Chowla's conjecture · analytic number theory
Take any number and break it into prime factors. Count the pieces. 12 = 2·2·3 is three pieces (odd). 15 = 3·5 is two (even). Write down H for even, T for odd, and do this for 1, 2, 3, 4, … forever. You get an endless string of coin flips: T T T H T H T T H H…
Every test anyone has ever run says this string behaves exactly like a fair coin with no memory: knowing today's flip tells you nothing about tomorrow's. That innocent-sounding claim is Chowla's conjecture, and it is one of the deepest unsolved problems in mathematics. Proving even the weakest version, that a flip and its immediate neighbour agree only half the time in the long run, has resisted everyone, because every known technique loses exactly this kind of "parity" information. The obstruction has a name: the parity barrier. It is the same wall that stands between us and the twin prime conjecture.
The figure below is the coin's memory, measured further than it has been measured before. Each curve tracks how much the string agrees with a shifted copy of itself, in units where a fair coin would wander between roughly −2 and +2. Nearly five orders of magnitude in x, no drift.
The unaveraged conjecture is open for every shift; the modern results (Tao 2016, Helfgott–Radziwiłł 2021, Pilatte 2023/25) control only a logarithmically averaged version. This note contributes, from one session:
Everything below is the technical version. The fastest route through it: the identity figure in §2, then §7.
Let λ be the Liouville function and Sh(x) = Σn≤x λ(n)λ(n+h). Chowla's conjecture asserts Sh(x) = o(x) for every h ≥ 1; it is open for every h. We prove elementary but organizing exact identities: a 2-adic descent Sh(x) = Uh(x) + Sh/2(⌊x/2⌋) (h even, Uh the odd-n restriction), which exhibits the pair-correlation field as an exact fixed point of a renormalization map with odd-restricted correlations as innovations, and a Walsh–Hadamard duality between the k-window sign-pattern census and multipoint correlations (PROVED). We compute a certified integer census of Sh (h ≤ 32), Uh, the triples {0,1,2}, {0,2,4}, {0,1,3}, the quadruple {0,1,2,3}, the full 256-cell 8-window census, and harmonic sums ℓh(x) = Σ λ(n)λ(n+h)/n, on a grid of spacing 2·107 to x = 1012, every row passing the descent and duality certificates exactly, with external anchors reproduced (L(10k) per OEIS A090410 for k = 9…12; the Pólya island 906,150,257…906,488,079 with max L = 829 at 906,316,571; BFM's 252 zeros and 133 sign changes of L below 109) (CERTIFIED). Every ±1 pattern of length ≤ 27 occurs; completion points Nk track the coupon-collector model across 27 doublings (mean ratio 1.007 for k ≥ 10, scatter matching the model's own Gumbel width) and are new to the OEIS (CERTIFIED by exhibition). Six predictions registered before the large computation ran resolve to 14 scored clauses: 10 hits and 4 misses, all reported, including a parameter-free cross-scale increment correlation 1/√2 forced by descent plus innovation-independence, measured +0.7085 ± 0.0046 across 1012 (NUMERICAL). At x = 1012, |S1|/x ≈ 1.6·10−7 while the best proven almost-all-scales bound (Pilatte 2025, Remark 2.8) is numerically ≥ 0.9: the parity barrier is six to seven orders of magnitude wide at real scales. Finally we audit arXiv:2211.09736, whose claimed unconditional equidistribution rests on transferring logarithmic-average control to natural averages, an implication refuted by f(n) = niα. A second pre-registered instrument set went 10-for-10: a certified 4-point census to 1011 (first data at an order where even averaged Chowla is open; all twelve quadruples square-root-sized, quad echo +0.721 ± 0.013), the full covariance web of the descent structure matching the fair-coin model in every entry, and coverage extended to every pattern of length ≤ 30 (N30 = 22,249,147,014). A second, independent session (§10) then reimplemented the coverage engine from scratch — different language, different machine, no shared code — and reproduced all thirty Nk values and last-completing pattern codes exactly, together with L(10j) for j = 1..8; coverage was then extended to every pattern of length ≤ 33 (N31, N32, N33 = 43,901,697,682, 99,494,377,311, 196,202,853,829, each CERTIFIED by exhibition and independently re-verified by trial division). The same session introduced the first-occurrence spectrum, scoring every one of the 224 sign patterns of length 24 against Conway's exact fair-coin waiting time: mean R = 1.000123 against 32 decorrelated control streams at 0.999953 ± 0.000319 (+0.53 control-σ), indistinguishable from a fair coin at a resolution of about 3·10−4 (NUMERICAL). The conjecture survives, its open core isolated in the odd-restricted correlations.
λ(n) = (−1)Ω(n), where Ω counts prime factors with multiplicity; λ is completely multiplicative with λ(p) = −1 at every prime.
For every fixed h ≥ 1: Sh(x) = Σn≤x λ(n)λ(n+h) = o(x).
Status of the art (each statement checked against the cited paper's text, not a summary of it):
Let d ≥ 1 and h = d·h′. Then Σn≤x, d|n λ(n)λ(n+h) = Sh′(⌊x/d⌋). In particular, for even h, with Uh(x) := Σn≤x, n odd λ(n)λ(n+h):
Sh(x) = Uh(x) + Sh/2(⌊x/2⌋).
Proof. For n = dm, complete multiplicativity gives λ(dm)λ(dm+dh′) = λ(d)²·λ(m)λ(m+h′) = λ(m)λ(m+h′), and n ↦ m bijects {n ≤ x : d|n} with {m ≤ ⌊x/d⌋}. ∎
The identity is checkable in the strongest possible sense: exactly, integer for integer, at any scale. The figure computes λ live and does so; the shuffle button destroys multiplicativity while keeping the exact same multiset of signs, and the identity collapses. That contrast is the content of the lemma: the self-similarity is carried by multiplicative structure, not by statistics.
Write h = 2ah₀ with h₀ odd. Then Sh(x) = Σj<a Uh/2^j(⌊x/2j⌋) + Sh₀(⌊x/2a⌋), and if |Ug(y)| ≤ δy for all even g and all y then |Sh(x)| ≤ 2δx + x/2a. Hence the two-point conjecture is equivalent to: Sh = o(x) for odd h, and Uh = o(x) for even h. The odd-restricted correlations are the irreducible content.
Proof. Iterate Lemma 1; bound the tail trivially; sum the geometric series. ∎
For ε ∈ {±1}k let Nε(x) count n ≤ x whose sign window matches ε, and let RS(x) = Σn≤x Πi∈S λ(n+i). Then Nε(x) = 2−k ΣS⊆[k] (Πi∈S εi) RS(x), and RS(x) = Σε (Πi∈S εi) Nε(x).
Proof. Expand Πi(1+εiλ(n+i))/2, the indicator of the window; invert by Walsh–Hadamard orthogonality. ∎
Let f be completely multiplicative with f(3) = −1 and f(p) = χ(p) for p ≠ 3, χ the quadratic character mod 3. Then Σn≤x f(n)f(n+1) = −x/3 + O(log x) and Σn≤x f(n)f(n+3) = (5/9)x + O(log x).
Proof. Exactly one of n, n+1 is divisible by 3. If neither would be (n ≡ 1 mod 3), f(n)f(n+1) = χ(1)χ(2) = −1: density 1/3, contribution −x/3. If 3 | n, write n = 3au: the factor (−1)aχ(u) summed over u ≤ x/3a is O(1) per level by the bounded character sum, O(log x) in all; symmetrically for 3 | n+1. For h = 3: the coprime class contributes χ(n)² = 1 at density 2/3, and the class 3|n reduces by Lemma 1's mechanism (with d = 3) to the h = 1 sum at scale x/3: (2/3) − (1/3)(1/3) = 5/9. ∎
Measured at 107: −0.3333334 and +0.5555556. A completely multiplicative ±1 function can fail Chowla maximally; the conjecture is precisely the assertion that λ is far from every such character-like ("pretentious") function. This is the known structural obstruction (Frantzikinakis–Host; MRT), rediscovered here from the computation before the literature was consulted; recorded as validation, not contribution. Note the failure constants themselves obey the descent recursion s₃ = 2/3 + s₁/3.
Engine: an exact segmented sieve for λ: for each n the parity of small-prime-power divisors plus an exact integer product of the small-prime part, which detects the single possible prime factor > √x. No floating point decides any certified value: correlation sums are float64 BLAS dot products of ±1 vectors in which every intermediate is an integer < 253, so every IEEE-754 addition is exact. Validated against brute-force trial division (in full on [1, 2·105]; windows near 109, 3·1010, 1012, 240) and against an algorithmically independent smallest-prime-factor sieve (all 310 columns equal at 106); two independent kernels agree on all rows ≤ 109.
Certificates, checked at every grid row:
External anchors, the published numbers the pipeline must and does reproduce:
| anchor | published | this census |
|---|---|---|
| L(109) … L(1012) (OEIS A090410) | −25216, −116026, −342224, −522626 | equal, exactly |
| first n with L(n) > 0 (Tanaka 1980) | 906,150,257 | 906,150,257 |
| max L below 109 (Tanaka) | 829 at 906,316,571 | 829 at 906,316,571 |
| zeros / sign changes of L below 109 (BFM 2008) | 252 / 133 | 252 / 133 |
| Turán's Σλ(k)/k > 0 up to 7.2·1013 (BFM) | positive | positive at every grid row |
BFM's L-computation to 2·1014 took ~2.5 CPU-years in 2008; this census covers 0.5% of that range but records 300+ statistics per grid row that were never computed at any scale: the pair-correlation and pattern-census columns are new measurements.
One more thing the certificates did: across the two long runs they caught and localized ten transient memory-corruption events (single zeroed pages, statistically invisible; one detectable only by the descent identity C1). Each was repaired from a deterministic recompute, raw pre-repair files are kept in the repository, and the two independently corrupted, independently repaired runs agree bit for bit on all 5000 overlapping rows, while L(109)…L(1012) match the published values exactly. Details: VERDICTS.md.
Certified values at decades (h = 1…8; each cell an exact integer, shown with Sh/√x underneath):
| x | S₁ | S₂ | S₃ | S₄ | S₅ | S₆ | S₇ | S₈ |
|---|---|---|---|---|---|---|---|---|
| 109 | −46,682 | 43,790 | 38,866 | 34,016 | 16,032 | 24,660 | 4,132 | −8,508 |
| 1010 | −71,450 | −118,586 | 18,508 | 153,922 | 83,042 | 1,176 | 132,516 | 137,364 |
| 1011 | −219,878 | 153,142 | 494,210 | 65,362 | −62,982 | 479,064 | 142,722 | −203,530 |
| 1012 | −163,274 | 57,386 | −604,438 | 600,878 | −407,794 | 1,672,602 | −996,980 | −299,864 |
At 1012, maxh≤32 |Sh|/√x = 2.149 (h = 13) and the log-log fit of RMSh Sh against x over [108, 1012] gives exponent α = 0.467, against 0.436 on the range ending at 1011: the estimate moves toward ½ as range is added, exactly as the single-path-noise reading predicts, though it stays marginally below the registered band (§5). The fluctuation level is square-root-sized across nearly five decades of x. NUMERICAL interpretation of CERTIFIED integers; the fit describes this range only and no asymptotic guarantee follows.
Every ±1 pattern of length k ≤ 33 occurs (k ≤ 30 from this session; k = 31, 32, 33 from the independent session-2 audit, §10). Completion points Nk (least x by which all 2k patterns have appeared; exact first occurrences on file):
At 1012 the 256 window counts deviate from x/256 by at most 2.98 standard deviations (4.21 at the 1011 cap; Gaussian-max expectation ≈ 3.5): no pattern is favoured beyond fair-coin fluctuation.
ℓh(x) = Σn≤x λ(n)λ(n+h)/n is what the modern theorems actually bound (trivial bound: log x ≈ 27.6 at 1012). Measured values are O(1) and frozen by small n: between 109 and 1012 no ℓh (h ≤ 32) moved by more than 9.63·10−5 (expected maximum over 32 shifts under the model: 9.2·10−5), against a fair-coin tail σ of 3.2·10−5. ℓ1(1012) = −0.842488. Convergence of the series is itself open (strictly stronger than logarithmic Chowla, strictly weaker than power-saving Chowla), and we propose it as a benchmark problem (§7). These columns are float64 with a rigorous < 10−8 rounding bound (every other column is exact).
The log-weighted mean of L(x)/√x over the fine grid is −0.671, against the conjectured density-mean 1/ζ(1/2) ≈ −0.6848 from the pole of ζ(2s)/ζ(s): agreement to 2%. Illustration only (NUMERICAL).
Six predictions were written to disk (PREDICTIONS.md, timestamped) when the census had only reached 109, each naming its generating model; the large run then ran through the targets. Verdicts:
| prediction (model) | registered | measured | verdict |
|---|---|---|---|
| RMSh Sh/√x at 1010, 1011, 1012 (fair coin) | 0.90 ± 0.35 / ± 0.40 | 1.155 / 0.794 / 0.898 | hit / hit / hit |
| exponent α on [108, xcap] | [0.47, 0.53] | 0.436 | miss |
| cross-scale corr, pooled (descent + independence) | [0.68, 0.73] | +0.688 ± 0.017 | hit |
| per-h corridor ±0.05 of 0.7071 | 6 of 6 inside | 4 of 6 inside | miss |
| census max|z| (Gumbel over 256 cells) | [2.7, 4.3] | 4.21 | hit |
| L(1010)…L(1012) (A090410); L<0; hL>0 | exact / all rows | exact / all rows | hit |
| hL(1012) band (from BFM local minima) | [5·10−8, 6·10−7] | +8.23·10−7 | miss (band calibrated low) |
| N25 (coupon collector, 2σ) | (6.0 ± 0.9)·108 | 7.228·108 (ratio 1.203) | miss (~2.6% tail) |
| N26, N27 | (1.25 ± 0.17)·109 / (2.6 ± 0.34)·109 | 1.313·109 / 2.795·109 | hit / hit |
| harmonic freeze 109 → 1012 (±2·10−4, model σ 3.2·10−5) | max drift in band | 9.63·10−5 (expected max-of-32: 9.2·10−5) | hit |
Scoring happened twice: at the pre-registered 1011 cap, and again when the full 1012 run resolved the remaining horizon clauses. Final ledger: 14 scored clauses, 10 hits, 4 misses; misses reported exactly as registered (full detail: VERDICTS.md in the repository).
The point of the protocol is not the hit rate: it is that the models named in advance (fair-coin scaling, coupon collector, descent + independence, published anchors) are exactly the models the data keeps electing.
Lemma 1 forces block increments to satisfy ΔS2h[2y, 2y+2Δ) = ΔU2h[2y, 2y+2Δ) + ΔSh[y, y+Δ): every fluctuation of Sh at scale y is replayed, exactly, inside S2h at scale 2y. If the innovation ΔU is uncorrelated with ΔSh and carries half the variance (the fair-coin hypothesis on the odd subsequence only), the correlation of the coupled increments is exactly 1/√2 = 0.7071, with no free parameter.
The identity contributes the coupling; the data contributes the 1/√2. A conspiracy in the odd innovations would move this number; with ten times the data it instead converged to within one standard error of the parameter-free value. To our knowledge this 2-adic self-similarity of the Liouville correlation field has not been measured before.
At x = 1012: |S1(x)|/x = 1.6·10−7. For the proven shape to reach that value at this x, its exponent would need c ≈ 9.4; the method that produced the theorem caps c at a small constant, and its author states the power-of-log saving is the limit of current techniques. Nothing in this census can close that gap, but it fixes its width, at real scales, for the first time we know of.
Three open problems, in increasing strength, each strictly between what is proved and what is believed:
A 2022/23 preprint (Carella, math.GM) claims unconditionally that the four double sign patterns of λ have natural density 1/4, equivalent to unaveraged two-point Chowla, which Pilatte (Dec 2025) lists as open. Reading the preprint in full: the claim reduces (its eq. 3.9) to Σn≤x λ(n)λ(n+t) = O(x/(log log x)1/2−ε) (its Thm 6.4), obtained from Helfgott–Radziwiłł's logarithmically averaged bound via its Lemmas 6.1–6.2, which assert: if Σn≤x f(n)/n = o(log x) then Σn≤x f(n) = o(x).
That implication is false for bounded sequences. With f(n) = niα (α ≠ 0): Σ f(n)/n = O(1), yet |Σn≤x f(n)|/x → 1/|1+iα| ≠ 0. The proof's contradiction step needs its assumed lower bound B(z) ≥ z(log log z)−1/2+ε pointwise across the whole integration range, while the negation of the conclusion supplies it only along a sequence of scales, with uncontrolled sign; ∫B(z)/z² dz can cancel between scales. The preprint's Theorems 1.1, 1.2, 6.4 and 6.6 are therefore unproven. (Our census shows the claim itself is empirically on track: the flaw is in the proof, not the belief. The audit keeps the record straight: the conjecture remains open.)
All code in conjectures/chowla/ of the Matematica repository; scripts run from that directory; no randomness in any certified path. Machine: Apple M3 Pro (6P+6E cores, 36 GB), Python 3.12, NumPy 2.3.5. Runtimes: fine run 231 s; harmonic baseline 267 s; first main run 2752 s wall to the 1011 cap; full 1012 run 23,362 s wall on 10 workers; coverage extension 978 s.
| artifact | contents |
|---|---|
| liouville.py + test_liouville.py | exact sieve; brute-force validation |
| census_run.py / certify.py | the census; exact certificate checks C1–C3 |
| coverage_ext.py / signchange.py | pattern coverage k ≤ 27; L sign-change counts |
| PREDICTIONS.md / VERDICTS.md | the out-of-sample protocol, timestamped |
| data/*_grid.csv, *_firstocc.npz, *_meta.json | certified grids; first occurrences; runtimes |
A second, unrelated session returned to this conjecture three days later, in a sandbox with four cores, 15 GB of RAM, and no outbound network. It also had no NumPy, SciPy or networkx — every script in §§1–9 imports NumPy, so none of the results above could be rerun on that machine at all. That is a reproducibility defect worth stating plainly rather than working around silently: a certified computation whose reproduction depends on a package set a clean machine may not have is less reproducible than it looks. The response was to rebuild the coverage half of the pipeline from scratch in C, sharing no code with the original — which turns the obstacle into the more useful thing, an independent audit.
lambda_coverage.c is an exact segmented sieve for λ, written
independently of the NumPy pipeline: Ω(n) mod 2 is tracked by flipping a parity
bit on the multiples of every prime power pe ≤ x, with a per-segment
residue that carries the one possible prime factor above √x, divided out
exactly by multiplication with p−1 mod 264 rather than by
hardware division. No floating point anywhere in the critical path.
Pointed at the previous session's certified output, it reproduces
all thirty values of Nk and all thirty last-completing
pattern codes, k = 1…30, across three separate files
(fineA_coverage.csv, coverage_ext.csv,
coverage_2830.csv), plus L(10j) for j = 1…8
(0, −2, −14, −94, −288, −530, −842, −3884) — exactly, integer for
integer. A third, independent implementation
(verify_coverage.py, pure Python standard library, trial
division) confirms Nk for k = 1…14. Reproduction log:
data/repro_k25_30.csv.
One run, 4 cores, 7387 s wall, peak RSS ≈ 2.5 GB, tracking k = 31, 32, 33 simultaneously against a single λ stream:
| k | Nk | last-completing code | Gumbel z |
|---|---|---|---|
| 31 | 43,901,697,682 | 1,784,492,180 | −1.26 |
| 32 | 99,494,377,311 | 2,930,773,200 | +0.32 |
| 33 | 196,202,853,829 | 3,712,643,644 | −0.48 |
Each of the three completing windows was independently re-verified by
pure-Python trial division, along with 10, 12 and 12 further endgame windows;
data/endgame_3133.txt ships the last 256 completions per k, every
one an independently checkable exhibition. As in §4.2, "certified" means
exhibited and re-verified, not proved minimal — see §10.4.
The unseen-pattern count tracked the model 2ke−x/2^k to
three significant figures the whole way down: at x = 1011 the k = 33
tracker had 75,925 patterns left against a predicted 75,600. Overlapping
windows are not literally uniform coupon draws — periodic patterns wait
longer, by Conway's formula (§10.3) — but the excess probability that a
low-period pattern is what's unseen at the collector scale is below 0.3% at
k = 32 and 33 (coverage_model.py), so the plain coupon-collector
model is the right one to read Nk against to well under a percent.
Nk is one number: the last pattern to appear. The
--firstocc mode instead records the first occurrence of
every one of the 2k patterns — 16,777,216 data points at
k = 24 instead of one — and each has an exact, parameter-free prediction under
the fair-coin null. Conway's leading number gives the expected waiting time to
the first occurrence of a word w as A(w) = Σj δj(w)·2j,
where δj = 1 iff w's length-j prefix equals its length-j suffix.
With s(w) the first-occurrence start index, the statistic
has model mean exactly 1 for every individual pattern, not just on average. A(w) itself ranges over a factor of two — 2k for a pattern with no self-overlap, up to 2k+1−2 for a constant pattern — and that clumping is a property of the word, not of λ; dividing it out is what leaves room to see anything else. Measured directly at k = 24 (all 16,777,216 patterns): mean R = 1.000123, sd R = 1.000251 — matching the Exp(1) shape an i.i.d. fair coin predicts, to four digits.
The 2k values of s(w) come from a single stream and are not independent, so σ/√2k is the wrong error bar, by a large factor. Every ± below is instead the dispersion of 32 independent i.i.d. fair-coin control streams, pushed through the identical window/bitmap/first-occurrence code path.
data/firstocc_k24_controls.csv); λ is the filled marker. Axis
zoomed to [0.9990, 1.0010]. λ = 1.000123; controls 0.999953 ± 0.000319 —
λ sits at +0.53 control-σ, rank 23 of 33, inside the pack.That popcount slope is not free-floating structure. Across the 32 controls
it correlates at r = +0.80 with the stream's own realised one-point bias
L(107)/107 — a density excess has to tilt a
popcount-resolved waiting time, and to first order the tilt is linear in that
bias. Regressing λ's slope through the control-calibrated relation leaves a
residual of −0.56 residual-σ. (Fitted coefficient 0.92 ± 0.12 against
the crude first-order value 2 — expected, since the relevant density is a
weighted average over the whole range of first-occurrence times, not its value
at one reference scale; this regression is a calibration, not a second
prediction.) The overlap-class breakdown (how many nontrivial self-overlaps
w has) is flat at 1.000 for both λ and the controls, which is the check that
the Conway normalisation does what it claims; no figure for this — the table
in data/firstocc_k24_analysis.txt is the whole content, and
there is also none for the endgame logs of §10.2, for the same reason: no
sentence to attach to either.
The first control ensemble was defective, and the control is what exposed
it. Streams were seeded sm64(n ⊕ seed); for small seeds,
XOR with n only permutes n inside a short block, so every seed produced a
stream with the same bit multiset — all 32 "controls" reported
L(108) = +16362, to within 2. Ensemble dispersion collapsed, and
against that fake error bar λ's popcount slope read as a 4.4σ anomaly —
a discovery that existed entirely inside the miscalibration. Reseeded as
sm64(sm64(seed·φ) + n), the streams decorrelate
(L(108) = −5552, +5176, −2372, +7388, −3638, +1142, …) and the
identical λ measurement is −0.44σ. Left in the record on purpose: an ensemble
of controls that is secretly one control is exactly the way this instrument
would have lied.
Machine: 4 cores, 15 GB RAM, no other load. The k = 31,32,33 run reached N31 at 1718 s, N32 at 3774 s, N33 at 7387 s (sieve 1701 s, bitmap probes 5677 s); steady state ≈ 390 s per 1010 values of n. Reproduce (no NumPy required):
| artifact | contents |
|---|---|
| lambda_coverage.c | exact segmented λ sieve + coverage/first-occurrence census; --prng control stream, --firstocc full spectrum, --lcheck L(10j) |
| verify_coverage.py | independent trial-division verifier: Nk from scratch, single-window, endgame batches |
| firstocc_stats.c / firstocc_analysis.py | Conway-normalised R(w); λ vs control ensemble: mean, popcount slope, overlap classes |
| coverage_model.py | coupon-collector comparison for Nk, self-overlap correction bound, Gumbel z |