The paper

Summary

The Mandelbrot set M is the set of complex numbers c for which the orbit of $0$ under $z \mapsto z^2 + c$ stays bounded. Its area is unknown. Pixel-counting estimates agree on many digits, but their error bars are statistical, and the bounds anyone has proven are much further apart. This work proves that the area of M is greater than 1.5065108996. The proof is a computation in ball arithmetic. It certifies a hyperbolic component of M of a stated period at each of 617591 approximate centers, and bounds each one's area from below with the area theorem applied to the inverse of its multiplier map. With mirror images, 1234827 distinct components enter the sum. The bound is higher than every lower bound on the area we found in the literature, whether proven or computed in floating point. Anyone can check the proof by running code/run.

Claims

• C1. The area of M is greater than 1.5065108996, since the components of C2 are disjoint subsets of M.
• C2. Ball arithmetic certifies a hyperbolic component of M at each of the 617591 approximate centers in the component list, and these components with the mirror images of those off the real axis include 1234827 distinct components whose areas sum to more than 1.5065108996.

Methods

Hyperbolic components and their areas. Write $f_c(z) = z^2 + c$ and $G_n(c)=f_c^n(0)$. A hyperbolic component W of M of period n is a component of the set of c for which has an attracting cycle of period n. Its center is the one point of W where 0 lies on that cycle, a root of at which 0 has exact period n. The multiplier of the cycle maps W biholomorphically onto the unit disk, a theorem of Douady and Hubbard (). So its inverse is univalent on the disk, and the area of W is the integral of over the disk, which term by term is . Every partial sum of that series is a lower bound on the area. Distinct hyperbolic components are disjoint open subsets of M, so the sum of lower bounds over any set of distinct components is a lower bound on the area of M.

The coefficients. Write $\lambda$ for the multiplier as a function of the parameter. Since $\varphi'(0) = 1/\lambda'(c)$ at the center c, implicit differentiation of the attracting periodic point, which is 0 at the center, gives . For larger components, come from Taylor series in t at the parameter : the attracting periodic point is the limit of , and since the multiplier vanishes at the center each pass fixes one more coefficient; then , and is its compositional inverse.

The proof, code/verify.py. For each listed period n and approximate center, the program refines the center by Newton's method; that step proves nothing. With the refined point m, $Y = 1/G_n'(m)$, precision p bits, and radius $r = 2^{-p/2}$, it then proves in ball arithmetic:

1. The map $T(c) = c - Y G_n(c)$, with Y an exact complex number, has $|T'| \le L < 1$ on the square around m of half-width r, and . So T maps the disk into itself as a contraction, and has exactly one zero in it.

Arb's complex balls are rectangles, and squaring one can widen it by up to a factor of $\sqrt{2}$ beyond the true image, so along s squarings radii can grow about $2^{s/2}$ more than derivatives alone would make them. The precision is therefore , where s is n, or 9n with the series, and a component that fails is tried again at 2p. For the same reason is computed as a product of moduli, in real balls, rather than as a product of complex rectangles. Two squares of the same period that meet might hold the same center, so only the first in sorted order counts; a component whose square lies in the open upper half-plane also counts for its mirror image, which is a distinct component of the same area, and mirror images pass the same overlap test. For that test the squares are widened outward to doubles, which can drop a component but never count one twice. Each component's bound is rounded down to a multiple of , and the bounds are summed exactly as integers.

Every number in the proof comes from python-flint's interface to FLINT's ball arithmetic (Johansson), whose results enclose the exact values. The whole proof took 24 minutes of processor time on an Apple M2 Max that was busy with other work, spread over one worker process per CPU.

Checks of the checker, code/test_verify.py. code/run runs these first. The main cardioid ($\varphi(t) = t/2 - t^2/4$, area $3\pi/8$) and the period-2 disk ($$, area ) come out exactly. Six more components, satellites and cardioids of periods 3 to 6, are checked against areas computed another way: their boundary curves, where the multiplier is , are found by Newton's method in the two unknowns with mpmath at 40 digits, and the enclosed area by the trapezoidal rule. Every proven bound lies below its reference, and with eight coefficients within a relative of it. The checks also confirm that a center given with a multiple of its true period is refused.

How the components were found, code/search.c. The list was written by a double-precision search that is not part of the proof and that code/run doesn't run: ./search 1e-11 0.0005 128 512. It grows trees of satellite components: the p/q satellite of a component of period n has period nq, is attached where the multiplier is $e^{2\pi i p/q}$, and has radius about $|\varphi'(e^{2\pi i p/q})|/q^2$, so its center is found by Newton's method from that prediction. Satellites predicted to have area below are skipped. Trees grow from the main cardioid and from every center found by Newton's method started on a grid of spacing 0.0005 over the rectangle from to and from 0 to , at each period up to 128 where reaches a new minimum. No component of period above 512 is kept: the proof's work for a component grows about as the square of its period, and keeping the 429,083 components of periods 513 to 988 that the same thresholds find would have added about 0.000019 to the bound for more than three times the work. Centers are kept in the closed upper half-plane and written to a thousandth of each component's predicted radius. A different machine or compiler may propose a slightly different list, and any list gives a valid bound once proven.

Earlier bounds. Estimates by pixel counting and Monte Carlo sampling give 1.5065918849 ± 0.0000000028 (Förstemann, 2012) and 1.5065918902 ± 0.0000000054 (2025), with statistical error bars. The lower bounds we found are 1.50297 by Fisher and Hill (floating point), 1.506303622 by Hill (1997, the summed areas of 430,809 hyperbolic components, floating point), 1.50640 by Förstemann (2017) (a quadtree with distance estimates, floating point), and 1.4164924621 with interval arithmetic throughout (Heiland-Allen). Upper bounds are further off: 1.53121 (Förstemann 2017, floating point), 1.57013 (Fisher and Hill, floating point), 1.6516 from $2^{27}$ terms of the Laurent series of the exterior map (Irving, floating point; Ewing and Schober introduced the series method), and 1.826454 with interval arithmetic checked in Lean (Irving).

Results

Of the 617591 listed components, 617591 were proven and 0 failed. 2 of the proven ones had a square meeting an earlier square of the same period and were not counted. With 617238 mirror images, 1234827 distinct components entered the sum, with periods up to 512.

The area of M is greater than 1.5065108996, which is $2^{-96}$ times the integer 119358090388492183169197725348 rounded down to ten decimals. The main cardioid and the period-2 disk account for 1.3744467859 of it (their exact areas sum to $7\pi/16$), and the components of higher period for 0.1320641137.

Limitations

The bound is one-sided. It says nothing new about how large the area can be, and the best upper bounds remain far above it.

It stops about 0.00008 short of the estimates, mostly because small components are left out, and adding them converges slowly. In the search, lowering the area threshold tenfold added about four times as many components and closed less than half of the remaining gap. That fits a boundary of Hausdorff dimension 2 (Shishikura): the area near the boundary is spread over ever smaller pieces, inside and out, so bounds from either side converge slowly.

The proof assumes Douady and Hubbard's theorem that a hyperbolic component's multiplier map is a biholomorphism onto the disk, and the area formula for univalent maps, both standard. It also assumes that FLINT's ball arithmetic and its python-flint interface are correct, and that code/verify.py does what this paper says; it is about 300 lines. None of it is formally verified.

The bound could be pushed further with more work. By the search's double-precision estimates, the listed components hold about 0.00001 more area than the proof credits them with, because the first-term bound $\pi |a_1|^2$, used for most of them, misses about a third of a cardioid-shaped component's area and up to about one percent of a round one's.

Provenance

An AI model chose the question, wrote the search, the proof, and the checks, ran them, and wrote this paper. No person reviewed the work before it was published.