Pith. sign in

REVIEW 2 major objections 3 minor 28 references

Overlapping Schwarz Preconditioners for Pose-Graph SLAM in Robotics

T0 review · 2 major / 3 minor · reviewed 2026-07-15 · grok-4.5

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

Pith's one-line read Overlapping Schwarz preconditioning keeps CG iterations bounded as pose-graph SLAM problems grow, on a controlled square-loop test family.

desk verdict Solid first numerical evidence that one-level overlapping Schwarz keeps CG iterations bounded on a regular pose-graph family, but the abstract promises two-level GDSW that the body never builds or tests. read the letter →

arxiv 2603.08975 v2 pith:45X346FN submitted 2026-03-09 math.NA cs.NA

Stephan Köhler , Oliver Rheinbach , Yue Xiang Tee , Sebastian Zug This is my paper · ORCID
classification math.NAcs.NA MSC 65F0865N5565H1068T40
keywords pose-graphSLAMoverlappingSchwarzdomaindecompositionGauss–Newtonconjugategradientsnumericalscalabilitynonlinearleastsquaresrobotics
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

Graph-based SLAM back-ends produce large sparse linear systems from Gauss–Newton linearization of a nonlinear least-squares pose-graph problem. Simple preconditioners (block-Jacobi, incomplete factorizations, subgraph methods) typically lose effectiveness as the trajectory and number of loop closures grow, so iteration counts climb and long-term autonomy becomes harder. This paper shows that an additive overlapping Schwarz domain-decomposition preconditioner, applied to a synthetic family of square-loop pose graphs, keeps the number of conjugate-gradient iterations small and essentially independent of problem size—at most 16 iterations while the unpreconditioned method exceeds 10 000. A deliberately elementary one-dimensional SLAM problem is rewritten as a chain of linear elastic bars, recovering the classical tridiagonal discrete Laplacian and thereby motivating the transfer of continuum-mechanics preconditioners to robotics. The result is a concrete demonstration that PDE-style domain decomposition can supply the numerical scalability that large-scale SLAM has been missing.

What carries the argument

Additive one-level overlapping Schwarz preconditioner M_AS^{-1} = sum R_i^T A_i^{-1} R_i, where each A_i is the Gauss–Newton Hessian restricted to a contiguous segment of the pose graph with minimal overlap that contains the long-range loop-closure couplings.

What would settle it

Replace the synthetic square loops by a standard irregular outdoor or indoor SLAM benchmark (for example KITTI or EuRoC), keep the same one-level minimal-overlap Schwarz construction, and check whether the CG iteration count still stays essentially constant as the trajectory length and number of loop closures increase.

Watch

Extended reading notes

Core claim

On a synthetic family of two-dimensional pose-graph SLAM problems in which a robot repeatedly traverses a unit square, the additive overlapping Schwarz preconditioner (minimal overlap, one subdomain per loop, loop-closure edges placed in the overlap) yields a number of conjugate-gradient iterations that remains bounded by 16 independently of both the number of subdomains and the subdomain size, while the unpreconditioned iteration count grows without bound past 10 000.

Load-bearing premise

That the bounded iteration counts seen on a perfectly regular square trajectory, with every loop defining a subdomain and every loop-closure edge deliberately sitting inside the minimal overlap, will continue to hold for irregular real-world pose graphs.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 3 minor

Summary. The paper applies additive overlapping Schwarz domain-decomposition preconditioners to the sparse positive-definite linear systems that arise from Gauss–Newton linearization of pose-graph SLAM. After reviewing the nonlinear least-squares formulation (residuals (2.3), objective (2.4), Gauss–Newton Hessian (2.8)) and the classical one-level additive Schwarz operator (3.2), it presents a one-dimensional toy problem that is exactly equivalent to a chain of linear elastic bars (Section 4). Numerical experiments on a synthetic family of square-loop trajectories (each loop defining one subdomain, loop-closure edges forced into the minimal overlap) show that the number of preconditioned CG iterations stays bounded by 16 while unpreconditioned CG grows beyond 10 000 and that the estimated condition numbers of the preconditioned operators remain O(1) (Tables 1–2, Fig. 6).

Significance. Scalable linear solvers for large pose-graph SLAM remain an acknowledged open challenge. Demonstrating that a classical overlapping Schwarz preconditioner can keep CG iteration counts independent of problem size on a controlled family of graphs is therefore of genuine interest to both the domain-decomposition and robotics communities. The exact equivalence of the one-dimensional SLAM model to a finite-element bar chain supplies a clean conceptual bridge that motivates the transfer of PDE-based preconditioners. The numerical evidence in Tables 1–2 is cleanly presented and reproducible in principle. These strengths survive even after the abstract’s over-statement of the method is corrected.

major comments (2)
  1. [Abstract / §3 / Tables 1–2] The abstract and title assert that “scalable two-level overlapping Schwarz au methods with energy-minimizing coarse spaces of GDSW type” are investigated and that “the two-level au preconditioner” keeps CG iterations bounded while overcoming “the typical limitations of au one-level Schwarz approaches.” The body, however, constructs and tests only the classical one-level operator (3.2); no coarse-space basis, no GDSW energy-minimizing extension, and no two-level operator ever appear. Tables 1–2 and Fig. 6 report results exclusively for that one-level method. The abstract claim is therefore unsupported by any experiment in the manuscript and must be rewritten to match the actual content.
  2. [§5.2 / §6] The observed numerical scalability of the one-level method relies on two deliberate modelling choices stated in §5.2: (i) each square loop defines exactly one subdomain, and (ii) every long-range loop-closure edge is placed inside the minimal-overlap region. The paper itself notes (§5.2, §6) that one-level Schwarz is generally not scalable for elliptic problems and that the bounded iteration counts may be geometry-specific. Without additional experiments that move loop closures out of the overlap or that use irregular real-world graphs, the claim that the method “overcomes the typical limitations of one-level Schwarz” cannot be sustained even for the one-level operator that was actually implemented.
minor comments (3)
  1. [throughout] Several grammatical slips remain (e.g., “our the Schwarz preconditioner”, “formulately algebraically”, “the numberofloopsisidenticaltothenumberof”). A careful proof-reading pass is needed.
  2. [§5.1 / Fig. 4] Figure 4’s caption and the surrounding text refer to “the system matrix H”, yet the Gauss–Newton matrix is denoted H_GN earlier; consistent notation would help.
  3. [§5.1] The absolute Gauss–Newton tolerance of 1e-6 is size-dependent; a brief remark on whether a relative residual or a gradient-norm criterion would alter the outer iteration counts would be useful.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: bounded CG iterations are direct empirical measurements on independently generated systems; the 1-D elastic-bar equivalence is an exact identity by construction of the toy residual, not a circular prediction.

full rationale

The paper's load-bearing claim is the observed numerical scalability of one-level additive overlapping Schwarz (Eq. 3.2) on a synthetic square-loop pose-graph family: CG iterations remain ≤16 while unpreconditioned CG grows past 10 000 (Tables 1–2, Fig. 6). These counts are obtained by assembling the Gauss–Newton matrix H_GN = J^T W J from noisy odometry + loop-closure residuals, applying the Boolean restriction operators R_i that extract contiguous trajectory segments, and running CG; nothing is fitted to the iteration counts themselves. The 1-D elastic-bar analogy (Section 4) equates the SLAM objective (4.1) to the total elastic energy of bars with rest length 0.9 by writing identical quadratic residuals; the resulting tridiagonal stencil is therefore identical by definition of the residual, which is an intentional pedagogical equivalence rather than a self-definitional prediction of a new quantity. Self-citations to the authors' prior domain-decomposition papers appear only as background on classical Schwarz theory and parallel implementations; they are not invoked to force uniqueness of the SLAM preconditioner or to supply any numerical result reported here. The abstract/body mismatch on two-level GDSW versus one-level experiments is an overclaim of scope, not a circular reduction of a derivation to its inputs. Consequently the central empirical claim stands independently of any fitted parameter or self-referential definition.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central numerical claim rests on standard positive-definiteness of the gauge-fixed Gauss–Newton Hessian, on the classical definition of the one-level additive Schwarz operator, and on a handful of modeling choices that define the synthetic test family (square trajectory, loop-per-subdomain partition, loop closures forced into the overlap, fixed weight matrices). No new physical entities or free parameters are fitted to data; the free parameters are purely experimental design choices.

free parameters (5)
  • odometry weight matrix Wij = 20 I_3
    Fixed by hand to 20 I_3 for sequential edges; not fitted but chosen ad hoc for the experiments.
  • loop-closure weight matrix Wij = 100 I_3
    Fixed by hand to 100 I_3; higher weight chosen to emphasize loop closures.
  • Gauss-Newton absolute tolerance = 1e-6
    Absolute Euclidean-norm stopping criterion 1e-6; absolute criteria can introduce mild size dependence.
  • CG relative tolerance = 1e-8
    Relative residual tolerance 1e-8 used for all linear solves.
  • overlap size = 3 dofs
    Minimal overlap of three degrees of freedom (one pose) chosen throughout; not varied.
assumptions (4)
  • domain assumption After fixing one pose the Gauss-Newton Hessian H = J^T W J is symmetric positive definite, so CG is applicable.
    Stated in §2.2; standard for gauge-fixed pose-graph SLAM with SPD weights.
  • standard math The one-level additive overlapping Schwarz operator M_AS^{-1} = sum R_i^T A_i^{-1} R_i is a valid SPD preconditioner for the global SPD matrix A.
    Classical result recalled in §3; used without proof.
  • ad hoc to paper Loop-closure edges may be placed inside the overlap region of the subdomain partition.
    Explicit design choice in §5.2 that may contribute to the observed scalability; flagged as a limitation in §6.
  • ad hoc to paper Each loop of the square trajectory defines one overlapping Schwarz subdomain.
    Partition rule used for all experiments (§5.2); not a general graph-partitioning strategy.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Overlapping Schwarz Preconditioners for Pose-Graph SLAM in Robotics." pith.science (2026). https://pith.science/paper/45X346FN

@misc{pith2026260308975,
  author       = {Pith},
  title        = {Pith review of: Overlapping Schwarz Preconditioners for Pose-Graph SLAM in Robotics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/45X346FN}},
  note         = {Machine review of arXiv:2603.08975}
}
read the original abstract

We investigate scalable two-level overlapping Schwarz domain decomposition methods with energy-minimizing coarse spaces of GDSW type (Generalized Dryja--Smith--Widlund type) as preconditioners for the sparse linear systems arising in graph-based nonlinear least-squares problems, specifically the pose-graph optimization back-end in Simultaneous Localization and Mapping (SLAM). After a brief introduction to SLAM and domain decomposition preconditioners, we describe the nonlinear least-squares formulation, its linearization, and the resulting matrix structure, to facilitate access for readers without prior knowledge of either field. Numerical experiments demonstrate the numerical scalability of the preconditioned conjugate gradient method (CG): Using the two-level overlapping Schwarz preconditioner, the number of CG iterations remains bounded independently of the problem size, overcoming the typical limitations of simple preconditioners, including one-level Schwarz approaches. We further show that a simplified SLAM problem can be interpreted as a finite element problem using linear elastic bars, reinforcing the analogy to continuum mechanics and motivating the use of scalable domain decomposition techniques.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 11 canonical work pages

  1. [1]

    Subgraph-preconditioned conjugate gradients for large scale SLAM

    Frank Dellaert, Justin Carlson, Viorela Ila, Kai Ni, and Charles E Thorpe. Subgraph-preconditioned conjugate gradients for large scale SLAM. In2010 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 2566–

  2. [2]

    IEEE, 2010.http://doi.org/10.1109/IROS.2010.5650422

    work page doi:10.1109/iros.2010.5650422 2010

  3. [3]

    Incremental subgraph preconditioned conjugate gra- dient method for graph slam

    Jian He and Frank Dellaert. Incremental subgraph preconditioned conjugate gra- dient method for graph slam. InProceedings of the IEEE/RSJ International Con- ference on Intelligent Robots and Systems (IROS), pages 3934–3941. IEEE, 2014. http://doi.org/10.1109/IROS.2014.6942924

    work page doi:10.1109/iros.2014.6942924 2014

  4. [4]

    Seitz, and Richard Szeliski

    Sameer Agarwal, Noah Snavely, Steven M. Seitz, and Richard Szeliski. Bundle adjustment in the large. InProceedings of the European Conference on Com- puter Vision (ECCV), pages 29–42. Springer, 2010.http://doi.org/10.1007/ 978-3-642-15552-9_3

    2010

  5. [5]

    Efficient sparse pose adjustment for 2d mapping

    Kurt Konolige, Giorgio Grisetti, Rainer Kümmerle, Wolfram Burgard, Benson Limketkai, and Regis Vincent. Efficient sparse pose adjustment for 2d mapping. In2010 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 22–29, 2010.http://doi.org/10.1109/IROS.2010.5649043

    doi:10.1109/iros.2010.5649043 2010

  6. [6]

    g2o: A general framework for graph optimization

    Rainer Kümmerle, Giorgio Grisetti, Hauke Strasdat, Kurt Konolige, and Wolfram Burgard. g2o: A general framework for graph optimization. InProceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 3607–

  7. [7]

    IEEE, 2011.http://doi.org/10.1109/ICRA.2011.5979949

    doi:10.1109/icra.2011.5979949 2011

  8. [8]

    Cesar Cadena, Luca Carlone, Henry Carrillo, Yasir Latif, Davide Scaramuzza, José Neira, Ian Reid, and John J. Leonard. Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age.IEEE Transactions on Robotics, 32(6):1309–1332, 2016.http://doi.org/10.1109/TRO.2016.2624754

    doi:10.1109/tro.2016.2624754 2016

Show all 28 references
  1. [9]

    Simultaneous localization and mapping: Part i.IEEE Robotics & Automation Magazine, 13(2):99–110, 2006.http://doi.org/ 10.1109/MRA.2006.1638022

    Hugh Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: Part i.IEEE Robotics & Automation Magazine, 13(2):99–110, 2006.http://doi.org/ 10.1109/MRA.2006.1638022

    2006 doi arXiv

  2. [10]

    Simultaneous localization and mapping: Part ii.IEEE Robotics & Automation Magazine, 13(3):108–117, 2006.http://doi.org/ 10.1109/MRA.2006.1678144

    Hugh Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: Part ii.IEEE Robotics & Automation Magazine, 13(3):108–117, 2006.http://doi.org/ 10.1109/MRA.2006.1678144

    2006 doi arXiv

  3. [11]

    Localization of autonomous guided vehicles

    Hugh Durrant-Whyte, David Rye, and Eduardo Nebot. Localization of autonomous guided vehicles. In Georges Giralt and Gerhard Hirzinger, editors,Robotics Re- search, pages 613–625, London, 1996. Springer London.http://doi.org/doi.org/ 10.1007/978-1-4471-0765-1_69

    1996 doi

  4. [12]

    The SLAM problem: asurvey

    Josep Aulinas, Yvan Petillot, Joaquim Salvi, and Xavier Lladó. The SLAM problem: asurvey. InArtificial Intelligence Research and Development: Proceedings of the 11th International Conference of the Catalan Association for Artificial Intelligence, pages 363–371. IOS Press, 2008...

    2008 doi

  5. [13]

    The simultaneous localization and mapping (SLAM): An overview.Journal of Applied Science and Technology Trends, 2(2):147– 158, 2021.http://doi.org/10.38094/jastt204117

    Bashar Alsadik and Samer Karam. The simultaneous localization and mapping (SLAM): An overview.Journal of Applied Science and Technology Trends, 2(2):147– 158, 2021.http://doi.org/10.38094/jastt204117

    2021 doi

  6. [14]

    SLAM overview: from single sensor to heterogeneous fusion.Remote Sensing, 14(23):6033, 2022.http://doi.org/10.3390/rs14236033

    Weifeng Chen, Chengjun Zhou, Guangtao Shang, Xiyang Wang, Zhenxiong Li, Chonghui Xu, and Kai Hu. SLAM overview: from single sensor to heterogeneous fusion.Remote Sensing, 14(23):6033, 2022.http://doi.org/10.3390/rs14236033

    2022 doi

  7. [15]

    A comprehensive survey of visual SLAM algorithms.Robotics, 11(1):24, 2022.http://doi.org/10.3390/robotics11010024

    Andréa Macario Barros, Maugan Michel, Yoann Moline, Gwenolé Corre, and Frédérick Carrel. A comprehensive survey of visual SLAM algorithms.Robotics, 11(1):24, 2022.http://doi.org/10.3390/robotics11010024

    2022 doi

  8. [16]

    Review on LiDAR-based SLAM techniques

    Leyao Huang. Review on LiDAR-based SLAM techniques. In2021 International conference on signal processing and machine learning (CONF-SPML), pages 163–

  9. [17]

    IEEE, 2021.http://doi.org/10.1109/CONF-SPML54095.2021.00040

    2021 doi

  10. [18]

    A tutorial on graph-based SLAM.IEEE Intelligent Transportation Systems Magazine, 2(4):31–43, 2011.http://doi.org/10.1109/MITS.2010.939925

    Giorgio Grisetti, Rainer Kümmerle, Cyrill Stachniss, and Wolfram Burgard. A tutorial on graph-based SLAM.IEEE Intelligent Transportation Systems Magazine, 2(4):31–43, 2011.http://doi.org/10.1109/MITS.2010.939925. Overlapping Schwarz in Robotics 21

    2011 doi

  11. [19]

    Springer- Verlag, Berlin, 2005.http://doi.org/10.1007/b137868

    Andrea Toselli and Olof Widlund.Domain decomposition methods—algorithms and theory, volume 34 ofSpringer Series in Computational Mathematics. Springer- Verlag, Berlin, 2005.http://doi.org/10.1007/b137868

    2005 doi

  12. [20]

    Smith, Petter E

    Barry F. Smith, Petter E. Bjørstad, and William D. Gropp.Domain decomposition. Parallel multilevel methods for elliptic partial differential equations. Cambridge: Cambridge University Press, 1996. ISBN 9780521602860

    1996

  13. [21]

    Domain decomposition methods in com- putational fluid dynamics.International journal for numerical methods in fluids, 14(2):147–165, 1992.http://doi.org/10.1002/fld.1650140203

    William D Gropp and David E Keyes. Domain decomposition methods in com- putational fluid dynamics.International journal for numerical methods in fluids, 14(2):147–165, 1992.http://doi.org/10.1002/fld.1650140203

    1992 doi

  14. [22]

    Non-overlapping domain decomposition meth- ods in structural mechanics.Archives of computational methods in engineering, 13(4):515, 2006.http://doi.org/10.1007/BF02905857

    Pierre Gosselet and Christian Rey. Non-overlapping domain decomposition meth- ods in structural mechanics.Archives of computational methods in engineering, 13(4):515, 2006.http://doi.org/10.1007/BF02905857

    2006 doi

  15. [23]

    Paralleliterativesubstructuringinstructuralmechanics.Archives of Computational Methods in Engineering, 16(4):425–463, 2009.http://doi.org/ 10.1007/s11831-009-9035-4

    OliverRheinbach. Paralleliterativesubstructuringinstructuralmechanics.Archives of Computational Methods in Engineering, 16(4):425–463, 2009.http://doi.org/ 10.1007/s11831-009-9035-4

    2009 doi

  16. [24]

    AlbertoCorigliano,MartinoDossi,andStefanoMariani. Domaindecompositionand model order reduction methods applied to the simulation of multi-physics problems in MEMS.Computers & Structures, 122:113–127, 2013.http://doi.org/10.1016/ j.compstruc.2012.12.012

    2013

  17. [25]

    Mono- lithic parallel overlapping Schwarz methods in fully-coupled nonlinear chemo- mechanics problems.Comput

    Bjoern Kiefer, Stefan Prüger, Oliver Rheinbach, and Friederike Röver. Mono- lithic parallel overlapping Schwarz methods in fully-coupled nonlinear chemo- mechanics problems.Comput. Mech., 71(4):765–788, 2023.http://doi.org/10. 1007/s00466-022-02254-y

    2023

  18. [26]

    Com- putational homogenization with million-way parallelism using domain decompo- sition methods.Comput

    Axel Klawonn, Stephan Köhler, Martin Lanser, and Oliver Rheinbach. Com- putational homogenization with million-way parallelism using domain decompo- sition methods.Comput. Mech., 65(1):1–22, 2020.http://doi.org/10.1007/ s00466-019-01749-5

    2020

  19. [27]

    Parallel scalability of three-level FROSch preconditioners to 220000 cores using the Theta supercomputer

    Alexander Heinlein, Oliver Rheinbach, and Friederike Röver. Parallel scalability of three-level FROSch preconditioners to 220000 cores using the Theta supercomputer. SIAM Journal on Scientific Computing, 45(3):s173–s198, 2023.http://doi.org/ 10.1137/21M1431205

    2023 doi

  20. [28]

    On the scalability of classical one-level domain-decomposition methods.Viet- nam Journal of Mathematics, 46(4):1053–1088, 2018.http://doi.org/10.1007/ s10013-018-0316-9

    Faycal Chaouqui, Gabriele Ciaramella, Martin J Gander, and Tommaso Vanzan. On the scalability of classical one-level domain-decomposition methods.Viet- nam Journal of Mathematics, 46(4):1053–1088, 2018.http://doi.org/10.1007/ s10013-018-0316-9

    2018

Pith tools

Reviewed July 15, 2026 · model on record in the stance chip above.