Fine, Regular & Star Triangulations

• Fine, Regular, and Star Triangulations (FRSTs) are defined on 4D reflexive polytopes by using all relevant lattice points, enforcing a convex lifting (regularity), and ensuring every maximal simplex contains the origin (star condition).
• They play a crucial role in toric constructions of Calabi–Yau threefold hypersurfaces by producing simplicial, projective ambient spaces with fixed Hodge numbers derived solely from the polytope data.
• FRSTs also facilitate enumeration, sampling, and algorithmic exploration of vast Calabi–Yau families, linking combinatorial geometry with practical applications in string compactification.

Fine, Regular, and Star Triangulations (FRSTs) are the triangulations that underlie Batyrev’s toric construction of Calabi–Yau threefold hypersurfaces from four-dimensional reflexive lattice polytopes. In the standard 4D setting, an FRST is a triangulation of the boundary of a reflexive polytope, or of its polar dual, such that every relevant lattice point is used, the triangulation is coherent in the sense of a convex lifting, and every maximal simplex contains the origin. Combinatorially, FRSTs encode maximal projective crepant partial subdivisions; geometrically, they produce simplicial, projective toric ambient spaces in which a generic anticanonical hypersurface is smooth. In contemporary work on the Kreuzer–Skarke landscape, FRSTs also serve as the basic objects for counting, sampling, classifying, and algorithmically exploring large families of Calabi–Yau threefolds (MacFadden et al., 18 Feb 2026).

1. Definitions in the reflexive-polytopal setting

Let $\Delta \subset N_{\mathbb{R}}$ be a four-dimensional reflexive lattice polytope, with $N \cong \mathbb{Z}^4$, and let its polar dual be

$$\Delta^\vee=\{y\in M_{\mathbb{R}}:\langle y,x\rangle\ge -1 \text{ for all } x\in \Delta\},$$

where $M=\mathrm{Hom}(N,\mathbb{Z})$. Reflexivity means that both $\Delta$ and $\Delta^\vee$ are lattice polytopes and that the origin is the unique interior lattice point of each. In the toric Calabi–Yau literature one may work on $\Delta$ or on $\Delta^\vee$; both viewpoints are standard, and the resulting fan data are related by polar duality (MacFadden et al., 18 Feb 2026).

The three defining conditions are structurally distinct. Fine means that every lattice point in the chosen point configuration appears as a vertex of the triangulation. In the 4D hypersurface setting, several papers restrict the point configuration to the boundary lattice points excluding points interior to facets, because these points are irrelevant to the generic anticanonical hypersurface in dimension four; accordingly, “fine” is often interpreted relative to this resolved set rather than literally all lattice points of the polytope. Regular means coherent: there exists a height function or convex piecewise-linear support function such that the triangulation is obtained by projecting the lower faces of the lifted convex hull. Star means that every maximal simplex contains the origin, so that coning over the simplices yields a complete simplicial fan compatible with the reflexive structure (MacFadden et al., 16 Dec 2025).

These conventions are not merely terminological. Fine controls which rays are present in the toric refinement; regularity is the projectivity condition; and star is the crepant, origin-centered condition relevant for the anticanonical hypersurface. A common source of confusion is that “fine” is not always synonymous with unimodular, and FRST validity does not require the ambient toric fourfold to be smooth. In dimension four, it is enough that the triangulation defines an MPCP refinement for the generic hypersurface to be smooth (Crinò et al., 2022).

2. Toric interpretation and Calabi–Yau hypersurfaces

Given a 4D reflexive pair $(\Delta,\Delta^\vee)$, one associates a toric Fano variety to the normal fan of $\Delta^\vee$, and a Calabi–Yau threefold $N \cong \mathbb{Z}^4$0 as the closure of a generic anticanonical hypersurface. To obtain a smooth Calabi–Yau hypersurface, one passes to a maximal projective crepant partial subdivision of the fan; combinatorially, this is achieved by an FRST of $N \cong \mathbb{Z}^4$1 or $N \cong \mathbb{Z}^4$2. In this way, FRSTs provide a combinatorial parametrization of smooth Calabi–Yau threefold hypersurfaces in the toric hypersurface class (MacFadden et al., 18 Feb 2026).

The anticanonical hypersurface is described by

$N \cong \mathbb{Z}^4$3

where the global sections correspond to lattice points of $N \cong \mathbb{Z}^4$4. Regularity implies projectivity of the ambient toric variety; star ensures that the triangulation defines a genuine fan refinement; and fine gives the maximality needed for an MPCP desingularization. In dimension $N \cong \mathbb{Z}^4$5, such a refinement suffices to ensure that a generic anticanonical hypersurface is smooth, because the remaining ambient singularities do not intersect the hypersurface in codimension one (MacFadden et al., 16 Dec 2025).

The Hodge numbers of the resulting Calabi–Yau threefold depend only on the reflexive polytope pair and not on the specific FRST. Batyrev’s formulas take the form

$N \cong \mathbb{Z}^4$6

and

$N \cong \mathbb{Z}^4$7

Here $N \cong \mathbb{Z}^4$8 counts lattice points, $N \cong \mathbb{Z}^4$9 counts interior lattice points in a face, and $$\Delta^\vee=\{y\in M_{\mathbb{R}}:\langle y,x\rangle\ge -1 \text{ for all } x\in \Delta\},$$0 denotes faces of codimension $$\Delta^\vee=\{y\in M_{\mathbb{R}}:\langle y,x\rangle\ge -1 \text{ for all } x\in \Delta\},$$1. These formulas explain why FRSTs are central for topology-changing but Hodge-number-preserving birational models: the Hodge numbers are fixed by the polytope data, whereas intersection numbers, second Chern class data, Stanley–Reisner ideals, and Kähler-theoretic features depend on the triangulation (Crinò et al., 2022).

3. Regularity, secondary fans, circuits, and flips

Regularity is the most algorithmically delicate of the three FRST conditions. The basic criterion is the lifting construction: a triangulation $$\Delta^\vee=\{y\in M_{\mathbb{R}}:\langle y,x\rangle\ge -1 \text{ for all } x\in \Delta\},$$2 is regular if there exists a weight vector $$\Delta^\vee=\{y\in M_{\mathbb{R}}:\langle y,x\rangle\ge -1 \text{ for all } x\in \Delta\},$$3 such that $$\Delta^\vee=\{y\in M_{\mathbb{R}}:\langle y,x\rangle\ge -1 \text{ for all } x\in \Delta\},$$4 is the projection of the lower convex hull of the lifted points $$\Delta^\vee=\{y\in M_{\mathbb{R}}:\langle y,x\rangle\ge -1 \text{ for all } x\in \Delta\},$$5. Equivalently, regular triangulations are chambers of the secondary fan, or vertices of the secondary polytope. In practice, regularity is often checked by linear feasibility in height space, by circuit inequalities, or by lower-hull computations (Kastner, 2024).

Circuits organize the local combinatorics of regular triangulations. A circuit is a minimally affinely dependent set, with affine dependence

$$\Delta^\vee=\{y\in M_{\mathbb{R}}:\langle y,x\rangle\ge -1 \text{ for all } x\in \Delta\},$$6

The sign pattern of the coefficients determines the two local triangulations on the circuit hull, and bistellar flips correspond to replacing one local triangulation by the other. In the secondary-fan language, the walls between coherent chambers are defined by the hyperplanes $$\Delta^\vee=\{y\in M_{\mathbb{R}}:\langle y,x\rangle\ge -1 \text{ for all } x\in \Delta\},$$7. This makes circuits the natural primitives for both flip-based enumeration and learned local search (Jordan et al., 2017).

Several modern algorithms exploit this structure. Down-flip reverse search in mptopcom enumerates regular triangulations by traversing the regular flip graph while filtering by coherence-preserving moves; the later “regular flips” refinement reduces the number and size of linear programs by testing whether a flip direction is an extremal ray of the local edge cone. TriSearch represents actions by circuit-supported local subtriangulations and uses certified bistellar flips to traverse the flip graph in 3D and 4D. The dualGNN sampler labels dual-graph edges by signed circuits and uses the fact that these circuit inequalities are necessary and sufficient for regularity. Across these approaches, regularity is not inferred heuristically: it is encoded by explicit geometric or oriented-matroid criteria (Wang et al., 28 May 2026).

This secondary-fan viewpoint also interfaces directly with FRST generation from 2-face data. In the NTFE framework, one fixes fine regular triangulations on all 2-faces and then searches for a global height vector in the strict interior of the intersection of the corresponding secondary cones. If such a vector exists, it induces a global regular triangulation whose 2-face restrictions are exactly the prescribed ones; the star condition can then be enforced by lowering the origin’s height sufficiently, without changing the 2-face triangulations (MacFadden, 2023).

4. Two-face restrictions, topological equivalence, and bounds

A major conceptual development is that FRSTs are often too fine a combinatorial invariant for classifying the resulting Calabi–Yau threefolds. For simply connected Calabi–Yau threefolds with torsion-free cohomology, Wall’s theorem implies that the diffeomorphism class is determined by the Hodge numbers, the second Chern class, and the triple intersection form. In the toric hypersurface setting, the Hodge numbers are determined purely by the reflexive polytope, while the second Chern class and the triple intersection numbers depend only on the restrictions of the triangulation to the 2-dimensional faces. This leads to the notion of 2-face equivalence: two FRSTs are 2-face equivalent if they induce the same triangulation on every 2-face, and any two 2-face-equivalent FRSTs give rise to diffeomorphic Calabi–Yau hypersurfaces (MacFadden et al., 18 Feb 2026).

This result has two immediate consequences. First, it replaces the problem of counting FRSTs by the smaller problem of counting 2-face equivalence classes. Second, it introduces a sharp distinction between upper and lower bounds. Distinct FRSTs can collapse to the same Calabi–Yau diffeomorphism class, and even distinct 2-face equivalence classes can sometimes yield diffeomorphic threefolds. Accordingly, counting 2-face classes gives an upper bound on diffeomorphism classes and a lower bound only on computationally distinct representatives, not on topology classes themselves (MacFadden, 2023).

The strongest quantitative statement to date is that there are at most $$\Delta^\vee=\{y\in M_{\mathbb{R}}:\langle y,x\rangle\ge -1 \text{ for all } x\in \Delta\},$$8 diffeomorphism classes of Calabi–Yau threefolds produced via FRSTs over the Kreuzer–Skarke database, improving the previous upper bound $$\Delta^\vee=\{y\in M_{\mathbb{R}}:\langle y,x\rangle\ge -1 \text{ for all } x\in \Delta\},$$9. The same work proves that there are at least $M=\mathrm{Hom}(N,\mathbb{Z})$0 distinct 2-face equivalence classes for the unique dual polytope with $M=\mathrm{Hom}(N,\mathbb{Z})$1, while emphasizing that this is not a lower bound on diffeomorphism classes. The upper bound is overwhelmingly dominated by that unique polytope, and the contribution of the next 400 polytopes with $M=\mathrm{Hom}(N,\mathbb{Z})$2 is only $M=\mathrm{Hom}(N,\mathbb{Z})$3, whereas all polytopes with $M=\mathrm{Hom}(N,\mathbb{Z})$4 contribute less than $M=\mathrm{Hom}(N,\mathbb{Z})$5 (MacFadden et al., 18 Feb 2026).

A related misconception is that all assignments of fine regular triangulations to the 2-faces automatically extend to an FRST. They do not. Extendability can fail, which is why lower bounds require explicit constructions of compatible families of 2-face triangulations. For the dominant $M=\mathrm{Hom}(N,\mathbb{Z})$6 polytope, a large extendable family is obtained via primary subdivisions and stripwise lifting arguments, yielding

$M=\mathrm{Hom}(N,\mathbb{Z})$7

This shows that the global upper bound is at most twenty orders of magnitude loose in the dominant region of the landscape (MacFadden et al., 18 Feb 2026).

5. Enumeration, sampling, and learned generation

The combinatorics of FRSTs are immense. The Kreuzer–Skarke database contains $M=\mathrm{Hom}(N,\mathbb{Z})$8 4D reflexive polytopes, and early estimates based on direct FRST counting produced a naive upper bound of $M=\mathrm{Hom}(N,\mathbb{Z})$9 on the number of Calabi–Yau threefold hypersurfaces arising this way. Later work reduced the problem to NTFE classes, and more recent work has separated the combinatorics of FRSTs, fine regular triangulations, and 2-face equivalence classes into distinct enumeration problems with different asymptotic scales (MacFadden et al., 18 Feb 2026).

Exact and exhaustive enumeration is therefore confined to low-complexity regions. For the Kreuzer–Skarke database with $\Delta$0, an exhaustive study found $\Delta$1 fine regular triangulations, of which $\Delta$2 are vex triangulations; over $\Delta$3 of all fine regular triangulations in that regime are vex rather than star. The same work gives an upper bound of $\Delta$4 for fine regular triangulations in the entire Kreuzer–Skarke database, while stressing that this vastly overcounts distinct Calabi–Yau diffeomorphism classes (MacFadden et al., 16 Dec 2025).

A parallel line of work seeks representative generation rather than exhaustive listing. Reinforcement-learning methods encode 2-face triangulations as states and use intersections of secondary cones to detect global regularity; in low-$\Delta$5 regimes these methods reproduced the complete set of non-2-face-equivalent FRSTs for $\Delta$6 and required fewer steps than a random walk. TriSearch instead navigates the flip graph directly by certified bistellar flips and, in 4D, discovers more distinct FRSTs of reflexive polytopes than existing samplers under a fixed budget. In the largest tested range for nearby-FRST search, it maintained a perfect success rate across unseen polytopes with $\Delta$7 within a 50-flip budget per start (Berglund et al., 2024).

Autoregressive and transformer-based models address a complementary task: sampling regular face triangulations or full FRSTs without explicit enumeration of the entire search space. dualGNN samples fine regular triangulations of 2-faces using circuit-labeled dual graphs and, when combined with the NTFE extension algorithm, gives uniform samples of Calabi–Yau threefolds modulo 2-face equivalence at $\Delta$8 and results consistent with uniformity at $\Delta$9. CYTransformer and related transformer architectures generate FRSTs directly across polytope sizes, use post hoc validation in CYTools, and can self-improve through retraining on validated outputs; one study reports up to approximately $\Delta^\vee$0 unique FRSTs out of $\Delta^\vee$1 candidates during training monitoring, while another reports up to $\Delta^\vee$2 higher average recovery of distinct FRSTs on the test set after iterative retraining from small seed data (MacFadden, 26 May 2026).

6. Variants, applications, and limitations

FRSTs are not the whole story of toric birational geometry. A decisive recent generalization is the notion of vex triangulations: fine, regular, simplicial, Gorenstein triangulations that are not star and typically yield non-weak-Fano toric varieties. All fine regular triangulations of a fixed 4D reflexive polytope, including both FRSTs and vex triangulations, give rise to smooth birational Calabi–Yau hypersurfaces. This shows that FRSTs do not exhaust the toric descriptions of a birational class or the torically visible chambers of the extended Kähler cone; instead, the union of nef cones over all fine regular fans gives the toric moving cone, an inner approximation to the extended Kähler cone (MacFadden et al., 16 Dec 2025).

There are also structural constructions beyond brute-force flipping or facewise assembly. For free sums of point configurations, webs of stars and the stabbing poset classify triangulations of the sum in terms of triangulations of the summands. This framework is not formulated specifically for reflexive 4-polytopes, but it provides a combinatorial mechanism for building triangulations from lower-dimensional data and conjecturally characterizes regularity of such sum-triangulations by total ordering of the web images under inclusion. A plausible implication is that decomposition structure can substantially reduce the effective search space for certain FRST problems (Assarf et al., 2015).

In string-compactification applications, FRSTs are the combinatorial input for large toric databases. One orientifold study constructed Calabi–Yau orientifolds from holomorphic reflection involutions for all favourable FRSTs with $\Delta^\vee$3 and random samples up to $\Delta^\vee$4, reporting $\Delta^\vee$5 favourable FRSTs, $\Delta^\vee$6 reflection involutions, and $\Delta^\vee$7 “smooth” orientifolds in the exhaustive low-$\Delta^\vee$8 regime. In that dataset, the largest D3-tadpole before worldvolume fluxes occurs for a Calabi–Yau threefold with $\Delta^\vee$9, with $\Delta$0 for the local D7 case and $\Delta$1 for the non-local Whitney brane case (Crinò et al., 2022).

Several limitations are intrinsic. Not every compatible-looking family of face triangulations extends to an FRST. Regularity is global and can be broken by a single flip. Distinct 2-face equivalence classes can still collapse to the same diffeomorphism class. Learned samplers can improve coverage or uniformity but do not provide completeness guarantees. Finally, FRST-based analyses probe only the star sector of fine regular triangulations; the broader birational geometry of toric Calabi–Yau hypersurfaces requires vex triangulations and related non-weak-Fano phases as well (MacFadden et al., 18 Feb 2026).