{"id":"293ae447-19e2-40e3-a8c4-7cb19529996d","arxiv_id":"1908.01231","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Nearest-neighbor tangent-plane regression on delay maps is claimed to give asymptotically consistent extreme-event predictions, supported by a new disjoint-embedding lemma and an unresolved extrapolation assumption.","lead":"This paper proposes that delay-coordinate maps plus nearest-neighbor local linear regression can predict extreme values in stationary chaotic systems. The argument combines classic embedding theorems with a new lemma, but the central extrapolation step is asserted rather than proved.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing step is the claim that a local tangent-plane fit extrapolates to extremes to first order; Taylor's theorem gives no error control at extreme points far outside the fitted neighborhood, so the central consistency claim is unsupported.","rationale":"The paper's strongest claim is that the embedding/immersion theorem lifts the k-NN local linear fit to an asymptotically consistent extreme predictor. The reader's rejection is justified: the missing piece is a bound on extrapolation outside the neighborhood where the Taylor approximation was fitted. An immersion is a local condition; Sauer et al. guarantee a tangent plane but not a radius of linearity over the whole attractor. The paper explicitly says 'extrapolation correct to the 1st order rather than the 0th order,' but '1st order' is a local rate, not a global error bound. For an extreme event the conditioning vector may lie in a low-density part of the attractor; ergodicity fills typical sets but gives no uniform control of the curvature term over the support. I agree with the reader's weakest-assumption diagnosis. I also note a separate unsupported statement: the claim that SRB measures are absolutely continuous with respect to Lebesgue measure is false for typical dissipative attractors, so the higher-dimensional extension's differentiability argument is not valid as written. Independent support in the paper is thin: it cites Stone's consistency of k-NN regression and Sauer et al.'s embedding theorem, both real, but neither supplies the missing extreme-extrapolation argument. The central theoretical claim therefore remains unproven, and the verdict should stay as the reader's REJECT.","tokens_in":4184,"tokens_out":8283,"duration_ms":98714,"concrete_test":"Use a known chaotic map with an accessible invariant measure and a nonlinear observable, e.g., the Hénon map with delay coordinates p=3. Choose a fixed extreme threshold u at the 99.9th percentile of y=x_{n+1}. For each forecast time t with y_{t+1}>u, fit the paper's k-NN local linear model to the preceding n observations at x_t=(x_t,x_{t-1},x_{t-2}), with k=n^α, α<1, and record the prediction error. Increase n from 10^4 to 10^7. If the median and 90th percentile of errors on these extreme cases do not decrease to zero as n grows, the claimed asymptotic consistency for extremes fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central assertion appears in the construction: 'because of the immersive nature the regression surface is converging to the tangent plane of the mapping, which will provide an extrapolation correct to the 1st order rather than the 0th order.' This is the step that would make extreme-value prediction asymptotically consistent, and it is not justified. Sauer et al.'s theorem guarantees that the delay-coordinate map is an embedding/immersion; it does not bound the difference between the tangent plane at one point and the map value at another point. Taylor's theorem with remainder shows the extrapolation error is at least of order (1/2) sup ||Hessian|| ||x - x0||^2 along an interpolating path. For an extreme y, the antecedent state x may lie outside the convex hull, or in a low-density tail, of the observed neighborhood; the distance ||x - x0|| need not shrink as n grows, and the remainder need not vanish. 'Correct to 1st order' is a statement about infinitesimal neighborhoods; it says nothing about a fixed finite extrapolation distance. Ergodic theorems supply density of observations in typical regions of the attractor, but not a uniform linearization radius over the whole attractor. The paper supplies no rate or bound connecting neighborhood radius, curvature, tail probability, and extreme threshold. Without such a bound, nearest-neighbor linear regression plus an embedding does not imply consistent prediction of 'new extremes'; it implies, at best, consistent local estimation at points with positive limiting density and adequate smoothness. The later appeal to SRB measures also relies on a false premise: SRB measures for dissipative chaotic systems are generally singular with respect to Lebesgue measure, not absolutely continuous, so the claimed preservation of differentiability in the higher-dimensional extension is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper claims to establish asymptotic consistency for the prediction of extreme values in stationary chaotic systems. The proposed method combines delay-coordinate embedding (Sauer et al.) with nearest-neighbor linear regression on the embedded variables. The central argument is that, because the delay-coordinate map is an immersion when the number of delay coordinates p exceeds twice the box dimension of the attractor, a locally fitted linear regression surface converges to the tangent plane of the target map, and therefore extrapolation to extreme values is correct to first order rather than zeroth order. The paper then extends the argument to the case of unknown attractor dimension by considering multiview embeddings, disjoint sampling, and conditional expectations based on SRB measures, and it illustrates the approach with a qualitative discussion of a precipitation prediction exercise.","tokens_in":4542,"tokens_out":3213,"duration_ms":32907,"significance":"If the central claim were rigorously established, the paper would provide a practically relevant bridge between embedding theory and extreme-value prediction in chaotic time series. The paper identifies a real problem and cites the relevant literature from nonlinear dynamics and nonparametric regression. Its strength is the proposal that local tangent-plane information, rather than local averaging, could enable extrapolation to extremes. However, the significance is severely limited because the central mathematical step is not proven, the main assumption about SRB measures is false for typical chaotic attractors, and the extension to unknown dimension rests on an unverified ad hoc assumption. The paper is more of a research proposal or position statement than a worked-out theory.","major_comments":[{"comment":"The load-bearing assertion is the sentence: 'because of the immersive nature the regression surface is converging to the tangent plane of the mapping, which will provide an extrapolation correct to the 1st order rather than the 0th order.' This step is not justified. Sauer et al.'s theorem guarantees that the delay-coordinate map is an immersion, but an immersion only controls the derivative locally; Taylor's theorem with remainder shows that the error in extrapolating from a fitted tangent plane at x0 to a point x is bounded by (1/2) sup||Hessian|| ||x - x0||^2 along a path. For an extreme value of y, the antecedent x may lie far outside the fitted neighborhood, and ||x - x0|| need not shrink as the sample size grows. The paper provides no bound connecting neighborhood radius, curvature, tail probability, and extreme threshold, so the claimed asymptotic consistency for extremes is unsupported.","section":"p. 1-2, construction paragraph"},{"comment":"The paper states: 'Since the SRB measure is absolutely continuous with respect to lebesgue measure, this preserves differentiability on the collapsed system.' This assumption is false for typical chaotic attractors, including the Lorenz, Hénon, and many other physically relevant systems, where the SRB measure is supported on a set of zero Lebesgue measure and is singular with respect to Lebesgue measure. Consequently, the claimed preservation of differentiability of the conditional expectation E(Y|x) after marginalizing over unobserved variables does not follow from the cited theory. The paper does not identify any class of systems for which the assumption holds, and it gives no numerical example satisfying it.","section":"p. 2, SRB measure paragraph"},{"comment":"In the extension to arbitrary p, the paper says: 'Assume the conditional system inherits the properties relevant to Sauer et al's theorem 2.10 with respect to this best approximation.' This is an unproved, ad hoc assumption that is load-bearing for the multiview embedding claim. The paper neither proves that such a best approximation exists with the required regularity nor that the conditional system indeed inherits the relevant properties. Without this step, the argument for the p < 2 boxdim(A) regime, and for the practical recommendation to use multiple p-variable regressions, is a leap rather than a derivation.","section":"p. 3, 'best approximation' assumption"},{"comment":"The proof of Corollary 1 is not rigorous. It invokes Sauer et al.'s theorem 2.7 for the product map (F_i, F_j) and argues that if the two self-intersection sets intersect, then the product map has a self-intersection, contradicting theorem 2.7 because the product dimension exceeds 2d. However, theorem 2.7 requires additional hypotheses (e.g., constraints on periodic orbits), and the proof does not carefully handle the distinction between a self-intersection of the product map and the intersection of the two self-intersection sets. The corollary is used to justify the 'disjoint sampling' multiview approach, so this gap affects a central part of the claimed extension.","section":"p. 2-3, Corollary 1"}],"minor_comments":[{"comment":"The citations are inconsistent: Sauer et al. is sometimes cited as [1] and sometimes as [2], and the text 'some extension [].7' contains a missing reference. The reference list should be carefully updated.","section":"References and citations"},{"comment":"The mathematical notation is heavily garbled, for example the box dimension formula is unreadable in the text: '𝑏𝑜𝑥 𝑑𝑖𝑚(𝐴)=𝑙𝑖𝑚ఌ→଴ቀ௟௢௚(ேഄ)ି ௟௢௚(ఌ)ቁ'. Equations need to be typeset cleanly.","section":"Notation and typography"},{"comment":"The precipitation prediction example refers to a figure that is not included in the manuscript, and the description is too vague to be reproducible. The paper should either include the figure and data-processing details or remove the example as non-essential.","section":"Figure and empirical example"},{"comment":"The sentence 'It turns out we can bound the predictability of the tangent planes a little bit based on a further result of Sauer et al[1], and the some extension [].7' promises a bound that is never stated. Either provide the bound or delete the sentence.","section":"p. 2, 'It turns out' sentence"}],"recommendation":"reject","confidential_remarks":"The manuscript reads as an early draft rather than a complete research paper. The central consistency claim is not derived, and the SRB absolute-continuity assumption is false for typical chaotic attractors. The mathematical gaps are too substantial to be repaired by minor revision, and the core result as stated is not correct at the level claimed. I recommend rejection, though the general research direction might be worth pursuing in a more rigorous framework."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a rough draft of a theory for asymptotically consistent extreme-value prediction in stationary chaotic systems. The one genuinely new piece is Corollary 1: for two strictly distinct delay maps with dimensions between boxdim(A) and 2 boxdim(A), their δ-distant self-intersection sets are disjoint. The proof sketch is plausible and looks correct. That is a real, citable lemma, and it is the part worth keeping.\n\nThe rest is a synthesis of known embedding results (Sauer et al.) and nearest-neighbor regression (Stone), applied to a worthwhile problem. The idea that a local linear fit on delay-coordinate neighbors might extrapolate to extremes is a natural framing, and the paper honestly notes that a simple mean of neighbors only gives 0th-order accuracy.\n\nThe load-bearing step does not hold up. The paper asserts that because the delay map is immersive, the regression surface converges to the tangent plane, and that this 'will provide an extrapolation correct to the 1st order rather than the 0th order.' That is not a consequence of the embedding theorem. An immersive map has an injective derivative at each point, which says nothing about the size of the remainder term when you move a finite distance away from the fitted neighborhood. For an extreme event, the antecedent state can be in a low-density tail of the observed neighborhood, and the extrapolation distance need not shrink as n grows. Without a rate bound tying neighborhood radius, curvature, tail probability, and extreme threshold, the asymptotic consistency claim is unsupported. The stress-test note is right on this.\n\nThe SRB measure argument also rests on a false premise: typical SRB measures for dissipative chaotic systems are singular with respect to Lebesgue measure, not absolutely continuous. So the claim that differentiability is preserved under marginalization is not established.\n\nThe empirical section is explicitly not the proposed method—it uses climate-model ensembles and post-selection—so it provides no validation. The writing is garbled in places, with missing references and unexplained notation, which makes the paper harder to assess than it should be.\n\nWho is this for? Researchers working on the theory of dynamic forecasting or embedding-based prediction. They would find Corollary 1 useful but the main theorem unconvincing. The paper deserves a serious referee because the problem is important and the lemma is new, but the central claim needs either a real proof or a substantial reframing as a conjecture with supporting numerics. If I were asked to referee it, I would engage, but I would not recommend acceptance without major revision.","headline":"A suggestive sketch with one worthwhile lemma, but the central consistency claim for extreme extrapolation is not established.","tokens_in":5034,"tokens_out":1801,"would_cite":false,"duration_ms":20936,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["37M10","37D45","62M10","62G32","62G08"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that delay-coordinate embedding plus nearest-neighbor linear regression makes extreme-value prediction in stationary chaotic systems asymptotically consistent.","keywords":["extreme value prediction","chaotic time series","delay-coordinate embedding","nearest neighbor regression","tangent-plane extrapolation","asymptotic consistency","stationary attractor","SRB measure"],"falsifier":"Run the proposed estimator on a stationary chaotic flow with a known attractor dimension: embed with more than twice the box dimension, fit nearest-neighbor linear regressions, and extrapolate to a future value far outside the sampled range; if the prediction error does not shrink as the observation record and neighbor count grow, the central consistency claim is false.","tokens_in":4000,"feed_emoji":"🌊","tokens_out":12956,"duration_ms":116012,"temperature":0.7,"pith_summary":"This paper tries to establish that extreme values in a stationary chaotic system can be predicted with asymptotic consistency, without any parametric model of extremes. The mechanism is delay-coordinate embedding followed by nearest-neighbor linear regression: when the embedding dimension is more than twice the box dimension of the attractor, the delay map is an immersion, meaning its derivative is one-to-one, so the fitted regression surface converges to the tangent plane of the underlying dynamics. Extrapolating along that tangent plane is first-order accurate, which the paper argues makes prediction of values far outside the observed range consistent as the record grows. For systems of unknown or higher dimension, the paper argues that a unique SRB measure (the natural invariant measure of the attractor) preserves differentiability of the collapsed conditional expectation, and that residual distributions around the fitted planes provide probabilistic predictions of extremes.","feed_headline":"Extreme events in chaotic systems can be predicted consistently","feed_subtitle":"Nearest-neighbor tangent planes turn ergodic attractors into valid extrapolations, not just averages.","key_machinery":"The central object is the delay-coordinate map $F(x)=(h_1(x),\\ldots,h_1(g^{p_1-1}(x)),\\ldots,h_J(x),\\ldots,h_J(g^{p_J-1}(x)))$, which sends a state on the attractor to a vector of time-lagged measurements. The load-bearing machinery is the embedding theorem in [1]: for $p>2\\,\\mathrm{boxdim}(A)$, $F$ is one-to-one and an immersion on smooth submanifolds of the attractor, and for $p\\le 2d$ its self-intersection set has dimension at most $2d-p$. These facts certify that a linear regression over nearest neighbors in the reconstructed space converges to the tangent plane of the underlying map, converting extrapolation from a zero-order average to a first-order linear prediction. The paper's corollary extends the self-intersection bound to two strictly distinct delay maps, and the SRB-measure argument preserves differentiability when variables are integrated out, so residual distributions around the fitted planes carry the probabilistic content for extremes.","core_discovery":"The central claim is that asymptotic consistency of extreme-value prediction is a geometric consequence of embedding, not an assumption about tail distributions. With $p$ time-lagged coordinates as inputs and one future coordinate as the target, the delay-coordinate map is one-to-one and immersive when $p>2\\,\\mathrm{boxdim}(A)$; the paper therefore treats the local regression of the target on the $p$ inputs as converging to the tangent plane of the map. Because the chaotic orbit samples the attractor ergodically and the number of nearest neighbors grows more slowly than the sample size, the tangent-plane extrapolation is claimed to be correct to first order, so a predicted extreme is an asymptotically consistent estimate rather than a nearest-neighbor average. In the intermediate regime where the reconstruction is not a full embedding, the paper proves a corollary that two strictly distinct delay-coordinate maps cannot share a self-intersection point, so their predictions cannot both be wrong at the same location; combined with a unique SRB measure (the attractor's natural invariant measure), the residual distribution around the plane gives a probabilistic forecast of extremes.","pith_inferences":["Editorial extension: if the tangent-plane extrapolation is the right error model, the dominant error for very large extremes should scale with attractor curvature rather than sampling noise, so a curvature-corrected variant is a natural next test.","Editorial extension: the residual-based predictive distribution can be compared with generalized Pareto tail fits on simulated chaotic data; disagreement would show where the geometric and statistical extreme-value pictures diverge.","Editorial extension: the disjoint-intersection corollary suggests a practical diagnostic that two agreeing embeddings identify a nonsingular point, while disagreement means neither tangent-plane forecast should be trusted alone.","Editorial extension: the same local-tangent argument may carry over to slowly non-stationary systems with short-window re-estimation, though the paper's consistency proof is deliberately limited to the stationary case."],"forward_implications":["For a stationary chaotic system with known attractor dimension, choosing an embedding with more than twice that dimension and a neighbor count growing more slowly than the sample size gives forecasts of extremes that improve as the record grows.","No extreme-value parametric model is needed; the same local regression machinery produces a probabilistic forecast from the residuals around the tangent plane when the embedding is incomplete.","Two strictly distinct embeddings in the intermediate dimension range cannot both be wrong at the same point, so disagreement between their predictions flags locations where the tangent-plane approximation is unreliable.","Building many random p-variable subsets from a candidate set of delay coordinates turns a single prediction into a predictive distribution of the extreme response."],"supporting_citations":[{"why":"supplies the embedding theorems: immersion for p>2 boxdim(A) and dimension bounds for self-intersection sets.","marker":"[1]"},{"why":"supplies the ergodic conditions that let long-time observations sample the attractor densely.","marker":"[2]"},{"why":"supplies the nearest-neighbor regression consistency result used to make the local fit converge.","marker":"[5]"},{"why":"supplies the empirical example of useful prediction with reconstruction dimension below the embedding threshold.","marker":"[3]"},{"why":"supplies the multiview embedding result that motivates combining several p-variable regressions.","marker":"[4]"},{"why":"supplies the SRB-measure condition used to preserve differentiability of the collapsed conditional expectation.","marker":"[6]"},{"why":"supplies the disjoint-sampling efficiency argument for multiview prediction.","marker":"[7]"}],"fun_headline_variants":["Geometric extremes: tangent planes yield consistent chaos predictions","Embedding makes extreme forecasts asymptotically unbiased","Chaos extremes: tangent-plane extrapolation beats averaging","From ergodic sampling to consistent extreme-value prediction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on the assumption that a tangent plane estimated from nearest neighbors inside a small observed neighborhood remains an accurate model when extrapolated to an extreme value far outside that neighborhood.","fun_headline_variants_meta":{"raw":{"variants":["Geometric extremes: tangent planes yield consistent chaos predictions","Embedding makes extreme forecasts asymptotically unbiased","Chaos extremes: tangent-plane extrapolation beats averaging","From ergodic sampling to consistent extreme-value prediction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000648,"raw_usage":{"total_tokens":2942,"prompt_tokens":879,"completion_tokens":2063,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":495,"completion_tokens_details":{"reasoning_tokens":2015}},"tokens_in":495,"tokens_out":2063,"duration_ms":14238,"temperature":1.0,"reasoning_tokens":2015,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:19:30.890859+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the proposed estimator on a stationary chaotic flow with a known attractor dimension: embed with more than twice the box dimension, fit nearest-neighbor linear regressions, and extrapolate to a future value far outside the sampled range; if the prediction error does not shrink as the observation record and neighbor count grow, the central consistency claim is false.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the embedding theorems: immersion for p>2 boxdim(A) and dimension bounds for self-intersection sets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the ergodic conditions that let long-time observations sample the attractor densely."},{"cited_title":"Embedology","cited_arxiv_id":null,"evidence_quote":"supplies the nearest-neighbor regression consistency result used to make the local fit converge."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the empirical example of useful prediction with reconstruction dimension below the embedding threshold."},{"cited_title":"This however is not much help in extrapolating to new extremes with new dependent variables","cited_arxiv_id":null,"evidence_quote":"supplies the multiview embedding result that motivates combining several p-variable regressions."},{"cited_title":"Ergodic Theory of chaos and strange attractors, Reviews of modern Physics, 57, 617-656, (1985)","cited_arxiv_id":null,"evidence_quote":"supplies the SRB-measure condition used to preserve differentiability of the collapsed conditional expectation."},{"cited_title":"Prediction in Projection","cited_arxiv_id":null,"evidence_quote":"supplies the disjoint-sampling efficiency argument for multiview prediction."}],"review_version":1}