Deterministic Grid Graphs (DGG) Overview

• Deterministic grid graphs are models built on regular lattice structures that replace random node placements with fixed grids, yielding explicit spectral formulas.
• They serve as surrogates for random geometric graphs by facilitating quantitative approximations in connectivity, transferability in network models, and insights into combinatorial structures.
• Variants like Cartesian and local-grid graphs provide clear frameworks for embedding, decomposition, and algorithmic complexity, highlighting their practical and theoretical significance.

Deterministic Grid Graphs (DGG) denotes a family of deterministic graph models built from grid structure, but the term is not used with a single universal meaning across the literature represented here. In the spectral theory of random geometric graphs, a DGG is the deterministic geometric graph obtained by replacing random node locations by a regular grid and retaining the same radius-threshold edge rule. In adjacent areas, closely related deterministic grid objects include Cartesian products of paths or cycles, graphs whose local neighborhoods are grid-shaped, and directed cylindrical grids that serve as width obstructions in digraph minor theory (Hamidouche et al., 2019, Millichap et al., 2021, Amarra et al., 2019, Kawarabayashi et al., 2014).

1. Terminology and scope

The main usages represented in the literature can be organized as follows.

Usage	Formal object	Representative papers
Deterministic geometric graph	$G(\mathcal D_n,r_n)$ on a regular lattice or torus grid	(Hamidouche et al., 2019, Hamidouche et al., 2019, Camargo et al., 2 Jun 2026)
Cartesian grid graph	$P_{n_1}\square\cdots\square P_{n_d}$ or $C_{n_1}\square\cdots\square C_{n_d}$	(Millichap et al., 2021, Danai et al., 1 Jan 2026, Sardroud et al., 2014, Hakim et al., 12 Jun 2026, Cam et al., 5 Jun 2026)
Local-grid graph	$\Gamma(x)\cong K_n\square K_n$ for every vertex $x$	(Amarra et al., 2019)
Directed grid obstruction	Cylindrical grid or cylindrical wall in a digraph	(Kawarabayashi et al., 2014)
Implicit labeled grid graph	Grid cells as vertices, edges induced by label equality	(Au, 26 Apr 2026)

The most explicit use of the phrase “deterministic grid graph” occurs in work on random geometric graphs. There the DGG is a deterministic surrogate for an RGG, introduced because its symmetry yields explicit spectral formulas and quantitative approximation results. By contrast, several graph-theoretic papers study deterministic grid families without foregrounding the DGG label itself. This terminological dispersion matters: a DGG in spectral random geometry is not the same object as a directed cylindrical grid, a locally $n\times n$ grid graph, or an implicit colored-grid graph.

2. Deterministic geometric graphs on regular lattices

In the random-geometric-graph literature, the deterministic grid graph is defined by replacing the random point set $\mathcal X_n$ with a deterministic grid

$$\mathcal D_n=\text{the set of } n \text{ grid points at intersections of all parallel hyperplanes with separation } n^{-1/d}.$$

The graph $G(\mathcal D_n,r_n)$ is then formed by connecting two grid points whenever their distance is at most $r_n$: $P_{n_1}\square\cdots\square P_{n_d}$0 The corresponding RGG uses $P_{n_1}\square\cdots\square P_{n_d}$1 i.i.d. uniform points on the $P_{n_1}\square\cdots\square P_{n_d}$2-dimensional torus $P_{n_1}\square\cdots\square P_{n_d}$3 and the same distance-threshold rule. The only difference is node placement: random for the RGG, deterministic and equally spaced for the DGG (Hamidouche et al., 2019).

This model is intrinsically geometric and typically uses torus geometry to avoid boundary effects. The DGG is exactly regular, or asymptotically homogeneous, because the lattice is translation-invariant. In the $P_{n_1}\square\cdots\square P_{n_d}$4-metric with $P_{n_1}\square\cdots\square P_{n_d}$5, its degree is

$P_{n_1}\square\cdots\square P_{n_d}$6

and its adjacency matrix is circulant when $P_{n_1}\square\cdots\square P_{n_d}$7 and block circulant with circulant blocks when $P_{n_1}\square\cdots\square P_{n_d}$8. That algebraic regularity is the main reason the DGG is analytically tractable (Hamidouche et al., 2019).

A related deterministic lattice model appears in wireless-network transferability analysis. There a DGG is defined by taking a regular lattice in Euclidean space with $P_{n_1}\square\cdots\square P_{n_d}$9 intersections, placing a node at each intersection, and connecting neighboring nodes following the grid. The paper then studies conflict graphs of these DGGs, where each conflict-graph vertex represents an original edge and two conflict-graph vertices are adjacent when the corresponding original edges are incident on the same node. In that setting the DGG functions as a deterministic baseline for sparse wireless topologies (Camargo et al., 2 Jun 2026).

3. Spectral approximation and transferability

The most developed theory of DGGs concerns approximation of random geometric graphs by deterministic ones. For the regularized normalized Laplacian, the key result is that in the connectivity regime

$C_{n_1}\square\cdots\square C_{n_d}$0

the empirical spectral distribution of the RGG converges in probability, in Lévy distance, to that of the corresponding DGG: $C_{n_1}\square\cdots\square C_{n_d}$1 In the thermodynamic regime,

$C_{n_1}\square\cdots\square C_{n_d}$2

the DGG remains an explicit approximation after regularization, with error threshold

$C_{n_1}\square\cdots\square C_{n_d}$3

In $C_{n_1}\square\cdots\square C_{n_d}$4, the DGG is simple enough to yield explicit eigenvalue formulas, and in the connectivity regime its limiting eigenvalue distribution collapses to the Dirac measure at $C_{n_1}\square\cdots\square C_{n_d}$5 (Hamidouche et al., 2019).

The same surrogate role appears for adjacency spectra. In the connectivity regime, the empirical spectral distribution of the RGG adjacency matrix converges in probability to that of the DGG adjacency matrix under appropriate growth conditions on the average degree and the bottleneck matching scale. In the $C_{n_1}\square\cdots\square C_{n_d}$6-metric, the DGG eigenvalues admit an explicit multidimensional DFT formula, reflecting the circulant or block-circulant structure of the periodic grid (Hamidouche et al., 2019).

In wireless resource allocation, this spectral and operator-level regularity is turned into a transferability theory for graph neural networks. The paper assumes that the conflict graph of an RGG is a perturbation of the conflict graph of a DGG, with zero-padded normalized adjacency matrices satisfying

$C_{n_1}\square\cdots\square C_{n_d}$7

For an $C_{n_1}\square\cdots\square C_{n_d}$8-layer GNN $C_{n_1}\square\cdots\square C_{n_d}$9 built from an integral Lipschitz graph filter with constant $\Gamma(x)\cong K_n\square K_n$0 and a normalized Lipschitz nonlinearity, the output perturbation is bounded by

$\Gamma(x)\cong K_n\square K_n$1

The same paper also proves a small-DGG to large-DGG transfer theorem based on the fact that, in 2D grids, interior nodes eventually have the same $\Gamma(x)\cong K_n\square K_n$2-hop neighborhoods and only boundary effects remain. DGGs therefore act as the deterministic middle layer in the chain

$\Gamma(x)\cong K_n\square K_n$3

which yields scale-transfer guarantees for sparse wireless conflict graphs (Camargo et al., 2 Jun 2026).

4. Cartesian grid graphs: paths, cycles, and surface embeddings

A second major strand treats deterministic grids as explicit Cartesian products. In this usage a $\Gamma(x)\cong K_n\square K_n$4-dimensional grid graph is

$\Gamma(x)\cong K_n\square K_n$5

with

$\Gamma(x)\cong K_n\square K_n$6

and

$\Gamma(x)\cong K_n\square K_n$7

This framework supports a substantial topological theory. For connected graphs of girth at least $\Gamma(x)\cong K_n\square K_n$8,

$\Gamma(x)\cong K_n\square K_n$9

and for grids this lower bound is attained exactly when the graph admits a quadrilateral embedding. In particular, if $x$0 and at least three grid parameters are odd, then $x$1 admits a quadrilateral embedding and hence

$x$2

The same paper gives a complete classification of planar grid graphs and toroidal grid graphs, showing that planarity occurs exactly for $x$3 or for certain essentially one-dimensional 3D cases, while toroidal embeddability survives only in a short explicit list (Millichap et al., 2021).

Rectangular grid graphs also admit exact length-constrained path and cycle theory. In a rectangular grid $x$4, a cycle of length $x$5 exists if and only if

$x$6

For two specified vertices $x$7, if $x$8 and $x$9 denote the lengths of a shortest and a longest $n\times n$0-$n\times n$1 path, then there is an $n\times n$2-$n\times n$3 path of length $n\times n$4 if and only if

$n\times n$5

The constructive side is equally sharp: the paper gives an $n\times n$6 algorithm for cycles of prescribed length and an $n\times n$7 algorithm for paths of prescribed length, and extends the method to $n\times n$8 3D grids (Sardroud et al., 2014).

A complementary design-theoretic direction studies decompositions of complete graphs into grid graphs. Here a grid graph is defined as a Cartesian product of path graphs $n\times n$9 and cycle graphs $\mathcal X_n$0, and a grid design is a $\mathcal X_n$1-design where $\mathcal X_n$2 is such a grid graph. The paper proves that when $\mathcal X_n$3 is an odd prime or the square of an odd prime, the toroidal grid $\mathcal X_n$4 admits a $\mathcal X_n$5-design. In the path-product setting it proves the opposite outcomes

$\mathcal X_n$6

with explicit finite-field constructions in the positive cases (Danai et al., 1 Jan 2026).

5. Decompositions, width, and representation theory

Deterministic Cartesian grids are also central examples in modern width theory. For

$\mathcal X_n$7

the rank-width lower bound proved via expansion, induced matchings, and cut-rank is

$\mathcal X_n$8

Since the same paper states

$\mathcal X_n$9

the resulting comparison is

$$\mathcal D_n=\text{the set of } n \text{ grid points at intersections of all parallel hyperplanes with separation } n^{-1/d}.$$0

For fixed $$\mathcal D_n=\text{the set of } n \text{ grid points at intersections of all parallel hyperplanes with separation } n^{-1/d}.$$1, grids therefore have sublinear but high rank-width,

$$\mathcal D_n=\text{the set of } n \text{ grid points at intersections of all parallel hyperplanes with separation } n^{-1/d}.$$2

which recovers the classical $$\mathcal D_n=\text{the set of } n \text{ grid points at intersections of all parallel hyperplanes with separation } n^{-1/d}.$$3 scale for square grids and extends it to higher dimensions (Cam et al., 5 Jun 2026).

Representation theory provides a different set of deterministic-grid invariants. Every $$\mathcal D_n=\text{the set of } n \text{ grid points at intersections of all parallel hyperplanes with separation } n^{-1/d}.$$4-dimensional grid

$$\mathcal D_n=\text{the set of } n \text{ grid points at intersections of all parallel hyperplanes with separation } n^{-1/d}.$$5

is proved to be a $$\mathcal D_n=\text{the set of } n \text{ grid points at intersections of all parallel hyperplanes with separation } n^{-1/d}.$$6-interval-PCG and also a $$\mathcal D_n=\text{the set of } n \text{ grid points at intersections of all parallel hyperplanes with separation } n^{-1/d}.$$7-OR-PCG. The construction decomposes the grid into hyperplanes of constant coordinate sum and uses a large-base encoding so that leaf-to-leaf distances in a weighted tree identify coordinate directions. The lower-bound side is equally explicit: the graph

$$\mathcal D_n=\text{the set of } n \text{ grid points at intersections of all parallel hyperplanes with separation } n^{-1/d}.$$8

is not a PCG. Consequently, the minimum number of intervals sufficient for all three-dimensional grid graphs is exactly $$\mathcal D_n=\text{the set of } n \text{ grid points at intersections of all parallel hyperplanes with separation } n^{-1/d}.$$9, and the same exact threshold holds for OR-PCGs in dimension $G(\mathcal D_n,r_n)$0 (Hakim et al., 12 Jun 2026).

Taken together, these results show that deterministic grids are neither trivial nor generic. They are rigid enough to permit exact algebraic and combinatorial constructions, yet rich enough to witness nontrivial width growth, sharp representability thresholds, and explicit separation between single-interval and multi-interval tree-metric models.

6. Local-grid and directed-grid analogues

A different meaning of “grid-like determinism” appears in locally $G(\mathcal D_n,r_n)$1 grid graphs. Here the graph $G(\mathcal D_n,r_n)$2 is not itself a global Cartesian rectangle; instead every vertex neighborhood is required to satisfy

$G(\mathcal D_n,r_n)$3

These graphs have valency $G(\mathcal D_n,r_n)$4, and for a distance-two pair $G(\mathcal D_n,r_n)$5 the $G(\mathcal D_n,r_n)$6-graph

$G(\mathcal D_n,r_n)$7

has even order between $G(\mathcal D_n,r_n)$8 and $G(\mathcal D_n,r_n)$9. The paper proves strong threshold phenomena: if every distance-two pair is joined by at least $r_n$0 paths of length $r_n$1, then the diameter is bounded by $r_n$2; if every such pair is joined by at least $r_n$3 paths of length $r_n$4, then

$r_n$5

and

$r_n$6

Equality forces distance-regular antipodal covers of complete graphs, and the paper constructs an infinite family for every odd prime power $r_n$7 (Amarra et al., 2019).

Directed graph theory uses yet another, nonstandardized analogue. The paper “The Directed Grid Theorem” does not use the phrase deterministic grid graph; the guaranteed object is a cylindrical grid $r_n$8, not an oriented $r_n$9 Cartesian grid. Its main theorem states that there is a function $P_{n_1}\square\cdots\square P_{n_d}$00 such that every digraph of directed tree-width at least $P_{n_1}\square\cdots\square P_{n_d}$01 contains a cylindrical grid of order $P_{n_1}\square\cdots\square P_{n_d}$02 as a butterfly minor. Equivalently, sufficiently large directed brambles force the same obstruction. The paper also proves a constructive algorithmic form: for fixed $P_{n_1}\square\cdots\square P_{n_d}$03, one can in polynomial time obtain either a cylindrical grid of order $P_{n_1}\square\cdots\square P_{n_d}$04 as a butterfly minor or a directed tree decomposition of bounded width. In directed minor theory, therefore, the canonical deterministic grid obstruction is cylindrical and minor-theoretic rather than rectangular and Cartesian (Kawarabayashi et al., 2014).

These two analogues show that “grid” can refer either to exact local neighborhood geometry or to a canonical obstruction object for width. In both cases the deterministic content lies in rigid incidence patterns, but the resulting theories differ sharply from the geometric DGG model of random geometric graph approximation.

7. Algorithmic and dynamical frontiers

Deterministic grid structure does not by itself fix algorithmic complexity; the precise access model and geometric restrictions are decisive. In an implicit colored-grid model, the input is an $P_{n_1}\square\cdots\square P_{n_d}$05 array

$P_{n_1}\square\cdots\square P_{n_d}$06

with graph edges only between geometrically adjacent cells of equal color: $P_{n_1}\square\cdots\square P_{n_d}$07 For this model the deterministic query complexity of cycle detection is maximal: for all $P_{n_1}\square\cdots\square P_{n_d}$08, any deterministic algorithm deciding whether the grid graph contains a cycle must read all $P_{n_1}\square\cdots\square P_{n_d}$09 cells in the worst case. The proof uses adaptive adversaries on $P_{n_1}\square\cdots\square P_{n_d}$10, $P_{n_1}\square\cdots\square P_{n_d}$11, $P_{n_1}\square\cdots\square P_{n_d}$12, and $P_{n_1}\square\cdots\square P_{n_d}$13 blocks, together with checkerboard isolation by disjoint alphabets (Au, 26 Apr 2026).

Hamiltonicity shows a similarly sharp dependence on structural restrictions. For lattice-based square, triangular, and hexagonal grid graphs, the paper proves that square polygonal grid graph Hamiltonicity and hexagonal thin grid graph Hamiltonicity are NP-complete, while thin polygonal grid graph Hamiltonicity is polynomial-time solvable for square, triangular, and hexagonal grids. In the square case the positive statement is stronger: every polygonal thin square grid graph is Hamiltonian. This suggests that deterministic geometric regularity alone does not imply tractability; the conjunction of thinness and polygonality is the critical rigidity in that setting (Demaine et al., 2017).

Deterministic dynamics on grids exhibit the same pattern. For 2-neighbour bootstrap percolation, deciding whether the maximum percolation time $P_{n_1}\square\cdots\square P_{n_d}$14 is NP-complete even for grid graphs with maximum degree $P_{n_1}\square\cdots\square P_{n_d}$15, while on solid grid graphs with maximum degree $P_{n_1}\square\cdots\square P_{n_d}$16 the value $P_{n_1}\square\cdots\square P_{n_d}$17 can be computed in $P_{n_1}\square\cdots\square P_{n_d}$18 time. The structural reason is that in degree-$P_{n_1}\square\cdots\square P_{n_d}$19 settings, long percolation time is equivalent to the existence of induced paths of specific degree patterns, and solidity collapses the relevant geometry to ladders and paths (Marcilon et al., 2015). In online exploration, even the known-topology, unknown-weight problem remains nontrivial on ladders $P_{n_1}\square\cdots\square P_{n_d}$20: every deterministic strategy has competitive ratio at least $P_{n_1}\square\cdots\square P_{n_d}$21, while a $P_{n_1}\square\cdots\square P_{n_d}$22-competitive deterministic algorithm exists; for directed ladders and square grids, a natural greedy strategy has linear lower bounds on competitive ratio (Böckenhauer et al., 2016).

Across these problems, the common theme is that deterministic grids are a family of highly explicit graphs rather than a single canonical complexity class. Some deterministic grid models yield exact spectral formulas, exact length spectra, or exact decomposition theorems; others support NP-completeness, worst-case query lower bounds, or only conditional tractability. The term DGG is therefore best read as a structured umbrella over several deterministic grid formalisms whose common feature is rigid spatial or combinatorial organization, not a single universally agreed graph class.