The case where scaling tells you nothing

Part of an honest, long-shot attempt at the Navier–Stokes blow-up problem. The two previous posts measured where viscosity starts winning. Both ended at the same sentence: "and at the critical point exactly, this argument returns no information." That point is where Navier–Stokes lives. This post is about going there anyway. Still a toy model. Still not a breakthrough.

Why the empty answer is the whole problem

The standard way to ask whether viscosity stops a blow-up is to compare two terms. The nonlinear term wants to concentrate; the dissipative term wants to spread. Write down how each scales as the solution collapses, compare the powers, and you get a threshold: above it viscosity wins, below it the blow-up survives.

For Navier–Stokes the two powers are equal. Not close: equal, by dimensional analysis, exactly. The comparison gives 0 = 0.

This is what people mean when they say NS is critical, and it is a much more specific statement than "the problem is hard". It says a particular, powerful, cheap tool returns a blank. Everything else about the problem is downstream of that blank.

The previous two legs of this project measured that threshold in a toy model where you can dial it around, confirmed it against an independent instrument, ported it to a 2D system, and found, as expected, that the toy hits the same blank at its own critical point. Which left the obvious question unasked: what actually happens there?

The move: make the viscosity a variable

Here's the trick, and it's the only real idea in this post.

When you zoom in on a collapsing solution (rescaling as it shrinks so the shape stays put) the viscous term doesn't vanish and doesn't stay fixed. It comes back multiplied by a number that depends on how far you've zoomed. Call that number μ. Most treatments regard it as a nuisance parameter.

Treat it as a coordinate instead. Then the zoomed-in system has one more equation:

dμ/dτ = (2s − α) μ

where s is the viscosity's exponent and α describes how the collapsing profile falls off far from the singularity. Now everything is a dynamical system in (shape, μ), and two things fall out immediately.

First: the threshold was an eigenvalue all along. The rate 2s − α₀ is the growth rate of μ near the inviscid solution. Negative means μ decays and viscosity is irrelevant; positive means it takes over. The threshold is where it changes sign. A scaling exponent has turned into a stability exponent: same number, but now it belongs to an operator instead of a dimensional-analysis argument.

Second: at criticality the eigenvalue is zero, which is not the end but the beginning. When a linear term vanishes, you look at the quadratic one. Expand α in μ:

dμ/dτ = −α₁ μ² + …

and the whole marginal case is now the sign of one number, α₁. Positive: μ still decays, but as 1/τ instead of e^{−cτ}, algebraically, because the fast term is gone. Negative: viscosity runs away and there's no self-similar collapse. Zero: a whole family of viscous blow-ups sitting side by side.

That's the reduction. One number instead of a blank.

Measuring it

Two places in this family let you pose the critical problem exactly, and at both of them the answer is checkable.

At a = 0 there's a closed form. The equation reduces to a complex Burgers equation whose characteristics are a constant shift into the imaginary direction, and out drops an exact viscous blow-up in elementary functions, for every viscosity, and every member of a one-parameter family. Which means α₁ = 0 there: a line of viscous self-similar blow-ups, neutral not just to leading order but exactly.

That's the gate. The numerical machinery, which does not know the closed form, solves the rescaled equation from a cold start and lands on it to fifteen digits, at values of μ where the viscous term is four times the size of everything else. The far-field exponent comes back as 1.00000000000000 at every single μ.

At the next point the degeneracy breaks, and there's a trap worth naming. α(μ) is curved, so if you fit a straight line to it over any finite window you get the chord, not the slope at zero. That's a 5.5% error in the one number the whole verdict rests on. Fitting the secants and extrapolating them to zero instead:

Stable in the fourth digit, and positive. So at criticality μ decays, and the reflex reading ("more dissipation means more dissipation") is backwards. There's a unit test that checks the sentence, in words, because I do not trust myself to keep the sign straight.

But look at the rate. Algebraic decay means taking μ from 0.2 down to 0.02 costs τ = 337, and going one more decade costs τ = 3703. Nine times as long per decade, forever, and τ is itself a logarithm of the time remaining. So the critical viscous solution does relax onto the inviscid one, and does so impossibly slowly. That's the texture of criticality: not a barrier, a tar pit.

The third point nearly shipped the opposite answer

There's a third place where the critical problem should be exactly posable, and this is the part of the leg I'd most want someone to read.

At a finer grid it looks like it works. The solve converges to about 4e−4, which is an unremarkable residual, and the slope comes back negative, the opposite verdict. That's a headline: "the answer flips at the third point."

It's not a measurement. The entire movement of the quantity being differentiated, across the whole window, is 0.00002: twenty times smaller than the error bar of the solve it came out of. I was fitting a derivative to something well below my own noise, and getting a confident sign out of it.

What makes this worth writing down is that the obvious guard doesn't catch it. "Did the solve converge?" is the check you reach for, and 4e−4 passes any threshold you'd have picked in advance without knowing the answer. The check that catches it is a different question:

Is the thing I'm measuring bigger than the error in what I measured it from?

That's now the gate, and it's unit-tested against these exact numbers: including a test that the sign it would have quoted is a real sign, so the trap stays live rather than becoming a story about a trap.

The underlying reason it fails is also a correction to something I'd written. The operator needed there is a fifth power, built by composing five copies of something that each time pushes a little energy off the end of the basis. At the fifth power the discarded part is three to four orders larger than the part you keep. The "lucky alignment" I put in the module's own documentation (that landing on these special points makes everything exact) is true about the operator being finite and false about it being accurate. It buys the first and third powers, not the fifth. Refining does help, roughly as the fourth power of the grid size, but from so far up that you'd need a grid around 3000 to get there. I corrected the docstring rather than quietly not mentioning it.

So there is no third data point. The trend from two points is a trend from two points, and an earlier leg in this project already paid for forgetting that.

The interesting negative

An earlier leg tried a different attack (looking for solutions that repeat themselves at geometrically shrinking scales) and shut it down with a mechanism rather than a shrug. The relevant operator turned out to have almost no isolated eigenvalues at all. Its spectrum is a continuum, and a continuum has nothing discrete in it that could move, collide, and spiral off into the complex plane the way a bifurcation needs.

Critical viscosity is exactly the perturbation that could repair that. So: does it?

Yes. Turning μ on takes the count of grid-converged eigenvalues from 2 to 8, and they arrange themselves on the negative integers: a clean ladder condensing out of what was a continuum. The mechanism that shut the lane is genuinely gone.

And the lane stays shut anyway, which is a better place to be than before. Everything that condenses out lands on the negative real axis. Nothing reaches the right half plane (worst value: 3e−13, i.e. zero). Nothing goes complex: the two values my code first flagged as complex turned out to be a degenerate real pair split by 0.000018, which is noise wearing a costume. I changed the driver to report the size of the imaginary part instead of a yes/no, because the yes/no said "yes" and it was wrong.

One mode does move: the amplitude mode slides from −1 out to −√(1+4μ), fitting to nine digits. It moves left. More stable, not less.

The claim "nothing crosses" is an absence, so it needs a control: plant a bump in the same operator and the same filter finds three extra eigenvalues, one of them at +1.58, firmly in the right half plane. The filter can see. It's just that there's nothing there.

What this is and isn't

It moves nothing toward the Clay problem, and I want to be exact about the gap. This is a one-dimensional toy. The exponent used at the second point is hyperviscosity, chosen because it's where the critical problem can be posed exactly, not because it resembles anything physical. The closed-form solution decays like 1/x, so it has infinite energy, which is the honest analogy to the classical theorem ruling out exactly self-similar Navier–Stokes blow-up in the finite-energy class. The escape route from that theorem is precisely the slowly-decaying profiles, and they were constructed decades ago. This is a toy version of the thing that escapes, not a counterexample to anything.

What it does do is convert "the argument returns no information" into a question with an answer, in a setting where the answer is checkable against a closed form. And it leaves the project in a slightly better epistemic position on one lane: the negative result now rests on a measurement with a working control, rather than on there being nothing to measure.

The novelty is unchecked, and probably thin. Explicit viscous solutions of this equation by complexification date to 1986. The threshold formula is likely in a 2022 paper I still cannot download. If anything here is new it's methodological (treating the viscosity coefficient as a dynamical coordinate so that the marginal case becomes a normal form with one coefficient) and that's exactly the sort of thing that turns out to be standard the moment you ask someone in the field.