# The exact variance of twin-candidate counts in windows over a primorial period
*Working note, research project, 2026-08-13; §7 added 2026-08-15; §§8–11
restructured 2026-08-28 around the model limit theorem; audited 2026-09-10
and 2026-09-19.
Companion code:
`paper/variance-note-numerics.js`, `research/06-variance-theorem.js`; for §7
`research/natal5-variance.js`, `research/natal-cap-16-fast-variance.js`,
`research/natal-cap-33-overnight.js`; for §8
`research/history/staging/varE-asymptotic.js`; for §9
`research/history/staging/varE-spectral.js` and
`research/history/staging/varE-limit-theorem.js`; for §10
`research/history/staging/varE-theta2-step.js`,
`research/history/staging/varE-theta2-proof.js` and
`research/varE-exact-ladder-01.js`; for §11
`research/history/staging/redteam-0828-varE.js`. Every numerical claim below
is reproduced by one of these, and each is named at the number it produced.*
## Abstract
Let $P = P_n\# = \prod_{p \le p_n} p$ and let $A \subset \mathbb{Z}_P$ be the
set of *twin candidates*: residues $r$ with $\gcd(r,P) = \gcd(r+2,P) = 1$.
We derive the exact pair-correlation function of $A$ (Theorem 1) and from it a
closed, finitely computable formula for the variance of the number of elements
of $A$ in a window of length $L$ whose starting point is uniform on
$\mathbb{Z}_P$ (Theorem 2). The formula is verified against brute-force
enumeration of every window position at two levels. Two consequences: certified
Chebyshev bounds, at each computed level, on the fraction of windows of "zone
length" $L = p_{n+1}^2$ with no twin candidate (below $1.3\times10^{-3}$ at
$p_n = 97$), and the measurement of
a strictly sub-Poisson variance whose ratio to the mean follows an empirical
scaling law in the window exponent. Section 7 restricts the correlation to two
of the three mod-30 houses and evaluates it on the diagonal $L = W = x\#$,
$y \approx \sqrt{W}$, at nine exact levels reaching $W = 7.4\times10^{12}$.
The limit of that ratio is the note's open question, and it stays open. What
§§8–11 add is a decomposition of it. On the diagonal
$\operatorname{Var}/\mathbb{E} = \delta X$ exactly, with
$\delta \ln^2 W \to 16C_2e^{-2\gamma}/3 = 1.109905$ and with the
Montgomery–Soundararajan main term of $X$ equal to $L$ prime by prime, so all
of the ratio is a discrepancy (§8, proven). A decoupled model of the
discrepancy has its own limit, and that limit is a theorem: the model's
conductor law converges to the generalized Dickman law $GD(2)$ and its value
converges to $\lambda_2(u) = \Pr[GD(2) > u]$ with no correction term, giving
the closed form $\lambda_2(2) = 1 - e^{-2\gamma}(9/2 - 4\ln 2) = 0.45546$ at
the diagonal exponent (§9, proven). At one excluded class per prime the same
construction returns Gorodetsky's $\lambda(u)$ as an identity, where the
corresponding asymptotic is published. That the true variance shares the
model's limit is **conjectured, not proven** (§10): the replacement error is
measured at eight exact levels, falling to $1.000471$ at $x = 31$, two of its
three lag groups are closed unconditionally, and the third reduces to a
statement about $y$-smooth divisors of $C(C^2-4)$, searched in its owning
convention on 2026-08-28 and 29 without a matching theorem among the results
inspected, the one candidate chapter still unread at the page (§10). Section 11
records that the earlier fitted reading
$0.611$ is refuted as an inference: a control sequence with a known, different
limit passes both halves of the protocol that produced it.
## 1. Setup
Fix $n \ge 2$ and write $P = P_n\# = \prod_{p\le p_n} p$. For every odd prime
$p \le p_n$ the condition $\gcd(r(r+2), p) = 1$ excludes exactly the two
residue classes $r \equiv 0$ and $r \equiv -2 \pmod p$; for $p = 2$ the two
classes coincide and exclude the even residues. By the Chinese Remainder
Theorem,
$$|A| \;=\; \prod_{2 < p \le p_n} (p-2), \qquad
\delta \;:=\; \frac{|A|}{P} \;=\; \frac{1}{2}\prod_{2
1$ when $p \mid d$, $p(p-3)/(p-2)^2 > 1$
when $d \equiv \pm 2$ and $p \ge 5$ (at $p = 3$ that factor is $0$), and
$p(p-4)/(p-2)^2 = 1 - 4/(p-2)^2 < 1$ generically for $p \ge 5$.
$W$ is the finite-level, two-class analogue of the Hardy–Littlewood singular
series.
**Lemma (sum rule).** $\sum_{d \bmod P} J(d) = \delta^2 P$, and hence
$\sum_{d \ne 0} \big(J(d) - \delta^2\big) = -\,\delta(1-\delta)$, exactly.
*Proof.* $\sum_d \#\{r: r \in A, r+d \in A\} = |A|^2$, since every ordered
pair $(r,s) \in A^2$ is counted once, at $d = s - r$. The second identity
follows from $J(0) = \delta$. $\square$
The sum rule says the total off-diagonal anticorrelation exactly balances the
diagonal: consistently, the full-period window $L = P$ has variance $0$
(the count is deterministic).
## 3. The exact variance
**Theorem 2.** For $1 \le L \le P$ and $t$ uniform on $\mathbb{Z}_P$,
$$\mathbb{E}[N_L] = \delta L, \qquad
\operatorname{Var}[N_L] \;=\; \sum_{|d| < L} \big(L - |d|\big)\,\big(J(d) - \delta^2\big).$$
*Proof.* Linearity gives the mean. For the second moment,
$\mathbb{E}[N_L^2] = \sum_{0\le i,j u]$, worth $0.45546$ at $u = 2$ (§9), while the
identification of the true variance with that model stays conjectural (§10).
So the open question has one named unproven step in place of an empty page,
and the note's main unfinished business is that step.
## 7. The comb-restricted correlation, and the diagonal window
Theorem 1's proof is indifferent to which admissible classes we keep modulo
the small primes. Restricting to the two classes $11, 17 \pmod{30}$ (the
Natal@5 comb: two of the three mod-30 houses of twin candidates, carrying two
thirds of the census) gives the working correlation of the anchored programme.
**Corollary 3 (comb-restricted pair correlation).** Fix a sieve level $y$ and
let $M = \prod_{p \le y} p$. Put
$$A_5 = \{\, r \in \mathbb{Z}_M : r \equiv 11 \text{ or } 17 \ (\mathrm{mod}\ 30),\ \
r \bmod p \notin \{0, -2\} \ \text{ for every prime } 7 \le p \le y \,\},$$
so that $\delta = |A_5|/M = \tfrac{2}{30}\prod_{7 \le p \le y}(1 - 2/p)$. Then
the pair correlation $J_5(d) = \frac{1}{M}\#\{r : r \in A_5,\ r + d \in A_5\}$
factors as
$$J_5(d) \;=\; \frac{\rho_{30}(d)}{30} \prod_{7 \le p \le y} \frac{\rho_p(d)}{p},
\qquad
\rho_{30}(d) = \begin{cases} 2, & d \equiv 0 \pmod{30},\\
1, & d \equiv \pm 6 \pmod{30},\\ 0, & \text{otherwise,}\end{cases}$$
with $\rho_p$ exactly as in Theorem 1.
*Proof.* Identical to Theorem 1 at the primes $7 \le p \le y$. At the modulus
$30$ the comb has two teeth, so $\rho_{30}(d)$ counts the teeth $r$ with
$r + d$ also a tooth: both when $30 \mid d$, one when $d \equiv \pm 6$ (the
teeth are $6$ apart), none otherwise. $\square$
The comb kills every gap class outside $d \equiv 0, \pm 6 \pmod{30}$, which is
why the correlation sum is a sum over one class in ten rather than over all
$d$, and it is also the source of an exact cancellation used elsewhere in the
programme: the first five correlation lags vanish identically, so the
adjacent-window correlation constant is $-1/2$ with no error term
(`research/natal-cap-26-minus-half.md`, Theorem 2).
**The sum rule, verified exactly.** The local sums are
$S_{30} = 4/30 = 2^2/30$ (one residue class at $d \equiv 0$ carrying $2/30$,
two at $d \equiv \pm 6$ carrying $1/30$ each) and
$S_p = \big((p-2) + 2(p-3) + (p-3)(p-4)\big)/p = (p-2)^2/p$. Hence
$M \prod_i S_i = \delta^2 M^2$ identically, which is the sum rule of §2 for
$A_5$, and it has been checked in exact BigInt arithmetic at $y = 13, 47, 101$
(`research/natal-cap-16-fast-variance.js`, Part 1). Equivalently
$\operatorname{Var}[N_M] = 0$: a whole-period window always holds exactly
$|A_5|$ comb members. All variance at $L < M$ is truncation variance.
**The diagonal.** The window length that the twin problem actually asks about
is $L = W = x\#$ with sieve level $y = $ the largest prime $\le \sqrt{W}$, so
that $u = \ln L / \ln y = 2$ in the limit, and $2.0847$, $2.0116$, then
$2.000x$ from $x = 13$ up, at the computed levels
(`research/history/staging/redteam-0828-varE.js`; the earlier reading "$= 2$
exactly" was wrong, since $y$ is the largest prime *below* $\sqrt{W}$, and
from $x = 13$ up the coordinate error moves the limit function $\lambda_2$ of
§9 by less than $0.001$). The diagonal is §6's zone exponent, evaluated at
levels far beyond the reach of the $y \le 2003$ table. Nine exact points, each
computed by the product-sieve route (the direct $O(W \cdot \pi(y))$ sum is
months of compute at the deepest levels; the sieve is
$\approx 0.3\,W \ln\ln\sqrt{W}$ and reproduces the direct values inside their
certified bars wherever both exist):
| $x$ | $W$ | $\mathbb{E}[N_W]$ | $\operatorname{Var}$ | $\operatorname{Var}/\mathbb{E}$ |
|---|---|---|---|---|
| 7 | 210 | 6.92 | 1.05 | 0.1521 |
| 11 | 2,310 | 39.27 | 10.06 | 0.2563 |
| 13 | 30,030 | 304.28 | 91.13 | 0.2995 |
| 17 | 510,510 | 3,245.51 | 1,060.54 | 0.3268 |
| 19 | 9,699,690 | 41,441.19 | 14,392.59 | 0.3473 |
| 23 | 223,092,870 | 669,028.80 | 243,740.37 | 0.3643 |
| 29 | 6,469,693,230 | 14,063,617.40 | 5,307,862.63 | 0.3774 |
| 31 | 200,560,490,130 | 328,601,798.62 | 127,363,168.00| 0.3876 |
| 37 | 7,420,738,134,810 | 9,377,228,928.8 | 3,711,451,136 | 0.3958 |
Certified roundoff bars: $\pm 3.1\mathrm{e}{-3}$ at $x = 23$, $\pm 1.4$ at
$29$, $\pm 7.4\mathrm{e}2$ at $31$, $\pm 8.2\mathrm{e}5$ at $37$ (relative to
$\operatorname{Var}$: $1.3\mathrm{e}{-8}$, $2.6\mathrm{e}{-7}$,
$5.8\mathrm{e}{-6}$, $2.2\mathrm{e}{-4}$).
The nine points are the note's deepest data, and what follows is an analysis of
them rather than more of them. The tenth level, $x = 41$, costs roughly forty
times the last one and is out of reach of the current engine. Section 10 records what
it would and would not decide.
## 8. The spectral form of $\operatorname{Var}/\mathbb{E}$ on the diagonal
Everything in this section is proven, and none of it produces a constant. It
says where the constant has to come from. The producer is
`research/history/staging/varE-asymptotic.js`, and the nine points of §7 are
read rather than recomputed. Notation, once: $W(d)$ with an argument is the singular-series
factor of §2, and $W = x\#$ without one is the window length of §7.
**The normalisation.** Since $\sum_{|d| L$ if and only
if $D_y > u_y$.
**Theorem 7.** For every fixed $s > 0$,
$$\ln \mathbb{E}\big[e^{-sD_y}\big] \;=\; 2\int_0^1 \frac{e^{-sw}-1}{w}\,dw \;+\; O_s\!\Big(\frac{\ln\ln y}{\ln y}\Big),$$
and consequently $D_y \to GD(2)$ in distribution, the generalized Dickman law
of index $2$.
*Proof.* Put $w_p = \ln p/\ln y \in (0,1]$ and $\phi_p = e^{-sw_p} - 1$, so
that $\mathbb{E}[e^{-sD_y}] = \prod_{p\le y}(1 + \pi_p\phi_p)$ with every
factor positive.
(i) *Linearisation.* $|\pi_p\phi_p| \le 1/3$ for $p \ge 7$, and
$|\pi_p\phi_p| \le s\ln 5/\ln y \le 1/2$ for $p \in \{2,3,5\}$ once
$\ln y \ge 2s\ln 5$, so every factor admits $\ln(1+z) = z + O(z^2)$. Since
$\sum_p (\pi_p\phi_p)^2 \le 3(s\ln5/\ln y)^2 + (s^2/\ln^2 y)\sum_{p\ge7}4\ln^2p/(p-1)^2
= O_s(1/\ln^2 y)$, we get
$\ln\mathbb{E}[e^{-sD_y}] = \sum_p \pi_p\phi_p + O_s(1/\ln^2 y)$.
(ii) *The exact $\pi_p$ against the rate $2/p$.* Write $\pi_p = 2/p + \epsilon_p$,
so $\epsilon_p = 2/(p(p-1))$ for $p \ge 7$ and $|\epsilon_p| \le 1$ for
$p \in \{2,3,5\}$. With $|\phi_p| \le \min(1, sw_p) \le s\ln p/\ln y$,
$|\sum_p \epsilon_p\phi_p| \le (s/\ln y)\sum_p|\epsilon_p|\ln p = O_s(1/\ln y)$,
the tail converging. This is the only place the value $2$ enters, and it enters
as a convergent sum: the deviation of the model's inclusion rate from $2/p$ is
summable against $\ln p$ and cannot move the limit.
(iii) *Mertens.* Fix $t_0 = 3$. The primes below $t_0$ contribute
$O_s(1/\ln y)$. Above $t_0$, write $\sum_{p\le t}1/p = \ln\ln t + M + r(t)$
with $|r(t)| \le c/\ln t$ unconditionally, and $\phi(w) = e^{-sw}-1$. Then
$\sum_{t_0 0$ gives convergence in
distribution. $\square$
Two remarks. The rate is measured as well as asserted: the model's exact
transform against $GD(2)$'s at $s = 0.5, 1, 2, 4$ on all nine levels has error
times $\ln y/\ln\ln y$ bounded, which is the predicted shape and not a settled
constant. And $GD(2)$ has a bounded density, $p_2(t) = e^{-2\gamma}t$ on
$(0,1]$, with a continuous distribution function, which is what Theorem 8 uses.
**Theorem 8.** Let $y \to \infty$ with $u_y \to u > 0$. Then
$$\mathbb{E}[g] \;\longrightarrow\; \Pr[GD(2) > u] \;=:\; \lambda_2(u),$$
with no correction term: the transition region $n \asymp L$, where $g$ is
neither $0$ nor $1$, contributes nothing in the limit.
*Proof.* Read the three bands off $g(n) = \{L/n\}(1-\{L/n\})n/L$. If $n \mid L$
then $g(n) = 0$ exactly. If $n > L$ then $\{L/n\} = L/n$ and $g(n) = 1 - L/n$
exactly. If $n < L$ and $n \nmid L$ then $0 < g(n) \le n/(4L)$. So with
$P_> = \Pr[n > L]$, $R = \mathbb{E}[(L/n)\mathbf 1_{n>L}]$ and
$B = \mathbb{E}[g\,\mathbf 1_{n \;-\; R \;+\; B, \qquad \text{all four exact.}$$
Since $n > L$ iff $D_y > u_y$, and $GD(2)$ has a continuous distribution
function, Theorem 7 and Pólya's theorem give $P_> \to \lambda_2(u)$. Both
remainders are dominated by
$\mathcal{M} := \mathbb{E}[\min(n/L, L/n)] = \mathbb{E}[e^{-\ln y\,|D_y - u_y|}]$,
because $R \le \mathcal{M}$ and $0 \le B \le \mathcal{M}/4$. Fix $\epsilon > 0$. On $|D_y - u_y| > \epsilon$
the integrand is at most $y^{-\epsilon} \to 0$; on $|D_y - u_y| \le \epsilon$ it
is at most $1$, and
$\limsup_y \Pr[|D_y - u_y| \le \epsilon] = \Pr[|GD(2) - u| \le \epsilon] \le 2\epsilon\sup p_2$.
So $\limsup \mathcal{M} \le 2\epsilon \sup p_2$ for every $\epsilon$, hence
$\mathcal{M} \to 0$ and both remainders vanish. $\square$
What this settles. The closed form omits nothing: the transition band
is a two-sided window of width $O(1/\ln y)$ in $D_y$ and dies against a bounded
limiting density, so no term is added to $\lambda_2(u)$ and none subtracted. No
local limit theorem is needed, only weak convergence and boundedness. And
nothing in either proof uses the value $\theta = 2$: step (ii) is the only place
it enters, and it enters as summability, so any excluded-class count $\theta$
gives $\lambda_\theta(u)$ by the same two arguments.
**Lemma 9 (block bound).** Let $h$ be multiplicative, supported on squarefree
integers whose prime factors lie in $[5,y]$, with $h(5) = 3$ and
$h(p) = 2/(p-3)$ for $7 \le p \le y$. Then
$\sum_{N < m \le eN} h(m) \le K(1 + \ln 2N)$ for every $N \ge 1$, with $K$
absolute and in particular uniform in $y$.
*Proof.* $h(m) \le c(m)/m$ with $c = m\,h(m)$ multiplicative, squarefree
supported, $c(5) = 15$ and $c(p) = 2 + 6/(p-3)$ for $7 \le p \le y$, so the
block is at most $N^{-1}\sum_{m\le eN}c(m)$. Let
$q_y = \mu^2\,2^{\omega}\,\mathbf 1[y\text{-smooth, coprime to }6]$; since
$\mu^2(m)2^{\omega(m)} = \sum_{ab=m}\mu^2(a)\mu^2(b)$ for squarefree $m$,
$\sum_{m\le M}q_y(m) \le \sum_{a \le M}\lfloor M/a\rfloor \le M(1+\ln M)$.
Write $c = q_y * b$; locally $\sum_k b(p^k)z^k = (1+c(p)z)/(1+2z)$, so
$b(p^k) = (-2)^{k-1}(c(p)-2)$ and
$\sum_k |b(p^k)|/p^k = |c(p)-2|/(p-2) = 6/((p-3)(p-2))$ for $7 \le p \le y$,
summable over primes, with the single prime $5$ contributing a finite factor.
Hence $\sum_m |b(m)|/m \le K_0 < \infty$ absolutely, and
$\sum_{m\le M}c(m) = \sum_{d\le M}b(d)\sum_{e\le M/d}q_y(e) \le K_0\,M(1+\ln M)$. $\square$
**Lemma 10 (local bound).** Uniformly in $T > 0$ and $y$,
$\Pr[\ln n \in (T-1, T]] \le K'(1+T)/\ln^2 y$.
*Proof.* The support is $n = 6m$ with $m$ as in Lemma 9, and the weight is
$w(6m) = C_y h(m)$ with
$C_y = \prod_{5\le p\le y}(1-\pi_p) = \frac14\prod_{7\le p\le y}\frac{p-3}{p-1}$.
Since $(p-3)/(p-1) = (1-1/p)^2\big(1 - (3p-1)/(p-1)^3\big)$ and
$\sum_p (3p-1)/(p-1)^3$ converges, Mertens gives $C_y \asymp 1/\ln^2 y$. Apply
Lemma 9 with $N = e^{T-1}/6$. $\square$
**Proposition 11 (the two transition bands).** With $L = y^{u+o(1)}$ and $u$
fixed, $R = O_u(1/\ln y)$ and $B = O_u(1/\ln y)$, hence
$\mathbb{E}[g] = \Pr[n > L] + O_u(1/\ln y)$.
*Proof.* Bin by $j \ge 0$ and use Lemma 10. For the lower band,
$\mathbb{E}[(n/L)\mathbf 1_{n u]$ a new function would be wrong. What
we have not found in print, searched in the owning convention (the variance of $y$-rough integers in short intervals, Gorodetsky's; `research/SEARCH-CONVENTIONS.md` §1) is the identification of that tail with a sifted
window variance at $\theta > 1$, and the search behind that sentence is
incomplete: Gorodetsky's paper is $\kappa = 1$ throughout (checked at source,
with `tuple`, `twin`, `admissible`, `generalized`, `dimension` and `Poisson`
absent from the full text), Aryan's Lemma 1.2 is the same object at general
tuple size with an upper bound and no asymptotic, and the successor question is
owed four channels. Nothing here is claimed as new.
**The three bands, exact at nine levels.** The decomposition
$\mathbb{E}[g] = P_> - R + B$ of Theorem 8's proof is computable without
enumerating anything above $L$: the conductors $n \le L$ are enumerated one at a
time by descent over the primes, and $\mathbb{E}[1/n] = \delta$ supplies the
rest. There is no Monte Carlo in the columns below
(`research/history/staging/varE-limit-theorem.js`, 4930 s, of which 4792 s is
the $x = 37$ level; `research/varE-exact-ladder-01.js` is an independent
second route).
| $x$ | $\Pr[n \mid L]$ | $P_>$ | $R$ | $B$ | $\mathbb{E}[g]$ |
|---|---|---|---|---|---|
| 7 | 0.666666667 | 0.283333333 | 0.108197358 | 0.003056943 | 0.178192918 |
| 11 | 0.417387328 | 0.335525028 | 0.081053290 | 0.012367138 | 0.266838875 |
| 13 | 0.296671934 | 0.350725586 | 0.066478692 | 0.012344728 | 0.296591623 |
| 17 | 0.212414912 | 0.368116926 | 0.052466671 | 0.009853657 | 0.325503912 |
| 19 | 0.160547361 | 0.382620145 | 0.043864663 | 0.008044066 | 0.346799547 |
| 23 | 0.123952461 | 0.394436926 | 0.037234816 | 0.006640313 | 0.363842423 |
| 29 | 0.096758571 | 0.403580143 | 0.032052128 | 0.005616554 | 0.3771445 |
| 31 | 0.078138150 | 0.410607924 | 0.028055939 | 0.004855050 | 0.387405 |
| 37 | 0.063810103 | 0.416225089 | 0.024823834 | 0.004252203 | 0.39566 |
The last three values of $\mathbb{E}[g]$ are quoted short because $R$ is a
difference of two quantities of size $L\delta = 9.4\times10^9$ at $x = 37$ and
carries a float floor. The witness spread across two independent routes is
$7.5\times10^{-8}$, $3.2\times10^{-6}$, $1.3\times10^{-5}$ at $x = 29, 31, 37$.
Every other column is a direct sum with no cancellation and agrees between the
two routes to every printed digit, and the recomposition $P_> - R + B$ returns
$\mathbb{E}[g]$ with residual $0$ at all nine levels.
Three readings. The inert band is not a small set, carrying $6.4\%$ of the
conductor mass at $x = 37$, and it contributes exactly zero at every finite
level rather than approximately zero. The lower transition band is
$B\ln y = 0.063616, 0.064678, 0.064691, 0.063820, 0.063440, 0.063175, 0.063008$
at $x = 13 \ldots 37$: bounded, as Proposition 11 requires, and settling on
$0.0630$ against the local-density prediction $0.062729$, while its share of
$\mathbb{E}[g]$ falls from $4.2\%$ to $1.1\%$. And the model's whole distance
from its own limit is $(\lambda_2(2) - \mathbb{E}[g])\ln y = 0.817, 0.847,
0.872, 0.880, 0.884, 0.885, 0.886$ over the same levels, settling on $0.886$.
That last number is a measurement, not a derived rate, and it prices how far
the computable levels sit from the constant: $\mathbb{E}[g] = 0.45$ needs
$\ln y \approx 162$, that is $W \approx e^{325}$. No level of the present kind
comes near $0.4555$, and no measurement of the constant exists to check it
against. The model's own value at the tenth level is on record,
$0.402368 \pm 3.4\times10^{-5}$ at $x = 41$ by Monte Carlo on the same conductor
law (`research/varE-exact-ladder-01.js`, SEC 3), the band a tenth diagonal
point would be read against.
**The model against the data, with no fitted parameter.** The model's only
inputs are $\alpha_p$, $y$ and $L$. Against §7's nine diagonal points the
residuals, measured minus model, are $-0.0261$ and $-0.0106$ at $x = 7, 11$,
then $+0.0029$, $+0.0013$, $+0.0005$, $+0.0005$, $+0.0003$, $+0.0002$ from
$x = 13$ to $31$, all exact on both sides, and $+0.0001$ at $x = 37$ where the
measured column's own rounding is $\pm 5\times10^{-5}$. Against §6's six-point $u$-sweep at $y = 401$, which is a
different object (the full set $A$, all three mod-30 houses), it returns
$0.8543, 0.7025, 0.4842, 0.2927, 0.1576, 0.0764$ (Monte Carlo, $N = 2\times10^5$,
bars $5$ to $10\times10^{-4}$; `research/history/staging/varE-spectral.js`,
Part 5) against the note's $0.845, 0.685, 0.477, 0.290, 0.157, 0.076$; against
§6's seven-point $u = 2$ level sweep it returns $0.2541 \ldots 0.3232$, same
Monte Carlo, against $0.251 \ldots 0.321$. The model is close over twenty-two
points and is exact nowhere.
## 10. The identification with the true variance: conjectured, with one named gap
**Conjecture 1.** On the diagonal $L = W = x\#$, $y$ the largest prime
$\le \sqrt W$, the decoupling replacement is asymptotically harmless:
$$\delta\,\big(X(L) - X_{\rm dec}(L)\big) \;\longrightarrow\; 0 ,
\qquad\text{hence}\qquad
\lim \frac{\operatorname{Var}}{\mathbb{E}} \;=\; \lambda_2(2) \;=\; 1 - e^{-2\gamma}\Big(\tfrac92 - 4\ln2\Big) \;=\; 0.45546 .$$
This is conjectured, not proven, and it is the one step between §9's theorem
and any statement about the variance of §7's table. The chain is: link 1,
$\operatorname{Var}/\mathbb{E} = \delta X$, proven (§8); link 2, this
conjecture, open; link 3, $\delta X_{\rm dec} = \mathbb{E}[g(n)]$, proven given
the replacement (§9); link 4, $\mathbb{E}[g] \to \lambda_2(u)$, proven
(Theorem 8); link 5, the closed form, proven. Exactly one link is unproven, and
it is the one that connects the model to the arithmetic.
**What the error is, exactly.** In real space the replacement flattens
$W$ to $V(h) = \prod_{p \le y}\big(1 + m_p\,c_p(h)\big)$ with $c_p(h) = p - 1$
if $p \mid h$ and $-1$ otherwise (the Ramanujan sum), that is
$$V(h) = 6\cdot\mathbf 1[6 \mid h]\Big(\tfrac58 + \tfrac{15}{8}\mathbf 1[5 \mid h]\Big)\,
C_y\prod_{p \mid h,\ 7\le p\le y}\frac{p-1}{p-3},
\qquad C_y = \prod_{7\le p\le y}\Big(1 - \frac{2}{(p-1)(p-2)}\Big);$$
the factors at $2$, $3$ and $5$ are part of the kernel (an earlier version of
this display kept only the product over $p \ge 7$; at $x = 7$ the all-prime
form returns $\delta X_{\rm dec} = 0.178192918$, the table's value, and the
truncated one $0.001459$), so
$$X - X_{\rm dec} \;=\; \sum_{|h| 2L\\ P^+(n) \le y}}\!\! w(n,C)
\;=\; L\sum_{n>2L}\sum_{c\ne0}\frac{w(n,c)}{n} \;+\; O\big(\ln y \cdot o(\ln y)\big),$$
that is: the weighted count of large $y$-smooth divisors of $C(C^2-4)$ agrees
with its expected count to one logarithm better than trivially, on average over
a window of length $L = y^2$. The owning conventions for that are the
distribution of divisors of a polynomial value in a dyadic range: Ford's
$H(x,y,z)$, Hooley's $\Delta$-function, and the Erdős multiplication-table
problem, applied to the reducible cubic $C(C-2)(C+2)$ with a smoothness
constraint at $u = 2$. That search was run on 2026-08-28 and 29
(`research/history/staging/lit-smooth-divisors.md`,
`research/history/staging/lit-scourfield-2008.md`; the row for large
$y$-smooth divisors of $C(C^2-4)$ in `research/SEARCH-CONVENTIONS.md` §1): the
no theorem among the results inspected gives the weighted, above-range
estimate needed. The located set counts (Tenenbaum's $H_F$, Ford's $H(x,y,z)$)
sit at divisor ranges strictly below the sampling range, while our divisors
satisfy $n > 2L$, above it. Scourfield's 2008 result is a divisor sum, every
divisor $m \le x$ of $f(n)$ counted with weight $1$ over $n \le x$, not a set
count; read second-hand through her 2016 restatement, with the chapter itself
unread at the page and a degree hypothesis unresolved between the two
statements, it carries relative error $O(1/\log x)$ where this step needs
$o(1/\ln y)$, one epsilon short. So the reduction is not claimed as new, and no
absence is asserted while the chapter is unread; the same one-logarithm gap is
usually the whole difficulty in divisor problems of this shape.
One route through the exact kernel is already lost. The only split it offers,
flat against active, has $\delta\,|{\rm flat}|$ rising on all six levels where
it is computed, $0.559, 0.912, 0.960, 1.087, 1.188, 1.301$ against a target of
zero, so every absolute-value argument through that split fails before it
starts.
**What would falsify Conjecture 1**, and whether the check has run: the
conjecture asserts $\delta(X - X_{\rm dec}) \to 0$, so it is killed by the
unscaled error staying away from $0$ along an unbounded sequence of levels, or
by a proof that the replacement shifts the limit; growth of the scaled quantity
$\delta(X - X_{\rm dec})\ln y$ at a level would not kill it (an error of size
$1/\sqrt{\ln y}$ grows when scaled and still vanishes), and eight exact levels
show the scaled quantity bounded and falling, which excludes nothing about the
limit; neither killer exists either way. The tenth level $x = 41$ has been priced
and declined, and by the registered forecast band it would retire one drift form
and separate nothing else.
## 11. Calibration: the fitted reading $0.611$, and why it is refuted
An earlier version of this note read the ninth level as favouring one drift law
over another and reported its intercept, $0.611$, as the value the open question
had come down to. That inference is refuted, on both halves of the protocol that
produced it, and a reader of the old version needs the reason.
The two forms were $r = 0.4435 - 1.509/\ln W$ (rms $1.8\mathrm{e}{-3}$) and
$r = 0.6106 - 0.729/\ln\ln W$ (rms $8.2\mathrm{e}{-4}$), fitted before $x = 37$
was computed. At $x = 37$ they predict $0.3926$ and $0.3955$ against the
measured $0.3958$, a factor of ten in residual, and refitting improves the
second while degrading the first. Two corrections to that account, then the
refutation. The published coefficients are the six points $x = 13 \ldots 31$ and
the seven points $x = 13 \ldots 37$, not "the first eight" and "all nine" as the
old text said; on the eight points $x = 7 \ldots 31$ the $1/\ln\ln W$ form is
the *worse* of the two. And on the six points actually used, two further
two-parameter forms fit at least as well, with intercepts $0.5387$ and $0.7219$,
while a three-parameter member of the $1/\ln W$ family fits better than all of
them with intercept $0.4642$.
The refutation is a control, not an argument. Section 9's model is a sequence
that tracks these nine points to $0.003$ and whose limit is $0.455456$ by
construction. Run the same in-sample protocol on it and $1/\ln\ln W$ wins, with
intercept $0.6164$ against $1/\ln W$'s $0.4468$: the same winner, the same
margin, and the same two numbers the data give, on a sequence whose limit is
neither. Run the frozen half, fitting on $x = 13 \ldots 31$ and forecasting
$x = 37$, and the control reproduces that too, $1/\ln\ln W$ residual $-0.00039$
against $1/\ln W$'s $+0.00276$, resolvable against the model's own
$0.395567 \pm 0.000106$. So neither the in-sample winner nor the out-of-sample
separation carries information about the limit
(`research/history/staging/redteam-0828-varE.js`). Bias-corrected against the
control the seven forms' intercepts cluster in $[0.4471, 0.4546]$ around
$\lambda_2(2)$, and that is not independent evidence either: the intercept is a
linear functional $a = \sum_i c_i y_i$ with $\sum_i c_i = 1$, so the corrected
column is bounded a priori by $\|c\|_1\max_i|{\rm data}-{\rm model}|$, and every
observed gap sits inside that bound. It restates that the model tracks the data
pointwise.
Stated plainly: this was model comparison, not measurement of a limit. Over the
entire computable range $1/\ln\ln W$ moves only from $0.295$ to $0.596$, so the
forms are separated on a short lever arm, and no finite computation of this kind
distinguishes one limit from another. Neither $0.611$ nor the $0.44$ reading is
measured here, and no value of $\lim\operatorname{Var}/\mathbb{E}$ is measured
anywhere in this note. What replaces the fitted number is the reduction of §8,
$\lim\operatorname{Var}/\mathbb{E} = \kappa\lim X/\ln^2 W$ with
$\kappa = 1.109905$ proven, together with §9's proven model limit and §10's one
open link.
## Authorship and AI disclosure
Sole author: Chris Benjaminsen.
> The framework, vocabulary, and driving questions are the author's,
> developed over six years of independent work. Formal derivations,
> literature audits, computations, and manuscript drafting were carried out
> using an AI assistant operating under the author's
> direction; all results were verified by explicit computation, with code
> and outputs published in the accompanying repository, and all refuted
> intermediate claims retained in the record.
Sections 8 to 11 were produced the same way, on 2026-08-28, and carry two
qualifications of their own. The derivations of Theorem 7, Theorem 8,
Lemma 9, Lemma 10 and Proposition 11, and both closed groups of §10, are
written out and checkable here, and each has been put against exact
computation at nine levels by the producers named at the number. None has yet
had an external referee's pass. And the earlier readings this restructure
replaces, the $0.611$ intercept of §11 and the "$u = 2$ exactly" of §7, were
both corrected by the project's own adversarial pass of the same date, whose
record is `research/history/staging/redteam-0828-varE.md`.
## References
- M. Hausman, H. N. Shapiro, *On the mean square distribution of primitive
roots of unity*, Comm. Pure Appl. Math. 26 (1973), 539–547.
- H. L. Montgomery, R. C. Vaughan, *On the distribution of reduced residues*,
Ann. of Math. (2) 123 (1986), 311–333.
- F. Aryan, *The distribution of k-tuples of reduced residues*, Mathematika 61
(2015), 72–88, DOI 10.1112/S0025579314000151; arXiv:1302.2296.
- T. Bloom, V. Kuperberg, *Odd moments and adding fractions*, arXiv:2312.09021.
- O. Gorodetsky, *The variance of integers without small prime factors in
short intervals*, Math. Z. 308 (2024), no. 4, Paper No. 59, DOI
10.1007/s00209-024-03601-w; arXiv:2111.00853.
- R. G. Pinsky, *On the strange domain of attraction to generalized Dickman
distributions for sums of independent random variables*,
arXiv:1611.07207v3.
- F. B. Holt, *On the counts of p-rough numbers*, arXiv:2308.07570.
- H. L. Montgomery, K. Soundararajan, *Primes in short intervals*,
Comm. Math. Phys. 252 (2004), 589–617.
- V. Schemmel, *Über relative Primzahlen*, J. reine angew. Math. 70 (1869);
OEIS A059861.
- E. J. Scourfield, *Smooth divisors of polynomials*, in *Number Theory and
Polynomials*, LMS Lecture Note Series 352, CUP 2008, 286–311; read through
her restatement in Funct. Approx. Comment. Math. 55 (2016), eq. (1.1).
- Project artifacts: `research/06-variance-theorem.js` (verification),
`paper/variance-note-numerics.js` (scaling data), `research/PRIOR-ART.md`
(novelty audit with sources). For §7: `research/natal5-variance.js` (the
$J_5$ closed form and its brute-force check), `research/natal-cap-16-fast-variance.js`
(the product-sieve engine, the BigInt sum rule, levels through $x = 31$),
`research/natal-cap-33-overnight.js` (level $x = 37$),
`research/natal-cap-26-minus-half.md` (the exact $-1/2$),
`research/natal-cap-29-sigma-plateau.js` (the spectral mass). For §§8–11:
`research/history/staging/varE-asymptotic.js` (the normalisation, $\kappa$,
the conductor facts), `research/history/staging/varE-spectral.js` (the
model, the closed form, the $\theta = 1$ identity),
`research/history/staging/varE-limit-theorem.js` (the three bands, exact at
nine levels), `research/history/staging/varE-theta2-step.js` and
`research/history/staging/varE-theta2-proof.js` (the replacement error and
its group split), `research/varE-exact-ladder-01.js` (the exact ratios at
$x = 29, 31$), `research/history/staging/redteam-0828-varE.js` (the three
routes to the constant and the fit control),
`research/history/staging/verify-0830-record-defects.md` (the $2X_2$
correction of §10), `research/history/staging/lit-smooth-divisors.md` and
`research/history/staging/lit-scourfield-2008.md` (the divisor search of
§10).