Change the ruler, change the answer
2026-08-18 · Scale
How long is the coast of Britain? It looks like a question you could settle by looking it up. It has no answer — not because nobody has measured it, but because the finer you measure, the longer it gets, with no end in sight.
The CIA World Factbook puts the British coastline at 12,429 km. The UK's own Ordnance Survey makes it 17,820 km. That is nearly a factor of two apart, and neither side has made an arithmetic error. They worked from maps of different resolution, and that is the whole story.
You might file this under measurement error. But error has a habit: the finer your instrument, the more tightly the answer closes in on some true value. Coastlines do not close in. Halve the ruler and the measured length jumps again; keep halving and it keeps climbing. What that really says is fairly radical: length is not a property of the coast. It is a property of the coast and the ruler together. Drop the second half and the number means nothing.
The next four topics all circle one thing: how a quantity changes as you change the scale you look at it from. This one is about shape — magnify a small piece and it looks like the whole, a property called scale invariance. Power laws, from Topic 19, are the same property in a distribution: stretch the horizontal axis tenfold and the curve keeps its form. Two outcrops of one property; here it is pinned to something you can see.
In 1967 Benoit Mandelbrot published a short paper in Science with exactly that title. He was not being cute; he had data.
The data came from the meteorologist Lewis Fry Richardson, who spent part of his life doing something distinctly odd: measuring national borders and coastlines with a pair of dividers. The method is charmingly crude. Open the dividers to a fixed span, then walk them end over end along the whole curve, count the steps, and multiply steps by span. Then close the dividers a little and walk it again.
What he found was that the shorter the span, the longer the total — with no hint of settling. Better still, the two quantities fell on a clean straight line on log-log paper. Mandelbrot read that line: if L is the measured length and ε the span, then L is proportional to ε raised to the power (1 − D), where D is some number greater than 1. For a smooth curve — a circle, say — D equals 1, and shortening the ruler walks the answer neatly down onto the circumference. For a coastline D is distinctly above 1, and the answer never lands.
D comes out differently for different coasts, and the ordering matches what your eye tells you from a map: the smooth South African coast around 1.05, the crinkled west coast of Britain around 1.25, the fjord-inside-a-fjord Norwegian coast around 1.52. It is not a metaphor — it is a number you can compute, and section 3 is about how.
So the CIA figure and the Ordnance Survey figure are both correct. Both simply left out a sentence: here is how long my ruler was. On curves like this, the word "length" needs an argument before it means anything.
Whenever you meet a total length, total area, or total count, ask first what resolution it was counted at. When two sources differ by a factor of two, do not open by arguing about who got it wrong — normalise both to the same resolution and compare then. If you cannot, drop the comparison: it is not evidence. More things run on this rule than you would guess: pipeline mileage, national borders, "how many sites we cover", lines of code, total complaints logged.
The last section said a coastline cannot be measured out. It did not say why. The answer is startlingly short: because every small piece of it hides the same structure as the whole.
In 1904 the Swedish mathematician Helge von Koch wrote down a rule. Take a line segment, cut out the middle third, and replace it with an outward spike made of two segments of the same length. One segment has become four, each a third as long. Now do the same thing to every segment you just made.
Each round multiplies the total length by 4/3. Keep going and the length has no ceiling. Yet the curve never leaves a patch you could cover with your hand — infinitely long, enclosing almost nothing. That combination is impossible in smooth geometry. Here it is what one casually written rule hands you by default.
This is the second key word: self-similarity. Magnify any small stretch of the Koch curve and you see the whole curve again. There is no magnification at which you finally reach the basic unit, because the rule contains no basic unit — only "again".
Real coasts are not like this, and the distinction matters. Exact self-similarity is a mathematical construction: magnify a piece and it coincides with the whole, point for point — the Koch curve, the Sierpinski triangle. Statistical self-similarity is nature's version: magnify a piece and it does not coincide with anything, but its statistics do — the same roughness, the same distribution of turns, the same mix of gap sizes. Coasts, cloud edges, ridgelines, river networks, blood vessels: all of them are the second kind.
This is not pedantry. An exactly self-similar figure has a dimension you can derive; a statistically self-similar object has one you can only measure, with error bars, and the answer depends on how wide a range of scales you measured over. Section 5 settles that bill.
One more thing deserves its own line, because it is this site's recurring claim rendered in geometry: the complexity of a fractal does not come from a complicated rule. The Koch rule fits in a sentence, and the shape it makes has no measurable length. That is the same conclusion cellular automata gave us in Topic 4 → ref · Game of Life & cellular automata — a complicated result does not imply a complicated cause. To generate endless detail a rule needs exactly one property: it must be able to act on its own output.
"That coast is more ragged than this one" is a sentence in human language. To make it something you can compute, compare, and have someone else check, you need a number.
Start by rethinking what dimension means. Instead of "how many coordinates", think "how many copies of itself appear when you magnify it":
Double the length of a line segment and it is made of 2 copies of the original. Double the side of a square and it is made of 4. A cube gives 8. Now, 2 = 2¹, 4 = 2², 8 = 2³ — the pattern lives in the exponent. Magnify by k and get N copies, and the dimension D satisfies N = k^D, that is, D = log N ÷ log k.
Apply that to the Koch curve: magnify threefold and it is made of 4 copies of itself. So D = log 4 ÷ log 3 ≈ 1.26. Not a whole number. That is where the word comes from — Mandelbrot coined "fractal" in 1975 from the Latin fractus, broken or fractured. A value between 1 and 2 means something plain enough: the curve fills the plane more thoroughly than a line does, and less thoroughly than a surface.
That derivation only works when you already know "magnify by how much gives how many copies". For a photograph of a real coast, or an image of retinal blood vessels, you need something clumsier and more general: box counting. Lay a grid of side ε over the figure and count how many boxes it touches, call that N(ε). Halve ε and count again. If N grows as N ∝ ε^(−D), that D is the fractal dimension. On a log-log plot it is simply the slope of the line. → ref · Box-counting dimension
Here is something easy to miss: measure the same curve with Richardson's dividers and with box counting, and the two numbers need not agree. The simulated coast on this page is a case in point — the divider slope comes out slightly above the box-counting one. "Dimension" is not one definition but a family of them (box-counting, Hausdorff, information dimension, and more). They coincide on tidy mathematical objects and need not coincide on real data. So a reported dimension has to say which one, measured over which range of scales.
Before quoting a dimension, a Hurst exponent, or any scaling exponent, look at how many orders of magnitude the "straight line" on your log-log plot actually spans. Under one order of magnitude, do not quote the number — over a window that narrow, almost any gently curving line can pass for straight. When you do quote it, quote three things together: which method, the upper and lower limits of the fitted range, and how many points lie inside it. Missing any one of those, nobody can check the number, and it does not belong in a conclusion.
Lungs, blood vessels, river networks, lightning, tree crowns, cloud edges, ridgelines, fracture surfaces. Fractals are not an occasional novelty in nature; they turn up far too often for coincidence. The reason is that several unrelated routes all arrive at the same family of shapes. At least three.
Route one: branching supply networks. A body has to deliver oxygen to every cell, which takes plumbing that reaches the whole volume. The plumbing itself occupies volume — too much and there is no room for anything else — while the exchange happens across surface area, and too little of that means nothing gets delivered. The two requirements pull against each other, and a branching fractal is precisely the solution.
Work it out. Suppose each level doubles the number of tubes while length and radius each shrink to 2^(−1/3) ≈ 0.79 of their previous value:
· Volume contributed per level = count × length × radius² ∝ 2 × 0.79 × 0.79² = 1, unchanged.
· Surface area per level = count × length × radius ∝ 2 × 0.79 × 0.79 ≈ 1.26, up 26% every level.
Stack up enough levels and the surface area becomes enormous while the volume stays under control. Human airways run about 23 levels from trachea to alveoli, and 1.26²³ ≈ 200. (That is the skeleton calculation only: real airways change their rules around level 17, where alveoli start appearing, so do not take the figure as a measurement — it gives the order of magnitude the rule implies.) Measured alveolar surface area runs to a hundred-odd square metres, packed into a few litres of chest. While we are here, a much-repeated line deserves correcting: the lung's surface is not "the size of a tennis court" — a singles court is 195 m².
Route two: criticality. Topics 16, 17 and 18 kept returning to one sentence — at a critical point a system has no characteristic scale. The immediate geometric consequence of that sentence is a fractal. The giant cluster that has just spanned a percolation lattice → ref · Percolation has a fractal dimension known exactly in two dimensions: 91/48 ≈ 1.90. It fills the plane in a way that sits between a line and a surface, riddled with holes that have holes in them. Its counterpart in the world of distributions is the power law → ref · Identifying (and misidentifying) power laws. So any system that climbs to criticality on its own — the Topic 18 family — grows fractal geometry as a matter of course. The two facts are not neighbours; they are one fact seen from two sides.
Route three: the shaping process itself has no preferred scale. Erosion, fracture, deposition, water splitting around obstacles — these follow the same rules at every size. What water does in a small gully is what water does in a large valley, only smaller. If the process that makes the shape does not recognise scale, the shape it makes will not have a characteristic one.
The three routes work by entirely different mechanisms: optimisation, criticality, and scale-free process. Yet they produce the same family of shapes. That explains why fractals are so common — and, in the same breath, why "this thing is a fractal" carries almost no information. It does not tell you which route was taken, and the three routes predict quite different things.
Fractal is the single most abused word in complexity science. It is pretty, it is easy to grasp, and you can "see" it everywhere — which is exactly where the trouble starts.
First, fractals in nature are short. Mathematical fractals span infinitely many scales; natural ones do not. The self-similarity of blood vessels stops downward at the capillary (the inner cutoff) and upward at the whole circulatory system (the outer cutoff). The real question is how narrow that window usually is. In 1998 Avnir and colleagues did something rather unkind in Science: they went through the papers in the Physical Review journals reporting fractals in nature and tabulated, one by one, the range of scales actually used. The average was about 1.3 orders of magnitude.
To put 1.3 orders of magnitude in perspective: that is from 1 to a bit past 20 on the horizontal axis. Over a stretch that short, a straight line and a gently bending curve are indistinguishable. So a great many "X is fractal" conclusions amount to a line fitted over a small window and then carried out to places it has never been.
Second, dimension is not a fingerprint. Two figures with the same dimension can have nothing else in common — an unremarkable fact that got overlooked in a famous authentication case. In 1999 Richard Taylor and colleagues reported in Nature that Jackson Pollock's drip paintings have a stable fractal dimension, usable for telling real from fake, and the method duly entered authentication practice. In 2006 Katherine Jones-Smith and Harsh Mathur replied in the same journal: they knocked out a few sketches made of stars and scribbles, and the doodles passed the same fractal test. The measure was too coarse to separate Pollock from an idle hand. Fractal analysis of a disputed group of Pollock attributions ran into controversy over the same period.
Third, and most important: fractal is a description, not an explanation. Saying "trees are fractal" explains nothing; it restates "the branching of a tree looks similar at several scales" in fewer syllables. The explanations are the three routes in section 4 — supply-network optimisation, criticality, or a scale-free shaping process — and they make quite different testable predictions. The first predicts a radius ratio near 2^(−1/3); the second predicts power-law fluctuations and critical slowing down in the same system; the third predicts nothing at all and merely says the shape will come out that way. If a sentence loses its force once you swap in "this thing looks about the same at several scales", it never had any.
Treat "this is fractal" as a hypothesis awaiting a test, never as a conclusion. Before it goes into any finding, supply three things: (1) the upper and lower bounds of the scaling window, in orders of magnitude; (2) which dimension and which fit; (3) one prediction that only holds if it really is fractal — for instance "double the resolution and I will count 2^D times as much" — and then go and count. If you cannot assemble all three, delete the word from the sentence and see whether what remains still stands up.
Yes, but the meaning shrinks to one very specific sentence: "within this one order of magnitude I can describe it with a single set of parameters." That is still useful — it means you can interpolate across that range. It supports no inference outside the window, least of all "so the same thing happens at larger scales". The way to check is to ask: if I widened the window by another half an order of magnitude, would the conclusion flip? If it would, the conclusion was a product of the window.
The rate at which detail accumulates as resolution improves, and nothing else. That rate is a summary statistic of the shape, in the way that mean height is a summary statistic of a crowd — knowing the mean height tells you nothing about what the crowd looks like. So dimension is suited to comparison (which of these objects, measured the same way over the same window, is rougher) and not to identification. The Pollock case stepped over exactly that line: a comparative quantity pressed into service as an identifier.
Part of it is how things get made. A fractal structure has to be grown recursively, and human manufacturing has long meant machining parts and then assembling them — assembly does not recurse. Additive manufacturing loosens that constraint, which is why fractal-ish structures (lattices, porous media, branching heatsinks) have only recently started appearing in engineering. The other part is maintainability: a fractal structure has no natural seams, so when it fails you cannot swap a part out. Evolution does not care about replacing parts; engineering does.
They are one property showing up in two kinds of object: a power law says a distribution keeps its shape when you stretch the horizontal axis, a fractal says a shape keeps its appearance when you magnify it. They frequently occur together — critical systems give you both — but not necessarily. A shape can be fractal while some of its statistics are not power-law distributed, and the reverse holds too. Finding one does not license assuming the other; that needs checking separately.
Almost always some physical smallest unit or some boundary. Downward it is the size of the stuff the thing is made of — a cell, a grain of sand, a pixel. Upward it is the size of the whole system — an organ, a catchment, the total size of a market. That makes finding the cutoffs a practical exercise: ask what the smallest indivisible thing in this system is, and what the largest container is. Quote those two numbers and the width of the scaling window is settled, and with it how much room this language has left here.