The Optimal Paper Moebius Band

arXiv:2308.12641v4 [[Link]] 10 Sep 2023

Richard Evan Schwartz ∗

September 12, 2023

Abstract
In this paper we prove that a smooth√embedded paper Moebius
band must have aspect ratio greater than 3. We also prove that any
sequence of smooth √ embedded paper Moebius bands whose aspect
ratio converges to 3 must converge, up to isometry, to the famous
triangular Moebius band. These results resolve the optimal paper
Moebius band conjecture of Halpern and Weaver from 1977.

1 Introduction
Here is how to make a paper Moebius band: Take a unit width strip of paper,
give it an odd number of twists, then tape the ends together. If you start
with a long strip, you can do the task easily. If you start with a shorter strip,
you will find the task more difficult. How short can you go?

fold
e
tap

3 fold rotate
e
tap

fold

1
Figure 1: The triangular Moebius band
∗
Supported by N.S.F. Grant DMS-2102802 and also a Mercator Fellowship.

1
 √
Figure 1 shows a famous example that is based on a 1 × 3 strip. The
strip in Figure 1 is colored red on one side and blue on the other. You are
supposed to fold and then somehow tape this thing as indicated in Figure 1.
The tape runs along the dotted line in the “inside” of the little triangular
“wallet” you are making. (You might enjoy finding other ways of making this
example in which the taping happens “on the outside” and is much easier to
manage.) The final rotation highlights a kind of “T -pattern” made from the
top edge and the dotted line – a pattern we will see again. Let’s call this
example the triangular Moebius band . Can you do better than this example?
√
That is, can you turn a 1 × λ strip into a paper Moebius band if λ < 3?
In order to answer this question we have to be more formal about what
we are doing. Formally speaking, an embedded paper Moebius band of aspect
ratio λ is a smooth isometric embedding I : Mλ → R3 , where Mλ is the flat
Mobius band

Mλ = ([0, 1] × [0, λ])/ ∼, (x, 0) ∼ (1 − x, λ) (1)

An isometric mapping is a map whose differential is an isometry. The map

is an embedding if it is injective. The map is always distance non-increasing
with repect to the ambient metric on R3 . The image M = I(Mλ ) is an
example of a developable surface (with boundary). I learned about paper
Moebius bands from the beautiful expository article [FT, Chapter 14] by
Dmitry Fuchs and Sergei Tabachnikov.
According to the formal definition, the triangular Moebius band does
not count as a paper Moebius band because the map is neither smooth nor
injective. However, as Benjamin Halpern and Charles Weaver prove in [HW],
it is the limit of smooth examples. That is, you can approximate it as closely
as you like by maps which are paper Moebius bands in the formal sense
above. Why bother with smooth maps? Well, if you just look at ways of
folding paper up to make a Moebius band you can get all kinds of weird
examples. For instance, you could take a square, fold it like an accordion
into a thin strip, twist, then tape. This monster is not approximable by
smooth examples. The smooth formalism rules out pathologies like this.
The Moebius band question is related to many topics. The early paper
[Sa] proves rigorously that smooth paper Moebius bands exist. (See [HF] for
a modern translation to english.) The paper [CF] gives a general framework
for the differential geometry of developable surfaces. Some authors have
discussed optimal shapes for Moebius bands from other perspectives, e.g.

2
algebraic or physical. See, e.g. [MK]. The Moebius band question has
connections to origami. See e.g. the beautiful examples of isometrically
embedded flat tori [AHLM]. The Moebius band question is also similar in
spirit to rope knot theory. See e.g. [CKS].
In [HW], Halpern and Weaver show that λ > π/2 for any smooth paper
Moebius band. This bound works more generally in the immersed case,
when I is not assumed to be injective. They also show that you can find an
immersed example having λ as close to√π/2 as you desire. The last line of
[HW] states the conjecture that λ > 3 for an embedded paper Moebius
band. In this paper I will prove this conjecture and show that the triangular
Moebius band is indeed the best you can do.

Theorem 1.1 (Main)√ A smooth embedded paper Moebius band has aspect
ratio greater than 3.

Theorem 1.2 (Triangular Limit) Let In : M√ λn → Mn be a sequence of

embedded paper Moebius bands such that λn → 3. Then, up to isometry,
In converges uniformly to the map giving the triangular Moebius band.

The work here supersedes my earlier paper [S] and also is independent
from it. Nonetheless, the work here is an outgrowth of [S]. In [S] I improved
the lower bound λ > π/2 in the√ embedded case to a bound λ > λ1 for some
complicated number λ1 ∈ (π/2, 3). Let me explain this in some detail.
An embedded paper Moebius band M = I(Mλ ) has 1 a foliation – i.e. a
continuously varying decomposition – by straight line segments which meet
∂M transversely at their endpoints. We call these segments the bends. They
are sometimes called the ruling lines. We call the corresponding pre-images of
these bends on Mλ the pre-bends. Every bend has an extending line, namely
the line that contains it. We say that a T -pattern is a pair of co-planar bends
whose extending lines are perpendicular. In [S] I proved the following result
under the additional hypothesis that λ < 7π/12.

Lemma 1.3 (T) An embedded paper Moebius band has a T -pattern.

I deduce easily from the version of Lemma T in [S] that λ ≥ φ, the golden
ratio. (This amounts to taking t = 0 in Equation 2 below.) Then I solve an
1
This is a classic result. See e.g. the bottom of p. 46 in [HW]. Here they explain that
the subset of points of non-zero mean curvature on a paper Moebius band has a foliation
by such segments. One can then extend the foliation continuously to the complement.

3
 √
optimization problem to get a lower bound λ1 ∈ (φ, 3). Embarrassingly,
I discovered recently that I made an error in setting up the optimization
problem in [S]. I mistakenly assumed that when you cut Mλ open along an
embedded line segment which joins two points in ∂Mλ you get a parallelogram
rather than a trapezoid. This mistake caused me to miscalculate λ1 . So,
really, all I can conclude from [S] is that λ1 ≥ φ. However, I was amazed and
delighted √to discover that when I did the optimization problem correctly I
got λ1 = 3 right on the nose! Here is the form the optimization calculation
takes in this paper.

Lemma 1.4 (A) Suppose that M√= I(Mλ ) is an embedded paper Moebius
band with a T -pattern. Then λ > 3.

The Main Theorem follows immediately from Lemma A and Lemma T.

In §2 and §3 respectively I will give short and self-contained proofs of these
results. (The side hypothesis for Lemma T that I had in [S] is not needed.)
The proof of the Triangular Limit Theorem, given in §4, amounts to exam-
ining what our proof of Lemma A says about a minimizing sequence. In §5
I will include some remarks about Lemmas A and T. Many of these remarks
were inspired by my conversations with other mathematicians who have read
earlier drafts of this paper.
I would like to thank Matei Coiculescu, Robert Connelly, Dan Cristofaro-
Gardiner, Dmitry Fuchs, Javier Gomez-Serrano, Anton Izosimov, Jeremy
Kahn, Stephen D. Miller, and Sergei Tabachnikov for helpful discussions
about this subject. I’d especially like to thank Matei for suggesting that I
try for a “mapping proof” of Lemma T as opposed to the kind of proof I had
previously. That suggestion led me to find a really nice proof of Lemma T
that greatly simplified this paper.

4
2 Proof of Lemma A
Let ∇ be a triangle with horizontal base. Let p(∇) be the perimeter of ∇
and let n(∇) be the sum of the lengths of the non-horizontal edges of ∇.
√ √
Lemma 2.1√If ∇ has √ base 1 + t2 and height h ≥ 1 then n(∇) ≥ 5 + t2
and p(∇) ≥ 1 + t2 + 5 + t2 . Equality occurs iff ∇ is isosceles and h = 1.
√
Proof: This is an extremely well known result. Let β = 1 + t2 .

v1 v2

2h
v3
v2'
Figure 2.1: The diagram for Lemma 2.1.

Let v1 , v2 , v3 be the vertices of ∇, with v3 the apex. Let v2′ be the reflection
of v2 through the horizontal line containing v3 . By symmetry, the triangle
inequality, and the Pythagorean Theorem,
p p √
n(∇) = kv1 −v3 k+kv3 −v2′ k ≥ kv1 −v2′ k = β 2 + 4h2 ≥ β 2 + 4 = 5 + t2 .

The bound for p(∇) follows immediately. In the case of Equality, h = 1 and
v1 , v3 , v2′ are collinear, meaning that ∇ is isosceles. ♠

We have I : Mλ → M. We write X ′ = I(X) for any relevant set X and

we let ℓ(X) denote the length of any relevant X. By definition, we have
ℓ(γ) = ℓ(γ ′ ) for any curve γ ⊂ Mλ . For instance, ℓ(∂Mλ ) = ℓ(∂M).
Let B ′ and T ′ be the pair of disjoint bends comprising a T -pattern of
M. We choose so that the line extending T ′ is disjoint from B ′ , then rotate
so that B ′ and T ′ are respectively vertical and horizontal segments in the
XY -plane and B ′ is strictly below the line extending T ′ . Let B and T be
the pre-bends corresponding to B ′ and T ′ . We cut Mλ open along B to get
a bilaterally symmetric trapezoid. See Figure 2.2.

5
 v B H2'
H2 b w' T' x'
u
w v' H1'
t T t D2
D1 x B'
H1 D1' D2'
b
v
u B u'
Figure 2.2: The trapezoid (left) and the T-pattern (right).

Here −t and b are respectively the slopes of T and the bottom choice of B.
The picture looks a bit different when the signs of t and b are different, but
it is always true that ℓ(H1 )√
= ℓ(D1 ) − t − b and ℓ(H2 ) = ℓ(D2 ) − t + b. The
yellow triangle ∇ has base 1 + t2 and height greater than 1.
√ √
First Bound: We have 2λ ≥ 1 + t2 +
5 + t2 . Here is the derivation:
√ √
2λ = ℓ(∂Mλ ) = ℓ(∂M) ≥ p(∇) ≥ 1 + t2 + 5 + t2 . (2)

The first inequality comes from the fact that ∂M is a (red and magenta) loop
containing all vertices of ∇. The second inequality is Lemma 2.1.
√
Second Bound: We have 2λ ≥ 2 5 + t2 − 2t. Here is the derivation.

2λ = ℓ(D1 ) + ℓ(D2 ) + ℓ(H1 ) + ℓ(H2 ) = 2ℓ(D1 ) + 2ℓ(D2 ) − 2t =

√
2ℓ(D1′ ) + 2ℓ(D2′ ) − 2t ≥ 2n(∇) − 2t ≥ 2 5 + t2 − 2t. (3)
The first inequality comes from the fact that D1′ ∪ D2′ is a (red) path that
connects w ′ to x′ and contains u′ . The second inequality is Lemma 2.1.
√
Combining√ the Bounds: Let t0 = 1/ 3. If t ≥ t0 then our √ first bound
gives√λ ≥ 3. √If t ≤ t0 then our second bound√gives λ ≥ 3. Hence
λ ≥ 3. If λ = 3 then t =√t0 and ∇ has base 2/ 3 and height 1. This is
a contradiction. Hence λ > 3. This completes the proof of Lemma A.

6
3 Proof of Lemma T
Let I : Mλ → M be a paper Moebius band. We parametrize the space of
pre-bends of Mλ with the circle R/λZ as follows: Each pre-bend corresponds
to its intersection with the centerline of Mλ , and this is a copy of R/λZ.
Using I, we simultaneously parametrize the space of bends of M by R/λZ.

The Cylinder: Let Υ be the cylinder obtained from (R/λZ)2 by delet-

ing the diagonal. A point (x, y) ∈ Υ corresponds to a pair (u, v) of distinct
bends. We let Υ be the compactification of Υ obtained by adding 2 boundary
components. The point (x, y) lies near one boundary component if y lies just
ahead of x in the cyclic order coming from R/λZ. The point (x, y) lies near
the other boundary component if y lies just behind of x in the same cyclic
order. Let ∂Υ be the boundary of Υ.

Oriented Bends: Let (x, y) ∈ Υ be arbitrary. There is a unique minimal

path xt ∈ R/λZ such that x0 = x and x1 = y and xt is locally increasing
with respect to the cyclic order on R/λZ. This path is short when (x, y)
is near one component of ∂Υ and long near the other. Let ut be the bend
associated to xt . Given an orientation on u0 = u, we extend it continuously
to an orientation on u1 = v. Let − →
u be vector parallel to our oriented u.
That is, −
→u points from the tail of u to the head of u. Likewise define −→v.
We write u→
− →
− →
−
v . Note that − u →
−
− v and, since we are on a Moebius
band, −→
v −−→u.

Two Functions: Let mu and mv be the midpoints of u and v. Define

g(x, y) = −
→
u ·−
→
v, h(x, y) = (mu − mv ) · (−
→
u ×−
→
v ). (4)

If we had started with the other orientation of u we would get the same value
for g and h because −− →
u −−
→v . Hence g and h are well defined. Note that
g and h extend continuously to Υ. We have g ≥ 1 on one component of ∂Υ
and g ≤ −1 on the other. We have h = 0 on ∂Υ. We compute

g(y, x) = −
→
v · (−−
→
u ) = −g(x, y). (5)

Likewise

h(y, x) = (mv − mu ) · (−
→
v × (−−
→
u )) = (mv − mv ) · (−
→
u ×−
→
v ) = −h(x, y). (6)

7
Lemma 3.1 If g(x, y) = h(x, y) = 0 then (u, v) make a T-pattern.

Proof: Since g(x, y) = 0 the vectors − →

u and − →v are orthogonal. Hence
→
− →
− →
−
n = u × v is nonzero. By construction u and v and the segment mu mv
all lie in planes orthogonal to −→n . But then they all lie in the same plane
orthogonal to −→n . In short, u and v are co-planar. ♠

To prove Lemma T, we just have to prove that g and h simultaneously

vanish somewhere in Υ. Suppose not. Since |g| ≥ 1 on ∂Υ, we know g and h
do not simultaneously vanish on Υ. Let S 1 be the unit circle. Let A = (f, g)
and B = A/kAk. Then B : Υ → S 1 is well-defined and continuous. B maps
one component of ∂Υ to (1, 0) and the other to (−1, 0).
Consider any path γ which connects a point in one component of ∂Υ to
a point in the other. The image B(γ), always oriented from (1, 0) to (−1, 0),
winds some half integer w(γ) times around the origin. All choices of γ are
homotopic to each other relative to ∂Υ. Thus w(γ) is independent of γ.
The function Σ(x, y) = (y, x) extends continuously to Υ and swaps the
components of ∂Υ. Our independence result says that w(Σ(γ)) = w(γ). On
the other hand, Equations 5 and 6 together imply that B ◦ Σ = −B. But
then, as Figure 3 illustrates, when we orient B(Σ(γ)) = −B(γ) from (1, 0)
to (−1, 0), the winding number is −w(γ). This contradiction completes the
proof of Lemma T.

B( )

w=1/2 w=-1/2

-B( )

Figure 3: The effect of negation: a cartoon

8
4 Proof of the Triangular Limit Theorem
We revisit Lemma A. Here is Figure 2.2 again.

v B H2'
H2 b w' T' x'
u
w v' H1'
t T t D2
D1 x B'
H1 D1' D2'
b
v
u B u'
Figure 2.2: The trapezoid (left) and the T-pattern (right).

Here is Equation 2 again.

√ √
2λ = ℓ(∂Mλ ) = ℓ(∂M) ≥ p(∇) ≥ 1 + t2 + 5 + t2 . (7)

Suppose we
√ have a sequence {Mn } of embedded paper Moebius bands
with λn → 3. We run the constructions from Lemma A for each one.
Looking at the analysis
√ done at the end of the proof of Lemma A, we see
that tn → t0 =p1/ 3. Also bn → 0, because otherwise the height of ∇n ,
which exceeds 1 + b2n , does not converge to 1.
Thanks to the uniqeness in Lemma 2.1, the triangle √ ∇n converges up to
isometry to an equilateral triangle ∇ of perimeter 2 3. We normalize by
isometries so that the the vertices of ∇n converge to ∇. Inspecting Equation
7, we see that
|ℓ(∂Mn ) − p(∇)| → 0. (8)
Since In is length perserving the convergence in Equation 8 implies that
In , when restricted to each of the 4 segments Dn,j and Hj,n in ∂Mλn , converges
uniformly to a linear isometry. Hence the restriction of In to ∂Mλn converges
uniformly to the map that comes from the triangular Moebius band. The
action of In on ∂Mλn determines the action of In on Mλn , so the convergence
on the boundary implies the convergence on the whole space. This completes
the proof of the Triangular Limit Theorem.

9
5 Discussion
5.1 Lemma A
The proof of Lemma A only requires the map I : Mλ → M to have the
following properties.

1. I is continuous.

2. The interior of Mλ has a continuous partition by open line segments

whose endpoints lie in the boundary.

3. Given an arbitrary line segment v in the partition the image I(v) is a

line segment in R3 that is at least as long as v.

4. The restriction of I to ∂Mλ is locally distance non-increasing with

respect to the ambient metric on R3 .

5. There exist 2 segments v, w in the partition such that I(v) and I(w)
are disjoint, co-planar, and lie on perpendicular lines.

The map I need not be an embedding; all we need is an embedded T -pattern.

It is worth pointing out that the Triangular Limit Theorem does not quite
work out the same way for the general kinds of maps defined above. The step
that goes wrong is the final one. In the generality above, the restriction of I
to ∂Mλ does not determine the action of I on all of Mλ . Nevertheless, we can
say that for a minimizing sequence {In }, the maps converge uniformly on the
boundary, up to isometry, to the triangular Moebius band map. Also, the
images In (Mλn ) do converge (e.g. in the Hausdorff metric) up to isometries
to the triangular paper Moebius band.

5.2 Lemma T
The proof I give of Lemma T is quite reminiscent of the proof of the Borsuk-
Ulam Theorem. Indeed, Jeremy Kahn pointed out to me that the endgame
of my proof really is the Borsuk-Ulam proof in disguise. To see this, note
that we obtain the 2-sphere S 2 by crushing each component of ∂Υ to a point.
Then B induces a map S 2 → S 1 with B ◦ Σ = −Σ. The map Σ, which is a
glide reflection on Υ, acts on S 2 as the antipodal map.

10
 In this context, it is more natural to redefine the vectors −
→
u and −
→
v to be
the unit vectors parallel to the orientations of u and v. Once this is done,
the functions g and h themselves restrict to our quotient S 2 and we really
have the exact conditions for the Borsuk-Ulam Theorem.

Anton Izosimov and Sergei Tabachnikov independently suggested to me

the following general formulation of Lemma T.

Lemma 5.1 Suppose {Lt | t ∈ [0, 1]} is a continuous family of oriented lines
in R3 such that L1 = Lopp0 , the same line as L0 but with the opposite ori-
entation. Then there exist parameters r, s ∈ [0, 1] such that Lr and Ls are
perpendicular intersecting lines.

This result immediately implies Lemma T, and it has essentially the same
proof. In particular, Lemma 5.1 applies to maps I : Mλ → M which satisfy
Conditions 1-4 above. The output is a T -pattern which might or might not
be embedded. If I is an embedding then, of course, the T -pattern will also
be embedded.

Sergei also suggests a beautiful alternate formalism for the proof of Lemma
T. One introduces the Study numbers. These have the form x + ǫy where
x, y ∈ R and ǫ2 = 0. Likewise, one introduces the Study vectors. These have
→
− →
−
the form − →
a + ǫ b , where −→
a , b ∈ R3 and again ǫ2 = 0. In this context, the
dot product of two Study vectors makes sense and is a Study number.
→
−
Each oriented line ℓ ⊂ R3 gives rise to a Study vector ξℓ = − →
a + ǫ b where
→
− →
−
a is the unit vector pointing in the direction of ℓ and b = ℓ′ × − →a . Here
→
−
ℓ ∈ ℓ is any point. All choices of ℓ give rise to the same b ; this vector is
′ ′

called the moment vector of ℓ. This formalism identifies the space of oriented
lines in R3 with the so-called study sphere consisting of Study vectors ξ such
that ξ · ξ = 1. The Study dot product ξℓ · ξm vanishes if and only if ℓ and m
are perpendicular and intersect. Thus our two functions g and h carry the
same information as the Study dot product. This makes the functions g and
h seem more canonical.

11
6 References
[AHLM] A. Malaga, S. Lelievre, E. Harriss, P. Arnoux,
ICERM website: [Link] (2019)
[CB], R. Connolly, A Back, Mathematics and Tensegrity, in American Sci-
entist, Vol 82 2, (March-April 1998) pp 142-151.
[CF] Y. Chen and E. Fried, Mobius bands, unstretchable material sheets and
developable surfaces, Proceedings of the Royal Society A, (2016)
[CKS] J. Cantarella, R. Kusner, J. Sullivan, On the minimum ropelength of
knots and links, Invent. Math. 150 (2) pp 257-286 (2003)
[FT], D. Fuchs, S. Tabachnikov, Mathematical Omnibus: Thirty Lectures on
Classic Mathematics, AMS 2007
[GS] G. Schwarz, A pretender to the title “canonical Moebius strip”, Pacific
J. of Math., 143 (1) pp. 195-200, (1990)
[HF, D.F. Hinz, E. Fried, Translation of Michael Sadowsky’s paper ‘An el-
ementary proof for the existence of a developable MÖBIUS band and the
attribution of the geometric problem to a variational problem’ . J. Elast. 119,
3–6 (2015)
[HW], B. Halpern and C. Weaver, Inverting a cylinder through isometric
immersions and embeddings, Trans. Am. Math. Soc 230, pp 41–70 (1977)
[MK] L. Mahadevan and J. B. Keller, The shape of a Mobius band , Proceed-
ings of the Royal Society A (1993)
[Sa], M. Sadowski, Ein elementarer Beweis für die Existenz eines abwickel-
baren MÖBIUSschen Bandes und die Zurückführung des geometrischen Prob-
lems auf einVariationsproblem. Sitzungsberichte der Preussischen Akad. der
Wissenschaften, physikalisch-mathematische Klasse 22, 412–415.2 (1930)
[S] R. E. Schwartz, An improved bound on the optimal paper Moebius band ,
Geometriae Dedicata, 2021

12