Wiki
Wiki

Research notes on every problem, and a library of the papers behind them. Built from the open erdos repository.

Updated

Erdős 699: a smaller coefficient bound at index 20

P699 research · Effective bounds at 15 and 20 · The interpolation nonvanishing lemma


Complete exclusions. Indices 15,2015,20 are fully excluded by the cutoff-15 proof; the bounds and constructions below are retained as earlier stages and inputs.

The existing index-20 interpolation construction gives the much smaller algebraic tail

n<106396\boxed{n<10^{6396}}

once its cofactor inequalities hold. This replaces the generic coefficient estimate that gave 10400000010^{4000000}. It is not an assertion that every source threshold is below 10639610^{6396}: the original near-limit asymmetric estimates still have larger numerical thresholds. The polynomial bound below can also be combined with another certified cofactor gain for the same row weights.

An explicit complete kernel basis

Keep the index-20 vectors

h=(34,32,30,27,25,22,19,16,13,9,5,1,0,0,0,0,0,0,0,0),m=(67,65,65,65,62,60,58,57,57,54,50,46,42,39,37,34,32,30,28,26),\begin{aligned} h={}&(34,32,30,27,25,22,19,16,13,9,5,1,0,0,0,0,0,0,0,0),\\ m={}&(67,65,65,65,62,60,58,57,57,54,50,46,42,39,37,34,32,30,28,26), \end{aligned}

and residual multiplicities Rxy=max⁡{0,mx+y−hx−hy}R_{xy}=\max\{0,m_{x+y}-h_x-h_y\} at x,y≥0x,y\ge0, x+y<20x+y<20. The previously certified symmetric degree-96 jet map is surjective, of rank 2399. There are 2450 symmetric monomials through degree 97, so its degree-97 kernel has dimension 51.

The compressed coefficient artifact contains 51 integer polynomials in the basis XaYb+XbYaX^aY^b+X^bY^a for a<ba<b and XaYaX^aY^a for a=ba=b. The direct verifier checks:

  • All 2399 required jets of every polynomial are exactly zero over the integers.
  • The coefficient minor on the last 51 symmetric monomials has determinant 13345(mod65521)13345\pmod{65521}, so the polynomials are independent.
  • Their largest coefficient has 121 decimal digits, and their largest coefficient norm has 123 decimal digits.

Thus these polynomials form a rational basis of the complete symmetric degree-97 kernel. The existing multiplication-by-X+YX+Y lemma guarantees that they have no common zero when X+Y>19X+Y>19.

The basis was obtained by an exact Dixon solve for the 51 free columns, followed by primitive normalization, saturation of the integer lattice, and LLL reduction. The initial primitive basis had a 1696-digit maximum coefficient norm. Saturation uses the row Hermite form of the transpose: if BB has the basis vectors as rows and HH is a row basis of the coordinate lattice of BTB^{\mathsf T}, then BTH−1B^{\mathsf T}H^{-1} is integral and generates the saturated column lattice. Exact identities verify this step and the unimodular LLL transformation. Verification of the retained result does not depend on any property of the search or reduction algorithm: integer jets and the nonzero coefficient minor suffice.

Evaluating the polynomials on a large row

For each retained polynomial GG, write its degree-97 part as

G97(X,Y)=∑a=097caXaY97−a.G_{97}(X,Y)=\sum_{a=0}^{97}c_aX^aY^{97-a}.

The exact largest Bernstein coefficient, across the entire basis, is

B:=max⁡G,a∣ca∣(97a)=7848927543503780125160732509336560.B:=\max_{G,a}\frac{|c_a|}{\binom{97}{a}} =7848927543503780125160732509336560.

For X,Y≥0X,Y\ge0 and X+Y=nX+Y=n, the binomial theorem gives ∣G97(X,Y)∣≤Bn97|G_{97}(X,Y)|\le Bn^{97}. All lower-degree terms together have magnitude less than n97n^{97} when n≥10123n\ge10^{123}, because the full coefficient norm of every GG is less than 1012310^{123}. Consequently

∣G(j,c)∣<(B+1)n97(n≥10123).|G(j,c)|<(B+1)n^{97}\qquad(n\ge10^{123}).

This bound uses only the nonnegative physical coordinates, not a bound for arbitrary complex evaluations.

The weighted divisibility constant

Set

H(X,Y)=∏a=019((X−a)(Y−a))ha.H(X,Y)=\prod_{a=0}^{19}((X-a)(Y-a))^{h_a}.

Here 2∑ha=4662\sum h_a=466. Pairing the factors gives the sharper direct bound

0<H(j,c)≤n4662466.0<H(j,c)\le\frac{n^{466}}{2^{466}}.

The exact off-maximum small-prime constant is

Coff=∏p<20pmax⁡r∑t≠rmtvp(t−r)=28853441517771241167136717671967.\begin{aligned} C_{\rm off} &=\prod_{p<20}p^{\max_r\sum_{t\ne r}m_t v_p(t-r)}\\ &=2^{885}3^{441}5^{177}7^{124}11^{67}13^{67}17^{67}19^{67}. \end{aligned}

For the normalized cofactor exponent vpv_p, retain

Pp≤n1−vp(1−19/n)−vpP_p\le n^{1-v_p}(1-19/n)^{-v_p}

at singleton positions instead of replacing the last factor by 2vp2^{v_p}. At a collision position the selected prime powers have product at most nn. Write L=∑tmt=7692L=\sum t m_t=7692. The numerator factors and all singleton normalization factors together are at least

1−L+8m0⋅19n≥9991000(n≥17876000).1-\frac{L+8m_0\cdot19}{n} \ge\frac{999}{1000} \qquad(n\ge17876000).

This follows by expanding products of factors in [0,1][0,1] and bounding the total singleton exponent by 8m08m_0.

For every bad pair beyond the cofactor thresholds, select a retained GG with G(j,c)≠0G(j,c)\ne0, which is possible by nonvanishing. Apply weighted Kummer divisibility to HGHG. The already certified uniform gain is 713/5000713/5000, so

n713/5000<1000 Coff(B+1)999 2466<10912.n^{713/5000} <\frac{1000\,C_{\rm off}(B+1)}{999\,2^{466}} <10^{912}.

The second inequality is checked by an integer comparison. Finally, 6396⋅713/5000>9126396\cdot713/5000>912. Since 10639610^{6396} also exceeds the elementary and evaluation thresholds, the claimed algebraic tail follows.

Retained verification

The small JSON report contains the coefficient norm, Bernstein bound, coefficient-artifact checksum, independence minor, exact off-maximum factors, and the resulting exponent comparison. Run the checker with the pinned dependencies installed in the repository's .venv/:

sh
.venv/bin/python wiki/research/erdos_699/computations/erdos_699_index_20_small_height.py

Ordinary verification checks the retained integer polynomials directly. The optional --regenerate flag repeats the exact solve, saturation, and LLL construction. No binomial rows are enumerated.