39]

IEEE Transactions on Robotics , volume =

Forster, Christian and Carlone, Luca and Dellaert, Frank and Scaramuzza, Davide , title =. IEEE Transactions on Robotics , volume =
[40]

Vision Algorithms: Theory and Practice , publisher =

Triggs, Bill and McLauchlan, Philip and Hartley, Richard and Fitzgibbon, Andrew , title =. Vision Algorithms: Theory and Practice , publisher =
[41]

The International Journal of Robotics Research , volume =

Thrun, Sebastian and Montemerlo, Michael , title =. The International Journal of Robotics Research , volume =
[42]

Chiuso, Alessandro and Soatto, Stefano , title =
[43]

Journal de Math\'

Rodrigues, Olinde , title =. Journal de Math\'. 1840 , note =
[44]

Cheng, Hui and Gupta, K. C. , title =. ASME Journal of Applied Mechanics , volume =. 1989 , doi =

1989
[45]

Novi Commentarii Academiae Scientiarum Petropolitanae , volume =

Euler, Leonhard , title =. Novi Commentarii Academiae Scientiarum Petropolitanae , volume =. 1775 , note =
[46]

Computer Methods in Applied Mechanics and Engineering , volume =

Argyris, John , title =. Computer Methods in Applied Mechanics and Engineering , volume =. 1982 , doi =

1982
[47]

and Mladenov, Iva\"

Mladenova, Clementina D. and Mladenov, Iva\". Vector Decomposition of Finite Rotations , journal =. 2011 , doi =

2011
[48]

Horn, Berthold K. P. , title =. Journal of the Optical Society of America A , volume =
[49]

Horn, Berthold K. P. and Hilden, Hugh M. and Negahdaripour, Shahriar , title =. Journal of the Optical Society of America A , volume =
[50]

, title =

Altman, Simon L. , title =
[51]

Proceedings of the European Conference on Computer Vision (ECCV) , pages =

Kanatani, Kenichi and Ohta, Naoya , title =. Proceedings of the European Conference on Computer Vision (ECCV) , pages =
[52]

IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =

Kanatani, Kenichi , title =. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =
[53]

Vision Algorithms: Theory and Practice , publisher =

Matei, Bogdan and Meer, Peter , title =. Vision Algorithms: Theory and Practice , publisher =
[54]

and Ioffe, Sergey and Haddon, John , title =

Forsyth, David A. and Ioffe, Sergey and Haddon, John , title =. Proceedings of the International Conference on Computer Vision (ICCV) , pages =
[55]

and Kanatani, Kenichi and Kanade, Takeo , title =

Morris, Daniel D. and Kanatani, Kenichi and Kanade, Takeo , title =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =
[56]

Proceedings of the International Conference on Computer Vision (ICCV) , year =

Bartoli, Adrien , title =. Proceedings of the International Conference on Computer Vision (ICCV) , year =
[57]

Close Range Photogrammetry and Machine Vision , publisher =
[58]

and Ahuja, Narendra , title =

Weng, Juyang and Huang, Thomas S. and Ahuja, Narendra , title =. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =
[59]

, title =

Minkler, George J. , title =
[60]

Griewank, Andreas and Walther, Andrea , title =
[61]

, title =

Jackson, James S. , title =
[62]

Ceres Solver , howpublished =

Agarwal, Sameer and Mierle, Keir and. Ceres Solver , howpublished =
[63]

Deray, Jeremie and Sol\`. Manif:. Journal of Open Source Software , volume =. 2020 , doi =

2020
[64]

Strasdat, Hauke , title =
[65]

, title =

Kuehnel, Frank O. , title =. Bayesian Inference and Maximum Entropy Methods in Science and Engineering (. 2004 , doi =

2004
[66]

, title =

Kuehnel, Frank O. , title =. 2005 , doi =

2005
[67]

, title =

Kuehnel, Frank O. , title =
[68]

, title =

Kuehnel, Frank O. , title =. 2026 , note =

2026
[69]

2012 , note =

Bry, Adam Parker , title =. 2012 , note =

2012
[70]

Bry, Adam and Bachrach, Abraham and Roy, Nicholas , title =. Proc. IEEE International Conference on Robotics and Automation (ICRA) , pages =. 2012 , doi =

2012