The Physical Seam · convergent form, divergent mechanism

Not One Branches the Same

A compound leaf, an antler, the lung, a river network, a bolt of lightning. Five things that branch — and the internet will tell you they are all the same thing. Fractals. Unified by self-similarity. That story is one of the most successful ideas of the last century, and it is mostly a costume.

Here is the claim, stated at full strength so we can take it seriously: nature branches because branching is scale-invariant — the same shape repeats at every magnification, from the whole tree down to the smallest twig, so one number, the fractal dimension, captures them all. It's beautiful. It's on documentaries. And when you make it operable — when you actually grow each of these five things by its own rule, then measure how far the self-similarity really reaches — it comes apart in two specific, demonstrable places. This page walks you to both, and lets you run them.

Grow them, and the family dissolves

Below are the five, each drawn not from a photograph but from the actual rule that generates its kind of branching. Move the controls. Nothing here is a picture of a fractal — each is a different process caught mid-act. Read the mechanism label under each: a rewrite grammar, a budding tip, a work-minimising pipe, an eroding basin, an electrical discharge. These processes share no equations. They share a silhouette.

Five silhouettes, five unrelated machines. The leaf runs on a chemical grammar; the antler on a growth plate under positional signals; the lung on fluid mechanics; the river on sediment and gravity; the bolt on the Laplace equation. If they all deserve the single word fractal, that word has to be earning its keep — it has to be telling us something true across all five. So let's test the two things the word actually claims.

A "fractal" that spans less than a decade

Self-similarity is a claim about range: the shape must repeat across many scales for "scale-invariant" to mean anything. A true mathematical fractal — the Koch curve, the Sierpiński gasket — repeats across all scales, infinitely. A photograph of a fern does not. So the real question is never "is it a fractal?" but "over how much scale does the self-similarity actually hold?"

You can measure it. Box-counting lays a grid over the shape and counts occupied boxes as the grid gets finer; if the shape is self-similar, the count grows as a power of the grid fineness, and the log–log plot is a straight line. The slope is the dimension. The length of the straight part is how far the fractal story reaches. Pick a form and watch where the line goes straight — and where it bends.

Try the Koch curve first: its line stays straight across the whole measured range, because it is self-similar at every scale by construction — that is what an actual fractal looks like. Now try the others. Each has a straight window that is finite and short — a decade or two, and no more — bounded by how many branch levels the real object actually has. The lightning cluster runs out of branches and its curve saturates after about a decade. The antler box-counts as a plain curve, dimension ≈ 1: there is no cascade to find, which is exactly the point of including it. Below the window the boxes are too coarse to resolve any branching; above it you have run out of branches. The "fractal" lives in a thin band in the middle, and for a physical object the band is narrow.

This is not a quirk of these five. In 1998, David Avnir, Ofer Biham, Daniel Lidar and Ofer Malcai went through 96 experimental reports of "fractal" objects published in the Physical Review journals between 1990 and 1996 and measured the scaling range each one actually claimed. The average was about one decade. Essentially none exceeded two. Their Science paper was titled, pointedly, "Is the Geometry of Nature Fractal?" — and the answer they gave was: over the range you can measure, not really. [contested] The paper drew published rebuttals in Science defending Mandelbrot; this is live terrain, not a closed case. But the core measurement — that real fractals are short — has held up, and you just reproduced its shape on five objects at once.

The one clean law — and how little of the lung obeys it

The lung is the best case the "nature is fractal" story has. Its airways really do split and shrink by a fixed ratio, and that ratio really does come from an optimisation, not a coincidence. Here is that optimisation, run live. A pipe of radius r carrying flow Q pays two costs: the power to push viscous fluid through it (which rewards a wide pipe) and the metabolic cost of keeping that volume of tissue and blood alive (which punishes a wide pipe). Minimise the sum:

Split one pipe into two, each carrying half the flow, and the optimisation says each daughter should be 2^(−1/3) ≈ 0.794 times the parent's diameter. Do it again and again and you get the lung's celebrated rule: diameter d(z) = d₀ · 2^(−z/3) at generation z. Ewald Weibel measured the real thing in 1963. Here is his idealised Model A against the law — slide through the 23 generations.

Watch what the ratio does. At the trachea and main bronchi it is nearer 0.68 — the big airways shrink faster than the law. Down in the respiratory zone it climbs toward 1.0 — the smallest airways barely shrink at all, because now the job is gas exchange, not plumbing. The clean cube-root law is a good description of the middle of the tree and nowhere else, and the whole diameter range — trachea to alveolus — is a factor of 44, about 1.6 decades. Twenty-three generations of the most fractal-looking organ you own, and its one exact scaling law holds tightly over less than a single decade. That is Avnir's point, sitting inside your chest.

Where the law does hold, chance already passes it

A defender of "it's all fractals" has a stronger move than the lung, and it is the rivers. River networks obey Horton's laws, discovered in 1945 and made rigorous by Strahler in 1957. Rank every stream by order — the smallest fingertip tributaries are order 1; where two order-1 streams meet you get an order-2 stream; two order-2s make an order-3, and so on. Then two remarkable regularities appear. The number of streams falls off by a constant bifurcation ratio Rb from each order to the next — almost always between 3 and 5. Their mean lengths grow by a constant length ratio RL — almost always between 1.5 and 3.5. Constant ratios across orders: that is self-similarity, and here it spans real range. Surely this is structure.

In 1993 James Kirchner published the sentence that takes it away. Horton's laws, he showed, are satisfied not just by rivers but by virtually every branching network — including trees built by pure chance, that know nothing about erosion, water, or optimality. If a random tree passes the test, then passing the test tells you almost nothing about how the river was made. Here is that claim, made operable. This instrument builds a uniformly random binary tree — every possible branching topology equally likely, Rémy's algorithm, no geomorphology anywhere in it — then Strahler-orders it and measures Horton's ratios, live.

Every tree here is generated by an algorithm that has never heard of a river, and nearly all of them land inside both river bands. Across 300 of them the mean bifurcation ratio sits near 3.8 and the mean length ratio near 2.0 — right where real basins live, and exactly the values Shreve found for the pure random-topology model he introduced in 1966: Rb → 4 (Shreve 1966), RL → 2 in its later development (Shreve 1967/1969). So when a geologist measures Horton's laws in a real catchment and finds Rb = 4.1, that number is not evidence the landscape self-organised into something special. It is evidence the landscape branches — which chance alone would have delivered. The regularity is real. Its explanatory content is close to nil.

Kirchner's result also drew a Comment (Costa-Cabral, 1997), and the debate over how much information Horton's laws carry continues at the margins. But the central finding — that the ratios are nearly insensitive to how the network was built — is broadly accepted. That is why it belongs here beside Avnir: not as a settled verdict handed down, but as a serious, well-supported knife held to a comfortable story, with its own scars shown.

Convergent form from divergent mechanism

Put the two turns together and a cleaner picture stands where "it's all fractals" used to be. Branching is a case of convergent form: many unrelated processes arrive at the same silhouette because the silhouette is what you get, cheaply, when a system has to fill space or move stuff through it. The leaf convergences on it through a developmental grammar; the vasculature and lung through constrained optimisation; the river through erosional dynamics; the lightning through a field equation; and the antler — pointedly — does not converge on it at all, because a growth tip under positional control has no reason to. Same-looking things, different machines. Different-looking things (the antler), same class of question.

"Fractal," in that light, is a description of form that has been quietly promoted to an explanation of mechanism. It answers "what does it look like?" — self-similar, over a decade or so, if you're generous — and then poses as an answer to "why is it shaped that way?", where it says nothing. Why is the lung shaped that way? Because pushing air through pipes has a cost and evolution found the minimum. Not because it is fractal. Why do rivers obey Horton's laws? Because branching networks do, whatever built them. Not because they are fractal. The word is a fingerprint, not a cause — and the fingerprint, measured honestly, is smudged after a single factor of ten.

None of which makes the branches less astonishing. It makes them more so: five different solutions, none copying another, converging on a shape so useful that mathematics needed a whole new kind of dimension just to talk about it — and then a physicist counting 96 papers, and a hydrologist shuffling random trees, quietly reminding everyone that the shape is the easy part.