Why It's Hard

Most equations in physics are linear — the output is proportional to the input. Double the cause, double the effect. Linear equations are well-understood and (relatively) easy to solve.

The Navier-Stokes equations are nonlinear. The fluid's velocity affects its own rate of change — the fluid pushes itself. It's like trying to predict where a crowd will go when every person's movement depends on what everyone else is doing.

This self-interaction term, $(u \cdot \nabla)u$, is what makes the equations so hard. It creates feedback loops where small disturbances can amplify into large ones, and it's why fluid turbulence is so complex (see subproblems for more).

The convective nonlinearity $(u \cdot \nabla)u$ is the fundamental obstacle. In the vorticity formulation $\omega = \nabla \times u$, the equation becomes

$$\partial_t \omega + (u \cdot \nabla)\omega = (\omega \cdot \nabla)u + \nu \Delta \omega$$

The vortex stretching term $(\omega \cdot \nabla)u$ has no sign — it can amplify vorticity without bound. In 2D, this term vanishes (since $\omega$ is a scalar perpendicular to the flow), which is why 2D global regularity is known (Ladyzhenskaya, 1969). In 3D, vortex stretching is the primary candidate mechanism for finite-time blowup.

Critically, the nonlinearity is quadratic in $u$: the $H^1$ energy estimate gives $\|\nabla u\|_{L^2}$, but controlling $(u \cdot \nabla)u$ in $L^2$ requires $u \in L^\infty$ or at least $u \in L^3$ — information not provided by the energy class.

Here's a key insight: the Navier-Stokes equations have a scaling symmetry. If you zoom into a solution (making everything smaller and faster by the right amounts), you get another valid solution.

This is a problem because the only quantity we can control — the total energy of the fluid — is at the "wrong scale." It tells us about the big picture, but not about what's happening at very small scales where a blowup would form.

Imagine monitoring a city's total electricity usage to detect a single spark. The measurement is real and useful, but it's not fine-grained enough to catch the thing you're worried about. That's the gap researchers need to bridge.

Under the natural scaling $u_\lambda(x,t) = \lambda u(\lambda x, \lambda^2 t)$, the critical Sobolev space is $\dot{H}^{1/2}(\mathbb{R}^3)$ (equivalently $L^3$). A quantity is:

• Subcritical if it grows under the scaling (favoring large scales) — e.g., $\|u\|_{L^2}$
• Critical if scale-invariant — e.g., $\|u\|_{L^3}$, $\|u\|_{\dot{H}^{1/2}}$
• Supercritical if it shrinks under scaling (favoring small scales)

The energy inequality gives control of $u \in L^\infty_t L^2_x \cap L^2_t \dot{H}^1_x$. Both components are subcritical:

$$\|u_\lambda\|_{L^2} = \lambda^{-1/2} \|u\|_{L^2}, \quad \|u_\lambda\|_{L^2_t \dot{H}^1_x} = \lambda^{1/2} \|u\|_{L^2_t \dot{H}^1_x}$$

This means the energy estimate provides no small-scale control — the nonlinearity can in principle overwhelm dissipation at fine scales. Bridging from the subcritical energy class to a critical norm is the central difficulty.

Anyone who has watched a river knows that fluid motion becomes chaotic — turbulent. Big eddies break into smaller ones, which break into even smaller ones, all the way down to microscopic scales where viscosity finally smooths things out.

This energy cascade (described by Kolmogorov in 1941) is beautifully captured by the Navier-Stokes equations. But it also hints at the danger: what if energy concentrates into smaller and smaller regions faster than viscosity can dissipate it? That would be a blowup.

Whether this can actually happen — or whether viscosity always wins in the end — is exactly the open question.

Kolmogorov's K41 theory predicts an energy spectrum $E(k) \sim \varepsilon^{2/3} k^{-5/3}$ in the inertial range $k_f \ll k \ll k_\eta$, where $k_\eta \sim (\varepsilon/\nu^3)^{1/4}$ is the Kolmogorov dissipation wavenumber. The energy flux is constant across scales in this range.

The regularity question asks whether this cascade can degenerate: can the dissipation scale $k_\eta^{-1}$ shrink to zero in finite time? This would require $\|\nabla u\|_{L^2} \to \infty$ (dissipation rate blow up) while the total energy remains finite.

The dissipation anomaly conjecture (Onsager, 1949) suggests that in the vanishing viscosity limit $\nu \to 0$, energy dissipation persists — weak solutions of Euler can dissipate energy. This has been confirmed for Hölder exponents below $1/3$ (Isett, 2018; Buckmaster et al., 2018), but the connection to Navier-Stokes regularity remains unclear.

In the Navier-Stokes equations, pressure plays a strange role. It's not an independent variable — it's determined entirely by the velocity through a constraint (the fluid is incompressible, meaning it can't be squeezed).

This makes pressure nonlocal: a change in velocity at one point instantly affects the pressure everywhere. It's as if every part of the fluid is connected by invisible springs to every other part.

This nonlocality makes the equations much harder to analyze. You can't study what happens at one point without considering the entire fluid at once.

The incompressibility constraint $\nabla \cdot u = 0$ determines the pressure through the Poisson equation

$$-\Delta p = \nabla \cdot ((u \cdot \nabla)u) = \partial_i \partial_j (u_i u_j)$$

so $p = (-\Delta)^{-1} \partial_i \partial_j (u_i u_j)$, involving Riesz transforms — singular integral operators. The pressure is a nonlocal function of velocity, and this nonlocality is the key obstacle to pointwise or local-in-space estimates.

In particular, standard maximum principle arguments fail: even though the viscous term $\nu \Delta u$ is dissipative, the pressure gradient $-\nabla p$ can concentrate energy from distant regions. The Caffarelli-Kohn-Nirenberg theory handles this via local energy inequalities on parabolic cylinders, but extracting pointwise regularity from these remains the hard step.

In two dimensions, the Navier-Stokes problem is solved — smooth solutions exist for all time. (Ladyzhenskaya proved this in 1969.)

So what goes wrong in three dimensions? The key difference is vortex stretching. In 2D, vortices can spin and merge, but they can't stretch. In 3D, fluid can pull vortex tubes thinner and thinner, potentially concentrating all the energy into an infinitely thin filament.

Whether this stretching process can run away to infinity in finite time — or whether viscosity always steps in to stop it — is the million-dollar question.

The dichotomy between 2D and 3D is sharp:

• 2D: Vorticity $\omega$ is a scalar satisfying $\partial_t \omega + u \cdot \nabla \omega = \nu \Delta \omega$. The maximum principle gives $\|\omega(t)\|_{L^\infty} \leq \|\omega_0\|_{L^\infty}$, and BKM implies global regularity. The vortex stretching term $(\omega \cdot \nabla)u$ is identically zero.
• 3D: Vorticity $\omega \in \mathbb{R}^3$ satisfies $\partial_t \omega + (u \cdot \nabla)\omega = (\omega \cdot \nabla)u + \nu \Delta \omega$. The stretching term $(\omega \cdot \nabla)u$ can amplify $|\omega|$ superlinearly (formally $\sim |\omega|^2$ via Biot-Savart), and no maximum principle is available.

The enstrophy $\|\omega\|_{L^2}^2$ satisfies

$$\frac{d}{dt}\|\omega\|_{L^2}^2 \leq C\|\omega\|_{L^2}^2 \|\nabla u\|_{L^\infty} - 2\nu \|\nabla \omega\|_{L^2}^2$$

but controlling $\|\nabla u\|_{L^\infty}$ requires $\omega \in L^\infty$, creating a circular dependency that no existing technique has broken.

These obstacles have led mathematicians to decompose the problem into subproblems and develop specialized approaches for each.

For a decomposition of the regularity problem into tractable components, see Subproblems. For the analytical tools developed to address these obstacles, see Approaches.