Toric NCCRs: Non-Commutative Crepant Resolutions

• Toric NCCRs are non-commutative crepant resolutions of affine toric singularities built from direct sums of rank-one reflexive modules or modules of covariants.
• They employ invariant-theoretic frameworks and tilting bundles on quotient stacks to yield algebras with finite global dimension and Cohen–Macaulay properties.
• The distinction between toric and general NCCRs highlights scenarios where toric constructions fail, necessitating non-toric modifications for achieving crepant resolutions.

Toric non-commutative crepant resolutions are non-commutative resolutions attached to affine toric singularities, typically rings of the form $R=k[\sigma^\vee\cap M]$ or, in invariant-theoretic form, $R=\operatorname{Sym}(W)^G$ for an abelian reductive group $G$. In the standard Van den Bergh sense, an NCCR of a normal Gorenstein domain $R$ is an algebra $\Lambda=\operatorname{End}_R(M)$ with $M$ nonzero, finitely generated, and reflexive, such that $\Lambda$ has finite global dimension and is Cohen–Macaulay as an $R$-module. In toric geometry, one often singles out the narrower notion of a toric NCCR, namely an NCCR obtained from a direct sum of rank-one reflexive modules, or equivalently from modules of covariants when the toric singularity is presented as a quotient. A central theme of the subject is that these two notions do not coincide: a toric singularity may fail to admit any toric NCCR while still admitting an NCCR of a more general, non-toric form (Špenko et al., 2017).

1. Foundational framework

The invariant-theoretic realization used throughout much of the modern literature starts with an abelian reductive group $G$ acting linearly on a finite-dimensional vector space $W$, with

$R=\operatorname{Sym}(W)^G$0

For a normal Gorenstein domain $R=\operatorname{Sym}(W)^G$1, the standard NCCR definition is

$R=\operatorname{Sym}(W)^G$2

where $R=\operatorname{Sym}(W)^G$3 is reflexive, $R=\operatorname{Sym}(W)^G$4 has finite global dimension, and $R=\operatorname{Sym}(W)^G$5 is Cohen–Macaulay over $R=\operatorname{Sym}(W)^G$6. In this setup, $R=\operatorname{Sym}(W)^G$7 is called generic if $R=\operatorname{Sym}(W)^G$8 acts generically on $R=\operatorname{Sym}(W)^G$9, meaning

$G$0

and unimodular if

$G$1

so that the determinant character is trivial; for toric invariant rings, this unimodularity is equivalent to the quotient being Gorenstein (Špenko et al., 2017).

For torus quotients, the basic toric modules are modules of covariants

$G$2

and for a character $G$3 one writes $G$4. A toric NCCR is then an algebra of the form

$G$5

with $G$6 a finite direct sum of characters. Because $G$7 is a torus, such modules are finite direct sums of rank-one pieces $G$8; this is the sense in which toric NCCRs are built from toric rank-one reflexive modules (Špenko et al., 2018).

A second foundational distinction is between general NCCRs and special subclasses such as splitting and steady NCCRs. A reflexive module $G$9 is splitting if it is a direct sum of rank-one reflexive modules, and steady if

$R$0

For splitting modules, steadiness is controlled by the divisor class group: if

$R$1

with $R$2 finite, then

$R$3

(Iyama et al., 2015).

2. Existence via GIT stacks and unstable loci

A decisive existence criterion for NCCRs of toric singularities was given by Špenko–Van den Bergh in the quotient-stack setting. For a character $R$4, let $R$5 be the semistable locus and consider the projective morphism

$R$6

If every point of $R$7 has finite stabilizer and the fibers of $R$8 have dimension at most $R$9, then $\Lambda=\operatorname{End}_R(M)$0 carries a tilting bundle $\Lambda=\operatorname{End}_R(M)$1, and its global sections

$\Lambda=\operatorname{End}_R(M)$2

yield

$\Lambda=\operatorname{End}_R(M)$3

When $\Lambda=\operatorname{End}_R(M)$4 is generic and unimodular, this endomorphism algebra is an NCCR of $\Lambda=\operatorname{End}_R(M)$5 (Špenko et al., 2017).

The geometric input is converted into a concrete toric criterion through the unstable locus

$\Lambda=\operatorname{End}_R(M)$6

If

$\Lambda=\operatorname{End}_R(M)$7

then the fibers of $\Lambda=\operatorname{End}_R(M)$8 have dimension at most $\Lambda=\operatorname{End}_R(M)$9, hence the NCCR exists. This criterion is especially effective because it reduces a homological question to a low-dimensionality condition on the nullcone (Špenko et al., 2017).

In dimension three, the criterion recovers Broomhead’s theorem: if $M$0 is generic and

$M$1

then the unstable locus has codimension at least $M$2, so

$M$3

and every affine Gorenstein toric singularity of dimension $M$4 has an NCCR. This proof is shorter than the original dimer-model argument and emphasizes the role of GIT and tilting on quotient stacks rather than quivers with potential (Špenko et al., 2017).

The same paper also shows where the toric and non-toric notions separate. In the four-dimensional example with

$M$5

and weights

$M$6

the invariant ring is

$M$7

a $M$8-dimensional toric Gorenstein singularity. Its unstable locus satisfies

$M$9

so the criterion produces an NCCR; however, earlier work had shown that this singularity admits no toric NCCR. The resulting NCCR is obtained only after adjoining a non-toric reflexive summand $\Lambda$0, defined by an exact sequence rather than by a direct sum of rank-one toric modules (Špenko et al., 2017).

3. Three-dimensional toric NCCRs: dimers, stacks, and derived equivalence

The three-dimensional theory was originally organized around consistent dimer models. For a consistent dimer on the two-torus, the Jacobian algebra

$\Lambda$1

has center $\Lambda$2 equal to a Gorenstein semigroup algebra, so

$\Lambda$3

is a $\Lambda$4-dimensional affine toric Gorenstein singularity, and $\Lambda$5 is an NCCR of $\Lambda$6 (Bocklandt et al., 2013). The derived equivalence between $\Lambda$7 and a crepant resolution $\Lambda$8 identifies the simple $\Lambda$9-modules with geometric objects on $R$0: for nonzero vertices, the images are pure sheaves, while the zero vertex maps to the dualizing complex of the compact exceptional fiber (Bocklandt et al., 2013).

Špenko–Van den Bergh later gave an alternative proof of Broomhead’s theorem using standard toric geometry rather than dimers. For a triangulation $R$1 of the lattice polygon $R$2 defining the threefold singularity, the associated smooth toric Deligne–Mumford stack $R$3 carries a split tilting bundle

$R$4

whose endomorphism algebra is

$R$5

This produces a toric NCCR because the summands are toric rank-one reflexive modules. The proof is specific to $R$6-dimensional Gorenstein affine toric varieties, but it shows that toric NCCRs can be read directly from line bundles on stacky crepant resolutions (Špenko et al., 2017).

A further structural result concerns uniqueness up to derived equivalence. If $R$7 is a generic, weakly symmetric, unimodular torus representation, then all toric NCCRs of

$R$8

are derived equivalent. More precisely, each toric NCCR is derived equivalent to the same Deligne–Mumford GIT quotient stack

$R$9

for a generic character $G$0, so any two toric NCCRs have equivalent derived categories (Špenko et al., 2018). This statement is explicitly restricted to toric NCCRs; it does not govern non-toric NCCRs of the same singularity.

Within the dimer world, additional module-theoretic refinements appear. A consistent dimer gives a splitting NCCR of the form

$G$1

where the $G$2 are rank-one reflexive modules read off from perfect matchings. Among these, steady NCCRs correspond to regular hexagonal dimers, while semi-steady but non-steady NCCRs correspond to square dimers; consequently, for isoradial dimers, semi-steadiness characterizes regular dimers (Nakajima, 2016).

4. Toric singularities versus toric NCCRs

One of the clearest lessons of the subject is that “toric singularity” and “toric NCCR” are not equivalent notions. In the sense used in the quotient and dimer literature, a toric NCCR is one obtained from a direct sum of rank-one reflexive modules, equivalently from modules of covariants when the acting group is abelian. The four-dimensional hypersurface

$G$3

shows that this class is strictly narrower than the class of all NCCRs of toric singularities: it has no toric NCCR, yet it does have an NCCR constructed from a nontrivial reflexive summand $G$4 fitting into

$G$5

(Špenko et al., 2017).

This distinction has categorical consequences. The derived-equivalence theorem for toric NCCRs applies only to NCCRs built from modules of covariants under the hypotheses of genericity, weak symmetry, and unimodularity. It does not imply that every NCCR of the singularity is toric, nor that every NCCR is derived equivalent to the toric ones. The four-dimensional example therefore marks a genuine boundary of current uniform results: toric NCCRs can fail to exist even when ordinary NCCRs do exist (Špenko et al., 2018).

A parallel classification phenomenon appears in the rigid subclass of steady splitting NCCRs. For a complete local CM normal domain over an algebraically closed field of characteristic $G$6, the following are equivalent: $G$7 is a quotient singularity by a finite abelian group, $G$8 has a steady splitting NCCR, and $G$9 is finite with

$W$0

giving an NCCR. In the toric case, this says that completed affine toric singularities admit steady splitting NCCRs exactly in the simplicial or finite-abelian-quotient case (Iyama et al., 2015). This result sharply separates a very rigid toric subclass from the broader NCCR landscape.

A plausible implication is that higher-dimensional toric geometry should not be organized solely around toric NCCRs. The existing examples and classification results consistently show that toric rank-one constructions are powerful but not exhaustive.

5. Conic modules, Hibi rings, and rank-one class groups

Another major line of work constructs non-commutative resolutions from conic modules. For an affine toric algebra

$W$1

a conic module is

$W$2

If

$W$3

is the direct sum over all isomorphism classes of conic modules, then

$W$4

always has

$W$5

Moreover, $W$6 is an NCCR if and only if the toric variety is simplicial (Faber et al., 2018). Thus complete sums of conic modules always yield NCRs, but they are crepant precisely in the simplicial case.

Recent work refines this by studying incomplete sums of conic modules. If

$W$7

then $W$8 is an NCR exactly when the chosen set is lockable, and an NCCR exactly when it is incredulous. The same paper links conic modules to the Bondal–Thomsen collection of line bundles on smooth toric DM stacks, reduces the existence problem to a torsion-free class-group case, and classifies when almost simplicial Gorenstein cones admit NCCRs via endomorphism algebras of conic modules (Malter, 25 Mar 2026).

Hibi rings provide a particularly explicit toric laboratory for this circle of ideas. For a Hibi ring $W$9, the divisor class group can be described from a spanning tree in the Hasse diagram of the augmented poset $R=\operatorname{Sym}(W)^G$00, and the conic divisorial ideals are exactly the lattice points of an explicitly defined polytope $R=\operatorname{Sym}(W)^G$01 cut out by inequalities indexed by circuits in that Hasse diagram. In the special case of the Segre product of $R=\operatorname{Sym}(W)^G$02-variable polynomial rings, the chosen subset

$R=\operatorname{Sym}(W)^G$03

inside the conic region yields a splitting NCCR

$R=\operatorname{Sym}(W)^G$04

(Higashitani et al., 2017).

The rank-one divisor-class-group case admits an especially sharp classification. For Gorenstein toric singularities with

$R=\operatorname{Sym}(W)^G$05

toric NCCRs exist, are classified by non-trivial upper sets in a certain quotient of $R=\operatorname{Sym}(W)^G$06 equipped with a partial order, and all toric NCCRs are connected by iterated Iyama–Wemyss mutations (Tomonaga, 30 Oct 2025). This places a substantial higher-dimensional toric class under explicit combinatorial control.

6. Stacky constructions, face descent, and current directions

A strong recent trend is to construct NCCRs from tilting or partial tilting objects on toric Deligne–Mumford stacks. For an affine toric Gorenstein singularity

$R=\operatorname{Sym}(W)^G$07

with

$R=\operatorname{Sym}(W)^G$08

one chooses a simplicial toric refinement $R=\operatorname{Sym}(W)^G$09, passes to the smooth toric DM stack $R=\operatorname{Sym}(W)^G$10, and studies

$R=\operatorname{Sym}(W)^G$11

If $R=\operatorname{Sym}(W)^G$12 is a partial tilting complex on $R=\operatorname{Sym}(W)^G$13 and $R=\operatorname{Sym}(W)^G$14 has finite global dimension, then $R=\operatorname{Sym}(W)^G$15 is an NCCR of $R=\operatorname{Sym}(W)^G$16. This framework yields concrete existence results for cones arising from canonical bundles, weighted and fake weighted projective spaces, and especially simplicial reflexive polytopes with at most $R=\operatorname{Sym}(W)^G$17 vertices (Malter et al., 15 Sep 2025).

A complementary structural theorem shows that toric NCCRs descend to faces. If $R=\operatorname{Sym}(W)^G$18 is lattice equivalent to a face of a polytope $R=\operatorname{Sym}(W)^G$19, and the toric algebra attached to

$R=\operatorname{Sym}(W)^G$20

has a toric NCCR, then so does the toric algebra attached to

$R=\operatorname{Sym}(W)^G$21

This yields short proofs of the existence of toric NCCRs for simplicial affine toric Gorenstein algebras and for almost simplicial ones, meaning cones with

$R=\operatorname{Sym}(W)^G$22

(Malter et al., 25 Feb 2026).

Positive-characteristic results now fit into the same broad picture. If $R=\operatorname{Sym}(W)^G$23 is a normal $R=\operatorname{Sym}(W)^G$24-dimensional toric and $R=\operatorname{Sym}(W)^G$25-factorial singularity over an algebraically closed field of characteristic $R=\operatorname{Sym}(W)^G$26, then

$R=\operatorname{Sym}(W)^G$27

for a finite diagonalizable subgroup scheme $R=\operatorname{Sym}(W)^G$28, and both

$R=\operatorname{Sym}(W)^G$29

for $R=\operatorname{Sym}(W)^G$30 are NCCRs. In dimension $R=\operatorname{Sym}(W)^G$31, the associated $R=\operatorname{Sym}(W)^G$32-blowups recover the minimal resolution, and in dimension $R=\operatorname{Sym}(W)^G$33, under Gorensteinness, they recover a crepant resolution (Liedtke et al., 2023).

The present state of the subject is therefore stratified. Three-dimensional Gorenstein toric singularities admit toric NCCRs uniformly. In higher dimension, existence is established for several substantial families—simplicial, almost simplicial, rank-one class group, and specific reflexive or canonical-bundle constructions—but toric NCCRs are not universal, and general NCCRs may exist beyond the toric rank-one regime. The literature increasingly treats toric stacks, derived categories, mutation, and class-group combinatorics as complementary rather than competing languages for organizing this distinction (Špenko et al., 2017).