{"id":"ec768a58-20cc-4603-b5f7-e00cc2dd7871","arxiv_id":"2607.22478","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":1,"one_line_summary":"Fixed-neuron ReLU^k ridge functions form an exact finite-dimensional de Rham complex in any dimension, provided the lowest-order family is linearly independent.","lead":"This paper builds a chain of function spaces made from simple ReLU^k neuron building blocks that exactly mimic the grad-curl-div structure of vector calculus. It proves the chain is mathematically exact, a first step toward stable neural-network solvers for PDEs.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 6.2 is correct as a conditional statement, but its key hypothesis—linear independence of {σ_{k-d}(s_i)}—is not verified for the quasi-uniform neuron sets used in the numerical experiments; many such sets violate Assumption 4.1, so the reported numerics may lie outside the theorem's scope.","rationale":"I examined the proof steps of the central theorem. The derivative identity (3.3), the linear-independence lifting in Lemma 4.3, and the Koszul-homotopy preimages in Theorem 6.2 are all correct. The exactness argument is a standard and valid linear-algebraic consequence of the stated assumption. Therefore I do not see a flaw in the theorem itself. The reader's identified weakest assumption—linear independence of the lowest-order ridge family—is indeed the load-bearing condition, and I agree that this is where the argument is most exposed. However, I would sharpen the failure mode: the critical issue is not non-uniqueness of representations per se, but the inability to infer coefficient-wise vanishing from a zero sum when the scalar family is dependent. More importantly, the paper does not verify the assumption for the parameter sets used in the numerical experiments, and the provided sufficient condition (Assumption 4.1) is not guaranteed by quasi-uniform sampling on the full sphere. This leaves a real gap between the algebraic theorem and the numerical evidence, but it does not change the verdict: the theorem is correctly proved under its explicit hypothesis, and the numerical concerns were already reflected in the CONDITIONAL verdict.","tokens_in":12950,"tokens_out":23337,"duration_ms":230196,"concrete_test":"For the actual neuron sets used in Figures 7.1–7.4, compute the numerical rank of the lowest-order ridge family {σ_{k-d}(s_i)} on Ω (e.g., via a fine-grid evaluation matrix or the Gram matrix M_{ij}=∫_Ω σ_{k-d}(s_i)σ_{k-d}(s_j) dx). If rank < n, verify whether the complex is still exact by checking the dimensions of ker d and im d at each degree. This single computational check would determine whether the numerical experiments are within the scope of Theorem 6.2 or need an extended argument.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The exactness proof in Theorem 6.2 is internally sound: Lemma 4.3 correctly lifts linear independence from order k−d to all higher orders, and the Koszul homotopy (Lemma 6.1) gives the required preimages. The load-bearing step is the inference, from dν=0, that each neuron-wise term ds_i∧α_i vanishes individually; this requires the scalar family {σ_{k-p-1}(s_i)} to be linearly independent. Lemma 4.4 gives a sufficient condition through Assumption 4.1, but that condition is not automatic for the quasi-uniform parameter sets used in the numerical sections. For a bounded domain such as Ω=(0,1)^d, a positive fraction of neurons sampled quasi-uniformly from S^d will have hyperplanes H_i disjoint from Ω; for those neurons σ_{k-d}(s_i) is identically zero or a polynomial on Ω, so the lowest-order family can be rank-deficient. Section 7 only says 'quasi-uniform neuron parameters excluding antipodal pairs' and does not state that H_i∩Ω is enforced or that the rank of {σ_{k-d}(s_i)} is checked. Without this hypothesis, the coefficient-wise decoupling in Theorem 6.2 is unjustified: the proof cannot conclude ds_i∧α_i=0 from the vanishing of the sum. This is not a flaw in the conditional theorem, but a gap between the theorem's hypothesis and its claimed numerical applicability.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper constructs finite-dimensional de Rham subcomplexes inside fixed-neuron shallow ReLU^k neural network spaces. For affine neurons s_i(x)=ω_i·x+b_i and k≥d, the authors define spaces C^p of p-forms with coefficients σ_{k-p}(s_i), noting that exterior differentiation lowers the ReLU degree and wedges with d s_i. The main result, Theorem 6.2, states that under linear independence of the lowest-order family {σ_{k-d}(s_i)}, the complex 0→C^0→⋯→C^d→0 is exact. The proof combines Lemma 4.3 (independence propagates upward via directional derivatives), Lemma 6.1 (Koszul homotopy for a nonzero covector), and a coefficient-wise decoupling argument. The paper also gives a sufficient geometric condition (Assumption 4.1 and Lemma 4.4) and presents numerical experiments for mixed Poisson, Hodge Laplacian, and eigenvalue problems. The conditional exactness theorem appears algebraically correct; the main gaps concern the scope of the numerical experiments and the weak/distributional interpretation at the k=d endpoint.","tokens_in":13381,"tokens_out":11755,"duration_ms":130974,"significance":"If the main theorem is accepted, the paper provides a clean and elementary construction of an exact de Rham subcomplex within a linearized shallow ReLU^k approximation space. This is a genuinely useful bridge between FEEC and neural-network approximation: exactness gives the discrete kernel–range identities that rule out algebraic spurious cohomology, and the proof is transparent, using only the ReLU derivative identity and the standard Koszul homotopy. No fitted constants or numerical tuning enter the theoretical argument, which is a strength. The paper is also careful to state that some ingredients of a full FEEC theory—uniform inf-sup bounds, bounded commuting projections, boundary compatibility—are not provided. However, the numerical section is currently not a rigorous validation of the theorem because it does not verify the key linear-independence hypothesis and uses k values outside the standing k≥d assumption. The central mathematical contribution is sound, but the applied claims require revision.","major_comments":[{"comment":"The numerical experiments are presented as supporting the exact neural complex, but they never verify the load-bearing hypothesis of Theorem 6.2, namely linear independence of {σ_{k-d}(s_i)}. Section 7 states only \"quasi-uniform neuron parameters excluding antipodal pairs.\" This does not enforce Assumption 4.1's requirement that each hyperplane H_i intersect Ω. For Ω=(0,1)^d, many quasi-uniform neurons on S^d have H_i∩Ω=∅; for those neurons σ_{k-d}(s_i) is identically zero or a polynomial on Ω, which can make the lowest-order family rank-deficient. Without an explicit rank check or a neuron-generation procedure that filters neurons with H_i∩Ω≠∅, the reported convergence rates and the absence of spurious modes cannot be attributed to the exactness theorem. This is not a flaw in the conditional theorem, but it is a real gap between the theorem's hypothesis and the numerical evidence claime","section":"§7 vs. Assumption 4.1 / Theorem 6.2"},{"comment":"The paper assumes k≥d throughout §3 and in Theorem 6.2, but the numerical sections use smaller values: §7.1 uses k=1 in d=2 and §7.2 uses k=2 in d=3. These cases fall outside the theorem's scope. Moreover, in §7.1 the discrete pair (L^1_n)^2 × L^0_n is not the H(div)–L^2 part of the complex for k=1; it is the C^1–C^2 pair for the complex with parameter k=2. The displayed rates in §7.1 use the exponent k=1 and are therefore not the correct approximation exponents for the spaces actually used. A clarification of which complex parameter and which truncated subcomplex is used in each experiment is needed, along with corrected rate statements or a restriction to k≥d.","section":"§3 and §7: k≥d standing assumption"},{"comment":"When k=d, the top space C^d contains σ_0(s_i), which is discontinuous (σ_0(t)=max{t,0}^0, an indicator-type function). Lemma 4.4 explicitly uses distribution derivatives and δ(s_i), and the exactness proof at C^d relies on representing top forms as ds_i∧ι_{v_i}α_i. However, the paper never states the ambient function-space setting in which the complex is exact: pointwise, L^2, or distributional. The directional-derivative argument in Lemma 4.3 and the final surjectivity step (6.26) need this clarification when k=d. This is a rigor gap in an endpoint case of the main theorem, not a fatal error, but it should be fixed.","section":"§2–§6, Lemma 4.3/4.4 and Theorem 6.2 for k=d"}],"minor_comments":[{"comment":"The definition σ_k(t)=max{t,0}^k for k∈N_0 does not specify the value of σ_0 at t=0. Please state a convention, e.g., σ_0(t)=1 for t>0 and σ_0(t)=0 otherwise (or define the relevant equivalence class).","section":"Eq. (2.2)"},{"comment":"The term \"quasi-uniform\" is used without definition. Since it is borrowed from [20], please either define it or give a precise reference to the definition used.","section":"§7.1 and §7.2"},{"comment":"The inf-sup constant for k=2 at n=16 is 4.346e-01, which then jumps to above 9.6e-01 for larger n. A brief comment on small-n behavior would be helpful, although it does not affect the asymptotic claim.","section":"§7, Table 7.1"},{"comment":"The figure captions do not fully describe axes, markers, and reference lines. Adding axis labels and legends would improve reproducibility and readability.","section":"Figures 7.1–7.4"},{"comment":"The discussion of possible future extensions (BGG, boundary-compatible spaces, nontrivial topology) is reasonable, but the sentence suggesting that the complex \"rules out algebraic spurious cohomology\" should be tied explicitly to the exactness theorem and its hypothesis, since the numerical evidence does not yet verify that hypothesis.","section":"§8"}],"recommendation":"major_revision","confidential_remarks":"The algebraic core of the paper—Theorem 6.2 and its Koszul-based proof—appears correct and is a nice contribution. The main risk is overclaiming numerical validation: the experiments do not check the linear-independence hypothesis and use k below the standing k≥d assumption, so the abstract's claim of numerical evidence for the complex is premature. I recommend major revision with the expectation that the authors either bring the experiments within the assumptions (e.g., rank checks or filtered neuron sets), explicitly label the experiments as heuristic, and correct the off-by-one notation in §7. The use of [20] (co-authored by a paper author) is limited to motivation and benchmark rates; it does not enter the exactness proof, so I see no circularity concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the exactness theorem is right, and the gap the stress-test flags is real. I read the paper carefully and the algebra checks out. Under linear independence of {σ_{k-d}(s_i)}, the complex (3.5) is exact. The proof is the right kind of proof: Lemma 4.3 transfers independence upward, Lemma 6.1 applies the standard Koszul homotopy neuron-by-neuron. No fitted constants, no circularity. The construction is new – [19] emulates FEEC spaces with deep nets, [20] gives approximation rates, [21] handles divergence-free only; none of them have the intrinsic fixed-neuron de Rham complex. Credit where due: this is a clean algebraic result.\n\nThe soft spot is exactly where the stress-test points. Theorem 6.2 is conditional on the lowest-order family being linearly independent. Lemma 4.4 gives a sufficient geometric condition: hyperplanes must intersect Ω and no antipodal pairs. But Section 7 says only \"quasi-uniform neuron parameters excluding antipodal pairs\" – it does not say the intersection condition is enforced, nor that independence is checked. For a bounded domain, many neurons drawn quasi-uniformly from the sphere will have hyperplanes that miss Ω entirely; for those, σ_{k-d}(s_i) is either identically zero or a polynomial on Ω, so the family can be rank-deficient. If it is, the coefficient-wise decoupling in Theorem 6.2 has no basis. The theorem itself is fine as a conditional statement; what is not fine is presenting the numerics as if they fall under it. That should be stated explicitly, or the experiments should enforce the condition.\n\nTwo smaller things. First, the k=d case: the lowest power is σ_0, which is a step function, and the derivative identity needs distribution interpretation; the paper doesn't flag this. Second, the numerical section has no code, no data, no error bars, so the convergence plots are not independently reproducible. Those are fixable.\n\nWho is this for? People building structure-preserving neural discretizations for mixed PDEs. It's a useful algebraic brick, not a full FEEC theory – the paper says as much in the conclusion. I'd send it to peer review: the theorem deserves referee time, and the gap between hypothesis and numerics needs to be addressed in revision, but it's not a desk reject.","headline":"A correct and genuinely new exactness theorem for fixed-neuron ReLU^k de Rham complexes, with a real but narrow gap between the theorem's hypothesis and the numerical experiments.","tokens_in":13808,"tokens_out":2220,"would_cite":true,"duration_ms":24991,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["41A30","41A46","55U15","58A10","65N25","65N30","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Fixed-neuron shallow ReLU^k networks contain a finite-dimensional de Rham subcomplex that is exact — the exterior derivative decouples neuron by neuron into Koszul complexes.","keywords":["ReLU neural networks","de Rham complex","Koszul complex","exact sequence","structure-preserving discretization","differential forms","mixed finite elements","linearized shallow networks"],"falsifier":"Take a quasi-uniform neuron set that avoids antipodal pairs and compute the numerical rank of the Gram matrix of {σ_{k-d}(s_i)} at dense sample points. A rank below n would violate the paper's assumption and break the exactness proof; a full rank for many random sets would support the assumption's genericity. The antipodal-pair example (4.3) shows the failure mode is real.","tokens_in":12839,"feed_emoji":"🧠","tokens_out":7471,"duration_ms":80571,"temperature":0.7,"pith_summary":"The paper shows that fixed-neuron shallow ReLU^k networks can be organized into a finite-dimensional de Rham complex — a chain of spaces of differential forms connected by the exterior derivative that satisfies the correct kernel–image identities. The construction assigns to each form degree p the span of the ridge functions σ_{k-p}(s_i) with constant p-form coefficients; because differentiating a ReLU power lowers its order by one and acts as multiplication by the fixed one-form ds_i, the derivative maps each space into the next. Under a linear independence condition on the lowest-order ridge family, the global complex is exact: every closed form is a derivative, with no spurious cohomology. The value is a mesh-free analogue of structure-preserving finite element spaces, with numerical experiments indicating stable mixed discretizations and spectra free of spurious modes.","feed_headline":"ReLU^k neurons form exact de Rham complexes","feed_subtitle":"Shallow fixed-neuron networks carry correct grad-curl-div structure, enabling stable, spurious-free PDE discretizations.","key_machinery":"The machinery is the neuron-wise Koszul complex: for a fixed nonzero covector λ, the sequence Λ^0 → Λ^1 → ... → Λ^d given by wedge product with λ is exact, with homotopy provided by the contraction identity ι_v(λ∧α)+λ∧ι_v α = α. The proof isolates each neuron by the linear independence of the ridge functions, which also transfers from the lowest degree to all higher degrees via directional derivatives (Lemma 4.3). A geometric sufficient condition for the independence — distinct, non-antipodal neuron parameters with hyperplanes meeting the domain — is given in Lemma 4.4.","core_discovery":"The central discovery is that the exterior derivative acts independently on each neuron: d(σ_{k-p}(s_i) α_i) = (k-p) σ_{k-p-1}(s_i) ds_i ∧ α_i. Consequently the global complex 0→C^0→...→C^d→0 splits coefficient-wise into the classical Koszul complex of the one-form ds_i for each neuron. Provided the lowest-order ridge family {σ_{k-d}(s_i)} is linearly independent on the domain, the paper proves (Theorem 6.2) that the complex is exact in every degree, for any dimension d and any integer k≥d. Exactness gives the discrete kernel–range identities: every closed discrete form is the derivative of one at the previous level, ruling out spurious closed fields and ensuring the discrete spaces represen","pith_inferences":["If the linear independence assumption holds generically for quasi-uniform neuron sets — as the paper's geometric condition suggests but does not fully characterize — then exactness is a generic property rather than a special construction; a probabilistic or measure-theoretic argument would settle this.","The mechanism 'differentiation lowers activation order and acts as a fixed algebraic operation' is not unique to ReLU powers: cosine or Fourier ridge functions, where derivatives cycle back, could yield exact complexes with nontrivial cohomology representing the domain's topology, as the paper hints in its closing remarks.","Exactness gives an immediate algebraic test: the alternating sum of the dimensions of C^p must vanish, and since each C^p is spanned by n independent ridge forms, the dimension count is automatically consistent; any failure of independence would manifest as an anomalous dimension drop, which is computable for a given neuron set.","A bounded cochain projection onto the complex is still missing; if it exists, it would upgrade the algebraic exactness to full stability estimates (inf-sup constants uniformly bounded in n), turning the numerical evidence into a theorem."],"forward_implications":["If the exactness theorem holds, fixed-neuron shallow ReLU^k spaces contain a bona fide discrete de Rham complex on contractible domains, giving the correct kernel–range identities and eliminating algebraic spurious cohomology.","Mixed Galerkin discretizations built from the complex inherit the exactness structure, which is the algebraic backbone of stable mixed methods; the reported inf-sup constants and eigenvalue spectra support this.","Combined with the optimal approximation rates already known for fixed-neuron ReLU^k spaces, the complex provides a path to convergent, mesh-free structure-preserving discretizations of Hodge-type and Maxwell-type boundary value problems.","Because the proof is dimension-independent, the same construction yields grad–curl–div sequences in 3D and higher-degree form complexes in any dimension, and it opens the door to derived complexes for elasticity-like systems."],"fun_headline_variants":["ReLU^k neurons form exact de Rham complexes","No spurious modes with ReLU^k de Rham complexes","Exact grad-curl-div from ReLU^k ridge functions","ReLU^k powers yield spurious-free discretizations"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"Everything rests on the assumption that the n ridge functions σ_{k-d}(s_i) are linearly independent on the domain; the paper supplies a sufficient geometric condition (distinct, non-antipodal neurons whose zero hyperplanes cross the domain), but does not prove that every quasi-uniform parameter set satisfies it, and if it fails the neuron-wise decoupling that drives the exactness proof collapses.","fun_headline_variants_meta":{"raw":{"variants":["ReLU^k neurons form exact de Rham complexes","No spurious modes with ReLU^k de Rham complexes","Exact grad-curl-div from ReLU^k ridge functions","ReLU^k powers yield spurious-free discretizations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000149,"raw_usage":{"total_tokens":1037,"prompt_tokens":756,"completion_tokens":281,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":214}},"tokens_in":500,"tokens_out":281,"duration_ms":3503,"temperature":1.0,"reasoning_tokens":214,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T04:35:25.580070+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a quasi-uniform neuron set that avoids antipodal pairs and compute the numerical rank of the Gram matrix of {σ_{k-d}(s_i)} at dense sample points. A rank below n would violate the paper's assumption and break the exactness proof; a full rank for many random sets would support the assumption's genericity. The antipodal-pair example (4.3) shows the failure mode is real.","supporting_citations":[],"review_version":1}