A mathematician at Oxford read the preprint and said it was not written for humans. That is a fair description. Its Section 2 tells the physical story and Section 3 lists the steps, but the layers a human author would normally supply are thin. Why this ansatz and not another? What does each mechanism look like in the simplest possible case? Which steps carry the new idea and which are bookkeeping? How do its many results depend on one another? This page supplies those layers, as questions a reader would ask.
The toy models below are classical and are not the paper's construction; each is checked by the script at the end. Where we reconstruct a reason the paper does not state, we say so.
Viscosity smooths. How can a viscous flow reach infinite speed?
Viscosity wins over a length ℓ in time ℓ²/ν. Run that backwards: a structure that shrinks like ℓ ∼ √(ν(1−t)) shrinks exactly as fast as diffusion can follow. Speeds of order ℓ/τ ∼ τ−1/2 are then "critical": the Reynolds number ℓ·speed/ν stays of order one, and viscosity keeps up. To beat viscosity, something has to move faster than that.
The paper's core does exactly that, but only in one direction. Its radial inflow stays at the critical scale, so viscosity and inflow remain in balance and the profile stays smooth and self-similar. Its swirl, the speed around the axis, runs faster by a factor τ−h, which is the swirl Reynolds number Reθ → ∞. The fluid makes more and more turns per diffusion time. Rotation is the part viscosity cannot keep up with, and rotation is what carries the blowup. COMPUTED
And why does the energy stay bounded?
Because the fast region is tiny. Its volume shrinks like τ3/2−h while the squared speed grows like τ−1−2h, so the energy in the core goes to zero, like τ1/2−3h. Infinite speed in a vanishing volume costs almost nothing.
A core that shrank equally in all directions would leave no small parameter to expand in. Making the height shrink more slowly than the radius (τ1/2−h against τ1/2) does three things at once. The column gets slender, ℓr/ℓz ∼ τh. Diffusion along the axis becomes weaker than diffusion across it, by a factor τ2h. And that factor, written q2h, becomes the expansion parameter: Section 5 corrects the profile order by order in powers of q2h. The paper says this in one line (§3.1); it is the reason h exists.
Why h < 1/100 rather than, say, 1/10? The energy argument only needs h < 1/6. We did not find where the paper explains the smaller bound; we read it as a technical margin used in the estimates, not as something essential. OPEN
Why the deliberate up–down asymmetry?
The pulses need shear to grow. With exact mirror symmetry about the mid-plane, the axial flow would vanish there and so would the rotational amplification, leaving a dead layer. Tilting the axial profile slightly upward separates the layer where one mechanism is weak from the layer where the other is weak, so some amplification is always present (§2.1).
A wave w = a cos(k·x), with its velocity a perpendicular to k so that it is divergence-free, averages to zero. But the force it exerts on the mean flow comes from the product w ⊗ w, and that does not average to zero:
⟨w ⊗ w⟩ = ½ a aᵀ.
This averaged product is the wave's momentum flux, the Reynolds stress of turbulence theory. Its divergence acts on the mean flow exactly like a force. That is the whole trick of the paper: the background vortex needs a force it cannot supply, so the paper places waves where that force is needed and lets their momentum flux supply it. "The fluid's own motion supplies the missing momentum transport" (§2.2). COMPUTED
Oscillations that realise a prescribed stress are also the engine of convex integration (De Lellis–Székelyhidi, Buckmaster–Vicol). The difference here is that the paper's waves are not imposed at every instant: each is seeded by an exponentially small push and then grows on its own, by drawing energy from the shear.
The force needed in the annulus has two components: azimuthal momentum and axial momentum, both transported radially. A single wave family carries flux in one fixed direction v₁, set by its orientation; turning up its amplitude only moves along that ray, and the amplitude enters squared, so it can never point backwards. Two families with directions v₁ and v₂ can reach any target T = c₁v₁ + c₂v₂ with c₁, c₂ ≥ 0: exactly the cone between them, and nothing outside it.
So the background must be built so that its required stress lies inside the cone everywhere in the annulus. That is the "admissible stress cone" condition (Lemma 4.5). It does not hold by luck. Appendix C forces it by adding a small radial ripple to the profile, with phase N log X, which changes the shear by an order-one amount while changing everything else by only O(1/N). This is one of the genuinely delicate steps.
The simplest exact example is Kelvin's shearing wave (1887). On a plane shear flow u = (Sy, 0), a single Fourier mode of vorticity is carried by the shear: its wavevector tilts, ky(t) = ky0 − S kx t, and viscosity damps it by exp(−ν∫|k|²). Remarkably, this is an exact solution of the full Navier–Stokes equations, nonlinear term included, because a single plane wave does not advect itself. Its energy is proportional to (amplitude)²/|k|²:
The paper's pulses do the same thing, with two differences. Their background is a rotating, axially sheared vortex, and part of the growth is centrifugal: a parcel moving outward with extra swirl keeps accelerating outward if the angular velocity falls off fast enough with radius, a version of Rayleigh's criterion (the paper cites the centrifugal-instability criteria of Leibovich–Stewartson and Billant–Gallaire). And the paper chooses each pulse's initial wavelength so that growth wins first and viscosity wins later, exactly as in the toy, so each pulse is a finite burst whose tails are exponentially small (Lemma 7.4; Figure 3(b)). CITED
Because the leading-order cancellation is only the start. Every wave that cancels one error creates others: interactions between waves, corrections from the cutoffs, terms from the varying amplitude, the effect of the waves on the pressure. The force has to be smooth to every order, so every one of these must be pushed down as well, with control of every derivative, uniformly. The paper does this with a fixed cycle of four corrections that improves the decay of the remainder by a fixed step each time (σj+1 = σj + 1/10), then sums infinitely many cycles with shrinking cutoffs. Most of the length is this bookkeeping.
| Where | What kind of step | In our reading |
|---|---|---|
| §2, §3.3, Prop. 7.5 | The idea | Shear-amplified pulses whose averaged flux supplies the missing force on a collapsing vortex. The paper names this as its new ingredient (§1.1). |
| §2.1, §3.1, Thm 4.6 | Design choices | Slender core with small h; deliberate asymmetry; the heat exterior that needs no force and allows cutting off. |
| Lemma 4.5, App. C | Delicate analysis | Making the required stress lie inside the cone everywhere. |
| §5, §9 | Delicate analysis | Corrections to every order with derivative losses that do not grow with the stage. |
| Lemmas 10.2, 10.3 | Delicate analysis | Showing every derivative of the force has a limit at t = 1, so it extends smoothly. |
| §6, §8, symbol table | Bookkeeping | Dyadic charts, the auxiliary torus that keeps pulses apart, coefficient classes, moment equations. |
The classification in the right-hand column is our judgment as readers, not the authors'.
The paper's Figures 5 and 6 are a flow chart; here is the same information as a tree, read from the top down. Each line names a result in the paper. CITED
This is our reconstruction of a plausible path, written to make the logic memorable. It is not the authors' account, and we do not know how the construction was actually found.
These are the questions a human referee or seminar audience would likely ask. We could not answer them from the text. OPEN
The script checks every toy on this page: the wave flux, the cone, that Kelvin's wave is an exact Navier–Stokes solution (symbolically, nonlinear term included), its growth and decay at the default settings, the speed comparison, and, with the PDF, that every result named in the dependency tree exists in the paper.
python3 human-layer-verify.py --pdf navier-stokes.pdf