The formula
In 1914, Srinivasa Ramanujan published this series for 1/π:
The first term alone (k = 0) already gives
- 2√2 · 1103 / 9801 = 0.31830987844…
- 1/π = 0.31830988618…
eight correct digits from one term. Each further term adds about eight more.
So where do 1103 and 26390 come from? I looked at them with one rule: use integers and addition only. This post collects two of my notes (originally written in Japanese on Zenn) into one story.
Part 1 — 1103 and 26390 are neighbours in an addition-only sequence
26390 = 24 × 1103 − 82
Multiply 1103 by 24 and you overshoot 26390 by exactly 82. That suggests the recurrence
a(n+1) = 24 · a(n) − a(n−1)
Run it backwards and forwards and you get one sequence:
…, 865, 82, 1103, 26390, 632257, 15147778, 362914415, …
1103 and 26390 are adjacent terms. The recurrence only needs addition: 24·a is a doubling ladder (a, 2a, 4a, 8a, 16a; then 16a + 8a), and subtraction is adding the negative.
I also checked one integer that stays fixed along the whole sequence:
a(n+1)² − 24 · a(n) · a(n+1) + a(n)² = −947371
It holds for every adjacent pair I computed (14 terms, from 865 up to a 19-digit number). This is what keeps the sequence on one track.
24 = 11 + 13, a pair of twin primes
The coefficient 24 is the sum of the twin primes 11 and 13. For twin primes there is a tidy identity:
(p + q)² − 4 = 4pq ⇔ |p − q| = 2
because (p + q)² − 4 − 4pq = (p − q)² − 4.
| (p, q) | (p+q)² − 4 | 4pq | |
|---|---|---|---|
| (3, 5) | 60 | 60 | √15 |
| (5, 7) | 140 | 140 | √35 |
| (11, 13) | 572 | 572 | √143 |
| (17, 19) | 1292 | 1292 | √323 |
| (29, 31) | 3596 | 3596 | √899 |
Pairs that are not twins miss: (11, 17) gives 780 vs 748, and (7, 13) gives 396 vs 364.
For our sequence this means the characteristic equation x² − 24x + 1 = 0 has the roots 12 ± √143. The ratio of neighbouring terms heads toward 12 + √143 = 23.9582607431…
Converging without reaching
The animation plots, step by step, how far three sequences still are from their limits (log scale). Every value was computed with integers scaled by 10²⁶⁰ (π itself from Machin's formula, added and subtracted term by term):
| Sequence | Limit | Digits gained per step |
|---|---|---|
| Fibonacci ratios F(n+1)/F(n) | φ | 0.42 |
| 24-recurrence ratios a(n+1)/a(n) | 12 + √143 | 2.8 |
| Ramanujan partial sums | 1/π | 8.0 |
All three are straight lines. Only the slope differs. None of them reaches its limit in finitely many steps; what is left at each step is the residual (in my framework I call it zansa). The residual is not noise. It is the signature of how the sequence is built.
Part 2 — two integers remember π√N
Divide b by a, then by π, then square
Take the two integers of the formula, a = 1103 and b = 26390:
- b / a = 23.92565729…
- π√58 = 23.92565684…
They agree to seven digits. Square b ÷ a ÷ π and you land next to 58:
(26390 ÷ 1103 ÷ π)² = 58.0000022…
58 is the level of this formula (its CM point is τ = i√58 / 2). The same trick works for other Ramanujan-type series:
| Formula | a | b | (b ÷ a ÷ π)² | N |
|---|---|---|---|---|
| Ramanujan (44) | 1103 | 26390 | 58.0000022… | 58 |
| Ramanujan (39), alternating | 1123 | 21460 | 36.9998682… | 37 |
| Ramanujan (43) | 19 | 280 | 22.0043789… | 22 |
| Chudnovsky | 13591409 | 545140134 | 162.99999999999494… | 163 |
(Values computed again for this post with integers scaled by 10²⁶⁰.)
Is that just luck? A control
I drew 20,000 random integer pairs (a up to 10⁵, b up to 10⁶, skipping pairs whose nearest integer is 0) and measured how far (b ÷ a ÷ π)² falls from the nearest integer:
- median distance: 0.25, exactly what you expect from chance
- within 10⁻³: 0.225 % of the pairs
- within 10⁻⁵: none
The real formulas sit at 10⁻²·⁴, 10⁻³·⁹, 10⁻⁵·⁷ and 10⁻¹¹·³, far outside the grey cloud in the animation.
The two parts meet at 24 and 82
Round π√58 to the nearest integer: m = 24. Now compute the "residue" of the formula's pair:
r = m · a − b = 24 × 1103 − 26390 = 82
These are the same 24 and 82 from Part 1. Read from the addition side, 26390 is "24 times 1103, minus 82". Read from the π side, 24 is the integer nearest π√58 and 82 is what is left over.
Where does b/a ≈ π√N come from?
This part is known mathematics, not mine. In the theory of Ramanujan-type series (Chudnovsky; Chan and Yang; the surveys by Zudilin and by Baruah, Berndt and Chan), the factor 1/π comes from the non-holomorphic part of the Eisenstein series E₂, and at the CM point the ratio b/a is fixed by modular data. In my Japanese note I wrote this as one identity valid for both the level-1 (j) and level-2 families, checked it at 50-digit precision on 22 known formulas, and expanded b/a in powers of q with no free parameters. That note has the details and its own negative controls.
What is known, and what is mine
- Known: Ramanujan's 1914 formulas, the Chudnovsky formula, the modular explanation of 1/π, and Y. Zhao's modular derivation of 1103 and 26390 (2020).
- What I add: reading 1103 and 26390 as adjacent terms of the addition-only sequence a(n+1) = 24a(n) − a(n−1), the twin-prime view of 24 and √143, the residual comparison above, and recovering N from two integers with a random control.
- Not claimed: that the twin-prime reading explains Ramanujan's formula. It is a different viewpoint on the same integers, computed with addition alone.
Read more (Japanese)
- The additive sequence (expanded edition): https://zenn.dev/morc_b13/articles/351794ef91db16
- Coefficient ratio and π√N: https://zenn.dev/morc_b13/articles/16abf5780e2c62
References
- S. Ramanujan, "Modular equations and approximations to π", Quart. J. Math. 45 (1914), 350–372.
- D. V. Chudnovsky and G. V. Chudnovsky, "Approximations and complex multiplication according to Ramanujan", in Ramanujan Revisited, Academic Press, 1988, 375–472.
- W. Zudilin, "Ramanujan-type formulae for 1/π: A second wind?", arXiv:0712.1332.
- N. D. Baruah, B. C. Berndt, H. H. Chan, "Ramanujan's Series for 1/π: A Survey", Amer. Math. Monthly 116 (2009), 567–587.
- Y. Zhao, "A modular proof of two of Ramanujan's formulae for 1/π", J. Aust. Math. Soc. 109 (2020), 131–144.
- J. M. Borwein and P. B. Borwein, Pi and the AGM, Wiley, 1987.
Summary
1103 and 26390 are two neighbouring terms of a sequence that needs nothing but addition: a(n+1) = 24a(n) − a(n−1), with 24 = 11 + 13. Its ratio creeps toward 12 + √143 without ever arriving, just as Fibonacci ratios creep toward φ and Ramanujan's partial sums creep toward 1/π. Only the speed differs.
The same two integers also remember the level of the formula: (26390 ÷ 1103 ÷ π)² = 58.0000022…, and random pairs never come that close. Both readings meet at the same numbers. 24 is the integer nearest π√58, and 82 is what is left over.


Top comments (0)