Orthogonal Folding in Computational Geometry

Updated 5 July 2026

Orthogonal folding is a family of geometric and algorithmic problems defined by axis-aligned structures and crease patterns in polyhedra and lattices.
The methodology involves grid refinement, spiral-path traversals, and heavy–light ordering to determine unfolding feasibility and computational complexity.
Implications span NP-completeness in folding models, universality in fold-and-cut designs, and applications in biomolecular lattice chains and origami.

Orthogonal folding is a family of geometric, combinatorial, and algorithmic problems in which the admissible structure is axis-aligned: polyhedra have edges parallel to the coordinate axes, crease patterns use horizontal and vertical folds, or lattice chains advance by orthogonal steps. In this literature, the term covers both direct folding problems and inverse unfolding problems. The common theme is the interaction between orthogonality, nonoverlap, topology, and computational complexity: one asks whether an object can be flattened or refolded, which cuts or creases are sufficient, how much grid refinement is needed, and whether feasibility is decidable efficiently or is computationally hard (Damian et al., 2011, Akitaya et al., 2023).

1. Formal scope and core models

An orthogonal polyhedron is a polyhedron whose edges are parallel to the $x$ -, $y$ -, or $z$ -axes; consequently, all dihedral angles between faces are $90^\circ$ . In the unfolding literature, the central topological assumption is often that the surface is homeomorphic to a sphere, i.e., genus $0$. An unfolding is a cut set on the surface such that the cut surface can be isometrically flattened into the plane without overlap, and the planar image is a net. Cuts are typically restricted to a vertex grid, or to a refined version of that grid obtained by inserting additional axis-parallel planes between consecutive vertex planes (Damian et al., 2011).

In the crease-pattern literature, the paper region is a line segment, a rectangle, or an arbitrary orthogonal polygon, and an orthogonal crease pattern is a finite set of axis-parallel crease segments. A mountain/valley assignment is a map $a : P \to \{M,V,?\}$ , where $M$ denotes mountain, $V$ valley, and $?$ an unassigned crease. A simple fold chooses a single line and rotates a contiguous set of layers by $\pm 180^\circ$ about that line; the principal models are one-layer, some-layers, and all-layers simple folds, together with finite and infinite fold-line variants (Akitaya et al., 2023).

A third usage appears in fold-and-cut. Here the sheet is an axis-aligned rectangle, the creases are horizontal and vertical lines, and a single straight cut is made after folding. The cut pattern is the preimage of that line under the folding map, and the geometry factors independently by $y$ 0- and $y$ 1-coordinates in the folded state (Ani et al., 2022).

A further, distinct usage appears in lattice models of chains and biomolecules. In these settings, orthogonal folding means embedding a self-avoiding walk or fixed-angle chain on a square or cubic lattice using axis-aligned steps, or axis-aligned steps together with constrained plane-diagonal moves (Demaine et al., 2022, P et al., 2023).

A standard inverse perspective on orthogonal folding is orthogonal unfolding: a nonoverlapping net immediately specifies a folding pattern along the uncut grid lines. Early positive results for genus-0 orthogonal polyhedra showed that gridding is often necessary. "Grid Vertex-Unfolding Orthogonal Polyhedra" proves that any orthogonal polyhedron of genus zero has a grid vertex-unfolding, and states an $y$ 2-time algorithm together with a simpler vertex-unfolding algorithm that requires a $y$ 3 refinement of the vertex grid [0509054].

The principal general genus-0 improvement is "Unfolding Orthogonal Polyhedra with Quadratic Refinement: The Delta-Unfolding Algorithm" (Damian et al., 2011). It shows that every orthogonal polyhedron homeomorphic to a sphere can be unfolded without overlap while using only polynomially many orthogonal cuts, whereas the preceding epsilon-unfolding required exponential refinement in the worst case. More precisely, if $y$ 4 has $y$ 5 vertices, Delta-Unfolding cuts only where the surface meets the vertex grid together with $y$ 6 additional coordinate planes inserted between every two consecutive vertex planes in each of $y$ 7, $y$ 8, and $y$ 9 (Damian et al., 2011).

The algorithm slices the polyhedron by $z$ 0-planes into slabs, represents each slab by a band with two rims, and links bands by vertical $z$ 1-beams. These beams induce an unfolding tree $z$ 2. A thin non-self-crossing spiral path $z$ 3 is then routed across the bands; after thickening in $z$ 4, $z$ 5 covers the band faces, and in the planar layout it becomes a monotone staircase strip. Forward and rearward faces are partitioned into strips by vertical illumination from rim segments and hang above or below the staircase (Damian et al., 2011).

The main combinatorial innovation is heavy–light ordering. For a band $z$ 6 with subtree size $z$ 7, a child $z$ 8 is heavy if $z$ 9. At each band, the heavy child, if present, is visited last, so the largest subtree is not retraced repeatedly. This yields the recurrence

$90^\circ$ 0

and the quadratic bound is tight: a perfect binary unfolding tree gives $90^\circ$ 1 (Damian et al., 2011). A common misconception is that the result is merely existential; in fact, the construction specifies where cuts lie: along rims, along the spiral path $90^\circ$ 2, along $90^\circ$ 3-beam boundaries, and along illumination partitions, all aligned to the refined orthogonal grid (Damian et al., 2011).

Several special classes admit stronger bounds than quadratic refinement. "Unfolding Orthogonal Terrains" proves that every orthogonal terrain based on a rectangular base admits a $90^\circ$ 4 grid unfolding: the surface unfolds to a single connected non-overlapping, weakly simple planar piece by cutting only along grid edges induced by coordinate planes through the vertices. The terrain hypothesis is the $90^\circ$ 5-monotonicity condition that every vertical line meeting the solid does so in a single segment with one endpoint on the rectangular base (0707.0610).

"Unfolding Orthogrids with Constant Refinement" defines orthogrids as genus-0 orthogonal polyhedra with no $90^\circ$ 6-dents and all left vertices exposed, and proves that they admit a constant- $90^\circ$ 7 refined grid unfolding for some constant $90^\circ$ 8 independent of input size. The construction again uses layers, bands, rims, an unfolding tree, and a spiral-like traversal, but enforces nonoverlap through invariant-driven forced turns and anchor-specific transitions (Damian et al., 2013).

A major extension beyond genus $90^\circ$ 9 is "Unfolding Genus-2 Orthogonal Polyhedra with Linear Refinement" (Damian et al., 2016). It proves that any orthogonal polyhedron of genus $0$0 may be unfolded to a planar simple orthogonal polygon by cutting along a linear grid-refinement. The method passes from a band tree to a rim graph $0$1, constructs a rim unfolding tree $0$2, and shows that the number of nonface leaves created is at most $0$3. For $0$4, this gives at most two special leaves, which can still be stitched into a single strip at the root. The paper emphasizes that genus $0$5 and beyond require new techniques, both because a face-node root may fail to exist and because the two-terminal stitching paradigm breaks down (Damian et al., 2016).

Recent work has also isolated edge-unfoldable subclasses that need no extra refinement at all. "Edge-Unfolding Polycubes with Orthogonally Convex Layers" proves that any genus-0 polycube whose horizontal unit-height layers are orthogonally convex can be edge unfolded. The construction is layer-based: it visits band segments, connects them with bridges, attaches top and bottom beams, and completes remaining band pieces by local case analysis. The implementation runs in time linear in the number of surface cells $0$6, and all cuts are along cube edges (Damian et al., 2024).

Taken together, these results show that orthogonality by itself does not determine the refinement threshold. Terrains permit $0$7 grid unfolding, orthogrids permit constant refinement, genus-$0$8 general orthogonal polyhedra are known to admit quadratic refinement, and genus $0$9 can be handled with linear refinement in a different framework (0707.0610, Damian et al., 2013, Damian et al., 2011, Damian et al., 2016).

4. Simple folds, map folding, and complexity classification

The algorithmic theory of orthogonal folding in the crease-pattern sense begins with simple folds. "When Can You Fold a Map?" proves a foundational 1D equivalence: a 1D crease pattern is flat-foldable by any means precisely if it is foldable by a sequence of one-layer simple folds. The same paper gives a linear-time algorithm for deciding foldability of an orthogonal crease pattern on a rectangular piece of paper, while proving weak NP-completeness for slight generalizations, including orthogonal crease patterns on orthogonal paper, axis-parallel plus $a : P \to \{M,V,?\}$ 0 crease patterns on a square, and crease patterns without a mountain/valley assignment [0011026].

"Complexity of Simple Folding of Mixed Orthogonal Crease Patterns" sharpens and nearly completes the classification for orthogonal patterns with assigned, unassigned, or mixed mountain/valley labels (Akitaya et al., 2023). In 1D, foldability of mixed crease patterns is polynomial-time for one-layer and some-layers simple folds, via a characterization in terms of suspicious and innocent intervals; the stated complexity is $a : P \to \{M,V,?\}$ 1, with improvements possible. For mixed 1D all-layers folding, the paper gives a refined greedy algorithm based on plausible creases nearest an end, again polynomial-time (Akitaya et al., 2023).

In 2D, the behavior splits sharply by paper geometry and layer model. For mixed orthogonal crease patterns on rectangular paper, one-layer simple foldability is polynomial-time: if both horizontal and vertical creases are present, no solution exists, and if all creases are parallel the problem reduces to the 1D mixed case. By contrast, mixed orthogonal crease patterns on rectangular paper are strongly NP-complete in the some-layers model, and finite simple folds of unassigned orthogonal crease patterns on arbitrary orthogonal paper are strongly NP-complete for all three layer models (Akitaya et al., 2023).

A central structural lemma in the hardness proofs states that, on rectangular paper, any fold whose direction differs from the previous fold must be an all-layers fold, or else it would tear the paper along previously folded creases (Akitaya et al., 2023). This helps explain why orthogonality does not imply tractability. A plausible implication is that the decisive boundary is not axis alignment alone, but the product structure of rectangular paper together with strict limitations on how fold directions can alternate 0011026.

5. Universal orthogonal patterns and fold-and-cut characterizations

Orthogonal folding also includes constructive universality results. "A Universal Crease Pattern for Folding Orthogonal Shapes" shows that tetrakis tilings, or box pleating, provide one universal finite crease pattern for each $a : P \to \{M,V,?\}$ 2 that can fold any polycube of size $a : P \to \{M,V,?\}$ 3. The main quantitative statement is that, for any polycube of $a : P \to \{M,V,?\}$ 4 unit cubes, a tetrakis crease pattern on a rectangle of size $a : P \to \{M,V,?\}$ 5 folds to the target with all faces seamless except one specified face; a corollary gives full seamlessness on $a : P \to \{M,V,?\}$ 6, and a square of side length $a : P \to \{M,V,?\}$ 7 also suffices (0909.5388). The novelty lies in fixing the crease geometry per $a : P \to \{M,V,?\}$ 8: the target shape is encoded in the angle assignment and overlap order, not in a new geometric pattern for each instance (0909.5388).

A different constructive line is "Orthogonal Fold & Cut" (Ani et al., 2022). Here one folds an axis-aligned rectangular sheet along horizontal and vertical creases and then makes a single straight cut at any angle. The paper gives a complete characterization of the resulting cut patterns. In the non-axis-aligned case, all cut segments must have coherent slope $a : P \to \{M,V,?\}$ 9; after scaling to $M$ 0, the rectangle decomposes into horizontal and vertical bands where creases are forbidden and stripes where creases are allowed. The canonical solution places one vertical crease at the center of each vertical stripe and one horizontal crease at the center of each horizontal stripe, including zero-width stripes, and this canonical crease pattern suffices whenever any orthogonal fold-and-cut solution exists (Ani et al., 2022).

The same paper solves orthogonal fold & punch: a set of holes on a rectangle is achievable if and only if it is a combinatorial rectangle $M$ 1. It also proves a scissor-cut equivalence for solvable orthogonal fold-and-cut instances: no crease needs positive-length overlap with a cut segment (Ani et al., 2022). These results show that orthogonality can support both universality and exact characterization, but only under highly structured crease models.

6. Lattice-chain and biomolecular interpretations

In discrete chain models, orthogonal folding denotes lattice embeddings rather than nets or crease patterns. "Computational Complexity of Flattening Fixed-Angle Orthogonal Chains" studies equilateral chains whose interior angles are fixed to be either $M$ 2 or $M$ 3. For open chains, there is always a planar noncrossing zig-zag embedding. For closed chains, deciding whether a planar noncrossing embedding exists is strongly NP-complete. The box-constrained version for open chains is also strongly NP-complete, and the fixed-angle HP optimization problem is strongly NP-complete even when the chain has exactly two hydrophobic vertices (Demaine et al., 2022).

A related square-lattice model appears in "Properties of a Two Dimensional Model of RNA Folding" (Maron, 2018). There a folding is a self-avoiding walk on $M$ 4, contacts are nonconsecutive Watson–Crick adjacencies, and each nucleotide participates in at most one contact. The paper proves that for $M$ 5, the sequence $M$ 6 has a unique optimal folding, that finding a folding with at least $M$ 7 bonds is NP-hard, and that for sequences consisting only of $M$ 8 and $M$ 9, a linear-time algorithm achieves at least $V$ 0 of the optimum (Maron, 2018).

In three-dimensional coarse-grained protein models, orthogonal folding is implemented as a turn-based lattice walk. "An approach to solve the coarse-grained Protein folding problem in a Quantum Computer" uses a cubic lattice with unit axis-parallel moves and axial-plane diagonals. The formulation introduces binary variables for turn components, a continuity constraint that can penalize body-diagonal moves, self-avoidance and no-crossing penalties, and a QUBO objective suitable for both classical optimization and gate-based quantum hardware. With body diagonals penalized, the model guides orthogonal folding to axes and plane diagonals; the implementation is reported up to $V$ 1 residues, using $V$ 2 qubits in the $V$ 3 case (P et al., 2023).

These lattice models are terminologically distinct from origami and polyhedral unfolding, but they preserve the central geometric restriction: motion is confined to orthogonal directions or to tightly controlled extensions of them. This suggests that orthogonal folding is best understood not as a single problem, but as a methodological regime in which axis alignment converts geometric feasibility into a combination of discrete structure, refinement control, and sharp complexity transitions (Demaine et al., 2022, Maron, 2018, P et al., 2023).