Graver Basis: Integer Optimization and Toric Ideals

• Graver basis is the finite set of conformally minimal nonzero integer kernel elements that serve as primitive augmentation directions in integer optimization.
• It underpins efficient algorithms in combinatorial optimization by decomposing lattice elements and certifying optimality through structured Graver moves.
• Applications span toric algebra, algebraic statistics, and GPU-based heuristics, although exact computation remains computationally challenging.

Searching arXiv for recent and foundational papers on Graver bases and related applications. A Graver basis is the finite set of conformally minimal nonzero integer kernel elements of an integer matrix, and it functions as a canonical collection of primitive augmentation directions for integer optimization and as a distinguished binomial basis in toric algebra. For a matrix $A \in \mathbb{Z}^{m\times n}$, it is defined through the lattice $\ker_{\mathbb{Z}}(A)$ and the conformal partial order that compares vectors within a common orthant coordinatewise by absolute value. Across integer programming, combinatorial optimization, toric ideals, algebraic statistics, and structured nonlinear optimization, the Graver basis serves both as a decomposition system for lattice elements and as an optimality certificate: if an objective can still be improved, then some Graver direction improves it (1207.1149).

1. Definition and core structure

Let $A \in \mathbb{Z}^{m\times n}$. The relevant lattice is the set of integer dependencies

$L(A)=\{z\in \mathbb{Z}^n\setminus\{0\}:Az=0\}.$

For $u,v\in \mathbb{Z}^n$, the conformal partial order is

$$u \sqsubseteq v \iff u_i v_i \ge 0 \text{ and } |u_i|\le |v_i| \text{ for every coordinate } i.$$

Thus $u$ and $v$ must lie in the same orthant, and $u$ must be coordinatewise dominated by $v$ in absolute value. The Graver basis is the set of conformally minimal nonzero lattice elements,

$\ker_{\mathbb{Z}}(A)$0

This finiteness holds for every integer matrix $\ker_{\mathbb{Z}}(A)$1 (1207.1149).

Equivalent formulations recur across the literature. In toric-ideal language, Graver elements are the primitive binomials associated with $\ker_{\mathbb{Z}}(A)$2, and they contain all circuits. In several sources, a nonzero lattice element is Graver precisely when it is conformally indecomposable, meaning it cannot be written as a nontrivial conformal sum of two nonzero kernel elements (Tatakis et al., 2013). For pointed cones, the Graver basis can also be described as the union of orthant-wise Hilbert bases, which is the formulation used in GPU-based extraction heuristics (Liu et al., 2024).

Three structural properties are central. First, every nonzero lattice element admits a conformal decomposition into Graver elements. Second, the Graver basis is a test set: if a feasible point is not optimal, then some Graver move yields improvement. Third, circuits form a subset of the Graver basis, but the inclusion can be strict and, in general, circuits do not control Graver complexity polynomially (Alghassi et al., 2019).

2. Fundamental optimization role

The Graver basis provides the combinatorial backbone for augmentation algorithms in integer optimization. For problems of the form

$\ker_{\mathbb{Z}}(A)$3

if $\ker_{\mathbb{Z}}(A)$4 is linear or separable convex, then non-optimality implies the existence of a feasible improving step along some $\ker_{\mathbb{Z}}(A)$5, possibly with an integer step length $\ker_{\mathbb{Z}}(A)$6 chosen by one-dimensional optimization along the ray $\ker_{\mathbb{Z}}(A)$7 (1207.1149). This gives a finite optimality certificate and yields iterative augmentation methods that terminate at an optimum.

For separable convex objectives, the univariate objective profile

$\ker_{\mathbb{Z}}(A)$8

is convex over the feasible integer interval of step sizes, so the best step along a Graver direction can be located by discrete convex line search. In the structured setting of $\ker_{\mathbb{Z}}(A)$9-fold 4-block decomposable integer programs, this mechanism combines with proximity results and dynamic programming to obtain polynomial-time algorithms when the blocks are fixed and $A \in \mathbb{Z}^{m\times n}$0 varies (1207.1149).

The same augmentation principle extends beyond separable convexity. One survey of algorithmic uses lists separable convex minimization, convex integer maximization, convex norm minimization, quadratic minimization in the dual of the quadratic Graver cone, and polynomial minimization in the dual of higher-degree Graver cones as classes where Graver augmentation is exact (Alghassi et al., 2019). A more specialized treatment of quadratic and higher-degree forms defines the quadratic Graver cone and its dual, and proves polynomial-time solvability when the quadratic form lies in the dual cone and the Graver basis is given (Lee et al., 2010).

A recurrent theme is that good norm bounds on Graver elements drive algorithm design. In block-structured integer programming, augmentation complexity depends critically on bounds on $A \in \mathbb{Z}^{m\times n}$1 or $A \in \mathbb{Z}^{m\times n}$2 for $A \in \mathbb{Z}^{m\times n}$3. For 4-block $A \in \mathbb{Z}^{m\times n}$4-fold matrices, improved bounds of order $A \in \mathbb{Z}^{m\times n}$5 on the $A \in \mathbb{Z}^{m\times n}$6-norm of Graver elements sharpened augmentation-based algorithms and matched new lower bounds up to fixed-parameter factors (Chen et al., 2018).

3. Block structure, proximity, and parameterized complexity

One of the most developed areas of Graver-basis theory concerns block-structured matrices. For fixed matrices $A \in \mathbb{Z}^{m\times n}$7, the $A \in \mathbb{Z}^{m\times n}$8-fold 4-block decomposable matrix

$A \in \mathbb{Z}^{m\times n}$9

simultaneously generalizes classical $L(A)=\{z\in \mathbb{Z}^n\setminus\{0\}:Az=0\}.$0-fold integer programs and two-stage stochastic integer programs. In this setting, structural bounds on the norms of Graver elements, together with a proximity theorem adapted from Hochbaum–Shanthikumar, yield polynomial-time algorithms for separable convex minimization when the blocks are fixed and $L(A)=\{z\in \mathbb{Z}^n\setminus\{0\}:Az=0\}.$1 varies (1207.1149).

The proximity statement is especially important. If $L(A)=\{z\in \mathbb{Z}^n\setminus\{0\}:Az=0\}.$2 is an optimal solution to the continuous relaxation, then there exists an optimal integer solution $L(A)=\{z\in \mathbb{Z}^n\setminus\{0\}:Az=0\}.$3 such that

$L(A)=\{z\in \mathbb{Z}^n\setminus\{0\}:Az=0\}.$4

with $L(A)=\{z\in \mathbb{Z}^n\setminus\{0\}:Az=0\}.$5 the constraint matrix. This restricts the search for an optimal integer point to a bounded neighborhood of a continuous optimum, after which Graver-best augmentation solves the reduced instance (1207.1149).

Later work improved the quantitative side of this picture. For 4-block $L(A)=\{z\in \mathbb{Z}^n\setminus\{0\}:Az=0\}.$6-fold matrices, the $L(A)=\{z\in \mathbb{Z}^n\setminus\{0\}:Az=0\}.$7-norm of Graver elements was improved from a previously known $L(A)=\{z\in \mathbb{Z}^n\setminus\{0\}:Az=0\}.$8 bound to $L(A)=\{z\in \mathbb{Z}^n\setminus\{0\}:Az=0\}.$9, and a matching lower bound of $u,v\in \mathbb{Z}^n$0 was established even for arbitrary nonzero lattice elements (Chen et al., 2018). This showed that augmentation frameworks relying on substantially smaller-norm moves cannot succeed in general for that class.

Another direction asks when Graver norms are bounded because the matrix is sparse only after row operations. A structural characterization in terms of matroid depth parameters proves that the $u,v\in \mathbb{Z}^n$1-norm of Graver elements is bounded by a function of the maximum $u,v\in \mathbb{Z}^n$2-norm of a circuit of $u,v\in \mathbb{Z}^n$3, and yields parameterized algorithms based on row-equivalent matrices of bounded primal or dual tree-depth and bounded entry complexity (Brianski et al., 2022). This suggests that, in some settings, sparse row-equivalent structure rather than the original matrix sparsity governs Graver complexity.

The mixed-integer setting behaves differently. For mixed kernels, the paper on separable convex mixed-integer hardness introduces the mixed Graver basis and obtains the first nontrivial norm bounds, including double-exponential $u,v\in \mathbb{Z}^n$4 bounds for 2-stage matrices, but also proves that these bounds do not lead to the same kind of tractability as in the pure-integer case (Brand et al., 2021). This establishes that the familiar intuition from integer Graver theory does not transfer unchanged to mixed-integer separable convex optimization.

4. Toric ideals, circuits, and Gröbner-theoretic relations

In commutative algebra, the Graver basis is the set of primitive binomials of a toric ideal. For $u,v\in \mathbb{Z}^n$5, the toric ideal

$u,v\in \mathbb{Z}^n$6

is generated by binomials corresponding to lattice relations. Circuits are support-minimal primitive relations, the universal Gröbner basis is the union of all reduced Gröbner bases, and the standard containments are

$u,v\in \mathbb{Z}^n$7

(Thoma et al., 2023).

A major negative result is that Graver degrees are not polynomially bounded by true circuit degrees in general. For toric ideals of graphs, an explicit family $u,v\in \mathbb{Z}^n$8 exhibits Graver elements whose degrees grow exponentially in $u,v\in \mathbb{Z}^n$9 while the maximal circuit degree grows only linearly. This disproves the conjectured polynomial control of Graver degrees by true circuit degrees and shows that circuit structure cannot, in general, be used to bound Graver complexity polynomially (Tatakis et al., 2013). Closely related work proves analogous non-polynomial separations in both degree and size between Graver bases, universal Gröbner bases, Markov bases, and circuits whenever the corresponding inclusions are strict (Tatakis et al., 2019).

At the opposite extreme, some classes collapse these bases. For self-dual projective toric varieties, the toric ideal is weakly robust: the Graver basis equals the union of all minimal binomial generating sets. In the non-pyramidal self-dual case, the ideal is strongly robust, and the Graver basis is itself a minimal generating set, hence it coincides with the universal Gröbner basis and with the unique minimal binomial generating set (Thoma et al., 2023). A 2025 result establishes analogous equalities of Graver basis, universal Gröbner basis, and a reduced Gröbner basis for toric ideals of certain weighted oriented graphs built from cycles sharing a path (Nanduri et al., 21 Apr 2025).

Equality with the universal Gröbner basis is also known in specific older families. For rational normal scrolls and homogeneous primitive colored partition identities, one paper classifies exactly when $$u \sqsubseteq v \iff u_i v_i \ge 0 \text{ and } |u_i|\le |v_i| \text{ for every coordinate } i.$$0, emphasizing that this equality is neither generic nor captured by simple notions such as normality alone (Bogart et al., 2010).

5. Graphical, statistical, and combinatorial incarnations

For graph toric ideals, Graver elements admit explicit combinatorial descriptions. In simple undirected graphs, a primitive binomial corresponds to an even closed walk satisfying a precise vertex condition: each vertex appears at most twice, and when a vertex appears twice, the two induced closed subwalks are odd and intersect only at that vertex (Ogawa et al., 2011). This gives an algorithmic route to generate primitive walks, and hence Graver moves, for Markov chain Monte Carlo over graph fibers.

That graphical viewpoint has direct statistical applications. In testing the beta model of random graphs, Graver moves connect all graphs with a fixed degree sequence, and the square-free part of the Graver basis connects the fiber of simple graphs. This makes Graver-based MCMC applicable even where smaller move sets such as 2-switches fail to connect the state space (Ogawa et al., 2011).

Connectivity improvements are also visible in fiber graphs. For a family of matrices $$u \sqsubseteq v \iff u_i v_i \ge 0 \text{ and } |u_i|\le |v_i| \text{ for every coordinate } i.$$1, Gröbner fiber graphs can have edge-connectivity equal to $$u \sqsubseteq v \iff u_i v_i \ge 0 \text{ and } |u_i|\le |v_i| \text{ for every coordinate } i.$$2 while their minimal degree is $$u \sqsubseteq v \iff u_i v_i \ge 0 \text{ and } |u_i|\le |v_i| \text{ for every coordinate } i.$$3, producing arbitrarily bad bottlenecks. The same fibers become optimally edge-connected when one uses Graver moves instead: the Graver fiber graphs satisfy

$$u \sqsubseteq v \iff u_i v_i \ge 0 \text{ and } |u_i|\le |v_i| \text{ for every coordinate } i.$$4

for all feasible right-hand sides in that family (Hemmecke et al., 2014). This supports the broader conjecture that Graver move sets may systematically yield best-possible edge-connectivity in fiber graphs.

Hierarchical models provide another combinatorial realization. For unimodular hierarchical models, the Graver basis coincides with the circuit set and the universal Gröbner basis, and it can be characterized via signed cycles or signed bonds in auxiliary directed graphs attached to the underlying simplicial complex (Bernstein et al., 2017). This gives a direct combinatorial description of all primitive moves in those unimodular statistical models.

In a different combinatorial direction, the Graver basis of a total-variation constraint matrix on a graph is characterized as the collection of connected-subgraph indicator moves. This identifies primitive TV-compatible augmentations with raising or lowering a connected plateau by one unit, linking Graver theory to $$u \sqsubseteq v \iff u_i v_i \ge 0 \text{ and } |u_i|\le |v_i| \text{ for every coordinate } i.$$5-convex steepest descent and to randomized augmentation heuristics for TV-regularized integer programs (Yang et al., 7 Aug 2025).

6. Computation, heuristics, and limitations

Exact computation of the Graver basis is difficult in general. Multiple sources emphasize that full Graver computation can be exponential or $$u \sqsubseteq v \iff u_i v_i \ge 0 \text{ and } |u_i|\le |v_i| \text{ for every coordinate } i.$$6-hard, which is why many practical methods rely either on strong matrix structure or on approximate extraction (Liu et al., 2024). Classical exact methods such as completion or project-and-lift remain expensive on large instances, although they are effective for moderate-scale toric computations (Liu et al., 2024).

This has motivated heuristic and hybrid approaches. One early quantum-classical proposal formulates kernel sampling and feasibility as QUBOs, uses a quantum annealer to sample lattice elements, and then filters them classically to recover conformally minimal directions and perform Graver augmentation for nonlinear integer programming (Alghassi et al., 2019). A fully classical descendant, GAMA, uses structured Graver bases and multi-seed augmentation for non-convex problems such as CBQP, QSAP, and QAP, relying on the observation that these objectives can behave as “quilted convex” over combinatorial slices (Alghassi et al., 2019).

More recently, MAPLE proposes GPU-based approximate extraction of Graver directions by optimizing a nonconvex continuous surrogate in coordinates of an integer kernel basis and then rounding to integer kernel vectors. The extracted set is used as a reusable approximate test set for augmentation across instances sharing the same constraint matrix, preserving feasibility by construction though without an optimality certificate when the extracted set is incomplete (Liu et al., 2024).

These computational developments coexist with strong structural limitations. The exponential-vs-linear separation between Graver and circuit degrees shows that circuit-based surrogates cannot control Graver complexity in general (Tatakis et al., 2013). Lower bounds for 4-block and $$u \sqsubseteq v \iff u_i v_i \ge 0 \text{ and } |u_i|\le |v_i| \text{ for every coordinate } i.$$7-fold matrices show that Graver norms can grow polynomially or exponentially with structural parameters, and these bounds can be tight or nearly tight