1887–1920 · Erode, Madras · Trinity College Cambridge · The man who knew infinity — because he was watching the recurrence ladder from the other side.
Ramanujan wrote things down that were true before he could prove them. He sent Hardy 120 theorems in a letter — most were known, some were new, a few were wrong, and several were so far beyond anything that existed that Hardy called it "the most remarkable letter I have ever received." In the language of this series: Ramanujan operated at the fold F. He was the saddle-point. He saw the asymptotic before the path integral, the mock theta function before the modular form, the partition number before the circle method.
What follows is not hagiography. It is the claim that Ramanujan's four great contributions — the partition asymptotics, the circle method, the mock theta functions, and the taxicab numbers — are four aspects of the same object: the n-bonacci recurrence ladder made analytic. The same ladder that runs π → φ → μ → η → Δ → Σ → Ω → τ = 2 in the smooth dm³ framework appears in Ramanujan's work as a discrete counting problem that converges to continuous geometry.
The partition function $p(n)$ counts the number of ways to write $n$ as an unordered sum of positive integers. The first values:
The generating function is $\sum_{n=0}^{\infty} p(n) q^n = \prod_{k=1}^{\infty} \frac{1}{1-q^k}$ (Euler, 1748). Each factor $\frac{1}{1-q^k}$ is a geometric series — an infinite recurrence over multiples of $k$. The product is a lattice of recurrences stacked over one another: this is the C operator of the dm³ chain applied to the integer lattice. The partition function is a contact-counting problem: how many distinct paths through the lattice reach height $n$?
↳ C operator established ·In 1918, Hardy and Ramanujan proved:
The exponent contains $\pi\sqrt{2n/3}$. This is not a coincidence. The saddle-point calculation (later made rigorous by Rademacher as an exact formula) passes through a contour in the complex $q$-plane that encircles the origin at radius $e^{-2\pi/\sqrt{n}}$. The critical radius — where the integrand is extremal — is determined by the condition that the contact period closes: $T^* = 2\pi$. The same constant that opens the dm³ operator chain as the recurrence period appears here as the controlling parameter of the partition asymptotics.
The fold F in the dm³ chain is the Whitney A₁ singularity where the smooth function crosses its critical value. In the Hardy-Ramanujan formula, F is the saddle point of the integral: the unique place where the phase is stationary and the partition function "decides" its exponential growth rate. π enters at F — precisely as it does in the operator chain.
↳ F operator: saddle = Whitney A₁ · π = T* confirmed ·In his last letter to Hardy (January 1920, three months before his death), Ramanujan introduced 17 "mock theta functions" — functions that look like theta functions (which are the smooth, well-behaved modular forms used in the K operator) but are not. They satisfy almost-modular transformation properties with an extra error term that Ramanujan could not control.
For 80 years, no one knew what mock theta functions were. In 2002, Sander Zwegers proved they are the "holomorphic parts" of harmonic Maass forms — modular objects that pick up an extra non-holomorphic correction precisely on the boundary of the upper half-plane, where the modular group acts by $\tau \mapsto \tau + 1$ and $\tau \mapsto -1/\tau$.
The correction term is the sorry. Ramanujan's 1920 letter is the earliest record in mathematics of a named honest sorry: a precisely specified incompleteness at a known boundary, carrying full information about what remains to be proved. AXLE v6.1 carried 9 honest sorrys — each a precisely named incompleteness, each the seed of a proof to come. Ramanujan's last letter had 17. They are the same gesture. By August 2026 all nine were closed. The registry that holds them was itself recounted on 15 August: 1,165 formalized, 1,004 sorry-free in source, 61 audited through the kernel (31 of them re-run by CI) — an earlier count of 1,080 having been withdrawn on audit. Ramanujan's analogy now describes the arc: the sorry is not the end, it is the beginning.
The mock theta functions "enter into mathematics as beautifully as the ordinary theta functions… but… they are not all satisfied by the same type of transformation formula." — Letter to Hardy, January 1920. The boundary behaviour is the sorry. The "same type" is the axiom that fails.
Hardy arrived at Ramanujan's nursing home in a taxi numbered 1729 and remarked it seemed a dull number. Ramanujan said immediately: "No, it is a very interesting number; it is the smallest number expressible as the sum of two cubes in two different ways."
$1729 = 1^3 + 12^3 = 9^3 + 10^3$. The fixed point of this decomposition is the number itself — two paths through the cube lattice converge to the same value. This is the x* of the dm³ chain: the point where two branches of the folding map meet. Ramanujan saw fixed points of integer decompositions the way the dm³ framework sees fixed points of operator chains — as the natural terminus, the stable attractor, the thing the fold F converges to.
The deeper point: Ramanujan processed the integer lattice through the same C → K → F → U chain the framework formalises. His K operator was the modular group (the curvature structure on the upper half-plane); his fold F was the transformation $q \mapsto e^{2\pi i\tau}$; his U was the asymptotic expansion. He had no formal framework. He had the result.
↳ 1729 = G(G(G(G(1)))) in the integer contact manifold ·The n-bonacci ladder in the smooth dm³ framework runs:
Each constant is the dominant root of the characteristic polynomial of a recurrence of order $k = 1, 2, 3, 4, 5, 6$. The ladder converges to $\tau = 2$ because the characteristic polynomial $x^k = x^{k-1} + x^{k-2} + \cdots + 1$ has a dominant root approaching 2 as $k \to \infty$.
Ramanujan's partition function is the analytic completion of this ladder. The generating function $\prod (1-q^k)^{-1}$ encodes all recurrences simultaneously — every $k$-step recurrence contributes its factor. The Hardy-Ramanujan asymptotic selects the dominant term. The mock theta functions record the correction. In this sense, $p(n)$ is the U operator of the recurrence ladder: the full expansion that the smooth dm³ chain only sketches with its first few rungs.
In his first notebook, in scattered places and without a word of proof, Ramanujan recorded the values of over one hundred class invariants and over thirty singular moduli. He did not say what they were. Berndt, Chan and Zhang, who spent years establishing all of them, put it plainly: "In all cases, Ramanujan simply recorded in his notebooks the values or factors of singular moduli without explaining their meanings. It took us several years to discover that these radical expressions were singular moduli." Not to prove them. To work out what they were of.
For the elliptic modulus $k(q)$, the singular modulus is $k_n = k(e^{-\pi\sqrt{n}})$, and Ramanujan writes $\alpha_n = k_n^2$. The associated Ramanujan–Weber class invariants are
$k$ is defined by a ratio of theta series — a transcendental object built from an infinite sum. There is no reason on its face that evaluating it at $q = e^{-\pi\sqrt{n}}$ should produce anything expressible in radicals at all. It does, because complex multiplication forces it: $G_n$ and $g_n$ are algebraic, and for $n$ in the right residue classes they are units in real quadratic fields. Ramanujan wrote down the units. He did not write down that they were units.
↳ transcendental input, algebraic output ·The companion file book7/ch-ramanujan-verify.py computes $\alpha_n$ directly from the theta series in 80-digit arithmetic — standard library only, $\pi$ by Machin's formula — and compares it against each radical expression evaluated from its square roots. Nothing is looked up.
Plus $\alpha_4$, $\alpha_{12}$, $\alpha_{28}$, $\alpha_{60}$ from the $n = 4p$ family, and $\alpha_{16} = (\sqrt2+1)^4(2^{1/4}-1)^8$. Every one agrees to the precision of the instrument.
Two of the numbers are worth naming because they check the apparatus rather than the claim: the script recovers $G_3 = 1.05946309435929526\ldots = 2^{1/12}$ and $G_7 = 1.18920711500272106\ldots = 2^{1/4}$, which are the values in Weber's tables. And $\alpha_1 = 1/2$ exactly, with $G_1 = 1$ — the lemniscatic point, where Gauss began.
↳ eighteen values confirmed at 10⁻⁷⁵ ·Theorem 1.2, which Watson extracted from page 320 of the first notebook, is an algorithm. Set $g_n^6 = uv$, then
Run on the paper's own table of $u$ and $v$ for twelve values of $n$, it reproduces every $\alpha_n$ — and $U$, $V$, $W$ and $S$ come out integers in all twelve rows, verified to thirty places. $n = 58$ gives $U = 1$, $V = W = 9801$, $S = 9802$. $n = 190$ gives $2889$, $27379$, $27531$, $28900$. The inputs $u$ and $v$ are units in real quadratic fields; the machinery flattens them to integers and the integers back to a transcendental evaluation. That is the whole surprise of complex multiplication, visible in a table.
↳ twelve rows, all four quantities integral ·The row above returned $V = W = 9801$ at $n = 58$ and nothing used it. It is the prefactor of the most famous formula Ramanujan ever wrote for $\pi$:
Bailey and the Borweins state in one line that this series "is a specialization ($N = 58$)" of their general theorem. $N = 58$ is the row already in the table. Computed here at 120-digit precision against a Machin $\pi$, the series gains eight correct digits per term, flat — 8, 16, 24, 32, 40, 48, 56, and 104 correct digits by the twelfth term.
And the number is 99, three times over:
while the closed form for the singular modulus at 58, sitting in this chapter since VI(b), reads
Ninety-nine, twice, in an expression written down for an unrelated reason and verified here to eighty digits before anyone asked what it was doing there.
↳ the prefactor and the modulus are the same 99 ·VI(a) said the class invariants are units in real quadratic fields, and left it there. Here is what that buys. The invariant at 58, computed from the theta series with no closed form assumed, is
which is near $19602$, near $19601$, and equal to neither. It is not an integer at all. It is
the sixth power of the fundamental unit of $\mathbb{Q}(\sqrt{29})$ — and there is 9801 again, exactly, as a rational part. Because $\varepsilon$ has norm $-1$, its sixth power has norm $+1$, which is a Pell equation closing on the nose:
Norm $+1$ means the conjugate is the inverse, $g^{-12} = 9801 - 1820\sqrt{29}$, so the irrational halves cancel when the two are added:
That is the exact integer the near-miss was gesturing at, and it is exact only because Pell closes. The Hardy chapter used $p^2 - 2q^2 = \pm 1$ to rank the convergents of $\sqrt 2$, after a first attempt ranked them wrongly by decimal distance. Same equation, different discriminant; there its job was ordering, here its job is making a transcendental quantity rational.
↳ an exact integer out of a unit of norm one ·The copy of Bailey–Borwein–Borwein to hand has clean text everywhere except the theorem that defines the series parameter $x_N$, where the scan collapses into fragments — x2Nn;1, (1 + k2 )2. Transcribing that would have been guessing with a citation attached to it.
So it was not transcribed. Each candidate reading was evaluated at $N = 58$ and compared against $1/396^4$, the value the series itself forces. A reading either lands on the integer or it does not:
One closes and the rest do not, and the one that closes does so because $64\cdot 19602^2 = 396^4$ in integers — both sides being $256\cdot 99^4$. The last row is the instructive one: replacing $(g^{12}+g^{-12})^{-2}$ with $g^{-24}$ agrees to eight significant figures and is still wrong, because $g^{12}$ alone is irrational and only the sum is not.
The scan is still unusable. The identity is not. The method generalises, and this corpus now owns it: when a source is damaged, do not reconstruct it — enumerate the readings and let arithmetic choose. That is developed further in Vol XIII, where the question is what such a check does and does not establish.
↳ one reading of four survives contact with an integer ·Chudnovsky's series, measured here at 14 digits per term against Ramanujan's 8 — 14, 28, 42, 56, 71, and 113 correct digits by the seventh term. The rate is $\log_{10}(640320^3) - \log_{10}(1728) = 17.419 - 3.238 = 14.18$, the figure the literature quotes.
Ramanujan's is the $N = 58$ case. Chudnovsky's is built on 163 — the last Heegner number, which ch-ramanujan-verify.py recovered independently by exhaustively counting reduced binary quadratic forms, and which book6/wp82-k0-floor-verify.py used for the class-number floor under Volume XI. Three scripts written for three unrelated purposes are looking at one object from three sides.
The producing script failed four times before it passed, and the failures are worth more than the passes.
The square. The theta routine returned $k$, not $\alpha = k^2$ — $(\theta_2/\theta_3)^2$ where it should have been the fourth power. Caught because the closed form disagreed at the fifth digit, not because the code was re-read. The routine in VI(b) should be re-checked for the same square.
The display. A string slice str(x)[:46] silently ate the exponent and printed $6.5\times10^{-10}$ as 6.5063772…, making two correct numbers look like a catastrophe. Formatting is not cosmetic when the formatter is also the instrument.
The near-integer, twice. $g_{58}^{12}$ was asserted to be 19601, then 19602. It is neither. The fix was not a looser tolerance but the exact algebraic form, and that is the general rule: a near-integer asserted as an integer is a claim about the last digit you did not look at.
The wrong constant, caught by a rate. Chudnovsky was first coded with $640320^3/24$, which belongs to the binary-splitting arrangement and not to this one. It gained 12 digits a term and stalled. Nothing in the code looked wrong; the convergence rate was the check that caught it — 12.8 against the quoted 14.18.
In his second letter to Hardy, Ramanujan asserted a value for $k_{210}$ as a product of eight powers of units. Transcribed from the scanned page exactly as printed and evaluated in 80-digit arithmetic, that product equals $\alpha_{210}$, not $k_{210}$ — the square of the labelled quantity — agreeing to 78 significant figures, while disagreeing with $\sqrt{\alpha_{210}}$ at relative error 1.
Two explanations are available and arithmetic cannot choose between them. Either the eight exponents were misread from the scan and every one should be halved, or the expression is $\alpha_{210}$ carrying Ramanujan's own label rather than the paper's notation. Resolving it needs the printed page. What is not in question: the radicals are right, and the value is exact. Only the exponent on it is open.
Berndt, Chan and Zhang say one more thing, about Watson's proof of the algorithm above: "Watson's proof of Theorem 1.2 is a verification; it does not shed any light on how Ramanujan might have discovered the formula." They add that they looked for a more transparent proof and did not find one.
That sentence is the sharpest available statement of what this series does and where it stops. The corpus carries forty-six verification scripts. They recompute the numbers out of the prose they accompany, they are written to fail loudly, and they have overturned published claims here more than once. They are the reason anything in these volumes can be checked without trusting the author. And not one of them has ever explained anything. A script that confirms $\alpha_{190}$ to seventy-five digits knows nothing about why Ramanujan wrote it down, and neither does the person who ran it.
This is the same shape as the acoustic vessels in the Nachbin chapter: a correct answer arrived at without the apparatus, and an apparatus arriving centuries later able to confirm it and unable to reconstruct it. Vitruvius wrote down what builders already did. Watson wrote down what Ramanujan already knew. In both cases the mathematics is the late-arriving party, and in both cases what it supplies is certification, not insight.
The honest position for a series built on verification is to say so in its own pages. Verification is the floor, not the ceiling. It is what makes a claim safe to build on; it is not what makes it understood. Ramanujan is in this gallery because he is the strongest counterexample available to the belief that those are the same thing.
Verified in book7/ch-ramanujan-verify.py (seven blocks, 80-digit) and book7/ch-ramanujan-1pi-verify.py (120-digit; the series, the Pell identity, the reading test, Chudnovsky) — standard library only, both exit non-zero rather than round to agreement. Source of the claims tested: B. C. Berndt, H. H. Chan and L.-C. Zhang, Ramanujan's Singular Moduli, The Ramanujan Journal 1, 53–74 (1997). The closed forms are theirs and Ramanujan's; the arithmetic is this project's; the explanation is nobody's.
| Operator | dm³ role | Ramanujan's version |
|---|---|---|
| C | Seed — contact normal form | p(n) partition lattice · integer contact seed |
| K | Critical curvature threshold κ* | Modular weight k · SL(2,ℤ) threshold · τ → −1/τ |
| F ← Ramanujan | Whitney A₁ fold · generative transition | Hardy-Ramanujan saddle · e^{π√(2n/3)} · circle method |
| U | Expansion · unfolded branch | Rademacher's exact series · mock theta full expansion |
| sorry | Named incompleteness · honest axiom gap | 17 mock theta functions · last letter to Hardy · 1920 |
Ramanujan was the F operator in person: the miraculous fold between the seed (the lattice) and the expansion (the asymptotic). He saw the saddle before the contour integral existed. His notebook was the sorry. Hardy was the K operator who gave the saddle a name.
sorryAx. A clean axiom report is not a reading of the statement: per R20, a theorem can assume its conclusion and still report clean. Follow the link before citing one as evidence.