{"id":"db49ab19-49e6-414a-aba3-f0be36954880","arxiv_id":"2509.05894","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A toric-geometry framework for ReLU networks yields a necessary and sufficient intersection-number criterion for exact realizability by unbiased shallow ReLU networks.","lead":"This paper builds a dictionary between ReLU neural networks and toric geometry, defining a ReLU fan and a ReLU Cartier divisor for every bias-free network with rational weights. The payoff is a new criterion, expressed through toric intersection numbers, for which functions can be built exactly by a shallow ReLU network.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed iff criterion in Theorems 3-4 is not necessary as stated: f(x,y)=max{0,x} is realizable by a shallow unbiased ReLU network but its nonlinear locus does not extend to a fan, so the criterion excludes it.","rationale":"The reader's weakest-assumption analysis identified exactly this concern: Remark 5.2 assumes the canonical complex is a fan, but this fails for realizable degenerate functions such as max{0,x} in the plane. My stress-test confirms that the failure is not cosmetic: it makes the central 'if and only if' statement false as written, because a function can be in ReLU_Q_0(2,1) while failing the hypothesis of Theorem 4. The paper's other weaknesses, such as the underjustified linearity step in the proof of Theorem 4, are secondary and likely repairable by a standard argument showing that matching bends across every wall force the difference of two support functions to be linear. The fan gap is the most load-bearing because it concerns the scope of the main theorem itself. Since the gap is concrete and repairable by adding a refinement lemma and restating the criterion existentially, the appropriate verdict remains CONDITIONAL: the framework is plausible and the non-degenerate case is coherent, but the stated classification must be revised before it can be accepted as a complete characterization of ReLU_Q_0(n0,1). The reader's CONDITIONAL verdict already reflects this, so no adjustment is needed.","tokens_in":21319,"tokens_out":18837,"duration_ms":185689,"concrete_test":"Apply the stated criterion to f(x,y)=max{0,x}. First verify f is realizable: L1=[1,0], L2=1 gives f(x,y)=max{0,x}. Then compute the nonlinear locus: it is the hyperplane {x=0}. Extending this locus to a full hyperplane yields cells {x>=0} and {x<=0}, neither of which is a strongly convex cone, so no fan exists and D_f is undefined. Hence Theorem 4's hypothesis fails even though f is realizable, disproving the claimed equivalence. As a control, refine the arrangement by adding the hyperplane y=0 with zero output weight; the cells become pointed quadrants, the fan condition is satisfied, and the intersection-number equality holds with value -1 on walls of {x=0} and 0 on walls of {y=0}. This confirms that a refinement/existence clause would repair the theorem, but the paper currently lacks it.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the equivalence, advertised in Example 7.6, between realizability in ReLU_Q_0(n0,1) and the intersection-number condition of Theorems 3 and 4. The load-bearing gap is the standing assumption in Remark 5.2 that the canonical polyhedral complex is a fan of strongly convex cones. The stated justification 'n1 > n0' does not imply pointed cells: in R^2, f(x,y)=max{0,x} is realized by the unbiased shallow network with L1=[1,0] and L2=1, so f is in ReLU_Q_0(2,1). Its canonical complex has cells {x>=0} and {x<=0}, which contain lines and are not strongly convex cones. Consequently the ReLU fan, the ReLU toric variety, the ReLU Cartier divisor, and the intersection numbers D_f.V(tau) used in Theorem 3 are not defined for this representation. The function also fails the hypothesis of Theorem 4: extending the nonlinear locus {x=0} to a full hyperplane gives a halfspace arrangement, not a fan. Thus the claimed necessary condition is not necessary for realizable functions, and the 'equivalent condition' in Example 7.6 is false as written. The gap is repairable: one can refine the halfspaces by adding redundant hyperplanes with zero output weight, e.g. y=0, obtaining a fan, and reformulate the criterion as an existence statement about a fan refinement. But this refinement lemma is absent from the paper, and the stated theorems do not cover the example.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a dictionary between unbiased rational-weight ReLU neural networks and toric geometry. For a network with parameters θ, it defines the \"ReLU fan\" Σ_{f_θ} as the canonical polyhedral complex of the network, the associated toric variety X_{Σ_{f_θ}}, and a Q-Cartier \"ReLU divisor\" D_f whose support function is the network output. The paper then connects this toric data to tropical geometry, proving that for tropical-polynomial outputs the divisor polytope equals the negative Newton polytope and that volumes agree. The main application is a claimed complete classification of functions realized by unbiased shallow ReLU networks with rational weights: Theorem 3 gives a necessary condition involving equality of intersection numbers D_f·V(τ) for walls from the same hyperplane, Theorem 4 gives a converse, and Example 7.6 presents the resulting \"equivalent condition.\"","tokens_in":21606,"tokens_out":5097,"duration_ms":49429,"significance":"If the classification were established, the paper would contribute a genuinely new algebraic-geometric invariant to the study of exact ReLU realizability, and the toric/tropical bridge (Theorems 6.9 and 6.14) is an elegant and potentially useful observation. The paper is self-contained, with many worked examples and a careful exposition of toric background, and it gives an explicit reduction algorithm for shallow ReLU representations. However, the central classification claim is not supported as stated: the standing fan assumption excludes realizable functions, and the converse proof contains a load-bearing unsupported assertion. The framework is promising and likely repairable, but the advertised necessary-and-sufficient characterization is currently false in its stated form.","major_comments":[{"comment":"In the proof of Theorem 4 (§7), the assertion that if the constructed network output f' is not exactly f then f − f' is linear is made without proof. The construction chooses the rows of L1 from the hyperplanes of Σ and the entries of L2 from the common intersection numbers t_i, but it is not shown that the piecewise linear function determined by these data agrees with f up to a linear function. Equality of wall intersection numbers alone does not obviously determine the Cartier data (and hence the support function) up to a global linear term; an argument is needed, for example by comparing the Cartier data cone-by-cone or by showing the difference vanishes on each maximal cone. This is load-bearing for the sufficiency direction, and without it Theorem 4 is not established.","section":"§5, Remark 5.2; §7, Theorems 3 and 4 and Example 7.6"}],"minor_comments":[{"comment":"Example 7.6 is too terse to be useful: the function is not actually defined in the text, and the displayed expressions \"y  0  x+2y  3y  −4y  2x−2y\" do not specify the piecewise linear function, the architecture, or the intended bent hyperplane arrangement. The example should be rewritten with an explicit formula or diagram.","section":"§7, Example 7.6"},{"comment":"There are several typos throughout, e.g., \"Negavie infinity\" in Remark 6.2, \"the funcion\" in Remark 4.48, \"finitely-piecewise\" in Theorem 4, and \"polytop\" in Lemma 6.13. These should be corrected.","section":"§6, Remark 6.2"},{"comment":"The terminology is potentially confusing: Definition 5.1 calls Σ_{f_θ} a \"ReLU fan\" even though it is defined as the canonical polyhedral complex, which the authors themselves note is only a generalized fan in general. It would be clearer to use a distinct name (e.g., \"ReLU complex\") until the fan condition is actually established or imposed.","section":"§5, Definition 5.1 and Remark 5.2"},{"comment":"The statement of Theorem 3 refers to \"the n1 hyperplanes H_1^{(1)},...,H_{n1}^{(1)}\", but if two rows of L1 are opposite normals they define the same hyperplane; the indexing should be clarified to avoid double-counting, especially since Lemma 7.4 explicitly handles this case.","section":"§7, Theorem 3"}],"recommendation":"major_revision","confidential_remarks":"The stress-test counterexample f(x,y)=max{0,x} is decisive: it is a realizable unbiased shallow ReLU function whose canonical complex is not a fan, so the stated Theorem 3/Theorem 4 equivalence cannot be correct as written. I nevertheless recommend major revision rather than rejection because the toric dictionary and the volume identities are of independent interest, and the classification can plausibly be repaired by adding a fan-refinement construction and a missing argument in the converse. The author should also address the unsupported statement in the proof of Theorem 4 before the manuscript can be considered for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The toric dictionary for ReLU networks is the real contribution here: the ReLU fan, the ReLU toric variety, the ReLU Cartier divisor, and the volume identity P_{-D} = -Newt(f) for tropical polynomials. That bridge between tropical and toric geometry is new and worth taking seriously. The paper also deserves credit for an honest attempt at a complete shallow realizability criterion via intersection numbers, and the intuition that \"amount of bending\" along walls is the right invariant is sound.\n\nThe soft spots are in the classification theorems, and they are load-bearing. Remark 5.2 assumes the canonical polyhedral complex is a fan of strongly convex cones, justified only by n1 > n0. That is false in simple realizable cases. Take f(x,y) = max{0,x}, realized by the unbiased shallow network with rows (1,0) and output weight 1. Its canonical complex has cells {x >= 0} and {x <= 0}, which contain lines and are not strongly convex cones. So the ReLU fan, the ReLU Cartier divisor, and the intersection numbers used in Theorem 3 are not defined for this representation. The criterion in Theorems 3 and 4 simply does not cover it, so the \"equivalent condition\" advertised in Example 7.6 is not necessary as stated. This is repairable: refine the halfspaces by adding redundant hyperplanes with zero output weight, as the stress-test note says, and reformulate as an existence statement about fan refinements. But the refinement lemma is absent, and the theorems as written are too strong.\n\nTheorem 4's proof also skips a step. After constructing the network l_theta from the intersection numbers, it asserts f - f' is linear. That is exactly what needs proof: equal wall intersection numbers ensure the local bends match, but you still have to show the differences glue to a single linear function across the whole fan. The argument doesn't demonstrate that. It might be true with a careful connectedness argument, but as written it is a gap.\n\nThe volume identity is the cleanest part of the paper and appears correct under the stated tropical-polynomial assumption. The reduced representation is introduced ad hoc but is not circular; it is a normalization for the proof.\n\nWho should read this: algebraic geometers curious about neural networks, and expressivity researchers who can tolerate toric vocabulary. The framework is promising and the flaws are repairable, but the classification claims need to be restated and reproved. I would send it to a serious referee, expecting major revision.","headline":"A genuinely new toric dictionary for ReLU networks, but the headline realizability criterion is not necessary as stated and the sufficiency proof has a gap.","tokens_in":22158,"tokens_out":1254,"would_cite":true,"duration_ms":13309,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["14M25","14T05","68T07","52B20"],"pacs":[],"model":"deepseek-v4-flash","headline":"A toric intersection equality classifies shallow ReLU realizability","keywords":["ReLU neural networks","toric geometry","exact function realization","Cartier divisor","tropical geometry","intersection numbers","piecewise linear functions","fan"],"falsifier":"Take $f = \\max\\{0, x\\}$ on $\\mathbb{R}^2$, which is trivially in $\\operatorname{ReLU}^{\\mathbb{Q}}_0(2,1)$; its canonical polyhedral complex has two halfspace cells along the line $x=0$, neither strongly convex. Writing down the ReLU fan, the polytope $P_D$, or the intersection numbers $D_f \\cdot V(\\tau)$ for this $f$ is impossible under the paper's definitions, so the claimed complete classification does not apply to it; any repair of this gap, by subdivision or by allowing generalized fans, can be tested by checking whether the equal-bend criterion still holds after the repair.","tokens_in":21063,"feed_emoji":"📐","tokens_out":8606,"duration_ms":73862,"temperature":0.7,"pith_summary":"This paper tries to turn exact function realization for unbiased ReLU networks with rational weights into a toric geometry question. It attaches to every such network a ReLU fan, a ReLU toric variety, and a ReLU Cartier divisor whose support function is the network output, and it proves that shallow, one-hidden-layer realizability is equivalent to a numerical symmetry: the divisor must have equal intersection numbers with the torus-invariant curves coming from walls of the same hyperplane. The paper also shows that for tropical polynomial outputs, the divisor polytope coincides with the negative Newton polytope, linking tropical and toric perspectives. A sympathetic reader would care because this turns an existence question about network widths and weights into finitely many computable numbers attached to a fixed function.","feed_headline":"Equal intersections on each hyperplane classify shallow ReLU functions","feed_subtitle":"A fan and equal bend numbers for every wall pair are necessary and sufficient for exact realization by one hidden layer.","key_machinery":"The load-bearing object is the toric encoding map, which sends a parameter $\\theta$ of an unbiased rational ReLU network to its ReLU fan $\\Sigma_{f_\\theta}$, defined as the canonical polyhedral complex on which the output and all hidden-layer functions are linear. On this fan one places the ReLU Cartier divisor $D_f$, a $\\mathbb{Q}$-divisor whose support function is the network output up to a linear term, and one computes its intersection numbers $D_f \\cdot V(\\tau)$ with the complete torus-invariant curves $V(\\tau)$ indexed by walls $\\tau$. These numbers are the amount by which the output function bends across a wall, and the classification is precisely that these bend numbers are constant along the walls of each hyperplane; the proof of sufficiency builds the hidden layer from the hyperplane normals and the last layer from the bend numbers.","core_discovery":"On its own terms, the paper's central claim is that a continuous finitely piecewise linear function $f: \\mathbb{R}^{n_0} \\to \\mathbb{R}$ is exactly realizable by an unbiased shallow ReLU network with rational weights if and only if, after extending the codimension-one part of its non-linear locus to full hyperplanes, those hyperplanes form a fan $\\Sigma$ in $\\mathbb{R}^{n_0}$ and the ReLU Cartier divisor $D_f$ supported on $\\Sigma$ satisfies $D_f \\cdot V(\\tau_1) = D_f \\cdot V(\\tau_2)$ for any two walls $\\tau_1, \\tau_2$ lying in the same hyperplane (Theorems 3 and 4). Here $D_f \\cdot V(\\tau)$ measures how much the output function bends along the wall $\\tau$. The paper further claims that when $f$ is a tropical polynomial, the polytope $P_{-D}$ equals $-\\operatorname{Newt}(f)$ and $\\operatorname{Vol}(\\operatorname{Newt}(f))$ equals the volume of the line bundle $\\mathcal{O}_{X_{\\Sigma_{f_\\theta}}}(-D)$, establishing a bridge between tropical and toric geometry of ReLU networks.","pith_inferences":["Editorial extension: if the equal-intersection condition is rephrased in terms of divisor classes modulo linear functions, realizability testing for fixed $n_0$ and $n_1$ would reduce to a linear-algebra membership problem, which could be implemented by evaluating $f$ on a few generic points.","Editorial extension: the strong-convexity assumption could likely be repaired by subdividing non-pointed cells; functions like $\\max\\{0,x\\}$ on $\\mathbb{R}^2$ would then enter the classification, and the same intersection-number criterion should survive after subdivision.","Editorial extension: the divisor polytope $P_D$ proposed as a Newton-polytope analogue may give an expressivity measure for non-convex tropical rational functions, testable by comparing volumes across architectures."],"forward_implications":["Exact realizability by a one-hidden-layer unbiased rational ReLU network can be checked by finitely many intersection-number computations, with no search over widths or weights.","Any function violating the equal-bend condition on some hyperplane is provably not realizable by any shallow unbiased rational ReLU network.","Adding a linear term to a realizable function preserves realizability, so the classification is really a classification of realizable $\\mathbb{Q}$-Cartier divisors on the ReLU fan.","For tropical-polynomial outputs, mixed volume of the Newton polytope equals the volume of the associated line bundle on the ReLU toric variety, giving a toric interpretation of tropical expressivity.","When the fan condition holds, the hidden width needed is at most the number of distinct hyperplanes of the fan, not the number of linear regions."],"supporting_citations":[{"why":"Supplies the standard toric-variety definitions of fans, Cartier divisors, support functions, and intersection numbers used throughout the framework.","marker":"[3]"},{"why":"Supplies the canonical polyhedral complex and bent hyperplane arrangement that define the ReLU fan for a network.","marker":"[5]"},{"why":"Establishes that ReLU network outputs are tropical rational functions, providing the tropical-geometry connection the paper extends to toric geometry.","marker":"[13]"}],"fun_headline_variants":["Toric geometry pins down exact shallow ReLU functions","Equal bends on each hyperplane decide ReLU realizability","ReLU fans and Cartier divisors solve exact realization","Shallow ReLU exactness: equal intersection numbers suffice"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The construction assumes the canonical polyhedral complex is a fan of strongly convex cones, justified only by taking more hidden neurons than input dimensions; when the hidden-layer normals do not span the input space, cells can contain lines and the ReLU toric variety, the ReLU Cartier divisor, and the intersection numbers are not defined.","fun_headline_variants_meta":{"raw":{"variants":["Toric geometry pins down exact shallow ReLU functions","Equal bends on each hyperplane decide ReLU realizability","ReLU fans and Cartier divisors solve exact realization","Shallow ReLU exactness: equal intersection numbers suffice"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00042,"raw_usage":{"total_tokens":2181,"prompt_tokens":988,"completion_tokens":1193,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":604,"completion_tokens_details":{"reasoning_tokens":1128}},"tokens_in":604,"tokens_out":1193,"duration_ms":8607,"temperature":1.0,"reasoning_tokens":1128,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:21:58.933139+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take $f = \\max\\{0, x\\}$ on $\\mathbb{R}^2$, which is trivially in $\\operatorname{ReLU}^{\\mathbb{Q}}_0(2,1)$; its canonical polyhedral complex has two halfspace cells along the line $x=0$, neither strongly convex. Writing down the ReLU fan, the polytope $P_D$, or the intersection numbers $D_f \\cdot V(\\tau)$ for this $f$ is impossible under the paper's definitions, so the claimed complete classification does not apply to it; any repair of this gap, by subdivision or by allowing generalized fans, can be tested by checking whether the equal-bend criterion still holds after the repair.","supporting_citations":[{"cited_title":"American Mathematical Soc., 2011","cited_arxiv_id":null,"evidence_quote":"Supplies the standard toric-variety definitions of fans, Cartier divisors, support functions, and intersection numbers used throughout the framework."},{"cited_title":"On transversality of bent hyperplane ar- rangements and the topological expressiveness of relu neural networks.SIAM Journal on Applied Algebra and Geometry, 6(2):216–242, 2022","cited_arxiv_id":null,"evidence_quote":"Supplies the canonical polyhedral complex and bent hyperplane arrangement that define the ReLU fan for a network."},{"cited_title":"Tropical geometry of deep neu- ral networks","cited_arxiv_id":null,"evidence_quote":"Establishes that ReLU network outputs are tropical rational functions, providing the tropical-geometry connection the paper extends to toric geometry."}],"review_version":1}