{"id":"56fabf27-c4ab-46d1-8759-52b5e8501b87","arxiv_id":"2607.06723","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Determinant-one SPD action fibers form a global analytic totally geodesic bundle whose AIRM nearest controller is computable by a residual-gradient iteration with posterior certificates and recovers inverse-Hessian shape at rank d-1.","lead":"The paper builds a geometric theory for the hidden positive operators inside adaptive optimizers, reducing them to visible actions via infimal pushforward and Schur curvature. It then gives an exact, globally convergent solver for the determinant-one SPD action map with residual certificates and a sharp multi-secant recovery threshold.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified to the strongest claim (Theorems 5 and 7).","rationale":"The manuscript’s central geometric and algorithmic claims for the determinant-one action map are supported by complete, self-contained appendix proofs that appear consistent on inspection. The reader correctly isolates the external scalar gauge c_H as load-bearing only for the finite-identification statement (Theorem 8) and correctly notes that the conditional inexact corollary propagates rather than constructs matrix-function majorants. Neither issue undermines Theorems 5 or 7, which form the strongest claim. Because the proofs are analytic and the paper is theory-only, the appropriate stress-test is a targeted re-derivation of the residual and first-return steps; that check is expected to pass. The CONDITIONAL verdict already flags the precise caveats a careful reader needs; no further adjustment is warranted.","tokens_in":22210,"tokens_out":526,"duration_ms":4942,"concrete_test":"Independently re-derive the residual identity (40) and the first-return sublevel argument of Theorem 7 from the block form of ΦA and Lemmas 1–2 alone (without invoking the multi-secant recovery material of §7); confirm that the same L0=ψ(D0/√2) and posterior residual certificates are recovered.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's strongest claim is the global analytic action bundle for full-column-rank A, closed totally geodesic fibers, unique analytic nearest-controller section via residual R22(Σ⋆)°=0, and global linear residual-gradient convergence with sharp non-asymptotic bounds and posterior certificates (Theorems 5 and 7). Appendix B constructs the explicit analytic trivialization ΦA, proves total geodesy by block geodesic splitting under AIRM, and obtains analyticity of the section by the analytic implicit-function theorem on the strongly convex fiber objective. Appendix C derives the residual gradient identity, applies the current-sublevel first-return argument with the sharp radius majorant L0=ψ(D0/√2) from AIRM curvature comparison (Lemmas 1–2, Proposition 1), and obtains exact rate inversion plus residual certificates. Classical Hadamard ingredients are correctly credited; the action-specific package is self-contained. The gauge assumption flagged by the reader is confined to Theorem 8 and does not underwrite Theorems 5 or 7. No internal inconsistency or missing step appears in the load-bearing chain for the strongest claim.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","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).","tokens_in":22453,"tokens_out":1112,"duration_ms":10633,"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":[{"comment":"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.","section":null},{"comment":"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.","section":null}],"minor_comments":[{"comment":"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.","section":null},{"comment":"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.","section":null},{"comment":"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.","section":null},{"comment":"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.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The manuscript is a solid theory contribution whose load-bearing claims (Theorems 5 and 7) appear sound on a careful reading of the appendices. The gauge caveat for Theorem 8 is already acknowledged by the authors and is not a hidden flaw; it only needs more balanced emphasis. Fit for a serious math.OC or geometric-optimization venue is good. I see no citation or novelty-disclosure issues that would require editorial intervention beyond ordinary review."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The load-bearing claims hold. For full-column-rank A the map P ↦ PA on the determinant-one SPD manifold is a global analytic bundle with closed totally geodesic fibers, a unique analytic nearest-controller section, and a residual-gradient iteration that converges globally and linearly from every feasible start with sharp non-asymptotic bounds and observable posterior certificates (Theorems 5 and 7). That package is new; the classical Schur, Hadamard, and projection ingredients are correctly credited.\n\nWhat the paper does well is close the geometry into computation. Appendix B gives an explicit analytic trivialization ΦA, proves total geodesy by block geodesic splitting, and gets analyticity of the section from the analytic IFT on a strongly convex fiber objective. Appendix C derives the exact residual gradient, runs a current-sublevel first-return argument with the sharp radius majorant ψ(D0/√2) from AIRM curvature comparison, and inverts the linear rate exactly. The active reduction to an r ≤ 2m subspace with an O(r³) spectral-arithmetic bound is clean and useful. The interaction-curvature story is standard Schur but correctly specialized so that the same vertical Hessian governs both response and the residual solver.\n\nSoft spots are real but limited. Theorem 8’s finite-identification claim needs the external scalar gauge c_H = (det H)^{1/d}; the paper itself supplies the 2-d counter-example showing ordinary secants do not determine the shape. That assumption does not underwrite Theorems 5 or 7. The conditional inexact corollary only propagates supplied matrix-function majorants; it does not construct them. There are no experiments, so reproducibility is re-implementation from the text. None of this breaks the central chain.\n\nThis is for people who work on Riemannian optimization, quasi-Newton geometry, or the theory of adaptive preconditioners. A serious referee should see it. I would engage with the bundle and solver results; the recovery theorem needs the gauge flag kept visible.","headline":"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.","tokens_in":23058,"tokens_out":509,"would_cite":true,"duration_ms":5209,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C25","53C35","65K10","15A24","49J53"],"pacs":[],"model":"grok-4.5","headline":"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.","keywords":["optimization geometrodynamics","infimal pushforward","interaction curvature","SPD action bundle","affine-invariant metric","canonical controller","multi-secant recovery","active spectral reduction"],"falsifier":"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.","tokens_in":23067,"feed_emoji":"📐","tokens_out":757,"duration_ms":7131,"temperature":0.7,"pith_summary":"Adaptive optimizers hide positive operators that decide how a visible gradient becomes a step. Many such operators can realize the same visible action, so the paper treats the choice of hidden geometry as a variational reduction problem: eliminate every hidden state that produces the same action, keep the least-deformation representative, and read off the curvature that remains after that elimination. The central construction is the determinant-one action map that multiplies a positive-definite matrix by a fixed full-rank matrix A. That map is shown to be a global analytic bundle whose fibers are closed and totally geodesic; each fiber therefore has a unique nearest controller. An explicit logarithmic residual turns the search for that controller into a strongly convex problem that converges linearly from every feasible start, supplies sharp non-asymptotic bounds, and yields observable stopping certificates. When the same idea is applied to nested multi-secant observations of a quadratic Hessian (with its determinant scale known), the projections recover the exact inverse shape after rank d-1. The result converts an abstract geometric description of adaptive optimizers into an exact iterative computation with posterior certificates and a finite-identification theory.","feed_headline":"Hidden optimizer geometry becomes a certified residual solver","feed_subtitle":"The SPD action map is a global bundle whose fibers admit a unique nearest controller and sharp linear rates.","key_machinery":"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.","core_discovery":"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.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["SPD action map is a global analytic bundle with unique nearest controller","Determinant-one fibers admit certified residual-gradient linear rates","Optimizer geometry reduces to residual equation on totally geodesic fibers","Affine-invariant section yields global linear convergence with certificates","Hidden states collapse to Schur curvature and exact residual solver"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["SPD action map is a global analytic bundle with unique nearest controller","Determinant-one fibers admit certified residual-gradient linear rates","Optimizer geometry reduces to residual equation on totally geodesic fibers","Affine-invariant section yields global linear convergence with certificates","Hidden states collapse to Schur curvature and exact residual solver"]},"model":"grok-4.5","effort":"low","cost_usd":0.003682,"raw_usage":{"total_tokens":1261,"prompt_tokens":875,"num_sources_used":0,"completion_tokens":82,"cost_in_usd_ticks":36820000,"prompt_tokens_details":{"text_tokens":875,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":304,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":875,"tokens_out":82,"duration_ms":3272,"temperature":1.0,"reasoning_tokens":304,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T15:57:38.943538+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"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.","supporting_citations":[],"review_version":2}