Pith. sign in

REVIEW 5 minor 24 references

Positive metrics cannot erase critical structure, but the cost of better conditioning is exactly the affine-invariant distance of the relative log-spectrum to a low-width set.

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

2026-07-10 22:40 UTC pith:M75IMBSA

load-bearing objection Clean theory paper: exact SPD complexity for conditioning is real, scoped as an oracle lower bound, and the proofs hold.

arxiv 2607.06723 v1 pith:M75IMBSA submitted 2026-07-07 math.OC cs.LGmath.DG

Optimization Geometrodynamics: A Framework for Dynamic Geometric Optimization

Zavier Li This is my paper
classification math.OC cs.LGmath.DG MSC 90C2553C2165K1049Q20
keywords optimization geometrodynamicsdynamic geometric complexityaffine-invariant SPD geometrycondition numberRiemannian metricspreconditioningMorse index invarianceHessian matching
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.

Most optimizers quietly reshape geometry through preconditioners and adaptive states, yet they are still described as motion in a fixed Euclidean space. This paper formalizes that idea as optimization geometrodynamics: a coupled evolution of a parameter path, a transported particle distribution, and a controlled time-varying Riemannian metric. It proves what positive metrics can never change—critical points, Morse indices, and global geodesic-convexity barriers—while showing what they can improve: local conditioning, transport, and exit flux away from exact critical points. The central quantitative object is dynamic geometric complexity, the minimum geometric deformation needed to lower a difficulty measure such as condition number. In the oracle model of strongly convex quadratics with full positive-definite metric control, that cost equals the affine-invariant distance from the relative log-spectrum to the set of metrics whose condition number is at most K. The framework is offered as a theory-only benchmark against which real adaptive methods can later be scored once their admissible metric families, curvature proxies, and discretization errors are stated.

Core claim

In the full SPD oracle for strongly convex quadratics, the length-type dynamic geometric complexity of reducing the relative condition number to K is exactly the affine-invariant distance from the relative metric S0 to the set CK of metrics with κ ≤ K; that distance equals the Euclidean distance of the sorted log-eigenvalues of S0 to the nearest interval of width log K.

What carries the argument

Dynamic geometric complexity CLen_AI(K; S0): the infimum of affine-invariant SPD path length needed to reach relative condition number ≤ K. In the full SPD benchmark it collapses to DK(S0), the ℓ2 distance of the relative log-spectrum to an interval of width log K.

Load-bearing premise

The exact complexity formula treats the optimizer as free to choose any continuous positive-definite metric path while knowing the true target curvature exactly; practical methods only approximate this ideal through restricted families, noisy proxies, and finite steps.

What would settle it

On a strongly convex quadratic with known Hessian H, compute DK(S0) for an initial metric and check whether any full-SPD continuous path of strictly shorter affine-invariant length can reach condition number ≤ K; if a shorter path exists, the equality CLen_AI = DK fails.

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 / 5 minor

Summary. The paper proposes optimization geometrodynamics as a formal benchmark language for gradient-based methods whose internal states act as evolving Riemannian metrics. The state is a triple (θ_t, μ_t, g_t) evolving by steepest descent in the current metric, weak transport of a particle distribution, and a controlled metric response law. Boundary theorems show that positive metrics leave critical sets and Morse indices invariant and cannot remove global geodesic-convexity obstructions arising from non-global critical points. Positive mechanisms include Hessian-matching flows that monotonically reduce the geometric condition number for strongly convex quadratics without a commutativity assumption. The central quantitative object is dynamic geometric complexity: in the full-SPD oracle for strongly convex quadratics it equals the affine-invariant distance from the relative metric S_0 to the set C_K of metrics with condition number at most K, which is exactly the Euclidean distance of the sorted log-spectrum to the nearest interval of width log K (Theorem 6). Restricted-family certificates (diagonal, block, Kronecker), gauge-invariant observables, and a fixed-time local Morse-saddle flux formula complete the framework. All claims are formal statements with proofs collected in Appendices A–C; the work is theory-only and scopes the exact complexity result as an oracle lower bound.

Significance. If the results hold, the paper supplies a clean, coordinate-free separation between invariant obstructions (critical sets, Morse indices, global geodesic convexity) and reducible geometric mismatch (conditioning, transport, flux), together with an exact, affine-invariant cost for the latter in the full-SPD quadratic oracle. Theorem 6 is a precise, parameter-free equality that can serve as a lower bound against which restricted adaptive preconditioners (diagonal, Kronecker, low-rank, etc.) can be measured once their admissible families, proxy errors, and discretization are specified. The restricted certificates (Propositions 6–8), the explicit Hessian-matching ODE without commutativity, and the gauge-invariant observable algebra are concrete, checkable contributions. The work is deliberately foundational rather than algorithmic; its value lies in providing invariants and benchmark costs that future algorithm-specific analyses can use. Strengths include fully written proofs, explicit model-scope statements, and a deterministic toy workflow for the small SPD certificates.

minor comments (5)
  1. Section 4.3 and the Discussion correctly flag the absence of a general well-posedness theorem for arbitrary nonlinear metric laws G; a short forward pointer in the introduction that every formal result is stated under its own finite-dimensional or fixed-time hypotheses would further reduce any residual ambiguity for readers.
  2. Proposition 2 (fixed-time local Morse-saddle flux) is carefully scoped as a leading-order local asymptotic; the Discussion already lists the needed caveats (degenerate saddles, long-time density evolution). A single sentence in the proposition statement itself restating that no claim is made about escape times or basin-transition probabilities would make the limitation even harder to miss.
  3. Notation for the relative metric S = H^{-1/2} G H^{-1/2} and the geometric condition number κ_g(f) is introduced cleanly in Section 3, but a brief reminder when the same symbols reappear in the complexity section (Section 7) would help readers who jump between sections.
  4. The two-dimensional diagonal certificate (Proposition 6) and the Kronecker residual invariance (Proposition 8) are useful; ensuring that the supplementary toy-workflow script is permanently archived with the paper (or linked from the arXiv abstract) would strengthen reproducibility of the numerical checks.
  5. A few typographical items: “geometrodynamics” is hyphenated inconsistently in early paragraphs; the arXiv identifier in the header is fine, but the journal submission should standardize the citation style for the continuous-time optimization references (Su et al., Wibisono et al., etc.).

Circularity Check

0 steps flagged

No significant circularity: dynamic geometric complexity is defined as an infimum of path lengths and then proved equal to a spectral distance via standard SPD geometry.

full rationale

The paper is a self-contained theory-only framework. Dynamic geometric complexity is introduced as an infimum of affine-invariant lengths to a condition-number set (Definition 1), and Theorem 6 proves that this equals the Euclidean distance of the sorted log-spectrum to the nearest interval of width log K. The proof (Appendix B.2) uses the standard 1-Lipschitz property of the sorted log-spectrum map under the affine-invariant metric (Lemma 1) plus an explicit shared-eigenvector projection construction; the equality is a theorem, not a redefinition of the cost as the spectral quantity. Boundary theorems follow directly from the definition of the Riemannian gradient and Sylvester's law of inertia. Hessian matching, Onsager relaxation, and restricted-family certificates are derived under stated model classes without fitted parameters or load-bearing self-citations. There are no predictions that reduce to inputs by construction, no uniqueness theorems imported from the same authors, and no renaming of known empirical patterns presented as new derivation. The full-SPD oracle is explicitly scoped as a lower-bound benchmark, not as a claim about practical optimizers. Score 0 is therefore appropriate.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 2 invented entities

The paper is pure theory built on standard Riemannian and SPD geometry. No data-fitted free parameters. Load-bearing background is classical differential geometry and matrix analysis; the paper’s own inventions are the coupled state language and the complexity functional, not new physical entities. The main modeling idealization is the full-SPD oracle with exact target curvature.

axioms (5)
  • standard math Riemannian gradient and Hessian are defined via a positive-definite metric g; at critical points Hess_g f equals the intrinsic second differential d²f (standard Riemannian geometry).
    Used throughout Sections 3–5 and Theorems 2–3; inertia independence via Sylvester’s law.
  • standard math The affine-invariant metric on SPD matrices has geodesic distance dAI(S0,S1)=||log(S0^{-1/2}S1 S0^{-1/2})||_F (Bhatia; Pennec et al.).
    Foundation of Definition 1 and Theorem 6; cited as standard SPD geometry.
  • standard math Sorted log-eigenvalue map is 1-Lipschitz from (S^d_++, dAI) to Euclidean space, including repeated eigenvalues (Hermitian perturbation / Lidskii-type bound).
    Lemma 1 in Appendix B; used to prove the lower bound in Theorem 6.
  • domain assumption Objective, metric path, density, and response law are smooth enough for the displayed transport and descent identities on compact domains or with finite flux (regularity and transport axiom, §4.3).
    Required for equations (1)–(4) and Propositions 3–4; general well-posedness for arbitrary G is not claimed.
  • ad hoc to paper Full positive-definite metric control with exact target curvature H is available as an oracle benchmark (model scope, §1 and §7).
    Defines the exact complexity class; paper treats restricted families as follow-up specializations.
invented entities (2)
  • optimization geometrodynamics state (θ_t, μ_t, g_t) no independent evidence
    purpose: Unified benchmark language for trajectory, transport, and controlled time-varying metric.
    New formal packaging of optimizer state; not an independent physical object.
  • dynamic geometric complexity CLen_AI / DK no independent evidence
    purpose: Minimum affine-invariant geometric cost to reduce a difficulty observable (condition number) to a target class.
    Central quantitative object of the paper; defined and characterized within the SPD oracle model.

pith-pipeline@v1.1.0-grok45 · 24062 in / 3135 out tokens · 40257 ms · 2026-07-10T22:40:44.013686+00:00 · methodology

0 comments
read the original abstract

Most gradient-based optimization methods move parameters through a fixed background geometry, even when their internal states implicitly define changing notions of length, curvature, and preconditioning. We introduce optimization geometrodynamics, a benchmark language in which optimization is a coupled evolution of a parameter trajectory, a transported distribution of particles, and a controlled time-varying Riemannian metric. The language separates invariant obstructions from improvable geometric mismatch: positive metrics preserve critical points and Morse indices, and cannot remove global geodesic-convexity obstructions, but can alter conditioning, distributional transport, and flux away from exact critical points. We introduce dynamic geometric complexity, the minimum geometric cost required to reduce an optimization difficulty observable. In the oracle benchmark model of strongly convex quadratic objectives with full positive-definite metric control, this complexity is exactly the affine-invariant distance from the relative log-spectrum to a low-condition-number set. We also analyze Hessian-matching flows, spectral Onsager relaxation, discrete exponential projection updates, gauge-invariant observables, and fixed-time local Morse-saddle flux. The paper is theory-only: its claims are formal statements with proofs, intended to provide invariants and benchmark costs against which implementable adaptive optimizers can be compared once their admissible metric families, curvature estimates, and discretization errors are specified.

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

24 extracted references · 24 canonical work pages · 7 internal anchors

  1. [1]

    Rohan Anil, Vineet Gupta, Tomer Koren, and Yoram Singer

    doi: 10.1162/089976698300017746. Rohan Anil, Vineet Gupta, Tomer Koren, and Yoram Singer. Memory efficient adaptive optimization. InAdvances in Neural Information Processing Systems, volume 32,

    work page doi:10.1162/089976698300017746

  2. [2]

    neurips.cc/paper_files/paper/2019/hash/8f1fa0193ca2b5d2fa0695827d8270e9-Abstract.html

    URLhttps://proceedings. neurips.cc/paper_files/paper/2019/hash/8f1fa0193ca2b5d2fa0695827d8270e9-Abstract.html. Amir Beck and Marc Teboulle. Mirror descent and nonlinear projected subgradient methods for convex optimization.Operations Research Letters, 31(3):167–175,

    work page 2019

  3. [3]

    Rajendra Bhatia.Matrix Analysis

    doi: 10.1016/S0167-6377(02)00231-6. Rajendra Bhatia.Matrix Analysis. Springer,

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

  4. [4]

    Rajendra Bhatia.Positive Definite Matrices

    doi: 10.1007/978-1-4612-0653-8. Rajendra Bhatia.Positive Definite Matrices. Princeton University Press,

    work page doi:10.1007/978-1-4612-0653-8

  5. [5]

    Nicolas Boumal.An Introduction to Optimization on Smooth Manifolds

    doi: 10.1137/16M1080173. Nicolas Boumal.An Introduction to Optimization on Smooth Manifolds. Cambridge University Press,

    work page doi:10.1137/16m1080173

  6. [6]

    Bregman dynamics, contact transformations and convex optimization

    doi: 10.1017/9781009166164. Alessandro Bravetti, Maria de Leon Daza-Torres, Hugo Flores-Arguedas, and Michael Betancourt. Bregman dynamics, contact transformations and convex optimization.arXiv preprint arXiv:1912.02928,

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1017/9781009166164 1912

  7. [7]

    URL https://arxiv.org/abs/1912.02928. J. Chen, H. Chen, M. Wang, G. Dai, I. W. Tsang, and Y. Liu. Learning discretized neural networks under ricci flow.Journal of Machine Learning Research, 25:1–39,

    work page internal anchor Pith review Pith/arXiv arXiv 1912

  8. [8]

    URLhttps://arxiv.org/abs/1206.6791

    doi: 10.1080/02331934.2012.733883. URLhttps://arxiv.org/abs/1206.6791. Dennis M. DeTurck. Deforming metrics in the direction of their ricci tensors.Journal of Differential Geometry, 18(1):157–162,

    work page doi:10.1080/02331934.2012.733883 2012

  9. [9]

    12 John Duchi, Elad Hazan, and Yoram Singer

    doi: 10.4310/jdg/1214509286. 12 John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization.Journal of Machine Learning Research, 12:2121–2159,

    work page doi:10.4310/jdg/1214509286

  10. [11]

    Vineet Gupta, Tomer Koren, and Yoram Singer

    URL https://arxiv.org/abs/ 2505.12553. Vineet Gupta, Tomer Koren, and Yoram Singer. Shampoo: Preconditioned stochastic tensor optimization. InProceedings of the 35th International Conference on Machine Learning, volume 80 ofProceedings of Machine Learning Research, pp. 1842–1850. PMLR,

    work page arXiv

  11. [12]

    Elad Hazan, Amit Agarwal, and Satyen Kale

    doi: 10.4310/jdg/1214436922. Elad Hazan, Amit Agarwal, and Satyen Kale. Logarithmic regret algorithms for online convex optimization. Machine Learning, 69(2–3):169–192,

    work page doi:10.4310/jdg/1214436922

  12. [13]

    Moritz Hehl, Max von Renesse, and Melanie Weber

    doi: 10.1007/s10994-006-5006-8. Moritz Hehl, Max von Renesse, and Melanie Weber. Neural feature geometry evolves as discrete ricci flow. arXiv preprint arXiv:2509.22362,

    work page doi:10.1007/s10994-006-5006-8

  13. [14]

    Diederik P

    URLhttps://arxiv.org/abs/2509.22362. Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. InInternational Conference on Learning Representations,

    work page arXiv

  14. [15]

    Adam: A Method for Stochastic Optimization

    URLhttps://arxiv.org/abs/1412.6980. arXiv:1412.6980. Philip A. Knight and Daniel Ruiz. A fast algorithm for matrix balancing.IMA Journal of Numerical Analysis, 33(3):1029–1047,

    work page internal anchor Pith review Pith/arXiv arXiv

  15. [16]

    James Martens and Roger Grosse

    doi: 10.1093/imanum/drs019. James Martens and Roger Grosse. Optimizing neural networks with kronecker-factored approximate curvature. InProceedings of the 32nd International Conference on Machine Learning, volume 37 ofProceedings of Machine Learning Research, pp. 2408–2417. PMLR,

    work page doi:10.1093/imanum/drs019

  16. [17]

    doi: 10.1007/s11263-005-3222-z. Boris T. Polyak. Some methods of speeding up the convergence of iteration methods.USSR Computational Mathematics and Mathematical Physics, 4(5):1–17,

    work page doi:10.1007/s11263-005-3222-z

  17. [18]

    Garvesh Raskutti and Sayan Mukherjee

    doi: 10.1016/0041-5553(64)90137-5. Garvesh Raskutti and Sayan Mukherjee. The information geometry of mirror descent.IEEE Transactions on Information Theory, 61(3):1451–1457,

    work page doi:10.1016/0041-5553(64)90137-5

  18. [19]

    URLhttps://arxiv.org/ abs/1310.7780

    doi: 10.1109/TIT.2015.2391243. URLhttps://arxiv.org/ abs/1310.7780. Noam Shazeer and Mitchell Stern. Adafactor: Adaptive learning rates with sublinear memory cost. In Proceedings of the 35th International Conference on Machine Learning, volume 80 ofProceedings of Machine Learning Research, pp. 4596–4604. PMLR,

    work page doi:10.1109/tit.2015.2391243 2015

  19. [20]

    Weijie Su, Stephen Boyd, and Emmanuel J

    doi: 10.2140/pjm.1967.21.343. Weijie Su, Stephen Boyd, and Emmanuel J. Candès. A differential equation for modeling nesterov’s accelerated gradient method: Theory and insights.Journal of Machine Learning Research, 17(153):1–43,

    work page doi:10.2140/pjm.1967.21.343 1967

  20. [21]

    A Differential Equation for Modeling Nesterov's Accelerated Gradient Method: Theory and Insights

    URL https://arxiv.org/abs/1503.01243. Abraham van der Sluis. Condition numbers and equilibration of matrices.Numerische Mathematik, 14:14–23,

    work page internal anchor Pith review Pith/arXiv arXiv

  21. [22]

    doi: 10.1007/BF02165096. X. Wang, V. Srinivasan, Q. Fu, S. Mitra, A. Wilson, and A. Wibisono. Accelerated convex optimization via hamiltonian dynamics with deterministic integration time.arXiv preprint arXiv:2606.17260,

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1007/bf02165096

  22. [23]

    Accelerated Convex Optimization via Hamiltonian Dynamics with Deterministic Integration Time

    URL https://arxiv.org/abs/2606.17260. Andre Wibisono, Ashia C. Wilson, and Michael I. Jordan. A variational perspective on accelerated methods in optimization.Proceedings of the National Academy of Sciences, 113(47):E7351–E7358,

    work page internal anchor Pith review Pith/arXiv arXiv

  23. [24]

    A Variational Perspective on Accelerated Methods in Optimization

    URL https://arxiv.org/abs/1603.04245. A Framework and Boundary Proofs This appendix gives the elementary geometric facts used in the main text. We state them for smooth finite-dimensional manifolds and smooth data. Boundary terms are omitted by assuming either compact support, no boundary, or the usual no-flux boundary condition. A.1 Transport and descent...

    work page internal anchor Pith review Pith/arXiv arXiv

  24. [25]

    Thusf(θ)≥f(θ⋆)for every θ, soθ⋆is a global minimizer

    For a convex one-dimensional function,ϕ(1)≥ϕ(0). Thusf(θ)≥f(θ⋆)for every θ, soθ⋆is a global minimizer. Therefore the existence of a non-global critical point rules out global geodesic convexity for every geodesically connected positive metric. A.3 Constrained Onsager response Proof of Theorem 1.By the Riesz representation theorem, there is a unique elemen...

    work page 1997