Erdős–Woods conjecture

Research log for issue #46. The conjecture is open.

Write rad(n)\operatorname{rad}(n) for the product of the distinct primes dividing a positive integer nn. The conjecture asserts that an absolute positive integer kk exists such that

rad(x+i)=rad(y+i)(0≤i<k)⇒x=y.\operatorname{rad}(x+i)=\operatorname{rad}(y+i)\quad(0\le i\lt k) \quad\Longrightarrow\quad x=y.

A matching block means these equalities for positive x<yx\lt y. Erdős Problem 850 asks for the stronger statement that k=3k=3 suffices.

Literature

  • Erdős, 1963, Problem 60 gives the two-term family with starts 2h−22^h-2 and 2h(2h−2)2^h(2^h-2), for h≥2h\ge2. Makowski's pair (75,1215)(75,1215) is another two-term example, as recorded by Hercher.
  • Balasubramanian, Shorey, and Waldschmidt, 1989 bound the possible length in terms of the smaller start and give lower bounds on the separation. Neither bound is absolute.
  • Shorey and Tijdeman, §9 recount the earlier conditional finiteness under ordinary abc and prove the three-term statement under Baker's explicit version. Waldschmidt's 2026 lecture still presents the general problem as a consequence of abc.
  • Lebowitz-Lockard, 2025, Theorems 4 and 6 studies unconditional bounds on the length and on the count of termwise matching blocks. The count bound is the comparison for the averaging argument in this project.
  • Hercher, 2025 proves a finite two-term census. A November 2026 computation reports a larger census and no three-term witness through its cutoff. The present research does not rely on either enumeration.
  • Venkat, 2026 studies equality of radicals of whole block products. That differs from equality of each aligned term. A construction with whole-product equality does not provide the termwise equalities above.

The primary sources, recent preprints, problem discussion and related claims were checked in November 2026. No unconditional proof of a universal length and no family with unbounded matching length were found.

Every fixed algebraic curve contains finitely many matches

Theorem. For any nonzero polynomial P∈ℚ[X,Y]P\in\mathbb{Q}[X,Y] and any k≥3k\ge3, there are only finitely many positive integer pairs x<yx\lt y with P(x,y)=0P(x,y)=0 and matching blocks of length kk.

Proof

Suppose an irreducible component of the plane curve contains infinitely many matching pairs. We can work over a number field over which that component is geometrically irreducible. Let CC be its smooth projective normalization, and write f,gf,g for the two coordinate functions. Discard the finitely many singular points of the original curve so that the remaining pairs lift to points on CC.

If one coordinate is constant, the radicals of the first two terms of the other coordinate are supported on a fixed finite set of primes. The pair of consecutive terms is then a solution of an SS-unit equation, of which there are finitely many. Assume both coordinates are nonconstant. The function f−gf-g is nonzero since all pairs have x<yx\lt y.

Let D0D_0 be the union of the geometric poles of ff and gg, and let t=|D0|t=|D_0|. For each ii with 0≤i<k0\le i\lt k, mark every point outside D0D_0 which is a zero of exactly one of f+if+i and g+ig+i. Let D1D_1 be the union of these marked points.

The matching pairs give SS-integral points on C\(D0∪D1)C\setminus(D_0\cup D_1) for a fixed finite set of places SS. To see this, enlarge the number field to split the finitely many poles and zeros involved. Choose a smooth proper model outside finitely many places, so that distinct special points have disjoint reductions and the zero and pole divisors of the functions have no vertical components. At any place outside that set, reducing to a pole would give a nonintegral coordinate. Reducing to a point of D1D_1 would make exactly one of the two corresponding integers divisible by the underlying rational prime. This contradicts the aligned equality of their radicals. Therefore all reductions outside SS avoid the deleted points.

By Siegel's theorem for SS-integral points the normalization has genus zero and t+|D1|≤2t+|D_1|\le2, since t>0t>0. Take d=max(degf,degg)d=\max(\deg f,\deg g) and exchange the functions in the proof if needed so that degf=d\deg f=d. If ff has tft_f distinct poles, the ramification contribution over infinity is d−tfd-t_f. Riemann–Hurwitz leaves at most d+tf−2d+t_f-2 ramification units on the finite fibers. Thus the kk fibers over 0,−1,…,−(k−1)0,-1,\ldots,-(k-1) contain at least (k−1)d+2−tf(k-1)d+2-t_f distinct points. At most t−tft-t_f of them lie in D0D_0. Of the remaining points, at most 2−t2-t lie in D1D_1. Consequently at least (k−1)d≥2d(k-1)d\ge 2d distinct points are shared with the corresponding fibers of gg. Each is a zero of f−gf-g.

There is a common pole of ff and gg. In fact, since both coordinate maps are nonconstant, only finitely many curve points can have either coordinate in a given finite set. The infinitely many distinct positive integer pairs therefore have a subsequence along which both coordinates tend to infinity. A limit point on the compact complex curve is a pole of both maps. At a common pole the order of the pole of f−gf-g is at most the maximum of the two pole orders. It follows that the total pole degree of f−gf-g is at most degf+degg−1≤2d−1\deg f + \deg g -1\le2d-1. A nonzero rational function cannot have more distinct zeros than its pole degree. This contradicts the shared-fiber count.

Consequence

If there are infinitely many length-three matches, their pairs are Zariski dense in the affine plane. The proof also gives a sufficient route to finiteness: an unconditional algebraic constraint placing all but finitely many matches on a proper closed subset would finish that step. The existing radical divisibility argument gives congruence constraints and does not establish such a containment.

Multiplicative dependence is excluded at every aligned position

Suppose the length is at least three. Then no aligned pair of terms can be two powers of a common positive integer base. In particular, neither block contains a prime power.

To prove this, consider aligned terms au,ava^u,a^v for integers a≥2a\ge2 and 1≤u<v1\le u\lt v. If they have a matching successor, a primitive prime divisor of a2v−1a^{2v}-1 divides av+1a^v+1 and cannot divide au+1a^u+1. By Zsigmondy's theorem the only applicable exception for 2v≥42v\ge4 is a=2,v=3a=2,v=3. In that case the only equal successor radicals occur at u=1u=1, giving aligned terms 2 and 8. Their predecessors 1 and 7 have different radicals. Their second successors 4 and 10 also have different radicals. Thus they cannot be part of a longer match.

If the aligned terms are last, use their predecessors. A primitive prime divisor of av−1a^v-1 distinguishes the predecessor radicals except when v=2v=2 and a+1a+1 is a power of two, or when a=2,v=6a=2,v=6. In the latter case the support of 63 is {3,7}\{3,7\}, different from each of the supports of 2u−12^u-1 for 1≤u<61\le u\lt6. In the former case u=1u=1, aa is odd, and gcd(a−2,a2−2)=gcd(a−2,2)=1\gcd(a-2,a^2-2)=\gcd(a-2,2)=1. The second predecessors therefore have different radicals because a2−2>1a^2-2>1. These cases exhaust all positions. A term equal to 1 also cannot occur in a matching pair of distinct starts because its partner would have radical 1.

Counting matches by averaging both coordinates

Let Pk(X)P_k(X) count pairs of positive integers x<y≤Xx\lt y\le X with matching blocks of length kk. For fixed k≥2k\ge2,

Pk(X)≪kX1/kexp(O((logXloglogX)2/3)).P_k(X)\ll_k X^{1/k} \exp\!\left(O\!\left( \left(\frac{\log X}{\log\log X}\right)^{2/3} \right)\right).

This gives a smaller subexponential factor than the one in Lebowitz-Lockard, Theorem 6. The power X1/kX^{1/k} is the same.

For a matched pair, put ri=rad(y+i)r_i=\operatorname{rad}(y+i). The least common multiple of the rir_i divides y−xy-x. A prime pp occurs in at most 1+⌊(k−1)/p⌋1+\lfloor(k-1)/p\rfloor terms. Consequently the product of the rir_i is at most XCkX C_k, where

Ck=∏p≤k−1p⌊(k−1)/p⌋.C_k=\prod_{p\le k-1}p^{\lfloor(k-1)/p\rfloor}.

There is at least one shift ii with ri≤B=(XCk)1/kr_i\le B=(X C_k)^{1/k}. Put Y=X+kY=X+k. For a fixed choice of shift, the number of candidate pairs is at most ∑r≤BN(Y,r)2\sum_{r\le B} N(Y,r)^2, where N(Y,r)N(Y,r) counts positive integers at most YY with radical exactly rr. The full count is at most kk times that sum.

For squarefree rr and η>0\eta>0, multiplying the geometric series for the exponents of its primes gives

N(Y,r)≤Yη∏p∣r(pη−1)−1.N(Y,r)\le Y^\eta \prod_{p\mid r}(p^\eta-1)^{-1}.

Using 1≤B/r1\le B/r in the sum and then expanding an Euler product gives the following bound for the candidate pairs at a fixed shift.

∑r≤BN(Y,r)2≤BY2η∏p(1+1p(pη−1)2).\sum_{r\le B}N(Y,r)^2 \le B Y^{2\eta}\prod_p\left(1+\frac{1}{p(p^\eta-1)^2}\right).

The product converges for every fixed positive η\eta. For small η\eta, put Z=1/ηZ=1/\eta. The logarithms of factors for p≤Zp\le Z sum to O(Z)O(Z): each is O(logZ)O(\log Z), and Chebyshev's bound gives O(Z/logZ)O(Z/\log Z) such primes. For p>Zp>Z, use pη−1≥ηlogpp^\eta-1\ge\eta\log p. Partial summation with Chebyshev's bound gives ∑p>Z1/(p(logp)2)=O((logZ)−2)\sum_{p>Z}1/(p(\log p)^2)=O((\log Z)^{-2}). The logarithm of the product is therefore O(η−2/log2(1/η))O(\eta^{-2}/\log^2(1/\eta)). Taking η=(logY(loglogY)2)−1/3\eta=(\log Y\,(\log\log Y)^2)^{-1/3} proves the estimate.

This bound controls the number of candidates below a height cutoff. A proof of an absolute upper bound on matching length still requires an argument covering arbitrarily sparse, unbounded pairs.