{"id":"955de39c-ac1d-4864-9ea8-bfbc85af552e","arxiv_id":"2501.03247","paper_version":4,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A sub-Riemannian distance on movement fragments is proposed to reproduce the eight motor-cortex neural states of Kadmon-Harpaz et al., but the reported tests are synthetic only.","lead":"A geometric model treats motor-cortex movement fragments as curves in a sub-Riemannian space and groups them into 'neural states' using spectral clustering. It claims to recover experimentally observed states, but the results only show synthetic examples and no comparison to the neural data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's central claim is not tested: Section 4 clusters only synthetic fragments generated from the model's own parametric family, with no comparison to the neural states of [11], so the claimed recovery is unsupported.","rationale":"The paper's stated goal is to show that the chosen kinematic variables and distance metric are sufficient to explain neural-state formation, by recovering the states of [11] from measured cortical activity. For that claim to hold, the model's clusters must be shown to match [11]'s empirical states. Section 4 never does this: it generates fragments from the model's own parametric family (16), applies the clustering algorithm, and visually asserts 'correct clusterization' in eight clusters. The target's eight states appear to be used to fix k, so part of the apparent match is by construction. This is a more fundamental gap than the spectral-to-biological equivalence highlighted by the reader: even if the eigenvector correspondence from [45] were valid for motor cortex, the results would still not validate the central claim without a comparison to [11]'s data. I therefore agree with the reader's REJECT verdict but emphasize a different assumption as the most load-bearing: the implicit assumption that clustering synthetic fragments generated by the model itself constitutes evidence of recovery of empirical neural states. The reader's rationale does mention the absence of reference to [11]'s data, so the agreement is partial rather than full.","tokens_in":12388,"tokens_out":3695,"duration_ms":33931,"concrete_test":"Apply the Section 3.5 pipeline (affinity matrix (21), normalized random-walk matrix P, thresholding, and k-means with k set to the number of states reported in [11]) to the actual fragments used in [11]'s random-target pursuit task, and compare the model's cluster assignments to the HMM-derived neural states of [11] using the adjusted Rand index. If the index is not substantially above chance (e.g., not above 0.5 for eight clusters), the claimed recovery is refuted; if the data from [11] are unavailable, the recovery claim should be explicitly downgraded to an untested conjecture.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and conclusion state that applying the grouping algorithm to the model recovers the neural states observed in Kadmon Harpaz et al. [11], which were based on measured cortical activity. The only evidence in Section 4 is clustering of synthetic curves generated from the model's own fragment family (16): uniform theta0 and j in Section 4.1, all parameters random in Section 4.2. There is no quantitative or visual comparison of the resulting clusters to the HMM-derived states of [11], no use of [11]'s neural data, and no evaluation metric such as an adjusted Rand index. The number of clusters appears to be set to eight in the k-means step, matching the target's eight states, and the results are described as 'correct' without stating the criterion. Therefore the central claim is unsupported as stated, regardless of the geometric construction's internal consistency. The theoretical step linking eigenvectors of K_F to biological states is imported from [45] and not validated for motor cortex, but the empirical gap alone is sufficient to undermine the claimed recovery.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a sub-Riemannian model of the primary motor cortex in which movement fragments are curves in a feature space, a pseudo-metric is defined on a reduced space independent of (x,y), and a connectivity kernel is used in a mean-field equation. The authors then apply a spectral clustering algorithm to the eigenvectors of the discretized kernel and claim that this procedure recovers the eight neural states found by Kadmon Harpaz et al. [11] from measured cortical activity. The only experimental evidence presented is clustering of synthetic fragments generated from the model's own family (16), with no comparison to the neural states or data of [11].","tokens_in":12590,"tokens_out":4100,"duration_ms":40275,"significance":"If the central claim were established, the model would provide a kinematic explanation for the formation of neural states without invoking neural data, which would be a valuable contribution to motor-control theory. The geometric construction is explicit, the exponential coordinates in Remark 1 are computed in closed form, and the synthetic experiments are reproducible in principle. However, the paper does not validate the claimed recovery of the biological states of [11]: the results in Section 4 cluster synthetic fragments only, and the link between eigenvectors of the connectivity kernel and neural states is imported from [45] without testing. The evidence is therefore circular relative to the main claim, and the manuscript currently does not deliver the promised biological validation.","major_comments":[{"comment":"The central claim of the paper, stated in the abstract and conclusion, is that the grouping algorithm recovers the neural states observed in [11]. Section 4 tests this only on synthetic fragments generated from the model's own family (16): uniform θ0 and j in Section 4.1 and fully random parameters in Section 4.2. There is no comparison, quantitative or visual, to the HMM-derived states of [11], no use of the neural data from [11], and no evaluation metric such as an adjusted Rand index. The text in Section 4.1 says the algorithm gives a 'correct clusterization' without defining the correctness criterion, and the number of clusters is set to eight, matching the target. The claimed recovery is therefore unsupported as stated.","section":"Section 4, Figures 6-11"},{"comment":"The fragment family in (16) already encodes the variables that separate the resulting clusters: the initial direction θ0 and the jerk coefficient j, which determines the sign of acceleration. Section 4.1 reports that the clusters are characterized by orientation quadrant and increasing versus decreasing velocity, so the algorithm partly redis covers labels that are built into the generative model. To break this circularity, the clustering should be applied to fragments whose direction and acceleration profile are not predetermined by the model, or to the real fragments of [11], and compared with the HMM states using a standard clustering agreement measure.","section":"Section 3.2-3.3, Eq. (16)"},{"comment":"Proposition 2 guarantees only a local estimate C0 dM1 ≤ |e| ≤ C1 dM1 with unspecified constants, but Eq. (15) is used as an exact distance formula in the kernel (17). The paper does not justify that the constants can be ignored, that the Heisenberg coordinate expression is valid globally, or that replacing e2 by 4 sin((θ0−θ1)/4) preserves the required estimate. Since the affinity matrix (21) depends exponentially on dF, uncontrolled constants in the distance estimate can materially change the spectral clustering output.","section":"Section 3.1, Proposition 2 and Eq. (15)"},{"comment":"The identification of stable neural states with the principal eigenvectors of the connectivity kernel KF is assumed from [45] and is neither derived nor tested for motor cortex. The manuscript does not show that the HMM-derived states of [11] correspond to the groupings of eigenvectors of the linearized mean-field equation (18). If this spectral-to-biological correspondence fails, the synthetic clusters in Section 4 have no bearing on real neural states. This is a load-bearing assumption that needs either a derivation in the motor-cortex setting or a direct empirical test.","section":"Sections 3.4-3.5, Eq. (20)"},{"comment":"The pseudo-distance dF in Definition 5 contains the term ||γ′1(t)−γ′2(t)||M1, but the manuscript defines a horizontal norm only at a single point via Eq. (9). The difference of tangent vectors at different base points requires a connection or a global coordinate identification, which is not introduced. This makes the definition of dF incomplete and directly affects the kernel (17) used in all subsequent computations.","section":"Definition 5, Eq. (17)"}],"minor_comments":[{"comment":"The word 'esperiment' should be 'experiment'.","section":"Section 1, first paragraph"},{"comment":"The word 'senslible' should be 'sensitive'.","section":"Section 2.1, paragraph on neural states"},{"comment":"In the system displayed in the proof, '˙v=e1a+e5' uses an undefined variable e5; it should be e4, consistent with the expression for e4 immediately below.","section":"Remark 1, proof"},{"comment":"The coefficient α3 is written as j(t−T/2) in Eq. (6) but as j(s−T/2) in Eq. (16), with T undefined in (16). Clarify that the interval is rescaled and define T consistently.","section":"Eqs. (6) and (16)"},{"comment":"The space M1 is written as R^+_t × S^1 × R^2, but the original space M in Eq. (2) has t∈R. Justify the restriction to nonnegative time or state that the half-line is only a notational convenience for the interval of definition of fragments.","section":"Section 3.1, definition of M1"}],"recommendation":"reject","confidential_remarks":"The paper's central claim is the recovery of the neural states of [11] from a kinematic model. Since Section 4 only clusters synthetic fragments generated from the model's own family, and the abstract and conclusion state the recovery as a fact, the manuscript does not meet the standard for publication in its current form. Even if the geometric construction is internally consistent, the missing empirical comparison with [11] is a load-bearing gap that cannot be fixed by rewriting alone. The authors would need to obtain the neural data or the state labels of [11] and perform a quantitative cluster comparison, or substantially weaken the claim to a proposal about possible kinematic mechanisms. This is a scope and novelty concern: the paper promises a validation that it does not perform."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The paper builds a genuinely new geometric construction: it reduces the feature space M to a submanifold M1 that drops (x,y), computes an explicit distance estimate on it, and uses it to define a pseudo-metric on the space of fragments. That part is careful and new relative to the prior fragment model [21]. The second thing is that the abstract and conclusion overclaim: they say the model recovers the neural states of Kadmon Harpaz et al. from measured cortical activity, but Section 4 never touches [11]'s data. It clusters synthetic fragments generated from the model's own parametric family (16), with k fixed to eight to match the target, and calls the results 'correct' without stating a criterion. No adjusted Rand index, no comparison to the HMM states. So the central empirical claim is unsupported as stated.\n\nWhat's good: the submanifold M1, the Heisenberg-group identification, and the distance estimate (15) are concrete and checkable. The modular hierarchy (features → fragments → states) is a reasonable way to think about motor coding, and the mean-field plus spectral-clustering pipeline is coherent if you accept the imported link from [45]. The math is mostly sound; I don't see an internal contradiction.\n\nSoft spots beyond the missing data: the Nagel-Stein-Wainger estimate only gives distance equivalence up to constants, but the explicit norm (15) is used as if exact. That's a gap, though probably not fatal if the clustering is robust. The spectral-to-biological equivalence is imported from the authors' own prior work and not validated for motor cortex. And the fragment family already encodes direction and acceleration sign—the variables that separate the clusters—so the synthetic test is close to a construction check, not a prediction.\n\nWho gets value: readers working on neurogeometry or on geometric models of motor cortex will find the construction interesting. But anyone citing it as evidence for how neural states emerge should wait.\n\nRecommendation: this deserves a serious referee, but not in its current form. The right fix is either to bring in [11]'s actual data or to rewrite the claims as 'we propose a kinematic grouping that qualitatively matches the fragment-level structure of [11]'. With that change, the geometric part could be a solid contribution. As is, the overclaim is the main problem.","headline":"A genuinely new geometric construction in service of an overstated empirical claim: the paper never compares its clustering to the data it claims to recover.","tokens_in":13142,"tokens_out":5554,"would_cite":false,"duration_ms":50463,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["92C20","53C17"],"pacs":[],"model":"deepseek-v4-flash","headline":"The eight neural states observed in motor cortex can be recovered from kinematics alone.","keywords":["sub-Riemannian geometry","motor cortex","neural states","movement fragments","spectral clustering","connectivity kernel","mean-field equation","kinematic features"],"falsifier":"Run the same spectral-clustering pipeline on the fragments of [11]—or on a fresh motor-cortex dataset—and compare the k-means labels on the leading eigenvectors of $K_F$ with the hidden-Markov-model states; if the match is at chance level, or if the eigenvalue threshold predicts a number of clusters different from eight, the central claim is falsified.","tokens_in":12174,"feed_emoji":"🧠","tokens_out":7912,"duration_ms":66414,"temperature":0.7,"pith_summary":"The paper aims to show that the eight neural states observed in the primary motor cortex during arm reaching can be produced from purely kinematic and geometric principles, with no access to neural recordings. It represents each movement fragment—a short hand trajectory with nearly constant direction and a bell-shaped speed profile—as a curve in a space of time, direction, speed, and acceleration, and equips that space with a sub-Riemannian distance that ignores hand position. From this distance it builds a connectivity kernel, inserts it into a mean-field equation for cortical activity, and groups the kernel's leading eigenvectors. The resulting eight groups reproduce the neural states that an earlier experimental study [11] found with a hidden Markov model on measured neural activity, a grouping that study could not recover from kinematics alone. The paper reads this match as evidence that the chosen kinematic variables and distance capture the mechanism behind neural-state formation.","feed_headline":"Eight brain states recreated from movement geometry alone","feed_subtitle":"A purely kinematic distance metric, without neural recordings, reproduces the motor cortex's eight neural-state clusters.","key_machinery":"The load-bearing mechanism is a pseudo-metric on the space of fragments, $$d_F(\\gamma_1,\\gamma_2)=\\$int_0^{1}$\\|\\gamma_1'(t)-\\gamma_2'(t)\\|_{M_1}\\,dt+d_{M_1}(\\gamma_1(1),\\gamma_2(1)),$$ built from a sub-Riemannian distance on a four-dimensional manifold that drops the position coordinates and factors locally as the product of the Heisenberg group and a circle; the distance is estimated through canonical coordinates using a standard local bound. This metric defines the Gaussian-type connectivity kernel $K_F(\\gamma,\\bar\\gamma)=e^{-d_F(\\gamma,\\bar\\gamma)^2}$, which enters a mean-field equation for cortical activity. Linearizing that equation reduces stable states to the leading eigenvectors of the kernel's discretization, and the argument uses a normalized Markov matrix followed by k-means to obtain the clusters. The modularity of the construction—features first, then fragments, then states—is itself part of the proposed neural architecture.","core_discovery":"The paper's central claim is that neural states are not hidden in the population activity but are already present in the geometry of hand movement, once the right kinematic variables and distance are used. Each fragment is modeled as an admissible horizontal curve in the sub-Riemannian feature space; projecting out the position variables defines a pseudo-metric on fragment space in which fragments that differ only by hand location are equivalent. The heat-kernel estimate $K_F(\\gamma,\\bar\\gamma)=e^{-d_F(\\gamma,\\bar\\gamma)^2}$ is then treated as the cortical connectivity kernel, and a linearized mean-field equation shows that stable activation patterns correspond to the principal eigenvectors of $K_F$. Applying spectral clustering and k-means to those eigenvectors yields eight clusters whose (x,y) direction histograms and (t,v) speed profiles coincide with the neural states of [11]. The paper concludes that this kinematic construction is sufficient to explain neural-state formation, and that its two-stage grouping mirrors the brain's hierarchical processing of motor primitives.","pith_inferences":["The paper leaves implicit that the eigenvalue spectrum of $K_F$ could predict the number of neural states directly, rather than fixing eight from the experimental data.","Because the pseudo-metric is invariant to hand position, the same construction should transfer to other task geometries or effectors once their symmetry variables are quotiented out.","If the eigenvector-to-HMM correspondence holds, the model offers a behavioral readout: neural-state structure could be inferred from kinematic recordings alone, which would be useful for brain-machine interface decoders."],"forward_implications":["Neural-state classification no longer requires neural recordings; a kinematic kernel and spectral clustering reproduce the same eight groups.","Because the pseudo-metric ignores hand position, the model explains why neural states are invariant to where in the work space the hand moves.","The same sub-Riemannian framework that describes fragments can be extended upward to describe states, giving one geometry for multiple scales of motor coding.","The linearized mean-field equation ties stable cortical activity to the kernel's leading eigenvectors, so state formation is a pattern-formation phenomenon rather than a purely statistical clustering.","The eight-cluster structure emerges from the eigenvalue threshold on the connectivity kernel, matching the experimentally observed number of states."],"supporting_citations":[{"why":"provides the experimentally measured neural-state clusters that the model must reproduce and that a purely kinematic distance previously failed to recover","marker":"[11]"},{"why":"defines the feature space, admissible curves as fragments, and the initial sub-Riemannian kernel on features","marker":"[21]"},{"why":"gives the local distance estimate in canonical coordinates that makes the pseudo-metric computable","marker":"[33]"},{"why":"supplies the argument that stable patterns of the mean-field equation correspond to principal eigenvectors of the connectivity kernel","marker":"[45]"},{"why":"provides the normalized Markov matrix used for spectral clustering of the affinity matrix","marker":"[50]"},{"why":"justifies that clustering in the eigenvector coordinate space is equivalent to clustering in fragment space","marker":"[52]"}],"fun_headline_variants":["Movement geometry alone yields motor cortex's eight states","Motor cortex states from pure kinematic distance","Sub-Riemannian geometry explains motor brain states","Hand-motion metric alone reproduces motor cortex states"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument collapses if the leading eigenvectors of the kinematic connectivity kernel do not actually give the same grouping as the hidden-Markov-model neural states, because that equivalence is imported from earlier work on visual cortex and is not re-derived or tested for motor cortex here.","fun_headline_variants_meta":{"raw":{"variants":["Movement geometry alone yields motor cortex's eight states","Motor cortex states from pure kinematic distance","Sub-Riemannian geometry explains motor brain states","Hand-motion metric alone reproduces motor cortex states"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000659,"raw_usage":{"total_tokens":3013,"prompt_tokens":946,"completion_tokens":2067,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":2009}},"tokens_in":562,"tokens_out":2067,"duration_ms":16225,"temperature":1.0,"reasoning_tokens":2009,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:43:28.016705+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same spectral-clustering pipeline on the fragments of [11]—or on a fresh motor-cortex dataset—and compare the k-means labels on the leading eigenvectors of $K_F$ with the hidden-Markov-model states; if the match is at chance level, or if the eigenvalue threshold predicts a number of clusters different from eight, the central claim is falsified.","supporting_citations":[{"cited_title":"Movement decomposition in the primary motor cortex,","cited_arxiv_id":null,"evidence_quote":"provides the experimentally measured neural-state clusters that the model must reproduce and that a purely kinematic distance previously failed to recover"},{"cited_title":"Functional architecture of m1 cells encoding movement direction,","cited_arxiv_id":null,"evidence_quote":"defines the feature space, admissible curves as fragments, and the initial sub-Riemannian kernel on features"},{"cited_title":"Balls and metrics defined by vector fields i: Basic properties,","cited_arxiv_id":null,"evidence_quote":"gives the local distance estimate in canonical coordinates that makes the pseudo-metric computable"},{"cited_title":"The constitution of visual perceptual units in the functional architecture of v1,","cited_arxiv_id":null,"evidence_quote":"supplies the argument that stable patterns of the mean-field equation correspond to principal eigenvectors of the connectivity kernel"},{"cited_title":"A random walks view of spectral segmentation,","cited_arxiv_id":null,"evidence_quote":"provides the normalized Markov matrix used for spectral clustering of the affinity matrix"},{"cited_title":"Diffusion maps,","cited_arxiv_id":null,"evidence_quote":"justifies that clustering in the eigenvector coordinate space is equivalent to clustering in fragment space"}],"review_version":1}