arxiv: 2604.02995 · v2 · submitted 2026-04-03 · 🧮 math.AG · cs.LG· math.CO

Recognition: 2 theorem links

· Lean Theorem

A semicontinuous relaxation of Saito's criterion and freeness as angular minimization

Tom\'as S. R. Silva

Authors on Pith no claims yet

Pith reviewed 2026-05-13 18:00 UTC · model grok-4.3

classification 🧮 math.AG cs.LGmath.CO

keywords line arrangementsfree arrangementsSaito criterionTerao conjecturelogarithmic derivationsprojective plane

0 comments

The pith

A nonnegative functional on line arrangements in the projective plane vanishes exactly when the arrangement is free.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper defines a functional S on arrangements of n lines in P^2 that returns zero precisely when the arrangement admits a free basis of logarithmic derivations with given exponents. The construction relaxes Saito's criterion by parameterizing candidate derivation spaces through null spaces of derivation matrices, forming a bilinear map whose image is compared to the arrangement's defining polynomial Q(A) in coefficient space. The value of S is the squared sine of the angle between that image and the direction of Q(A), yielding a concrete distance to freeness. Upper semicontinuity of S on strata produces a functional version of Terao's conjecture. The same functional then serves as a reward or filter in computational searches that locate verified free arrangements for every admissible exponent pair with n up to 20.

Core claim

Given an arrangement A with candidate exponents (d1, d2), the spaces of logarithmic derivations of degrees d1 and d2 are identified with null spaces of the corresponding derivation matrices. The Saito determinant is realized as a bilinear map from the product of these spaces into the vector space of degree-n homogeneous polynomials. The functional S is the squared sine of the angle between the image of this bilinear map and the one-dimensional subspace spanned by Q(A). This quantity is nonnegative, vanishes if and only if A is free with those exponents, and is upper semicontinuous on the natural strata of the space of arrangements.

What carries the argument

The bilinear Saito determinant map from the product of two derivation spaces to the space of degree-n polynomials, whose angular misalignment with Q(A) supplies the value of the functional S.

If this is right

Terao's conjecture becomes the statement that the global minimum of S is attained exactly on free arrangements.
Reinforcement learning guided by S discovers hundreds of distinct free arrangements for every admissible exponent pair with n at most 13.
A hybrid algebraic extension procedure, seeded by S as a pre-filter, produces at least one verified free arrangement for every admissible exponent pair with n up to 20.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Because S is defined via a continuous angle in a fixed vector space, its level sets could be used to deform arrangements continuously while tracking proximity to freeness.
The same angular construction might supply a quantitative measure of freeness for arrangements in higher-dimensional projective spaces or for other classes of hypersurface singularities.
Once S is available as a smooth objective, standard numerical optimization routines could locate candidate free arrangements without exhaustive enumeration of integer lattices.

Load-bearing premise

The angle measured in coefficient space between the image of the bilinear map and Q(A) is zero precisely when a free basis of the expected degrees exists.

What would settle it

An explicit arrangement of lines in P^2 for which S evaluates to zero yet no basis of logarithmic derivations of the candidate degrees exists, or a known free arrangement on which S is strictly positive.

read the original abstract

We introduce a nonnegative functional $\mathfrak{S}$ on the space of line arrangements in $\mathbb{P}^2$ that vanishes precisely on free arrangements, obtained as a semicontinuous relaxation of Saito's criterion. Given an arrangement $\mathcal{A}$ of $n$ lines with candidate exponents $(d_1, d_2)$, we parameterize the spaces of logarithmic derivations of degrees $d_1$ and $d_2$ via the null spaces of the associated derivation matrices and express the Saito determinant as a bilinear map into the space of degree-$n$ polynomials. The functional admits a natural geometric interpretation: it measures the squared sine of the angle between the image of this bilinear map and the direction of the defining polynomial $Q(\mathcal{A})$ in coefficient space, providing a computable measure of how far an arrangement is from admitting a free basis of logarithmic derivations of the expected degrees. We prove that $\mathfrak{S}$ is upper semicontinuous on natural strata, and use this to give a functional reformulation of Terao's conjecture. Beyond its theoretical interest, $\mathfrak{S}$ provides a viable computational handle on the landscape of free arrangements. We illustrate this through two complementary roles: as a smooth reward signal driving a reinforcement learning search for moderate $n$, and as a fast pre-filter accelerating an algebraic extension procedure for larger $n$. For $n \leq 13$, the reinforcement learning system discovers hundreds of verified free arrangements spanning all admissible exponent types. For $n \geq 14$, where the reinforcement learning reward signal becomes insufficient, the hybrid extension procedure -- combined with classical supersolvable constructions -- produces at least one verified free arrangement for every admissible exponent pair $(d_1, d_2)$ with $n \leq 20$.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This paper gives a workable functional S that vanishes exactly on free line arrangements and uses it to computationally verify free examples for all admissible exponents up to n=20.

read the letter

The core contribution is a nonnegative functional S on arrangements in P^2 that is zero precisely when the arrangement is free with given exponents. It comes from relaxing Saito's criterion into an angular distance in coefficient space between the image of the bilinear Saito map and the actual defining polynomial. That construction is direct and does not introduce extra zeros or miss genuine free cases, since the null spaces recover the logarithmic derivations and the determinant condition matches the classical criterion exactly. They prove upper semicontinuity on the natural strata and rewrite Terao's conjecture as the statement that S reaches zero on every admissible pair. On the practical side, they run reinforcement learning with S as reward to locate hundreds of verified free arrangements for n up to 13, then switch to a hybrid algebraic extension procedure (plus supersolvable constructions) that produces at least one verified free example for every admissible (d1, d2) when n reaches 20. The computational reach is the clearest advance over prior work. The math checks out from the definitions, and the stress-test note confirms there are no false zeros or missed cases built into S. The main soft spot is that the abstract and summary give little detail on how the RL outputs and extension steps were independently verified or on error bounds for the hybrid procedure; a referee would want explicit checks or code for the hundreds of small-n examples. The semicontinuity argument is stated but not expanded here, so the full write-up needs to show the strata and limits carefully. This is aimed at people already working on hyperplane arrangements and Terao's conjecture. A reader who wants new computational handles or a functional view of freeness will find concrete value in the search method and the reformulation. It is solid enough to deserve a serious referee, even if the computational sections need more documentation. I would send it to review rather than desk-reject.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces a nonnegative functional 𝔖 on the space of line arrangements in ℙ² that vanishes precisely on free arrangements, constructed as a semicontinuous relaxation of Saito's criterion. Given an arrangement 𝒜 of n lines with candidate exponents (d₁, d₂), the spaces of logarithmic derivations are parameterized via null spaces of derivation matrices, and the Saito determinant is expressed as a bilinear map; 𝔖 is then the squared sine of the angle between the image of this map and Q(𝒜) in coefficient space. The paper proves that 𝔖 is upper semicontinuous on natural strata and gives a functional reformulation of Terao's conjecture. It further applies 𝔖 as a reward signal in reinforcement learning to discover hundreds of free arrangements for n ≤ 13 and in a hybrid algebraic extension procedure (combined with supersolvable constructions) to produce verified free arrangements for every admissible exponent pair with n ≤ 20.

Significance. If the central claims hold, the functional provides a direct, parameter-free, and computable measure of distance from freeness that is derived exactly from the classical Saito determinant and derivation spaces. The upper semicontinuity and reformulation of Terao's conjecture as a minimization problem constitute a genuine theoretical contribution. The computational illustrations—RL discovery of hundreds of examples for moderate n and complete coverage of admissible exponents up to n=20—are strengths that demonstrate practical utility, provided the verifications are fully rigorous.

major comments (2)

[Computational Applications] § on computational results (RL search and hybrid extension): the claim that the RL system discovers 'hundreds of verified free arrangements' for n ≤ 13 and that the hybrid procedure yields at least one verified free arrangement for every admissible (d₁, d₂) with n ≤ 20 is load-bearing for the practical contribution. No explicit description is given of the algebraic verification procedure (e.g., explicit computation of the Saito determinant or Gröbner-basis checks) nor of numerical error bounds on the angular minimization step; without these, the verification statements cannot be independently confirmed.
[Proof of upper semicontinuity] Proof of upper semicontinuity: the abstract asserts that 𝔖 is upper semicontinuous on natural strata, but the manuscript should specify the topology (Zariski or classical) and the precise stratification (e.g., fixed n and fixed candidate exponents (d₁, d₂)). If the strata are defined by the rank of the derivation matrices, the semicontinuity argument must be checked against possible jumps in null-space dimension.

minor comments (2)

[Introduction / Definition of 𝔖] The notation 𝔖 is introduced without an earlier non-fraktur symbol; a brief remark on why the fraktur font is chosen would prevent confusion with any auxiliary scalar S that might appear in intermediate calculations.
[Abstract] The abstract states that the RL search spans 'all admissible exponent types'; a short table or explicit list of the discovered exponent pairs (d₁, d₂) with their multiplicities would make the coverage claim immediately verifiable.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading, positive evaluation of the theoretical contribution, and constructive suggestions on the computational and semicontinuity sections. We address the two major comments point by point below, indicating the revisions that will be incorporated.

read point-by-point responses

Referee: [Computational Applications] § on computational results (RL search and hybrid extension): the claim that the RL system discovers 'hundreds of verified free arrangements' for n ≤ 13 and that the hybrid procedure yields at least one verified free arrangement for every admissible (d₁, d₂) with n ≤ 20 is load-bearing for the practical contribution. No explicit description is given of the algebraic verification procedure (e.g., explicit computation of the Saito determinant or Gröbner-basis checks) nor of numerical error bounds on the angular minimization step; without these, the verification statements cannot be independently confirmed.

Authors: We agree that an explicit description of the verification pipeline is necessary for independent confirmation. In the revised manuscript we will insert a dedicated subsection (new §5.3) that details the post-discovery algebraic verification: for each candidate arrangement found by RL or the hybrid procedure, we compute the derivation matrices over ℚ, extract bases for the null spaces via exact linear algebra, form the Saito bilinear map, and check that the resulting determinant polynomial is identically zero by Gröbner-basis reduction (using Macaulay2 or Singular). Because all steps are performed in exact arithmetic, no numerical error bounds are required for the final verification step; the angular-minimization stage of 𝔖 is used only as a heuristic filter and is never relied upon for the published claims of freeness. We will also tabulate the precise exponent pairs and the software scripts used, so that every listed example can be reproduced. revision: yes
Referee: [Proof of upper semicontinuity] Proof of upper semicontinuity: the abstract asserts that 𝔖 is upper semicontinuous on natural strata, but the manuscript should specify the topology (Zariski or classical) and the precise stratification (e.g., fixed n and fixed candidate exponents (d₁, d₂)). If the strata are defined by the rank of the derivation matrices, the semicontinuity argument must be checked against possible jumps in null-space dimension.

Authors: We will revise the statement and proof to specify that upper semicontinuity holds in the classical (Euclidean) topology on the strata 𝒮_{n,d₁,d₂} consisting of all arrangements of exactly n lines whose candidate exponents are the fixed pair (d₁,d₂). These strata are further partitioned into locally closed pieces according to the rank of the two derivation matrices; on each such piece the dimensions of the logarithmic derivation spaces are constant, the bilinear Saito map varies continuously, and the squared-sine functional 𝔖 is therefore continuous. At the boundaries where rank drops occur, the image of the bilinear map can only shrink or stay the same, so the distance to Q(𝒜) cannot decrease; hence 𝔖 is upper semicontinuous across the whole stratum. A short remark will be added after the proof to record this stratification and the handling of rank jumps. revision: yes

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The functional 𝔖 is explicitly constructed from the null spaces of the derivation matrices and the bilinear map reproducing the Saito determinant; by this definition 𝔖 vanishes precisely when the classical Saito criterion holds for the given exponents. The upper semicontinuity proof and the functional reformulation of Terao's conjecture are direct consequences of this construction rather than reductions to fitted parameters or self-referential loops. No load-bearing self-citations, imported uniqueness theorems, or ansatzes smuggled via prior work appear in the derivation chain. The computational uses (RL reward, pre-filter) are applications of the defined object, not circular claims.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The construction rests on the classical theory of logarithmic derivations and Saito's criterion; no new free parameters or invented entities beyond the functional S itself are introduced.

axioms (1)

standard math Standard properties of the module of logarithmic derivations and the Saito determinant for line arrangements
Invoked throughout the definition of S and the parameterization of derivation spaces.

invented entities (1)

Functional S no independent evidence
purpose: Nonnegative measure of distance to freeness via angular deviation
Newly defined in the paper as the squared sine of the angle between the bilinear image and Q(A).

pith-pipeline@v0.9.0 · 5629 in / 1260 out tokens · 95717 ms · 2026-05-13T18:00:57.030808+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel echoes

?

echoes
ECHOES: this paper passage has the same mathematical shape or conceptual pattern as the Recognition theorem, but is not a direct formal dependency.

S(A) = 1−sup ⟨T(α1,α2),q⟩² / (∥T∥² ∥q∥²) ... vanishes precisely on free arrangements ... squared sine of the angle

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

53 extracted references · 53 canonical work pages · 3 internal anchors

[1]

Journal of the Faculty of Science, University of Tokyo, Section IA, Mathematics 27(2), 265–291 (1980)

Saito, K.: Theory of logarithmic differential forms and logarithmic vector fields. Journal of the Faculty of Science, University of Tokyo, Section IA, Mathematics 27(2), 265–291 (1980)

work page 1980
[2]

Springer, Berlin (1992)

Orlik, P., Terao, H.: Arrangements of Hyperplanes. Springer, Berlin (1992). https:// doi.org/10.1007/978-3-662-02772-1 .http://dx.doi.org/10.1007/978-3-662-02772-1

work page doi:10.1007/978-3-662-02772-1 1992
[3]

Springer, Cham (2017)

Dimca, A.: Hyperplane Arrangements. Springer, Cham (2017). https://doi.org/10. 1007/978-3-319-56221-6 .http://dx.doi.org/10.1007/978-3-319-56221-6

work page doi:10.1007/978-3-319-56221-6 2017
[4]

Annales de la Facult´ e des sciences de Toulouse: Math´ ematiques23(2), 483–512 (2014) https://doi.org/10.5802/afst.1413

Yoshinaga, M.: Freeness of hyperplane arrangements and related topics. Annales de la Facult´ e des sciences de Toulouse: Math´ ematiques23(2), 483–512 (2014) https://doi.org/10.5802/afst.1413

work page doi:10.5802/afst.1413 2014
[5]

Transactions of the American Mathematical Society377(11), 7633–7655 (2024) https://doi.org/10.1090/tran/9293

Jardim, M., Nasrollah Nejad, A., Simis, A.: The Bourbaki degree of a plane projective curve. Transactions of the American Mathematical Society377(11), 7633–7655 (2024) https://doi.org/10.1090/tran/9293

work page doi:10.1090/tran/9293 2024
[6]

Proceedings of the American Mathematical Society118(3), 745–752 (1993) https: //doi.org/10.1090/s0002-9939-1993-1160307-6

Yuzvinsky, S.: Free and locally free arrangements with a given intersection lattice. Proceedings of the American Mathematical Society118(3), 745–752 (1993) https: //doi.org/10.1090/s0002-9939-1993-1160307-6

work page doi:10.1090/s0002-9939-1993-1160307-6 1993
[7]

Terao, H.: Arrangements of hyperplanes and their freeness I, II. J. Fac. Sci. Univ. Tokyo Sect. IA Math.27(1980)

work page 1980
[8]

https://arxiv.org/abs/1101.0356

Schenck, H.: Hyperplane arrangements: computations and conjectures (2012). https://arxiv.org/abs/1101.0356

work page arXiv 2012
[9]

Michigan Mathematical Journal66(2) (2017) https://doi.org/10.1307/mmj/1490639821 15

Dimca, A.: Curve arrangements, pencils, and jacobian syzygies. Michigan Mathematical Journal66(2) (2017) https://doi.org/10.1307/mmj/1490639821 15

work page doi:10.1307/mmj/1490639821 2017
[10]

Revista Matem´ atica Complutense30(2), 259–268 (2017) https: //doi.org/10.1007/s13163-017-0228-3

Dimca, A., Sticlaru, G.: On the exponents of free and nearly free projective plane curves. Revista Matem´ atica Complutense30(2), 259–268 (2017) https: //doi.org/10.1007/s13163-017-0228-3

work page doi:10.1007/s13163-017-0228-3 2017
[11]

Mathematical Proceedings of the Cambridge Philosophical Society163(1), 161–172 (2016) https://doi.org/10.1017/s0305004116000803

Dimca, A.: Freeness versus maximal global tjurina number for plane curves. Mathematical Proceedings of the Cambridge Philosophical Society163(1), 161–172 (2016) https://doi.org/10.1017/s0305004116000803

work page doi:10.1017/s0305004116000803 2016
[12]

´Epijournal de G´ eom´ etrie Alg´ ebriqueVolume 7(2023) https://doi.org/10.46298/ epiga.2023.7323

Marchesi, S., Vall` es, J.: Triangular arrangements on the projective plane. ´Epijournal de G´ eom´ etrie Alg´ ebriqueVolume 7(2023) https://doi.org/10.46298/ epiga.2023.7323

work page arXiv 2023
[13]

https://arxiv.org/abs/2509.19011

Abe, T., K¨ uhne, L., Pokora, P.: Addition theorems for Ziegler pairs of hyperplane arrangements (2025). https://arxiv.org/abs/2509.19011

work page arXiv 2025
[14]

https://arxiv.org/abs/2505.01733

Dimca, A., K¨ uhne, L., Pokora, P.: Free line arrangements with low maximal multiplicity (2025). https://arxiv.org/abs/2505.01733

work page arXiv 2025
[15]

https://arxiv.org/abs/2407.07070

K¨ uhne, L., Luber, D., Pokora, P.: On the numerical Terao’s conjecture and Ziegler pairs for line arrangements (2024). https://arxiv.org/abs/2407.07070

work page arXiv 2024
[16]

Rendiconti del Circolo Matematico di Palermo Series 273(7), 2501–2512 (2024) https://doi.org/10.1007/s12215-024-01047-x

K¨ uhne, L., Szemberg, T., Tutaj-Gasi´ nska, H.: Line arrangements with many triple points. Rendiconti del Circolo Matematico di Palermo Series 273(7), 2501–2512 (2024) https://doi.org/10.1007/s12215-024-01047-x

work page doi:10.1007/s12215-024-01047-x 2024
[17]

https://arxiv.org/abs/2403.20024

Pokora, P., Roulleau, X.: Construction of free arrangements using point-line operators (2024). https://arxiv.org/abs/2403.20024

work page arXiv 2024
[18]

Taiwanese Journal of Mathematics1(1) (2024) https: //doi.org/10.11650/tjm/241105

Pokora, P.: Freeness of arrangements of lines and one conic with ordinary quasi- homogeneous singularities. Taiwanese Journal of Mathematics1(1) (2024) https: //doi.org/10.11650/tjm/241105

work page doi:10.11650/tjm/241105 2024
[19]

Journal of Algebra552, 38–51 (2020) https://doi.org/10.1016/j.jalgebra.2020.02.007

Abe, T., Dimca, A.: On complex supersolvable line arrangements. Journal of Algebra552, 38–51 (2020) https://doi.org/10.1016/j.jalgebra.2020.02.007

work page doi:10.1016/j.jalgebra.2020.02.007 2020
[20]

Periodica Mathematica Hungarica78(1), 31–37 (2018) https://doi.org/10.1007/s10998-018-0250-3

Kabat, J.: On the freeness of B¨ or¨ oczky line arrangements. Periodica Mathematica Hungarica78(1), 31–37 (2018) https://doi.org/10.1007/s10998-018-0250-3

work page doi:10.1007/s10998-018-0250-3 2018
[21]

Discrete & Computational Geometry53(4), 965–970 (2015) https://doi.org/10

Mu, L., Stanley, R.P.: Supersolvability and freeness for ψ-graphical arrangements. Discrete & Computational Geometry53(4), 965–970 (2015) https://doi.org/10. 1007/s00454-015-9684-z

work page 2015
[22]

SIAM Journal on Discrete Mathematics35(2), 1201–1223 (2021) https://doi.org/ 10.1137/19m1296744

Barakat, M., Behrends, R., Jefferson, C., K¨ uhne, L., Leuner, M.: On the generation of rank 3 simple matroids with an application to Terao’s freeness conjecture. SIAM Journal on Discrete Mathematics35(2), 1201–1223 (2021) https://doi.org/ 10.1137/19m1296744

work page doi:10.1137/19m1296744 2021
[23]

International Journal of 16 Data Science in the Mathematical Sciences01(01), 23–47 (2022) https://doi.org/ 10.1142/s2810939222500010

He, Y.-H.: Machine-learning mathematical structures. International Journal of 16 Data Science in the Mathematical Sciences01(01), 23–47 (2022) https://doi.org/ 10.1142/s2810939222500010

work page doi:10.1142/s2810939222500010 2022
[24]

In: Krantz, S.G., Rao, A.S.R.S., Rao, C.R

He, Y.-H., Heyes, E., Hirst, E.: Chapter 2 - Machine learning in physics and geometry. In: Krantz, S.G., Rao, A.S.R.S., Rao, C.R. (eds.) Artificial Intelligence. Handbook of Statistics, vol. 49, pp. 47–81. Elsevier, Amsterdam (2023). https://doi.org/10.1016/bs.host.2023.06.002 . https://www.sciencedirect.com/science/article/pii/S0169716123000263

work page doi:10.1016/bs.host.2023.06.002 2023
[25]

Garc´ ıa-Tecocoatzi, A

He, Y.-H., Jejjala, V., Silva, T.S.R.: Metaheuristic generation of brane tilings. Physics Letters B862, 139365 (2025) https://doi.org/10.1016/j.physletb.2025. 139365

work page doi:10.1016/j.physletb.2025 2025
[26]

Nature600(7887), 70–74 (2021) https://doi.org/10.1038/s41586-021-04086-x

Davies, A., Veliˇ ckovi´ c, P., Buesing, L., Blackwell, S., Zheng, D., Tomaˇ sev, N., Tanburn, R., Battaglia, P., Blundell, C., Juh´ asz, A., Lackenby, M., Williamson, G., Hassabis, D., Kohli, P.: Advancing mathematics by guiding human intuition with AI. Nature600(7887), 70–74 (2021) https://doi.org/10.1038/s41586-021-04086-x

work page doi:10.1038/s41586-021-04086-x 2021
[27]

Physics Letters B850, 138504 (2024) https: //doi.org/10.1016/j.physletb.2024.138504

Berglund, P., He, Y.-H., Heyes, E., Hirst, E., Jejjala, V., Lukas, A.: New Calabi–Yau manifolds from genetic algorithms. Physics Letters B850, 138504 (2024) https: //doi.org/10.1016/j.physletb.2024.138504

work page doi:10.1016/j.physletb.2024.138504 2024
[28]

Machine Learning: Science and Technology2(2), 025035 (2021) https://doi.org/10.1088/ 2632-2153/abe91f

Gukov, S., Halverson, J., Ruehle, F., Su lkowski, P.: Learning to unknot. Machine Learning: Science and Technology2(2), 025035 (2021) https://doi.org/10.1088/ 2632-2153/abe91f

work page 2021
[29]

Machine Learning: Science and Technology6(2), 025065 (2025) https://doi.org/10.1088/2632-2153/ade362

Gukov, S., Halverson, J., Manolescu, C., Ruehle, F.: Searching for ribbons with machine learning. Machine Learning: Science and Technology6(2), 025065 (2025) https://doi.org/10.1088/2632-2153/ade362

work page doi:10.1088/2632-2153/ade362 2025
[30]

https://arxiv.org/ abs/2602.12438

Heyes, E., Hirst, E., Earp, H.N.S., Silva, T.S.R.: Neural and numerical methods for G2-structures on contact Calabi-Yau 7-manifolds (2026). https://arxiv.org/ abs/2602.12438

work page arXiv 2026
[31]

Minimising Willmore Energy via Neural Flow

Hirst, E., Earp, H.N.S., Silva, T.S.R.: Minimising Willmore energy via neural flow (2026). https://arxiv.org/abs/2604.04321

work page internal anchor Pith review Pith/arXiv arXiv 2026
[32]

Physics Letters B850, 138517 (2024) https://doi.org/10.1016/j.physletb.2024.138517

Aggarwal, D., He, Y.-H., Heyes, E., Hirst, E., S´ a Earp, H.N., Silva, T.S.R.: Machine learning Sasakian and G2 topology on contact Calabi-Yau 7-manifolds. Physics Letters B850, 138517 (2024) https://doi.org/10.1016/j.physletb.2024.138517

work page doi:10.1016/j.physletb.2024.138517 2024
[33]

Inventiones Mathematicae63(1), 159–179 (1981) https://doi.org/10.1007/bf01389197

Terao, H.: Generalized exponents of a free arrangement of hyperplanes and shepherd-todd-brieskorn formula. Inventiones Mathematicae63(1), 159–179 (1981) https://doi.org/10.1007/bf01389197

work page doi:10.1007/bf01389197 1981
[34]

Adaptive Computa- tion and Machine Learning series

Sutton, R.S., Barto, A.G.: Reinforcement Learning, 2nd edn. Adaptive Computa- tion and Machine Learning series. Bradford Books, Cambridge, MA (2018) 17

work page 2018
[35]

rational cuspidal plane curves

Dimca, A., Sticlaru, G.: Free and nearly free curves vs. rational cuspidal plane curves. Publications of the Research Institute for Mathematical Sciences54(1), 163–179 (2018) https://doi.org/10.4171/prims/54-1-6

work page doi:10.4171/prims/54-1-6 2018
[36]

https://arxiv.org/abs/1505.00666

Dimca, A., Sticlaru, G.: Nearly free divisors and rational cuspidal curves (2015). https://arxiv.org/abs/1505.00666

work page arXiv 2015
[37]

International Mathematics Research Notices2021(12), 9233–9261 (2019) https: //doi.org/10.1093/imrn/rnz099

Abe, T.: Plus-one generated and next to free arrangements of hyperplanes. International Mathematics Research Notices2021(12), 9233–9261 (2019) https: //doi.org/10.1093/imrn/rnz099

work page doi:10.1093/imrn/rnz099 2019
[38]

Inventiones mathematicae 204(1), 317–346 (2015) https://doi.org/10.1007/s00222-015-0615-7

Abe, T.: Divisionally free arrangements of hyperplanes. Inventiones mathematicae 204(1), 317–346 (2015) https://doi.org/10.1007/s00222-015-0615-7

work page doi:10.1007/s00222-015-0615-7 2015
[39]

Bishop, C.: Pattern Recognition and Machine Learning

M. Bishop, C.: Pattern Recognition and Machine Learning. Information Science and Statistics. Springer, New York, NY (2016)

work page 2016
[40]

MIT Press, Cambridge, MA (2016)

Goodfellow, I., Bengio, Y., Courville, A.: Deep Learning. MIT Press, Cambridge, MA (2016). http://www.deeplearningbook.org

work page 2016
[41]

The MIT Press, Cambridge, MA (2023).http://udlbook.com

Prince, S.J.D.: Understanding Deep Learning. The MIT Press, Cambridge, MA (2023).http://udlbook.com

work page 2023
[42]

Proximal Policy Optimization Algorithms

Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms (2017). https://arxiv.org/abs/1707.06347

work page internal anchor Pith review Pith/arXiv arXiv 2017
[43]

SIAM Journal on Control and Optimization42(4), 1143–1166 (2003) https://doi.org/10.1137/ s0363012901385691

Konda, V.R., Tsitsiklis, J.N.: On actor-critic algorithms. SIAM Journal on Control and Optimization42(4), 1143–1166 (2003) https://doi.org/10.1137/ s0363012901385691

work page 2003
[44]

In: Proceedings of the 31st International Conference on Neural Information Processing Systems

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Proceedings of the 31st International Conference on Neural Information Processing Systems. NIPS’17, pp. 6000–6010. Curran Associates Inc., Red Hook, NY, USA (2017)

work page 2017
[45]

In: Proceedings of the 29th International Conference on Neural Information Processing Systems - Volume

Vinyals, O., Fortunato, M., Jaitly, N.: Pointer networks. In: Proceedings of the 29th International Conference on Neural Information Processing Systems - Volume

work page
[46]

2692–2700

NIPS’15, pp. 2692–2700. MIT Press, Cambridge, MA, USA (2015)

work page 2015
[47]

https://openreview.net/forum?id=ByxBFsRqYm

Kool, W., Hoof, H., Welling, M.: Attention, learn to solve routing prob- lems! In: International Conference on Learning Representations (2019). https://openreview.net/forum?id=ByxBFsRqYm

work page 2019
[48]

In: ICLR (Workshop) (2017)

Bello, I., Pham, H., Le, Q.V., Norouzi, M., Bengio, S.: Neural combinato- rial optimization with reinforcement learning. In: ICLR (Workshop) (2017). https://openreview.net/forum?id=Bk9mxlSFx

work page 2017
[49]

High-Dimensional Continuous Control Using Generalized Advantage Estimation

Schulman, J., Moritz, P., Levine, S., Jordan, M., Abbeel, P.: High-dimensional 18 continuous control using generalized advantage estimation (2018). https://arxiv. org/abs/1506.02438 Appendix A Background on machine learning We give a concise overview of the machine learning concepts used in this paper, with emphasis on reinforcement learning. For in-depth...

work page internal anchor Pith review Pith/arXiv arXiv 2018
[50]

Initialization.Choose α(0) 2 ∈R k2 at random (e.g., from a standard normal distribution)

work page
[51]

Write T(α 1,α 2) =A 1 α1, whereA 1 ∈R Nout×k1 is obtained by contractingTwithα 2

Step 1: solve for α1.With α2 = α(ℓ) 2 fixed, the map α1 7→T (α1,α 2) is linear. Write T(α 1,α 2) =A 1 α1, whereA 1 ∈R Nout×k1 is obtained by contractingTwithα 2. The subproblem min α1, c ∥A1 α1 −cq∥ 2 is a homogeneous least squares problem in the augmented variablew= ( α1, c) ∈ Rk1+1: one seeks the unit-norm vector minimizing ∥[A1 | −q]w ∥2, whose solutio...

work page
[52]

4.Iterate.Repeat Steps 1 and 2 forℓ= 0,1,

Step 2: solve for α2.With α1 = α(ℓ+1) 1 fixed, write T (α1,α 2) = A2 α2 and solve the analogous problem for (α 2, c). 4.Iterate.Repeat Steps 1 and 2 forℓ= 0,1, . . . , L−1. 20 5.Evaluate.After the final iteration, compute S= 1− ⟨T(α 1,α 2),q⟩ 2 ∥T(α 1,α 2)∥2 ∥q∥2 . Since ALS may converge to a local minimum, we run R independent restarts with random initia...

work page
[53]

at level n + 1, only candidates L satisfying ∆ b2 = (n+d ′ 1d′ 2)−b 2(A) are considered. For cells not reached by the extension cascade, we use direct supersolvable construc- tions: two pencils sharing a common line yield a free arrangement with any prescribed admissible exponents. These are classical and provide one example per cell instantly. 23 Appendi...

work page