ERDŐS'S LONGEST CURVE, PROVED AT LAST
Take a polynomial p whose leading term is zⁿ, such as z⁶ − z, and let z roam over the plane of complex numbers. Colour in every point where the value p(z) lies at distance exactly 1 from zero. The result is a curve called a lemniscate, written {|p(z)| = 1}. For z⁶ − 1 it is a flower: six petals meeting at the origin.
In 1958, P. Erdős, F. Herzog and G. Piranian asked how long such a curve can be. Their conjecture: among all polynomials of degree n with leading coefficient 1, the longest lemniscate belongs to zⁿ − 1. Erdős listed it again in 1983 as the first problem of a collection on polynomials.
Sixty-eight years of upper bounds
The length of the model curve has an exact formula, 2^(1/n) B(1/2, 1/(2n)), where B is Euler’s beta function. For n = 1 it gives 2π, the circumference of the unit circle.
The paper’s first table retraces how upper bounds closed in on it — for large n, the model curve measures just over 2n (202.78 for n = 100). Upper bounds on the length went from 74n² (Pommerenke, 1961) to 8eπn, 4πn, 2πn, πn plus a correction, then 2n plus smaller and smaller corrections. Degree 2 was settled by Alexandre Eremenko and W. Hayman in 1999, degree 3 by B. E. Dahlke in 2026. Terence Tao had proved the conjecture for every sufficiently large degree. The degrees in between remained open.
Teng Zhang, of Xi’an Jiaotong University in China, now proves the conjecture in every degree: no lemniscate is longer than that of zⁿ − 1.
Bending a circle into a line
The strategy starts from an observation. zⁿ − 1 has a single critical value (a value taken where the derivative vanishes): −1. A longest curve can always be chosen with its critical values on the unit circle; the task is to show that such a maximiser has only one of them.
To compare lengths, the author bends the target: the curve is made of the points where p lands on the unit circle, and it is this circle that gets deformed. A Möbius transformation, w ↦ w/(1 − βw) — a classical way of distorting the complex plane that turns circles into circles or straight lines — is applied both to the values of the polynomial and to the unit circle. Pushed to its limit, the circle opens into a straight line.
On a line, length becomes a matter of counting. For zⁿ, the points sent onto the real axis form n straight lines through the origin, and almost every line drawn in the plane crosses them exactly n times. Another polynomial sometimes offers fewer crossings. A classical formula of integral geometry, Crofton’s formula — which measures a curve by counting how often randomly drawn lines cross it — turns these missing crossings into missing length. With two distinct critical values, the paper proves that enough length is lost on the line to survive the return to the circle.
Degrees 4 to 255, checked by machine
For degrees of 256 and above, the estimates hold for all degrees at once, by pure analysis. Between 4 and 255, they are sharpened degree by degree, and the remaining inequalities over whole regions of parameters are verified by computer with rational or interval arithmetic, which encloses every quantity between guaranteed bounds. The code is public, written in Python and C#, and most checks are run by two independent programs; one test for degree 5 alone builds a tree of 347,691 internal nodes. Degrees 2 and 3 rely on the earlier proofs.
Only one winner
The last section settles the equality case: a polynomial reaches the maximal length only if it is zⁿ − 1 moved across the plane or turned, (z − z₀)ⁿ − e^(iθ). That is exactly the answer Tao had suggested.
The paper is a 63-page preprint, and much of it rests on certified computation, so other mathematicians still have to go through it. The author reports using ChatGPT for language editing, figures and exploratory discussions, and states that the proofs were developed, checked and written independently.
