{"id":"e9453e94-bb93-4aac-bd27-171ef95c892c","arxiv_id":"2501.06197","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The ANM Hessian is shown to be the Laplacian of an 'anisotropic sheaf', zero eigenmodes are its global sections, and Delaunay-based 1-skeleta guarantee exactly six trivial modes.","lead":"The paper recasts the standard anisotropic network model of protein motion as a cellular sheaf, and proves that the model's Hessian matrix is exactly the sheaf Laplacian. It then shows that graph constructions based on Delaunay triangulation, and a minimal greedy construction, guarantee the six rigid-body zero modes that protein dynamics models want.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1's step 6 existence is unproven: Corollary B.6 applies to a single simplex, not the multi-tetrahedron complex, and no convexity argument is given; the six-mode/minimality claim for Algorithm 1 is unsupported as stated.","rationale":"Independent review confirms Theorem 3.1 (sheaf Laplacian = ANM Hessian) and the global-section interpretation are correct. The rank-induction in Theorem 5.6 for admissible complexes also checks out: each newly exposed vertex in the face-adjacency expansion has three edges to the previous layer, and those three rows are linearly independent because the new vertex is the apex of a nondegenerate tetrahedron. The Delaunay corollary follows if Proposition 5.5 is accepted. The unresolved piece is Algorithm 1: the existence of the 2-simplex in step 6 is asserted with a citation to a lemma that applies to a single simplex, and no argument shows that the growing complex remains convex or that the nearest-point selection prevents unintended intersections. The reader's weakest_assumption identifies exactly this gap. A failed run would refute the minimal-graph construction but would not damage the sheaf/Hessian results, so the appropriate verdict remains CONDITIONAL.","tokens_in":42160,"tokens_out":41100,"duration_ms":373136,"concrete_test":"Implement Algorithm 1 exactly as written for several 3D point clouds in general position, including clouds with points inside the convex hull of the first four selected vertices (e.g., random clouds with |V| = 8..20). At every iteration, verify that the chosen 2-simplex σ satisfies conv(σ, vi) ∩ K_{i-1} = σ. If any run fails to find such σ, or if the final Hessian nullity exceeds 6, the constructive claim in Algorithm 1 is refuted. As a complementary analytical check, attempt to prove the existence lemma for a nonconvex K built by one prior external tetrahedron attachment; a concrete counterexample there would settle the gap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing gap is in Algorithm 1, step 6 (and the paragraph after it): the algorithm must choose a boundary 2-simplex σ such that conv(σ, vi) ∩ K_{i-1} = σ. The paper asserts this 'exists by the proof ... using Corollary B.6' and 'holds due to the minimal distance property.' Corollary B.6 proves only that a point outside a single 3-simplex can be separated by one of its face hyperplanes; it says nothing about a multi-tetrahedron complex. Moreover, K_{i-1} is not shown to be convex: the incremental construction adds exactly one tetrahedron per vertex, and attaching a tetrahedron over one boundary triangle of a convex tetrahedral complex can produce a nonconvex complex. For a nonconvex complex, an external point's nearest point may lie on an edge or in a concavity where every candidate tetrahedron conv(σ, vi) intersects some non-adjacent part of K_{i-1}, so Eq (19) can fail. If step 6 fails, the constructed complex need not be admissible, and the six-mode and minimality conclusions for Algorithm 1 (Theorem 4.5) are unsupported. The core sheaf/Hessian equivalence and the Delaunay theorems 5.6-5.7 do not depend on this step.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a cellular sheaf, the 'anisotropic sheaf,' on an undirected graph whose vertices carry position data in R3, and shows that its 0-th sheaf Laplacian equals the Hessian matrix of the anisotropic network model (ANM) when edge weights are chosen as gamma^{1/2}/s_ij (Theorem 3.1). It then identifies the kernel of this Laplacian, equivalently the space of global sections of the sheaf, with the zero-eigenvalue normal modes of the ANM, proves that this space has dimension at least six for any non-collinear configuration (Theorem 4.1), analyzes the rank of the coboundary map for complete graphs (Appendix A), and proves existence of an edge-minimal graph with exactly six zero modes (Theorem 4.5). The final section introduces 'admissible' homogeneous simplicial 3-complexes, proves that the 1-skeleton of any such complex induces a Hessian with exactly six zero eigenvalues (Theorem 5.6), derives that every 3D Delaunay triangulation of a general-position point cloud is admissible (Corollary 5.7), and gives an incremental algorithm (Algorithm 1) intended to construct a minimal graph with exactly six zero modes.","tokens_in":42526,"tokens_out":23244,"duration_ms":221183,"significance":"If the main results hold, the paper gives a clean sheaf-theoretic reformulation of ANM: the Hessian is a sheaf Laplacian, and the six trivial modes are exactly the global sections of the anisotropic sheaf. The explicit identification in Theorem 3.1 is a direct, checkable computation, and Theorem 4.1 provides an explicit basis of six global sections. The Delaunay-based result (Corollary 5.7) is potentially useful in practice, as it gives a topological guarantee that the 1-skeleton of a Delaunay triangulation produces the desired six zero modes without a cutoff-distance search. The connection to force cosheaves and Maxwell's rule is also suggestive. However, the paper's constructive claims, especially Algorithm 1, are not yet supported by the arguments given. The central sheaf-Hessian equivalence is solid, but the algorithmic contribution as stated has serious gaps.","major_comments":[{"comment":"As written, Eq. (19) cannot be satisfied: v_i is a newly chosen point and is not a vertex of K_{i-1}, so the right-hand side {v_j1, v_j2, v_j3, v_i} cannot equal the intersection with K_{i-1}. If the intended condition is conv(v_j1,v_j2,v_j3,v_i) ∩ K_{i-1} = {v_j1,v_j2,v_j3}, or equivalently that the new tetrahedron meets K_{i-1} only in the 2-simplex σ, then the existence of such a 2-simplex is not established. The citation to Corollary B.6 is insufficient because that result separates a point from a single 3-simplex, not from a multi-tetrahedron complex, and K_{i-1} is not shown to be convex. More seriously, the nearest-point rule in step 5 can select an unprocessed point lying in the interior of the current tetrahedral complex (distance zero); for such a point no boundary 2-simplex satisfies the intended non-overlap condition. Thus Algorithm 1 does not provide the promised construction for a general-position point cloud, and the six-mode and minimality conclusions attributed to Algorithm 1 are unsupported.","section":"§5, Algorithm 1 and Eq. (19)"},{"comment":"The proof of Theorem 5.6 relies on the assertion that 'each vertex in V2 \\ V1 contributes at least three additional linearly independent rows to the extension matrix C(2) from matrix C(1).' This is a load-bearing rank-increase lemma, but it is only stated, not proved. While the statement is plausible and can be justified by considering the three new rows whose nonzero entries at the new vertex are linearly independent vectors (because the shared face is a nondegenerate 2-simplex), the proof as written is a one-sentence assertion rather than an argument. This gap should be fixed by stating and proving the lemma explicitly, since Theorem 5.6 depends on it at every filtration step.","section":"§5, Theorem 5.6, rank-increase step"},{"comment":"In the proof that a Delaunay triangulation is admissible, the finite walk across 2-adjacent tetrahedra is terminated by the statement 'This process terminates because L is a finite simplicial complex.' Finiteness alone does not rule out a cycle in the dual graph of the component L. A rigorous termination argument is needed, for example by showing that the chosen separating face moves the walk monotonically with respect to a potential such as distance from the point x0, or by using a boundary-adjacency tree argument. Without this, the proof that a boundary face with deg_L,u(τ)=1 is reached is incomplete. This matters because Proposition 5.5 is the basis for Corollary 5.7.","section":"§B, Proposition 5.5"}],"minor_comments":[{"comment":"The heading 'Anisotropic network and Hassien matrix' contains a typo; it should read 'Hessian matrix.'","section":"§3, section heading"},{"comment":"The convention Fvi,[vj,vk] = [0 0 0] is written with a single zero row; it may be clearer to write the 1×3 zero matrix explicitly, and to state that this convention is used only for ordered pairs where the vertex is not a face of the edge.","section":"§2, Eq. (9) and surrounding text"},{"comment":"The caption contains a stray expression 'and = Fvj,[vi,vj]'; the equality should be written cleanly as Fvi,[vi,vj] = Fvj,[vi,vj].","section":"Figure 2 caption"},{"comment":"In the proof of linear independence of the six global sections, the phrase 'without loss of generality' is used to assume that the first three vertices are affinely independent; this is fine but should be stated explicitly before the matrix calculation, since the displayed matrix uses vertices 1, 2, 3.","section":"§4, Theorem 4.1 proof"}],"recommendation":"major_revision","confidential_remarks":"The central identification of the ANM Hessian as a sheaf Laplacian is elegant and the linear algebra in Theorems 3.1 and 4.1 appears correct. The paper's main weakness is the overclaiming of Algorithm 1 as a 'systematic method' for arbitrary general-position point clouds; the nearest-point insertion rule can fail when a remaining point lies inside the current tetrahedral complex, and the existence step is not proven. I would suggest the authors either significantly revise Algorithm 1, add the missing hypotheses, or reframe it as a heuristic with clearly stated limitations. The Delaunay corollary is promising but needs the rank-increase lemma and the termination argument in Proposition 5.5 to be made rigorous before the result can be considered fully established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper deserves a referee's time, but with one real gap to close. The central identity is sound: Theorem 3.1 checks out, and the interpretation of zero modes as global sections of the anisotropic sheaf is a neat, correct dictionary. The strongest genuinely new result is Corollary 5.7, proving that the 1-skeleton of a Delaunay triangulation gives exactly six zero modes for a point cloud in general position. That is a clean and practically useful theorem for ENM practitioners, even if rigidity theory has known similar conclusions via Maxwell counting and generic rigidity; the paper does not cite that literature, which mildly weakens the novelty claim.\n\nThe soft spot is Algorithm 1, step 6. The algorithm needs to pick a boundary 2-simplex sigma in the growing complex K_{i-1} such that conv(sigma, v_i) intersects K_{i-1} only in sigma. The paper justifies this by Corollary B.6. But that corollary separates a point from a single 3-simplex; it says nothing about a union of tetrahedra. The growing complex is not necessarily convex, so a new vertex's nearest point could lie in a concavity or on an edge, and no such sigma may exist. As written, the existence claim is unsupported, and the six-mode/minimality conclusion for Algorithm 1 (Theorem 4.5) rests on it. This is a real proof gap, not a stylistic complaint. The core Theorems 3.1, 5.6, and 5.7 do not depend on this step, so the paper's central contribution survives.\n\nAlso minor: the rank-increase lemma in Theorem 5.6 is asserted rather than fully proven; it looks fixable, but a referee should ask for more detail.\n\nBottom line: send this to a serious referee. It is a credible sheaf-theoretic restatement of ANM, and the Delaunay theorem is worth publishing. The authors just need to either prove step 6 properly (perhaps via a careful closest-point argument on a star-shaped domain) or downgrade Algorithm 1 to a heuristic with numerical verification. As is, I would not accept without revision, but I would not desk-reject.","headline":"Solid sheaf/Hessian identification and a useful Delaunay six-mode theorem, but Algorithm 1's existence step is unproven and needs either a real proof or a downgrade.","tokens_in":42988,"tokens_out":2792,"would_cite":true,"duration_ms":25113,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["55N30","55U10","05C50","92C40"],"pacs":[],"model":"deepseek-v4-flash","headline":"By modeling a protein's interaction network as a cellular sheaf, this paper proves that the ANM Hessian is a sheaf Laplacian and that its six rigid-body zero modes are exactly the sheaf's global sections.","keywords":["normal mode analysis","anisotropic network model","cellular sheaf","sheaf Laplacian","global sections","Delaunay triangulation","zero modes","elastic network model"],"falsifier":"Take five points in $\\mathbb{R}^3$ in general position: an initial tetrahedron and a fifth point that is strictly closest to the tetrahedron but positioned so that every boundary triangle would make the new tetrahedron intersect the existing complex in more than that triangle; if Algorithm 1 cannot complete step 6 for this input, the claimed minimal-graph guarantee fails for that configuration. The nullity of the resulting Hessian can then be computed directly and compared with six.","tokens_in":41980,"feed_emoji":"🧬","tokens_out":6788,"duration_ms":62741,"temperature":0.7,"pith_summary":"This paper tries to give normal mode analysis a rigorous geometric foundation by recasting it in sheaf theory. It introduces a cellular sheaf, called the anisotropic sheaf, on the graph of atoms and interactions, and proves that with edge weights $w_{ij}=\\gamma^{1/2}/s_{ij}$ the sheaf Laplacian is exactly the ANM Hessian. Under this identification, the six zero-eigenvalue rigid motions become the sheaf's global sections: local atomic displacements that agree consistently along every edge. The paper further proves that any point cloud in general position admits admissible triangulations, including the 3D Delaunay triangulation, whose 1-skeletons induce Hessians with exactly six zero modes, and gives an algorithm for an edge-minimal such graph. If true, this replaces the usual cutoff-distance heuristic in anisotropic network models with a topology-backed construction guarantee.","feed_headline":"Delaunay graph guarantees six zero modes in protein networks","feed_subtitle":"Read as a cellular sheaf, the ANM Hessian's six trivial modes are exactly its global sections.","key_machinery":"The carrying object is the anisotropic sheaf: vertices get stalk $\\mathbb{R}^3$, edges get stalk $\\mathbb{R}$, and each edge's restriction map is the rank-one row vector pointing along the edge, scaled by $w_{ij}$. The 0-th sheaf coboundary matrix $C$ has one row per edge with signed copies of these maps, and the sheaf Laplacian $\\Delta_0 = C^T C$ is what the proof works with. Its kernel is $H^0$, the global section space, so the six rigid motions appear as harmonic signals on the network. The dimension count $3|V|-\\operatorname{rank}(C)$ then turns the six-mode question into a rank condition on $C$, and the admissibility and Delaunay arguments are precisely constructions that force $\\operatorname{rank}(C) = 3|V|-6$.","core_discovery":"The central discovery is a dictionary between the anisotropic network model and cellular sheaf cohomology. For a molecular graph with coordinates, each vertex carries a stalk $\\mathbb{R}^3$, each edge carries a stalk $\\mathbb{R}$, and the restriction map along an edge is the row vector $w_{ij}(x_j-x_i,\\,y_j-y_i,\\,z_j-z_i)$. The 0-th coboundary matrix $C$ is built from these signed edge maps, and its sheaf Laplacian $C^T C$ equals the ANM Hessian precisely when $w_{ij}=\\gamma^{1/2}/s_{ij}$ (Theorem 3.1). Consequently, by the Hodge identification of $H^0$ with the kernel of the sheaf Laplacian, the zero modes of the Hessian are exactly the global sections of the sheaf, with the six rigid translations and rotations forming a canonical basis (Theorems 2.3 and 4.1). The paper then characterizes when the global section space has dimension exactly six: for a point cloud in general position, the 1-skeleton of any admissible homogeneous 3-complex, in particular the 3D Delaunay triangulation, induces an ANM Hessian with exactly six zero eigenvalues (Theorem 5.6, Corollary 5.7). Algorithm 1 constructs a graph with $3|V|-6$ edges that is minimal with this property, and deleting any edge raises the nullity above six.","pith_inferences":["Beyond the paper, this dictionary suggests that the low-frequency nonzero modes of an ANM Hessian could be read as the first obstructions to gluing in sheaf cohomology, connecting vibrational softness to cohomological invariants; the paper does not develop this.","The minimal-graph construction may extend to non-protein point clouds, such as ligand or material configurations, where the same six-mode guarantee is desired, but Algorithm 1's greedy attachment step would need empirical testing on such distributions.","The Delaunay admissibility proof (Corollary 5.7) and the greedy Algorithm 1 construction are logically independent; a counterexample to the greedy attachment step would not automatically falsify the Delaunay-based six-mode claim.","Because the sheaf Laplacian is $C^T C$, adding edges imposes interlacing-type constraints on the nonzero spectrum; a testable extension is that Delaunay-based ANM low-frequency modes converge to complete-graph ANM modes as the point cloud is refined."],"forward_implications":["Any ANM built from the 1-skeleton of a Delaunay triangulation of a protein's $C_\\alpha$ atoms in general position has exactly six trivial modes, with no cutoff distance to tune.","The nullity of an ANM Hessian is a sheaf-cohomology invariant, so adding or removing edges changes the number of trivial modes in a controlled way: a subgraph cannot have fewer trivial modes than the graph it sits in.","The greedy Algorithm 1 produces a graph with $3|V|-6$ edges that is minimal for the six-mode property; removing any edge yields at least seven zero eigenvalues.","Because the sheaf-to-Hessian equality holds for any nonzero edge weights, the six-mode guarantee is independent of the $\\gamma^{1/2}/s_{ij}$ scaling used in practice.","The six zero-mode eigenvectors can be written down explicitly as translation and rotation vector fields on the atomic coordinates, without diagonalizing the Hessian (Theorem 4.1)."],"supporting_citations":[{"why":"Defines the ANM potential, Hessian matrices, and spring-constant setting that Theorem 3.1 equates with the sheaf Laplacian.","marker":"[3]"},{"why":"Supplies the cellular sheaf, sheaf cohomology, and signed-incidence coboundary formalism used throughout Section 2.","marker":"[22]"},{"why":"Provides the spectral/Hodge theorem identifying $H^0$ with the kernel of the sheaf Laplacian, which is the bridge to zero modes.","marker":"[36]"},{"why":"Introduces the force cosheaf whose dual structure motivates the anisotropic sheaf and its edge weights.","marker":"[17]"},{"why":"Gives the truss/cosheaf Maxwell-rule interpretation that frames the six-dimensional rigid-motion count.","marker":"[16]"},{"why":"Defines the Delaunay triangulation whose admissibility underlies Corollary 5.7.","marker":"[24]"},{"why":"Supplies the alpha-shape/Delaunay-based construction of elastic network graphs that Algorithm 1 streamlines.","marker":"[65]"},{"why":"Provides the protein-application context for using Delaunay triangulation to build interaction graphs.","marker":"[70]"}],"fun_headline_variants":["Sheaf theory matches ANM Hessian: zero modes are global sections","Delaunay graph ensures six zero modes via sheaf Laplacian","A sheaf view of normal modes: exact six trivial motions guaranteed","ANM zero modes equal sheaf global sections on Delaunay graphs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the greedy attachment step in Algorithm 1 can always find a boundary triangle such that the new tetrahedron touches the existing complex only along that triangle; the paper argues this from a separation lemma that is proved only for a single simplex, so the guarantee for a growing multi-tetrahedron complex is assumed rather than fully established.","fun_headline_variants_meta":{"raw":{"variants":["Sheaf theory matches ANM Hessian: zero modes are global sections","Delaunay graph ensures six zero modes via sheaf Laplacian","A sheaf view of normal modes: exact six trivial motions guaranteed","ANM zero modes equal sheaf global sections on Delaunay graphs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000592,"raw_usage":{"total_tokens":2850,"prompt_tokens":1098,"completion_tokens":1752,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":714,"completion_tokens_details":{"reasoning_tokens":1674}},"tokens_in":714,"tokens_out":1752,"duration_ms":12007,"temperature":1.0,"reasoning_tokens":1674,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:37:44.365613+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take five points in $\\mathbb{R}^3$ in general position: an initial tetrahedron and a fifth point that is strictly closest to the tetrahedron but positioned so that every boundary triangle would make the new tetrahedron intersect the existing complex in more than that triangle; if Algorithm 1 cannot complete step 6 for this input, the claimed minimal-graph guarantee fails for that configuration. The nullity of the resulting Hessian can then be computed directly and compared with six.","supporting_citations":[{"cited_title":"Delaunay, S","cited_arxiv_id":null,"evidence_quote":"Defines the Delaunay triangulation whose admissibility underlies Corollary 5.7."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the ANM potential, Hessian matrices, and spring-constant setting that Theorem 3.1 equates with the sheaf Laplacian."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the alpha-shape/Delaunay-based construction of elastic network graphs that Algorithm 1 streamlines."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the protein-application context for using Delaunay triangulation to build interaction graphs."}],"review_version":1}