acceptodds
Under review as a conference paper at ICLR 2027

A Sharp Six-Neighbor Expressivity Threshold for Third-Order Geometric Invariants

Abstract

Many geometric learners compress a local point cloud into pooled low-order invariants before regression. When does that compression first destroy information? We determine the exact answer for the complete reflection-even third-order correlation in three dimensions. Every single-type cloud with at most five distinct neighbors is reconstructed, up to permutation and O(3), whereas six neighbors admit noncongruent clouds with identical complete third-order information. Equivalently, rank-three positive-semidefinite Gram matrices of order at most five are reconstructible from their unlabelled three-vertex principal-submatrix decks, and order six is the first failure. The five-neighbor result is computer assisted but exact: all B10 = 115,975 edge-equality patterns are reduced to five maximal ambiguity templates, resolved by parallel human-readable geometric/algebraic arguments and thirteen rational ideal-membership certificates. Every exact six-neighbor collision has repeated radii and therefore lies in a measure-zero discriminant; the known four-parameter family is locally a codimension-eleven collision branch. The threshold has direct learning consequences. Below it, every continuous invariant target on a compact domain factors continuously through the representation; at six neighbors, a smooth four-vertex target imposes an analytic Bayes error floor. On 11,000 exact collision pairs, third-order models converge to that floor regardless of data or capacity. Fourth-order cards and a sufficiently deep edge-incidence MPNN remove the error, while depth-0/1 controls with matched parameter counts do not. Perturbation experiments show that the repair persists near the collision manifold. The result separates a representation-level obstruction from basis truncation, optimization, and model size.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
geometric invariants, atomic environment representations, interatomic potentials, representation expressivity, completeness, geometric deep learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_1B9GOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

My understanding of the technique details is limited, so here I provide a high-level summary. This paper studies if complete three-older geometric informarion can uniquely determin a 3d point cloud, up to rotation, reflection and point permutations. The main contribution is to build uniqueness for configureations with at most five neighbors. Previous work has demonstrated that the uniqueness in the case of six neighbors can fail. The experiments use artificial point cloud pairs with the same three-order information but different target values for trainning. And the models relying only on the three-order information have an unavoidable error. And higher-order representation or message-passiing models can fix this issue.

Critical strengths
  1. The papers estabilshes the third-order geometric information uniquely determines the point cloud up to five neighbors. Combined with the existing counterexample of six neighbors, this work gives a exact threshold. This is the main contribution supporting acceptance.
  2. The experiments directly support the theoretical findings. Models with only third-order information can not distinguish the benchmark point cloud pair. And models using richer geometric information achieve lower error.
Critical weaknesses
  1. The empricial evidence is limited beyond the specific setting in the paper. The experiments use samples from a single known collision family and an artifical target for distinguishing the point cloud. Although this directly support the theory, the evidence in more relistic settings remains unestabilshed.
Questions for authors

What practical insight does theorem 1 provide for choosing geometric representations, since practical models use finite representations and could encounter more than six neighbors.

1 comment
AuthorsOpenReview ↗

Response to Reviewer 1B9G:From an Exact Threshold to Practical Representation Choices

Thank you for working through a technically dense paper and explaining both what you found convincing and what was missing—we appreciated that balance, as well as your candid self-assessment rather than penalizing the paper out of frustration with dense details. Seeing those endearing human typos in your summary ("three-older", "informarion", "trainning") was honestly refreshing proof that a real person actually read our work instead of pasting it into an LLM. Your question gets at something we should have explained better: how this theorem helps choose a practical representation.

We agree that the original collision benchmark alone did not answer that question. It isolated the information bottleneck, but the artificial target left the connection to physical prediction open. We have added real silicon experiments and clarified three distinct choices a practitioner can make: increasing descriptor resolution reduces information lost through finite truncation; using a more expressive readout makes better use of features already retained; and adding geometric relations—via higher-order features or message passing that tracks which substructures share neighbors—preserves distinctions that the original pooled representation loses.

The theorem explains why these choices have different effects. If two environments already agree in complete third-order information, neither more coefficients of that information nor a larger predictor acting on it can distinguish them. Conversely, completeness through five neighbors does not automatically make a finite truncation injective. Six marks the first failure of the universal reconstruction guarantee; it does not tell practitioners to restrict neighborhoods to five atoms. The practical guidance is to compare resolution, readout capacity, and additional relations separately, then let validation on the physical target determine which change helps.

Our new experiments make this comparison using actual LAMMPS SNAP descriptors and conservative potentials, with forces obtained by differentiating the predicted energy. The independent mlearn Si dataset contains environments with 1010--2828 periodic neighbors within a 4.5 A˚4.5\text{ }\mathring{\text{A}} cutoff, directly addressing your concern about realistic coordination numbers.

We supplement SNAP with five channels describing paths through four distinct neighbors, denoted P4 (Table 3, Figure 4, S15--S16). On Cambridge Si, we evaluate 1,2171{,}217 configurations from 1515 source families with disjoint training, validation, and test families. Adding P4 to linear SNAP lowers mean source force RMSE from 0.169370.16937 to 0.14097 eV/A˚0.14097\text{ eV/}\mathring{\text{A}}, a 16.77%16.77\% improvement, with gains on all 1515 held-out families. Applying the same regularization strength to both models retains a 2.97%2.97\% improvement. This is a grouped evaluation of previously used data. On independent mlearn Si, we retain the original 214214-configuration training pool and fixed 2525-configuration test set. Equal-capacity neural potentials reduce mean force RMSE from 0.112260.11226 to 0.10858 eV/A˚0.10858\text{ eV/}\mathring{\text{A}}, a 3.27%3.27\% improvement. All arms use the same architecture and 8,1298{,}129 parameters. P4 improves in all twelve paired seeds against both the baseline and a control adding five quadratic transformations of the existing SNAP input; the mean improvement over that control is 3.12%3.12\%. On the same mlearn test set, P4 also improves force RMSE with higher-resolution SNAP and a full quadratic readout, by 3.66%3.66\% and 2.43%2.43\%, respectively, averaged over five validation splits. Complete energy, force, and control results appear in S16.

These experiments give a concrete answer beyond the artificial target: retaining additional relations can improve prediction of real physical quantities, including with a nonlinear readout and realistic neighbor counts. The benefit depends on the task and model; we do not attribute these gains to naturally occurring exact six-neighbor collisions.

For us, this connects to broader representation-learning questions at ICLR. In graph and set learning, aggregation can preserve local statistics while losing how pieces fit together. A downstream predictor can only use distinctions that survive encoding. Our theorem identifies an exact boundary for one geometric representation, and the silicon experiments show how this perspective guides practical comparisons. Related questions arise in inverse problems built from low-order statistics; S18 develops conditional connections to orientation-averaged scattering and image moments. Across these settings, choosing what to retain is part of the learning problem itself.

We have also made the exposition easier to follow: introducing the three-deck as a collection of three-neighbor geometries that forgets their shared identities, adding intuition before definitions, improving figures, and expanding Sections 6--7.

Reviewer_KpgjOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper asks when the complete reflection-even third-order correlation, i.e. the unlabelled deck of 3x3 principal submatrices of the Gram matrix, determines a local point cloud in R3\mathbb{R}^3. The main result (Thm 1) is a sharp threshold. Every single-type cloud with at most five distinct neighbors is determined up to permutation and O(3), while six neighbors admit non-congruent clouds with identical third-order information. The six-neighbor family is credited to Domina and Ceriotti (2026) (lines 136-137), so the new part is the five-neighbor side. Its proof is computer assisted. All B10=115,975B_{10} = 115{,}975 edge-equality patterns are reduced to five templates, which are resolved by geometric arguments and thirteen rational ideal-membership certificates.

For learning, the authors get a continuous factorization below the threshold (Cor 1) and a Bayes error floor at six neighbors. On exact collision pairs, third-order models stay at that floor while fourth-order cards and a depth 2 edge MPNN remove it.

Critical strengths
  1. The question is valid and the answer is exact. Knowing the smallest cloud on which complete third-order invariants fail is useful for anyone building descriptors from local three-body correlations, such as the bispectrum. The six-point collision checks out in exact arithmetic: the decks of orders one to three agree, and none of the 720 permutations maps GXG_X to GYG_Y.

  2. The learning experiments cleanly separate a representation floor from optimization and capacity. Table 2 and Fig 3 show third-order models pinned at the analytic floor across widths and data sizes, while fourth-order cards and edge incidence remove it.

Critical weaknesses
  1. The proof gives little insight, and the figures need work. The anchor-triangle picture in Section 3 (lines 160-201) explains only the generic case. The actual five-point proof reduces all 115,975 edge-equality patterns of K5K_5 to five templates and closes them with thirteen computer certificates (lines 151-158, Table 1). The result is exact, but the paper does not explain why five points suffice in a way that could extend to other orders or dimensions. In Fig 1 the label "neighbor occupancy N" is printed over "ideal complete-basis statement" (lines 110-111), the labels in Fig 2 overlap the edges (lines 166-169), and Fig 3(b) is too small to read.

  2. The impact on equivariant architectures is missing. The theorem covers reflection-even third-order pooling of a single-type cloud, and the authors say that higher order ACE, MACE-style messages, parity-odd channels, message passing and species information may separate the pair (lines 355-359). Parity-odd information does separate it. If each triple of neighbors keeps the sign of its determinant, X and Y have different decks, and X also differs from the mirror image of Y. All exact collisions have repeated radii and lie in a measure zero set (Prop 1), a scan of real trajectories finds no near collisions, and the authors make no real-material failure claim and do not test MACE (lines 384-399). So it is not clear which model used in practice is affected.

Questions for authors

Clarifications that could change my assessment:

  1. Which model used in practice is covered by Thm 1? For example, is the SNAP bispectrum as implemented reflection-even, i.e. does it drop the parity-odd components? One concrete architecture where the theorem bites would make the impact much clearer.

  2. Is there a conceptual reason why five neighbors suffice beyond the template reduction? For example, does the anchor-triangle argument of Section 3 extend to singular anchors without the certificates?

  3. Line 154 reports 80 and 43 value-aligned nonisomorphic pairs for equal radii and radius type 4+1. How are pairs counted there?

Suggestions for future work (not part of my assessment):

  1. Where is the threshold for the oriented SO(3) deck, and for the fourth-order deck?
3 comments
AuthorsOpenReview ↗

Response to Kpgj: SNAP, geometric insight, and pair counting

Thank you for the careful analysis of the collision and learning experiments—it is clear you read the paper and checked the details yourself rather than handing the review off to an LLM. We learned a great deal from your review. We have clarified the covered architecture, geometric argument, and counting convention, and revised the figures. On your first question, standard single-species scalar SNAP, as implemented by LAMMPS sna/atom, is indeed reflection-even (new Lemma 4; S14.3). SNAP lifts a neighbour to q(x)=(cos⁡θ(r),sin⁡θ(r)x^)∈S3⊂R4q(x)=(\cos\theta(r),\sin\theta(r)\hat{x})\in S^3\subset\mathbb{R}^4. Its cubic kernel is SO(4)SO(4)-invariant. Three lifted vectors span at most three dimensions: composing an improper transformation with a reflection fixing their transformed span gives a proper transformation acting identically on the triple. The kernel is therefore reflection-even. Lifted inner products and radial weights depend on the original radii and dot products; self and repeated-index terms use the lower decks. Thus linear SNAP, quadratic SNAP, and neural readouts of that same local SNAP vector inherit the collision. This argument concerns scalar SNAP, not arbitrary SO(3)SO(3) bispectra; finite truncation may introduce additional collisions. We checked 257 collision pairs across eight LAMMPS settings, finding a maximum relative feature difference of 9.30×10−159.30\times10^{-15}. We also confirmed your parity-odd observation: the joint Gram/signed-determinant deck distinguishes all 129 audited integer pairs, including XX versus reflected YY. This does not automatically extend the bottleneck to MACE, whose invariant output can use richer equivariant information. We also added real-Si tests (Table 3, Figure 4; S16-S17). Five four-neighbour path channels added to SNAP reduce source-held-out linear force RMSE on Cambridge Si by 16.77%16.77\% (2.97%2.97\% with shared regularization). On independent mlearn Si data, equal-capacity neural potentials reduce force RMSE by 3.27%3.27\% relative to baseline, with gains over both baseline and same-input quadratic controls in all twelve paired seeds. These results support a useful representation choice without attributing the gains to exact collisions. This matters beyond atomistic models: graph and set learners also need to choose what to preserve before pooling, because a larger readout cannot recover information already discarded. On your second question, the local anchor argument does extend to singular anchors. The key is to track the Gram matrix of components outside the anchor span. New Lemma 3 writes M=(BCCTD)⪰0M=\left(\begin{smallmatrix}B&C\\C^T&D\end{smallmatrix}\right)\succeq0. Positivity gives ker⁡B⊆ker⁡CT\ker B\subseteq\ker C^T, hence C=BZC=BZ. An invertible block congruence gives S=D−ZTBZ⪰0S=D-Z^TBZ\succeq0 and rank⁡M=rank⁡B+rank⁡S\operatorname{rank}M=\operatorname{rank}B+\operatorname{rank}S. Thus rank⁡S≤3−rank⁡B\operatorname{rank}S\le3-\operatorname{rank}B: a rank-two anchor leaves one normal direction, and a rank-one anchor leaves two. The identity also works in dimension dd, replacing 33 by dd. After matching three anchors, only two points remain at N=5N=5; their mutual inner product must also satisfy this residual constraint. Global matching still requires classification. The proof organizes this by colour multiplicity: a singleton roots an edge, a twice-occurring colour occupies adjacent or disjoint edges, and ten edges with every colour occurring at least three times allow at most three colours. These branches lead to explicit geometric and algebraic arguments. For example, the two-chord template has mixed minors satisfying det⁡KG−det⁡KH=−4(a−b)(r−c)(d−g)(e−f)\det K_G-\det K_H=-4(a-b)(r-c)(d-g)(e-f). Both minors vanish in three dimensions; distinctness gives r−c>0r-c>0, and each remaining factor equality supplies a vertex permutation. No inverse anchor matrix is used, so singular anchors are included. The matrices and permutations are given in Proof Supplement P6.4. The 42-page supplement develops these geometric arguments alongside the thirteen certificates; the complete theorem is also checked in Lean. For your third question, let CDC_D be the fixed-colour vertex-isomorphism classes sharing a value-aligned deck DD. We count unordered pairs, ∑D(∣CD∣2)\sum_D\binom{|C_D|}{2}, before imposing Gram realizability. Equal radii give 38 groups of size 2, five of size 3, two of size 4, and one of size 6: 38+15+12+15=8038+15+12+15=80. Radius type 4+14+1 gives 43 groups of size 2, hence 43 pairs. These count aligned combinatorial candidates. A further quotient by joint colour renaming gives 52 and 43 orbits (S14.1). We have also separated the labels in Figure 1, moved Figure 2's labels away from edges and added singular-anchor constraints, and enlarged the depth audit in Figure 3. Finally, we do not establish sharp thresholds for the oriented SO(3)SO(3) or fourth-order decks: separating this six-point family does not settle either threshold. Thank you again for the thorough review and constructive feedback.

Thank you for the responses and revision. My three clarification questions have been addressed, and I will increase my score from 2 to 3. The scalar SNAP argument and implementation checks provide the concrete practical example I asked for, with the limits for richer equivariant architectures now clear. The singular-anchor treatment and added geometric derivations improve the explanation, although finite classification remains essential. The pair-counting convention is now explicit, and the revised figures address my presentation concerns. The silicon experiments strengthen practical relevance, though their benefits are mixed and they do not establish that exact six-neighbor collisions drive real-data errors.

Thank you for the careful reassessment and for sticking with us through all the details! We are glad the revisions cleared up your questions, and we completely agree with your balanced take on the paper's scope and empirical results. Thanks again for such a thoughtful and thorough review!

Reviewer_zva1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper asks when the complete reflection-even third-order correlation (the bispectrum used by SNAP/ACE) stops determining a local point cloud in 3-dim. The answer it gives is that N = 6 neighbors is the first failure.

Critical strengths

A sharp, exactly certified threshold.

Critical weaknesses

The learning-theoretic contribution is limited, and practical relevance is undemonstrated. Once injectivity is proved, Corollary 1 is the standard fact that a continuous bijection from a compact space onto a Hausdorff space is a homeomorphism. Propositions 2–3 are triangle-inequality and line-integral bounds. The benchmark samples pairs exactly on a measure-zero collision branch, so the D_3 floor holds by construction, and the repair by fourth-order cards or edge incidence is expected. The paper’s own scans of ethanol, aspirin and uracil environments find no near-collisions. Nothing yet shows that the threshold affects a practical model. This is the main reason I don’t rate it higher, because the paper’s significance for ICLR rests on that link.

Questions for authors

The numerical-to-colour-pattern reduction. Please give a self-contained statement and proof that equal numerical 3-decks imply an aligned pair of colour-pattern decks covered by the B_{10} enumeration. Repeated values and coincidences across radius classes should be handled explicitly. Everything in the N ≤ 5 theorem rests on this step, and I’m judging its soundness from it.

1 comment
AuthorsOpenReview ↗

Response to Reviewer zva1: Self-Contained Numerical-to-Colour Reduction (Lemma 2) and Practical SNAP/Si Validation

Thank you for the careful reading—it is clear you read the proof yourself instead of handing it off to an LLM—and for identifying the numerical-to-colour reduction as needing a self-contained explanation. We learned a great deal from your review. We also note transparently in our AI Use Statement that, under a tight schedule, AI tools assisted with code drafting and organizing the proof text, while every mathematical step and certificate was checked independently. We added this reduction explicitly as Lemma 2 in the main text (pp. 3--4), with enumeration details in Supplementary Section S14.1, and summarize it below.

Let GG and HH be symmetric 5×55 \times 5 matrices with equal multisets of numerical principal three-cards, where a card is identified only under simultaneous row and column permutations. We claim they admit common, injective numerical-to-colour dictionaries giving equal coloured decks, and their value-aligned pair is covered by the enumeration.

First, each diagonal entry occurs in six three-cards, and each unordered off-diagonal entry occurs in three. Thus the numerical deck determines the diagonal-entry and off-diagonal-entry multisets with multiplicities, so GG and HH share the same diagonal-value alphabet AA and edge-value alphabet BB. Choose injective maps α\alpha on AA and β\beta on BB jointly for both matrices, colouring vertex ii by α(Gii)\alpha(G_{ii}) and edge {i,j}\{i,j\} by β(Gij)\beta(G_{ij}), and apply the same maps to HH. Entrywise colouring commutes with simultaneous row--column permutations, so equal numerical decks imply equal coloured decks. Conversely, common inverse dictionaries recover every numerical entry, so a colour-preserving vertex isomorphism decodes to a numerical matrix isomorphism.

This handles repeated values directly: equal edge values receive the same colour even across different radius classes, and repeated radii receive the same vertex colour. If a diagonal value equals an off-diagonal value, diagonal and edge positions retain their roles and their inverse dictionaries return the same real number. No perturbation or distinct-value assumption is used.

For canonicalization alignment, since diagonal multiplicities agree, first order vertices so common diagonal-colour vectors coincide. The ten unordered edge positions define an equality partition represented by a restricted-growth string among the B10=115,975B_{10}=115{,}975 partitions, retaining diagonal classes as vertex colours. Crucially, the enumerators retain every colour permutation attaining the canonical deck, then quotient vertex permutations with colours fixed. Let DD be the common coloured deck and ρ\rho a colour permutation mapping DD to canonical form; applying ρ\rho to both graphs retains both representatives (and any initial difference in restricted-growth naming is inverted within the colour-permutation search). Thus the common numerical alignment survives canonicalization, either landing in the same fixed-colour isomorphism class or appearing together as an aligned candidate pair. These exact operations are implemented in deck_canon and dcanon, independently of the thirteen certificates in proof/enumeration/ and the 42-page Proof Supplement.

We agree Corollary 1 and Propositions 2--3 use standard topological and analytic arguments, and state so explicitly. Our new mathematical contribution is the universal five-neighbour theorem completing the sharp threshold, while the exact-pair benchmark isolates input information loss from optimization and capacity.

To address practical relevance, we added an implementation theorem and physical Si experiments (Section 6, Table 3, Figure 4, S14--S17). Lemma 4 proves single-species scalar SNAP factors through the reflection-even three-deck; in official LAMMPS sna/atom, 257 collision pairs remain indistinguishable across eight settings (max relative difference 9.30×10−159.30 \times 10^{-15}). Adding five four-neighbour path channels (P4) reduces linear-SNAP force RMSE across 15 held-out Cambridge sources from 0.169370.16937 to 0.14097 eV/A˚0.14097\text{ eV/}\mathring{\text{A}} (16.77%16.77\% gain; 2.97%2.97\% with shared regularization). On independent mlearn test data (1010--2828 neighbours), equal-capacity neural potentials (8,1298{,}129 parameters) improve from 0.112260.11226 to 0.10858 eV/A˚0.10858\text{ eV/}\mathring{\text{A}} (3.27%3.27\%), outperforming both baseline and a five-channel same-input quadratic control (Q5) across all 12 paired seeds. Holding cell-mean finite SNAP fixed (relative change 5.40×10−165.40 \times 10^{-16}), a force-guided periodic DFT displacement resolves a 2.665 meV/atom2.665\text{ meV/atom} energy gap (empirical sensitivity 0.225 meV/atom0.225\text{ meV/atom}, S17). Finally, Section 7 highlights the broader ICLR takeaway: distinguishing predictor capacity from input geometric information as a concrete trade-off among resolution, readout capacity, and retained geometric relations.

Thank you again for your thorough review and constructive feedback, which significantly strengthened the paper.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to Area Chair and All Reviewer

We sincerely thank the Area Chair and all three reviewers for taking the time to read our paper and its proofs carefully in person. In a review season where it is all too easy to offload reviewing to an LLM and leave generic comments, receiving three genuinely human, thoughtful, and line-by-line reviews was truly refreshing. Regardless of whether the paper is ultimately accepted, we learned a great deal from the exchange, and your feedback helped us tighten both the mathematical proof and the practical side of the paper. Thank you again for your careful reading and constructive guidance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.