{"id":"45bea5bf-9349-4aa1-8d99-863cfbaec8fc","arxiv_id":"1908.06730","paper_version":2,"verdict":"UNVERDICTED","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"low","formal_verification":"none","parameter_count":3,"one_line_summary":"An exposition of the flow-map formulation and Jacobian-free Newton-Krylov method for computing equilibria and periodic orbits, with problem-independent code demonstrated on the Lorenz system.","lead":"This is a tutorial on finding equilibria and periodic orbits of dynamical systems using the Jacobian-free Newton-Krylov method, with a reusable code template demonstrated on the Lorenz system. It is a readable, code-supported introduction for researchers who want to compute invariant solutions without writing their own solver.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing assumption is the finite-difference Jacobian-vector product in Eq. (3.8); it is standard and supported by the Lorenz example, so no significant objection lands.","rationale":"The paper is a tutorial, not a research contribution. Its central claim is that JFNK with the flow map requires only a time-stepping code and a finite-difference Jacobian-vector product. The unspoken condition is that Phi must be smooth enough and the finite-difference step size appropriately scaled. Eq. (3.8) is a standard first-order approximation with a recommended step size that balances truncation and roundoff error. For the Lorenz example, which is smooth and uses fixed-step timestepping, this is adequate, and the provided code gives a concrete demonstration. No internal inconsistencies, circular reasoning, or unsupported research claims were found. The only caveat is that the phrase 'problem-independent' should not be read as 'universally robust', but the paper does not promise that. Therefore the reader's verdict of UNVERDICTED remains appropriate; the concern about Eq. (3.8) is valid but does not change the tutorial's status.","tokens_in":10001,"tokens_out":4832,"duration_ms":55385,"concrete_test":"Using the Lorenz ODE, compute the flow-map Jacobian-vector product J delta x by integrating the variational equation alongside the trajectory, and compare it to the finite-difference approximation in Eq. (3.8) over 100 random unit vectors delta x with the recommended step size epsilon = 1e-6 ||x||/||delta x||; if the maximum relative error exceeds 1e-3, the recommended step-size rule is questionable even for this smooth test case.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central practical claim is that invariant solutions can be found using little more than an existing time-stepping code, with the flow-map Jacobian-vector product approximated by the finite difference in Eq. (3.8), taking (epsilon ||delta x||)/||x|| = 1e-6. If Phi is not sufficiently smooth, or if the timestepper is adaptive or event-driven, this first-order approximation can be inaccurate and the GMRES-computed Newton step may fail. The paper gives no error bound and no diagnostic for detecting a poor step size. However, this is a well-established method (Knoll and Keyes, 2004), the Lorenz system is smooth, and the supplied code converges on the provided initial guesses. The paper is an exposition, not a claim of universal robustness, so this concern does not undermine its central claim for the systems it targets.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a short pedagogical exposition of how equilibria, periodic orbits, relative periodic orbits, and pre-periodic orbits can be formulated as fixed-point problems for the flow map Phi, and how these fixed points can be computed with the Jacobian-free Newton-Krylov (JFNK) method. The first half defines the flow map, invariant solutions, Poincaré sections, and Fourier slicing for continuous symmetries. The second half develops the Newton-Raphson method, the matrix-free approximation of Jacobian-vector products (Eq. (3.8)), the GMRES iteration, the hookstep trust-region strategy, constraints for the period T and shift l, optional preconditioning, and a practical template code applied to the Lorenz system using known periodic orbits from Viswanath (2003).","tokens_in":1123,"tokens_out":1259,"duration_ms":89417,"significance":"If taken as a tutorial/reference rather than a research contribution, the paper is successful: the mathematical descriptions are accurate and consistent with the cited literature, the practical pitfalls (initial guesses, tolerances, norm choices) are addressed, and the accompanying code (openpipeflow.org) directly illustrates the method on a nontrivial example. The paper does not claim new mathematical results, but it fills a gap by giving a compact, self-contained description of a technique widely used in computational fluid dynamics for finding invariant solutions. The JFNK formulation is standard, and the explanations are clear enough to be used by graduate students.","major_comments":[{"comment":"The paper states that the method requires little more than calls to an existing time stepping code and suggests a typical finite-difference step with (epsilon ||delta x||)/||x|| = 10^-6. This is reasonable for smooth systems integrated with fixed time steps, but it is not robust for adaptive or event-driven integrators, where the discrete flow map Phi_T may be non-smooth or only piecewise differentiable with respect to the initial condition. The authors should add a sentence cautioning that when adapting the code to an existing timestepper, the user must ensure that the map being differentiated is sufficiently smooth (e.g., by fixing the time step), and should perhaps mention a simple diagnostic (e.g., checking that the finite-difference quotient converges as epsilon is reduced). This does not invalidate the examples in the paper, but it qualifies the claim that little more than calls to a time stepping code is needed.","section":"Section 3.2, Eq. (3.8)"}],"minor_comments":[{"comment":"The phrase 'we can consider an equilibrium to be a special case of a periodic orbit where T may be arbitrarily chosen' is slightly misleading: for a periodic orbit T is the minimal period, while for an equilibrium any T works; this is a minor wording issue.","section":"Section 1.2"},{"comment":"The GMRES description says 'It is common to start with K1 = b/||b||'; this assumes a zero initial guess for delta x.  It would be clearer to state that the initial guess is taken as zero.","section":"Section 3.2"},{"comment":"In the expression for \tilde{A} = e^A, the scalar '1' should be the identity matrix I.","section":"Section 3.5.1"},{"comment":"Typo: 'Z = 27 in call cases' should be 'in all cases'.","section":"Section 4"},{"comment":"The constraint ang<dot(x)_i, delta x> = 0 requires that dot(x)_i is nonzero; the degenerate case of an equilibrium is not discussed.","section":"Section 3.4.1"},{"comment":"Several typos should be corrected, e.g., 'Kyrlov' in Section 3.5.1, 'pe riodic' in Section 3.4.2, 'section s' in Section 3.4.1, and the missing period after 'is required .' in Section 3.2.","section":"Various"}],"recommendation":"minor_revision","confidential_remarks":"This is a well-written pedagogical paper. The mathematical content is standard and correctly presented. My only substantive concern is the unqualified claim about 'little more than calls to an existing time stepping code' in the presence of adaptive timesteppers. The manuscript would benefit from a brief cautionary remark. I recommend minor revision; no additional experiments are needed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a teaching document, not a research contribution, and it should be judged on that basis. Its genuinely useful piece is the problem-independent Newton-Krylov template with the Lorenz system as a runnable example. The exposition of flow maps, equilibria, periodic orbits, relative orbits, Poincaré sections, slicing, JFNK, the hookstep, and the constraints is accurate and pleasantly direct. I checked Sections 1.3 through 3.4 against the standard references it cites and found no hand-waving where it matters. The Lorenz example reproduces Viswanath's orbits from published initial guesses, which is a real, reproducible demonstration that the code works.\n\nThe soft spots are mostly minor. There is no new mathematics, and the paper does not try to hide that; every ingredient is attributed. The finite-difference Jacobian-vector product in (3.8) is presented as a heuristic with a typical epsilon value, with no error analysis and no discussion of when adaptive or event-driven timesteppers will break it. For the target audience—people with smooth, fixed-step timesteppers—this is fine; it is standard Knoll–Keyes practice. A sentence saying 'if convergence stalls, check this epsilon first' would have helped, but its absence is not a flaw in what the paper claims. The blanket statement that chaotic attractors are dense in periodic orbits is a bit breezy, though for the Lorenz system at the standard parameters it is the right intuition and the cited Viswanath computations support it.\n\nOn the reader's take: the identified 'weakest assumption' about (3.8) is real but not close to load-bearing for a tutorial of this scope. The method is standard, the example converges, and the paper explicitly frames the template for smooth large systems such as pipe flow, not for every conceivable timestepper.\n\nWho is this for: graduate students and researchers who already have a timestepper and want invariant solutions without writing a solver from scratch. It deserves a serious referee—not because it breaks new ground, but because its pedagogical clarity and reusable code justify the time. I would not cite it in my own papers; I would cite openpipeflow.org and the original sources. But I would hand it to a student without hesitation.","headline":"A clear, honest tutorial that packages the standard Newton-Krylov machinery for invariant solutions into a reusable code template; it earns referee time as an exposition, not as a research result.","tokens_in":10682,"tokens_out":1867,"would_cite":false,"duration_ms":22839,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Time stepping alone finds periodic orbits via a Jacobian-free Newton method","keywords":["periodic orbits","equilibria","flow map","Jacobian-free Newton-Krylov","GMRES","Lorenz system","travelling waves","symmetry reduction"],"falsifier":"On a system with an analytically known Jacobian, such as the Lorenz equations at the paper's parameters, compare the finite-difference product of equation (3.8) with the exact Jacobian-vector product over the directions $\\delta x$ that arise during a Newton solve; if the relative error exceeds the GMRES tolerance, the step-size rule is the point of failure and the method's convergence claim fails for that system.","tokens_in":9820,"feed_emoji":"🌀","tokens_out":8582,"duration_ms":86110,"temperature":0.7,"pith_summary":"This short exposition sets out a single recipe for finding invariant solutions: write the dynamics as a flow map $\\Phi$, look for fixed points of $F(x) = \\Phi(x) - x$, and solve that equation with a Jacobian-free Newton–Krylov method. The claim is that this recipe needs little more than calls to an existing time-stepping code, because the Jacobian itself is never formed: only finite-difference Jacobian-vector products are required. A sympathetic reader cares because this turns “I can integrate the equations” into “I can find the equilibria, periodic orbits, and travelling waves that organise the dynamics,” with the same problem-independent solver applicable to systems with millions of degrees of freedom. The paper also shows how to adapt the fixed-point equation to Poincaré sections, unknown periods, spatial shifts, and symmetry reduction, with a worked Lorenz-system template.","feed_headline":"Time stepping alone finds periodic orbits via Jacobian-free Newton","feed_subtitle":"The method needs only an existing time stepper, so any simulation code can be searched for hidden orbits.","key_machinery":"The carrying object is the flow map $\\Phi$, which advances a state by a prescribed time, turning equilibria and periodic orbits into fixed points of the map; the residual $F(x)=\\Phi(x)-x$ is the equation the solver attacks. The carrying numerical mechanism is the Jacobian-free Newton–Krylov method: Newton's linear system $(J-I)\\delta x = -F(x)$ is solved by GMRES, and GMRES only needs products of the Jacobian with vectors, each approximated by a finite difference of $F$. Constraints for unknown period and shift are folded into the linear system by augmenting the state and adding orthogonality conditions, and a hookstep/trust-region refinement widens the basin of convergence.","core_discovery":"An equilibrium is a fixed point of the flow map for any time, a periodic orbit is a fixed point for its period $T$, and travelling waves and relative periodic orbits become ordinary equilibria and periodic orbits after a symmetry-reducing slice transformation. The paper's central claim is that all of these can be computed by the Jacobian-free Newton–Krylov method applied to $F(x) = \\Phi(x) - x = 0$, with Newton steps obtained from GMRES using the finite-difference product in place of the exact Jacobian. Because the method requires only calls to $F$, which is just an existing time stepper, the same code can be pointed at new problems by supplying a “step by $T$” routine. The paper demonstrates this on the Lorenz system, using previously computed periodic orbits as checks, and reports that the same approach has found periodic orbits and travelling waves in a 154755-dimensional pipe-flow simulation.","pith_inferences":["The paper leaves implicit that the practical bottleneck is the initial guess: recurrence plots and symmetry reduction supply candidates, but the automated discovery of good starting points is still open and could be combined with systematic recurrence scanning.","A natural extension not discussed is replacing the finite-difference Jacobian-vector product with complex-step or automatic differentiation where available, which could remove the step-size sensitivity that is the method's load-bearing assumption.","The same fixed-point formulation with augmentations could be aimed at other invariant objects such as heteroclinic connections or unstable manifolds, not just equilibria and periodic orbits."],"forward_implications":["Anyone with a working time stepper can search for equilibria and periodic orbits by plugging that stepper into the supplied template; no Jacobian evaluation or storage is needed.","Travelling waves and relative periodic orbits become ordinary equilibria and periodic orbits under Fourier slicing, so the same solver finds them automatically.","Unknown periods and spatial shifts can be handled by augmenting the unknowns with constraints that keep the Newton update orthogonal to the trajectory and to the shift direction.","In chaotic systems, the computed periodic orbits can serve as an organising skeleton: near-recurrences shadow them, and their relative stability organises statistical averages.","The method carries over to large systems; the same code was used to find periodic and travelling-wave solutions of a 154755-dimensional pipe-flow model."],"supporting_citations":[{"why":"Supplies the Fourier slicing procedure that removes continuous translation symmetry so relative solutions become ordinary fixed points.","marker":"Budanur et al., 2015"},{"why":"Demonstrates the same JFNK code finding periodic and travelling-wave solutions in a 154755-dimensional pipe-flow model.","marker":"Willis et al., 2016"},{"why":"Provides the survey of Jacobian-free Newton–Krylov methods on which the exposition's numerical core is based.","marker":"Knoll and Keyes, 2004"},{"why":"Provides the catalogue of Lorenz-attractor periodic orbits used as test data and worked examples.","marker":"Viswanath, 2003"},{"why":"Gives the details of computing the hookstep within GMRES to limit the Newton step.","marker":"Viswanath, 2007"},{"why":"Supplies the trust-region step-size adjustment used to compare predicted and actual reduction.","marker":"Dennis and Schnabel, 1996"},{"why":"Shows the recurrence-plot search for near-recurrences that supplies initial guesses in pipe flow.","marker":"Willis et al., 2013"},{"why":"Provides the three-variable convection model used as the paper's worked example and demonstration template.","marker":"Lorenz, 1963"}],"fun_headline_variants":["Find periodic orbits with just a time stepper","No Jacobian? Just time-step to find orbits","Use your existing simulator to locate orbits","Locate hidden orbits from any time stepper"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the finite-difference approximation of the Jacobian-vector product, with step set so that $(\\varepsilon \\|\\delta x\\|)/\\|x\\|$ is about $10^{-6}$, is accurate enough for GMRES to produce a usable Newton step; if the flow map is insufficiently smooth or the step is badly scaled, the search fails.","fun_headline_variants_meta":{"raw":{"variants":["Find periodic orbits with just a time stepper","No Jacobian? Just time-step to find orbits","Use your existing simulator to locate orbits","Locate hidden orbits from any time stepper"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001046,"raw_usage":{"total_tokens":4338,"prompt_tokens":827,"completion_tokens":3511,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":443,"completion_tokens_details":{"reasoning_tokens":3453}},"tokens_in":443,"tokens_out":3511,"duration_ms":26945,"temperature":1.0,"reasoning_tokens":3453,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:14:20.416262+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a system with an analytically known Jacobian, such as the Lorenz equations at the paper's parameters, compare the finite-difference product of equation (3.8) with the exact Jacobian-vector product over the directions $\\delta x$ that arise during a Newton solve; if the relative error exceeds the GMRES tolerance, the step-size rule is the point of failure and the method's convergence claim fails for that system.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Fourier slicing procedure that removes continuous translation symmetry so relative solutions become ordinary fixed points."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Demonstrates the same JFNK code finding periodic and travelling-wave solutions in a 154755-dimensional pipe-flow model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the survey of Jacobian-free Newton–Krylov methods on which the exposition's numerical core is based."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the catalogue of Lorenz-attractor periodic orbits used as test data and worked examples."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the details of computing the hookstep within GMRES to limit the Newton step."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the trust-region step-size adjustment used to compare predicted and actual reduction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows the recurrence-plot search for near-recurrences that supplies initial guesses in pipe flow."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the three-variable convection model used as the paper's worked example and demonstration template."}],"review_version":1}