acceptodds
Under review as a conference paper at ICLR 2027

A Group-Theoretic Perspective on Path Integration and The Origin of Grid Cells

Abstract

We develop a group-theoretic framework predicting that the neural representations supporting path integration reflect the mathematical structure of the space being navigated. Path integration is the process through which an animal estimates its allocentric position by integrating egocentric movements and is thought to be a core computation of the entorhinal–hippocampal system. We formalize path integration as _sequential group composition_ over finite groups, with different groups describing different spaces. Using group Fourier analysis we analytically construct recurrent neural networks that perform the computation exactly. These circuits decompose into modules indexed by irreducible representations (irreps) and achieve asymptotically optimal hidden width. For a group coupling translations and rotations in 2D space, our construction reproduces key features of the entorhinal grid-cell code: hexagonal spatial tuning, modules with distinct spatial scales, conjunctive position–heading tuning, and toroidal population geometry. Extending the construction to 3D navigation predicts a grid-cell code with irregular, yet locally structured, spatial tuning built from irreps of a 3D navigation group. Lastly, we introduce _irrep selectivity_, a group-theoretic generalization of gridness that identifies new units aligned to path integration tasks without requiring strict hexagonal symmetry. Taken together, our results provide simple accounts of detailed grid cell phenomenology and new ways for analyzing grid cell data.

Primary area:
applications to neuroscience & cognitive science
Keywords:
neuroscience, navigation, neural representations, group theory, Fourier analysis

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ykbxOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes a mathematical framework based on group theory, the group Fourier transform, and irreducible representations to understand why grid cells are a suitable representation for path integration. The authors formulate path integration as sequential composition of group elements, construct an RNN with interpretable computations and a provably asymptotically optimal number of hidden neurons under specific conditions, and then connect its neuronal activity patterns to grid cells, linking the mathematical description to a neural interpretation. The paper also offers predictions for 3D navigation and introduces irrep selectivity as a method for analyzing neural activity. I like this work overall. What I appreciate most is that the mathematical perspective is elegant and makes sense intuitively: on the one hand, we think of grid cells as supporting path integration; on the other, their periodicity and multiple spatial scales bring Fourier decomposition to mind. This work brings these two ideas, along with their relationship to group operations, into a single framework that can be analyzed mathematically. However, the paper really could be written more clearly, and the figures are also difficult to follow. Some symbols, definitions, and key steps are not explained well enough, so understanding the paper takes quite a bit of effort. It also needs to be clearer about which structures are chosen in advance, which results are derived, and what exactly “optimal” means. I would mainly view Section 5 as predictions about 3D representations, while Section 6 provides an interesting analysis tool and a preliminary application; their biological implications still need further validation. So, I would currently give the paper a 3. I appreciate the overall idea, and if these aspects are properly worked through, I would be willing to give it a 4.

Critical strengths
  1. What I appreciate most is that the overall mathematical perspective is elegant and makes sense intuitively. Grid cells support path integration, while their periodicity and multiple spatial scales bring Fourier decomposition to mind. This work brings these two ideas, along with their relationship to group operations, into a single framework that can be analyzed mathematically, offering an insightful way to understand the problem.

  2. The paper also turns this mathematical perspective into a concrete neural network construction. The network’s computations have an explicit interpretation, and its number of hidden neurons is asymptotically optimal under specific conditions. The authors then connect the activity patterns of these neurons to grid cells. I think this makes the link between the mathematical framework and neural representations more concrete, connecting the mathematical description to a neural interpretation.

Critical weaknesses

My biggest concern is that the hexagonal structure is already built into the model through the choice of a triangular lattice and sixfold rotational symmetry, so the resulting hexagonal tuning cannot serve as evidence explaining why grid cells have a hexagonal arrangement. What this shows is how to construct the corresponding representation given hexagonal symmetry; it does not answer why the navigation system would choose this symmetry. The paper also reports the emergence of multiple grid-cell modules, but this depends on how the model is divided into modules. The authors therefore need to explain this division more clearly, including which aspects are determined by the mathematical structure of the group and which depend on other choices made by the authors. For the latter, they need to explain why these choices would correspond to the actual neural system.

Questions for authors

I have questions about the following aspects:

  1. How exactly are the two reference activity vectors determined, and why are they chosen this way? The allocentric neural representation seems to be constructed as localized bump activity, which the paper connects to place cells and head-direction cells, but there is no corresponding explanation for the egocentric representation. Since these choices affect the subsequent network construction and experiments, the main text needs to explain them and their rationale clearly. When the vectors are first introduced, it should also point to the specific appendix location containing their detailed definitions.

  2. Fig. 2 needs a clearer explanation. The top and bottom rows look the same apart from their colors, yet the egocentric reference vector used in the actual experiments is not a Gaussian bump, while the figure makes it look like one. This distinction needs to be made explicit.

  3. Some symbols and key steps are not explained sufficiently, for example in Eq. (5). Could the authors clarify what the symbols represent when they first appear and fill in the reasoning connecting group representations, the group Fourier transform, and the neural activity encoding?

  4. Besides the illustrated position-selective, conjunctive, and heading-only units, what other activity types are present in the network? What proportion does each type account for, and are these distributions fully reported? The statistics described in the appendix cover only a subset with fixed labels. Please provide the existing counts and proportions for each category and explain how well this subset represents the full network.

2 comments

Thank you for your response. While I appreciate the clarification, my main concern is not fully resolved. Even if a triangular lattice is not sufficient by itself to produce hexagonal tuning, it already imposes a geometric structure closely related to the observed pattern. It therefore remains difficult to argue that hexagonal tuning emerges from path integration independently of the model’s assumptions.

My other questions have not yet been addressed. As most of my concerns remain unresolved, I will maintain my current rating for now.

Thank you again very much for your thoughtful review and for your follow up.

Hexagonal Tuning

We are glad we could partially resolve your main concern with the normative arguments for using a triangular lattice. Again we want to emphasize that simply using a triangular lattice is not sufficient to see hexagonal tuning, a Fourier-basis is necessary. While we too still have questions, for example as we discuss future work should try to extend our framework from finite to continuous groups, we think we have provided a very clean and elegant framework that is a step in the right direction.

Multiple grid-cell modules

In our framework a grid-cell module is a group of neurons selective to the same irrep of GG. The group GG determines the available modules and their spatial structure. The allocentric code xallorefx_{\mathrm{allo}}^{\mathrm{ref}} determines which are required for perfect path integration. For the reference code used in our experiments, modules at the ten largest distinct spacings and irreps with nρ<6n_\rho<6 together account for 94%94\% of its power. We have made this clear in the main and added a new panel D to Figure 7.

Additionally, we should emphasize that if we allow for the RNN to accumulate error over long sequences (which is natural and observed biologically), then we can get away with using less irreps than needed for perfect reconstruction. Specifically, for the Gaussian bump allocentric code we use in our simulations, we find that restricting our construction to only the first few low-frequency irrep modules maintains very low error. See our new Fig. 15 in App. F. These low-frequency modules are also the modules with the largest and most distinct spacing. In this updated Appendix F section we also discuss how irrep modules relate to the lattice spacing. From an irrep we can determine a mathematical expression for the average spacing, defined as the distance between neighboring firing fields. We can then sort irreps from large spacing (low-frequency) to small spacing (high-frequency). Lastly we can then compute the ratio of spacing between consecutive modules, as done in Stensola et al. (2012). We find two really interesting results: (1) the ratio is independent of nn the discretization of the lattice in Cn2⋊CmC_n^2 \rtimes C_m. (2) The average spacing ratio for the top four low-frequency modules (the most important ones for reducing path integration error) is 1.4031.403 with a standard deviation of 0.242. This is very close to the experimental estimate of 1.42±0.171.42 \pm 0.17 from Stensola et al. (2012), as reanalyzed by Wei et al. (2015). While this is just a connection for now – we think this is worthy of future exploration.

Reference codes and original Fig. 2.

We choose the allocentric reference as a localized spatial bump with nonuniform heading amplitudes, motivated by position- and heading-selective population activity. The egocentric reference is a generic random code, sampled to have invertible Fourier coefficients on the retained irreps. Its particular tuning shape is a modeling choice, without a claimed biological correspondence. Appendix F gives the definitions and sampling procedure, and we added a visualization in Fig.14. We would like to emphasize that while the egocentric reference will affect the weights, it actually doesn’t affect the tuning curves because the weights effectively remove it as seen in equation 7 by x^egoref[ρ]−†\widehat{x}_{\mathrm{ego}}^{\mathrm{ref}}[\rho]^{-\dagger}.

Fig. 2 illustrates the regular group action schematically: the egocentric bump shown there is not the random reference code used in the experiments. To avoid confusion we have removed it leaving just the allocentric reference code.

Notation

A full introduction to group theory and representation theory in the main would take too much space. Around Eq. (5) we now point to Appendix A “Groups and Their Fourier Transforms” which serves as a compact primer to these mathematical topics. We have added citations here to textbooks with formal introductions.

Activity types and sampling

Our classification distinguishes position-only, heading-only, conjunctive, and units unmodulated in both marginals. The counts and proportions are now shown in Fig. 18.

Thank you again for your review and for your follow up. We hope in light of these additional experiments and explanations you will raise your score and continue to advocate for the acceptance of our paper.

Reviewer_u7PYOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper casts path integration as the composition of group elements, where the same finite group indexes both allocentric states and egocentric updates. Using the group Fourier transform it constructs, without training, a squared-ReLU recurrent network that composes any sequence exactly, with hidden units organised into modules indexed by irreducible representations and a block-diagonal recurrent matrix. For the planar group, with six headings, the hidden units show hexagonal maps, modules of distinct scale, phase-shifted copies within a module, conjunctive position–heading tuning and a toroidal population geometry. It also sketches 3D predictions from orbits of plane waves under the tetrahedral group, and proposes irrep selectivity, an entropy of the power across irreps, and applies it to mouse entorhinal recordings.

Critical strengths

The central theorem is correct, and it could be checked all the way down. Summing the squared ReLU over the sign index makes each step bilinear in the two inputs, and the sum over the phase index removes the unwanted terms exactly. Eqs. 6 to 8 on C7² ⋊ C6 compose sequences exactly to 6 × 10⁻¹⁴ with three phases and failed with two, so the stated condition is both needed and sufficient.

Critical weaknesses

The hexagonal tuning follows from the chosen six-fold heading group, not from the structure of 2D space

Irrep selectivity is not validated, is read on a geometry that differs from the model's, and scores the paper's own clearest grid cells near the minimum

For the retained irreps the score is a spatial power spectrum binned into orbits, so it is not "a group-theoretic generalization of gridness"

The module claim does not hold as stated. With "all irreps retained." the network has 82 modules and about 50 scales, against the few discrete modules with a near-constant ratio of Stensola et al. (2012), and which modules appear is set by the Fourier content of a reference code that a Gaussian bump fills entirely

The torus of Fig. 6F is the periodic arena at the lowest frequency

Conjunctive tuning is partly built in

"naturally learned" (p. 5) rests on citations, and the construction is credited to prior work only in the appendix

The comparison with continuous attractor networks overstates the similarity. With all irreps retained, every direction of the decoded state is neutral, so nothing corrects noise off a low-dimensional manifold.

Questions for authors
  1. What does the predicted single-unit tuning look like for the spectral truncation of SE(2) proposed on p. 9, or for increasing heading resolution off the lattice? Do hexagons survive?
  2. Under the exact E.2 pipeline with your exclusion rule, what are the irrep-selectivity values of a synthetic ideal grid, a single field, a band cell and a smooth random map, and of a spatial-shuffle null?
  3. What is n in Fig. 8, which sessions and arenas were used, and what is the bin size in cm?
  4. Is the hypothesis of Proposition C.5 that units exist for every irrep, as in the simulation, rather than for the irreps in S?
  5. What is new in Theorem 3.1 relative to Marchetti et al. (2026) and Kunin et al. (2026)?
  6. Does the persistent homology survive for module ρ = 21 on a non-periodic sub-window of the arena?
Reviewer_3sbnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper develops a group-theoretic framework for path integration by representing allocentric states and egocentric movements as elements of a finite group. Using group Fourier analysis, the authors construct a squared-ReLU recurrent neural network that exactly performs sequential group composition. The circuit decomposes into modules associated with irreducible representations and achieves asymptotically optimal hidden-width scaling for the navigation-group families considered.

The authors specialize this construction to planar translation–rotation groups and demonstrate grid-like spatial tuning, multiple spatial scales, conjunctive position–heading responses, and toroidal geometry of averaged population representations. They further discuss predictions for three-dimensional navigation and introduce irrep selectivity, a measure of spectral concentration over irreducible representations, which they compare with conventional gridness using existing MEC recordings.

Critical strengths

1.An explicit and interpretable connection between navigation structure and recurrent computation. The translation–rotation formulation captures the dependence of an egocentric movement’s effect on the current heading. The Fourier-based construction then translates this structure into a concrete recurrent circuit with interpretable modules. The width analysis in Appendix C.3 provides an efficiency guarantee within the specified encoding and linear-readout setting. Together, these results offer a useful analytical model for studying how structured navigation computations can be implemented.

2.A coherent framework connecting several neural phenomena and suggesting a complementary analysis measure. The paper relates spatial periodicity, module-dependent scales, conjunctive tuning, and population geometry within one construction. Irrep selectivity is also a potentially useful extension beyond strictly hexagonal tuning, especially for studying heterogeneous navigation-related responses. The appendices transparently document several modeling and estimation choices, making the scope of these results easier to assess.

Critical weaknesses

1.The navigation-specific explanatory contribution is not yet sufficiently established. The planar construction assumes a triangular translation lattice and C6C_6 rotations, while Figures 11–12 show that choosing C4C_4 instead produces square periodicity. The results therefore establish how a selected symmetry generates corresponding tuning, but do not establish why path integration should favor hexagonal organization. The width-optimality result does not distinguish these alternatives. This limitation matters because the cited SGC work of Marchetti et al. (2026) already develops an irrep-organized recurrent construction based on sequential composition, as acknowledged in Appendix C. The paper would benefit from more clearly identifying which additional findings explain biological organization beyond this inherited construction and the chosen geometric assumptions.

2.The empirical analyses do not yet sufficiently validate the proposed biological interpretation. Appendix E.2 directly interprets rectangular spatial-bin indices as triangular-lattice coordinates, without transforming the physical positions. The resulting group action therefore does not directly correspond to Euclidean rotations in the recorded arena. IS and gridness also use different preprocessing, and their percentile cutoffs are descriptive rather than significance thresholds. Consequently, high-IS/low-gridness examples do not yet establish reliable additional navigation-related structure. Separately, Figure 6 uses local movement averaging for single-cell tuning but global averaging for population geometry; the torus obtained from averaged positional tuning does not establish the same geometry for activity during local navigation. These choices limit how strongly the empirical results support a unified account of biological tuning and population organization.

Questions for authors
  1. Clarification of the explanatory scope and incremental contribution. What principle, if any, favors C3/C6C_3/C_6 over C4C_4 within the proposed framework? Please distinguish results inherited from the existing SGC construction from new navigation-specific results, and clarify which aspects of grid-cell organization are predicted rather than assumed through the choice of group and representation. This would materially affect my assessment of the explanatory contribution.

  2. Validation of irrep selectivity under geometry-consistent and matched analysis conditions. Does the high-IS/low-gridness population persist when the spatial representation is made consistent with physical rotation geometry and the two scores use matched sample selection? Can the authors report split-half reliability and an occupancy-aware null analysis for these cells? These checks would help distinguish reproducible tuning structure from effects of coordinates, preprocessing, and sampling.

  3. Clarification of the population-geometry result. Does the toroidal structure persist when population activity is analyzed under the local movement distribution used for the single-cell results? Please distinguish the geometry of averaged positional tuning from that of instantaneous activity, including variation associated with heading and movement. This would clarify whether the different phenomena in Figure 6 are supported under a common behavioral condition.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revised Manuscript

We thank all three reviewers for their careful reviews and follow-up discussion, and apologize for the delay in providing our revision and detailed responses. Your feedback has led to the following changes:

  • Scope and assumptions: We removed "and the origin of grid cells" from the paper title clarified the assumptions behind hexagonal tuning, and added the two normative arguments for choosing a triangular lattice (Section 5).

  • Irrep selectivity: We rewrote this material as an exploratory proposal in the Discussion and removed it from the abstract and list of contributions.

  • Angular path integration: A new section applies the same framework to head direction, connecting sinusoidal Fourier weights to EPG-to-Δ7 connectivity in Drosophila (Section 4; Appendix E).

  • Planar analyses: We added calculations of module spacing and spectral power, visualized the reference codes, and compared population geometry across modules under local and global movement distributions (Appendix F).

  • Presentation and attribution: We clarified notation, the relationship to prior SGC constructions, and the distinction between our circuit and continuous attractor networks.

Our individual responses address your specific questions in more detail. Thank you again for helping improve the paper. We hope you will reconsider your scores in light of these revisions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.