Meander geometry is remarkably consistent across scales. The bends in a rivulet running down a windowpane, a creek behind a suburb, and the lower Mississippi all have the same shape, up to a scale factor, and the scale factor is the width of the channel. Once you know that, you cannot look at a river the same way. It isn't wandering. It's working.

Eleven widths

Luna Leopold and Gordon Wolman measured this in 1960 across streams from a few feet wide to the Mississippi, and the result is close to a law. The wavelength of a meander, the distance from one bend to the next bend going the same way, is about ten to eleven times the channel width. The radius of curvature of a bend is about two to three widths. These hold across four orders of magnitude of size. They hold for meltwater streams on top of glaciers, where there is no sediment at all. They hold for the Gulf Stream, which has no banks.

So whatever produces meanders lives in the flow itself, not in the mud.

Why a straight river cannot stay straight

Start with a perfectly straight channel. Any tiny irregularity, a rock, a soft patch of bank, deflects the current slightly. On the outside of that deflection, water moves faster and hits the bank harder, eroding it. On the inside, water slows and drops its sediment, building a bar. The bend gets bigger. The bigger bend deflects the flow harder. Positive feedback. A straight river is an unstable equilibrium, and no real river is straight for longer than about ten widths, which is, not coincidentally, the meander wavelength.

The mechanism inside the bend is a helix. Water on the surface is moving faster than water on the bottom, so when the flow turns, the fast surface water is flung outward by its momentum while the slow bottom water is pushed inward by the pressure gradient. The result is a corkscrew: outward on top, inward on the bottom, scouring the outer bank and depositing on the inner. This is why the outside of a river bend is deep and the inside is a beach.

Einstein wrote a short paper on this in 1926, explaining why tea leaves collect in the center of a stirred cup by the same mechanism, and then applying it to rivers. His son Hans Albert became a hydraulic engineer and spent his career on sediment transport. There's something in that.

Least work, not least distance

Here is the part that changed how I think about it.

Meanders are inefficient if you think the river's goal is to get to the sea. A meandering river travels much farther than a straight one would. The Mississippi below Cairo, Illinois, is roughly twice the straight-line distance.

But the river doesn't minimize distance. Leopold and Langbein argued in 1966 that it minimizes something else: the variance of the work it does along its length. Every stretch of river has to dissipate the potential energy of falling water. A straight, steep channel dissipates it violently, with rapids and scour, concentrated in a few places. A meandering channel spreads the drop over a longer path, lowering the gradient everywhere and dissipating the energy evenly. The curve that does this most evenly is a specific one, the sine-generated curve, where the direction of the channel varies sinusoidally with distance along it. It is also the curve a thin steel spring takes if you hold both ends and push them together. Minimum total bending.

So a meander is what you get when a system with a fixed amount of energy to dissipate searches for the shape that dissipates it most uniformly. That's an optimization, and the river runs it continuously, with the bank as the gradient.

The cutoff

The optimization doesn't converge. Bends keep growing, the neck between two adjacent bends gets thinner, and during a flood the river jumps the neck and abandons the loop. The loop becomes an oxbow lake. The channel is suddenly shorter and steeper, the gradient is higher, erosion accelerates, and new bends start forming to bring the gradient back down.

This is a limit cycle, not a steady state. The river is always approaching the optimum and always overshooting it. Mark Twain, who piloted the Mississippi, noticed that cutoffs had shortened the lower river by well over a hundred miles in his lifetime and extrapolated, with a straight face, that in a few hundred years it would be a mile and a half long. The joke is a real observation about a real rate.

Gradient descent with memory

If you want the engineering analogy, a river is gradient descent. It moves downhill locally, at every point, with no global plan. It has momentum, in the literal sense, which is why it overshoots bends instead of turning cleanly. And it has memory: the channel it carved yesterday constrains where it can flow today, so the current solution depends on the whole path of previous solutions.

Which means it also has the failure modes. It gets stuck in local minima. A river will sit in a valley it cut a million years ago long after the tectonics that made that valley the best route have changed. It will cut through a mountain range rather than go around, because it was there first and the mountains rose slowly under it. The Colorado didn't find the Grand Canyon. The Grand Canyon is a river refusing to update.

Arguing with the Mississippi

The Mississippi, currently, would like to leave. Its main channel reaches the Gulf through a long, low, sediment-clogged delta near New Orleans. About 300 miles upstream there's a distributary called the Atchafalaya that reaches the Gulf in half the distance at twice the slope. By every principle in this post, the river should switch. It has been trying since the 1950s.

The Army Corps of Engineers built the Old River Control Structure in 1963 to stop it, because the switch would leave New Orleans and Baton Rouge on a stagnant backwater and strand a large fraction of American petrochemical infrastructure. The structure lets about 30% of the flow go down the Atchafalaya and forces the rest to stay. In the 1973 flood it nearly failed. A scour hole opened under it that engineers found by dropping a weighted line and getting no bottom.

They reinforced it. It has held. It's a concrete argument with a system that solves for least resistance and has an unlimited number of floods to make its case. Every hydrologist I have read on the subject uses some version of the same phrase: not if, when.

What it means

A river is the clearest example I know of an optimizer you can walk beside. It has no model of the sea. It has no plan. It has a local rule, go down, and a shape that emerges from applying that rule to itself over time, and the shape is the same on a windowpane and a continent because the rule is the same.

We think of rivers as landscape. They're closer to algorithms. The landscape is the output.