{"id":"e0bab817-beae-4ef1-acc1-77647f51188a","arxiv_id":"2411.09508","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For an arrangement of hypersurfaces, the likelihood ideal of its likelihood correspondence is the Rees ideal of the likelihood module, and an arrangement is gentle exactly when the pre-likelihood ideal is prime.","lead":"This paper proves that the likelihood ideal of an arrangement of hypersurfaces equals the Rees ideal of its likelihood module, linking algebraic statistics with the theory of logarithmic derivations. The new notion of a gentle arrangement makes maximum likelihood degrees easy to compute, with applications to statistical models and scattering amplitudes in physics.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The likelihood ideal is not always prime: a concrete arrangement in P2 with n=3, m=4 has a reducible likelihood correspondence, so Theorem 1.1 and Proposition 2.9 are false as stated.","rationale":"In good faith, the paper's goal is to identify the likelihood ideal with the Rees ideal of the likelihood module. For this to hold, I(A) must be prime. The reader correctly identified the proof of primeness as the weak point, noting circularity in Proposition 2.9 and the unjustified vector-bundle rank in Lemma 2.4. My stress-test finds the issue is not merely a missing proof: the statement itself is false. The explicit arrangement above has a rank drop of the logarithmic derivative matrix on x=0 while all defining polynomials remain nonzero and the image point is smooth, so the dropped-rank locus is a genuine part of the likelihood correspondence. The generic component and the rank-drop component have the same dimension and are distinct, so LA is reducible. This directly contradicts Theorem 1.1 and the saturation formula in Proposition 2.9 and Remark 2.10. Credit is due for the many correct examples and the software, but the central theorem needs an additional hypothesis such as the Jacobian of F having constant maximal rank on the open locus where all fi are nonzero, or F being an immersion. Without such a condition, the paper's main identification cannot hold.","tokens_in":18428,"tokens_out":37441,"duration_ms":346505,"concrete_test":"Using Macaulay2, define R=QQ[x,y,z,s1,s2,s3,s4], F=(x^2+2*y^2, 2*x^2+y^2, x^2+3*y^2, z), and form the ideal J generated by the cleared likelihood numerators N_x = 2*x*(s1*f2*f3 + 2*s2*f1*f3 + s3*f1*f2), N_y = 2*y*(2*s1*f2*f3 + s2*f1*f3 + 3*s3*f1*f2), and N_z = s4. Compute I = saturate(J, f1*f2*f3*z). If I is not prime and has a minimal prime (x, s1+s2+s3, s4), the central theorem is refuted. Equivalently, check that the point (x,y,z,s)=(0,1,1,1,-1,0,0) satisfies the defining equations of LA and is not the limit of the generic solution curve.","verdict_should_be":"REJECT","load_bearing_attack":"Lemma 2.4 asserts I(A) is prime because LA is a vector bundle of rank m-n over the maximal-rank locus of the likelihood equations. The gap is that the rank can drop on a locus where every fi is nonzero and F(x) is still a smooth point of X; such points are genuine solutions, and they can form extra irreducible components of the closure. Concretely, take n=3, m=4 and A = {x^2+2y^2, 2x^2+y^2, x^2+3y^2, z} in P2. On x=0 with y,z nonzero, the x-derivative likelihood equation vanishes identically, and the remaining equations force s4=0 and s1+s2+s3=0. Thus the set {x=0, s1+s2+s3=0, s4=0} is contained in LA. The generic graph over x != 0 sends x to the single projective point [5f1:-f2:-3f3:0], whose limit over x=0 is the one point [10:-1:-9:0]. Hence the extra line over x=0 is not in the closure of the generic component, LA is reducible, and I(A) is not prime. Since every Rees algebra R[s]/I' is a domain, R[s]/I(A) cannot equal the Rees algebra of M(A). The saturation step I(A)=I0:p^infinity in Proposition 2.9 also fails here: the maximal minors of the Jacobian are multiples of xy, so the paper's p contains x and the saturation removes the genuine component. The root cause is that F(x) in Xreg does not imply dF_x has maximal rank.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a new algebraic framework for the likelihood correspondence of an arrangement of hypersurfaces. It defines a likelihood module M(A) via the kernel of a matrix containing the polynomials and their partial derivatives, constructs the pre-likelihood ideal I0(A) and the Rees ideal of M(A), and claims (Theorem 1.1) that the Rees ideal equals the likelihood ideal I(A) of the likelihood correspondence. The notion of a 'gentle' arrangement is introduced to describe when I0(A) = I(A). The paper also studies the relationship between gentleness, freeness, and tameness, gives a combinatorial description for graphic arrangements, and provides Macaulay2 code. However, the central theorem is false as stated, and the proof of a key proposition is circular.","tokens_in":18689,"tokens_out":26681,"duration_ms":241626,"significance":"If correct, the main theorem would give a powerful and computable description of likelihood correspondences via modules of logarithmic derivations, and would unify hyperplane and nonlinear arrangements. The paper contains several concrete and reproducible computations, a new concept (gentleness) with intriguing conjectures, and interesting connections to scattering equations and toric models. Nevertheless, the central claim fails for a simple explicit arrangement, so the foundational results as stated cannot be accepted. The computational and combinatorial parts may still be valuable, but the main theorem and its proof require substantial revision.","major_comments":[{"comment":"The assertion that I(A) is prime is false. Consider the arrangement in P^2 given by f1 = x^2+2y^2, f2 = 2x^2+y^2, f3 = x^2+3y^2, f4 = z. For x ≠ 0 the critical equations have the unique solution s = [5f1 : -f2 : -3f3 : 0]. The Zariski closure of this graph intersects the hyperplane {x=0} in a 1-dimensional set, namely ({x=0} × [10:-1:-9:0]) ∪ ({[0:0:1]} × L), where L = {s1+s2+s3=0, s4=0}. On the other hand, every point of E = {x=0} × L satisfies the defining conditions of L_A: for y,z ≠ 0 one has ∂ℓ/∂x = 0 identically, ∂ℓ/∂y = 2(s1+s2+s3)/y = 0, ∂ℓ/∂z = s4/z = 0, all f_i are nonzero, and F(x) is a smooth point of the image (a linear space). Since E has dimension 2 and the generic component meets {x=0} in a 1-dimensional set, E is an additional irreducible component of L_A. Hence L_A is reducible and I(A) is not prime. As the Rees algebra of M(A) is a domain, R[s]/I(A) cannot equal the Rees algebra, contradicting Theorem 1.1. The flaw in Lemma 2.4 is that F(x) ∈ X_reg does not imply that the Jacobian dF has maximal rank; here the rank drops on {x=0} while F(x) is still smooth.","section":"§2, Lemma 2.4 and Theorem 1.1"},{"comment":"The proof of Proposition 2.9 is circular. It states: 'The likelihood ideal I(A) is always prime, since the Rees algebra is a domain whenever R is.' The Rees algebra being a domain implies that the Rees ideal I' (the kernel in (3)) is prime; the equality I(A) = I' is precisely the content of Theorem 1.1, which is what Proposition 2.9 is being used to prove. Moreover, the saturation formula I(A) = (I0(A) : p^∞) fails in the counterexample of the previous comment: the maximal minors of the Jacobian are multiples of xy, so the element p from Remark 2.10 contains x, and saturating at p removes the genuine component {x=0} × L of L_A. Thus Proposition 2.9 is false as stated.","section":"§2, Proposition 2.9"}],"minor_comments":[{"comment":"The notation s_+ is used in the displayed decomposition of the pre-likelihood ideal but is never defined; it should be introduced as s_+ = s1+s2+s3+s4.","section":"Example 3.5"},{"comment":"The text 'Computing P = I0 : f reveals the second minimal prime' refers to an additional generator f of degree (3,3) with 3092 terms, but f is not explicitly defined at that point; please state its construction or refer clearly to the earlier computation.","section":"Example 5.1"},{"comment":"In item 3, the phrase 'the homological degree is shifted by one' is not fully explained; a sentence clarifying that the same matrix A serves as the first map in the resolution of coker(A) and as the presentation of im(A) would improve readability.","section":"Proof of Lemma 2.2"}],"recommendation":"reject","confidential_remarks":"The explicit counterexample in the major comments appears to be decisive: the likelihood correspondence is reducible for an arrangement of three conics and a line in P^2, and the paper's main theorem fails for a case that lies entirely within the stated scope. The authors could potentially repair the paper by adding hypotheses (for instance, requiring the Jacobian to have maximal rank at all smooth points of the image, or restricting to birational parametrizations), but such assumptions would exclude several of the paper's own examples and would change the nature of the main contribution. The combinatorial and computational sections (Sections 4-6) retain independent interest and could form the basis of a revised submission. I recommend rejection of the current version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Thanks for the report. I read the paper and checked the stress-test example. The stress-test is right: the main theorem is false as stated.\n\nThe paper's core idea—that the likelihood ideal of an arrangement is the Rees ideal of the likelihood module—is attractive, and the gentle/tame/free analysis is a useful new perspective. The octahedron example, the separator-based generators, and the Macaulay2 package are genuine contributions. The exposition is clear and the authors know the literature.\n\nBut the load-bearing claim, Lemma 2.4, that I(A) is always prime, is false. The counterexample from the stress-test works. Take A = {x²+2y², 2x²+y², x²+3y², z} in P². The image X is the hyperplane 5u−v−3w=0, so every point is smooth. At x=0, the Jacobian of F has rank 2, not 3, and the likelihood equations force s4=0 and s1+s2+s3=0. This gives a whole line of s-values over the line x=0, which is not in the closure of the generic graph: over x≠0 the unique solution is s = [−5f1 : f2 : 3f3 : 0], which limits to the single point [−10:1:9:0] over x=0. So LA is reducible and I(A) is not prime. This also kills Proposition 2.9: saturating I0 by the maximal minors (which contain x) removes the genuine component. Remark 2.10's claim that X_reg can be replaced by maximal Jacobian rank without changing the closure is exactly wrong.\n\nThe proof of Proposition 2.9 is circular, as the reader noted, but the counterexample is the deeper problem. The theorem needs an additional hypothesis, such as requiring the parametrization to be generically finite with maximal Jacobian rank on the domain, or redefining the likelihood correspondence via the maximal-rank locus. Under such a definition the Rees-algebra isomorphism might survive, but not for the setup as written.\n\nMinor issues: Theorem 5.2's proof says 'fewer than six vertices' while the statement says n ≤ 6, and the exhaustive computation is not shown. Example 6.4 relies on numerical associated primes. These are secondary.\n\nThis paper deserves a serious referee, but not acceptance in its current form. The examples and tools are worth keeping; the main theorem needs restatement and a correct proof. I'd bring it to a reading group to discuss the counterexample, but I wouldn't cite the main theorem as reliable until it is fixed.","headline":"A good idea with a false main theorem: the likelihood ideal is not always prime, so Theorem 1.1 needs a corrected hypothesis.","tokens_in":19286,"tokens_out":14528,"would_cite":false,"duration_ms":123505,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["13A30","13N15","52C35","62R01"],"pacs":[],"model":"deepseek-v4-flash","headline":"For every hypersurface arrangement, the likelihood ideal equals the Rees ideal of a module of logarithmic derivations.","keywords":["likelihood correspondence","Rees algebra","logarithmic derivations","likelihood module","gentle arrangements","graphic arrangements","maximum likelihood degree","scattering equations"],"falsifier":"Take the octahedron arrangement of Example 5.1 and compute the saturation $I_0(\\mathcal{A}) : (x_1-x_2)^\\infty$ and, from the same matrix $A$, the kernel of the map $R[s] \\to \\mathcal{R}(M(\\mathcal{A}))$. The theorem predicts both equal the prime likelihood ideal with the extra degree-$(3,3)$ generator reported in the paper; any difference between the two ideals, in generators or in multidegree, would falsify Theorem 1.1.","tokens_in":18157,"feed_emoji":"📐","tokens_out":19812,"duration_ms":183640,"temperature":0.7,"pith_summary":"This paper builds a bridge from the classical theory of hypersurface arrangements to likelihood geometry. Its central claim is that for any arrangement $\\mathcal{A}$ of hypersurfaces in $\\mathbb{P}^{n-1}$, the ideal $I(\\mathcal{A})$ of the likelihood correspondence is the Rees ideal of a module $M(\\mathcal{A})$ built from logarithmic derivations; equivalently, $R[s]/I(\\mathcal{A})$ is the Rees algebra of $M(\\mathcal{A})$. If that is right, the usually troublesome step of clearing denominators in the likelihood equations is handled automatically by the module of derivations, and the maximum likelihood degree can be read off from the multidegree of the correspondence. The paper introduces gentle arrangements, for which the likelihood ideal is simply the pre-likelihood ideal generated by evaluating derivations on the log-likelihood; tame linear arrangements are gentle, while the octahedron gives the smallest non-gentle graphic arrangement. The identification is used to compute likelihood degrees in parametric statistical models, toric models, and scattering equations.","feed_headline":"A single module decides the likelihood geometry of any arrangement","feed_subtitle":"If true, the hard denominator-clearing step disappears: maximum-likelihood degrees come from logarithmic derivations.","key_machinery":"Central machinery is the matrix $Q = [\\operatorname{diag}(f_i), \\operatorname{Jac}(F)]$ of size $m \\times (m+n)$ together with the graded modules it defines: the Jacobian syzygy module $S(\\mathcal{A})$, the log-derivation module $D_{\\log}(\\mathcal{A})$, and the likelihood module $M(\\mathcal{A}) = \\operatorname{coker}(A)$. The Rees algebra of a module, the quotient of its symmetric algebra by its $R$-torsion, is the object that records the polynomial relations among the module's generators after localization. The paper proves that this Rees algebra is the coordinate ring of the likelihood correspondence, so the ideal of the correspondence is the saturation $I(\\mathcal{A}) = (I_0(\\mathcal{A}) : p^\\infty)$ of the pre-likelihood ideal. Gentleness of $\\mathcal{A}$ means the symmetric algebra is already torsion-free, making the saturation step unnecessary.","core_discovery":"The paper's central discovery is Theorem 1.1: for any arrangement $\\mathcal{A}$ of hypersurfaces, the quotient $R[s]/I(\\mathcal{A})$ is the Rees algebra of the likelihood module $M(\\mathcal{A})$, and therefore the likelihood ideal $I(\\mathcal{A})$ equals the Rees ideal of $M(\\mathcal{A})$. The likelihood module is $\\operatorname{coker}(A)$, where the columns of $\\binom{A}{B}$ generate the kernel of the matrix $Q$ whose rows are $(f_i, \\partial f_i/\\partial x_1,\\dots,\\partial f_i/\\partial x_n)$; its first syzygy module is the module of logarithmic derivations $D_{\\log}(\\mathcal{A})$. The pre-likelihood ideal $I_0(\\mathcal{A}) = \\langle (s_1,\\dots,s_m) A \\rangle$ presents the symmetric algebra of $M(\\mathcal{A})$, and the likelihood ideal is its saturation at an element $p$ for which $M(\\mathcal{A})[p^{-1}]$ is free. An arrangement is gentle when no saturation is needed, that is, when $I_0(\\mathcal{A})$ is already prime and hence equal to $I(\\mathcal{A})$; in that case the likelihood ideal is generated by the evaluations of logarithmic derivations on the log-likelihood, as stated in Theorem 2.11. In the hyperplane case tame arrangements are gentle, and the octahedron arrangement is the unique smallest non-gentle graphic arrangement.","pith_inferences":["Editorial inference: the saturation step that separates $I_0(\\mathcal{A})$ from $I(\\mathcal{A})$ can be read as deleting spurious critical points supported on the singular locus of the parametrization; a testable prediction is that saturating only at the Jacobian ideal of $F$ already produces $I(\\mathcal{A})$ whenever the parametrization is generically finite.","Editorial inference: because gentleness depends on the chosen presentation of $M(\\mathcal{A})$, the same statistical model can be gentle in one parametrization and non-gentle in another, as with the independence model in Examples 3.5 and 3.6; this suggests asking whether every model admits a gentle parametrization and whether a minimal gentle degree over parametrizations is a computable invariant.","Editorial inference: Theorem 1.1 implies non-gentleness is exactly the $R$-torsion of $\\operatorname{Sym}(M(\\mathcal{A}))$; if so, depth tests on the Fitting ideals of $M(\\mathcal{A})$ can certify gentleness before any saturation is attempted, which would make a search for non-gentle arrangements beyond the octahedron feasible for much larger graphs."],"forward_implications":["For gentle arrangements, the likelihood ideal is generated by the evaluations $\\theta(\\ell_\\mathcal{A})$ of logarithmic derivations, so the maximum likelihood degree is computable from the pre-likelihood ideal without saturation; for free gentle arrangements it equals the product of the positive column degrees of $A$.","Every tame linear arrangement is gentle, so for hyperplane arrangements gentleness is common; in particular all line arrangements in $\\mathbb{P}^2$ are gentle.","The octahedron is the unique smallest non-gentle graphic arrangement, and any graph containing the octahedron as an induced subgraph is not gentle; the paper conjectures a contraction-based characterization of gentle graphic arrangements.","The identification converts several concrete models into arrangements with explicit likelihood ideals: the two-coin model has ML degree 24, the binary independence model has ML degree 1, the no-three-way interaction model has ML degree 3, and the braid arrangement attached to $M_{0,n}$ has ML degree $(n-3)!$."],"supporting_citations":[{"why":"Supplies the module of logarithmic derivations and the derivation basis for the braid arrangement used in the proof of Theorem 2.11 and Example 5.7.","marker":"[27]"},{"why":"Provides the theory of Rees and symmetric algebras of modules on which the pre-likelihood ideal and Theorem 1.1 rest.","marker":"[28]"},{"why":"Establishes in the linear case that evaluations of derivations generate the (pre-)likelihood ideal and supplies the tameness results behind Theorem 4.3.","marker":"[8]"},{"why":"Sets up parametric likelihood equations and the coin model whose ML degree 24 is recomputed in Example 3.4.","marker":"[18]"},{"why":"Defines the likelihood correspondence and the implicit maximum likelihood degree that the paper reformulates in parametric arrangement terms.","marker":"[20]"},{"why":"Foundational study of freeness for hyperplane arrangements, invoked for free graphic arrangements and their derivation modules.","marker":"[31]"},{"why":"Provides the separator-based derivations that Theorem 5.8 uses to generate the pre-likelihood ideal of every graphic arrangement.","marker":"[25]"},{"why":"Connects likelihood equations to scattering amplitudes on $M_{0,n}$, giving the $(n-3)!$ ML degree and the reality of critical points used in Section 3.","marker":"[30]"},{"why":"Gives the signed-Euler-characteristic formula for ML degrees of very affine varieties used in the toric-model discussion.","marker":"[19]"}],"fun_headline_variants":["Logarithmic derivations determine likelihood ideals of arrangements","Gentle arrangements skip saturation in likelihood computation","Rees algebra of likelihood module encodes arrangement likelihood","Nonlinear arrangements: likelihood from logarithmic derivations","Likelihood ideal as Rees ideal: a single module decides geometry"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument requires the likelihood ideal to have a single irreducible component, and the supplied proof of that fact cites the Rees algebra being a domain, which is the content of Theorem 1.1 itself; without an independent primeness proof the saturation step is unsupported.","fun_headline_variants_meta":{"raw":{"variants":["Logarithmic derivations determine likelihood ideals of arrangements","Gentle arrangements skip saturation in likelihood computation","Rees algebra of likelihood module encodes arrangement likelihood","Nonlinear arrangements: likelihood from logarithmic derivations","Likelihood ideal as Rees ideal: a single module decides geometry"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000559,"raw_usage":{"total_tokens":2628,"prompt_tokens":884,"completion_tokens":1744,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":1670}},"tokens_in":500,"tokens_out":1744,"duration_ms":12326,"temperature":1.0,"reasoning_tokens":1670,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:37:18.024525+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the octahedron arrangement of Example 5.1 and compute the saturation $I_0(\\mathcal{A}) : (x_1-x_2)^\\infty$ and, from the same matrix $A$, the kernel of the map $R[s] \\to \\mathcal{R}(M(\\mathcal{A}))$. The theorem predicts both equal the prime likelihood ideal with the extra degree-$(3,3)$ generator reported in the paper; any difference between the two ideals, in generators or in multidegree, would falsify Theorem 1.1.","supporting_citations":[{"cited_title":"Saito: Theory of logarithmic differential forms and logarithmic vector fields , Journal of the Faculty of Science, University of Tokyo, Section IA Math","cited_arxiv_id":null,"evidence_quote":"Supplies the module of logarithmic derivations and the derivation basis for the braid arrangement used in the proof of Theorem 2.11 and Example 5.7."},{"cited_title":"Simis, B","cited_arxiv_id":null,"evidence_quote":"Provides the theory of Rees and symmetric algebras of modules on which the pre-likelihood ideal and Theorem 1.1 rest."},{"cited_title":"Cohen, G","cited_arxiv_id":null,"evidence_quote":"Establishes in the linear case that evaluations of derivations generate the (pre-)likelihood ideal and supplies the tameness results behind Theorem 4.3."},{"cited_title":"Ho¸ sten, A","cited_arxiv_id":null,"evidence_quote":"Sets up parametric likelihood equations and the coin model whose ML degree 24 is recomputed in Example 3.4."},{"cited_title":"Huh and B","cited_arxiv_id":null,"evidence_quote":"Defines the likelihood correspondence and the implicit maximum likelihood degree that the paper reformulates in parametric arrangement terms."},{"cited_title":"Terao: Arrangements of hyperplanes and their freeness","cited_arxiv_id":null,"evidence_quote":"Foundational study of freeness for hyperplane arrangements, invoked for free graphic arrangements and their derivation modules."},{"cited_title":"Sturmfels and S","cited_arxiv_id":null,"evidence_quote":"Connects likelihood equations to scattering amplitudes on $M_{0,n}$, giving the $(n-3)!$ ML degree and the reality of critical points used in Section 3."},{"cited_title":"Huh: The maximum likelihood degree of a very affine variety , Compositio Mathemat- ica, 149 (2013) 1245–1266","cited_arxiv_id":null,"evidence_quote":"Gives the signed-Euler-characteristic formula for ML degrees of very affine varieties used in the toric-model discussion."}],"review_version":1}