{"id":"8a8f938e-7278-4010-8032-ae15b2851060","arxiv_id":"2608.05882","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"EEG activity is more stable and lower-dimensional in Alzheimer's disease and mild cognitive impairment, while healthy aging shows the opposite pattern of expansion and reduced stability.","lead":"This paper measures how similar brain activity patterns remain over time across EEG recordings, using Wasserstein distance between windowed activity distributions and intrinsic dimensionality as a complexity measure. The two measures move in opposite directions in healthy aging versus Alzheimer's disease, which the authors propose as a biomarker-relevant axis for cognitive decline.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim that intrinsic dimensionality governs stability is partly circular: the paper's own Weed-Bach bound makes larger Wasserstein distance a mathematical consequence of higher dimension, so the observed coupling is not independent evidence; group differences must be checked against…","rationale":"Reader's verdict is reasonable; I agree conditionality is appropriate, but the most load-bearing problem is slightly different from the reader's emphasis on demographic/artifact confounds. The paper's own mathematical appendix makes the main coupling prediction unavoidable: once windows are modeled as iid draws from a distribution, the expected two-sample Wasserstein distance is controlled by the same effective dimension that the Levina-Bickel estimator approximates. Therefore the central empirical 'association' between dimensionality and stability does not validate a neural geometric constraint; it validates the sampling theorem. That does not destroy the clinical observations, but it changes what they mean: the aging and disease effects on both measures could be downstream of a single latent variable such as signal variance or spectral content. The surrogate tests proposed are decisive because phase randomization preserves spectral power but removes any representational geometry, whereas variance-matched white noise tests whether amplitude variability alone suffices. If the topographies and group differences survive both surrogates, the interpretation is strengthened; if not, the conditional verdict should be revised downward. Given the available evidence, CONDITIONAL (as the reader said) remains the appropriate verdict pending those controls.","tokens_in":13203,"tokens_out":5077,"duration_ms":54419,"concrete_test":"For each channel, generate surrogate EEG by randomizing the Fourier phases of the 1-s windows (preserving each window's power spectrum while destroying cross-window temporal structure), and independently generate amplitude-matched Gaussian white-noise surrogates with the same per-window variance. Recompute Levina-Bickel intrinsic dimensionality and the within-condition Wasserstein distance on these surrogates with the same K=50 protocol. If the posterior-anterior gradients, the age-related increases, or the HC/MCI/AD reductions (Figs. 3-5) are reproduced by either surrogate family, then the reported 'representational expansion/collapse' is a property of generic signal statistics, not of neural geometric organization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Appendix A.5 states Theorem A.1 (Weed-Bach): for n samples from a distribution with upper Wasserstein dimension d*, E Wp(mu, mu_hat_n) <= C n^{-1/s}. The paper's within-condition Wasserstein distance is computed between two empirical distributions drawn from the same recording; its expected value is bounded by 2 C n^{-1/s}, so it must increase with intrinsic dimension under the null that windows are iid draws. Thus the across-region and across-condition correlations in Figs. 3-5 are not independent evidence that 'richer representational spaces are less reproducible'; they are partly a finite-sample-statistics consequence of the estimator definitions. The aging/MCI/AD dissociation could be produced by any factor that changes the scale, smoothness, or noise level of the single-channel amplitude distribution - e.g., spectral slowing, reduced signal-to-noise ratio, or medication - without any change in a neural 'representational region'. The manuscript does not report age/education/medication balances for HC/MCI/AD (Section 3.4), does not provide phase-randomized or amplitude-matched surrogate controls, and does not compare against a simple variance or spectral-power baseline. Consequently, the interpretation of reduced Wasserstein distance in AD as 'pathological over-stability' rather than as a generic change in signal statistics is not uniquely supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a distribution-level framework for EEG analysis in which each channel's activity is represented as an empirical distribution of windowed patterns, stability is quantified by Wasserstein distance between time-segment distributions, and representational complexity is quantified by intrinsic dimensionality. Using three datasets (multi-task, lifespan, and clinical HC/MCI/AD), the authors report that within-condition Wasserstein distance is bounded and condition-specific, that higher intrinsic dimensionality is associated with larger Wasserstein distances across regions and conditions, that healthy aging increases both measures, and that MCI and AD decrease both measures. The paper interprets these joint changes as healthy representational expansion versus pathological collapse/over-stability.","tokens_in":13351,"tokens_out":5100,"duration_ms":54713,"significance":"If the interpretation is valid, the framework offers a clinically relevant, easily computable biomarker candidate for cognitive aging and neurodegeneration, and it provides a unified geometric axis linking dimensionality to stability. The study's strengths include the use of multiple independent EEG datasets, explicit theoretical grounding in empirical-process results, and robustness checks across neighborhood sizes K. However, the central empirical association between dimensionality and Wasserstein distance is, to a substantial degree, a mathematical consequence of the very convergence bound the authors invoke, and the clinical group comparisons lack essential demographic and medication controls. These issues are load-bearing because they directly affect the paper's main claims, so the significance is conditional on addressing them.","major_comments":[{"comment":"The reported positive association between intrinsic dimensionality and within-condition Wasserstein distance is largely a finite-sample-statistics consequence of the cited Weed-Bach bound, not an independent empirical discovery. For two independent empirical samples of size n from the same distribution μ, the triangle inequality gives E Wp(μ̂_n, ν̂_n) ≤ 2 C n^{-1/s} with s = d*_p(μ)+ε, so the expected within-condition distance must increase with the upper Wasserstein dimension even under the null of i.i.d. sampling. Since the paper does not report the number of windows per segment, the quantitative predictions of this bound cannot be evaluated. To make the 'geometric constraint' claim non-circular, the authors should compare the observed Wasserstein distances against a surrogate null that holds the estimated dimension fixed (e.g., phase-randomized or amplitude-matched data), or explicitly test whether the empirical dimension–stability relationship exceeds the n^{-1/s} scaling predicted by the theorem.","section":"Section 2.1.4 and Appendix A.5, Theorem A.1"},{"comment":"The clinical comparisons between HC, MCI, and AD do not report or statistically control for age, sex, education, medication, or comorbidities, despite Section 3.3 showing that age alone increases both intrinsic dimensionality and Wasserstein distance. Because AD patients are typically older than healthy controls, the observed 'collapse' of both measures in MCI/AD could be confounded by age, medication effects, or other clinical variables, and the direction of the age effect makes simple intuition unreliable. The authors should report the demographic and clinical characteristics of each group and repeat the group comparisons with age as a covariate or with age-matched subgroups.","section":"Section 3.4 and Section 2.1"},{"comment":"The paper asserts that temporal dependence does not undermine the theoretical link, but Theorem A.1 assumes independent samples, whereas consecutive 1-second EEG windows are strongly autocorrelated. The effective number of independent samples is likely far smaller than the number of windows, which can change the constant and effective rate in the bound and may itself differ across age or disease groups. The authors should quantify the effective degrees of freedom (e.g., via block bootstrap or spectral estimates) or provide surrogate analyses that preserve autocorrelation, to confirm that the dimension–stability relationship is not an artifact of group differences in temporal dependence.","section":"Appendix A.7"},{"comment":"The empirical estimator of intrinsic dimensionality is the Levina-Bickel nearest-neighbor MLE, while the theoretical bound in Theorem A.1 concerns the upper Wasserstein dimension defined through covering numbers. The manuscript states that the MLE is a 'proxy' but does not establish any formal connection between the two quantities for finite, noisy, temporally dependent EEG samples. Without such a link, invoking Theorem A.1 to interpret the empirical correlations is an unproven assumption; the authors should either prove or cite a quantitative relationship, or soften the theoretical justification accordingly.","section":"Sections 2.1.2 and Appendix A.4"}],"minor_comments":[{"comment":"The text says participants were grouped according to the '2011 NIA-AA criteria' but cites Jack Jr et al. (2024), which describes revised criteria; please clarify which criteria version was used and update the citation or text accordingly.","section":"Section 2.1"},{"comment":"The panel labels contain unusual spacing artifacts, e.g., 'Te m p o r a lFluctuations' and 'AgingAndNeurodegenerativeDisease'; please typeset the figure text properly.","section":"Figure 1"},{"comment":"The posterior–anterior gradient is described qualitatively from topographies; providing quantitative summaries (e.g., effect sizes or cluster statistics) would strengthen the claim of reproducibility across conditions.","section":"Section 3.2"},{"comment":"The definition d_ε(μ,τ) = log N_ε(μ,τ) / (−log ε) is only positive for ε < 1; please state the range of ε for which this definition is intended, and clarify how it behaves as ε → 0.","section":"Appendix A.2, Definition 3"},{"comment":"The description of window standardization ('after standardization') is ambiguous: clarify whether each window is standardized independently or each channel is standardized globally across the recording, as this affects the interpretation of Wasserstein distances.","section":"Section 2.1.1"}],"recommendation":"major_revision","confidential_remarks":"The central concern is circularity: the paper's key dimension–stability coupling may be an inevitable consequence of the Weed-Bach bound it cites. The clinical comparison also lacks essential controls. These are fixable with surrogate analyses and demographic adjustment, but until then the main claims are not uniquely supported. The manuscript fit to the journal's scope is good, but the novelty of the 'expansion vs. collapse' framing depends on resolving these issues."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper has one genuinely new empirical result: healthy aging pushes EEG intrinsic dimensionality and within-condition Wasserstein distance up together, while MCI and AD push both down. That dissociation, if it survives controls, could be a cheap clinical tracking signal. The framework is clearly described, the three datasets give the claim breadth, and the sensitivity analyses across neighborhood sizes K are a real mark of care. I also credit the bounded-drift analysis and the within- versus cross-condition separation as sensible checks.\n\nBut the central interpretation is softer than the abstract suggests. The paper's own Theorem A.1 (Weed–Bach) makes larger Wasserstein distance a mathematical consequence of higher intrinsic dimension: with fixed sample size, the expected distance between empirical measures grows with the upper Wasserstein dimension. Since both measures are estimated from the same EEG windows, the across-region and across-condition correlations in Figures 3–5 are largely a finite-sample-statistics effect, not independent evidence that \"richer representational spaces are less reproducible.\" The paper never confronts this circularity head-on; it states the theorem and then treats the empirical coupling as a discovery.\n\nThat said, not everything reduces to the theorem. The aging versus disease dissociation is not forced by the math—it depends on the group differences in the estimated dimension. But those differences are exactly where the clinical analysis is weakest. The HC/MCI/AD comparison in Section 3.4 does not report age, education, or medication balance, and the groups are presumably not age-matched. No artifact or signal-quality metrics are compared across groups, so spectral slowing or reduced SNR could plausibly produce the same pattern without any change in a neural \"representational region.\" The authors mention ICA and manual inspection, but that is not a quantitative control. They also do not compare against simpler baselines like variance or spectral power, so it is unclear what the distributional measures add.\n\nTwo smaller issues: the Weed–Bach theorem assumes independent samples, and the Appendix's hand-wave about weak temporal dependence is not a proof. And neither code nor the lifespan/clinical data are public, which limits reproducibility beyond the multi-task dataset.\n\nWho is this for? Researchers working on EEG complexity biomarkers will want to know about the dissociation, but they should treat the geometric interpretation cautiously until the clinical comparisons are age-adjusted and noise-controlled. The paper deserves a serious referee; a good one will ask for those analyses and for a clear statement that the dimension–stability coupling is a mathematical consequence rather than a neural finding.","headline":"Large EEG study reports an aging/disease dissociation in dimensionality and Wasserstein stability, but the central dimension–stability coupling is partly a finite-sample consequence of the estimator definitions and the clinical comparisons need age and noise controls.","tokens_in":13995,"tokens_out":2043,"would_cite":false,"duration_ms":23916,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Healthy aging expands the brain's representational space; MCI and Alzheimer's disease collapse it.","keywords":["EEG","intrinsic dimensionality","Wasserstein distance","neural stability","healthy aging","mild cognitive impairment","Alzheimer's disease","representational geometry"],"falsifier":"A concrete test is to repeat the clinical comparison with an age- and education-matched healthy control group, keeping the same ICA artifact-removal pipeline and blind automated rejection; the collapse claim would be refuted if lower intrinsic dimensionality and lower Wasserstein distance in MCI/AD disappear once confounds are controlled, and strongly supported if they survive. A second check is computational: simulate EEG noise with known intrinsic dimension and confirm the estimators recover the prescribed dimensionality differences; if they cannot, the group differences could be noise artifacts.","tokens_in":12877,"feed_emoji":"🧠","tokens_out":5641,"duration_ms":55687,"temperature":0.7,"pith_summary":"This paper tries to establish that the brain under a fixed cognitive state does not drift freely but keeps re-entering a condition-specific region of its activity space, and that how stable that region looks over time is governed by how many effective dimensions the activity occupies. If true, the same two numbers—intrinsic dimensionality and Wasserstein distance between time-segment distributions—would describe both normal aging and neurodegeneration on one axis. Healthy aging appears to push representations into a higher-dimensional, less reproducible regime; mild cognitive impairment and Alzheimer's disease push them into a lower-dimensional, abnormally rigid regime. This matters clinically because it turns the question 'is the signal more or less variable?' into 'is the brain's representational geometry expanded or collapsed?', which is a more specific target for tracking cognitive decline.","feed_headline":"Aging stretches brain activity; dementia shrinks it","feed_subtitle":"EEG geometry links richer neural codes to less stability, a possible early signpost for MCI and Alzheimer's.","key_machinery":"The central object is the empirical distribution of EEG activity patterns: per channel, per subject, per condition, the EEG is cut into non-overlapping 1-second windows, each a 100-dimensional vector, and the set of windows is treated as a sample from an underlying distribution. Two descriptors are computed on these distributions. Intrinsic dimensionality, estimated with a maximum-likelihood local estimator, is the effective number of degrees of freedom in the activity patterns. Wasserstein distance, computed by optimal transport, is the minimal transport cost between empirical distributions from different time segments; smaller values mean the same patterns keep reappearing. The link between them is supplied by the empirical Wasserstein convergence theorem, which says the expected distance between independently sampled empirical measures shrinks at a rate controlled by the upper Wasserstein dimension of the underlying measure. That theorem is the load-bearing identity: it makes dimensionality a predictor of stability before any neuroscience assumption is added.","core_discovery":"The central claim is that neural representations, measured as empirical distributions of 1-second multichannel EEG windows, show constrained condition-specific stability rather than unconstrained drift: within-condition Wasserstein distances stay bounded, do not grow with time, and are smaller than cross-condition distances. The paper further claims that intrinsic dimensionality—the effective number of directions in which the windowed patterns vary—constrains stability in a quantitative way, through the empirical-Wasserstein convergence bound $E W_p(\\mu,\\hat\\mu_n) \\le C n^{-1/s}$, where $s$ is tied to the upper Wasserstein dimension of the underlying distribution. Lower-dimensional distributions therefore yield smaller expected Wasserstein distances between repeated samples. Across multi-task, lifespan, and clinical datasets, the authors find a consistent positive coupling: regions and conditions with higher intrinsic dimensionality show larger within-condition displacement. The clinical result is the sharpest claim: healthy aging expands the representation (higher dimension, lower stability), while MCI and AD show a joint collapse (lower dimension, lower Wasserstein distance), which the paper interprets as pathological over-stability rather than preserved function.","pith_inferences":["A natural next test is longitudinal: does an individual's dimensionality first rise with age and then fall as MCI converts to AD? The paper's cross-sectional design leaves that trajectory open, but its shared geometric axis predicts exactly that ordering.","The scalp-level gradient could be checked against source-reconstructed EEG or MEG; if posterior regions remain higher-dimensional with larger Wasserstein distances after source modeling, the result is less likely to be an artifact of volume conduction.","Because the clinical groups are compared without age matching, a direct extension is to re-run the HC–MCI–AD comparison against age-, education-, and medication-matched controls; the collapse interpretation would be substantially strengthened if it survives, and weakened if it does not."],"forward_implications":["Within a fixed cognitive state, the brain's activity-pattern distribution is bounded and condition-specific; progressive representational drift is not the dominant mode of dynamics in these EEG datasets.","A channel's intrinsic dimensionality predicts its temporal reproducibility: richer representational spaces are less reproducible, and this coupling survives across tasks, age groups, and clinical groups, though with reduced spatial extent in AD.","Healthy aging can be described as representational expansion: both dimensionality and within-condition Wasserstein distance increase with age and are positively correlated across channels.","MCI and AD can be described as representational collapse: both measures drop relative to healthy controls, and the paper argues the lowered Wasserstein distance should be read as pathological over-stability, not healthy maintenance.","The two measures together give a single geometric axis on which aging and neurodegeneration move in opposite directions, which could support longitudinal tracking in the clinic."],"supporting_citations":[{"why":"Supplies the maximum-likelihood intrinsic-dimension estimator that defines the paper's complexity measure.","marker":"Levina and Bickel (2004)"},{"why":"Provides the empirical-Wasserstein convergence bound that formally connects intrinsic dimensionality to stability.","marker":"Weed and Bach (2019)"},{"why":"Provides the optimal-transport solver used to compute Wasserstein distances.","marker":"Flamary et al. (2021)"},{"why":"Supplies the multi-task EEG dataset used for within- and cross-condition comparisons.","marker":"Wang et al. (2022)"},{"why":"Supplies the NIA-AA criteria used to classify HC, MCI, and AD in the clinical dataset.","marker":"Jack Jr et al. (2024)"},{"why":"Frames representational drift, the alternative account the paper argues against.","marker":"Driscoll et al. (2022)"},{"why":"Provides the representational-geometry framing that motivates treating activity patterns as distributions in a neural space.","marker":"Kriegeskorte and Kievit (2013)"}],"fun_headline_variants":["Healthy aging expands neural codes; dementia collapses them","EEG geometry: aging widens complexity, dementia narrows it","Neural codes: aging increases dimensionality, disease collapses both","Aging expands brain geometry; dementia flattens it"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that group differences in Wasserstein distance and intrinsic dimensionality reflect the geometry of neural representations themselves rather than differences in age, education, medication, or EEG signal quality across the compared groups.","fun_headline_variants_meta":{"raw":{"variants":["Healthy aging expands neural codes; dementia collapses them","EEG geometry: aging widens complexity, dementia narrows it","Neural codes: aging increases dimensionality, disease collapses both","Aging expands brain geometry; dementia flattens it"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000466,"raw_usage":{"total_tokens":2331,"prompt_tokens":957,"completion_tokens":1374,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":1316}},"tokens_in":573,"tokens_out":1374,"duration_ms":11469,"temperature":1.0,"reasoning_tokens":1316,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T21:44:39.987051+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test is to repeat the clinical comparison with an age- and education-matched healthy control group, keeping the same ICA artifact-removal pipeline and blind automated rejection; the collapse claim would be refuted if lower intrinsic dimensionality and lower Wasserstein distance in MCI/AD disappear once confounds are controlled, and strongly supported if they survive. A second check is computational: simulate EEG noise with known intrinsic dimension and confirm the estimators recover the prescribed dimensionality differences; if they cannot, the group differences could be noise artifacts.","supporting_citations":[],"review_version":1}