Weak, Strong, and Smooth Solutions to the Navier-Stokes Equations

Here's the situation. The Millennium Prize offers a million dollars for resolving whether 3D Navier-Stokes always has smooth solutions that last forever, or whether things can blow up. Smooth means the velocity field is perfectly well-behaved: no sudden jumps, no infinite speeds, no points where the math breaks down. But the best existence result anyone has ever proven, in nearly a century of trying, only guarantees something weaker. These are called weak solutions.

So what's a weak solution? It's not an approximation. It's not "almost right." It's an exact solution to the equations, but one that plays by relaxed rules. A normal ("classical") solution requires the velocity to be smooth enough that you can compute its rate of change at every single point. A weak solution skips that requirement. Instead of checking the equations point by point, you check them "on average" across regions of space.

Here's an analogy. A classical solution is a student who solves every exam problem by showing all their work, step by step. A weak solution is a student who can't show you the intermediate steps, but whose final answers are provably correct for every possible question you could ask. You can't watch them work, but the answers always check out.

Why would you accept that? Because sometimes the equations are too wild for classical solutions. The fluid might develop regions where the velocity changes so sharply that you simply can't compute a rate of change there. The math breaks. Weak solutions let you keep going where classical solutions give up. They're the safety net that keeps the equations alive when things get rough.

The catch: weak solutions might not be unique. You could get multiple weak solutions starting from the exact same flow, and nobody can tell you which one is "the real answer." That's a problem, because physics says the fluid should do one specific thing, not several. And weak solutions might not be smooth. Smoothness is what the Millennium Prize demands, and it's what nobody can prove.

A weak solution to the Navier-Stokes equations replaces the pointwise PDE with a distributional formulation. Consider the incompressible system on $\mathbb{R}^3 \times (0,T)$:

$$\partial_t u + (u \cdot \nabla)u = -\nabla p + \nu \Delta u, \qquad \nabla \cdot u = 0.$$

A divergence-free vector field $u \in L^2_{\mathrm{loc}}(\mathbb{R}^3 \times [0,T))$ is a weak solution if for every smooth, compactly supported, divergence-free test function $\varphi$:

$$\int_0^T \!\int_{\mathbb{R}^3} \bigl[-u \cdot \partial_t \varphi - (u \otimes u) : \nabla \varphi + \nu \, u \cdot \Delta \varphi \bigr] \, dx \, dt = \int_{\mathbb{R}^3} u_0 \cdot \varphi(x,0) \, dx.$$

The pressure vanishes from this formulation because $\varphi$ is divergence-free. All derivatives have been moved off $u$ and onto $\varphi$ via integration by parts. The point: $u$ doesn't need to be classically differentiable. It just needs to be integrable enough for these integrals to converge.

The weak formulation is not an approximation. A classical solution satisfying the PDE pointwise also satisfies the weak formulation for every test function (integrate by parts in the other direction). The converse fails: a weak solution need not be smooth enough to satisfy the PDE pointwise.

This matters because: (1) weak solutions exist globally for $L^2$ initial data (Leray 1934), while classical solutions are only known to exist locally in time for 3D; (2) weak solutions in general are not known to be unique; (3) the question of whether weak solutions are always smooth is equivalent to the Clay Millennium Problem for appropriate initial data.

In 1934, Jean Leray did something that still defines the field. In a single 73-page paper, he proved that weak solutions to the 3D Navier-Stokes equations exist for all time, starting from any reasonable initial flow. Any. As long as the starting velocity isn't infinitely energetic or physically nonsensical, Leray guarantees you'll get a solution that lasts forever. This was the first time anyone proved a global existence result for the 3D equations, and over ninety years later, it's still the strongest unconditional existence theorem we have.

His strategy was clever. The actual equations are too nasty to solve directly because of how the fluid's velocity feeds back into itself (that's the nonlinearity). So Leray blurred the equations slightly, like adding a tiny Gaussian filter to an image. The blurred equations are tame enough to solve. Then he dialed the blur down toward zero and showed that the solutions don't fly apart. They settle into something that satisfies the original, unblurred equations in the weak sense.

But here's what Leray did NOT prove. Uniqueness. His method produces at least one weak solution, but there might be others starting from the same flow. He couldn't rule that out. He also didn't prove smoothness. His solutions have finite energy and satisfy an energy inequality: friction can drain energy away, but energy can't spontaneously appear from nowhere. That's it. Nothing more.

Leray himself suspected that singularities might form. He sketched what one might look like: the fluid collapsing toward a point, faster and faster, concentrating all its energy into a tinier and tinier region, like a whirlpool shrinking to a point at infinite speed. In 1996, Nečas, Růžička, and Šverák proved that this exact self-similar collapse can't happen. Leray's guess about the shape of potential blowup was wrong. Whether blowup happens at all, in any form? Nobody knows.

In 1951, Eberhard Hopf extended Leray's construction to fluids in bounded containers (not just all of infinite space), and the resulting class became known as Leray-Hopf weak solutions: weak solutions that satisfy the energy inequality. This is the standard notion. When researchers say "weak solutions" without further qualification, they almost always mean this.

One more thing. Even within Leray-Hopf weak solutions, there's a pickier subclass called suitable weak solutions. These don't just satisfy the energy inequality globally (total energy doesn't grow). They satisfy it locally too: energy can't secretly pile up in one corner of the fluid while draining from another. Caffarelli, Kohn, and Nirenberg (CKN) proved their famous partial regularity result in 1982 specifically for this smaller class. Don't confuse the two: CKN applies to suitable weak solutions, not to all Leray-Hopf solutions.

Leray's 1934 paper established the following: for any divergence-free $u_0 \in L^2(\mathbb{R}^3)$, there exists at least one weak solution $u$ to the Navier-Stokes equations on $\mathbb{R}^3 \times (0,\infty)$ satisfying:

• $u \in L^\infty(0,\infty; L^2(\mathbb{R}^3)) \cap L^2(0,\infty; \dot{H}^1(\mathbb{R}^3))$
• The energy inequality: $\|u(t)\|_{L^2}^2 + 2\nu \int_0^t \|\nabla u(s)\|_{L^2}^2 \, ds \leq \|u_0\|_{L^2}^2$ for a.e. $t > 0$

The construction proceeds by mollification. Replace the nonlinearity $(u \cdot \nabla)u$ with $(u_\varepsilon \cdot \nabla)u$ where $u_\varepsilon = J_\varepsilon * u$ is a spatial mollification. The regularized system has global smooth solutions (the mollification kills the worst of the nonlinear interactions). Leray obtained uniform energy bounds for the regularized solutions, then extracted a weakly convergent subsequence. The limit satisfies the weak formulation and the energy inequality.

Leray did not establish uniqueness. The compactness argument gives existence of at least one accumulation point; different subsequences might converge to different limits. Uniqueness of Leray-Hopf weak solutions in 3D remains open to this day.

Hopf (1951) adapted the construction to bounded domains $\Omega \subset \mathbb{R}^3$ with Dirichlet boundary conditions, using Galerkin approximation (projection onto finite-dimensional subspaces) rather than mollification. The resulting class, weak solutions satisfying the energy inequality, carries both names: Leray-Hopf weak solutions.

The Caffarelli-Kohn-Nirenberg theorem (1982) concerns a more restrictive class: suitable weak solutions, which additionally satisfy a local energy inequality of the form

$$\partial_t \left(\frac{|u|^2}{2}\right) + \nabla \cdot \left(\left(\frac{|u|^2}{2} + p\right)u\right) + \nu |\nabla u|^2 \leq \nu \Delta \left(\frac{|u|^2}{2}\right)$$

in the sense of distributions. CKN proved that for any suitable weak solution, the one-dimensional parabolic Hausdorff measure of the singular set in spacetime is zero. This means singularities, if they exist, are extremely sparse (they have zero one-dimensional parabolic Hausdorff measure). But the theorem says nothing about whether singularities actually occur, and it applies only to suitable weak solutions, not to all Leray-Hopf solutions.

Leray himself considered the possibility of self-similar blowup of the form $u(x,t) = (T-t)^{-1/2} U(x / (T-t)^{1/2})$. Nečas, Růžička, and Šverák (1996) proved that no such self-similar blowup exists for solutions in $L^3(\mathbb{R}^3)$, and Tsai (1998) ruled out certain asymptotically self-similar blowup scenarios under corresponding hypotheses. The shape of potential singularities, if any, remains unknown.

Weak solutions exist globally. But they might not be unique, and they might not be smooth. Can we do better?

Yes, but only temporarily. Strong solutions are the upgrade: solutions where the equations hold exactly at every point, not just "on average." For smooth initial data in 3D, strong solutions exist for a short time. How short? That depends on how wild the starting flow is. Calm, gentle flows get longer guarantees. Violent, turbulent starting conditions? Microseconds.

And nobody can prove that these strong solutions don't eventually blow up.

In 1962, James Serrin proved something like a promotion rule. It goes like this: if a weak solution happens to stay well-behaved enough (not too large, not concentrating its energy into smaller and smaller regions), then it was secretly smooth the whole time. You can promote it. And by a principle called weak-strong uniqueness, it's also the only weak solution with those starting conditions. One solution, smooth and unique, case closed. But if you can't verify that the solution stays tame? Nothing. You're stuck.

This is a conditional result. IF the solution isn't too wild, THEN it's perfectly well-behaved. The entire difficulty is proving the IF.

In two dimensions, the energy estimates are strong enough that every weak solution automatically passes Serrin's test. Done. That's why 2D is solved. In 3D, the estimates fall just barely short of what you'd need, and closing that gap is the whole game.

Researchers have found other conditional tests too, each one a different angle of attack: "Prove this one specific thing about the solution, and I'll give you smoothness for free." Proving any single one of them unconditionally would solve the Millennium Problem. Nobody has managed it. For a survey of the different proof strategies people have tried, there's a whole page on that.

A strong solution to the Navier-Stokes equations is one with enough regularity that the PDE holds pointwise (a.e.) and the nonlinear term $(u \cdot \nabla)u$ is well-defined as a function rather than merely a distribution. Typically this means $u \in L^\infty(0,T; H^1(\mathbb{R}^3)) \cap L^2(0,T; H^2(\mathbb{R}^3))$. A related but distinct framework is Fujita-Kato (1964), which constructs local mild solutions in critical spaces.

Local existence of strong solutions for sufficiently regular Sobolev data is well established. In critical spaces, the Fujita-Kato (1964) framework constructs local mild solutions via a fixed-point argument:

$$u(t) = e^{\nu t \Delta} u_0 - \int_0^t e^{\nu(t-s)\Delta} \mathbb{P} \nabla \cdot (u \otimes u)(s) \, ds,$$

where $\mathbb{P}$ is the Leray projection onto divergence-free fields. This integral equation has a unique local solution by the contraction mapping principle in suitable function spaces. For small data in critical spaces ($L^3$, $\dot{H}^{1/2}$, $BMO^{-1}$), the solution is global.

The question is whether large-data strong solutions persist for all time. The key conditional result is Serrin's (1962): if a Leray-Hopf weak solution satisfies $u \in L^p(0,T; L^q(\mathbb{R}^3))$ with

$$\frac{2}{p} + \frac{3}{q} \leq 1, \qquad q > 3,$$

then $u$ is smooth on $(0,T] \times \mathbb{R}^3$ and is the unique Leray-Hopf weak solution with the given initial data. These are called the Prodi-Serrin conditions (Prodi 1959 established a related result).

The endpoint $q = 3$ ($p = \infty$) was resolved by Escauriaza, Seregin, and Šverák (2003): if $u \in L^\infty(0,T; L^3(\mathbb{R}^3))$, then $u$ doesn't blow up at time $T$. This is the endpoint of the Prodi-Serrin scale and one of the sharpest known continuation criteria.

In two dimensions, the energy bound $u \in L^\infty_t L^2_x \cap L^2_t H^1_x$ combined with the Ladyzhenskaya inequality $\|f\|_{L^4}^2 \leq C\|f\|_{L^2}\|\nabla f\|_{L^2}$ (specific to 2D) gives $u \in L^4_t L^4_x$, which satisfies the Serrin condition $2/4 + 2/4 = 1$ (in the 2D version with $2/p + 2/q \leq 1$). In 3D, the energy bound gives $u \in L^{10/3}_t L^{10/3}_x$ by Sobolev embedding, which satisfies $2/(10/3) + 3/(10/3) = 3/2 > 1$. The Serrin condition fails by exactly the margin corresponding to the supercriticality gap. See Why Navier-Stokes Is Hard for more on this structural obstruction.

Smooth solutions are the gold standard. The velocity field is perfectly well-behaved everywhere, for all time. No sudden jumps. No infinite speeds. Zoom in as far as you want, and the solution just keeps being nice.

The Clay Millennium Prize Problem, formulated by Charles Fefferman in 2000, asks a question that fits on an index card. Start with any smooth, physically reasonable velocity field filling three-dimensional space. Does the Navier-Stokes equation always produce a smooth solution that lasts forever, or can you find a starting flow where the solution eventually blows up?

Either answer is worth a million dollars.

Here's where we stand. Nobody has proven that smooth solutions always exist globally in 3D, and nobody has constructed a blowup either. We've been stuck in between since Leray's 1934 paper, over ninety years of one of the hardest open questions in all of mathematics, and we still don't know which side the answer falls on.

Short-term? Fine. For smooth starting data, the equations do produce a smooth solution for some stretch of time. The fluid starts moving, the math works, everything is clean. But what happens later? Does the solution stay smooth forever, or does it hit a point where the velocity rockets off to infinity?

If it does stay smooth, something nice happens. That smooth solution automatically satisfies the relaxed rules for weak solutions too, so it's a weak solution. And by weak-strong uniqueness, no other weak solution with those starting conditions can exist. So if someone proved global smoothness, the entire hierarchy would collapse: weak, strong, and smooth would all turn out to be the same thing, a single unique solution that's perfectly well-behaved for all time. That's what makes this problem so appealing and so hard. The gap between what we can prove exists (weak solutions) and what we want (smooth solutions) is exactly the content of the million-dollar question.

The Clay Millennium Problem (Fefferman 2000) asks: for any divergence-free $u_0 \in C^\infty(\mathbb{R}^3)$ satisfying $|\partial^\alpha u_0(x)| \leq C_{\alpha K}(1 + |x|)^{-K}$ for all $\alpha, K$, does there exist $u \in C^\infty(\mathbb{R}^3 \times [0,\infty))$ and $p \in C^\infty(\mathbb{R}^3 \times [0,\infty))$ satisfying the Navier-Stokes equations with $\int_{\mathbb{R}^3} |u(x,t)|^2 \, dx < C$ for all $t \geq 0$?

The alternative (also prize-worthy): find $u_0$ in the above class such that no such smooth solution exists.

What's known:

• Local existence: For $u_0 \in H^s(\mathbb{R}^3)$, $s \geq 1/2$, a unique local mild/strong solution exists on $[0,T^*)$ for some $T^* > 0$ depending on $\|u_0\|_{H^s}$, and it is smooth for every positive time $t > 0$. If $T^* < \infty$, then $\|u(t)\|_{H^s} \to \infty$ as $t \to T^*$.
• Small data global existence: If $\|u_0\|_{\dot{H}^{1/2}}$ (or $\|u_0\|_{L^3}$, or $\|u_0\|_{BMO^{-1}}$) is smaller than a universal constant, the solution is global and smooth. Key references: $\dot{H}^{1/2}$ (Fujita-Kato 1964), $BMO^{-1}$ (Koch-Tataru 2001), with analogous results in $L^3$ (Kato 1984).
• Weak-strong uniqueness: If a strong solution exists on $[0,T]$, then every Leray-Hopf weak solution with the same initial data coincides with it on $[0,T]$. This was established by Serrin (1962) and refined by subsequent work. It means proving regularity also settles uniqueness within the Leray-Hopf class.

The hierarchy collapses upward: smooth $\Rightarrow$ strong $\Rightarrow$ weak, and weak-strong uniqueness means a smooth solution, if it exists, is the unique Leray-Hopf weak solution. So the Millennium Problem is equivalent to asking: are all Leray-Hopf weak solutions with smooth initial data themselves smooth? The gap between "weak solution exists" (Leray 1934) and "smooth solution exists" (open) is exactly the prize question.

If weak solutions exist and describe the fluid, why should anyone care about smoothness?

Three reasons.

First, uniqueness. Physics demands one answer. Give me the initial state of a fluid, and I should be able to tell you exactly what it does next. Not "here are several possibilities, pick whichever you like." But weak solutions don't guarantee that. Multiple weak solutions might emerge from the same starting flow with no way to tell which one the real fluid follows. The equations would become a menu instead of a recipe. That's not physics.

Second, numerical reliability. Many important fluid simulations are based on Navier-Stokes or closely related models: weather forecasts, aerodynamics, blood flow through arteries, and more. Make the grid finer and the simulation should converge toward the true answer. Without a smoothness-and-uniqueness guarantee? No theorem says that actually happens in every 3D scenario. The simulations work. We can't fully explain why.

Third, extreme physics. If singularities can form, that's nature sending us a message. The Navier-Stokes equations, our best model of fluid motion, would have a built-in expiration date: at some extreme scale the model itself stops working, and the equations are telling us, "You need new physics."

This isn't a technicality. It's the fault line running through everything. Existence on one side (Leray, 1934, done). Smoothness on the other (open, one million dollars). Why is crossing so hard? The energy estimates in 3D fall just barely short of what's needed, and ninety years of effort, hundreds of papers, entire careers spent trying, nobody has closed that gap.

Every proof strategy being pursued right now is an attempt to bridge this divide. Prove weak solutions are smooth. Or prove they aren't. Two words or three.

The solution hierarchy for the 3D incompressible Navier-Stokes equations is:

$$\text{smooth} \subset \text{strong} \subset \text{Leray-Hopf weak} \subset \text{distributional weak}$$

What's known at each level:

The Millennium Problem sits at the gap between global Leray-Hopf weak existence and global smooth regularity. The core difficulty: the energy inequality provides $u \in L^\infty_t L^2_x \cap L^2_t \dot{H}^1_x$, which is a half-derivative below the critical scaling $\dot{H}^{1/2}$. Bridging this supercriticality gap is equivalent to proving global regularity.

Weak-strong uniqueness implies that the hierarchy collapses if smooth solutions exist globally. More precisely: if for given $u_0 \in C^\infty$ a smooth solution $u$ exists on $[0,T]$, then every Leray-Hopf weak solution with the same data equals $u$ on $[0,T]$. So global smoothness $\Rightarrow$ global uniqueness within the Leray-Hopf class.

Conversely, non-uniqueness of Leray-Hopf weak solutions for smooth admissible initial data would imply that global smooth solutions cannot persist for that data class (since a smooth solution would force uniqueness). Recent work on convex integration (building on De Lellis-Székelyhidi for Euler, extended by Buckmaster-Vicol 2019 to construct non-unique weak solutions of Navier-Stokes below the Leray-Hopf regularity) shows that distributional weak solutions can be highly non-unique. Whether this non-uniqueness extends to the Leray-Hopf class is a major open question with direct implications for the Millennium Problem.

For the current state of proof strategies attacking this gap, and for the structural reasons it's so resistant, see Why Navier-Stokes Is Hard.