on finite-time blowup for forced Navier-Stokes

Overview

OpenAI’s 8 September 2026 preprint claims a finite-time singularity for the three-dimensional incompressible Navier—Stokes equations with smooth compactly supported external forcing. For every viscosity $\nu >0$, it constructs a smooth solution on $[0,1)$ starting from rest, with uniformly bounded kinetic energy but $\Vert u(t){\Vert }_{L^{\infty }}\to \infty$ as $t\uparrow 1$ (OpenAI, 2026). This is the forced whole-space breakdown alternative C in the Clay formulation. Because the fields are compactly supported in a fixed spatial set, the same construction periodizes to the three-torus, giving alternative D as well (OpenAI, 2026; Fefferman, 2000).

The central problem is not merely to make a velocity field large. Given any smooth $u$ before a terminal time, one can define the force to be its Navier—Stokes residual. To meet C/D, that residual and every one of its space—time derivatives must remain smooth through the proposed singular time. The paper designs an anisotropically collapsing axisymmetric vortex, represents its singular annular residual as a target momentum stress, realizes that stress with two families of sheared oscillatory pulses, and then iterates corrections until the remaining residual is flat to arbitrary order. Localization and a bounded-energy comparison argument turn the local construction into the claimed global obstruction.

Status and scope

This is a guide to an unrefereed September 2026 preprint, not an independent verification of its 166-page analysis. Clay has said that the problem has apparently been settled and that its evaluation and credit process is deliberately unhurried; it has not announced a prize award (Clay Mathematics Institute, 2026). The paper concerns a carefully constructed nonzero force. It does not prove finite-time blowup for the unforced Cauchy problem, nor does it prove the global regularity alternatives A/B. The accompanying Lean repository is valuable mechanical evidence about its formalized statement, but it is not a substitute for auditing the analytic construction and the correspondence between manuscript and formal statement (OpenAI, 2026b).

🏷️ The C/D Breakdown Target

For velocity $u:{\mathbb{R}}^3\times [0,\infty )\to {\mathbb{R}}^3$, pressure $p$, viscosity $\nu >0$, and force $f$, the incompressible Navier—Stokes system is

$${\partial }_{t}u+(u\cdot \nabla )u-\nu \Delta u+\nabla p=f,\nabla \cdot u=0.$$

The Clay problem permits a proof of any one of four alternatives. Alternatives A and B ask for global smoothness in the whole space and on the torus, respectively, with $f=0$. Alternatives C and D instead ask for one smooth datum and one smooth force for which no solution satisfying the stated global regularity conditions can exist (Fefferman, 2000). C is the whole-space version: the datum and force must decay faster than every power at spatial and temporal infinity, and a permitted global solution must have uniformly bounded kinetic energy. D is the spatially periodic version, with rapid time decay of the force.

Claimed whole-space construction

For every $\nu >0$, the preprint constructs

$$f\in C_c^{\infty }({\mathbb{R}}^3\times (0,\infty );{\mathbb{R}}^3),$$

a compact $K\subset {\mathbb{R}}^3$, and smooth $u,p$ on ${\mathbb{R}}^3\times [0,1)$ such that $u(\cdot ,0)=0$, the Navier—Stokes equations hold, and

$$\mathrm{supp}u(\cdot ,t)\cup \mathrm{supp}p(\cdot ,t)\subset K,\underset{0\le t<1}{\sup }\Vert u(t){\Vert }_{L^2({\mathbb{R}}^3)}<\infty ,$$

while

$$\underset{t\uparrow 1}{\mathrm{limsup}}\Vert u(t){\Vert }_{L^{\infty }({\mathbb{R}}^3)}=\infty .$$

The claimed consequence is that no global smooth whole-space solution with the same force and initial datum can have uniformly bounded energy. Fixed compact support allows periodization and yields the analogous torus statement (OpenAI, 2026a).

Compact support is stronger than the decay hypotheses in C. It is also exactly what makes the passage to D straightforward: choose a period cell containing $K$ in its interior, extend $u,p,f$ periodically, and retain the same local blowup before $t=1$.

The word consequence carries real weight. Constructing a blowing-up local solution does not by itself exclude some other admissible global solution with the same data; that step is a uniqueness argument, carried out in the paper’s §10.3 under hypotheses weak enough to admit an arbitrary competitor pressure. The construction and the exclusion are separate pieces of work, and they are discussed separately below.

🏷️ Concentration Without Energy Blowup

Put $\tau =1-t$. The leading field is axisymmetric in cylindrical coordinates $(r,\theta ,z)$ and is anisotropic: its radial and axial core lengths are claimed to obey

$${\ell }_r\asymp {\tau }^{1/2},{\ell }_z\asymp {\tau }^{1/2-h},0<h<\frac{1}{100}.$$

Thus the core becomes increasingly slender, since ${\ell }_r/{\ell }_z\asymp {\tau }^h\to 0$. The azimuthal and axial speeds grow on the scale ${\tau }^{-1/2-h}$, while the radial speed is only $O({\tau }^{-1/2})$:

$$|u_{\theta }|,|u_z|\asymp {\tau }^{-1/2-h},|u_r|=O({\tau }^{-1/2}).$$

The choice is designed to make the large pointwise velocity compatible with finite energy. The core volume is of order ${\ell }_r^2{\ell }_z\asymp {\tau }^{3/2-h}$, so the leading energy scaling is

$$({\tau }^{-1/2-h}{)}^2{\tau }^{3/2-h}={\tau }^{1/2-3h}⟶0.$$

☘️ The Reynolds Budget and the Expansion Parameter

This dimensional computation is only a consistency check; the paper later proves the actual localized energy estimate. Its dynamical role is more specific, and it is worth stating quantitatively, because the exponent $h$ is chosen to satisfy several competing inequalities at once. Measuring against the core radius ${\ell }_r$, the two Reynolds numbers behave differently:

$${\mathrm{Re}}_{\theta }=\frac{|u_{\theta }|{\ell }_r}{\nu }\asymp {\tau }^{-h}⟶\infty ,{\mathrm{Re}}_r=\frac{|u_r|{\ell }_r}{\nu }=O(1).$$

The angular Reynolds number diverges: fluid completes ever more turns during one radial diffusion time, which is what supplies enough shear to amplify oscillatory perturbations despite viscosity. The radial Reynolds number stays bounded, so viscosity continues to compete with radial inflow and the core cannot simply be advected into a genuine self-similar Euler singularity. The corresponding rates are

$$\frac{|u_r|}{{\ell }_r}=O({\tau }^{-1}),\frac{|u_z|}{{\ell }_z}\asymp {\tau }^{-1},\frac{\nu }{{\ell }_r^2}\asymp {\tau }^{-1},\frac{\nu /{\ell }_z^2}{\nu /{\ell }_r^2}=\frac{{\ell }_r^2}{{\ell }_z^2}\asymp {\tau }^{2h}.$$

Radial diffusion and both transport terms therefore remain in the leading balance, while axial diffusion is smaller by ${\tau }^{2h}$. That last ratio is not incidental bookkeeping: $q^{2h}$ becomes the formal expansion parameter in which the background corrections of §5 are organized, so the anisotropy exponent $h$ simultaneously sets the shear budget, the energy budget, and the small parameter of the background expansion (OpenAI, 2026a§2—3).

☘️ Similarity Coordinates and the Leading Profile

The anisotropy is encoded before any waves are added. Set

$$A=\frac{1}{2}+h,D=\frac{1}{2}-h,$$

and introduce $q>0$, $-1<\eta <1$, and $X\ge 0$ by

$$\tau =q(1-{\eta }^2),z=q^D\eta ,X=\frac{r^2}{2q}.$$

Eliminating $\eta$ gives $q-z^2{q}^{2h}=\tau$. Its derivative in $q$ is $1-2h{\eta }^2\ge 1-2h>0$, so this really defines one scale $q=q(z,\tau )$, and in fact $q\asymp \tau +|z{|}^{1/D}$. In the core $q\asymp \tau$; at $t=1$ away from the origin, the coordinate instead approaches $\eta =\pm 1$ with $q>0$. This separates the singular point from the regular terminal slice, and it is the reason a single scalar $q$ can serve as the flatness variable in the residual estimates (OpenAI, 2026a).

The leading axisymmetric field is then written in three fixed profiles:

$$u_{\theta }^{(0)}=q^{-A}E(X,\eta ),u_z^{(0)}=q^{-A}U(X,\eta ),r{u}_r^{(0)}=V_0(X,\eta ),p^{(0)}=q^{-2A}\Pi (X,\eta ).$$

This ansatz makes incompressibility an equation determining $V_0$ from $U$, with the regular-axis condition $V_0(0,\eta )=0$. The leading radial momentum balance determines the pressure profile from the swirl:

$${\partial }_r{p}^{(0)}=\frac{(u_{\theta }^{(0)}{)}^2}{r},\Pi (X,\eta )=-{\int }_X^{\infty }\frac{E(x,\eta {)}^2}{2x}dx.$$

Thus pressure is not an afterthought: its radial integral ties the axis, annulus, and exterior together. Smoothness in Cartesian coordinates requires $E/\sqrt{2X}$, $U$, and $V_0/X$ to be smooth at $X=0$, so the swirl vanishes on the axis even though it becomes unbounded on shrinking circles with $X$ fixed and positive (OpenAI, 2026a, sec.3.1).

That last point is where the unboundedness actually lives, and the paper isolates it as a single explicit path. Fixing $0<X_{\ast }<X_c$ inside the core and taking the circle $z=0$, $r=\sqrt{2X_{\ast }\tau }$ gives $q=\tau$, $\eta =0$, and

$$u_{\theta }^{(0)}=E(X_{\ast },0){\tau }^{-1/2-h}⟶+\infty .$$

Everything downstream — the corrections, the cutoffs, the periodization — is arranged so that this one asymptotic survives verbatim, with a relative error $O({\tau }^{2h})$ (OpenAI, 2026a).

☘️ Broken Reflection Symmetry and the Shear Budget

An axisymmetric vortex of this kind would naturally be built with exact reflection symmetry about $z=0$, and the manuscript deliberately declines to do so. The reason is a conflict between two jobs the core must perform simultaneously.

Away from the middle plane, axial flow carries angular momentum from faster-rotating layers, sustaining a steep radial decrease of angular velocity; this is the rotational amplification mechanism that feeds the pulses. Under exact reflection symmetry that transport vanishes at $z=0$. Symmetry would also force $u_z(r,0,t)=0$, removing the radial shear of axial velocity that might otherwise compensate. The middle plane would then have no amplification at all, and pulses placed there would simply be damped.

The construction therefore chooses a slightly asymmetric axial profile with a small upward bias and nonzero velocity at $z=0$. This separates the layer where rotational amplification is weak from the layer where axial shear vanishes, so that at every height one of the two mechanisms supplies the needed growth (OpenAI, 2026a, sec.2.1). The nonzero axis datum is also what produces the axial lower bound $\Vert u_z^{(0)}{\Vert }_{L^{\infty }(C_{\tau })}\asymp {\tau }^{-1/2-h}$.

☘️ Energy and Dissipation Budgets

The core $C_{\tau }$ is the region $0\le X\le X_c$, $|\eta |\le {\eta }_c$; it is a genuinely time-dependent spatial set whose boundary is fixed in similarity coordinates. Its leading kinetic energy and its leading radial-derivative Dirichlet integral scale as

$$E_{\mathrm{core}}\asymp {\tau }^{3/2-h}{\tau }^{-1-2h}={\tau }^{1/2-3h},D_{\mathrm{core}}\asymp {\tau }^{3/2-h}{\tau }^{-2-2h}={\tau }^{-1/2-3h}.$$

So the energy vanishes at the singular time, while the dissipation rate diverges. What matters for admissibility is that the divergence be integrable, and

$${\int }_0^{{\tau }_0}{\tau }^{-1/2-3h}d\tau =\frac{{\tau }_0^{1/2-3h}}{1/2-3h}<\infty \text{exactly when }h<\frac{1}{6}.$$

The constraint $h<1/100$ is far stronger than $h<1/6$, so this budget is comfortable; but it explains why the anisotropy exponent cannot simply be taken large to make the core more slender. Beyond $h=1/6$ the construction would produce a flow of vanishing energy whose total dissipation on $[0,1)$ is infinite, which no Leray-type admissibility statement would tolerate (OpenAI, 2026a, sec.3.5).

The force is not prescribed by hand near the blowup

Radial inflow carries angular momentum to smaller radii, spinning up the swirl. Incompressibility sends fluid away axially rather than allowing it to pile up on the axis. But this background vortex does not itself solve the forced problem with a smooth force: its momentum residual diverges in the annular transition to the exterior. The proof must make the fluid’s own quadratic momentum transport cancel that divergent residual.

🏷️ The Residual-Cancellation Problem

At viscosity one, write the momentum residual as

$$\mathcal{R}(u,p)={\partial }_{t}u+(u\cdot \nabla )u-\Delta u+\nabla p.$$

If the desired force is $f=\mathcal{R}(u,p)$, then it is not enough that $\mathcal{R}$ have a pointwise limit at $t=1$. Smooth extension across $t=1$ requires compatible limits for every ${\partial }_x^{\alpha }{\partial }_t^j\mathcal{R}$. Near the concentrating point, the construction seeks the much stronger flatness estimate

$$\left|{\partial }_x^{\alpha }{\partial }_t^j\mathcal{R}(u,p)\right|\le C_{\alpha ,j,N}{q}^N\text{for every }\alpha ,j,N,$$

uniformly on $0\le X\le X_1$ for each finite $X_1$, where $q$ is the similarity-scale variable above. Flatness converts the singular-time problem from “does the residual remain bounded?” to “can all its jets be made to vanish?” (OpenAI, 2026a).

The background construction arranges that the leading residual vanishes in the inner core and in an exterior exact heat flow. Only the intervening cylindrical annulus remains. There the tangential residuals ${\mathcal{R}}_{\theta }^{(0)},{\mathcal{R}}_z^{(0)}$ are written as divergences of a two-component physical stress $T=(T_{r\theta },T_{rz})$ in the cylindrical weights appropriate to angular momentum and axial momentum:

$${\mathcal{R}}_{\theta }^{(0)}=-({\partial }_r+\frac{2}{r})T_{r\theta },{\mathcal{R}}_z^{(0)}=-({\partial }_r+\frac{1}{r})T_{rz}.$$

Regularity at the axis fixes the integration constants, so the stress is not a free antiderivative but an explicit cumulative integral:

$$T_{r\theta }(r)=-\frac{1}{r^2}{\int }_0^r{s}^2{\mathcal{R}}_{\theta }^{(0)}(s)ds,T_{rz}(r)=-\frac{1}{r}{\int }_0^{r}s{\mathcal{R}}_z^{(0)}(s)ds.$$

Requiring $T$ to vanish beyond the annulus as well as inside the core is therefore a pair of global moment conditions on the residual:

$${\int }_0^{\infty }{r}^2{\mathcal{R}}_{\theta }^{(0)}dr=0,{\int }_0^{\infty }r{\mathcal{R}}_z^{(0)}dr=0.$$

These are exactly the zero angular-momentum and zero axial-flux identities supplied by the prescribed exterior moments, and they are what confine all forcing work to one annulus $X_a<X<X_b$ (OpenAI, 2026a, sec.3.2 and Lemma A.8).

☘️ Profile Matching and the Five Cumulative Radial Integrals

The profile construction has to solve two conflicting boundary-value problems at once. Near $X=0$, analytic profiles make the leading tangential residual and stress vanish. Far from the core, the field is a purely azimuthal heat exterior — a solution of the radial swirl heat equation

$$-{\partial }_{\tau }K={\partial }_r^{2}K+r^{-1}{\partial }_{r}K-r^{-2}K,K(r,\tau )=r^{-1-2h}{H}_{\mathrm{ext}}(\tau /r^2),$$

whose residual vanishes exactly, so no force is needed to sustain it. Joining the two cannot be done by matching profile values alone, because pressure, radial velocity, and both stress components depend on cumulative radial integrals. The manuscript identifies exactly five such integrals that suffice:

$$M={\int }_0^{X}Udx,I={\int }_0^{X}Hdx,J={\int }_0^{X}UHdx,$$ $$S={\int }_0^X(U^2-E^2/2)dx,C_p={\int }_0^X\frac{E^2}{2x}dx,\Pi =\Pi (0,\eta )+C_p,$$

with $H=\sqrt{2X}E$. If two profile pairs agree beyond a joining radius and their five integrals agree there, then the induced pressure, radial velocity, shear coefficients, and stress agree for all larger radii. The accompanying stability estimate bounds differences of the derived quantities by differences of the profiles and integrals in one higher $\eta$-derivative — crucially without requiring radial derivatives of the difference to be small. That asymmetry is what later permits a profile repair that changes the shear at order one while moving the five integrals only slightly (OpenAI, 2026a).

☘️ The Admissible Stress Cone

The remaining requirement on the background is geometric, and it is the sharpest constraint in the profile construction. The two wave families will deliver two covariance directions $v_1,v_2\in {\mathbb{R}}^2$ in the tangential $(\theta ,z)$ plane; because the coefficients multiplying them are squared wave amplitudes, the target stress must lie in their positive cone,

$$T=c_1{v}_1+c_2{v}_2,c_1,c_2>0.$$

Positivity is not cosmetic. A signed representation would not be realizable by the selected oscillations at all, since the leading covariance of a real pulse enters with a definite sign.

The manuscript converts this into explicit inequalities on the profiles. At each $(X,\eta )$ the relevant data are the radial shear $(a,-b_s)$ and the integrated inviscid contribution $p_s=(p_{s,1},p_{s,2})$. Where $a>0$ one sets

$$t_s=-\frac{b_s}{a},v_s=a(1+t_s^2),P_c=p_{s,1}+t_s{p}_{s,2},J_c=p_{s,2}-t_s{p}_{s,1},$$

so that $P_c,J_c$ are the components of $p_s$ along the orthogonal unnormalized pair $(1,t_s)$ and $(-t_s,1)$. The admissible stress cone condition is then

$$v_s>2,P_c>v_s,(v_s-2)J_c^2<2(P_c-v_s{)}^2.$$

Dropping $v_s>2$ gives the relaxed condition, which is all that is needed while joining the axis profile to the outer profile; the extra inequality is imposed by the viscous waves, and its meaning is taken up below. Written directly in stress coordinates the same condition reads

$$T_{0,\theta }+t_s{T}_{0,z}>0,(v_s-2)(T_{0,z}-t_s{T}_{0,\theta }{)}^2<2(T_{0,\theta }+t_s{T}_{0,z}{)}^2.$$

Both inequalities are homogeneous in $T_0$. That is the point of writing them this way: as the stress decays smoothly to zero at an annular edge, its unit direction can still retain a strict margin inside the cone, so the edge condition is a statement about direction rather than magnitude (OpenAI, 2026a, sec.4.3 and Thm. 4.6(iii)).

Proof: the equivalence of the two forms of the cone condition

The manuscript’s first form asks for $P_c>2$ together with $v_s<U(P_c,J_c)$, where

$$U(P_c,J_c)=P_c+\frac{J_c^2}{4}-|J_c|\sqrt{\frac{P_c-2}{2}+\frac{J_c^2}{16}}.$$

Consider the quadratic in $v$

$$\mathfrak{p}(v)=2(P_c-v{)}^2-(v-2)J_c^2=2v^2-(4P_c+J_c^2)v+(2P_c^2+2J_c^2).$$

Its discriminant is $J_c^2(8P_c+J_c^2-16)$, so its roots are

$$v_{\pm }=P_c+\frac{J_c^2}{4}\pm |J_c|\sqrt{\frac{P_c-2}{2}+\frac{J_c^2}{16}},$$

and $U=v_{-}$. The leading coefficient is $2>0$, so $\mathfrak{p}$ opens upward. Evaluating at the two natural points gives $\mathfrak{p}(2)=2(P_c-2{)}^2>0$ and $\mathfrak{p}(P_c)=-(P_c-2)J_c^2\le 0$ whenever $P_c>2$, which places $v_{-}$ in the interval $(2,P_c]$. Hence on $2<v<P_c$ one has $\mathfrak{p}(v)>0$ precisely when $v<v_{-}$.

Now assume $v_s>2$. If $P_c>2$ and $v_s<U=v_{-}$, then $v_s<v_{-}\le P_c$ gives $P_c>v_s$, and $\mathfrak{p}(v_s)>0$ is the inequality $(v_s-2)J_c^2<2(P_c-v_s{)}^2$. Conversely, $P_c>v_s>2$ forces $P_c>2$, and $\mathfrak{p}(v_s)>0$ with $2<v_s<P_c$ forces $v_s<v_{-}=U$. The two forms therefore agree, which is the content of Lemma 4.5 (OpenAI, 2026a, p.L emma 4.5).

The construction enforces this condition by a device worth recording, because it exploits precisely the asymmetry of the matching estimate noted above. On a compact subinterval of the annulus the manuscript inserts a radial oscillation with phase $N\log X$ for a large fixed integer $N$. Its amplitude is $O(N^{-1})$, so profile values and all five radial moments move only by $O(N^{-1})$; but applying $X{\partial }_X$ to it produces an order-one change in the radial derivatives, hence in the leading tangential shear. The shear then follows a periodic curve arranged to satisfy the cone inequalities, and a separate localized correction restores the five integrals exactly. The inner solution and the heat exterior are untouched (OpenAI, 2026a, sec.3.2 and Prop. C.2—C.3).

🏷️ Oscillatory Stress Realization

flowchart TB
    CORE["Anisotropic axisymmetric core<br/>unbounded swirl and axial speed"]
    ANNULUS["Annular residual<br/>minus divergence of target stress T"]
    WAVES["Two sheared oscillatory wave families<br/>positive quadratic fluxes span T"]
    CORRECT["Correction cycle<br/>waves, mean flow, pressure, moments"]
    FLAT["Residual flat to every order<br/>at the singular point"]
    FORCE["Localization and smooth compact force"]
    CD["C on R3; periodization gives D on T3"]
    CORE --> ANNULUS --> WAVES --> CORRECT --> FLAT --> FORCE --> CD
    CORRECT --> CORRECT

Locally, a divergence-free oscillation looks schematically like

$$w_{\mathrm{lead}}(x)=a\cos (\xi \cdot x+\varphi ),a\cdot \xi =0.$$

Its mean is zero, but the average of $w\otimes w$ is nonzero. In a sheared background, the divergence of that quadratic flux acts on the mean flow as a momentum stress. The paper chooses the pulse geometry so that two families deliver two independent tangential flux directions and their positive span contains the target $T$. Physically, outward motion carrying an azimuthal velocity surplus and inward motion carrying a deficit both increase the outward flux of angular momentum; reversing both components leaves the product unchanged, which is why a zero-mean oscillation can transport momentum at all.

The scaling exposes why this has a chance to match the singular background. The pulse amplitude and wavelength are chosen as

$$A_{\mathrm{wave}}\asymp q^{-1/2-h/2},{\ell }_{\mathrm{wave}}\asymp q^{1/2+h/2},$$

so that the pulses are weaker and shorter than the background by the same factor:

$$\frac{A_{\mathrm{wave}}}{q^{-1/2-h}}\asymp \frac{{\ell }_{\mathrm{wave}}}{q^{1/2}}\asymp q^{h/2}.$$

Hence $A_{\mathrm{wave}}^2\asymp q^{-1-h}=|u_{\theta }^{(0)}|/q^{1/2}$, and the divergence of the averaged stress has scale $A_{\mathrm{wave}}^2/{\ell }_r\asymp q^{-3/2-h}$ — exactly the scale of the leading time derivative, transport, and radial viscosity in the tangential momentum equations (OpenAI, 2026a§2.2 and 3.3).

☘️ The Pulse Amplitude Equation and Its Gaussian Envelope

The manuscript does not treat the pulses as a formal ansatz; it solves an explicit linear equation for their amplitudes along a phase. Fix a dyadic band where $q\asymp Q$, write $\epsilon =Q^h$, and let $\Phi$ be a phase that approximately follows transport by the background, with normal $n_{\Phi }={\nabla }_{\ast }\Phi$. For a harmonic $e^{ikm\Phi }$, incompressibility to leading order is $n_{\Phi }\cdot t_m=0$, and the principal amplitude problem is to find $t_m\in {\mathbb{C}}^3$ and a pressure coefficient ${\pi }_m$ with

$$t_m^{\prime }+\mathcal{K}{t}_m+m^{2}d{t}_m+ikm{n}_{\Phi }{\pi }_m=-f_m,d=\epsilon k^2|n_{\Phi }{|}^2,$$

where ${}^{\prime }$ is differentiation along the pulse coordinate $v$. The matrix $\mathcal{K}$ carries the base tangential velocity and the radial shear, together with the cylindrical connection terms — radial motion rotates the tangential frame as well as advecting the base swirl. Eliminating the normal pressure while differentiating the constraint gives the projected evolution operator

$${\mathcal{A}}_{\Phi }=-\mathcal{K}+\frac{n_{\Phi }{n}_{\Phi }^{\mathsf{T}}(\mathcal{K}-(n_{\Phi }^{\prime }{)}^{\mathsf{T}})}{|n_{\Phi }{|}^2}.$$

In a moving orthonormal frame $B$ spanning $n_{\Phi }^{\perp }$, this two-dimensional system is compared with a diagonal reference,

$$B^{\ell }({\mathcal{A}}_{\Phi }B-B^{\prime })=\mathrm{diag}(\lambda ,-\lambda )+E,\lambda (v)=\frac{{\lambda }_0}{\sqrt{1+s(v{)}^2}},|E|\le C/S_{\ast },$$

with one growing and one decaying direction. The parameter $s(v)$ is the normalized radial component of the phase normal; it increases linearly along the pulse, because shear tilts the wavevector and steadily shortens the radial wavelength.

That single fact drives the whole envelope. Growth is governed by $\lambda (v)$, which decreases as $s$ grows, while viscous damping is governed by $d\approx \epsilon k^2{B}_s^2(1+s^2)$, which increases. The net exponential rate is

$$a_{\mathrm{net}}(y)=\frac{{\lambda }_0}{\sqrt{1+y^2}}-\frac{{\lambda }_0(1+y^2)}{(1+u_{\ast }^2{)}^{3/2}},y=|s(v)|,$$

with $B_s$ fixed so that the two terms balance at $y=u_{\ast }$. Along the pulse $|s(v)|$ sweeps the interval $[u_{\ast }/2,3u_{\ast }/2]$, and

$$\frac{d}{dv}{a}_{\mathrm{net}}(|s(v)|)=-\frac{u_{\ast }{\lambda }_0|s(v)|}{L_s}[(1+s(v{)}^2{)}^{-3/2}+2(1+u_{\ast }^2{)}^{-3/2}]\in [-\frac{C}{L_s},-\frac{c}{L_s}].$$

Since $(\log P{)}^{\prime }=a_{\mathrm{net}}$ and both $\log P$ and its derivative vanish at the midpoint $v=L_s/2$, integrating these bounds twice from the midpoint yields a two-sided Gaussian envelope:

$$e^{-C(v-L_s/2{)}^2/L_s}\le P(v)\le e^{-c(v-L_s/2{)}^2/L_s}\le 1.$$

So “shear amplifies, then viscosity wins” is not a qualitative remark but a quantitative statement: the pulse envelope is squeezed between two Gaussians in the pulse coordinate, peaked at the midpoint and exponentially small at both ends. The temporal cutoff is placed in those tails, and the errors it creates, together with all their derivatives, vanish at the singularity (OpenAI, 2026a, p.L emma 7.1, Eq. (7.16) and Lemma 7.4).

☘️ The Threshold as an Instability Criterion

The two conditions on the background — a profile inequality in §4 and an instability requirement in §7 — turn out to be the same inequality, and this is the structural hinge of the argument.

In the frame adapted to the radial shear $g_0$, set $N=g_0/|g_0|$ and let $K=N^{\perp }$ be the quarter turn. (This tangential frame vector $K$ is unrelated to the compact support set $K\subset {\mathbb{R}}^3$ of the theorem; both follow the manuscript’s notation.) The reference growth rate is defined by

$${\lambda }_0^2=-2F_0{N}_{\theta }(2F_0{N}_{\theta }+|g_0|),c_0=\frac{{\lambda }_0}{2F_0{N}_{\theta }}<0,$$

where $F_0=V/R$ is the angular velocity. Substituting $g_0=F_0(-a,b_s)$ and the definitions of $t_s,v_s$ collapses this to

$${\lambda }_0^2=2a{F}_0^2(1-\frac{2}{v_s}),c_0^2=\frac{v_s-2}{2}.$$

Hence ${\lambda }_0$ is real and positive precisely when $v_s>2$. The inequality that looked in §4 like an arbitrary algebraic threshold in the profile variables is in fact the statement that the background shear is unstable enough for the pulses to grow at all; it is a centrifugal-instability criterion in the tradition of Lifschitz—Hameiri and Friedlander—Vishik, specialized to this frame (OpenAI, 2026a, sec.7.1; Lifschitz & Hameiri, 1991; Friedlander & Vishik, 1991). Correspondingly, the frame constant $c_0$ appearing in the stress inequality is not independent data: $c_0^2=(v_s-2)/2$ is exactly the factor converting the pointwise condition

$$T_{0,\ast }\cdot N<0,\left|c_0\frac{T_{0,\ast }\cdot K}{T_{0,\ast }\cdot N}\right|<1$$

into the profile inequalities displayed earlier.

The instability that makes ${\lambda }_0>0$ also constrains which stresses the pulses are able to deliver. Taking the scalar product of the amplitude equation with the real amplitude $t$ gives

$$\frac{1}{2}\frac{d}{dv}|t{|}^2=-g\cdot (t_r(t_{\theta },t_z))-\epsilon k^2|n_{\Phi }{|}^2|t{|}^2.$$

The first term is energy exchange with the base shear and the second is viscous dissipation. But the vector $t_r(t_{\theta },t_z)$ is the same object that must match the prescribed momentum flux: it records radial transport of azimuthal and axial momentum. Decomposing a flux vector as $T=T_{N}N+T_{K}K$, the energy transfer into the pulse is $-g_0\cdot T=-|g_0|T_N$. So the wave can only grow if $T_N<0$, while the covariance construction independently prescribes the transverse component $T_K$. The pulses cannot be tuned freely to produce whatever stress is wanted; what they can produce is constrained by what allows them to exist (OpenAI, 2026a).

☘️ Positive Squared Amplitudes from the Covariance Matrix

The two homogeneous pulses $t_{h\pm }$ obtained from the growing frame coordinate determine two covariance directions per slow box. Writing $\mathcal{H}=[\mathcal{C}(b_{+})\mid \mathcal{C}(b_{-})]$ for the $2\times 2$ matrix of their covariances in the component order $(r\theta ,rz)$, disjointness of the two signs’ auxiliary supports gives exactly

$$\mathcal{C}(a_{+}{b}_{+}+a_{-}{b}_{-})=\mathcal{H}(a_{+}^2,a_{-}^2{)}^{\mathsf{T}},$$

so the achievable stresses are precisely the nonnegative span of the columns. To leading order the columns are computable: with $A_c=-c_0\sqrt{1+u_{\ast }^2}>0$,

$${\mathcal{H}}_{\sigma }=h_{\sigma }(-A_{c}N-\sigma u_{\ast }K+e_{\sigma }),|e_{\sigma }|\le C{S}_{\ast }^{-1/2},h_{\sigma }\asymp S_{\ast }^{-1/2},$$

where the positive scalar $h_{\sigma }$ absorbs the transverse and pulse-coordinate integrals together with the exact angular average $⟨{\cos }^2(k{\Phi }_{\sigma }){⟩}_{\theta }=1/2$ — exact, rather than approximate, because $kp$ is a nonzero integer. Dropping the errors, the two scalar equations for a target $T=T_{N}N+T_{K}K$ solve explicitly:

$$h_{+}{y}_{+}=\frac{1}{2}(-\frac{T_N}{A_c}-\frac{T_K}{u_{\ast }}),h_{-}{y}_{-}=\frac{1}{2}(-\frac{T_N}{A_c}+\frac{T_K}{u_{\ast }}).$$

Both are positive exactly when $T_N<0$ and $|T_K|<(u_{\ast }/A_c)(-T_N)$ — the reference cone. The margin built into the profile at §4 is uniform, so it survives freezing the frame at a representative point, and the errors $e_{\sigma }$ preserve the strict inequality. The square roots $a_{\sigma }=\sqrt{y_{\sigma }}$ are then the real wave amplitudes. Differentiating the relation $y={\mathcal{H}}^{-1}{T}_{0,\ast }$ at these fixed amplitudes gives a linear map whose image contains stress increments of either sign, which is what later corrections use (OpenAI, 2026a).

☘️ Separation of Supports on the Auxiliary Torus

The construction must stop distinct pulses from creating uncontrolled cross terms. It introduces an auxiliary two-torus parameter $Y\in {\mathbb{T}}^2={\mathbb{R}}^2/{\mathbb{Z}}^2$, independent of the physical coordinates, builds an extended field $\tilde{w}(r,\theta ,z,t,Y)$, and recovers the physical field by evaluation $w(r,\theta ,z,t)=\tilde{w}(r,\theta ,z,t,Y(r,t))$. Covariances are computed before evaluation. Pulses whose slow space—time supports overlap receive disjoint supports in $Y$, so their products vanish pointwise even after evaluation; harmonics and corrections carrying the same label still interact. This is a bookkeeping device for exact separation, not the physical torus of alternative D.

The specific choices are more than notation, and they are where a small-divisor problem is quietly solved. Let

$$J_g=\left(\begin{array}{cc}3& 1\\ 1& 5\end{array}\right),{\Lambda }_g=4-\sqrt{2},T_g=4+\sqrt{2},b_g=\sqrt{2}-1,$$

with eigenvectors $v_r=(1,-b_g)$ and $v_t=(b_g,1)$ for ${\Lambda }_g$ and $T_g$ respectively, and define the phase map

$$Y=v_r{r}^{d_r}+v_{t}t(\mathrm{mod}{\mathbb{Z}}^2).$$

Two features are being bought at once. First, because $1<{\Lambda }_g<T_g$, an integer power $J_g^i$ expands the temporal direction much faster than the radial one; choosing $i=i(\ell )$ per dyadic band makes the fast time derivative of order $S_{\ast }^{-1}$ while the fast radial derivative carries only the small factor ${\epsilon }^{-{\kappa }_s}$. This is precisely the requirement that the pulses vary rapidly in time without generating radial derivatives large enough to spoil the residual estimates.

Second, the correction cycle must invert the fast auxiliary-time derivative on zero-mean torus functions, which on a Fourier mode $n$ means dividing by $v_t\cdot n$. Both directions satisfy

$$|v_r\cdot n|\ge \frac{c}{1+|n|},|v_t\cdot n|\ge \frac{c}{1+|n|}(n\in {\mathbb{Z}}^2\setminus \{0\}),$$

and the proof is a one-line Diophantine argument: $v_r\cdot n=(n_1+n_2)-\sqrt{2}{n}_2$, and multiplying by its algebraic conjugate gives the nonzero integer $(n_1+n_2{)}^2-2n_2^2$, while the conjugate itself is $O(|n|)$. The same computation for $v_t$ gives $(n_2-n_1{)}^2-2n_1^2$. A quadratic irrational is badly approximable, so the inverse directional operators lose only finitely many torus derivatives rather than an unbounded number (OpenAI, 2026a).

☘️ From a Formal Average to an Exact Divergence-Free Field

The average is not inserted by hand. Each physical pulse is defined as a curl $w=\nabla \times A_{\mathrm{wave}}$ of a localized vector potential evaluated at $Y(r,t)$, so it is exactly divergence-free; the chain-rule terms from the phase map and the cutoff derivatives are retained rather than discarded. Its angular frequency is a nonzero integer, so the angular mean of the pulse is zero and the angular average of the leading cosine square is $1/2$. In the extended variables, the construction proves

$$\left(\begin{array}{c}⟨{\tilde{w}}_r{\tilde{w}}_{\theta }⟩\\ ⟨{\tilde{w}}_r{\tilde{w}}_z⟩\end{array}\right)=T+\text{higher-order terms},$$

where the brackets average both the angle and the auxiliary torus, and the higher-order terms — which include the curl corrections and cutoff tails — are smaller by positive powers of $q$ with proved derivative bounds. The radial divergence of this leading covariance cancels the leading negative stress divergence of the background (OpenAI, 2026a, sec.3.3, Lemmas 6.1, 7.7 and Cor. 7.8).

It is worth noting why viscosity does not degenerate out of the problem. The carrier wavelength is $\sqrt{Q}/k\asymp Q^{1/2+h/2}$ while the physical pulse amplitude is $Q^{-A}\sqrt{\epsilon }=Q^{-1/2-h/2}$, so the powers of $Q$ in their product cancel and $\epsilon k^2\asymp 1$ throughout. Viscosity therefore remains a leading-order term in the pulse equation at every stage of the correction process, rather than being a subdominant perturbation as in the inviscid convex-integration constructions (OpenAI, 2026a, sec.7.2; Daneri & Székelyhidi, 2017).

🏷️ Infinite Correction and Smooth Forcing

Canceling the leading stress does not finish the argument: curl corrections, cutoffs, pressure adjustments, and wave interactions leave further residuals. At stage $j$, the paper adds a divergence-free increment $\delta u_j$ and a pressure increment $\delta p_j$. The full, rather than linearized, residual obeys

$$\mathcal{R}(u^{[j+1]},p^{[j+1]})=\mathcal{R}(u^{[j]},p^{[j]})+L_{u^{[j]}}(\delta u_j,\delta p_j)+\nabla \cdot (\delta u_j\otimes \delta u_j),$$

with

$$L_v(w,\pi ):={\partial }_{t}w+(v\cdot \nabla )w+(w\cdot \nabla )v-\Delta w+\nabla \pi .$$

The last term is why a one-shot cancellation is insufficient. The cycle recomputes the entire residual after every operation, so all newly created linear and quadratic terms become the source for the next cycle.

The four components of one cycle are deliberately different.

• Inhomogeneous wave-amplitude equations cancel the supported nonzero angular Fourier modes, using the linear inverse constructed for the pulse equation with a prescribed harmonic source.

• Signed amplitude increments have cross covariance with the fixed positive leading pulses; this corrects the mean stress without destroying its initial positive-cone realization, because the linearized map $\mathcal{D}\mathcal{C}(w)[v]=\mathcal{B}(w,v)$ produces increments of either sign around a strictly interior point.

• An angularly averaged, zero-auxiliary-mean residual is removed by inverting the fast auxiliary-time derivative — the step that requires the small-divisor bound above. A vector potential realizes the resulting axial increment together with its required radial component.

• Five radial moment equations preserve the zero angular-momentum and axial-flux integrals and correct the pressure and tangential-momentum defects introduced by the preceding operations.

The important invariant is therefore not merely smallness. It is simultaneous preservation of incompressibility, support, the moment conditions needed for the heat exterior, and a residual class stable under the next correction (OpenAI, 2026a, sec.3.4).

☘️ Origin of the Five Moment Equations

The count is forced, and the manuscript derives it rather than asserting it. Writing the actual velocity in a chart as a slow base $(b,V,G)$, an angularly invariant correction $(\beta ,v,\gamma )$, and the divergence-free wave field $w$ with $⟨w{⟩}_{\theta }=0$, the angular mean of the momentum equations can be put in conservative form. Two integral constraints are imposed on the correction at every $(Z,T)$:

$$M_{\theta }={\int }_0^{\infty }{R}^2⟨v{⟩}_{Y}dR=0,M_z={\int }_0^{\infty }R⟨\gamma {⟩}_{Y}dR=0,$$

the zero angular-momentum moment and the zero axial-flux integral. Three further scalars measure the obstruction to representing the mean residual by a compactly supported stress: the pressure defect $\mathcal{P}$ from the radial reconstruction, and the two flux defects

$${\mathcal{J}}_{\theta }=\int R^2⟨Gv+V\gamma +\gamma v+W_{z\theta }{⟩}_{Y}dR,$$ $${\mathcal{J}}_z=\int R⟨2G\gamma +{\gamma }^2+W_{zz}{⟩}_{Y}dR-\frac{1}{2}\int R^2⟨g_r{⟩}_{Y}dR,$$

where $W_{ab}=⟨w_a{w}_b{⟩}_{\theta }$ retains every curl remainder. The key computation is that once $M_{\theta }=M_z=0$, the weighted radial integrals of the two tangential mean residuals collapse to axial derivatives of these defects alone:

$$\int R^2⟨{\mathcal{E}}_{\theta }{⟩}_{Y}dR=\epsilon {\partial }_Z{\mathcal{J}}_{\theta },\int R⟨{\mathcal{E}}_z{⟩}_{Y}dR=\epsilon {\partial }_Z({\mathcal{J}}_z+c_{\rho }\mathcal{P}).$$

The mechanism is clean: every radial divergence becomes an endpoint term that vanishes by compact support, the radial viscous integrals cancel after two integrations by parts, and the time and axial-viscosity terms are derivatives of the two zero moments. So the system to be solved at each stage has exactly five unknowns — two to preserve $(M_{\theta },M_z)$ and three to cancel the linear contributions to $(\mathcal{P},{\mathcal{J}}_{\theta },{\mathcal{J}}_z)$ — and the nonlinear remainders it leaves behind have strictly improved decay (OpenAI, 2026a§8.1—8.4, Eq. (8.16) and (8.25)).

☘️ The Gain per Cycle and the Summation

For a Cartesian derivative of order $m$, the paper measures residual gain by an exponent ${\sigma }_j$ in a bound of the form $q^{h{\sigma }_j-K_m}$, with $K_m$ independent of $j$ and logarithmic factors allowed. It starts from ${\sigma }_0=1/5$ and proves

$${\sigma }_{j+1}={\sigma }_j+\frac{1}{10},{\sigma }_j=\frac{1}{5}+\frac{j}{10}⟶\infty .$$

The constant gain $1/10$ is itself the conclusion of a bookkeeping argument in which several competing losses — a factor ${\kappa }_s=10^{-5}$ from each fast radial derivative, a factor from the radial-integral remainders, and the loss in the compatibility defects — are each shown to leave at least $0.1$ of improvement. Because $K_m$ does not depend on the stage, enough cycles beat the loss associated with any fixed finite derivative order.

The stages are then multiplied by shrinking cutoffs at the vector-potential level and summed; taking curls after multiplication keeps each partial sum exactly divergence-free. Each cutoff equals one sufficiently close to $q=0$ and its support shrinks with the stage, so the sum is locally finite for $q>0$, keeps the leading vortex intact, and yields the all-order flatness estimate at $q=0$ (OpenAI, 2026a, sec.3.4, Prop. 9.6 and 9.9).

☘️ Extension of the Residual to a Compact Smooth Force

The final localization is also designed at the vector-potential level. With $u_{\mathrm{loc}}=\mathrm{curl}A+\mathcal{B}{e}_{\theta }$ and a product cutoff $c={\chi }_x{\chi }_t$ equal to one near $(0,1)$, one sets

$$u=\mathrm{curl}(cA)+c\mathcal{B}{e}_{\theta }=c{u}_{\mathrm{loc}}+\nabla c\times A,p=c{p}_{\mathrm{loc}},$$

so both pieces are separately divergence-free and the fields extend smoothly by zero. Setting $f=\mathcal{R}(u,p)$ for $t<1$, the difference $f-c\mathcal{R}(u_{\mathrm{loc}},p_{\mathrm{loc}})$ consists of cutoff commutator terms such as $({\partial }_{t}c-\Delta c)u_{\mathrm{loc}}$, $p_{\mathrm{loc}}\nabla c$, the advection products generated by $\nabla c\times A$, and $(c^2-c)(u_{\mathrm{loc}}\cdot \nabla )u_{\mathrm{loc}}$ — all supported in the transition regions, away from a neighborhood of $(0,1)$.

The endpoint is then controlled in three separate regimes: near the origin by the all-order flatness estimate; on the cutoff transition regions where $q$ stays positive by uniform derivative bounds on the potentials; and in the region where $r$ stays positive while $q\to 0$ by the explicit heat exterior, where $A=0$ and the normalized pressure integral is available in closed form. These give compatible limits uniformly on ${\mathbb{R}}^3$ for every derivative of $f$, which are then realized as Taylor data from $t>1$ followed by a temporal cutoff. The asserted output is

$$f\in C_c^{\infty }({\mathbb{R}}^3\times (0,\infty );{\mathbb{R}}^3),\mathrm{supp}f\subset K\times [0,2].$$

That final compact-force statement is a theorem to audit, not an assumption added after the singular flow is built (OpenAI, 2026a).

🏷️ The C/D Endgame

The local solution already has unbounded velocity, but alternative C concerns nonexistence of a global admissible solution with the same data. The paper uses two separate ingredients.

First, it derives an energy estimate for its localized solution. With $F(t)={\int }_0^t\Vert f(s){\Vert }_{2}ds$,

$$\Vert u(t){\Vert }_2^2+2{\int }_0^t\Vert \nabla u(s){\Vert }_2^{2}ds\le F(t{)}^2(0\le t<1),$$

and $F(1)<\infty$ because the force is smooth with compact support. The derivation is short but uses a standard regularization that is worth recording: pairing the equation with $u$ gives the identity $\frac{1}{2}\frac{d}{dt}\Vert u{\Vert }_2^2+\Vert \nabla u{\Vert }_2^2=⟨f,u⟩$; dividing by $(\Vert u{\Vert }_2^2+{\delta }^2{)}^{1/2}$, discarding the dissipation, applying Cauchy—Schwarz, integrating from the zero datum and letting $\delta \downarrow 0$ gives $\Vert u(t){\Vert }_2\le F(t)$; substituting back into the identity and integrating gives the displayed bound. Monotone convergence then yields finite total dissipation on $[0,1)$ from the estimates for $t<1$ alone (OpenAI, 2026a, p.L emma 10.4).

☘️ The Comparison Argument and the Free Pressure

Second, on every $[0,T]$ with $T<1$, a smooth solution with the same force, zero datum, and bounded $L^2$ norm is shown to equal the constructed one. The comparison has one genuine technical obstacle: the hypotheses impose no spatial growth condition on the competitor’s pressure $P$, so the pressure flux through a large sphere cannot be discarded. For $w=v-u$ and $\pi =P-p$, the common force cancels and

$${\partial }_{t}w+(v\cdot \nabla )w+(w\cdot \nabla )u=\Delta w-\nabla \pi ,\nabla \cdot w=0.$$

The pressure is first identified from the quadratic difference $g_{ij}=v_i{v}_j-u_i{u}_j$ using Riesz transforms,

$${\pi }_{\ast }:=\sum _{i,j}{R}_i{R}_j{g}_{ij},$$

which lies uniformly in $H^{-s}$ for $s>3/2$ since $g\in L_t^{\infty }{L}_x^1$ and the Fourier transform of an $L^1$ function is bounded. Testing the conservative form of the difference equation against a compactly supported time cutoff puts $\int a\nabla \pi dt$ into $H^{-2}+L^2+H^{-3}$, while $\Delta \pi =\Delta {\pi }_{\ast }$ distributionally; the difference is a harmonic tempered distribution whose Fourier transform is a weighted $L^2$ function supported at the origin, hence zero. So $\nabla \pi =\nabla {\pi }_{\ast }$ with the spatial growth of $P$ unrestricted.

The pressure flux is then absorbed rather than dropped. Writing ${\chi }_R={\phi }_R^8$ for a cutoff at radius $R$ and $E_R,A_R,B_R$ for the localized energy, dissipation, and a weighted $L^6$ norm, the commutator decomposition

$${\phi }_R^4{\pi }_{\ast }=\sum _{i,j}{R}_i{R}_j\left({\phi }_R^4{g}_{ij}\right)+\sum _{i,j}[{\phi }_R^4,R_i{R}_j]g_{ij}$$

separates a term controlled by $L^{3/2}$ boundedness of the Riesz transforms from a commutator whose kernel obeys $|K_R(x-y)|\le C|x-y{|}^{-3}\min \{|x-y|/R,1\}$, giving $\Vert K_R{\Vert }_{4/3}^{4/3}\le C{R}^{-1}$ by direct radial integration. The resulting bound

$$\left|\int \pi w\cdot \nabla {\chi }_R\right|\le C{R}^{-1}(B_R+1)[B_R^{1/2}+R^{-3/4}{B}_R^{3/4}]$$

involves only powers of $A_R$ strictly below $2$, so Young’s inequality absorbs them into the dissipation. What survives is

$$\frac{1}{2}{E}_R^{\prime }+\frac{1}{2}{A}_R^2\le \Vert \nabla u{\Vert }_{\infty }{E}_R+\frac{C_T}{R},$$

and Gronwall with $E_R(0)=0$ gives $E_R(t)\le C_T^{\prime }/R\to 0$, hence $w\equiv 0$ on $[0,T]$ (OpenAI, 2026a, p.L emma 10.5).

A hypothetical global smooth bounded-energy solution would therefore agree with $u$ on every interval before $1$. Yet along the explicit path $x_{\tau }=(\sqrt{2X_{\mathrm{in}}\tau },0,0)$ at time $1-\tau$ — where both cutoffs equal one for small $\tau$ — the constructed field satisfies

$$u_{\theta }(x_{\tau },1-\tau )={\tau }^{-A}(e_0+O({\tau }^{2h}))⟶+\infty ,$$

contradicting boundedness of a smooth $v$ on a compact neighborhood of $(0,1)$. Since $H^3({\mathbb{R}}^3)\hookrightarrow L^{\infty }({\mathbb{R}}^3)$, the same asymptotic identifies $[0,1)$ as the maximal classical lifespan in $C([0,T];H^3)$. A small but pleasing corollary: the force must be nonzero, since the energy inequality with $f=0$ would give $u\equiv 0$ (OpenAI, 2026a).

☘️ Viscosity Rescaling and Periodization

Finally, a spatial rescaling transfers the viscosity-one construction to each fixed $\nu >0$:

$$u_{\nu }(x,t)=\sqrt{\nu }u\left(\frac{x}{\sqrt{\nu }},t\right),p_{\nu }(x,t)=\nu p\left(\frac{x}{\sqrt{\nu }},t\right),f_{\nu }(x,t)=\sqrt{\nu }f\left(\frac{x}{\sqrt{\nu }},t\right).$$

The singular time stays $1$, the compact support rescales to $\sqrt{\nu }K$, the force derivatives pick up the harmless factor ${\nu }^{(1-|\alpha |)/2}$, and the energy and dissipation both rescale by ${\nu }^{5/2}$. A competitor for the rescaled data would unscale to a competitor at viscosity one, which has already been excluded. This completes the claimed whole-space C construction.

For D, the paper first uses the parabolic Navier—Stokes scaling to fit the common support strictly inside one fundamental cube $Q_0=(-1/2,1/2{)}^3$. Choosing $\lambda >1$ with ${\lambda }^{-1}{K}_{\nu }⋐Q_0$ and $t_0=1-{\lambda }^{-2}$, it sets

$$\tilde{u}(x,t)=\lambda u(\lambda x,{\lambda }^2(t-t_0)),\tilde{p}(x,t)={\lambda }^{2}p(\lambda x,{\lambda }^2(t-t_0)),\tilde{f}(x,t)={\lambda }^{3}f(\lambda x,{\lambda }^2(t-t_0)),$$

extended by zero for $t<t_0$; every term of the momentum equation scales by ${\lambda }^3$, so the viscosity is unchanged and the original time one is reached at $t=1$. It then sums integer translates. The supports of distinct translates are disjoint with a positive gap, so at every point at most one translate is nonzero and even the nonlinear term has no cross-terms:

$$(U\cdot \nabla )U=\sum _{k\in {\mathbb{Z}}^3}(\tilde{u}(x+k)\cdot \nabla )\tilde{u}(x+k).$$

This explicit disjoint-support argument, not a formal appeal to “periodicity,” produces the periodic velocity, pressure, and force. The pressure is periodic as well, as the erratum to the problem statement requires, and the rescaled growth path ${\tilde{x}}_{\tau }={\lambda }^{-1}\sqrt{\nu }{x}_{\tau }$ at ${\tilde{t}}_{\tau }=1-{\lambda }^{-2}\tau$ stays inside $Q_0$, carrying the same blowup to D (OpenAI, 2026a, p.C or. 10.6).

📊 Numerical Verification

The analytic content of this construction — profile existence, the correction-cycle estimates, the flatness bound — is not reducible to finite computation, and nothing below tests it. What can be checked mechanically is that the algebraic identities the manuscript quotes are mutually consistent, since several of them are compressed enough in the text that a transcription error would be easy to miss. The script navier_stokes_blowup_verification.py (in codes/2026 Fall/) performs seven such checks in SymPy and NumPy, all of which pass.

Internal consistency checks

Cone polynomial. Symbolically, $2(P_c-v{)}^2-(v-2)J_c^2=2(v-v_{-})(v-v_{+})$ with the quoted roots, leading coefficient $2$, $\mathfrak{p}(2)=2(P_c-2{)}^2$, and $\mathfrak{p}(P_c)=-(P_c-2)J_c^2$. Over $2\times 10^5$ random triples $(P_c,J_c,v_s)$ with $v_s>2$, the relaxed form and the direct form of the admissible condition agree in all $200000$ cases ($35005$ inside the cone).

Growth rate. ${\lambda }_0^2=-2F_0{N}_{\theta }(2F_0{N}_{\theta }+|g_0|)$ reduces symbolically to $2a{F}_0^2(1-2/v_s)$, and $c_0^2=(v_s-2)/2$, for $a,F_0>0$. Numerically $\mathrm{sign}{\lambda }_0^2=\mathrm{sign}(v_s-2)$ pointwise, and $2F_0{N}_{\theta }<0$ always, so $c_0<0$ whenever ${\lambda }_0>0$.

Stress frame. With $N=g_0/|g_0|$ and $K=N^{\perp }$, the frame inequalities $T\cdot N<0$, $|c_0(T\cdot K)/(T\cdot N)|<1$ agree with the profile inequalities in all $195155$ admissible samples.

Positive amplitudes. The explicit solve reproduces the target stress to $8\times 10^{-15}$, and $y_{\pm }>0$ holds exactly on $T_N<0$, $|T_K|<(u_{\ast }/A_c)(-T_N)$.

Pulse envelope. $a_{\mathrm{net}}(u_{\ast })=0$; the closed form for $\frac{d}{dv}{a}_{\mathrm{net}}(|s(v)|)$ matches numerical differentiation to $7\times 10^{-13}$ and stays in $[-1.52/L_s,-0.73/L_s]$. Integrating $(\log P{)}^{\prime }=a_{\mathrm{net}}$ gives an envelope peaked at $v=L_s/2$ and squeezed between $e^{-0.651(v-L_s/2{)}^2/L_s}$ and $e^{-0.527(v-L_s/2{)}^2/L_s}$; at $L_s=400$ the endpoint value is $\approx 1.3\times 10^{-23}$, confirming that the temporal cutoff acts only in exponentially small tails.

Auxiliary torus. $J_g$ has spectrum $\{4-\sqrt{2},4+\sqrt{2}\}$ with the stated eigenvectors, ${\rho }_g\approx 0.5625$ and $d_r\approx 1.13>0$ for $h=1/200$, ${\kappa }_s=10^{-5}$. The conjugate identities $(n_1+n_2{)}^2-2n_2^2$ and $(n_2-n_1{)}^2-2n_1^2$ hold symbolically, and over all $0<|n{|}_{\infty }\le 300$ the quantity $(1+|n|)|v\cdot n|$ stays above $0.384$ for both directions.

Exponents. ${\ell }_r^2{\ell }_z={\tau }^{3/2-h}$, $E_{\mathrm{core}}={\tau }^{1/2-3h}$, $D_{\mathrm{core}}={\tau }^{-1/2-3h}$; ${\int }_0^1{\tau }^{-1/2-3h}d\tau$ is finite for $h=1/200$ and diverges at $h=1/6$. The wave relations $A_{\mathrm{wave}}^2=q^{-1-h}$, $A_{\mathrm{wave}}^2/{\ell }_r=q^{-3/2-h}$, and $A_{\mathrm{wave}}/|u^{(0)}|={\ell }_{\mathrm{wave}}/{\ell }_r=q^{h/2}$ hold identically, and ${\sigma }_j=1/5+j/10$.

The checks that carry the most information are the second and third. They show that the threshold $v_s>2$, which appears in §4 as an algebraic inequality on the profiles and in §7 as the positivity of a squared growth rate, is one condition and not two — a coincidence that would be invisible from either section alone.

🏷️ Reading the Claim Correctly

The manuscript’s novelty is the simultaneous control of three requirements that are easy to confuse:

• The velocity must become unbounded, not merely exhibit large but finite norms.

• The energy may remain bounded because the fast flow occupies a rapidly shrinking anisotropic core — but the dissipation must also stay integrable, which is the real content of the constraint on $h$.

• The applied force must remain smooth through the terminal time, so the singular residual must be canceled to all orders rather than merely in leading order.

A fourth requirement is structural rather than quantitative, and it is where most of the 166 pages go: the same quadratic object must serve two masters. The averaged momentum flux $t_r(t_{\theta },t_z)$ that supplies the missing stress is also the quantity whose sign decides whether the pulse producing it can grow at all. The admissible stress cone is precisely the set of targets for which both demands can be met at once, and the profile construction exists to steer the background into that cone at every point of the annulus.

The C/D route is a legitimate route in the Clay problem statement, but it is a forced counterexample route. It leaves the unforced regularity question embodied in alternatives A/B untouched. At the current stage, the responsible reading is conditional: if the profile construction, cone realization, all-order correction bounds, force extension, and comparison argument withstand independent review, then the paper proves C and D. The official paper, its public formalization, and the review process make those interfaces unusually inspectable; they do not remove the need for that inspection (OpenAI, 2026a; OpenAI, 2026b; Clay Mathematics Institute, 2026).

Links

• on the continuity of the percolation phase transition — another report on a recent formal mathematical artifact, useful here for the distinction between a checked formal derivation, a statement audit, and independent mathematical acceptance.

🐻  Clay Mathematics Institute 2026. Navier–Stokes Announcement.

🐻  Daneri, S. & Székelyhidi, L., Jr. 2017. Non-uniqueness and h-Principle for Hölder-Continuous Weak Solutions of the Euler Equations. Archive for Rational Mechanics and Analysis 224, 471–514.

🐻  Fefferman, C.L. 2000. Existence and Smoothness of the Navier–Stokes Equation.

🐻  Friedlander, S. & Vishik, M.M. 1991. Instability Criteria for the Flow of an Inviscid Incompressible Fluid. Physical Review Letters 66(17), 2204–2206.

🐻  Lifschitz, A. & Hameiri, E. 1991. Local Stability Conditions in Fluid Dynamics. Physics of Fluids A: Fluid Dynamics 3(11), 2644–2651.

🐻  OpenAI 2026a. Finite Time Blowup for Navier–Stokes.

🐻  OpenAI 2026b. On the Navier–Stokes Millennium Prize Problem.