{"id":"29f13112-802f-4dc7-85a2-875eb943a020","arxiv_id":"2510.16216","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Single-module grid cell activity can be decoded into local movement trajectories by lifting toroidal coordinates from the population activity to the plane, up to an affine transformation.","lead":"Grid-cell population activity can be mapped to a torus; this paper unwraps the resulting toroidal path to reconstruct the animal's movement trajectory up to an affine transformation, using no training data or external position information. The method suggests a single grid-cell module carries enough information for local path integration, a question of long-standing interest in neuroscience and navigation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Per-segment affine fits weaken the 'up to affine' claim in 2D; test with a single held-out affine map.","rationale":"The reader's weakest assumption focused on the dense-sampling/nearest-tile rule, which is indeed a genuine limitation and is acknowledged by the authors. However, the most load-bearing issue for the central claim as worded ('reconstructed paths differ from the original by an affine transformation') is the evaluation protocol: in 2D, a separate affine transformation is fit for each local segment, so the reported low errors do not demonstrate a single global affine relationship. The paper itself states the global reconstruction differs in shape, so the abstract overstates the claim. A held-out affine test would directly assess whether the reconstruction captures a stable metric. If the held-out error is high, the method's contribution reduces to showing local smooth distortion rather than a true affine decoding, which would merit a CONDITIONAL verdict with stronger caveats. Since the paper's actual results (local reconstructions, simulation with global affine) still provide support, the reader's CONDITIONAL verdict remains appropriate; no change needed, but the affine-fit ambiguity should be resolved for the claims as stated.","tokens_in":19236,"tokens_out":12177,"duration_ms":109477,"concrete_test":"Use the 2D experimental dataset (rat R, module 1, day 2, OF). Divide the recording into non-overlapping 40-second local segments. Hold out a random subset (e.g., 20%). On the remaining segments, jointly estimate a single affine transformation that best maps the lifted toroidal coordinates to the true positions (pool all training segments and solve a least-squares affine problem). Apply this fixed transformation to the held-out lifted segments and compute the mean reconstruction error. Compare this held-out error to (i) the per-segment affine-aligned error reported in Fig. 7K and (ii) the mismatched-pair baseline. If the held-out error is similar to the per-segment error, the global-affine claim is supported; if it jumps toward the baseline, the method only provides local smooth alignment and the 'up to an affine transformation' claim should be restricted accordingly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In the two-dimensional experimental analysis (Section 2.4, Fig. 7), reconstruction errors are computed after fitting an optimal affine transformation separately for each 40-second local segment (Section 4.3.2). This means the statement that 'reconstructed paths differ from the original by an affine transformation' is not a single global map: the global reconstruction visibly fails (Fig. 7B), and the reported low errors rely on a different 6-DOF affine correction for every segment. For short, smooth segments, an affine fit can absorb substantial shape errors, so the absolute error values are not a strong test of whether the toroidal lift recovers metric information. The mismatched-pair baseline (Fig. 7K) partially controls for this, but it also applies per-pair affine alignment and does not test whether a single affine relationship holds across segments. The abstract's unqualified phrasing 'The reconstructed paths differ from the original by an affine transformation' is therefore overstated; at best the data support locally affine alignment, not a global affine equivalence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a training-free, topology-based method for reconstructing an animal's movement trajectory from the population activity of a single grid-cell module. The pipeline uses persistent cohomology to confirm toroidal structure and to assign toroidal coordinates to each population vector, then 'lifts' the resulting discrete path on the torus to the plane via a greedy nearest-tile rule. The authors validate the method on continuous-attractor-network simulations (with and without added spontaneous firing) and on two published experimental datasets: a 1D virtual linear track and a 2D open field. For simulations and the 1D recording, reconstruction errors are low and far below random baselines; for the 2D recording, only local 40-second segments are reconstructed accurately, while the global reconstruction fails. The central claim is that the reconstructed path differs from the original by an affine transformation, and that co-modular grid cells therefore carry sufficient information for path integration.","tokens_in":19519,"tokens_out":3345,"duration_ms":29601,"significance":"If the central claim holds, the paper offers a conceptually novel decoding strategy that requires no external position information, no training, and no knowledge of grid phases, and it provides a concrete computational demonstration that a single grid module can support path integration. The simulation results are clean: the correct environment topology is recovered, and reconstruction errors are far below random baselines (z-scores between -5.6 and -8.6). The 1D experimental result (median error 3.6% over 617 runs) is a solid proof-of-concept. The paper ships code and uses public data, which are important strengths. However, the 2D experimental evidence is restricted to local segments and, as discussed below, the 'up to affine' claim is weakened by the per-segment affine fitting procedure.","major_comments":[{"comment":"The abstract states that 'reconstructed paths differ from the original by an affine transformation,' but in the 2D experimental analysis the reported errors are computed after fitting a separate 6-DOF affine transformation for each 40-second local segment (Section 4.3.2). The global reconstruction visibly fails (Fig. 7B), so no single affine map is demonstrated to relate the reconstructed path to the true trajectory. For short, smooth segments an affine fit can absorb substantial shape error, so the low mean error (0.120) and the t-test against mismatched pairs (Fig. 7K) do not establish that the toroidal lift recovers metric information beyond a local shape similarity. Please test whether a single affine transformation, fitted on one segment (or on a training set of segments), generalizes to held-out segments; report the reconstruction error under that fixed map, and qualify the abstrac","section":"Section 2.4 / Section 4.3.2"},{"comment":"The greedy lifting rule assumes that between consecutive time samples the true path on the torus crosses at most one torus edge in each coordinate; otherwise the wrong tile is selected and errors accumulate. The footnote notes this assumption but does not validate it. This is load-bearing: the authors attribute the global 2D failure precisely to such lifting errors (Fig. 8, Section 2.4). Please quantify how often the nearest-tile rule makes a correct decision, using the CAN simulations where the ground-truth lift is known, and show how errors accumulate over time. Without this, it is unclear whether the 1D and local 2D successes reflect a generally valid property of grid-cell paths or favorable sampling conditions.","section":"Section 2.1, Eq. (1), footnote 3"},{"comment":"The proximity parameter ε controls every edge-crossing decision in the lift, yet it is selected by an ad-hoc procedure: restrict maximal coordinate differences to [2,2π], reverse their order, and take the value at which the CDF first exceeds α=0.99. No sensitivity analysis is reported for either ε or α. Since the method's ability to distinguish a genuine edge crossing from noise depends on this threshold, please report how reconstruction errors (or the success/failure of toroidal coordinate computation) vary with ε and α, at least for the simulated data. If the results are highly sensitive to these choices, the robustness claims in Section 2.2 need to be substantially qualified.","section":"Section 4.3.3"},{"comment":"The 2D experimental analysis is based on a single dataset (rat R, module 1, day 2, OF). The claim that 'local trajectories can be reliably reconstructed' from 2D recordings would be much stronger if the method were applied to additional animals/modules/days from the same public dataset [15]. Given that the global reconstruction fails, the generality of the local result remains uncertain.","section":"Section 2.4"}],"minor_comments":[{"comment":"Typo: 'Accumulated errors of this type can can cause' should read 'can cause'.","section":"Section 2.4"},{"comment":"The subheading 'F rom grid cell activity to path on a torus' contains a spacing error ('F rom').","section":"Section 2.1"},{"comment":"Reference [1] is incomplete ('523 Neural Syst. 2012.' lacks the journal title and page/volume details).","section":"Reference list"},{"comment":"In Table 1, the entries with N/A indicate failure to compute toroidal coordinates, but it is not stated whether these trials were excluded from the summary statistics or from the z-scores in Fig. 4E. Please clarify how failed trials are handled.","section":"Section 2.2 / Table 1"}],"recommendation":"major_revision","confidential_remarks":"The core idea is interesting and the simulation/1D results are convincing, but the abstract's unqualified 'up to an affine transformation' claim is not supported by the 2D experimental evidence, where only per-segment affine fits are used and the global reconstruction fails. The requested sensitivity analyses and a global-affine test are feasible within the manuscript's scope and would substantially strengthen the paper. I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a solid empirical demonstration of something the reader should know: with the right toroidal coordinates, phase unwrapping (they call it path lifting) is enough to reconstruct short trajectories from a single grid module, without training. That is a useful result, and the 2D local results are real, even if the math itself is not new. The novelty is the packaging: persistent cohomology plus toroidal coordinates plus greedy unwrapping into one pipeline, validated on both simulations and real data.\n\nWhat works well: the simulation study is clean. The reconstructions preserve the topology of the environment (holes included), and the z-scores against random baselines are strong. The noise robustness tables are systematic and convince me that moderate noise is tolerable. The 1D experimental result — median 3.6% error over 617 runs on the buildup track — is the best evidence in the paper and, for me, the main reason to take the work seriously. The authors also deserve credit for public code and data.\n\nNow the soft spots, in proportion. The stress-test note is correct: the 'up to an affine transformation' claim in the abstract is overstated for the 2D experiment. The global reconstruction visibly fails, and the reported low errors come from fitting a separate 6-DOF affine map per 40-second segment. That is not a single affine equivalence; it is a shape-similarity test on short smooth segments. The mismatched-pair baseline partially controls for this, but it does not test whether a single affine relationship holds across segments. The fix is easy: qualify the claim as 'locally, up to a segment-wise affine transformation.' The abstract already says 'local trajectories' later, so the paper is internally inconsistent.\n\nThe other weaknesses are real but not fatal. The epsilon threshold is chosen from the data histogram and cannot distinguish a fast genuine crossing from noise; the paper's own footnote admits the key assumption that time points are dense enough. The 2D experimental evidence rests on one dataset, one module, one day. And the paper does not compare against standard phase unwrapping or other decoding baselines, which would clarify what the topological framing adds beyond the obvious. None of these sink the central claim of the paper, but they matter for interpretation.\n\nWho should read this: anyone working on grid-cell decoding, neural manifolds, or spatial navigation. It provides a useful reference result and a public implementation. It deserves a serious referee. I would ask the authors to fix the abstract, add a held-out affine test for 2D, and compare with a baseline unwrapping method. Then I think the paper is publishable.","headline":"A clean, honest empirical paper showing that phase-unwrapping toroidal coordinates recovers local trajectories from a single grid module; the abstract overstates the affine claim in 2D, but the core result is worth engaging with seriously.","tokens_in":19961,"tokens_out":2607,"would_cite":true,"duration_ms":25390,"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":"The trajectory of a moving animal can be reconstructed from the activity of a single module of grid cells, without external position information or training, by lifting the population activity path from a torus to the plane; the reconstruct","keywords":["grid cells","path integration","toroidal manifold","persistent cohomology","path lifting","covering space","neural decoding","topological data analysis"],"falsifier":"Take a CAN-simulated trajectory in a 2D environment and deliberately downsample the recorded time points—or speed up the simulated motion—so that the consecutive toroidal angles skip past a tile boundary in at least one coordinate. If the reconstruction error then jumps to the level of random baseline pairs, the greedy lifting rule is the load-bearing assumption; if the error stays low, the method does not actually depend on the nearest-tile premise.","tokens_in":19109,"feed_emoji":"🧠","tokens_out":3865,"duration_ms":34132,"temperature":0.7,"pith_summary":"The paper claims that a single module of grid cells carries enough information to reconstruct the path an animal walks, even though each module's periodic code maps many physical locations to the same neural state. The key idea is to view grid-cell population activity as a path on a torus—the surface of a donut—and then 'unwrap' that torus into the plane using a standard topological operation called path lifting. No training, no external position data, and no knowledge of grid phases are needed; the recovered path matches the true trajectory up to a linear transformation (rotation, scaling, translation). Validation on both simulated and experimental recordings shows reliable reconstructions on short local segments and on one-dimensional tracks, suggesting that co-modular grid cells are sufficient for path integration. This is the first explicit demonstration that the toroidal organization of grid cells can be functionally leveraged for decoding.","feed_headline":"Grid cells alone can rebuild a tracked path","feed_subtitle":"No position data or training: lifting torus activity to the plane recovers the trajectory.","key_machinery":"Path lifting to a covering space: the torus S¹ × S¹ is covered by the plane R² via the map p(x, y) = ((cos 2πx, sin 2πx), (cos 2πy, sin 2πy)), which tiles the plane into congruent 'tiles'—copies of the torus. A path on the torus lifts, from a chosen starting tile, to a unique path in R², provided each segment between consecutive samples crosses at most one tile boundary in each coordinate. The greedy lifting rule compares consecutive toroidal coordinates: if their difference is below a threshold ε, keep the same tile; otherwise move to the adjacent tile that minimizes the distance. The integer offsets (M_t, N_t) accumulate the tile choices and define the lifted coordinates (θ_x + 2πM_t, θ_y","core_discovery":"The paper introduces a two-stage decoding framework. First, persistent cohomology confirms that grid-cell population vectors lie on a torus and assigns each time point a pair of circular coordinates, producing a discrete path on the grid-cell torus. Second, this path is lifted to the plane via the covering map p: R² → S¹ × S¹, which tiles the plane into copies of the torus. Consecutive samples are placed in the same tile if their toroidal coordinates are close; otherwise they are placed in adjacent tiles, with the direction chosen to minimize the distance to the previous lifted point. The resulting lifted path is the reconstructed movement trajectory. In idealized noiseless settings the lift","pith_inferences":["If the greedy nearest-tile rule is the true mechanism, reconstruction error should grow monotonically with movement speed or with downsampling of the neural time series; this is a direct, testable prediction that the current data only partially address.","Because the reconstruction is only up to an affine transformation, the method recovers the shape of the environment but not absolute scale or orientation; anchoring the lift to a known landmark or combining two modules with incommensurate scales could fix these factors—an extension the paper mentions but does not explore.","The method's success on a one-dimensional virtual track and on short 2D segments suggests that interpolation or probabilistic unwrapping could push global 2D reconstructions from local to whole-session accuracy, which would make the framework a candidate for brain-machine interfaces that require continuous spatial awareness.","The path-lifting framework generalizes beyond grid cells: any neural population whose activity is organized on a torus or other compact manifold with a universal cover could use the same 'unwrap-and-lift' idea to decode internally represented states."],"forward_implications":["The toroidal organization of grid-cell population activity is functionally relevant: it directly supports path integration without a learned decoder or external positional reference.","A single grid module—not multiple modules or phase information—is sufficient for local trajectory reconstruction, which narrows the requirements for neural decoding pipelines.","The method extends to realistic noise: moderate spontaneous firing preserves reconstructions, while noise that destroys the toroidal structure causes failure, setting a clear robustness boundary.","In one-dimensional environments, the pipeline recovers the topology and geometry of the track across hundreds of runs, suggesting it can serve as a model of how the brain solves the 'teleport' discontinuity problem in virtual navigation.","Local two-dimensional reconstructions are possible, but global reconstructions fail, indicating that error accumulation in the lift is the main limitation to be addressed."],"fun_headline_variants":["Grid cell torus lifting decodes paths without training","No external cues: grid cells decode path via covering maps","From torus to plane: grid cells alone rebuild trajectories","Path lifting on grid cell tori reconstructs movement","Topological lift of grid cell activity yields spatial path"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"Between two consecutive time samples, the true path on the torus crosses at most one edge in each coordinate, so that the greedy nearest-tile rule never mistakes a fast crossing for noise or misses a needed tile change.","fun_headline_variants_meta":{"raw":{"variants":["Grid cell torus lifting decodes paths without training","No external cues: grid cells decode path via covering maps","From torus to plane: grid cells alone rebuild trajectories","Path lifting on grid cell tori reconstructs movement","Topological lift of grid cell activity yields spatial path"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000236,"raw_usage":{"total_tokens":1315,"prompt_tokens":690,"completion_tokens":625,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":434,"completion_tokens_details":{"reasoning_tokens":547}},"tokens_in":434,"tokens_out":625,"duration_ms":5959,"temperature":1.0,"reasoning_tokens":547,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T09:16:27.561164+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a CAN-simulated trajectory in a 2D environment and deliberately downsample the recorded time points—or speed up the simulated motion—so that the consecutive toroidal angles skip past a tile boundary in at least one coordinate. If the reconstruction error then jumps to the level of random baseline pairs, the greedy lifting rule is the load-bearing assumption; if the error stays low, the method does not actually depend on the nearest-tile premise.","supporting_citations":[],"review_version":1}