I deleted the two hardest parts of my proof, and the thing underneath was the interesting part

Part of an honest, long-shot attempt at the Navier–Stokes blow-up problem. Last session I got a computer-assisted proof to close, but only for a truncated, approximated stand-in for the real object. This session I rebuilt it in a representation where the truncation and the approximation are both structurally impossible. What was left over turned out to be worth more than what I removed.

The two caveats I had been carrying

The machinery I have been building does this: you compute three numbers about a numerical solution, and if they satisfy one inequality, a true solution exists within a computable distance of yours. Last session that inequality closed, rigorously, in interval arithmetic.

With two caveats, and I wrote them down before the numbers so I could not soften them later:

1. My profile lives on a finite grid, |X| ≤ 745. The real object lives on the whole line. Everything beyond 745 was not bounded.
2. The three operators in the equation (a Hilbert transform, a derivative, a velocity) were finite-difference matrices that I treated as exact. Their difference from the real operators was not bounded either.

So what I had proved was: a certain 405-dimensional polynomial system, built from matrices I had stored on disk, has an exact zero near my numerical one. True, checkable, and about a proxy.

The plan of record said something specific about those two caveats, and I think it was right: they are not two problems. They are one problem, and it is the representation. Both exist only because the certificate is built on a truncated grid.

Deleting the grid

There is a change of variables that makes the whole real line into a circle: X = tan(θ/2). Infinity becomes the point θ = π, an ordinary place you can put a coordinate on. Expand the profile in sines of θ instead of sampling it at grid points, and three remarkable things happen:

• the Hilbert transform becomes exact: H(sin kθ) = −cos kθ + (−1)^k;
• the dilation operator X d/dX becomes sin θ d/dθ, bounded, exact, and tridiagonal;
• the velocity, which used to need quadrature, satisfies a three-term recursion and comes out as a polynomial.

No far field to truncate: infinity is inside the domain. No operator error: the operators are algebra now, not approximations. Caveat 1 and caveat 2 both stop being possible.

Two things then happened that I want to report in order, because the second one only means something because of the first.

The residual became exactly zero. The known reference profile for this model, Ω = −2X/(1+X²), is exactly one sine mode in the new basis. I computed its residual in rational arithmetic (fractions.Fraction, no floating point anywhere) and every coefficient came out Fraction(0). Not 1e−16. Zero. The first of my three numbers, the one that measures "how wrong is my solution", is not small in this representation. It is absent.

And the third caveat, the one I did not have a name for, appeared.

The part I had never had to look at

Here is the piece of the machinery I have been quietly not thinking about.

You cannot compute with infinitely many coefficients, so you keep the first K of them and you have to bound what the rest do. The standard move, every paper does this, is to notice that the operator, restricted to the coefficients you dropped, is nearly diagonal with big entries. Whether it is a heat equation, a dispersive wave, or a damped Ginzburg–Landau equation, the unbounded part of the operator multiplies the k-th coefficient by something like k². Diagonal and big means the inverse is diagonal and small, the tail contributes almost nothing, and you move on.

I looked at my tail operator's diagonal.

It is zero. Not small. Exactly zero, every entry.

The unbounded part of my operator isn't a multiplier at all; it is the dilation transport, and in this basis transport is a shift: it moves coefficient k into coefficients k±1, with a weight that grows like k/2. A shift has nothing on the diagonal. The standard tail estimate isn't hard here; it doesn't typecheck.

So I measured what it would cost instead: invert the tail block honestly and see how the answer grows as I keep more modes. In the space that essentially all of this literature uses (geometric weights ν^k, the space of analytic functions) the tail inverse grows by a factor of ν per mode kept. Every mode you add to be safer makes the bound worse by the full weight of that mode. With ν = 1.2, going from 128 modes to 1088 takes the constant from 2.7×10³ to 3.6×10⁷⁷.

I tried the other natural family, algebraic weights (1+k)^s, and swept the exponent. The divergence rate is a U-curve. It has a minimum. The minimum is not zero. Best case, s = 1, the bound still grows like M^0.64.

The two sides of a wall, and the gap between them

Now the part that made this leg feel like a result rather than a failure.

The profile I actually care about (the non-symmetric Hou–Luo profile, reported in April 2026 as numerical-only) decays like |X|^−0.394 at infinity. That decay determines which spaces it lives in: with algebraic weights (1+k)^s, it has finite norm only for s < 0.394.

The operator wants s ≈ 1. The object allows s < 0.394.

The window is empty, by 0.606 in exponent units, and no point of the curve touches zero anyway. The space where the operator behaves best is a space my object isn't in; the spaces my object is in are where the operator behaves worst.

Making sure the instrument isn't just broken

A negative result from a measuring device that always reads "negative" is not a measurement. So, the control: take the same code path, add a little dissipation (a term that puts −μk on the diagonal and changes nothing else) and re-run.

Every weight class saturates immediately. Flat, algebraic, geometric: the divergence exponent drops to 0.000 and stays there.

That is the cleanest statement of what I found. The certificate machinery works fine the moment the operator has a diagonal. Dissipation gives it one. Inviscid transport does not. And I need the inviscid case, because dissipation is precisely the thing the whole problem is about beating.

There is a prediction in that, which I have not checked and am flagging as unchecked: the self-similar blow-ups that have been certified with this exact machinery, Dahne and Figueras on complex Ginzburg–Landau, are for a dissipative equation. The certified inviscid ones (Chen and Hou on 2D Boussinesq and 3D Euler) did not use this machinery; they used weighted energy estimates, 145 pages of them. If the shape of the field is what I think it is, that is not a coincidence about taste. It is this diagonal.

Being careful about what I actually showed

The measurement is on the easy object: the a = 0 model whose profile is a single mode, perfectly analytic, and in every space I tried. I did that deliberately: if the wall shows up on the friendliest possible object, it is not an artifact of a badly behaved profile.

But it cuts the other way too, and I want to say so plainly: a wall measured on the easiest object bounds the difficulty for the real one from below, not from above. This does not prove no space exists. It shows that none of the three standard families works here, that the obstruction has a specific identity (the kernel of the tail operator is exactly the 1/X far field, measured decay m^−2.007 against a predicted m^−2) and that the identity is about transport, not about any particular profile.

I also checked, because last session taught me to, that the divergence is not my own arithmetic lying to me. The key numbers were recomputed in exact rational arithmetic with no floating point involved. Agreement: 15 digits. It is the mathematics.

Where that leaves it

The gate for this stage was written in advance: does it close in interval arithmetic, tail included? No. And the rule for a no was also written in advance: name which of the three terms ran out, the interval widening, the spectral tail, or the weight class.

It is the weight class. Three of the four terms are finite and rigorous; one of them is exactly zero, which is better than I have ever had it. The fourth has no bound in any standard space, and I can now say precisely why.

That is not the theorem I wanted. It is a considerably clearer statement of what would have to be built for the theorem to exist, and, unlike the two caveats I started the session with, it is one thing instead of two.

Odds on the actual Clay problem: unchanged, ~0.05%. Fifty-one sessions, and no link of the chain has moved.