{"id":"3bb504d6-14f0-4f65-81a3-b0b599ba645e","arxiv_id":"2606.22039","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"A proposed three-domain partition of cardiac-attractor information is not rigorously proven; the empirical validation is modest, partly circular, and shows chance-level per-record performance.","lead":"This paper proposes Attractor Domain Theory, which claims the information in a reconstructed heart-signal attractor splits into three non-redundant domains — geometry, ergodic statistics, and local stretching — each natively supporting a different clinical readout. The authors test the geometry domain for tachypnea detection on PPG data, reporting AUC 0.757 after correcting evaluation leaks, but the proofs are largely heuristic.","discovery_kind":"paradigm_shift","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4's decomposition is internally inconsistent: T_S and T_V are deterministic functions of T_G[x]=X, so conditioning on T_G[x] forces the asserted non-redundancy terms in (24) to be exactly zero; the sufficiency partition fails even before ε≈0 is addressed.","rationale":"The reader's rejection is correct, but the weakest assumption identified there—ε≈0—is not the most fundamental problem. Even granting ε≈0, Theorem 4's decomposition is internally inconsistent because T_S and T_V are defined as functions of X=T_G[x], so conditioning on T_G[x] annihilates their conditional mutual information. This is not an empirical gap; it is an algebraic contradiction with the positivity claims in (24). The paper does contain useful empirical evaluation corrections (record-level CV, train-only normalization, per-record AUC reporting) and the artifact analysis is a genuine contribution. However, the central advertised contribution—a proof that the three domains form a sufficient and non-redundant partition—does not survive scrutiny. The reader's emphasis on ε is partially aligned but too narrow; my concern is upstream. Verdict remains REJECT, hence UNCHANGED from the reader's assessment.","tokens_in":21417,"tokens_out":6993,"duration_ms":70423,"concrete_test":"Analytic check: for any fixed y and any realization x, compute X=T_G[x], A=T_S[X], B=T_V[X,Δn]. Since A and B are deterministic functions of X, H(A|X)=0 and H(B|X,A)=0, so the two conditional mutual information terms in Eq. (23) are identically zero. To make this concrete, simulate a small attractor (e.g., Lorenz), generate a scalar observation x_t=h(Z_t)+ε_t, define a target y=g(Z_t), and estimate all terms in Eq. (23) with fine binning. The conditional terms will be exactly zero in the discrete limit and the claimed positive non-redundancy in (24) cannot appear; the only way to balance the equation is to reinterpret the left side as I[X;y] or to let ε absorb the loss from T_G. This settles whether the decomposition has the form asserted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theorem fails before ε enters. Definitions 1–3 define T_S and T_V as deterministic functions of X = T_G[x]: T_S maps X to a transition matrix/stationary distribution by binning and counting; T_V maps X (and Δn) to an FTLE field. Hence H(T_S[X] | X) = 0 and H(T_V[X,Δn] | X, T_S[X]) = 0, so for any y, I[T_S[X]; y | T_G[x]] = 0 and I[T_V[X,Δn]; y | T_G[x], T_S[X]] = 0 identically. Eq. (23) therefore cannot have the positive terms demanded by (24). If the left side is I[X;y], the equation collapses to I[X;y] = I[X;y] and the 'partition' contributes nothing beyond the first term. If the left side is I[x_n;y] (the observed scalar), the equality is not a chain rule: because T_G is lossy, I[x;y] = I[X;y] + I[x;y|X], and the residual cannot be recovered by functions of X. The proof's Phase 2 argument about infinite-horizon versus finite-window limits describes a conceptual distinction but cannot create conditional dependence for deterministic functions of the conditioning variable. Thus the sufficiency/non-redundancy claim is internally inconsistent, independent of whether ε≈0. This is the load-bearing defect; the unvalidated ε in Remark 6 and the conditional completeness of Theorem 2 are secondary.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Attractor Domain Theory (ADT), claiming that the delay-reconstructed cardiac attractor's information partitions exactly into three mutually non-redundant domains: Geometry (G), Ergodic (S), and Variational (V). Theorem 4 (Eq. 23) asserts a chain-rule decomposition I[X;y] = I[T_G[x];y] + I[T_S[X];y|T_G[x]] + I[T_V[X,Δn];y|T_G[x],T_S[X]] + ε, with the conditional terms claimed strictly positive (Eq. 24). The paper also claims Theorem 2 derives the CSI as the unique Ergodic-Domain functional, Theorem 3 proves affine BP from phase-aggregated FTLE, and Proposition 6 proves {G,S,V} minimal and maximal. The empirical component instantiates T_G as the SCSI index, evaluates tachypnea detection on 176,742 PPG segments, and reports AUC=0.757 after correcting three evaluation artifacts.","tokens_in":21861,"tokens_out":5327,"duration_ms":54627,"significance":"If Theorem 4 were correct, ADT would be a substantial contribution: it would replace feature search in cardiovascular attractor analysis with principled domain identification and would explain cross-modal ECG/PPG transfer asymmetries. The empirical protocol is a genuine strength: record-level GroupKFold, train-only normalization, per-record AUC reporting, explicit artifact quantification, and honest disclosure that only 4/42 CapnoBase records are evaluable. However, the central theoretical claim is internally inconsistent for a fundamental reason: T_S and T_V are deterministic functions of T_G[x]=X, so the conditional mutual-information terms in Eq. (23) are identically zero. The empirical validation, however careful, instantiates only T_G and only for one endpoint (tachypnea), and thus cannot establish the general sufficiency/non-redundancy theorem.","major_comments":[{"comment":"The theorem's core decomposition is internally inconsistent. T_S is defined as a deterministic function of X (binning and counting transitions in X, Definition 2), and T_V is defined as a deterministic function of X and Δn (FTLE from X, Definition 3). Since T_G[x]=X, we have I[T_S[X]; y | X] = 0 and I[T_V[X,Δn]; y | X, T_S[X]] = 0 for any y, because conditioning on X fully determines both quantities. Thus the positivity claim in Eq. (24) is false, identically. The chain rule applied to (T_G[x], T_S[X], T_V[X,Δn]) collapses to I[X;y] = I[X;y], since the joint entropy of the triple equals H(X). The asserted non-redundant partition contributes nothing beyond the first term.","section":"Section IV-E, Eq. (23)-(24); Definitions 2-3"},{"comment":"The proof's Phase 2 cannot repair the collapse. The distinction between the 'infinite-horizon limit' of T_S and the 'finite forward-time window' of T_V is conceptual; operationally, both are functions of the finite trajectory matrix X. Even if one imagined limiting procedures, the limiting object would still be a deterministic function of X, so conditional entropy remains zero. Remark 5's claim that ADT 'avoids this collapse' because T_S and T_V 'impose mutually exclusive physical and temporal limits' is not a mathematical argument: a deterministic function of the conditioning variable cannot produce positive conditional mutual information. The comparison with Fourier/wavelet domains actually highlights the defect, since Fourier coefficients are also deterministic functions of the signal and the conditional terms vanish for the same reason.","section":"Section IV-E, 'Phase 2' and Remark 5"},{"comment":"There is a further ambiguity in the left side of Eq. (23). If the left side is I[X_n;y] with X_n = T_G[x], the equality reduces to I[X;y] = I[X;y] + ε, forcing ε=0. If the left side is instead the scalar I[x_n;y], then T_G is lossy, and the residual I[x;y|X] is not bounded by I[x;ε_n], contrary to the claim after Eq. (23). Phase 1's invocation of DPI for a diffeomorphism only establishes I[Z;y]=I[X;y] when X=Φ(Z) is a bijective reconstruction; it does not generate positive conditional terms for T_S and T_V. The sufficiency claim therefore fails independently of whether ε≈0.","section":"Section IV-E, Eq. (23) and Phase 1"},{"comment":"The paper overstates the scope of its own results. Theorem 2's proof admits that the completeness claim is 'conditional on Axiom 3, not absolute' and that the three invariants are 'canonical representatives' rather than a unique complete set; yet the abstract and conclusion present Theorem 2 as a derivation of the CSI as the unique Ergodic-Domain functional. Proposition 6's maximality is explicitly acknowledged to be 'a classification argument supported by standard dynamical systems theory rather than a derivation.' Remark 6 concedes that Part B ('ε≈0 for the cardiovascular system') is only 'empirically supported' by a single ablation on tachypnea. The central 'necessary and sufficient' claim is therefore not established even setting aside the collapse of Eq. (23).","section":"Theorem 2, Proposition 6, and Remark 6"}],"minor_comments":[{"comment":"The domain assignment of C_NL is inconsistent. Table X labels C_NL as 'G(dominant)', but Section VIII-C says removing it 'quantifies I[T_V;y]', and Remark 3 says C_NL 'encodes T_V-domain local strain, not an Ergodic Domain quantity.' The ablation's interpretation depends on resolving this contradiction.","section":"Table X / Section VIII-C / Remark 3"},{"comment":"Eq. (23) uses I[X_n;y] and I[T_G[x];y] interchangeably. Since x_n denotes the scalar observable and X=T_G[x] the trajectory matrix, the notation should distinguish I[x_n;y] from I[X;y] throughout; the current usage obscures the collapse of the decomposition.","section":"Section IV-E, notation"},{"comment":"The text says SCSI features are 'geometric estimators of the underlying Ergodic Domain invariants', while the framework claims to instantiate only T_G. The relationship between Geometry-Domain estimators and Ergodic-Domain invariants should be stated precisely, since Theorem 2's completeness argument applies to (P,μ), not to finite-sample estimators computed without constructing P.","section":"Section VI-B and Table X"},{"comment":"Theorem 2's 'uniqueness' is also undermined by the free weights w in Eq. (12), which are taken from prior work [3] rather than derived. The statement 'the CSI is the unique bounded monotone functional' requires either a derivation of the weights or a clear statement that uniqueness is up to the empirically fitted weight vector.","section":"Section III, Axiom 3 and Theorem 2"},{"comment":"The CapnoBase external validation AUC=0.621 is based on 4 evaluable records; the paper discloses this, but the abstract and conclusion list it without the same caveat. The main-text abstract states AUC=0.757 and NPV=0.966 without mentioning the 4/42 limitation until Section VIII-B, which may mislead readers.","section":"Section VIII-B / Table IX"}],"recommendation":"reject","confidential_remarks":"The empirical work on SCSI and evaluation artifacts is above average in care and could possibly support a separate applied paper. However, the paper's headline contribution is the Domain Sufficiency Theorem, and that theorem fails at Eq. (23) because T_S and T_V are deterministic functions of T_G[x]. This is not a local fix: restoring a nontrivial partition would require redefining the operators so they are not functions of X, which would change the entire framework. The remaining claims (uniqueness of Theorem 2, necessity/sufficiency of Proposition 6) are also explicitly conditional or heuristic in the body. In my view the manuscript is not salvageable by revision within its stated scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The stress-test note holds up. The central theorem, Theorem 4, is false as stated. T_S and T_V are deterministic functions of X = T_G[x], so conditioning on T_G[x] makes both conditional mutual information terms in (23) exactly zero, not positive. This is not a subtle gap; it is a direct contradiction with Definition 2 and Definition 3. Phase 2's argument about infinite-horizon versus finite-window limits describes a conceptual distinction but cannot create conditional dependence for deterministic functions. The sufficiency partition therefore collapses before ε is even considered.\n\nWhat the paper does well: the evaluation-artifact corrections are genuinely useful. Record-level GroupKFold, train-only normalization, per-record reporting, and the honest CapnoBase limitation (4/42 evaluable) are good practice. The three-domain categorization is a reasonable way to organize the existing catalog of invariants—Lyapunov exponents, recurrence, entropy, FTLE—as a heuristic map, even if it is not a rigorous decomposition.\n\nSoft spots in proportion: the theoretical core fails, so the stated contribution is not met. Beyond the identity issue, Theorem 2's 'unique' is conditional on an unproven axiom, Proposition 6 is admitted to be a classification argument, and Remark 6 defers ε≈0 to empirical support. The empirical support is thin: per-record mean AUC is chance, external validation is on four records, and the Wilcoxon versus CNN is not significant. The predictions are confirmed on the dev set or on the authors' own prior papers, so the circularity burden is real.\n\nIf this were submitted to me, I would tell the authors to separate the artifact-correction work into a standalone empirical paper, and to rewrite the theory as a proposal or taxonomy rather than a theorem. The math is not salvageable as stated. Still, the paper deserves a referee: the empirical methodology is good enough to warrant scrutiny, and the refutation of the theorem is instructive. Send it to peer review, but the verdict should be reject and redirect.","headline":"The evaluation protocol is the best part, but the central theorem is false: the non-redundancy terms are identically zero.","tokens_in":22372,"tokens_out":5816,"would_cite":false,"duration_ms":51593,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["37M10","37C45","94A17"],"pacs":[],"model":"deepseek-v4-flash","headline":"Attractor Domain Theory claims the reconstructed cardiac attractor's information splits exactly into three non-redundant domains—geometry, ergodic statistics, local deformation—making feature selection a matter of domain identification rath","keywords":["attractor domain theory","cardiac attractor","delay embedding","Takens embedding theorem","finite-time Lyapunov exponent","photoplethysmography","domain sufficiency theorem","feature selection"],"falsifier":"A concrete falsifier: take a clinical endpoint not yet tested (sepsis severity, cardiac output, or fluid responsiveness), build the full three-domain feature set, and test whether a model using features outside G, S, V, or the raw waveform's full morphology, beats the three-domain model; equivalently, if removing all three domain representatives does not drop performance to chance, the sufficiency claim fails. More directly, estimate the conditional mutual information I[W;y|T_G,T_S,T_V] for a candidate fourth feature family W; a measurably positive value contradicts Theorem 4.","tokens_in":21215,"feed_emoji":"🫀","tokens_out":13561,"duration_ms":107771,"temperature":0.7,"pith_summary":"Attractor Domain Theory (ADT) claims that everything a wearable photoplethysmography (PPG) or ECG signal can reveal about the cardiovascular system lives in three non-redundant domains of the reconstructed cardiac attractor: the Geometry Domain (the delay-embedded trajectory, which enables artifact rejection), the Ergodic Domain (asymptotic statistics like Lyapunov exponents, recurrence, and entropy, which enable stability estimation), and the Variational Domain (finite-time local expansion rates phase-locked to systole and diastole, which enable hemodynamic inference). The paper proves a Domain Sufficiency Theorem: the total predictive information decomposes by the chain rule of mutual information into the sum of the three domains' contributions plus a remainder bounded by sensor noise, and argues the three domains are minimal and maximal. If correct, the consequence is that choosing features for a cardiovascular endpoint becomes labeling which domain the endpoint belongs to, not searching a feature space. The paper further validates the Geometry Domain with a screening index called SCSI on 176,742 PPG segments, reporting AUC 0.757 and NPV 0.966 after correcting three evaluation artifacts. The open empirical load is the claim that the remainder term is negligible for all clinically relevant targets; the paper's evidence for that is an ablation on one endpoint (tachypnea).","feed_headline":"Three domains carry all cardiac attractor information","feed_subtitle":"Feature selection becomes domain identification: three domains, validated on 176,742 PPG segments.","key_machinery":"The load-bearing mechanism is the sequential operator chain {T_G, T_S, T_V} and the identity of Theorem 4. T_G (Geometry Domain) builds the delay-embedding trajectory matrix X, the faithful Takens reconstruction of the attractor. T_S (Ergodic Domain) turns X into a transition matrix P and stationary distribution μ, yielding the largest Lyapunov exponent, recurrence determinism, and stationary entropy. T_V (Variational Domain) computes a finite-time Lyapunov exponent field λ_n from local deformation gradients and aggregates it over systolic and diastolic phases. The theorem states that these three codomains form a sufficient, mutually non-redundant partition of I[X;y], with remainder bounded","core_discovery":"The central discovery is the Domain Sufficiency Theorem: for any target y, the mutual information between the observed signal and y decomposes as I[T_G;y] + I[T_S;y|T_G] + I[T_V;y|T_G,T_S] + ε, with ε bounded by the information in the acquisition noise. T_G is the delay-embedding trajectory matrix; T_S builds a transition matrix and stationary distribution over binned phase space; T_V computes a finite-time Lyapunov exponent field localized to systolic and diastolic phase. The proof uses Takens' theorem to identify the reconstructed trajectory with the true attractor, then applies the chain rule of mutual information; non-redundancy comes from the operators' mutually exclusive limits (T_S's","pith_inferences":["If the sufficiency claim holds, the same three-domain partition is likely to apply to any bounded dissipative physiological system observed through a scalar sensor, since the argument depends on Takens embedding, ergodic limits, and finite-time deformation rather than cardiac specifics; respiratory or neurological attractors would be natural targets.","The paper's own bound ε ≤ I[x;ε_n] implies that sufficiency degrades with sensor noise; in high-motion wearable conditions a noise-aware version of the decomposition may be needed, and the paper's quality gates can be reinterpreted as a way of keeping ε small.","The ablation recipe—adding a second same-domain feature does not improve a model once the dominant domain representative is present—offers a cheap sparsity test on any PPG dataset: a model with exactly one feature per domain should match a full feature search.","If the domain partition is the right representational unit, then ECG+PPG fusion should be planned per domain rather than per modality: fuse Ergodic estimates, keep Geometry estimates separate, and expect no benefit from fusing two observations of the same domain."],"forward_implications":["Feature selection in cardiovascular attractor analysis stops being a search: the physiological timescale of an endpoint identifies its native domain, and the paper's Bayesian-optimization prediction (short windows for within-beat Geometry targets, long windows for Ergodic targets) becomes a testable design rule.","The ECG-to-PPG transfer asymmetry is explained: ergodic invariants such as the largest Lyapunov exponent transfer across modalities (ρ = 0.559), while geometry-sensitive recurrence does not (ρ = 0.065), because the PPG observation operator corrupts fine orbit structure through peripheral state.","The affine blood-pressure-from-waveform mapping is derived from the Variational Domain: phase-aggregated expansion rates over systole and diastole determine blood pressure up to an affine calibration whose constants are fixed by one reference measurement.","A validated wearable screening use-case follows immediately: the Geometry-Domain instantiation reaches AUC 0.757 and NPV 0.966 on held-out records after correcting three evaluation artifacts, and the corrected protocol establishes an unbiased baseline (AUC 0.573) for future wearable PPG models.","The three-domain partition is minimal and maximal: any fourth feature family either reduces to G, S, or V, or requires information outside the attractor, so redundancy among features becomes detectable in principle."],"fun_headline_variants":["Three domains capture all cardiac attractor info","Attractor info splits into exactly three domains","Cardiac attractor info needs three proven domains","Domain Sufficiency Theorem: three info domains"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The central claim collapses if the remainder term ε in the decomposition is not negligible for some clinically relevant target—that is, if a real cardiovascular quantity is not captured by any combination of Geometry, Ergodic, and Variational features; the paper proves the decomposition identity but supports ε ≈ 0 only with an ablation on tachypnea.","fun_headline_variants_meta":{"raw":{"variants":["Three domains capture all cardiac attractor info","Attractor info splits into exactly three domains","Cardiac attractor info needs three proven domains","Domain Sufficiency Theorem: three info domains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000437,"raw_usage":{"total_tokens":2105,"prompt_tokens":840,"completion_tokens":1265,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":1218}},"tokens_in":584,"tokens_out":1265,"duration_ms":8132,"temperature":1.0,"reasoning_tokens":1218,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T10:35:24.948143+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete falsifier: take a clinical endpoint not yet tested (sepsis severity, cardiac output, or fluid responsiveness), build the full three-domain feature set, and test whether a model using features outside G, S, V, or the raw waveform's full morphology, beats the three-domain model; equivalently, if removing all three domain representatives does not drop performance to chance, the sufficiency claim fails. More directly, estimate the conditional mutual information I[W;y|T_G,T_S,T_V] for a candidate fourth feature family W; a measurably positive value contradicts Theorem 4.","supporting_citations":[],"review_version":2}