{"id":"30d0670a-e7b8-477f-9d0f-467906db3b26","arxiv_id":"2411.10377","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SynGait generates synthetic hip-rotation gait data from small quaternion time-series datasets, reporting good fidelity and geometry preservation on 27 multiple sclerosis patients.","lead":"The paper introduces a pipeline that turns small sets of real hip-rotation recordings, stored as unit quaternion time series, into larger sets of realistic synthetic gait recordings for multiple sclerosis patients. This matters because clinical datasets are small and privacy-restricted, and synthetic data could let researchers develop and test gait analysis tools without sharing patient records.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"All fidelity/geometry metrics are computed on the score matrix, and synthetic scores are convex combinations of real neighbors by Eq. (8); no independent QTS-space or clinical check supports the claim of preserving gait geometry.","rationale":"The paper presents a coherent and useful pipeline: tangent-space fPCA is a principled way to reduce QTS to tabular scores, and the flexible avatar method is a reasonable privacy-oriented generator. The reader's verdict of CONDITIONAL is appropriate. My stress-test focuses on the single most load-bearing premise: that score-space fidelity, as measured in Section 4.2, transfers to the reconstructed unit QTS. Because Eq. (8) generates each synthetic score vector as a convex combination of the nearest real score vectors, the synthetic scores are guaranteed to be close to the original scores in Euclidean space, making k-NNG, RV, and SDV metrics on F partly self-confirming. The paper's own observation that synthetic curves occupy a smaller space (Section 5) is consistent with this shrinkage and indicates reduced variability that is not captured by the average-score metrics. The absence of any evaluation on the reconstructed QTS, or of any clinical/biological plausibility check, means the central claim of high-fidelity synthetic gait data is not yet independently supported. This is addressable: the proposed concrete test computes the same graph and distance metrics directly on the QTS using the geodesic distance, which would reveal whether the geometry preservation is real or an artifact of score-space evaluation. If the direct QTS metrics match the score-space results, the concern does not land; if they diverge, the paper needs to temper its claims or add a post-processing step. Either way, the verdict remains CONDITIONAL because the issue is empirical and fixable rather than a fundamental flaw in the framework.","tokens_in":17185,"tokens_out":7462,"duration_ms":77147,"concrete_test":"Evaluate the same geometry metrics directly on the reconstructed unit QTS rather than on the score matrix. For each of the 100 synthetic datasets, compute (i) the k-NNG on the QTS using the geodesic distance d_g(q1,q2)=||log(q1^{-1}q2)|| integrated over the 101-point grid, and (ii) the mean pairwise geodesic distance between original and synthetic QTS sets, normalized by the original mean pairwise distance. If the QTS-space Frobenius distances are substantially larger than the score-space values (e.g., >50% increase) or the normalized mean pairwise distance shrinks by more than ~30%, the score-space metrics are not a reliable proxy and the geometry-preservation claim fails. A secondary check: have a clinician blind-rate synthetic IGPs as plausible/implausible compared with real IGPs.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that SynGait 'produces high-fidelity synthetic QTS data' while respecting 'the initial geometry of the data' (Abstract; Section 3). The supporting evidence in Section 4.2 and Section 5 is computed exclusively on the functional score matrix F: k-NNG Frobenius distances, RV coefficient, SDV metrics, and local cloaking/hidden rate all use rows of F. But Eq. (8) defines each synthetic score vector as a convex combination of the 𝛾 nearest real score vectors, so the synthetic scores live inside the convex hull of the original scores. Metrics such as k-NNG and RV are therefore partly entailed by the generative process rather than measuring fidelity of the reconstructed QTS. The paper even observes in Section 5 (Figure 5) that synthetic curves 'seem to be living on a smaller space,' a shrinkage that is a direct consequence of convex averaging. No experiment tests whether the reconstructed exp(log-QFD) curves (Eq. 6) are clinically plausible IGPs or lie on the true gait manifold; visual inspection of q_x, q_y, q_z is not a quantitative validation. Unless closeness in score space is shown to transfer to the QTS space, the paper's headline claim of high-fidelity synthetic gait data is not established by the reported metrics.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SynGait, a two-part framework for generating synthetic unit quaternion time series (QTS) gait data. The first part transforms QTS into a tabular score matrix via Fréchet centering, logarithm map to the tangent space, B-spline smoothing, and multivariate functional PCA. The second part synthesizes new scores using a flexible avatar method: for each observation, the γ nearest neighbors are found in the first τ principal-component scores, weights are sampled from a Dirichlet distribution whose concentration parameters depend on distances to those neighbors, and the new score is a weighted average of the neighbors' scores (Eq. 8). Synthetic QTS are reconstructed by exponentiating the linear combination of principal functions and adding back the mean QTS (Eq. 6). The method is applied to 27 multiple sclerosis patients' gait data and compared with copula and CTGAN baselines using k-nearest-neighbor graph Frobenius distances, the RV coefficient, SDV fidelity metrics, and local cloaking/hidden rate; all evaluation metrics are computed on the score matrix.","tokens_in":17441,"tokens_out":3316,"duration_ms":32276,"significance":"If the claims are supported, the framework would be a valuable contribution: it adapts manifold-aware functional data analysis to the synthesis of rotation-valued time series, works with very small samples, and could enable privacy-preserving sharing and stability analysis of gait data. The geometric treatment of unit quaternions and the use of fPCA on the tangent space are appropriate and well motivated. The method is simple, does not require large training datasets, and the comparison with standard tabular synthesizers is a useful benchmark. However, the current evaluation is not sufficient to establish the central claims of high-fidelity QTS generation and geometry preservation, because the metrics are computed in the same score space in which the synthetic data are constructed as convex combinations of real neighbors.","major_comments":[{"comment":"All fidelity and geometry metrics are computed on the functional score matrix F, as stated at the start of Section 4.2. However, Eq. (8) generates each synthetic score vector as a convex combination of the γ nearest real score vectors. Consequently, the k-NNG Frobenius distances (Section 4.2.1) and the RV coefficient (Section 4.2.2) measure similarity in the very space where the generative process is defined, so high scores are partly entailed by construction and do not independently confirm that the reconstructed QTS (Eq. 6) preserve the geometry of the original gait manifold. Please add quantitative validation directly on the reconstructed QTS space, for example by computing geodesic distances between original and synthetic QTS, or by comparing clinically meaningful gait descriptors derived from the QTS (e.g., range of motion, timing of gait phases).","section":"§4.2 and §5, Eq. (8)"},{"comment":"The hyperparameters α0, γ, and τ are selected by the procedure in Section 3.4 using the same 27-patient dataset that is later used to report all fidelity and geometry metrics in Section 5. This selection-on-test-set protocol risks overfitting the reported performance to the particular sample. Please provide a resampling-based evaluation, such as leave-one-out or a train/validation split, to demonstrate that the method's performance generalizes beyond the tuning sample.","section":"§3.4 and §5"},{"comment":"The paper acknowledges in Section 5 that the synthetic curves 'seem to be living on a smaller space,' which is a direct consequence of the convex averaging in Eq. (8). This shrinkage implies that the synthetic QTS may not cover the full variability of the original gait patterns, but no quantitative QTS-space metric is provided to assess the impact of this shrinkage on fidelity. The claim that SynGait produces 'high-fidelity synthetic QTS data' is therefore supported only by score-space metrics, not by any direct evaluation of the reconstructed unit quaternion time series.","section":"§5, Figure 5"},{"comment":"The stated motivation for the work includes using synthetic data to study the stability of clustering methods on the IGP data (Abstract and Section 6), but no clustering stability experiment is reported. Adding such an analysis, for example by clustering the synthetic QTS and comparing the resulting groups with those obtained on the original data, would provide a utility-based validation that the synthetic data preserve clinically relevant structure.","section":"§6 and Abstract"}],"minor_comments":[{"comment":"The text says quaternions were formalized by Sir William Hamilton in 1943; the correct year is 1843.","section":"§2.3.2"},{"comment":"There is a typo in the second paragraph: 'n this article' should read 'In this article'.","section":"§1"},{"comment":"The description of the nearest-neighbor weighting step would be clearer if the Dirichlet sampling were written explicitly, e.g., w_i ∼ Dir(α_i) with α_i defined as in Eq. (7), and if the notation for the weights w_ij were defined consistently with Eq. (8).","section":"§3.3"},{"comment":"The abbreviation 'IGP' is used in the results section but is not defined there; it appears earlier in the text, but a reminder at first use in Section 5 would help readability.","section":"§5"},{"comment":"The statement 'the mean R V coefficient is 0.84' reports a single point estimate; please include the variability across the 100 synthetic datasets, such as a standard deviation or a boxplot.","section":"§5"},{"comment":"The recommended threshold for d_min is set to 10% of the smallest distance between observations in the original dataset, but the rationale for this particular percentage is not explained.","section":"§3.4"}],"recommendation":"major_revision","confidential_remarks":"The paper is potentially a good fit for the journal, but the evaluation is the main weakness. The authors should be encouraged to address the circularity concern by adding direct validation on the reconstructed QTS space and by using a proper resampling scheme for hyperparameter selection. The code is not currently public, which limits reproducibility; I would encourage the editor to request a public release of the implementation as part of the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe one thing to know: this is a real novelty in a niche area. Le Gall et al. are, as far as I can tell, the first to combine tangent-space fPCA with a nearest-neighbor weighted avatar method to synthesize unit quaternion time series. That is a useful contribution for anyone working with small clinical gait datasets, and the framework cleanly separates the geometric reduction (log map, fPCA) from the tabular synthesis step, so other generators could be plugged in. The paper also gives sensible hyperparameter guidance and honestly reports that synthetic curves visibly shrink toward a smaller space.\n\nThe soft spot is real, though, and it is in the evaluation. The fidelity and geometry metrics are all computed on the functional score matrix, not on the reconstructed QTS. And the synthetic scores are generated as convex combinations of the nearest neighbors' real scores (Eq. 8), so the synthetic scores live inside the convex hull of the original scores. Measures like the k-NNG Frobenius distance and the RV coefficient are therefore partly entailed by the generative process. The paper never tests whether the reconstructed quaternion curves are clinically plausible or lie on the true gait manifold; visual inspection of q_x, q_y, q_z is not quantitative. The authors even observe the shrinkage effect themselves, which is exactly what convex averaging predicts, but they do not connect that observation to the metric problem.\n\nThere are also the expected secondary issues: hyperparameters are selected on the same 27-patient data used for evaluation, code and data are not public, and the motivating downstream application (clustering stability) is not actually evaluated. None of this kills the method; the pipeline is coherent and the empirical results are consistent with the claim that SynGait produces plausible synthetic gait data. But the strong claim about geometry preservation is not established by the reported metrics.\n\nWho is this for? Researchers in gait analysis, functional data synthesis, or privacy-preserving sharing of clinical movement data. It deserves a serious referee, but the referee should ask for code/data, a QTS-space or clinically grounded validation, and a downstream utility test before accepting the headline claim.\n\nMy recommendation: send it to peer review. It is novel and the framework is reusable, but the current evaluation needs substantial revision.","headline":"A genuinely novel pipeline for synthetic unit quaternion gait data, but the geometry-preservation claim rests on metrics that are partly built into the construction.","tokens_in":17977,"tokens_out":2233,"would_cite":false,"duration_ms":19899,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H25","62R30"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that synthetic gait data faithful to quaternion geometry can be generated from a small clinical sample of 27 MS patients.","keywords":["synthetic data","quaternion time series","functional principal component analysis","gait analysis","multiple sclerosis","nearest neighbor weighting","Dirichlet distribution","data privacy"],"falsifier":"Compute geodesic distances between reconstructed synthetic QTS and real QTS on the unit-quaternion sphere and compare them with the range seen among real gait cycles, or ask clinicians to label synthetic IGPs as plausible or implausible gait; if the synthetic curves fall far outside the real gait manifold or are flagged as implausible, the geometry-preservation claim fails.","tokens_in":16978,"feed_emoji":"🚶","tokens_out":5699,"duration_ms":54669,"temperature":0.7,"pith_summary":"Multiple sclerosis gait studies often have too few patients to build stable statistical tools, and sharing clinical data is privacy-sensitive. This paper claims that unit quaternion time series recording hip rotation during a step can be transformed into functional PCA scores, synthesized in that tabular space with a nearest-neighbor weighting method called SynGait, and mapped back into unit-quaternion curves that stay on the same geometric manifold as the original data. Applied to 27 MS patients, the generated synthetic gait patterns score well on fidelity metrics and preserve the data's geometry better than copula-based and GAN-based tabular synthesizers. If correct, small clinical samples can be expanded into privacy-aware synthetic datasets for developing and testing clustering and other gait-analysis tools.","feed_headline":"Synthetic gait pipeline preserves geometry from just 27 MS patients","feed_subtitle":"A quaternion-to-PCA-to-nearest-neighbor pipeline lets small clinical studies generate private, high-fidelity gait data.","key_machinery":"The central object is the unit quaternion time series (QTS), which represents hip rotation during a gait cycle; the key identity is the logarithmic/exponential map pair linking the unit-quaternion space $\nmathcal{H}_u$ to its tangent space $\\mathbb{R}^3$. The machinery combines the log map, cubic B-spline functionalization, multivariate functional PCA that produces eigenfunctions and a score matrix, and the flexible avatar method: synthetic scores are convex combinations of the scores of the $\\gamma$ nearest neighbors, with Dirichlet weights whose concentrations are inversely proportional to neighbor distances. Equation (6) recombines synthetic scores with the original principal functions and the mean QTS to produce synthetic QTS.","core_discovery":"The central claim is that the SynGait framework produces high-fidelity synthetic unit quaternion time series for small and private datasets. The pipeline centers each QTS at the Fréchet mean, applies the logarithmic map to the tangent space, interpolates cubic B-splines, runs multivariate functional PCA to obtain a score matrix, synthesizes new scores by taking Dirichlet-weighted averages of each observation's nearest neighbors, and reconstructs QTS with the exponential map plus the mean. On the 27-patient MS dataset, the synthetic data achieves a mean RV coefficient of 0.84, Frobenius distances between k-nearest-neighbor graphs that are smaller than the copula and CTGAN baselines, and a hidden rate of 85%, supporting the claim that the generated curves are new, faithful, and geometry-preserving.","pith_inferences":["Because synthetic scores are convex combinations of real nearest-neighbor scores, geometric closeness in score space is partly enforced by construction; the paper does not independently verify that the reconstructed QTS are clinically plausible gait curves, so a clinician review or geodesic-distance check would strengthen the claim.","The approach could serve as a privacy-preserving data-sharing mechanism for multi-site MS cohorts: each site can release synthetic QTS built from its own patients, and the hidden-rate and local-cloaking metrics suggest re-identification is nontrivial.","Neighboring problems such as synthetic accelerometer or EMG time series could reuse the same tangent-space fPCA plus nearest-neighbor score synthesis, with the same geometry-preservation claim being testable."],"forward_implications":["Small-sample gait studies can generate many synthetic QTS datasets for bootstrapping and stability analysis of clustering, without collecting more patients.","Because the framework reduces to tabular score generation, any existing tabular synthesizer can be plugged in, while SynGait's nearest-neighbor weighting offers better geometry preservation.","The method provides a practical privacy workflow: hyperparameters can be tuned so synthetic points are new and the hidden rate is high, supporting sharing of synthetic gait data.","The same pipeline applies to any motion describable as unit QTS, not just MS hip rotation, and to additional pathologies.","Default hyperparameter settings (components covering about 95% of inertia, concentration around 5, and a number of neighbors near one tenth of the sample size) make the method usable on other unit-QTS datasets."],"supporting_citations":[{"why":"Supplies the functional PCA method that decomposes log-quaternion functions into principal modes and scores.","marker":"[22]"},{"why":"The original avatar method that SynGait generalizes by using Dirichlet-distributed nearest-neighbor weights.","marker":"[23]"},{"why":"Provides the multivariate functional PCA approach used to turn QTS data into a tabular score matrix.","marker":"[28]"},{"why":"The implementation used to compute the multivariate functional PCA scores and eigenfunctions.","marker":"[29]"},{"why":"Gives the Dirichlet distribution properties used for sampling the nearest-neighbor weights.","marker":"[30]"},{"why":"Supplies the copula and CTGAN baseline synthesizers and the evaluation metrics used for comparison.","marker":"[31]"},{"why":"Sklar's theorem is the foundation of the copula baseline for modeling the joint distribution of scores.","marker":"[35]"},{"why":"Provides the k-nearest-neighbor graph theory underlying the geometry-preservation metric.","marker":"[40]"}],"fun_headline_variants":["Synthetic gait data from 27 MS patients keeps geometry intact","Quaternion-to-PCA method produces private MS gait data","Small-sample gait synthesis stays faithful to geometry","SynGait: high-fidelity synthetic gait for MS studies","Dirichlet-weighted neighbors yield private gait data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that closeness of synthetic scores to the nearest neighbors' real scores in PCA space implies the reconstructed quaternion curves lie on the true gait manifold and are clinically plausible, which the paper does not independently test.","fun_headline_variants_meta":{"raw":{"variants":["Synthetic gait data from 27 MS patients keeps geometry intact","Quaternion-to-PCA method produces private MS gait data","Small-sample gait synthesis stays faithful to geometry","SynGait: high-fidelity synthetic gait for MS studies","Dirichlet-weighted neighbors yield private gait data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000671,"raw_usage":{"total_tokens":3044,"prompt_tokens":920,"completion_tokens":2124,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":2047}},"tokens_in":536,"tokens_out":2124,"duration_ms":14020,"temperature":1.0,"reasoning_tokens":2047,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:40:51.061282+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute geodesic distances between reconstructed synthetic QTS and real QTS on the unit-quaternion sphere and compare them with the range seen among real gait cycles, or ask clinicians to label synthetic IGPs as plausible or implausible gait; if the synthetic curves fall far outside the real gait manifold or are flagged as implausible, the geometry-preservation claim fails.","supporting_citations":[{"cited_title":"Ramsay and Bernard W","cited_arxiv_id":null,"evidence_quote":"Supplies the functional PCA method that decomposes log-quaternion functions into principal modes and scores."},{"cited_title":"Patient-centric synthetic data generation, no reason to risk re-identification in biomedical data analysis","cited_arxiv_id":null,"evidence_quote":"The original avatar method that SynGait generalizes by using Dirichlet-distributed nearest-neighbor weights."},{"cited_title":"Multivariate functional principal component analysis for data observed on different (dimensional) domains","cited_arxiv_id":null,"evidence_quote":"Provides the multivariate functional PCA approach used to turn QTS data into a tabular score matrix."},{"cited_title":"MFPCA: Multivariate Functional Principal Component Analysis for Data Observed on Different Dimensional Domains , 2022","cited_arxiv_id":null,"evidence_quote":"The implementation used to compute the multivariate functional PCA scores and eigenfunctions."},{"cited_title":"Dirichlet and Related Distri- butions: Theory, Methods and Applications","cited_arxiv_id":null,"evidence_quote":"Gives the Dirichlet distribution properties used for sampling the nearest-neighbor weights."},{"cited_title":"The synthetic data vault","cited_arxiv_id":null,"evidence_quote":"Supplies the copula and CTGAN baseline synthesizers and the evaluation metrics used for comparison."},{"cited_title":"Computational geometry: an introduc- tion","cited_arxiv_id":null,"evidence_quote":"Provides the k-nearest-neighbor graph theory underlying the geometry-preservation metric."}],"review_version":1}