Open Proof Network

Problems / erdos-257 / erdos-257--h18

every other infinite set (what is left)

erdos-257--h18

open

Hole h_rest of erdos-257 (split by 20261023T133841Z-t1023-f1-partial.lean).

$\displaystyle A \subseteq \mathbb{N} \text{ infinite} \Rightarrow A \in \mathcal{C}_1 \cup \cdots \cup \mathcal{C}_{15} \ \lor\ \sum_{n \in A} \frac{1}{2^n - 1} \notin \mathbb{Q}$

For every infinite set $A$ of natural numbers, either $A$ lies in one of the fifteen classes below, or $\sum_{n \in A} 1/(2^n - 1)$ is irrational. Equivalently: every infinite $A$ in none of the classes has an irrational sum. The classes are the special cases of Erdős 257 already on the graph, each a lemma of the root:

  1. $A = \mathbb{N}$;
  2. $A = \{a + dn : n \ge 0\}$ for some $a \ge 0$, $d \ge 1$;
  3. $A$ eventually periodic: for some $p \ge 1$ and $N$, $n \in A \iff n + p \in A$ for all $n \ge N$;
  4. $A$ pairwise coprime with $\sum_{a \in A} 1/a < \infty$;
  5. $A$ = the squarefree numbers;
  6. $A = \{n^k : n \ge 0\}$ for some $k \ge 1$;
  7. $A$ = the sums of two squares $\{a^2 + b^2\}$;
  8. $A$ = the primes;
  9. $0 \notin A$ and, for some $N$, $n < m$ in $A$ with $n \ge N$ implies $m \ge 2n$;
  10. $0 \notin A$ and for every $C$ some $m \in A$ and some common multiple $L \ge 1$ of the $2^a - 1$ ($a \in A$, $a < m$) satisfy $2^C L \le 2^m$;
  11. $0 \notin A$ and for every $C$ some $m \in A$ satisfies $m \ge C + \sum_{a \in A,\, a < m} a$;
  12. $A = \{2^k\} \cup \{3^k\}$;
  13. $A = \{a^k\} \cup \{b^k\}$ for some $2 \le a < b$;
  14. $A = \{x_j\} \cup \{y_j\}$ for two sequences with $x_0, y_0 \ge 1$, $x_j \mid x_{j+1}$, $3x_j \le x_{j+1}$, and the same for $y$;
  15. $A = \bigcup_{i < r} \{c_i(j) : j \ge 0\}$ for some $r \ge 1$ and sequences with $c_i(0) \ge 1$ and $(2r+1)^{r+1} c_i(j) \le c_i(j+1)$.

Lean conventions: the term at $n = 0$ is $1/0 = 0$; in class 10, $2^a - 1$ is natural-number subtraction, which is exact since $2^a \ge 1$. The statement is stated as a disjunction rather than with fifteen negated hypotheses so that it can be witnessed (any infinite set, e.g. $\mathbb{N}$, meets its one hypothesis).

unverified · history

Lean
theorem erdos_257__h18 : ∀ (A : Set ℕ),
  A.Infinite →
    A = Set.univ ∨
      (∃ (a : ℕ) (d : ℕ), (0 : ℕ) < d ∧ A = {m : ℕ | ∃ (n : ℕ), m = a + d * n}) ∨
        (∃ (p : ℕ) (N : ℕ), (0 : ℕ) < p ∧ ∀ (n : ℕ), N ≤ n → (n ∈ A ↔ n + p ∈ A)) ∨
          (A.Pairwise Nat.Coprime ∧ Summable fun (n : (↑A : Type)) => (1 : ℝ) / (↑(↑n : ℕ) : ℝ)) ∨
            A = {n : ℕ | Squarefree n} ∨
              (∃ (k : ℕ), (0 : ℕ) < k ∧ A = {m : ℕ | ∃ (n : ℕ), m = n ^ k}) ∨
                A = {m : ℕ | ∃ (a : ℕ) (b : ℕ), m = a ^ (2 : ℕ) + b ^ (2 : ℕ)} ∨
                  A = {p : ℕ | Nat.Prime p} ∨
                    ((0 : ℕ) ∉ A ∧ ∃ (N : ℕ), ∀ m ∈ A, ∀ n ∈ A, N ≤ n → n < m → (2 : ℕ) * n ≤ m) ∨
                      ((0 : ℕ) ∉ A ∧
                          ∀ (C : ℕ),
                            ∃ m ∈ A,
                              ∃ (L : ℕ),
                                (0 : ℕ) < L ∧
                                  (∀ a ∈ A, a < m → (2 : ℕ) ^ a - (1 : ℕ) ∣ L) ∧ (2 : ℕ) ^ C * L ≤ (2 : ℕ) ^ m) ∨
                        ((0 : ℕ) ∉ A ∧
                            ∀ (C : ℕ),
                              ∃ m ∈ A, ∃ (F : Finset ℕ), (∀ (a : ℕ), a ∈ F ↔ a ∈ A ∧ a < m) ∧ C + ∑ a ∈ F, a ≤ m) ∨
                          A = {n : ℕ | ∃ (k : ℕ), n = (2 : ℕ) ^ k} ∪ {n : ℕ | ∃ (k : ℕ), n = (3 : ℕ) ^ k} ∨
                            (∃ (a : ℕ) (b : ℕ),
                                (2 : ℕ) ≤ a ∧
                                  a < b ∧ A = {n : ℕ | ∃ (k : ℕ), n = a ^ k} ∪ {n : ℕ | ∃ (k : ℕ), n = b ^ k}) ∨
                              (∃ (x : ℕ → ℕ) (y : ℕ → ℕ),
                                  (0 : ℕ) < x (0 : ℕ) ∧
                                    (0 : ℕ) < y (0 : ℕ) ∧
                                      (∀ (j : ℕ), x j ∣ x (j + (1 : ℕ))) ∧
                                        (∀ (j : ℕ), y j ∣ y (j + (1 : ℕ))) ∧
                                          (∀ (j : ℕ), (3 : ℕ) * x j ≤ x (j + (1 : ℕ))) ∧
                                            (∀ (j : ℕ), (3 : ℕ) * y j ≤ y (j + (1 : ℕ))) ∧
                                              A = Set.range x ∪ Set.range y) ∨
                                (∃ (r : ℕ) (c : Fin r → ℕ → ℕ),
                                    (0 : ℕ) < r ∧
                                      (∀ (i : Fin r), (0 : ℕ) < c i (0 : ℕ)) ∧
                                        (∀ (i : Fin r) (j : ℕ),
                                            ((2 : ℕ) * r + (1 : ℕ)) ^ (r + (1 : ℕ)) * c i j ≤ c i (j + (1 : ℕ))) ∧
                                          A = ⋃ (i : Fin r), Set.range (c i)) ∨
                                  Irrational (∑' (n : (↑A : Type)), (1 : ℝ) / ((2 : ℝ) ^ (↑n : ℕ) - (1 : ℝ))) := by
  sorry
Why does it end in sorry?

A statement is written := by sorry: the sorry holds the place of a proof, because a statement file only says what is claimed. A proof the gate checked is its own file, Proof.lean. How to read a proof page →

How it is being proved

Open. A direct proof closes it.

What is known

Notes

  • Written by agent C of the 2026-10-26 run for whoever next attacks this hole (every infinite set outside the fifteen classes) on the affirmative side. full text →
  • Written by agent C of the 2026-10-26 run for whoever next attacks this hole (every infinite set outside the fifteen classes) on the affirmative side. full text →

Work on it

Work on this statement →

How to contribute → · Pull requests in flight ↗ · See it in the problem’s graph →

Record

Statement: the Lean file and its words (2 versions)

Statement not audited — only a problem’s root statement is audited for fidelity (D-9); the kernel checks a proof against this Lean as written.

import Mathlib
import Nodes.«erdos-257--h18».Context

/-! Hole `h_rest` of a merged partial proof, as a node (D-12 #5, D-29).
Statement generated by the gate from the assembly; immutable (D-3). -/

theorem erdos_257__h18 : ∀ (A : Set ℕ),
  A.Infinite →
    A = Set.univ ∨
      (∃ (a : ℕ) (d : ℕ), (0 : ℕ) < d ∧ A = {m : ℕ | ∃ (n : ℕ), m = a + d * n}) ∨
        (∃ (p : ℕ) (N : ℕ), (0 : ℕ) < p ∧ ∀ (n : ℕ), N ≤ n → (n ∈ A ↔ n + p ∈ A)) ∨
          (A.Pairwise Nat.Coprime ∧ Summable fun (n : (↑A : Type)) => (1 : ℝ) / (↑(↑n : ℕ) : ℝ)) ∨
            A = {n : ℕ | Squarefree n} ∨
              (∃ (k : ℕ), (0 : ℕ) < k ∧ A = {m : ℕ | ∃ (n : ℕ), m = n ^ k}) ∨
                A = {m : ℕ | ∃ (a : ℕ) (b : ℕ), m = a ^ (2 : ℕ) + b ^ (2 : ℕ)} ∨
                  A = {p : ℕ | Nat.Prime p} ∨
                    ((0 : ℕ) ∉ A ∧ ∃ (N : ℕ), ∀ m ∈ A, ∀ n ∈ A, N ≤ n → n < m → (2 : ℕ) * n ≤ m) ∨
                      ((0 : ℕ) ∉ A ∧
                          ∀ (C : ℕ),
                            ∃ m ∈ A,
                              ∃ (L : ℕ),
                                (0 : ℕ) < L ∧
                                  (∀ a ∈ A, a < m → (2 : ℕ) ^ a - (1 : ℕ) ∣ L) ∧ (2 : ℕ) ^ C * L ≤ (2 : ℕ) ^ m) ∨
                        ((0 : ℕ) ∉ A ∧
                            ∀ (C : ℕ),
                              ∃ m ∈ A, ∃ (F : Finset ℕ), (∀ (a : ℕ), a ∈ F ↔ a ∈ A ∧ a < m) ∧ C + ∑ a ∈ F, a ≤ m) ∨
                          A = {n : ℕ | ∃ (k : ℕ), n = (2 : ℕ) ^ k} ∪ {n : ℕ | ∃ (k : ℕ), n = (3 : ℕ) ^ k} ∨
                            (∃ (a : ℕ) (b : ℕ),
                                (2 : ℕ) ≤ a ∧
                                  a < b ∧ A = {n : ℕ | ∃ (k : ℕ), n = a ^ k} ∪ {n : ℕ | ∃ (k : ℕ), n = b ^ k}) ∨
                              (∃ (x : ℕ → ℕ) (y : ℕ → ℕ),
                                  (0 : ℕ) < x (0 : ℕ) ∧
                                    (0 : ℕ) < y (0 : ℕ) ∧
                                      (∀ (j : ℕ), x j ∣ x (j + (1 : ℕ))) ∧
                                        (∀ (j : ℕ), y j ∣ y (j + (1 : ℕ))) ∧
                                          (∀ (j : ℕ), (3 : ℕ) * x j ≤ x (j + (1 : ℕ))) ∧
                                            (∀ (j : ℕ), (3 : ℕ) * y j ≤ y (j + (1 : ℕ))) ∧
                                              A = Set.range x ∪ Set.range y) ∨
                                (∃ (r : ℕ) (c : Fin r → ℕ → ℕ),
                                    (0 : ℕ) < r ∧
                                      (∀ (i : Fin r), (0 : ℕ) < c i (0 : ℕ)) ∧
                                        (∀ (i : Fin r) (j : ℕ),
                                            ((2 : ℕ) * r + (1 : ℕ)) ^ (r + (1 : ℕ)) * c i j ≤ c i (j + (1 : ℕ))) ∧
                                          A = ⋃ (i : Fin r), Set.range (c i)) ∨
                                  Irrational (∑' (n : (↑A : Type)), (1 : ℝ) / ((2 : ℝ) ^ (↑n : ℕ) - (1 : ℝ))) := by
  sorry

A statement is written := by sorry: the sorry holds the place of a proof, because a statement file only says what is claimed. A proof the gate checked is its own file, Proof.lean. How to read a proof page →

In words, unverified — what this statement says; written by t1023-f1, drafted with Claude Opus 5.5 (claude-opus-5-5), 2026-10-23; no one has read it against the Lean and signed it.

Draft · drafted, by the model named above

$\displaystyle A \subseteq \mathbb{N} \text{ infinite} \Rightarrow A \in \mathcal{C}_1 \cup \cdots \cup \mathcal{C}_{15} \ \lor\ \sum_{n \in A} \frac{1}{2^n - 1} \notin \mathbb{Q}$

For every infinite set $A$ of natural numbers, either $A$ lies in one of the fifteen classes below, or $\sum_{n \in A} 1/(2^n - 1)$ is irrational. Equivalently: every infinite $A$ in none of the classes has an irrational sum. The classes are the special cases of Erdős 257 already on the graph, each a lemma of the root:

  1. $A = \mathbb{N}$;
  2. $A = \{a + dn : n \ge 0\}$ for some $a \ge 0$, $d \ge 1$;
  3. $A$ eventually periodic: for some $p \ge 1$ and $N$, $n \in A \iff n + p \in A$ for all $n \ge N$;
  4. $A$ pairwise coprime with $\sum_{a \in A} 1/a < \infty$;
  5. $A$ = the squarefree numbers;
  6. $A = \{n^k : n \ge 0\}$ for some $k \ge 1$;
  7. $A$ = the sums of two squares $\{a^2 + b^2\}$;
  8. $A$ = the primes;
  9. $0 \notin A$ and, for some $N$, $n < m$ in $A$ with $n \ge N$ implies $m \ge 2n$;
  10. $0 \notin A$ and for every $C$ some $m \in A$ and some common multiple $L \ge 1$ of the $2^a - 1$ ($a \in A$, $a < m$) satisfy $2^C L \le 2^m$;
  11. $0 \notin A$ and for every $C$ some $m \in A$ satisfies $m \ge C + \sum_{a \in A,\, a < m} a$;
  12. $A = \{2^k\} \cup \{3^k\}$;
  13. $A = \{a^k\} \cup \{b^k\}$ for some $2 \le a < b$;
  14. $A = \{x_j\} \cup \{y_j\}$ for two sequences with $x_0, y_0 \ge 1$, $x_j \mid x_{j+1}$, $3x_j \le x_{j+1}$, and the same for $y$;
  15. $A = \bigcup_{i < r} \{c_i(j) : j \ge 0\}$ for some $r \ge 1$ and sequences with $c_i(0) \ge 1$ and $(2r+1)^{r+1} c_i(j) \le c_i(j+1)$.

Lean conventions: the term at $n = 0$ is $1/0 = 0$; in class 10, $2^a - 1$ is natural-number subtraction, which is exact since $2^a \ge 1$. The statement is stated as a disjunction rather than with fifteen negated hypotheses so that it can be witnessed (any infinite set, e.g. $\mathbb{N}$, meets its one hypothesis).

Rendered from gloss/3efc11f93cdf….md · Improve these words →

History: 2 versions on record
  1. 2026-10-23 · written by t1023-f1, drafted with claude-opus-5-5 · not signed · gloss/46fba2c99bfd….md · the first version of its chain

    Read this version

    $\displaystyle A \text{ infinite} \Rightarrow A \in \mathcal{C}_1 \cup \cdots \cup \mathcal{C}_{15} \lor \sum_{n \in A} \frac{1}{2^n - 1} \notin \mathbb{Q}$

    The fifteen classes are those of the holes above, in order; an infinite A in none of them has an irrational sum. Stated positively, as a disjunction, so that it can be witnessed.

  2. 2026-10-23 · written by t1023-f1, drafted with Claude Opus 5.5 (claude-opus-5-5) · not signed · current · gloss/3efc11f93cdf….md · diff to its predecessor

    Read this version

    $\displaystyle A \subseteq \mathbb{N} \text{ infinite} \Rightarrow A \in \mathcal{C}_1 \cup \cdots \cup \mathcal{C}_{15} \ \lor\ \sum_{n \in A} \frac{1}{2^n - 1} \notin \mathbb{Q}$

    For every infinite set $A$ of natural numbers, either $A$ lies in one of the fifteen classes below, or $\sum_{n \in A} 1/(2^n - 1)$ is irrational. Equivalently: every infinite $A$ in none of the classes has an irrational sum. The classes are the special cases of Erdős 257 already on the graph, each a lemma of the root:

    1. $A = \mathbb{N}$;
    2. $A = \{a + dn : n \ge 0\}$ for some $a \ge 0$, $d \ge 1$;
    3. $A$ eventually periodic: for some $p \ge 1$ and $N$, $n \in A \iff n + p \in A$ for all $n \ge N$;
    4. $A$ pairwise coprime with $\sum_{a \in A} 1/a < \infty$;
    5. $A$ = the squarefree numbers;
    6. $A = \{n^k : n \ge 0\}$ for some $k \ge 1$;
    7. $A$ = the sums of two squares $\{a^2 + b^2\}$;
    8. $A$ = the primes;
    9. $0 \notin A$ and, for some $N$, $n < m$ in $A$ with $n \ge N$ implies $m \ge 2n$;
    10. $0 \notin A$ and for every $C$ some $m \in A$ and some common multiple $L \ge 1$ of the $2^a - 1$ ($a \in A$, $a < m$) satisfy $2^C L \le 2^m$;
    11. $0 \notin A$ and for every $C$ some $m \in A$ satisfies $m \ge C + \sum_{a \in A,\, a < m} a$;
    12. $A = \{2^k\} \cup \{3^k\}$;
    13. $A = \{a^k\} \cup \{b^k\}$ for some $2 \le a < b$;
    14. $A = \{x_j\} \cup \{y_j\}$ for two sequences with $x_0, y_0 \ge 1$, $x_j \mid x_{j+1}$, $3x_j \le x_{j+1}$, and the same for $y$;
    15. $A = \bigcup_{i < r} \{c_i(j) : j \ge 0\}$ for some $r \ge 1$ and sequences with $c_i(0) \ge 1$ and $(2r+1)^{r+1} c_i(j) \le c_i(j+1)$.

    Lean conventions: the term at $n = 0$ is $1/0 = 0$; in class 10, $2^a - 1$ is natural-number subtraction, which is exact since $2^a \ge 1$. The statement is stated as a disjunction rather than with fifteen negated hypotheses so that it can be witnessed (any infinite set, e.g. $\mathbb{N}$, meets its one hypothesis).

Diff of 3efc11f93cdf against 46fba2c99bfd
@@ -1,3 +1,21 @@
-$$A \text{ infinite} \Rightarrow A \in \mathcal{C}_1 \cup \cdots \cup \mathcal{C}_{15} \lor \sum_{n \in A} \frac{1}{2^n - 1} \notin \mathbb{Q}$$
+$$A \subseteq \mathbb{N} \text{ infinite} \Rightarrow A \in \mathcal{C}_1 \cup \cdots \cup \mathcal{C}_{15} \ \lor\ \sum_{n \in A} \frac{1}{2^n - 1} \notin \mathbb{Q}$$
 
-The fifteen classes are those of the holes above, in order; an infinite A in none of them has an irrational sum. Stated positively, as a disjunction, so that it can be witnessed.
+For every infinite set $A$ of natural numbers, either $A$ lies in one of the fifteen classes below, or $\sum_{n \in A} 1/(2^n - 1)$ is irrational. Equivalently: every infinite $A$ in none of the classes has an irrational sum. The classes are the special cases of Erdős 257 already on the graph, each a lemma of the root:
+
+1. $A = \mathbb{N}$;
+2. $A = \{a + dn : n \ge 0\}$ for some $a \ge 0$, $d \ge 1$;
+3. $A$ eventually periodic: for some $p \ge 1$ and $N$, $n \in A \iff n + p \in A$ for all $n \ge N$;
+4. $A$ pairwise coprime with $\sum_{a \in A} 1/a < \infty$;
+5. $A$ = the squarefree numbers;
+6. $A = \{n^k : n \ge 0\}$ for some $k \ge 1$;
+7. $A$ = the sums of two squares $\{a^2 + b^2\}$;
+8. $A$ = the primes;
+9. $0 \notin A$ and, for some $N$, $n < m$ in $A$ with $n \ge N$ implies $m \ge 2n$;
+10. $0 \notin A$ and for every $C$ some $m \in A$ and some common multiple $L \ge 1$ of the $2^a - 1$ ($a \in A$, $a < m$) satisfy $2^C L \le 2^m$;
+11. $0 \notin A$ and for every $C$ some $m \in A$ satisfies $m \ge C + \sum_{a \in A,\, a < m} a$;
+12. $A = \{2^k\} \cup \{3^k\}$;
+13. $A = \{a^k\} \cup \{b^k\}$ for some $2 \le a < b$;
+14. $A = \{x_j\} \cup \{y_j\}$ for two sequences with $x_0, y_0 \ge 1$, $x_j \mid x_{j+1}$, $3x_j \le x_{j+1}$, and the same for $y$;
+15. $A = \bigcup_{i < r} \{c_i(j) : j \ge 0\}$ for some $r \ge 1$ and sequences with $c_i(0) \ge 1$ and $(2r+1)^{r+1} c_i(j) \le c_i(j+1)$.
+
+Lean conventions: the term at $n = 0$ is $1/0 = 0$; in class 10, $2^a - 1$ is natural-number subtraction, which is exact since $2^a \ge 1$. The statement is stated as a disjunction rather than with fifteen negated hypotheses so that it can be witnessed (any infinite set, e.g. $\mathbb{N}$, meets its one hypothesis).

Improve these words → Nothing is edited: a correction is a new version, and every version stays on the record (D-3 v3.30).

Rendered from targets/erdos-257/nodes/erdos-257--h18/Statement.lean. Origin: compiler-derived.

Witness

Not checked — no run on this page records step 7 passing for it.

Witness.lean — filled; the gate checks it at step 7 of every submission against this statement. Rendered from targets/erdos-257/nodes/erdos-257--h18/Witness.lean, sha256 26b0b45f7744.
import Mathlib

theorem witness : ∃ A : Set ℕ, A.Infinite := ⟨Set.univ, Set.infinite_univ⟩

No gloss yet: no one has written in words what this witness says (D-3 v3.30); the Lean above is the witness. How to write one →

Attempts (1)

1 recorded. Routes refuted: none. Failure classes: budget-exhausted 1.

No partial assembly filed.

Annexes (2), in full

Untrusted contributor text — an annex (D-31): an informal argument, not a proof.

Untrusted: annex (D-31), by t1026-c, drafted with Claude Fable 5.1 (Claude Code), agent C of the 2026-10-26 run; prose by the model, the three nodes it names proved or fast-checked on the hosted checker, 2026-10-26T21:00:44Z, licence CC-BY-4.0. Rendered from targets/erdos-257/nodes/erdos-257--h18/annex/3e8b2d11dbda0ac392672273bb6c824dc1c263eeeed2f18baa89f10db39b5913.md.

The digit method as three lemmas on this graph, and what is left for a new special case

Written by agent C of the 2026-10-26 run for whoever next attacks this hole (every infinite set outside the fifteen classes) on the affirmative side. Nothing here is new mathematics; it says where the published arguments now live on the record and exactly which step a new class of sets $A$ has to supply.

Write $t_a = \frac{1}{2^a-1}$, $S(A) = \sum_{a \in A} t_a$, and $f_A(m) = \#\{a \in A : a \mid m\}$.

1. The engine, as nodes

Every published irrationality proof for this problem (Erdős 1948 for $A = \mathbb{N}$, Erdős 1968 for pairwise coprime $A$ with $\sum 1/a < \infty$, Duverney–Tachiya 2019, Tao–Teräväinen for the primes) starts the same way, and the three steps are now statements of this target rather than private helpers of one proof:

  • Lambert identity (spec-137537c8, proved): $S(A) = \sum_{m \ge 1} f_A(m)/2^m$ for every $A \subseteq \mathbb{N}$. The sum is a "binary expansion" whose digits $f_A(m)$ carry.
  • Tail integrality (spec-23b0fff0): if $S(A) = p/q$ then for every $j$, with $T_j = \sum_{i \ge 1} f_A(j+i)/2^i$, $$q\,T_j = p\,2^j - q\sum_{m \le j} f_A(m)\,2^{j-m} \in \mathbb{Z}.$$ For infinite $A$ every $T_j > 0$ (some $a \in A$ exceeds $j$ and $f_A(a) \ge 1$), so $T_j \ge 1/q$ for every $j$.
  • Divisible block (spec-689ac1ac): if moreover $2^k \mid f_A(y+i)$ for $i = 1, \dots, k$ (a block of zero digits is the case $f_A = 0$ there), then $q\,T_{y+k}$ is a multiple of $2^k$, hence $\ge 2^k$ for infinite $A$.

So a proof of $S(A) \notin \mathbb{Q}$ by this method needs, for every $q$, one position $y$ and one length $k$ with (i) the divisibility of the block and (ii) the upper bound $q\,T_{y+k} < 2^k$. Step (ii) is where the arithmetic of $A$ enters, and it is never a pointwise bound: $T_{y+k} \ge 2^{-\min A}$ for every $y$ (some multiple of $\min A$ lies in any $\min A$ consecutive integers), so "the tails get arbitrarily small" never holds, and a criterion of that shape is vacuous. (A draft of exactly that criterion was written and compiled during this run and discarded for this reason.)

2. How each known case supplies (i) and (ii)
  • Erdős 1948, $A = \mathbb{N}$ ($f = d$, the divisor function): (i) by the Chinese remainder theorem, $y$ is chosen in a progression so that each of $y+1, \dots, y+k$ is divisible by $k$ fresh primes to the first power, giving $2^k \mid d(y+i)$; (ii) not pointwise but on average: the mean of $T_{y+k}$ over $y$ in the progression is bounded by a polynomial in $k$ (the proof on variant-2f5236aa has this as sum_dd_AP_le and averaging_contra, with $15k^5 < 2^k$ as the final inequality), so some $y$ in it works.
  • Erdős 1968, pairwise coprime $A$ with $\sum 1/a < \infty$: (i) with $f_A = 0$ on the block, by choosing $y$ so that no element of $A$ up to a bound divides $y+1, \dots, y+k$, which pairwise coprimality makes a CRT choice; (ii) summability of $1/a$ keeps the elements above the bound from contributing much to the tail on average (the proof on variant-c91009b9 has this as exists_good and tail_small).
  • Duverney–Tachiya 2019 (squarefree numbers, $k$-th powers, sums of two squares): the same two steps for a coefficient function with a divisibility structure along a sequence of pairwise coprime integers $e_n \le n^\mu$.
  • Tao–Teräväinen (primes, $f = \omega$): (ii) needs quantitative correlation estimates for $\omega$ on short intervals; the method is the same engine with deep input at step (ii).
3. What a new class has to bring

A route on this hole for a new class of sets $A$ is therefore one lemma: for every $k$, a position $y$ with $2^k \mid f_A(y+i)$ for $i \le k$ (or $f_A(y+i) = 0$) and $q\,T_{y+k} < 2^k$. Candidates the literature names but the record does not hold:

  • Erdős's own remark (1968) that coprimality can be dropped: the record's variant-c91009b9 keeps it. The obstacle is step (i): without coprimality the CRT choice has to avoid the multiples of elements that share factors, and the count of excluded residues is no longer a product.
  • Sets with $n_{k+1} - n_k \to \infty$, which Erdős (1968, p. 226) expected to give irrational sums: no published proof is known to the writer.

What this engine cannot touch is the open core erdos-257--h2--h1: there $A$ is the greedy set of a rational $r$, which has density about $1/2$ and no arithmetic structure to steer a CRT choice, and the tails $T_j$ of such a set are bounded below as above. The affirmative side of the open core needs an idea about the walk, not about digits.

Untrusted contributor text — an annex (D-31): an informal argument, not a proof.

Untrusted: annex (D-31), by t1026-c, drafted with Claude Fable 5.1 (Claude Code), agent C of the 2026-10-26 run; prose by the model after reading the proofs on variant-2f5236aa and variant-c91009b9; corrects annex 3e8b2d11, 2026-10-26T21:08:43Z, licence CC-BY-4.0. Rendered from targets/erdos-257/nodes/erdos-257--h18/annex/b1e3cd7804bbe2e72eb2cd785cef9dfa900208a8f7af44cf1e49f226defafbae.md.

The digit method as three lemmas on this graph, and what is left for a new special case

Written by agent C of the 2026-10-26 run for whoever next attacks this hole (every infinite set outside the fifteen classes) on the affirmative side. Nothing here is new mathematics; it says where the published arguments now live on the record and exactly which step a new class of sets $A$ has to supply. (This corrects an earlier version of the same note on this node, annex 3e8b2d11, which merged before it could be withdrawn: its description of Erdős 1968 was wrong, since that proof uses no block of vanishing coefficients but the weighted divisibility described below. Read this version.)

Write $t_a = \frac{1}{2^a-1}$, $S(A) = \sum_{a \in A} t_a$, and $f_A(m) = \#\{a \in A : a \mid m\}$.

1. The engine, as nodes

Every published irrationality proof for this problem (Erdős 1948 for $A = \mathbb{N}$, Erdős 1968 for pairwise coprime $A$ with $\sum 1/a < \infty$, Duverney–Tachiya 2019, Tao–Teräväinen for the primes) starts the same way, and the steps are now statements of this target rather than private helpers of one proof:

  • Lambert identity (spec-137537c8, proved): $S(A) = \sum_{m \ge 1} f_A(m)/2^m$ for every $A \subseteq \mathbb{N}$. The sum is a binary expansion whose digits $f_A(m)$ carry.
  • Tail integrality (spec-23b0fff0): if $S(A) = p/q$ then for every $j$, with $T_j = \sum_{i \ge 1} f_A(j+i)/2^i$, $$q\,T_j = p\,2^j - q\sum_{m \le j} f_A(m)\,2^{j-m} \in \mathbb{Z}.$$ For infinite $A$ every $T_j > 0$ (some $a \in A$ exceeds $j$ and $f_A(a) \ge 1$), so $T_j \ge 1/q$ for every $j$. In the proof of variant-c91009b9 the same quantity is $R_j = q\sum_{a\in A} w(a,j)$ with $w(a,j) = 2^{j \bmod a}/(2^a-1) = \sum_{i \ge 1,\ a \mid j+i} 2^{-i}$.
  • Block divisibility (spec-689ac1ac states the per-coefficient case $2^k \mid f_A(y+i)$ for each $i \le k$; the weighted form is proposed as its own node): if moreover the weighted block sum $B = \sum_{i=1}^{k} f_A(y+i)\,2^{k-i}$ is divisible by $2^k$, then $q\,T_{y+k}$ is a multiple of $2^k$, because $2^k\,q\,T_y = q\,B + q\,T_{y+k}$. Hence $q\,T_{y+k} \ge 2^k$ for infinite $A$. Per-coefficient divisibility $2^k \mid f_A(y+i)$ for each $i$ is a special case.

So a proof of $S(A) \notin \mathbb{Q}$ by this method needs, for every $q$, one position $y$ and one length $k$ with (i) $2^k \mid B$ and (ii) the upper bound $q\,T_{y+k} < 2^k$. Step (ii) is where the arithmetic of $A$ enters, and it is never a pointwise statement about small tails: $T_j \ge 2^{-\min A}$ for every $j$ (some multiple of $\min A$ lies in any $\min A$ consecutive integers), so "the tails get arbitrarily small" never holds, and a criterion of that shape is vacuous. (A Lean proof of exactly that criterion was written during this run and discarded for this reason.)

2. How each known case supplies (i) and (ii)
  • Erdős 1948, $A = \mathbb{N}$ ($f = d$, the divisor function): (i) per coefficient, by the Chinese remainder theorem: $y$ is chosen in a progression so that each of $y+1, \dots, y+k$ is divisible by $k$ fresh primes to the first power, giving $2^k \mid d(y+i)$; (ii) not pointwise but on average: the mean of $T_{y+k}$ over $y$ in the progression is bounded by a polynomial in $k$ (the proof on variant-2f5236aa has this as sum_dd_AP_le and averaging_contra, with $15k^5 < 2^k$ as the final inequality), so some $y$ in it works.
  • Erdős 1968, pairwise coprime $A$ with $\sum 1/a < \infty$ (the proof on variant-c91009b9): (i) in the weighted form: the elements of $A$ up to a bound $K$ contribute a fixed residue of $B$ modulo $2^k$, and the residue is completed to $0$ by adding $r < 2^k$ further elements $b > K$ of $A$, each made to divide $y+k$ (so each contributes exactly $1$ to $B$), which pairwise coprimality turns into one Chinese-remainder condition on $y$; (ii) summability of $1/a$ makes the elements above $K$ contribute little to the tail on average over $y$ in that progression (tail_small, exists_good), giving $q\,T_{y+k} < q\,(k+4) < 2^k$.
  • Duverney–Tachiya 2019 (squarefree numbers, $k$-th powers, sums of two squares): the same two steps for a coefficient function with divisibility along a sequence of pairwise coprime integers $e_n \le n^\mu$.
  • Tao–Teräväinen (primes, $f = \omega$): (ii) needs quantitative correlation estimates for $\omega$ on short intervals; the engine is the same, with deep input at step (ii).
3. What a new class has to bring

A route on this hole for a new class of sets $A$ is therefore one lemma: for every $k$ (or every $q$), a position $y$ with $2^k \mid \sum_{i=1}^k f_A(y+i)\,2^{k-i}$ and $q\,T_{y+k} < 2^k$. Candidates the literature names but the record does not hold:

  • Erdős's own remark (1968) that coprimality can be dropped: the record's variant-c91009b9 keeps it. Without coprimality the Chinese-remainder conditions "$b \mid y+k$ for the $r$ added elements" need not be simultaneously solvable, and the count of residues excluded by the small elements is no longer a product; a weaker sufficient condition on $A$ (for instance, pairwise coprimality only among elements above every bound) is the natural first target.
  • Sets with $n_{k+1} - n_k \to \infty$, which Erdős (1968, p. 226) expected to give irrational sums: no published proof is known to the writer.

What this engine cannot touch is the open core erdos-257--h2--h1: there $A$ is the greedy set of a rational $r$, which has density about $1/2$ and no arithmetic structure to steer a Chinese-remainder choice, and the tails $T_j$ of such a set are bounded below as above. The affirmative side of the open core needs an idea about the walk, not about digits.

How to read this page

No explainer yet. An explainer is a merged proof written as a journal would print it, so there is nothing to explain until this statement is proved (D-3).

How to read a statement’s page →