pith. sign in

arxiv: 2606.31738 · v2 · pith:HVMKOVGAnew · submitted 2026-06-30 · 🧮 math.OC · cs.NA· math.NA

A Geometry-Adaptive Regularized Newton-Type Method for Manifold-Affine Intersection Problems

Dengyu Zheng , Shixiang Chen This is my paper

Pith reviewed 2026-07-03 22:04 UTC · model grok-4.3

classification 🧮 math.OC cs.NAmath.NA
keywords regularized Newton methodmanifold-affine intersectionstructured low-rank approximationtransversalitylocal convergencequadratic convergence
0
0 comments X

The pith

Regularized Newton method converges linearly to manifold-affine intersections under intrinsic transversality.

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

The paper introduces RN-SLRA as a regularized Newton-type method for local manifold-affine intersection problems that arise in structured low-rank approximation. It replaces the exact tangent-space intersection step of classical Newton-SLRA with a regularized quadratic subproblem over the affine space to avoid singularity or ill-conditioning when transversality fails. Under intrinsic transversality the method is shown to converge locally linearly to the intersection. Under full transversality a residual-dependent choice of the regularization parameter produces higher-order local convergence, including quadratic convergence for the linear residual rule.

Core claim

RN-SLRA solves a regularized quadratic subproblem over the affine space at each step. Under intrinsic transversality this yields local linear convergence to the intersection. Under transversality a residual-dependent regularization parameter produces higher-order convergence, with quadratic convergence for the linear residual rule. An inexact variant based on quasioptimal manifold projections retains local residual convergence when the quasioptimality constant is accurate enough.

What carries the argument

Regularized quadratic subproblem over the affine space, with residual-dependent regularization parameter that replaces the exact tangent-space intersection step.

If this is right

  • Local linear convergence holds under intrinsic transversality at the limit point.
  • Quadratic convergence holds under transversality when the linear residual rule selects the regularization parameter.
  • The inexact variant preserves local residual convergence provided the quasioptimality constant meets a sufficient accuracy threshold.

Where Pith is reading between the lines

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

  • The same regularization device could stabilize other Newton-type methods on manifolds that encounter ill-conditioned tangent intersections.
  • The inexact projection variant may enable scaling to larger structured low-rank problems by trading projection accuracy for speed.

Load-bearing premise

The manifold must be at least C^2 smooth and an intrinsic transversality condition must hold at the limit point.

What would settle it

An example manifold-affine pair satisfying intrinsic transversality where the iterates of RN-SLRA fail to converge linearly, or where the linear residual rule fails to produce quadratic convergence despite full transversality.

read the original abstract

We propose Regularized Newton-SLRA (RN-SLRA), a regularized Newton-type method for local manifold--affine intersection problems motivated by structured low-rank approximation. Classical Newton-SLRA achieves fast local convergence under transversality, but its tangent-space intersection step may become ill-defined, singular, or severely ill-conditioned when transversality fails. RN-SLRA overcomes this difficulty by replacing the exact tangent-space intersection step with a regularized quadratic subproblem over the affine space. Under intrinsic transversality, we prove local linear convergence to the intersection. Under transversality, we show that a residual-dependent choice of the regularization parameter yields higher-order local convergence; in particular, the method converges quadratically for the linear residual rule. We also analyze an inexact variant based on quasioptimal manifold projections. When the quasioptimality constant is sufficiently accurate, the inexact method retains local residual convergence. Numerical experiments on constructed degenerate SLRA instances and Hankel-structured examples illustrate the robustness of RN-SLRA in settings where Newton-SLRA may fail, and show that the inexact variant can reduce the projection cost in large-scale problems.

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.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes RN-SLRA, a regularized Newton-type method for local manifold-affine intersection problems motivated by structured low-rank approximation. It replaces the tangent-space intersection step of classical Newton-SLRA with a regularized quadratic subproblem to handle ill-conditioning when transversality fails. Under intrinsic transversality the method is shown to converge locally linearly to the intersection; under transversality a residual-dependent regularization parameter is proved to yield higher-order local convergence, including quadratic convergence for the linear residual rule. An inexact variant based on quasi-optimal manifold projections is analyzed, and numerical experiments on constructed degenerate SLRA instances and Hankel-structured examples are presented.

Significance. If the stated convergence results hold, the geometry-adaptive regularization supplies a practical and theoretically grounded way to extend Newton-type methods to degenerate manifold-affine problems. The explicit treatment of both intrinsic transversality (linear convergence) and full transversality (higher-order rates via residual-dependent regularization) together with the inexact variant analysis constitute a coherent contribution to manifold optimization. The numerical tests on deliberately constructed degenerate instances provide concrete evidence of robustness where standard Newton-SLRA may fail.

major comments (3)
  1. [§4 (convergence analysis)] The central convergence claims rest on the C^2 smoothness assumption and the intrinsic transversality condition at the limit point; the manuscript should state explicitly in the main theorem (presumably §4 or §5) whether these conditions are also necessary or only sufficient, and whether any quantitative modulus of transversality appears in the linear rate.
  2. [§5 (higher-order convergence)] The quadratic convergence result for the linear residual rule is asserted in the abstract and presumably proved in §5; the proof sketch should clarify whether the regularization parameter is chosen exactly as λ_k = c·r_k or whether an additional safeguard is required to guarantee that the quadratic term dominates the linear term in the error recursion.
  3. [Numerical experiments section] Table 1 (or the corresponding numerical table) reports success on degenerate instances but does not list the condition numbers of the tangent-space intersection matrices or the observed iteration counts for Newton-SLRA on the same problems; without these quantities the claim that RN-SLRA succeeds where Newton-SLRA fails remains qualitative.
minor comments (2)
  1. [§6 (inexact variant)] The notation for the quasi-optimality constant in the inexact variant should be introduced once and used consistently; its dependence on the manifold curvature is mentioned only in passing.
  2. A short paragraph comparing the computational cost of the regularized subproblem solve versus the exact tangent-space projection would help readers assess the practical overhead of RN-SLRA.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the positive evaluation and the constructive comments, which will help clarify the presentation of the convergence results and numerical comparisons. We address each major comment below.

read point-by-point responses

  1. Referee: [§4 (convergence analysis)] The central convergence claims rest on the C^2 smoothness assumption and the intrinsic transversality condition at the limit point; the manuscript should state explicitly in the main theorem (presumably §4 or §5) whether these conditions are also necessary or only sufficient, and whether any quantitative modulus of transversality appears in the linear rate.

    Authors: The C^2 smoothness and intrinsic transversality assumptions in Theorem 4.1 are sufficient for local linear convergence; the manuscript does not claim necessity. The linear rate is quantitative and depends explicitly on the modulus of intrinsic transversality (appearing in the contraction factor derived in the proof). We will add a remark immediately after Theorem 4.1 stating that the conditions are sufficient and noting the explicit dependence of the rate on the transversality modulus. revision: yes

  2. Referee: [§5 (higher-order convergence)] The quadratic convergence result for the linear residual rule is asserted in the abstract and presumably proved in §5; the proof sketch should clarify whether the regularization parameter is chosen exactly as λ_k = c·r_k or whether an additional safeguard is required to guarantee that the quadratic term dominates the linear term in the error recursion.

    Authors: The proof of quadratic convergence under the linear residual rule (Theorem 5.2) uses exactly λ_k = c r_k for a sufficiently small fixed c > 0. The error recursion analysis shows that this choice alone ensures the quadratic term dominates for all large k; no additional safeguard is required beyond the smallness of c. We will expand the proof presentation in §5 to state this choice and the role of c explicitly. revision: yes

  3. Referee: [Numerical experiments section] Table 1 (or the corresponding numerical table) reports success on degenerate instances but does not list the condition numbers of the tangent-space intersection matrices or the observed iteration counts for Newton-SLRA on the same problems; without these quantities the claim that RN-SLRA succeeds where Newton-SLRA fails remains qualitative.

    Authors: We agree that reporting the condition numbers of the tangent-space intersection matrices together with the iteration counts (or failure indicators) for Newton-SLRA on the same degenerate instances would make the comparison quantitative rather than qualitative. In the revised manuscript we will augment the numerical experiments section and the corresponding table with these quantities computed on the identical test problems. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The derivation chain consists of a proposed regularized Newton method (RN-SLRA) whose local convergence rates are proved from first principles under stated C^2 smoothness and transversality (or intrinsic transversality) assumptions at the limit point. The residual-dependent regularization rule is introduced as an independent algorithmic device, and the quadratic convergence claim for the linear residual rule follows from the analysis rather than from any self-definition, fitted input renamed as prediction, or load-bearing self-citation. No equations reduce the claimed rates to quantities defined by the method's own outputs, and the supporting analysis is presented as self-contained against the external benchmarks of the transversality conditions.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

Review performed on abstract alone; full paper would be needed to enumerate all background assumptions on manifold smoothness and transversality definitions.

free parameters (1)
  • regularization parameter
    Chosen via residual-dependent rules whose exact functional form and scaling constants are part of the method but not numerically fitted to data in the abstract description.
axioms (2)
  • domain assumption The manifold is at least twice continuously differentiable near the solution point.
    Required for the Newton step and local convergence analysis to be well-defined.
  • domain assumption Intrinsic transversality or transversality holds at the limit point.
    Central hypothesis under which the linear and higher-order convergence statements are proved.

pith-pipeline@v0.9.1-grok · 5735 in / 1395 out tokens · 26615 ms · 2026-07-03T22:04:39.052949+00:00 · methodology

Review history (2 revisions) →

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

84 extracted references · 84 canonical work pages

  1. [1]

    2008 , issn =

    Structured low-rank approximation and its applications , journal =. 2008 , issn =

    work page 2008

  2. [2]

    Drusvyatskiy and A

    D. Drusvyatskiy and A. S. Lewis , title =. Vietnam J. Math. , year =

    work page

  3. [3]

    2016 , pages =

    A Quadratically Convergent Algorithm for Structured Low-Rank Approximation , journal =. 2016 , pages =

    work page 2016

  4. [4]

    Ioffe and Adrian S

    Dmitriy Drusvyatskiy and Alexander D. Ioffe and Adrian S. Lewis , title =. Foundations of Computational Mathematics , year =

    work page

  5. [5]

    Foundations of Computational Mathematics , year =

    Dominikus Noll and Aude Rondepierre , title =. Foundations of Computational Mathematics , year =

    work page

  6. [6]

    Kruger and D

    Alexander Y. Kruger and D. Russell Luke and Nguyen H. Thao , title =. Set-Valued and Variational Analysis , year =

    work page

  7. [7]

    Kruger , title =

    Alexander Y. Kruger , title =. Set-Valued and Variational Analysis , year =

    work page

  8. [8]

    Kruger and D

    Alexander Y. Kruger and D. Russell Luke and Nguyen H. Thao , title =. Mathematical Programming , year =

    work page

  9. [9]

    A Family of Inexact SQA Methods for Non-Smooth Convex Minimization with Provable Convergence Guarantees Based on the Luo-Tseng Error Bound Property , author=. Math. Program , year=

    work page

  10. [10]

    2025 , eprint=

    A Quadratically Convergent Alternating Projection Method for Nonconvex Sets , author=. 2025 , eprint=

    work page 2025

  11. [11]

    2001 , isbn =

    Frank Deutsch , title =. 2001 , isbn =

    work page 2001

  12. [12]

    Lewis and J

    Adrian S. Lewis and J. Alternating Projections on Manifolds , journal =

    work page

  13. [13]

    and Luke, D

    Lewis, Adrian S. and Luke, D. Russell and Malick, J. Local Linear Convergence for Alternating and Averaged Nonconvex Projections , journal =

    work page

  14. [14]

    and Luke, D

    Bauschke, Heinz H. and Luke, D. Russell and Phan, Hung M. and Wang, Xianfu , title =. Set-Valued and Variational Analysis , volume =

    work page

  15. [15]

    Russell , title =

    Hesse, Robert and Luke, D. Russell , title =. SIAM Journal on Optimization , volume =

    work page

  16. [16]

    Drusvyatskiy, Dmitriy , title =

    work page

  17. [17]

    and Thao, Nguyen H

    Kruger, Alexander Y. and Thao, Nguyen H. , title =. Journal of Convex Analysis , volume =

    work page

  18. [18]

    Russell and Teboulle, Marc and Thao, Nguyen H

    Luke, D. Russell and Teboulle, Marc and Thao, Nguyen H. , title =. Mathematical Programming , volume =

    work page

  19. [19]

    and Cuong, Nguyen Duy and Verhaegen, Michel , title =

    Thao, Nguyen Hieu and Bui, Hoa T. and Cuong, Nguyen Duy and Verhaegen, Michel , title =. Set-Valued and Variational Analysis , volume =

    work page

  20. [20]

    and Borwein, Jonathan M

    Bauschke, Heinz H. and Borwein, Jonathan M. , title =. Set-Valued Analysis , volume =

    work page

  21. [21]

    , title =

    Dolecki, S. , title =. Annali di Matematica Pura ed Applicata , volume =

    work page

  22. [22]

    , title =

    Ioffe, Alexander D. , title =. Journal of the Australian Mathematical Society , volume =

    work page

  23. [23]

    , title =

    Ioffe, Alexander D. , title =. Journal of Optimization Theory and Applications , volume =

    work page

  24. [24]

    , title =

    Kruger, Alexander Y. , title =. Set-Valued Analysis , volume =

    work page

  25. [25]

    Schwarz, H. A. , title =. Journal f

    work page

  26. [26]

    von Neumann, John , title =

    work page

  27. [27]

    , journal=

    Cadzow, J.A. , journal=. Signal enhancement-a composite property mapping algorithm , year=

    work page

  28. [28]

    Statistics and Its Interface , volume =

    Gillard, Jonathan , title =. Statistics and Its Interface , volume =

    work page

  29. [29]

    and Funderlic, Robert E

    Chu, Moody T. and Funderlic, Robert E. and Plemmons, Robert J. , title =. Linear Algebra and its Applications , volume =

    work page

  30. [30]

    Rosen, J. B. and Park, H. and Glick, J. , title =. SIAM Journal on Matrix Analysis and Applications , volume =

    work page

  31. [31]

    Park, Haesun and Zhang, Lei and Rosen, J. B. , title =. BIT Numerical Mathematics , volume =

    work page

  32. [32]

    and Amodei, Luca and Meyer, Gilles , title =

    Absil, P.-A. and Amodei, Luca and Meyer, Gilles , title =. Computational Statistics , volume =

    work page

  33. [33]

    SIAM Journal on Optimization , volume =

    Vandereycken, Bart , title =. SIAM Journal on Optimization , volume =

    work page

  34. [34]

    Exact Matrix Completion via Convex Optimization , journal =

    Cand. Exact Matrix Completion via Convex Optimization , journal =

    work page

  35. [35]

    The Power of Convex Relaxation: Near-Optimal Matrix Completion , journal =

    Cand. The Power of Convex Relaxation: Near-Optimal Matrix Completion , journal =

    work page

  36. [36]

    Journal of Machine Learning Research , volume =

    Recht, Benjamin , title =. Journal of Machine Learning Research , volume =

    work page

  37. [37]

    Matrix Completion with Noise , journal =

    Cand. Matrix Completion with Noise , journal =

    work page

  38. [38]

    Proceedings of the Forty-Fifth Annual ACM Symposium on Theory of Computing , pages =

    Jain, Prateek and Netrapalli, Praneeth and Sanghavi, Sujay , title =. Proceedings of the Forty-Fifth Annual ACM Symposium on Theory of Computing , pages =

    work page

  39. [39]

    SIAM Journal on Matrix Analysis and Applications , volume =

    Ottaviani, Giorgio and Spaenlehauer, Pierre-Jean and Sturmfels, Bernd , title =. SIAM Journal on Matrix Analysis and Applications , volume =

    work page

  40. [40]

    Symbolic-Numeric Computation , series =

    Kaltofen, Erich and Yang, Zhengfeng and Zhi, Lihong , title =. Symbolic-Numeric Computation , series =

    work page

  41. [41]

    Quarterly of Applied Mathematics , volume =

    Levenberg, Kenneth , title =. Quarterly of Applied Mathematics , volume =

    work page

  42. [42]

    , title =

    Marquardt, Donald W. , title =. Journal of the Society for Industrial and Applied Mathematics , volume =

    work page

  43. [43]

    The Levenberg--Marquardt Algorithm: Implementation and Theory , booktitle =

    Mor. The Levenberg--Marquardt Algorithm: Implementation and Theory , booktitle =

    work page

  44. [44]

    , title =

    Nesterov, Yurii and Polyak, Boris T. , title =. Mathematical Programming , volume =

    work page

  45. [45]

    Inverse Problems , volume =

    Hanke, Martin , title =. Inverse Problems , volume =

    work page

  46. [46]

    Quasi-GCD Computations , journal =

    Sch. Quasi-GCD Computations , journal =

    work page

  47. [47]

    Karmarkar, Nitin and Lakshman, Y. N. , title =. Proceedings of the 1996 International Symposium on Symbolic and Algebraic Computation , pages =

    work page 1996

  48. [48]

    Karmarkar, Nitin and Lakshman, Y. N. , title =. Journal of Symbolic Computation , volume =

    work page

  49. [49]

    , title =

    Pan, Victor Y. , title =. Information and Computation , volume =

    work page

  50. [50]

    and Watt, Stephen M

    Corless, Robert M. and Watt, Stephen M. and Zhi, Lihong , title =. IEEE Transactions on Signal Processing , volume =

    work page

  51. [51]

    Proceedings of the 2004 International Symposium on Symbolic and Algebraic Computation , pages =

    Zeng, Zhonggang and Dayton, Bruce , title =. Proceedings of the 2004 International Symposium on Symbolic and Algebraic Computation , pages =

    work page 2004

  52. [52]

    Proceedings of the 2006 International Symposium on Symbolic and Algebraic Computation , pages =

    Kaltofen, Erich and Yang, Zhengfeng and Zhi, Lihong , title =. Proceedings of the 2006 International Symposium on Symbolic and Algebraic Computation , pages =

    work page 2006

  53. [53]

    Symbolic-Numeric Computation , pages =

    Kaltofen, Erich and Yang, Zhengfeng and Zhi, Lihong , title =. Symbolic-Numeric Computation , pages =

    work page

  54. [54]

    and Boito, Paola , title =

    Bini, Dario A. and Boito, Paola , title =. Proceedings of the 2007 International Symposium on Symbolic and Algebraic Computation , pages =

    work page 2007

  55. [55]

    and Allan, Stephen , title =

    Winkler, Joab R. and Allan, Stephen , title =. Electronic Transactions on Numerical Analysis , volume =

    work page

  56. [56]

    and Galligo, Andr

    Emiris, Ioannis Z. and Galligo, Andr. Certified Approximate Univariate GCDs , journal =

    work page

  57. [57]

    Computing Nearest GCD with Certification , booktitle =

    Ch. Computing Nearest GCD with Certification , booktitle =

    work page

  58. [58]

    Proceedings of the 2009 International Symposium on Symbolic and Algebraic Computation , pages =

    Terui, Akira , title =. Proceedings of the 2009 International Symposium on Symbolic and Algebraic Computation , pages =

    work page 2009

  59. [59]

    Approximate GCD

    Yakoubsohn, Jean-Claude and Masmoudi, Mouna and Ch. Approximate GCD. Applied Mathematics E-Notes , volume =

    work page

  60. [60]

    Journal of Computational and Applied Mathematics , volume =

    Markovsky, Ivan and Usevich, Konstantin , title =. Journal of Computational and Applied Mathematics , volume =

    work page

  61. [61]

    Ben and Park, Haesun and Glick, John , title =

    Rosen, J. Ben and Park, Haesun and Glick, John , title =. SIAM Journal on Matrix Analysis and Applications , volume =

    work page

  62. [62]

    Linear Algebra and its Applications , volume =

    Fazzi, Antonio and Guglielmi, Nicola and Markovsky, Ivan , title =. Linear Algebra and its Applications , volume =

    work page

  63. [63]

    2012 , series =

    Ivan Markovsky , title =. 2012 , series =

    work page 2012

  64. [64]

    Latent Variable Analysis and Signal Separation , series =

    Ivan Markovsky , title =. Latent Variable Analysis and Signal Separation , series =

    work page

  65. [65]

    Latent Variable Analysis and Signal Separation , series =

    Ivan Markovsky and Pier Luigi Dragotti , title =. Latent Variable Analysis and Signal Separation , series =

    work page

  66. [66]

    Symbolic-Numeric Computation , series =

    Erich Kaltofen and Zhengfeng Yang and Lihong Zhi , title =. Symbolic-Numeric Computation , series =

    work page

  67. [67]

    Computational Optimization and Applications , volume =

    Dong-Hui Li and Masao Fukushima and Liqun Qi and Nobuo Yamashita , title =. Computational Optimization and Applications , volume =

    work page

  68. [68]

    An improved GLMNET for l1-regularized logistic regression , volume =

    Yuan, Guo-Xun and Ho, Chia-Hua and Lin, Chih-Jen , year =. An improved GLMNET for l1-regularized logistic regression , volume =

    work page

  69. [69]

    Absil and J

    P.-A. Absil and J. Projection-Like Retractions on Matrix Manifolds , journal =

    work page

  70. [70]

    R. A. Poliquin and R. T. Rockafellar and L. Thibault , title =. Transactions of the American Mathematical Society , volume =

    work page

  71. [71]

    Florian Feppon and Pierre F. J. Lermusiaux , title =. SIAM Journal on Matrix Analysis and Applications , volume =

    work page

  72. [72]

    BIT Numerical Mathematics , volume =

    Gianluca Ceruti and Christian Lubich and Helmut Walach , title =. BIT Numerical Mathematics , volume =

    work page

  73. [73]

    , title =

    Halko, Nathan and Martinsson, Per-Gunnar and Tropp, Joel A. , title =. SIAM Review , year =

    work page

  74. [74]

    Larsen, Rasmus Munk , title =

    work page

  75. [75]

    IEEE Journal of Selected Topics in Signal Processing , volume =

    An Overview of Low-Rank Matrix Recovery from Incomplete Observations , author =. IEEE Journal of Selected Topics in Signal Processing , volume =

    work page

  76. [76]

    Constructive Approximation , volume =

    Andersson, Fredrik and Carlsson, Marcus , title =. Constructive Approximation , volume =

    work page

  77. [77]

    and Mahony, R

    Absil, P.-A. and Mahony, R. and Sepulchre, R. , title =. 2008 , pages =

    work page 2008

  78. [78]

    2023 , isbn =

    Boumal, Nicolas , title =. 2023 , isbn =

    work page 2023

  79. [79]

    2026 , journal=

    Retractions by Alternating Projections , author=. 2026 , journal=

    work page 2026

  80. [80]

    SIAM Journal on Scientific Computing , volume =

    Baglama, James and Reichel, Lothar , title =. SIAM Journal on Scientific Computing , volume =

    work page

Showing first 80 references.