The cascade that runs backwards

The picture of turbulence everyone carries is a cascade: big eddies break into smaller ones, which break into smaller ones, until viscosity turns the last of them into heat. It is the site’s own account and it is right.

It is also three-dimensional. Remove a dimension and the direction reverses.

The term that disappears #

The vorticity equation carries a term $(\boldsymbol{\omega}\cdot\nabla)\mathbf{u}$ that is the whole of three-dimensional turbulence: stretching a vortex tube intensifies it, which makes vorticity at smaller and smaller scales and is how the cascade proceeds.

In a plane flow it is identically zero. The vorticity points out of the plane and nothing varies along that direction, so the derivative the term asks for does not exist.

That single fact produces everything else in this essay. With no stretching, the vorticity of every fluid parcel is simply carried about unchanged — which means the integral of any function of the vorticity is conserved, and in particular the enstrophy $\int\omega^2\,\mathrm{d}A$ is.

So two-dimensional flow conserves two quadratic quantities in the inviscid limit where three dimensions conserves one, and that is the whole difference.

Two laws, three modes, one answer #

Fjørtoft’s argument from 1953 needs nothing else. Take three wavenumbers and remove a unit of energy from the middle one. Where can it go?

Energy conservation says $\delta E_1 + \delta E_3 = \delta E$. Enstrophy conservation weights each wavenumber by $k^2$ and says $k_1^2\delta E_1 + k_3^2\delta E_3 = k_2^2\delta E$. Two equations, two unknowns, one answer — and for octave spacing it is $\delta E_1 = 0.8$, $\delta E_3 = 0.2$.

There is no model of turbulence in that calculation. No closure, no eddy viscosity, no assumption about how the transfer happens; only two conservation laws and some arithmetic. Whatever mechanism moves energy between scales in a two-dimensional flow, it moves four-fifths of it upwards.

Where the textbook statement stops being a theorem #

The energy half of Fjørtoft’s result holds for every ordered triple. The enstrophy half does not, and the boundary is exact.

Requiring more enstrophy to go to $k_3$ than to $k_1$ and simplifying gives

$$k_2^2 > \frac{2k_1^2k_3^2}{k_1^2+k_3^2},$$

which says that $k_2^2$ must exceed the harmonic mean of $k_1^2$ and $k_3^2$. For octave spacing it does, comfortably. For a lopsided triple such as (0.5, 0.7, 9) it does not, and 50.9 per cent of the enstrophy goes up in scale rather than down: 0.4985 against a k₂² of 0.49, which is how narrow the failure is.

This is the kind of caveat that only appears when a result is computed rather than quoted, and it is worth being clear about its size: it does not overturn the physical picture, because a real cascade proceeds through neighbouring scales rather than lopsided ones. What it overturns is the statement that the enstrophy always goes downscale, which is repeated as though it followed from conservation alone. It follows from conservation and a condition on the triple.

Two ranges, two slopes #

Kraichnan’s picture from 1967 puts a forcing scale in the middle and two inertial ranges either side. Below it the energy flows upwards at a constant rate ε, and dimensional analysis with ε and k alone gives $E(k)\propto \varepsilon^{2/3}k^{-5/3}$ — the same exponent as three dimensions, obtained the same way, describing a flux in the opposite direction.

Above the forcing the quantity flowing is the enstrophy, at a rate η whose dimensions are different, and the same argument gives $E(k)\propto \eta^{2/3}k^{-3}$.

Both exponents here come out of the same dimension-matrix machinery the site uses elsewhere, so they are computed rather than remembered. What dimensional analysis cannot supply is the constants in front, and this site does not have them: they are measurements.

Why three dimensions needs a closure and two does not #

The triad calculation is worth doing again with one law removed, because the comparison explains something about the whole subject.

In three dimensions only the energy is conserved, so the triad gives one equation in two unknowns:

$$\delta E_1 + \delta E_3 = \delta E.$$

That is not enough to determine anything. The direction of the three-dimensional cascade is therefore not a consequence of conservation: it is an extra fact, supplied by the physics of vortex stretching, and every quantitative statement about it needs a model — which is exactly what a closure is, and why the subject has spent seventy years on them.

In two dimensions the second conservation law closes the system, and the answer falls out with no model at all. The peculiarity is that the harder-looking case is the one that is solved, and it is solved because it has more constraints rather than fewer.

Not two conserved quantities, but infinitely many #

There is a stronger statement available and it is worth having, because it explains why the two-dimensional case is so much more constrained than a count of two suggests.

In an inviscid two-dimensional flow the vorticity of every fluid parcel is materially conserved: $\mathrm{D}\omega/\mathrm{D}t = 0$, which is Kelvin’s theorem in its local form with the stretching term absent. A quantity carried unchanged by every parcel makes the integral of any function of it conserved — $\int f(\omega)\,\mathrm{d}A$ for every f — so there are not two invariants but an infinite family of them, of which the enstrophy is simply the one that is quadratic and therefore usable in a spectral argument.

That is why two-dimensional turbulence is less free than three-dimensional turbulence rather than merely different. Almost every rearrangement of the vorticity field is forbidden by one invariant or another, and what is left is the slow organisation into large vortices that the inverse cascade describes.

What a backwards cascade does to a flow #

The physical consequence is that a two-dimensional flow organises itself. Small vortices merge into larger ones, the large ones grow until they meet the size of the domain, and what would be a mess in three dimensions becomes a small number of long-lived coherent structures.

Three places this is visible:

A soap film. A film a few microns thick is two-dimensional for any structure larger than that, and a film stirred by a grid develops large vortices rather than fine-grained turbulence. It is the standard laboratory realisation and the pictures are unmistakable.

The atmosphere at planetary scale. The troposphere is 10 km deep and weather systems are 1,000 km across, so the largest motions are effectively two-dimensional. Cyclones are the merged product of smaller disturbances, and the inverse cascade is why weather has features of that size rather than being smooth.

Jupiter’s bands and its Great Red Spot. A vortex that has persisted for centuries is not something three-dimensional turbulence would allow — it would be torn apart within a few turnover times — and its longevity is the inverse cascade’s most conspicuous advertisement.

How it actually happens, which is not spectral #

Everything above is written in wavenumbers, and a real two-dimensional flow does not look like a spectrum. It looks like a scattering of round vortices moving about in an otherwise quiet field, and watching one is the best way to understand what the triad argument is describing.

The mechanism is merger. Two vortices of the same sign closer than about three core radii do not orbit each other indefinitely — they wrap around one another and combine into a single larger vortex. That is the inverse cascade happening, once, visibly: two structures of one size are replaced by one structure of a larger size, and the energy has moved up the scales without any spectral transfer having to be imagined.

And the same event does the other half of Fjørtoft’s statement at the same time. A merger is not tidy. As the two cores wind together they throw out long thin filaments of vorticity, which the surrounding strain field stretches and thins further — steepening the vorticity gradients, carrying enstrophy to smaller and smaller scales, and eventually delivering it to a scale where viscosity can act. So one event sends energy up and enstrophy down, which is exactly what the two conservation laws demanded, and the filaments are where the enstrophy dissipation the model limit below mentions actually happens.

What emerges from repeated mergers is a vortex gas: a dilute population of long-lived coherent structures, growing steadily larger and steadily fewer, in a background that has been stripped of almost everything else. The number of vortices decays as a power of time — the exponent is near 0.7, measured rather than derived — and the flow’s evolution is a sequence of discrete pairwise events rather than a continuous flux.

That has a consequence for the spectra above. A field of isolated coherent vortices has a spectrum set by the shape of a vortex rather than by any cascade, and the two are not the same. Part of why measured enstrophy-range slopes come out steeper than the $-3$ the dimensional argument gives is that the argument assumes the transfer is local — each scale talking to its neighbours — and a strong vortex strains everything around it at once, which is about as non-local as a mechanism can be.

None of that damages the theorem, and the distinction is worth keeping straight. The conservation laws hold whatever the mechanism, and their consequence for the direction of transfer holds too — that is the whole strength of an argument with no model in it. What the vortex picture supplies is the thing the triad calculation deliberately refused to say: how. And having refused to say it, the argument cannot be surprised by the answer.

What stops it at the top #

An inverse cascade has a problem a forward one does not: there is no viscosity at the large-scale end. A forward cascade terminates at the dissipation scale, where the eddies are small enough for viscosity to convert them to heat. Energy going the other way runs into no such limit — it simply reaches the largest scale available and stays there.

What happens then is called a condensate: energy accumulates in the gravest mode of the domain, which for a box is a single vortex pair filling it, and the flow becomes dominated by a structure the box chose rather than the physics. It is the standard nuisance of two-dimensional simulations and it is not an artefact of the numerics; it is what the equations do.

Real quasi-two-dimensional flows escape it through friction that acts on the largest scales rather than the smallest. In the atmosphere and the ocean that is drag on the ground and the sea floor, which removes energy at a rate proportional to the velocity itself and therefore bites hardest on the biggest, slowest structures. In a soap film it is the air on either side. In a rotating tank experiment it is the Ekman layer on the base — and its spin-down time is the one number that decides how large the vortices get.

That is a genuinely different picture of what “dissipation” means. In three dimensions the energy takes a long route to a small scale and is dissipated there; in two dimensions it goes to a large scale and is dissipated by something outside the cascade entirely.

Where two dimensions comes from #

Nothing in nature is two-dimensional, so the question is what makes a flow behave as though it were, and there are three answers with three different numbers.

Geometry. A soap film is a few microns thick, so any structure larger than that has nowhere to go in the third direction. This is the only case where two-dimensionality is literal, and it is why soap films are the standard laboratory realisation.

Stratification. Moving a parcel vertically in a density-stratified fluid costs energy against buoyancy, so vertical motion is suppressed while horizontal motion is not. The atmosphere and the ocean are two-dimensional in this sense at large scales, and the number that decides how strongly is the one the next essay is about.

Rotation. In a rapidly rotating fluid the Taylor–Proudman theorem forbids variation along the rotation axis, so the flow becomes columnar and behaves two-dimensionally in planes perpendicular to it. This is why Jupiter’s bands are what they are, and it is measured by the Rossby number — the same number a bath vortex fails to reach.

What the two cascades do to predictability #

There is a consequence for forecasting that follows from the direction of the cascade and is worth stating, because it is the reason two-dimensionality matters to anybody outside fluid mechanics.

In three dimensions, errors move up the scales: an error at small scale grows and contaminates larger ones, so a forecast’s useful lifetime is set by how fast the smallest resolved scales corrupt the largest — the argument that gives weather a predictability limit of a couple of weeks.

In two dimensions the energy moves up but the enstrophy — the gradients, the fine structure — moves down, and the large scales are correspondingly better behaved. A two-dimensional flow’s largest structures are long-lived and comparatively predictable, which is why planetary-scale circulation patterns are forecastable for far longer than the weather within them.

The atmosphere is neither, and the crossover between the two behaviours sits somewhere around a few hundred kilometres, where the flow stops being quasi-two-dimensional. Where that crossover lies and what the spectrum does there is an active question, and it decides how far ahead a forecast at a given scale can be believed.

What the model does not contain #

No simulation of anything. This site cannot compute a turbulent field and says so as its seventh invariant; nothing in these figures is a solved two-dimensional turbulent flow. What is computed is the part that is a theorem — the conservation laws, the triad, the centroids, the vanishing term — and the spectra are drawn as scaling arguments with no amplitudes in them.

The dissipation of enstrophy is not addressed. In the inviscid limit both quantities are conserved; with viscosity the enstrophy is dissipated at a rate that does not vanish as the viscosity does, while the energy dissipation does. That asymmetry is what makes the two cascades possible and it is a subtle result this essay states rather than derives.

No forcing, no boundaries, no domain size. The inverse cascade must stop somewhere, and in practice it stops at the size of the domain — where energy piles up in a “condensate” unless something removes it. Real geophysical flows have friction at the bottom that does exactly that. None of that is here.

Nothing is really two-dimensional. A soap film has a thickness; the atmosphere has 10 km of depth and stratification doing the flattening rather than geometry; a laboratory flow is made quasi-two-dimensional by a magnetic field or by rotation. The crossover from three-dimensional to two-dimensional behaviour is its own subject, and the honest statement is that this essay describes an idealisation that everything approaches and nothing is.

The enstrophy range’s slope is not what is measured. Observed spectra in the enstrophy range are usually steeper than −3 and the discrepancy is attributed to the non-locality of the transfer there. The −3 is what dimensional analysis gives, and dimensional analysis assumes locality.

Who found it, and when #

Ragnar Fjørtoft, a Norwegian meteorologist, published the triad argument in 1953 while working on numerical weather prediction — which is a fair indication of where the question came from, since the first computer forecasts modelled the atmosphere as a single two-dimensional layer and had to know what such a layer would do.

Robert Kraichnan gave the two-range picture and the −5/3 and −3 spectra in 1967, and George Batchelor independently in 1969, which is why the enstrophy range is sometimes the Batchelor range. The peculiar thing about the history is its order: the two-dimensional theory is more nearly complete than the three-dimensional one, because two conservation laws pin down what one cannot, and the result is that the case nobody can build is better understood than the case everybody meets.

Where the ladder goes next #

Two dimensions can be enforced by geometry, as in a soap film, and it can be approached by physics — a stratified fluid resists vertical motion because moving a parcel up or down costs energy against buoyancy. That resistance has its own number, its own threshold at exactly one quarter, and a dispersion relation in which the frequency of a wave decides its direction rather than its wavelength.