The number that says why Navier–Stokes is hard

Route-F v1, the critical dissipation exponent, and a cross-check between two computations that share nothing.

Here is the sentence everyone writes about the Navier–Stokes problem, and almost nobody makes quantitative:

any real proof has to beat viscosity at small scales.

It is true, and as stated it is useless. So this leg turns it into an equation with a measured right-hand side, in a model where the arithmetic is checkable.

The dial

Take gCLM (a 1D caricature of the vorticity equation with an advection dial a) and give it adjustable dissipation:

ω_t + a u ω_x = ω u_x − ν (−Δ)^s ω ,       u_x = H(ω)

s = 1 is ordinary viscosity. s is the knob. The question is where, as s rises, dissipation stops losing.

The scaling argument is three lines and it is worth doing slowly, because the answer turns out to be something I had already measured for an unrelated reason.

A self-similar blow-up has amplitude ω ~ (T−t)^{−1} and a length scale L ~ (T−t)^β. The nonlinearity is ~ ω². The dissipation is ~ ν ω / L^{2s}. So

dissipation / nonlinearity  ~  ν (T − t)^{1 − 2sβ}

and the blow-up wins exactly when 1 − 2sβ > 0, i.e. s < 1/(2β).

Now, β. In the previous leg I built a dynamically-rescaled solver to ask a completely different question (whether a periodic orbit could bifurcate off the self-similar profile, it can't). Its by-product was α, the profile's far-field decay exponent: Ω ~ X^{−α}. And the rescaling ODEs give β = 1/α. So:

s_c(a) = α(a) / 2

The critical dissipation exponent is half the far-field decay exponent. The rate at which the profile decays in space determines whether the blow-up beats viscosity in time.

Why this is the Navier–Stokes sentence

Navier–Stokes has a natural scaling: L ~ (T−t)^{1/2}, i.e. β = 1/2, i.e. α = 2. Put that in:

s_c = 1 .

Exactly the ordinary Laplacian. Navier–Stokes sits precisely on the line where neither term wins, which is not a coincidence and is not bad luck; it is what "critical" means, and it is why the problem is hard. Every scaling argument you can make about NS returns exactly zero information, because the two sides balance identically.

In gCLM, by contrast, α is a measured function of a dial. It runs from 1 upward. So the family walks through the point where NS is stuck, and you can watch what happens on both sides. That is the whole value of the toy: not that it blows up, but that it is off-critical in a controlled way.

Measuring an exponent instead of a threshold

The obvious experiment (sweep s, see where blow-up stops) is exactly the experiment this project has learned not to run. Near a critical exponent the blow-up is only asymptotically dissipation-free, so at finite compute the apparent threshold is biased, resolution-dependent, and biased in the direction you expect. That is three ways to fool yourself in one measurement.

So instead: track D/N, the ratio of the dissipative to the nonlinear term at the peak, and fit

D/N ~ (T − t)^p ,     prediction:  p = 1 − 2s/α

This predicts a whole line, not a threshold. Its slope, its intercept and its zero crossing are separately checkable, and the zero crossing is s_c. Measuring a line is a far stronger test than locating a transition by eye.

One thing had to be fixed first. Dissipation delays the blow-up, so fitting against the inviscid singular time T₀ biases everything: every point came out above its prediction, with a shallower slope and a zero crossing ~10% high. That pattern, a uniform offset with a slope error, is the signature of a wrong singular time, not a wrong exponent. The fix is that each run supplies its own T: for ω ~ 1/(T−t), 1/amp is linear in t, so extrapolate it to zero. On the inviscid case, where T is known exactly, that recovers it to 3e−5.

The result, at the point where nothing is fitted

At a = 0 the model is exactly solvable: z = z₀/(1 − t z₀/2) with z = H(ω) + iω, which holds on the circle too (checked against an independent RK4 to 2e−14). And α = 1 exactly: the self-similar profile is a single Fourier mode. So the prediction p = 1 − 2s has no fitted input at all.

Slope −2.068 against −2; zero crossing at s = 0.5033 against 1/2. And s_c(0) = 1/2 is the classical critical exponent for dissipative CLM, so this is a known-answer gate rather than a self-consistency check.

The honest error bar

The dominant systematic is the fit window, and it is swept rather than chosen, because p is an asymptotic statement, an early window hasn't got there and a late one is noise:

slope = −2.02 ± 0.09 (predicted −2) · s_c = 0.51 ± 0.05 (predicted 0.500)

Individual exponents move by ±0.07 across windows, approaching the prediction monotonically from above, which is what entering an asymptotic regime looks like. The slope is much steadier, because every s shares the window and the bias cancels in the difference. That is why the claims here are about the slope and the zero, not about any single number.

The part I actually like

α was measured by a steady spectral solve, compactified, on the whole line, in a different module, for a different question. The slope dp/ds is measured by time-dependent pseudo-spectral simulation on a periodic domain with dissipation. These two computations share no grid, no basis, no formulation, and no fitted constant. The scaling relation says the second should be −2/α of the first.

The ratio is 1.022 ± 0.014 while α itself doubles. A uniform 2% bias, not an a-dependent failure: the shape of the relation is confirmed to sub-percent by an instrument that knows nothing about the one that produced α.

That is worth more than either leg's internal error bar. Refining one computation can't test the other; agreement across two unrelated ones tests both.

The map, and the sentence to be careful with

Putting the measured α(a) into s_c = α/2:

s_c crosses 1, the ordinary Laplacian, at a ≈ 0.383. Above that, the scaling says the blow-up beats ordinary viscosity.

Now the careful version, because this is the sentence that would be misquoted.

It is a statement about gCLM's own scaling, and it is not a statement about Navier–Stokes. It does not say that a viscous gCLM blow-up exists above a ≈ 0.383: the scaling says which term dominates given the self-similar form; showing a solution actually reaches it is the entire difficulty, here as everywhere. And NS's α is not a dial: it is pinned at 2 by dimensional analysis, which is precisely why NS gets no free information from this kind of argument.

What the map is genuinely good for is orientation: it shows what the NS difficulty is made of, by exhibiting a family that walks through it. NS is the marginal member of a family whose non-marginal members are computable.

What it cost

One module, one experiment, six gates, and three attempts at the resolution guard before I had an honest one. (Energy above ⅔ of k_max reads exactly 0.0 at some grid sizes, because the dealiasing already zeroed that band: a guard that is zero by construction reads as "perfectly resolved". Energy above n/6 reads 0.37 for every run, resolved or not, because a near-singular spectrum genuinely is fat. What discriminates is the amplitude at the cutoff relative to the peak: 1.0 at n = 1024, 1.4e−2 at n = 4096, 2.0e−4 at n = 8192. Now the run refuses rather than returning a number off an unresolved state.)

And the standing honesty, unchanged: none of this is Clay progress. It moves no link of the chain. It does not certify anything. What it does is take one sentence that everybody writes about why the problem is hard, and give it a measured right-hand side in a place where the measurement is possible. Odds unchanged, ~0.05%.

The lesson I'll carry: look for a second, structurally different route to a number you already have. α had been sitting in the previous leg's data file as a by-product. Used once, it was a curiosity. Used twice, by two instruments with nothing in common, it became a check on both.