Computational research note · Gilbreath's conjecture
Write down the prime numbers. Underneath each pair, write the gap between them — always as a positive number. Then do the same thing to that row. And again. And again.
Something strange happens. Every single row you make starts with the number 1. Not sometimes. Every time, as far as anyone has ever checked — which is now hundreds of millions of rows deep.
Nobody knows why, and nobody has been able to prove it always will. That is Gilbreath's conjecture, noticed by Norman Gilbreath in 1958 while doodling on a napkin. Press play below and watch it happen.
The primes are hard, so in July 2026 Chase, Hunter and Tao published a practice version: replace the primes with random numbers and do the same subtracting. Now you can ask a clean question — after i rows, how big is a typical number? Call that average ci. If it shrinks fast enough, rows of ones are inevitable. They could compute the first four of these numbers exactly and guessed the rest shrink like 1/i.
This note does four things.
The Tao paper reduced the whole conjecture to one job: show that a few very specific rigid patterns never occur among the prime gaps. Proving that needs a tool that can see individual arrangements of primes inside microscopically thin windows of the number line — a tool nobody has. We started building it, and three more things came out.
Everything below is the technical version. The interactive figures in sections 3, 5 and 7 are the fastest way to see points 2, 4 and 6.
For the continuous model of Gilbreath's conjecture introduced by Chase, Hunter and Tao (arXiv 2607.08712, July 2026) — iterated absolute differences of i.i.d. Exp(1) data — we compute the expected depth-i entry ci exactly for i = 4, 5, 6, extending CHT's exact values, each with a computer-verified exactness certificate in ℚ. We prove the model's sign arrangement attains the maximal 2i(i+1)/2 chambers. A Monte Carlo to depth 4095, anchored on the exact values, shows that ci is governed by the binary expansion of i through a two-parameter submask law that explains the Lucas-theorem oscillation CHT observed, fits with R² ≈ 0.98 where a pure power law fails entirely, and renders their tentative conjecture ci ≍ 1/i untenable in the accessible range: partial sums grow like n1/5, not log n. On the rigorous side we prove a two-channel sharpening of CHT's key inequality: cn ≥ 2 exp(−Σk<n ck) for n ≥ 2, whence Σi≤n ci ≥ log(2n − 2 + e²).
July 29 update. Against CHT's deterministic criterion (their Theorem 1.6) we prove thirteen covering-congruence lenses on the actual primes: depth-1 zero-blocks are progressions of consecutive primes and die by the primorial bound (length ≤ 1.25 log x); top-row {0,d}-blocks are impossible for every d of odd part > 1, with the bounds attained exactly in the data; a dichotomy confines depth-2 structures to residue-class prisons or second differences that are exact powers of two. The survivors provably lie in the kernel of every covering argument and below the resolution of every sieve. We then prove: (R1) a Cramér-type bound plus a residual pattern axiom — strictly narrower than CHT's hypotheses — implies Gilbreath for all but finitely many rows; (R2) no axiom system of fixed-order gap statistics implies eventual Gilbreath: a sequence exists with perfect fixed-order Cramér statistics whose leads fail infinitely often. Whether fixed-order statistics plus o(n) entries suffice is posed as the sharp open problem (R3).
CHT define the Gilbreath array by a(0,j) = aj (i.i.d. Exp(1)) and a(i+1,j) = |a(i,j) − a(i,j+1)|; by stationarity ci := 𝐄 a(i,j) depends only on i. They compute c0 = c1 = 1, c2 = 7/9, c3 = 227/288, prove Σi≤n ci ≥ log(n+e) and the key inequality cn ≥ exp(−Σi<n ci), state that they cannot prove boundedness of (ci), and tentatively conjecture ci ≍ 1/i from a 106-sample simulation to depth 20 (their Remark 1.5).
Fixing the sign of every intermediate difference splits ℝi+1≥0 into polyhedral cones on which the depth-i entry is linear; exponential integrals over simplicial cones have closed rational forms. Rays come from exact integer double description; triangulations use floating Delaunay for combinatorics only, cross-checked by a second independent projection and repaired, on disagreement, by an exact pulling triangulation via integer rank tests. Certificate: the exponential measures of all chambers must sum to exactly 1 in ℚ — this identity holds exactly in every reported run, and the pipeline reproduces CHT's c2 and c3.
Every one of the 2i(i+1)/2 sign patterns of the intermediate differences is realized on a full-dimensional cone of initial data.
Proof. Build the array bottom-up: set the bottom entry to 1; given row t+1 > 0 and the desired signs, put At1 = Tt and Atj+1 = Atj − st+1jAt+1j; taking Tt large makes row t positive and all constraints strict. ∎
The continuous model therefore has no combinatorial obstruction: cooling is irreducibly statistical. Counts 2, 8, 64, 1024, 32768, 2097152 confirmed computationally.
| i | ci | Status |
|---|---|---|
| 2 | 7 / 9 = 0.7778 | CHT |
| 3 | 227 / 288 = 0.78819 | CHT |
| 4 | 778959731701 / 1447295850000 = 0.53821735… | NewCertified |
| 5 | 14008668886481596262550223816901 / 25320304994525128311856832700000 = 0.55325830… | NewCertified |
| 6 | 0.44838867213… exact 153-digit numerator in c6_certified.txt | NewCertified |
The c6 computation covered 2,097,152 chambers; 685,625 needed the exact fallback. A first run without it silently lost 1.7×10−7 of measure and produced a value wrong by 1.25×10−7 — caught by the algebraic certificate, invisible to the Monte Carlo (both values sat 0.8σ from it). Denominators are smooth (largest prime factors 17, 47, 331); numerators contain enormous primes, so no closed form is apparent. None of these numbers appear in the OEIS. Note c6 < c4 < c5 < c3: the dyadic non-monotonicity deepens.
A tiered Monte Carlo (40M samples to depth 63, decreasing to 150k at 4095; all exact anchors reproduced within 1.2σ) shows that log(i·ci) regressed on the individual binary digits of i yields R² = 0.95 with a cascade of per-bit factors:
| Bit j (scale 2j) | 0 | 1 | 2 | 3 | 4 | ≥ 5 |
|---|---|---|---|---|---|---|
| Measured factor | 1.696 | 1.461 | 1.214 | 1.071 | 1.024 | ≈ 1 |
| 1 + q2^j, q = 0.685 | 1.685 | 1.469 | 1.220 | 1.048 | 1.002 | 1.000 |
The one-parameter match in the second row is the statement that a binary submask m of i contributes weight qm, via the identity Σm⊆i qm = Πj∈i(1 + q2^j). Fitting directly:
Partial sums fit Sn = 10.84·n1/5 − 11.43 to three decimals at every dyadic checkpoint n = 63 … 1023 (where log(n+e) ≤ 6.94, the CHT lower bound). Fitted on n ≤ 1023, the same formula then predicted S2047 = 38.37 and S4095 = 45.78; the deep run measured 38.33 and 45.60 — 0.1–0.4% accuracy two octaves out of sample.
Under ci ≍ 1/i one would have Sn ~ log n; the observed growth is a clean power, consistent with the typical exponent 1 − α ≈ 0.2. The conjecture of Remark 1.5 is untenable in the computationally accessible range — though CHT's theorems only force Sn ≳ log n, so a turnover beyond this range remains logically possible.
A large initial value at offset m propagates through |diff| dynamics as an exact Pascal-mod-2 pattern and reaches the apex iff C(i, m) is odd — iff m ⊆ i bitwise (Lucas). The submask sum in (⋆) is a sum over these open channels, with qm the measured attenuation of the channel at lateral offset m.
Two confirmations of the mechanism:
For every n ≥ 2, cn ≥ 𝐄 e−(SL−+B) + 𝐄 e−(SR−+A) ≥ 2 exp(−Σk<n ck), where A = a(n−1,1), B = a(n−1,2), SL− = Σk≤n−2 a(k,2), SR− = Σk≤n−2 a(k,n−k). Consequently Σi≤n ci ≥ log(2n − 2 + e²), an asymptotic additive log 2 over CHT's Theorem 1.4.
Proof. The apex is exactly |A − B| = (A−B)+ + (B−A)+. CHT's telescoping at depth n−1 gives A ≥ a1 − SL−; its mirror (the triangle is reversal-invariant) gives B ≥ an+1 − SR−. Now W = SL− + B involves only a2,…,an+1, so it is independent of a1, and 𝐄[(a1 − W)+ | W] = e−W; Jensen with 𝐄SL− + 𝐄B = Σk<n ck gives the first term, and the mirrored argument the second. ∎
Elementary — CHT's own techniques plus reversal symmetry — but a strict sharpening of their Prop. 2.1, and the first bound using two Lucas channels simultaneously. Every step was additionally verified numerically at n = 6 on 4×106 samples. The natural conjecture suggested by (⋆): all open channels contribute, cn ≳ (Σm⊆n qm)·exp(−Σk<n ck).
CHT's deterministic criterion (their Theorem 1.6, whose proof we read in full) reduces Gilbreath at row N to excluding two families of rigid structures from the gap array: 0-valued blocks of length ≈ log10N at any depth, and {0,d}-valued blocks vastly longer than they are deep. The following are unconditional theorems of this program (proofs in MICROSCOPE.md; every bound verified against the primes below 10⁹):
Why the survivors survive — two structural principles. Covering arguments run on the one residue forbidden to primes, 0 mod q; modulo powers of two, primes forbid only evenness, so no covering argument can touch a pattern whose modulus support is 2-adic. And sieve bounds exceed trivial counts in polylog-length windows, so density methods are structurally blind at exactly the scale where the dangerous patterns live. Every surviving enemy sits in the intersection of those two blind spots — which is a precise statement of what a new tool must do.
Theorem R1 (conditional reduction; proved). Cramér's gap bound plus a residual pattern axiom P — strictly narrower than CHT's own hypotheses, because the lenses of section 6 discharge depth-1 zero-blocks and all odd-part-d top-row blocks unconditionally — implies Gilbreath's conjecture for all but finitely many rows. Statement and proof in REDUCTION.md.
Theorem R2 (insufficiency of fixed-order statistics; proved). There is a gap sequence whose every fixed-order pattern count matches the i.i.d. Cramér model to within Ok(log log x) — so every Hardy–Littlewood-analogue asymptotic holds, at any error quality — and whose Gilbreath leads exceed 1 infinitely often. The construction plants a single value V = m² at doubly-exponentially sparse positions m: one entry changed per plant leaves all fixed-order statistics intact, but the plant rides the always-open diagonal Pascal channel (C(t,t) = 1), and the mirrored CHT telescope bounds its erosion by the background sum, so it reaches the left edge with room to spare. Consequently no axiom system consisting of fixed-order gap statistics can imply eventual Gilbreath — the first rigorous form of CHT's remark that their hypotheses "look difficult to establish rigorously, even if one assumes strong conjectures on the primes such as the Hardy–Littlewood prime tuples conjecture."
Open Problem R3 (the frontier). The plant is forced to be huge — a gap comparable to its position, grossly non-Cramér — and this is provable: under any o(n) entry bound the plant erodes away before arriving, while extended "highway" structures leave fixed-order twin-density fingerprints of exactly the kind the statistics forbid. So: does fixed-order statistics + an = o(n) imply eventual Gilbreath? The o(n) bound is known unconditionally for primes (Baker–Harman–Pintz), so an affirmative answer would make Gilbreath a consequence of the standard conjectural statistics of prime gaps — conditional, since those contain the twin prime conjecture, but the first route of any kind from standard conjectures to Gilbreath. A negative answer would strengthen R2. Either resolution is a paper.
The same session verified Gilbreath's conjecture rigorously for the first 455,052,510 rows (primes < 1010, via the propagation-lemma certificate), audited the two dangerous structures of CHT's deterministic Theorem 1.6 on the real primes — worst zero-block 31 vs threshold ~3×1012 and worst {0,d}-block ratio 11 vs ~1026, both growing at coin-flip rates — and measured the prime array's defect dynamics (×0.936/row decay under an exactly critical mean-field, with period-4 Sierpiński modulation). Details in the companion session write-up.