A concise geometric proof of the three distance theorem

Hamada Tadahisa

Abstract
The three distance theorem states that for any given irrational number α and a natural number n, when
the interval (0, 1) is divided into n + 1 subintervals by integer multiples of α, namely, {0}, {α}, {2 α},
. . . , {n α}, then each subinterval is limited to at most three different lengths. Steinhaus conjectured this
theorem in the 1950s, and many researchers have given various proofs since then.
This paper aims to improve the perspective by showing a two-dimensional map of how the unit
interval is divided by continuously changing α, and give a concise proof of the theorem. We also present
a simple proof of the three gap theorem, a dual of the three distance theorem.
arXiv:2308.11999v5 [math.NT] 16 Sep 2023

1. Introduction
Pythagoras is credited by Nicomachus, Boethius, and others, though possibly untrue, with having
discovered that when the ratio of string lengths is 3 : 2, the notes produced by the strings resonate
harmoniously together (for example, [6], [3], [11]). We can determine the pitch of the notes F → C →
G → D → A → E → B by starting with the note F and successively multiplying the length of the string
2
by the factor of , and doubling the length of the string when it was less than half its original length,
3
known as the Pythagorean temperament.

2 notes
1
0 0.585 0.415 1
3 notes
2 1
0 0.170 0.415 1
5 notes
2 4 1 3
0 0.170 0.245 1
7 notes
2 4 6 1 3 5
0 0.170 0.075 1
12 notes
7 2 9 4 11 6 1 8 3 10 5
0 0.095 0.075 1
17 notes 14 16 13 15
12 7 2 9 4 11 6 1 8 3 10 5
0 0.075 1
0.020

Figure 1: Formation of Pythagorean temperament

Next, Figure 1 illustrates how the Pythagorean temperament subdivides an octave. We obtain this by
multiplying log2 3 by rank numbers 1, 2, 3, . . . , while marking the fractional parts since 1 and 2 correspond
to 0 and 1, respectively, on a logarithmic scale with a base of 2.
If we arrange the values of these divisions in magnitude order and take the difference (interval)
between adjacent values, we can see that the difference values are limited to two or three. Figure 1
shows the divisions for two different values of difference, in order of increasing number of divisions.
Temperaments with 5, 7, or 12 divisions of an octave are typical in traditional and contemporary music.
We can also see that the number of divisions when there are two values of difference is the denom-
inators of the convergents obtained by expanding log2 3 into continued fractions and that the interval
adjacent to 0 is different from the interval adjacent to 1. We will refer to one as the large interval and

1
the other as the small interval. Furthermore, the difference in the rank number of adjacent values is
equal to one of the two types of interval numbers modulo n, and the sign differs depending on whether
the difference in adjacent values is a large or small interval. If we set the rank number corresponding to
0 and 1 at both ends of an octave to 0, this relationship holds, including both ends.
These observations lead to a natural deduction of the three distance theorem since it is understood
that the above holds true even if log2 3 is replaced by an arbitrary, irrational number, given a continued
fraction expansion. This fact can also be interpreted as an expression of regularity of the image of a
finite arithmetic progression with the initial term 0 in the projection from the set of real numbers R to
the unit interval R/Z(= [0, 1)). Steinhaus formulated this as a conjecture only in the second half of the
20th century, and many researchers have provided various proofs since then.
Here we give a brief survey of the different approaches adopted by various authors to prove the three
distance theorem and its dual, the three gap theorem.
The previous proofs of the three distance theorem all show that for a given α, the length of the line
segment formed by {iα} dividing the unit interval is determined by min({iα}) and max(1 − {iα}).
Throughout the 20th century, an intriguing array of proofs appeared, leaning heavily on the foun-
dational pillars of elementary number theory. Concepts like continued fractions and combinatorics fre-
quently used, with notable examples being highlighted in works like [9], [10], [4], [8], [12]. Our paper
finds its roots in this classical tradition. Pioneers Alessandri and Berthé recognizing the significance of
these endeavors, meticulously wrote a survey paper [1], encapsulating the essence of previous works.
In the 21st century, the community began to embrace more contemporary tools and methodologies.
For instance, Biringer and Schmidt [2] ventured into the intricate realms of the Riemannian manifold.
Haynes and team [5] delved into the multifaceted world of CW-complexes. Not to be left behind, Marklof
and Strömbergsson [7] amalgamated their understanding of the modular group with lattice structures,
presenting yet another groundbreaking approach.
We give in this paper an elementary, short proof of the three distance theorem. We also show a simple
proof of the three gap theorem, which is dual to it. Our unique point is that we showed how the unit
interval is divided by continuously changing the value of α with α as a variable on a two-dimensional
map.

2. Preparation
We denote by ⌊x⌋ (x ∈ R) the largest integer not exceeding x, and by {x} the fractional part of
x ∈ R, i.e., {x} := x − ⌊x⌋.
a
The Farey sequence Fn for any positive integer n is the set of irreducible rational numbers with
b
a c

0 ≤ a ≤ b ≤ n and gcd(a, b) = 1 arranged in increasing order. Any pair of consecutive fractions ,
b d
a
in Fn is called a Farey pair. If a, b, c, d are four positive integers, the mediant of two fractions and
b
c a+c
is defined as . In the Farey sequence Fn for n, if we interpolate the mediants of the Farey
d b+d
pairs where the sum of the denominators are n + 1 we can get Fn+1 , the Farey sequence for n + 1. An
a c
a c
interval , is called a Farey interval if and are a Farey pair in Fn for some n.
b d b d
Next, we define the unit interval integral partition of order n (Un ) as a two-dimensional region with
α ∈ (0, 1) on the horizontal axis and height h ∈ [0, 1) on the vertical axis, with the group of lines
h = {iα}, (0 < α < 1, i = 1, . . . , n) drawn. Once the value of α is determined, the line segment
representing h ∈ [0, 1) is divided into, at most n + 1 subintervals. For convenience, separated individual
intervals are assumed to be semi-open intervals, closed to the left and open to the right. Figure 2 shows
the case n = 3.

2
 h
1

0
0 1 1 2 1α
3 2 3

Figure 2: Unit interval integral partition of order 3 (U3 ): h = {iα}, (0 < α < 1, i = 1, 2, 3)

3. The three distance theorem

We state the three distance theorem by extending it to the case where α is real.
Theorem 1 (Three distance theorem). Let 0 < α < 1 be a real number and n a positive integer. The
points {i α}, for 0 ≤ i ≤ n, partition the unit interval into at most n + 1 subintervals, the lengths of
which take at most three values, one being the sum of the other two.
More precisely, if α ∈ Fn , then the subintervals of the unit interval divided by {i α}, for 0 ≤ i ≤ n, all
a c
have the same length. Otherwise, let < be the Farey pair surrounding α. Then, the unit interval
b d
is divided by the points {i α}, for 0 ≤ i ≤ n, into n + 1 intervals which satisfy:
• n + 1 − b of them have length b α − a,
• n + 1 − d of them have length c − d α,
• b + d − (n + 1) of them have length (b − d) α + c − a.
The last case occurs when b + d > n + 1.
Proof. First, every line segment in Un can generally be expressed as h = i α − j (0 < i ≤ n, 0 ≤
j < i, 0 ≤ h < 1). Thus, if two line segments h = i α − j and h = i′ α − j ′ have an intersection, its
α-coordinate is an irreducible fraction with a denominator less than n. That is, all α-coordinates of the
intersections of line segments in Un belong to Fn−1 .
p
If α = ∈ Fn , the indeterminate equation p i − q j = k (0 < p < q ≤ n, 0 ≤ k ≤ q) has
q
solutions for each k in the range 0 < i ≤ n, 0 ≤ j < i, so the set of intersections with each line
p
segment h = i α − j (0 < i ≤ n, 0 ≤ j < i, 0 ≤ h ≤ 1) and the line segment α = (0 ≤ h ≤ 1) is
q
 p k

, 0≤k≤q .
q q
Otherwise, consider a rectangular region with the corresponding Farey interval on the horizontal axis
a c

and the unit interval on the vertical axis , ×[0, 1). From the previous discussion, it is clear that all
b d
a k a
points ( , ) (0 ≤ k < b) that divide the leftmost unit line segment of this Farey interval × [0, 1)
b b b
c k
into b equal parts and all points ( , ) (0 < k ≤ d) that divide the rightmost unit line segment of this
 c d d
Farey interval × [0, 1) into d equal parts, pass through one of the lines h = {i α} (i = 1, . . . , n),
d
and these lines do not intersect inside this rectangle.
Therefore, the rectangular region is divided into the following three types of regions by h = i α−j (0 <
i ≤ n, 0 ≤ j < i, 0 ≤ h < 1):

3
 1
• triangles with a base of length at the left end and a vertex at the right end of the rectangular
b
region.
1
• triangles with a base of length at the right end and a vertex at the left end of the rectangular
d
region.
and if b + d > n + 1, then,
1 1
• trapezoids with one base of length at the left end and the other base of length at the right
b d
end of the rectangular region.
Given α, we immediately see that the length of the subintervals separating the unit interval {α} × [0, 1)
are constant for each of the above types of regions and that the length of the section through the trapezoid
is equal to the sum of the lengths of the sections through the two types of triangles above, and that the
specific lengths of each are also as stated in the theorem.
Note that when α is the mediant of the Farey pair that surrounds it, the lengths of the subinterval
delimited by {i α} (0 ≤ i ≤ n) is one less since the lengths of the subinterval delimiting each triangle of
the first and second type is equal.
Remark. From the argument with Un in this proof, it follows that when α ∈
/ Fn ,

n ≥ max(b, d), (1)

n + 1 ≤ b + d. (2)

Next, when {1α}, . . . , {nα} are sorted in increasing order, we denote that {i α} is followed by {j α}
as suc(i) = j. The minimum value among them is {b α}, which we denote by s, and the maximum value
is {d α}, which we denote by 1 − t, then the following holds.

n + 1 − b intervals of length s = b α − a, suc(i) = i + b, for 0 ≤ i ≤ n − b, (3)

n + 1 − d intervals of length t = c − d α, suc(i) = i − d, for d ≤ i ≤ n, (4)
b + d − (n + 1) intervals of length s + t = (b − d) α + (c − a),
(5)
suc(i) = i + b − d, for n − b < i < d.

4. The three gap theorem

The three gap theorem, described next, is a dual to the three distance theorem. The decision param-
eter n in the three distance theorem is replaced by a real number β.


Theorem 2 (Three gap theorem). Let α, β ∈ 0, 1 with α being irrational. The gaps between successive
integers i for which that {i α} < β have at most three distinct values, one of which is the sum of the
other two.
More precisely, if β ≥ max(α, 1 − α), the possible gaps are 1 and 2:
• the gap of 1 occurs with frequency (2 β − 1)/β,
• the gap of 2 occurs with frequency (1 − β)/β.
If β < max(α, 1 − α), let b and d be the smallest positive integers satisfying

{b α} < β, {d α} > 1 − β.

Then, the possible gaps are b, d and b + d. Letting s = {b α} and t = 1 − {d α}:

• the gap b occurs with frequency (β − s)/β,
• the gap d occurs with frequency (β − t)/β,
• the gap b + d occurs with frequency (s + t − β)/β.
The last case arises when s + t > β.

4
Proof. Case 1: For β ≥ max(α, 1 − α), since both {0α} and {1α} are less than β, 1 is one of the gaps.
Given that 0 < 1 − β ≤ min(α, 1 − α), there are clearly two possible gaps: 1 and 2.
In detail, if {i α} ∈ [0, β − α) ∪ [1 − α, β), the subsequent gap is 1. If {i α} ∈ [β − α, 1 − α), the next
gap is 2. From number theory, we know that the numbers {i α}(i ≥ 0) are uniformly distributed over
the interval [0, 1). Hence, the frequency of each gap matches the length of the associated interval.
Case 2: If β < max(α, 1 − α), clearly, b ̸= d. Hence, we have

β > max{s, t}. (6)

It’s straightforward to see that β ≤ {(b − d) α}. If {(b − d) α} < β, then {(d − b) α} > 1 − β, leading to
a contradiction either when b > d or b < d. Given that 0 < s + t < 1, we have:

β ≤ {(b − d)α} = s + t. (7)

Thus, for {i α} ∈ [0, β), the gaps depend on the value of {i α}.

If {i α} ∈ [0, β − s), then the gap after i is b, (8)

because {(i + b) α} ∈ [s, β). If there exists an e ∈ Z such that 0 < e < b and {(i + e) α} ∈ [0, β), then
considering {b α} = s, {e α} ∈ [0, β) or {(b − e) α} ∈ [0, β) contradicts the definition of b.

If {i α} ∈ [t, β), then the gap after i is d, (9)

because {(i + d) α} ∈ [0, β − t). If there exists an e ∈ Z such that 0 < e < d and {(i + e) α} ∈ [0, β), then
considering {d α} = 1 − t, {e α} ∈ [1 − β, 1) or {(d − e) α} ∈ [1 − β, 1) contradicts the definition of d.

If {i α} ∈ [β − s, t), then the gap after i is b + d, (10)

Because, first of all, {(i + b + d) α} ∈ [β − t, s). There does not exist an e ∈ Z such that 0 < e < b + d
and {(i + e) α} ∈ [0, β). Let u = {iα} ∈ [β − s, t), and first consider the range [u, β). For 0 < e < b,
{(i+e) α} ∈/ [u, u+β). Also, since {(i+b) α} = u+s ∈ [β, s+t), for b < e < b+d, {(i+e)α} ∈ / [s+t−β, β).
For the range [0, u), it can be shown in the same way by swapping b and d.
As in Case 1, the frequency of each gap is equal to the length of the corresponding interval.

Remark. Expressions (6) through (10) in the three gap theorem are duals of expressions (1) through (5)
in the three distance theorem, respectively.

References
[1] Pascal Alessandri and Valérie Berthé, Three distance theorems and combinatorics on words, En-
seignement Mathématique 44 (1998) 103–132.
[2] Ian Biringer and Benjamin Schmidt, The three gap theorem and Riemannian geometry, Geometriae
Dedicata 136 (2008) no. 1, 175–190.

[3] Walter Burkert et al., Lore and science in ancient Pythagoreanism, Harvard University Press, 1972.
[4] John H. Halton, The distribution of the sequence {n ξ}(n= 0, 1, 2,. . . ), Mathematical Proceedings
of the Cambridge Philosophical Society, vol. 61, Cambridge University Press, 1965, pp. 665–670.
[5] Alan Haynes, Henna Koivusalo, James Walton, and Lorenzo Sadun, Gaps problems and frequencies
of patches in cut and project sets, Mathematical Proceedings of the Cambridge Philosophical Society,
vol. 161, Cambridge University Press 2016, pp. 65–85.
[6] Chalcidensis Iamblichus and August Nauck, [De vita Pythagorica liber]; Iamblichi de vita Pythagor-
ica liber, Hakkert, 1965.
[7] Jens Marklof and Andreas Strömbergsson, The three gap theorem and the space of lattices, The
American Mathematical Monthly 124 (2017), no. 8, 741–745.
[8] Noel B. Slater, Gaps and steps for the sequence n θ mod 1, Mathematical Proceedings of the Cam-
bridge Philosophical Society, vol. 63, Cambridge University Press, 1967, pp. 1115–1123.

5
 [9] Vera T. Sós, On the distribution mod 1 of the sequence n α, Ann. Univ. Sci. Budapest, Eötvös Sect.
Math, 1 (1958), 127–134.
[10] Stanislaw Świerczkowski, On successive settings of an arc on the circumference of a circle, Funda-
menta Mathematicae 46 (1958), no. 2, 187–189.

[11] Bartel van der Waerden, Die Harmonielehre der Pythagoreer, Hermes 78 (1943), no. 2, 163–199.
[12] Tony Van Ravenstein, The three gap theorem (Steinhaus conjecture), Journal of the Australian
Mathematical Society 45 (1988), no. 3, 360–370.