betaFeedback

WorksMathematics

A Universal 88-Vertex Origami Torus Triangulation

Abstract

An origami torus is a piecewise affine isometric embedding of a flat torus into R3\R^3. We prove that every flat torus is realized as an 88-vertex origami torus based on the triangulation with degree sequence 4,6,6,6,6,6,7,74,6,6,6,6,6,7,7. This provides an 88-vertex universal triangulation for origami tori.

Introduction

The Main Result

A flat torus is a quotient of the form T=R2/ΛT = \mathbb{R}^{2}/\Lambda, where Λ\Lambda is a lattice of translations of R2\mathbb{R}^{2}. The lattice Λ\Lambda is called the lattice of TT. A triangulation T\mathcal{T} of a flat torus TT is a decomposition of TT into finitely many triangles, such that every two triangles are either disjoint, or share one vertex, or share one edge.

Two triangulations T1\mathcal{T}_{1} and T2\mathcal{T}_{2}, respectively of flat tori T1T_{1} and T2T_{2}, are combinatorially equivalent if there is a homeomorphism h:T1→T2h : T_{1} \to T_{2} which affinely maps each triangle of T1\mathcal{T}_{1} to a triangle of T2\mathcal{T}_{2}. The equivalence class of T\mathcal{T} is called its combinatorial type. We will sometimes blur the distinction between a triangulation and its combinatorial type. There are 7 different combinatorial types of 8-vertex triangulations of tori. Figure 1.1 below shows the 8-vertex triangulations T(46666677)\mathcal{T}(46666677) and T(66666666)\mathcal{T}(66666666). We are naming these particular triangulations by their degree sequence. In each of these two cases, the degree sequence determines the combinatorial type.

Figure 1.1. The triangulations T(46666677)\mathcal{T}(46666677) and T(666666666)\mathcal{T}(666666666).

An origami torus is an embedded topological torus in R3\mathbb{R}^{3} that is made by fitting together finitely many triangles so that the cone angle around each vertex is 2π2\pi. Equivalently, an origami torus is the image of a piecewise affine isometric embedding ϕ:T→R3\phi:T\to\mathbb{R}^{3} of a flat torus TT. This is to say that TT has a triangulation T\mathcal{T} such that the restriction of ϕ\phi to each triangle of T\mathcal{T} is an isometric affine map. (The specific triangulation is part of the data.) We call a lattice Λ\Lambda realized by T\mathcal{T} if there exists an origami torus associated to the lattice Λ\Lambda and the combinatorial type T\mathcal{T}. We call a combinatorial type T\mathcal{T} universal if every lattice Λ\Lambda is realized by T\mathcal{T}. Here is our main result.

Theorem 1.1 (Universality). T(46666677)\mathcal{T}(46666677) is universal.

The Universality Theorem says that there is a single 8-vertex combinatorial type out of which you can build an origami torus of any shape you like. Our construction yields 2 continuous families of 8-vertex origami tori based on T(46666677)\mathcal{T}(46666677) which together realize all shapes. We do not know if a single continuous family suffices, but see Remark 3.3. Here 8 is as good as it gets: In [2028] it is proved that 7-vertex origami tori do not exist. See also [2029]. Figure 1.2 below shows a representative example of one of our tori.

Three stages of an origami torus: flattened, inflated, and corrected

Figure 1.2. A flattened origami torus, inflated then corrected

History and Context

The Universality Theorem is the culmination of a series of results about origami tori, some quite recent.

  • In 1960, Y. Burago and V. Zalgaller [1960] give the first construction of origami tori.

  • In 1995, Y. Burago and V. Zalgaller [1995] prove that every lattice is realized by some combinatorial type.

  • In 2020, T. Tsuboi [T] gives explicit constructions which realize every lattice as an origami torus called a diplotorus (For the rectangular lattices one needs a modified construction.) P. Arnoux, S. Lelièvre, and A. Málaga [2027] give an independent proof of the same result. The construction requires infinitely many combinatorial types. Diplotori were discovered by U. Brehm [1978] in 1978. See also [2016 §6].

  • In 2023, F. Lazarus and F. Tallerie [2024] give a universal triangulation having 1217 vertices. Their paper also has an excellent discussion of the various attempts made to realize all lattices as origami tori.

  • In 2025, R.E.S. [2028] constructs an 8-vertex origami torus and proves that a 7-vertex origami torus does not exist. See also [2029]. The 8-vertex origami tori in [2028] and [2029], called pup tents, have combinatorial type T(66666666)T(66666666) and are symmetric with respect to order 2 rotation about the ZZ-axis.

  • In 2025, P.G.D. and R.E.S. [2025] prove that a lattice is realized by a pup tent provided that it is neither rectangular nor sharp rhombic. A lattice is rectangular if it is generated by a pair of perpendicular vectors and sharp rhombic if it is generated by a pair of equal-length vectors which make an angle of at most π/3\pi/3 with each other. When the angle is π/3\pi/3 you get the famous hexagonal lattice. Thus, almost every lattice is realized by a pup tent. In [2025] the authors speculate that perhaps the remaining lattices can be realized by 8-vertex origami tori without 2-fold symmetry.

  • In 2026, F. Lander [2026] proves that the square lattice is realized by T(66666666)T(66666666) and the hex lattice is realized by T(46666677)T(46666677). These origami tori have no symmetries.

    The result in [2026] grew out of some experiments that F.L. and S.T. did in the summer of 2026. These authors made a poster for R.E.S.’s birthday conference and, in an exciting development, their software could realize some rectangular and sharp rhombic lattices. Upon further experimentation, F.L. and S.T. discovered that the folding pattern used to realize the hexagonal lattice from T(46666677)T(46666677) turned out to be extremely robust. This fact led to the Universality Theorem.

Proof Strategy and Organization

Writing down explicit origami tori seems quite difficult, partly because there are too many choices to make. The idea that emerged in [2025], and was used again in [2026], is to work instead on the boundary of the space of origami tori, where faces collapse together and so make some of these choices for us. The most collapsed examples lie entirely in a plane; we find these first and then push them back inside. Here we carry this out for T(46666677)\mathcal{T}(46666677) and every lattice. In § 2 we give a geometric account of the strategy, ending in a proposition that reduces the proof to finitely many checks. We verify these checks in three steps.

Step 1: We give an explicit rational parametrization of a 5-dimensional family of flattened origami tori: flat tori which lie entirely in the plane! Within it we find two 2-parameter families which together realize every modulus. We do this in § 3.

Step 2: For each family we give a single vertical lift, the same heights for every member, which is embedded throughout. We call this inflating, and we do it in § 4.

Step 3: We check that horizontal motions can correct the cone angles and the modulus, so that Proposition 2.2 promotes each (inflated) flattened origami torus to an origami torus of the same modulus. We do this in in § 5.

The two families came out of extending the searches performed in [2026]. On several occasions we need certain polynomials not to vanish on certain domains. To keep the story clean, we leave the description of these certification techniques to § 7. Figure 1.2 shows the three step process which produces our examples.

Computer Calculations

We perform our calculations in Mathematica [W]. These calculations are all fairly short and easy to reproduce. One file contains everything. Here are two places to get the file.

http://www.math.brown.edu/~res/Java/UniversalityProof

https://github.com/FabianLander/universal-8-vertex-origami-torus-triangulation

Artificial Intelligence Use

We used Artificial Intelligence in a moderate way. The conceptual framework for the proof is ours, and evolved from the papers [2028], [2025], and [2026]. At various stages, ChatGPT 5.6 implemented some of our search algorithms and optimized some of our code. Perhaps needless to say, the AI also helped format equations, provide references, and proofread the paper. Also, we mention that ChatGPT had an excellent idea for simplifying the angle-derivative formulas for the Jacobian calculation in § 5.

Acknowledgements

R.E.S. would like to thank the I.H.E.S. for their support, in the form of the Gretchen and Barry Mazur Chair, during the writing of this paper.

Producing Origami Tori

We now describe the method we use to produce origami tori, following [2025] and [2026]: we construct degenerate example and then inflate them. In this paper, our degenerate examples are “completely flattened”: they lie in the plane.

Looking on the Boundary

Fix the triangulation T=T(46666677)\mathcal{T} = \mathcal{T}(46666677), though the idea is general. A polyhedral torus of type T\mathcal{T} is determined by its 8 vertices in R3\mathbb{R}^{3}, so these tori form the open subset of R24\mathbb{R}^{24} where no triangle is degenerate. The origami tori are its flat, embedded members, O=E∩F\mathcal{O} = \mathcal{E} \cap\mathcal{F}, where E\mathcal{E} is the open set of embedded tori and F\mathcal{F} is the set of intrinsically flat tori, whose triangle angles sum to 2π2\pi at each vertex.

Origami tori are quite hard to produce: although 8-vertex origami tori are known to exist [2028], [2025], [2026], none is known in closed form. The two sets of constraints are hard in different ways. Embeddedness is an open condition, so embedded tori can at least be found by sampling, though they are rare: for our triangulation they make up approximately 0.009% of all tori.3 Flatness is a system of equations: every cone angle must be exactly 2π2\pi. Sampling will never find a flat torus; they must instead be constructed by solving a nonlinear system exactly, which means making many choices. One way forward is to impose extra conditions that make some of these choices for us. But the resulting tori must still land in E\mathcal{E}, which is small. The idea is to draw the extra conditions from E\mathcal{E} itself.

Turning some of the inequalities of embeddedness into equalities puts us on the boundary of E\mathcal{E}. This boundary is stratified by how embeddedness fails: a vertex can touch a face, an edge can lie in a face, faces can coincide, and each further contact imposes more conditions. Carried to the end, this suggests a natural family: tori whose triangles have all collapsed into a single plane, which we take to be R2×{0}\mathbb{R}^{2} \times\{0\}. We call these planar tori, write P\mathcal{P} for the space of them, and call the flat ones flattened origami tori.

Planarity gives us enough structure to hold on to. In space, a triangle can rotate about an edge through a whole circle of dihedral angles; in the plane, it is either folded over or not. Once we choose which triangles are folded over, the only remaining condition is that the folding closes up around each vertex and each loop of the torus. These are easy equations in the Euclidean group of the plane, and in § 3 we give explicit solutions for T\mathcal{T} across all of moduli space.

Inflating from the Boundary

We took the equations for planar tori from the boundary ∂E\partial\mathcal{E}, but some solutions may not lie in ∂E\partial\mathcal{E}. A simple sufficient condition for a planar torus Ψ\Psi to lie in ∂E\partial\mathcal{E} is this: If some vertical lift Ψ+z\Psi+ z, giving the vertices heights z∈R8z \in\mathbb{R}^{8}, is embedded, then so are the scaled lifts Ψ+tz\Psi+ tz, which converge to Ψ\Psi as t→0t \to0.

If Ψ\Psi is also flat, it sits in F\mathcal{F} right at the edge of E\mathcal{E}, and origami tori should lie just inside. To reach them, we push Ψ\Psi a little way into E\mathcal{E}, to Ψ+tz\Psi+ tz. This leaves F\mathcal{F}, so we correct by moving the vertices horizontally, parallel to the plane, by some h∈R16h \in\mathbb{R}^{16} that brings us back to F\mathcal{F}. The following result assures us that the correction does not push us back out of E\mathcal{E}.

Lemma 2.1. Let Ψ\Psi be a planar torus such that Ψ+z\Psi+ z is embedded. Then there is η>0\eta> 0 such that Ψ+tz+h\Psi+ tz + h is embedded for every t>0t > 0 and every horizontal motion hh with ∣h∣<η\lvert h\rvert< \eta.

Proof. The vertical scaling (X,Y,Z)↦(X,Y,Z/t)(X,Y,Z) \mapsto(X,Y,Z/t) is an affine bijection of R3\mathbb{R}^{3}, and it sends Ψ+tz+h\Psi+ tz + h to Ψ+z+h\Psi+ z + h. Since embeddedness is open, Ψ+z+h\Psi+ z + h is embedded once ∣h∣\lvert h\rvert is small enough. ♠

This is a lot of room: as t→0t \to0 the lifted torus collapses onto the plane, yet it tolerates horizontal motions of a fixed size. So any horizontal correction that tends to 00 with tt eventually fits inside it. For flatness, the implicit function theorem supplies such a correction whenever horizontal motions can adjust the cone angles at Ψ\Psi. But nothing in this argument is special to cone angles. The same reasoning inflates Ψ\Psi while keeping fixed any quantity that depends smoothly on the vertices, provided horizontal motions can adjust it at Ψ\Psi.

Proposition 2.2 Let Ψ\Psi be a planar torus and Φ\Phi a smooth map from a neighborhood of Ψ\Psi to Rm\mathbb{R}^{m}. Suppose that

  1. Ψ\Psi has an embedded vertical lift Ψ+z\Psi+z;

  1. Φ\Phi restricted to P\mathcal{P} is a submersion at Ψ\Psi.

Then there are embedded tori arbitrarily close to Ψ\Psi with the same value of Φ\Phi.

Proof. By hypothesis 2 there are horizontal vectors e1,…,eme_1,\ldots,e_m with dΦΨ(ei)\mathrm{d}\Phi_{\Psi}(e_i) the standard basis of Rm\mathbb{R}^{m}. Let

G(t,ϵ)=Φ(Ψ+tz+ϵ1e1+⋯+ϵmem).G(t,\epsilon)=\Phi(\Psi+tz+\epsilon_1e_1+\cdots+\epsilon_me_m).

Then G(0,0)=Φ(Ψ)G(0,0)=\Phi(\Psi) and ∂G/∂ϵ\partial G/\partial\epsilon is the identity at (0,0)(0,0), so the implicit function theorem gives a smooth ϵ(t)\epsilon(t) with ϵ(0)=0\epsilon(0)=0 and G(t,ϵ(t))=Φ(Ψ)G(t,\epsilon(t))=\Phi(\Psi) for small tt. The horizontal motion h(t)=∑ϵi(t)eih(t)=\sum\epsilon_i(t)e_i tends to 00 with tt, so by Lemma 2.1 the torus Ψ+tz+h(t)\Psi+tz+h(t) is embedded for small t>0t>0, and it has the same value of Φ\Phi as Ψ\Psi. ♠

Proposition 2.2 provides easily checkable conditions that let us promote a planar torus with the properties we want to an inflated torus with the same properties. In this paper we want origami tori whose modulus we know exactly. So we must control both the cone angles and the modulus, and in §5 we take Φ\Phi to record both. This makes Proposition 2.2 the engine of the paper, and the rest of the paper checks its hypotheses.

The Flattened Tori

The Families

Universality requires an origami torus for every lattice. Scaling in R3\mathbb{R}^{3} takes origami tori to origami tori, and a lattice Λ\Lambda and its mirror image Λ‾\overline{\Lambda} give isometric tori, so it is enough to realize one lattice in each class up to scaling and reflection. Every such class contains Λζ=Z+Zζ\Lambda_{\zeta} = \mathbb{Z} + \mathbb{Z}\zeta for some ζ\zeta in

M={x+iy:0≤x≤12, x2+y2≥1},\mathcal{M} = \{x + iy : 0 \le x \le\frac{1}{2},\ x^{2} + y^{2} \ge1\},

the closed (2,3,∞)(2, 3, \infty) triangle in H2\mathbb{H}^{2}, which is half of the standard fundamental domain for the modular group. So for each ζ∈M\zeta\in\mathcal{M} we need a flattened origami torus with lattice Λζ\Lambda_{\zeta} satisfying the hypotheses of Proposition 2.2.

Since M\mathcal{M} is a continuum, we work with families of flattened tori Ψ(ζ)\Psi(\zeta) with lattice Λζ\Lambda_{\zeta}, depending smoothly on ζ\zeta; along a family, each hypothesis of Proposition 2.2 becomes a statement about functions of ζ\zeta, checked over a whole region at once. A few families whose regions cover M\mathcal{M} then suffice.

Our families come from fixing how the torus folds. When a torus is folded into the plane, each triangle lands either face up or face down, so every planar torus has a folding pattern, recording which triangles are reversed. For a fixed pattern, the flattened origami tori are found by solving the equations that make the folding close up. For the Universality Theorem we need only a single folding pattern, shown in Figure 3.1. We spoil the story and give the formulas right away. They form a 5-parameter family of flattened origami tori, with parameters ζ=x+iy\zeta= x + iy and u,k,λu, k, \lambda. Fixing u,k,λu, k, \lambda as functions of ζ\zeta in two ways then gives two families, which together cover M\mathcal{M}.

Figure 3.1. The square torus and its flattened origami realization.

Each member has two planar pictures: the flat torus before folding, described by vertices wjw_j and their lattice translates, and its folded image, described by vertices pjp_j. The folding maps corresponding triangles isometrically. We begin with the folded vertices.

The Flattened Origami Torus: Here are the points of the corresponding flattened origami torus Ψ\Psi. With the auxiliary variable b=1−x−2kb = 1 - x - 2k, they are

p0=(0,0),p1=(u,u(b+2k/λ)y),p2=(u−12,(1−2u)x2y),p3=(−k−bλ,λ(y2−b2)−2bk2y),p4=(x+k−12,y2+bx2y),p5=(u−12,y2+b(b+2u−1)2y),p6=(u−k,b(u−k)y),p7=(u,y2+b2+2bu2y−2u−2x−2k+12y(1−λ)).\begin{aligned} p_0 &= (0,0), & p_1 &= \left(u,\frac{u(b+2k/\lambda)}{y}\right),\\ p_2 &= \left(u-\frac{1}{2},\frac{(1-2u)x}{2y}\right), & p_3 &= \left(-k-b\lambda,\frac{\lambda(y^2-b^2)-2bk}{2y}\right),\\ p_4 &= \left(x+k-\frac{1}{2},\frac{y^2+bx}{2y}\right), & p_5 &= \left(u-\frac{1}{2},\frac{y^2+b(b+2u-1)}{2y}\right),\\ p_6 &= \left(u-k,\frac{b(u-k)}{y}\right), & p_7 &= \left(u,\frac{y^2+b^2+2bu}{2y}-\frac{2u-2x-2k+1}{2y(1-\lambda)}\right). \end{aligned}

The Flat Torus: We will describe 16 triangles. We get a tiling of R2\mathbb{R}^2 by taking the orbit of these triangles under the action of the lattice Λζ\Lambda_\zeta defined above. The quotient then gives us our 8-vertex triangulation of the flat torus T=R2/ΛζT = \mathbb{R}^2/\Lambda_\zeta. Here are the triangles

012(w0,w1,w2)167(w1,w6,w7)031(w0,w3,w1)264(w2+e,w6,w4+e−τ)024(w0,w2,w4−τ)257(w2+e,w5+e,w7)043(w0,w4−τ,w3)276(w2+e,w7,w6)152(w1,w5,w2)347(w3+τ,w4,w7)136(w1,w3,w6)356(w3,w5+e−τ,w6)145(w1,w4,w5)375(w3+τ,w7,w5+e)174(w1,w7,w4)465(w4+e−τ,w6,w5+e−τ)(1)\begin{aligned} 012\quad&(w_0,w_1,w_2) & 167\quad&(w_1,w_6,w_7)\\ 031\quad&(w_0,w_3,w_1) & 264\quad&(w_2+e,w_6,w_4+e-\tau)\\ 024\quad&(w_0,w_2,w_4-\tau) & 257\quad&(w_2+e,w_5+e,w_7)\\ 043\quad&(w_0,w_4-\tau,w_3) & 276\quad&(w_2+e,w_7,w_6)\\ 152\quad&(w_1,w_5,w_2) & 347\quad&(w_3+\tau,w_4,w_7)\\ 136\quad&(w_1,w_3,w_6) & 356\quad&(w_3,w_5+e-\tau,w_6)\\ 145\quad&(w_1,w_4,w_5) & 375\quad&(w_3+\tau,w_7,w_5+e)\\ 174\quad&(w_1,w_7,w_4) & 465\quad&(w_4+e-\tau,w_6,w_5+e-\tau) \tag*{(1)} \end{aligned}

Here e=(1,0)e=(1,0) and τ=(x,y)\tau=(x,y). We have wj=pjw_j=p_j for j≠3,6j\ne3,6, and

w3=(k,−2bk+λ(b2+y2)2y),w6=(u+k,b(u−k)y).w_3=\left(k,-\frac{2bk+\lambda(b^2+y^2)}{2y}\right),\qquad w_6=\left(u+k,\frac{b(u-k)}{y}\right).

Preliminary Isometric Mapping: Let WjW_j denote the jjth triangle above. We call the set {W0,…,W15}\{W_0,\ldots,W_{15}\} the fundamental triangles. Let PjP_j denote the corresponding triangle whose vertices are the pp points. Let ϕ:⋃Wj→⋃Pj\phi:\bigcup W_j\to\bigcup P_j be the piecewise affine map from the one set of 16 triangles to the other.

Lemma 3.1 ϕ\phi is an isometry on each of the 16 triangles.

Proof: For each of the 16 triangles, one checks that the three squared edge lengths of WjW_j and of PjP_j agree; these are rational functions of xx, yy, uu, kk, λ\lambda, so this is a direct symbolic computation. We do this exactly with the file UniversalityProof. ♠

Remark 3.2 For ζ∈M\zeta\in\mathcal{M}, there are some algebraic constraints on the fiber (u,k,λ)(u,k,\lambda) which suffice to ensure that the triangles above have positive area:

u, k, x+k−u, λ/2∈(0,1/2),u−x−k+1/21−λ+2kuλ<b+2k(x+k).(2)u,\ k,\ x+k-u,\ \lambda/2 \in(0,1/2), \qquad\frac{u-x-k+1/2}{1-\lambda}+\frac{2ku}{\lambda}<b+2k(x+k). \tag*{(2)}

We will never need to check or work with these constraints.

Remark 3.3 By the constraint x+k−u∈(0,12)x+k-u \in(0,\frac{1}{2}), k>uk>u when x=0x=0 and k<uk<u when x=12x=\frac{1}{2}. So a continuous family over all of M\mathcal{M} must pass through k=uk=u, where p0=p6p_0=p_6, and there it seems impossible to add heights that give an embedded lift. This is why we cover M\mathcal{M} with two families.

The Two Families: We now pick out two subfamilies, parametrized by the modulus ζ=x+iy\zeta=x+iy, whose domains together cover M\mathcal{M}:

Domain 1 ={ζ∈M:x≤1/6}=\{\zeta\in\mathcal{M}: x \le1/6\}, Family 1: (u,k,λ)=(3/128,11/32,13/32)(u,k,\lambda)=(3/128,11/32,13/32),

Domain 2 ={ζ∈M:x≥1/6}=\{\zeta\in\mathcal{M}: x \ge1/6\}, Family 2: (u,k,λ)=(x/2−1/2048,1/16,1/2)(u,k,\lambda)=(x/2-1/2048,1/16,1/2).

Both formulas make sense for ζ\zeta in an open neighborhood of their domains: the denominators never vanish there.

These choices of family give the vertices above simple coordinates. Every first coordinate is linear in xx, and every second coordinate is a rational function (f(x)+cy2)/y(f(x)+cy^2)/y, where ff is at most quadratic and cc is constant. This greatly simplifies the polynomials in our checks: the triangle areas (Lemma 3.4), the volumes of § 4 (Lemma 4.1), and the Jacobian (Lemma 5.3).

In § 3.2 we check that the members of the two families are flattened origami tori; this is all the proof needs. In § 3.3 we show how one can derive these formulas from the folding pattern.

Checking the Families Are Flattened Origami Tori

The left side of Figure 3.1 shows the member of Family 1 at ζ=i\zeta= i. The right side shows the flattened origami torus built one triangle at a time, in an ordering that is compatible with our corresponding lift. Figure 3.2 does the same thing for the somewhat randomly chosen member ζ=8/17+i\zeta= 8/17 + i of Family 2. (It would be more mathematically satisfying to show the picture for the hex torus, but the picture does not look that good: two of the vertices are quite close together.) The coloring convention in both figures is as follows. Fixing an orientation of TT, light triangles keep their orientation and dark ones are reversed. Read down the left column of the table in § 3.1 and then the right, with + for kept and − for reversed, our folding pattern is + − − + + + + + − + − − − − +−.

Figure 3.2. The torus of modulus 8/17+i8/17 + i and its flattened origami realization.

In both figures, the sixteen triangles form a polygon whose boundary edges are paired by the lattice: ee pairs the edges 42 and 25 on the left with those on the right, τ\tau pairs 43 and 35 on the bottom with those on the top, and e−τe-\tau pairs the two edges 45. The table shows that these pairings hold for every member, not only the two pictured: for instance, triangles 145 and 465 share the edge 45, shifted by e−τe-\tau.

Lemma 3.4 Let d=1,2d=1,2.

  1. For all parameters in a suitably chosen open neighborhood of Domain dd, every triangle WjW_j in Family dd has positive area.

  1. Throughout a connected open neighborhood UdU_d of Domain dd, ϕ\phi gives a flattened origami torus ϕ:T→Ψ\phi:T\to\Psi.

Proof: (1) Twice the signed area of a triangle is the 2×22 \times2 determinant of two of its edge vectors, so we must show that these sixteen determinants do not vanish. UniversalityProof certifies this exactly via strip dominance (§ 7), on each family’s domain.

One may also check this by hand. As the vertex coordinates are simple (§ 3.1), so are these determinants: each has the form A(x)/y+B(x)yA(x)/y+B(x)y, so yy times it is A(x)+B(x)y2A(x)+B(x)y^2, linear in y2y^2. Here Lemma 7.1 reduces each triangle to a pair of sign checks, for two at most quadratic polynomials on the family’s domain, and all of them hold. This open condition persists in an open neighborhood of Domain dd.

(2) Let Π=⋃Wj\Pi=\bigcup W_j and Ψ=⋃Pj\Psi=\bigcup P_j. We have a map ϕ:Π→Ψ\phi:\Pi\to\Psi which is an isometry on each triangle, by Lemma 3.1. For each of the figures we plotted above, Π\Pi is an embedded solid polygon and the three elements e,τ,e−τe,\tau,e-\tau of Λζ\Lambda_\zeta pair the sides of Π\Pi. Let TT be the quotient of Π\Pi by this side-pairing. TT is a topological torus. Because the side-pairings are translations, we have the familiar result that TT is a flat torus. This picture persists for an open set of parameters in Domain dd, for d=1,2d=1,2.

We let UdU_d be some connected open neighborhood of Domain dd for which (1) holds. Then (1) tells us that the triangles retain their orientations and never degenerate in UdU_d. Also, all the angles vary analytically. Hence, if we form the abstract topological space by gluing together the triangles in the same pattern as we do when we know that Π\Pi is embedded, we get a topological torus whose cone angles vary analytically. Since these cone angles equal 2π2\pi on an open set, they are always 2π2\pi. Hence, our torus TT is always a flat torus. (The expert reader may enjoy proving the unnecessary fact that Π\Pi is always an embedded polygon and a fundamental domain for Λζ\Lambda_\zeta.) Furthermore, ϕ\phi respects all the gluings needed to make TT and therefore gives us a flattened origami torus ϕ:T→Ψ\phi:T\to\Psi throughout UdU_d for each choice of dd. ♠\spadesuit

Producing Families by Folding

In this subsection, which is not needed for the proof, we derive the formulas in § 3.1 from our folding pattern. There are very many folding patterns, 2162^{16} for each triangulation, and ours has several properties that let the proof go through; we explain how we found it in § 6.

Let Ψ=ϕ(T)\Psi=\phi(T) be any flattened origami torus with our folding pattern, where ϕ\phi is isometric on each triangle of its intrinsic flat torus TT. An edge is a crease if its two triangles have opposite signs. Two triangles of the same sign sharing an edge are moved by the same isometry. So if we cut the universal cover R2\mathbb{R}^2 of TT along the creases into islands, then ϕ\phi, lifted to R2\mathbb{R}^2, moves each island by a single isometry, and the isometries of two islands meeting along a crease differ by the reflection in it. For our folding pattern the creases cut R2\mathbb{R}^2 into ten islands A,…,JA,\ldots,J, up to translation by the lattice.

Figure 3.3 below shows these islands, drawn in the universal cover and shaded according to their orientation as in Figure 2.1 and 2.2. More precisely, the folding keeps the orientation of the light triangles and reverses the orientation of the dark ones. Our subtle shading in the figure is designed to distinguish the 16 fundamental triangles from their images under the lattice action. We also note one additional bookkeeping feature of Figure 3.3: A shaded vertex labeled vv is the translate of wvw_v by the indicated vector.

Inflating from the boundary

Figure 2.1. Inflating from the boundary.

The islands and the sixteen triangles, drawn in the universal cover

Figure 3.3. The islands and the sixteen triangles, drawn in the universal cover.

Our aim is to find all flattened origami tori with this folding pattern, up to similarity. Knowing how ϕ\phi moves one island determines how it moves all the others, by crossing creases, so rather than start with TT and look for ϕ\phi, we build both together, one island at a time. Normalize the lattice of TT to be generated by e=(1,0)e=(1,0) and τ=(x,y)\tau=(x,y) with y>0y>0, and put w0=0w_0=0. Then TT is determined by lifts wj∈R2w_j\in\mathbb{R}^2 of its vertices, which are our unknowns. The table in § 3.1 lists the sixteen triangles with the lifts of their vertices.

Composing with an isometry of the plane, we may take ϕ\phi to be the identity on the island AA, and we write pj=ϕ(wj)p_j=\phi(w_j) for the vertices of Ψ\Psi. Crossing creases then determines ϕ\phi everywhere, and the only requirement is that it be well defined on TT: all copies of a vertex in R2\mathbb{R}^2 must go to the same point. These closing conditions are the only equations to solve, and they are easy. Each says that a composition of reflections in creases sends one copy PP of a vertex to another copy QQ. When it is a single reflection, the crease is the perpendicular bisector of PP and QQ, so each point cc on it satisfies ∣c−P∣2=∣c−Q∣2\lvert c-P\rvert^2=\lvert c-Q\rvert^2, which is linear in cc. For our folding pattern the conditions can be ordered so that each is linear in the unknowns not yet determined. So we can solve them one at a time, and the solutions form a family, rational in xx, yy and the choices left free.

We illustrate, through two examples, how the closing conditions are solved, starting from AA. The first example shows both how a condition determines unknowns and how a free parameter arises. Since ϕ\phi is the identity on AA, pj=wjp_j = w_j for its vertices 0, 1, 2, 4, 5, 7. The island CC meets AA along the crease 17, so ϕ\phi moves CC by the reflection in 17 (Figure 3.4). CC contains the copies w2+ew_2 + e and w5+ew_5 + e of the vertices 2 and 5, while AA contains w2w_2 and w5w_5, so this reflection sends w2+ew_2 + e to w2w_2 and w5+ew_5 + e to w5w_5. Thus 17 is the perpendicular bisector of w2w_2 and w2+ew_2 + e, a vertical line X=uX = u, and nothing determines uu: it is our first free choice. It places w1w_1 and w7w_7 on X=uX = u, and w2w_2 and w5w_5 on X=u−12X = u - \frac{1}{2}.

A, B, and C unfolded; folding C across 17; folding B across 02

Figure 3.4. (a) AA, BB, CC, unfolded. (b) Fold CC across 17. (c) Fold BB across 02.

Crossing into BB along the crease 02 works the same way (Figure 3.4c): the reflection in 02 must send w4−τw_4 - \tau to w4w_4, so 02, which passes through w0=0w_0 = 0, is perpendicular to τ\tau, and this pins down w2w_2 on X=u−12X = u - \frac{1}{2}. The rest goes the same way once w3w_3 is placed freely, which gives our other two choices: its first coordinate kk and its second coordinate, which we parametrize by λ\lambda. Comparing EE with DD then fixes w4w_4 and w5w_5, comparing EE with CC fixes w6w_6, the islands FF and GG fix w1w_1 and w7w_7, and HH, II and JJ add no new equations. Three free choices is what one expects: up to similarity, eight points in the plane have 12 degrees of freedom, and flatness imposes 7 independent conditions, leaving the modulus and three more.

Working out all the closing conditions in this way led to the formulas of §3.1 for the vertices pj=ϕ(wj)p_j = \phi(w_j) of Ψ\Psi. The flat torus TT is given by the lifts wjw_j. Since ϕ\phi is the identity on AA, wj=pjw_j = p_j for the vertices 0, 1, 2, 4, 5, 7 of AA. The other two vertices lie on islands moved by a single reflection: w3w_3 on FF, reflected in the crease 01, and w6w_6 on CC, reflected in the crease 17. This gives the formulas for w3w_3 and w6w_6.

The Lifting Property

Hypothesis 1 of Proposition 2.2 asks for an embedded vertical lift. For each family we propose one set of heights z=(z0,…,z7)z=(z_0,\ldots,z_7) for every member:

z0z_0z1z_1z2z_2z3z_3z4z_4z5z_5z6z_6z7z_7
Family 100−11-11−1-1−3-3−15-150977
Family 20−2017-20170−2-2−1-1−3-301430

If Ψ\Psi has vertices pjp_j, the vertical lift Ψ+z\Psi+z has vertices p^j=(pj,zj)\widehat{p}_j=(p_j,z_j). To see that it is embedded, we must check that no two of its triangles meet except along a common edge or vertex. We do this in two steps: we check that the lift is embedded at a single modulus, and then that nothing changes across the family.

Throughout we use tetrahedra. Any four vertices span a tetrahedron, and the sign of its volume records their relative position: on which side of the plane through three of them the fourth one lies. We first show that none of these 70 volumes vanishes on the domain, so their signs are constant across the family. We then explain how the signs certify embeddedness.

Lemma 4.1 For both families and every modulus in the domain, no four vertices of the lift Ψ+z\Psi+z are coplanar, that is, they are in general position.

Proof: For each quadruple α,β,γ,δ\alpha,\beta,\gamma,\delta of vertices we must show that

V=det⁡(p^α−p^δ,p^β−p^δ,p^γ−p^δ)V=\det(\widehat{p}_\alpha-\widehat{p}_\delta,\widehat{p}_\beta-\widehat{p}_\delta,\widehat{p}_\gamma-\widehat{p}_\delta)

does not vanish on Domain dd. This is certified in UniversalityProof using the tweaked method of strip dominance from § 7.

One may also check this by hand. The columns of VV are differences of XX-, YY- and ZZ-coordinates, so by § 3.1, multiplying the middle column by yy and expanding along it gives

yV=A(x)+B(x)y2.yV=A(x)+B(x)y^2.

For each of the 70 determinants, Lemma 7.1 then reduces nonvanishing to the signs of A+B(1−x2)A+B(1-x^2) and BB, two polynomials of degree at most two, on the family’s domain. All these checks hold. ♠

Lemma 4.2 Suppose that the vertices of a polyhedral torus are in general position, and that every edge is disjoint from every triangle with which it shares no vertices. Then the torus is embedded.

Proof: We show that any two triangles Δ1\Delta_1 and Δ2\Delta_2 intersect precisely in their common vertices and edge. If Δ1\Delta_1 and Δ2\Delta_2 share an edge, their planes are distinct by general position. Hence they intersect exactly along their common edge. Now suppose Δ1\Delta_1 and Δ2\Delta_2 do not share an edge. If they have an unwanted intersection, some edge of one triangle must intersect the other triangle outside their common vertices. An edge having a vertex in common with the other triangle cannot do this: by general position, it intersects the plane of that triangle only at the common vertex. Thus the unwanted intersection must involve an edge and a triangle with no vertices in common. This contradicts our hypothesis. ♠

Next we test whether a segment (a0a1)(a_0a_1) is disjoint from a triangle (a2a3a4)(a_2a_3a_4). Deleting aka_k from the list leaves the vertices of a tetrahedron. Let VkV_k be its signed volume, with the points taken in increasing order. We assume the five points are in general position.

Lemma 4.3 (01)(01) and (234)(234) are disjoint provided max⁡(V0V1,V2V3,V3V4)>0\max(V_0V_1, V_2V_3, V_3V_4) > 0.

Proof: Let [abcd][abcd] denote the signed volume of the tetrahedron with vertices listed in the order a,b,c,da,b,c,d. If V1=[0234]V_1 = [0234] and V0=[1234]V_0 = [1234] have the same sign, then (01)(01) lies strictly on one side of the plane containing (234)(234). This gives disjointness. If (01)(01) intersects (234)(234), then [0123][0123], [0134][0134] and [0142][0142] have the same sign. Since [0123]=V4[0123] = V_4, [0134]=V2[0134] = V_2 and [0142]=−V3[0142] = -V_3, this fails if V2V3>0V_2V_3 > 0 or V3V4>0V_3V_4 > 0. ♠

Corollary 4.4 For both families and every modulus in the domain, the lift Ψ+z\Psi+z is embedded.

Proof: There are 138 pairs (e,Δ)(e,\Delta) where ee is an edge and Δ\Delta a triangle of T\mathcal{T} with no common vertex. By Lemmas 4.1 and 4.2, Ψ+z\Psi+z is embedded if all of them satisfy the criterion of Lemma 4.3. Using the file UniversalityProof we check this at ζ=i\zeta= i in Family 1 and at ζ=8/17+i\zeta= 8/17 + i in Family 2 (Figure 3.1 and Figure 3.2). Each VkV_k in the criterion is the volume of four lifted vertices, so by Lemma 4.1 it never vanishes on the domain. Since the domain is connected, its sign is constant there, and the criterion holds throughout. ♠

Remark 4.5 Constant heights that work for a whole family are not automatic: we searched for them alongside the choice of family. At a single modulus the checks above take a computer almost no time, so candidate heights are easy to test at many moduli. Even so, they are delicate: changing a single height by ±1\pm1 usually breaks the argument. See §6.

Universality

The Map Φ\Phi

By §4, each of our flattened origami tori has an embedded vertical lift, so it can be inflated. Proposition 2.2 assures us we can keep the values of any smooth map Φ\Phi fixed during inflation, provided it is a submersion on the planar tori at our starting point. So we build a Φ\Phi to suit our purpose, recording both the cone angles and the modulus of polyhedral tori.

For a polyhedral torus, let θv\theta_v be the cone angle at the vertex vv, the sum of the angles of the triangles meeting there, and let δv=2π−θv\delta_v = 2\pi- \theta_v be its angle deficit. Each angle of a nondegenerate triangle is a smooth function of its edge lengths, so the δv\delta_v are smooth functions on the space of polyhedral tori. By Gauss–Bonnet, the eight deficits of a torus sum to zero. So a polyhedral torus is flat exactly when δ0,…,δ6\delta_0,\ldots,\delta_6 vanish, and we record these seven deficits as δ=(δ0,…,δ6)\delta= (\delta_0,\ldots,\delta_6).

There is one subtlety: the modulus is only defined for flat tori, but Proposition 2.2 needs Φ\Phi on a neighborhood of Ψ(ζ)\Psi(\zeta), which contains tori that are not flat. The modulus of a flat torus is read off from its development into the plane: the holonomy along two generating loops is a pair of translations, and the modulus is their ratio. Other tori are cone surfaces, and technically the developing map is only defined on the universal cover of the punctured surface we get by removing the cone points. However, we can define a kind of pseudo modulus, sufficient for our purposes, by choosing (once and for all) flat annuli bounding each generating loop and then defining the holonomy relative to these annuli.

Lemma 5.1 Developing along a fixed choice of annuli defines a smooth function μ~\widetilde{\mu} of the edge lengths, on an open set of polyhedral tori containing F\mathcal{F}, which equals the modulus on F\mathcal{F}.

Proof. When we lift these annuli to the universal cover of T\mathcal{T} we get strips along which we can properly define the developing map. In the universal cover, choose a vertex vv, a triangle Δ\Delta containing vv, and their translates v1v_1, Δ1\Delta_1 by ee and v2v_2, Δ2\Delta_2 by τ\tau. For each jj, choose a strip of triangles from Δ\Delta to Δj\Delta_j, each sharing an edge with the next (Figure 5.1 shows our choice).

Figure 5.1. The modulus defined by strip development.

Given a polyhedral torus, develop each strip into the plane. Place Δ\Delta by a fixed rule, for instance with vv at the origin, a fixed edge of Δ\Delta at vv along the positive YY-axis, and Δ\Delta on the side given by the orientation of the torus. Then attach each subsequent triangle along its shared edge, on the opposite side. Each placement is determined by the edge lengths and depends smoothly on them, because the triangles are nondegenerate. So, identifying R2\mathbb{R}^{2} with C\mathbb{C},

μ~=dev⁡2(v2)−dev⁡2(v)dev⁡1(v1)−dev⁡1(v)\widetilde{\mu}=\frac{\operatorname{dev}_{2}(v_{2})-\operatorname{dev}_{2}(v)}{\operatorname{dev}_{1}(v_{1})-\operatorname{dev}_{1}(v)}

is a smooth function of the edge lengths wherever the denominator is nonzero. When the torus is flat, the two strips are part of its development, so the differences are its holonomy translations, up to a common rotation, and μ~\widetilde{\mu} is the modulus. In particular the denominator is nonzero on F\mathcal{F}, and so on an open set containing it, where μ~\widetilde{\mu} is the required function. ♠\spadesuit

Figure 5.1 shows our choice of strips, and also illustrates our construction of this pseudo modulus. The strips for T(46666677)T(46666677), with vv at vertex 1 and Δ=167\Delta=167 (dark), and the edge 17 vertical:

167,276,257,152,145,174,167167,276,257,152,145,174,167

from Δ\Delta to Δ1\Delta_{1}, and

167,276,257,375,356,136,167167,276,257,375,356,136,167

from Δ\Delta to Δ2\Delta_{2}. They are developed for the square torus (left) and for a nearby torus that is not flat (right).

Set Φ=(μ~,δ)\Phi=(\widetilde{\mu},\delta), a smooth map from a neighborhood of each Ψ(ζ)\Psi(\zeta) to C×R7\mathbb{C}\times\mathbb{R}^{7}. Near Ψ(ζ)\Psi(\zeta), the tori with Φ=(ζ,0)\Phi=(\zeta,0) are exactly the flat tori of modulus ζ\zeta: δ=0\delta=0 means flat, and on flat tori μ~\widetilde{\mu} is the modulus.

To apply Proposition 2.2 we must show that Φ\Phi restricted to P\mathcal{P} is a submersion at Ψ(ζ)\Psi(\zeta). Along our families the modulus part is automatic, so this reduces to a statement about δ\delta alone:

Lemma 5.2 Let Ψ(ξ)\Psi(\xi) be a smooth family of flattened origami tori of modulus ξ\xi, for ξ\xi in an open set U⊂H2U \subset\mathbb{H}^{2} containing ζ\zeta. If δ\delta restricted to P\mathcal{P} is a submersion at Ψ(ζ)\Psi(\zeta), then so is Φ\Phi restricted to P\mathcal{P}.

Proof: Since each Ψ(ξ)\Psi(\xi) is planar, the velocities f1=∂Ψ/∂xf_1 = \partial\Psi/\partial x and f2=∂Ψ/∂yf_2 = \partial\Psi/\partial y at ζ\zeta are horizontal. Since δ(Ψ(ξ))=0\delta(\Psi(\xi)) = 0 and μ~(Ψ(ξ))=ξ\widetilde{\mu}(\Psi(\xi)) = \xi, we have

dδ(fj)=0,dμ~(f1)=1,μ~(f2)=i.d\delta(f_j)=0,\qquad d\widetilde{\mu}(f_1)=1,\qquad\widetilde{\mu}(f_2)=i.

By hypothesis there are horizontal vectors e1,…,e7e_1,\ldots,e_7 with dδ(ei)d\delta(e_i) the standard basis of R7\mathbb{R}^7. On these nine vectors the derivative of Φ\Phi has the block form

(I2B0I7),\begin{pmatrix} I_2 & B \\ 0 & I_7 \end{pmatrix},

which is invertible. So dΦd\Phi maps the horizontal vectors onto C×R7\mathbb{C}\times\mathbb{R}^7. ♠\spadesuit

Completing the Proof

By Lemma 5.2, it remains to show that δ\delta restricted to P\mathcal{P} is a submersion at each Ψ(ζ)\Psi(\zeta). In fact the first coordinates of vertices 0,…,60,\ldots,6 already suffice. Given ϵ∈R7\epsilon\in\mathbb{R}^7, let Ψ(ϵ)\Psi(\epsilon) move each vertex pip_i of Ψ\Psi by (ϵi,0)(\epsilon_i,0) for i=0,…,6i=0,\ldots,6, leaving vertex 7 fixed, and for Family dd set

Jd=det⁡∂δ∂ϵ∣ϵ=0.J_d=\det\left.\frac{\partial\delta}{\partial\epsilon}\right|_{\epsilon=0}.

We show that JdJ_d does not vanish.

Writing ζ=x+iy\zeta=x+iy, JdJ_d is a rational function of xx and yy, even though the angles themselves are not. The angle α\alpha between vectors V=(v1,v2)V=(v_1,v_2) and W=(w1,w2)W=(w_1,w_2) is

α=arccos⁡(V⋅W∥V∥∥W∥).\alpha=\arccos\left(\frac{V\cdot W}{\lVert V\rVert\lVert W\rVert}\right).

This formula is far from rational, but its derivatives are, for example

∂α∂v1=σv2v12+v22,σ=sign⁡det⁡(V,W).\frac{\partial\alpha}{\partial v_1}=\sigma\frac{v_2}{v_1^2+v_2^2},\qquad\sigma=\operatorname{sign}\det(V,W).

Taking VV and WW in the order given by the orientation of the torus, σ\sigma is the sign of the triangle in our folding pattern. The vertex coordinates of our flattened tori are rational functions of xx and yy, so each such derivative is too. The entries of ∂δ/∂ϵ\partial\delta/\partial\epsilon are sums of these derivatives, and JdJ_d is a polynomial in the entries, so JdJ_d is rational in xx and yy. Its denominators are squared edge lengths v12+v22v_1^2+v_2^2, which are positive because our triangles are nondegenerate (Lemma 3.4), so JdJ_d is defined on all of Domain dd.

Lemma 5.3. The function JdJ_d does not vanish on Domain dd.

Proof. This is certified in UniversalityProof using the tweaked method of strip dominance from § 7. Alternatively, with a computer algebra system one may verify that the numerator of JdJ_d in lowest terms factors as

Cdy13bd(Ad(x)+Bd(x)y2)C_d y^{13} b_d\left(A_d(x)+B_d(x)y^2\right)

where CdC_d is a nonzero constant, b1=516−xb_1=\frac{5}{16}-x, b2=78−xb_2=\frac{7}{8}-x, and AdA_d and BdB_d are polynomials of degree 5 and 3. For example, for Family 1,

A1(x)=13791139987456x5−29311757123584x4−347158265856x3+16581166028288x2−5729028705984x+420099304545,A_1(x)=13791139987456x^5-29311757123584x^4-347158265856x^3 +16581166028288x^2-5729028705984x+420099304545,
B1(x)=−13791139987456x3+20692294631424x2+16351578693632x+715752410880.B_1(x)=-13791139987456x^3+20692294631424x^2+16351578693632x+715752410880.

On Domain dd, both yy and bdb_d are positive. By Lemma 7.1, the remaining factor is nonzero because the quintic Ad+Bd(1−x2)A_d+B_d(1-x^2) and the cubic BdB_d are both positive for Family 1 and both negative for Family 2. ♠

This completes the proof of the Universality Theorem. Indeed, let ζ∈M\zeta\in\mathcal{M}. Since our two domains cover M\mathcal{M}, ζ\zeta lies in one of them, say Domain dd, and Ψ(ζ)\Psi(\zeta), the flattened origami torus of Family dd with modulus ζ\zeta, satisfies both hypotheses of Proposition 2.2. Hypothesis 1 holds because its lift is embedded (Corollary 4.4). Hypothesis 2 holds by Lemma 5.2 applied to Family dd, since δ\delta restricted to PP is a submersion at Ψ(ζ)\Psi(\zeta) by Lemma 5.3. So by Proposition 2.2 there are embedded tori near Ψ(ζ)\Psi(\zeta) with Φ=(ζ,0)\Phi=(\zeta,0), that is, origami tori modeled on T\mathcal{T} with modulus ζ\zeta. As this holds for every ζ∈M\zeta\in\mathcal{M}, T\mathcal{T} is universal by § 3.1.

Remark 5.4. These same computations serve a more geometric proof, which needs no smooth extension of the modulus. By Lemma 5.3, near our family the flat tori form a 17-dimensional manifold F\mathcal{F}, and since μ(Ψ(ξ))=ξ\mu(\Psi(\xi))=\xi, the modulus μ:F→H2\mu:\mathcal{F}\to\mathbb{H}^2 is a submersion there, with 15-dimensional fibers. Vertical motions of a planar torus change the edge lengths only to second order, so the lift Ψ+tz\Psi+ tz is tangent to the fiber through Ψ(ζ)\Psi(\zeta), and stays within O(t2)O(t^2) of it, while the embeddedness determinants of § 4 grow linearly in tt. So a curve in the fiber leaving Ψ(ζ)\Psi(\zeta) with velocity zz differs from the lift by O(t2)O(t^2), and is embedded for small t>0t > 0: an origami torus of modulus ζ\zeta.

How We Found the Family

As described in § 1.2, our family grew out of numerical experiments by F.L. and S.T., which searched all seven 8-vertex triangulations and found many approximate origami tori. F.L. took tori with moduli very near the square and hexagonal lattices, drove them numerically onto those lattices, and then down into the plane, keeping them flat with fixed modulus. This did not always work, but when it did, it produced flattened tori with simple incidences; on the hexagonal one, for instance, the vertices 0, 2, 4, 5, 6 lie on a circle [2026 §6.1]. Imposing these incidences as exact equations gave the exact flattened tori of [2026].

F.L.’s incidences generalized to families of flattened tori along vertical rays in M\mathcal{M}. Trying to explain the incidences, we found their source: they are what it takes for the folding to close up in the Euclidean group of the plane. Solving these equations step by step, as in § 3.3, gives the full family of flattened tori with a given folding pattern. The family of F.L.’s square example does not reach the hexagonal lattice, but that of his hexagonal example covers M\mathcal{M}. It is the family of this paper.

Families that cover M\mathcal{M} seem to be rare. Up to symmetry, the seven 8-vertex triangulations have 67,645 folding patterns, and three simple conditions rule out all but 2,940. The signed angles of the triangles around each vertex must sum to 0 or ±2π\pm2\pi, which constrains how many triangles can be face up. No island may wrap around the torus, since ϕ\phi moves it by a single isometry. And every vertex must lie on a crease, since the cone angle at a vertex inside an island cannot change under horizontal motions, so δ\delta restricted to PP cannot be a submersion. For the remaining 2,940 patterns, we attempted to solve the folding equations numerically at many moduli. In the end, only 18 appeared to fold up across all of M\mathcal{M}.

The existence of an embedded vertical lift is also quite restrictive: for universality, even a single modulus where it fails rules a pattern out. At a given flattened torus, the conditions describing embedded lifts cut out an open cone, and when this is nonempty it is usually easy to numerically exhibit a point. So either a pattern lifted quickly, and we moved on to other moduli, or it failed at many hundreds of flattened tori of one modulus, and we discarded it. Only two patterns survived: the pattern of F.L.’s hexagonal torus, on T(46666677)\mathcal{T}(46666677), and one on T(55666677)\mathcal{T}(55666677).

These two were our candidates for universality, and for both we then confirmed everything exactly. To keep the polynomial checks simple, we restrict to 2-dimensional sections over M\mathcal{M}, taking uu, kk, λ\lambda to be functions of ζ\zeta, so that every polynomial is in xx and yy rather than five variables. We chose the sections and the heights together, by trial and error: we tried many candidate sections, and for each searched for heights that work along the whole section. The closer the sections and heights are to constant, the lower the degrees of the polynomials in the final checks.

On T(46666677)\mathcal{T}(46666677) the vertex coordinates are simple (§ 3.1), so the polynomials have low degree and every check reduces to one in a single variable. On T(55666677)\mathcal{T}(55666677) they are much more complicated, and even the simplest sections we know lead to polynomials that are irreducibly multivariable and of high degree: the lift proof rests on the signs of 131 irreducible polynomials in two variables, of degree up to 10 in one and 4 in the other, and the Jacobian has irreducible factors of degrees (14,8)(14, 8) and (22,11)(22, 11), the latter with 220 terms. So we wrote the paper around the simpler one.

Polynomial Checks

Covering M\mathcal{M} with families promoted the pointwise numerical checks required by Proposition 2.2 to familywise polynomial checks: each says that some polynomial in xx and yy does not vanish on a strip M∣I={ζ∈M:x∈I}\mathcal{M}_{\lvert I} = \{\zeta\in\mathcal{M} : x \in I\}. Our two families were chosen so that each check comes down to a polynomial of the form A(x)+B(x)y2A(x) + B(x)y^2 (§ 6), which is linear in y2y^2 along each vertical line, and Lemma 7.1 below reduces it to checks on polynomials in xx alone. We also describe a second method, strip dominance, which is the one used by the file UniversalityProof. It needs none of this structure: it applies directly to the rational functions in our lemmas, without first putting them in this form, and certifies each check exactly.

Lemma 7.1 Let I⊂[0,12]I \subset[0,\frac{1}{2}] be an interval and P(x,y)=A(x)+B(x)y2P(x,y) = A(x) + B(x)y^2, with AA and BB polynomials. Suppose that for every x∈Ix \in I,

  1. A(x)+B(x)(1−x2)≠0A(x) + B(x)(1 - x^2) \ne0, and

  1. B(x)B(x) is either zero or has the same sign as A(x)+B(x)(1−x2)A(x) + B(x)(1 - x^2).

Then PP has no zero on M∣I\mathcal{M}_{\lvert I}.

Proof: Fix x∈Ix \in I. In M∣I\mathcal{M}_{\lvert I} the vertical line through xx is the ray y2≥1−x2y^2 \ge1 - x^2, and along it PP is linear in y2y^2, with slope B(x)B(x). At the bottom of the ray its value is A(x)+B(x)(1−x2)≠0A(x) + B(x)(1 - x^2) \ne0, and the slope is zero or has the same sign, so PP never vanishes on the ray. ♠

We now describe the general-purpose method. Its basic tool is a sign test for polynomials on the region Ω=[0,1]×[0,∞)\Omega= [0,1] \times[0,\infty).

Strip Dominance: Given a polynomial g(t)=a0+⋯+antng(t) = a_0 + \cdots+ a_n t^n we define

⟨g⟩=min⁡(A0,…,An),Ak=a0+⋯+ak.(3)\langle g\rangle= \min(A_0,\ldots,A_n), \qquad A_k = a_0+\cdots+a_k. \tag*{(3)}

Then g>0g > 0 on [0,1][0,1] if ⟨g⟩>0\langle g\rangle> 0, and g≥0g \ge0 on [0,1][0,1] if ⟨g⟩≥0\langle g\rangle\ge0, since

g(t)=(1−t)(A0+A1t+⋯+An−1tn−1)+Antn.(4)g(t) = (1-t)(A_0+A_1t+\cdots+A_{n-1}t^{n-1})+A_nt^n. \tag*{(4)}

We call G(t,h)G(t,h) strip dominant if for one choice of sign

±G(t,h)=g0(t)+g1(t)h+⋯+gm(t)hm,⟨g0⟩>0,⟨g1⟩,…,⟨gm⟩≥0.(5)\pm G(t,h) = g_0(t)+g_1(t)h+\cdots+g_m(t)h^m,\qquad\langle g_0\rangle> 0,\qquad\langle g_1\rangle,\ldots,\langle g_m\rangle\ge0. \tag*{(5)}

A strip dominant GG does not vanish on Ω\Omega. More generally, GG does not vanish on Ω\Omega if G(t/2,h)G(t/2,h) and G(1−t/2,h)G(1-t/2,h) are strip dominant; we call this the tweaked criterion.

Rational Parametrization: To apply this to a strip M∣[x0,x1]\mathcal{M}|_{[x_0,x_1]}, which is bounded below by an arc of the unit circle, we change coordinates to turn it into Ω\Omega. As uu goes from 00 to 11, the point

(x(u),y(u))=(2u1+u2,1−u21+u2)(6)(x(u),y(u))=\left(\frac{2u}{1+u^2},\frac{1-u^2}{1+u^2}\right) \tag*{(6)}

runs along the unit circle from (0,1)(0,1) to (1,0)(1,0), with rational coefficients. We choose rationals r0<r1r_0<r_1 with

x(r0)≤x0<x1≤x(r1),(7)x(r_0)\le x_0<x_1\le x(r_1), \tag*{(7)}

so that the arc from u=r0u=r_0 to u=r1u=r_1 lies over [x0,x1][x_0,x_1], and rescale it to t∈[0,1]t\in[0,1] by u=(1−t)r0+tr1u=(1-t)r_0+tr_1. Then

F(t,h)=(x(u),y(u)+h)(8)F(t,h)=(x(u),y(u)+h) \tag*{(8)}

maps Ω\Omega onto the part of M\mathcal{M} above this arc, which contains the strip. We use r0=0r_0=0 and r1=43/512r_1=43/512 for Family 1, which covers 0≤x≤0.166790\le x\le0.16679, and r0=143/1704r_0=143/1704 and r1=144/512r_1=144/512 for Family 2, which covers 0.1666666≤x≤0.52130.1666666\le x\le0.5213. The choice of 143/1704143/1704 is extremely tight: just left of x=16x=\frac{1}{6}, at x≈0.1666641x\approx0.1666641, the lift of Family 2 stops being embedded.

Certificate: Given a rational function ff, we write f∘F=P/Qf \circ F = P/Q with PP and QQ integer polynomials in tt and hh, and check that each of PP and QQ satisfies the strip dominance criterion or its tweaked version. Then neither vanishes on Ω\Omega, so ff is defined and nonzero on M∣[x0,x1]\mathcal{M}|_{[x_0,x_1]}.

References

  1. [1960][BZ1] Y. D. Burago and V. A. Zalgaller, Polyhedral realizations of developments (in Russian) Vestnik Leningrad Univ. 15, pp 66–80 (1960)
  2. [1978][Br] U. Brehm, Oberwolfach report (1978)
  3. [1995][BZ2] Y. D. Burago and V. A. Zalgaller, Isometric piecewise-linear embeddings of two-dimensional manifolds with a polyhedral metric into R³, Algebra i Analiz 7(3), pp 76–95 (1995). English translation: St. Petersburg Math. J. 7(3), pp 369–385 (1996).
  4. [2016][Se] H. Segerman, Visualizing Mathematics with 3D Printing, Johns Hopkins U. Press (2016)DOI
  5. [2024][LT] F. Lazarus, F. Tallerie, A Universal Triangulation for Flat Tori, Discrete and Computational Geometry 71 (2024) pp 278–307arxiv.org/abs/2203.05496
  6. [2025][DS] P. G. Doyle and R. E. Schwartz, Collapsibility and Near Universality for Vertex Minimal Paper Tori, arXiv:2510.12623 (2025)arxiv.org/abs/2510.12623
  7. [2026][L] F. Lander, The Two Most Symmetric Flat Tori as Eight-Vertex Paper Tori, arXiv:2609.10703 (2026)arxiv.org/abs/2609.10703
  8. [2027][ALM] P. Arnoux, S. Lelièvre, A. Málaga, Diplotori: a family of polyhedral flat tori. in preparation
  9. [2028][S1] R. E. Schwartz, Vertex Minimal Paper Tori, arXiv 2507.14998 (2025)arxiv.org/abs/2507.14998
  10. [2029][S2] R. E. Schwartz, The Most Efficient Origami Torus, P.N.A.S. 123 (21) (2026)DOI

Paper details

vlnc:W_hwab9hgw
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

ChatGPT 5.6

View statement
Looking for related papers…
Contents