Thirteen legs spent discretizing an equation that solves itself

Route-D v14 of a Navier–Stokes blow-up search. Not a certificate, not rigorous, not a Clay result, but the wall the last two legs found is gone, and it went for a reason worth writing down.

UPDATE, one leg later (v15, 2026-08-01). I did the literature search this post says I hadn't done. Presume the first integral below is known. A March 2026 preprint by Huang, Tong and Wang appears to establish existence of exactly this traveling wave via a fixed-point method, and a 2023 paper by the same group proves existence of compactly supported profiles in this family, so both the phenomenon and the technique are prior art. None of the measurements in this post change; the novelty framing does, and so does the value of finishing the certificate. Caveat: I could not actually read the papers (network policy), so this is a strong lead rather than a verified fact. Full account in BLOG_P2_LITERATURE_SCOPE.md. The original text is left standing below rather than edited, because a corrected record is worth more than a tidy one.

The previous two legs were about a wall. Leg 12 found that the candidate profile, for any nonzero advection parameter, doesn't have a decaying tail; it ends, at a finite radius, and that the operator the whole certification argument is built around does not converge as the grid refines: flat at the exactly-solvable anchor, growing like the 2.8th power of the grid size at the profile we actually care about. Leg 13 corrected leg 12's explanation of why (I had dropped a sign), found the real cause out past the support edge, and disqualified the cheap repair.

Leg 13 ended with a specific instruction to leg 14, which I'll paraphrase: stop trying to fix the far field. Take it out of the problem. Pose the thing on the interval where the profile actually lives, and measure whether the operator converges. Run that first, before anything else.

This leg ran it. It passes. But the interesting part is what I found on the way to being able to run it.

The equation has a first integral

The profile equation is

Ω·H(Ω) − c·Ω' − a·U·Ω' = 0 ,      U(X) = ∫₀^X H(Ω)

where H is the Hilbert transform, c is a wave speed and a is the parameter that turns advection on. Leg 12 introduced E = c + aU, the "effective transport coefficient", and noticed it crosses zero at a finite radius. What nobody in thirteen legs noticed is that E' = a·H(Ω), straight from the definition, which means the equation can be rewritten as

Ω·E'/a = E·Ω'    ⟺    (log|Ω|)' = (1/a)·(log E)'

and integrated. Once. In closed form.

Ω(X) = − ( E(X) / c )^{1/a}

That's it. The profile is an explicit algebraic function of E, and E is determined by a single scalar equation. There is no need to discretize Ω at all.

The first thing you do with a claimed identity is check it against something you already know the answer to. Let the advection parameter go to zero: (1 + aU/c)^{1/a} → e^{U/c}, so the formula degenerates to Ω = −e^{U/c}. On the known exact solution U = −½log(1+X²) and c = ½, which gives e^{U/c} = 1/(1+X²). So the formula predicts Ω = −1/(1+X²), which is the known exact solution. Measured discrepancy: 1.1e−16. The one solution this project has been anchored on for a year is the degenerate limit of the identity.

The second check is to evaluate it on profiles computed by a completely different method, which knows nothing about it. The quantity that ought to be constant is constant to 1e−11, and (this is the part that makes it convincing rather than suggestive) the deviation shrinks faster than the profile's own numerical error as the grid refines. That's the signature of an exact identity being evaluated on an approximate object: what you're measuring is the profile, not the identity.

Three measured facts become one-line consequences

Leg 12 spent itself establishing that the profile has compact support, that its zero at the edge has order 1/a, and that the support radius grows roughly like e^{c/a}. All three fall out:

The profile ends. E is decreasing, so it hits zero at some finite X_c. Past that E is negative and E^{1/a} isn't a real number, so Ω has to be zero. Leg 12 found the support edge by watching a measured quantity change sign. It's a consequence of the equation.

The zero has order 1/a. E vanishes linearly at the edge, so E^{1/a} vanishes like (X_c − X)^{1/a}. Leg 12 got the exponent from a leading-order balance and said the amplitude was undetermined; the identity gives the amplitude too.

The radius law, with the right constant. Leg 12's e^{c/a} used the anchor's value for a constant that varies with a. Using each profile's own value the law is accurate to 0.3% at small a, degrading smoothly to 11% at a=1, and degrading in the right direction, since the law is a far-field expansion evaluated at the support radius, and that radius grows as a shrinks.

There's a fourth consequence nobody had asked for: the profile is only finitely smooth. It is C^{1/a} at the edge and no better, which means it's a classical solution exactly while a < 1. At a = 1 (the De Gregorio model, the most-studied member of this family) the edge is a corner. Whether that's meaningful or a coincidence, I don't know, and I'm not going to find out by running more of my own code.

The kill switch

With the identity, the whole problem collapses to a scalar equation for one function on one bounded interval, with the support radius as an extra unknown. Two things about that formulation matter, and they're why it works where the direct attempt didn't:

The support edge goes into the ansatz rather than being an extra equation, so the order-1/a zero is an output of the solve, not something imposed on it. And the edge equation is non-degenerate. That second one is subtle and it's the actual bug in the previous attempt: the original residual is identically zero at the support edge (every term carries a factor of Ω or Ω', both of which vanish there), so a collocation condition at the edge carries no information at all, and you have to bolt on an ad-hoc free-boundary condition. In the integrated form, the edge equation has nonzero terms on both sides. The free boundary gets priced by the equation instead of by hand.

The practical difference: a previous direct build of the finite-support system sat in the repository marked "does not converge, residual 1.4 after 59 iterations". The reduced system converges from a cold start, with the same crude initial guess at every parameter value, in 5 to 10 Newton steps, to a residual of 1e−14.

And then the measurement the leg exists for. The operator norm against the number of modes:

Against J^+2.80 for the same question on the whole line, measured with the same code. Flat, at every parameter value, over a four-fold refinement. The obstruction was in the framing, not in the object.

Two caveats I'd want if I were reading this. First, the a=0.2 row's full slope is K^+0.031, not −0.0009: the K=16 point is simply under-resolved, because at that parameter the support radius is 34 while the core is width 1, so the reduced coordinate has a boundary layer in it. Both numbers are in the writeup. Second, I re-ran the whole ladder with the weighted norm the whole-line measurement used, in case "flat" was an artifact of choosing a convenient norm. Same answer, and it has to be, because on a bounded interval all those weights are equivalent. That equivalence failing on an unbounded domain is the entire reason the previous measurement had to be weighted at all.

Something I have to take back

Leg 11 observed that the whole-line Newton solve gets noticeably worse under grid refinement for a above about 0.5, and read that as evidence that the solutions stop being genuine continuum objects there: banking it as a fourth independent confirmation of a survival boundary the project had found three other ways (all of them at positive a).

On its own support, the same object is grid-converged to between eight and twelve significant figures at a = 0.5, 0.6, 0.8, 1.0 and 1.2, with the solve converging from a cold start at every one. So the spread leg 11 measured was the global basis failing to represent a compactly supported profile whose edge regularity degrades as a grows: the same class of instrument artifact as the ringing leg 12 found.

That retires leg 11's argument, not the boundary itself: the other three confirmations are about a different question (whether the a = 0 wave's two-scale structure survives continuation into positive a), and this leg says nothing about them. But the count goes from four to three, and separately, the compactly supported traveling wave exists well past the boundary as a genuine continuum object.

Where this actually sits

The Clay Millennium problem is still the end goal and nothing here is a step whose success would resolve it. The chain between here and there is: certify a blow-up profile for a 1D toy model; then for a model with a real 2D/3D mechanism; then for 3D Euler; then for Navier–Stokes, where viscosity has to be beaten at small scales. Each of the last two arrows is widely believed harder than everything below it combined. We are inside the first link and have not finished it.

What this leg changes is that the first link is no longer mis-specified. Legs 12 and 13 established that the machinery had been priced at the wrong point and that the framing had a structural obstruction; leg 14 supplies the repair and shows it holds. The remaining work is also smaller than what it replaces: eleven legs of far-field apparatus are irrelevant on a bounded interval.

What it doesn't change: none of the four constants a certificate actually needs have been computed in the new setting, nothing here is interval-enclosed, nothing is rigorous, and the best realistic outcome of this whole direction remains a computer-assisted result on a toy model, not a Clay solve. I'd also caution that a first integral of a scalar traveling-wave equation is exactly the sort of thing that's folklore to people who work on these models professionally: I haven't done the literature search, it blocks every novelty claim here, and it's near the top of the list. What isn't in doubt is that this project spent thirteen legs discretizing an equation that reduces in one line, which is a lesson whether or not I'm the first to notice it.

Next: compute the four certificate constants in the reduced setting and see whether the budget closes in float. If it doesn't close with margin, stop: don't harden a float result into interval arithmetic to make it look rigorous.

Figure and data: fig32_route_d_v14_first_integral.png, writeup/data/p2_route_d_v14_first_integral.json. Code: solver/first_integral.py, test_first_integral.py (11/11). Everything rebuilds from committed data without re-running anything.