{"id":"ff1565ff-d77b-4584-bbf9-b6030c86e5f8","arxiv_id":"2501.02395","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":3,"one_line_summary":"For uniformly hyperbolic systems, the optimal linear response perturbation is the Riesz representative of the response functional divided by its norm, with Fourier coefficients given by linear responses of basis perturbations.","lead":"This paper shows how to find the smallest allowable push to a chaotic, uniformly hyperbolic dynamical system that most increases a chosen long-term average. It proves the best push is unique and computable from a response vector, and demonstrates the computation in 2, 3 and 21 dimensions.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract's uniqueness theorem omits boundedness of P; as stated, it is false for unbounded strictly convex closed P, so the advertised central claim overreaches.","rationale":"The most load-bearing claim of the paper is the existence/uniqueness theorem advertised in the abstract and proved in Section 4.1. The proof in the body explicitly requires boundedness, but the abstract omits it, making the central claim false as stated. This is an internal inconsistency, not a matter of external consensus. The reader's weakest assumption—reliance on the fast adjoint response formula and decay of correlations—is a legitimate concern about the numerical examples and the dependence on prior work, but it does not identify the missing boundedness hypothesis. The strict convexity definition in the paper is also problematic as written, since it appears to require every boundary point to be interior; this is secondary to the boundedness issue. The unit-ball optimal perturbation formula v/||v||_H remains correct because the unit ball is bounded. Overall, the paper's computational contribution is plausible and the mathematical core can be repaired by adding 'bounded' to the abstract and clarifying the definition of strict convexity. The conditional verdict stands, as the advertised theorem needs correction before the paper is definitive.","tokens_in":17227,"tokens_out":15301,"duration_ms":140183,"concrete_test":"Verify the counterexample directly: set H = R^2, P = {(x,y) : y ≥ x^2}, and define I:H → C^3(T^1, TT^1) by I(x,y) = y·V for a fixed nonzero C^3 vector field V. Choose a uniformly hyperbolic map and an observable Φ with R(V) ≠ 0 (e.g., a small perturbation of an Anosov map in the direction V). Then R(I(x,y)) = R(V)·y is unbounded above on P, so no maximizer exists, despite P being closed, strictly convex, and continuously mapped into C^3. If the authors instead add 'bounded' to the theorem statement, confirm that Proposition 3 then matches the abstract.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract claims that for any strictly convex, closed subset P of a Hilbert space H continuously mapped into C^3 vector fields, there is a unique optimal perturbation. The proof in Section 4.1 (Proposition 3), however, explicitly assumes P is bounded: 'Let P be a bounded, convex and closed set of H, then the problem (20) has a solution.' Boundedness is not a technical convenience: a continuous linear functional need not attain its supremum on an unbounded closed strictly convex set. For example, take H = R^2 and P = {(x,y) : y ≥ x^2}, which is closed, convex, strictly convex, and unbounded. The continuous linear functional L(x,y) = y has sup_P L = ∞ and no maximizer exists. This example can be embedded in the paper's framework by mapping (x,y) to the constant vector field y·V for a nonzero V in C^3 and choosing an observable whose response at V is nonzero. Thus the theorem as stated in the abstract is false; the paper's actual bounded version is correct and the main unit-ball applications are unaffected, but the central advertised claim needs the boundedness hypothesis.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the optimal linear response problem for uniformly hyperbolic (mixing Axiom A) diffeomorphisms. Given an observable and a convex feasible set P of infinitesimal perturbations contained in a Hilbert space H continuously embedded into C^3 vector fields, it seeks the perturbation maximizing the linear response R. Using the authors' fast adjoint response formula, Lemma 1 establishes a bound on R in the C^{1,α} norm, and Section 4.1 gives existence and uniqueness for bounded strictly convex closed feasible sets. When P is the unit ball, the optimizer is identified with the normalized Riesz representative v of R, and Section 4.3 expands v in a Fourier basis on tori. Section 5 implements the algorithm on 2-, 3-, and 21-dimensional solenoid-like maps, validating the computed optimal response against finite-difference estimates of the perturbed measure. Section 6 sketches a PDE-based approach for manifolds with boundary.","tokens_in":17495,"tokens_out":12836,"duration_ms":132195,"significance":"If the main estimates are correct, the paper offers a general Hilbert-space framework for optimal linear response in hyperbolic systems and a practical adjoint-based algorithm that scales favorably when the feasible perturbation set is small relative to the phase-space dimension. The numerical coefficients are computed, not fitted, and the validation against direct finite-difference approximations of μ_γ(Φ) is an external check. The main theorem in Proposition 3 is defensible once boundedness is included; the issues identified below are local to the statements and definitions, not to the central numerical strategy.","major_comments":[{"comment":"The uniqueness theorem as advertised is false because boundedness of P is omitted. For example, take H=R^2 and P={(x,y): y≥x^2}, which is closed, convex, strictly convex, and unbounded; the continuous linear functional L(x,y)=y is unbounded above on P, so no maximizer exists. This example can be embedded in the paper's setting by mapping (x,y) to the constant vector field yV for a nonzero V∈C^3 and choosing an observable with R(V)≠0. Proposition 3 correctly assumes P is bounded; the abstract and the informal statement in Section 1.2 must also state boundedness of P, together with the condition that R is not identically zero on P. The unit-ball applications are unaffected by this correction, but the advertised central claim overreaches as written.","section":"Abstract; §4.1, Proposition 3"},{"comment":"Definition 1 as written requires γx+(1−γ)y∈int(A) for all pairs x,y∈A, including x=y. For x=y the condition forces every point of A to be interior, so no closed convex set with nonempty boundary, including the unit ball which the text explicitly says is strictly convex, satisfies the definition. The intended definition must be for all x≠y. Proposition 3's uniqueness conclusion relies on this corrected reading.","section":"Definition 1"},{"comment":"The statement of Proposition 4 omits the hypothesis that P is the unit ball of H. As written it is applied to the general optimization problem (11) over an arbitrary bounded strictly convex closed set, but the formula X_opt=v/||v||_H is only correct for the unit ball; for a ball of radius r the maximizer is r v/||v||_H. The preceding paragraph indicates the intended setting, but the proposition itself should state explicitly that P is the unit ball.","section":"Proposition 4"}],"minor_comments":[{"comment":"The statement writes |R(X)|≤ C||X||_{C^{1,α}}≤||X||_{C^3}; the second inequality is not generally true with constant 1 and should read ||X||_{C^{1,α}}≤ C||X||_{C^3} (or the C^3 norm should be defined to dominate the C^{1,α} norm by construction).","section":"Lemma 1"},{"comment":"The estimate writes a sum of C λ^n for n from −W to W; for negative n this should be C λ^{|n|}, otherwise the summand grows as n becomes negative.","section":"Proof of Lemma 1, estimate for R_2W"},{"comment":"Proposition 3 refers to 'problem (20)', but at that point the relevant problem is (11); the reference should be corrected.","section":"Proposition 3"},{"comment":"The feasible set is written as P:={||X'||_{H^5}=1}, which is the unit sphere rather than the unit ball used in Propositions 5 and 6. For a linear objective the maximizer is the same, but the feasible set should be written consistently with the theoretical statement.","section":"§5.1 and §5.2"},{"comment":"The three maps are described as 'solenoid-like' and asserted to be hyperbolic, but no verification of uniform hyperbolicity or exponential decay of correlations is given. Since Lemma 1 and the fast adjoint response formula assume such a regime, the numerical claims should either include a check of hyperbolicity or be phrased as illustrative computations whose theoretical guarantees require the stated hypotheses.","section":"§5.1–5.3"},{"comment":"The text says 'M = TM is the M dimensional torus'; this should be M = T^M (or the notation should be introduced consistently).","section":"§4.3"}],"recommendation":"major_revision","confidential_remarks":"The paper's main theoretical framework is reasonable, but the abstract's theorem is false without a boundedness hypothesis, and Definition 1 is formally self-contradictory. These are fixable locally, and the unit-ball results are unaffected. The numerical examples are suggestive, but a rigorous check of hyperbolicity would strengthen the demonstration. I would recommend major revision rather than rejection, since the core derivation and computational method appear sound once the statements are corrected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is the first paper I know that solves the optimal linear response problem for general uniformly hyperbolic systems, and the 21-dimensional example is a genuine step beyond the earlier Markov-chain and expanding-circle work. The abstract, however, states a theorem that is false as written because it drops the boundedness hypothesis on the feasible set P.\n\nWhat is genuinely new: the C^{1,alpha} boundedness of the response operator, proved via the fast adjoint formula and decay of correlations, and the Hilbert-space optimization argument that turns a computed Riesz representative into the optimal perturbation. The numerical section is careful to compute the response of each Fourier basis element and then verify the predicted slope against direct finite-difference estimates of mu_gamma(Phi). That verification is external to the optimization, not fitted. The code is public. The treatment of both diffeomorphism perturbations and additive perturbations is a nice structural point.\n\nNow the soft spots, in order of size. First, the stress-test is right: Proposition 3 assumes P is bounded, the abstract does not, and the theorem is false without boundedness—a continuous linear functional need not attain its supremum on an unbounded closed strictly convex set (e.g., the epigraph of x^2 with L(x,y)=y). All the applications use the unit ball, which is bounded, so the main results survive; the abstract just needs the word 'bounded.' Second, Definition 1 of strict convexity literally fails for x=y on the boundary; it needs x≠y. Third, Lemma 1 displays '|R(X)| ≤ C||X||C1,alpha ≤ ||X||C3', which has the norm inequality backwards—surely meant C||X||C1,alpha ≤ C'||X||C3. Minor, but confusing. Fourth, the numerical examples assert uniform hyperbolicity without proof, and the figures have no error bars, even though the text identifies the two error sources (decorrelation truncation W and sampling length T). A convergence table or error bars would materially strengthen the empirical claim.\n\nNone of this is fatal. The central idea is sound, the numerics are genuinely informative, and the authors are honest about what is proved versus assumed (see the remark after Proposition 2). This deserves a serious referee. With the abstract fixed and the typos corrected, I would take it. If I were working on control or response of chaotic systems, I would cite it, and I would put it on a reading group list as an example of a well-executed response computation in high dimension.","headline":"First general optimal-response result for hyperbolic systems, with a genuinely high-dimensional numerical demonstration; the abstract overreaches by omitting boundedness, but the core result is sound and worth refereeing.","tokens_in":17991,"tokens_out":3381,"would_cite":true,"duration_ms":32586,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["37D20","37C30","37M25"],"pacs":[],"model":"deepseek-v4-flash","headline":"A unique optimal perturbation exists for hyperbolic systems, and its Fourier coefficients are linear responses of a basis.","keywords":["linear response","uniformly hyperbolic systems","Axiom A attractors","optimal perturbation","fast adjoint response formula","Riesz representation theorem","Sobolev spaces","numerical linear response"],"falsifier":"For one of the toral examples, compute $\\mu_\\gamma(\\Phi)$ for several small $\\gamma$ by direct long-orbit averages for a perturbation $X$, fit the slope at $\\gamma=0$, and compare it with $R(X)$ from the fast adjoint formula; a systematic mismatch beyond the stated sampling error would falsify the representation (6) and hence the computed $v$.","tokens_in":17036,"feed_emoji":"🎯","tokens_out":7326,"duration_ms":66171,"temperature":0.7,"pith_summary":"This paper asks a control question for chaotic systems that have a well-defined statistical steady state: among all small allowed perturbations, which one makes the long-time average of a given observation grow the most? For uniformly hyperbolic (Axiom A) systems, the answer is shown to be unique whenever the set of allowed perturbations is a strictly convex closed subset of a Hilbert space embedded in $C^3$ vector fields. The proof uses the fast adjoint response formula to show the linear response operator is bounded, then represents it by a vector $v$ via Riesz representation; on the unit ball the best perturbation is simply $v/\\|v\\|_H$. This turns an infinite-dimensional optimization into a Fourier-coefficient computation, and the authors demonstrate it numerically in dimensions 2, 3, and 21.","feed_headline":"The best small nudge for a hyperbolic system is unique and computable","feed_subtitle":"Linear response plus a Riesz representative turns the search into Fourier coefficients of one vector v.","key_machinery":"The fast adjoint response formula, Eq. (6), $$R(X) = \\lim_{W\\to\\infty} \\mu\\!\\left[ \\big(S(d\\Phi) + \\varphi_W\\, S(\\operatorname{div}_v f_*)\\big) X + \\varphi_W\\, \\operatorname{div}_v X \\right],$$ combines the adjoint shadowing operator $S$ (whose fixed-point equation is $\\omega = f^*\\omega + d\\Phi$) with the equivariant divergence formula $\\operatorname{div}_v X = \\tilde\\varepsilon \\nabla_{\\tilde e} X$. Because every term is a pointwise function on the attractor, this formula converts response estimates into decay-of-correlation estimates, which is what yields the $C^{1,\\alpha}$ bound in Lemma 1 and ultimately the Hilbert-space representative $v$ via Riesz representation.","core_discovery":"The paper's central claim is that the optimal response problem for a uniformly hyperbolic (Axiom A) system has a unique solution under natural convexity assumptions, and that the solution is given explicitly by the Riesz representative of the linear response operator. Lemma 1 shows that the response operator $R$ extends to $C^{1,\\alpha}$ and satisfies $|R(X)| \\le C\\|X\\|_{C^{1,\\alpha}}$, so $R$ is a continuous linear functional on any Hilbert space $H$ continuously embedded in $C^3$. Riesz representation then yields a unique $v\\in H$ with $R(w)=\\langle w,v\\rangle_H$; when the feasible set is the unit ball, the unique maximizer is $X_{\\mathrm{opt}} = v/\\|v\\|_H$, and its Fourier coefficients are $c_i = R(b_i)$ for an orthonormal basis $\\{b_i\\}$. The same conclusion holds both for perturbations by composition with a diffeomorphism near the identity and for additive perturbations of the map, and the paper verifies the construction numerically on 2-, 3-, and 21-dimensional toral examples.","pith_inferences":["The paper leaves open whether the linear response itself exists for perturbations that are only $C^{1,\\alpha}$; if it does, the $C^{1,\\alpha}$ bound in Lemma 1 suggests the optimizer could be defined over rougher perturbation classes than $C^3$, but this requires a separate proof.","Because each coefficient $c_i = R(b_i)$ is an independent orbit-based computation, the method is embarrassingly parallel; the reported wall-clock times suggest that with enough cores the 21-dimensional example could be pushed to much larger basis sets.","The cost of the Fourier approach grows with the number of basis functions, not with the phase-space dimension, so choosing a deliberately small Hilbert space $H$ (e.g., perturbations depending only on one coordinate) is a natural route to control high-dimensional systems, as the 21-dimensional example illustrates.","For applied settings with a prescribed family of controls, the strict-convexity result means the optimal infinitesimal control is unique; testing whether that uniqueness survives finite-time or nonlinear effects would be a natural next step."],"forward_implications":["When the feasible set is the unit ball of $H$, the unique optimal perturbation is $v/\\|v\\|_H$, so computing the optimizer reduces to computing the linear response of each basis element, $c_i = R(b_i)$.","The same statement holds for two perturbation mechanisms: composing the map with a diffeomorphism near the identity, and adding a deterministic perturbation to the map, with the additive case represented by $I(X') = X'\\circ f^{-1}$.","For any strictly convex, bounded, closed feasible set with a nonzero response functional, the optimal perturbation exists and is unique, extending the unit-ball result to constrained control sets such as perturbations acting only on selected coordinates.","On the torus, the Fourier expansion of $v$ is explicitly constructed from trigonometric basis functions, and on more general manifolds $v$ solves a high-order Laplacian equation with boundary conditions given in Section 6.","The numerical experiments in dimensions 2, 3, and 21 confirm that the computed optimal perturbation produces a larger linear response than every single basis element and matches the finite-difference trend of $\\mu_\\gamma(\\Phi)$."],"supporting_citations":[{"why":"Supplies the adjoint shadowing operator and its Hölder regularity, used in Lemma 1 and in the fast adjoint response decomposition.","marker":"[20]"},{"why":"Supplies the equivariant divergence formula for the unstable contribution, the second ingredient of formula (6).","marker":"[22]"},{"why":"Establishes that linear response exists for C^3 perturbations of uniformly hyperbolic maps, the starting point of the optimization.","marker":"[27]"},{"why":"Provides the fast adjoint response algorithm whose orbit-based estimates compute the coefficients R(b_i) in the numerical examples.","marker":"[19]"},{"why":"Gives the general strict-convexity argument that Proposition 3 invokes for existence and uniqueness of the optimizer on a Hilbert ball.","marker":"[2]"},{"why":"The Fourier-basis-plus-response method for expanding circle maps that this paper simplifies and generalizes to hyperbolic systems.","marker":"[10]"},{"why":"Supplies exponential decay of correlations for mixing Axiom A systems, which gives the geometric factor in the Lemma 1 bounds.","marker":"[7]"},{"why":"Provides the Sobolev embedding theorem and the Lax-Milgram setup used for the Hilbert space and the non-torus Laplacian formulation.","marker":"[8]"}],"fun_headline_variants":["Fast adjoint method finds unique optimal perturbation for hyperbolic systems","Riesz representative yields the best small nudge for hyperbolic flows","Optimal response is unique and explicit via fast adjoint method","Hyperbolic system's best perturbation: one vector's Fourier coefficients","Unique optimal nudge from Riesz representative in high dimensions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the fast adjoint response formula exactly represents the linear response of the unperturbed Axiom A system, which requires exponential decay of correlations; for the numerical examples, the required uniform hyperbolicity is asserted rather than rigorously verified.","fun_headline_variants_meta":{"raw":{"variants":["Fast adjoint method finds unique optimal perturbation for hyperbolic systems","Riesz representative yields the best small nudge for hyperbolic flows","Optimal response is unique and explicit via fast adjoint method","Hyperbolic system's best perturbation: one vector's Fourier coefficients","Unique optimal nudge from Riesz representative in high dimensions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000412,"raw_usage":{"total_tokens":2176,"prompt_tokens":1035,"completion_tokens":1141,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":651,"completion_tokens_details":{"reasoning_tokens":1071}},"tokens_in":651,"tokens_out":1141,"duration_ms":9266,"temperature":1.0,"reasoning_tokens":1071,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:15:11.334046+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For one of the toral examples, compute $\\mu_\\gamma(\\Phi)$ for several small $\\gamma$ by direct long-orbit averages for a perturbation $X$, fit the slope at $\\gamma=0$, and compare it with $R(X)$ from the fast adjoint formula; a systematic mismatch beyond the stated sampling error would falsify the representation (6) and hence the computed $v$.","supporting_citations":[{"cited_title":"Antown, G","cited_arxiv_id":null,"evidence_quote":"Gives the general strict-convexity argument that Proposition 3 invokes for existence and uniqueness of the optimizer on a Hilbert ball."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies exponential decay of correlations for mixing Axiom A systems, which gives the geometric factor in the Lemma 1 bounds."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Sobolev embedding theorem and the Lax-Milgram setup used for the Hilbert space and the non-torus Laplacian formulation."}],"review_version":1}