The irrationality exponent measures how fast the error of fractional approximations to an irrational number shrinks, as a power of the denominator. After its definition and examples (√2, Liouville's number, almost every real number), with proofs, we read the OpenAI model's proof that the irrationality exponent of π is 2, and how the convergence of the Flint–Hills series Σ1/(n³sin²n) follows from it, with figures and formulas.
A number that measures how well an irrational number can be approximated by fractions: the exponent of the denominator q in the error. It is at least 2 for every irrational number, and 2 means that only “ordinary” approximations exist.
The new result
A manuscript written by an OpenAI model proves that the irrationality exponent of π is exactly 2. The best upper bound before was about 7.10, and the value 2 had long been conjectured.
A series
It was known that the Flint–Hills series ∑1/(n3sin2n) converges if the irrationality exponent of π is less than 5/2. The new result settles its convergence.
How sure
The main theorem is formalized in Lean and is among the statements the repository checks. The manuscript is a preprint and has not been peer reviewed.
The PDF is a copy of the version on GitHub as of October 9, 2026, shared under the Apache License 2.0 (full text). If a revised version appears, the one on GitHub will be newer.
The “Proof” under a statement opens into a list of steps. A closer look asides use first- and second-year university mathematics; For specialists asides go into the manuscript itself. You can read on without opening either.
A conjecture about approximating π, proved
On October 6, 2026, OpenAI released mathematical manuscripts written by an unreleased internal model, in the GitHub repository openai/math[1]. Number 017 is the one this article reads, “The irrationality exponent of π is 2”[2].
Its claim fits in one sentence: the irrationality exponent of π is 2. As a consequence, it also shows that the Flint–Hills series converges, a question that had long been open.
The irrationality exponent rarely comes up even in university courses. So this article first explains its definition and examples with care. It then explains the new theorem and how it decides the convergence of the series, with proofs.
The manuscript is not among those corrected or withdrawn in the update of October 7. The repository's updates are recorded in our record of corrections and withdrawals.
Approximating by fractions
Two fractions often used for π are 3.14=157/50 and 22/7. Their errors are as follows.
The fraction with denominator 7 is closer to π than the one with denominator 50. The smaller the denominator for the same closeness, the better the approximation. So we measure the error against the size of the denominator.
First, the closeness any real number x gets for free. Once the denominator q is fixed, the fractions p/q lie on the number line 1/q apart. Taking the one nearest to x, the error is at most half the spacing.
p∈Zminx−qp≤2q1p∈Zminx−qp≤2q1
So anyone can make the error about 1/q. The question is how many denominators give a much smaller error than that.
In this article, ∥t∥ denotes the distance from a real number t to the nearest integer.
∥t∥=p∈Zmin∣t−p∣∥t∥=p∈Zmin∣t−p∣
For example, the integer nearest to 7π=21.991… is 22, so ∥7π∥=∣7π−22∣=0.00885….
Since ∣x−p/q∣=∣qx−p∣/q, the best error with denominator q is ∥qx∥/q. The value ∥qx∥ is always at most 1/2, and is usually somewhere between 0.1 and 0.5.
Here are the best errors for π with denominators q=1,…,8. The last column, ν, is the exponent when the error is written as q−ν, that is ν=−log∣π−p/q∣/logq.
q
best p
error ∣π−p/q∣
q2× error
exponent ν
1
3
0.1416
0.142
—
2
6
0.1416
0.566
2.82
3
9
0.1416
1.27
1.78
4
13
0.1084
1.73
1.60
5
16
0.0584
1.46
1.77
6
19
0.0251
0.903
2.06
7
22
0.00126
0.0620
3.43
8
25
0.0166
1.06
1.97
Only q=7 stands out, with an error of about 7−3.43. For q=2,3 the best fractions are 6/2 and 9/3, which reduce to 3/1. Since the exponent ν scatters like this while the denominators are small, the irrationality exponent is defined as an exponent achieved “infinitely often, for large denominators”.
The question quotes this part of the article as it is.
Question copied
Let x be irrational. We say that ν>0 is achieved infinitely often by x if there are infinitely many pairs of coprime integers (p,q) with q≥2 such that
0<x−qp<qν10<x−qp<qν1
The supremum of the ν achieved infinitely often is the irrationality exponent (or irrationality measure) of x, written μ(x). If the supremum is infinite, μ(x)=∞.
Since q≥2, a larger ν makes q−ν smaller and the condition stricter. If some ν is achieved infinitely often, so is every smaller exponent. So the set of exponents achieved is an interval of the form (0,μ(x)) or (0,μ(x)].
The larger the irrationality exponent, the better x can be approximated by fractions. The condition 0< is there so that p/q=x is not counted when x itself is a fraction. This article only considers irrational numbers, for which it holds automatically.
Example 222/7 and 355/113
Ask an AI about this part
The question quotes this part of the article as it is.
Question copied
Since ∣π−22/7∣=0.001264…, solving 7−ν=0.001264 gives
In the same way, ∣π−355/113∣=2.668…×10−7 gives ν=15.14/4.727=3.20.
These two do not show μ(π)≥3.2. The definition asks for infinitely many fractions, so two good fractions do not affect the irrationality exponent. It measures where the approximations end up as the denominators grow.
The new theorem is stated as a lower bound for the error. So we rephrase the irrationality exponent in that form.
Proposition 3a reformulation
Ask an AI about this part
The question quotes this part of the article as it is.
Question copied
Let x be irrational and ν>0.
(1) If ν>μ(x), there is an integer Q such that ∣x−p/q∣≥q−ν for all integers p,q with q≥Q. (2) Conversely, if there is such a Q, then μ(x)≤ν.
Hence μ(x) is the infimum of the ν for which ∣x−p/q∣≥q−ν holds for all large enough denominators q.
Ask an AI about this part
The question quotes this part of the article as it is.
Question copied
ProofOpenClose
We prove (1). Since ν>μ(x), ν is not achieved infinitely often, so only finitely many coprime pairs (with q≥2) satisfy ∣x−p/q∣<q−ν.
We show that the denominators of all pairs (p,q), reduced or not, with ∣x−p/q∣<q−ν are bounded.DetailsClose
Suppose there were such pairs with arbitrarily large q. Reduce p/q to p′/q′ with 1≤q′≤q. Then
If q′≥2, then (p′,q′) is one of the finitely many pairs above. If q′=1, then ∣x−p′∣<1, which leaves finitely many p′. So p′/q′ is one of finitely many rational numbers.
Since x is irrational, the smallest distance δ from x to these rational numbers is positive. On the other hand, q−ν→0 as q→∞, so eventually ∣x−p/q∣<q−ν<δ. This is a contradiction.
Taking Q larger than the bound on the denominators, ∣x−p/q∣≥q−ν for q≥Q. This proves (1).
We prove (2). Let ν′>ν. The pairs with ∣x−p/q∣<q−ν′≤q−ν all have q<Q.DetailsClose
For each q, only finitely many p satisfy ∣x−p/q∣<1. The denominators q<Q are also finitely many, so there are finitely many pairs in all.
So ν′ is not achieved infinitely often, and μ(x)≤ν′. Since ν′>ν was arbitrary, μ(x)≤ν.
Every irrational number reaches exponent 2
In the table, the column q2× error has many values around 1. In fact, for every irrational number there are infinitely many denominators for which q2× error is less than 1. The pigeonhole principle is all the proof needs.
Proposition 4Dirichlet's approximation theorem
Ask an AI about this part
The question quotes this part of the article as it is.
Question copied
Let x be irrational. For each integer N≥1 there are integers p and q with 1≤q≤N such that
Moreover, infinitely many reduced fractions p/q satisfy ∣x−p/q∣<1/q2. In particular μ(x)≥2.
Ask an AI about this part
The question quotes this part of the article as it is.
Question copied
ProofOpenClose
Consider the N+1 numbers {0⋅x},{1⋅x},…,{Nx}, where {t}=t−⌊t⌋ is the fractional part. All of them lie in [0,1).
Cut [0,1) into N boxes [j/N,(j+1)/N) of length 1/N. Two of the N+1 numbers fall into the same box (the pigeonhole principle); call them {k1x},{k2x} with k1<k2.
Put q=k2−k1 and p=⌊k2x⌋−⌊k1x⌋. Then 1≤q≤N and ∣qx−p∣<1/N.DetailsClose
We have qx−p=(k2x−⌊k2x⌋)−(k1x−⌊k1x⌋)={k2x}−{k1x}, the difference of two numbers in one box, so its absolute value is less than 1/N.
Dividing by q gives ∣x−p/q∣<1/(qN), and 1/(qN)≤1/q2 since q≤N.
The fractions obtained this way are infinitely many.DetailsClose
Since x is irrational, ∣qx−p∣>0. Doing the same with an N′ such that 1/N′<∣qx−p∣ gives a pair with ∣q′x−p′∣<1/N′<∣qx−p∣, which differs from the first.
Repeating this gives infinitely many pairs with strictly decreasing ∣qx−p∣. Reducing a fraction keeps ∣x−p/q∣<1/q2 (the denominator gets smaller and the right-hand side larger). For each denominator only finitely many fractions are near x, so the denominators grow without bound.
If ν<2, then 1/q2<q−ν for q≥2, so ν is achieved infinitely often. Hence μ(x)≥2.
The figure below runs this proof for x=π. Moving the slider changes the number of boxes N, and the two points in the same box (orange) give a fraction p/q.
0
1
Putting the fractional parts of 0,π,2π,…,Nπ into 8 boxes, kπ for k=1 and k=8 land in the same box. Their difference gives q=7 and p=22, with ∣qπ−p∣=0.00885<1/N=0.12500.
Figure 1ExampleThe proof of Proposition 4 for x=π. Of the pairs sharing a box, the closest pair is orange. For N from 8 to 105 it finds 22/7, from 106 to 113 it finds 333/106, and from 114 on 355/113. Since ∥7π∥=0.00885 is very small, 22/7 keeps being chosen for a long time.
By Proposition 4, the irrationality exponent is always at least 2. The question is how far above 2 it can be.
Examples
Here are numbers whose irrationality exponents are known, starting with the ones we can compute.
Example 5the irrationality exponent of 2 is 2
Ask an AI about this part
The question quotes this part of the article as it is.
Question copied
If 1≤p/q≤2, then p2−2q2 is a nonzero integer, so ∣p2−2q2∣≥1. Hence
If p/q lies outside [1,2], the error is at least 0.4. In both cases, if ν>2 then ∣2−p/q∣≥q−ν for all large q. By Proposition 3 and Proposition 4, μ(2)=2.
For example, for 99/70 we have 992−2⋅702=9801−9800=1, so the error is exactly
This argument only uses that 2 is a root of the integer polynomial t2−2 of degree 2. The same idea gives the following for algebraic numbers in general.
Proposition 6Liouville's theorem
Ask an AI about this part
The question quotes this part of the article as it is.
Question copied
If x is an algebraic number of degree d≥2, then μ(x)≤d.
Ask an AI about this part
The question quotes this part of the article as it is.
Question copied
ProofOpenClose
Let f be an irreducible polynomial with integer coefficients of degree d with root x. Since d≥2, f has no rational roots.
For a fraction p/q, qdf(p/q) is a nonzero integer, so ∣f(p/q)∣≥q−d.
If ∣x−p/q∣≤1, the mean value theorem gives ∣f(p/q)∣=∣f(p/q)−f(x)∣≤M∣x−p/q∣, where M is the maximum of ∣f′∣ on [x−1,x+1].
Together, ∣x−p/q∣≥1/(Mqd) (fractions with ∣x−p/q∣>1 need no attention, since q−ν<1 for q≥2). If ν>d, then 1/(Mqd)≥q−ν for all large q, so μ(x)≤ν by Proposition 3(2). Hence μ(x)≤d.
The bound d in Liouville's theorem was later improved a great deal. In 1955 Roth proved that the irrationality exponent of every irrational algebraic number is 2[3].
The question quotes this part of the article as it is.
Question copied
If x is an irrational algebraic number, then μ(x)=2.
Conversely, a number with irrationality exponent ∞ cannot be algebraic of any degree. Liouville used this to construct the first examples of transcendental numbers.
Example 8Liouville's number
Ask an AI about this part
The question quotes this part of the article as it is.
Question copied
Let L=∑k≥110−k!=0.110001000000000000000001…. Put qn=10n! and pn=qn∑k=1n10−k!. The last digit of pn is 1, so pn/qn is reduced, and
For any ν, the right-hand side is less than qn−ν for large n with n+1>ν. So μ(L)=∞, and by Liouville's theorem L is not algebraic.
Numbers with irrationality exponent ∞ are very rare, though. The irrationality exponent of almost every real number is 2.
Proposition 9almost every real number has exponent 2
Ask an AI about this part
The question quotes this part of the article as it is.
Question copied
The set of real numbers x with μ(x)=2 has Lebesgue measure 0.
Ask an AI about this part
The question quotes this part of the article as it is.
Question copied
ProofOpenClose
Fix ν>2 and work in [0,1] (adding an integer does not change the irrationality exponent, so this is enough). Let Eq be the set of points within q−ν of a fraction with denominator q. Its length is at most 4q1−ν.DetailsClose
The fractions p/q relevant to [0,1] are the q+1 ones with p=0,…,q, and Eq is the union of the intervals of length 2q−ν around them. Its length is at most (q+1)⋅2q−ν≤4q1−ν.
Since ν>2, ∑q4q1−ν<∞. By the Borel–Cantelli lemma, the set of points lying in infinitely many Eq has measure 0.
At a point x lying in only finitely many Eq, ν is not achieved infinitely often, so μ(x)≤ν.
The countably many exceptional sets for ν=2+1/m (m=1,2,…) together still have measure 0. So do the rationals, so the remaining x are irrational with μ(x)≤2, and μ(x)=2 by Proposition 4.
To sum up: the irrationality exponent of Euler's number e is also known to be 2, from its regular continued fraction e=[2;1,2,1,1,4,1,1,6,…].
Since almost every real number has exponent 2, π was expected to have exponent 2 as well, unless there was a special reason otherwise. But little is known about π that can be used here, and proving it was hard.
The convergents of π and their exponents
To compute an irrationality exponent, it is enough to look at the good approximations. These come from the continued fraction. The continued fraction of π is
The fractions pk/qk obtained by cutting it off are the convergents. The convergents of π are 3/1, 22/7, 333/106, 355/113, 103993/33102, …, and each satisfies ∣π−pk/qk∣<1/qk2.
Computing the exponent ν=−log∣π−pk/qk∣/logqk for each convergent gives the following.
convergent p/q
error ∣π−p/q∣
exponent ν
22/7
1.26×10−3
3.43
333/106
8.32×10−5
2.01
355/113
2.67×10−7
3.20
103993/33102
5.78×10−10
2.04
104348/33215
3.32×10−10
2.10
5419351/1725033
2.21×10−14
2.19
21053343141/6701487259
2.62×10−22
2.20
100
102
104
106
108
1010
2
5/2
3
3.5
ν=2
ν=5/2
22/7
355/113
denominator q (ticks a factor 10 apart)
ν
Ask an AI about this part
The question quotes this part of the article as it is.
Question copied
Figure 2ExampleFor the convergents of π with denominator from 2 up to 1010 (twenty of them), the exponent ν with error q−ν. Only 22/7 and 355/113 are above 5/2; the others lie between 2 and 2.25.
22/7 and 355/113 stand out because the next terms of the continued fraction, 15 and 292, are large. The larger the next term, the better the convergent before it.
Figure 2 has finitely many points, so it does not decide the irrationality exponent. That the exponent is 2 means that above any line at height ν>2 there are only finitely many points. No amount of numerical computation can check this.
A closer lookWhy the convergents are enough, and the terms of the continued fraction
Ask an AI about this part
The question quotes this part of the article as it is.
Question copied
Two facts from the theory of continued fractions. First, the error of a convergent is essentially set by the next term ak+1:
Second, if a reduced fraction p/q satisfies ∣x−p/q∣<1/(2q2), then p/q is a convergent (Legendre's theorem). If ν>2, then q−ν<1/(2q2) for large q, so the fractions achieving ν are eventually all convergents.
Together they express the irrationality exponent through the growth of the terms:
So μ(π)=2 says that the terms ak+1 of the continued fraction of π grow more slowly than any positive power qkε of the denominators. This is compatible with the terms being unbounded.
Upper bounds so far
Mahler first proved in 1953 that the irrationality exponent of π is finite[4]. The bound has come down step by step since[5,6,7,8,9].
Year
By
Upper bound for μ(π)
1953
Mahler
42 (30 for large denominators)
1974
Mignotte
21 (20 for large denominators)
1993
Hata
8.0160…
2008
Salikhov
7.6063…
2020
Zeilberger, Zudilin
7.1032…
September 2026
Bai (preprint)
7.1018…
September 2026
OpenAI (this result)
2
The improvements since 1993 refine sequences of approximations built from integrals, shaving digits off the bound from 8 to 7.1. The new manuscript does not use these bounds; it goes straight to 2 by a different method. The manuscript itself says that the earlier bounds are context, not inputs to its proof[2, §1.1].
Writing ε=ν−2, this says: for all large denominators, ∣π−p/q∣≥q−2−ε. In Figure 2, above any line at height 2+ε there are only finitely many points.
Three points are worth keeping in mind when reading the theorem. The first two are also made in the manuscript's introduction[2, §1].
The value of Q(ν) is not known. The proof is by contradiction and shows that Q(ν) exists without giving a value. For ν=2.1, say, we do not know from which denominator on the bound holds.
It is not a bound of the form c/q2. “There is a constant c>0 with ∣π−p/q∣≥c/q2 for all q” is a stronger claim. It is equivalent to the terms of the continued fraction being bounded, and does not follow from the theorem.
Good approximations do not stop. Convergents like 355/113 can still appear. What the theorem says is that only finitely many approximations beat any exponent ν>2.
The Flint–Hills series
As an application of the irrationality exponent, the manuscript proves that the Flint–Hills series converges[2, Corollary 1.2]. The question was popularized by Pickover; angles are in radians[10].
Definition 11the Flint–Hills series
Ask an AI about this part
The question quotes this part of the article as it is.
Question copied
n=1∑∞n3sin2n1
With 1/n3 alone the series would converge. The trouble is sin2n: when the integer n is close to a multiple of π, sinn is very small and the term is large.
n=355 is such a case. Since 355 is very close to 113π,
This one term is 24.6, and the partial sum jumps from 4.8 to 29.4 at n=355.
100
101
102
103
104
105
106
0
10
20
30
n=355:+24.6
n=22
30.3145…
N (ticks a factor 10 apart)
SN
Ask an AI about this part
The question quotes this part of the article as it is.
Question copied
Figure 3ExampleThe partial sums SN=∑n≤N1/(n3sin2n) of the Flint–Hills series, with N on a scale where each tick is a factor 10. The sum jumps by 24.6 at N=355, and the partial sum up to N=106 is 30.3145…. This value is a computation; it has not been shown that the limit is 30.3145….
At the next good convergent, 103993/33102, sin103993 is also small, −1.9×10−5. But n3 wins, and the term is only 2.4×10−6. Whether a term gets large is a race between how small sinn is and how large n3 is.
Why 5/2 is the threshold
Let us compare the race in exponents. If ∣π−p/q∣≈q−ν, the sine at the numerator n=p is about
If ν>5/2, the exponent 2ν−5 is positive and the terms grow. If ν<5/2, the terms tend to 0. This suggests that the threshold is irrationality exponent 5/2.
For 355/113, ν=3.202, so 1132⋅3.202−5/π3=1131.404/31.0=24.6, matching the term computed above.
For the necessary condition, this estimate turns directly into a proof[10, Corollary 4, p.3].
The question quotes this part of the article as it is.
Question copied
If the Flint–Hills series converges, then μ(π)≤5/2.
Ask an AI about this part
The question quotes this part of the article as it is.
Question copied
ProofOpenClose
We prove the contrapositive. Suppose μ(π)>5/2 and take 5/2<ν<μ(π). Since ν is achieved infinitely often, infinitely many reduced fractions p/q satisfy ∣π−p/q∣<q−ν.
For these fractions, ∣sinp∣<q1−ν and 1≤p≤4q.DetailsClose
From sin(p−qπ)=±sinp and ∣sint∣≤∣t∣, we get ∣sinp∣≤∣p−qπ∣=q∣π−p/q∣<q1−ν.
Also q≥2 and ν>5/2 give ∣π−p/q∣<2−5/2<1/4, so 2.8<p/q<3.4; in particular 1≤p≤4q.
The denominators q are unbounded and p>2q, so there are infinitely many such numerators p. The terms do not tend to 0, and the series diverges.
The sufficient condition in the other direction was proved by Meiburg in 2022[11, Theorem 2.5, p.4]. The idea of applying approximation bounds to series with sin in the denominator also appears in Mahler's 1953 paper[4]. The new manuscript includes just the form it needs, with a short proof[2, Lemma 5.1].
The question quotes this part of the article as it is.
Question copied
ProofOpenClose
Fix a positive integer K and put d=c(2K)1−ν. If K≤q<2K, then ∥qα∥≥d (since 1−ν<0, q1−ν>(2K)1−ν).
On the circle R/Z of length 1, the points qα with K≤q<2K are also at distance at least d from each other.DetailsClose
For q=q′, the distance between the two points on the circle is ∥(q−q′)α∥. Since 0<∣q−q′∣<2K, the hypothesis gives ∥(q−q′)α∥≥c∣q−q′∣1−ν>c(2K)1−ν=d.
As in Figure 4, two points being close means that their difference q−q′ is the denominator of a good approximation, which the hypothesis forbids.
On each of the two half-circles starting at 0, order the points by their distance from 0. The j-th point is at distance at least jd.DetailsClose
The first point is at least d from 0. For neighbouring points on the same half-circle, the difference of their distances from 0 is the arc between them, which is at least their distance on the circle, d. So the distances are at least d,2d,3d,….
So the sum over this range is bounded by a constant times K2ν−5.DetailsClose
The ranges for K=1,2,4,8,… cover all q≥1. Since 2ν−5<0, ∑m(2m)2ν−5 is a convergent geometric series, and so is the whole sum.
0
8π
9π
10π
11π
12π
13π
14π
15π
∥7π∥
R/Z
(circumference 1)
Ask an AI about this part
The question quotes this part of the article as it is.
Question copied
Figure 4ExampleThe second step of the proof of Lemma 13 for α=π and K=8. On a circle of length 1, the fractional parts of 8π,9π,…,15π. The closest pair is 8π and 15π, at distance ∥7π∥=0.0088 for their difference 7. That 22/7 is a good approximation shows up as two points being close.
Corollary 14the Flint–Hills series converges[2, Corollary 1.2]
Ask an AI about this part
The question quotes this part of the article as it is.
Question copied
n=1∑∞n3sin2n1 converges.
Ask an AI about this part
The question quotes this part of the article as it is.
Question copied
ProofOpenClose
Fix 2<ν<5/2. By Theorem 10, ∥qπ∥≥q1−ν for q≥Q(ν).DetailsClose
With p the integer nearest to qπ, the theorem gives ∥qπ∥=∣qπ−p∣=q∣π−p/q∣≥q⋅q−ν=q1−ν.
Putc=min({1}∪{qν−1∥qπ∥:1≤q<Q(ν)})Then c>0 and ∥qπ∥≥cq1−ν for all q≥1 (since π is irrational, ∥qπ∥>0).
Group each n≥1 by the integer q≥0 nearest to n/π. Only n=1 has q=0. For q≥1, the group has at most four n, each with ∣n−qπ∣≤π/2.DetailsClose
That q is the integer nearest to n/π means ∣n/π−q∣≤1/2, that is ∣n−qπ∣≤π/2. The integers n satisfying this lie in an interval of length π<4, so there are at most four.
Each term in the group q≥1 is at most π2⋅q3∥qπ∥21.DetailsClose
We have n≥π(q−1/2)≥πq/2. Also ∣sint∣≥(2/π)∣t∣ for ∣t∣≤π/2 (since sin is concave on [0,π/2]). As n is an integer, ∣n−qπ∣≥∥qπ∥, and
Convergence follows from the irrationality exponent being less than 5/2; it does not need to be 2. There was a wide gap between the previous bound 7.1 and 5/2, and the new theorem closes it in one go.
A closer lookThe generalization: when ∑1/(na∣sinn∣b) converges
The manuscript also settles exactly when the series with general exponents converges[2, Corollary 5.2].
The question quotes this part of the article as it is.
Question copied
Ask an AI about this part
The question quotes this part of the article as it is.
Question copied
Let a,b>0. The series n=1∑∞na∣sinn∣b1 converges if and only if a>max{1,b}.
Ask an AI about this part
The question quotes this part of the article as it is.
Question copied
ProofOpenClose
If a>max{1,b}, it converges. Take ε>0 with a>max{1,b}+bε, and from ∥qπ∥≥cq−1−ε bound each range as in Lemma 13.DetailsClose
With d=c(2K)−1−ε, the sum over K≤q<2K is at most 2K−ad−b∑j=1Kj−b. For b<1, b=1 and b>1 this is a constant times K1−a+bε, K1−a+εlog2K and Kb−a+bε, all with negative exponents. Then return to the sum over n by the same grouping as in Corollary 14.
If a≤1, it diverges. Since ∣sinn∣≤1, each term is at least n−a, and ∑n−a diverges.
If 1<a≤b, it diverges: the terms at the numerators of the convergents pj/qj do not tend to 0.DetailsClose
From ∣sinpj∣≤∣pj−πqj∣<1/qj and pj≤(π+1)qj for large j,
The question quotes this part of the article as it is.
Question copied
Figure 5ExampleThe range in Corollary 15, drawn in the (b,a) plane. The series converges in the green region a>max{1,b} and diverges outside it, including on its boundary. The Flint–Hills series is the point (2,3).
Divergence for 1<a≤b follows from the property ∣π−pj/qj∣<1/qj2 of convergents alone and does not use the irrationality exponent. Where exponent 2 matters is in pushing the convergent side all the way to the boundary.
How the main theorem is proved
The proof of the main theorem takes up most of the manuscript, and this article does not reproduce it. Here is only its outline. The basic idea is a principle long used in proofs of transcendence: a nonzero integer has absolute value at least 1.
1
Assume many good approximations
Assume that for some ν>2 there are arbitrarily large denominators with ∣π−p/q∣≤q−ν, and derive a contradiction.
2
Pick finitely many with widely separated denominators
Fix a dimension m first, then choose approximations p1/q1,…,pm/qm whose denominators are successively far apart.
3
Build a nonzero determinant by interpolation
Prove an interpolation theorem: the coefficients of a polynomial in several variables can be chosen to prescribe Taylor coefficients along logarithmic curves. Use it to build a nonzero determinant with entries in Q(i).
4
Compare a lower and an upper bound
Clearing denominators makes the determinant a nonzero algebraic integer, bounded below. On the other hand, since 2pi/qi is close to the period 2π, the determinant is very small. The two cannot both hold.
The manuscript relates its method to Laurent's interpolation determinants[2, §1.2]. It also notes that choosing denominators successively far apart is the same kind of device as in the proof of Roth's theorem.
For specialistsInterpolation and the determinant bounds
Ask an AI about this part
The question quotes this part of the article as it is.
Question copied
Give the coordinates Y,X1,…,Xm positive weights and keep only the monomials of weighted degree up to a cutoff; their exponent vectors fill a simplex. At each point (1,cj1,…,cjm), expand after the substitution
The variable t follows the logarithmic curve and the ui are transverse to it. The manuscript's interpolation theorem says that every packet of Taylor coefficients below a weighted threshold can be prescribed at all the centres at once[2, §2]. Its proof uses positivity on a blowup and a weighted curve inequality on possibly singular schemes.
The centres are cji=2ijpi/qi (here i is the imaginary unit), approximations of the periods 2πij of the exponential function. Interpolation gives a matrix over Q(i); clearing the denominators of a nonzero minor gives an arithmetic lower bound.
For the analytic upper bound, each row is moved to the exact period 2πij and expanded. Rows with the same transverse index test Taylor coefficients of the same entire functions, and repeated Taylor degrees make terms of the determinant vanish. This cancellation saves a factor quadratic in the number of rows[2, §3].
Finally, the dimension m, the weights and the centres are fixed first, and only the polynomial degree goes to infinity. Keeping this order avoids a circular choice in which thresholds would depend on the centres[2, §4].
How sure is it?
The manuscript is a preprint dated September 24, 2026, and has not been peer reviewed[2].
The repository's Comparator setup checks the main theorem OAI.PiExponent.main[12]. The permitted axioms are the three standard ones in Lean: propext, Quot.sound and Classical.choice.
The description of the formalization's scope says that the convergence of the Flint–Hills series is outside the selected statement[13]. The same file has a theorem flint_hills_summable, but it is not among the checked statements.
The update of October 7, 2026 neither corrected nor withdrew this manuscript[14].
The statement checked in Lean is below. Its first half is the bound in Theorem 10, and its second half is the definition itself: the supremum of the ν achieved infinitely often is 2.
The set in the second half consists of the reduced fractions r with denominator at least 2 and 0<∣π−r∣<(denr)−ν, which reads as Definition 1 written out. We have not rerun the Lean check ourselves at the time of writing.
A 2025 manuscript also claimed that the Flint–Hills series converges, but it was withdrawn in September 2026 because its proof did not hold[2, §1.1]. The corollary here follows from Theorem 10 by the short argument above, so it stands if the main theorem does.
Open questions
An explicit Q(ν): there is still no bound with values, such as “∣π−p/q∣≥q−2.1 for denominators at least 10100”.
Are the terms of the continued fraction bounded? Whether the terms of the continued fraction of π are bounded (whether some c gives ∣π−p/q∣≥c/q2) is unknown.
Other constants: for many constants, such as log2 and ζ(3), only upper bounds for the irrationality exponent are known. Whether the new method applies to them remains to be seen.