Pith. sign in

REVIEW 4 minor 2 cited by

The geometric effort of reaching a useful preconditioner is a path-space value whose responses are generated by one Green–Jacobi inverse.

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

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

T0 review · grok-4.5

2026-07-14 15:49 UTC pith:PXLVNNNO

load-bearing objection Solid path-space theory for restricted preconditioner metrics: global Green–Jacobi and bordered hard-target response under Hadamard/convexity hypotheses, with clean closed-form checks.

arxiv 2607.07204 v2 pith:PXLVNNNO submitted 2026-07-08 math.OC cs.LG

Restricted Dynamic Geometric Complexity: Path-Space Reduction and M\"obius--Jacobi Response

Zavier Li This is my paper
classification math.OC cs.LG MSC 49L2049K4053C2365K1090C25
keywords restricted dynamic geometric complexityaffine-invariant geometrypath-space reductionJacobi operatorMöbius interactionstructured preconditioningHadamard manifoldbordered KKT
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Structured preconditioners only let optimizers move among a small family of positive metrics. Reachability of a target condition number says whether a useful metric exists, but not how much intrinsic motion is needed to get there. This paper defines restricted dynamic geometric complexity as the least affine-invariant length of an admissible metric path that ends inside a Hessian-relative condition set. Path elimination is exactly min-plus composition and therefore obeys a Bellman principle; for fixed horizon the kinetic energy of the path equals squared complexity over twice the horizon. On a Hadamard state space with geodesically convex bulk and terminal potentials, a single uniformly coercive Green–Jacobi inverse produces the value Hessian, exact force-to-curvature bounds, Möbius pair and higher-order interaction effects, and every prescribed finite-order response on an intervention cube. When the condition target is hard, a bordered Jacobi–KKT operator differentiates the moving projection, its multiplier, and the constrained value Hessian, and can produce interaction signs that the unconstrained Gram law forbids. Explicit diagonal models confirm both layers in closed form, and a sequential update protocol shows that the declared path class can make complexity strictly larger than ordinary projection distance.

Core claim

Restricted dynamic geometric complexity is a genuine path-space value: its global elimination law is Bellman composition of affine-invariant lengths, its smooth local response is the Schur complement of a coercive Jacobi operator, and its finite intervention effects are exact integrals of one Green or bordered-Green kernel. On a Hadamard manifold with geodesically convex potentials the Green inverse is global on a whole intervention cube; on a regular active spectral stratum the bordered inverse differentiates the moving hard condition target and explains why hard interactions need not share the unconstrained sign.

What carries the argument

The Green–Jacobi inverse of the path second variation (and its bordered Jacobi–KKT counterpart for hard targets). One uniformly coercive operator solves bulk and terminal forcing, yields the value Hessian as a negative Gram, and by recursion generates every finite-order Möbius response; the bordered version jointly differentiates path, moving projection endpoint, and multiplier.

Load-bearing premise

All bulk and terminal potentials must stay geodesically convex on a whole neighborhood of the intervention cube, and the state space must be Hadamard (nonpositive curvature), so that a single coercive Green inverse exists cube-wide; hard-target claims further need a fixed smooth active spectral stratum with simple extreme eigenvalues.

What would settle it

In the two-dimensional determinant-one diagonal model, compute the interaction matrix of the forced kinetic action both from the closed-form second derivatives and by finite differences of the value; if the matrix is not strictly negative definite with the stated determinant 1/2160, or if the sequential three-dimensional protocol fails to exceed ambient AIRM distance by the factor 2/√3, the explicit claims collapse.

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

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

0 major / 4 minor

Summary. The paper defines restricted dynamic geometric complexity (RDGC) as the least affine-invariant length of an admissible metric path whose endpoint meets a Hessian-relative generalized-eigenvalue condition target. Path elimination yields an exact min-plus semigroup and Bellman principle; fixed-horizon kinetic energy equals squared complexity over twice the horizon. On a Hadamard state space with geodesically convex bulk and terminal potentials, a single uniformly coercive Green–Jacobi inverse produces the value Hessian, force-to-curvature bounds, exact Möbius pair/conditional effects, and arbitrary finite-order responses (Theorem 4.11, Corollary 4.12). For hard condition targets on a regular active spectral stratum, a bordered Jacobi–KKT theorem differentiates the moving projection, multiplier, and constrained Hessian, allowing interaction signs that need not match the unconstrained Gram law (Theorem 4.13). The theory specializes to AIRM geometry, with closed-form diagonal models (Theorems 5.1–5.4) and a sequential protocol whose RDGC strictly exceeds ambient projection distance (Proposition 5.3). Appendices supply detailed proofs; Section 6 reports deterministic numerical verification of discrete identities.

Significance. If the results hold under the stated hypotheses, the paper supplies a coherent path-space foundation for measuring geometric effort of structured preconditioners, together with explicit differential response laws (Green and bordered) that convert path elimination into Möbius interaction formulas. Strengths include the clean separation of algebraic (min-plus), variational (Tonelli/HJB), and differential (Jacobi–Schur) layers; the global Hadamard derivation of cube-wide uniqueness and coercivity; closed-form diagonal realizations that instantiate both unconstrained and hard-target responses with both interaction signs; and the sequential gap showing that the declared path class can change the value. The appendices give full proofs, and Section 6 provides reproducible deterministic checks of discrete algebraic identities. The work is self-contained at the path-space level and correctly scopes its classical differentiability to smooth active strata.

minor comments (4)
  1. The companion papers [29,30] are cited as arXiv preprints with 2026 dates; ensure final bibliographic entries are complete and that the self-contained claim for the path-space layer is preserved if those works remain unpublished.
  2. Notation for the relative spectrum and κ_gen is introduced carefully, but occasional shorthand κ(G^{-1}H) could still be flagged once more in the introduction to prevent misreading as the Euclidean condition number of a nonnormal product.
  3. Table 1 and the verification script are useful; a one-sentence note that the coordinate eigenvalue is mesh-scaled (already present) could be repeated in the table caption for readers who skip the surrounding text.
  4. A few long sentences in the abstract and contributions list could be split for readability without changing content.

Circularity Check

1 steps flagged

No significant circularity: path-elimination, length–energy, and global Green/bordered responses are derived under stated hypotheses; companion [30] is infrastructure only and is re-proved in path-Hilbert form.

specific steps
  1. self citation load bearing [Introduction Contributions / §4.4–4.7 (Thm 4.9, Cor 4.15); companion [30]]
    "The finite-dimensional companion [30] proves generic Schur response and affine-intervention Gram formulas for hidden controller states. Those formulas are the finite-dimensional template here. The present paper’s new response results begin with their path-Hilbert realization… The local formulas in Theorem 4.9 and Corollary 4.15 are their Hilbert-path lift and are used here as infrastructure."

    Same-author arXiv [30] is cited as the finite-dimensional Schur/Gram template for the local path-response formulas. This is not load-bearing for the paper’s central claims: Thm 4.9 is proved independently via the Hilbert IFT and coercive second variation, Cor 4.15 is derived from it, and the global cube-wide Green and bordered hard-target theorems are new under Hadamard/KKT hypotheses with full proofs. Mild self-citation only; does not force the main results by construction.

full rationale

The paper’s derivation chain is self-contained. RDGC is defined as an infimum of affine-invariant path length to a spectral target; the min-plus semigroup (Thm 4.2) follows from path concatenation and action additivity alone; the length–energy identity (Thm 4.4) is Cauchy–Schwarz plus reparameterization, not a smuggled prediction. The global Möbius–Jacobi theorem (Thm 4.11) and bordered hard-target theorem (Thm 4.13) are proved from geodesic convexity (MJ2), Hadamard nonpositive curvature, and KKT regularity on a fixed active stratum, with full appendix proofs (implicit-function, Lax–Milgram, envelope, Boolean integration). Closed-form diagonal models (Thms 5.1–5.2, 5.4; Prop 5.3) instantiate those formulas by direct calculation without fitted parameters. Companion [30] supplies the finite-dimensional Schur/Gram template; the present paper re-proves the path-Hilbert lift (Thm 4.9, Cor 4.15) and explicitly separates infrastructure from the new Hadamard/cube-wide and bordered results. No self-definitional loop, no fitted-input-as-prediction, and no uniqueness imported solely from an unverified self-citation. Score 1 only for the minor same-author template citation, which is not load-bearing for the central claims.

Axiom & Free-Parameter Ledger

0 free parameters · 6 axioms · 2 invented entities

The work is pure variational geometry: no data-fitted constants. Load-bearing structure is standard Hadamard/SPD geometry plus domain assumptions that restrict the path class and potentials so that a single coercive Green operator exists on an intervention cube. RDGC itself is the central invented measure; companions supply finite-dimensional templates but are not used as numerical fits.

axioms (6)
  • domain assumption State space is a finite-dimensional Hadamard manifold (nonpositive curvature, complete, simply connected), so geodesics and projections behave well and curvature contributes nonnegatively to the index form.
    Used throughout Theorem 4.11, Corollary 4.10, and the SPD specialization; without it cube-wide coercivity need not hold.
  • domain assumption Bulk and terminal potentials are geodesically convex (Hess W_u ⪰ 0, Hess Φ_u ⪰ 0) with C^{r+2} regularity and uniform base-point bounds on a neighborhood of the intervention cube (MJ2).
    Produces unique global minimizers and the uniformly coercive Jacobi form that makes one Green inverse control all responses.
  • standard math Admissible paths form a composable system with additive action (Assumption 4.1), enabling exact min-plus path elimination without smoothness.
    Standard dynamic-programming hypothesis; Theorem 4.2 is pure algebra under this axiom.
  • domain assumption Fixed-horizon path class is closed under absolutely continuous reparameterization so length and kinetic energy share minimizers (Theorem 4.4).
    Gauge choice that identifies RDGC length with the smooth action used for response analysis.
  • domain assumption For hard targets, the active condition-number stratum has simple extreme eigenvalues, LICQ, positive multiplier, and bordered invertibility on the face of interest.
    Corollary 3.7 / Theorem 4.13; classical differentiability stops at collisions or active-set changes.
  • standard math Structured families used in corollaries (diagonal, fixed block-diagonal, det-one diagonal) are complete simply connected totally geodesic AIRM submanifolds.
    Standard SPD geometry facts that let the abstract Hadamard theorems specialize without extra embedding theory.
invented entities (2)
  • Restricted dynamic geometric complexity (RDGC) D_{K,F}(S_0;H) no independent evidence
    purpose: Quantifies least affine-invariant metric-path effort to a Hessian-relative generalized-eigenvalue condition target under a restricted family F.
    Central new value function; reduces to ordinary Hadamard projection only for closed geodesically convex unrestricted path classes.
  • Intervention-cube Green–Jacobi / bordered Jacobi–KKT response operators for metric paths no independent evidence
    purpose: Convert bulk/terminal or hard-boundary interventions into value Hessians and exact Möbius interaction integrals via one inverse.
    Path-space lift of Schur response specialized to RDGC; independent evidence is only internal closed-form checks, not external measurements.

pith-pipeline@v1.1.0-grok45 · 30319 in / 3797 out tokens · 41474 ms · 2026-07-14T15:49:51.973126+00:00 · methodology

0 comments
read the original abstract

Structured preconditioners restrict optimization to a small family of positive metrics, but endpoint condition-number reachability does not measure the geometric effort required to reach a useful metric. We formulate this effort as a path-space value problem. Restricted dynamic geometric complexity is the least affine-invariant length of an admissible metric path whose endpoint reaches a Hessian-relative generalized-eigenvalue condition target. Path elimination gives an exact min-plus semigroup and Bellman principle, while fixed-horizon kinetic energy is exactly squared complexity divided by twice the horizon. The main response result is global on a Hadamard state space: geodesic convexity produces a smooth intervention-cube path branch and a uniformly coercive Jacobi form, while one Green inverse generates the value Hessian, two-sided force-to-curvature bounds, exact M\"obius effects, and arbitrary prescribed finite-order responses. For the hard condition target, a bordered Jacobi--KKT theorem differentiates the moving projection endpoint and multiplier on every regular active spectral stratum; its indefinite inverse also explains why hard-target interactions need not share the unconstrained sign. The theory specializes to affine-invariant positive-definite geometry. A determinant-one two-dimensional diagonal model has an exact target interval, a closed-form forced path, and a strictly negative-definite interaction matrix. A moving diagonal Hessian gives a closed-form hard-target projection, multiplier, and pair effects of either sign, while a coordinate-sequential three-dimensional protocol yields an exact path metric strictly larger than the ambient projection distance. Thus the global Green and bordered hard-target responses are explicit laws of restricted metric-path elimination built on Bellman composition.

Review history (2 revisions) →

discussion (0)

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

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Causal Optimizer Interaction Calculus: Hidden Geometric Relaxation and Identifiable Interventions

    cs.LG 2026-07 conditional novelty 6.0

    Under fixed innovation coupling, finite-horizon optimizers admit minimal pathwise realizations and incidence-identifiable Möbius effects, with a five-term readout transfer from hidden relaxation and a closed reduced-v...

  2. Causal Optimizer Interaction Calculus: Hidden Geometric Relaxation and Identifiable Interventions

    cs.LG 2026-07 conditional novelty 5.0

    A modular calculus decomposes optimizer updates into geometric preconditioning plus structured nongeometric mechanisms, with a direction-expressivity theorem showing full SPD geometry captures exactly strict descent d...

Reference graph

Works this paper leans on

42 extracted references · 3 canonical work pages · cited by 1 Pith paper

  1. [1]

    A. A. Agrachev and Y. L. Sachkov,Control Theory from the Geometric Viewpoint, vol. 87 of Encyclopaedia of Mathematical Sciences, Springer, Berlin, 2004, https://doi.org/10.1007/ 978-3-662-06404-7

    2004

  2. [2]

    Amari,Natural gradient works efficiently in learning, Neural Computation, 10 (1998), pp

    S.-i. Amari,Natural gradient works efficiently in learning, Neural Computation, 10 (1998), pp. 251–276, https://doi.org/10.1162/089976698300017746

    doi:10.1162/089976698300017746 1998

  3. [3]

    Amos and J

    B. Amos and J. Z. Kolter,OptNet: Differentiable optimization as a layer in neural networks, in Proceedings of the 34th International Conference on Machine Learning, vol. 70 of Proceedings of Machine Learning Research, PMLR, 2017, pp. 136–145, https://arxiv.org/abs/1703.00443

    Pith/arXiv arXiv 2017

  4. [4]

    R. Anil, V. Gupta, T. Koren, K. Regan, and Y. Singer,Scalable second order optimization for deep learning, arXiv preprint arXiv:2002.09018, (2020), https://arxiv.org/abs/2002. 09018

    Pith/arXiv arXiv 2002

  5. [5]

    Baˇcak,Convex Analysis and Optimization in Hadamard Spaces, De Gruyter, 2014, https: //doi.org/10.1515/9783110361629

    M. Baˇcak,Convex Analysis and Optimization in Hadamard Spaces, De Gruyter, 2014, https: //doi.org/10.1515/9783110361629

    work page doi:10.1515/9783110361629 2014

  6. [6]

    Beck and M

    A. Beck and M. Teboulle,Mirror descent and nonlinear projected subgradient methods for convex optimization, Operations Research Letters, 31 (2003), pp. 167–175, https: //doi.org/10.1016/S0167-6377(02)00231-6. [7]R. Bhatia,Positive Definite Matrices, Princeton University Press, 2007

    doi:10.1016/s0167-6377(02)00231-6 2003

  7. [7]

    Blondel, Q

    M. Blondel, Q. Berthet, M. Cuturi, R. Frostig, S. Hoyer, F. Llinares-L ´opez, F. Pe- dregosa, and J.-P. Vert,Efficient and modular implicit differentiation, in Advances in Neural Information Processing Systems, vol. 35, 2022, https://arxiv.org/abs/2105.15183

    Pith/arXiv arXiv 2022

  8. [8]

    J. F. Bonnans and A. Shapiro,Perturbation Analysis of Optimization Problems, Springer New York, 2000, https://doi.org/10.1007/978-1-4612-1394-9

    doi:10.1007/978-1-4612-1394-9 2000

  9. [9]

    Dacorogna,Direct Methods in the Calculus of Variations, vol

    B. Dacorogna,Direct Methods in the Calculus of Variations, vol. 78 of Applied Mathe- matical Sciences, Springer New York, New York, 2 ed., 2008, https://doi.org/10.1007/ 978-0-387-55249-1

    2008

  10. [10]

    Dasgupta, N

    T. Dasgupta, N. S. Pillai, and D. B. Rubin,Causal inference from2 K factorial designs 20Z. LI by using potential outcomes, Journal of the Royal Statistical Society: Series B (Statistical Methodology), 77 (2015), pp. 727–753, https://doi.org/10.1111/rssb.12085

    work page doi:10.1111/rssb.12085 2015

  11. [11]

    Dhamdhere, M

    K. Dhamdhere, M. Sundararajan, and A. Agarwal,The Shapley–Taylor interaction index, in Proceedings of the 37th International Conference on Machine Learning, vol. 119 of Proceedings of Machine Learning Research, PMLR, 2020, pp. 2709–2718, https://arxiv.org/ abs/1902.05622

    Pith/arXiv arXiv 2020

  12. [12]

    Diamond and S

    S. Diamond and S. Boyd,Stochastic matrix-free equilibration, Journal of Optimization Theory and Applications, 172 (2016), pp. 436–454, https://doi.org/10.1007/s10957-016-0990-2

    work page doi:10.1007/s10957-016-0990-2 2016

  13. [13]

    M. L. Do˘gan, A. Erg¨ur, and E. Tsigaridas,Optimal preconditioning is a geodesically convex optimization problem, 2025, https://arxiv.org/abs/2512.06618, https://arxiv.org/abs/2512. 06618

    arXiv 2025

  14. [14]

    A. L. Dontchev and R. T. Rockafellar,Implicit Functions and Solution Mappings, Springer Series in Operations Research and Financial Engineering, Springer New York, 2014, https: //doi.org/10.1007/978-1-4939-1037-3

    doi:10.1007/978-1-4939-1037-3 2014

  15. [15]

    Duchi, E

    J. Duchi, E. Hazan, and Y. Singer,Adaptive subgradient methods for online learning and stochastic optimization, Journal of Machine Learning Research, 12 (2011), pp. 2121–2159, https://jmlr.org/papers/v12/duchi11a.html

    2011

  16. [16]

    W. H. Fleming and H. M. Soner,Controlled Markov Processes and Viscosity Solutions, vol. 25 of Stochastic Modelling and Applied Probability, Springer, New York, 2 ed., 2006, https://doi.org/10.1007/0-387-31071-1

    doi:10.1007/0-387-31071-1 2006

  17. [17]

    Franceschi, P

    L. Franceschi, P. Frasconi, S. Salzo, R. Grazzi, and M. Pontil,Bilevel programming for hyperparameter optimization and meta-learning, in Proceedings of the 35th International Conference on Machine Learning, vol. 80 of Proceedings of Machine Learning Research, PMLR, 2018, pp. 1568–1577, https://arxiv.org/abs/1806.04910

    Pith/arXiv arXiv 2018

  18. [18]

    G. H. Golub and V. Pereyra,The differentiation of pseudo-inverses and nonlinear least squares problems whose variables separate, SIAM Journal on Numerical Analysis, 10 (1973), pp. 413–432, https://doi.org/10.1137/0710036

    doi:10.1137/0710036 1973

  19. [19]

    Grosse and J

    R. Grosse and J. Martens,A Kronecker-factored approximate Fisher matrix for convolution layers, in Proceedings of the 33rd International Conference on Machine Learning, vol. 48 of Proceedings of Machine Learning Research, PMLR, 2016, pp. 573–582, https://proceedings. mlr.press/v48/grosse16.html, https://arxiv.org/abs/1602.01407

    Pith/arXiv arXiv 2016

  20. [20]

    Gupta, T

    V. Gupta, T. Koren, and Y. Singer,Shampoo: Preconditioned stochastic tensor opti- mization, in Proceedings of the 35th International Conference on Machine Learning, vol. 80 of Proceedings of Machine Learning Research, PMLR, 2018, pp. 1842–1850, https://proceedings.mlr.press/v80/gupta18a.html

    2018

  21. [21]

    N. J. Higham,Functions of Matrices: Theory and Computation, Society for Industrial and Applied Mathematics, 2008, https://doi.org/10.1137/1.9780898717778

    doi:10.1137/1.9780898717778 2008

  22. [22]

    J. D. Janizek, P. Sturmfels, and S.-I. Lee,Explaining explanations: Axiomatic feature interactions for deep networks, Journal of Machine Learning Research, 22 (2021), pp. 1–54, https://arxiv.org/abs/2002.04138

    Pith/arXiv arXiv 2021

  23. [23]

    D. P. Kingma and J. Ba,Adam: A method for stochastic optimization, in International Confer- ence on Learning Representations, 2015, https://arxiv.org/abs/1412.6980. arXiv:1412.6980

    Pith/arXiv arXiv 2015

  24. [24]

    Krichene, A

    W. Krichene, A. M. Bayen, and P. L. Bartlett,Accelerated mirror de- scent in continuous and discrete time, in Advances in Neural Information Processing Systems, 2015, https://proceedings.neurips.cc/paper/2015/hash/ f60bb6bb4c96d4df93c51bd69dcc15a0-Abstract.html

    2015

  25. [25]

    A. S. Lewis,Derivatives of spectral functions, Mathematics of Operations Research, 21 (1996), pp. 576–588, https://doi.org/10.1287/moor.21.3.576

    doi:10.1287/moor.21.3.576 1996

  26. [26]

    A. S. Lewis,The mathematics of eigenvalue optimization, Mathematical Programming, 97 (2003), pp. 155–176, https://doi.org/10.1007/s10107-003-0441-3

    doi:10.1007/s10107-003-0441-3 2003

  27. [27]

    A. S. Lewis and H. S. Sendov,Twice differentiable spectral functions, SIAM Journal on Matrix Analysis and Applications, 23 (2001), pp. 368–386, https://doi.org/10.1137/ S089547980036838X

    2001

  28. [28]

    Li,Information-Induced Training Geometry: Exact Reduction, Canonical Completion, and Structured Expressivity, 2026

    Z. Li,Information-Induced Training Geometry: Exact Reduction, Canonical Completion, and Structured Expressivity, 2026. arXiv preprint

    2026

  29. [29]

    Li,Optimization Geometrodynamics: Variational Reduction and Interaction Curvature, 2026, https://arxiv.org/abs/2607.06723, https://arxiv.org/abs/2607.06723

    Z. Li,Optimization Geometrodynamics: Variational Reduction and Interaction Curvature, 2026, https://arxiv.org/abs/2607.06723, https://arxiv.org/abs/2607.06723. arXiv:2607.06723

    Pith/arXiv arXiv 2026

  30. [30]

    Lim,Best approximation in riemannian geodesic submanifolds of positive definite matri- ces, Canadian Journal of Mathematics, 56 (2004), pp

    Y. Lim,Best approximation in riemannian geodesic submanifolds of positive definite matri- ces, Canadian Journal of Mathematics, 56 (2004), pp. 776–793, https://doi.org/10.4153/ CJM-2004-035-5

    2004

  31. [31]

    Lu and T

    Z. Lu and T. K. Pong,Minimizing condition number via convex programming, SIAM Journal on Matrix Analysis and Applications, 32 (2011), pp. 1193–1211, https://doi.org/10.1137/ RESTRICTED DYNAMIC GEOMETRIC COMPLEXITY21 100795097

    2011

  32. [32]

    Mar´echal and J

    P. Mar´echal and J. J. Ye,Optimizing condition numbers, SIAM Journal on Optimization, 20 (2009), pp. 935–947, https://doi.org/10.1137/080740544

    doi:10.1137/080740544 2009

  33. [33]

    Martens and R

    J. Martens and R. Grosse,Optimizing neural networks with Kronecker-factored approximate curvature, in Proceedings of the 32nd International Conference on Machine Learning, vol. 37 of Proceedings of Machine Learning Research, PMLR, 2015, pp. 2408–2417, https: //proceedings.mlr.press/v37/martens15.html

    2015

  34. [34]

    Pennec, P

    X. Pennec, P. Fillard, and N. Ayache,A riemannian framework for tensor computing, International Journal of Computer Vision, 66 (2006), pp. 41–66, https://doi.org/10.1007/ s11263-005-3222-z

    2006

  35. [35]

    Pukelsheim,Optimal Design of Experiments, Society for Industrial and Applied Mathematics, Philadelphia, 2006, https://doi.org/10.1137/1.9780898719109

    F. Pukelsheim,Optimal Design of Experiments, Society for Industrial and Applied Mathematics, Philadelphia, 2006, https://doi.org/10.1137/1.9780898719109

    doi:10.1137/1.9780898719109 2006

  36. [36]

    Z. Qu, W. Gao, O. Hinder, Y. Ye, and Z. Zhou,Optimal diagonal preconditioning, Operations Research, 73 (2025), pp. 1479–1495, https://doi.org/10.1287/opre.2022.0592

    doi:10.1287/opre.2022.0592 2025

  37. [37]

    Shazeer and M

    N. Shazeer and M. Stern,Adafactor: Adaptive learning rates with sublinear memory cost, in Proceedings of the 35th International Conference on Machine Learning, vol. 80 of Proceedings of Machine Learning Research, PMLR, 2018, pp. 4596–4604, https://proceedings.mlr.press/ v80/shazeer18a.html

    2018

  38. [38]

    Sra and R

    S. Sra and R. Hosseini,Conic geometric optimization on the manifold of positive definite matrices, SIAM Journal on Optimization, 25 (2015), pp. 713–739, https://doi.org/10.1137/ 140978168

    2015

  39. [39]

    W. Su, S. Boyd, and E. J. Cand `es,A differential equation for modeling Nesterov’s accelerated gradient method: Theory and insights, Journal of Machine Learning Research, 17 (2016), pp. 1–43, https://arxiv.org/abs/1503.01243

    Pith/arXiv arXiv 2016

  40. [40]

    Tanaka and K

    M. Tanaka and K. Nakata,Positive definite matrix approximation with condition num- ber constraint, Optimization Letters, 8 (2014), pp. 939–947, https://doi.org/10.1007/ s11590-013-0632-7. Published online in 2013

    2014

  41. [41]

    van der Sluis,Condition numbers and equilibration of matrices, Numerische Mathematik, 14 (1969), pp

    A. van der Sluis,Condition numbers and equilibration of matrices, Numerische Mathematik, 14 (1969), pp. 14–23, https://doi.org/10.1007/BF02165096

    doi:10.1007/bf02165096 1969

  42. [42]

    Wibisono, A

    A. Wibisono, A. C. Wilson, and M. I. Jordan,A variational perspective on accelerated methods in optimization, Proceedings of the National Academy of Sciences, 113 (2016), pp. E7351–E7358, https://arxiv.org/abs/1603.04245. Appendix A. Proofs for Path-Space Reduction. Proof of Theorem4.2. Fix an admissible path γ∈P s,t(x, z) and put y = γ(r). Restriction an...

    Pith/arXiv arXiv 2016