{"id":"b4342b12-a0da-40ba-938e-81a7b9e530c0","arxiv_id":"1908.10789","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Applying a variable forgetting factor λ(k)=1+1.02/k to a recursive least squares estimator changes largest Lyapunov exponent estimates, improving accuracy in three of five chaotic systems but worsening it in two.","lead":"This paper tests whether a variable weighting scheme, which trusts early data more than later data, improves estimates of the largest Lyapunov exponent for chaotic systems. Across five test systems, the method was closer to literature values in three cases but increased variability in two.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The three reported accuracy gains are small relative to the reported standard deviations, so the central claim is not statistically supported; the heuristic choice of 1.02 in λ(k)=1+1.02/k adds unquantified selection risk.","rationale":"The central claim is deliberately modest: in five tested systems, three sample-mean estimates moved closer to literature values, with increased variance in two systems. Read literally as a description of Table 2, the claim is true. The load-bearing question is whether the evidence supports the broader inference that variable weighting improves LLE estimation. It does not, for two compounding reasons. First, the reported differences are within the spread of the 100 initial-condition estimates: the largest apparent gain, Hénon, is 0.0086 in mean with σ≈0.085, so the effect is about 0.1 standard deviation; a paired test on 100 paired runs would not reject the null. Second, the constant 1.02 in λ(k)=1+1.02/k was set heuristically on the same five systems, and no sensitivity or out-of-sample check is reported. The reader's conditional verdict already captures the need for these checks, so I keep the verdict unchanged rather than escalating. The proposed concrete check—sweeping c, applying paired inference, and testing held-out systems—directly separates the tuning hypothesis from the error-accumulation hypothesis.","tokens_in":9347,"tokens_out":5553,"duration_ms":61375,"concrete_test":"Recompute Table 2 with λ(k)=1+c/k for c∈{0.5, 0.75, 1.02, 1.25, 1.5} on the same five systems, storing per-initial-condition LLEs so that a paired bootstrap or permutation test can compare uniform and variable-factor mean differences. Then repeat the best c on at least three held-out chaotic systems (e.g., Ikeda map, Lorenz-63, Rössler) using the same rounding-mode protocol. If the 95% bootstrap intervals for Hénon, Sine, and Tent include zero, or if the improvement sign is not stable across c and does not replicate out-of-sample, the central claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's only quantitative evidence is Table 2, which reports means and standard deviations over 100 initial conditions. For each of the three systems where the variable-factor mean moves toward the literature value, the shift is a small fraction of a standard deviation: Hénon moves 0.4088→0.4174 (target 0.4180) with σ≈0.085; Sine moves 0.7487→0.7598 (target 0.7730) with σ≈0.078; Tent moves 0.6898→0.6891 (target 0.6880) with σ≈0.005. In all three cases the effect is ≲0.15σ, so a paired or two-sample test would not detect it, and the two systems that worsen (Logistic and Mackey-Glass) shift by comparable amounts. Thus 'more precise in three systems' is a statement about point estimates, not an established effect. This is compounded by Section 3, where λ(k)=1+1.02/k is described as 'ajustado heuristicamente'; without a sensitivity sweep over the constant or evaluation on systems not used while tuning 1.02, the modest observed shifts may reflect overfitting to these five examples rather than the stated error-accumulation hypothesis.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a small modification to the recursive least squares (RLS) estimator for the largest Lyapunov exponent (LLE) introduced by Peixoto et al. (2018). Instead of weighting all observations equally, the authors introduce a time-varying factor λ(k) = 1 + 1.02/k that gives higher weight to earlier data, motivated by the hypothesis that later data carry more accumulated floating-point error. The method is applied to five chaotic systems (Logistic map, Hénon map, Sine map, Tent map, and Mackey-Glass system). The authors report that, compared with the uniform-weighting RLS estimator, the proposed method gives mean estimates closer to literature values for three systems (Hénon, Sine, Tent) and larger variance for two systems (Tent and Mackey-Glass). The paper concludes that the variable-weighting method is more accurate in three systems and has only small variation in the other two.","tokens_in":9567,"tokens_out":4663,"duration_ms":42240,"significance":"If the claimed improvement were real and robust, the contribution would be a simple, computationally cheap enhancement to an existing LLE estimator, with potential value in applications where only finite-precision implementations are available. The paper has a commendable feature: it reports means and standard deviations over 100 initial conditions for every system and both methods, which allows the reader to assess variability directly. However, as presented, the evidence is not statistically sufficient to support the central claim, and the manuscript contains an internal contradiction about the standard deviations. The work is better read as a preliminary numerical experiment than as an established methodological result; with significant additional analysis it could become publishable.","major_comments":[{"comment":"The text states that 'os valores calculados do desvio padrão são menores quando o fator λ é empregado' (the computed standard deviations are smaller when λ is employed). This is contradicted by the table itself: the standard deviation for the Tent map increases from 0.0051 to 0.0054 and for Mackey-Glass from 0.0020 to 0.0023 when the variable factor is used. The abstract and the conclusion correctly acknowledge increases in variance for two systems, so Section 4's blanket statement is internally inconsistent and should be corrected.","section":"Section 4, Table 2"},{"comment":"The central claim that 'the method obtained more accurate results in three systems' is supported only by point-estimate shifts that are small relative to the reported standard deviations. For Hénon, the mean moves from 0.4088 to 0.4174 (target 0.4180) with σ ≈ 0.085; for Sine, from 0.7487 to 0.7598 (target 0.7730) with σ ≈ 0.078; for Tent, from 0.6898 to 0.6891 (target 0.6880) with σ ≈ 0.005. Each shift is roughly 0.1σ to 0.15σ, and the two systems that move away from the literature values (Logistic and Mackey-Glass) shift by comparable amounts. No significance tests, confidence intervals, or paired comparisons are provided. The observed differences are fully consistent with sampling noise, so the stated conclusion is not statistically supported.","section":"Section 4, Table 2"},{"comment":"The forgetting-factor schedule λ(k) = 1 + 1.02/k is introduced as 'ajustado heuristicamente' (heuristically adjusted), but no sensitivity analysis is given and no validation is performed on systems other than the five used in the comparison. If the constant 1.02 was chosen with knowledge of these systems' literature values, then Table 2 compares the tuned method against the untuned uniform-weighting baseline on the same data, making the comparison partly self-confirming. A sensitivity sweep over the constant (e.g., λ(k) = 1 + c/k for a range of c) and evaluation on at least one system not used during tuning are necessary to establish that the reported improvements are not an artifact of the heuristic choice.","section":"Section 3"}],"minor_comments":[{"comment":"The terms 'accurate' (preciso) and 'precise' are used interchangeably, but the paper's quantitative evidence concerns closeness to literature values (accuracy) and standard deviation (precision). The wording should be aligned with the statistical meaning, especially because the abstract says 'more accurate' while the conclusion emphasizes 'small variation.'","section":"Abstract / Section 4"},{"comment":"In the recursive least squares equations, λ appears as a scalar in the denominator of the gain K_k and as λ(k) in the covariance update P_k. The notation should be made consistent, and the dependence of λ on k should be stated explicitly in all occurrences.","section":"Equation (2)"},{"comment":"The Mackey-Glass system is a delay differential equation, but the manuscript does not state the integration scheme, step size, or time delay discretization used. Without this information, the reported LLE values are not reproducible.","section":"Table 1 / Section 4"},{"comment":"The caption states 'a linha vermelha é o ajuste de mínimos quadrados' (the red line is the least squares fit), but the printed figure appears in grayscale and the line is not clearly identified in panels (f)–(j). The caption or the figure should make the fitted line visually distinct.","section":"Figure 1"},{"comment":"The phrase 'com pequena variação nos outros dois' (with small variation in the other two) minimizes the fact that for the Logistic map and the Mackey-Glass system the mean estimate moves away from the literature value, not just varies. This should be stated explicitly as a deterioration or as a statistically indistinguishable change.","section":"Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is very short and reads more like an extended abstract than a full journal paper. The main issues are statistical: the reported improvements are within the noise of the reported standard deviations, and the heuristic constant is not subjected to sensitivity analysis or independent validation. The internal contradiction about the standard deviations in Section 4 should be caught by the authors before resubmission. I do not see the results as impossible to repair, but the current evidence is insufficient for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper takes Peixoto et al.'s recursive least-squares estimator for the largest Lyapunov exponent and replaces the uniform weighting with a time-decaying factor λ(k)=1+1.02/k, motivated by the idea that early data carry less accumulated computer arithmetic error. That hypothesis is reasonable, the modification is simple, and the authors are unusually honest in reporting the numbers across five chaotic systems, including the two cases where variance went up.\n\nWhat is genuinely new here is the specific schedule and the application to the rounding-mode LLE estimator; the underlying RLS with variable forgetting factor is standard and properly cited. The linking of precision loss to differential weighting is an interesting idea, even if the results do not yet establish it.\n\nThe main problem is that the evidence for \"more accurate in three systems\" is statistically thin. Look at Table 2: Hénon moves from 0.4088 to 0.4174 (target 0.4180) with σ≈0.085, a shift of about 0.1σ. Sine moves from 0.7487 to 0.7598 (target 0.7730) with σ≈0.078, still a fraction of a standard deviation. Tent moves from 0.6898 to 0.6891 (target 0.6880) with σ≈0.005, again under 0.15σ. Meanwhile Logistic goes the wrong way (0.7014→0.7098, target 0.6930) and Mackey-Glass also worsens (0.0074→0.0082, target 0.0074). No significance test or confidence interval is reported, so \"more accurate\" is a statement about point estimates, not an established effect.\n\nThe second issue is the heuristic tuning. λ(k)=1+1.02/k is described as adjusted heuristically, with no sensitivity analysis and no evaluation on systems held out from tuning. Since the same five systems are used both to set 1.02 and to measure the improvement, the already-small gains are partly self-confirming.\n\nI also think the conclusion is slightly spun: it says the other two systems showed \"small variation,\" but the Logistic and Mackey-Glass errors got larger, not just different. No code or data are provided, which limits independent checking.\n\nI would not send this to a serious journal peer review in its current form. The idea could earn a short communication if the authors added a sensitivity sweep over the constant, an out-of-sample test, and a basic significance test or bootstrap CI. As is, it is a credible conference contribution, not a journal-level result.","headline":"A sensible but unsupported incremental tweak to an LLE estimator; the reported gains are within noise and the one free parameter is tuned on the same systems used for evaluation.","tokens_in":10090,"tokens_out":2253,"would_cite":false,"duration_ms":23385,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["37M25","65G50"],"pacs":[],"model":"deepseek-v4-flash","headline":"A variable weighting factor inside a recursive least squares estimator shifts largest Lyapunov exponent estimates closer to literature values in three of five chaotic systems.","keywords":["largest Lyapunov exponent","recursive least squares","variable forgetting factor","chaotic systems","finite precision","computer arithmetic","rounding mode","lower bound error"],"falsifier":"Run the same five-system comparison with $\\lambda(k) = 1 + c/k$ for several values of $c$, and also on a chaotic system not used in this study (for example, the Lorenz system); if no single $c$ outperforms uniform weighting across the new set, or if the winning $c$ differs sharply from 1.02, the claim that variable weighting is broadly beneficial fails.","tokens_in":9166,"feed_emoji":"📈","tokens_out":7283,"duration_ms":63940,"temperature":0.7,"pith_summary":"Computing the largest Lyapunov exponent from finite-precision simulations is sensitive to error that accumulates as iteration proceeds. This paper proposes a variable weighting factor, $\\lambda(k) = 1 + 1.02/k$, inside the recursive least squares estimator so that early, more trustworthy data count more than later data. On five classic chaotic systems, the weighted estimator returns means closer to literature values for the Hénon, Sine, and Tent maps, while being slightly worse for the Logistic and Mackey-Glass maps. The paper's point is that treating all data equally is not the only defensible choice, and that a simple forgetting schedule can improve accuracy, though it also raises variance in some cases.","feed_headline":"Variable weighting improves Lyapunov estimates in 3 of 5 chaotic maps","feed_subtitle":"Early data count more in the recursive fit, beating uniform weighting on Hénon, Sine, and Tent, at a variance cost in two systems.","key_machinery":"The central object is the variable forgetting factor $\\lambda(k) = 1 + 1.02/k$ inserted into the recursive least squares estimator, replacing the fixed denominator gain with a time-dependent weight that gives early samples more influence. It is applied to the natural-log lower bound error (LBE) signal produced by simulating the system twice under different IEEE 754-2008 rounding modes (round-to-nearest and round-toward-plus-infinity). The slope of the least-squares line fit to this error signal is the estimate of the largest Lyapunov exponent, and the paper claims that the variable factor adjusts that slope toward the literature value in most of the tested systems.","core_discovery":"Working from the lower-bound-error approach and the earlier rounding-mode method for computing the largest Lyapunov exponent, the paper argues that the recursive least squares line fit used to extract the exponent should not weight all observations equally. Because floating-point errors accumulate, earlier samples are more precise, so the estimator is modified with a forgetting factor $\\lambda(k) = 1 + 1.02/k$ that decreases toward one as $k$ grows. Applied to the Logistic, Hénon, Sine, Tent, and Mackey-Glass systems, the modified estimator yields average LLE values closer to published values in three of the five systems (Hénon, Sine, Tent) and lower standard deviation in three (Logistic, Hénon, Sine), while standard deviation increases for Tent and Mackey-Glass. The paper presents this as evidence that variable weighting can improve recursive LLE computation, with the caveat that the schedule was chosen heuristically.","pith_inferences":["If the constant 1.02 was selected by trying values on the same five systems, the reported gains may partly reflect tuning rather than a general principle; testing on systems not used in the heuristic search would separate the two.","A natural extension is to make $\\lambda(k)$ depend on an error estimate rather than only on $k$, so systems with slower error accumulation would not down-weight later data as aggressively.","The variance increase for Tent and Mackey-Glass suggests a bias-variance trade-off: the schedule improves the mean but can make individual estimates less stable, which matters when LLE is computed from a single trajectory."],"forward_implications":["For the Hénon, Sine, and Tent maps, the variable forgetting factor shifts the mean LLE estimate closer to the literature value than uniform weighting.","For the Logistic and Mackey-Glass systems, uniform weighting remains closer to the literature value, so the benefit is not universal.","The standard deviation of the LLE estimate falls for Logistic, Hénon, and Sine when the factor is used, but rises for Tent and Mackey-Glass.","The method inherits the simplicity of the rounding-mode approach: it needs no parameterization, embedding dimension, or linearized flow, only the original equations and a chosen forgetting schedule."],"supporting_citations":[{"why":"Supplies the base method: computing LLE from rounding modes and recursive least squares without variable weighting, which this paper modifies.","marker":"Peixoto et al. (2018)"},{"why":"Provides the simple lower-bound-error LLE method that the recursive estimator fits.","marker":"Mendes and Nepomuceno (2016)"},{"why":"Introduces the lower bound error concept used to define the error signal whose slope is the LLE.","marker":"Nepomuceno and Martins (2016)"},{"why":"Classic Lyapunov exponent estimation framework whose literature values are used as comparison targets.","marker":"Wolf et al. (1985)"},{"why":"Alternative robust LLE estimator whose reported values serve as comparison targets.","marker":"Kantz (1994)"},{"why":"Textbook derivation of the recursive least squares estimator with forgetting factor that the paper adapts.","marker":"Aguirre (2015)"},{"why":"Prior use of variable forgetting factor in recursive least squares that motivates the time-varying $\\lambda(k)$.","marker":"Beza and Bongiorno (2014)"}],"fun_headline_variants":["Variable weighting boosts Lyapunov accuracy in 3 of 5 systems","Recursive LLE with forgetting factor: better in some, worse in others","Weighted early data improves Lyapunov estimates in three chaotic maps","Adaptive weighting for Lyapunov exponents: gains in three systems","Not all points equal: variable weight LLE helps, with variance cost"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the constant 1.02 in the forgetting schedule $\\lambda(k) = 1 + 1.02/k$ is a sound choice; the paper states it was adjusted heuristically, so if it was tuned on the same five systems used for evaluation, the observed improvements may reflect fitting rather than a general rule.","fun_headline_variants_meta":{"raw":{"variants":["Variable weighting boosts Lyapunov accuracy in 3 of 5 systems","Recursive LLE with forgetting factor: better in some, worse in others","Weighted early data improves Lyapunov estimates in three chaotic maps","Adaptive weighting for Lyapunov exponents: gains in three systems","Not all points equal: variable weight LLE helps, with variance cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000549,"raw_usage":{"total_tokens":2592,"prompt_tokens":887,"completion_tokens":1705,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":1620}},"tokens_in":503,"tokens_out":1705,"duration_ms":12446,"temperature":1.0,"reasoning_tokens":1620,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:34:57.005015+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same five-system comparison with $\\lambda(k) = 1 + c/k$ for several values of $c$, and also on a chaotic system not used in this study (for example, the Lorenz system); if no single $c$ outperforms uniform weighting across the new set, or if the winning $c$ differs sharply from 1.02, the claim that variable weighting is broadly beneficial fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the base method: computing LLE from rounding modes and recursive least squares without variable weighting, which this paper modifies."},{"cited_title":"and Nepomuceno, E.G","cited_arxiv_id":null,"evidence_quote":"Provides the simple lower-bound-error LLE method that the recursive estimator fits."},{"cited_title":"and Martins, S.A.M","cited_arxiv_id":null,"evidence_quote":"Introduces the lower bound error concept used to define the error signal whose slope is the LLE."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Classic Lyapunov exponent estimation framework whose literature values are used as comparison targets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Alternative robust LLE estimator whose reported values serve as comparison targets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Textbook derivation of the recursive least squares estimator with forgetting factor that the paper adapts."},{"cited_title":"and Bongiorno, M","cited_arxiv_id":null,"evidence_quote":"Prior use of variable forgetting factor in recursive least squares that motivates the time-varying $\\lambda(k)$."}],"review_version":1}