Erdős Problem a Daythe open tracker

Erdős Problem #522

OPEN LEANdatabase updated 2026-09-25

No proof or disproof is known, and it isn't known whether a finite computation could settle it. A solution has been formally verified in Lean, but no human reader has digested it yet — which is why the database doesn't call it solved.

What the community database records, not our verdict. erdosproblems.com no longer shows statuses (why).

A sequence of *Kac coefficients* over a subset `S` of a field `k` is a countably infinite sequence of independent random variables, each uniformly distributed over `S` with respect to the reference measure `μ`. The default reference measure is the counting measure, so that for a finite set `S` each coefficient takes every value of `S` with probability `1 / |S|`. Such a sequence determines a *Kac polynomial* of degree `n` for each `n`, which is the random polynomial given by `KacCoefficients.poly

We can always view a Kac polynomial as a random variable on `ℕ`. -/ instance : FunLike (KacCoefficients S Ω μ) ℕ (Ω → k) where coe P := P.toFun coe_injective P Q h := by aesop namespace KacCoefficients open scoped Polynomial variable {S Ω} {μ : Measure k} /-- The random polynomial associated to a sequence `c : KacCoefficients S Ω μ` of Kac coefficients given by `∑ i ∈ Finset.range (n + 1), c i z^i`.

The random multiset of roots associated to a Kac polynomial

Let \(f(z)=\sum_{0\leq k\leq n} \epsilon_k z^k\) be a random polynomial, where \(\epsilon_k\in \{-1,1\}\) independently uniformly at random for \(0\leq k\leq n\). Is it true that, if \(R_n\) is the number of roots of \(f(z)\) in \(\{ z\in \mathbb{C} : \lvert z\rvert \leq 1\}\), then \[ \frac{R_n}{n/2}\to 1 \] almost surely? There is some ambiguity as to whether the intended coefficient set is \(\{-1, 1\}\) or \(\{0, 1\}\), see `erdos_522.variants.zero_one` for the alternate version. This is true. Proofs were posted on the erdosproblems.com forum in April 2026 [Ch26], [KZ26], and Lean proofs were given independently in September 2026 [Ka26], [Ki26]. More generally, [Ka26] proves an almost-sure radial law: the number of roots in \(\{\lvert z\rvert \le 1 + x/n\}\) is \(\Phi(x)\,n + o(n)\) for every \(x\), where \(\Phi(x) = \tfrac12\left(1 + \coth x - \tfrac1x\right)\). This problem is the case \(x = 0\).
Statement as formalized in Formal Conjectures (Google DeepMind, Apache-2.0) · their label: research solved · 3 further variants / known partial results in the Lean file. The authoritative wording is on erdosproblems.com/522.
Prizenone
Topicsanalysis, polynomials, probability
Statement in LeanFormal Conjectures
Solution in Leanyes
Original pageerdosproblems.com/522
WORKING REPORT BY THIS SITE · JUL 31, 2026

Read our working report →

Live page status is OPEN (0 formal claimed proofs; “Currently working” marker: None), but its April 2026 comments explicitly claim possible/complete solutions and state that Brayden Letwin is actively

SKIPPED