Navier–Stokes Existence or Breakdown

This is an AI-generated survey distilled from a talk by Javier Gómez-Serrano, lightly edited but not independently checked line by line. Treat it as a dense orientation map, not as a reference text.

Problem

The Navier–Stokes existence and smoothness problem asks whether smooth divergence-free data for 3D incompressible Navier–Stokes on ${ℝ}^3$ or ${𝕋}^3$ generate smooth solutions for all time, or whether some smooth datum develops a finite-time singularity. It is one of the Clay Millennium Prize Problems.

For velocity $𝑢(𝑥,𝑡)$, pressure $𝑝(𝑥,𝑡)$, density $𝜌$, viscosity $𝜈$, and force $𝑓$,

With $𝜈=0$ this is incompressible Euler; with $𝜈>0$ it is incompressible Navier–Stokes. The nonlinear term $(𝑢\cdot \nabla )𝑢$ lets the flow advect itself. Incompressibility removes simple compression blowup. Pressure is nonlocal: taking divergence gives

and hence, in ${ℝ}^3$,

Thus $𝑝$ is determined instantaneously by $𝑢$ through a singular integral.

The Clay alternatives are:

• global regularity: every smooth divergence-free ${𝑢}_0$ yields a smooth solution for all time;
• finite-time breakdown: some smooth divergence-free ${𝑢}_0$ yields blowup of $𝑢$ or its derivatives in finite time.

One can also separate three logical possibilities: global smooth uniqueness; singularity with uniqueness; singularity with nonuniqueness. Ladyzhenskaya’s emphasis was often uniqueness versus nonuniqueness.

Historical scale

Da Vinci already described eddies nested across scales and used turbolenza. Bernoulli’s Hydrodynamica (1738) linked velocity and pressure; d’Alembert’s inviscid drag calculation led to the 1752 d’Alembert paradox Euler wrote the inviscid incompressible equations around 1757; Prandtl’s 1904 boundary layer theory explained how small viscosity can still produce drag near boundaries. The modern problem sits on this tension: inviscid structure, viscous smoothing, boundary effects, and multiscale transfer.

Weak solutions

For unforced Navier–Stokes,

a weak solution satisfies the equation after testing against smooth compactly supported divergence-free fields $𝜑$:

A strong solution is weak; not conversely.

A Leray–Hopf weak solution also satisfies the energy inequality

Leray (1934) proved global existence for divergence-free ${𝑢}_0\in {𝐿}^2({ℝ}^3)$:

with the energy inequality and smoothness away from a small exceptional set; the set of singular times has zero $\frac{1}{2}$-dimensional Hausdorff measure.¹ Hopf (1951) extended the theory to bounded domains with no-slip boundary conditions.² Global weak solutions exist; smoothness and uniqueness in 3D remain open.

Two dimensions

Ladyzhenskaya proved global well-posedness for 2D incompressible Navier–Stokes: for divergence-free ${𝑢}_0\in {𝐿}^2({ℝ}^2)$ there is a unique global solution

smooth for $𝑡>0$ and continuous in the data.³

The 2D estimate

closes the energy method. In 3D the analogous

puts too much weight on $\nabla 𝑢$; the same proof fails.

Partial regularity

A suitable weak solution satisfies the local energy inequality

in distributions. Scheffer (1977) proved the singular set has parabolic Hausdorff dimension at most $\frac{5}{3}$ (later improved to at most $2$).⁴ Caffarelli–Kohn–Nirenberg (1982) proved its one-dimensional parabolic Hausdorff measure is zero.⁵ Thus singularities, if present, cannot contain a spacetime curve of positive parabolic length.

Vorticity

Let

Using

and

with $\nabla \cdot 𝑢=0$ and $\nabla \cdot 𝜔=0$, the curl equation is

The term $(𝜔\cdot \nabla )𝑢$ is vortex stretching.

The Biot–Savart law recovers velocity from vorticity:

Hence $\nabla 𝑢$ is a Calderón–Zygmund singular integral of $𝜔$, and stretching has the schematic form $𝜔\cdot 𝑇(𝜔)$: quadratic, nonlocal, and sign-indefinite.

In 2D, $𝑢=({𝑢}_1,{𝑢}_2,0)$ is independent of ${𝑥}_3$ and

Then $(𝜔\cdot \nabla )𝑢={𝜔}_3{𝜕}_3𝑢=0$, so

2D vorticity is transported and diffused. In 3D, stretching is active. Writing $𝛼=\frac{𝜔}{|𝜔|}$,

The first right-hand term is strain-driven amplification; the second records geometric variation of vorticity direction.

Scaling and energy

Navier–Stokes is invariant under

Then

Energy is supercritical: it becomes weaker under zooming. Critical norms include ${\dot{𝐻}}^{\frac{1}{2}}$, ${𝐿}^3$, and ${\text{BMO}}^{-1}$. The energy inequality controls $𝑢\in {𝐿}_{𝑡}^{\infty }{𝐿}_{𝑥}^2$ and $\nabla 𝑢\in {𝐿}_{𝑡}^2{𝐿}_{𝑥}^2$, but not the critical quantities needed to rule out concentration. Energy methods alone are therefore structurally insufficient; model systems with the same soft energy properties can blow up.

Classical local theory and criteria

Fujita–Kato local well-posedness gives, for ${𝑢}_0\in {\dot{𝐻}}^{\frac{1}{2}}({ℝ}^3)$, a unique mild solution on $\left[0,𝑇\right)$,

with Duhamel formula

where $\mathbb{P}$ is the Leray projection. For ${𝑢}_0\in {𝐻}^{𝑠}$, $𝑠>\frac{1}{2}$, one has $𝑢\in 𝐶(\left[0,𝑇\right);{𝐻}^{𝑠})\cap {𝐿}^2((0,𝑇);{𝐻}^{𝑠+1})$. If ${𝑇}_{max}<\infty$, then

equivalently by Beale–Kato–Majda type criteria,

The Prodi–Serrin–Ladyzhenskaya criterion: a Leray–Hopf solution is smooth if

For $𝑞<6$ one combines

with Gronwall estimates. Escauriaza–Seregin–Šverák (2003) proved the endpoint $𝑢\in {𝐿}_{𝑡}^{\infty }{𝐿}_{𝑥}^3$ implies smoothness, using backward uniqueness for vorticity.⁶

Beale–Kato–Majda for Euler says blowup at $𝑇$ forces

⁷

Constantin–Fefferman adds geometry: for $𝜉=\frac{𝜔}{|𝜔|}$, if

then no singularity occurs by time $𝑇$.⁸ Blowup requires not only large vorticity but sufficiently violent directional oscillation.

Critical spaces

Koch–Tataru (2001) proved global well-posedness for small data in ${\text{BMO}}^{-1}$: if ${\Vert {𝑢}_0\Vert }_{{\text{BMO}}^{-1}}<𝜀$, then there is a unique global mild solution

solving

where

Moreover ${𝑡}^{\frac{1}{2}}{\Vert 𝑢(\cdot ,𝑡)\Vert }_{{𝐿}^{\infty }}\le 𝐶{\Vert {𝑢}_0\Vert }_{{\text{BMO}}^{-1}}$, and $𝑢$ is smooth for $𝑡>0$.⁹

Bourgain–Pavlović (2008) showed ill-posedness in the larger critical Besov space ${\dot{𝐵}}_{\infty }^{-1,\infty }$: for every $𝛿,𝜀>0$, smooth ${𝑢}_0$ can satisfy ${\Vert {𝑢}_0\Vert }_{{\dot{𝐵}}_{\infty }^{-1,\infty }}<𝛿$ while ${\Vert 𝑢(𝜀)\Vert }_{{\dot{𝐵}}_{\infty }^{-1,\infty }}>\frac{1}{𝛿}$.¹⁰ This norm inflation comes from high-frequency data near $𝑁$ whose quadratic interaction transfers energy to frequencies near ${𝑁}^2$ before viscosity dominates. Thus ${\text{BMO}}^{-1}$ is essentially the largest critical well-posedness space.

Germain–Pavlović–Staffilani (2007) proved that Koch–Tataru solutions are real analytic in space for $𝑡>0$ and satisfy

for every multi-index $𝛼$.¹¹ The proof expands the mild solution as a convergent power series in ${𝑢}_0$ using the heat semigroup. They also proved any self-similar solution in ${\text{BMO}}^{-1}$ is smooth, complementing Nečas–Růžička–Šverák.

Numerical and computer-assisted singularity search

Kerr (1993) simulated perturbed anti-parallel vortex tubes and saw rapid maximum-vorticity growth consistent with Euler blowup; the BKM integral appeared divergent.¹² Hou–Li later used adaptive high-resolution computation and found depletion rather than blowup; Kerr’s growth was an underresolution artifact near the vortex core, saturating later with algebraic scaling.¹³

Other numerical scenarios include folded vortex sheets, interacting vortex rings, multiscale structures, axisymmetric Euler with swirl (Grauer–Sideris; Cichocki), and tornado-type Navier–Stokes boundary structures. The obstacles are resolution, truncation error, artificial viscosity, and the fact that numerics alone do not prove blowup.

Computer-assisted proof strategy: compute an approximate solution $\overline{𝑢}$; write the PDE as $𝐹(𝑢)=0$ in a Banach space; verify Newton–Kantorovich hypotheses by bounding $𝐹(\overline{𝑢})$ and ${𝐹}^{\prime }{(\overline{𝑢})}^{-1}$; use interval arithmetic so floating-point operations enclose exact values. This turns a numerical candidate into a theorem when the analytic estimates close.

Tao’s model warnings

Hyperdissipative Navier–Stokes replaces $𝜈\mathrm{\Delta }$ by $𝜈{(-\mathrm{\Delta })}^{𝛼}$. In 3D the critical threshold is $𝛼=\frac{5}{4}$: $𝛼>\frac{5}{4}$ is subcritical and globally regular by standard methods; $𝛼=1$ is classical Navier–Stokes and supercritical. Tao (2009) proved global regularity at a logarithmically supercritical borderline, using a Fourier multiplier with symbol

just stronger than critical dissipation.¹⁴

Tao also built averaged Navier–Stokes-type systems preserving energy identity, Sobolev estimates, symmetries, and scaling, yet blowing up in finite time. The nonlinearity is averaged/truncated, retaining soft features while removing cancellations; the blowup uses a self-similar ansatz and an Ornstein–Uhlenbeck-type stochastic construction. Moral: any proof for classical Navier–Stokes must use the exact nonlinearity, not only soft estimates.

Onsager and convex integration

Onsager’s conjecture for 3D Euler: weak solutions with Hölder/Besov regularity above $\frac{1}{3}$ conserve energy; below $\frac{1}{3}$ anomalous dissipation can occur. Constantin–E–Titi proved conservation if

¹⁵

De Lellis–Székelyhidi imported convex integration into fluids from Nash–Gromov geometry. Subsequent work reached Isett’s theorem: for every $𝛼<\frac{1}{3}$, there exists a nonzero compactly time-supported weak Euler solution $𝑢\in {𝐶}_{𝑡}{𝐶}_{𝑥}^{𝛼}$, hence with nonconserved energy.¹⁶ The proof uses convex integration, gluing approximation, and Mikado flows (Daneri–Székelyhidi). Buckmaster–De Lellis–Székelyhidi–Vicol further prescribed arbitrary nonnegative energy profiles $𝑒(𝑡)$.

Navier–Stokes nonuniqueness

Buckmaster–Vicol (2019) proved nonuniqueness for rough weak 3D Navier–Stokes solutions: two distinct global weak solutions

can share the same finite-energy initial data ${𝑢}_0\in {𝐿}^2$. They may be ${𝐶}^{𝛼}$ for $𝛼<\frac{1}{3}$ and satisfy energy equality.¹⁷ The construction uses convex integration with intermittent Beltrami flows: periodic, divergence-free, almost Beltrami fields satisfying $\nabla \times 𝑣\approx 𝜆𝑣$, frequency-separated, spatially intermittent, and equipped with a third scale to handle diffusion and nonlinear errors. The solutions are too rough to be Leray–Hopf because they do not satisfy the energy inequality.

Jia–Šverák proposed Leray–Hopf nonuniqueness via self-similar scale-invariant solutions: if the linearized operator around such a profile has an unstable eigenvalue, one constructs another solution on the unstable manifold.¹⁸

Albritton–Brué–Colombo (2022) realized this for forced Leray–Hopf solutions: two distinct Leray–Hopf weak solutions with ${𝑢}_0=0$ and the same smooth force $𝑓$, built around an unstable self-similar compactly supported vortex ring in similarity variables.¹⁹

Hou–Wang–Yang announced in 2025 a computer-assisted proof of unforced Leray–Hopf nonuniqueness: a self-similar Leray–Hopf solution plus rigorous unstable eigenpair certification for the linearized operator, using high-precision computation and a decomposition into a coercive part plus compact finite-rank perturbation. The announced conclusion is infinitely many Leray–Hopf solutions for the same smooth compactly supported initial data and zero force.²⁰

Euler blowup and boundaries

Luo–Hou (2014) numerically studied axisymmetric 3D Euler in a cylinder with no-flow solid wall and axial periodicity. A hybrid sixth-order Galerkin/finite-difference adaptive method saw a $3\times {10}^8$ increase in maximum vorticity and predicted ${𝑡}_{𝑠}\approx 0.0035056$, checked against BKM, Constantin–Fefferman, and Deng–Hou–Yu criteria.²¹ The boundary creates a shear layer driving amplification.

Chen–Hou proved finite-time blowup for 3D incompressible Euler in a cylindrical domain with no-penetration boundary: smooth finite-energy initial data develop a singularity while the velocity remains ${𝐶}^{1,𝛼}$ and finite-energy up to blowup.²² The proof is computer-assisted: approximate self-similar profile; spectral stability of the rescaled linearized equation; nonlinear stability via fixed point; interval arithmetic. This does not solve Clay: it is Euler, bounded-domain, and boundary-driven.

Elgindi (2021) proved finite-time blowup for 3D Euler in ${ℝ}^3$ from rough data: axisymmetric no-swirl ${𝑢}_0\in {𝐶}^{1,𝛼}$ produces blowup with $𝜔(𝑡)\approx {(𝑇-𝑡)}^{-1}$ and $\nabla 𝑢(𝑡)\approx {(𝑇-𝑡)}^{-1}$.²³ The proof uses dynamic rescaling and convergence to a stable nontrivial stationary profile. Elgindi–Ghoul–Masmoudi proved stability under small ${𝐶}^{1,𝛼}$ perturbations within axisymmetric no-swirl data, using spectral analysis in similarity variables and weighted nonlinear estimates.²⁴ Huang, Chen, Hou, and collaborators proved related blowup results for models including 2D Boussinesq and axisymmetric Euler with boundary. The open upgrade is smooth-data ${𝐶}^{\infty }$ Euler blowup in the whole space.

Dimension, neural search, self-similarity

Hou’s generalized axisymmetric Navier–Stokes numerics analytically continue dimension $𝑑$ by replacing the Biot–Savart kernel ${|𝑥-𝑦|}^{-(𝑑-1)}$ with a $𝑑$-dependent kernel. Self-similar singularities appear numerically above ${𝑑}_{𝑐}\approx 3.188$, while $𝑑<{𝑑}_{𝑐}$ appears regular; 3D may lie just below a critical dimension, with nonlinear depletion/cancellation separating regularity from blowup.²⁵

Neural networks and PINNs use nonlinear parametrizations and PDE-residual losses to discover candidate self-similar profiles. Applications mentioned include axisymmetric 3D Euler models such as De Gregorio-type 1D models, the Constantin–Cordoba–Fontelos equation, incompressible porous media (IPM), and analogues related to SQG. These methods find profiles; proof still requires Newton–Kantorovich/interval-arithmetic verification or other rigorous analysis.

For a putative singularity at $𝑇$, a general ansatz is

Balancing ${𝜕}_{𝑡}𝑢$ with $𝜈\mathrm{\Delta }𝑢$ gives $2𝛽+1=𝛾$. With

the rescaled Navier–Stokes equation is

At critical scaling this becomes autonomous:

Self-similar blowup becomes convergence to a steady profile in similarity variables. A proof program: find approximate ${𝑈}_{\text{app}}$; linearize ${𝜕}_{𝜏}𝑉=𝐿𝑉+𝑁(𝑉)$; prove one unstable eigenvalue and stable complement; close nonlinear estimates in weighted spaces; verify constants by interval arithmetic; return to physical variables.

Restrictions are severe. Nečas–Růžička–Šverák (1996) ruled out nontrivial Leray self-similar Navier–Stokes blowup

with divergence-free $𝑈\in {𝐿}^3({ℝ}^3)$; then $𝑈\equiv 0$.²⁶ Thus the naive finite-energy self-similar scenario is excluded. Constantin–Ignatova–Vicol proved restrictions on Euler self-similar exponents: finite-energy data require $𝛾\ge \frac{2}{5}$ generally and $𝛾\ge \frac{1}{2}$ in axisymmetry.²⁷

Status

Known:

• global Leray–Hopf weak solutions in 3D;
• complete global well-posedness in 2D;
• partial regularity of suitable weak solutions;
• conditional regularity criteria: Prodi–Serrin, endpoint ${𝐿}_{𝑡}^{\infty }{𝐿}_{𝑥}^3$, BKM, Constantin–Fefferman;
• small-data critical well-posedness in ${\text{BMO}}^{-1}$ and ill-posedness beyond it;
• nonuniqueness for very rough Navier–Stokes weak solutions;
• nonuniqueness for forced Leray–Hopf solutions;
• announced unforced Leray–Hopf nonuniqueness;
• finite-time Euler blowup in a cylinder;
• finite-time Euler blowup in ${ℝ}^3$ from ${𝐶}^{1,𝛼}$ data;
• numerical, computer-assisted, and AI-assisted tools for candidate discovery.

Unknown:

• smooth 3D Navier–Stokes global regularity versus finite-time blowup;
• smooth-data 3D Euler blowup in the whole space;
• classical unforced Leray–Hopf uniqueness, pending confirmation of announced results;
• the exact cancellation/depletion mechanism, if any, separating 3D from nearby blowup models.

Formula sheet

Navier–Stokes on ${ℝ}^3$ or ${𝕋}^3$:

Euler:

Vorticity:

3D Navier–Stokes vorticity:

2D Navier–Stokes vorticity:

Leray–Hopf energy inequality:

Prodi–Serrin:

BKM:

is necessary for blowup at $𝑇$.