REVIEW 2 major objections 4 minor 4 cited by
The map that turns hidden SPD controllers into visible actions is a global analytic bundle whose fibers admit a unique nearest controller and a certified residual solver.
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 →
T0 review · grok-4.5
2026-07-14 15:57 UTC pith:M75IMBSA
load-bearing objection Solid theory paper: global analytic PA=B bundle plus certified residual solver are real and well-proved; gauge caveat is confined to the recovery theorem. the 2 major comments →
Optimization Geometrodynamics: Variational Reduction and Interaction Curvature
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
For full-column-rank A the determinant-one SPD action map P ↦ PA is a global real-analytic bundle whose fibers are closed, totally geodesic and isometric to a lower-dimensional determinant-one SPD manifold; the unique affine-invariant nearest controller is the analytic section characterized by the residual equation R_{22}(Σ★)° = 0, and the residual-gradient iteration converges globally and linearly from every feasible initializer with sharp value, distance and residual bounds together with observable posterior certificates.
What carries the argument
The global analytic action bundle Φ_A together with its explicit logarithmic residual W_B(Σ) = R_{22}(Σ)°. The bundle trivializes every fiber of P ↦ PA; the residual is the exact whitened negative gradient of the squared affine-invariant distance, converting nearest-controller selection into a strongly convex fiber problem with sharp curvature majorants and posterior certificates.
Load-bearing premise
The finite multi-secant recovery of inverse-Hessian shape requires that the overall determinant scale of the Hessian is already known from outside the action observations; without that scalar the rank-d-1 threshold no longer identifies a unique shape.
What would settle it
In dimension d=3 construct an admissible action B of rank m=1 for which the residual-gradient iteration either leaves the current sublevel or fails to satisfy the claimed linear rate with the stated majorant L_0=ψ(D_0/√2); or, for nested multi-secants of rank d-1 with known gauge, exhibit a determinant-one completion distinct from the true inverse shape.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper develops a variational theory of hidden geometry in adaptive optimizers via infimal pushforward of deformation energies, yielding composition across hierarchies, smooth response (hidden susceptibility), and Schur effective curvature. For affine pre-reduction mechanism amplitudes the induced interaction curvature is the negative-semidefinite operator −G∗H−1G. The main realization is the determinant-one affine-invariant SPD action map πA(P)=PA. For full-column-rank A the authors construct a global analytic trivialization ΦA of the action bundle, prove that every fiber is closed and totally geodesic under AIRM, and obtain a unique analytic nearest-controller section characterized by the residual equation R22(Σ⋆)°=0. A strongly-convex fiber engine with the sharp current-sublevel majorant L0=ψ(D0/√2) produces a residual-gradient iteration that converges globally and linearly from every feasible initializer, with non-asymptotic value, distance and residual bounds and observable posterior certificates (Theorems 5–7). An active realization reduces the spectral kernel to dimension r≤2m, and nested shape-normalized multi-secant projections obey a CAT(0) Pythagorean law that recovers the determinant-one inverse Hessian shape at the sharp rank threshold d−1 when the scalar gauge cH=(det H)1/d is known (Theorem 8).
Significance. If the claims hold, the work supplies a clean geometric closure of the multi-secant completion problem under the AIRM metric: an explicit global analytic bundle, total geodesy of the determinant-one fibers, an exact residual, a globally linearly convergent solver with sharp non-asymptotic and posterior certificates, an active spectral reduction of cost O(r3) with r≤2m, and a sharp finite-identification threshold for nested normalized actions. Classical Hadamard first-order and projection ingredients are correctly credited; the action-specific package (bundle, residual, current-sublevel majorant, active realization, conditional error propagation) is self-contained and fully proved in the appendices. The theory-only character, complete analytic proofs, and explicit flagging of the external scalar gauge are strengths. The results turn an abstract reduction principle into an exact iterative computation with certificates and a finite-identification theory, which is of genuine interest for geometric optimization and quasi-Newton analysis.
major comments (2)
- Theorem 8 and the surrounding discussion in §7 correctly flag that the scalar gauge cH=(det H)1/d must be known independently of the action observations; the two-dimensional counter-example Ht=diag(1,t) shows that ordinary Hessian-vector secants of rank d−1 do not determine the inverse shape. This external information is load-bearing for the finite-identification claim and is not supplied by the action oracle. The abstract and contributions list should state the gauge requirement with equal prominence to the rank threshold itself, so that the recovery result is not read as a scale-free identification theorem.
- The interaction-curvature development (§4, Theorem 4, Corollary 1) is formally correct under the affine pre-reduction hypothesis, but its concrete link to the main action-bundle solver remains local and illustrative. The paper would be stronger if it either (i) exhibited one explicit optimizer-relevant family of Δi for which the mixed entries of −G∗H−1G are computed and interpreted, or (ii) more clearly demoted the interaction material to a supporting corollary of the same vertical Hessian that governs the residual solver. As written, the two halves of the title sit somewhat loosely together.
minor comments (4)
- Notation for the whitened residual WB(Σ)=R22(Σ)° and the active residual (Wa,wc) is introduced in several places; a single display early in §5 or §6 would reduce the reader’s bookkeeping load.
- The companion compression map P↦A⊤PA is cited as Li (2026). A one-sentence contrast of the two fiber geometries in the introduction or related-work section would help readers who encounter only one of the papers.
- Proposition 1 (action-family attainment of the radius majorant) is useful; a brief forward pointer from the statement of L0=ψ(D0/√2) in Theorem 7 would make the sharpness claim easier to locate.
- A few long sentences in the abstract and introduction (e.g., the multi-clause description of the solver) could be split for readability without changing content.
Circularity Check
No significant circularity: action-bundle geometry, residual solver, and rates are derived from explicit AIRM trivialization and classical Hadamard descent, not from fitted inputs or load-bearing self-citation.
full rationale
The paper is a self-contained pure-theory derivation. Infimal pushforward, Schur response, and the strongly-convex fiber engine are classical and correctly credited; the action-specific claims (global analytic trivialization Φ_A of P ↦ PA, total geodesy of determinant-one fibers, exact logarithmic residual R_22(Σ)°, sharp current-sublevel majorant L_0=ψ(D_0/√2), active r≤2m realization, and nested CAT(0) recovery) are proved from the AIRM metric and the determinant-one constraint in Appendices B–C without free parameters fitted to data. The companion self-citation (Li, 2026) is scoped to a different fiber (A⊤PA) and is not used to underwrite Theorems 5 or 7. Theorem 8 explicitly flags the external scalar gauge c_H rather than smuggling it. No step reduces a claimed prediction to its own definition or to an unverified self-citation chain.
Axiom & Free-Parameter Ledger
axioms (5)
- domain assumption The ambient space of hidden controllers is the determinant-one SPD manifold equipped with the affine-invariant Riemannian metric (AIRM).
- standard math Sectional curvatures of Sd++ and its totally geodesic submanifolds lie in [-1/2,0] (Lemma 1).
- standard math Every nonempty closed geodesically convex subset of a finite-dimensional Hadamard manifold admits a unique metric projection.
- ad hoc to paper The scalar gauge c_H = (det H)^{1/d} is known independently of the action observations.
- domain assumption A is full column rank and B lies in the admissible base BA = {B : A⊤B ≻ 0 and symmetric}.
invented entities (2)
-
Optimization geometrodynamics / action bundle ΦA
no independent evidence
-
Interaction curvature -G*H^{-1}G
no independent evidence
read the original abstract
Adaptive optimizers carry hidden states that change how visible gradients become parameter motion. We develop optimization geometrodynamics as a variational theory of this hidden geometry. Infimal pushforward eliminates all hidden states realizing the same action and composes across optimizer hierarchies. Under smooth nondegeneracy, it yields hidden susceptibility and the Schur-complement curvature seen after relaxation. For affine pre-reduction perturbations, the induced interaction curvature is the negative-semidefinite operator $-G^*H^{-1}G$, whose mixed entries integrate to finite mechanism contrasts. Our main realization is the determinant-one affine-invariant SPD action map $P\mapsto PA$. We prove a global analytic bundle with closed totally geodesic fibers and a unique analytic nearest-controller section. A strongly convex fiber theorem and an explicit logarithmic action residual give a globally linearly convergent solver from every feasible initializer, together with nonasymptotic value, controller-distance, and residual bounds and observable posterior stopping certificates. A conditional inexact result propagates supplied rigorous residual-error and radius majorants. The dense spectral kernel is confined to an active subspace of dimension $r\le 2m$, yielding an explicit spectral-arithmetic operation bound and a strict dimensional reduction when $r<d$. For nested shape-normalized quadratic actions, canonical multi-secant projections satisfy an exact CAT(0) Pythagorean decrease and recover the determinant-one inverse Hessian shape at the sharp rank threshold $d-1$, provided the scalar gauge $c_H=(\det H)^{1/d}$ is known. These results turn the action bundle into an exact iterative computation with posterior certificates and a finite-identification theory.
Forward citations
Cited by 4 Pith papers
-
Causal Optimizer Interaction Calculus: Hidden Geometric Relaxation and Identifiable Interventions
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...
-
Restricted Dynamic Geometric Complexity: Path-Space Reduction and M\"obius--Jacobi Response
RDGC is the least affine-invariant length of an admissible metric path to a Hessian-relative condition target, with global Green and bordered Jacobi–KKT response laws on Hadamard path space.
-
Causal Optimizer Interaction Calculus: Hidden Geometric Relaxation and Identifiable Interventions
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...
-
Restricted Dynamic Geometric Complexity: Path-Space Reduction and M\"obius--Jacobi Response
Restricted dynamic geometric complexity measures the intrinsic affine-invariant path distance from an initial metric to a condition-number target when the metric family is structurally constrained, with exact LMI and ...
Reference graph
Works this paper leans on
-
[1]
12 Miroslav Bačak.Convex Analysis and Optimization in Hadamard Spaces
doi: 10.1162/089976698300017746. 12 Miroslav Bačak.Convex Analysis and Optimization in Hadamard Spaces. De Gruyter,
-
[2]
doi: 10.15352/afa/1396833507. J. Frédéric Bonnans and Alexander Shapiro.Perturbation Analysis of Optimization Problems. Springer New York,
-
[3]
Nicolas Boumal.An Introduction to Optimization on Smooth Manifolds
doi: 10.1007/978-1-4612-1394-9. Nicolas Boumal.An Introduction to Optimization on Smooth Manifolds. Cambridge University Press,
-
[4]
Nicolas Boutet, Joris Degroote, and Rob Haelterman
doi: 10.1017/9781009166164. Nicolas Boutet, Joris Degroote, and Rob Haelterman. A symmetric grouped and ordered multi-secant quasi-newton update formula.Optimization Methods and Software, 37(6):1965–1986,
-
[5]
doi: 10.1080/ 10556788.2022.2053970. Paul H. Calamai and Jorge J. Moré. Quasi-newton updates with bounds.SIAM Journal on Numerical Analysis, 24(6):1434–1441,
arXiv 2022
-
[6]
Michael Crawshaw, Chirag Modi, Mingrui Liu, and Robert M
doi: 10.1137/0724092. Michael Crawshaw, Chirag Modi, Mingrui Liu, and Robert M. Gower. An exploration of non-euclidean gradient descent: Muon and its many variants,
-
[7]
Yu-Hong Dai and Nobuo Yamashita
URLhttps://arxiv.org/abs/2510.09827. Yu-Hong Dai and Nobuo Yamashita. Analysis of sparse quasi-newton updates with positive definite matrix completion.Journal of the Operations Research Society of China, 2(1):39–56,
-
[8]
doi: 10.1137/1021091. M. Levent Doğan, Alperen Ergür, and Elias Tsigaridas. Optimal preconditioning is a geodesically convex optimization problem,
-
[9]
John Duchi, Elad Hazan, and Yoram Singer
URLhttps://arxiv.org/abs/2512.06618. John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization.Journal of Machine Learning Research, 12:2121–2159,
-
[10]
URLhttps://arxiv.org/abs/2505.02222. Roger Fletcher. A new variational result for quasi-newton formulae.SIAM Journal on Optimization, 1(1): 18–21,
-
[11]
doi: 10.1137/0801002. Wenbo Gao and Donald Goldfarb. Block BFGS methods.SIAM Journal on Optimization, 28(2):1205–1231,
-
[12]
Wenzhi Gao, Zhaonan Qu, Madeleine Udell, and Yinyu Ye
doi: 10.1137/16M1092106. Wenzhi Gao, Zhaonan Qu, Madeleine Udell, and Yinyu Ye. Scalable approximate optimal diagonal preconditioning.Computational Optimization and Applications, 94(2):439–473,
-
[13]
URLhttps://doi.org/10.1007/s10589-026-00770-8
doi: 10.1007/ s10589-026-00770-8. URLhttps://doi.org/10.1007/s10589-026-00770-8. Gene H. Golub and Victor Pereyra. The differentiation of pseudo-inverses and nonlinear least squares problems whose variables separate.SIAM Journal on Numerical Analysis, 10(2):413–432,
-
[14]
doi: 10.1137/0710036. Serge Gratton, V. Malmedy, and Philippe L. Toint. Quasi-newton updates with weighted secant equations. Optimization Methods and Software, 30(4):748–755,
-
[15]
doi: 10.1080/10556788.2014.971025. 13 Robert Grone, Charles R. Johnson, Eduardo M. Sá, and Henry Wolkowicz. Positive definite completions of partial hermitian matrices.Linear Algebra and its Applications, 58:109–124,
-
[16]
doi: 10.1137/1.9780898717778. Keller Jordan. Muon: An optimizer for hidden layers in neural networks.https://kellerjordan.github. io/posts/muon/,
-
[17]
doi: 10.1080/10556788. 2011.613073. Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. InInternational Conference on Learning Representations,
doi:10.1080/10556788 2011
-
[18]
URLhttps://arxiv.org/abs/1412.6980. arXiv:1412.6980. Zavier Li. Information-induced training geometry: Exact reduction, canonical completion, and structured expressivity,
-
[19]
doi: 10.4153/CJM-2004-035-5. Zhaosong Lu and Ting Kei Pong. Minimizing condition number via convex programming.SIAM Journal on Matrix Analysis and Applications, 32(4):1193–1211,
-
[20]
doi: 10.1137/100795097. Pierre Maréchal and Jane J. Ye. Optimizing condition numbers.SIAM Journal on Optimization, 20(2): 935–947,
-
[21]
James Martens and Roger Grosse
doi: 10.1137/080740544. 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,
-
[22]
URLhttps://arxiv.org/abs/2406.17748. U. Passy and E. Z. Prisman. Secant relations versus positive definiteness in quasi-newton methods.Journal of Optimization Theory and Applications, 44(4):681–687,
-
[23]
Xavier Pennec, Pierre Fillard, and Nicholas Ayache
doi: 10.1007/BF00938401. Xavier Pennec, Pierre Fillard, and Nicholas Ayache. A riemannian framework for tensor computing.Interna- tional Journal of Computer Vision, 66(1):41–66,
-
[24]
Zhaonan Qu, Wenzhi Gao, Oliver Hinder, Yinyu Ye, and Zhengyuan Zhou
doi: 10.1007/s11263-005-3222-z. Zhaonan Qu, Wenzhi Gao, Oliver Hinder, Yinyu Ye, and Zhengyuan Zhou. Optimal diagonal preconditioning. Operations Research, 73(3):1479–1495,
-
[25]
doi: 10.1287/opre.2022.0592. R. Tyrrell Rockafellar.Convex Analysis. Princeton University Press,
-
[26]
doi: 10.1007/978-3-642-02431-3. Robert B. Schnabel. Quasi-newton methods using multiple secant equations. Technical report, Defense Technical Information Center,
-
[27]
14 Mirai Tanaka and Kazuhide Nakata
doi: 10.1137/140978168. 14 Mirai Tanaka and Kazuhide Nakata. Positive definite matrix approximation with condition number constraint. Optimization Letters, 8(3):939–947,
-
[28]
doi: 10.1007/s11590-013-0632-7. Yongge Tian. Equalities and inequalities for hermitian solutions and hermitian definite solutions of the two matrix equationsAX =B and AXA∗=B.Aequationes Mathematicae, 86(1–2):107–135,
-
[29]
Alice Barbara Tumpach and Gabriel Larotonda
doi: 10.1007/s00010-012-0179-1. Alice Barbara Tumpach and Gabriel Larotonda. Totally geodesic submanifolds in the manifold SPD of symmetric positive-definite real matrices.Information Geometry, 7(S2):913–942,
-
[30]
URLhttps://arxiv. org/abs/2409.11321. Nobuo Yamashita. Sparse quasi-newton updates with positive definite matrix completion.Mathematical Programming, 115(1):1–30,
-
[31]
doi: 10.1007/s10107-007-0137-1. Hongyi Zhang and Suvrit Sra. First-order methods for geodesically convex optimization. In29th Annual Conference on Learning Theory, volume 49 ofProceedings of Machine Learning Research, pp. 1617–1638. PMLR,
-
[32]
It is closed, and block distance splitting gives dAI(diag(M,ρΣ0),diag(M,ρΣ1)) =d AI(Σ 0,Σ 1)
Hence the fiber is totally geodesic. It is closed, and block distance splitting gives dAI(diag(M,ρΣ0),diag(M,ρΣ1)) =d AI(Σ 0,Σ 1). Congruence byQLis an AIRM isometry, proving all fiber claims. 18 Step 5: canonical projection and analyticity.The determinant-one SPD slice is a finite-dimensional Hadamard manifold. A nonempty closed geodesically convex set h...
2007
-
[33]
Projection, proximal, and convex-analytic foundations on closed geodesically convex sets Supplies abstract projection principles; does not construct or solve the action fiber Definite matrix equations and geodesic SPD projec- tion (Tian, 2013; Lim, 2004; Tumpach&Larotonda,2024; Sra & Hosseini,
2013
-
[34]
Algebraic definite-solution identities, pro- jection results, total-geodesy criteria, and algorithms for geodesic SPD or conic models No global determinant-one action bundle together with its residual, active realization, and action posterior certificate Least-change and multi- secant updates (Dennis & Schnabel, 1979; Schnabel, 1983; Fletcher, 1991; Gao &...
1979
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.