{"id":"8a191264-83b4-442f-a87b-020e59a36832","arxiv_id":"2603.20756","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Sub-Riemannian horizontal curves on (x,y,t,θ,v,a) automatically satisfy the speed-curvature relation, and Wasserstein distance clusters real movement fragments into the neural states of Kadmon et al.","lead":"A sub-Riemannian model of motor-cortex fragments encodes the observed speed-curvature law via horizontal curves, and Wasserstein clustering of those fragments recovers experimental neural states better than Sobolev distance. The work offers a geometric account of how M1 organizes short hand trajectories into coherent motor primitives.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"The recovered neural-state partition is driven by a hand-tuned time weight that discards absolute time, so the match to experiment may be an artifact of that choice rather than of the Wasserstein geometry.","rationale":"The Reader correctly isolates the weakest link: the specific measure and the hand-tuned weight c5. The geometric derivation of the speed-curvature relation is tautological once α2 is free, so the only non-trivial claim is the clustering result on real data. That result is obtained only after the time coordinate has been effectively nullified and fragments have been shifted to a common origin—precisely the preprocessing that makes the subsequent Wasserstein comparison easy. Because no sensitivity analysis or public code is supplied, the match to experiment cannot yet be regarded as independent confirmation of the proposed geometry. The concrete sweep of c5 (and of the measure itself) would settle whether the agreement is robust or an artifact of the chosen weights; until that check is performed the CONDITIONAL verdict remains appropriate.","tokens_in":12911,"tokens_out":644,"duration_ms":7648,"concrete_test":"Re-run the spectral clustering of Section 4.3 on the same 329 real fragments while sweeping c5 over {0, 0.008, 0.05, 0.1, 0.5, 1.0} (and, separately, replacing the vector measure by the real-valued density of Eq. (13)). Report the adjusted Rand index (or silhouette) against the Kadmon-Harpaz partition for each setting. If the match collapses for any c5 \neq 0.008, the empirical support for the model is parameter-dependent rather than structural.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central empirical claim is that Wasserstein distance on the vector-valued measures μ=(sinθ,cosθ,a)/∫√(1+a^{2}) recovers the same neural states as Kadmon-Harpaz et al. (Section 4.3, Figures 9–11). That recovery, however, rests on the metric weights of Definition 7 / Eq. (11), and especially on the explicit choice c5=0.008 for the temporal coordinate. The authors state that this value “completely discard[s] the time variable” and then reparametrize every fragment to a common initial time. Once absolute timing is removed, the remaining features (orientation and acceleration profile) are precisely the ones already used by the experimental clustering; the Wasserstein step therefore operates on a heavily pre-processed representation whose free parameters have been set to match the target partition. If a different, non-vanishing c5 (or a different normalization of the measure) is used, the affinity matrix and the resulting spectral clusters can change, so the claimed agreement with experiment is not an independent validation of the sub-Riemannian + Wasserstein model.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper models primary motor cortex (M1) features as the 6-dimensional manifold M = R^{2}_{(x,y)} \times R^{+}_t \times S^{1}_\theta \times R_v \times R_a equipped with the sub-Riemannian structure generated by the horizontal fields X1, X2, X3 of Eq. (3). Horizontal (admissible) curves are proposed as models of movement fragments; the differential constraints immediately yield the curvature–speed relation k = \theta'/v (Eq. 8) and hence R = v/\theta' (Eq. 9), recovering the experimentally observed speed–curvature law (and the 2/3-power law by the special choice \theta' = v^{-2}). Fragments are then represented by real- or vector-valued probability measures (Eqs. 13–14) and clustered by spectral analysis of a Gaussian kernel of the Wasserstein distance. On synthetic data the method recovers coherent groups (silhouette 0.65); on the real reaching trajectories of Kadmon-Harpaz et al. it recovers a partition visually comparable to the experimentally reported neural states (silhouette 0.35), outperforming the earlier Sobolev distance that required identical time domains.","tokens_in":13222,"tokens_out":1085,"duration_ms":9045,"significance":"If the geometric derivation and the clustering results hold, the work supplies a single sub-Riemannian framework that simultaneously (i) explains the classical speed–curvature relation from first principles and (ii) organizes experimentally recorded fragments into the same neural states previously obtained from multi-electrode recordings. The introduction of Wasserstein distance into cortical connectivity modelling is novel and removes the artificial re-parametrization required by Sobolev-type metrics. The paper therefore offers a concrete, falsifiable link between differential geometry, optimal transport and motor-cortex physiology that can be tested on further data sets and with alternative measures.","major_comments":[{"comment":"Section 4.3 and Definition 7 / Eq. (11): the recovery of the Kadmon-Harpaz neural states rests on the explicit choice c5 = 0.008 for the temporal weight, which the authors state “completely discard[s] the time variable,” followed by re-parametrization of every fragment to a common initial time. Once absolute timing is removed, the remaining features (orientation and acceleration profile) are precisely those already used by the experimental clustering. No sensitivity analysis or ablation with non-vanishing c5 (or with alternative normalizations of the measure μ) is provided; without it the claimed agreement with experiment cannot be regarded as an independent validation of the sub-Riemannian + Wasserstein model.","section":null},{"comment":"Section 3.1.3, Eq. (14) and Section 4.3: the vector-valued density μ = (sin \theta, cos \theta, a)/∫√(1+a^{2}) is introduced without physiological or information-theoretic justification, and the kernel widths σ1 = 0.7, σ2 = 0.09 are hand-tuned. The silhouette scores (0.65 synthetic, 0.35 real) are reported, yet no quantitative comparison against alternative distances (e.g., dynamic time warping, plain L^{2} on the same features) or against shuffled baselines is given. Consequently the claim that Wasserstein distance is “much more efficient” remains only qualitative.","section":null}],"minor_comments":[{"comment":"Eq. (14) writes the components of μ as (sin \theta, cos \theta, a) while the surrounding text lists (cos \theta, sin \theta, a); the order should be made consistent.","section":null},{"comment":"Figure 7 caption and the accompanying text refer to a “center-out task,” yet the data provenance (Hatsopoulos laboratory) is only acknowledged in the acknowledgments; a brief methods paragraph describing the recording and pre-processing pipeline would improve reproducibility.","section":null},{"comment":"The parameters c1…c4 of the ball-box norm (Eq. 11) are said to be “tuned \to experimentally observed clusterization,” but their concrete values are never listed; they should be reported for the real-data experiment.","section":null},{"comment":"Typographical inconsistencies appear throughout (e.g., “Hatsoupolos,” “real-valued probability: speed and curvature,” missing spaces after commas in several equations).","section":null}],"recommendation":"major_revision","confidential_remarks":"The geometric derivation of the speed–curvature law is clean and publishable on its own. The clustering claim, however, is currently under-supported by the hand-tuned temporal weight and the absence of quantitative baselines; a major revision that either removes the free parameter or demonstrates robustness would make the paper substantially stronger. Fit for a computational-neuroscience or mathematical-biology journal is good once the empirical validation is tightened."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing worth knowing is that this is a clean, usable fix to the authors’ own earlier model. Their 2023–24 sub-Riemannian feature space and admissible curves already gave the speed-curvature relation R = v/α2 (and the 2/3-power version by choice of control). What is new is replacing the Sobolev distance with a vector-valued Wasserstein metric so that fragments of unequal length can be compared without forced re-parametrization. On the Hatsopoulos data they recover essentially the same neural-state partition that Kadmon-Harpaz et al. found from the neural activity itself; the Sobolev version simply fails on the same set. That is a concrete empirical advance, and the math is correctly derived.\n\nThey do the geometry carefully: the horizontal generators force k = α2/v, the ball-box metric on the reduced space is standard, and the one-dimensional Wasserstein formula is applied without hand-waving. The synthetic experiments (spirals, random accel/decel trajectories) produce coherent clusters with a respectable silhouette of 0.65. On real fragments the silhouette drops to 0.35, but the visual match to the experimental states is still clear.\n\nThe soft spots are real but proportionate. The metric weights, especially c5 = 0.008 that “completely discards the time variable,” are tuned so that absolute timing is removed and every fragment is shifted to a common start—exactly the normalization already used by Kadmon. The particular density μ = (sin θ, cos θ, a)/∫√(1+a^{2}) is also an author choice. Change either and the affinity matrix can shift. There are no shuffled baselines, no systematic sensitivity sweep, and no released code or processed fragments. Those are the usual mid-stage gaps, not load-bearing contradictions.\n\nThis is for people already working in neurogeometry or motor-primitive decoding. If you care about sub-Riemannian models of cortex or about distances that respect variable-duration trajectories, read it; otherwise it is optional. The thinking is coherent and the literature engagement is honest. I would send it to referees; the core claim is interesting enough and formally grounded enough to deserve a proper review, even if the free-parameter story needs tightening.","headline":"Solid incremental step: Wasserstein lets their prior sub-Riemannian fragment model finally work on real variable-duration M1 data and recover the Kadmon states, while the speed-curvature relation falls out of the horizontal distribution by construction.","tokens_in":13823,"tokens_out":601,"would_cite":false,"duration_ms":13115,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["53C17","92C20","49Q22"],"pacs":[],"model":"grok-4.5","headline":"Horizontal curves in a sub-Riemannian model of motor cortex automatically satisfy the observed speed-curvature law, and Wasserstein clustering of fragments recovers experimental neural states.","keywords":["primary motor cortex","movement fragments","sub-Riemannian geometry","Wasserstein distance","neural states","speed-curvature law","neurogeometry"],"falsifier":"Apply the same Wasserstein affinity matrix (same density definition and metric weights) to a new multi-electrode reaching dataset recorded under comparable conditions; if the resulting spectral clusters fail to match the experimentally reported neural states, the central claim is falsified.","tokens_in":13777,"feed_emoji":"🧠","tokens_out":877,"duration_ms":20347,"temperature":0.7,"pith_summary":"The paper models the primary motor cortex as a six-dimensional feature space of hand position, time, orientation, velocity and acceleration equipped with a sub-Riemannian structure. Its horizontal curves are forced by the differential constraints to obey the experimentally observed relation between trajectory curvature and movement speed. Short hand trajectories called fragments are turned into vector-valued probability measures; clustering them with the Wasserstein distance recovers the same grouping into neural states previously extracted from multi-electrode recordings, and does so without the artificial time reparametrization required by earlier Sobolev distances. A reader who accepts the model therefore obtains a single geometric account of both the kinematics of reaching and the hierarchical organization of motor primitives in cortex.","feed_headline":"Motor-cortex geometry forces the speed-curvature law","feed_subtitle":"Wasserstein clustering of trajectory fragments recovers the neural states seen in real cortical recordings.","key_machinery":"The sub-Riemannian structure generated by the vector fields X1 = v cos θ ∂x + v sin θ ∂y + a ∂v + ∂t, X2 = ∂θ, X3 = ∂a, whose horizontal curves automatically obey R = v/α₂, together with the Wasserstein distance on the vector-valued densities μ = (sin θ, cos θ, a)/∫√(1+a²) that supplies the affinity matrix for spectral clustering into neural states.","core_discovery":"A sub-Riemannian geometry on the feature space of position, time, orientation, velocity and acceleration makes its horizontal curves satisfy the speed-curvature relation observed in reaching movements; the Wasserstein distance between the vector-valued measures associated with those curves then clusters real movement fragments into the same neural states reported from cortical recordings, outperforming the Sobolev distance that required all trajectories to share a common time interval.","pith_inferences":["The same Wasserstein-on-horizontal-curves construction may apply to other cortical areas whose cells are selective to short spatiotemporal trajectories.","Metric weights that effectively discard absolute time imply a short-term shape memory that is shift-invariant but not scale-invariant in time.","Replacing hand-tuned coefficients with anatomically or data-derived weights would yield a fully parameter-free prediction of neural-state boundaries.","Population codes in motor cortex should be more sensitive to the joint distribution of orientation and acceleration than to absolute position or absolute time."],"forward_implications":["The speed-curvature law need not be imposed by hand; it is a differential consequence of the chosen feature geometry.","Cortical connectivity among fragments can be modelled by a Wasserstein kernel that tolerates variable durations without reparametrization.","Spectral clustering of that kernel recovers the experimentally observed neural states from real reaching data.","The same geometric setting that organizes elementary features into fragments also organizes fragments into higher-order neural states."],"fun_headline_variants":["Sub-Riemannian geometry forces motor speed-curvature law","Horizontal curves enforce observed speed-curvature relation","Wasserstein distance clusters fragments into neural states","Sub-Riemannian model links kinematics to cortical recordings","Trajectory geometry recovers motor-cortex neural fragments"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The model assumes that the particular vector-valued density built from orientation and acceleration, together with metric weights that nearly discard absolute time, correctly represents how motor-cortex cells compare movement fragments.","fun_headline_variants_meta":{"raw":{"variants":["Sub-Riemannian geometry forces motor speed-curvature law","Horizontal curves enforce observed speed-curvature relation","Wasserstein distance clusters fragments into neural states","Sub-Riemannian model links kinematics to cortical recordings","Trajectory geometry recovers motor-cortex neural fragments"]},"model":"grok-4.5","effort":"low","cost_usd":0.004078,"raw_usage":{"total_tokens":1179,"prompt_tokens":653,"num_sources_used":0,"completion_tokens":76,"cost_in_usd_ticks":40780000,"prompt_tokens_details":{"text_tokens":653,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":450,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":653,"tokens_out":76,"duration_ms":22715,"temperature":1.0,"reasoning_tokens":450,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T21:16:08.214969+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Apply the same Wasserstein affinity matrix (same density definition and metric weights) to a new multi-electrode reaching dataset recorded under comparable conditions; if the resulting spectral clusters fail to match the experimentally reported neural states, the central claim is falsified.","supporting_citations":[],"review_version":1}