⟵ ARTIFICIAL WASTELAND / STRATUM 030 · PATTERN

How Big Is the Mandelbrot Set?

AREA = 1.506 591 8849…? · BEST ESTIMATE, NOT A PROOF · EXACT VALUE UNKNOWN

It is the most recognized object in modern mathematics — printed on posters, tattooed on arms, set as a billion screensavers. We can compute it to a trillion points. And we do not know its area.

Take a complex number c. Start at zero and repeat z → z² + c. For some c the numbers stay bounded forever; for others they fly off to infinity. The Mandelbrot set is just the collection of c that stay bounded — a single, connected black island in the plane, ringed by a coastline of infinite, self-similar intricacy.

Everything about it has been studied to exhaustion. Its boundary is the most-rendered curve in history. And yet one of the simplest questions you could ask — what is its area? — has no known answer. Not "we haven't computed enough digits." There is no proven exact value, and no formula. The best number anyone has, from counting pixels on enormous grids, is

A ≈ 1.506 591 8849 ± 0.000 000 0028

— Thorsten Förstemann, 2012, from roughly 88 trillion sample points. But pixel-counting is an estimate, not a theorem. Meanwhile the one method that gives a rigorous answer — a 1914 theorem about areas and a power series — has been pushed to five million terms and can only prove the area is below 1.6829. That bound and that estimate have never met.

This page is three instruments. The first counts the area in your browser and shows it converging. The second shows you the two pieces of the set whose areas we do know, exactly. The third runs the rigorous method and shows you, live, why it stalls. None of the numbers are mine to assert — they recompute in front of you or carry a citation.

I · Count it yourself

The honest way to measure an area is to lay a grid over it and count the cells that fall inside. For each cell we run z → z²+c from its centre, up to an iteration cap (hundreds of steps at first, tens of thousands later): if z never escapes, the cell is (probably) inside. Press refine to make the grid 1.4 times finer in each direction and double the iteration cap, then re-run. Two interior regions are short-circuited analytically (the next instrument explains them); everything else is counted by brute iteration, folded across the real axis by symmetry.

INSTRUMENT 01 · the area census, computed livegrid —

inside (orbit bounded) undecided boundary band escaped (outside)

Press refine the grid. The estimate counts every cell whose centre has not escaped within the iteration cap. Points just outside the set can take thousands of steps to escape, so a low cap counts them as inside: that is most of the over-count, and it shrinks because each refine doubles the cap. Testing only the centre can also misjudge a boundary cell either way, so the count wobbles by a few ten-thousandths. Watch the blue boundary band: it thins as the grid tightens, but more slowly than it would for a smooth curve.

As you refine, the estimate falls (at the default start: 1.5118 → 1.5092 → 1.5081 → 1.5070 → 1.5072 → 1.5067), homing in on 1.5066. Most of the excess is not the grid at all: it is slow-escaping points near the edge that the iteration cap has not yet caught. Refine the grid alone at a fixed cap and the count barely moves; raise the cap and it drops. The blue band of cells on the edge thins too, but more slowly than a smooth shape's would: halving the grid spacing leaves about 60% of the band's area here, not 50%. That blue band, and the slow escapes inside it, are the whole problem, and Section IV says why.

II · The two pieces we know exactly

Not everything is unknown. The set's two largest features have areas in perfect closed form. The big heart-shaped body is the main cardioid — the parameters where the orbit settles to a single point. Its boundary is the curve c = e^iθ/2 − e^2iθ/4, and a one-line integral gives its area as exactly 3π/8. The disk bolted to its left, centered at −1 with radius ¼ — the period-2 region — has area exactly π/16.

Together that is 7π/16 ≈ 1.3744 — already 91% of the whole set. Tap a row to light it up.

INSTRUMENT 02 · what is known in closed formof est. 1.50659

The main cardioid (3π/8) and the period-2 disk (π/16) are exact. Every smaller bulb has an area too, but for period ≥ 3 there is no known closed form: they are measured numerically, one by one. Jay Hill measured 430,809 of them (the components of period up to 16, plus their attached buds followed out to period 256) and summed them to a numerical lower bound of 1.5063. The last sliver, and whether the lacy boundary contributes any area at all, is open.

III · The bound that won't come down

Here is the rigorous attack. Because the set is connected (Douady & Hubbard, 1982), the region outside it can be mapped conformally from the outside of a unit disk by a single function

Ψ(w) = w + b₀ + b₁/w + b₂/w² + b₃/w³ + ⋯

and a 1914 theorem of Gronwall — the area theorem — turns the coefficients of that map directly into the area of what it encloses:

Area(M) = π · ( 1 − Σ_m≥1 m · b_m² )

Every term you add is subtracted, so each partial sum is a guaranteed upper bound that only ever decreases. The coefficients b_m are exact dyadic fractions (b₁=1/8, b₃=15/128, b₅=−47/1024…). This page ships the first 4,096 of them, recomputed from scratch, and sums the bound live in your browser. Drag the slider: the ceiling comes down fast at first, then refuses to come down fast enough, and every point your own browser computes lands exactly on the published record (the amber trail sits right on the dim published dots).

INSTRUMENT 03 · the Gronwall area bound, summed liven = 1

terms n 1

bound π(1−Σ m b_m²), computed here published, to 5,000,000 terms best estimate 1.50659

At n = 1 the bound is just under π. Drag right and it falls fast at first, then crawls. The amber points are summed in your browser from 4,096 coefficients computed here; they land right on the dim published dots (Scott Huddleston's 1990 table, to 65,536 terms). The published record runs on past them: 240,000 terms give 1.7274 (Ewing & Schober), and after five million the bound is still 1.6829. The green line is where the truth is believed to be. The rigorous ceiling has never touched it.

Five million terms of a convergent series, and the proven answer is still 0.18 too big. The series is honest. It is just unspeakably slow.

IV · Why it's this hard

The two methods fail in mirror-image ways, and both failures trace to a single theorem. In 1991 Mitsuhiro Shishikura proved that the boundary of the Mandelbrot set has Hausdorff dimension 2 — the maximum possible for a curve in the plane. The coastline is so violently crinkled that, by the measure of dimension, it is as big as an area.

That is exactly why the pixel count has a stubborn error bar. For a smooth boundary, the cells straddling the edge are a one-dimensional sliver: refine the grid and their total area falls like the grid spacing. For a dimension-2 boundary, the straddling cells behave almost two-dimensionally: that blue band in Instrument 01 thins more slowly than a smooth edge's would, and every cell in it is an unresolved ±. And it is why the series crawls: the coefficients b_m fall off only about as fast as 1/m (no faster), so the corrections m·b_m² are a near-harmonic dribble that takes millions of terms to add up. That slowness is not a coincidence: if the b_m decayed quickly the boundary would have to be small, and it isn't. The fat boundary sabotages both attacks at once.

There is a second, quieter difficulty, and it is about computing those coefficients in the first place. The map Ψ has a twin, the forward map Φ that carries the outside of the set to the outside of a disk, and Φ's own Laurent coefficients grow like 2^m (the set reaches out to c = −2, which pins the radius of that series at one half). The b_m we actually want stay small, but the obvious route to them runs straight through those exploding forward coefficients, so pulling a number near 1 back out of numbers near 2^m costs about 0.3 m digits of accuracy. That is why an ordinary double-precision attempt collapses within fifty terms, and why the record computations are careful numerical analysis rather than a brute loop. The coefficients summed above were made with a recursion that never forms the exploding numbers at all: written in the disk variable the whole thing becomes a fixed point whose every intermediate stays near 1, solved by Newton's method to 262,144 terms in about a minute, and every partial bound it produces reproduces the published checkpoints to the digit. The recomputation and its checks are open in /checks/research/mandelbrot-area/series_newton.py (it needs Python 3 with numpy: python3 series_newton.py 65536 exits 0 only if every check passes).

V · The deeper unknown

Dimension 2 is not the same as positive area. A dimension-2 set can still have zero area — and so there is a stranger question hiding under the first one: does the boundary of the Mandelbrot set itself have positive area? Nobody knows. It is consistent with everything proven that the black coastline, taken alone, has measurable thickness — Buff and Chéritat showed in 2012 that the closely-related Julia sets can have positive area, so it is not a wild idea. If the boundary does have positive area, then the very question "what is the area of the set" sharpens into "and how much of that lives in the infinitely-thin-looking edge."

So the honest state of the art is a nested pair of open problems wrapped around a shape on a million T-shirts: we cannot prove its area, and we cannot even decide whether its edge has any area to give. The estimate 1.5065918849 is almost certainly close. Almost is doing real work in that sentence — and closing it is genuine, unfinished mathematics.

There's a coda worth knowing: in 1992 Keith Briggs idly wondered whether the area might be exactly 3/2. It's a pretty guess, and the pixel estimate, 1.5066, sits about two million error-bars away from 1.5000. If you trust the counting, the area is not 3/2. But counting isn't proof, so even that tidy little conjecture isn't formally dead. That is how thin the ice is here.